1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopInfo.h"
62 #include "llvm/Analysis/LoopIterator.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolution.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
67 #include "llvm/Analysis/TargetTransformInfo.h"
68 #include "llvm/Analysis/ValueTracking.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/DiagnosticInfo.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/Function.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Module.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Value.h"
84 #include "llvm/IR/ValueHandle.h"
85 #include "llvm/IR/Verifier.h"
86 #include "llvm/Pass.h"
87 #include "llvm/Support/BranchProbability.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Debug.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Transforms/Scalar.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm::PatternMatch;
102 #define LV_NAME "loop-vectorize"
103 #define DEBUG_TYPE LV_NAME
105 STATISTIC(LoopsVectorized, "Number of loops vectorized");
106 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
108 static cl::opt<unsigned>
109 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
110 cl::desc("Sets the SIMD width. Zero is autoselect."));
112 static cl::opt<unsigned>
113 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
114 cl::desc("Sets the vectorization interleave count. "
115 "Zero is autoselect."));
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
129 /// This enables versioning on the strides of symbolically striding memory
130 /// accesses in code like the following.
131 /// for (i = 0; i < N; ++i)
132 /// A[i * Stride1] += B[i * Stride2] ...
134 /// Will be roughly translated to
135 /// if (Stride1 == 1 && Stride2 == 1) {
136 /// for (i = 0; i < N; i+=4)
140 static cl::opt<bool> EnableMemAccessVersioning(
141 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
142 cl::desc("Enable symblic stride memory access versioning"));
144 /// We don't unroll loops with a known constant trip count below this number.
145 static const unsigned TinyTripCountUnrollThreshold = 128;
147 /// When performing memory disambiguation checks at runtime do not make more
148 /// than this number of comparisons.
149 static const unsigned RuntimeMemoryCheckThreshold = 8;
151 /// Maximum simd width.
152 static const unsigned MaxVectorWidth = 64;
154 static cl::opt<unsigned> ForceTargetNumScalarRegs(
155 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's number of scalar registers."));
158 static cl::opt<unsigned> ForceTargetNumVectorRegs(
159 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of vector registers."));
162 /// Maximum vectorization interleave count.
163 static const unsigned MaxInterleaveFactor = 16;
165 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
166 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's max interleave factor for "
170 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
171 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
172 cl::desc("A flag that overrides the target's max interleave factor for "
173 "vectorized loops."));
175 static cl::opt<unsigned> ForceTargetInstructionCost(
176 "force-target-instruction-cost", cl::init(0), cl::Hidden,
177 cl::desc("A flag that overrides the target's expected cost for "
178 "an instruction to a single constant value. Mostly "
179 "useful for getting consistent testing."));
181 static cl::opt<unsigned> SmallLoopCost(
182 "small-loop-cost", cl::init(20), cl::Hidden,
183 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
185 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
186 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
187 cl::desc("Enable the use of the block frequency analysis to access PGO "
188 "heuristics minimizing code growth in cold regions and being more "
189 "aggressive in hot regions."));
191 // Runtime unroll loops for load/store throughput.
192 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
193 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
194 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
196 /// The number of stores in a loop that are allowed to need predication.
197 static cl::opt<unsigned> NumberOfStoresToPredicate(
198 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
199 cl::desc("Max number of stores to be predicated behind an if."));
201 static cl::opt<bool> EnableIndVarRegisterHeur(
202 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
203 cl::desc("Count the induction variable only once when unrolling"));
205 static cl::opt<bool> EnableCondStoresVectorization(
206 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
207 cl::desc("Enable if predication of stores during vectorization."));
209 static cl::opt<unsigned> MaxNestedScalarReductionUF(
210 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
211 cl::desc("The maximum unroll factor to use when unrolling a scalar "
212 "reduction in a nested loop."));
216 // Forward declarations.
217 class LoopVectorizationLegality;
218 class LoopVectorizationCostModel;
219 class LoopVectorizeHints;
221 /// Optimization analysis message produced during vectorization. Messages inform
222 /// the user why vectorization did not occur.
225 raw_string_ostream Out;
229 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
230 Out << "loop not vectorized: ";
233 template <typename A> Report &operator<<(const A &Value) {
238 Instruction *getInstr() { return Instr; }
240 std::string &str() { return Out.str(); }
241 operator Twine() { return Out.str(); }
244 /// InnerLoopVectorizer vectorizes loops which contain only one basic
245 /// block to a specified vectorization factor (VF).
246 /// This class performs the widening of scalars into vectors, or multiple
247 /// scalars. This class also implements the following features:
248 /// * It inserts an epilogue loop for handling loops that don't have iteration
249 /// counts that are known to be a multiple of the vectorization factor.
250 /// * It handles the code generation for reduction variables.
251 /// * Scalarization (implementation using scalars) of un-vectorizable
253 /// InnerLoopVectorizer does not perform any vectorization-legality
254 /// checks, and relies on the caller to check for the different legality
255 /// aspects. The InnerLoopVectorizer relies on the
256 /// LoopVectorizationLegality class to provide information about the induction
257 /// and reduction variables that were found to a given vectorization factor.
258 class InnerLoopVectorizer {
260 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
261 DominatorTree *DT, const DataLayout *DL,
262 const TargetLibraryInfo *TLI, unsigned VecWidth,
263 unsigned UnrollFactor)
264 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
265 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
266 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
269 // Perform the actual loop widening (vectorization).
270 void vectorize(LoopVectorizationLegality *L) {
272 // Create a new empty loop. Unlink the old loop and connect the new one.
274 // Widen each instruction in the old loop to a new one in the new loop.
275 // Use the Legality module to find the induction and reduction variables.
277 // Register the new loop and update the analysis passes.
281 virtual ~InnerLoopVectorizer() {}
284 /// A small list of PHINodes.
285 typedef SmallVector<PHINode*, 4> PhiVector;
286 /// When we unroll loops we have multiple vector values for each scalar.
287 /// This data structure holds the unrolled and vectorized values that
288 /// originated from one scalar instruction.
289 typedef SmallVector<Value*, 2> VectorParts;
291 // When we if-convert we need create edge masks. We have to cache values so
292 // that we don't end up with exponential recursion/IR.
293 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
294 VectorParts> EdgeMaskCache;
296 /// \brief Add code that checks at runtime if the accessed arrays overlap.
298 /// Returns a pair of instructions where the first element is the first
299 /// instruction generated in possibly a sequence of instructions and the
300 /// second value is the final comparator value or NULL if no check is needed.
301 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
303 /// \brief Add checks for strides that where assumed to be 1.
305 /// Returns the last check instruction and the first check instruction in the
306 /// pair as (first, last).
307 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
309 /// Create an empty loop, based on the loop ranges of the old loop.
310 void createEmptyLoop();
311 /// Copy and widen the instructions from the old loop.
312 virtual void vectorizeLoop();
314 /// \brief The Loop exit block may have single value PHI nodes where the
315 /// incoming value is 'Undef'. While vectorizing we only handled real values
316 /// that were defined inside the loop. Here we fix the 'undef case'.
320 /// A helper function that computes the predicate of the block BB, assuming
321 /// that the header block of the loop is set to True. It returns the *entry*
322 /// mask for the block BB.
323 VectorParts createBlockInMask(BasicBlock *BB);
324 /// A helper function that computes the predicate of the edge between SRC
326 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
328 /// A helper function to vectorize a single BB within the innermost loop.
329 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
331 /// Vectorize a single PHINode in a block. This method handles the induction
332 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
333 /// arbitrary length vectors.
334 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
335 unsigned UF, unsigned VF, PhiVector *PV);
337 /// Insert the new loop to the loop hierarchy and pass manager
338 /// and update the analysis passes.
339 void updateAnalysis();
341 /// This instruction is un-vectorizable. Implement it as a sequence
342 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
343 /// scalarized instruction behind an if block predicated on the control
344 /// dependence of the instruction.
345 virtual void scalarizeInstruction(Instruction *Instr,
346 bool IfPredicateStore=false);
348 /// Vectorize Load and Store instructions,
349 virtual void vectorizeMemoryInstruction(Instruction *Instr);
351 /// Create a broadcast instruction. This method generates a broadcast
352 /// instruction (shuffle) for loop invariant values and for the induction
353 /// value. If this is the induction variable then we extend it to N, N+1, ...
354 /// this is needed because each iteration in the loop corresponds to a SIMD
356 virtual Value *getBroadcastInstrs(Value *V);
358 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
359 /// to each vector element of Val. The sequence starts at StartIndex.
360 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
362 /// When we go over instructions in the basic block we rely on previous
363 /// values within the current basic block or on loop invariant values.
364 /// When we widen (vectorize) values we place them in the map. If the values
365 /// are not within the map, they have to be loop invariant, so we simply
366 /// broadcast them into a vector.
367 VectorParts &getVectorValue(Value *V);
369 /// Generate a shuffle sequence that will reverse the vector Vec.
370 virtual Value *reverseVector(Value *Vec);
372 /// This is a helper class that holds the vectorizer state. It maps scalar
373 /// instructions to vector instructions. When the code is 'unrolled' then
374 /// then a single scalar value is mapped to multiple vector parts. The parts
375 /// are stored in the VectorPart type.
377 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
379 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
381 /// \return True if 'Key' is saved in the Value Map.
382 bool has(Value *Key) const { return MapStorage.count(Key); }
384 /// Initializes a new entry in the map. Sets all of the vector parts to the
385 /// save value in 'Val'.
386 /// \return A reference to a vector with splat values.
387 VectorParts &splat(Value *Key, Value *Val) {
388 VectorParts &Entry = MapStorage[Key];
389 Entry.assign(UF, Val);
393 ///\return A reference to the value that is stored at 'Key'.
394 VectorParts &get(Value *Key) {
395 VectorParts &Entry = MapStorage[Key];
398 assert(Entry.size() == UF);
403 /// The unroll factor. Each entry in the map stores this number of vector
407 /// Map storage. We use std::map and not DenseMap because insertions to a
408 /// dense map invalidates its iterators.
409 std::map<Value *, VectorParts> MapStorage;
412 /// The original loop.
414 /// Scev analysis to use.
423 const DataLayout *DL;
424 /// Target Library Info.
425 const TargetLibraryInfo *TLI;
427 /// The vectorization SIMD factor to use. Each vector will have this many
432 /// The vectorization unroll factor to use. Each scalar is vectorized to this
433 /// many different vector instructions.
436 /// The builder that we use
439 // --- Vectorization state ---
441 /// The vector-loop preheader.
442 BasicBlock *LoopVectorPreHeader;
443 /// The scalar-loop preheader.
444 BasicBlock *LoopScalarPreHeader;
445 /// Middle Block between the vector and the scalar.
446 BasicBlock *LoopMiddleBlock;
447 ///The ExitBlock of the scalar loop.
448 BasicBlock *LoopExitBlock;
449 ///The vector loop body.
450 SmallVector<BasicBlock *, 4> LoopVectorBody;
451 ///The scalar loop body.
452 BasicBlock *LoopScalarBody;
453 /// A list of all bypass blocks. The first block is the entry of the loop.
454 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
456 /// The new Induction variable which was added to the new block.
458 /// The induction variable of the old basic block.
459 PHINode *OldInduction;
460 /// Holds the extended (to the widest induction type) start index.
462 /// Maps scalars to widened vectors.
464 EdgeMaskCache MaskCache;
466 LoopVectorizationLegality *Legal;
469 class InnerLoopUnroller : public InnerLoopVectorizer {
471 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
472 DominatorTree *DT, const DataLayout *DL,
473 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
474 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
477 void scalarizeInstruction(Instruction *Instr,
478 bool IfPredicateStore = false) override;
479 void vectorizeMemoryInstruction(Instruction *Instr) override;
480 Value *getBroadcastInstrs(Value *V) override;
481 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
482 Value *reverseVector(Value *Vec) override;
485 /// \brief Look for a meaningful debug location on the instruction or it's
487 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
492 if (I->getDebugLoc() != Empty)
495 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
496 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
497 if (OpInst->getDebugLoc() != Empty)
504 /// \brief Set the debug location in the builder using the debug location in the
506 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
507 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
508 B.SetCurrentDebugLocation(Inst->getDebugLoc());
510 B.SetCurrentDebugLocation(DebugLoc());
514 /// \return string containing a file name and a line # for the given loop.
515 static std::string getDebugLocString(const Loop *L) {
518 raw_string_ostream OS(Result);
519 const DebugLoc LoopDbgLoc = L->getStartLoc();
520 if (!LoopDbgLoc.isUnknown())
521 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
523 // Just print the module name.
524 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
531 /// \brief Propagate known metadata from one instruction to another.
532 static void propagateMetadata(Instruction *To, const Instruction *From) {
533 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
534 From->getAllMetadataOtherThanDebugLoc(Metadata);
536 for (auto M : Metadata) {
537 unsigned Kind = M.first;
539 // These are safe to transfer (this is safe for TBAA, even when we
540 // if-convert, because should that metadata have had a control dependency
541 // on the condition, and thus actually aliased with some other
542 // non-speculated memory access when the condition was false, this would be
543 // caught by the runtime overlap checks).
544 if (Kind != LLVMContext::MD_tbaa &&
545 Kind != LLVMContext::MD_alias_scope &&
546 Kind != LLVMContext::MD_noalias &&
547 Kind != LLVMContext::MD_fpmath)
550 To->setMetadata(Kind, M.second);
554 /// \brief Propagate known metadata from one instruction to a vector of others.
555 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
557 if (Instruction *I = dyn_cast<Instruction>(V))
558 propagateMetadata(I, From);
561 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
562 /// to what vectorization factor.
563 /// This class does not look at the profitability of vectorization, only the
564 /// legality. This class has two main kinds of checks:
565 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
566 /// will change the order of memory accesses in a way that will change the
567 /// correctness of the program.
568 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
569 /// checks for a number of different conditions, such as the availability of a
570 /// single induction variable, that all types are supported and vectorize-able,
571 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
572 /// This class is also used by InnerLoopVectorizer for identifying
573 /// induction variable and the different reduction variables.
574 class LoopVectorizationLegality {
578 unsigned NumPredStores;
580 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
581 DominatorTree *DT, TargetLibraryInfo *TLI,
582 AliasAnalysis *AA, Function *F,
583 const TargetTransformInfo *TTI)
584 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
585 DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), 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 = C.
607 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
610 // This enum represents the kind of minmax reduction.
611 enum MinMaxReductionKind {
621 /// This struct holds information about reduction variables.
622 struct ReductionDescriptor {
623 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
624 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
626 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
627 MinMaxReductionKind MK)
628 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
630 // The starting value of the reduction.
631 // It does not have to be zero!
632 TrackingVH<Value> StartValue;
633 // The instruction who's value is used outside the loop.
634 Instruction *LoopExitInstr;
635 // The kind of the reduction.
637 // If this a min/max reduction the kind of reduction.
638 MinMaxReductionKind MinMaxKind;
641 /// This POD struct holds information about a potential reduction operation.
642 struct ReductionInstDesc {
643 ReductionInstDesc(bool IsRedux, Instruction *I) :
644 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
646 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
647 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
649 // Is this instruction a reduction candidate.
651 // The last instruction in a min/max pattern (select of the select(icmp())
652 // pattern), or the current reduction instruction otherwise.
653 Instruction *PatternLastInst;
654 // If this is a min/max pattern the comparison predicate.
655 MinMaxReductionKind MinMaxKind;
658 /// This struct holds information about the memory runtime legality
659 /// check that a group of pointers do not overlap.
660 struct RuntimePointerCheck {
661 RuntimePointerCheck() : Need(false) {}
663 /// Reset the state of the pointer runtime information.
670 DependencySetId.clear();
674 /// Insert a pointer and calculate the start and end SCEVs.
675 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
676 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
678 /// This flag indicates if we need to add the runtime check.
680 /// Holds the pointers that we need to check.
681 SmallVector<TrackingVH<Value>, 2> Pointers;
682 /// Holds the pointer value at the beginning of the loop.
683 SmallVector<const SCEV*, 2> Starts;
684 /// Holds the pointer value at the end of the loop.
685 SmallVector<const SCEV*, 2> Ends;
686 /// Holds the information if this pointer is used for writing to memory.
687 SmallVector<bool, 2> IsWritePtr;
688 /// Holds the id of the set of pointers that could be dependent because of a
689 /// shared underlying object.
690 SmallVector<unsigned, 2> DependencySetId;
691 /// Holds the id of the disjoint alias set to which this pointer belongs.
692 SmallVector<unsigned, 2> AliasSetId;
695 /// A struct for saving information about induction variables.
696 struct InductionInfo {
697 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
698 : StartValue(Start), IK(K), StepValue(Step) {
699 assert(IK != IK_NoInduction && "Not an induction");
700 assert(StartValue && "StartValue is null");
701 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
702 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
703 "StartValue is not a pointer for pointer induction");
704 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
705 "StartValue is not an integer for integer induction");
706 assert(StepValue->getType()->isIntegerTy() &&
707 "StepValue is not an integer");
710 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
712 /// Get the consecutive direction. Returns:
713 /// 0 - unknown or non-consecutive.
714 /// 1 - consecutive and increasing.
715 /// -1 - consecutive and decreasing.
716 int getConsecutiveDirection() const {
717 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
718 return StepValue->getSExtValue();
722 /// Compute the transformed value of Index at offset StartValue using step
724 /// For integer induction, returns StartValue + Index * StepValue.
725 /// For pointer induction, returns StartValue[Index * StepValue].
726 /// FIXME: The newly created binary instructions should contain nsw/nuw
727 /// flags, which can be found from the original scalar operations.
728 Value *transform(IRBuilder<> &B, Value *Index) const {
730 case IK_IntInduction:
731 assert(Index->getType() == StartValue->getType() &&
732 "Index type does not match StartValue type");
733 if (StepValue->isMinusOne())
734 return B.CreateSub(StartValue, Index);
735 if (!StepValue->isOne())
736 Index = B.CreateMul(Index, StepValue);
737 return B.CreateAdd(StartValue, Index);
739 case IK_PtrInduction:
740 if (StepValue->isMinusOne())
741 Index = B.CreateNeg(Index);
742 else if (!StepValue->isOne())
743 Index = B.CreateMul(Index, StepValue);
744 return B.CreateGEP(StartValue, Index);
752 TrackingVH<Value> StartValue;
756 ConstantInt *StepValue;
759 /// ReductionList contains the reduction descriptors for all
760 /// of the reductions that were found in the loop.
761 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
763 /// InductionList saves induction variables and maps them to the
764 /// induction descriptor.
765 typedef MapVector<PHINode*, InductionInfo> InductionList;
767 /// Returns true if it is legal to vectorize this loop.
768 /// This does not mean that it is profitable to vectorize this
769 /// loop, only that it is legal to do so.
772 /// Returns the Induction variable.
773 PHINode *getInduction() { return Induction; }
775 /// Returns the reduction variables found in the loop.
776 ReductionList *getReductionVars() { return &Reductions; }
778 /// Returns the induction variables found in the loop.
779 InductionList *getInductionVars() { return &Inductions; }
781 /// Returns the widest induction type.
782 Type *getWidestInductionType() { return WidestIndTy; }
784 /// Returns True if V is an induction variable in this loop.
785 bool isInductionVariable(const Value *V);
787 /// Return true if the block BB needs to be predicated in order for the loop
788 /// to be vectorized.
789 bool blockNeedsPredication(BasicBlock *BB);
791 /// Check if this pointer is consecutive when vectorizing. This happens
792 /// when the last index of the GEP is the induction variable, or that the
793 /// pointer itself is an induction variable.
794 /// This check allows us to vectorize A[idx] into a wide load/store.
796 /// 0 - Stride is unknown or non-consecutive.
797 /// 1 - Address is consecutive.
798 /// -1 - Address is consecutive, and decreasing.
799 int isConsecutivePtr(Value *Ptr);
801 /// Returns true if the value V is uniform within the loop.
802 bool isUniform(Value *V);
804 /// Returns true if this instruction will remain scalar after vectorization.
805 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
807 /// Returns the information that we collected about runtime memory check.
808 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
810 /// This function returns the identity element (or neutral element) for
812 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
814 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
816 bool hasStride(Value *V) { return StrideSet.count(V); }
817 bool mustCheckStrides() { return !StrideSet.empty(); }
818 SmallPtrSet<Value *, 8>::iterator strides_begin() {
819 return StrideSet.begin();
821 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
823 /// Returns true if the target machine supports masked store operation
824 /// for the given \p DataType and kind of access to \p Ptr.
825 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
826 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
828 /// Returns true if the target machine supports masked load operation
829 /// for the given \p DataType and kind of access to \p Ptr.
830 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
831 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
833 /// Returns true if vector representation of the instruction \p I
835 bool isMaskRequired(const Instruction* I) {
836 return (MaskedOp.count(I) != 0);
839 /// Check if a single basic block loop is vectorizable.
840 /// At this point we know that this is a loop with a constant trip count
841 /// and we only need to check individual instructions.
842 bool canVectorizeInstrs();
844 /// When we vectorize loops we may change the order in which
845 /// we read and write from memory. This method checks if it is
846 /// legal to vectorize the code, considering only memory constrains.
847 /// Returns true if the loop is vectorizable
848 bool canVectorizeMemory();
850 /// Return true if we can vectorize this loop using the IF-conversion
852 bool canVectorizeWithIfConvert();
854 /// Collect the variables that need to stay uniform after vectorization.
855 void collectLoopUniforms();
857 /// Return true if all of the instructions in the block can be speculatively
858 /// executed. \p SafePtrs is a list of addresses that are known to be legal
859 /// and we know that we can read from them without segfault.
860 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
862 /// Returns True, if 'Phi' is the kind of reduction variable for type
863 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
864 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
865 /// Returns a struct describing if the instruction 'I' can be a reduction
866 /// variable of type 'Kind'. If the reduction is a min/max pattern of
867 /// select(icmp()) this function advances the instruction pointer 'I' from the
868 /// compare instruction to the select instruction and stores this pointer in
869 /// 'PatternLastInst' member of the returned struct.
870 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
871 ReductionInstDesc &Desc);
872 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
873 /// pattern corresponding to a min(X, Y) or max(X, Y).
874 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
875 ReductionInstDesc &Prev);
876 /// Returns the induction kind of Phi and record the step. This function may
877 /// return NoInduction if the PHI is not an induction variable.
878 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
880 /// \brief Collect memory access with loop invariant strides.
882 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
884 void collectStridedAccess(Value *LoadOrStoreInst);
886 /// Report an analysis message to assist the user in diagnosing loops that are
888 void emitAnalysis(Report &Message) {
889 DebugLoc DL = TheLoop->getStartLoc();
890 if (Instruction *I = Message.getInstr())
891 DL = I->getDebugLoc();
892 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
893 *TheFunction, DL, Message.str());
896 /// The loop that we evaluate.
900 /// DataLayout analysis.
901 const DataLayout *DL;
904 /// Target Library Info.
905 TargetLibraryInfo *TLI;
909 Function *TheFunction;
910 /// Target Transform Info
911 const TargetTransformInfo *TTI;
913 // --- vectorization state --- //
915 /// Holds the integer induction variable. This is the counter of the
918 /// Holds the reduction variables.
919 ReductionList Reductions;
920 /// Holds all of the induction variables that we found in the loop.
921 /// Notice that inductions don't need to start at zero and that induction
922 /// variables can be pointers.
923 InductionList Inductions;
924 /// Holds the widest induction type encountered.
927 /// Allowed outside users. This holds the reduction
928 /// vars which can be accessed from outside the loop.
929 SmallPtrSet<Value*, 4> AllowedExit;
930 /// This set holds the variables which are known to be uniform after
932 SmallPtrSet<Instruction*, 4> Uniforms;
933 /// We need to check that all of the pointers in this list are disjoint
935 RuntimePointerCheck PtrRtCheck;
936 /// Can we assume the absence of NaNs.
937 bool HasFunNoNaNAttr;
939 unsigned MaxSafeDepDistBytes;
941 ValueToValueMap Strides;
942 SmallPtrSet<Value *, 8> StrideSet;
944 /// While vectorizing these instructions we have to generate a
945 /// call to the appropriate masked intrinsic
946 SmallPtrSet<const Instruction*, 8> MaskedOp;
949 /// LoopVectorizationCostModel - estimates the expected speedups due to
951 /// In many cases vectorization is not profitable. This can happen because of
952 /// a number of reasons. In this class we mainly attempt to predict the
953 /// expected speedup/slowdowns due to the supported instruction set. We use the
954 /// TargetTransformInfo to query the different backends for the cost of
955 /// different operations.
956 class LoopVectorizationCostModel {
958 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
959 LoopVectorizationLegality *Legal,
960 const TargetTransformInfo &TTI,
961 const DataLayout *DL, const TargetLibraryInfo *TLI,
962 AssumptionCache *AC, const Function *F,
963 const LoopVectorizeHints *Hints)
964 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
965 TheFunction(F), Hints(Hints) {
966 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
969 /// Information about vectorization costs
970 struct VectorizationFactor {
971 unsigned Width; // Vector width with best cost
972 unsigned Cost; // Cost of the loop with that width
974 /// \return The most profitable vectorization factor and the cost of that VF.
975 /// This method checks every power of two up to VF. If UserVF is not ZERO
976 /// then this vectorization factor will be selected if vectorization is
978 VectorizationFactor selectVectorizationFactor(bool OptForSize);
980 /// \return The size (in bits) of the widest type in the code that
981 /// needs to be vectorized. We ignore values that remain scalar such as
982 /// 64 bit loop indices.
983 unsigned getWidestType();
985 /// \return The most profitable unroll factor.
986 /// If UserUF is non-zero then this method finds the best unroll-factor
987 /// based on register pressure and other parameters.
988 /// VF and LoopCost are the selected vectorization factor and the cost of the
990 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
992 /// \brief A struct that represents some properties of the register usage
994 struct RegisterUsage {
995 /// Holds the number of loop invariant values that are used in the loop.
996 unsigned LoopInvariantRegs;
997 /// Holds the maximum number of concurrent live intervals in the loop.
998 unsigned MaxLocalUsers;
999 /// Holds the number of instructions in the loop.
1000 unsigned NumInstructions;
1003 /// \return information about the register usage of the loop.
1004 RegisterUsage calculateRegisterUsage();
1007 /// Returns the expected execution cost. The unit of the cost does
1008 /// not matter because we use the 'cost' units to compare different
1009 /// vector widths. The cost that is returned is *not* normalized by
1010 /// the factor width.
1011 unsigned expectedCost(unsigned VF);
1013 /// Returns the execution time cost of an instruction for a given vector
1014 /// width. Vector width of one means scalar.
1015 unsigned getInstructionCost(Instruction *I, unsigned VF);
1017 /// A helper function for converting Scalar types to vector types.
1018 /// If the incoming type is void, we return void. If the VF is 1, we return
1019 /// the scalar type.
1020 static Type* ToVectorTy(Type *Scalar, unsigned VF);
1022 /// Returns whether the instruction is a load or store and will be a emitted
1023 /// as a vector operation.
1024 bool isConsecutiveLoadOrStore(Instruction *I);
1026 /// Report an analysis message to assist the user in diagnosing loops that are
1028 void emitAnalysis(Report &Message) {
1029 DebugLoc DL = TheLoop->getStartLoc();
1030 if (Instruction *I = Message.getInstr())
1031 DL = I->getDebugLoc();
1032 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
1033 *TheFunction, DL, Message.str());
1036 /// Values used only by @llvm.assume calls.
1037 SmallPtrSet<const Value *, 32> EphValues;
1039 /// The loop that we evaluate.
1042 ScalarEvolution *SE;
1043 /// Loop Info analysis.
1045 /// Vectorization legality.
1046 LoopVectorizationLegality *Legal;
1047 /// Vector target information.
1048 const TargetTransformInfo &TTI;
1049 /// Target data layout information.
1050 const DataLayout *DL;
1051 /// Target Library Info.
1052 const TargetLibraryInfo *TLI;
1053 const Function *TheFunction;
1054 // Loop Vectorize Hint.
1055 const LoopVectorizeHints *Hints;
1058 /// Utility class for getting and setting loop vectorizer hints in the form
1059 /// of loop metadata.
1060 /// This class keeps a number of loop annotations locally (as member variables)
1061 /// and can, upon request, write them back as metadata on the loop. It will
1062 /// initially scan the loop for existing metadata, and will update the local
1063 /// values based on information in the loop.
1064 /// We cannot write all values to metadata, as the mere presence of some info,
1065 /// for example 'force', means a decision has been made. So, we need to be
1066 /// careful NOT to add them if the user hasn't specifically asked so.
1067 class LoopVectorizeHints {
1074 /// Hint - associates name and validation with the hint value.
1077 unsigned Value; // This may have to change for non-numeric values.
1080 Hint(const char * Name, unsigned Value, HintKind Kind)
1081 : Name(Name), Value(Value), Kind(Kind) { }
1083 bool validate(unsigned Val) {
1086 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1088 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1096 /// Vectorization width.
1098 /// Vectorization interleave factor.
1100 /// Vectorization forced
1103 /// Return the loop metadata prefix.
1104 static StringRef Prefix() { return "llvm.loop."; }
1108 FK_Undefined = -1, ///< Not selected.
1109 FK_Disabled = 0, ///< Forcing disabled.
1110 FK_Enabled = 1, ///< Forcing enabled.
1113 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1114 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1115 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1116 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1118 // Populate values with existing loop metadata.
1119 getHintsFromMetadata();
1121 // force-vector-interleave overrides DisableInterleaving.
1122 if (VectorizationInterleave.getNumOccurrences() > 0)
1123 Interleave.Value = VectorizationInterleave;
1125 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1126 << "LV: Interleaving disabled by the pass manager\n");
1129 /// Mark the loop L as already vectorized by setting the width to 1.
1130 void setAlreadyVectorized() {
1131 Width.Value = Interleave.Value = 1;
1132 Hint Hints[] = {Width, Interleave};
1133 writeHintsToMetadata(Hints);
1136 /// Dumps all the hint information.
1137 std::string emitRemark() const {
1139 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1140 R << "vectorization is explicitly disabled";
1142 R << "use -Rpass-analysis=loop-vectorize for more info";
1143 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1144 R << " (Force=true";
1145 if (Width.Value != 0)
1146 R << ", Vector Width=" << Width.Value;
1147 if (Interleave.Value != 0)
1148 R << ", Interleave Count=" << Interleave.Value;
1156 unsigned getWidth() const { return Width.Value; }
1157 unsigned getInterleave() const { return Interleave.Value; }
1158 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1161 /// Find hints specified in the loop metadata and update local values.
1162 void getHintsFromMetadata() {
1163 MDNode *LoopID = TheLoop->getLoopID();
1167 // First operand should refer to the loop id itself.
1168 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1169 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1171 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1172 const MDString *S = nullptr;
1173 SmallVector<Metadata *, 4> Args;
1175 // The expected hint is either a MDString or a MDNode with the first
1176 // operand a MDString.
1177 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1178 if (!MD || MD->getNumOperands() == 0)
1180 S = dyn_cast<MDString>(MD->getOperand(0));
1181 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1182 Args.push_back(MD->getOperand(i));
1184 S = dyn_cast<MDString>(LoopID->getOperand(i));
1185 assert(Args.size() == 0 && "too many arguments for MDString");
1191 // Check if the hint starts with the loop metadata prefix.
1192 StringRef Name = S->getString();
1193 if (Args.size() == 1)
1194 setHint(Name, Args[0]);
1198 /// Checks string hint with one operand and set value if valid.
1199 void setHint(StringRef Name, Metadata *Arg) {
1200 if (!Name.startswith(Prefix()))
1202 Name = Name.substr(Prefix().size(), StringRef::npos);
1204 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1206 unsigned Val = C->getZExtValue();
1208 Hint *Hints[] = {&Width, &Interleave, &Force};
1209 for (auto H : Hints) {
1210 if (Name == H->Name) {
1211 if (H->validate(Val))
1214 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1220 /// Create a new hint from name / value pair.
1221 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1222 LLVMContext &Context = TheLoop->getHeader()->getContext();
1223 Metadata *MDs[] = {MDString::get(Context, Name),
1224 ConstantAsMetadata::get(
1225 ConstantInt::get(Type::getInt32Ty(Context), V))};
1226 return MDNode::get(Context, MDs);
1229 /// Matches metadata with hint name.
1230 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1231 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1235 for (auto H : HintTypes)
1236 if (Name->getString().endswith(H.Name))
1241 /// Sets current hints into loop metadata, keeping other values intact.
1242 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1243 if (HintTypes.size() == 0)
1246 // Reserve the first element to LoopID (see below).
1247 SmallVector<Metadata *, 4> MDs(1);
1248 // If the loop already has metadata, then ignore the existing operands.
1249 MDNode *LoopID = TheLoop->getLoopID();
1251 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1252 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1253 // If node in update list, ignore old value.
1254 if (!matchesHintMetadataName(Node, HintTypes))
1255 MDs.push_back(Node);
1259 // Now, add the missing hints.
1260 for (auto H : HintTypes)
1261 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1263 // Replace current metadata node with new one.
1264 LLVMContext &Context = TheLoop->getHeader()->getContext();
1265 MDNode *NewLoopID = MDNode::get(Context, MDs);
1266 // Set operand 0 to refer to the loop id itself.
1267 NewLoopID->replaceOperandWith(0, NewLoopID);
1269 TheLoop->setLoopID(NewLoopID);
1272 /// The loop these hints belong to.
1273 const Loop *TheLoop;
1276 static void emitMissedWarning(Function *F, Loop *L,
1277 const LoopVectorizeHints &LH) {
1278 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1279 L->getStartLoc(), LH.emitRemark());
1281 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1282 if (LH.getWidth() != 1)
1283 emitLoopVectorizeWarning(
1284 F->getContext(), *F, L->getStartLoc(),
1285 "failed explicitly specified loop vectorization");
1286 else if (LH.getInterleave() != 1)
1287 emitLoopInterleaveWarning(
1288 F->getContext(), *F, L->getStartLoc(),
1289 "failed explicitly specified loop interleaving");
1293 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1295 return V.push_back(&L);
1297 for (Loop *InnerL : L)
1298 addInnerLoop(*InnerL, V);
1301 /// The LoopVectorize Pass.
1302 struct LoopVectorize : public FunctionPass {
1303 /// Pass identification, replacement for typeid
1306 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1308 DisableUnrolling(NoUnrolling),
1309 AlwaysVectorize(AlwaysVectorize) {
1310 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1313 ScalarEvolution *SE;
1314 const DataLayout *DL;
1316 TargetTransformInfo *TTI;
1318 BlockFrequencyInfo *BFI;
1319 TargetLibraryInfo *TLI;
1321 AssumptionCache *AC;
1322 bool DisableUnrolling;
1323 bool AlwaysVectorize;
1325 BlockFrequency ColdEntryFreq;
1327 bool runOnFunction(Function &F) override {
1328 SE = &getAnalysis<ScalarEvolution>();
1329 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1330 DL = DLP ? &DLP->getDataLayout() : nullptr;
1331 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1332 TTI = &getAnalysis<TargetTransformInfo>();
1333 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1334 BFI = &getAnalysis<BlockFrequencyInfo>();
1335 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1336 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1337 AA = &getAnalysis<AliasAnalysis>();
1338 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1340 // Compute some weights outside of the loop over the loops. Compute this
1341 // using a BranchProbability to re-use its scaling math.
1342 const BranchProbability ColdProb(1, 5); // 20%
1343 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1345 // If the target claims to have no vector registers don't attempt
1347 if (!TTI->getNumberOfRegisters(true))
1351 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1352 << ": Missing data layout\n");
1356 // Build up a worklist of inner-loops to vectorize. This is necessary as
1357 // the act of vectorizing or partially unrolling a loop creates new loops
1358 // and can invalidate iterators across the loops.
1359 SmallVector<Loop *, 8> Worklist;
1362 addInnerLoop(*L, Worklist);
1364 LoopsAnalyzed += Worklist.size();
1366 // Now walk the identified inner loops.
1367 bool Changed = false;
1368 while (!Worklist.empty())
1369 Changed |= processLoop(Worklist.pop_back_val());
1371 // Process each loop nest in the function.
1375 bool processLoop(Loop *L) {
1376 assert(L->empty() && "Only process inner loops.");
1379 const std::string DebugLocStr = getDebugLocString(L);
1382 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1383 << L->getHeader()->getParent()->getName() << "\" from "
1384 << DebugLocStr << "\n");
1386 LoopVectorizeHints Hints(L, DisableUnrolling);
1388 DEBUG(dbgs() << "LV: Loop hints:"
1390 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1392 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1394 : "?")) << " width=" << Hints.getWidth()
1395 << " unroll=" << Hints.getInterleave() << "\n");
1397 // Function containing loop
1398 Function *F = L->getHeader()->getParent();
1400 // Looking at the diagnostic output is the only way to determine if a loop
1401 // was vectorized (other than looking at the IR or machine code), so it
1402 // is important to generate an optimization remark for each loop. Most of
1403 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1404 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1405 // less verbose reporting vectorized loops and unvectorized loops that may
1406 // benefit from vectorization, respectively.
1408 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1409 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1410 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1411 L->getStartLoc(), Hints.emitRemark());
1415 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1416 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1417 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1418 L->getStartLoc(), Hints.emitRemark());
1422 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1423 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1424 emitOptimizationRemarkAnalysis(
1425 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1426 "loop not vectorized: vector width and interleave count are "
1427 "explicitly set to 1");
1431 // Check the loop for a trip count threshold:
1432 // do not vectorize loops with a tiny trip count.
1433 const unsigned TC = SE->getSmallConstantTripCount(L);
1434 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1435 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1436 << "This loop is not worth vectorizing.");
1437 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1438 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1440 DEBUG(dbgs() << "\n");
1441 emitOptimizationRemarkAnalysis(
1442 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1443 "vectorization is not beneficial and is not explicitly forced");
1448 // Check if it is legal to vectorize the loop.
1449 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1450 if (!LVL.canVectorize()) {
1451 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1452 emitMissedWarning(F, L, Hints);
1456 // Use the cost model.
1457 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1460 // Check the function attributes to find out if this function should be
1461 // optimized for size.
1462 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1463 F->hasFnAttribute(Attribute::OptimizeForSize);
1465 // Compute the weighted frequency of this loop being executed and see if it
1466 // is less than 20% of the function entry baseline frequency. Note that we
1467 // always have a canonical loop here because we think we *can* vectoriez.
1468 // FIXME: This is hidden behind a flag due to pervasive problems with
1469 // exactly what block frequency models.
1470 if (LoopVectorizeWithBlockFrequency) {
1471 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1472 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1473 LoopEntryFreq < ColdEntryFreq)
1477 // Check the function attributes to see if implicit floats are allowed.a
1478 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1479 // an integer loop and the vector instructions selected are purely integer
1480 // vector instructions?
1481 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1482 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1483 "attribute is used.\n");
1484 emitOptimizationRemarkAnalysis(
1485 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1486 "loop not vectorized due to NoImplicitFloat attribute");
1487 emitMissedWarning(F, L, Hints);
1491 // Select the optimal vectorization factor.
1492 const LoopVectorizationCostModel::VectorizationFactor VF =
1493 CM.selectVectorizationFactor(OptForSize);
1495 // Select the unroll factor.
1497 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1499 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1500 << DebugLocStr << '\n');
1501 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1503 if (VF.Width == 1) {
1504 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1507 emitOptimizationRemarkAnalysis(
1508 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1509 "not beneficial to vectorize and user disabled interleaving");
1512 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1514 // Report the unrolling decision.
1515 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1516 Twine("unrolled with interleaving factor " +
1518 " (vectorization not beneficial)"));
1520 // We decided not to vectorize, but we may want to unroll.
1522 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1523 Unroller.vectorize(&LVL);
1525 // If we decided that it is *legal* to vectorize the loop then do it.
1526 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1530 // Report the vectorization decision.
1531 emitOptimizationRemark(
1532 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1533 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1534 ", unrolling interleave factor: " + Twine(UF) + ")");
1537 // Mark the loop as already vectorized to avoid vectorizing again.
1538 Hints.setAlreadyVectorized();
1540 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1544 void getAnalysisUsage(AnalysisUsage &AU) const override {
1545 AU.addRequired<AssumptionCacheTracker>();
1546 AU.addRequiredID(LoopSimplifyID);
1547 AU.addRequiredID(LCSSAID);
1548 AU.addRequired<BlockFrequencyInfo>();
1549 AU.addRequired<DominatorTreeWrapperPass>();
1550 AU.addRequired<LoopInfoWrapperPass>();
1551 AU.addRequired<ScalarEvolution>();
1552 AU.addRequired<TargetTransformInfo>();
1553 AU.addRequired<AliasAnalysis>();
1554 AU.addPreserved<LoopInfoWrapperPass>();
1555 AU.addPreserved<DominatorTreeWrapperPass>();
1556 AU.addPreserved<AliasAnalysis>();
1561 } // end anonymous namespace
1563 //===----------------------------------------------------------------------===//
1564 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1565 // LoopVectorizationCostModel.
1566 //===----------------------------------------------------------------------===//
1568 static Value *stripIntegerCast(Value *V) {
1569 if (CastInst *CI = dyn_cast<CastInst>(V))
1570 if (CI->getOperand(0)->getType()->isIntegerTy())
1571 return CI->getOperand(0);
1575 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1577 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1579 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1580 ValueToValueMap &PtrToStride,
1581 Value *Ptr, Value *OrigPtr = nullptr) {
1583 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1585 // If there is an entry in the map return the SCEV of the pointer with the
1586 // symbolic stride replaced by one.
1587 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1588 if (SI != PtrToStride.end()) {
1589 Value *StrideVal = SI->second;
1592 StrideVal = stripIntegerCast(StrideVal);
1594 // Replace symbolic stride by one.
1595 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1596 ValueToValueMap RewriteMap;
1597 RewriteMap[StrideVal] = One;
1600 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1601 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1606 // Otherwise, just return the SCEV of the original pointer.
1607 return SE->getSCEV(Ptr);
1610 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1611 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1612 unsigned ASId, ValueToValueMap &Strides) {
1613 // Get the stride replaced scev.
1614 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1615 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1616 assert(AR && "Invalid addrec expression");
1617 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1618 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1619 Pointers.push_back(Ptr);
1620 Starts.push_back(AR->getStart());
1621 Ends.push_back(ScEnd);
1622 IsWritePtr.push_back(WritePtr);
1623 DependencySetId.push_back(DepSetId);
1624 AliasSetId.push_back(ASId);
1627 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1628 // We need to place the broadcast of invariant variables outside the loop.
1629 Instruction *Instr = dyn_cast<Instruction>(V);
1631 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1632 Instr->getParent()) != LoopVectorBody.end());
1633 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1635 // Place the code for broadcasting invariant variables in the new preheader.
1636 IRBuilder<>::InsertPointGuard Guard(Builder);
1638 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1640 // Broadcast the scalar into all locations in the vector.
1641 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1646 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1648 assert(Val->getType()->isVectorTy() && "Must be a vector");
1649 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1650 "Elem must be an integer");
1651 assert(Step->getType() == Val->getType()->getScalarType() &&
1652 "Step has wrong type");
1653 // Create the types.
1654 Type *ITy = Val->getType()->getScalarType();
1655 VectorType *Ty = cast<VectorType>(Val->getType());
1656 int VLen = Ty->getNumElements();
1657 SmallVector<Constant*, 8> Indices;
1659 // Create a vector of consecutive numbers from zero to VF.
1660 for (int i = 0; i < VLen; ++i)
1661 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1663 // Add the consecutive indices to the vector value.
1664 Constant *Cv = ConstantVector::get(Indices);
1665 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1666 Step = Builder.CreateVectorSplat(VLen, Step);
1667 assert(Step->getType() == Val->getType() && "Invalid step vec");
1668 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1669 // which can be found from the original scalar operations.
1670 Step = Builder.CreateMul(Cv, Step);
1671 return Builder.CreateAdd(Val, Step, "induction");
1674 /// \brief Find the operand of the GEP that should be checked for consecutive
1675 /// stores. This ignores trailing indices that have no effect on the final
1677 static unsigned getGEPInductionOperand(const DataLayout *DL,
1678 const GetElementPtrInst *Gep) {
1679 unsigned LastOperand = Gep->getNumOperands() - 1;
1680 unsigned GEPAllocSize = DL->getTypeAllocSize(
1681 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1683 // Walk backwards and try to peel off zeros.
1684 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1685 // Find the type we're currently indexing into.
1686 gep_type_iterator GEPTI = gep_type_begin(Gep);
1687 std::advance(GEPTI, LastOperand - 1);
1689 // If it's a type with the same allocation size as the result of the GEP we
1690 // can peel off the zero index.
1691 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1699 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1700 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1701 // Make sure that the pointer does not point to structs.
1702 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1705 // If this value is a pointer induction variable we know it is consecutive.
1706 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1707 if (Phi && Inductions.count(Phi)) {
1708 InductionInfo II = Inductions[Phi];
1709 return II.getConsecutiveDirection();
1712 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1716 unsigned NumOperands = Gep->getNumOperands();
1717 Value *GpPtr = Gep->getPointerOperand();
1718 // If this GEP value is a consecutive pointer induction variable and all of
1719 // the indices are constant then we know it is consecutive. We can
1720 Phi = dyn_cast<PHINode>(GpPtr);
1721 if (Phi && Inductions.count(Phi)) {
1723 // Make sure that the pointer does not point to structs.
1724 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1725 if (GepPtrType->getElementType()->isAggregateType())
1728 // Make sure that all of the index operands are loop invariant.
1729 for (unsigned i = 1; i < NumOperands; ++i)
1730 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1733 InductionInfo II = Inductions[Phi];
1734 return II.getConsecutiveDirection();
1737 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1739 // Check that all of the gep indices are uniform except for our induction
1741 for (unsigned i = 0; i != NumOperands; ++i)
1742 if (i != InductionOperand &&
1743 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1746 // We can emit wide load/stores only if the last non-zero index is the
1747 // induction variable.
1748 const SCEV *Last = nullptr;
1749 if (!Strides.count(Gep))
1750 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1752 // Because of the multiplication by a stride we can have a s/zext cast.
1753 // We are going to replace this stride by 1 so the cast is safe to ignore.
1755 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1756 // %0 = trunc i64 %indvars.iv to i32
1757 // %mul = mul i32 %0, %Stride1
1758 // %idxprom = zext i32 %mul to i64 << Safe cast.
1759 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1761 Last = replaceSymbolicStrideSCEV(SE, Strides,
1762 Gep->getOperand(InductionOperand), Gep);
1763 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1765 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1769 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1770 const SCEV *Step = AR->getStepRecurrence(*SE);
1772 // The memory is consecutive because the last index is consecutive
1773 // and all other indices are loop invariant.
1776 if (Step->isAllOnesValue())
1783 bool LoopVectorizationLegality::isUniform(Value *V) {
1784 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1787 InnerLoopVectorizer::VectorParts&
1788 InnerLoopVectorizer::getVectorValue(Value *V) {
1789 assert(V != Induction && "The new induction variable should not be used.");
1790 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1792 // If we have a stride that is replaced by one, do it here.
1793 if (Legal->hasStride(V))
1794 V = ConstantInt::get(V->getType(), 1);
1796 // If we have this scalar in the map, return it.
1797 if (WidenMap.has(V))
1798 return WidenMap.get(V);
1800 // If this scalar is unknown, assume that it is a constant or that it is
1801 // loop invariant. Broadcast V and save the value for future uses.
1802 Value *B = getBroadcastInstrs(V);
1803 return WidenMap.splat(V, B);
1806 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1807 assert(Vec->getType()->isVectorTy() && "Invalid type");
1808 SmallVector<Constant*, 8> ShuffleMask;
1809 for (unsigned i = 0; i < VF; ++i)
1810 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1812 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1813 ConstantVector::get(ShuffleMask),
1817 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1818 // Attempt to issue a wide load.
1819 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1820 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1822 assert((LI || SI) && "Invalid Load/Store instruction");
1824 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1825 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1826 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1827 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1828 // An alignment of 0 means target abi alignment. We need to use the scalar's
1829 // target abi alignment in such a case.
1831 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1832 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1833 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1834 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1836 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1837 !Legal->isMaskRequired(SI))
1838 return scalarizeInstruction(Instr, true);
1840 if (ScalarAllocatedSize != VectorElementSize)
1841 return scalarizeInstruction(Instr);
1843 // If the pointer is loop invariant or if it is non-consecutive,
1844 // scalarize the load.
1845 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1846 bool Reverse = ConsecutiveStride < 0;
1847 bool UniformLoad = LI && Legal->isUniform(Ptr);
1848 if (!ConsecutiveStride || UniformLoad)
1849 return scalarizeInstruction(Instr);
1851 Constant *Zero = Builder.getInt32(0);
1852 VectorParts &Entry = WidenMap.get(Instr);
1854 // Handle consecutive loads/stores.
1855 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1856 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1857 setDebugLocFromInst(Builder, Gep);
1858 Value *PtrOperand = Gep->getPointerOperand();
1859 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1860 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1862 // Create the new GEP with the new induction variable.
1863 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1864 Gep2->setOperand(0, FirstBasePtr);
1865 Gep2->setName("gep.indvar.base");
1866 Ptr = Builder.Insert(Gep2);
1868 setDebugLocFromInst(Builder, Gep);
1869 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1870 OrigLoop) && "Base ptr must be invariant");
1872 // The last index does not have to be the induction. It can be
1873 // consecutive and be a function of the index. For example A[I+1];
1874 unsigned NumOperands = Gep->getNumOperands();
1875 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1876 // Create the new GEP with the new induction variable.
1877 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1879 for (unsigned i = 0; i < NumOperands; ++i) {
1880 Value *GepOperand = Gep->getOperand(i);
1881 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1883 // Update last index or loop invariant instruction anchored in loop.
1884 if (i == InductionOperand ||
1885 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1886 assert((i == InductionOperand ||
1887 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1888 "Must be last index or loop invariant");
1890 VectorParts &GEPParts = getVectorValue(GepOperand);
1891 Value *Index = GEPParts[0];
1892 Index = Builder.CreateExtractElement(Index, Zero);
1893 Gep2->setOperand(i, Index);
1894 Gep2->setName("gep.indvar.idx");
1897 Ptr = Builder.Insert(Gep2);
1899 // Use the induction element ptr.
1900 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1901 setDebugLocFromInst(Builder, Ptr);
1902 VectorParts &PtrVal = getVectorValue(Ptr);
1903 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1906 VectorParts Mask = createBlockInMask(Instr->getParent());
1909 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1910 "We do not allow storing to uniform addresses");
1911 setDebugLocFromInst(Builder, SI);
1912 // We don't want to update the value in the map as it might be used in
1913 // another expression. So don't use a reference type for "StoredVal".
1914 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1916 for (unsigned Part = 0; Part < UF; ++Part) {
1917 // Calculate the pointer for the specific unroll-part.
1918 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1921 // If we store to reverse consecutive memory locations then we need
1922 // to reverse the order of elements in the stored value.
1923 StoredVal[Part] = reverseVector(StoredVal[Part]);
1924 // If the address is consecutive but reversed, then the
1925 // wide store needs to start at the last vector element.
1926 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1927 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1928 Mask[Part] = reverseVector(Mask[Part]);
1931 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1932 DataTy->getPointerTo(AddressSpace));
1935 if (Legal->isMaskRequired(SI))
1936 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1939 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1940 propagateMetadata(NewSI, SI);
1946 assert(LI && "Must have a load instruction");
1947 setDebugLocFromInst(Builder, LI);
1948 for (unsigned Part = 0; Part < UF; ++Part) {
1949 // Calculate the pointer for the specific unroll-part.
1950 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1953 // If the address is consecutive but reversed, then the
1954 // wide load needs to start at the last vector element.
1955 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1956 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1957 Mask[Part] = reverseVector(Mask[Part]);
1961 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1962 DataTy->getPointerTo(AddressSpace));
1963 if (Legal->isMaskRequired(LI))
1964 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1965 UndefValue::get(DataTy),
1966 "wide.masked.load");
1968 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1969 propagateMetadata(NewLI, LI);
1970 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1974 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1975 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1976 // Holds vector parameters or scalars, in case of uniform vals.
1977 SmallVector<VectorParts, 4> Params;
1979 setDebugLocFromInst(Builder, Instr);
1981 // Find all of the vectorized parameters.
1982 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1983 Value *SrcOp = Instr->getOperand(op);
1985 // If we are accessing the old induction variable, use the new one.
1986 if (SrcOp == OldInduction) {
1987 Params.push_back(getVectorValue(SrcOp));
1991 // Try using previously calculated values.
1992 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1994 // If the src is an instruction that appeared earlier in the basic block
1995 // then it should already be vectorized.
1996 if (SrcInst && OrigLoop->contains(SrcInst)) {
1997 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1998 // The parameter is a vector value from earlier.
1999 Params.push_back(WidenMap.get(SrcInst));
2001 // The parameter is a scalar from outside the loop. Maybe even a constant.
2002 VectorParts Scalars;
2003 Scalars.append(UF, SrcOp);
2004 Params.push_back(Scalars);
2008 assert(Params.size() == Instr->getNumOperands() &&
2009 "Invalid number of operands");
2011 // Does this instruction return a value ?
2012 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2014 Value *UndefVec = IsVoidRetTy ? nullptr :
2015 UndefValue::get(VectorType::get(Instr->getType(), VF));
2016 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2017 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2019 Instruction *InsertPt = Builder.GetInsertPoint();
2020 BasicBlock *IfBlock = Builder.GetInsertBlock();
2021 BasicBlock *CondBlock = nullptr;
2024 Loop *VectorLp = nullptr;
2025 if (IfPredicateStore) {
2026 assert(Instr->getParent()->getSinglePredecessor() &&
2027 "Only support single predecessor blocks");
2028 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2029 Instr->getParent());
2030 VectorLp = LI->getLoopFor(IfBlock);
2031 assert(VectorLp && "Must have a loop for this block");
2034 // For each vector unroll 'part':
2035 for (unsigned Part = 0; Part < UF; ++Part) {
2036 // For each scalar that we create:
2037 for (unsigned Width = 0; Width < VF; ++Width) {
2040 Value *Cmp = nullptr;
2041 if (IfPredicateStore) {
2042 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2043 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2044 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2045 LoopVectorBody.push_back(CondBlock);
2046 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2047 // Update Builder with newly created basic block.
2048 Builder.SetInsertPoint(InsertPt);
2051 Instruction *Cloned = Instr->clone();
2053 Cloned->setName(Instr->getName() + ".cloned");
2054 // Replace the operands of the cloned instructions with extracted scalars.
2055 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2056 Value *Op = Params[op][Part];
2057 // Param is a vector. Need to extract the right lane.
2058 if (Op->getType()->isVectorTy())
2059 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2060 Cloned->setOperand(op, Op);
2063 // Place the cloned scalar in the new loop.
2064 Builder.Insert(Cloned);
2066 // If the original scalar returns a value we need to place it in a vector
2067 // so that future users will be able to use it.
2069 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2070 Builder.getInt32(Width));
2072 if (IfPredicateStore) {
2073 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2074 LoopVectorBody.push_back(NewIfBlock);
2075 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2076 Builder.SetInsertPoint(InsertPt);
2077 Instruction *OldBr = IfBlock->getTerminator();
2078 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2079 OldBr->eraseFromParent();
2080 IfBlock = NewIfBlock;
2086 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2090 if (Instruction *I = dyn_cast<Instruction>(V))
2091 return I->getParent() == Loc->getParent() ? I : nullptr;
2095 std::pair<Instruction *, Instruction *>
2096 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2097 Instruction *tnullptr = nullptr;
2098 if (!Legal->mustCheckStrides())
2099 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2101 IRBuilder<> ChkBuilder(Loc);
2104 Value *Check = nullptr;
2105 Instruction *FirstInst = nullptr;
2106 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2107 SE = Legal->strides_end();
2109 Value *Ptr = stripIntegerCast(*SI);
2110 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2112 // Store the first instruction we create.
2113 FirstInst = getFirstInst(FirstInst, C, Loc);
2115 Check = ChkBuilder.CreateOr(Check, C);
2120 // We have to do this trickery because the IRBuilder might fold the check to a
2121 // constant expression in which case there is no Instruction anchored in a
2123 LLVMContext &Ctx = Loc->getContext();
2124 Instruction *TheCheck =
2125 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2126 ChkBuilder.Insert(TheCheck, "stride.not.one");
2127 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2129 return std::make_pair(FirstInst, TheCheck);
2132 std::pair<Instruction *, Instruction *>
2133 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2134 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2135 Legal->getRuntimePointerCheck();
2137 Instruction *tnullptr = nullptr;
2138 if (!PtrRtCheck->Need)
2139 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2141 unsigned NumPointers = PtrRtCheck->Pointers.size();
2142 SmallVector<TrackingVH<Value> , 2> Starts;
2143 SmallVector<TrackingVH<Value> , 2> Ends;
2145 LLVMContext &Ctx = Loc->getContext();
2146 SCEVExpander Exp(*SE, "induction");
2147 Instruction *FirstInst = nullptr;
2149 for (unsigned i = 0; i < NumPointers; ++i) {
2150 Value *Ptr = PtrRtCheck->Pointers[i];
2151 const SCEV *Sc = SE->getSCEV(Ptr);
2153 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2154 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2156 Starts.push_back(Ptr);
2157 Ends.push_back(Ptr);
2159 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2160 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2162 // Use this type for pointer arithmetic.
2163 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2165 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2166 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2167 Starts.push_back(Start);
2168 Ends.push_back(End);
2172 IRBuilder<> ChkBuilder(Loc);
2173 // Our instructions might fold to a constant.
2174 Value *MemoryRuntimeCheck = nullptr;
2175 for (unsigned i = 0; i < NumPointers; ++i) {
2176 for (unsigned j = i+1; j < NumPointers; ++j) {
2177 // No need to check if two readonly pointers intersect.
2178 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2181 // Only need to check pointers between two different dependency sets.
2182 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2184 // Only need to check pointers in the same alias set.
2185 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2188 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2189 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2191 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2192 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2193 "Trying to bounds check pointers with different address spaces");
2195 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2196 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2198 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2199 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2200 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2201 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2203 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2204 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2205 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2206 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2207 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2208 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2209 if (MemoryRuntimeCheck) {
2210 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2212 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2214 MemoryRuntimeCheck = IsConflict;
2218 // We have to do this trickery because the IRBuilder might fold the check to a
2219 // constant expression in which case there is no Instruction anchored in a
2221 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2222 ConstantInt::getTrue(Ctx));
2223 ChkBuilder.Insert(Check, "memcheck.conflict");
2224 FirstInst = getFirstInst(FirstInst, Check, Loc);
2225 return std::make_pair(FirstInst, Check);
2228 void InnerLoopVectorizer::createEmptyLoop() {
2230 In this function we generate a new loop. The new loop will contain
2231 the vectorized instructions while the old loop will continue to run the
2234 [ ] <-- Back-edge taken count overflow check.
2237 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2240 || [ ] <-- vector pre header.
2244 || [ ]_| <-- vector loop.
2247 | >[ ] <--- middle-block.
2250 -|- >[ ] <--- new preheader.
2254 | [ ]_| <-- old scalar loop to handle remainder.
2257 >[ ] <-- exit block.
2261 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2262 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2263 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2264 assert(BypassBlock && "Invalid loop structure");
2265 assert(ExitBlock && "Must have an exit block");
2267 // Some loops have a single integer induction variable, while other loops
2268 // don't. One example is c++ iterators that often have multiple pointer
2269 // induction variables. In the code below we also support a case where we
2270 // don't have a single induction variable.
2271 OldInduction = Legal->getInduction();
2272 Type *IdxTy = Legal->getWidestInductionType();
2274 // Find the loop boundaries.
2275 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2276 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2278 // The exit count might have the type of i64 while the phi is i32. This can
2279 // happen if we have an induction variable that is sign extended before the
2280 // compare. The only way that we get a backedge taken count is that the
2281 // induction variable was signed and as such will not overflow. In such a case
2282 // truncation is legal.
2283 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2284 IdxTy->getPrimitiveSizeInBits())
2285 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2287 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2288 // Get the total trip count from the count by adding 1.
2289 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2290 SE->getConstant(BackedgeTakeCount->getType(), 1));
2292 // Expand the trip count and place the new instructions in the preheader.
2293 // Notice that the pre-header does not change, only the loop body.
2294 SCEVExpander Exp(*SE, "induction");
2296 // We need to test whether the backedge-taken count is uint##_max. Adding one
2297 // to it will cause overflow and an incorrect loop trip count in the vector
2298 // body. In case of overflow we want to directly jump to the scalar remainder
2300 Value *BackedgeCount =
2301 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2302 BypassBlock->getTerminator());
2303 if (BackedgeCount->getType()->isPointerTy())
2304 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2305 "backedge.ptrcnt.to.int",
2306 BypassBlock->getTerminator());
2307 Instruction *CheckBCOverflow =
2308 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2309 Constant::getAllOnesValue(BackedgeCount->getType()),
2310 "backedge.overflow", BypassBlock->getTerminator());
2312 // The loop index does not have to start at Zero. Find the original start
2313 // value from the induction PHI node. If we don't have an induction variable
2314 // then we know that it starts at zero.
2315 Builder.SetInsertPoint(BypassBlock->getTerminator());
2316 Value *StartIdx = ExtendedIdx = OldInduction ?
2317 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2319 ConstantInt::get(IdxTy, 0);
2321 // We need an instruction to anchor the overflow check on. StartIdx needs to
2322 // be defined before the overflow check branch. Because the scalar preheader
2323 // is going to merge the start index and so the overflow branch block needs to
2324 // contain a definition of the start index.
2325 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2326 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2327 BypassBlock->getTerminator());
2329 // Count holds the overall loop count (N).
2330 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2331 BypassBlock->getTerminator());
2333 LoopBypassBlocks.push_back(BypassBlock);
2335 // Split the single block loop into the two loop structure described above.
2336 BasicBlock *VectorPH =
2337 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2338 BasicBlock *VecBody =
2339 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2340 BasicBlock *MiddleBlock =
2341 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2342 BasicBlock *ScalarPH =
2343 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2345 // Create and register the new vector loop.
2346 Loop* Lp = new Loop();
2347 Loop *ParentLoop = OrigLoop->getParentLoop();
2349 // Insert the new loop into the loop nest and register the new basic blocks
2350 // before calling any utilities such as SCEV that require valid LoopInfo.
2352 ParentLoop->addChildLoop(Lp);
2353 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2354 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2355 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2357 LI->addTopLevelLoop(Lp);
2359 Lp->addBasicBlockToLoop(VecBody, *LI);
2361 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2363 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2365 // Generate the induction variable.
2366 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2367 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2368 // The loop step is equal to the vectorization factor (num of SIMD elements)
2369 // times the unroll factor (num of SIMD instructions).
2370 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2372 // This is the IR builder that we use to add all of the logic for bypassing
2373 // the new vector loop.
2374 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2375 setDebugLocFromInst(BypassBuilder,
2376 getDebugLocFromInstOrOperands(OldInduction));
2378 // We may need to extend the index in case there is a type mismatch.
2379 // We know that the count starts at zero and does not overflow.
2380 if (Count->getType() != IdxTy) {
2381 // The exit count can be of pointer type. Convert it to the correct
2383 if (ExitCount->getType()->isPointerTy())
2384 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2386 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2389 // Add the start index to the loop count to get the new end index.
2390 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2392 // Now we need to generate the expression for N - (N % VF), which is
2393 // the part that the vectorized body will execute.
2394 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2395 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2396 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2397 "end.idx.rnd.down");
2399 // Now, compare the new count to zero. If it is zero skip the vector loop and
2400 // jump to the scalar loop.
2402 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2404 BasicBlock *LastBypassBlock = BypassBlock;
2406 // Generate code to check that the loops trip count that we computed by adding
2407 // one to the backedge-taken count will not overflow.
2409 auto PastOverflowCheck =
2410 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2411 BasicBlock *CheckBlock =
2412 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2414 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2415 LoopBypassBlocks.push_back(CheckBlock);
2416 Instruction *OldTerm = LastBypassBlock->getTerminator();
2417 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2418 OldTerm->eraseFromParent();
2419 LastBypassBlock = CheckBlock;
2422 // Generate the code to check that the strides we assumed to be one are really
2423 // one. We want the new basic block to start at the first instruction in a
2424 // sequence of instructions that form a check.
2425 Instruction *StrideCheck;
2426 Instruction *FirstCheckInst;
2427 std::tie(FirstCheckInst, StrideCheck) =
2428 addStrideCheck(LastBypassBlock->getTerminator());
2430 // Create a new block containing the stride check.
2431 BasicBlock *CheckBlock =
2432 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2434 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2435 LoopBypassBlocks.push_back(CheckBlock);
2437 // Replace the branch into the memory check block with a conditional branch
2438 // for the "few elements case".
2439 Instruction *OldTerm = LastBypassBlock->getTerminator();
2440 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2441 OldTerm->eraseFromParent();
2444 LastBypassBlock = CheckBlock;
2447 // Generate the code that checks in runtime if arrays overlap. We put the
2448 // checks into a separate block to make the more common case of few elements
2450 Instruction *MemRuntimeCheck;
2451 std::tie(FirstCheckInst, MemRuntimeCheck) =
2452 addRuntimeCheck(LastBypassBlock->getTerminator());
2453 if (MemRuntimeCheck) {
2454 // Create a new block containing the memory check.
2455 BasicBlock *CheckBlock =
2456 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2458 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2459 LoopBypassBlocks.push_back(CheckBlock);
2461 // Replace the branch into the memory check block with a conditional branch
2462 // for the "few elements case".
2463 Instruction *OldTerm = LastBypassBlock->getTerminator();
2464 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2465 OldTerm->eraseFromParent();
2467 Cmp = MemRuntimeCheck;
2468 LastBypassBlock = CheckBlock;
2471 LastBypassBlock->getTerminator()->eraseFromParent();
2472 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2475 // We are going to resume the execution of the scalar loop.
2476 // Go over all of the induction variables that we found and fix the
2477 // PHIs that are left in the scalar version of the loop.
2478 // The starting values of PHI nodes depend on the counter of the last
2479 // iteration in the vectorized loop.
2480 // If we come from a bypass edge then we need to start from the original
2483 // This variable saves the new starting index for the scalar loop.
2484 PHINode *ResumeIndex = nullptr;
2485 LoopVectorizationLegality::InductionList::iterator I, E;
2486 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2487 // Set builder to point to last bypass block.
2488 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2489 for (I = List->begin(), E = List->end(); I != E; ++I) {
2490 PHINode *OrigPhi = I->first;
2491 LoopVectorizationLegality::InductionInfo II = I->second;
2493 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2494 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2495 MiddleBlock->getTerminator());
2496 // We might have extended the type of the induction variable but we need a
2497 // truncated version for the scalar loop.
2498 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2499 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2500 MiddleBlock->getTerminator()) : nullptr;
2502 // Create phi nodes to merge from the backedge-taken check block.
2503 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2504 ScalarPH->getTerminator());
2505 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2507 PHINode *BCTruncResumeVal = nullptr;
2508 if (OrigPhi == OldInduction) {
2510 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2511 ScalarPH->getTerminator());
2512 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2515 Value *EndValue = nullptr;
2517 case LoopVectorizationLegality::IK_NoInduction:
2518 llvm_unreachable("Unknown induction");
2519 case LoopVectorizationLegality::IK_IntInduction: {
2520 // Handle the integer induction counter.
2521 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2523 // We have the canonical induction variable.
2524 if (OrigPhi == OldInduction) {
2525 // Create a truncated version of the resume value for the scalar loop,
2526 // we might have promoted the type to a larger width.
2528 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2529 // The new PHI merges the original incoming value, in case of a bypass,
2530 // or the value at the end of the vectorized loop.
2531 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2532 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2533 TruncResumeVal->addIncoming(EndValue, VecBody);
2535 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2537 // We know what the end value is.
2538 EndValue = IdxEndRoundDown;
2539 // We also know which PHI node holds it.
2540 ResumeIndex = ResumeVal;
2544 // Not the canonical induction variable - add the vector loop count to the
2546 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2547 II.StartValue->getType(),
2549 EndValue = II.transform(BypassBuilder, CRD);
2550 EndValue->setName("ind.end");
2553 case LoopVectorizationLegality::IK_PtrInduction: {
2554 EndValue = II.transform(BypassBuilder, CountRoundDown);
2555 EndValue->setName("ptr.ind.end");
2560 // The new PHI merges the original incoming value, in case of a bypass,
2561 // or the value at the end of the vectorized loop.
2562 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2563 if (OrigPhi == OldInduction)
2564 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2566 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2568 ResumeVal->addIncoming(EndValue, VecBody);
2570 // Fix the scalar body counter (PHI node).
2571 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2573 // The old induction's phi node in the scalar body needs the truncated
2575 if (OrigPhi == OldInduction) {
2576 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2577 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2579 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2580 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2584 // If we are generating a new induction variable then we also need to
2585 // generate the code that calculates the exit value. This value is not
2586 // simply the end of the counter because we may skip the vectorized body
2587 // in case of a runtime check.
2589 assert(!ResumeIndex && "Unexpected resume value found");
2590 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2591 MiddleBlock->getTerminator());
2592 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2593 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2594 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2597 // Make sure that we found the index where scalar loop needs to continue.
2598 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2599 "Invalid resume Index");
2601 // Add a check in the middle block to see if we have completed
2602 // all of the iterations in the first vector loop.
2603 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2604 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2605 ResumeIndex, "cmp.n",
2606 MiddleBlock->getTerminator());
2608 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2609 // Remove the old terminator.
2610 MiddleBlock->getTerminator()->eraseFromParent();
2612 // Create i+1 and fill the PHINode.
2613 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2614 Induction->addIncoming(StartIdx, VectorPH);
2615 Induction->addIncoming(NextIdx, VecBody);
2616 // Create the compare.
2617 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2618 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2620 // Now we have two terminators. Remove the old one from the block.
2621 VecBody->getTerminator()->eraseFromParent();
2623 // Get ready to start creating new instructions into the vectorized body.
2624 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2627 LoopVectorPreHeader = VectorPH;
2628 LoopScalarPreHeader = ScalarPH;
2629 LoopMiddleBlock = MiddleBlock;
2630 LoopExitBlock = ExitBlock;
2631 LoopVectorBody.push_back(VecBody);
2632 LoopScalarBody = OldBasicBlock;
2634 LoopVectorizeHints Hints(Lp, true);
2635 Hints.setAlreadyVectorized();
2638 /// This function returns the identity element (or neutral element) for
2639 /// the operation K.
2641 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2646 // Adding, Xoring, Oring zero to a number does not change it.
2647 return ConstantInt::get(Tp, 0);
2648 case RK_IntegerMult:
2649 // Multiplying a number by 1 does not change it.
2650 return ConstantInt::get(Tp, 1);
2652 // AND-ing a number with an all-1 value does not change it.
2653 return ConstantInt::get(Tp, -1, true);
2655 // Multiplying a number by 1 does not change it.
2656 return ConstantFP::get(Tp, 1.0L);
2658 // Adding zero to a number does not change it.
2659 return ConstantFP::get(Tp, 0.0L);
2661 llvm_unreachable("Unknown reduction kind");
2665 /// This function translates the reduction kind to an LLVM binary operator.
2667 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2669 case LoopVectorizationLegality::RK_IntegerAdd:
2670 return Instruction::Add;
2671 case LoopVectorizationLegality::RK_IntegerMult:
2672 return Instruction::Mul;
2673 case LoopVectorizationLegality::RK_IntegerOr:
2674 return Instruction::Or;
2675 case LoopVectorizationLegality::RK_IntegerAnd:
2676 return Instruction::And;
2677 case LoopVectorizationLegality::RK_IntegerXor:
2678 return Instruction::Xor;
2679 case LoopVectorizationLegality::RK_FloatMult:
2680 return Instruction::FMul;
2681 case LoopVectorizationLegality::RK_FloatAdd:
2682 return Instruction::FAdd;
2683 case LoopVectorizationLegality::RK_IntegerMinMax:
2684 return Instruction::ICmp;
2685 case LoopVectorizationLegality::RK_FloatMinMax:
2686 return Instruction::FCmp;
2688 llvm_unreachable("Unknown reduction operation");
2692 Value *createMinMaxOp(IRBuilder<> &Builder,
2693 LoopVectorizationLegality::MinMaxReductionKind RK,
2696 CmpInst::Predicate P = CmpInst::ICMP_NE;
2699 llvm_unreachable("Unknown min/max reduction kind");
2700 case LoopVectorizationLegality::MRK_UIntMin:
2701 P = CmpInst::ICMP_ULT;
2703 case LoopVectorizationLegality::MRK_UIntMax:
2704 P = CmpInst::ICMP_UGT;
2706 case LoopVectorizationLegality::MRK_SIntMin:
2707 P = CmpInst::ICMP_SLT;
2709 case LoopVectorizationLegality::MRK_SIntMax:
2710 P = CmpInst::ICMP_SGT;
2712 case LoopVectorizationLegality::MRK_FloatMin:
2713 P = CmpInst::FCMP_OLT;
2715 case LoopVectorizationLegality::MRK_FloatMax:
2716 P = CmpInst::FCMP_OGT;
2721 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2722 RK == LoopVectorizationLegality::MRK_FloatMax)
2723 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2725 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2727 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2732 struct CSEDenseMapInfo {
2733 static bool canHandle(Instruction *I) {
2734 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2735 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2737 static inline Instruction *getEmptyKey() {
2738 return DenseMapInfo<Instruction *>::getEmptyKey();
2740 static inline Instruction *getTombstoneKey() {
2741 return DenseMapInfo<Instruction *>::getTombstoneKey();
2743 static unsigned getHashValue(Instruction *I) {
2744 assert(canHandle(I) && "Unknown instruction!");
2745 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2746 I->value_op_end()));
2748 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2749 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2750 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2752 return LHS->isIdenticalTo(RHS);
2757 /// \brief Check whether this block is a predicated block.
2758 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2759 /// = ...; " blocks. We start with one vectorized basic block. For every
2760 /// conditional block we split this vectorized block. Therefore, every second
2761 /// block will be a predicated one.
2762 static bool isPredicatedBlock(unsigned BlockNum) {
2763 return BlockNum % 2;
2766 ///\brief Perform cse of induction variable instructions.
2767 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2768 // Perform simple cse.
2769 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2770 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2771 BasicBlock *BB = BBs[i];
2772 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2773 Instruction *In = I++;
2775 if (!CSEDenseMapInfo::canHandle(In))
2778 // Check if we can replace this instruction with any of the
2779 // visited instructions.
2780 if (Instruction *V = CSEMap.lookup(In)) {
2781 In->replaceAllUsesWith(V);
2782 In->eraseFromParent();
2785 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2786 // ...;" blocks for predicated stores. Every second block is a predicated
2788 if (isPredicatedBlock(i))
2796 /// \brief Adds a 'fast' flag to floating point operations.
2797 static Value *addFastMathFlag(Value *V) {
2798 if (isa<FPMathOperator>(V)){
2799 FastMathFlags Flags;
2800 Flags.setUnsafeAlgebra();
2801 cast<Instruction>(V)->setFastMathFlags(Flags);
2806 void InnerLoopVectorizer::vectorizeLoop() {
2807 //===------------------------------------------------===//
2809 // Notice: any optimization or new instruction that go
2810 // into the code below should be also be implemented in
2813 //===------------------------------------------------===//
2814 Constant *Zero = Builder.getInt32(0);
2816 // In order to support reduction variables we need to be able to vectorize
2817 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2818 // stages. First, we create a new vector PHI node with no incoming edges.
2819 // We use this value when we vectorize all of the instructions that use the
2820 // PHI. Next, after all of the instructions in the block are complete we
2821 // add the new incoming edges to the PHI. At this point all of the
2822 // instructions in the basic block are vectorized, so we can use them to
2823 // construct the PHI.
2824 PhiVector RdxPHIsToFix;
2826 // Scan the loop in a topological order to ensure that defs are vectorized
2828 LoopBlocksDFS DFS(OrigLoop);
2831 // Vectorize all of the blocks in the original loop.
2832 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2833 be = DFS.endRPO(); bb != be; ++bb)
2834 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2836 // At this point every instruction in the original loop is widened to
2837 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2838 // that we vectorized. The PHI nodes are currently empty because we did
2839 // not want to introduce cycles. Notice that the remaining PHI nodes
2840 // that we need to fix are reduction variables.
2842 // Create the 'reduced' values for each of the induction vars.
2843 // The reduced values are the vector values that we scalarize and combine
2844 // after the loop is finished.
2845 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2847 PHINode *RdxPhi = *it;
2848 assert(RdxPhi && "Unable to recover vectorized PHI");
2850 // Find the reduction variable descriptor.
2851 assert(Legal->getReductionVars()->count(RdxPhi) &&
2852 "Unable to find the reduction variable");
2853 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2854 (*Legal->getReductionVars())[RdxPhi];
2856 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2858 // We need to generate a reduction vector from the incoming scalar.
2859 // To do so, we need to generate the 'identity' vector and override
2860 // one of the elements with the incoming scalar reduction. We need
2861 // to do it in the vector-loop preheader.
2862 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2864 // This is the vector-clone of the value that leaves the loop.
2865 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2866 Type *VecTy = VectorExit[0]->getType();
2868 // Find the reduction identity variable. Zero for addition, or, xor,
2869 // one for multiplication, -1 for And.
2872 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2873 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2874 // MinMax reduction have the start value as their identify.
2876 VectorStart = Identity = RdxDesc.StartValue;
2878 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2883 // Handle other reduction kinds:
2885 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2886 VecTy->getScalarType());
2889 // This vector is the Identity vector where the first element is the
2890 // incoming scalar reduction.
2891 VectorStart = RdxDesc.StartValue;
2893 Identity = ConstantVector::getSplat(VF, Iden);
2895 // This vector is the Identity vector where the first element is the
2896 // incoming scalar reduction.
2897 VectorStart = Builder.CreateInsertElement(Identity,
2898 RdxDesc.StartValue, Zero);
2902 // Fix the vector-loop phi.
2904 // Reductions do not have to start at zero. They can start with
2905 // any loop invariant values.
2906 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2907 BasicBlock *Latch = OrigLoop->getLoopLatch();
2908 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2909 VectorParts &Val = getVectorValue(LoopVal);
2910 for (unsigned part = 0; part < UF; ++part) {
2911 // Make sure to add the reduction stat value only to the
2912 // first unroll part.
2913 Value *StartVal = (part == 0) ? VectorStart : Identity;
2914 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2915 LoopVectorPreHeader);
2916 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2917 LoopVectorBody.back());
2920 // Before each round, move the insertion point right between
2921 // the PHIs and the values we are going to write.
2922 // This allows us to write both PHINodes and the extractelement
2924 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2926 VectorParts RdxParts;
2927 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2928 for (unsigned part = 0; part < UF; ++part) {
2929 // This PHINode contains the vectorized reduction variable, or
2930 // the initial value vector, if we bypass the vector loop.
2931 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2932 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2933 Value *StartVal = (part == 0) ? VectorStart : Identity;
2934 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2935 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2936 NewPhi->addIncoming(RdxExitVal[part],
2937 LoopVectorBody.back());
2938 RdxParts.push_back(NewPhi);
2941 // Reduce all of the unrolled parts into a single vector.
2942 Value *ReducedPartRdx = RdxParts[0];
2943 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2944 setDebugLocFromInst(Builder, ReducedPartRdx);
2945 for (unsigned part = 1; part < UF; ++part) {
2946 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2947 // Floating point operations had to be 'fast' to enable the reduction.
2948 ReducedPartRdx = addFastMathFlag(
2949 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2950 ReducedPartRdx, "bin.rdx"));
2952 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2953 ReducedPartRdx, RdxParts[part]);
2957 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2958 // and vector ops, reducing the set of values being computed by half each
2960 assert(isPowerOf2_32(VF) &&
2961 "Reduction emission only supported for pow2 vectors!");
2962 Value *TmpVec = ReducedPartRdx;
2963 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2964 for (unsigned i = VF; i != 1; i >>= 1) {
2965 // Move the upper half of the vector to the lower half.
2966 for (unsigned j = 0; j != i/2; ++j)
2967 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2969 // Fill the rest of the mask with undef.
2970 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2971 UndefValue::get(Builder.getInt32Ty()));
2974 Builder.CreateShuffleVector(TmpVec,
2975 UndefValue::get(TmpVec->getType()),
2976 ConstantVector::get(ShuffleMask),
2979 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2980 // Floating point operations had to be 'fast' to enable the reduction.
2981 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2982 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2984 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2987 // The result is in the first element of the vector.
2988 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2989 Builder.getInt32(0));
2992 // Create a phi node that merges control-flow from the backedge-taken check
2993 // block and the middle block.
2994 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2995 LoopScalarPreHeader->getTerminator());
2996 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2997 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2999 // Now, we need to fix the users of the reduction variable
3000 // inside and outside of the scalar remainder loop.
3001 // We know that the loop is in LCSSA form. We need to update the
3002 // PHI nodes in the exit blocks.
3003 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3004 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3005 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3006 if (!LCSSAPhi) break;
3008 // All PHINodes need to have a single entry edge, or two if
3009 // we already fixed them.
3010 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3012 // We found our reduction value exit-PHI. Update it with the
3013 // incoming bypass edge.
3014 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
3015 // Add an edge coming from the bypass.
3016 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3019 }// end of the LCSSA phi scan.
3021 // Fix the scalar loop reduction variable with the incoming reduction sum
3022 // from the vector body and from the backedge value.
3023 int IncomingEdgeBlockIdx =
3024 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3025 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3026 // Pick the other block.
3027 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3028 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3029 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
3030 }// end of for each redux variable.
3034 // Remove redundant induction instructions.
3035 cse(LoopVectorBody);
3038 void InnerLoopVectorizer::fixLCSSAPHIs() {
3039 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3040 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3041 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3042 if (!LCSSAPhi) break;
3043 if (LCSSAPhi->getNumIncomingValues() == 1)
3044 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3049 InnerLoopVectorizer::VectorParts
3050 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3051 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3054 // Look for cached value.
3055 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3056 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3057 if (ECEntryIt != MaskCache.end())
3058 return ECEntryIt->second;
3060 VectorParts SrcMask = createBlockInMask(Src);
3062 // The terminator has to be a branch inst!
3063 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3064 assert(BI && "Unexpected terminator found");
3066 if (BI->isConditional()) {
3067 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3069 if (BI->getSuccessor(0) != Dst)
3070 for (unsigned part = 0; part < UF; ++part)
3071 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3073 for (unsigned part = 0; part < UF; ++part)
3074 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3076 MaskCache[Edge] = EdgeMask;
3080 MaskCache[Edge] = SrcMask;
3084 InnerLoopVectorizer::VectorParts
3085 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3086 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3088 // Loop incoming mask is all-one.
3089 if (OrigLoop->getHeader() == BB) {
3090 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3091 return getVectorValue(C);
3094 // This is the block mask. We OR all incoming edges, and with zero.
3095 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3096 VectorParts BlockMask = getVectorValue(Zero);
3099 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3100 VectorParts EM = createEdgeMask(*it, BB);
3101 for (unsigned part = 0; part < UF; ++part)
3102 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3108 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3109 InnerLoopVectorizer::VectorParts &Entry,
3110 unsigned UF, unsigned VF, PhiVector *PV) {
3111 PHINode* P = cast<PHINode>(PN);
3112 // Handle reduction variables:
3113 if (Legal->getReductionVars()->count(P)) {
3114 for (unsigned part = 0; part < UF; ++part) {
3115 // This is phase one of vectorizing PHIs.
3116 Type *VecTy = (VF == 1) ? PN->getType() :
3117 VectorType::get(PN->getType(), VF);
3118 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3119 LoopVectorBody.back()-> getFirstInsertionPt());
3125 setDebugLocFromInst(Builder, P);
3126 // Check for PHI nodes that are lowered to vector selects.
3127 if (P->getParent() != OrigLoop->getHeader()) {
3128 // We know that all PHIs in non-header blocks are converted into
3129 // selects, so we don't have to worry about the insertion order and we
3130 // can just use the builder.
3131 // At this point we generate the predication tree. There may be
3132 // duplications since this is a simple recursive scan, but future
3133 // optimizations will clean it up.
3135 unsigned NumIncoming = P->getNumIncomingValues();
3137 // Generate a sequence of selects of the form:
3138 // SELECT(Mask3, In3,
3139 // SELECT(Mask2, In2,
3141 for (unsigned In = 0; In < NumIncoming; In++) {
3142 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3144 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3146 for (unsigned part = 0; part < UF; ++part) {
3147 // We might have single edge PHIs (blocks) - use an identity
3148 // 'select' for the first PHI operand.
3150 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3153 // Select between the current value and the previous incoming edge
3154 // based on the incoming mask.
3155 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3156 Entry[part], "predphi");
3162 // This PHINode must be an induction variable.
3163 // Make sure that we know about it.
3164 assert(Legal->getInductionVars()->count(P) &&
3165 "Not an induction variable");
3167 LoopVectorizationLegality::InductionInfo II =
3168 Legal->getInductionVars()->lookup(P);
3170 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3171 // which can be found from the original scalar operations.
3173 case LoopVectorizationLegality::IK_NoInduction:
3174 llvm_unreachable("Unknown induction");
3175 case LoopVectorizationLegality::IK_IntInduction: {
3176 assert(P->getType() == II.StartValue->getType() && "Types must match");
3177 Type *PhiTy = P->getType();
3179 if (P == OldInduction) {
3180 // Handle the canonical induction variable. We might have had to
3182 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3184 // Handle other induction variables that are now based on the
3186 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3188 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3189 Broadcasted = II.transform(Builder, NormalizedIdx);
3190 Broadcasted->setName("offset.idx");
3192 Broadcasted = getBroadcastInstrs(Broadcasted);
3193 // After broadcasting the induction variable we need to make the vector
3194 // consecutive by adding 0, 1, 2, etc.
3195 for (unsigned part = 0; part < UF; ++part)
3196 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3199 case LoopVectorizationLegality::IK_PtrInduction:
3200 // Handle the pointer induction variable case.
3201 assert(P->getType()->isPointerTy() && "Unexpected type.");
3202 // This is the normalized GEP that starts counting at zero.
3203 Value *NormalizedIdx =
3204 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3205 // This is the vector of results. Notice that we don't generate
3206 // vector geps because scalar geps result in better code.
3207 for (unsigned part = 0; part < UF; ++part) {
3209 int EltIndex = part;
3210 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3211 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3212 Value *SclrGep = II.transform(Builder, GlobalIdx);
3213 SclrGep->setName("next.gep");
3214 Entry[part] = SclrGep;
3218 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3219 for (unsigned int i = 0; i < VF; ++i) {
3220 int EltIndex = i + part * VF;
3221 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3222 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3223 Value *SclrGep = II.transform(Builder, GlobalIdx);
3224 SclrGep->setName("next.gep");
3225 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3226 Builder.getInt32(i),
3229 Entry[part] = VecVal;
3235 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3236 // For each instruction in the old loop.
3237 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3238 VectorParts &Entry = WidenMap.get(it);
3239 switch (it->getOpcode()) {
3240 case Instruction::Br:
3241 // Nothing to do for PHIs and BR, since we already took care of the
3242 // loop control flow instructions.
3244 case Instruction::PHI: {
3245 // Vectorize PHINodes.
3246 widenPHIInstruction(it, Entry, UF, VF, PV);
3250 case Instruction::Add:
3251 case Instruction::FAdd:
3252 case Instruction::Sub:
3253 case Instruction::FSub:
3254 case Instruction::Mul:
3255 case Instruction::FMul:
3256 case Instruction::UDiv:
3257 case Instruction::SDiv:
3258 case Instruction::FDiv:
3259 case Instruction::URem:
3260 case Instruction::SRem:
3261 case Instruction::FRem:
3262 case Instruction::Shl:
3263 case Instruction::LShr:
3264 case Instruction::AShr:
3265 case Instruction::And:
3266 case Instruction::Or:
3267 case Instruction::Xor: {
3268 // Just widen binops.
3269 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3270 setDebugLocFromInst(Builder, BinOp);
3271 VectorParts &A = getVectorValue(it->getOperand(0));
3272 VectorParts &B = getVectorValue(it->getOperand(1));
3274 // Use this vector value for all users of the original instruction.
3275 for (unsigned Part = 0; Part < UF; ++Part) {
3276 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3278 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3279 VecOp->copyIRFlags(BinOp);
3284 propagateMetadata(Entry, it);
3287 case Instruction::Select: {
3289 // If the selector is loop invariant we can create a select
3290 // instruction with a scalar condition. Otherwise, use vector-select.
3291 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3293 setDebugLocFromInst(Builder, it);
3295 // The condition can be loop invariant but still defined inside the
3296 // loop. This means that we can't just use the original 'cond' value.
3297 // We have to take the 'vectorized' value and pick the first lane.
3298 // Instcombine will make this a no-op.
3299 VectorParts &Cond = getVectorValue(it->getOperand(0));
3300 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3301 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3303 Value *ScalarCond = (VF == 1) ? Cond[0] :
3304 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3306 for (unsigned Part = 0; Part < UF; ++Part) {
3307 Entry[Part] = Builder.CreateSelect(
3308 InvariantCond ? ScalarCond : Cond[Part],
3313 propagateMetadata(Entry, it);
3317 case Instruction::ICmp:
3318 case Instruction::FCmp: {
3319 // Widen compares. Generate vector compares.
3320 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3321 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3322 setDebugLocFromInst(Builder, it);
3323 VectorParts &A = getVectorValue(it->getOperand(0));
3324 VectorParts &B = getVectorValue(it->getOperand(1));
3325 for (unsigned Part = 0; Part < UF; ++Part) {
3328 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3330 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3334 propagateMetadata(Entry, it);
3338 case Instruction::Store:
3339 case Instruction::Load:
3340 vectorizeMemoryInstruction(it);
3342 case Instruction::ZExt:
3343 case Instruction::SExt:
3344 case Instruction::FPToUI:
3345 case Instruction::FPToSI:
3346 case Instruction::FPExt:
3347 case Instruction::PtrToInt:
3348 case Instruction::IntToPtr:
3349 case Instruction::SIToFP:
3350 case Instruction::UIToFP:
3351 case Instruction::Trunc:
3352 case Instruction::FPTrunc:
3353 case Instruction::BitCast: {
3354 CastInst *CI = dyn_cast<CastInst>(it);
3355 setDebugLocFromInst(Builder, it);
3356 /// Optimize the special case where the source is the induction
3357 /// variable. Notice that we can only optimize the 'trunc' case
3358 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3359 /// c. other casts depend on pointer size.
3360 if (CI->getOperand(0) == OldInduction &&
3361 it->getOpcode() == Instruction::Trunc) {
3362 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3364 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3365 LoopVectorizationLegality::InductionInfo II =
3366 Legal->getInductionVars()->lookup(OldInduction);
3368 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3369 for (unsigned Part = 0; Part < UF; ++Part)
3370 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3371 propagateMetadata(Entry, it);
3374 /// Vectorize casts.
3375 Type *DestTy = (VF == 1) ? CI->getType() :
3376 VectorType::get(CI->getType(), VF);
3378 VectorParts &A = getVectorValue(it->getOperand(0));
3379 for (unsigned Part = 0; Part < UF; ++Part)
3380 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3381 propagateMetadata(Entry, it);
3385 case Instruction::Call: {
3386 // Ignore dbg intrinsics.
3387 if (isa<DbgInfoIntrinsic>(it))
3389 setDebugLocFromInst(Builder, it);
3391 Module *M = BB->getParent()->getParent();
3392 CallInst *CI = cast<CallInst>(it);
3393 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3394 assert(ID && "Not an intrinsic call!");
3396 case Intrinsic::assume:
3397 case Intrinsic::lifetime_end:
3398 case Intrinsic::lifetime_start:
3399 scalarizeInstruction(it);
3402 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3403 for (unsigned Part = 0; Part < UF; ++Part) {
3404 SmallVector<Value *, 4> Args;
3405 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3406 if (HasScalarOpd && i == 1) {
3407 Args.push_back(CI->getArgOperand(i));
3410 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3411 Args.push_back(Arg[Part]);
3413 Type *Tys[] = {CI->getType()};
3415 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3417 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3418 Entry[Part] = Builder.CreateCall(F, Args);
3421 propagateMetadata(Entry, it);
3428 // All other instructions are unsupported. Scalarize them.
3429 scalarizeInstruction(it);
3432 }// end of for_each instr.
3435 void InnerLoopVectorizer::updateAnalysis() {
3436 // Forget the original basic block.
3437 SE->forgetLoop(OrigLoop);
3439 // Update the dominator tree information.
3440 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3441 "Entry does not dominate exit.");
3443 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3444 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3445 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3447 // Due to if predication of stores we might create a sequence of "if(pred)
3448 // a[i] = ...; " blocks.
3449 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3451 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3452 else if (isPredicatedBlock(i)) {
3453 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3455 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3459 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3460 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3461 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3462 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3464 DEBUG(DT->verifyDomTree());
3467 /// \brief Check whether it is safe to if-convert this phi node.
3469 /// Phi nodes with constant expressions that can trap are not safe to if
3471 static bool canIfConvertPHINodes(BasicBlock *BB) {
3472 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3473 PHINode *Phi = dyn_cast<PHINode>(I);
3476 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3477 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3484 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3485 if (!EnableIfConversion) {
3486 emitAnalysis(Report() << "if-conversion is disabled");
3490 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3492 // A list of pointers that we can safely read and write to.
3493 SmallPtrSet<Value *, 8> SafePointes;
3495 // Collect safe addresses.
3496 for (Loop::block_iterator BI = TheLoop->block_begin(),
3497 BE = TheLoop->block_end(); BI != BE; ++BI) {
3498 BasicBlock *BB = *BI;
3500 if (blockNeedsPredication(BB))
3503 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3504 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3505 SafePointes.insert(LI->getPointerOperand());
3506 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3507 SafePointes.insert(SI->getPointerOperand());
3511 // Collect the blocks that need predication.
3512 BasicBlock *Header = TheLoop->getHeader();
3513 for (Loop::block_iterator BI = TheLoop->block_begin(),
3514 BE = TheLoop->block_end(); BI != BE; ++BI) {
3515 BasicBlock *BB = *BI;
3517 // We don't support switch statements inside loops.
3518 if (!isa<BranchInst>(BB->getTerminator())) {
3519 emitAnalysis(Report(BB->getTerminator())
3520 << "loop contains a switch statement");
3524 // We must be able to predicate all blocks that need to be predicated.
3525 if (blockNeedsPredication(BB)) {
3526 if (!blockCanBePredicated(BB, SafePointes)) {
3527 emitAnalysis(Report(BB->getTerminator())
3528 << "control flow cannot be substituted for a select");
3531 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3532 emitAnalysis(Report(BB->getTerminator())
3533 << "control flow cannot be substituted for a select");
3538 // We can if-convert this loop.
3542 bool LoopVectorizationLegality::canVectorize() {
3543 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3544 // be canonicalized.
3545 if (!TheLoop->getLoopPreheader()) {
3547 Report() << "loop control flow is not understood by vectorizer");
3551 // We can only vectorize innermost loops.
3552 if (!TheLoop->getSubLoopsVector().empty()) {
3553 emitAnalysis(Report() << "loop is not the innermost loop");
3557 // We must have a single backedge.
3558 if (TheLoop->getNumBackEdges() != 1) {
3560 Report() << "loop control flow is not understood by vectorizer");
3564 // We must have a single exiting block.
3565 if (!TheLoop->getExitingBlock()) {
3567 Report() << "loop control flow is not understood by vectorizer");
3571 // We only handle bottom-tested loops, i.e. loop in which the condition is
3572 // checked at the end of each iteration. With that we can assume that all
3573 // instructions in the loop are executed the same number of times.
3574 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3576 Report() << "loop control flow is not understood by vectorizer");
3580 // We need to have a loop header.
3581 DEBUG(dbgs() << "LV: Found a loop: " <<
3582 TheLoop->getHeader()->getName() << '\n');
3584 // Check if we can if-convert non-single-bb loops.
3585 unsigned NumBlocks = TheLoop->getNumBlocks();
3586 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3587 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3591 // ScalarEvolution needs to be able to find the exit count.
3592 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3593 if (ExitCount == SE->getCouldNotCompute()) {
3594 emitAnalysis(Report() << "could not determine number of loop iterations");
3595 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3599 // Check if we can vectorize the instructions and CFG in this loop.
3600 if (!canVectorizeInstrs()) {
3601 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3605 // Go over each instruction and look at memory deps.
3606 if (!canVectorizeMemory()) {
3607 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3611 // Collect all of the variables that remain uniform after vectorization.
3612 collectLoopUniforms();
3614 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3615 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3618 // Okay! We can vectorize. At this point we don't have any other mem analysis
3619 // which may limit our maximum vectorization factor, so just return true with
3624 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3625 if (Ty->isPointerTy())
3626 return DL.getIntPtrType(Ty);
3628 // It is possible that char's or short's overflow when we ask for the loop's
3629 // trip count, work around this by changing the type size.
3630 if (Ty->getScalarSizeInBits() < 32)
3631 return Type::getInt32Ty(Ty->getContext());
3636 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3637 Ty0 = convertPointerToIntegerType(DL, Ty0);
3638 Ty1 = convertPointerToIntegerType(DL, Ty1);
3639 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3644 /// \brief Check that the instruction has outside loop users and is not an
3645 /// identified reduction variable.
3646 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3647 SmallPtrSetImpl<Value *> &Reductions) {
3648 // Reduction instructions are allowed to have exit users. All other
3649 // instructions must not have external users.
3650 if (!Reductions.count(Inst))
3651 //Check that all of the users of the loop are inside the BB.
3652 for (User *U : Inst->users()) {
3653 Instruction *UI = cast<Instruction>(U);
3654 // This user may be a reduction exit value.
3655 if (!TheLoop->contains(UI)) {
3656 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3663 bool LoopVectorizationLegality::canVectorizeInstrs() {
3664 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3665 BasicBlock *Header = TheLoop->getHeader();
3667 // Look for the attribute signaling the absence of NaNs.
3668 Function &F = *Header->getParent();
3669 if (F.hasFnAttribute("no-nans-fp-math"))
3670 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3671 AttributeSet::FunctionIndex,
3672 "no-nans-fp-math").getValueAsString() == "true";
3674 // For each block in the loop.
3675 for (Loop::block_iterator bb = TheLoop->block_begin(),
3676 be = TheLoop->block_end(); bb != be; ++bb) {
3678 // Scan the instructions in the block and look for hazards.
3679 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3682 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3683 Type *PhiTy = Phi->getType();
3684 // Check that this PHI type is allowed.
3685 if (!PhiTy->isIntegerTy() &&
3686 !PhiTy->isFloatingPointTy() &&
3687 !PhiTy->isPointerTy()) {
3688 emitAnalysis(Report(it)
3689 << "loop control flow is not understood by vectorizer");
3690 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3694 // If this PHINode is not in the header block, then we know that we
3695 // can convert it to select during if-conversion. No need to check if
3696 // the PHIs in this block are induction or reduction variables.
3697 if (*bb != Header) {
3698 // Check that this instruction has no outside users or is an
3699 // identified reduction value with an outside user.
3700 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3702 emitAnalysis(Report(it) << "value could not be identified as "
3703 "an induction or reduction variable");
3707 // We only allow if-converted PHIs with exactly two incoming values.
3708 if (Phi->getNumIncomingValues() != 2) {
3709 emitAnalysis(Report(it)
3710 << "control flow not understood by vectorizer");
3711 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3715 // This is the value coming from the preheader.
3716 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3717 ConstantInt *StepValue = nullptr;
3718 // Check if this is an induction variable.
3719 InductionKind IK = isInductionVariable(Phi, StepValue);
3721 if (IK_NoInduction != IK) {
3722 // Get the widest type.
3724 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3726 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3728 // Int inductions are special because we only allow one IV.
3729 if (IK == IK_IntInduction && StepValue->isOne()) {
3730 // Use the phi node with the widest type as induction. Use the last
3731 // one if there are multiple (no good reason for doing this other
3732 // than it is expedient).
3733 if (!Induction || PhiTy == WidestIndTy)
3737 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3738 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3740 // Until we explicitly handle the case of an induction variable with
3741 // an outside loop user we have to give up vectorizing this loop.
3742 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3743 emitAnalysis(Report(it) << "use of induction value outside of the "
3744 "loop is not handled by vectorizer");
3751 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3752 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3755 if (AddReductionVar(Phi, RK_IntegerMult)) {
3756 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3759 if (AddReductionVar(Phi, RK_IntegerOr)) {
3760 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3763 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3764 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3767 if (AddReductionVar(Phi, RK_IntegerXor)) {
3768 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3771 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3772 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3775 if (AddReductionVar(Phi, RK_FloatMult)) {
3776 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3779 if (AddReductionVar(Phi, RK_FloatAdd)) {
3780 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3783 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3784 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3789 emitAnalysis(Report(it) << "value that could not be identified as "
3790 "reduction is used outside the loop");
3791 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3793 }// end of PHI handling
3795 // We still don't handle functions. However, we can ignore dbg intrinsic
3796 // calls and we do handle certain intrinsic and libm functions.
3797 CallInst *CI = dyn_cast<CallInst>(it);
3798 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3799 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3800 DEBUG(dbgs() << "LV: Found a call site.\n");
3804 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3805 // second argument is the same (i.e. loop invariant)
3807 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3808 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3809 emitAnalysis(Report(it)
3810 << "intrinsic instruction cannot be vectorized");
3811 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3816 // Check that the instruction return type is vectorizable.
3817 // Also, we can't vectorize extractelement instructions.
3818 if ((!VectorType::isValidElementType(it->getType()) &&
3819 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3820 emitAnalysis(Report(it)
3821 << "instruction return type cannot be vectorized");
3822 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3826 // Check that the stored type is vectorizable.
3827 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3828 Type *T = ST->getValueOperand()->getType();
3829 if (!VectorType::isValidElementType(T)) {
3830 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3833 if (EnableMemAccessVersioning)
3834 collectStridedAccess(ST);
3837 if (EnableMemAccessVersioning)
3838 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3839 collectStridedAccess(LI);
3841 // Reduction instructions are allowed to have exit users.
3842 // All other instructions must not have external users.
3843 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3844 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3853 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3854 if (Inductions.empty()) {
3855 emitAnalysis(Report()
3856 << "loop induction variable could not be identified");
3864 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3865 /// return the induction operand of the gep pointer.
3866 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3867 const DataLayout *DL, Loop *Lp) {
3868 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3872 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3874 // Check that all of the gep indices are uniform except for our induction
3876 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3877 if (i != InductionOperand &&
3878 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3880 return GEP->getOperand(InductionOperand);
3883 ///\brief Look for a cast use of the passed value.
3884 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3885 Value *UniqueCast = nullptr;
3886 for (User *U : Ptr->users()) {
3887 CastInst *CI = dyn_cast<CastInst>(U);
3888 if (CI && CI->getType() == Ty) {
3898 ///\brief Get the stride of a pointer access in a loop.
3899 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3900 /// pointer to the Value, or null otherwise.
3901 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3902 const DataLayout *DL, Loop *Lp) {
3903 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3904 if (!PtrTy || PtrTy->isAggregateType())
3907 // Try to remove a gep instruction to make the pointer (actually index at this
3908 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3909 // pointer, otherwise, we are analyzing the index.
3910 Value *OrigPtr = Ptr;
3912 // The size of the pointer access.
3913 int64_t PtrAccessSize = 1;
3915 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3916 const SCEV *V = SE->getSCEV(Ptr);
3920 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3921 V = C->getOperand();
3923 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3927 V = S->getStepRecurrence(*SE);
3931 // Strip off the size of access multiplication if we are still analyzing the
3933 if (OrigPtr == Ptr) {
3934 DL->getTypeAllocSize(PtrTy->getElementType());
3935 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3936 if (M->getOperand(0)->getSCEVType() != scConstant)
3939 const APInt &APStepVal =
3940 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3942 // Huge step value - give up.
3943 if (APStepVal.getBitWidth() > 64)
3946 int64_t StepVal = APStepVal.getSExtValue();
3947 if (PtrAccessSize != StepVal)
3949 V = M->getOperand(1);
3954 Type *StripedOffRecurrenceCast = nullptr;
3955 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3956 StripedOffRecurrenceCast = C->getType();
3957 V = C->getOperand();
3960 // Look for the loop invariant symbolic value.
3961 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3965 Value *Stride = U->getValue();
3966 if (!Lp->isLoopInvariant(Stride))
3969 // If we have stripped off the recurrence cast we have to make sure that we
3970 // return the value that is used in this loop so that we can replace it later.
3971 if (StripedOffRecurrenceCast)
3972 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3977 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3978 Value *Ptr = nullptr;
3979 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3980 Ptr = LI->getPointerOperand();
3981 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3982 Ptr = SI->getPointerOperand();
3986 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3990 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3991 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3992 Strides[Ptr] = Stride;
3993 StrideSet.insert(Stride);
3996 void LoopVectorizationLegality::collectLoopUniforms() {
3997 // We now know that the loop is vectorizable!
3998 // Collect variables that will remain uniform after vectorization.
3999 std::vector<Value*> Worklist;
4000 BasicBlock *Latch = TheLoop->getLoopLatch();
4002 // Start with the conditional branch and walk up the block.
4003 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4005 // Also add all consecutive pointer values; these values will be uniform
4006 // after vectorization (and subsequent cleanup) and, until revectorization is
4007 // supported, all dependencies must also be uniform.
4008 for (Loop::block_iterator B = TheLoop->block_begin(),
4009 BE = TheLoop->block_end(); B != BE; ++B)
4010 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4012 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4013 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4015 while (!Worklist.empty()) {
4016 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4017 Worklist.pop_back();
4019 // Look at instructions inside this loop.
4020 // Stop when reaching PHI nodes.
4021 // TODO: we need to follow values all over the loop, not only in this block.
4022 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4025 // This is a known uniform.
4028 // Insert all operands.
4029 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4034 /// \brief Analyses memory accesses in a loop.
4036 /// Checks whether run time pointer checks are needed and builds sets for data
4037 /// dependence checking.
4038 class AccessAnalysis {
4040 /// \brief Read or write access location.
4041 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4042 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4044 /// \brief Set of potential dependent memory accesses.
4045 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4047 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4048 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4050 /// \brief Register a load and whether it is only read from.
4051 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4052 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4053 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4054 Accesses.insert(MemAccessInfo(Ptr, false));
4056 ReadOnlyPtr.insert(Ptr);
4059 /// \brief Register a store.
4060 void addStore(AliasAnalysis::Location &Loc) {
4061 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4062 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4063 Accesses.insert(MemAccessInfo(Ptr, true));
4066 /// \brief Check whether we can check the pointers at runtime for
4067 /// non-intersection.
4068 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4069 unsigned &NumComparisons, ScalarEvolution *SE,
4070 Loop *TheLoop, ValueToValueMap &Strides,
4071 bool ShouldCheckStride = false);
4073 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4074 /// and builds sets of dependent accesses.
4075 void buildDependenceSets() {
4076 processMemAccesses();
4079 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4081 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4082 void resetDepChecks() { CheckDeps.clear(); }
4084 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4087 typedef SetVector<MemAccessInfo> PtrAccessSet;
4089 /// \brief Go over all memory access and check whether runtime pointer checks
4090 /// are needed /// and build sets of dependency check candidates.
4091 void processMemAccesses();
4093 /// Set of all accesses.
4094 PtrAccessSet Accesses;
4096 /// Set of accesses that need a further dependence check.
4097 MemAccessInfoSet CheckDeps;
4099 /// Set of pointers that are read only.
4100 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4102 const DataLayout *DL;
4104 /// An alias set tracker to partition the access set by underlying object and
4105 //intrinsic property (such as TBAA metadata).
4106 AliasSetTracker AST;
4108 /// Sets of potentially dependent accesses - members of one set share an
4109 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4110 /// dependence check.
4111 DepCandidates &DepCands;
4113 bool IsRTCheckNeeded;
4116 } // end anonymous namespace
4118 /// \brief Check whether a pointer can participate in a runtime bounds check.
4119 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4121 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4122 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4126 return AR->isAffine();
4129 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4130 /// the address space.
4131 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4132 const Loop *Lp, ValueToValueMap &StridesMap);
4134 bool AccessAnalysis::canCheckPtrAtRT(
4135 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4136 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4137 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4138 // Find pointers with computable bounds. We are going to use this information
4139 // to place a runtime bound check.
4140 bool CanDoRT = true;
4142 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4145 // We assign a consecutive id to access from different alias sets.
4146 // Accesses between different groups doesn't need to be checked.
4148 for (auto &AS : AST) {
4149 unsigned NumReadPtrChecks = 0;
4150 unsigned NumWritePtrChecks = 0;
4152 // We assign consecutive id to access from different dependence sets.
4153 // Accesses within the same set don't need a runtime check.
4154 unsigned RunningDepId = 1;
4155 DenseMap<Value *, unsigned> DepSetId;
4158 Value *Ptr = A.getValue();
4159 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4160 MemAccessInfo Access(Ptr, IsWrite);
4163 ++NumWritePtrChecks;
4167 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4168 // When we run after a failing dependency check we have to make sure we
4169 // don't have wrapping pointers.
4170 (!ShouldCheckStride ||
4171 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4172 // The id of the dependence set.
4175 if (IsDepCheckNeeded) {
4176 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4177 unsigned &LeaderId = DepSetId[Leader];
4179 LeaderId = RunningDepId++;
4182 // Each access has its own dependence set.
4183 DepId = RunningDepId++;
4185 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4187 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4193 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4194 NumComparisons += 0; // Only one dependence set.
4196 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4197 NumWritePtrChecks - 1));
4203 // If the pointers that we would use for the bounds comparison have different
4204 // address spaces, assume the values aren't directly comparable, so we can't
4205 // use them for the runtime check. We also have to assume they could
4206 // overlap. In the future there should be metadata for whether address spaces
4208 unsigned NumPointers = RtCheck.Pointers.size();
4209 for (unsigned i = 0; i < NumPointers; ++i) {
4210 for (unsigned j = i + 1; j < NumPointers; ++j) {
4211 // Only need to check pointers between two different dependency sets.
4212 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4214 // Only need to check pointers in the same alias set.
4215 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4218 Value *PtrI = RtCheck.Pointers[i];
4219 Value *PtrJ = RtCheck.Pointers[j];
4221 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4222 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4224 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4225 " different address spaces\n");
4234 void AccessAnalysis::processMemAccesses() {
4235 // We process the set twice: first we process read-write pointers, last we
4236 // process read-only pointers. This allows us to skip dependence tests for
4237 // read-only pointers.
4239 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4240 DEBUG(dbgs() << " AST: "; AST.dump());
4241 DEBUG(dbgs() << "LV: Accesses:\n");
4243 for (auto A : Accesses)
4244 dbgs() << "\t" << *A.getPointer() << " (" <<
4245 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4246 "read-only" : "read")) << ")\n";
4249 // The AliasSetTracker has nicely partitioned our pointers by metadata
4250 // compatibility and potential for underlying-object overlap. As a result, we
4251 // only need to check for potential pointer dependencies within each alias
4253 for (auto &AS : AST) {
4254 // Note that both the alias-set tracker and the alias sets themselves used
4255 // linked lists internally and so the iteration order here is deterministic
4256 // (matching the original instruction order within each set).
4258 bool SetHasWrite = false;
4260 // Map of pointers to last access encountered.
4261 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4262 UnderlyingObjToAccessMap ObjToLastAccess;
4264 // Set of access to check after all writes have been processed.
4265 PtrAccessSet DeferredAccesses;
4267 // Iterate over each alias set twice, once to process read/write pointers,
4268 // and then to process read-only pointers.
4269 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4270 bool UseDeferred = SetIteration > 0;
4271 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4273 for (auto AV : AS) {
4274 Value *Ptr = AV.getValue();
4276 // For a single memory access in AliasSetTracker, Accesses may contain
4277 // both read and write, and they both need to be handled for CheckDeps.
4279 if (AC.getPointer() != Ptr)
4282 bool IsWrite = AC.getInt();
4284 // If we're using the deferred access set, then it contains only
4286 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4287 if (UseDeferred && !IsReadOnlyPtr)
4289 // Otherwise, the pointer must be in the PtrAccessSet, either as a
4291 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4292 S.count(MemAccessInfo(Ptr, false))) &&
4293 "Alias-set pointer not in the access set?");
4295 MemAccessInfo Access(Ptr, IsWrite);
4296 DepCands.insert(Access);
4298 // Memorize read-only pointers for later processing and skip them in
4299 // the first round (they need to be checked after we have seen all
4300 // write pointers). Note: we also mark pointer that are not
4301 // consecutive as "read-only" pointers (so that we check
4302 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
4303 if (!UseDeferred && IsReadOnlyPtr) {
4304 DeferredAccesses.insert(Access);
4308 // If this is a write - check other reads and writes for conflicts. If
4309 // this is a read only check other writes for conflicts (but only if
4310 // there is no other write to the ptr - this is an optimization to
4311 // catch "a[i] = a[i] + " without having to do a dependence check).
4312 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4313 CheckDeps.insert(Access);
4314 IsRTCheckNeeded = true;
4320 // Create sets of pointers connected by a shared alias set and
4321 // underlying object.
4322 typedef SmallVector<Value *, 16> ValueVector;
4323 ValueVector TempObjects;
4324 GetUnderlyingObjects(Ptr, TempObjects, DL);
4325 for (Value *UnderlyingObj : TempObjects) {
4326 UnderlyingObjToAccessMap::iterator Prev =
4327 ObjToLastAccess.find(UnderlyingObj);
4328 if (Prev != ObjToLastAccess.end())
4329 DepCands.unionSets(Access, Prev->second);
4331 ObjToLastAccess[UnderlyingObj] = Access;
4340 /// \brief Checks memory dependences among accesses to the same underlying
4341 /// object to determine whether there vectorization is legal or not (and at
4342 /// which vectorization factor).
4344 /// This class works under the assumption that we already checked that memory
4345 /// locations with different underlying pointers are "must-not alias".
4346 /// We use the ScalarEvolution framework to symbolically evalutate access
4347 /// functions pairs. Since we currently don't restructure the loop we can rely
4348 /// on the program order of memory accesses to determine their safety.
4349 /// At the moment we will only deem accesses as safe for:
4350 /// * A negative constant distance assuming program order.
4352 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4353 /// a[i] = tmp; y = a[i];
4355 /// The latter case is safe because later checks guarantuee that there can't
4356 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4357 /// the same variable: a header phi can only be an induction or a reduction, a
4358 /// reduction can't have a memory sink, an induction can't have a memory
4359 /// source). This is important and must not be violated (or we have to
4360 /// resort to checking for cycles through memory).
4362 /// * A positive constant distance assuming program order that is bigger
4363 /// than the biggest memory access.
4365 /// tmp = a[i] OR b[i] = x
4366 /// a[i+2] = tmp y = b[i+2];
4368 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4370 /// * Zero distances and all accesses have the same size.
4372 class MemoryDepChecker {
4374 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4375 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4377 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4378 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4379 ShouldRetryWithRuntimeCheck(false) {}
4381 /// \brief Register the location (instructions are given increasing numbers)
4382 /// of a write access.
4383 void addAccess(StoreInst *SI) {
4384 Value *Ptr = SI->getPointerOperand();
4385 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4386 InstMap.push_back(SI);
4390 /// \brief Register the location (instructions are given increasing numbers)
4391 /// of a write access.
4392 void addAccess(LoadInst *LI) {
4393 Value *Ptr = LI->getPointerOperand();
4394 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4395 InstMap.push_back(LI);
4399 /// \brief Check whether the dependencies between the accesses are safe.
4401 /// Only checks sets with elements in \p CheckDeps.
4402 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4403 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4405 /// \brief The maximum number of bytes of a vector register we can vectorize
4406 /// the accesses safely with.
4407 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4409 /// \brief In same cases when the dependency check fails we can still
4410 /// vectorize the loop with a dynamic array access check.
4411 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4414 ScalarEvolution *SE;
4415 const DataLayout *DL;
4416 const Loop *InnermostLoop;
4418 /// \brief Maps access locations (ptr, read/write) to program order.
4419 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4421 /// \brief Memory access instructions in program order.
4422 SmallVector<Instruction *, 16> InstMap;
4424 /// \brief The program order index to be used for the next instruction.
4427 // We can access this many bytes in parallel safely.
4428 unsigned MaxSafeDepDistBytes;
4430 /// \brief If we see a non-constant dependence distance we can still try to
4431 /// vectorize this loop with runtime checks.
4432 bool ShouldRetryWithRuntimeCheck;
4434 /// \brief Check whether there is a plausible dependence between the two
4437 /// Access \p A must happen before \p B in program order. The two indices
4438 /// identify the index into the program order map.
4440 /// This function checks whether there is a plausible dependence (or the
4441 /// absence of such can't be proved) between the two accesses. If there is a
4442 /// plausible dependence but the dependence distance is bigger than one
4443 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4444 /// distance is smaller than any other distance encountered so far).
4445 /// Otherwise, this function returns true signaling a possible dependence.
4446 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4447 const MemAccessInfo &B, unsigned BIdx,
4448 ValueToValueMap &Strides);
4450 /// \brief Check whether the data dependence could prevent store-load
4452 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4455 } // end anonymous namespace
4457 static bool isInBoundsGep(Value *Ptr) {
4458 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4459 return GEP->isInBounds();
4463 /// \brief Check whether the access through \p Ptr has a constant stride.
4464 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4465 const Loop *Lp, ValueToValueMap &StridesMap) {
4466 const Type *Ty = Ptr->getType();
4467 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4469 // Make sure that the pointer does not point to aggregate types.
4470 const PointerType *PtrTy = cast<PointerType>(Ty);
4471 if (PtrTy->getElementType()->isAggregateType()) {
4472 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4477 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4479 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4481 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4482 << *Ptr << " SCEV: " << *PtrScev << "\n");
4486 // The accesss function must stride over the innermost loop.
4487 if (Lp != AR->getLoop()) {
4488 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4489 *Ptr << " SCEV: " << *PtrScev << "\n");
4492 // The address calculation must not wrap. Otherwise, a dependence could be
4494 // An inbounds getelementptr that is a AddRec with a unit stride
4495 // cannot wrap per definition. The unit stride requirement is checked later.
4496 // An getelementptr without an inbounds attribute and unit stride would have
4497 // to access the pointer value "0" which is undefined behavior in address
4498 // space 0, therefore we can also vectorize this case.
4499 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4500 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4501 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4502 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4503 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4504 << *Ptr << " SCEV: " << *PtrScev << "\n");
4508 // Check the step is constant.
4509 const SCEV *Step = AR->getStepRecurrence(*SE);
4511 // Calculate the pointer stride and check if it is consecutive.
4512 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4514 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4515 " SCEV: " << *PtrScev << "\n");
4519 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4520 const APInt &APStepVal = C->getValue()->getValue();
4522 // Huge step value - give up.
4523 if (APStepVal.getBitWidth() > 64)
4526 int64_t StepVal = APStepVal.getSExtValue();
4529 int64_t Stride = StepVal / Size;
4530 int64_t Rem = StepVal % Size;
4534 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4535 // know we can't "wrap around the address space". In case of address space
4536 // zero we know that this won't happen without triggering undefined behavior.
4537 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4538 Stride != 1 && Stride != -1)
4544 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4545 unsigned TypeByteSize) {
4546 // If loads occur at a distance that is not a multiple of a feasible vector
4547 // factor store-load forwarding does not take place.
4548 // Positive dependences might cause troubles because vectorizing them might
4549 // prevent store-load forwarding making vectorized code run a lot slower.
4550 // a[i] = a[i-3] ^ a[i-8];
4551 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4552 // hence on your typical architecture store-load forwarding does not take
4553 // place. Vectorizing in such cases does not make sense.
4554 // Store-load forwarding distance.
4555 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4556 // Maximum vector factor.
4557 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4558 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4559 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4561 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4563 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4564 MaxVFWithoutSLForwardIssues = (vf >>=1);
4569 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4570 DEBUG(dbgs() << "LV: Distance " << Distance <<
4571 " that could cause a store-load forwarding conflict\n");
4575 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4576 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4577 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4581 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4582 const MemAccessInfo &B, unsigned BIdx,
4583 ValueToValueMap &Strides) {
4584 assert (AIdx < BIdx && "Must pass arguments in program order");
4586 Value *APtr = A.getPointer();
4587 Value *BPtr = B.getPointer();
4588 bool AIsWrite = A.getInt();
4589 bool BIsWrite = B.getInt();
4591 // Two reads are independent.
4592 if (!AIsWrite && !BIsWrite)
4595 // We cannot check pointers in different address spaces.
4596 if (APtr->getType()->getPointerAddressSpace() !=
4597 BPtr->getType()->getPointerAddressSpace())
4600 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4601 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4603 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4604 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4606 const SCEV *Src = AScev;
4607 const SCEV *Sink = BScev;
4609 // If the induction step is negative we have to invert source and sink of the
4611 if (StrideAPtr < 0) {
4614 std::swap(APtr, BPtr);
4615 std::swap(Src, Sink);
4616 std::swap(AIsWrite, BIsWrite);
4617 std::swap(AIdx, BIdx);
4618 std::swap(StrideAPtr, StrideBPtr);
4621 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4623 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4624 << "(Induction step: " << StrideAPtr << ")\n");
4625 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4626 << *InstMap[BIdx] << ": " << *Dist << "\n");
4628 // Need consecutive accesses. We don't want to vectorize
4629 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4630 // the address space.
4631 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4632 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4636 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4638 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4639 ShouldRetryWithRuntimeCheck = true;
4643 Type *ATy = APtr->getType()->getPointerElementType();
4644 Type *BTy = BPtr->getType()->getPointerElementType();
4645 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4647 // Negative distances are not plausible dependencies.
4648 const APInt &Val = C->getValue()->getValue();
4649 if (Val.isNegative()) {
4650 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4651 if (IsTrueDataDependence &&
4652 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4656 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4660 // Write to the same location with the same size.
4661 // Could be improved to assert type sizes are the same (i32 == float, etc).
4665 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4669 assert(Val.isStrictlyPositive() && "Expect a positive value");
4671 // Positive distance bigger than max vectorization factor.
4674 "LV: ReadWrite-Write positive dependency with different types\n");
4678 unsigned Distance = (unsigned) Val.getZExtValue();
4680 // Bail out early if passed-in parameters make vectorization not feasible.
4681 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4682 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4684 // The distance must be bigger than the size needed for a vectorized version
4685 // of the operation and the size of the vectorized operation must not be
4686 // bigger than the currrent maximum size.
4687 if (Distance < 2*TypeByteSize ||
4688 2*TypeByteSize > MaxSafeDepDistBytes ||
4689 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4690 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4691 << Val.getSExtValue() << '\n');
4695 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4696 Distance : MaxSafeDepDistBytes;
4698 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4699 if (IsTrueDataDependence &&
4700 couldPreventStoreLoadForward(Distance, TypeByteSize))
4703 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4704 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4709 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4710 MemAccessInfoSet &CheckDeps,
4711 ValueToValueMap &Strides) {
4713 MaxSafeDepDistBytes = -1U;
4714 while (!CheckDeps.empty()) {
4715 MemAccessInfo CurAccess = *CheckDeps.begin();
4717 // Get the relevant memory access set.
4718 EquivalenceClasses<MemAccessInfo>::iterator I =
4719 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4721 // Check accesses within this set.
4722 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4723 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4725 // Check every access pair.
4727 CheckDeps.erase(*AI);
4728 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4730 // Check every accessing instruction pair in program order.
4731 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4732 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4733 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4734 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4735 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4737 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4748 bool LoopVectorizationLegality::canVectorizeMemory() {
4750 typedef SmallVector<Value*, 16> ValueVector;
4751 typedef SmallPtrSet<Value*, 16> ValueSet;
4753 // Holds the Load and Store *instructions*.
4757 // Holds all the different accesses in the loop.
4758 unsigned NumReads = 0;
4759 unsigned NumReadWrites = 0;
4761 PtrRtCheck.Pointers.clear();
4762 PtrRtCheck.Need = false;
4764 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4765 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4768 for (Loop::block_iterator bb = TheLoop->block_begin(),
4769 be = TheLoop->block_end(); bb != be; ++bb) {
4771 // Scan the BB and collect legal loads and stores.
4772 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4775 // If this is a load, save it. If this instruction can read from memory
4776 // but is not a load, then we quit. Notice that we don't handle function
4777 // calls that read or write.
4778 if (it->mayReadFromMemory()) {
4779 // Many math library functions read the rounding mode. We will only
4780 // vectorize a loop if it contains known function calls that don't set
4781 // the flag. Therefore, it is safe to ignore this read from memory.
4782 CallInst *Call = dyn_cast<CallInst>(it);
4783 if (Call && getIntrinsicIDForCall(Call, TLI))
4786 LoadInst *Ld = dyn_cast<LoadInst>(it);
4787 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4788 emitAnalysis(Report(Ld)
4789 << "read with atomic ordering or volatile read");
4790 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4794 Loads.push_back(Ld);
4795 DepChecker.addAccess(Ld);
4799 // Save 'store' instructions. Abort if other instructions write to memory.
4800 if (it->mayWriteToMemory()) {
4801 StoreInst *St = dyn_cast<StoreInst>(it);
4803 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4806 if (!St->isSimple() && !IsAnnotatedParallel) {
4807 emitAnalysis(Report(St)
4808 << "write with atomic ordering or volatile write");
4809 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4813 Stores.push_back(St);
4814 DepChecker.addAccess(St);
4819 // Now we have two lists that hold the loads and the stores.
4820 // Next, we find the pointers that they use.
4822 // Check if we see any stores. If there are no stores, then we don't
4823 // care if the pointers are *restrict*.
4824 if (!Stores.size()) {
4825 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4829 AccessAnalysis::DepCandidates DependentAccesses;
4830 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4832 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4833 // multiple times on the same object. If the ptr is accessed twice, once
4834 // for read and once for write, it will only appear once (on the write
4835 // list). This is okay, since we are going to check for conflicts between
4836 // writes and between reads and writes, but not between reads and reads.
4839 ValueVector::iterator I, IE;
4840 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4841 StoreInst *ST = cast<StoreInst>(*I);
4842 Value* Ptr = ST->getPointerOperand();
4844 if (isUniform(Ptr)) {
4847 << "write to a loop invariant address could not be vectorized");
4848 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4852 // If we did *not* see this pointer before, insert it to the read-write
4853 // list. At this phase it is only a 'write' list.
4854 if (Seen.insert(Ptr).second) {
4857 AliasAnalysis::Location Loc = AA->getLocation(ST);
4858 // The TBAA metadata could have a control dependency on the predication
4859 // condition, so we cannot rely on it when determining whether or not we
4860 // need runtime pointer checks.
4861 if (blockNeedsPredication(ST->getParent()))
4862 Loc.AATags.TBAA = nullptr;
4864 Accesses.addStore(Loc);
4868 if (IsAnnotatedParallel) {
4870 << "LV: A loop annotated parallel, ignore memory dependency "
4875 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4876 LoadInst *LD = cast<LoadInst>(*I);
4877 Value* Ptr = LD->getPointerOperand();
4878 // If we did *not* see this pointer before, insert it to the
4879 // read list. If we *did* see it before, then it is already in
4880 // the read-write list. This allows us to vectorize expressions
4881 // such as A[i] += x; Because the address of A[i] is a read-write
4882 // pointer. This only works if the index of A[i] is consecutive.
4883 // If the address of i is unknown (for example A[B[i]]) then we may
4884 // read a few words, modify, and write a few words, and some of the
4885 // words may be written to the same address.
4886 bool IsReadOnlyPtr = false;
4887 if (Seen.insert(Ptr).second ||
4888 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4890 IsReadOnlyPtr = true;
4893 AliasAnalysis::Location Loc = AA->getLocation(LD);
4894 // The TBAA metadata could have a control dependency on the predication
4895 // condition, so we cannot rely on it when determining whether or not we
4896 // need runtime pointer checks.
4897 if (blockNeedsPredication(LD->getParent()))
4898 Loc.AATags.TBAA = nullptr;
4900 Accesses.addLoad(Loc, IsReadOnlyPtr);
4903 // If we write (or read-write) to a single destination and there are no
4904 // other reads in this loop then is it safe to vectorize.
4905 if (NumReadWrites == 1 && NumReads == 0) {
4906 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4910 // Build dependence sets and check whether we need a runtime pointer bounds
4912 Accesses.buildDependenceSets();
4913 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4915 // Find pointers with computable bounds. We are going to use this information
4916 // to place a runtime bound check.
4917 unsigned NumComparisons = 0;
4918 bool CanDoRT = false;
4920 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4923 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4924 " pointer comparisons.\n");
4926 // If we only have one set of dependences to check pointers among we don't
4927 // need a runtime check.
4928 if (NumComparisons == 0 && NeedRTCheck)
4929 NeedRTCheck = false;
4931 // Check that we did not collect too many pointers or found an unsizeable
4933 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4939 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4942 if (NeedRTCheck && !CanDoRT) {
4943 emitAnalysis(Report() << "cannot identify array bounds");
4944 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4945 "the array bounds.\n");
4950 PtrRtCheck.Need = NeedRTCheck;
4952 bool CanVecMem = true;
4953 if (Accesses.isDependencyCheckNeeded()) {
4954 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4955 CanVecMem = DepChecker.areDepsSafe(
4956 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4957 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4959 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4960 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4963 // Clear the dependency checks. We assume they are not needed.
4964 Accesses.resetDepChecks();
4967 PtrRtCheck.Need = true;
4969 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4970 TheLoop, Strides, true);
4971 // Check that we did not collect too many pointers or found an unsizeable
4973 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4974 if (!CanDoRT && NumComparisons > 0)
4975 emitAnalysis(Report()
4976 << "cannot check memory dependencies at runtime");
4978 emitAnalysis(Report()
4979 << NumComparisons << " exceeds limit of "
4980 << RuntimeMemoryCheckThreshold
4981 << " dependent memory operations checked at runtime");
4982 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4992 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4994 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4995 " need a runtime memory check.\n");
5000 static bool hasMultipleUsesOf(Instruction *I,
5001 SmallPtrSetImpl<Instruction *> &Insts) {
5002 unsigned NumUses = 0;
5003 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
5004 if (Insts.count(dyn_cast<Instruction>(*Use)))
5013 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
5014 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
5015 if (!Set.count(dyn_cast<Instruction>(*Use)))
5020 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
5021 ReductionKind Kind) {
5022 if (Phi->getNumIncomingValues() != 2)
5025 // Reduction variables are only found in the loop header block.
5026 if (Phi->getParent() != TheLoop->getHeader())
5029 // Obtain the reduction start value from the value that comes from the loop
5031 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
5033 // ExitInstruction is the single value which is used outside the loop.
5034 // We only allow for a single reduction value to be used outside the loop.
5035 // This includes users of the reduction, variables (which form a cycle
5036 // which ends in the phi node).
5037 Instruction *ExitInstruction = nullptr;
5038 // Indicates that we found a reduction operation in our scan.
5039 bool FoundReduxOp = false;
5041 // We start with the PHI node and scan for all of the users of this
5042 // instruction. All users must be instructions that can be used as reduction
5043 // variables (such as ADD). We must have a single out-of-block user. The cycle
5044 // must include the original PHI.
5045 bool FoundStartPHI = false;
5047 // To recognize min/max patterns formed by a icmp select sequence, we store
5048 // the number of instruction we saw from the recognized min/max pattern,
5049 // to make sure we only see exactly the two instructions.
5050 unsigned NumCmpSelectPatternInst = 0;
5051 ReductionInstDesc ReduxDesc(false, nullptr);
5053 SmallPtrSet<Instruction *, 8> VisitedInsts;
5054 SmallVector<Instruction *, 8> Worklist;
5055 Worklist.push_back(Phi);
5056 VisitedInsts.insert(Phi);
5058 // A value in the reduction can be used:
5059 // - By the reduction:
5060 // - Reduction operation:
5061 // - One use of reduction value (safe).
5062 // - Multiple use of reduction value (not safe).
5064 // - All uses of the PHI must be the reduction (safe).
5065 // - Otherwise, not safe.
5066 // - By one instruction outside of the loop (safe).
5067 // - By further instructions outside of the loop (not safe).
5068 // - By an instruction that is not part of the reduction (not safe).
5070 // * An instruction type other than PHI or the reduction operation.
5071 // * A PHI in the header other than the initial PHI.
5072 while (!Worklist.empty()) {
5073 Instruction *Cur = Worklist.back();
5074 Worklist.pop_back();
5077 // If the instruction has no users then this is a broken chain and can't be
5078 // a reduction variable.
5079 if (Cur->use_empty())
5082 bool IsAPhi = isa<PHINode>(Cur);
5084 // A header PHI use other than the original PHI.
5085 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5088 // Reductions of instructions such as Div, and Sub is only possible if the
5089 // LHS is the reduction variable.
5090 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5091 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5092 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5095 // Any reduction instruction must be of one of the allowed kinds.
5096 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5097 if (!ReduxDesc.IsReduction)
5100 // A reduction operation must only have one use of the reduction value.
5101 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5102 hasMultipleUsesOf(Cur, VisitedInsts))
5105 // All inputs to a PHI node must be a reduction value.
5106 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5109 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5110 isa<SelectInst>(Cur)))
5111 ++NumCmpSelectPatternInst;
5112 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5113 isa<SelectInst>(Cur)))
5114 ++NumCmpSelectPatternInst;
5116 // Check whether we found a reduction operator.
5117 FoundReduxOp |= !IsAPhi;
5119 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5120 // onto the stack. This way we are going to have seen all inputs to PHI
5121 // nodes once we get to them.
5122 SmallVector<Instruction *, 8> NonPHIs;
5123 SmallVector<Instruction *, 8> PHIs;
5124 for (User *U : Cur->users()) {
5125 Instruction *UI = cast<Instruction>(U);
5127 // Check if we found the exit user.
5128 BasicBlock *Parent = UI->getParent();
5129 if (!TheLoop->contains(Parent)) {
5130 // Exit if you find multiple outside users or if the header phi node is
5131 // being used. In this case the user uses the value of the previous
5132 // iteration, in which case we would loose "VF-1" iterations of the
5133 // reduction operation if we vectorize.
5134 if (ExitInstruction != nullptr || Cur == Phi)
5137 // The instruction used by an outside user must be the last instruction
5138 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5139 // operations on the value.
5140 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5143 ExitInstruction = Cur;
5147 // Process instructions only once (termination). Each reduction cycle
5148 // value must only be used once, except by phi nodes and min/max
5149 // reductions which are represented as a cmp followed by a select.
5150 ReductionInstDesc IgnoredVal(false, nullptr);
5151 if (VisitedInsts.insert(UI).second) {
5152 if (isa<PHINode>(UI))
5155 NonPHIs.push_back(UI);
5156 } else if (!isa<PHINode>(UI) &&
5157 ((!isa<FCmpInst>(UI) &&
5158 !isa<ICmpInst>(UI) &&
5159 !isa<SelectInst>(UI)) ||
5160 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5163 // Remember that we completed the cycle.
5165 FoundStartPHI = true;
5167 Worklist.append(PHIs.begin(), PHIs.end());
5168 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5171 // This means we have seen one but not the other instruction of the
5172 // pattern or more than just a select and cmp.
5173 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5174 NumCmpSelectPatternInst != 2)
5177 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5180 // We found a reduction var if we have reached the original phi node and we
5181 // only have a single instruction with out-of-loop users.
5183 // This instruction is allowed to have out-of-loop users.
5184 AllowedExit.insert(ExitInstruction);
5186 // Save the description of this reduction variable.
5187 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5188 ReduxDesc.MinMaxKind);
5189 Reductions[Phi] = RD;
5190 // We've ended the cycle. This is a reduction variable if we have an
5191 // outside user and it has a binary op.
5196 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5197 /// pattern corresponding to a min(X, Y) or max(X, Y).
5198 LoopVectorizationLegality::ReductionInstDesc
5199 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5200 ReductionInstDesc &Prev) {
5202 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5203 "Expect a select instruction");
5204 Instruction *Cmp = nullptr;
5205 SelectInst *Select = nullptr;
5207 // We must handle the select(cmp()) as a single instruction. Advance to the
5209 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5210 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5211 return ReductionInstDesc(false, I);
5212 return ReductionInstDesc(Select, Prev.MinMaxKind);
5215 // Only handle single use cases for now.
5216 if (!(Select = dyn_cast<SelectInst>(I)))
5217 return ReductionInstDesc(false, I);
5218 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5219 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5220 return ReductionInstDesc(false, I);
5221 if (!Cmp->hasOneUse())
5222 return ReductionInstDesc(false, I);
5227 // Look for a min/max pattern.
5228 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5229 return ReductionInstDesc(Select, MRK_UIntMin);
5230 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5231 return ReductionInstDesc(Select, MRK_UIntMax);
5232 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5233 return ReductionInstDesc(Select, MRK_SIntMax);
5234 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5235 return ReductionInstDesc(Select, MRK_SIntMin);
5236 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5237 return ReductionInstDesc(Select, MRK_FloatMin);
5238 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5239 return ReductionInstDesc(Select, MRK_FloatMax);
5240 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5241 return ReductionInstDesc(Select, MRK_FloatMin);
5242 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5243 return ReductionInstDesc(Select, MRK_FloatMax);
5245 return ReductionInstDesc(false, I);
5248 LoopVectorizationLegality::ReductionInstDesc
5249 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5251 ReductionInstDesc &Prev) {
5252 bool FP = I->getType()->isFloatingPointTy();
5253 bool FastMath = FP && I->hasUnsafeAlgebra();
5254 switch (I->getOpcode()) {
5256 return ReductionInstDesc(false, I);
5257 case Instruction::PHI:
5258 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5259 Kind != RK_FloatMinMax))
5260 return ReductionInstDesc(false, I);
5261 return ReductionInstDesc(I, Prev.MinMaxKind);
5262 case Instruction::Sub:
5263 case Instruction::Add:
5264 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5265 case Instruction::Mul:
5266 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5267 case Instruction::And:
5268 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5269 case Instruction::Or:
5270 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5271 case Instruction::Xor:
5272 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5273 case Instruction::FMul:
5274 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5275 case Instruction::FSub:
5276 case Instruction::FAdd:
5277 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5278 case Instruction::FCmp:
5279 case Instruction::ICmp:
5280 case Instruction::Select:
5281 if (Kind != RK_IntegerMinMax &&
5282 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5283 return ReductionInstDesc(false, I);
5284 return isMinMaxSelectCmpPattern(I, Prev);
5288 LoopVectorizationLegality::InductionKind
5289 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
5290 ConstantInt *&StepValue) {
5291 Type *PhiTy = Phi->getType();
5292 // We only handle integer and pointer inductions variables.
5293 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5294 return IK_NoInduction;
5296 // Check that the PHI is consecutive.
5297 const SCEV *PhiScev = SE->getSCEV(Phi);
5298 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5300 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5301 return IK_NoInduction;
5304 const SCEV *Step = AR->getStepRecurrence(*SE);
5305 // Calculate the pointer stride and check if it is consecutive.
5306 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5308 return IK_NoInduction;
5310 ConstantInt *CV = C->getValue();
5311 if (PhiTy->isIntegerTy()) {
5313 return IK_IntInduction;
5316 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5317 Type *PointerElementType = PhiTy->getPointerElementType();
5318 // The pointer stride cannot be determined if the pointer element type is not
5320 if (!PointerElementType->isSized())
5321 return IK_NoInduction;
5323 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
5324 int64_t CVSize = CV->getSExtValue();
5326 return IK_NoInduction;
5327 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
5328 return IK_PtrInduction;
5331 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5332 Value *In0 = const_cast<Value*>(V);
5333 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5337 return Inductions.count(PN);
5340 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5341 assert(TheLoop->contains(BB) && "Unknown block used");
5343 // Blocks that do not dominate the latch need predication.
5344 BasicBlock* Latch = TheLoop->getLoopLatch();
5345 return !DT->dominates(BB, Latch);
5348 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5349 SmallPtrSetImpl<Value *> &SafePtrs) {
5351 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5352 // Check that we don't have a constant expression that can trap as operand.
5353 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5355 if (Constant *C = dyn_cast<Constant>(*OI))
5359 // We might be able to hoist the load.
5360 if (it->mayReadFromMemory()) {
5361 LoadInst *LI = dyn_cast<LoadInst>(it);
5364 if (!SafePtrs.count(LI->getPointerOperand())) {
5365 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
5366 MaskedOp.insert(LI);
5373 // We don't predicate stores at the moment.
5374 if (it->mayWriteToMemory()) {
5375 StoreInst *SI = dyn_cast<StoreInst>(it);
5376 // We only support predication of stores in basic blocks with one
5381 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5382 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5384 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5385 !isSinglePredecessor) {
5386 // Build a masked store if it is legal for the target, otherwise scalarize
5388 bool isLegalMaskedOp =
5389 isLegalMaskedStore(SI->getValueOperand()->getType(),
5390 SI->getPointerOperand());
5391 if (isLegalMaskedOp) {
5393 MaskedOp.insert(SI);
5402 // The instructions below can trap.
5403 switch (it->getOpcode()) {
5405 case Instruction::UDiv:
5406 case Instruction::SDiv:
5407 case Instruction::URem:
5408 case Instruction::SRem:
5416 LoopVectorizationCostModel::VectorizationFactor
5417 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5418 // Width 1 means no vectorize
5419 VectorizationFactor Factor = { 1U, 0U };
5420 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5421 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5422 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5426 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5427 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5428 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5432 // Find the trip count.
5433 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5434 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5436 unsigned WidestType = getWidestType();
5437 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5438 unsigned MaxSafeDepDist = -1U;
5439 if (Legal->getMaxSafeDepDistBytes() != -1U)
5440 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5441 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5442 WidestRegister : MaxSafeDepDist);
5443 unsigned MaxVectorSize = WidestRegister / WidestType;
5444 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5445 DEBUG(dbgs() << "LV: The Widest register is: "
5446 << WidestRegister << " bits.\n");
5448 if (MaxVectorSize == 0) {
5449 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5453 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5454 " into one vector!");
5456 unsigned VF = MaxVectorSize;
5458 // If we optimize the program for size, avoid creating the tail loop.
5460 // If we are unable to calculate the trip count then don't try to vectorize.
5462 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5463 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5467 // Find the maximum SIMD width that can fit within the trip count.
5468 VF = TC % MaxVectorSize;
5473 // If the trip count that we found modulo the vectorization factor is not
5474 // zero then we require a tail.
5476 emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
5477 "same time. Enable vectorization of this loop "
5478 "with '#pragma clang loop vectorize(enable)' "
5479 "when compiling with -Os");
5480 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5485 int UserVF = Hints->getWidth();
5487 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5488 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5490 Factor.Width = UserVF;
5494 float Cost = expectedCost(1);
5496 const float ScalarCost = Cost;
5499 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5501 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5502 // Ignore scalar width, because the user explicitly wants vectorization.
5503 if (ForceVectorization && VF > 1) {
5505 Cost = expectedCost(Width) / (float)Width;
5508 for (unsigned i=2; i <= VF; i*=2) {
5509 // Notice that the vector loop needs to be executed less times, so
5510 // we need to divide the cost of the vector loops by the width of
5511 // the vector elements.
5512 float VectorCost = expectedCost(i) / (float)i;
5513 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5514 (int)VectorCost << ".\n");
5515 if (VectorCost < Cost) {
5521 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5522 << "LV: Vectorization seems to be not beneficial, "
5523 << "but was forced by a user.\n");
5524 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5525 Factor.Width = Width;
5526 Factor.Cost = Width * Cost;
5530 unsigned LoopVectorizationCostModel::getWidestType() {
5531 unsigned MaxWidth = 8;
5534 for (Loop::block_iterator bb = TheLoop->block_begin(),
5535 be = TheLoop->block_end(); bb != be; ++bb) {
5536 BasicBlock *BB = *bb;
5538 // For each instruction in the loop.
5539 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5540 Type *T = it->getType();
5542 // Ignore ephemeral values.
5543 if (EphValues.count(it))
5546 // Only examine Loads, Stores and PHINodes.
5547 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5550 // Examine PHI nodes that are reduction variables.
5551 if (PHINode *PN = dyn_cast<PHINode>(it))
5552 if (!Legal->getReductionVars()->count(PN))
5555 // Examine the stored values.
5556 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5557 T = ST->getValueOperand()->getType();
5559 // Ignore loaded pointer types and stored pointer types that are not
5560 // consecutive. However, we do want to take consecutive stores/loads of
5561 // pointer vectors into account.
5562 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5565 MaxWidth = std::max(MaxWidth,
5566 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5574 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5576 unsigned LoopCost) {
5578 // -- The unroll heuristics --
5579 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5580 // There are many micro-architectural considerations that we can't predict
5581 // at this level. For example, frontend pressure (on decode or fetch) due to
5582 // code size, or the number and capabilities of the execution ports.
5584 // We use the following heuristics to select the unroll factor:
5585 // 1. If the code has reductions, then we unroll in order to break the cross
5586 // iteration dependency.
5587 // 2. If the loop is really small, then we unroll in order to reduce the loop
5589 // 3. We don't unroll if we think that we will spill registers to memory due
5590 // to the increased register pressure.
5592 // Use the user preference, unless 'auto' is selected.
5593 int UserUF = Hints->getInterleave();
5597 // When we optimize for size, we don't unroll.
5601 // We used the distance for the unroll factor.
5602 if (Legal->getMaxSafeDepDistBytes() != -1U)
5605 // Do not unroll loops with a relatively small trip count.
5606 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5607 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5610 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5611 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5615 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5616 TargetNumRegisters = ForceTargetNumScalarRegs;
5618 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5619 TargetNumRegisters = ForceTargetNumVectorRegs;
5622 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5623 // We divide by these constants so assume that we have at least one
5624 // instruction that uses at least one register.
5625 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5626 R.NumInstructions = std::max(R.NumInstructions, 1U);
5628 // We calculate the unroll factor using the following formula.
5629 // Subtract the number of loop invariants from the number of available
5630 // registers. These registers are used by all of the unrolled instances.
5631 // Next, divide the remaining registers by the number of registers that is
5632 // required by the loop, in order to estimate how many parallel instances
5633 // fit without causing spills. All of this is rounded down if necessary to be
5634 // a power of two. We want power of two unroll factors to simplify any
5635 // addressing operations or alignment considerations.
5636 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5639 // Don't count the induction variable as unrolled.
5640 if (EnableIndVarRegisterHeur)
5641 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5642 std::max(1U, (R.MaxLocalUsers - 1)));
5644 // Clamp the unroll factor ranges to reasonable factors.
5645 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5647 // Check if the user has overridden the unroll max.
5649 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5650 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5652 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5653 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5656 // If we did not calculate the cost for VF (because the user selected the VF)
5657 // then we calculate the cost of VF here.
5659 LoopCost = expectedCost(VF);
5661 // Clamp the calculated UF to be between the 1 and the max unroll factor
5662 // that the target allows.
5663 if (UF > MaxInterleaveSize)
5664 UF = MaxInterleaveSize;
5668 // Unroll if we vectorized this loop and there is a reduction that could
5669 // benefit from unrolling.
5670 if (VF > 1 && Legal->getReductionVars()->size()) {
5671 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5675 // Note that if we've already vectorized the loop we will have done the
5676 // runtime check and so unrolling won't require further checks.
5677 bool UnrollingRequiresRuntimePointerCheck =
5678 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5680 // We want to unroll small loops in order to reduce the loop overhead and
5681 // potentially expose ILP opportunities.
5682 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5683 if (!UnrollingRequiresRuntimePointerCheck &&
5684 LoopCost < SmallLoopCost) {
5685 // We assume that the cost overhead is 1 and we use the cost model
5686 // to estimate the cost of the loop and unroll until the cost of the
5687 // loop overhead is about 5% of the cost of the loop.
5688 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5690 // Unroll until store/load ports (estimated by max unroll factor) are
5692 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5693 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5695 // If we have a scalar reduction (vector reductions are already dealt with
5696 // by this point), we can increase the critical path length if the loop
5697 // we're unrolling is inside another loop. Limit, by default to 2, so the
5698 // critical path only gets increased by one reduction operation.
5699 if (Legal->getReductionVars()->size() &&
5700 TheLoop->getLoopDepth() > 1) {
5701 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5702 SmallUF = std::min(SmallUF, F);
5703 StoresUF = std::min(StoresUF, F);
5704 LoadsUF = std::min(LoadsUF, F);
5707 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5708 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5709 return std::max(StoresUF, LoadsUF);
5712 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5716 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5720 LoopVectorizationCostModel::RegisterUsage
5721 LoopVectorizationCostModel::calculateRegisterUsage() {
5722 // This function calculates the register usage by measuring the highest number
5723 // of values that are alive at a single location. Obviously, this is a very
5724 // rough estimation. We scan the loop in a topological order in order and
5725 // assign a number to each instruction. We use RPO to ensure that defs are
5726 // met before their users. We assume that each instruction that has in-loop
5727 // users starts an interval. We record every time that an in-loop value is
5728 // used, so we have a list of the first and last occurrences of each
5729 // instruction. Next, we transpose this data structure into a multi map that
5730 // holds the list of intervals that *end* at a specific location. This multi
5731 // map allows us to perform a linear search. We scan the instructions linearly
5732 // and record each time that a new interval starts, by placing it in a set.
5733 // If we find this value in the multi-map then we remove it from the set.
5734 // The max register usage is the maximum size of the set.
5735 // We also search for instructions that are defined outside the loop, but are
5736 // used inside the loop. We need this number separately from the max-interval
5737 // usage number because when we unroll, loop-invariant values do not take
5739 LoopBlocksDFS DFS(TheLoop);
5743 R.NumInstructions = 0;
5745 // Each 'key' in the map opens a new interval. The values
5746 // of the map are the index of the 'last seen' usage of the
5747 // instruction that is the key.
5748 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5749 // Maps instruction to its index.
5750 DenseMap<unsigned, Instruction*> IdxToInstr;
5751 // Marks the end of each interval.
5752 IntervalMap EndPoint;
5753 // Saves the list of instruction indices that are used in the loop.
5754 SmallSet<Instruction*, 8> Ends;
5755 // Saves the list of values that are used in the loop but are
5756 // defined outside the loop, such as arguments and constants.
5757 SmallPtrSet<Value*, 8> LoopInvariants;
5760 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5761 be = DFS.endRPO(); bb != be; ++bb) {
5762 R.NumInstructions += (*bb)->size();
5763 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5765 Instruction *I = it;
5766 IdxToInstr[Index++] = I;
5768 // Save the end location of each USE.
5769 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5770 Value *U = I->getOperand(i);
5771 Instruction *Instr = dyn_cast<Instruction>(U);
5773 // Ignore non-instruction values such as arguments, constants, etc.
5774 if (!Instr) continue;
5776 // If this instruction is outside the loop then record it and continue.
5777 if (!TheLoop->contains(Instr)) {
5778 LoopInvariants.insert(Instr);
5782 // Overwrite previous end points.
5783 EndPoint[Instr] = Index;
5789 // Saves the list of intervals that end with the index in 'key'.
5790 typedef SmallVector<Instruction*, 2> InstrList;
5791 DenseMap<unsigned, InstrList> TransposeEnds;
5793 // Transpose the EndPoints to a list of values that end at each index.
5794 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5796 TransposeEnds[it->second].push_back(it->first);
5798 SmallSet<Instruction*, 8> OpenIntervals;
5799 unsigned MaxUsage = 0;
5802 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5803 for (unsigned int i = 0; i < Index; ++i) {
5804 Instruction *I = IdxToInstr[i];
5805 // Ignore instructions that are never used within the loop.
5806 if (!Ends.count(I)) continue;
5808 // Ignore ephemeral values.
5809 if (EphValues.count(I))
5812 // Remove all of the instructions that end at this location.
5813 InstrList &List = TransposeEnds[i];
5814 for (unsigned int j=0, e = List.size(); j < e; ++j)
5815 OpenIntervals.erase(List[j]);
5817 // Count the number of live interals.
5818 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5820 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5821 OpenIntervals.size() << '\n');
5823 // Add the current instruction to the list of open intervals.
5824 OpenIntervals.insert(I);
5827 unsigned Invariant = LoopInvariants.size();
5828 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5829 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5830 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5832 R.LoopInvariantRegs = Invariant;
5833 R.MaxLocalUsers = MaxUsage;
5837 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5841 for (Loop::block_iterator bb = TheLoop->block_begin(),
5842 be = TheLoop->block_end(); bb != be; ++bb) {
5843 unsigned BlockCost = 0;
5844 BasicBlock *BB = *bb;
5846 // For each instruction in the old loop.
5847 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5848 // Skip dbg intrinsics.
5849 if (isa<DbgInfoIntrinsic>(it))
5852 // Ignore ephemeral values.
5853 if (EphValues.count(it))
5856 unsigned C = getInstructionCost(it, VF);
5858 // Check if we should override the cost.
5859 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5860 C = ForceTargetInstructionCost;
5863 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5864 VF << " For instruction: " << *it << '\n');
5867 // We assume that if-converted blocks have a 50% chance of being executed.
5868 // When the code is scalar then some of the blocks are avoided due to CF.
5869 // When the code is vectorized we execute all code paths.
5870 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5879 /// \brief Check whether the address computation for a non-consecutive memory
5880 /// access looks like an unlikely candidate for being merged into the indexing
5883 /// We look for a GEP which has one index that is an induction variable and all
5884 /// other indices are loop invariant. If the stride of this access is also
5885 /// within a small bound we decide that this address computation can likely be
5886 /// merged into the addressing mode.
5887 /// In all other cases, we identify the address computation as complex.
5888 static bool isLikelyComplexAddressComputation(Value *Ptr,
5889 LoopVectorizationLegality *Legal,
5890 ScalarEvolution *SE,
5891 const Loop *TheLoop) {
5892 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5896 // We are looking for a gep with all loop invariant indices except for one
5897 // which should be an induction variable.
5898 unsigned NumOperands = Gep->getNumOperands();
5899 for (unsigned i = 1; i < NumOperands; ++i) {
5900 Value *Opd = Gep->getOperand(i);
5901 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5902 !Legal->isInductionVariable(Opd))
5906 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5907 // can likely be merged into the address computation.
5908 unsigned MaxMergeDistance = 64;
5910 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5914 // Check the step is constant.
5915 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5916 // Calculate the pointer stride and check if it is consecutive.
5917 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5921 const APInt &APStepVal = C->getValue()->getValue();
5923 // Huge step value - give up.
5924 if (APStepVal.getBitWidth() > 64)
5927 int64_t StepVal = APStepVal.getSExtValue();
5929 return StepVal > MaxMergeDistance;
5932 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5933 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5939 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5940 // If we know that this instruction will remain uniform, check the cost of
5941 // the scalar version.
5942 if (Legal->isUniformAfterVectorization(I))
5945 Type *RetTy = I->getType();
5946 Type *VectorTy = ToVectorTy(RetTy, VF);
5948 // TODO: We need to estimate the cost of intrinsic calls.
5949 switch (I->getOpcode()) {
5950 case Instruction::GetElementPtr:
5951 // We mark this instruction as zero-cost because the cost of GEPs in
5952 // vectorized code depends on whether the corresponding memory instruction
5953 // is scalarized or not. Therefore, we handle GEPs with the memory
5954 // instruction cost.
5956 case Instruction::Br: {
5957 return TTI.getCFInstrCost(I->getOpcode());
5959 case Instruction::PHI:
5960 //TODO: IF-converted IFs become selects.
5962 case Instruction::Add:
5963 case Instruction::FAdd:
5964 case Instruction::Sub:
5965 case Instruction::FSub:
5966 case Instruction::Mul:
5967 case Instruction::FMul:
5968 case Instruction::UDiv:
5969 case Instruction::SDiv:
5970 case Instruction::FDiv:
5971 case Instruction::URem:
5972 case Instruction::SRem:
5973 case Instruction::FRem:
5974 case Instruction::Shl:
5975 case Instruction::LShr:
5976 case Instruction::AShr:
5977 case Instruction::And:
5978 case Instruction::Or:
5979 case Instruction::Xor: {
5980 // Since we will replace the stride by 1 the multiplication should go away.
5981 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5983 // Certain instructions can be cheaper to vectorize if they have a constant
5984 // second vector operand. One example of this are shifts on x86.
5985 TargetTransformInfo::OperandValueKind Op1VK =
5986 TargetTransformInfo::OK_AnyValue;
5987 TargetTransformInfo::OperandValueKind Op2VK =
5988 TargetTransformInfo::OK_AnyValue;
5989 TargetTransformInfo::OperandValueProperties Op1VP =
5990 TargetTransformInfo::OP_None;
5991 TargetTransformInfo::OperandValueProperties Op2VP =
5992 TargetTransformInfo::OP_None;
5993 Value *Op2 = I->getOperand(1);
5995 // Check for a splat of a constant or for a non uniform vector of constants.
5996 if (isa<ConstantInt>(Op2)) {
5997 ConstantInt *CInt = cast<ConstantInt>(Op2);
5998 if (CInt && CInt->getValue().isPowerOf2())
5999 Op2VP = TargetTransformInfo::OP_PowerOf2;
6000 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6001 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6002 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6003 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6005 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6006 if (CInt && CInt->getValue().isPowerOf2())
6007 Op2VP = TargetTransformInfo::OP_PowerOf2;
6008 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6012 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6015 case Instruction::Select: {
6016 SelectInst *SI = cast<SelectInst>(I);
6017 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6018 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6019 Type *CondTy = SI->getCondition()->getType();
6021 CondTy = VectorType::get(CondTy, VF);
6023 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6025 case Instruction::ICmp:
6026 case Instruction::FCmp: {
6027 Type *ValTy = I->getOperand(0)->getType();
6028 VectorTy = ToVectorTy(ValTy, VF);
6029 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6031 case Instruction::Store:
6032 case Instruction::Load: {
6033 StoreInst *SI = dyn_cast<StoreInst>(I);
6034 LoadInst *LI = dyn_cast<LoadInst>(I);
6035 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
6037 VectorTy = ToVectorTy(ValTy, VF);
6039 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6040 unsigned AS = SI ? SI->getPointerAddressSpace() :
6041 LI->getPointerAddressSpace();
6042 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6043 // We add the cost of address computation here instead of with the gep
6044 // instruction because only here we know whether the operation is
6047 return TTI.getAddressComputationCost(VectorTy) +
6048 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6050 // Scalarized loads/stores.
6051 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6052 bool Reverse = ConsecutiveStride < 0;
6053 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
6054 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
6055 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
6056 bool IsComplexComputation =
6057 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6059 // The cost of extracting from the value vector and pointer vector.
6060 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6061 for (unsigned i = 0; i < VF; ++i) {
6062 // The cost of extracting the pointer operand.
6063 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6064 // In case of STORE, the cost of ExtractElement from the vector.
6065 // In case of LOAD, the cost of InsertElement into the returned
6067 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6068 Instruction::InsertElement,
6072 // The cost of the scalar loads/stores.
6073 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6074 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6079 // Wide load/stores.
6080 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6081 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6084 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6088 case Instruction::ZExt:
6089 case Instruction::SExt:
6090 case Instruction::FPToUI:
6091 case Instruction::FPToSI:
6092 case Instruction::FPExt:
6093 case Instruction::PtrToInt:
6094 case Instruction::IntToPtr:
6095 case Instruction::SIToFP:
6096 case Instruction::UIToFP:
6097 case Instruction::Trunc:
6098 case Instruction::FPTrunc:
6099 case Instruction::BitCast: {
6100 // We optimize the truncation of induction variable.
6101 // The cost of these is the same as the scalar operation.
6102 if (I->getOpcode() == Instruction::Trunc &&
6103 Legal->isInductionVariable(I->getOperand(0)))
6104 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6105 I->getOperand(0)->getType());
6107 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6108 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6110 case Instruction::Call: {
6111 CallInst *CI = cast<CallInst>(I);
6112 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6113 assert(ID && "Not an intrinsic call!");
6114 Type *RetTy = ToVectorTy(CI->getType(), VF);
6115 SmallVector<Type*, 4> Tys;
6116 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6117 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6118 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6121 // We are scalarizing the instruction. Return the cost of the scalar
6122 // instruction, plus the cost of insert and extract into vector
6123 // elements, times the vector width.
6126 if (!RetTy->isVoidTy() && VF != 1) {
6127 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6129 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6132 // The cost of inserting the results plus extracting each one of the
6134 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6137 // The cost of executing VF copies of the scalar instruction. This opcode
6138 // is unknown. Assume that it is the same as 'mul'.
6139 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6145 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6146 if (Scalar->isVoidTy() || VF == 1)
6148 return VectorType::get(Scalar, VF);
6151 char LoopVectorize::ID = 0;
6152 static const char lv_name[] = "Loop Vectorization";
6153 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6154 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6155 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6156 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
6157 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6158 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6159 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6160 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6161 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
6162 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6163 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6166 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6167 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6171 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6172 // Check for a store.
6173 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6174 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6176 // Check for a load.
6177 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6178 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6184 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6185 bool IfPredicateStore) {
6186 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6187 // Holds vector parameters or scalars, in case of uniform vals.
6188 SmallVector<VectorParts, 4> Params;
6190 setDebugLocFromInst(Builder, Instr);
6192 // Find all of the vectorized parameters.
6193 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6194 Value *SrcOp = Instr->getOperand(op);
6196 // If we are accessing the old induction variable, use the new one.
6197 if (SrcOp == OldInduction) {
6198 Params.push_back(getVectorValue(SrcOp));
6202 // Try using previously calculated values.
6203 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6205 // If the src is an instruction that appeared earlier in the basic block
6206 // then it should already be vectorized.
6207 if (SrcInst && OrigLoop->contains(SrcInst)) {
6208 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6209 // The parameter is a vector value from earlier.
6210 Params.push_back(WidenMap.get(SrcInst));
6212 // The parameter is a scalar from outside the loop. Maybe even a constant.
6213 VectorParts Scalars;
6214 Scalars.append(UF, SrcOp);
6215 Params.push_back(Scalars);
6219 assert(Params.size() == Instr->getNumOperands() &&
6220 "Invalid number of operands");
6222 // Does this instruction return a value ?
6223 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6225 Value *UndefVec = IsVoidRetTy ? nullptr :
6226 UndefValue::get(Instr->getType());
6227 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6228 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6230 Instruction *InsertPt = Builder.GetInsertPoint();
6231 BasicBlock *IfBlock = Builder.GetInsertBlock();
6232 BasicBlock *CondBlock = nullptr;
6235 Loop *VectorLp = nullptr;
6236 if (IfPredicateStore) {
6237 assert(Instr->getParent()->getSinglePredecessor() &&
6238 "Only support single predecessor blocks");
6239 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6240 Instr->getParent());
6241 VectorLp = LI->getLoopFor(IfBlock);
6242 assert(VectorLp && "Must have a loop for this block");
6245 // For each vector unroll 'part':
6246 for (unsigned Part = 0; Part < UF; ++Part) {
6247 // For each scalar that we create:
6249 // Start an "if (pred) a[i] = ..." block.
6250 Value *Cmp = nullptr;
6251 if (IfPredicateStore) {
6252 if (Cond[Part]->getType()->isVectorTy())
6254 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6255 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6256 ConstantInt::get(Cond[Part]->getType(), 1));
6257 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6258 LoopVectorBody.push_back(CondBlock);
6259 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
6260 // Update Builder with newly created basic block.
6261 Builder.SetInsertPoint(InsertPt);
6264 Instruction *Cloned = Instr->clone();
6266 Cloned->setName(Instr->getName() + ".cloned");
6267 // Replace the operands of the cloned instructions with extracted scalars.
6268 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6269 Value *Op = Params[op][Part];
6270 Cloned->setOperand(op, Op);
6273 // Place the cloned scalar in the new loop.
6274 Builder.Insert(Cloned);
6276 // If the original scalar returns a value we need to place it in a vector
6277 // so that future users will be able to use it.
6279 VecResults[Part] = Cloned;
6282 if (IfPredicateStore) {
6283 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6284 LoopVectorBody.push_back(NewIfBlock);
6285 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
6286 Builder.SetInsertPoint(InsertPt);
6287 Instruction *OldBr = IfBlock->getTerminator();
6288 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6289 OldBr->eraseFromParent();
6290 IfBlock = NewIfBlock;
6295 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6296 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6297 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6299 return scalarizeInstruction(Instr, IfPredicateStore);
6302 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6306 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6310 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
6311 // When unrolling and the VF is 1, we only need to add a simple scalar.
6312 Type *ITy = Val->getType();
6313 assert(!ITy->isVectorTy() && "Val must be a scalar");
6314 Constant *C = ConstantInt::get(ITy, StartIdx);
6315 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");