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.
223 class VectorizationReport {
225 raw_string_ostream Out;
229 VectorizationReport(Instruction *I = nullptr) : Out(Message), Instr(I) {
230 Out << "loop not vectorized: ";
233 template <typename A> VectorizationReport &operator<<(const A &Value) {
238 Instruction *getInstr() { return Instr; }
240 std::string &str() { return Out.str(); }
241 operator Twine() { return Out.str(); }
243 /// \brief Emit an analysis note with the debug location from the instruction
244 /// in \p Message if available. Otherwise use the location of \p TheLoop.
245 static void emitAnalysis(VectorizationReport &Message,
246 const Function *TheFunction,
247 const Loop *TheLoop);
250 /// InnerLoopVectorizer vectorizes loops which contain only one basic
251 /// block to a specified vectorization factor (VF).
252 /// This class performs the widening of scalars into vectors, or multiple
253 /// scalars. This class also implements the following features:
254 /// * It inserts an epilogue loop for handling loops that don't have iteration
255 /// counts that are known to be a multiple of the vectorization factor.
256 /// * It handles the code generation for reduction variables.
257 /// * Scalarization (implementation using scalars) of un-vectorizable
259 /// InnerLoopVectorizer does not perform any vectorization-legality
260 /// checks, and relies on the caller to check for the different legality
261 /// aspects. The InnerLoopVectorizer relies on the
262 /// LoopVectorizationLegality class to provide information about the induction
263 /// and reduction variables that were found to a given vectorization factor.
264 class InnerLoopVectorizer {
266 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
267 DominatorTree *DT, const DataLayout *DL,
268 const TargetLibraryInfo *TLI, unsigned VecWidth,
269 unsigned UnrollFactor)
270 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
271 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
272 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
275 // Perform the actual loop widening (vectorization).
276 void vectorize(LoopVectorizationLegality *L) {
278 // Create a new empty loop. Unlink the old loop and connect the new one.
280 // Widen each instruction in the old loop to a new one in the new loop.
281 // Use the Legality module to find the induction and reduction variables.
283 // Register the new loop and update the analysis passes.
287 virtual ~InnerLoopVectorizer() {}
290 /// A small list of PHINodes.
291 typedef SmallVector<PHINode*, 4> PhiVector;
292 /// When we unroll loops we have multiple vector values for each scalar.
293 /// This data structure holds the unrolled and vectorized values that
294 /// originated from one scalar instruction.
295 typedef SmallVector<Value*, 2> VectorParts;
297 // When we if-convert we need create edge masks. We have to cache values so
298 // that we don't end up with exponential recursion/IR.
299 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
300 VectorParts> EdgeMaskCache;
302 /// \brief Add code that checks at runtime if the accessed arrays overlap.
304 /// Returns a pair of instructions where the first element is the first
305 /// instruction generated in possibly a sequence of instructions and the
306 /// second value is the final comparator value or NULL if no check is needed.
307 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
309 /// \brief Add checks for strides that where assumed to be 1.
311 /// Returns the last check instruction and the first check instruction in the
312 /// pair as (first, last).
313 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
315 /// Create an empty loop, based on the loop ranges of the old loop.
316 void createEmptyLoop();
317 /// Copy and widen the instructions from the old loop.
318 virtual void vectorizeLoop();
320 /// \brief The Loop exit block may have single value PHI nodes where the
321 /// incoming value is 'Undef'. While vectorizing we only handled real values
322 /// that were defined inside the loop. Here we fix the 'undef case'.
326 /// A helper function that computes the predicate of the block BB, assuming
327 /// that the header block of the loop is set to True. It returns the *entry*
328 /// mask for the block BB.
329 VectorParts createBlockInMask(BasicBlock *BB);
330 /// A helper function that computes the predicate of the edge between SRC
332 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
334 /// A helper function to vectorize a single BB within the innermost loop.
335 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
337 /// Vectorize a single PHINode in a block. This method handles the induction
338 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
339 /// arbitrary length vectors.
340 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
341 unsigned UF, unsigned VF, PhiVector *PV);
343 /// Insert the new loop to the loop hierarchy and pass manager
344 /// and update the analysis passes.
345 void updateAnalysis();
347 /// This instruction is un-vectorizable. Implement it as a sequence
348 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
349 /// scalarized instruction behind an if block predicated on the control
350 /// dependence of the instruction.
351 virtual void scalarizeInstruction(Instruction *Instr,
352 bool IfPredicateStore=false);
354 /// Vectorize Load and Store instructions,
355 virtual void vectorizeMemoryInstruction(Instruction *Instr);
357 /// Create a broadcast instruction. This method generates a broadcast
358 /// instruction (shuffle) for loop invariant values and for the induction
359 /// value. If this is the induction variable then we extend it to N, N+1, ...
360 /// this is needed because each iteration in the loop corresponds to a SIMD
362 virtual Value *getBroadcastInstrs(Value *V);
364 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
365 /// to each vector element of Val. The sequence starts at StartIndex.
366 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
368 /// When we go over instructions in the basic block we rely on previous
369 /// values within the current basic block or on loop invariant values.
370 /// When we widen (vectorize) values we place them in the map. If the values
371 /// are not within the map, they have to be loop invariant, so we simply
372 /// broadcast them into a vector.
373 VectorParts &getVectorValue(Value *V);
375 /// Generate a shuffle sequence that will reverse the vector Vec.
376 virtual Value *reverseVector(Value *Vec);
378 /// This is a helper class that holds the vectorizer state. It maps scalar
379 /// instructions to vector instructions. When the code is 'unrolled' then
380 /// then a single scalar value is mapped to multiple vector parts. The parts
381 /// are stored in the VectorPart type.
383 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
385 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
387 /// \return True if 'Key' is saved in the Value Map.
388 bool has(Value *Key) const { return MapStorage.count(Key); }
390 /// Initializes a new entry in the map. Sets all of the vector parts to the
391 /// save value in 'Val'.
392 /// \return A reference to a vector with splat values.
393 VectorParts &splat(Value *Key, Value *Val) {
394 VectorParts &Entry = MapStorage[Key];
395 Entry.assign(UF, Val);
399 ///\return A reference to the value that is stored at 'Key'.
400 VectorParts &get(Value *Key) {
401 VectorParts &Entry = MapStorage[Key];
404 assert(Entry.size() == UF);
409 /// The unroll factor. Each entry in the map stores this number of vector
413 /// Map storage. We use std::map and not DenseMap because insertions to a
414 /// dense map invalidates its iterators.
415 std::map<Value *, VectorParts> MapStorage;
418 /// The original loop.
420 /// Scev analysis to use.
429 const DataLayout *DL;
430 /// Target Library Info.
431 const TargetLibraryInfo *TLI;
433 /// The vectorization SIMD factor to use. Each vector will have this many
438 /// The vectorization unroll factor to use. Each scalar is vectorized to this
439 /// many different vector instructions.
442 /// The builder that we use
445 // --- Vectorization state ---
447 /// The vector-loop preheader.
448 BasicBlock *LoopVectorPreHeader;
449 /// The scalar-loop preheader.
450 BasicBlock *LoopScalarPreHeader;
451 /// Middle Block between the vector and the scalar.
452 BasicBlock *LoopMiddleBlock;
453 ///The ExitBlock of the scalar loop.
454 BasicBlock *LoopExitBlock;
455 ///The vector loop body.
456 SmallVector<BasicBlock *, 4> LoopVectorBody;
457 ///The scalar loop body.
458 BasicBlock *LoopScalarBody;
459 /// A list of all bypass blocks. The first block is the entry of the loop.
460 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
462 /// The new Induction variable which was added to the new block.
464 /// The induction variable of the old basic block.
465 PHINode *OldInduction;
466 /// Holds the extended (to the widest induction type) start index.
468 /// Maps scalars to widened vectors.
470 EdgeMaskCache MaskCache;
472 LoopVectorizationLegality *Legal;
475 class InnerLoopUnroller : public InnerLoopVectorizer {
477 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
478 DominatorTree *DT, const DataLayout *DL,
479 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
480 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
483 void scalarizeInstruction(Instruction *Instr,
484 bool IfPredicateStore = false) override;
485 void vectorizeMemoryInstruction(Instruction *Instr) override;
486 Value *getBroadcastInstrs(Value *V) override;
487 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
488 Value *reverseVector(Value *Vec) override;
491 /// \brief Look for a meaningful debug location on the instruction or it's
493 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
498 if (I->getDebugLoc() != Empty)
501 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
502 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
503 if (OpInst->getDebugLoc() != Empty)
510 /// \brief Set the debug location in the builder using the debug location in the
512 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
513 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
514 B.SetCurrentDebugLocation(Inst->getDebugLoc());
516 B.SetCurrentDebugLocation(DebugLoc());
520 /// \return string containing a file name and a line # for the given loop.
521 static std::string getDebugLocString(const Loop *L) {
524 raw_string_ostream OS(Result);
525 const DebugLoc LoopDbgLoc = L->getStartLoc();
526 if (!LoopDbgLoc.isUnknown())
527 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
529 // Just print the module name.
530 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
537 /// \brief Propagate known metadata from one instruction to another.
538 static void propagateMetadata(Instruction *To, const Instruction *From) {
539 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
540 From->getAllMetadataOtherThanDebugLoc(Metadata);
542 for (auto M : Metadata) {
543 unsigned Kind = M.first;
545 // These are safe to transfer (this is safe for TBAA, even when we
546 // if-convert, because should that metadata have had a control dependency
547 // on the condition, and thus actually aliased with some other
548 // non-speculated memory access when the condition was false, this would be
549 // caught by the runtime overlap checks).
550 if (Kind != LLVMContext::MD_tbaa &&
551 Kind != LLVMContext::MD_alias_scope &&
552 Kind != LLVMContext::MD_noalias &&
553 Kind != LLVMContext::MD_fpmath)
556 To->setMetadata(Kind, M.second);
560 void VectorizationReport::emitAnalysis(VectorizationReport &Message,
561 const Function *TheFunction,
562 const Loop *TheLoop) {
563 DebugLoc DL = TheLoop->getStartLoc();
564 if (Instruction *I = Message.getInstr())
565 DL = I->getDebugLoc();
566 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
567 *TheFunction, DL, Message.str());
570 /// \brief Propagate known metadata from one instruction to a vector of others.
571 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
573 if (Instruction *I = dyn_cast<Instruction>(V))
574 propagateMetadata(I, From);
578 /// This struct holds information about the memory runtime legality
579 /// check that a group of pointers do not overlap.
580 struct RuntimePointerCheck {
581 RuntimePointerCheck() : Need(false) {}
583 /// Reset the state of the pointer runtime information.
590 DependencySetId.clear();
594 /// Insert a pointer and calculate the start and end SCEVs.
595 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
596 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
598 /// This flag indicates if we need to add the runtime check.
600 /// Holds the pointers that we need to check.
601 SmallVector<TrackingVH<Value>, 2> Pointers;
602 /// Holds the pointer value at the beginning of the loop.
603 SmallVector<const SCEV*, 2> Starts;
604 /// Holds the pointer value at the end of the loop.
605 SmallVector<const SCEV*, 2> Ends;
606 /// Holds the information if this pointer is used for writing to memory.
607 SmallVector<bool, 2> IsWritePtr;
608 /// Holds the id of the set of pointers that could be dependent because of a
609 /// shared underlying object.
610 SmallVector<unsigned, 2> DependencySetId;
611 /// Holds the id of the disjoint alias set to which this pointer belongs.
612 SmallVector<unsigned, 2> AliasSetId;
614 } // end anonymous namespace
616 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
617 /// to what vectorization factor.
618 /// This class does not look at the profitability of vectorization, only the
619 /// legality. This class has two main kinds of checks:
620 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
621 /// will change the order of memory accesses in a way that will change the
622 /// correctness of the program.
623 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
624 /// checks for a number of different conditions, such as the availability of a
625 /// single induction variable, that all types are supported and vectorize-able,
626 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
627 /// This class is also used by InnerLoopVectorizer for identifying
628 /// induction variable and the different reduction variables.
629 class LoopVectorizationLegality {
633 unsigned NumPredStores;
635 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
636 DominatorTree *DT, TargetLibraryInfo *TLI,
637 AliasAnalysis *AA, Function *F,
638 const TargetTransformInfo *TTI)
639 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
640 DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), Induction(nullptr),
641 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
644 /// This enum represents the kinds of reductions that we support.
646 RK_NoReduction, ///< Not a reduction.
647 RK_IntegerAdd, ///< Sum of integers.
648 RK_IntegerMult, ///< Product of integers.
649 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
650 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
651 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
652 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
653 RK_FloatAdd, ///< Sum of floats.
654 RK_FloatMult, ///< Product of floats.
655 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
658 /// This enum represents the kinds of inductions that we support.
660 IK_NoInduction, ///< Not an induction variable.
661 IK_IntInduction, ///< Integer induction variable. Step = C.
662 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
665 // This enum represents the kind of minmax reduction.
666 enum MinMaxReductionKind {
676 /// This struct holds information about reduction variables.
677 struct ReductionDescriptor {
678 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
679 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
681 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
682 MinMaxReductionKind MK)
683 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
685 // The starting value of the reduction.
686 // It does not have to be zero!
687 TrackingVH<Value> StartValue;
688 // The instruction who's value is used outside the loop.
689 Instruction *LoopExitInstr;
690 // The kind of the reduction.
692 // If this a min/max reduction the kind of reduction.
693 MinMaxReductionKind MinMaxKind;
696 /// This POD struct holds information about a potential reduction operation.
697 struct ReductionInstDesc {
698 ReductionInstDesc(bool IsRedux, Instruction *I) :
699 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
701 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
702 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
704 // Is this instruction a reduction candidate.
706 // The last instruction in a min/max pattern (select of the select(icmp())
707 // pattern), or the current reduction instruction otherwise.
708 Instruction *PatternLastInst;
709 // If this is a min/max pattern the comparison predicate.
710 MinMaxReductionKind MinMaxKind;
713 /// A struct for saving information about induction variables.
714 struct InductionInfo {
715 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
716 : StartValue(Start), IK(K), StepValue(Step) {
717 assert(IK != IK_NoInduction && "Not an induction");
718 assert(StartValue && "StartValue is null");
719 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
720 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
721 "StartValue is not a pointer for pointer induction");
722 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
723 "StartValue is not an integer for integer induction");
724 assert(StepValue->getType()->isIntegerTy() &&
725 "StepValue is not an integer");
728 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
730 /// Get the consecutive direction. Returns:
731 /// 0 - unknown or non-consecutive.
732 /// 1 - consecutive and increasing.
733 /// -1 - consecutive and decreasing.
734 int getConsecutiveDirection() const {
735 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
736 return StepValue->getSExtValue();
740 /// Compute the transformed value of Index at offset StartValue using step
742 /// For integer induction, returns StartValue + Index * StepValue.
743 /// For pointer induction, returns StartValue[Index * StepValue].
744 /// FIXME: The newly created binary instructions should contain nsw/nuw
745 /// flags, which can be found from the original scalar operations.
746 Value *transform(IRBuilder<> &B, Value *Index) const {
748 case IK_IntInduction:
749 assert(Index->getType() == StartValue->getType() &&
750 "Index type does not match StartValue type");
751 if (StepValue->isMinusOne())
752 return B.CreateSub(StartValue, Index);
753 if (!StepValue->isOne())
754 Index = B.CreateMul(Index, StepValue);
755 return B.CreateAdd(StartValue, Index);
757 case IK_PtrInduction:
758 if (StepValue->isMinusOne())
759 Index = B.CreateNeg(Index);
760 else if (!StepValue->isOne())
761 Index = B.CreateMul(Index, StepValue);
762 return B.CreateGEP(StartValue, Index);
767 llvm_unreachable("invalid enum");
771 TrackingVH<Value> StartValue;
775 ConstantInt *StepValue;
778 /// ReductionList contains the reduction descriptors for all
779 /// of the reductions that were found in the loop.
780 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
782 /// InductionList saves induction variables and maps them to the
783 /// induction descriptor.
784 typedef MapVector<PHINode*, InductionInfo> InductionList;
786 /// Returns true if it is legal to vectorize this loop.
787 /// This does not mean that it is profitable to vectorize this
788 /// loop, only that it is legal to do so.
791 /// Returns the Induction variable.
792 PHINode *getInduction() { return Induction; }
794 /// Returns the reduction variables found in the loop.
795 ReductionList *getReductionVars() { return &Reductions; }
797 /// Returns the induction variables found in the loop.
798 InductionList *getInductionVars() { return &Inductions; }
800 /// Returns the widest induction type.
801 Type *getWidestInductionType() { return WidestIndTy; }
803 /// Returns True if V is an induction variable in this loop.
804 bool isInductionVariable(const Value *V);
806 /// Return true if the block BB needs to be predicated in order for the loop
807 /// to be vectorized.
808 bool blockNeedsPredication(BasicBlock *BB);
810 /// Check if this pointer is consecutive when vectorizing. This happens
811 /// when the last index of the GEP is the induction variable, or that the
812 /// pointer itself is an induction variable.
813 /// This check allows us to vectorize A[idx] into a wide load/store.
815 /// 0 - Stride is unknown or non-consecutive.
816 /// 1 - Address is consecutive.
817 /// -1 - Address is consecutive, and decreasing.
818 int isConsecutivePtr(Value *Ptr);
820 /// Returns true if the value V is uniform within the loop.
821 bool isUniform(Value *V);
823 /// Returns true if this instruction will remain scalar after vectorization.
824 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
826 /// Returns the information that we collected about runtime memory check.
827 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
829 /// This function returns the identity element (or neutral element) for
831 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
833 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
835 bool hasStride(Value *V) { return StrideSet.count(V); }
836 bool mustCheckStrides() { return !StrideSet.empty(); }
837 SmallPtrSet<Value *, 8>::iterator strides_begin() {
838 return StrideSet.begin();
840 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
842 /// Returns true if the target machine supports masked store operation
843 /// for the given \p DataType and kind of access to \p Ptr.
844 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
845 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
847 /// Returns true if the target machine supports masked load operation
848 /// for the given \p DataType and kind of access to \p Ptr.
849 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
850 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
852 /// Returns true if vector representation of the instruction \p I
854 bool isMaskRequired(const Instruction* I) {
855 return (MaskedOp.count(I) != 0);
858 /// Check if a single basic block loop is vectorizable.
859 /// At this point we know that this is a loop with a constant trip count
860 /// and we only need to check individual instructions.
861 bool canVectorizeInstrs();
863 /// When we vectorize loops we may change the order in which
864 /// we read and write from memory. This method checks if it is
865 /// legal to vectorize the code, considering only memory constrains.
866 /// Returns true if the loop is vectorizable
867 bool canVectorizeMemory();
869 /// Return true if we can vectorize this loop using the IF-conversion
871 bool canVectorizeWithIfConvert();
873 /// Collect the variables that need to stay uniform after vectorization.
874 void collectLoopUniforms();
876 /// Return true if all of the instructions in the block can be speculatively
877 /// executed. \p SafePtrs is a list of addresses that are known to be legal
878 /// and we know that we can read from them without segfault.
879 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
881 /// Returns True, if 'Phi' is the kind of reduction variable for type
882 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
883 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
884 /// Returns a struct describing if the instruction 'I' can be a reduction
885 /// variable of type 'Kind'. If the reduction is a min/max pattern of
886 /// select(icmp()) this function advances the instruction pointer 'I' from the
887 /// compare instruction to the select instruction and stores this pointer in
888 /// 'PatternLastInst' member of the returned struct.
889 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
890 ReductionInstDesc &Desc);
891 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
892 /// pattern corresponding to a min(X, Y) or max(X, Y).
893 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
894 ReductionInstDesc &Prev);
895 /// Returns the induction kind of Phi and record the step. This function may
896 /// return NoInduction if the PHI is not an induction variable.
897 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
899 /// \brief Collect memory access with loop invariant strides.
901 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
903 void collectStridedAccess(Value *LoadOrStoreInst);
905 /// Report an analysis message to assist the user in diagnosing loops that are
907 void emitAnalysis(VectorizationReport &Message) {
908 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
911 /// The loop that we evaluate.
915 /// DataLayout analysis.
916 const DataLayout *DL;
919 /// Target Library Info.
920 TargetLibraryInfo *TLI;
924 Function *TheFunction;
925 /// Target Transform Info
926 const TargetTransformInfo *TTI;
928 // --- vectorization state --- //
930 /// Holds the integer induction variable. This is the counter of the
933 /// Holds the reduction variables.
934 ReductionList Reductions;
935 /// Holds all of the induction variables that we found in the loop.
936 /// Notice that inductions don't need to start at zero and that induction
937 /// variables can be pointers.
938 InductionList Inductions;
939 /// Holds the widest induction type encountered.
942 /// Allowed outside users. This holds the reduction
943 /// vars which can be accessed from outside the loop.
944 SmallPtrSet<Value*, 4> AllowedExit;
945 /// This set holds the variables which are known to be uniform after
947 SmallPtrSet<Instruction*, 4> Uniforms;
948 /// We need to check that all of the pointers in this list are disjoint
950 RuntimePointerCheck PtrRtCheck;
951 /// Can we assume the absence of NaNs.
952 bool HasFunNoNaNAttr;
954 unsigned MaxSafeDepDistBytes;
956 ValueToValueMap Strides;
957 SmallPtrSet<Value *, 8> StrideSet;
959 /// While vectorizing these instructions we have to generate a
960 /// call to the appropriate masked intrinsic
961 SmallPtrSet<const Instruction*, 8> MaskedOp;
964 /// LoopVectorizationCostModel - estimates the expected speedups due to
966 /// In many cases vectorization is not profitable. This can happen because of
967 /// a number of reasons. In this class we mainly attempt to predict the
968 /// expected speedup/slowdowns due to the supported instruction set. We use the
969 /// TargetTransformInfo to query the different backends for the cost of
970 /// different operations.
971 class LoopVectorizationCostModel {
973 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
974 LoopVectorizationLegality *Legal,
975 const TargetTransformInfo &TTI,
976 const DataLayout *DL, const TargetLibraryInfo *TLI,
977 AssumptionCache *AC, const Function *F,
978 const LoopVectorizeHints *Hints)
979 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
980 TheFunction(F), Hints(Hints) {
981 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
984 /// Information about vectorization costs
985 struct VectorizationFactor {
986 unsigned Width; // Vector width with best cost
987 unsigned Cost; // Cost of the loop with that width
989 /// \return The most profitable vectorization factor and the cost of that VF.
990 /// This method checks every power of two up to VF. If UserVF is not ZERO
991 /// then this vectorization factor will be selected if vectorization is
993 VectorizationFactor selectVectorizationFactor(bool OptForSize);
995 /// \return The size (in bits) of the widest type in the code that
996 /// needs to be vectorized. We ignore values that remain scalar such as
997 /// 64 bit loop indices.
998 unsigned getWidestType();
1000 /// \return The most profitable unroll factor.
1001 /// If UserUF is non-zero then this method finds the best unroll-factor
1002 /// based on register pressure and other parameters.
1003 /// VF and LoopCost are the selected vectorization factor and the cost of the
1005 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
1007 /// \brief A struct that represents some properties of the register usage
1009 struct RegisterUsage {
1010 /// Holds the number of loop invariant values that are used in the loop.
1011 unsigned LoopInvariantRegs;
1012 /// Holds the maximum number of concurrent live intervals in the loop.
1013 unsigned MaxLocalUsers;
1014 /// Holds the number of instructions in the loop.
1015 unsigned NumInstructions;
1018 /// \return information about the register usage of the loop.
1019 RegisterUsage calculateRegisterUsage();
1022 /// Returns the expected execution cost. The unit of the cost does
1023 /// not matter because we use the 'cost' units to compare different
1024 /// vector widths. The cost that is returned is *not* normalized by
1025 /// the factor width.
1026 unsigned expectedCost(unsigned VF);
1028 /// Returns the execution time cost of an instruction for a given vector
1029 /// width. Vector width of one means scalar.
1030 unsigned getInstructionCost(Instruction *I, unsigned VF);
1032 /// A helper function for converting Scalar types to vector types.
1033 /// If the incoming type is void, we return void. If the VF is 1, we return
1034 /// the scalar type.
1035 static Type* ToVectorTy(Type *Scalar, unsigned VF);
1037 /// Returns whether the instruction is a load or store and will be a emitted
1038 /// as a vector operation.
1039 bool isConsecutiveLoadOrStore(Instruction *I);
1041 /// Report an analysis message to assist the user in diagnosing loops that are
1043 void emitAnalysis(VectorizationReport &Message) {
1044 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
1047 /// Values used only by @llvm.assume calls.
1048 SmallPtrSet<const Value *, 32> EphValues;
1050 /// The loop that we evaluate.
1053 ScalarEvolution *SE;
1054 /// Loop Info analysis.
1056 /// Vectorization legality.
1057 LoopVectorizationLegality *Legal;
1058 /// Vector target information.
1059 const TargetTransformInfo &TTI;
1060 /// Target data layout information.
1061 const DataLayout *DL;
1062 /// Target Library Info.
1063 const TargetLibraryInfo *TLI;
1064 const Function *TheFunction;
1065 // Loop Vectorize Hint.
1066 const LoopVectorizeHints *Hints;
1069 /// Utility class for getting and setting loop vectorizer hints in the form
1070 /// of loop metadata.
1071 /// This class keeps a number of loop annotations locally (as member variables)
1072 /// and can, upon request, write them back as metadata on the loop. It will
1073 /// initially scan the loop for existing metadata, and will update the local
1074 /// values based on information in the loop.
1075 /// We cannot write all values to metadata, as the mere presence of some info,
1076 /// for example 'force', means a decision has been made. So, we need to be
1077 /// careful NOT to add them if the user hasn't specifically asked so.
1078 class LoopVectorizeHints {
1085 /// Hint - associates name and validation with the hint value.
1088 unsigned Value; // This may have to change for non-numeric values.
1091 Hint(const char * Name, unsigned Value, HintKind Kind)
1092 : Name(Name), Value(Value), Kind(Kind) { }
1094 bool validate(unsigned Val) {
1097 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1099 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1107 /// Vectorization width.
1109 /// Vectorization interleave factor.
1111 /// Vectorization forced
1114 /// Return the loop metadata prefix.
1115 static StringRef Prefix() { return "llvm.loop."; }
1119 FK_Undefined = -1, ///< Not selected.
1120 FK_Disabled = 0, ///< Forcing disabled.
1121 FK_Enabled = 1, ///< Forcing enabled.
1124 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1125 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1126 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1127 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1129 // Populate values with existing loop metadata.
1130 getHintsFromMetadata();
1132 // force-vector-interleave overrides DisableInterleaving.
1133 if (VectorizationInterleave.getNumOccurrences() > 0)
1134 Interleave.Value = VectorizationInterleave;
1136 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1137 << "LV: Interleaving disabled by the pass manager\n");
1140 /// Mark the loop L as already vectorized by setting the width to 1.
1141 void setAlreadyVectorized() {
1142 Width.Value = Interleave.Value = 1;
1143 Hint Hints[] = {Width, Interleave};
1144 writeHintsToMetadata(Hints);
1147 /// Dumps all the hint information.
1148 std::string emitRemark() const {
1149 VectorizationReport R;
1150 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1151 R << "vectorization is explicitly disabled";
1153 R << "use -Rpass-analysis=loop-vectorize for more info";
1154 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1155 R << " (Force=true";
1156 if (Width.Value != 0)
1157 R << ", Vector Width=" << Width.Value;
1158 if (Interleave.Value != 0)
1159 R << ", Interleave Count=" << Interleave.Value;
1167 unsigned getWidth() const { return Width.Value; }
1168 unsigned getInterleave() const { return Interleave.Value; }
1169 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1172 /// Find hints specified in the loop metadata and update local values.
1173 void getHintsFromMetadata() {
1174 MDNode *LoopID = TheLoop->getLoopID();
1178 // First operand should refer to the loop id itself.
1179 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1180 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1182 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1183 const MDString *S = nullptr;
1184 SmallVector<Metadata *, 4> Args;
1186 // The expected hint is either a MDString or a MDNode with the first
1187 // operand a MDString.
1188 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1189 if (!MD || MD->getNumOperands() == 0)
1191 S = dyn_cast<MDString>(MD->getOperand(0));
1192 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1193 Args.push_back(MD->getOperand(i));
1195 S = dyn_cast<MDString>(LoopID->getOperand(i));
1196 assert(Args.size() == 0 && "too many arguments for MDString");
1202 // Check if the hint starts with the loop metadata prefix.
1203 StringRef Name = S->getString();
1204 if (Args.size() == 1)
1205 setHint(Name, Args[0]);
1209 /// Checks string hint with one operand and set value if valid.
1210 void setHint(StringRef Name, Metadata *Arg) {
1211 if (!Name.startswith(Prefix()))
1213 Name = Name.substr(Prefix().size(), StringRef::npos);
1215 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1217 unsigned Val = C->getZExtValue();
1219 Hint *Hints[] = {&Width, &Interleave, &Force};
1220 for (auto H : Hints) {
1221 if (Name == H->Name) {
1222 if (H->validate(Val))
1225 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1231 /// Create a new hint from name / value pair.
1232 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1233 LLVMContext &Context = TheLoop->getHeader()->getContext();
1234 Metadata *MDs[] = {MDString::get(Context, Name),
1235 ConstantAsMetadata::get(
1236 ConstantInt::get(Type::getInt32Ty(Context), V))};
1237 return MDNode::get(Context, MDs);
1240 /// Matches metadata with hint name.
1241 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1242 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1246 for (auto H : HintTypes)
1247 if (Name->getString().endswith(H.Name))
1252 /// Sets current hints into loop metadata, keeping other values intact.
1253 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1254 if (HintTypes.size() == 0)
1257 // Reserve the first element to LoopID (see below).
1258 SmallVector<Metadata *, 4> MDs(1);
1259 // If the loop already has metadata, then ignore the existing operands.
1260 MDNode *LoopID = TheLoop->getLoopID();
1262 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1263 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1264 // If node in update list, ignore old value.
1265 if (!matchesHintMetadataName(Node, HintTypes))
1266 MDs.push_back(Node);
1270 // Now, add the missing hints.
1271 for (auto H : HintTypes)
1272 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1274 // Replace current metadata node with new one.
1275 LLVMContext &Context = TheLoop->getHeader()->getContext();
1276 MDNode *NewLoopID = MDNode::get(Context, MDs);
1277 // Set operand 0 to refer to the loop id itself.
1278 NewLoopID->replaceOperandWith(0, NewLoopID);
1280 TheLoop->setLoopID(NewLoopID);
1283 /// The loop these hints belong to.
1284 const Loop *TheLoop;
1287 static void emitMissedWarning(Function *F, Loop *L,
1288 const LoopVectorizeHints &LH) {
1289 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1290 L->getStartLoc(), LH.emitRemark());
1292 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1293 if (LH.getWidth() != 1)
1294 emitLoopVectorizeWarning(
1295 F->getContext(), *F, L->getStartLoc(),
1296 "failed explicitly specified loop vectorization");
1297 else if (LH.getInterleave() != 1)
1298 emitLoopInterleaveWarning(
1299 F->getContext(), *F, L->getStartLoc(),
1300 "failed explicitly specified loop interleaving");
1304 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1306 return V.push_back(&L);
1308 for (Loop *InnerL : L)
1309 addInnerLoop(*InnerL, V);
1312 /// The LoopVectorize Pass.
1313 struct LoopVectorize : public FunctionPass {
1314 /// Pass identification, replacement for typeid
1317 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1319 DisableUnrolling(NoUnrolling),
1320 AlwaysVectorize(AlwaysVectorize) {
1321 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1324 ScalarEvolution *SE;
1325 const DataLayout *DL;
1327 TargetTransformInfo *TTI;
1329 BlockFrequencyInfo *BFI;
1330 TargetLibraryInfo *TLI;
1332 AssumptionCache *AC;
1333 bool DisableUnrolling;
1334 bool AlwaysVectorize;
1336 BlockFrequency ColdEntryFreq;
1338 bool runOnFunction(Function &F) override {
1339 SE = &getAnalysis<ScalarEvolution>();
1340 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1341 DL = DLP ? &DLP->getDataLayout() : nullptr;
1342 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1343 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1344 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1345 BFI = &getAnalysis<BlockFrequencyInfo>();
1346 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1347 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1348 AA = &getAnalysis<AliasAnalysis>();
1349 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1351 // Compute some weights outside of the loop over the loops. Compute this
1352 // using a BranchProbability to re-use its scaling math.
1353 const BranchProbability ColdProb(1, 5); // 20%
1354 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1356 // If the target claims to have no vector registers don't attempt
1358 if (!TTI->getNumberOfRegisters(true))
1362 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1363 << ": Missing data layout\n");
1367 // Build up a worklist of inner-loops to vectorize. This is necessary as
1368 // the act of vectorizing or partially unrolling a loop creates new loops
1369 // and can invalidate iterators across the loops.
1370 SmallVector<Loop *, 8> Worklist;
1373 addInnerLoop(*L, Worklist);
1375 LoopsAnalyzed += Worklist.size();
1377 // Now walk the identified inner loops.
1378 bool Changed = false;
1379 while (!Worklist.empty())
1380 Changed |= processLoop(Worklist.pop_back_val());
1382 // Process each loop nest in the function.
1386 bool processLoop(Loop *L) {
1387 assert(L->empty() && "Only process inner loops.");
1390 const std::string DebugLocStr = getDebugLocString(L);
1393 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1394 << L->getHeader()->getParent()->getName() << "\" from "
1395 << DebugLocStr << "\n");
1397 LoopVectorizeHints Hints(L, DisableUnrolling);
1399 DEBUG(dbgs() << "LV: Loop hints:"
1401 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1403 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1405 : "?")) << " width=" << Hints.getWidth()
1406 << " unroll=" << Hints.getInterleave() << "\n");
1408 // Function containing loop
1409 Function *F = L->getHeader()->getParent();
1411 // Looking at the diagnostic output is the only way to determine if a loop
1412 // was vectorized (other than looking at the IR or machine code), so it
1413 // is important to generate an optimization remark for each loop. Most of
1414 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1415 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1416 // less verbose reporting vectorized loops and unvectorized loops that may
1417 // benefit from vectorization, respectively.
1419 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1420 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1421 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1422 L->getStartLoc(), Hints.emitRemark());
1426 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1427 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1428 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1429 L->getStartLoc(), Hints.emitRemark());
1433 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1434 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1435 emitOptimizationRemarkAnalysis(
1436 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1437 "loop not vectorized: vector width and interleave count are "
1438 "explicitly set to 1");
1442 // Check the loop for a trip count threshold:
1443 // do not vectorize loops with a tiny trip count.
1444 const unsigned TC = SE->getSmallConstantTripCount(L);
1445 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1446 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1447 << "This loop is not worth vectorizing.");
1448 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1449 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1451 DEBUG(dbgs() << "\n");
1452 emitOptimizationRemarkAnalysis(
1453 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1454 "vectorization is not beneficial and is not explicitly forced");
1459 // Check if it is legal to vectorize the loop.
1460 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1461 if (!LVL.canVectorize()) {
1462 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1463 emitMissedWarning(F, L, Hints);
1467 // Use the cost model.
1468 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1471 // Check the function attributes to find out if this function should be
1472 // optimized for size.
1473 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1474 F->hasFnAttribute(Attribute::OptimizeForSize);
1476 // Compute the weighted frequency of this loop being executed and see if it
1477 // is less than 20% of the function entry baseline frequency. Note that we
1478 // always have a canonical loop here because we think we *can* vectoriez.
1479 // FIXME: This is hidden behind a flag due to pervasive problems with
1480 // exactly what block frequency models.
1481 if (LoopVectorizeWithBlockFrequency) {
1482 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1483 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1484 LoopEntryFreq < ColdEntryFreq)
1488 // Check the function attributes to see if implicit floats are allowed.a
1489 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1490 // an integer loop and the vector instructions selected are purely integer
1491 // vector instructions?
1492 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1493 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1494 "attribute is used.\n");
1495 emitOptimizationRemarkAnalysis(
1496 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1497 "loop not vectorized due to NoImplicitFloat attribute");
1498 emitMissedWarning(F, L, Hints);
1502 // Select the optimal vectorization factor.
1503 const LoopVectorizationCostModel::VectorizationFactor VF =
1504 CM.selectVectorizationFactor(OptForSize);
1506 // Select the unroll factor.
1508 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1510 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1511 << DebugLocStr << '\n');
1512 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1514 if (VF.Width == 1) {
1515 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1518 emitOptimizationRemarkAnalysis(
1519 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1520 "not beneficial to vectorize and user disabled interleaving");
1523 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1525 // Report the unrolling decision.
1526 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1527 Twine("unrolled with interleaving factor " +
1529 " (vectorization not beneficial)"));
1531 // We decided not to vectorize, but we may want to unroll.
1533 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1534 Unroller.vectorize(&LVL);
1536 // If we decided that it is *legal* to vectorize the loop then do it.
1537 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1541 // Report the vectorization decision.
1542 emitOptimizationRemark(
1543 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1544 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1545 ", unrolling interleave factor: " + Twine(UF) + ")");
1548 // Mark the loop as already vectorized to avoid vectorizing again.
1549 Hints.setAlreadyVectorized();
1551 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1555 void getAnalysisUsage(AnalysisUsage &AU) const override {
1556 AU.addRequired<AssumptionCacheTracker>();
1557 AU.addRequiredID(LoopSimplifyID);
1558 AU.addRequiredID(LCSSAID);
1559 AU.addRequired<BlockFrequencyInfo>();
1560 AU.addRequired<DominatorTreeWrapperPass>();
1561 AU.addRequired<LoopInfoWrapperPass>();
1562 AU.addRequired<ScalarEvolution>();
1563 AU.addRequired<TargetTransformInfoWrapperPass>();
1564 AU.addRequired<AliasAnalysis>();
1565 AU.addPreserved<LoopInfoWrapperPass>();
1566 AU.addPreserved<DominatorTreeWrapperPass>();
1567 AU.addPreserved<AliasAnalysis>();
1572 } // end anonymous namespace
1574 //===----------------------------------------------------------------------===//
1575 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1576 // LoopVectorizationCostModel.
1577 //===----------------------------------------------------------------------===//
1579 static Value *stripIntegerCast(Value *V) {
1580 if (CastInst *CI = dyn_cast<CastInst>(V))
1581 if (CI->getOperand(0)->getType()->isIntegerTy())
1582 return CI->getOperand(0);
1586 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1588 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1590 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1591 ValueToValueMap &PtrToStride,
1592 Value *Ptr, Value *OrigPtr = nullptr) {
1594 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1596 // If there is an entry in the map return the SCEV of the pointer with the
1597 // symbolic stride replaced by one.
1598 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1599 if (SI != PtrToStride.end()) {
1600 Value *StrideVal = SI->second;
1603 StrideVal = stripIntegerCast(StrideVal);
1605 // Replace symbolic stride by one.
1606 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1607 ValueToValueMap RewriteMap;
1608 RewriteMap[StrideVal] = One;
1611 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1612 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1617 // Otherwise, just return the SCEV of the original pointer.
1618 return SE->getSCEV(Ptr);
1621 void RuntimePointerCheck::insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr,
1622 bool WritePtr, unsigned DepSetId,
1623 unsigned ASId, ValueToValueMap &Strides) {
1624 // Get the stride replaced scev.
1625 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1626 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1627 assert(AR && "Invalid addrec expression");
1628 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1629 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1630 Pointers.push_back(Ptr);
1631 Starts.push_back(AR->getStart());
1632 Ends.push_back(ScEnd);
1633 IsWritePtr.push_back(WritePtr);
1634 DependencySetId.push_back(DepSetId);
1635 AliasSetId.push_back(ASId);
1638 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1639 // We need to place the broadcast of invariant variables outside the loop.
1640 Instruction *Instr = dyn_cast<Instruction>(V);
1642 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1643 Instr->getParent()) != LoopVectorBody.end());
1644 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1646 // Place the code for broadcasting invariant variables in the new preheader.
1647 IRBuilder<>::InsertPointGuard Guard(Builder);
1649 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1651 // Broadcast the scalar into all locations in the vector.
1652 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1657 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1659 assert(Val->getType()->isVectorTy() && "Must be a vector");
1660 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1661 "Elem must be an integer");
1662 assert(Step->getType() == Val->getType()->getScalarType() &&
1663 "Step has wrong type");
1664 // Create the types.
1665 Type *ITy = Val->getType()->getScalarType();
1666 VectorType *Ty = cast<VectorType>(Val->getType());
1667 int VLen = Ty->getNumElements();
1668 SmallVector<Constant*, 8> Indices;
1670 // Create a vector of consecutive numbers from zero to VF.
1671 for (int i = 0; i < VLen; ++i)
1672 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1674 // Add the consecutive indices to the vector value.
1675 Constant *Cv = ConstantVector::get(Indices);
1676 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1677 Step = Builder.CreateVectorSplat(VLen, Step);
1678 assert(Step->getType() == Val->getType() && "Invalid step vec");
1679 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1680 // which can be found from the original scalar operations.
1681 Step = Builder.CreateMul(Cv, Step);
1682 return Builder.CreateAdd(Val, Step, "induction");
1685 /// \brief Find the operand of the GEP that should be checked for consecutive
1686 /// stores. This ignores trailing indices that have no effect on the final
1688 static unsigned getGEPInductionOperand(const DataLayout *DL,
1689 const GetElementPtrInst *Gep) {
1690 unsigned LastOperand = Gep->getNumOperands() - 1;
1691 unsigned GEPAllocSize = DL->getTypeAllocSize(
1692 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1694 // Walk backwards and try to peel off zeros.
1695 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1696 // Find the type we're currently indexing into.
1697 gep_type_iterator GEPTI = gep_type_begin(Gep);
1698 std::advance(GEPTI, LastOperand - 1);
1700 // If it's a type with the same allocation size as the result of the GEP we
1701 // can peel off the zero index.
1702 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1710 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1711 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1712 // Make sure that the pointer does not point to structs.
1713 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1716 // If this value is a pointer induction variable we know it is consecutive.
1717 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1718 if (Phi && Inductions.count(Phi)) {
1719 InductionInfo II = Inductions[Phi];
1720 return II.getConsecutiveDirection();
1723 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1727 unsigned NumOperands = Gep->getNumOperands();
1728 Value *GpPtr = Gep->getPointerOperand();
1729 // If this GEP value is a consecutive pointer induction variable and all of
1730 // the indices are constant then we know it is consecutive. We can
1731 Phi = dyn_cast<PHINode>(GpPtr);
1732 if (Phi && Inductions.count(Phi)) {
1734 // Make sure that the pointer does not point to structs.
1735 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1736 if (GepPtrType->getElementType()->isAggregateType())
1739 // Make sure that all of the index operands are loop invariant.
1740 for (unsigned i = 1; i < NumOperands; ++i)
1741 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1744 InductionInfo II = Inductions[Phi];
1745 return II.getConsecutiveDirection();
1748 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1750 // Check that all of the gep indices are uniform except for our induction
1752 for (unsigned i = 0; i != NumOperands; ++i)
1753 if (i != InductionOperand &&
1754 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1757 // We can emit wide load/stores only if the last non-zero index is the
1758 // induction variable.
1759 const SCEV *Last = nullptr;
1760 if (!Strides.count(Gep))
1761 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1763 // Because of the multiplication by a stride we can have a s/zext cast.
1764 // We are going to replace this stride by 1 so the cast is safe to ignore.
1766 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1767 // %0 = trunc i64 %indvars.iv to i32
1768 // %mul = mul i32 %0, %Stride1
1769 // %idxprom = zext i32 %mul to i64 << Safe cast.
1770 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1772 Last = replaceSymbolicStrideSCEV(SE, Strides,
1773 Gep->getOperand(InductionOperand), Gep);
1774 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1776 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1780 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1781 const SCEV *Step = AR->getStepRecurrence(*SE);
1783 // The memory is consecutive because the last index is consecutive
1784 // and all other indices are loop invariant.
1787 if (Step->isAllOnesValue())
1794 bool LoopVectorizationLegality::isUniform(Value *V) {
1795 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1798 InnerLoopVectorizer::VectorParts&
1799 InnerLoopVectorizer::getVectorValue(Value *V) {
1800 assert(V != Induction && "The new induction variable should not be used.");
1801 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1803 // If we have a stride that is replaced by one, do it here.
1804 if (Legal->hasStride(V))
1805 V = ConstantInt::get(V->getType(), 1);
1807 // If we have this scalar in the map, return it.
1808 if (WidenMap.has(V))
1809 return WidenMap.get(V);
1811 // If this scalar is unknown, assume that it is a constant or that it is
1812 // loop invariant. Broadcast V and save the value for future uses.
1813 Value *B = getBroadcastInstrs(V);
1814 return WidenMap.splat(V, B);
1817 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1818 assert(Vec->getType()->isVectorTy() && "Invalid type");
1819 SmallVector<Constant*, 8> ShuffleMask;
1820 for (unsigned i = 0; i < VF; ++i)
1821 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1823 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1824 ConstantVector::get(ShuffleMask),
1828 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1829 // Attempt to issue a wide load.
1830 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1831 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1833 assert((LI || SI) && "Invalid Load/Store instruction");
1835 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1836 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1837 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1838 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1839 // An alignment of 0 means target abi alignment. We need to use the scalar's
1840 // target abi alignment in such a case.
1842 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1843 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1844 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1845 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1847 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1848 !Legal->isMaskRequired(SI))
1849 return scalarizeInstruction(Instr, true);
1851 if (ScalarAllocatedSize != VectorElementSize)
1852 return scalarizeInstruction(Instr);
1854 // If the pointer is loop invariant or if it is non-consecutive,
1855 // scalarize the load.
1856 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1857 bool Reverse = ConsecutiveStride < 0;
1858 bool UniformLoad = LI && Legal->isUniform(Ptr);
1859 if (!ConsecutiveStride || UniformLoad)
1860 return scalarizeInstruction(Instr);
1862 Constant *Zero = Builder.getInt32(0);
1863 VectorParts &Entry = WidenMap.get(Instr);
1865 // Handle consecutive loads/stores.
1866 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1867 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1868 setDebugLocFromInst(Builder, Gep);
1869 Value *PtrOperand = Gep->getPointerOperand();
1870 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1871 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1873 // Create the new GEP with the new induction variable.
1874 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1875 Gep2->setOperand(0, FirstBasePtr);
1876 Gep2->setName("gep.indvar.base");
1877 Ptr = Builder.Insert(Gep2);
1879 setDebugLocFromInst(Builder, Gep);
1880 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1881 OrigLoop) && "Base ptr must be invariant");
1883 // The last index does not have to be the induction. It can be
1884 // consecutive and be a function of the index. For example A[I+1];
1885 unsigned NumOperands = Gep->getNumOperands();
1886 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1887 // Create the new GEP with the new induction variable.
1888 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1890 for (unsigned i = 0; i < NumOperands; ++i) {
1891 Value *GepOperand = Gep->getOperand(i);
1892 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1894 // Update last index or loop invariant instruction anchored in loop.
1895 if (i == InductionOperand ||
1896 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1897 assert((i == InductionOperand ||
1898 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1899 "Must be last index or loop invariant");
1901 VectorParts &GEPParts = getVectorValue(GepOperand);
1902 Value *Index = GEPParts[0];
1903 Index = Builder.CreateExtractElement(Index, Zero);
1904 Gep2->setOperand(i, Index);
1905 Gep2->setName("gep.indvar.idx");
1908 Ptr = Builder.Insert(Gep2);
1910 // Use the induction element ptr.
1911 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1912 setDebugLocFromInst(Builder, Ptr);
1913 VectorParts &PtrVal = getVectorValue(Ptr);
1914 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1917 VectorParts Mask = createBlockInMask(Instr->getParent());
1920 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1921 "We do not allow storing to uniform addresses");
1922 setDebugLocFromInst(Builder, SI);
1923 // We don't want to update the value in the map as it might be used in
1924 // another expression. So don't use a reference type for "StoredVal".
1925 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1927 for (unsigned Part = 0; Part < UF; ++Part) {
1928 // Calculate the pointer for the specific unroll-part.
1929 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1932 // If we store to reverse consecutive memory locations then we need
1933 // to reverse the order of elements in the stored value.
1934 StoredVal[Part] = reverseVector(StoredVal[Part]);
1935 // If the address is consecutive but reversed, then the
1936 // wide store needs to start at the last vector element.
1937 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1938 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1939 Mask[Part] = reverseVector(Mask[Part]);
1942 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1943 DataTy->getPointerTo(AddressSpace));
1946 if (Legal->isMaskRequired(SI))
1947 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1950 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1951 propagateMetadata(NewSI, SI);
1957 assert(LI && "Must have a load instruction");
1958 setDebugLocFromInst(Builder, LI);
1959 for (unsigned Part = 0; Part < UF; ++Part) {
1960 // Calculate the pointer for the specific unroll-part.
1961 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1964 // If the address is consecutive but reversed, then the
1965 // wide load needs to start at the last vector element.
1966 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1967 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1968 Mask[Part] = reverseVector(Mask[Part]);
1972 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1973 DataTy->getPointerTo(AddressSpace));
1974 if (Legal->isMaskRequired(LI))
1975 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1976 UndefValue::get(DataTy),
1977 "wide.masked.load");
1979 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1980 propagateMetadata(NewLI, LI);
1981 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1985 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1986 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1987 // Holds vector parameters or scalars, in case of uniform vals.
1988 SmallVector<VectorParts, 4> Params;
1990 setDebugLocFromInst(Builder, Instr);
1992 // Find all of the vectorized parameters.
1993 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1994 Value *SrcOp = Instr->getOperand(op);
1996 // If we are accessing the old induction variable, use the new one.
1997 if (SrcOp == OldInduction) {
1998 Params.push_back(getVectorValue(SrcOp));
2002 // Try using previously calculated values.
2003 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2005 // If the src is an instruction that appeared earlier in the basic block
2006 // then it should already be vectorized.
2007 if (SrcInst && OrigLoop->contains(SrcInst)) {
2008 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2009 // The parameter is a vector value from earlier.
2010 Params.push_back(WidenMap.get(SrcInst));
2012 // The parameter is a scalar from outside the loop. Maybe even a constant.
2013 VectorParts Scalars;
2014 Scalars.append(UF, SrcOp);
2015 Params.push_back(Scalars);
2019 assert(Params.size() == Instr->getNumOperands() &&
2020 "Invalid number of operands");
2022 // Does this instruction return a value ?
2023 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2025 Value *UndefVec = IsVoidRetTy ? nullptr :
2026 UndefValue::get(VectorType::get(Instr->getType(), VF));
2027 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2028 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2030 Instruction *InsertPt = Builder.GetInsertPoint();
2031 BasicBlock *IfBlock = Builder.GetInsertBlock();
2032 BasicBlock *CondBlock = nullptr;
2035 Loop *VectorLp = nullptr;
2036 if (IfPredicateStore) {
2037 assert(Instr->getParent()->getSinglePredecessor() &&
2038 "Only support single predecessor blocks");
2039 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2040 Instr->getParent());
2041 VectorLp = LI->getLoopFor(IfBlock);
2042 assert(VectorLp && "Must have a loop for this block");
2045 // For each vector unroll 'part':
2046 for (unsigned Part = 0; Part < UF; ++Part) {
2047 // For each scalar that we create:
2048 for (unsigned Width = 0; Width < VF; ++Width) {
2051 Value *Cmp = nullptr;
2052 if (IfPredicateStore) {
2053 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2054 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2055 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2056 LoopVectorBody.push_back(CondBlock);
2057 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2058 // Update Builder with newly created basic block.
2059 Builder.SetInsertPoint(InsertPt);
2062 Instruction *Cloned = Instr->clone();
2064 Cloned->setName(Instr->getName() + ".cloned");
2065 // Replace the operands of the cloned instructions with extracted scalars.
2066 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2067 Value *Op = Params[op][Part];
2068 // Param is a vector. Need to extract the right lane.
2069 if (Op->getType()->isVectorTy())
2070 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2071 Cloned->setOperand(op, Op);
2074 // Place the cloned scalar in the new loop.
2075 Builder.Insert(Cloned);
2077 // If the original scalar returns a value we need to place it in a vector
2078 // so that future users will be able to use it.
2080 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2081 Builder.getInt32(Width));
2083 if (IfPredicateStore) {
2084 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2085 LoopVectorBody.push_back(NewIfBlock);
2086 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2087 Builder.SetInsertPoint(InsertPt);
2088 Instruction *OldBr = IfBlock->getTerminator();
2089 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2090 OldBr->eraseFromParent();
2091 IfBlock = NewIfBlock;
2097 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2101 if (Instruction *I = dyn_cast<Instruction>(V))
2102 return I->getParent() == Loc->getParent() ? I : nullptr;
2106 std::pair<Instruction *, Instruction *>
2107 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2108 Instruction *tnullptr = nullptr;
2109 if (!Legal->mustCheckStrides())
2110 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2112 IRBuilder<> ChkBuilder(Loc);
2115 Value *Check = nullptr;
2116 Instruction *FirstInst = nullptr;
2117 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2118 SE = Legal->strides_end();
2120 Value *Ptr = stripIntegerCast(*SI);
2121 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2123 // Store the first instruction we create.
2124 FirstInst = getFirstInst(FirstInst, C, Loc);
2126 Check = ChkBuilder.CreateOr(Check, C);
2131 // We have to do this trickery because the IRBuilder might fold the check to a
2132 // constant expression in which case there is no Instruction anchored in a
2134 LLVMContext &Ctx = Loc->getContext();
2135 Instruction *TheCheck =
2136 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2137 ChkBuilder.Insert(TheCheck, "stride.not.one");
2138 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2140 return std::make_pair(FirstInst, TheCheck);
2143 std::pair<Instruction *, Instruction *>
2144 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2145 RuntimePointerCheck *PtrRtCheck = Legal->getRuntimePointerCheck();
2147 Instruction *tnullptr = nullptr;
2148 if (!PtrRtCheck->Need)
2149 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2151 unsigned NumPointers = PtrRtCheck->Pointers.size();
2152 SmallVector<TrackingVH<Value> , 2> Starts;
2153 SmallVector<TrackingVH<Value> , 2> Ends;
2155 LLVMContext &Ctx = Loc->getContext();
2156 SCEVExpander Exp(*SE, "induction");
2157 Instruction *FirstInst = nullptr;
2159 for (unsigned i = 0; i < NumPointers; ++i) {
2160 Value *Ptr = PtrRtCheck->Pointers[i];
2161 const SCEV *Sc = SE->getSCEV(Ptr);
2163 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2164 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2166 Starts.push_back(Ptr);
2167 Ends.push_back(Ptr);
2169 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2170 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2172 // Use this type for pointer arithmetic.
2173 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2175 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2176 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2177 Starts.push_back(Start);
2178 Ends.push_back(End);
2182 IRBuilder<> ChkBuilder(Loc);
2183 // Our instructions might fold to a constant.
2184 Value *MemoryRuntimeCheck = nullptr;
2185 for (unsigned i = 0; i < NumPointers; ++i) {
2186 for (unsigned j = i+1; j < NumPointers; ++j) {
2187 // No need to check if two readonly pointers intersect.
2188 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2191 // Only need to check pointers between two different dependency sets.
2192 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2194 // Only need to check pointers in the same alias set.
2195 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2198 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2199 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2201 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2202 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2203 "Trying to bounds check pointers with different address spaces");
2205 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2206 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2208 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2209 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2210 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2211 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2213 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2214 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2215 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2216 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2217 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2218 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2219 if (MemoryRuntimeCheck) {
2220 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2222 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2224 MemoryRuntimeCheck = IsConflict;
2228 // We have to do this trickery because the IRBuilder might fold the check to a
2229 // constant expression in which case there is no Instruction anchored in a
2231 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2232 ConstantInt::getTrue(Ctx));
2233 ChkBuilder.Insert(Check, "memcheck.conflict");
2234 FirstInst = getFirstInst(FirstInst, Check, Loc);
2235 return std::make_pair(FirstInst, Check);
2238 void InnerLoopVectorizer::createEmptyLoop() {
2240 In this function we generate a new loop. The new loop will contain
2241 the vectorized instructions while the old loop will continue to run the
2244 [ ] <-- Back-edge taken count overflow check.
2247 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2250 || [ ] <-- vector pre header.
2254 || [ ]_| <-- vector loop.
2257 | >[ ] <--- middle-block.
2260 -|- >[ ] <--- new preheader.
2264 | [ ]_| <-- old scalar loop to handle remainder.
2267 >[ ] <-- exit block.
2271 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2272 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2273 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2274 assert(BypassBlock && "Invalid loop structure");
2275 assert(ExitBlock && "Must have an exit block");
2277 // Some loops have a single integer induction variable, while other loops
2278 // don't. One example is c++ iterators that often have multiple pointer
2279 // induction variables. In the code below we also support a case where we
2280 // don't have a single induction variable.
2281 OldInduction = Legal->getInduction();
2282 Type *IdxTy = Legal->getWidestInductionType();
2284 // Find the loop boundaries.
2285 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2286 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2288 // The exit count might have the type of i64 while the phi is i32. This can
2289 // happen if we have an induction variable that is sign extended before the
2290 // compare. The only way that we get a backedge taken count is that the
2291 // induction variable was signed and as such will not overflow. In such a case
2292 // truncation is legal.
2293 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2294 IdxTy->getPrimitiveSizeInBits())
2295 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2297 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2298 // Get the total trip count from the count by adding 1.
2299 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2300 SE->getConstant(BackedgeTakeCount->getType(), 1));
2302 // Expand the trip count and place the new instructions in the preheader.
2303 // Notice that the pre-header does not change, only the loop body.
2304 SCEVExpander Exp(*SE, "induction");
2306 // We need to test whether the backedge-taken count is uint##_max. Adding one
2307 // to it will cause overflow and an incorrect loop trip count in the vector
2308 // body. In case of overflow we want to directly jump to the scalar remainder
2310 Value *BackedgeCount =
2311 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2312 BypassBlock->getTerminator());
2313 if (BackedgeCount->getType()->isPointerTy())
2314 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2315 "backedge.ptrcnt.to.int",
2316 BypassBlock->getTerminator());
2317 Instruction *CheckBCOverflow =
2318 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2319 Constant::getAllOnesValue(BackedgeCount->getType()),
2320 "backedge.overflow", BypassBlock->getTerminator());
2322 // The loop index does not have to start at Zero. Find the original start
2323 // value from the induction PHI node. If we don't have an induction variable
2324 // then we know that it starts at zero.
2325 Builder.SetInsertPoint(BypassBlock->getTerminator());
2326 Value *StartIdx = ExtendedIdx = OldInduction ?
2327 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2329 ConstantInt::get(IdxTy, 0);
2331 // We need an instruction to anchor the overflow check on. StartIdx needs to
2332 // be defined before the overflow check branch. Because the scalar preheader
2333 // is going to merge the start index and so the overflow branch block needs to
2334 // contain a definition of the start index.
2335 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2336 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2337 BypassBlock->getTerminator());
2339 // Count holds the overall loop count (N).
2340 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2341 BypassBlock->getTerminator());
2343 LoopBypassBlocks.push_back(BypassBlock);
2345 // Split the single block loop into the two loop structure described above.
2346 BasicBlock *VectorPH =
2347 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2348 BasicBlock *VecBody =
2349 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2350 BasicBlock *MiddleBlock =
2351 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2352 BasicBlock *ScalarPH =
2353 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2355 // Create and register the new vector loop.
2356 Loop* Lp = new Loop();
2357 Loop *ParentLoop = OrigLoop->getParentLoop();
2359 // Insert the new loop into the loop nest and register the new basic blocks
2360 // before calling any utilities such as SCEV that require valid LoopInfo.
2362 ParentLoop->addChildLoop(Lp);
2363 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2364 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2365 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2367 LI->addTopLevelLoop(Lp);
2369 Lp->addBasicBlockToLoop(VecBody, *LI);
2371 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2373 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2375 // Generate the induction variable.
2376 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2377 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2378 // The loop step is equal to the vectorization factor (num of SIMD elements)
2379 // times the unroll factor (num of SIMD instructions).
2380 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2382 // This is the IR builder that we use to add all of the logic for bypassing
2383 // the new vector loop.
2384 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2385 setDebugLocFromInst(BypassBuilder,
2386 getDebugLocFromInstOrOperands(OldInduction));
2388 // We may need to extend the index in case there is a type mismatch.
2389 // We know that the count starts at zero and does not overflow.
2390 if (Count->getType() != IdxTy) {
2391 // The exit count can be of pointer type. Convert it to the correct
2393 if (ExitCount->getType()->isPointerTy())
2394 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2396 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2399 // Add the start index to the loop count to get the new end index.
2400 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2402 // Now we need to generate the expression for N - (N % VF), which is
2403 // the part that the vectorized body will execute.
2404 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2405 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2406 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2407 "end.idx.rnd.down");
2409 // Now, compare the new count to zero. If it is zero skip the vector loop and
2410 // jump to the scalar loop.
2412 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2414 BasicBlock *LastBypassBlock = BypassBlock;
2416 // Generate code to check that the loops trip count that we computed by adding
2417 // one to the backedge-taken count will not overflow.
2419 auto PastOverflowCheck =
2420 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2421 BasicBlock *CheckBlock =
2422 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2424 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2425 LoopBypassBlocks.push_back(CheckBlock);
2426 Instruction *OldTerm = LastBypassBlock->getTerminator();
2427 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2428 OldTerm->eraseFromParent();
2429 LastBypassBlock = CheckBlock;
2432 // Generate the code to check that the strides we assumed to be one are really
2433 // one. We want the new basic block to start at the first instruction in a
2434 // sequence of instructions that form a check.
2435 Instruction *StrideCheck;
2436 Instruction *FirstCheckInst;
2437 std::tie(FirstCheckInst, StrideCheck) =
2438 addStrideCheck(LastBypassBlock->getTerminator());
2440 // Create a new block containing the stride check.
2441 BasicBlock *CheckBlock =
2442 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2444 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2445 LoopBypassBlocks.push_back(CheckBlock);
2447 // Replace the branch into the memory check block with a conditional branch
2448 // for the "few elements case".
2449 Instruction *OldTerm = LastBypassBlock->getTerminator();
2450 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2451 OldTerm->eraseFromParent();
2454 LastBypassBlock = CheckBlock;
2457 // Generate the code that checks in runtime if arrays overlap. We put the
2458 // checks into a separate block to make the more common case of few elements
2460 Instruction *MemRuntimeCheck;
2461 std::tie(FirstCheckInst, MemRuntimeCheck) =
2462 addRuntimeCheck(LastBypassBlock->getTerminator());
2463 if (MemRuntimeCheck) {
2464 // Create a new block containing the memory check.
2465 BasicBlock *CheckBlock =
2466 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2468 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2469 LoopBypassBlocks.push_back(CheckBlock);
2471 // Replace the branch into the memory check block with a conditional branch
2472 // for the "few elements case".
2473 Instruction *OldTerm = LastBypassBlock->getTerminator();
2474 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2475 OldTerm->eraseFromParent();
2477 Cmp = MemRuntimeCheck;
2478 LastBypassBlock = CheckBlock;
2481 LastBypassBlock->getTerminator()->eraseFromParent();
2482 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2485 // We are going to resume the execution of the scalar loop.
2486 // Go over all of the induction variables that we found and fix the
2487 // PHIs that are left in the scalar version of the loop.
2488 // The starting values of PHI nodes depend on the counter of the last
2489 // iteration in the vectorized loop.
2490 // If we come from a bypass edge then we need to start from the original
2493 // This variable saves the new starting index for the scalar loop.
2494 PHINode *ResumeIndex = nullptr;
2495 LoopVectorizationLegality::InductionList::iterator I, E;
2496 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2497 // Set builder to point to last bypass block.
2498 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2499 for (I = List->begin(), E = List->end(); I != E; ++I) {
2500 PHINode *OrigPhi = I->first;
2501 LoopVectorizationLegality::InductionInfo II = I->second;
2503 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2504 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2505 MiddleBlock->getTerminator());
2506 // We might have extended the type of the induction variable but we need a
2507 // truncated version for the scalar loop.
2508 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2509 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2510 MiddleBlock->getTerminator()) : nullptr;
2512 // Create phi nodes to merge from the backedge-taken check block.
2513 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2514 ScalarPH->getTerminator());
2515 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2517 PHINode *BCTruncResumeVal = nullptr;
2518 if (OrigPhi == OldInduction) {
2520 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2521 ScalarPH->getTerminator());
2522 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2525 Value *EndValue = nullptr;
2527 case LoopVectorizationLegality::IK_NoInduction:
2528 llvm_unreachable("Unknown induction");
2529 case LoopVectorizationLegality::IK_IntInduction: {
2530 // Handle the integer induction counter.
2531 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2533 // We have the canonical induction variable.
2534 if (OrigPhi == OldInduction) {
2535 // Create a truncated version of the resume value for the scalar loop,
2536 // we might have promoted the type to a larger width.
2538 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2539 // The new PHI merges the original incoming value, in case of a bypass,
2540 // or the value at the end of the vectorized loop.
2541 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2542 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2543 TruncResumeVal->addIncoming(EndValue, VecBody);
2545 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2547 // We know what the end value is.
2548 EndValue = IdxEndRoundDown;
2549 // We also know which PHI node holds it.
2550 ResumeIndex = ResumeVal;
2554 // Not the canonical induction variable - add the vector loop count to the
2556 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2557 II.StartValue->getType(),
2559 EndValue = II.transform(BypassBuilder, CRD);
2560 EndValue->setName("ind.end");
2563 case LoopVectorizationLegality::IK_PtrInduction: {
2564 EndValue = II.transform(BypassBuilder, CountRoundDown);
2565 EndValue->setName("ptr.ind.end");
2570 // The new PHI merges the original incoming value, in case of a bypass,
2571 // or the value at the end of the vectorized loop.
2572 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2573 if (OrigPhi == OldInduction)
2574 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2576 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2578 ResumeVal->addIncoming(EndValue, VecBody);
2580 // Fix the scalar body counter (PHI node).
2581 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2583 // The old induction's phi node in the scalar body needs the truncated
2585 if (OrigPhi == OldInduction) {
2586 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2587 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2589 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2590 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2594 // If we are generating a new induction variable then we also need to
2595 // generate the code that calculates the exit value. This value is not
2596 // simply the end of the counter because we may skip the vectorized body
2597 // in case of a runtime check.
2599 assert(!ResumeIndex && "Unexpected resume value found");
2600 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2601 MiddleBlock->getTerminator());
2602 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2603 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2604 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2607 // Make sure that we found the index where scalar loop needs to continue.
2608 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2609 "Invalid resume Index");
2611 // Add a check in the middle block to see if we have completed
2612 // all of the iterations in the first vector loop.
2613 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2614 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2615 ResumeIndex, "cmp.n",
2616 MiddleBlock->getTerminator());
2618 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2619 // Remove the old terminator.
2620 MiddleBlock->getTerminator()->eraseFromParent();
2622 // Create i+1 and fill the PHINode.
2623 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2624 Induction->addIncoming(StartIdx, VectorPH);
2625 Induction->addIncoming(NextIdx, VecBody);
2626 // Create the compare.
2627 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2628 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2630 // Now we have two terminators. Remove the old one from the block.
2631 VecBody->getTerminator()->eraseFromParent();
2633 // Get ready to start creating new instructions into the vectorized body.
2634 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2637 LoopVectorPreHeader = VectorPH;
2638 LoopScalarPreHeader = ScalarPH;
2639 LoopMiddleBlock = MiddleBlock;
2640 LoopExitBlock = ExitBlock;
2641 LoopVectorBody.push_back(VecBody);
2642 LoopScalarBody = OldBasicBlock;
2644 LoopVectorizeHints Hints(Lp, true);
2645 Hints.setAlreadyVectorized();
2648 /// This function returns the identity element (or neutral element) for
2649 /// the operation K.
2651 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2656 // Adding, Xoring, Oring zero to a number does not change it.
2657 return ConstantInt::get(Tp, 0);
2658 case RK_IntegerMult:
2659 // Multiplying a number by 1 does not change it.
2660 return ConstantInt::get(Tp, 1);
2662 // AND-ing a number with an all-1 value does not change it.
2663 return ConstantInt::get(Tp, -1, true);
2665 // Multiplying a number by 1 does not change it.
2666 return ConstantFP::get(Tp, 1.0L);
2668 // Adding zero to a number does not change it.
2669 return ConstantFP::get(Tp, 0.0L);
2671 llvm_unreachable("Unknown reduction kind");
2675 /// This function translates the reduction kind to an LLVM binary operator.
2677 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2679 case LoopVectorizationLegality::RK_IntegerAdd:
2680 return Instruction::Add;
2681 case LoopVectorizationLegality::RK_IntegerMult:
2682 return Instruction::Mul;
2683 case LoopVectorizationLegality::RK_IntegerOr:
2684 return Instruction::Or;
2685 case LoopVectorizationLegality::RK_IntegerAnd:
2686 return Instruction::And;
2687 case LoopVectorizationLegality::RK_IntegerXor:
2688 return Instruction::Xor;
2689 case LoopVectorizationLegality::RK_FloatMult:
2690 return Instruction::FMul;
2691 case LoopVectorizationLegality::RK_FloatAdd:
2692 return Instruction::FAdd;
2693 case LoopVectorizationLegality::RK_IntegerMinMax:
2694 return Instruction::ICmp;
2695 case LoopVectorizationLegality::RK_FloatMinMax:
2696 return Instruction::FCmp;
2698 llvm_unreachable("Unknown reduction operation");
2702 Value *createMinMaxOp(IRBuilder<> &Builder,
2703 LoopVectorizationLegality::MinMaxReductionKind RK,
2706 CmpInst::Predicate P = CmpInst::ICMP_NE;
2709 llvm_unreachable("Unknown min/max reduction kind");
2710 case LoopVectorizationLegality::MRK_UIntMin:
2711 P = CmpInst::ICMP_ULT;
2713 case LoopVectorizationLegality::MRK_UIntMax:
2714 P = CmpInst::ICMP_UGT;
2716 case LoopVectorizationLegality::MRK_SIntMin:
2717 P = CmpInst::ICMP_SLT;
2719 case LoopVectorizationLegality::MRK_SIntMax:
2720 P = CmpInst::ICMP_SGT;
2722 case LoopVectorizationLegality::MRK_FloatMin:
2723 P = CmpInst::FCMP_OLT;
2725 case LoopVectorizationLegality::MRK_FloatMax:
2726 P = CmpInst::FCMP_OGT;
2731 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2732 RK == LoopVectorizationLegality::MRK_FloatMax)
2733 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2735 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2737 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2742 struct CSEDenseMapInfo {
2743 static bool canHandle(Instruction *I) {
2744 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2745 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2747 static inline Instruction *getEmptyKey() {
2748 return DenseMapInfo<Instruction *>::getEmptyKey();
2750 static inline Instruction *getTombstoneKey() {
2751 return DenseMapInfo<Instruction *>::getTombstoneKey();
2753 static unsigned getHashValue(Instruction *I) {
2754 assert(canHandle(I) && "Unknown instruction!");
2755 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2756 I->value_op_end()));
2758 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2759 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2760 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2762 return LHS->isIdenticalTo(RHS);
2767 /// \brief Check whether this block is a predicated block.
2768 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2769 /// = ...; " blocks. We start with one vectorized basic block. For every
2770 /// conditional block we split this vectorized block. Therefore, every second
2771 /// block will be a predicated one.
2772 static bool isPredicatedBlock(unsigned BlockNum) {
2773 return BlockNum % 2;
2776 ///\brief Perform cse of induction variable instructions.
2777 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2778 // Perform simple cse.
2779 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2780 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2781 BasicBlock *BB = BBs[i];
2782 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2783 Instruction *In = I++;
2785 if (!CSEDenseMapInfo::canHandle(In))
2788 // Check if we can replace this instruction with any of the
2789 // visited instructions.
2790 if (Instruction *V = CSEMap.lookup(In)) {
2791 In->replaceAllUsesWith(V);
2792 In->eraseFromParent();
2795 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2796 // ...;" blocks for predicated stores. Every second block is a predicated
2798 if (isPredicatedBlock(i))
2806 /// \brief Adds a 'fast' flag to floating point operations.
2807 static Value *addFastMathFlag(Value *V) {
2808 if (isa<FPMathOperator>(V)){
2809 FastMathFlags Flags;
2810 Flags.setUnsafeAlgebra();
2811 cast<Instruction>(V)->setFastMathFlags(Flags);
2816 void InnerLoopVectorizer::vectorizeLoop() {
2817 //===------------------------------------------------===//
2819 // Notice: any optimization or new instruction that go
2820 // into the code below should be also be implemented in
2823 //===------------------------------------------------===//
2824 Constant *Zero = Builder.getInt32(0);
2826 // In order to support reduction variables we need to be able to vectorize
2827 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2828 // stages. First, we create a new vector PHI node with no incoming edges.
2829 // We use this value when we vectorize all of the instructions that use the
2830 // PHI. Next, after all of the instructions in the block are complete we
2831 // add the new incoming edges to the PHI. At this point all of the
2832 // instructions in the basic block are vectorized, so we can use them to
2833 // construct the PHI.
2834 PhiVector RdxPHIsToFix;
2836 // Scan the loop in a topological order to ensure that defs are vectorized
2838 LoopBlocksDFS DFS(OrigLoop);
2841 // Vectorize all of the blocks in the original loop.
2842 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2843 be = DFS.endRPO(); bb != be; ++bb)
2844 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2846 // At this point every instruction in the original loop is widened to
2847 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2848 // that we vectorized. The PHI nodes are currently empty because we did
2849 // not want to introduce cycles. Notice that the remaining PHI nodes
2850 // that we need to fix are reduction variables.
2852 // Create the 'reduced' values for each of the induction vars.
2853 // The reduced values are the vector values that we scalarize and combine
2854 // after the loop is finished.
2855 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2857 PHINode *RdxPhi = *it;
2858 assert(RdxPhi && "Unable to recover vectorized PHI");
2860 // Find the reduction variable descriptor.
2861 assert(Legal->getReductionVars()->count(RdxPhi) &&
2862 "Unable to find the reduction variable");
2863 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2864 (*Legal->getReductionVars())[RdxPhi];
2866 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2868 // We need to generate a reduction vector from the incoming scalar.
2869 // To do so, we need to generate the 'identity' vector and override
2870 // one of the elements with the incoming scalar reduction. We need
2871 // to do it in the vector-loop preheader.
2872 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2874 // This is the vector-clone of the value that leaves the loop.
2875 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2876 Type *VecTy = VectorExit[0]->getType();
2878 // Find the reduction identity variable. Zero for addition, or, xor,
2879 // one for multiplication, -1 for And.
2882 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2883 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2884 // MinMax reduction have the start value as their identify.
2886 VectorStart = Identity = RdxDesc.StartValue;
2888 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2893 // Handle other reduction kinds:
2895 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2896 VecTy->getScalarType());
2899 // This vector is the Identity vector where the first element is the
2900 // incoming scalar reduction.
2901 VectorStart = RdxDesc.StartValue;
2903 Identity = ConstantVector::getSplat(VF, Iden);
2905 // This vector is the Identity vector where the first element is the
2906 // incoming scalar reduction.
2907 VectorStart = Builder.CreateInsertElement(Identity,
2908 RdxDesc.StartValue, Zero);
2912 // Fix the vector-loop phi.
2914 // Reductions do not have to start at zero. They can start with
2915 // any loop invariant values.
2916 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2917 BasicBlock *Latch = OrigLoop->getLoopLatch();
2918 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2919 VectorParts &Val = getVectorValue(LoopVal);
2920 for (unsigned part = 0; part < UF; ++part) {
2921 // Make sure to add the reduction stat value only to the
2922 // first unroll part.
2923 Value *StartVal = (part == 0) ? VectorStart : Identity;
2924 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2925 LoopVectorPreHeader);
2926 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2927 LoopVectorBody.back());
2930 // Before each round, move the insertion point right between
2931 // the PHIs and the values we are going to write.
2932 // This allows us to write both PHINodes and the extractelement
2934 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2936 VectorParts RdxParts;
2937 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2938 for (unsigned part = 0; part < UF; ++part) {
2939 // This PHINode contains the vectorized reduction variable, or
2940 // the initial value vector, if we bypass the vector loop.
2941 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2942 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2943 Value *StartVal = (part == 0) ? VectorStart : Identity;
2944 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2945 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2946 NewPhi->addIncoming(RdxExitVal[part],
2947 LoopVectorBody.back());
2948 RdxParts.push_back(NewPhi);
2951 // Reduce all of the unrolled parts into a single vector.
2952 Value *ReducedPartRdx = RdxParts[0];
2953 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2954 setDebugLocFromInst(Builder, ReducedPartRdx);
2955 for (unsigned part = 1; part < UF; ++part) {
2956 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2957 // Floating point operations had to be 'fast' to enable the reduction.
2958 ReducedPartRdx = addFastMathFlag(
2959 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2960 ReducedPartRdx, "bin.rdx"));
2962 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2963 ReducedPartRdx, RdxParts[part]);
2967 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2968 // and vector ops, reducing the set of values being computed by half each
2970 assert(isPowerOf2_32(VF) &&
2971 "Reduction emission only supported for pow2 vectors!");
2972 Value *TmpVec = ReducedPartRdx;
2973 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2974 for (unsigned i = VF; i != 1; i >>= 1) {
2975 // Move the upper half of the vector to the lower half.
2976 for (unsigned j = 0; j != i/2; ++j)
2977 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2979 // Fill the rest of the mask with undef.
2980 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2981 UndefValue::get(Builder.getInt32Ty()));
2984 Builder.CreateShuffleVector(TmpVec,
2985 UndefValue::get(TmpVec->getType()),
2986 ConstantVector::get(ShuffleMask),
2989 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2990 // Floating point operations had to be 'fast' to enable the reduction.
2991 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2992 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2994 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2997 // The result is in the first element of the vector.
2998 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2999 Builder.getInt32(0));
3002 // Create a phi node that merges control-flow from the backedge-taken check
3003 // block and the middle block.
3004 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3005 LoopScalarPreHeader->getTerminator());
3006 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
3007 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3009 // Now, we need to fix the users of the reduction variable
3010 // inside and outside of the scalar remainder loop.
3011 // We know that the loop is in LCSSA form. We need to update the
3012 // PHI nodes in the exit blocks.
3013 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3014 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3015 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3016 if (!LCSSAPhi) break;
3018 // All PHINodes need to have a single entry edge, or two if
3019 // we already fixed them.
3020 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3022 // We found our reduction value exit-PHI. Update it with the
3023 // incoming bypass edge.
3024 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
3025 // Add an edge coming from the bypass.
3026 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3029 }// end of the LCSSA phi scan.
3031 // Fix the scalar loop reduction variable with the incoming reduction sum
3032 // from the vector body and from the backedge value.
3033 int IncomingEdgeBlockIdx =
3034 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3035 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3036 // Pick the other block.
3037 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3038 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3039 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
3040 }// end of for each redux variable.
3044 // Remove redundant induction instructions.
3045 cse(LoopVectorBody);
3048 void InnerLoopVectorizer::fixLCSSAPHIs() {
3049 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3050 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3051 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3052 if (!LCSSAPhi) break;
3053 if (LCSSAPhi->getNumIncomingValues() == 1)
3054 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3059 InnerLoopVectorizer::VectorParts
3060 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3061 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3064 // Look for cached value.
3065 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3066 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3067 if (ECEntryIt != MaskCache.end())
3068 return ECEntryIt->second;
3070 VectorParts SrcMask = createBlockInMask(Src);
3072 // The terminator has to be a branch inst!
3073 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3074 assert(BI && "Unexpected terminator found");
3076 if (BI->isConditional()) {
3077 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3079 if (BI->getSuccessor(0) != Dst)
3080 for (unsigned part = 0; part < UF; ++part)
3081 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3083 for (unsigned part = 0; part < UF; ++part)
3084 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3086 MaskCache[Edge] = EdgeMask;
3090 MaskCache[Edge] = SrcMask;
3094 InnerLoopVectorizer::VectorParts
3095 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3096 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3098 // Loop incoming mask is all-one.
3099 if (OrigLoop->getHeader() == BB) {
3100 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3101 return getVectorValue(C);
3104 // This is the block mask. We OR all incoming edges, and with zero.
3105 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3106 VectorParts BlockMask = getVectorValue(Zero);
3109 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3110 VectorParts EM = createEdgeMask(*it, BB);
3111 for (unsigned part = 0; part < UF; ++part)
3112 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3118 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3119 InnerLoopVectorizer::VectorParts &Entry,
3120 unsigned UF, unsigned VF, PhiVector *PV) {
3121 PHINode* P = cast<PHINode>(PN);
3122 // Handle reduction variables:
3123 if (Legal->getReductionVars()->count(P)) {
3124 for (unsigned part = 0; part < UF; ++part) {
3125 // This is phase one of vectorizing PHIs.
3126 Type *VecTy = (VF == 1) ? PN->getType() :
3127 VectorType::get(PN->getType(), VF);
3128 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3129 LoopVectorBody.back()-> getFirstInsertionPt());
3135 setDebugLocFromInst(Builder, P);
3136 // Check for PHI nodes that are lowered to vector selects.
3137 if (P->getParent() != OrigLoop->getHeader()) {
3138 // We know that all PHIs in non-header blocks are converted into
3139 // selects, so we don't have to worry about the insertion order and we
3140 // can just use the builder.
3141 // At this point we generate the predication tree. There may be
3142 // duplications since this is a simple recursive scan, but future
3143 // optimizations will clean it up.
3145 unsigned NumIncoming = P->getNumIncomingValues();
3147 // Generate a sequence of selects of the form:
3148 // SELECT(Mask3, In3,
3149 // SELECT(Mask2, In2,
3151 for (unsigned In = 0; In < NumIncoming; In++) {
3152 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3154 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3156 for (unsigned part = 0; part < UF; ++part) {
3157 // We might have single edge PHIs (blocks) - use an identity
3158 // 'select' for the first PHI operand.
3160 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3163 // Select between the current value and the previous incoming edge
3164 // based on the incoming mask.
3165 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3166 Entry[part], "predphi");
3172 // This PHINode must be an induction variable.
3173 // Make sure that we know about it.
3174 assert(Legal->getInductionVars()->count(P) &&
3175 "Not an induction variable");
3177 LoopVectorizationLegality::InductionInfo II =
3178 Legal->getInductionVars()->lookup(P);
3180 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3181 // which can be found from the original scalar operations.
3183 case LoopVectorizationLegality::IK_NoInduction:
3184 llvm_unreachable("Unknown induction");
3185 case LoopVectorizationLegality::IK_IntInduction: {
3186 assert(P->getType() == II.StartValue->getType() && "Types must match");
3187 Type *PhiTy = P->getType();
3189 if (P == OldInduction) {
3190 // Handle the canonical induction variable. We might have had to
3192 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3194 // Handle other induction variables that are now based on the
3196 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3198 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3199 Broadcasted = II.transform(Builder, NormalizedIdx);
3200 Broadcasted->setName("offset.idx");
3202 Broadcasted = getBroadcastInstrs(Broadcasted);
3203 // After broadcasting the induction variable we need to make the vector
3204 // consecutive by adding 0, 1, 2, etc.
3205 for (unsigned part = 0; part < UF; ++part)
3206 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3209 case LoopVectorizationLegality::IK_PtrInduction:
3210 // Handle the pointer induction variable case.
3211 assert(P->getType()->isPointerTy() && "Unexpected type.");
3212 // This is the normalized GEP that starts counting at zero.
3213 Value *NormalizedIdx =
3214 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3215 // This is the vector of results. Notice that we don't generate
3216 // vector geps because scalar geps result in better code.
3217 for (unsigned part = 0; part < UF; ++part) {
3219 int EltIndex = part;
3220 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3221 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3222 Value *SclrGep = II.transform(Builder, GlobalIdx);
3223 SclrGep->setName("next.gep");
3224 Entry[part] = SclrGep;
3228 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3229 for (unsigned int i = 0; i < VF; ++i) {
3230 int EltIndex = i + part * VF;
3231 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3232 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3233 Value *SclrGep = II.transform(Builder, GlobalIdx);
3234 SclrGep->setName("next.gep");
3235 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3236 Builder.getInt32(i),
3239 Entry[part] = VecVal;
3245 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3246 // For each instruction in the old loop.
3247 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3248 VectorParts &Entry = WidenMap.get(it);
3249 switch (it->getOpcode()) {
3250 case Instruction::Br:
3251 // Nothing to do for PHIs and BR, since we already took care of the
3252 // loop control flow instructions.
3254 case Instruction::PHI: {
3255 // Vectorize PHINodes.
3256 widenPHIInstruction(it, Entry, UF, VF, PV);
3260 case Instruction::Add:
3261 case Instruction::FAdd:
3262 case Instruction::Sub:
3263 case Instruction::FSub:
3264 case Instruction::Mul:
3265 case Instruction::FMul:
3266 case Instruction::UDiv:
3267 case Instruction::SDiv:
3268 case Instruction::FDiv:
3269 case Instruction::URem:
3270 case Instruction::SRem:
3271 case Instruction::FRem:
3272 case Instruction::Shl:
3273 case Instruction::LShr:
3274 case Instruction::AShr:
3275 case Instruction::And:
3276 case Instruction::Or:
3277 case Instruction::Xor: {
3278 // Just widen binops.
3279 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3280 setDebugLocFromInst(Builder, BinOp);
3281 VectorParts &A = getVectorValue(it->getOperand(0));
3282 VectorParts &B = getVectorValue(it->getOperand(1));
3284 // Use this vector value for all users of the original instruction.
3285 for (unsigned Part = 0; Part < UF; ++Part) {
3286 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3288 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3289 VecOp->copyIRFlags(BinOp);
3294 propagateMetadata(Entry, it);
3297 case Instruction::Select: {
3299 // If the selector is loop invariant we can create a select
3300 // instruction with a scalar condition. Otherwise, use vector-select.
3301 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3303 setDebugLocFromInst(Builder, it);
3305 // The condition can be loop invariant but still defined inside the
3306 // loop. This means that we can't just use the original 'cond' value.
3307 // We have to take the 'vectorized' value and pick the first lane.
3308 // Instcombine will make this a no-op.
3309 VectorParts &Cond = getVectorValue(it->getOperand(0));
3310 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3311 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3313 Value *ScalarCond = (VF == 1) ? Cond[0] :
3314 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3316 for (unsigned Part = 0; Part < UF; ++Part) {
3317 Entry[Part] = Builder.CreateSelect(
3318 InvariantCond ? ScalarCond : Cond[Part],
3323 propagateMetadata(Entry, it);
3327 case Instruction::ICmp:
3328 case Instruction::FCmp: {
3329 // Widen compares. Generate vector compares.
3330 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3331 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3332 setDebugLocFromInst(Builder, it);
3333 VectorParts &A = getVectorValue(it->getOperand(0));
3334 VectorParts &B = getVectorValue(it->getOperand(1));
3335 for (unsigned Part = 0; Part < UF; ++Part) {
3338 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3340 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3344 propagateMetadata(Entry, it);
3348 case Instruction::Store:
3349 case Instruction::Load:
3350 vectorizeMemoryInstruction(it);
3352 case Instruction::ZExt:
3353 case Instruction::SExt:
3354 case Instruction::FPToUI:
3355 case Instruction::FPToSI:
3356 case Instruction::FPExt:
3357 case Instruction::PtrToInt:
3358 case Instruction::IntToPtr:
3359 case Instruction::SIToFP:
3360 case Instruction::UIToFP:
3361 case Instruction::Trunc:
3362 case Instruction::FPTrunc:
3363 case Instruction::BitCast: {
3364 CastInst *CI = dyn_cast<CastInst>(it);
3365 setDebugLocFromInst(Builder, it);
3366 /// Optimize the special case where the source is the induction
3367 /// variable. Notice that we can only optimize the 'trunc' case
3368 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3369 /// c. other casts depend on pointer size.
3370 if (CI->getOperand(0) == OldInduction &&
3371 it->getOpcode() == Instruction::Trunc) {
3372 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3374 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3375 LoopVectorizationLegality::InductionInfo II =
3376 Legal->getInductionVars()->lookup(OldInduction);
3378 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3379 for (unsigned Part = 0; Part < UF; ++Part)
3380 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3381 propagateMetadata(Entry, it);
3384 /// Vectorize casts.
3385 Type *DestTy = (VF == 1) ? CI->getType() :
3386 VectorType::get(CI->getType(), VF);
3388 VectorParts &A = getVectorValue(it->getOperand(0));
3389 for (unsigned Part = 0; Part < UF; ++Part)
3390 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3391 propagateMetadata(Entry, it);
3395 case Instruction::Call: {
3396 // Ignore dbg intrinsics.
3397 if (isa<DbgInfoIntrinsic>(it))
3399 setDebugLocFromInst(Builder, it);
3401 Module *M = BB->getParent()->getParent();
3402 CallInst *CI = cast<CallInst>(it);
3403 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3404 assert(ID && "Not an intrinsic call!");
3406 case Intrinsic::assume:
3407 case Intrinsic::lifetime_end:
3408 case Intrinsic::lifetime_start:
3409 scalarizeInstruction(it);
3412 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3413 for (unsigned Part = 0; Part < UF; ++Part) {
3414 SmallVector<Value *, 4> Args;
3415 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3416 if (HasScalarOpd && i == 1) {
3417 Args.push_back(CI->getArgOperand(i));
3420 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3421 Args.push_back(Arg[Part]);
3423 Type *Tys[] = {CI->getType()};
3425 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3427 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3428 Entry[Part] = Builder.CreateCall(F, Args);
3431 propagateMetadata(Entry, it);
3438 // All other instructions are unsupported. Scalarize them.
3439 scalarizeInstruction(it);
3442 }// end of for_each instr.
3445 void InnerLoopVectorizer::updateAnalysis() {
3446 // Forget the original basic block.
3447 SE->forgetLoop(OrigLoop);
3449 // Update the dominator tree information.
3450 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3451 "Entry does not dominate exit.");
3453 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3454 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3455 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3457 // Due to if predication of stores we might create a sequence of "if(pred)
3458 // a[i] = ...; " blocks.
3459 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3461 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3462 else if (isPredicatedBlock(i)) {
3463 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3465 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3469 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3470 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3471 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3472 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3474 DEBUG(DT->verifyDomTree());
3477 /// \brief Check whether it is safe to if-convert this phi node.
3479 /// Phi nodes with constant expressions that can trap are not safe to if
3481 static bool canIfConvertPHINodes(BasicBlock *BB) {
3482 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3483 PHINode *Phi = dyn_cast<PHINode>(I);
3486 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3487 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3494 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3495 if (!EnableIfConversion) {
3496 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3500 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3502 // A list of pointers that we can safely read and write to.
3503 SmallPtrSet<Value *, 8> SafePointes;
3505 // Collect safe addresses.
3506 for (Loop::block_iterator BI = TheLoop->block_begin(),
3507 BE = TheLoop->block_end(); BI != BE; ++BI) {
3508 BasicBlock *BB = *BI;
3510 if (blockNeedsPredication(BB))
3513 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3514 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3515 SafePointes.insert(LI->getPointerOperand());
3516 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3517 SafePointes.insert(SI->getPointerOperand());
3521 // Collect the blocks that need predication.
3522 BasicBlock *Header = TheLoop->getHeader();
3523 for (Loop::block_iterator BI = TheLoop->block_begin(),
3524 BE = TheLoop->block_end(); BI != BE; ++BI) {
3525 BasicBlock *BB = *BI;
3527 // We don't support switch statements inside loops.
3528 if (!isa<BranchInst>(BB->getTerminator())) {
3529 emitAnalysis(VectorizationReport(BB->getTerminator())
3530 << "loop contains a switch statement");
3534 // We must be able to predicate all blocks that need to be predicated.
3535 if (blockNeedsPredication(BB)) {
3536 if (!blockCanBePredicated(BB, SafePointes)) {
3537 emitAnalysis(VectorizationReport(BB->getTerminator())
3538 << "control flow cannot be substituted for a select");
3541 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3542 emitAnalysis(VectorizationReport(BB->getTerminator())
3543 << "control flow cannot be substituted for a select");
3548 // We can if-convert this loop.
3552 bool LoopVectorizationLegality::canVectorize() {
3553 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3554 // be canonicalized.
3555 if (!TheLoop->getLoopPreheader()) {
3557 VectorizationReport() <<
3558 "loop control flow is not understood by vectorizer");
3562 // We can only vectorize innermost loops.
3563 if (!TheLoop->getSubLoopsVector().empty()) {
3564 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3568 // We must have a single backedge.
3569 if (TheLoop->getNumBackEdges() != 1) {
3571 VectorizationReport() <<
3572 "loop control flow is not understood by vectorizer");
3576 // We must have a single exiting block.
3577 if (!TheLoop->getExitingBlock()) {
3579 VectorizationReport() <<
3580 "loop control flow is not understood by vectorizer");
3584 // We only handle bottom-tested loops, i.e. loop in which the condition is
3585 // checked at the end of each iteration. With that we can assume that all
3586 // instructions in the loop are executed the same number of times.
3587 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3589 VectorizationReport() <<
3590 "loop control flow is not understood by vectorizer");
3594 // We need to have a loop header.
3595 DEBUG(dbgs() << "LV: Found a loop: " <<
3596 TheLoop->getHeader()->getName() << '\n');
3598 // Check if we can if-convert non-single-bb loops.
3599 unsigned NumBlocks = TheLoop->getNumBlocks();
3600 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3601 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3605 // ScalarEvolution needs to be able to find the exit count.
3606 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3607 if (ExitCount == SE->getCouldNotCompute()) {
3608 emitAnalysis(VectorizationReport() <<
3609 "could not determine number of loop iterations");
3610 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3614 // Check if we can vectorize the instructions and CFG in this loop.
3615 if (!canVectorizeInstrs()) {
3616 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3620 // Go over each instruction and look at memory deps.
3621 if (!canVectorizeMemory()) {
3622 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3626 // Collect all of the variables that remain uniform after vectorization.
3627 collectLoopUniforms();
3629 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3630 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3633 // Okay! We can vectorize. At this point we don't have any other mem analysis
3634 // which may limit our maximum vectorization factor, so just return true with
3639 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3640 if (Ty->isPointerTy())
3641 return DL.getIntPtrType(Ty);
3643 // It is possible that char's or short's overflow when we ask for the loop's
3644 // trip count, work around this by changing the type size.
3645 if (Ty->getScalarSizeInBits() < 32)
3646 return Type::getInt32Ty(Ty->getContext());
3651 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3652 Ty0 = convertPointerToIntegerType(DL, Ty0);
3653 Ty1 = convertPointerToIntegerType(DL, Ty1);
3654 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3659 /// \brief Check that the instruction has outside loop users and is not an
3660 /// identified reduction variable.
3661 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3662 SmallPtrSetImpl<Value *> &Reductions) {
3663 // Reduction instructions are allowed to have exit users. All other
3664 // instructions must not have external users.
3665 if (!Reductions.count(Inst))
3666 //Check that all of the users of the loop are inside the BB.
3667 for (User *U : Inst->users()) {
3668 Instruction *UI = cast<Instruction>(U);
3669 // This user may be a reduction exit value.
3670 if (!TheLoop->contains(UI)) {
3671 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3678 bool LoopVectorizationLegality::canVectorizeInstrs() {
3679 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3680 BasicBlock *Header = TheLoop->getHeader();
3682 // Look for the attribute signaling the absence of NaNs.
3683 Function &F = *Header->getParent();
3684 if (F.hasFnAttribute("no-nans-fp-math"))
3685 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3686 AttributeSet::FunctionIndex,
3687 "no-nans-fp-math").getValueAsString() == "true";
3689 // For each block in the loop.
3690 for (Loop::block_iterator bb = TheLoop->block_begin(),
3691 be = TheLoop->block_end(); bb != be; ++bb) {
3693 // Scan the instructions in the block and look for hazards.
3694 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3697 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3698 Type *PhiTy = Phi->getType();
3699 // Check that this PHI type is allowed.
3700 if (!PhiTy->isIntegerTy() &&
3701 !PhiTy->isFloatingPointTy() &&
3702 !PhiTy->isPointerTy()) {
3703 emitAnalysis(VectorizationReport(it)
3704 << "loop control flow is not understood by vectorizer");
3705 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3709 // If this PHINode is not in the header block, then we know that we
3710 // can convert it to select during if-conversion. No need to check if
3711 // the PHIs in this block are induction or reduction variables.
3712 if (*bb != Header) {
3713 // Check that this instruction has no outside users or is an
3714 // identified reduction value with an outside user.
3715 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3717 emitAnalysis(VectorizationReport(it) <<
3718 "value could not be identified as "
3719 "an induction or reduction variable");
3723 // We only allow if-converted PHIs with exactly two incoming values.
3724 if (Phi->getNumIncomingValues() != 2) {
3725 emitAnalysis(VectorizationReport(it)
3726 << "control flow not understood by vectorizer");
3727 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3731 // This is the value coming from the preheader.
3732 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3733 ConstantInt *StepValue = nullptr;
3734 // Check if this is an induction variable.
3735 InductionKind IK = isInductionVariable(Phi, StepValue);
3737 if (IK_NoInduction != IK) {
3738 // Get the widest type.
3740 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3742 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3744 // Int inductions are special because we only allow one IV.
3745 if (IK == IK_IntInduction && StepValue->isOne()) {
3746 // Use the phi node with the widest type as induction. Use the last
3747 // one if there are multiple (no good reason for doing this other
3748 // than it is expedient).
3749 if (!Induction || PhiTy == WidestIndTy)
3753 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3754 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3756 // Until we explicitly handle the case of an induction variable with
3757 // an outside loop user we have to give up vectorizing this loop.
3758 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3759 emitAnalysis(VectorizationReport(it) <<
3760 "use of induction value outside of the "
3761 "loop is not handled by vectorizer");
3768 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3769 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3772 if (AddReductionVar(Phi, RK_IntegerMult)) {
3773 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3776 if (AddReductionVar(Phi, RK_IntegerOr)) {
3777 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3780 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3781 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3784 if (AddReductionVar(Phi, RK_IntegerXor)) {
3785 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3788 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3789 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3792 if (AddReductionVar(Phi, RK_FloatMult)) {
3793 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3796 if (AddReductionVar(Phi, RK_FloatAdd)) {
3797 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3800 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3801 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3806 emitAnalysis(VectorizationReport(it) <<
3807 "value that could not be identified as "
3808 "reduction is used outside the loop");
3809 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3811 }// end of PHI handling
3813 // We still don't handle functions. However, we can ignore dbg intrinsic
3814 // calls and we do handle certain intrinsic and libm functions.
3815 CallInst *CI = dyn_cast<CallInst>(it);
3816 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3817 emitAnalysis(VectorizationReport(it) <<
3818 "call instruction cannot be vectorized");
3819 DEBUG(dbgs() << "LV: Found a call site.\n");
3823 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3824 // second argument is the same (i.e. loop invariant)
3826 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3827 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3828 emitAnalysis(VectorizationReport(it)
3829 << "intrinsic instruction cannot be vectorized");
3830 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3835 // Check that the instruction return type is vectorizable.
3836 // Also, we can't vectorize extractelement instructions.
3837 if ((!VectorType::isValidElementType(it->getType()) &&
3838 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3839 emitAnalysis(VectorizationReport(it)
3840 << "instruction return type cannot be vectorized");
3841 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3845 // Check that the stored type is vectorizable.
3846 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3847 Type *T = ST->getValueOperand()->getType();
3848 if (!VectorType::isValidElementType(T)) {
3849 emitAnalysis(VectorizationReport(ST) <<
3850 "store instruction cannot be vectorized");
3853 if (EnableMemAccessVersioning)
3854 collectStridedAccess(ST);
3857 if (EnableMemAccessVersioning)
3858 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3859 collectStridedAccess(LI);
3861 // Reduction instructions are allowed to have exit users.
3862 // All other instructions must not have external users.
3863 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3864 emitAnalysis(VectorizationReport(it) <<
3865 "value cannot be used outside the loop");
3874 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3875 if (Inductions.empty()) {
3876 emitAnalysis(VectorizationReport()
3877 << "loop induction variable could not be identified");
3885 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3886 /// return the induction operand of the gep pointer.
3887 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3888 const DataLayout *DL, Loop *Lp) {
3889 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3893 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3895 // Check that all of the gep indices are uniform except for our induction
3897 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3898 if (i != InductionOperand &&
3899 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3901 return GEP->getOperand(InductionOperand);
3904 ///\brief Look for a cast use of the passed value.
3905 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3906 Value *UniqueCast = nullptr;
3907 for (User *U : Ptr->users()) {
3908 CastInst *CI = dyn_cast<CastInst>(U);
3909 if (CI && CI->getType() == Ty) {
3919 ///\brief Get the stride of a pointer access in a loop.
3920 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3921 /// pointer to the Value, or null otherwise.
3922 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3923 const DataLayout *DL, Loop *Lp) {
3924 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3925 if (!PtrTy || PtrTy->isAggregateType())
3928 // Try to remove a gep instruction to make the pointer (actually index at this
3929 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3930 // pointer, otherwise, we are analyzing the index.
3931 Value *OrigPtr = Ptr;
3933 // The size of the pointer access.
3934 int64_t PtrAccessSize = 1;
3936 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3937 const SCEV *V = SE->getSCEV(Ptr);
3941 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3942 V = C->getOperand();
3944 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3948 V = S->getStepRecurrence(*SE);
3952 // Strip off the size of access multiplication if we are still analyzing the
3954 if (OrigPtr == Ptr) {
3955 DL->getTypeAllocSize(PtrTy->getElementType());
3956 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3957 if (M->getOperand(0)->getSCEVType() != scConstant)
3960 const APInt &APStepVal =
3961 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3963 // Huge step value - give up.
3964 if (APStepVal.getBitWidth() > 64)
3967 int64_t StepVal = APStepVal.getSExtValue();
3968 if (PtrAccessSize != StepVal)
3970 V = M->getOperand(1);
3975 Type *StripedOffRecurrenceCast = nullptr;
3976 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3977 StripedOffRecurrenceCast = C->getType();
3978 V = C->getOperand();
3981 // Look for the loop invariant symbolic value.
3982 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3986 Value *Stride = U->getValue();
3987 if (!Lp->isLoopInvariant(Stride))
3990 // If we have stripped off the recurrence cast we have to make sure that we
3991 // return the value that is used in this loop so that we can replace it later.
3992 if (StripedOffRecurrenceCast)
3993 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3998 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3999 Value *Ptr = nullptr;
4000 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4001 Ptr = LI->getPointerOperand();
4002 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4003 Ptr = SI->getPointerOperand();
4007 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
4011 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4012 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4013 Strides[Ptr] = Stride;
4014 StrideSet.insert(Stride);
4017 void LoopVectorizationLegality::collectLoopUniforms() {
4018 // We now know that the loop is vectorizable!
4019 // Collect variables that will remain uniform after vectorization.
4020 std::vector<Value*> Worklist;
4021 BasicBlock *Latch = TheLoop->getLoopLatch();
4023 // Start with the conditional branch and walk up the block.
4024 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4026 // Also add all consecutive pointer values; these values will be uniform
4027 // after vectorization (and subsequent cleanup) and, until revectorization is
4028 // supported, all dependencies must also be uniform.
4029 for (Loop::block_iterator B = TheLoop->block_begin(),
4030 BE = TheLoop->block_end(); B != BE; ++B)
4031 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4033 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4034 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4036 while (!Worklist.empty()) {
4037 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4038 Worklist.pop_back();
4040 // Look at instructions inside this loop.
4041 // Stop when reaching PHI nodes.
4042 // TODO: we need to follow values all over the loop, not only in this block.
4043 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4046 // This is a known uniform.
4049 // Insert all operands.
4050 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4055 /// \brief Analyses memory accesses in a loop.
4057 /// Checks whether run time pointer checks are needed and builds sets for data
4058 /// dependence checking.
4059 class AccessAnalysis {
4061 /// \brief Read or write access location.
4062 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4063 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4065 /// \brief Set of potential dependent memory accesses.
4066 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4068 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4069 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4071 /// \brief Register a load and whether it is only read from.
4072 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4073 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4074 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4075 Accesses.insert(MemAccessInfo(Ptr, false));
4077 ReadOnlyPtr.insert(Ptr);
4080 /// \brief Register a store.
4081 void addStore(AliasAnalysis::Location &Loc) {
4082 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4083 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4084 Accesses.insert(MemAccessInfo(Ptr, true));
4087 /// \brief Check whether we can check the pointers at runtime for
4088 /// non-intersection.
4089 bool canCheckPtrAtRT(RuntimePointerCheck &RtCheck, unsigned &NumComparisons,
4090 ScalarEvolution *SE, Loop *TheLoop,
4091 ValueToValueMap &Strides,
4092 bool ShouldCheckStride = false);
4094 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4095 /// and builds sets of dependent accesses.
4096 void buildDependenceSets() {
4097 processMemAccesses();
4100 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4102 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4103 void resetDepChecks() { CheckDeps.clear(); }
4105 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4108 typedef SetVector<MemAccessInfo> PtrAccessSet;
4110 /// \brief Go over all memory access and check whether runtime pointer checks
4111 /// are needed /// and build sets of dependency check candidates.
4112 void processMemAccesses();
4114 /// Set of all accesses.
4115 PtrAccessSet Accesses;
4117 /// Set of accesses that need a further dependence check.
4118 MemAccessInfoSet CheckDeps;
4120 /// Set of pointers that are read only.
4121 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4123 const DataLayout *DL;
4125 /// An alias set tracker to partition the access set by underlying object and
4126 //intrinsic property (such as TBAA metadata).
4127 AliasSetTracker AST;
4129 /// Sets of potentially dependent accesses - members of one set share an
4130 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4131 /// dependence check.
4132 DepCandidates &DepCands;
4134 bool IsRTCheckNeeded;
4137 } // end anonymous namespace
4139 /// \brief Check whether a pointer can participate in a runtime bounds check.
4140 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4142 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4143 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4147 return AR->isAffine();
4150 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4151 /// the address space.
4152 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4153 const Loop *Lp, ValueToValueMap &StridesMap);
4155 bool AccessAnalysis::canCheckPtrAtRT(
4156 RuntimePointerCheck &RtCheck,
4157 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4158 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4159 // Find pointers with computable bounds. We are going to use this information
4160 // to place a runtime bound check.
4161 bool CanDoRT = true;
4163 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4166 // We assign a consecutive id to access from different alias sets.
4167 // Accesses between different groups doesn't need to be checked.
4169 for (auto &AS : AST) {
4170 unsigned NumReadPtrChecks = 0;
4171 unsigned NumWritePtrChecks = 0;
4173 // We assign consecutive id to access from different dependence sets.
4174 // Accesses within the same set don't need a runtime check.
4175 unsigned RunningDepId = 1;
4176 DenseMap<Value *, unsigned> DepSetId;
4179 Value *Ptr = A.getValue();
4180 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4181 MemAccessInfo Access(Ptr, IsWrite);
4184 ++NumWritePtrChecks;
4188 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4189 // When we run after a failing dependency check we have to make sure we
4190 // don't have wrapping pointers.
4191 (!ShouldCheckStride ||
4192 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4193 // The id of the dependence set.
4196 if (IsDepCheckNeeded) {
4197 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4198 unsigned &LeaderId = DepSetId[Leader];
4200 LeaderId = RunningDepId++;
4203 // Each access has its own dependence set.
4204 DepId = RunningDepId++;
4206 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4208 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4214 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4215 NumComparisons += 0; // Only one dependence set.
4217 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4218 NumWritePtrChecks - 1));
4224 // If the pointers that we would use for the bounds comparison have different
4225 // address spaces, assume the values aren't directly comparable, so we can't
4226 // use them for the runtime check. We also have to assume they could
4227 // overlap. In the future there should be metadata for whether address spaces
4229 unsigned NumPointers = RtCheck.Pointers.size();
4230 for (unsigned i = 0; i < NumPointers; ++i) {
4231 for (unsigned j = i + 1; j < NumPointers; ++j) {
4232 // Only need to check pointers between two different dependency sets.
4233 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4235 // Only need to check pointers in the same alias set.
4236 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4239 Value *PtrI = RtCheck.Pointers[i];
4240 Value *PtrJ = RtCheck.Pointers[j];
4242 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4243 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4245 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4246 " different address spaces\n");
4255 void AccessAnalysis::processMemAccesses() {
4256 // We process the set twice: first we process read-write pointers, last we
4257 // process read-only pointers. This allows us to skip dependence tests for
4258 // read-only pointers.
4260 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4261 DEBUG(dbgs() << " AST: "; AST.dump());
4262 DEBUG(dbgs() << "LV: Accesses:\n");
4264 for (auto A : Accesses)
4265 dbgs() << "\t" << *A.getPointer() << " (" <<
4266 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4267 "read-only" : "read")) << ")\n";
4270 // The AliasSetTracker has nicely partitioned our pointers by metadata
4271 // compatibility and potential for underlying-object overlap. As a result, we
4272 // only need to check for potential pointer dependencies within each alias
4274 for (auto &AS : AST) {
4275 // Note that both the alias-set tracker and the alias sets themselves used
4276 // linked lists internally and so the iteration order here is deterministic
4277 // (matching the original instruction order within each set).
4279 bool SetHasWrite = false;
4281 // Map of pointers to last access encountered.
4282 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4283 UnderlyingObjToAccessMap ObjToLastAccess;
4285 // Set of access to check after all writes have been processed.
4286 PtrAccessSet DeferredAccesses;
4288 // Iterate over each alias set twice, once to process read/write pointers,
4289 // and then to process read-only pointers.
4290 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4291 bool UseDeferred = SetIteration > 0;
4292 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4294 for (auto AV : AS) {
4295 Value *Ptr = AV.getValue();
4297 // For a single memory access in AliasSetTracker, Accesses may contain
4298 // both read and write, and they both need to be handled for CheckDeps.
4300 if (AC.getPointer() != Ptr)
4303 bool IsWrite = AC.getInt();
4305 // If we're using the deferred access set, then it contains only
4307 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4308 if (UseDeferred && !IsReadOnlyPtr)
4310 // Otherwise, the pointer must be in the PtrAccessSet, either as a
4312 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4313 S.count(MemAccessInfo(Ptr, false))) &&
4314 "Alias-set pointer not in the access set?");
4316 MemAccessInfo Access(Ptr, IsWrite);
4317 DepCands.insert(Access);
4319 // Memorize read-only pointers for later processing and skip them in
4320 // the first round (they need to be checked after we have seen all
4321 // write pointers). Note: we also mark pointer that are not
4322 // consecutive as "read-only" pointers (so that we check
4323 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
4324 if (!UseDeferred && IsReadOnlyPtr) {
4325 DeferredAccesses.insert(Access);
4329 // If this is a write - check other reads and writes for conflicts. If
4330 // this is a read only check other writes for conflicts (but only if
4331 // there is no other write to the ptr - this is an optimization to
4332 // catch "a[i] = a[i] + " without having to do a dependence check).
4333 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4334 CheckDeps.insert(Access);
4335 IsRTCheckNeeded = true;
4341 // Create sets of pointers connected by a shared alias set and
4342 // underlying object.
4343 typedef SmallVector<Value *, 16> ValueVector;
4344 ValueVector TempObjects;
4345 GetUnderlyingObjects(Ptr, TempObjects, DL);
4346 for (Value *UnderlyingObj : TempObjects) {
4347 UnderlyingObjToAccessMap::iterator Prev =
4348 ObjToLastAccess.find(UnderlyingObj);
4349 if (Prev != ObjToLastAccess.end())
4350 DepCands.unionSets(Access, Prev->second);
4352 ObjToLastAccess[UnderlyingObj] = Access;
4361 /// \brief Checks memory dependences among accesses to the same underlying
4362 /// object to determine whether there vectorization is legal or not (and at
4363 /// which vectorization factor).
4365 /// This class works under the assumption that we already checked that memory
4366 /// locations with different underlying pointers are "must-not alias".
4367 /// We use the ScalarEvolution framework to symbolically evalutate access
4368 /// functions pairs. Since we currently don't restructure the loop we can rely
4369 /// on the program order of memory accesses to determine their safety.
4370 /// At the moment we will only deem accesses as safe for:
4371 /// * A negative constant distance assuming program order.
4373 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4374 /// a[i] = tmp; y = a[i];
4376 /// The latter case is safe because later checks guarantuee that there can't
4377 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4378 /// the same variable: a header phi can only be an induction or a reduction, a
4379 /// reduction can't have a memory sink, an induction can't have a memory
4380 /// source). This is important and must not be violated (or we have to
4381 /// resort to checking for cycles through memory).
4383 /// * A positive constant distance assuming program order that is bigger
4384 /// than the biggest memory access.
4386 /// tmp = a[i] OR b[i] = x
4387 /// a[i+2] = tmp y = b[i+2];
4389 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4391 /// * Zero distances and all accesses have the same size.
4393 class MemoryDepChecker {
4395 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4396 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4398 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4399 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4400 ShouldRetryWithRuntimeCheck(false) {}
4402 /// \brief Register the location (instructions are given increasing numbers)
4403 /// of a write access.
4404 void addAccess(StoreInst *SI) {
4405 Value *Ptr = SI->getPointerOperand();
4406 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4407 InstMap.push_back(SI);
4411 /// \brief Register the location (instructions are given increasing numbers)
4412 /// of a write access.
4413 void addAccess(LoadInst *LI) {
4414 Value *Ptr = LI->getPointerOperand();
4415 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4416 InstMap.push_back(LI);
4420 /// \brief Check whether the dependencies between the accesses are safe.
4422 /// Only checks sets with elements in \p CheckDeps.
4423 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4424 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4426 /// \brief The maximum number of bytes of a vector register we can vectorize
4427 /// the accesses safely with.
4428 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4430 /// \brief In same cases when the dependency check fails we can still
4431 /// vectorize the loop with a dynamic array access check.
4432 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4435 ScalarEvolution *SE;
4436 const DataLayout *DL;
4437 const Loop *InnermostLoop;
4439 /// \brief Maps access locations (ptr, read/write) to program order.
4440 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4442 /// \brief Memory access instructions in program order.
4443 SmallVector<Instruction *, 16> InstMap;
4445 /// \brief The program order index to be used for the next instruction.
4448 // We can access this many bytes in parallel safely.
4449 unsigned MaxSafeDepDistBytes;
4451 /// \brief If we see a non-constant dependence distance we can still try to
4452 /// vectorize this loop with runtime checks.
4453 bool ShouldRetryWithRuntimeCheck;
4455 /// \brief Check whether there is a plausible dependence between the two
4458 /// Access \p A must happen before \p B in program order. The two indices
4459 /// identify the index into the program order map.
4461 /// This function checks whether there is a plausible dependence (or the
4462 /// absence of such can't be proved) between the two accesses. If there is a
4463 /// plausible dependence but the dependence distance is bigger than one
4464 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4465 /// distance is smaller than any other distance encountered so far).
4466 /// Otherwise, this function returns true signaling a possible dependence.
4467 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4468 const MemAccessInfo &B, unsigned BIdx,
4469 ValueToValueMap &Strides);
4471 /// \brief Check whether the data dependence could prevent store-load
4473 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4476 } // end anonymous namespace
4478 static bool isInBoundsGep(Value *Ptr) {
4479 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4480 return GEP->isInBounds();
4484 /// \brief Check whether the access through \p Ptr has a constant stride.
4485 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4486 const Loop *Lp, ValueToValueMap &StridesMap) {
4487 const Type *Ty = Ptr->getType();
4488 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4490 // Make sure that the pointer does not point to aggregate types.
4491 const PointerType *PtrTy = cast<PointerType>(Ty);
4492 if (PtrTy->getElementType()->isAggregateType()) {
4493 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4498 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4500 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4502 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4503 << *Ptr << " SCEV: " << *PtrScev << "\n");
4507 // The accesss function must stride over the innermost loop.
4508 if (Lp != AR->getLoop()) {
4509 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4510 *Ptr << " SCEV: " << *PtrScev << "\n");
4513 // The address calculation must not wrap. Otherwise, a dependence could be
4515 // An inbounds getelementptr that is a AddRec with a unit stride
4516 // cannot wrap per definition. The unit stride requirement is checked later.
4517 // An getelementptr without an inbounds attribute and unit stride would have
4518 // to access the pointer value "0" which is undefined behavior in address
4519 // space 0, therefore we can also vectorize this case.
4520 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4521 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4522 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4523 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4524 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4525 << *Ptr << " SCEV: " << *PtrScev << "\n");
4529 // Check the step is constant.
4530 const SCEV *Step = AR->getStepRecurrence(*SE);
4532 // Calculate the pointer stride and check if it is consecutive.
4533 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4535 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4536 " SCEV: " << *PtrScev << "\n");
4540 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4541 const APInt &APStepVal = C->getValue()->getValue();
4543 // Huge step value - give up.
4544 if (APStepVal.getBitWidth() > 64)
4547 int64_t StepVal = APStepVal.getSExtValue();
4550 int64_t Stride = StepVal / Size;
4551 int64_t Rem = StepVal % Size;
4555 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4556 // know we can't "wrap around the address space". In case of address space
4557 // zero we know that this won't happen without triggering undefined behavior.
4558 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4559 Stride != 1 && Stride != -1)
4565 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4566 unsigned TypeByteSize) {
4567 // If loads occur at a distance that is not a multiple of a feasible vector
4568 // factor store-load forwarding does not take place.
4569 // Positive dependences might cause troubles because vectorizing them might
4570 // prevent store-load forwarding making vectorized code run a lot slower.
4571 // a[i] = a[i-3] ^ a[i-8];
4572 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4573 // hence on your typical architecture store-load forwarding does not take
4574 // place. Vectorizing in such cases does not make sense.
4575 // Store-load forwarding distance.
4576 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4577 // Maximum vector factor.
4578 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4579 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4580 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4582 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4584 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4585 MaxVFWithoutSLForwardIssues = (vf >>=1);
4590 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4591 DEBUG(dbgs() << "LV: Distance " << Distance <<
4592 " that could cause a store-load forwarding conflict\n");
4596 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4597 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4598 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4602 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4603 const MemAccessInfo &B, unsigned BIdx,
4604 ValueToValueMap &Strides) {
4605 assert (AIdx < BIdx && "Must pass arguments in program order");
4607 Value *APtr = A.getPointer();
4608 Value *BPtr = B.getPointer();
4609 bool AIsWrite = A.getInt();
4610 bool BIsWrite = B.getInt();
4612 // Two reads are independent.
4613 if (!AIsWrite && !BIsWrite)
4616 // We cannot check pointers in different address spaces.
4617 if (APtr->getType()->getPointerAddressSpace() !=
4618 BPtr->getType()->getPointerAddressSpace())
4621 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4622 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4624 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4625 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4627 const SCEV *Src = AScev;
4628 const SCEV *Sink = BScev;
4630 // If the induction step is negative we have to invert source and sink of the
4632 if (StrideAPtr < 0) {
4635 std::swap(APtr, BPtr);
4636 std::swap(Src, Sink);
4637 std::swap(AIsWrite, BIsWrite);
4638 std::swap(AIdx, BIdx);
4639 std::swap(StrideAPtr, StrideBPtr);
4642 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4644 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4645 << "(Induction step: " << StrideAPtr << ")\n");
4646 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4647 << *InstMap[BIdx] << ": " << *Dist << "\n");
4649 // Need consecutive accesses. We don't want to vectorize
4650 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4651 // the address space.
4652 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4653 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4657 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4659 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4660 ShouldRetryWithRuntimeCheck = true;
4664 Type *ATy = APtr->getType()->getPointerElementType();
4665 Type *BTy = BPtr->getType()->getPointerElementType();
4666 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4668 // Negative distances are not plausible dependencies.
4669 const APInt &Val = C->getValue()->getValue();
4670 if (Val.isNegative()) {
4671 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4672 if (IsTrueDataDependence &&
4673 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4677 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4681 // Write to the same location with the same size.
4682 // Could be improved to assert type sizes are the same (i32 == float, etc).
4686 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4690 assert(Val.isStrictlyPositive() && "Expect a positive value");
4692 // Positive distance bigger than max vectorization factor.
4695 "LV: ReadWrite-Write positive dependency with different types\n");
4699 unsigned Distance = (unsigned) Val.getZExtValue();
4701 // Bail out early if passed-in parameters make vectorization not feasible.
4702 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4703 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4705 // The distance must be bigger than the size needed for a vectorized version
4706 // of the operation and the size of the vectorized operation must not be
4707 // bigger than the currrent maximum size.
4708 if (Distance < 2*TypeByteSize ||
4709 2*TypeByteSize > MaxSafeDepDistBytes ||
4710 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4711 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4712 << Val.getSExtValue() << '\n');
4716 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4717 Distance : MaxSafeDepDistBytes;
4719 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4720 if (IsTrueDataDependence &&
4721 couldPreventStoreLoadForward(Distance, TypeByteSize))
4724 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4725 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4730 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4731 MemAccessInfoSet &CheckDeps,
4732 ValueToValueMap &Strides) {
4734 MaxSafeDepDistBytes = -1U;
4735 while (!CheckDeps.empty()) {
4736 MemAccessInfo CurAccess = *CheckDeps.begin();
4738 // Get the relevant memory access set.
4739 EquivalenceClasses<MemAccessInfo>::iterator I =
4740 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4742 // Check accesses within this set.
4743 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4744 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4746 // Check every access pair.
4748 CheckDeps.erase(*AI);
4749 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4751 // Check every accessing instruction pair in program order.
4752 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4753 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4754 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4755 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4756 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4758 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4769 bool LoopVectorizationLegality::canVectorizeMemory() {
4771 typedef SmallVector<Value*, 16> ValueVector;
4772 typedef SmallPtrSet<Value*, 16> ValueSet;
4774 // Holds the Load and Store *instructions*.
4778 // Holds all the different accesses in the loop.
4779 unsigned NumReads = 0;
4780 unsigned NumReadWrites = 0;
4782 PtrRtCheck.Pointers.clear();
4783 PtrRtCheck.Need = false;
4785 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4786 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4789 for (Loop::block_iterator bb = TheLoop->block_begin(),
4790 be = TheLoop->block_end(); bb != be; ++bb) {
4792 // Scan the BB and collect legal loads and stores.
4793 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4796 // If this is a load, save it. If this instruction can read from memory
4797 // but is not a load, then we quit. Notice that we don't handle function
4798 // calls that read or write.
4799 if (it->mayReadFromMemory()) {
4800 // Many math library functions read the rounding mode. We will only
4801 // vectorize a loop if it contains known function calls that don't set
4802 // the flag. Therefore, it is safe to ignore this read from memory.
4803 CallInst *Call = dyn_cast<CallInst>(it);
4804 if (Call && getIntrinsicIDForCall(Call, TLI))
4807 LoadInst *Ld = dyn_cast<LoadInst>(it);
4808 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4809 emitAnalysis(VectorizationReport(Ld)
4810 << "read with atomic ordering or volatile read");
4811 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4815 Loads.push_back(Ld);
4816 DepChecker.addAccess(Ld);
4820 // Save 'store' instructions. Abort if other instructions write to memory.
4821 if (it->mayWriteToMemory()) {
4822 StoreInst *St = dyn_cast<StoreInst>(it);
4824 emitAnalysis(VectorizationReport(it) <<
4825 "instruction cannot be vectorized");
4828 if (!St->isSimple() && !IsAnnotatedParallel) {
4829 emitAnalysis(VectorizationReport(St)
4830 << "write with atomic ordering or volatile write");
4831 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4835 Stores.push_back(St);
4836 DepChecker.addAccess(St);
4841 // Now we have two lists that hold the loads and the stores.
4842 // Next, we find the pointers that they use.
4844 // Check if we see any stores. If there are no stores, then we don't
4845 // care if the pointers are *restrict*.
4846 if (!Stores.size()) {
4847 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4851 AccessAnalysis::DepCandidates DependentAccesses;
4852 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4854 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4855 // multiple times on the same object. If the ptr is accessed twice, once
4856 // for read and once for write, it will only appear once (on the write
4857 // list). This is okay, since we are going to check for conflicts between
4858 // writes and between reads and writes, but not between reads and reads.
4861 ValueVector::iterator I, IE;
4862 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4863 StoreInst *ST = cast<StoreInst>(*I);
4864 Value* Ptr = ST->getPointerOperand();
4866 if (isUniform(Ptr)) {
4868 VectorizationReport(ST)
4869 << "write to a loop invariant address could not be vectorized");
4870 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4874 // If we did *not* see this pointer before, insert it to the read-write
4875 // list. At this phase it is only a 'write' list.
4876 if (Seen.insert(Ptr).second) {
4879 AliasAnalysis::Location Loc = AA->getLocation(ST);
4880 // The TBAA metadata could have a control dependency on the predication
4881 // condition, so we cannot rely on it when determining whether or not we
4882 // need runtime pointer checks.
4883 if (blockNeedsPredication(ST->getParent()))
4884 Loc.AATags.TBAA = nullptr;
4886 Accesses.addStore(Loc);
4890 if (IsAnnotatedParallel) {
4892 << "LV: A loop annotated parallel, ignore memory dependency "
4897 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4898 LoadInst *LD = cast<LoadInst>(*I);
4899 Value* Ptr = LD->getPointerOperand();
4900 // If we did *not* see this pointer before, insert it to the
4901 // read list. If we *did* see it before, then it is already in
4902 // the read-write list. This allows us to vectorize expressions
4903 // such as A[i] += x; Because the address of A[i] is a read-write
4904 // pointer. This only works if the index of A[i] is consecutive.
4905 // If the address of i is unknown (for example A[B[i]]) then we may
4906 // read a few words, modify, and write a few words, and some of the
4907 // words may be written to the same address.
4908 bool IsReadOnlyPtr = false;
4909 if (Seen.insert(Ptr).second ||
4910 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4912 IsReadOnlyPtr = true;
4915 AliasAnalysis::Location Loc = AA->getLocation(LD);
4916 // The TBAA metadata could have a control dependency on the predication
4917 // condition, so we cannot rely on it when determining whether or not we
4918 // need runtime pointer checks.
4919 if (blockNeedsPredication(LD->getParent()))
4920 Loc.AATags.TBAA = nullptr;
4922 Accesses.addLoad(Loc, IsReadOnlyPtr);
4925 // If we write (or read-write) to a single destination and there are no
4926 // other reads in this loop then is it safe to vectorize.
4927 if (NumReadWrites == 1 && NumReads == 0) {
4928 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4932 // Build dependence sets and check whether we need a runtime pointer bounds
4934 Accesses.buildDependenceSets();
4935 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4937 // Find pointers with computable bounds. We are going to use this information
4938 // to place a runtime bound check.
4939 unsigned NumComparisons = 0;
4940 bool CanDoRT = false;
4942 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4945 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4946 " pointer comparisons.\n");
4948 // If we only have one set of dependences to check pointers among we don't
4949 // need a runtime check.
4950 if (NumComparisons == 0 && NeedRTCheck)
4951 NeedRTCheck = false;
4953 // Check that we did not collect too many pointers or found an unsizeable
4955 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4961 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4964 if (NeedRTCheck && !CanDoRT) {
4965 emitAnalysis(VectorizationReport() << "cannot identify array bounds");
4966 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4967 "the array bounds.\n");
4972 PtrRtCheck.Need = NeedRTCheck;
4974 bool CanVecMem = true;
4975 if (Accesses.isDependencyCheckNeeded()) {
4976 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4977 CanVecMem = DepChecker.areDepsSafe(
4978 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4979 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4981 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4982 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4985 // Clear the dependency checks. We assume they are not needed.
4986 Accesses.resetDepChecks();
4989 PtrRtCheck.Need = true;
4991 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4992 TheLoop, Strides, true);
4993 // Check that we did not collect too many pointers or found an unsizeable
4995 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4996 if (!CanDoRT && NumComparisons > 0)
4997 emitAnalysis(VectorizationReport()
4998 << "cannot check memory dependencies at runtime");
5000 emitAnalysis(VectorizationReport()
5001 << NumComparisons << " exceeds limit of "
5002 << RuntimeMemoryCheckThreshold
5003 << " dependent memory operations checked at runtime");
5004 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
5014 emitAnalysis(VectorizationReport() <<
5015 "unsafe dependent memory operations in loop");
5017 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
5018 " need a runtime memory check.\n");
5023 static bool hasMultipleUsesOf(Instruction *I,
5024 SmallPtrSetImpl<Instruction *> &Insts) {
5025 unsigned NumUses = 0;
5026 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
5027 if (Insts.count(dyn_cast<Instruction>(*Use)))
5036 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
5037 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
5038 if (!Set.count(dyn_cast<Instruction>(*Use)))
5043 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
5044 ReductionKind Kind) {
5045 if (Phi->getNumIncomingValues() != 2)
5048 // Reduction variables are only found in the loop header block.
5049 if (Phi->getParent() != TheLoop->getHeader())
5052 // Obtain the reduction start value from the value that comes from the loop
5054 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
5056 // ExitInstruction is the single value which is used outside the loop.
5057 // We only allow for a single reduction value to be used outside the loop.
5058 // This includes users of the reduction, variables (which form a cycle
5059 // which ends in the phi node).
5060 Instruction *ExitInstruction = nullptr;
5061 // Indicates that we found a reduction operation in our scan.
5062 bool FoundReduxOp = false;
5064 // We start with the PHI node and scan for all of the users of this
5065 // instruction. All users must be instructions that can be used as reduction
5066 // variables (such as ADD). We must have a single out-of-block user. The cycle
5067 // must include the original PHI.
5068 bool FoundStartPHI = false;
5070 // To recognize min/max patterns formed by a icmp select sequence, we store
5071 // the number of instruction we saw from the recognized min/max pattern,
5072 // to make sure we only see exactly the two instructions.
5073 unsigned NumCmpSelectPatternInst = 0;
5074 ReductionInstDesc ReduxDesc(false, nullptr);
5076 SmallPtrSet<Instruction *, 8> VisitedInsts;
5077 SmallVector<Instruction *, 8> Worklist;
5078 Worklist.push_back(Phi);
5079 VisitedInsts.insert(Phi);
5081 // A value in the reduction can be used:
5082 // - By the reduction:
5083 // - Reduction operation:
5084 // - One use of reduction value (safe).
5085 // - Multiple use of reduction value (not safe).
5087 // - All uses of the PHI must be the reduction (safe).
5088 // - Otherwise, not safe.
5089 // - By one instruction outside of the loop (safe).
5090 // - By further instructions outside of the loop (not safe).
5091 // - By an instruction that is not part of the reduction (not safe).
5093 // * An instruction type other than PHI or the reduction operation.
5094 // * A PHI in the header other than the initial PHI.
5095 while (!Worklist.empty()) {
5096 Instruction *Cur = Worklist.back();
5097 Worklist.pop_back();
5100 // If the instruction has no users then this is a broken chain and can't be
5101 // a reduction variable.
5102 if (Cur->use_empty())
5105 bool IsAPhi = isa<PHINode>(Cur);
5107 // A header PHI use other than the original PHI.
5108 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5111 // Reductions of instructions such as Div, and Sub is only possible if the
5112 // LHS is the reduction variable.
5113 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5114 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5115 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5118 // Any reduction instruction must be of one of the allowed kinds.
5119 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5120 if (!ReduxDesc.IsReduction)
5123 // A reduction operation must only have one use of the reduction value.
5124 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5125 hasMultipleUsesOf(Cur, VisitedInsts))
5128 // All inputs to a PHI node must be a reduction value.
5129 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5132 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5133 isa<SelectInst>(Cur)))
5134 ++NumCmpSelectPatternInst;
5135 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5136 isa<SelectInst>(Cur)))
5137 ++NumCmpSelectPatternInst;
5139 // Check whether we found a reduction operator.
5140 FoundReduxOp |= !IsAPhi;
5142 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5143 // onto the stack. This way we are going to have seen all inputs to PHI
5144 // nodes once we get to them.
5145 SmallVector<Instruction *, 8> NonPHIs;
5146 SmallVector<Instruction *, 8> PHIs;
5147 for (User *U : Cur->users()) {
5148 Instruction *UI = cast<Instruction>(U);
5150 // Check if we found the exit user.
5151 BasicBlock *Parent = UI->getParent();
5152 if (!TheLoop->contains(Parent)) {
5153 // Exit if you find multiple outside users or if the header phi node is
5154 // being used. In this case the user uses the value of the previous
5155 // iteration, in which case we would loose "VF-1" iterations of the
5156 // reduction operation if we vectorize.
5157 if (ExitInstruction != nullptr || Cur == Phi)
5160 // The instruction used by an outside user must be the last instruction
5161 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5162 // operations on the value.
5163 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5166 ExitInstruction = Cur;
5170 // Process instructions only once (termination). Each reduction cycle
5171 // value must only be used once, except by phi nodes and min/max
5172 // reductions which are represented as a cmp followed by a select.
5173 ReductionInstDesc IgnoredVal(false, nullptr);
5174 if (VisitedInsts.insert(UI).second) {
5175 if (isa<PHINode>(UI))
5178 NonPHIs.push_back(UI);
5179 } else if (!isa<PHINode>(UI) &&
5180 ((!isa<FCmpInst>(UI) &&
5181 !isa<ICmpInst>(UI) &&
5182 !isa<SelectInst>(UI)) ||
5183 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5186 // Remember that we completed the cycle.
5188 FoundStartPHI = true;
5190 Worklist.append(PHIs.begin(), PHIs.end());
5191 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5194 // This means we have seen one but not the other instruction of the
5195 // pattern or more than just a select and cmp.
5196 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5197 NumCmpSelectPatternInst != 2)
5200 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5203 // We found a reduction var if we have reached the original phi node and we
5204 // only have a single instruction with out-of-loop users.
5206 // This instruction is allowed to have out-of-loop users.
5207 AllowedExit.insert(ExitInstruction);
5209 // Save the description of this reduction variable.
5210 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5211 ReduxDesc.MinMaxKind);
5212 Reductions[Phi] = RD;
5213 // We've ended the cycle. This is a reduction variable if we have an
5214 // outside user and it has a binary op.
5219 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5220 /// pattern corresponding to a min(X, Y) or max(X, Y).
5221 LoopVectorizationLegality::ReductionInstDesc
5222 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5223 ReductionInstDesc &Prev) {
5225 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5226 "Expect a select instruction");
5227 Instruction *Cmp = nullptr;
5228 SelectInst *Select = nullptr;
5230 // We must handle the select(cmp()) as a single instruction. Advance to the
5232 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5233 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5234 return ReductionInstDesc(false, I);
5235 return ReductionInstDesc(Select, Prev.MinMaxKind);
5238 // Only handle single use cases for now.
5239 if (!(Select = dyn_cast<SelectInst>(I)))
5240 return ReductionInstDesc(false, I);
5241 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5242 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5243 return ReductionInstDesc(false, I);
5244 if (!Cmp->hasOneUse())
5245 return ReductionInstDesc(false, I);
5250 // Look for a min/max pattern.
5251 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5252 return ReductionInstDesc(Select, MRK_UIntMin);
5253 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5254 return ReductionInstDesc(Select, MRK_UIntMax);
5255 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5256 return ReductionInstDesc(Select, MRK_SIntMax);
5257 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5258 return ReductionInstDesc(Select, MRK_SIntMin);
5259 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5260 return ReductionInstDesc(Select, MRK_FloatMin);
5261 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5262 return ReductionInstDesc(Select, MRK_FloatMax);
5263 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5264 return ReductionInstDesc(Select, MRK_FloatMin);
5265 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5266 return ReductionInstDesc(Select, MRK_FloatMax);
5268 return ReductionInstDesc(false, I);
5271 LoopVectorizationLegality::ReductionInstDesc
5272 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5274 ReductionInstDesc &Prev) {
5275 bool FP = I->getType()->isFloatingPointTy();
5276 bool FastMath = FP && I->hasUnsafeAlgebra();
5277 switch (I->getOpcode()) {
5279 return ReductionInstDesc(false, I);
5280 case Instruction::PHI:
5281 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5282 Kind != RK_FloatMinMax))
5283 return ReductionInstDesc(false, I);
5284 return ReductionInstDesc(I, Prev.MinMaxKind);
5285 case Instruction::Sub:
5286 case Instruction::Add:
5287 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5288 case Instruction::Mul:
5289 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5290 case Instruction::And:
5291 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5292 case Instruction::Or:
5293 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5294 case Instruction::Xor:
5295 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5296 case Instruction::FMul:
5297 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5298 case Instruction::FSub:
5299 case Instruction::FAdd:
5300 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5301 case Instruction::FCmp:
5302 case Instruction::ICmp:
5303 case Instruction::Select:
5304 if (Kind != RK_IntegerMinMax &&
5305 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5306 return ReductionInstDesc(false, I);
5307 return isMinMaxSelectCmpPattern(I, Prev);
5311 LoopVectorizationLegality::InductionKind
5312 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
5313 ConstantInt *&StepValue) {
5314 Type *PhiTy = Phi->getType();
5315 // We only handle integer and pointer inductions variables.
5316 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5317 return IK_NoInduction;
5319 // Check that the PHI is consecutive.
5320 const SCEV *PhiScev = SE->getSCEV(Phi);
5321 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5323 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5324 return IK_NoInduction;
5327 const SCEV *Step = AR->getStepRecurrence(*SE);
5328 // Calculate the pointer stride and check if it is consecutive.
5329 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5331 return IK_NoInduction;
5333 ConstantInt *CV = C->getValue();
5334 if (PhiTy->isIntegerTy()) {
5336 return IK_IntInduction;
5339 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5340 Type *PointerElementType = PhiTy->getPointerElementType();
5341 // The pointer stride cannot be determined if the pointer element type is not
5343 if (!PointerElementType->isSized())
5344 return IK_NoInduction;
5346 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
5347 int64_t CVSize = CV->getSExtValue();
5349 return IK_NoInduction;
5350 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
5351 return IK_PtrInduction;
5354 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5355 Value *In0 = const_cast<Value*>(V);
5356 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5360 return Inductions.count(PN);
5363 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5364 assert(TheLoop->contains(BB) && "Unknown block used");
5366 // Blocks that do not dominate the latch need predication.
5367 BasicBlock* Latch = TheLoop->getLoopLatch();
5368 return !DT->dominates(BB, Latch);
5371 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5372 SmallPtrSetImpl<Value *> &SafePtrs) {
5374 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5375 // Check that we don't have a constant expression that can trap as operand.
5376 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5378 if (Constant *C = dyn_cast<Constant>(*OI))
5382 // We might be able to hoist the load.
5383 if (it->mayReadFromMemory()) {
5384 LoadInst *LI = dyn_cast<LoadInst>(it);
5387 if (!SafePtrs.count(LI->getPointerOperand())) {
5388 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
5389 MaskedOp.insert(LI);
5396 // We don't predicate stores at the moment.
5397 if (it->mayWriteToMemory()) {
5398 StoreInst *SI = dyn_cast<StoreInst>(it);
5399 // We only support predication of stores in basic blocks with one
5404 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5405 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5407 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5408 !isSinglePredecessor) {
5409 // Build a masked store if it is legal for the target, otherwise scalarize
5411 bool isLegalMaskedOp =
5412 isLegalMaskedStore(SI->getValueOperand()->getType(),
5413 SI->getPointerOperand());
5414 if (isLegalMaskedOp) {
5416 MaskedOp.insert(SI);
5425 // The instructions below can trap.
5426 switch (it->getOpcode()) {
5428 case Instruction::UDiv:
5429 case Instruction::SDiv:
5430 case Instruction::URem:
5431 case Instruction::SRem:
5439 LoopVectorizationCostModel::VectorizationFactor
5440 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5441 // Width 1 means no vectorize
5442 VectorizationFactor Factor = { 1U, 0U };
5443 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5444 emitAnalysis(VectorizationReport() <<
5445 "runtime pointer checks needed. Enable vectorization of this "
5446 "loop with '#pragma clang loop vectorize(enable)' when "
5447 "compiling with -Os");
5448 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5452 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5453 emitAnalysis(VectorizationReport() <<
5454 "store that is conditionally executed prevents vectorization");
5455 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5459 // Find the trip count.
5460 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5461 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5463 unsigned WidestType = getWidestType();
5464 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5465 unsigned MaxSafeDepDist = -1U;
5466 if (Legal->getMaxSafeDepDistBytes() != -1U)
5467 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5468 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5469 WidestRegister : MaxSafeDepDist);
5470 unsigned MaxVectorSize = WidestRegister / WidestType;
5471 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5472 DEBUG(dbgs() << "LV: The Widest register is: "
5473 << WidestRegister << " bits.\n");
5475 if (MaxVectorSize == 0) {
5476 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5480 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5481 " into one vector!");
5483 unsigned VF = MaxVectorSize;
5485 // If we optimize the program for size, avoid creating the tail loop.
5487 // If we are unable to calculate the trip count then don't try to vectorize.
5490 (VectorizationReport() <<
5491 "unable to calculate the loop count due to complex control flow");
5492 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5496 // Find the maximum SIMD width that can fit within the trip count.
5497 VF = TC % MaxVectorSize;
5502 // If the trip count that we found modulo the vectorization factor is not
5503 // zero then we require a tail.
5505 emitAnalysis(VectorizationReport() <<
5506 "cannot optimize for size and vectorize at the "
5507 "same time. Enable vectorization of this loop "
5508 "with '#pragma clang loop vectorize(enable)' "
5509 "when compiling with -Os");
5510 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5515 int UserVF = Hints->getWidth();
5517 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5518 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5520 Factor.Width = UserVF;
5524 float Cost = expectedCost(1);
5526 const float ScalarCost = Cost;
5529 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5531 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5532 // Ignore scalar width, because the user explicitly wants vectorization.
5533 if (ForceVectorization && VF > 1) {
5535 Cost = expectedCost(Width) / (float)Width;
5538 for (unsigned i=2; i <= VF; i*=2) {
5539 // Notice that the vector loop needs to be executed less times, so
5540 // we need to divide the cost of the vector loops by the width of
5541 // the vector elements.
5542 float VectorCost = expectedCost(i) / (float)i;
5543 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5544 (int)VectorCost << ".\n");
5545 if (VectorCost < Cost) {
5551 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5552 << "LV: Vectorization seems to be not beneficial, "
5553 << "but was forced by a user.\n");
5554 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5555 Factor.Width = Width;
5556 Factor.Cost = Width * Cost;
5560 unsigned LoopVectorizationCostModel::getWidestType() {
5561 unsigned MaxWidth = 8;
5564 for (Loop::block_iterator bb = TheLoop->block_begin(),
5565 be = TheLoop->block_end(); bb != be; ++bb) {
5566 BasicBlock *BB = *bb;
5568 // For each instruction in the loop.
5569 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5570 Type *T = it->getType();
5572 // Ignore ephemeral values.
5573 if (EphValues.count(it))
5576 // Only examine Loads, Stores and PHINodes.
5577 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5580 // Examine PHI nodes that are reduction variables.
5581 if (PHINode *PN = dyn_cast<PHINode>(it))
5582 if (!Legal->getReductionVars()->count(PN))
5585 // Examine the stored values.
5586 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5587 T = ST->getValueOperand()->getType();
5589 // Ignore loaded pointer types and stored pointer types that are not
5590 // consecutive. However, we do want to take consecutive stores/loads of
5591 // pointer vectors into account.
5592 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5595 MaxWidth = std::max(MaxWidth,
5596 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5604 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5606 unsigned LoopCost) {
5608 // -- The unroll heuristics --
5609 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5610 // There are many micro-architectural considerations that we can't predict
5611 // at this level. For example, frontend pressure (on decode or fetch) due to
5612 // code size, or the number and capabilities of the execution ports.
5614 // We use the following heuristics to select the unroll factor:
5615 // 1. If the code has reductions, then we unroll in order to break the cross
5616 // iteration dependency.
5617 // 2. If the loop is really small, then we unroll in order to reduce the loop
5619 // 3. We don't unroll if we think that we will spill registers to memory due
5620 // to the increased register pressure.
5622 // Use the user preference, unless 'auto' is selected.
5623 int UserUF = Hints->getInterleave();
5627 // When we optimize for size, we don't unroll.
5631 // We used the distance for the unroll factor.
5632 if (Legal->getMaxSafeDepDistBytes() != -1U)
5635 // Do not unroll loops with a relatively small trip count.
5636 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5637 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5640 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5641 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5645 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5646 TargetNumRegisters = ForceTargetNumScalarRegs;
5648 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5649 TargetNumRegisters = ForceTargetNumVectorRegs;
5652 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5653 // We divide by these constants so assume that we have at least one
5654 // instruction that uses at least one register.
5655 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5656 R.NumInstructions = std::max(R.NumInstructions, 1U);
5658 // We calculate the unroll factor using the following formula.
5659 // Subtract the number of loop invariants from the number of available
5660 // registers. These registers are used by all of the unrolled instances.
5661 // Next, divide the remaining registers by the number of registers that is
5662 // required by the loop, in order to estimate how many parallel instances
5663 // fit without causing spills. All of this is rounded down if necessary to be
5664 // a power of two. We want power of two unroll factors to simplify any
5665 // addressing operations or alignment considerations.
5666 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5669 // Don't count the induction variable as unrolled.
5670 if (EnableIndVarRegisterHeur)
5671 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5672 std::max(1U, (R.MaxLocalUsers - 1)));
5674 // Clamp the unroll factor ranges to reasonable factors.
5675 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5677 // Check if the user has overridden the unroll max.
5679 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5680 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5682 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5683 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5686 // If we did not calculate the cost for VF (because the user selected the VF)
5687 // then we calculate the cost of VF here.
5689 LoopCost = expectedCost(VF);
5691 // Clamp the calculated UF to be between the 1 and the max unroll factor
5692 // that the target allows.
5693 if (UF > MaxInterleaveSize)
5694 UF = MaxInterleaveSize;
5698 // Unroll if we vectorized this loop and there is a reduction that could
5699 // benefit from unrolling.
5700 if (VF > 1 && Legal->getReductionVars()->size()) {
5701 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5705 // Note that if we've already vectorized the loop we will have done the
5706 // runtime check and so unrolling won't require further checks.
5707 bool UnrollingRequiresRuntimePointerCheck =
5708 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5710 // We want to unroll small loops in order to reduce the loop overhead and
5711 // potentially expose ILP opportunities.
5712 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5713 if (!UnrollingRequiresRuntimePointerCheck &&
5714 LoopCost < SmallLoopCost) {
5715 // We assume that the cost overhead is 1 and we use the cost model
5716 // to estimate the cost of the loop and unroll until the cost of the
5717 // loop overhead is about 5% of the cost of the loop.
5718 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5720 // Unroll until store/load ports (estimated by max unroll factor) are
5722 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5723 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5725 // If we have a scalar reduction (vector reductions are already dealt with
5726 // by this point), we can increase the critical path length if the loop
5727 // we're unrolling is inside another loop. Limit, by default to 2, so the
5728 // critical path only gets increased by one reduction operation.
5729 if (Legal->getReductionVars()->size() &&
5730 TheLoop->getLoopDepth() > 1) {
5731 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5732 SmallUF = std::min(SmallUF, F);
5733 StoresUF = std::min(StoresUF, F);
5734 LoadsUF = std::min(LoadsUF, F);
5737 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5738 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5739 return std::max(StoresUF, LoadsUF);
5742 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5746 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5750 LoopVectorizationCostModel::RegisterUsage
5751 LoopVectorizationCostModel::calculateRegisterUsage() {
5752 // This function calculates the register usage by measuring the highest number
5753 // of values that are alive at a single location. Obviously, this is a very
5754 // rough estimation. We scan the loop in a topological order in order and
5755 // assign a number to each instruction. We use RPO to ensure that defs are
5756 // met before their users. We assume that each instruction that has in-loop
5757 // users starts an interval. We record every time that an in-loop value is
5758 // used, so we have a list of the first and last occurrences of each
5759 // instruction. Next, we transpose this data structure into a multi map that
5760 // holds the list of intervals that *end* at a specific location. This multi
5761 // map allows us to perform a linear search. We scan the instructions linearly
5762 // and record each time that a new interval starts, by placing it in a set.
5763 // If we find this value in the multi-map then we remove it from the set.
5764 // The max register usage is the maximum size of the set.
5765 // We also search for instructions that are defined outside the loop, but are
5766 // used inside the loop. We need this number separately from the max-interval
5767 // usage number because when we unroll, loop-invariant values do not take
5769 LoopBlocksDFS DFS(TheLoop);
5773 R.NumInstructions = 0;
5775 // Each 'key' in the map opens a new interval. The values
5776 // of the map are the index of the 'last seen' usage of the
5777 // instruction that is the key.
5778 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5779 // Maps instruction to its index.
5780 DenseMap<unsigned, Instruction*> IdxToInstr;
5781 // Marks the end of each interval.
5782 IntervalMap EndPoint;
5783 // Saves the list of instruction indices that are used in the loop.
5784 SmallSet<Instruction*, 8> Ends;
5785 // Saves the list of values that are used in the loop but are
5786 // defined outside the loop, such as arguments and constants.
5787 SmallPtrSet<Value*, 8> LoopInvariants;
5790 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5791 be = DFS.endRPO(); bb != be; ++bb) {
5792 R.NumInstructions += (*bb)->size();
5793 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5795 Instruction *I = it;
5796 IdxToInstr[Index++] = I;
5798 // Save the end location of each USE.
5799 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5800 Value *U = I->getOperand(i);
5801 Instruction *Instr = dyn_cast<Instruction>(U);
5803 // Ignore non-instruction values such as arguments, constants, etc.
5804 if (!Instr) continue;
5806 // If this instruction is outside the loop then record it and continue.
5807 if (!TheLoop->contains(Instr)) {
5808 LoopInvariants.insert(Instr);
5812 // Overwrite previous end points.
5813 EndPoint[Instr] = Index;
5819 // Saves the list of intervals that end with the index in 'key'.
5820 typedef SmallVector<Instruction*, 2> InstrList;
5821 DenseMap<unsigned, InstrList> TransposeEnds;
5823 // Transpose the EndPoints to a list of values that end at each index.
5824 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5826 TransposeEnds[it->second].push_back(it->first);
5828 SmallSet<Instruction*, 8> OpenIntervals;
5829 unsigned MaxUsage = 0;
5832 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5833 for (unsigned int i = 0; i < Index; ++i) {
5834 Instruction *I = IdxToInstr[i];
5835 // Ignore instructions that are never used within the loop.
5836 if (!Ends.count(I)) continue;
5838 // Ignore ephemeral values.
5839 if (EphValues.count(I))
5842 // Remove all of the instructions that end at this location.
5843 InstrList &List = TransposeEnds[i];
5844 for (unsigned int j=0, e = List.size(); j < e; ++j)
5845 OpenIntervals.erase(List[j]);
5847 // Count the number of live interals.
5848 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5850 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5851 OpenIntervals.size() << '\n');
5853 // Add the current instruction to the list of open intervals.
5854 OpenIntervals.insert(I);
5857 unsigned Invariant = LoopInvariants.size();
5858 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5859 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5860 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5862 R.LoopInvariantRegs = Invariant;
5863 R.MaxLocalUsers = MaxUsage;
5867 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5871 for (Loop::block_iterator bb = TheLoop->block_begin(),
5872 be = TheLoop->block_end(); bb != be; ++bb) {
5873 unsigned BlockCost = 0;
5874 BasicBlock *BB = *bb;
5876 // For each instruction in the old loop.
5877 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5878 // Skip dbg intrinsics.
5879 if (isa<DbgInfoIntrinsic>(it))
5882 // Ignore ephemeral values.
5883 if (EphValues.count(it))
5886 unsigned C = getInstructionCost(it, VF);
5888 // Check if we should override the cost.
5889 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5890 C = ForceTargetInstructionCost;
5893 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5894 VF << " For instruction: " << *it << '\n');
5897 // We assume that if-converted blocks have a 50% chance of being executed.
5898 // When the code is scalar then some of the blocks are avoided due to CF.
5899 // When the code is vectorized we execute all code paths.
5900 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5909 /// \brief Check whether the address computation for a non-consecutive memory
5910 /// access looks like an unlikely candidate for being merged into the indexing
5913 /// We look for a GEP which has one index that is an induction variable and all
5914 /// other indices are loop invariant. If the stride of this access is also
5915 /// within a small bound we decide that this address computation can likely be
5916 /// merged into the addressing mode.
5917 /// In all other cases, we identify the address computation as complex.
5918 static bool isLikelyComplexAddressComputation(Value *Ptr,
5919 LoopVectorizationLegality *Legal,
5920 ScalarEvolution *SE,
5921 const Loop *TheLoop) {
5922 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5926 // We are looking for a gep with all loop invariant indices except for one
5927 // which should be an induction variable.
5928 unsigned NumOperands = Gep->getNumOperands();
5929 for (unsigned i = 1; i < NumOperands; ++i) {
5930 Value *Opd = Gep->getOperand(i);
5931 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5932 !Legal->isInductionVariable(Opd))
5936 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5937 // can likely be merged into the address computation.
5938 unsigned MaxMergeDistance = 64;
5940 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5944 // Check the step is constant.
5945 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5946 // Calculate the pointer stride and check if it is consecutive.
5947 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5951 const APInt &APStepVal = C->getValue()->getValue();
5953 // Huge step value - give up.
5954 if (APStepVal.getBitWidth() > 64)
5957 int64_t StepVal = APStepVal.getSExtValue();
5959 return StepVal > MaxMergeDistance;
5962 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5963 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5969 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5970 // If we know that this instruction will remain uniform, check the cost of
5971 // the scalar version.
5972 if (Legal->isUniformAfterVectorization(I))
5975 Type *RetTy = I->getType();
5976 Type *VectorTy = ToVectorTy(RetTy, VF);
5978 // TODO: We need to estimate the cost of intrinsic calls.
5979 switch (I->getOpcode()) {
5980 case Instruction::GetElementPtr:
5981 // We mark this instruction as zero-cost because the cost of GEPs in
5982 // vectorized code depends on whether the corresponding memory instruction
5983 // is scalarized or not. Therefore, we handle GEPs with the memory
5984 // instruction cost.
5986 case Instruction::Br: {
5987 return TTI.getCFInstrCost(I->getOpcode());
5989 case Instruction::PHI:
5990 //TODO: IF-converted IFs become selects.
5992 case Instruction::Add:
5993 case Instruction::FAdd:
5994 case Instruction::Sub:
5995 case Instruction::FSub:
5996 case Instruction::Mul:
5997 case Instruction::FMul:
5998 case Instruction::UDiv:
5999 case Instruction::SDiv:
6000 case Instruction::FDiv:
6001 case Instruction::URem:
6002 case Instruction::SRem:
6003 case Instruction::FRem:
6004 case Instruction::Shl:
6005 case Instruction::LShr:
6006 case Instruction::AShr:
6007 case Instruction::And:
6008 case Instruction::Or:
6009 case Instruction::Xor: {
6010 // Since we will replace the stride by 1 the multiplication should go away.
6011 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
6013 // Certain instructions can be cheaper to vectorize if they have a constant
6014 // second vector operand. One example of this are shifts on x86.
6015 TargetTransformInfo::OperandValueKind Op1VK =
6016 TargetTransformInfo::OK_AnyValue;
6017 TargetTransformInfo::OperandValueKind Op2VK =
6018 TargetTransformInfo::OK_AnyValue;
6019 TargetTransformInfo::OperandValueProperties Op1VP =
6020 TargetTransformInfo::OP_None;
6021 TargetTransformInfo::OperandValueProperties Op2VP =
6022 TargetTransformInfo::OP_None;
6023 Value *Op2 = I->getOperand(1);
6025 // Check for a splat of a constant or for a non uniform vector of constants.
6026 if (isa<ConstantInt>(Op2)) {
6027 ConstantInt *CInt = cast<ConstantInt>(Op2);
6028 if (CInt && CInt->getValue().isPowerOf2())
6029 Op2VP = TargetTransformInfo::OP_PowerOf2;
6030 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6031 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6032 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6033 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6035 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6036 if (CInt && CInt->getValue().isPowerOf2())
6037 Op2VP = TargetTransformInfo::OP_PowerOf2;
6038 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6042 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6045 case Instruction::Select: {
6046 SelectInst *SI = cast<SelectInst>(I);
6047 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6048 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6049 Type *CondTy = SI->getCondition()->getType();
6051 CondTy = VectorType::get(CondTy, VF);
6053 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6055 case Instruction::ICmp:
6056 case Instruction::FCmp: {
6057 Type *ValTy = I->getOperand(0)->getType();
6058 VectorTy = ToVectorTy(ValTy, VF);
6059 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6061 case Instruction::Store:
6062 case Instruction::Load: {
6063 StoreInst *SI = dyn_cast<StoreInst>(I);
6064 LoadInst *LI = dyn_cast<LoadInst>(I);
6065 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
6067 VectorTy = ToVectorTy(ValTy, VF);
6069 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6070 unsigned AS = SI ? SI->getPointerAddressSpace() :
6071 LI->getPointerAddressSpace();
6072 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6073 // We add the cost of address computation here instead of with the gep
6074 // instruction because only here we know whether the operation is
6077 return TTI.getAddressComputationCost(VectorTy) +
6078 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6080 // Scalarized loads/stores.
6081 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6082 bool Reverse = ConsecutiveStride < 0;
6083 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
6084 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
6085 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
6086 bool IsComplexComputation =
6087 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6089 // The cost of extracting from the value vector and pointer vector.
6090 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6091 for (unsigned i = 0; i < VF; ++i) {
6092 // The cost of extracting the pointer operand.
6093 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6094 // In case of STORE, the cost of ExtractElement from the vector.
6095 // In case of LOAD, the cost of InsertElement into the returned
6097 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6098 Instruction::InsertElement,
6102 // The cost of the scalar loads/stores.
6103 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6104 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6109 // Wide load/stores.
6110 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6111 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6114 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6118 case Instruction::ZExt:
6119 case Instruction::SExt:
6120 case Instruction::FPToUI:
6121 case Instruction::FPToSI:
6122 case Instruction::FPExt:
6123 case Instruction::PtrToInt:
6124 case Instruction::IntToPtr:
6125 case Instruction::SIToFP:
6126 case Instruction::UIToFP:
6127 case Instruction::Trunc:
6128 case Instruction::FPTrunc:
6129 case Instruction::BitCast: {
6130 // We optimize the truncation of induction variable.
6131 // The cost of these is the same as the scalar operation.
6132 if (I->getOpcode() == Instruction::Trunc &&
6133 Legal->isInductionVariable(I->getOperand(0)))
6134 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6135 I->getOperand(0)->getType());
6137 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6138 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6140 case Instruction::Call: {
6141 CallInst *CI = cast<CallInst>(I);
6142 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6143 assert(ID && "Not an intrinsic call!");
6144 Type *RetTy = ToVectorTy(CI->getType(), VF);
6145 SmallVector<Type*, 4> Tys;
6146 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6147 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6148 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6151 // We are scalarizing the instruction. Return the cost of the scalar
6152 // instruction, plus the cost of insert and extract into vector
6153 // elements, times the vector width.
6156 if (!RetTy->isVoidTy() && VF != 1) {
6157 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6159 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6162 // The cost of inserting the results plus extracting each one of the
6164 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6167 // The cost of executing VF copies of the scalar instruction. This opcode
6168 // is unknown. Assume that it is the same as 'mul'.
6169 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6175 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6176 if (Scalar->isVoidTy() || VF == 1)
6178 return VectorType::get(Scalar, VF);
6181 char LoopVectorize::ID = 0;
6182 static const char lv_name[] = "Loop Vectorization";
6183 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6184 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
6185 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6186 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
6187 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6188 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6189 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6190 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6191 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
6192 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6193 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6196 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6197 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6201 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6202 // Check for a store.
6203 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6204 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6206 // Check for a load.
6207 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6208 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6214 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6215 bool IfPredicateStore) {
6216 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6217 // Holds vector parameters or scalars, in case of uniform vals.
6218 SmallVector<VectorParts, 4> Params;
6220 setDebugLocFromInst(Builder, Instr);
6222 // Find all of the vectorized parameters.
6223 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6224 Value *SrcOp = Instr->getOperand(op);
6226 // If we are accessing the old induction variable, use the new one.
6227 if (SrcOp == OldInduction) {
6228 Params.push_back(getVectorValue(SrcOp));
6232 // Try using previously calculated values.
6233 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6235 // If the src is an instruction that appeared earlier in the basic block
6236 // then it should already be vectorized.
6237 if (SrcInst && OrigLoop->contains(SrcInst)) {
6238 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6239 // The parameter is a vector value from earlier.
6240 Params.push_back(WidenMap.get(SrcInst));
6242 // The parameter is a scalar from outside the loop. Maybe even a constant.
6243 VectorParts Scalars;
6244 Scalars.append(UF, SrcOp);
6245 Params.push_back(Scalars);
6249 assert(Params.size() == Instr->getNumOperands() &&
6250 "Invalid number of operands");
6252 // Does this instruction return a value ?
6253 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6255 Value *UndefVec = IsVoidRetTy ? nullptr :
6256 UndefValue::get(Instr->getType());
6257 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6258 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6260 Instruction *InsertPt = Builder.GetInsertPoint();
6261 BasicBlock *IfBlock = Builder.GetInsertBlock();
6262 BasicBlock *CondBlock = nullptr;
6265 Loop *VectorLp = nullptr;
6266 if (IfPredicateStore) {
6267 assert(Instr->getParent()->getSinglePredecessor() &&
6268 "Only support single predecessor blocks");
6269 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6270 Instr->getParent());
6271 VectorLp = LI->getLoopFor(IfBlock);
6272 assert(VectorLp && "Must have a loop for this block");
6275 // For each vector unroll 'part':
6276 for (unsigned Part = 0; Part < UF; ++Part) {
6277 // For each scalar that we create:
6279 // Start an "if (pred) a[i] = ..." block.
6280 Value *Cmp = nullptr;
6281 if (IfPredicateStore) {
6282 if (Cond[Part]->getType()->isVectorTy())
6284 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6285 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6286 ConstantInt::get(Cond[Part]->getType(), 1));
6287 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6288 LoopVectorBody.push_back(CondBlock);
6289 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
6290 // Update Builder with newly created basic block.
6291 Builder.SetInsertPoint(InsertPt);
6294 Instruction *Cloned = Instr->clone();
6296 Cloned->setName(Instr->getName() + ".cloned");
6297 // Replace the operands of the cloned instructions with extracted scalars.
6298 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6299 Value *Op = Params[op][Part];
6300 Cloned->setOperand(op, Op);
6303 // Place the cloned scalar in the new loop.
6304 Builder.Insert(Cloned);
6306 // If the original scalar returns a value we need to place it in a vector
6307 // so that future users will be able to use it.
6309 VecResults[Part] = Cloned;
6312 if (IfPredicateStore) {
6313 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6314 LoopVectorBody.push_back(NewIfBlock);
6315 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
6316 Builder.SetInsertPoint(InsertPt);
6317 Instruction *OldBr = IfBlock->getTerminator();
6318 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6319 OldBr->eraseFromParent();
6320 IfBlock = NewIfBlock;
6325 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6326 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6327 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6329 return scalarizeInstruction(Instr, IfPredicateStore);
6332 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6336 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6340 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
6341 // When unrolling and the VF is 1, we only need to add a simple scalar.
6342 Type *ITy = Val->getType();
6343 assert(!ITy->isVectorTy() && "Val must be a scalar");
6344 Constant *C = ConstantInt::get(ITy, StartIdx);
6345 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");