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/BlockFrequencyInfo.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugInfo.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/DiagnosticInfo.h"
72 #include "llvm/IR/Dominators.h"
73 #include "llvm/IR/Function.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/Instructions.h"
76 #include "llvm/IR/IntrinsicInst.h"
77 #include "llvm/IR/LLVMContext.h"
78 #include "llvm/IR/Module.h"
79 #include "llvm/IR/PatternMatch.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/Value.h"
82 #include "llvm/IR/ValueHandle.h"
83 #include "llvm/IR/Verifier.h"
84 #include "llvm/Pass.h"
85 #include "llvm/Support/BranchProbability.h"
86 #include "llvm/Support/CommandLine.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/Local.h"
92 #include "llvm/Transforms/Utils/VectorUtils.h"
98 using namespace llvm::PatternMatch;
100 #define LV_NAME "loop-vectorize"
101 #define DEBUG_TYPE LV_NAME
103 STATISTIC(LoopsVectorized, "Number of loops vectorized");
104 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
106 static cl::opt<unsigned>
107 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
108 cl::desc("Sets the SIMD width. Zero is autoselect."));
110 static cl::opt<unsigned>
111 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
112 cl::desc("Sets the vectorization unroll count. "
113 "Zero is autoselect."));
116 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
117 cl::desc("Enable if-conversion during vectorization."));
119 /// We don't vectorize loops with a known constant trip count below this number.
120 static cl::opt<unsigned>
121 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
123 cl::desc("Don't vectorize loops with a constant "
124 "trip count that is smaller than this "
127 /// This enables versioning on the strides of symbolically striding memory
128 /// accesses in code like the following.
129 /// for (i = 0; i < N; ++i)
130 /// A[i * Stride1] += B[i * Stride2] ...
132 /// Will be roughly translated to
133 /// if (Stride1 == 1 && Stride2 == 1) {
134 /// for (i = 0; i < N; i+=4)
138 static cl::opt<bool> EnableMemAccessVersioning(
139 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
140 cl::desc("Enable symblic stride memory access versioning"));
142 /// We don't unroll loops with a known constant trip count below this number.
143 static const unsigned TinyTripCountUnrollThreshold = 128;
145 /// When performing memory disambiguation checks at runtime do not make more
146 /// than this number of comparisons.
147 static const unsigned RuntimeMemoryCheckThreshold = 8;
149 /// Maximum simd width.
150 static const unsigned MaxVectorWidth = 64;
152 static cl::opt<unsigned> ForceTargetNumScalarRegs(
153 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
154 cl::desc("A flag that overrides the target's number of scalar registers."));
156 static cl::opt<unsigned> ForceTargetNumVectorRegs(
157 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
158 cl::desc("A flag that overrides the target's number of vector registers."));
160 /// Maximum vectorization unroll count.
161 static const unsigned MaxUnrollFactor = 16;
163 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
164 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's max unroll factor for scalar "
168 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
169 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
170 cl::desc("A flag that overrides the target's max unroll factor for "
171 "vectorized loops."));
173 static cl::opt<unsigned> ForceTargetInstructionCost(
174 "force-target-instruction-cost", cl::init(0), cl::Hidden,
175 cl::desc("A flag that overrides the target's expected cost for "
176 "an instruction to a single constant value. Mostly "
177 "useful for getting consistent testing."));
179 static cl::opt<unsigned> SmallLoopCost(
180 "small-loop-cost", cl::init(20), cl::Hidden,
181 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
183 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
184 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
185 cl::desc("Enable the use of the block frequency analysis to access PGO "
186 "heuristics minimizing code growth in cold regions and being more "
187 "aggressive in hot regions."));
189 // Runtime unroll loops for load/store throughput.
190 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
191 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
192 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
194 /// The number of stores in a loop that are allowed to need predication.
195 static cl::opt<unsigned> NumberOfStoresToPredicate(
196 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
197 cl::desc("Max number of stores to be predicated behind an if."));
199 static cl::opt<bool> EnableIndVarRegisterHeur(
200 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
201 cl::desc("Count the induction variable only once when unrolling"));
203 static cl::opt<bool> EnableCondStoresVectorization(
204 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
205 cl::desc("Enable if predication of stores during vectorization."));
209 // Forward declarations.
210 class LoopVectorizationLegality;
211 class LoopVectorizationCostModel;
213 /// Optimization analysis message produced during vectorization. Messages inform
214 /// the user why vectorization did not occur.
217 raw_string_ostream Out;
221 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
222 Out << "loop not vectorized: ";
225 template <typename A> Report &operator<<(const A &Value) {
230 Instruction *getInstr() { return Instr; }
232 std::string &str() { return Out.str(); }
233 operator Twine() { return Out.str(); }
236 /// InnerLoopVectorizer vectorizes loops which contain only one basic
237 /// block to a specified vectorization factor (VF).
238 /// This class performs the widening of scalars into vectors, or multiple
239 /// scalars. This class also implements the following features:
240 /// * It inserts an epilogue loop for handling loops that don't have iteration
241 /// counts that are known to be a multiple of the vectorization factor.
242 /// * It handles the code generation for reduction variables.
243 /// * Scalarization (implementation using scalars) of un-vectorizable
245 /// InnerLoopVectorizer does not perform any vectorization-legality
246 /// checks, and relies on the caller to check for the different legality
247 /// aspects. The InnerLoopVectorizer relies on the
248 /// LoopVectorizationLegality class to provide information about the induction
249 /// and reduction variables that were found to a given vectorization factor.
250 class InnerLoopVectorizer {
252 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
253 DominatorTree *DT, const DataLayout *DL,
254 const TargetLibraryInfo *TLI, unsigned VecWidth,
255 unsigned UnrollFactor)
256 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
257 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
258 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
261 // Perform the actual loop widening (vectorization).
262 void vectorize(LoopVectorizationLegality *L) {
264 // Create a new empty loop. Unlink the old loop and connect the new one.
266 // Widen each instruction in the old loop to a new one in the new loop.
267 // Use the Legality module to find the induction and reduction variables.
269 // Register the new loop and update the analysis passes.
273 virtual ~InnerLoopVectorizer() {}
276 /// A small list of PHINodes.
277 typedef SmallVector<PHINode*, 4> PhiVector;
278 /// When we unroll loops we have multiple vector values for each scalar.
279 /// This data structure holds the unrolled and vectorized values that
280 /// originated from one scalar instruction.
281 typedef SmallVector<Value*, 2> VectorParts;
283 // When we if-convert we need create edge masks. We have to cache values so
284 // that we don't end up with exponential recursion/IR.
285 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
286 VectorParts> EdgeMaskCache;
288 /// \brief Add code that checks at runtime if the accessed arrays overlap.
290 /// Returns a pair of instructions where the first element is the first
291 /// instruction generated in possibly a sequence of instructions and the
292 /// second value is the final comparator value or NULL if no check is needed.
293 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
295 /// \brief Add checks for strides that where assumed to be 1.
297 /// Returns the last check instruction and the first check instruction in the
298 /// pair as (first, last).
299 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
301 /// Create an empty loop, based on the loop ranges of the old loop.
302 void createEmptyLoop();
303 /// Copy and widen the instructions from the old loop.
304 virtual void vectorizeLoop();
306 /// \brief The Loop exit block may have single value PHI nodes where the
307 /// incoming value is 'Undef'. While vectorizing we only handled real values
308 /// that were defined inside the loop. Here we fix the 'undef case'.
312 /// A helper function that computes the predicate of the block BB, assuming
313 /// that the header block of the loop is set to True. It returns the *entry*
314 /// mask for the block BB.
315 VectorParts createBlockInMask(BasicBlock *BB);
316 /// A helper function that computes the predicate of the edge between SRC
318 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
320 /// A helper function to vectorize a single BB within the innermost loop.
321 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
323 /// Vectorize a single PHINode in a block. This method handles the induction
324 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
325 /// arbitrary length vectors.
326 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
327 unsigned UF, unsigned VF, PhiVector *PV);
329 /// Insert the new loop to the loop hierarchy and pass manager
330 /// and update the analysis passes.
331 void updateAnalysis();
333 /// This instruction is un-vectorizable. Implement it as a sequence
334 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
335 /// scalarized instruction behind an if block predicated on the control
336 /// dependence of the instruction.
337 virtual void scalarizeInstruction(Instruction *Instr,
338 bool IfPredicateStore=false);
340 /// Vectorize Load and Store instructions,
341 virtual void vectorizeMemoryInstruction(Instruction *Instr);
343 /// Create a broadcast instruction. This method generates a broadcast
344 /// instruction (shuffle) for loop invariant values and for the induction
345 /// value. If this is the induction variable then we extend it to N, N+1, ...
346 /// this is needed because each iteration in the loop corresponds to a SIMD
348 virtual Value *getBroadcastInstrs(Value *V);
350 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
351 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
352 /// The sequence starts at StartIndex.
353 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
355 /// When we go over instructions in the basic block we rely on previous
356 /// values within the current basic block or on loop invariant values.
357 /// When we widen (vectorize) values we place them in the map. If the values
358 /// are not within the map, they have to be loop invariant, so we simply
359 /// broadcast them into a vector.
360 VectorParts &getVectorValue(Value *V);
362 /// Generate a shuffle sequence that will reverse the vector Vec.
363 virtual Value *reverseVector(Value *Vec);
365 /// This is a helper class that holds the vectorizer state. It maps scalar
366 /// instructions to vector instructions. When the code is 'unrolled' then
367 /// then a single scalar value is mapped to multiple vector parts. The parts
368 /// are stored in the VectorPart type.
370 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
372 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
374 /// \return True if 'Key' is saved in the Value Map.
375 bool has(Value *Key) const { return MapStorage.count(Key); }
377 /// Initializes a new entry in the map. Sets all of the vector parts to the
378 /// save value in 'Val'.
379 /// \return A reference to a vector with splat values.
380 VectorParts &splat(Value *Key, Value *Val) {
381 VectorParts &Entry = MapStorage[Key];
382 Entry.assign(UF, Val);
386 ///\return A reference to the value that is stored at 'Key'.
387 VectorParts &get(Value *Key) {
388 VectorParts &Entry = MapStorage[Key];
391 assert(Entry.size() == UF);
396 /// The unroll factor. Each entry in the map stores this number of vector
400 /// Map storage. We use std::map and not DenseMap because insertions to a
401 /// dense map invalidates its iterators.
402 std::map<Value *, VectorParts> MapStorage;
405 /// The original loop.
407 /// Scev analysis to use.
416 const DataLayout *DL;
417 /// Target Library Info.
418 const TargetLibraryInfo *TLI;
420 /// The vectorization SIMD factor to use. Each vector will have this many
425 /// The vectorization unroll factor to use. Each scalar is vectorized to this
426 /// many different vector instructions.
429 /// The builder that we use
432 // --- Vectorization state ---
434 /// The vector-loop preheader.
435 BasicBlock *LoopVectorPreHeader;
436 /// The scalar-loop preheader.
437 BasicBlock *LoopScalarPreHeader;
438 /// Middle Block between the vector and the scalar.
439 BasicBlock *LoopMiddleBlock;
440 ///The ExitBlock of the scalar loop.
441 BasicBlock *LoopExitBlock;
442 ///The vector loop body.
443 SmallVector<BasicBlock *, 4> LoopVectorBody;
444 ///The scalar loop body.
445 BasicBlock *LoopScalarBody;
446 /// A list of all bypass blocks. The first block is the entry of the loop.
447 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
449 /// The new Induction variable which was added to the new block.
451 /// The induction variable of the old basic block.
452 PHINode *OldInduction;
453 /// Holds the extended (to the widest induction type) start index.
455 /// Maps scalars to widened vectors.
457 EdgeMaskCache MaskCache;
459 LoopVectorizationLegality *Legal;
462 class InnerLoopUnroller : public InnerLoopVectorizer {
464 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
465 DominatorTree *DT, const DataLayout *DL,
466 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
467 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
470 void scalarizeInstruction(Instruction *Instr,
471 bool IfPredicateStore = false) override;
472 void vectorizeMemoryInstruction(Instruction *Instr) override;
473 Value *getBroadcastInstrs(Value *V) override;
474 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
475 Value *reverseVector(Value *Vec) override;
478 /// \brief Look for a meaningful debug location on the instruction or it's
480 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
485 if (I->getDebugLoc() != Empty)
488 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
489 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
490 if (OpInst->getDebugLoc() != Empty)
497 /// \brief Set the debug location in the builder using the debug location in the
499 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
500 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
501 B.SetCurrentDebugLocation(Inst->getDebugLoc());
503 B.SetCurrentDebugLocation(DebugLoc());
507 /// \return string containing a file name and a line # for the given loop.
508 static std::string getDebugLocString(const Loop *L) {
511 raw_string_ostream OS(Result);
512 const DebugLoc LoopDbgLoc = L->getStartLoc();
513 if (!LoopDbgLoc.isUnknown())
514 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
516 // Just print the module name.
517 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
524 /// \brief Propagate known metadata from one instruction to another.
525 static void propagateMetadata(Instruction *To, const Instruction *From) {
526 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
527 From->getAllMetadataOtherThanDebugLoc(Metadata);
529 for (auto M : Metadata) {
530 unsigned Kind = M.first;
532 // These are safe to transfer (this is safe for TBAA, even when we
533 // if-convert, because should that metadata have had a control dependency
534 // on the condition, and thus actually aliased with some other
535 // non-speculated memory access when the condition was false, this would be
536 // caught by the runtime overlap checks).
537 if (Kind != LLVMContext::MD_tbaa &&
538 Kind != LLVMContext::MD_fpmath)
541 To->setMetadata(Kind, M.second);
545 /// \brief Propagate known metadata from one instruction to a vector of others.
546 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
548 if (Instruction *I = dyn_cast<Instruction>(V))
549 propagateMetadata(I, From);
552 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
553 /// to what vectorization factor.
554 /// This class does not look at the profitability of vectorization, only the
555 /// legality. This class has two main kinds of checks:
556 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
557 /// will change the order of memory accesses in a way that will change the
558 /// correctness of the program.
559 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
560 /// checks for a number of different conditions, such as the availability of a
561 /// single induction variable, that all types are supported and vectorize-able,
562 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
563 /// This class is also used by InnerLoopVectorizer for identifying
564 /// induction variable and the different reduction variables.
565 class LoopVectorizationLegality {
569 unsigned NumPredStores;
571 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
572 DominatorTree *DT, TargetLibraryInfo *TLI,
573 AliasAnalysis *AA, Function *F)
574 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
575 DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
576 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
579 /// This enum represents the kinds of reductions that we support.
581 RK_NoReduction, ///< Not a reduction.
582 RK_IntegerAdd, ///< Sum of integers.
583 RK_IntegerMult, ///< Product of integers.
584 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
585 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
586 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
587 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
588 RK_FloatAdd, ///< Sum of floats.
589 RK_FloatMult, ///< Product of floats.
590 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
593 /// This enum represents the kinds of inductions that we support.
595 IK_NoInduction, ///< Not an induction variable.
596 IK_IntInduction, ///< Integer induction variable. Step = 1.
597 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
598 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
599 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
602 // This enum represents the kind of minmax reduction.
603 enum MinMaxReductionKind {
613 /// This struct holds information about reduction variables.
614 struct ReductionDescriptor {
615 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
616 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
618 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
619 MinMaxReductionKind MK)
620 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
622 // The starting value of the reduction.
623 // It does not have to be zero!
624 TrackingVH<Value> StartValue;
625 // The instruction who's value is used outside the loop.
626 Instruction *LoopExitInstr;
627 // The kind of the reduction.
629 // If this a min/max reduction the kind of reduction.
630 MinMaxReductionKind MinMaxKind;
633 /// This POD struct holds information about a potential reduction operation.
634 struct ReductionInstDesc {
635 ReductionInstDesc(bool IsRedux, Instruction *I) :
636 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
638 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
639 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
641 // Is this instruction a reduction candidate.
643 // The last instruction in a min/max pattern (select of the select(icmp())
644 // pattern), or the current reduction instruction otherwise.
645 Instruction *PatternLastInst;
646 // If this is a min/max pattern the comparison predicate.
647 MinMaxReductionKind MinMaxKind;
650 /// This struct holds information about the memory runtime legality
651 /// check that a group of pointers do not overlap.
652 struct RuntimePointerCheck {
653 RuntimePointerCheck() : Need(false) {}
655 /// Reset the state of the pointer runtime information.
662 DependencySetId.clear();
666 /// Insert a pointer and calculate the start and end SCEVs.
667 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
668 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
670 /// This flag indicates if we need to add the runtime check.
672 /// Holds the pointers that we need to check.
673 SmallVector<TrackingVH<Value>, 2> Pointers;
674 /// Holds the pointer value at the beginning of the loop.
675 SmallVector<const SCEV*, 2> Starts;
676 /// Holds the pointer value at the end of the loop.
677 SmallVector<const SCEV*, 2> Ends;
678 /// Holds the information if this pointer is used for writing to memory.
679 SmallVector<bool, 2> IsWritePtr;
680 /// Holds the id of the set of pointers that could be dependent because of a
681 /// shared underlying object.
682 SmallVector<unsigned, 2> DependencySetId;
683 /// Holds the id of the disjoint alias set to which this pointer belongs.
684 SmallVector<unsigned, 2> AliasSetId;
687 /// A struct for saving information about induction variables.
688 struct InductionInfo {
689 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
690 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
692 TrackingVH<Value> StartValue;
697 /// ReductionList contains the reduction descriptors for all
698 /// of the reductions that were found in the loop.
699 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
701 /// InductionList saves induction variables and maps them to the
702 /// induction descriptor.
703 typedef MapVector<PHINode*, InductionInfo> InductionList;
705 /// Returns true if it is legal to vectorize this loop.
706 /// This does not mean that it is profitable to vectorize this
707 /// loop, only that it is legal to do so.
710 /// Returns the Induction variable.
711 PHINode *getInduction() { return Induction; }
713 /// Returns the reduction variables found in the loop.
714 ReductionList *getReductionVars() { return &Reductions; }
716 /// Returns the induction variables found in the loop.
717 InductionList *getInductionVars() { return &Inductions; }
719 /// Returns the widest induction type.
720 Type *getWidestInductionType() { return WidestIndTy; }
722 /// Returns True if V is an induction variable in this loop.
723 bool isInductionVariable(const Value *V);
725 /// Return true if the block BB needs to be predicated in order for the loop
726 /// to be vectorized.
727 bool blockNeedsPredication(BasicBlock *BB);
729 /// Check if this pointer is consecutive when vectorizing. This happens
730 /// when the last index of the GEP is the induction variable, or that the
731 /// pointer itself is an induction variable.
732 /// This check allows us to vectorize A[idx] into a wide load/store.
734 /// 0 - Stride is unknown or non-consecutive.
735 /// 1 - Address is consecutive.
736 /// -1 - Address is consecutive, and decreasing.
737 int isConsecutivePtr(Value *Ptr);
739 /// Returns true if the value V is uniform within the loop.
740 bool isUniform(Value *V);
742 /// Returns true if this instruction will remain scalar after vectorization.
743 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
745 /// Returns the information that we collected about runtime memory check.
746 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
748 /// This function returns the identity element (or neutral element) for
750 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
752 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
754 bool hasStride(Value *V) { return StrideSet.count(V); }
755 bool mustCheckStrides() { return !StrideSet.empty(); }
756 SmallPtrSet<Value *, 8>::iterator strides_begin() {
757 return StrideSet.begin();
759 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
762 /// Check if a single basic block loop is vectorizable.
763 /// At this point we know that this is a loop with a constant trip count
764 /// and we only need to check individual instructions.
765 bool canVectorizeInstrs();
767 /// When we vectorize loops we may change the order in which
768 /// we read and write from memory. This method checks if it is
769 /// legal to vectorize the code, considering only memory constrains.
770 /// Returns true if the loop is vectorizable
771 bool canVectorizeMemory();
773 /// Return true if we can vectorize this loop using the IF-conversion
775 bool canVectorizeWithIfConvert();
777 /// Collect the variables that need to stay uniform after vectorization.
778 void collectLoopUniforms();
780 /// Return true if all of the instructions in the block can be speculatively
781 /// executed. \p SafePtrs is a list of addresses that are known to be legal
782 /// and we know that we can read from them without segfault.
783 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
785 /// Returns True, if 'Phi' is the kind of reduction variable for type
786 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
787 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
788 /// Returns a struct describing if the instruction 'I' can be a reduction
789 /// variable of type 'Kind'. If the reduction is a min/max pattern of
790 /// select(icmp()) this function advances the instruction pointer 'I' from the
791 /// compare instruction to the select instruction and stores this pointer in
792 /// 'PatternLastInst' member of the returned struct.
793 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
794 ReductionInstDesc &Desc);
795 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
796 /// pattern corresponding to a min(X, Y) or max(X, Y).
797 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
798 ReductionInstDesc &Prev);
799 /// Returns the induction kind of Phi. This function may return NoInduction
800 /// if the PHI is not an induction variable.
801 InductionKind isInductionVariable(PHINode *Phi);
803 /// \brief Collect memory access with loop invariant strides.
805 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
807 void collectStridedAcccess(Value *LoadOrStoreInst);
809 /// Report an analysis message to assist the user in diagnosing loops that are
811 void emitAnalysis(Report &Message) {
812 DebugLoc DL = TheLoop->getStartLoc();
813 if (Instruction *I = Message.getInstr())
814 DL = I->getDebugLoc();
815 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
816 *TheFunction, DL, Message.str());
819 /// The loop that we evaluate.
823 /// DataLayout analysis.
824 const DataLayout *DL;
827 /// Target Library Info.
828 TargetLibraryInfo *TLI;
832 Function *TheFunction;
834 // --- vectorization state --- //
836 /// Holds the integer induction variable. This is the counter of the
839 /// Holds the reduction variables.
840 ReductionList Reductions;
841 /// Holds all of the induction variables that we found in the loop.
842 /// Notice that inductions don't need to start at zero and that induction
843 /// variables can be pointers.
844 InductionList Inductions;
845 /// Holds the widest induction type encountered.
848 /// Allowed outside users. This holds the reduction
849 /// vars which can be accessed from outside the loop.
850 SmallPtrSet<Value*, 4> AllowedExit;
851 /// This set holds the variables which are known to be uniform after
853 SmallPtrSet<Instruction*, 4> Uniforms;
854 /// We need to check that all of the pointers in this list are disjoint
856 RuntimePointerCheck PtrRtCheck;
857 /// Can we assume the absence of NaNs.
858 bool HasFunNoNaNAttr;
860 unsigned MaxSafeDepDistBytes;
862 ValueToValueMap Strides;
863 SmallPtrSet<Value *, 8> StrideSet;
866 /// LoopVectorizationCostModel - estimates the expected speedups due to
868 /// In many cases vectorization is not profitable. This can happen because of
869 /// a number of reasons. In this class we mainly attempt to predict the
870 /// expected speedup/slowdowns due to the supported instruction set. We use the
871 /// TargetTransformInfo to query the different backends for the cost of
872 /// different operations.
873 class LoopVectorizationCostModel {
875 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
876 LoopVectorizationLegality *Legal,
877 const TargetTransformInfo &TTI,
878 const DataLayout *DL, const TargetLibraryInfo *TLI)
879 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
881 /// Information about vectorization costs
882 struct VectorizationFactor {
883 unsigned Width; // Vector width with best cost
884 unsigned Cost; // Cost of the loop with that width
886 /// \return The most profitable vectorization factor and the cost of that VF.
887 /// This method checks every power of two up to VF. If UserVF is not ZERO
888 /// then this vectorization factor will be selected if vectorization is
890 VectorizationFactor selectVectorizationFactor(bool OptForSize,
892 bool ForceVectorization);
894 /// \return The size (in bits) of the widest type in the code that
895 /// needs to be vectorized. We ignore values that remain scalar such as
896 /// 64 bit loop indices.
897 unsigned getWidestType();
899 /// \return The most profitable unroll factor.
900 /// If UserUF is non-zero then this method finds the best unroll-factor
901 /// based on register pressure and other parameters.
902 /// VF and LoopCost are the selected vectorization factor and the cost of the
904 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
907 /// \brief A struct that represents some properties of the register usage
909 struct RegisterUsage {
910 /// Holds the number of loop invariant values that are used in the loop.
911 unsigned LoopInvariantRegs;
912 /// Holds the maximum number of concurrent live intervals in the loop.
913 unsigned MaxLocalUsers;
914 /// Holds the number of instructions in the loop.
915 unsigned NumInstructions;
918 /// \return information about the register usage of the loop.
919 RegisterUsage calculateRegisterUsage();
922 /// Returns the expected execution cost. The unit of the cost does
923 /// not matter because we use the 'cost' units to compare different
924 /// vector widths. The cost that is returned is *not* normalized by
925 /// the factor width.
926 unsigned expectedCost(unsigned VF);
928 /// Returns the execution time cost of an instruction for a given vector
929 /// width. Vector width of one means scalar.
930 unsigned getInstructionCost(Instruction *I, unsigned VF);
932 /// A helper function for converting Scalar types to vector types.
933 /// If the incoming type is void, we return void. If the VF is 1, we return
935 static Type* ToVectorTy(Type *Scalar, unsigned VF);
937 /// Returns whether the instruction is a load or store and will be a emitted
938 /// as a vector operation.
939 bool isConsecutiveLoadOrStore(Instruction *I);
941 /// The loop that we evaluate.
945 /// Loop Info analysis.
947 /// Vectorization legality.
948 LoopVectorizationLegality *Legal;
949 /// Vector target information.
950 const TargetTransformInfo &TTI;
951 /// Target data layout information.
952 const DataLayout *DL;
953 /// Target Library Info.
954 const TargetLibraryInfo *TLI;
957 /// Utility class for getting and setting loop vectorizer hints in the form
958 /// of loop metadata.
959 class LoopVectorizeHints {
962 FK_Undefined = -1, ///< Not selected.
963 FK_Disabled = 0, ///< Forcing disabled.
964 FK_Enabled = 1, ///< Forcing enabled.
967 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
968 : Width(VectorizationFactor),
969 Unroll(DisableUnrolling),
971 LoopID(L->getLoopID()) {
973 // force-vector-unroll overrides DisableUnrolling.
974 if (VectorizationUnroll.getNumOccurrences() > 0)
975 Unroll = VectorizationUnroll;
977 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
978 << "LV: Unrolling disabled by the pass manager\n");
981 /// Return the loop metadata prefix.
982 static StringRef Prefix() { return "llvm.loop."; }
984 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
985 SmallVector<Value*, 2> Vals;
986 Vals.push_back(MDString::get(Context, Name));
987 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
988 return MDNode::get(Context, Vals);
991 /// Mark the loop L as already vectorized by setting the width to 1.
992 void setAlreadyVectorized(Loop *L) {
993 LLVMContext &Context = L->getHeader()->getContext();
997 // Create a new loop id with one more operand for the already_vectorized
998 // hint. If the loop already has a loop id then copy the existing operands.
999 SmallVector<Value*, 4> Vals(1);
1001 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
1002 Vals.push_back(LoopID->getOperand(i));
1005 createHint(Context, Twine(Prefix(), "vectorize.width").str(), Width));
1007 createHint(Context, Twine(Prefix(), "interleave.count").str(), 1));
1009 MDNode *NewLoopID = MDNode::get(Context, Vals);
1010 // Set operand 0 to refer to the loop id itself.
1011 NewLoopID->replaceOperandWith(0, NewLoopID);
1013 L->setLoopID(NewLoopID);
1015 LoopID->replaceAllUsesWith(NewLoopID);
1020 std::string emitRemark() const {
1022 R << "vectorization ";
1024 case LoopVectorizeHints::FK_Disabled:
1025 R << "is explicitly disabled";
1027 case LoopVectorizeHints::FK_Enabled:
1028 R << "is explicitly enabled";
1029 if (Width != 0 && Unroll != 0)
1030 R << " with width " << Width << " and interleave count " << Unroll;
1031 else if (Width != 0)
1032 R << " with width " << Width;
1033 else if (Unroll != 0)
1034 R << " with interleave count " << Unroll;
1036 case LoopVectorizeHints::FK_Undefined:
1037 R << "was not specified";
1043 unsigned getWidth() const { return Width; }
1044 unsigned getUnroll() const { return Unroll; }
1045 enum ForceKind getForce() const { return Force; }
1046 MDNode *getLoopID() const { return LoopID; }
1049 /// Find hints specified in the loop metadata.
1050 void getHints(const Loop *L) {
1054 // First operand should refer to the loop id itself.
1055 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1056 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1058 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1059 const MDString *S = nullptr;
1060 SmallVector<Value*, 4> Args;
1062 // The expected hint is either a MDString or a MDNode with the first
1063 // operand a MDString.
1064 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1065 if (!MD || MD->getNumOperands() == 0)
1067 S = dyn_cast<MDString>(MD->getOperand(0));
1068 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1069 Args.push_back(MD->getOperand(i));
1071 S = dyn_cast<MDString>(LoopID->getOperand(i));
1072 assert(Args.size() == 0 && "too many arguments for MDString");
1078 // Check if the hint starts with the loop metadata prefix.
1079 StringRef Hint = S->getString();
1080 if (!Hint.startswith(Prefix()))
1082 // Remove the prefix.
1083 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1085 if (Args.size() == 1)
1086 getHint(Hint, Args[0]);
1090 // Check string hint with one operand.
1091 void getHint(StringRef Hint, Value *Arg) {
1092 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1094 unsigned Val = C->getZExtValue();
1096 if (Hint == "vectorize.width") {
1097 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1100 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1101 } else if (Hint == "vectorize.enable") {
1102 if (C->getBitWidth() == 1)
1103 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1104 : LoopVectorizeHints::FK_Disabled;
1106 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1107 } else if (Hint == "interleave.count") {
1108 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1111 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1113 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1117 /// Vectorization width.
1119 /// Vectorization unroll factor.
1121 /// Vectorization forced
1122 enum ForceKind Force;
1127 static void emitMissedWarning(Function *F, Loop *L,
1128 const LoopVectorizeHints &LH) {
1129 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1130 L->getStartLoc(), LH.emitRemark());
1132 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1133 if (LH.getWidth() != 1)
1134 emitLoopVectorizeWarning(
1135 F->getContext(), *F, L->getStartLoc(),
1136 "failed explicitly specified loop vectorization");
1137 else if (LH.getUnroll() != 1)
1138 emitLoopInterleaveWarning(
1139 F->getContext(), *F, L->getStartLoc(),
1140 "failed explicitly specified loop interleaving");
1144 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1146 return V.push_back(&L);
1148 for (Loop *InnerL : L)
1149 addInnerLoop(*InnerL, V);
1152 /// The LoopVectorize Pass.
1153 struct LoopVectorize : public FunctionPass {
1154 /// Pass identification, replacement for typeid
1157 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1159 DisableUnrolling(NoUnrolling),
1160 AlwaysVectorize(AlwaysVectorize) {
1161 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1164 ScalarEvolution *SE;
1165 const DataLayout *DL;
1167 TargetTransformInfo *TTI;
1169 BlockFrequencyInfo *BFI;
1170 TargetLibraryInfo *TLI;
1172 bool DisableUnrolling;
1173 bool AlwaysVectorize;
1175 BlockFrequency ColdEntryFreq;
1177 bool runOnFunction(Function &F) override {
1178 SE = &getAnalysis<ScalarEvolution>();
1179 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1180 DL = DLP ? &DLP->getDataLayout() : nullptr;
1181 LI = &getAnalysis<LoopInfo>();
1182 TTI = &getAnalysis<TargetTransformInfo>();
1183 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1184 BFI = &getAnalysis<BlockFrequencyInfo>();
1185 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1186 AA = &getAnalysis<AliasAnalysis>();
1188 // Compute some weights outside of the loop over the loops. Compute this
1189 // using a BranchProbability to re-use its scaling math.
1190 const BranchProbability ColdProb(1, 5); // 20%
1191 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1193 // If the target claims to have no vector registers don't attempt
1195 if (!TTI->getNumberOfRegisters(true))
1199 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1200 << ": Missing data layout\n");
1204 // Build up a worklist of inner-loops to vectorize. This is necessary as
1205 // the act of vectorizing or partially unrolling a loop creates new loops
1206 // and can invalidate iterators across the loops.
1207 SmallVector<Loop *, 8> Worklist;
1210 addInnerLoop(*L, Worklist);
1212 LoopsAnalyzed += Worklist.size();
1214 // Now walk the identified inner loops.
1215 bool Changed = false;
1216 while (!Worklist.empty())
1217 Changed |= processLoop(Worklist.pop_back_val());
1219 // Process each loop nest in the function.
1223 bool processLoop(Loop *L) {
1224 assert(L->empty() && "Only process inner loops.");
1227 const std::string DebugLocStr = getDebugLocString(L);
1230 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1231 << L->getHeader()->getParent()->getName() << "\" from "
1232 << DebugLocStr << "\n");
1234 LoopVectorizeHints Hints(L, DisableUnrolling);
1236 DEBUG(dbgs() << "LV: Loop hints:"
1238 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1240 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1242 : "?")) << " width=" << Hints.getWidth()
1243 << " unroll=" << Hints.getUnroll() << "\n");
1245 // Function containing loop
1246 Function *F = L->getHeader()->getParent();
1248 // Looking at the diagnostic output is the only way to determine if a loop
1249 // was vectorized (other than looking at the IR or machine code), so it
1250 // is important to generate an optimization remark for each loop. Most of
1251 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1252 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1253 // less verbose reporting vectorized loops and unvectorized loops that may
1254 // benefit from vectorization, respectively.
1256 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1257 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1258 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1259 L->getStartLoc(), Hints.emitRemark());
1263 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1264 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1265 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1266 L->getStartLoc(), Hints.emitRemark());
1270 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1271 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1272 emitOptimizationRemarkAnalysis(
1273 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1274 "loop not vectorized: vector width and interleave count are "
1275 "explicitly set to 1");
1279 // Check the loop for a trip count threshold:
1280 // do not vectorize loops with a tiny trip count.
1281 BasicBlock *Latch = L->getLoopLatch();
1282 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1283 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1284 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1285 << "This loop is not worth vectorizing.");
1286 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1287 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1289 DEBUG(dbgs() << "\n");
1290 emitOptimizationRemarkAnalysis(
1291 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1292 "vectorization is not beneficial and is not explicitly forced");
1297 // Check if it is legal to vectorize the loop.
1298 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1299 if (!LVL.canVectorize()) {
1300 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1301 emitMissedWarning(F, L, Hints);
1305 // Use the cost model.
1306 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1308 // Check the function attributes to find out if this function should be
1309 // optimized for size.
1310 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1311 F->hasFnAttribute(Attribute::OptimizeForSize);
1313 // Compute the weighted frequency of this loop being executed and see if it
1314 // is less than 20% of the function entry baseline frequency. Note that we
1315 // always have a canonical loop here because we think we *can* vectoriez.
1316 // FIXME: This is hidden behind a flag due to pervasive problems with
1317 // exactly what block frequency models.
1318 if (LoopVectorizeWithBlockFrequency) {
1319 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1320 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1321 LoopEntryFreq < ColdEntryFreq)
1325 // Check the function attributes to see if implicit floats are allowed.a
1326 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1327 // an integer loop and the vector instructions selected are purely integer
1328 // vector instructions?
1329 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1330 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1331 "attribute is used.\n");
1332 emitOptimizationRemarkAnalysis(
1333 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1334 "loop not vectorized due to NoImplicitFloat attribute");
1335 emitMissedWarning(F, L, Hints);
1339 // Select the optimal vectorization factor.
1340 const LoopVectorizationCostModel::VectorizationFactor VF =
1341 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1343 LoopVectorizeHints::FK_Enabled);
1345 // Select the unroll factor.
1347 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1349 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1350 << DebugLocStr << '\n');
1351 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1353 if (VF.Width == 1) {
1354 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1357 emitOptimizationRemarkAnalysis(
1358 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1359 "not beneficial to vectorize and user disabled interleaving");
1362 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1364 // Report the unrolling decision.
1365 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1366 Twine("unrolled with interleaving factor " +
1368 " (vectorization not beneficial)"));
1370 // We decided not to vectorize, but we may want to unroll.
1372 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1373 Unroller.vectorize(&LVL);
1375 // If we decided that it is *legal* to vectorize the loop then do it.
1376 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1380 // Report the vectorization decision.
1381 emitOptimizationRemark(
1382 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1383 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1384 ", unrolling interleave factor: " + Twine(UF) + ")");
1387 // Mark the loop as already vectorized to avoid vectorizing again.
1388 Hints.setAlreadyVectorized(L);
1390 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1394 void getAnalysisUsage(AnalysisUsage &AU) const override {
1395 AU.addRequiredID(LoopSimplifyID);
1396 AU.addRequiredID(LCSSAID);
1397 AU.addRequired<BlockFrequencyInfo>();
1398 AU.addRequired<DominatorTreeWrapperPass>();
1399 AU.addRequired<LoopInfo>();
1400 AU.addRequired<ScalarEvolution>();
1401 AU.addRequired<TargetTransformInfo>();
1402 AU.addRequired<AliasAnalysis>();
1403 AU.addPreserved<LoopInfo>();
1404 AU.addPreserved<DominatorTreeWrapperPass>();
1405 AU.addPreserved<AliasAnalysis>();
1410 } // end anonymous namespace
1412 //===----------------------------------------------------------------------===//
1413 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1414 // LoopVectorizationCostModel.
1415 //===----------------------------------------------------------------------===//
1417 static Value *stripIntegerCast(Value *V) {
1418 if (CastInst *CI = dyn_cast<CastInst>(V))
1419 if (CI->getOperand(0)->getType()->isIntegerTy())
1420 return CI->getOperand(0);
1424 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1426 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1428 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1429 ValueToValueMap &PtrToStride,
1430 Value *Ptr, Value *OrigPtr = nullptr) {
1432 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1434 // If there is an entry in the map return the SCEV of the pointer with the
1435 // symbolic stride replaced by one.
1436 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1437 if (SI != PtrToStride.end()) {
1438 Value *StrideVal = SI->second;
1441 StrideVal = stripIntegerCast(StrideVal);
1443 // Replace symbolic stride by one.
1444 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1445 ValueToValueMap RewriteMap;
1446 RewriteMap[StrideVal] = One;
1449 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1450 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1455 // Otherwise, just return the SCEV of the original pointer.
1456 return SE->getSCEV(Ptr);
1459 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1460 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1461 unsigned ASId, ValueToValueMap &Strides) {
1462 // Get the stride replaced scev.
1463 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1464 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1465 assert(AR && "Invalid addrec expression");
1466 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1467 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1468 Pointers.push_back(Ptr);
1469 Starts.push_back(AR->getStart());
1470 Ends.push_back(ScEnd);
1471 IsWritePtr.push_back(WritePtr);
1472 DependencySetId.push_back(DepSetId);
1473 AliasSetId.push_back(ASId);
1476 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1477 // We need to place the broadcast of invariant variables outside the loop.
1478 Instruction *Instr = dyn_cast<Instruction>(V);
1480 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1481 Instr->getParent()) != LoopVectorBody.end());
1482 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1484 // Place the code for broadcasting invariant variables in the new preheader.
1485 IRBuilder<>::InsertPointGuard Guard(Builder);
1487 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1489 // Broadcast the scalar into all locations in the vector.
1490 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1495 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1497 assert(Val->getType()->isVectorTy() && "Must be a vector");
1498 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1499 "Elem must be an integer");
1500 // Create the types.
1501 Type *ITy = Val->getType()->getScalarType();
1502 VectorType *Ty = cast<VectorType>(Val->getType());
1503 int VLen = Ty->getNumElements();
1504 SmallVector<Constant*, 8> Indices;
1506 // Create a vector of consecutive numbers from zero to VF.
1507 for (int i = 0; i < VLen; ++i) {
1508 int64_t Idx = Negate ? (-i) : i;
1509 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1512 // Add the consecutive indices to the vector value.
1513 Constant *Cv = ConstantVector::get(Indices);
1514 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1515 return Builder.CreateAdd(Val, Cv, "induction");
1518 /// \brief Find the operand of the GEP that should be checked for consecutive
1519 /// stores. This ignores trailing indices that have no effect on the final
1521 static unsigned getGEPInductionOperand(const DataLayout *DL,
1522 const GetElementPtrInst *Gep) {
1523 unsigned LastOperand = Gep->getNumOperands() - 1;
1524 unsigned GEPAllocSize = DL->getTypeAllocSize(
1525 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1527 // Walk backwards and try to peel off zeros.
1528 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1529 // Find the type we're currently indexing into.
1530 gep_type_iterator GEPTI = gep_type_begin(Gep);
1531 std::advance(GEPTI, LastOperand - 1);
1533 // If it's a type with the same allocation size as the result of the GEP we
1534 // can peel off the zero index.
1535 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1543 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1544 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1545 // Make sure that the pointer does not point to structs.
1546 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1549 // If this value is a pointer induction variable we know it is consecutive.
1550 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1551 if (Phi && Inductions.count(Phi)) {
1552 InductionInfo II = Inductions[Phi];
1553 if (IK_PtrInduction == II.IK)
1555 else if (IK_ReversePtrInduction == II.IK)
1559 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1563 unsigned NumOperands = Gep->getNumOperands();
1564 Value *GpPtr = Gep->getPointerOperand();
1565 // If this GEP value is a consecutive pointer induction variable and all of
1566 // the indices are constant then we know it is consecutive. We can
1567 Phi = dyn_cast<PHINode>(GpPtr);
1568 if (Phi && Inductions.count(Phi)) {
1570 // Make sure that the pointer does not point to structs.
1571 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1572 if (GepPtrType->getElementType()->isAggregateType())
1575 // Make sure that all of the index operands are loop invariant.
1576 for (unsigned i = 1; i < NumOperands; ++i)
1577 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1580 InductionInfo II = Inductions[Phi];
1581 if (IK_PtrInduction == II.IK)
1583 else if (IK_ReversePtrInduction == II.IK)
1587 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1589 // Check that all of the gep indices are uniform except for our induction
1591 for (unsigned i = 0; i != NumOperands; ++i)
1592 if (i != InductionOperand &&
1593 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1596 // We can emit wide load/stores only if the last non-zero index is the
1597 // induction variable.
1598 const SCEV *Last = nullptr;
1599 if (!Strides.count(Gep))
1600 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1602 // Because of the multiplication by a stride we can have a s/zext cast.
1603 // We are going to replace this stride by 1 so the cast is safe to ignore.
1605 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1606 // %0 = trunc i64 %indvars.iv to i32
1607 // %mul = mul i32 %0, %Stride1
1608 // %idxprom = zext i32 %mul to i64 << Safe cast.
1609 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1611 Last = replaceSymbolicStrideSCEV(SE, Strides,
1612 Gep->getOperand(InductionOperand), Gep);
1613 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1615 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1619 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1620 const SCEV *Step = AR->getStepRecurrence(*SE);
1622 // The memory is consecutive because the last index is consecutive
1623 // and all other indices are loop invariant.
1626 if (Step->isAllOnesValue())
1633 bool LoopVectorizationLegality::isUniform(Value *V) {
1634 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1637 InnerLoopVectorizer::VectorParts&
1638 InnerLoopVectorizer::getVectorValue(Value *V) {
1639 assert(V != Induction && "The new induction variable should not be used.");
1640 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1642 // If we have a stride that is replaced by one, do it here.
1643 if (Legal->hasStride(V))
1644 V = ConstantInt::get(V->getType(), 1);
1646 // If we have this scalar in the map, return it.
1647 if (WidenMap.has(V))
1648 return WidenMap.get(V);
1650 // If this scalar is unknown, assume that it is a constant or that it is
1651 // loop invariant. Broadcast V and save the value for future uses.
1652 Value *B = getBroadcastInstrs(V);
1653 return WidenMap.splat(V, B);
1656 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1657 assert(Vec->getType()->isVectorTy() && "Invalid type");
1658 SmallVector<Constant*, 8> ShuffleMask;
1659 for (unsigned i = 0; i < VF; ++i)
1660 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1662 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1663 ConstantVector::get(ShuffleMask),
1667 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1668 // Attempt to issue a wide load.
1669 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1670 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1672 assert((LI || SI) && "Invalid Load/Store instruction");
1674 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1675 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1676 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1677 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1678 // An alignment of 0 means target abi alignment. We need to use the scalar's
1679 // target abi alignment in such a case.
1681 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1682 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1683 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1684 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1686 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1687 return scalarizeInstruction(Instr, true);
1689 if (ScalarAllocatedSize != VectorElementSize)
1690 return scalarizeInstruction(Instr);
1692 // If the pointer is loop invariant or if it is non-consecutive,
1693 // scalarize the load.
1694 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1695 bool Reverse = ConsecutiveStride < 0;
1696 bool UniformLoad = LI && Legal->isUniform(Ptr);
1697 if (!ConsecutiveStride || UniformLoad)
1698 return scalarizeInstruction(Instr);
1700 Constant *Zero = Builder.getInt32(0);
1701 VectorParts &Entry = WidenMap.get(Instr);
1703 // Handle consecutive loads/stores.
1704 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1705 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1706 setDebugLocFromInst(Builder, Gep);
1707 Value *PtrOperand = Gep->getPointerOperand();
1708 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1709 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1711 // Create the new GEP with the new induction variable.
1712 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1713 Gep2->setOperand(0, FirstBasePtr);
1714 Gep2->setName("gep.indvar.base");
1715 Ptr = Builder.Insert(Gep2);
1717 setDebugLocFromInst(Builder, Gep);
1718 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1719 OrigLoop) && "Base ptr must be invariant");
1721 // The last index does not have to be the induction. It can be
1722 // consecutive and be a function of the index. For example A[I+1];
1723 unsigned NumOperands = Gep->getNumOperands();
1724 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1725 // Create the new GEP with the new induction variable.
1726 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1728 for (unsigned i = 0; i < NumOperands; ++i) {
1729 Value *GepOperand = Gep->getOperand(i);
1730 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1732 // Update last index or loop invariant instruction anchored in loop.
1733 if (i == InductionOperand ||
1734 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1735 assert((i == InductionOperand ||
1736 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1737 "Must be last index or loop invariant");
1739 VectorParts &GEPParts = getVectorValue(GepOperand);
1740 Value *Index = GEPParts[0];
1741 Index = Builder.CreateExtractElement(Index, Zero);
1742 Gep2->setOperand(i, Index);
1743 Gep2->setName("gep.indvar.idx");
1746 Ptr = Builder.Insert(Gep2);
1748 // Use the induction element ptr.
1749 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1750 setDebugLocFromInst(Builder, Ptr);
1751 VectorParts &PtrVal = getVectorValue(Ptr);
1752 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1757 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1758 "We do not allow storing to uniform addresses");
1759 setDebugLocFromInst(Builder, SI);
1760 // We don't want to update the value in the map as it might be used in
1761 // another expression. So don't use a reference type for "StoredVal".
1762 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1764 for (unsigned Part = 0; Part < UF; ++Part) {
1765 // Calculate the pointer for the specific unroll-part.
1766 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1769 // If we store to reverse consecutive memory locations then we need
1770 // to reverse the order of elements in the stored value.
1771 StoredVal[Part] = reverseVector(StoredVal[Part]);
1772 // If the address is consecutive but reversed, then the
1773 // wide store needs to start at the last vector element.
1774 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1775 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1778 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1779 DataTy->getPointerTo(AddressSpace));
1781 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1782 propagateMetadata(NewSI, SI);
1788 assert(LI && "Must have a load instruction");
1789 setDebugLocFromInst(Builder, LI);
1790 for (unsigned Part = 0; Part < UF; ++Part) {
1791 // Calculate the pointer for the specific unroll-part.
1792 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1795 // If the address is consecutive but reversed, then the
1796 // wide store needs to start at the last vector element.
1797 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1798 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1801 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1802 DataTy->getPointerTo(AddressSpace));
1803 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1804 propagateMetadata(NewLI, LI);
1805 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1809 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1810 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1811 // Holds vector parameters or scalars, in case of uniform vals.
1812 SmallVector<VectorParts, 4> Params;
1814 setDebugLocFromInst(Builder, Instr);
1816 // Find all of the vectorized parameters.
1817 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1818 Value *SrcOp = Instr->getOperand(op);
1820 // If we are accessing the old induction variable, use the new one.
1821 if (SrcOp == OldInduction) {
1822 Params.push_back(getVectorValue(SrcOp));
1826 // Try using previously calculated values.
1827 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1829 // If the src is an instruction that appeared earlier in the basic block
1830 // then it should already be vectorized.
1831 if (SrcInst && OrigLoop->contains(SrcInst)) {
1832 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1833 // The parameter is a vector value from earlier.
1834 Params.push_back(WidenMap.get(SrcInst));
1836 // The parameter is a scalar from outside the loop. Maybe even a constant.
1837 VectorParts Scalars;
1838 Scalars.append(UF, SrcOp);
1839 Params.push_back(Scalars);
1843 assert(Params.size() == Instr->getNumOperands() &&
1844 "Invalid number of operands");
1846 // Does this instruction return a value ?
1847 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1849 Value *UndefVec = IsVoidRetTy ? nullptr :
1850 UndefValue::get(VectorType::get(Instr->getType(), VF));
1851 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1852 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1854 Instruction *InsertPt = Builder.GetInsertPoint();
1855 BasicBlock *IfBlock = Builder.GetInsertBlock();
1856 BasicBlock *CondBlock = nullptr;
1859 Loop *VectorLp = nullptr;
1860 if (IfPredicateStore) {
1861 assert(Instr->getParent()->getSinglePredecessor() &&
1862 "Only support single predecessor blocks");
1863 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1864 Instr->getParent());
1865 VectorLp = LI->getLoopFor(IfBlock);
1866 assert(VectorLp && "Must have a loop for this block");
1869 // For each vector unroll 'part':
1870 for (unsigned Part = 0; Part < UF; ++Part) {
1871 // For each scalar that we create:
1872 for (unsigned Width = 0; Width < VF; ++Width) {
1875 Value *Cmp = nullptr;
1876 if (IfPredicateStore) {
1877 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1878 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1879 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1880 LoopVectorBody.push_back(CondBlock);
1881 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1882 // Update Builder with newly created basic block.
1883 Builder.SetInsertPoint(InsertPt);
1886 Instruction *Cloned = Instr->clone();
1888 Cloned->setName(Instr->getName() + ".cloned");
1889 // Replace the operands of the cloned instructions with extracted scalars.
1890 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1891 Value *Op = Params[op][Part];
1892 // Param is a vector. Need to extract the right lane.
1893 if (Op->getType()->isVectorTy())
1894 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1895 Cloned->setOperand(op, Op);
1898 // Place the cloned scalar in the new loop.
1899 Builder.Insert(Cloned);
1901 // If the original scalar returns a value we need to place it in a vector
1902 // so that future users will be able to use it.
1904 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1905 Builder.getInt32(Width));
1907 if (IfPredicateStore) {
1908 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1909 LoopVectorBody.push_back(NewIfBlock);
1910 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1911 Builder.SetInsertPoint(InsertPt);
1912 Instruction *OldBr = IfBlock->getTerminator();
1913 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1914 OldBr->eraseFromParent();
1915 IfBlock = NewIfBlock;
1921 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1925 if (Instruction *I = dyn_cast<Instruction>(V))
1926 return I->getParent() == Loc->getParent() ? I : nullptr;
1930 std::pair<Instruction *, Instruction *>
1931 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1932 Instruction *tnullptr = nullptr;
1933 if (!Legal->mustCheckStrides())
1934 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1936 IRBuilder<> ChkBuilder(Loc);
1939 Value *Check = nullptr;
1940 Instruction *FirstInst = nullptr;
1941 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1942 SE = Legal->strides_end();
1944 Value *Ptr = stripIntegerCast(*SI);
1945 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1947 // Store the first instruction we create.
1948 FirstInst = getFirstInst(FirstInst, C, Loc);
1950 Check = ChkBuilder.CreateOr(Check, C);
1955 // We have to do this trickery because the IRBuilder might fold the check to a
1956 // constant expression in which case there is no Instruction anchored in a
1958 LLVMContext &Ctx = Loc->getContext();
1959 Instruction *TheCheck =
1960 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1961 ChkBuilder.Insert(TheCheck, "stride.not.one");
1962 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1964 return std::make_pair(FirstInst, TheCheck);
1967 std::pair<Instruction *, Instruction *>
1968 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1969 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1970 Legal->getRuntimePointerCheck();
1972 Instruction *tnullptr = nullptr;
1973 if (!PtrRtCheck->Need)
1974 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1976 unsigned NumPointers = PtrRtCheck->Pointers.size();
1977 SmallVector<TrackingVH<Value> , 2> Starts;
1978 SmallVector<TrackingVH<Value> , 2> Ends;
1980 LLVMContext &Ctx = Loc->getContext();
1981 SCEVExpander Exp(*SE, "induction");
1982 Instruction *FirstInst = nullptr;
1984 for (unsigned i = 0; i < NumPointers; ++i) {
1985 Value *Ptr = PtrRtCheck->Pointers[i];
1986 const SCEV *Sc = SE->getSCEV(Ptr);
1988 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1989 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1991 Starts.push_back(Ptr);
1992 Ends.push_back(Ptr);
1994 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1995 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1997 // Use this type for pointer arithmetic.
1998 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2000 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2001 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2002 Starts.push_back(Start);
2003 Ends.push_back(End);
2007 IRBuilder<> ChkBuilder(Loc);
2008 // Our instructions might fold to a constant.
2009 Value *MemoryRuntimeCheck = nullptr;
2010 for (unsigned i = 0; i < NumPointers; ++i) {
2011 for (unsigned j = i+1; j < NumPointers; ++j) {
2012 // No need to check if two readonly pointers intersect.
2013 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2016 // Only need to check pointers between two different dependency sets.
2017 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2019 // Only need to check pointers in the same alias set.
2020 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2023 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2024 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2026 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2027 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2028 "Trying to bounds check pointers with different address spaces");
2030 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2031 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2033 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2034 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2035 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2036 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2038 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2039 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2040 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2041 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2042 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2043 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2044 if (MemoryRuntimeCheck) {
2045 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2047 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2049 MemoryRuntimeCheck = IsConflict;
2053 // We have to do this trickery because the IRBuilder might fold the check to a
2054 // constant expression in which case there is no Instruction anchored in a
2056 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2057 ConstantInt::getTrue(Ctx));
2058 ChkBuilder.Insert(Check, "memcheck.conflict");
2059 FirstInst = getFirstInst(FirstInst, Check, Loc);
2060 return std::make_pair(FirstInst, Check);
2063 void InnerLoopVectorizer::createEmptyLoop() {
2065 In this function we generate a new loop. The new loop will contain
2066 the vectorized instructions while the old loop will continue to run the
2069 [ ] <-- Back-edge taken count overflow check.
2072 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2075 || [ ] <-- vector pre header.
2079 || [ ]_| <-- vector loop.
2082 | >[ ] <--- middle-block.
2085 -|- >[ ] <--- new preheader.
2089 | [ ]_| <-- old scalar loop to handle remainder.
2092 >[ ] <-- exit block.
2096 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2097 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2098 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2099 assert(BypassBlock && "Invalid loop structure");
2100 assert(ExitBlock && "Must have an exit block");
2102 // Some loops have a single integer induction variable, while other loops
2103 // don't. One example is c++ iterators that often have multiple pointer
2104 // induction variables. In the code below we also support a case where we
2105 // don't have a single induction variable.
2106 OldInduction = Legal->getInduction();
2107 Type *IdxTy = Legal->getWidestInductionType();
2109 // Find the loop boundaries.
2110 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2111 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2113 // The exit count might have the type of i64 while the phi is i32. This can
2114 // happen if we have an induction variable that is sign extended before the
2115 // compare. The only way that we get a backedge taken count is that the
2116 // induction variable was signed and as such will not overflow. In such a case
2117 // truncation is legal.
2118 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2119 IdxTy->getPrimitiveSizeInBits())
2120 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2122 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2123 // Get the total trip count from the count by adding 1.
2124 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2125 SE->getConstant(BackedgeTakeCount->getType(), 1));
2127 // Expand the trip count and place the new instructions in the preheader.
2128 // Notice that the pre-header does not change, only the loop body.
2129 SCEVExpander Exp(*SE, "induction");
2131 // We need to test whether the backedge-taken count is uint##_max. Adding one
2132 // to it will cause overflow and an incorrect loop trip count in the vector
2133 // body. In case of overflow we want to directly jump to the scalar remainder
2135 Value *BackedgeCount =
2136 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2137 BypassBlock->getTerminator());
2138 if (BackedgeCount->getType()->isPointerTy())
2139 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2140 "backedge.ptrcnt.to.int",
2141 BypassBlock->getTerminator());
2142 Instruction *CheckBCOverflow =
2143 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2144 Constant::getAllOnesValue(BackedgeCount->getType()),
2145 "backedge.overflow", BypassBlock->getTerminator());
2147 // The loop index does not have to start at Zero. Find the original start
2148 // value from the induction PHI node. If we don't have an induction variable
2149 // then we know that it starts at zero.
2150 Builder.SetInsertPoint(BypassBlock->getTerminator());
2151 Value *StartIdx = ExtendedIdx = OldInduction ?
2152 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2154 ConstantInt::get(IdxTy, 0);
2156 // We need an instruction to anchor the overflow check on. StartIdx needs to
2157 // be defined before the overflow check branch. Because the scalar preheader
2158 // is going to merge the start index and so the overflow branch block needs to
2159 // contain a definition of the start index.
2160 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2161 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2162 BypassBlock->getTerminator());
2164 // Count holds the overall loop count (N).
2165 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2166 BypassBlock->getTerminator());
2168 LoopBypassBlocks.push_back(BypassBlock);
2170 // Split the single block loop into the two loop structure described above.
2171 BasicBlock *VectorPH =
2172 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2173 BasicBlock *VecBody =
2174 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2175 BasicBlock *MiddleBlock =
2176 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2177 BasicBlock *ScalarPH =
2178 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2180 // Create and register the new vector loop.
2181 Loop* Lp = new Loop();
2182 Loop *ParentLoop = OrigLoop->getParentLoop();
2184 // Insert the new loop into the loop nest and register the new basic blocks
2185 // before calling any utilities such as SCEV that require valid LoopInfo.
2187 ParentLoop->addChildLoop(Lp);
2188 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2189 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2190 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2192 LI->addTopLevelLoop(Lp);
2194 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2196 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2198 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2200 // Generate the induction variable.
2201 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2202 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2203 // The loop step is equal to the vectorization factor (num of SIMD elements)
2204 // times the unroll factor (num of SIMD instructions).
2205 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2207 // This is the IR builder that we use to add all of the logic for bypassing
2208 // the new vector loop.
2209 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2210 setDebugLocFromInst(BypassBuilder,
2211 getDebugLocFromInstOrOperands(OldInduction));
2213 // We may need to extend the index in case there is a type mismatch.
2214 // We know that the count starts at zero and does not overflow.
2215 if (Count->getType() != IdxTy) {
2216 // The exit count can be of pointer type. Convert it to the correct
2218 if (ExitCount->getType()->isPointerTy())
2219 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2221 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2224 // Add the start index to the loop count to get the new end index.
2225 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2227 // Now we need to generate the expression for N - (N % VF), which is
2228 // the part that the vectorized body will execute.
2229 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2230 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2231 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2232 "end.idx.rnd.down");
2234 // Now, compare the new count to zero. If it is zero skip the vector loop and
2235 // jump to the scalar loop.
2237 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2239 BasicBlock *LastBypassBlock = BypassBlock;
2241 // Generate code to check that the loops trip count that we computed by adding
2242 // one to the backedge-taken count will not overflow.
2244 auto PastOverflowCheck =
2245 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2246 BasicBlock *CheckBlock =
2247 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2249 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2250 LoopBypassBlocks.push_back(CheckBlock);
2251 Instruction *OldTerm = LastBypassBlock->getTerminator();
2252 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2253 OldTerm->eraseFromParent();
2254 LastBypassBlock = CheckBlock;
2257 // Generate the code to check that the strides we assumed to be one are really
2258 // one. We want the new basic block to start at the first instruction in a
2259 // sequence of instructions that form a check.
2260 Instruction *StrideCheck;
2261 Instruction *FirstCheckInst;
2262 std::tie(FirstCheckInst, StrideCheck) =
2263 addStrideCheck(LastBypassBlock->getTerminator());
2265 // Create a new block containing the stride check.
2266 BasicBlock *CheckBlock =
2267 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2269 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2270 LoopBypassBlocks.push_back(CheckBlock);
2272 // Replace the branch into the memory check block with a conditional branch
2273 // for the "few elements case".
2274 Instruction *OldTerm = LastBypassBlock->getTerminator();
2275 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2276 OldTerm->eraseFromParent();
2279 LastBypassBlock = CheckBlock;
2282 // Generate the code that checks in runtime if arrays overlap. We put the
2283 // checks into a separate block to make the more common case of few elements
2285 Instruction *MemRuntimeCheck;
2286 std::tie(FirstCheckInst, MemRuntimeCheck) =
2287 addRuntimeCheck(LastBypassBlock->getTerminator());
2288 if (MemRuntimeCheck) {
2289 // Create a new block containing the memory check.
2290 BasicBlock *CheckBlock =
2291 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2293 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2294 LoopBypassBlocks.push_back(CheckBlock);
2296 // Replace the branch into the memory check block with a conditional branch
2297 // for the "few elements case".
2298 Instruction *OldTerm = LastBypassBlock->getTerminator();
2299 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2300 OldTerm->eraseFromParent();
2302 Cmp = MemRuntimeCheck;
2303 LastBypassBlock = CheckBlock;
2306 LastBypassBlock->getTerminator()->eraseFromParent();
2307 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2310 // We are going to resume the execution of the scalar loop.
2311 // Go over all of the induction variables that we found and fix the
2312 // PHIs that are left in the scalar version of the loop.
2313 // The starting values of PHI nodes depend on the counter of the last
2314 // iteration in the vectorized loop.
2315 // If we come from a bypass edge then we need to start from the original
2318 // This variable saves the new starting index for the scalar loop.
2319 PHINode *ResumeIndex = nullptr;
2320 LoopVectorizationLegality::InductionList::iterator I, E;
2321 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2322 // Set builder to point to last bypass block.
2323 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2324 for (I = List->begin(), E = List->end(); I != E; ++I) {
2325 PHINode *OrigPhi = I->first;
2326 LoopVectorizationLegality::InductionInfo II = I->second;
2328 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2329 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2330 MiddleBlock->getTerminator());
2331 // We might have extended the type of the induction variable but we need a
2332 // truncated version for the scalar loop.
2333 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2334 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2335 MiddleBlock->getTerminator()) : nullptr;
2337 // Create phi nodes to merge from the backedge-taken check block.
2338 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2339 ScalarPH->getTerminator());
2340 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2342 PHINode *BCTruncResumeVal = nullptr;
2343 if (OrigPhi == OldInduction) {
2345 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2346 ScalarPH->getTerminator());
2347 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2350 Value *EndValue = nullptr;
2352 case LoopVectorizationLegality::IK_NoInduction:
2353 llvm_unreachable("Unknown induction");
2354 case LoopVectorizationLegality::IK_IntInduction: {
2355 // Handle the integer induction counter.
2356 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2358 // We have the canonical induction variable.
2359 if (OrigPhi == OldInduction) {
2360 // Create a truncated version of the resume value for the scalar loop,
2361 // we might have promoted the type to a larger width.
2363 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2364 // The new PHI merges the original incoming value, in case of a bypass,
2365 // or the value at the end of the vectorized loop.
2366 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2367 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2368 TruncResumeVal->addIncoming(EndValue, VecBody);
2370 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2372 // We know what the end value is.
2373 EndValue = IdxEndRoundDown;
2374 // We also know which PHI node holds it.
2375 ResumeIndex = ResumeVal;
2379 // Not the canonical induction variable - add the vector loop count to the
2381 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2382 II.StartValue->getType(),
2384 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2387 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2388 // Convert the CountRoundDown variable to the PHI size.
2389 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2390 II.StartValue->getType(),
2392 // Handle reverse integer induction counter.
2393 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2396 case LoopVectorizationLegality::IK_PtrInduction: {
2397 // For pointer induction variables, calculate the offset using
2399 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2403 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2404 // The value at the end of the loop for the reverse pointer is calculated
2405 // by creating a GEP with a negative index starting from the start value.
2406 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2407 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2409 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2415 // The new PHI merges the original incoming value, in case of a bypass,
2416 // or the value at the end of the vectorized loop.
2417 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2418 if (OrigPhi == OldInduction)
2419 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2421 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2423 ResumeVal->addIncoming(EndValue, VecBody);
2425 // Fix the scalar body counter (PHI node).
2426 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2428 // The old induction's phi node in the scalar body needs the truncated
2430 if (OrigPhi == OldInduction) {
2431 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2432 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2434 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2435 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2439 // If we are generating a new induction variable then we also need to
2440 // generate the code that calculates the exit value. This value is not
2441 // simply the end of the counter because we may skip the vectorized body
2442 // in case of a runtime check.
2444 assert(!ResumeIndex && "Unexpected resume value found");
2445 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2446 MiddleBlock->getTerminator());
2447 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2448 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2449 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2452 // Make sure that we found the index where scalar loop needs to continue.
2453 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2454 "Invalid resume Index");
2456 // Add a check in the middle block to see if we have completed
2457 // all of the iterations in the first vector loop.
2458 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2459 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2460 ResumeIndex, "cmp.n",
2461 MiddleBlock->getTerminator());
2463 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2464 // Remove the old terminator.
2465 MiddleBlock->getTerminator()->eraseFromParent();
2467 // Create i+1 and fill the PHINode.
2468 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2469 Induction->addIncoming(StartIdx, VectorPH);
2470 Induction->addIncoming(NextIdx, VecBody);
2471 // Create the compare.
2472 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2473 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2475 // Now we have two terminators. Remove the old one from the block.
2476 VecBody->getTerminator()->eraseFromParent();
2478 // Get ready to start creating new instructions into the vectorized body.
2479 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2482 LoopVectorPreHeader = VectorPH;
2483 LoopScalarPreHeader = ScalarPH;
2484 LoopMiddleBlock = MiddleBlock;
2485 LoopExitBlock = ExitBlock;
2486 LoopVectorBody.push_back(VecBody);
2487 LoopScalarBody = OldBasicBlock;
2489 LoopVectorizeHints Hints(Lp, true);
2490 Hints.setAlreadyVectorized(Lp);
2493 /// This function returns the identity element (or neutral element) for
2494 /// the operation K.
2496 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2501 // Adding, Xoring, Oring zero to a number does not change it.
2502 return ConstantInt::get(Tp, 0);
2503 case RK_IntegerMult:
2504 // Multiplying a number by 1 does not change it.
2505 return ConstantInt::get(Tp, 1);
2507 // AND-ing a number with an all-1 value does not change it.
2508 return ConstantInt::get(Tp, -1, true);
2510 // Multiplying a number by 1 does not change it.
2511 return ConstantFP::get(Tp, 1.0L);
2513 // Adding zero to a number does not change it.
2514 return ConstantFP::get(Tp, 0.0L);
2516 llvm_unreachable("Unknown reduction kind");
2520 /// This function translates the reduction kind to an LLVM binary operator.
2522 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2524 case LoopVectorizationLegality::RK_IntegerAdd:
2525 return Instruction::Add;
2526 case LoopVectorizationLegality::RK_IntegerMult:
2527 return Instruction::Mul;
2528 case LoopVectorizationLegality::RK_IntegerOr:
2529 return Instruction::Or;
2530 case LoopVectorizationLegality::RK_IntegerAnd:
2531 return Instruction::And;
2532 case LoopVectorizationLegality::RK_IntegerXor:
2533 return Instruction::Xor;
2534 case LoopVectorizationLegality::RK_FloatMult:
2535 return Instruction::FMul;
2536 case LoopVectorizationLegality::RK_FloatAdd:
2537 return Instruction::FAdd;
2538 case LoopVectorizationLegality::RK_IntegerMinMax:
2539 return Instruction::ICmp;
2540 case LoopVectorizationLegality::RK_FloatMinMax:
2541 return Instruction::FCmp;
2543 llvm_unreachable("Unknown reduction operation");
2547 Value *createMinMaxOp(IRBuilder<> &Builder,
2548 LoopVectorizationLegality::MinMaxReductionKind RK,
2551 CmpInst::Predicate P = CmpInst::ICMP_NE;
2554 llvm_unreachable("Unknown min/max reduction kind");
2555 case LoopVectorizationLegality::MRK_UIntMin:
2556 P = CmpInst::ICMP_ULT;
2558 case LoopVectorizationLegality::MRK_UIntMax:
2559 P = CmpInst::ICMP_UGT;
2561 case LoopVectorizationLegality::MRK_SIntMin:
2562 P = CmpInst::ICMP_SLT;
2564 case LoopVectorizationLegality::MRK_SIntMax:
2565 P = CmpInst::ICMP_SGT;
2567 case LoopVectorizationLegality::MRK_FloatMin:
2568 P = CmpInst::FCMP_OLT;
2570 case LoopVectorizationLegality::MRK_FloatMax:
2571 P = CmpInst::FCMP_OGT;
2576 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2577 RK == LoopVectorizationLegality::MRK_FloatMax)
2578 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2580 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2582 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2587 struct CSEDenseMapInfo {
2588 static bool canHandle(Instruction *I) {
2589 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2590 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2592 static inline Instruction *getEmptyKey() {
2593 return DenseMapInfo<Instruction *>::getEmptyKey();
2595 static inline Instruction *getTombstoneKey() {
2596 return DenseMapInfo<Instruction *>::getTombstoneKey();
2598 static unsigned getHashValue(Instruction *I) {
2599 assert(canHandle(I) && "Unknown instruction!");
2600 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2601 I->value_op_end()));
2603 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2604 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2605 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2607 return LHS->isIdenticalTo(RHS);
2612 /// \brief Check whether this block is a predicated block.
2613 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2614 /// = ...; " blocks. We start with one vectorized basic block. For every
2615 /// conditional block we split this vectorized block. Therefore, every second
2616 /// block will be a predicated one.
2617 static bool isPredicatedBlock(unsigned BlockNum) {
2618 return BlockNum % 2;
2621 ///\brief Perform cse of induction variable instructions.
2622 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2623 // Perform simple cse.
2624 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2625 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2626 BasicBlock *BB = BBs[i];
2627 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2628 Instruction *In = I++;
2630 if (!CSEDenseMapInfo::canHandle(In))
2633 // Check if we can replace this instruction with any of the
2634 // visited instructions.
2635 if (Instruction *V = CSEMap.lookup(In)) {
2636 In->replaceAllUsesWith(V);
2637 In->eraseFromParent();
2640 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2641 // ...;" blocks for predicated stores. Every second block is a predicated
2643 if (isPredicatedBlock(i))
2651 /// \brief Adds a 'fast' flag to floating point operations.
2652 static Value *addFastMathFlag(Value *V) {
2653 if (isa<FPMathOperator>(V)){
2654 FastMathFlags Flags;
2655 Flags.setUnsafeAlgebra();
2656 cast<Instruction>(V)->setFastMathFlags(Flags);
2661 void InnerLoopVectorizer::vectorizeLoop() {
2662 //===------------------------------------------------===//
2664 // Notice: any optimization or new instruction that go
2665 // into the code below should be also be implemented in
2668 //===------------------------------------------------===//
2669 Constant *Zero = Builder.getInt32(0);
2671 // In order to support reduction variables we need to be able to vectorize
2672 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2673 // stages. First, we create a new vector PHI node with no incoming edges.
2674 // We use this value when we vectorize all of the instructions that use the
2675 // PHI. Next, after all of the instructions in the block are complete we
2676 // add the new incoming edges to the PHI. At this point all of the
2677 // instructions in the basic block are vectorized, so we can use them to
2678 // construct the PHI.
2679 PhiVector RdxPHIsToFix;
2681 // Scan the loop in a topological order to ensure that defs are vectorized
2683 LoopBlocksDFS DFS(OrigLoop);
2686 // Vectorize all of the blocks in the original loop.
2687 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2688 be = DFS.endRPO(); bb != be; ++bb)
2689 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2691 // At this point every instruction in the original loop is widened to
2692 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2693 // that we vectorized. The PHI nodes are currently empty because we did
2694 // not want to introduce cycles. Notice that the remaining PHI nodes
2695 // that we need to fix are reduction variables.
2697 // Create the 'reduced' values for each of the induction vars.
2698 // The reduced values are the vector values that we scalarize and combine
2699 // after the loop is finished.
2700 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2702 PHINode *RdxPhi = *it;
2703 assert(RdxPhi && "Unable to recover vectorized PHI");
2705 // Find the reduction variable descriptor.
2706 assert(Legal->getReductionVars()->count(RdxPhi) &&
2707 "Unable to find the reduction variable");
2708 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2709 (*Legal->getReductionVars())[RdxPhi];
2711 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2713 // We need to generate a reduction vector from the incoming scalar.
2714 // To do so, we need to generate the 'identity' vector and override
2715 // one of the elements with the incoming scalar reduction. We need
2716 // to do it in the vector-loop preheader.
2717 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2719 // This is the vector-clone of the value that leaves the loop.
2720 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2721 Type *VecTy = VectorExit[0]->getType();
2723 // Find the reduction identity variable. Zero for addition, or, xor,
2724 // one for multiplication, -1 for And.
2727 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2728 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2729 // MinMax reduction have the start value as their identify.
2731 VectorStart = Identity = RdxDesc.StartValue;
2733 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2738 // Handle other reduction kinds:
2740 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2741 VecTy->getScalarType());
2744 // This vector is the Identity vector where the first element is the
2745 // incoming scalar reduction.
2746 VectorStart = RdxDesc.StartValue;
2748 Identity = ConstantVector::getSplat(VF, Iden);
2750 // This vector is the Identity vector where the first element is the
2751 // incoming scalar reduction.
2752 VectorStart = Builder.CreateInsertElement(Identity,
2753 RdxDesc.StartValue, Zero);
2757 // Fix the vector-loop phi.
2758 // We created the induction variable so we know that the
2759 // preheader is the first entry.
2760 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2762 // Reductions do not have to start at zero. They can start with
2763 // any loop invariant values.
2764 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2765 BasicBlock *Latch = OrigLoop->getLoopLatch();
2766 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2767 VectorParts &Val = getVectorValue(LoopVal);
2768 for (unsigned part = 0; part < UF; ++part) {
2769 // Make sure to add the reduction stat value only to the
2770 // first unroll part.
2771 Value *StartVal = (part == 0) ? VectorStart : Identity;
2772 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2773 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2774 LoopVectorBody.back());
2777 // Before each round, move the insertion point right between
2778 // the PHIs and the values we are going to write.
2779 // This allows us to write both PHINodes and the extractelement
2781 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2783 VectorParts RdxParts;
2784 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2785 for (unsigned part = 0; part < UF; ++part) {
2786 // This PHINode contains the vectorized reduction variable, or
2787 // the initial value vector, if we bypass the vector loop.
2788 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2789 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2790 Value *StartVal = (part == 0) ? VectorStart : Identity;
2791 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2792 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2793 NewPhi->addIncoming(RdxExitVal[part],
2794 LoopVectorBody.back());
2795 RdxParts.push_back(NewPhi);
2798 // Reduce all of the unrolled parts into a single vector.
2799 Value *ReducedPartRdx = RdxParts[0];
2800 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2801 setDebugLocFromInst(Builder, ReducedPartRdx);
2802 for (unsigned part = 1; part < UF; ++part) {
2803 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2804 // Floating point operations had to be 'fast' to enable the reduction.
2805 ReducedPartRdx = addFastMathFlag(
2806 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2807 ReducedPartRdx, "bin.rdx"));
2809 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2810 ReducedPartRdx, RdxParts[part]);
2814 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2815 // and vector ops, reducing the set of values being computed by half each
2817 assert(isPowerOf2_32(VF) &&
2818 "Reduction emission only supported for pow2 vectors!");
2819 Value *TmpVec = ReducedPartRdx;
2820 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2821 for (unsigned i = VF; i != 1; i >>= 1) {
2822 // Move the upper half of the vector to the lower half.
2823 for (unsigned j = 0; j != i/2; ++j)
2824 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2826 // Fill the rest of the mask with undef.
2827 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2828 UndefValue::get(Builder.getInt32Ty()));
2831 Builder.CreateShuffleVector(TmpVec,
2832 UndefValue::get(TmpVec->getType()),
2833 ConstantVector::get(ShuffleMask),
2836 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2837 // Floating point operations had to be 'fast' to enable the reduction.
2838 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2839 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2841 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2844 // The result is in the first element of the vector.
2845 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2846 Builder.getInt32(0));
2849 // Create a phi node that merges control-flow from the backedge-taken check
2850 // block and the middle block.
2851 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2852 LoopScalarPreHeader->getTerminator());
2853 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2854 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2856 // Now, we need to fix the users of the reduction variable
2857 // inside and outside of the scalar remainder loop.
2858 // We know that the loop is in LCSSA form. We need to update the
2859 // PHI nodes in the exit blocks.
2860 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2861 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2862 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2863 if (!LCSSAPhi) break;
2865 // All PHINodes need to have a single entry edge, or two if
2866 // we already fixed them.
2867 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2869 // We found our reduction value exit-PHI. Update it with the
2870 // incoming bypass edge.
2871 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2872 // Add an edge coming from the bypass.
2873 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2876 }// end of the LCSSA phi scan.
2878 // Fix the scalar loop reduction variable with the incoming reduction sum
2879 // from the vector body and from the backedge value.
2880 int IncomingEdgeBlockIdx =
2881 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2882 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2883 // Pick the other block.
2884 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2885 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2886 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2887 }// end of for each redux variable.
2891 // Remove redundant induction instructions.
2892 cse(LoopVectorBody);
2895 void InnerLoopVectorizer::fixLCSSAPHIs() {
2896 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2897 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2898 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2899 if (!LCSSAPhi) break;
2900 if (LCSSAPhi->getNumIncomingValues() == 1)
2901 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2906 InnerLoopVectorizer::VectorParts
2907 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2908 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2911 // Look for cached value.
2912 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2913 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2914 if (ECEntryIt != MaskCache.end())
2915 return ECEntryIt->second;
2917 VectorParts SrcMask = createBlockInMask(Src);
2919 // The terminator has to be a branch inst!
2920 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2921 assert(BI && "Unexpected terminator found");
2923 if (BI->isConditional()) {
2924 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2926 if (BI->getSuccessor(0) != Dst)
2927 for (unsigned part = 0; part < UF; ++part)
2928 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2930 for (unsigned part = 0; part < UF; ++part)
2931 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2933 MaskCache[Edge] = EdgeMask;
2937 MaskCache[Edge] = SrcMask;
2941 InnerLoopVectorizer::VectorParts
2942 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2943 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2945 // Loop incoming mask is all-one.
2946 if (OrigLoop->getHeader() == BB) {
2947 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2948 return getVectorValue(C);
2951 // This is the block mask. We OR all incoming edges, and with zero.
2952 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2953 VectorParts BlockMask = getVectorValue(Zero);
2956 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2957 VectorParts EM = createEdgeMask(*it, BB);
2958 for (unsigned part = 0; part < UF; ++part)
2959 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2965 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2966 InnerLoopVectorizer::VectorParts &Entry,
2967 unsigned UF, unsigned VF, PhiVector *PV) {
2968 PHINode* P = cast<PHINode>(PN);
2969 // Handle reduction variables:
2970 if (Legal->getReductionVars()->count(P)) {
2971 for (unsigned part = 0; part < UF; ++part) {
2972 // This is phase one of vectorizing PHIs.
2973 Type *VecTy = (VF == 1) ? PN->getType() :
2974 VectorType::get(PN->getType(), VF);
2975 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2976 LoopVectorBody.back()-> getFirstInsertionPt());
2982 setDebugLocFromInst(Builder, P);
2983 // Check for PHI nodes that are lowered to vector selects.
2984 if (P->getParent() != OrigLoop->getHeader()) {
2985 // We know that all PHIs in non-header blocks are converted into
2986 // selects, so we don't have to worry about the insertion order and we
2987 // can just use the builder.
2988 // At this point we generate the predication tree. There may be
2989 // duplications since this is a simple recursive scan, but future
2990 // optimizations will clean it up.
2992 unsigned NumIncoming = P->getNumIncomingValues();
2994 // Generate a sequence of selects of the form:
2995 // SELECT(Mask3, In3,
2996 // SELECT(Mask2, In2,
2998 for (unsigned In = 0; In < NumIncoming; In++) {
2999 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3001 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3003 for (unsigned part = 0; part < UF; ++part) {
3004 // We might have single edge PHIs (blocks) - use an identity
3005 // 'select' for the first PHI operand.
3007 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3010 // Select between the current value and the previous incoming edge
3011 // based on the incoming mask.
3012 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3013 Entry[part], "predphi");
3019 // This PHINode must be an induction variable.
3020 // Make sure that we know about it.
3021 assert(Legal->getInductionVars()->count(P) &&
3022 "Not an induction variable");
3024 LoopVectorizationLegality::InductionInfo II =
3025 Legal->getInductionVars()->lookup(P);
3028 case LoopVectorizationLegality::IK_NoInduction:
3029 llvm_unreachable("Unknown induction");
3030 case LoopVectorizationLegality::IK_IntInduction: {
3031 assert(P->getType() == II.StartValue->getType() && "Types must match");
3032 Type *PhiTy = P->getType();
3034 if (P == OldInduction) {
3035 // Handle the canonical induction variable. We might have had to
3037 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3039 // Handle other induction variables that are now based on the
3041 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3043 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3044 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3047 Broadcasted = getBroadcastInstrs(Broadcasted);
3048 // After broadcasting the induction variable we need to make the vector
3049 // consecutive by adding 0, 1, 2, etc.
3050 for (unsigned part = 0; part < UF; ++part)
3051 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3054 case LoopVectorizationLegality::IK_ReverseIntInduction:
3055 case LoopVectorizationLegality::IK_PtrInduction:
3056 case LoopVectorizationLegality::IK_ReversePtrInduction:
3057 // Handle reverse integer and pointer inductions.
3058 Value *StartIdx = ExtendedIdx;
3059 // This is the normalized GEP that starts counting at zero.
3060 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3063 // Handle the reverse integer induction variable case.
3064 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3065 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3066 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3068 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3071 // This is a new value so do not hoist it out.
3072 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3073 // After broadcasting the induction variable we need to make the
3074 // vector consecutive by adding ... -3, -2, -1, 0.
3075 for (unsigned part = 0; part < UF; ++part)
3076 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3081 // Handle the pointer induction variable case.
3082 assert(P->getType()->isPointerTy() && "Unexpected type.");
3084 // Is this a reverse induction ptr or a consecutive induction ptr.
3085 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3088 // This is the vector of results. Notice that we don't generate
3089 // vector geps because scalar geps result in better code.
3090 for (unsigned part = 0; part < UF; ++part) {
3092 int EltIndex = (part) * (Reverse ? -1 : 1);
3093 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3096 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3098 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3100 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3102 Entry[part] = SclrGep;
3106 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3107 for (unsigned int i = 0; i < VF; ++i) {
3108 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3109 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3112 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3114 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3116 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3118 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3119 Builder.getInt32(i),
3122 Entry[part] = VecVal;
3128 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3129 // For each instruction in the old loop.
3130 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3131 VectorParts &Entry = WidenMap.get(it);
3132 switch (it->getOpcode()) {
3133 case Instruction::Br:
3134 // Nothing to do for PHIs and BR, since we already took care of the
3135 // loop control flow instructions.
3137 case Instruction::PHI:{
3138 // Vectorize PHINodes.
3139 widenPHIInstruction(it, Entry, UF, VF, PV);
3143 case Instruction::Add:
3144 case Instruction::FAdd:
3145 case Instruction::Sub:
3146 case Instruction::FSub:
3147 case Instruction::Mul:
3148 case Instruction::FMul:
3149 case Instruction::UDiv:
3150 case Instruction::SDiv:
3151 case Instruction::FDiv:
3152 case Instruction::URem:
3153 case Instruction::SRem:
3154 case Instruction::FRem:
3155 case Instruction::Shl:
3156 case Instruction::LShr:
3157 case Instruction::AShr:
3158 case Instruction::And:
3159 case Instruction::Or:
3160 case Instruction::Xor: {
3161 // Just widen binops.
3162 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3163 setDebugLocFromInst(Builder, BinOp);
3164 VectorParts &A = getVectorValue(it->getOperand(0));
3165 VectorParts &B = getVectorValue(it->getOperand(1));
3167 // Use this vector value for all users of the original instruction.
3168 for (unsigned Part = 0; Part < UF; ++Part) {
3169 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3171 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3172 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3173 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3174 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3175 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3177 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3178 VecOp->setIsExact(BinOp->isExact());
3180 // Copy the fast-math flags.
3181 if (VecOp && isa<FPMathOperator>(V))
3182 VecOp->setFastMathFlags(it->getFastMathFlags());
3187 propagateMetadata(Entry, it);
3190 case Instruction::Select: {
3192 // If the selector is loop invariant we can create a select
3193 // instruction with a scalar condition. Otherwise, use vector-select.
3194 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3196 setDebugLocFromInst(Builder, it);
3198 // The condition can be loop invariant but still defined inside the
3199 // loop. This means that we can't just use the original 'cond' value.
3200 // We have to take the 'vectorized' value and pick the first lane.
3201 // Instcombine will make this a no-op.
3202 VectorParts &Cond = getVectorValue(it->getOperand(0));
3203 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3204 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3206 Value *ScalarCond = (VF == 1) ? Cond[0] :
3207 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3209 for (unsigned Part = 0; Part < UF; ++Part) {
3210 Entry[Part] = Builder.CreateSelect(
3211 InvariantCond ? ScalarCond : Cond[Part],
3216 propagateMetadata(Entry, it);
3220 case Instruction::ICmp:
3221 case Instruction::FCmp: {
3222 // Widen compares. Generate vector compares.
3223 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3224 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3225 setDebugLocFromInst(Builder, it);
3226 VectorParts &A = getVectorValue(it->getOperand(0));
3227 VectorParts &B = getVectorValue(it->getOperand(1));
3228 for (unsigned Part = 0; Part < UF; ++Part) {
3231 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3233 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3237 propagateMetadata(Entry, it);
3241 case Instruction::Store:
3242 case Instruction::Load:
3243 vectorizeMemoryInstruction(it);
3245 case Instruction::ZExt:
3246 case Instruction::SExt:
3247 case Instruction::FPToUI:
3248 case Instruction::FPToSI:
3249 case Instruction::FPExt:
3250 case Instruction::PtrToInt:
3251 case Instruction::IntToPtr:
3252 case Instruction::SIToFP:
3253 case Instruction::UIToFP:
3254 case Instruction::Trunc:
3255 case Instruction::FPTrunc:
3256 case Instruction::BitCast: {
3257 CastInst *CI = dyn_cast<CastInst>(it);
3258 setDebugLocFromInst(Builder, it);
3259 /// Optimize the special case where the source is the induction
3260 /// variable. Notice that we can only optimize the 'trunc' case
3261 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3262 /// c. other casts depend on pointer size.
3263 if (CI->getOperand(0) == OldInduction &&
3264 it->getOpcode() == Instruction::Trunc) {
3265 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3267 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3268 for (unsigned Part = 0; Part < UF; ++Part)
3269 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3270 propagateMetadata(Entry, it);
3273 /// Vectorize casts.
3274 Type *DestTy = (VF == 1) ? CI->getType() :
3275 VectorType::get(CI->getType(), VF);
3277 VectorParts &A = getVectorValue(it->getOperand(0));
3278 for (unsigned Part = 0; Part < UF; ++Part)
3279 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3280 propagateMetadata(Entry, it);
3284 case Instruction::Call: {
3285 // Ignore dbg intrinsics.
3286 if (isa<DbgInfoIntrinsic>(it))
3288 setDebugLocFromInst(Builder, it);
3290 Module *M = BB->getParent()->getParent();
3291 CallInst *CI = cast<CallInst>(it);
3292 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3293 assert(ID && "Not an intrinsic call!");
3295 case Intrinsic::lifetime_end:
3296 case Intrinsic::lifetime_start:
3297 scalarizeInstruction(it);
3300 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3301 for (unsigned Part = 0; Part < UF; ++Part) {
3302 SmallVector<Value *, 4> Args;
3303 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3304 if (HasScalarOpd && i == 1) {
3305 Args.push_back(CI->getArgOperand(i));
3308 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3309 Args.push_back(Arg[Part]);
3311 Type *Tys[] = {CI->getType()};
3313 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3315 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3316 Entry[Part] = Builder.CreateCall(F, Args);
3319 propagateMetadata(Entry, it);
3326 // All other instructions are unsupported. Scalarize them.
3327 scalarizeInstruction(it);
3330 }// end of for_each instr.
3333 void InnerLoopVectorizer::updateAnalysis() {
3334 // Forget the original basic block.
3335 SE->forgetLoop(OrigLoop);
3337 // Update the dominator tree information.
3338 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3339 "Entry does not dominate exit.");
3341 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3342 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3343 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3345 // Due to if predication of stores we might create a sequence of "if(pred)
3346 // a[i] = ...; " blocks.
3347 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3349 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3350 else if (isPredicatedBlock(i)) {
3351 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3353 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3357 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3358 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3359 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3360 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3362 DEBUG(DT->verifyDomTree());
3365 /// \brief Check whether it is safe to if-convert this phi node.
3367 /// Phi nodes with constant expressions that can trap are not safe to if
3369 static bool canIfConvertPHINodes(BasicBlock *BB) {
3370 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3371 PHINode *Phi = dyn_cast<PHINode>(I);
3374 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3375 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3382 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3383 if (!EnableIfConversion) {
3384 emitAnalysis(Report() << "if-conversion is disabled");
3388 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3390 // A list of pointers that we can safely read and write to.
3391 SmallPtrSet<Value *, 8> SafePointes;
3393 // Collect safe addresses.
3394 for (Loop::block_iterator BI = TheLoop->block_begin(),
3395 BE = TheLoop->block_end(); BI != BE; ++BI) {
3396 BasicBlock *BB = *BI;
3398 if (blockNeedsPredication(BB))
3401 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3402 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3403 SafePointes.insert(LI->getPointerOperand());
3404 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3405 SafePointes.insert(SI->getPointerOperand());
3409 // Collect the blocks that need predication.
3410 BasicBlock *Header = TheLoop->getHeader();
3411 for (Loop::block_iterator BI = TheLoop->block_begin(),
3412 BE = TheLoop->block_end(); BI != BE; ++BI) {
3413 BasicBlock *BB = *BI;
3415 // We don't support switch statements inside loops.
3416 if (!isa<BranchInst>(BB->getTerminator())) {
3417 emitAnalysis(Report(BB->getTerminator())
3418 << "loop contains a switch statement");
3422 // We must be able to predicate all blocks that need to be predicated.
3423 if (blockNeedsPredication(BB)) {
3424 if (!blockCanBePredicated(BB, SafePointes)) {
3425 emitAnalysis(Report(BB->getTerminator())
3426 << "control flow cannot be substituted for a select");
3429 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3430 emitAnalysis(Report(BB->getTerminator())
3431 << "control flow cannot be substituted for a select");
3436 // We can if-convert this loop.
3440 bool LoopVectorizationLegality::canVectorize() {
3441 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3442 // be canonicalized.
3443 if (!TheLoop->getLoopPreheader()) {
3445 Report() << "loop control flow is not understood by vectorizer");
3449 // We can only vectorize innermost loops.
3450 if (TheLoop->getSubLoopsVector().size()) {
3451 emitAnalysis(Report() << "loop is not the innermost loop");
3455 // We must have a single backedge.
3456 if (TheLoop->getNumBackEdges() != 1) {
3458 Report() << "loop control flow is not understood by vectorizer");
3462 // We must have a single exiting block.
3463 if (!TheLoop->getExitingBlock()) {
3465 Report() << "loop control flow is not understood by vectorizer");
3469 // We need to have a loop header.
3470 DEBUG(dbgs() << "LV: Found a loop: " <<
3471 TheLoop->getHeader()->getName() << '\n');
3473 // Check if we can if-convert non-single-bb loops.
3474 unsigned NumBlocks = TheLoop->getNumBlocks();
3475 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3476 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3480 // ScalarEvolution needs to be able to find the exit count.
3481 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3482 if (ExitCount == SE->getCouldNotCompute()) {
3483 emitAnalysis(Report() << "could not determine number of loop iterations");
3484 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3488 // Check if we can vectorize the instructions and CFG in this loop.
3489 if (!canVectorizeInstrs()) {
3490 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3494 // Go over each instruction and look at memory deps.
3495 if (!canVectorizeMemory()) {
3496 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3500 // Collect all of the variables that remain uniform after vectorization.
3501 collectLoopUniforms();
3503 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3504 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3507 // Okay! We can vectorize. At this point we don't have any other mem analysis
3508 // which may limit our maximum vectorization factor, so just return true with
3513 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3514 if (Ty->isPointerTy())
3515 return DL.getIntPtrType(Ty);
3517 // It is possible that char's or short's overflow when we ask for the loop's
3518 // trip count, work around this by changing the type size.
3519 if (Ty->getScalarSizeInBits() < 32)
3520 return Type::getInt32Ty(Ty->getContext());
3525 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3526 Ty0 = convertPointerToIntegerType(DL, Ty0);
3527 Ty1 = convertPointerToIntegerType(DL, Ty1);
3528 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3533 /// \brief Check that the instruction has outside loop users and is not an
3534 /// identified reduction variable.
3535 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3536 SmallPtrSet<Value *, 4> &Reductions) {
3537 // Reduction instructions are allowed to have exit users. All other
3538 // instructions must not have external users.
3539 if (!Reductions.count(Inst))
3540 //Check that all of the users of the loop are inside the BB.
3541 for (User *U : Inst->users()) {
3542 Instruction *UI = cast<Instruction>(U);
3543 // This user may be a reduction exit value.
3544 if (!TheLoop->contains(UI)) {
3545 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3552 bool LoopVectorizationLegality::canVectorizeInstrs() {
3553 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3554 BasicBlock *Header = TheLoop->getHeader();
3556 // Look for the attribute signaling the absence of NaNs.
3557 Function &F = *Header->getParent();
3558 if (F.hasFnAttribute("no-nans-fp-math"))
3559 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3560 AttributeSet::FunctionIndex,
3561 "no-nans-fp-math").getValueAsString() == "true";
3563 // For each block in the loop.
3564 for (Loop::block_iterator bb = TheLoop->block_begin(),
3565 be = TheLoop->block_end(); bb != be; ++bb) {
3567 // Scan the instructions in the block and look for hazards.
3568 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3571 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3572 Type *PhiTy = Phi->getType();
3573 // Check that this PHI type is allowed.
3574 if (!PhiTy->isIntegerTy() &&
3575 !PhiTy->isFloatingPointTy() &&
3576 !PhiTy->isPointerTy()) {
3577 emitAnalysis(Report(it)
3578 << "loop control flow is not understood by vectorizer");
3579 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3583 // If this PHINode is not in the header block, then we know that we
3584 // can convert it to select during if-conversion. No need to check if
3585 // the PHIs in this block are induction or reduction variables.
3586 if (*bb != Header) {
3587 // Check that this instruction has no outside users or is an
3588 // identified reduction value with an outside user.
3589 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3591 emitAnalysis(Report(it) << "value that could not be identified as "
3592 "reduction is used outside the loop");
3596 // We only allow if-converted PHIs with more than two incoming values.
3597 if (Phi->getNumIncomingValues() != 2) {
3598 emitAnalysis(Report(it)
3599 << "control flow not understood by vectorizer");
3600 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3604 // This is the value coming from the preheader.
3605 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3606 // Check if this is an induction variable.
3607 InductionKind IK = isInductionVariable(Phi);
3609 if (IK_NoInduction != IK) {
3610 // Get the widest type.
3612 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3614 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3616 // Int inductions are special because we only allow one IV.
3617 if (IK == IK_IntInduction) {
3618 // Use the phi node with the widest type as induction. Use the last
3619 // one if there are multiple (no good reason for doing this other
3620 // than it is expedient).
3621 if (!Induction || PhiTy == WidestIndTy)
3625 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3626 Inductions[Phi] = InductionInfo(StartValue, IK);
3628 // Until we explicitly handle the case of an induction variable with
3629 // an outside loop user we have to give up vectorizing this loop.
3630 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3631 emitAnalysis(Report(it) << "use of induction value outside of the "
3632 "loop is not handled by vectorizer");
3639 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3640 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3643 if (AddReductionVar(Phi, RK_IntegerMult)) {
3644 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3647 if (AddReductionVar(Phi, RK_IntegerOr)) {
3648 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3651 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3652 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3655 if (AddReductionVar(Phi, RK_IntegerXor)) {
3656 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3659 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3660 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3663 if (AddReductionVar(Phi, RK_FloatMult)) {
3664 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3667 if (AddReductionVar(Phi, RK_FloatAdd)) {
3668 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3671 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3672 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3677 emitAnalysis(Report(it) << "unvectorizable operation");
3678 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3680 }// end of PHI handling
3682 // We still don't handle functions. However, we can ignore dbg intrinsic
3683 // calls and we do handle certain intrinsic and libm functions.
3684 CallInst *CI = dyn_cast<CallInst>(it);
3685 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3686 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3687 DEBUG(dbgs() << "LV: Found a call site.\n");
3691 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3692 // second argument is the same (i.e. loop invariant)
3694 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3695 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3696 emitAnalysis(Report(it)
3697 << "intrinsic instruction cannot be vectorized");
3698 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3703 // Check that the instruction return type is vectorizable.
3704 // Also, we can't vectorize extractelement instructions.
3705 if ((!VectorType::isValidElementType(it->getType()) &&
3706 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3707 emitAnalysis(Report(it)
3708 << "instruction return type cannot be vectorized");
3709 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3713 // Check that the stored type is vectorizable.
3714 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3715 Type *T = ST->getValueOperand()->getType();
3716 if (!VectorType::isValidElementType(T)) {
3717 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3720 if (EnableMemAccessVersioning)
3721 collectStridedAcccess(ST);
3724 if (EnableMemAccessVersioning)
3725 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3726 collectStridedAcccess(LI);
3728 // Reduction instructions are allowed to have exit users.
3729 // All other instructions must not have external users.
3730 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3731 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3740 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3741 if (Inductions.empty()) {
3742 emitAnalysis(Report()
3743 << "loop induction variable could not be identified");
3751 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3752 /// return the induction operand of the gep pointer.
3753 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3754 const DataLayout *DL, Loop *Lp) {
3755 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3759 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3761 // Check that all of the gep indices are uniform except for our induction
3763 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3764 if (i != InductionOperand &&
3765 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3767 return GEP->getOperand(InductionOperand);
3770 ///\brief Look for a cast use of the passed value.
3771 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3772 Value *UniqueCast = nullptr;
3773 for (User *U : Ptr->users()) {
3774 CastInst *CI = dyn_cast<CastInst>(U);
3775 if (CI && CI->getType() == Ty) {
3785 ///\brief Get the stride of a pointer access in a loop.
3786 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3787 /// pointer to the Value, or null otherwise.
3788 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3789 const DataLayout *DL, Loop *Lp) {
3790 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3791 if (!PtrTy || PtrTy->isAggregateType())
3794 // Try to remove a gep instruction to make the pointer (actually index at this
3795 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3796 // pointer, otherwise, we are analyzing the index.
3797 Value *OrigPtr = Ptr;
3799 // The size of the pointer access.
3800 int64_t PtrAccessSize = 1;
3802 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3803 const SCEV *V = SE->getSCEV(Ptr);
3807 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3808 V = C->getOperand();
3810 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3814 V = S->getStepRecurrence(*SE);
3818 // Strip off the size of access multiplication if we are still analyzing the
3820 if (OrigPtr == Ptr) {
3821 DL->getTypeAllocSize(PtrTy->getElementType());
3822 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3823 if (M->getOperand(0)->getSCEVType() != scConstant)
3826 const APInt &APStepVal =
3827 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3829 // Huge step value - give up.
3830 if (APStepVal.getBitWidth() > 64)
3833 int64_t StepVal = APStepVal.getSExtValue();
3834 if (PtrAccessSize != StepVal)
3836 V = M->getOperand(1);
3841 Type *StripedOffRecurrenceCast = nullptr;
3842 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3843 StripedOffRecurrenceCast = C->getType();
3844 V = C->getOperand();
3847 // Look for the loop invariant symbolic value.
3848 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3852 Value *Stride = U->getValue();
3853 if (!Lp->isLoopInvariant(Stride))
3856 // If we have stripped off the recurrence cast we have to make sure that we
3857 // return the value that is used in this loop so that we can replace it later.
3858 if (StripedOffRecurrenceCast)
3859 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3864 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3865 Value *Ptr = nullptr;
3866 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3867 Ptr = LI->getPointerOperand();
3868 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3869 Ptr = SI->getPointerOperand();
3873 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3877 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3878 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3879 Strides[Ptr] = Stride;
3880 StrideSet.insert(Stride);
3883 void LoopVectorizationLegality::collectLoopUniforms() {
3884 // We now know that the loop is vectorizable!
3885 // Collect variables that will remain uniform after vectorization.
3886 std::vector<Value*> Worklist;
3887 BasicBlock *Latch = TheLoop->getLoopLatch();
3889 // Start with the conditional branch and walk up the block.
3890 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3892 // Also add all consecutive pointer values; these values will be uniform
3893 // after vectorization (and subsequent cleanup) and, until revectorization is
3894 // supported, all dependencies must also be uniform.
3895 for (Loop::block_iterator B = TheLoop->block_begin(),
3896 BE = TheLoop->block_end(); B != BE; ++B)
3897 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3899 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3900 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3902 while (Worklist.size()) {
3903 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3904 Worklist.pop_back();
3906 // Look at instructions inside this loop.
3907 // Stop when reaching PHI nodes.
3908 // TODO: we need to follow values all over the loop, not only in this block.
3909 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3912 // This is a known uniform.
3915 // Insert all operands.
3916 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3921 /// \brief Analyses memory accesses in a loop.
3923 /// Checks whether run time pointer checks are needed and builds sets for data
3924 /// dependence checking.
3925 class AccessAnalysis {
3927 /// \brief Read or write access location.
3928 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3929 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3931 /// \brief Set of potential dependent memory accesses.
3932 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3934 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
3935 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
3937 /// \brief Register a load and whether it is only read from.
3938 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
3939 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3940 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3941 Accesses.insert(MemAccessInfo(Ptr, false));
3943 ReadOnlyPtr.insert(Ptr);
3946 /// \brief Register a store.
3947 void addStore(AliasAnalysis::Location &Loc) {
3948 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3949 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3950 Accesses.insert(MemAccessInfo(Ptr, true));
3953 /// \brief Check whether we can check the pointers at runtime for
3954 /// non-intersection.
3955 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3956 unsigned &NumComparisons, ScalarEvolution *SE,
3957 Loop *TheLoop, ValueToValueMap &Strides,
3958 bool ShouldCheckStride = false);
3960 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3961 /// and builds sets of dependent accesses.
3962 void buildDependenceSets() {
3963 processMemAccesses();
3966 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3968 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3969 void resetDepChecks() { CheckDeps.clear(); }
3971 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3974 typedef SetVector<MemAccessInfo> PtrAccessSet;
3976 /// \brief Go over all memory access and check whether runtime pointer checks
3977 /// are needed /// and build sets of dependency check candidates.
3978 void processMemAccesses();
3980 /// Set of all accesses.
3981 PtrAccessSet Accesses;
3983 /// Set of accesses that need a further dependence check.
3984 MemAccessInfoSet CheckDeps;
3986 /// Set of pointers that are read only.
3987 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3989 const DataLayout *DL;
3991 /// An alias set tracker to partition the access set by underlying object and
3992 //intrinsic property (such as TBAA metadata).
3993 AliasSetTracker AST;
3995 /// Sets of potentially dependent accesses - members of one set share an
3996 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3997 /// dependence check.
3998 DepCandidates &DepCands;
4000 bool IsRTCheckNeeded;
4003 } // end anonymous namespace
4005 /// \brief Check whether a pointer can participate in a runtime bounds check.
4006 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4008 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4009 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4013 return AR->isAffine();
4016 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4017 /// the address space.
4018 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4019 const Loop *Lp, ValueToValueMap &StridesMap);
4021 bool AccessAnalysis::canCheckPtrAtRT(
4022 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4023 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4024 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4025 // Find pointers with computable bounds. We are going to use this information
4026 // to place a runtime bound check.
4027 bool CanDoRT = true;
4029 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4032 // We assign a consecutive id to access from different alias sets.
4033 // Accesses between different groups doesn't need to be checked.
4035 for (auto &AS : AST) {
4036 unsigned NumReadPtrChecks = 0;
4037 unsigned NumWritePtrChecks = 0;
4039 // We assign consecutive id to access from different dependence sets.
4040 // Accesses within the same set don't need a runtime check.
4041 unsigned RunningDepId = 1;
4042 DenseMap<Value *, unsigned> DepSetId;
4045 Value *Ptr = A.getValue();
4046 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4047 MemAccessInfo Access(Ptr, IsWrite);
4050 ++NumWritePtrChecks;
4054 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4055 // When we run after a failing dependency check we have to make sure we
4056 // don't have wrapping pointers.
4057 (!ShouldCheckStride ||
4058 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4059 // The id of the dependence set.
4062 if (IsDepCheckNeeded) {
4063 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4064 unsigned &LeaderId = DepSetId[Leader];
4066 LeaderId = RunningDepId++;
4069 // Each access has its own dependence set.
4070 DepId = RunningDepId++;
4072 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4074 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4080 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4081 NumComparisons += 0; // Only one dependence set.
4083 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4084 NumWritePtrChecks - 1));
4090 // If the pointers that we would use for the bounds comparison have different
4091 // address spaces, assume the values aren't directly comparable, so we can't
4092 // use them for the runtime check. We also have to assume they could
4093 // overlap. In the future there should be metadata for whether address spaces
4095 unsigned NumPointers = RtCheck.Pointers.size();
4096 for (unsigned i = 0; i < NumPointers; ++i) {
4097 for (unsigned j = i + 1; j < NumPointers; ++j) {
4098 // Only need to check pointers between two different dependency sets.
4099 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4101 // Only need to check pointers in the same alias set.
4102 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4105 Value *PtrI = RtCheck.Pointers[i];
4106 Value *PtrJ = RtCheck.Pointers[j];
4108 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4109 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4111 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4112 " different address spaces\n");
4121 void AccessAnalysis::processMemAccesses() {
4122 // We process the set twice: first we process read-write pointers, last we
4123 // process read-only pointers. This allows us to skip dependence tests for
4124 // read-only pointers.
4126 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4127 DEBUG(dbgs() << " AST: "; AST.dump());
4128 DEBUG(dbgs() << "LV: Accesses:\n");
4130 for (auto A : Accesses)
4131 dbgs() << "\t" << *A.getPointer() << " (" <<
4132 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4133 "read-only" : "read")) << ")\n";
4136 // The AliasSetTracker has nicely partitioned our pointers by metadata
4137 // compatibility and potential for underlying-object overlap. As a result, we
4138 // only need to check for potential pointer dependencies within each alias
4140 for (auto &AS : AST) {
4141 // Note that both the alias-set tracker and the alias sets themselves used
4142 // linked lists internally and so the iteration order here is deterministic
4143 // (matching the original instruction order within each set).
4145 bool SetHasWrite = false;
4147 // Map of pointers to last access encountered.
4148 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4149 UnderlyingObjToAccessMap ObjToLastAccess;
4151 // Set of access to check after all writes have been processed.
4152 PtrAccessSet DeferredAccesses;
4154 // Iterate over each alias set twice, once to process read/write pointers,
4155 // and then to process read-only pointers.
4156 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4157 bool UseDeferred = SetIteration > 0;
4158 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4161 Value *Ptr = A.getValue();
4162 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4164 // If we're using the deferred access set, then it contains only reads.
4165 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4166 if (UseDeferred && !IsReadOnlyPtr)
4168 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4170 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4171 S.count(MemAccessInfo(Ptr, false))) &&
4172 "Alias-set pointer not in the access set?");
4174 MemAccessInfo Access(Ptr, IsWrite);
4175 DepCands.insert(Access);
4177 // Memorize read-only pointers for later processing and skip them in the
4178 // first round (they need to be checked after we have seen all write
4179 // pointers). Note: we also mark pointer that are not consecutive as
4180 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4181 // the second check for "!IsWrite".
4182 if (!UseDeferred && IsReadOnlyPtr) {
4183 DeferredAccesses.insert(Access);
4187 // If this is a write - check other reads and writes for conflicts. If
4188 // this is a read only check other writes for conflicts (but only if
4189 // there is no other write to the ptr - this is an optimization to
4190 // catch "a[i] = a[i] + " without having to do a dependence check).
4191 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4192 CheckDeps.insert(Access);
4193 IsRTCheckNeeded = true;
4199 // Create sets of pointers connected by a shared alias set and
4200 // underlying object.
4201 typedef SmallVector<Value*, 16> ValueVector;
4202 ValueVector TempObjects;
4203 GetUnderlyingObjects(Ptr, TempObjects, DL);
4204 for (Value *UnderlyingObj : TempObjects) {
4205 UnderlyingObjToAccessMap::iterator Prev =
4206 ObjToLastAccess.find(UnderlyingObj);
4207 if (Prev != ObjToLastAccess.end())
4208 DepCands.unionSets(Access, Prev->second);
4210 ObjToLastAccess[UnderlyingObj] = Access;
4218 /// \brief Checks memory dependences among accesses to the same underlying
4219 /// object to determine whether there vectorization is legal or not (and at
4220 /// which vectorization factor).
4222 /// This class works under the assumption that we already checked that memory
4223 /// locations with different underlying pointers are "must-not alias".
4224 /// We use the ScalarEvolution framework to symbolically evalutate access
4225 /// functions pairs. Since we currently don't restructure the loop we can rely
4226 /// on the program order of memory accesses to determine their safety.
4227 /// At the moment we will only deem accesses as safe for:
4228 /// * A negative constant distance assuming program order.
4230 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4231 /// a[i] = tmp; y = a[i];
4233 /// The latter case is safe because later checks guarantuee that there can't
4234 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4235 /// the same variable: a header phi can only be an induction or a reduction, a
4236 /// reduction can't have a memory sink, an induction can't have a memory
4237 /// source). This is important and must not be violated (or we have to
4238 /// resort to checking for cycles through memory).
4240 /// * A positive constant distance assuming program order that is bigger
4241 /// than the biggest memory access.
4243 /// tmp = a[i] OR b[i] = x
4244 /// a[i+2] = tmp y = b[i+2];
4246 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4248 /// * Zero distances and all accesses have the same size.
4250 class MemoryDepChecker {
4252 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4253 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4255 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4256 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4257 ShouldRetryWithRuntimeCheck(false) {}
4259 /// \brief Register the location (instructions are given increasing numbers)
4260 /// of a write access.
4261 void addAccess(StoreInst *SI) {
4262 Value *Ptr = SI->getPointerOperand();
4263 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4264 InstMap.push_back(SI);
4268 /// \brief Register the location (instructions are given increasing numbers)
4269 /// of a write access.
4270 void addAccess(LoadInst *LI) {
4271 Value *Ptr = LI->getPointerOperand();
4272 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4273 InstMap.push_back(LI);
4277 /// \brief Check whether the dependencies between the accesses are safe.
4279 /// Only checks sets with elements in \p CheckDeps.
4280 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4281 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4283 /// \brief The maximum number of bytes of a vector register we can vectorize
4284 /// the accesses safely with.
4285 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4287 /// \brief In same cases when the dependency check fails we can still
4288 /// vectorize the loop with a dynamic array access check.
4289 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4292 ScalarEvolution *SE;
4293 const DataLayout *DL;
4294 const Loop *InnermostLoop;
4296 /// \brief Maps access locations (ptr, read/write) to program order.
4297 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4299 /// \brief Memory access instructions in program order.
4300 SmallVector<Instruction *, 16> InstMap;
4302 /// \brief The program order index to be used for the next instruction.
4305 // We can access this many bytes in parallel safely.
4306 unsigned MaxSafeDepDistBytes;
4308 /// \brief If we see a non-constant dependence distance we can still try to
4309 /// vectorize this loop with runtime checks.
4310 bool ShouldRetryWithRuntimeCheck;
4312 /// \brief Check whether there is a plausible dependence between the two
4315 /// Access \p A must happen before \p B in program order. The two indices
4316 /// identify the index into the program order map.
4318 /// This function checks whether there is a plausible dependence (or the
4319 /// absence of such can't be proved) between the two accesses. If there is a
4320 /// plausible dependence but the dependence distance is bigger than one
4321 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4322 /// distance is smaller than any other distance encountered so far).
4323 /// Otherwise, this function returns true signaling a possible dependence.
4324 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4325 const MemAccessInfo &B, unsigned BIdx,
4326 ValueToValueMap &Strides);
4328 /// \brief Check whether the data dependence could prevent store-load
4330 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4333 } // end anonymous namespace
4335 static bool isInBoundsGep(Value *Ptr) {
4336 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4337 return GEP->isInBounds();
4341 /// \brief Check whether the access through \p Ptr has a constant stride.
4342 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4343 const Loop *Lp, ValueToValueMap &StridesMap) {
4344 const Type *Ty = Ptr->getType();
4345 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4347 // Make sure that the pointer does not point to aggregate types.
4348 const PointerType *PtrTy = cast<PointerType>(Ty);
4349 if (PtrTy->getElementType()->isAggregateType()) {
4350 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4355 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4357 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4359 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4360 << *Ptr << " SCEV: " << *PtrScev << "\n");
4364 // The accesss function must stride over the innermost loop.
4365 if (Lp != AR->getLoop()) {
4366 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4367 *Ptr << " SCEV: " << *PtrScev << "\n");
4370 // The address calculation must not wrap. Otherwise, a dependence could be
4372 // An inbounds getelementptr that is a AddRec with a unit stride
4373 // cannot wrap per definition. The unit stride requirement is checked later.
4374 // An getelementptr without an inbounds attribute and unit stride would have
4375 // to access the pointer value "0" which is undefined behavior in address
4376 // space 0, therefore we can also vectorize this case.
4377 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4378 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4379 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4380 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4381 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4382 << *Ptr << " SCEV: " << *PtrScev << "\n");
4386 // Check the step is constant.
4387 const SCEV *Step = AR->getStepRecurrence(*SE);
4389 // Calculate the pointer stride and check if it is consecutive.
4390 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4392 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4393 " SCEV: " << *PtrScev << "\n");
4397 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4398 const APInt &APStepVal = C->getValue()->getValue();
4400 // Huge step value - give up.
4401 if (APStepVal.getBitWidth() > 64)
4404 int64_t StepVal = APStepVal.getSExtValue();
4407 int64_t Stride = StepVal / Size;
4408 int64_t Rem = StepVal % Size;
4412 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4413 // know we can't "wrap around the address space". In case of address space
4414 // zero we know that this won't happen without triggering undefined behavior.
4415 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4416 Stride != 1 && Stride != -1)
4422 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4423 unsigned TypeByteSize) {
4424 // If loads occur at a distance that is not a multiple of a feasible vector
4425 // factor store-load forwarding does not take place.
4426 // Positive dependences might cause troubles because vectorizing them might
4427 // prevent store-load forwarding making vectorized code run a lot slower.
4428 // a[i] = a[i-3] ^ a[i-8];
4429 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4430 // hence on your typical architecture store-load forwarding does not take
4431 // place. Vectorizing in such cases does not make sense.
4432 // Store-load forwarding distance.
4433 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4434 // Maximum vector factor.
4435 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4436 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4437 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4439 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4441 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4442 MaxVFWithoutSLForwardIssues = (vf >>=1);
4447 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4448 DEBUG(dbgs() << "LV: Distance " << Distance <<
4449 " that could cause a store-load forwarding conflict\n");
4453 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4454 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4455 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4459 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4460 const MemAccessInfo &B, unsigned BIdx,
4461 ValueToValueMap &Strides) {
4462 assert (AIdx < BIdx && "Must pass arguments in program order");
4464 Value *APtr = A.getPointer();
4465 Value *BPtr = B.getPointer();
4466 bool AIsWrite = A.getInt();
4467 bool BIsWrite = B.getInt();
4469 // Two reads are independent.
4470 if (!AIsWrite && !BIsWrite)
4473 // We cannot check pointers in different address spaces.
4474 if (APtr->getType()->getPointerAddressSpace() !=
4475 BPtr->getType()->getPointerAddressSpace())
4478 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4479 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4481 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4482 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4484 const SCEV *Src = AScev;
4485 const SCEV *Sink = BScev;
4487 // If the induction step is negative we have to invert source and sink of the
4489 if (StrideAPtr < 0) {
4492 std::swap(APtr, BPtr);
4493 std::swap(Src, Sink);
4494 std::swap(AIsWrite, BIsWrite);
4495 std::swap(AIdx, BIdx);
4496 std::swap(StrideAPtr, StrideBPtr);
4499 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4501 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4502 << "(Induction step: " << StrideAPtr << ")\n");
4503 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4504 << *InstMap[BIdx] << ": " << *Dist << "\n");
4506 // Need consecutive accesses. We don't want to vectorize
4507 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4508 // the address space.
4509 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4510 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4514 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4516 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4517 ShouldRetryWithRuntimeCheck = true;
4521 Type *ATy = APtr->getType()->getPointerElementType();
4522 Type *BTy = BPtr->getType()->getPointerElementType();
4523 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4525 // Negative distances are not plausible dependencies.
4526 const APInt &Val = C->getValue()->getValue();
4527 if (Val.isNegative()) {
4528 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4529 if (IsTrueDataDependence &&
4530 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4534 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4538 // Write to the same location with the same size.
4539 // Could be improved to assert type sizes are the same (i32 == float, etc).
4543 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4547 assert(Val.isStrictlyPositive() && "Expect a positive value");
4549 // Positive distance bigger than max vectorization factor.
4552 "LV: ReadWrite-Write positive dependency with different types\n");
4556 unsigned Distance = (unsigned) Val.getZExtValue();
4558 // Bail out early if passed-in parameters make vectorization not feasible.
4559 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4560 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4562 // The distance must be bigger than the size needed for a vectorized version
4563 // of the operation and the size of the vectorized operation must not be
4564 // bigger than the currrent maximum size.
4565 if (Distance < 2*TypeByteSize ||
4566 2*TypeByteSize > MaxSafeDepDistBytes ||
4567 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4568 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4569 << Val.getSExtValue() << '\n');
4573 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4574 Distance : MaxSafeDepDistBytes;
4576 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4577 if (IsTrueDataDependence &&
4578 couldPreventStoreLoadForward(Distance, TypeByteSize))
4581 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4582 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4587 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4588 MemAccessInfoSet &CheckDeps,
4589 ValueToValueMap &Strides) {
4591 MaxSafeDepDistBytes = -1U;
4592 while (!CheckDeps.empty()) {
4593 MemAccessInfo CurAccess = *CheckDeps.begin();
4595 // Get the relevant memory access set.
4596 EquivalenceClasses<MemAccessInfo>::iterator I =
4597 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4599 // Check accesses within this set.
4600 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4601 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4603 // Check every access pair.
4605 CheckDeps.erase(*AI);
4606 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4608 // Check every accessing instruction pair in program order.
4609 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4610 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4611 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4612 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4613 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4615 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4626 bool LoopVectorizationLegality::canVectorizeMemory() {
4628 typedef SmallVector<Value*, 16> ValueVector;
4629 typedef SmallPtrSet<Value*, 16> ValueSet;
4631 // Holds the Load and Store *instructions*.
4635 // Holds all the different accesses in the loop.
4636 unsigned NumReads = 0;
4637 unsigned NumReadWrites = 0;
4639 PtrRtCheck.Pointers.clear();
4640 PtrRtCheck.Need = false;
4642 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4643 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4646 for (Loop::block_iterator bb = TheLoop->block_begin(),
4647 be = TheLoop->block_end(); bb != be; ++bb) {
4649 // Scan the BB and collect legal loads and stores.
4650 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4653 // If this is a load, save it. If this instruction can read from memory
4654 // but is not a load, then we quit. Notice that we don't handle function
4655 // calls that read or write.
4656 if (it->mayReadFromMemory()) {
4657 // Many math library functions read the rounding mode. We will only
4658 // vectorize a loop if it contains known function calls that don't set
4659 // the flag. Therefore, it is safe to ignore this read from memory.
4660 CallInst *Call = dyn_cast<CallInst>(it);
4661 if (Call && getIntrinsicIDForCall(Call, TLI))
4664 LoadInst *Ld = dyn_cast<LoadInst>(it);
4665 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4666 emitAnalysis(Report(Ld)
4667 << "read with atomic ordering or volatile read");
4668 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4672 Loads.push_back(Ld);
4673 DepChecker.addAccess(Ld);
4677 // Save 'store' instructions. Abort if other instructions write to memory.
4678 if (it->mayWriteToMemory()) {
4679 StoreInst *St = dyn_cast<StoreInst>(it);
4681 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4684 if (!St->isSimple() && !IsAnnotatedParallel) {
4685 emitAnalysis(Report(St)
4686 << "write with atomic ordering or volatile write");
4687 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4691 Stores.push_back(St);
4692 DepChecker.addAccess(St);
4697 // Now we have two lists that hold the loads and the stores.
4698 // Next, we find the pointers that they use.
4700 // Check if we see any stores. If there are no stores, then we don't
4701 // care if the pointers are *restrict*.
4702 if (!Stores.size()) {
4703 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4707 AccessAnalysis::DepCandidates DependentAccesses;
4708 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4710 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4711 // multiple times on the same object. If the ptr is accessed twice, once
4712 // for read and once for write, it will only appear once (on the write
4713 // list). This is okay, since we are going to check for conflicts between
4714 // writes and between reads and writes, but not between reads and reads.
4717 ValueVector::iterator I, IE;
4718 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4719 StoreInst *ST = cast<StoreInst>(*I);
4720 Value* Ptr = ST->getPointerOperand();
4722 if (isUniform(Ptr)) {
4725 << "write to a loop invariant address could not be vectorized");
4726 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4730 // If we did *not* see this pointer before, insert it to the read-write
4731 // list. At this phase it is only a 'write' list.
4732 if (Seen.insert(Ptr)) {
4735 AliasAnalysis::Location Loc = AA->getLocation(ST);
4736 // The TBAA metadata could have a control dependency on the predication
4737 // condition, so we cannot rely on it when determining whether or not we
4738 // need runtime pointer checks.
4739 if (blockNeedsPredication(ST->getParent()))
4740 Loc.AATags.TBAA = nullptr;
4742 Accesses.addStore(Loc);
4746 if (IsAnnotatedParallel) {
4748 << "LV: A loop annotated parallel, ignore memory dependency "
4753 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4754 LoadInst *LD = cast<LoadInst>(*I);
4755 Value* Ptr = LD->getPointerOperand();
4756 // If we did *not* see this pointer before, insert it to the
4757 // read list. If we *did* see it before, then it is already in
4758 // the read-write list. This allows us to vectorize expressions
4759 // such as A[i] += x; Because the address of A[i] is a read-write
4760 // pointer. This only works if the index of A[i] is consecutive.
4761 // If the address of i is unknown (for example A[B[i]]) then we may
4762 // read a few words, modify, and write a few words, and some of the
4763 // words may be written to the same address.
4764 bool IsReadOnlyPtr = false;
4765 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4767 IsReadOnlyPtr = true;
4770 AliasAnalysis::Location Loc = AA->getLocation(LD);
4771 // The TBAA metadata could have a control dependency on the predication
4772 // condition, so we cannot rely on it when determining whether or not we
4773 // need runtime pointer checks.
4774 if (blockNeedsPredication(LD->getParent()))
4775 Loc.AATags.TBAA = nullptr;
4777 Accesses.addLoad(Loc, IsReadOnlyPtr);
4780 // If we write (or read-write) to a single destination and there are no
4781 // other reads in this loop then is it safe to vectorize.
4782 if (NumReadWrites == 1 && NumReads == 0) {
4783 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4787 // Build dependence sets and check whether we need a runtime pointer bounds
4789 Accesses.buildDependenceSets();
4790 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4792 // Find pointers with computable bounds. We are going to use this information
4793 // to place a runtime bound check.
4794 unsigned NumComparisons = 0;
4795 bool CanDoRT = false;
4797 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4800 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4801 " pointer comparisons.\n");
4803 // If we only have one set of dependences to check pointers among we don't
4804 // need a runtime check.
4805 if (NumComparisons == 0 && NeedRTCheck)
4806 NeedRTCheck = false;
4808 // Check that we did not collect too many pointers or found an unsizeable
4810 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4816 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4819 if (NeedRTCheck && !CanDoRT) {
4820 emitAnalysis(Report() << "cannot identify array bounds");
4821 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4822 "the array bounds.\n");
4827 PtrRtCheck.Need = NeedRTCheck;
4829 bool CanVecMem = true;
4830 if (Accesses.isDependencyCheckNeeded()) {
4831 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4832 CanVecMem = DepChecker.areDepsSafe(
4833 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4834 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4836 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4837 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4840 // Clear the dependency checks. We assume they are not needed.
4841 Accesses.resetDepChecks();
4844 PtrRtCheck.Need = true;
4846 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4847 TheLoop, Strides, true);
4848 // Check that we did not collect too many pointers or found an unsizeable
4850 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4851 if (!CanDoRT && NumComparisons > 0)
4852 emitAnalysis(Report()
4853 << "cannot check memory dependencies at runtime");
4855 emitAnalysis(Report()
4856 << NumComparisons << " exceeds limit of "
4857 << RuntimeMemoryCheckThreshold
4858 << " dependent memory operations checked at runtime");
4859 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4869 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4871 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4872 " need a runtime memory check.\n");
4877 static bool hasMultipleUsesOf(Instruction *I,
4878 SmallPtrSet<Instruction *, 8> &Insts) {
4879 unsigned NumUses = 0;
4880 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4881 if (Insts.count(dyn_cast<Instruction>(*Use)))
4890 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4891 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4892 if (!Set.count(dyn_cast<Instruction>(*Use)))
4897 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4898 ReductionKind Kind) {
4899 if (Phi->getNumIncomingValues() != 2)
4902 // Reduction variables are only found in the loop header block.
4903 if (Phi->getParent() != TheLoop->getHeader())
4906 // Obtain the reduction start value from the value that comes from the loop
4908 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4910 // ExitInstruction is the single value which is used outside the loop.
4911 // We only allow for a single reduction value to be used outside the loop.
4912 // This includes users of the reduction, variables (which form a cycle
4913 // which ends in the phi node).
4914 Instruction *ExitInstruction = nullptr;
4915 // Indicates that we found a reduction operation in our scan.
4916 bool FoundReduxOp = false;
4918 // We start with the PHI node and scan for all of the users of this
4919 // instruction. All users must be instructions that can be used as reduction
4920 // variables (such as ADD). We must have a single out-of-block user. The cycle
4921 // must include the original PHI.
4922 bool FoundStartPHI = false;
4924 // To recognize min/max patterns formed by a icmp select sequence, we store
4925 // the number of instruction we saw from the recognized min/max pattern,
4926 // to make sure we only see exactly the two instructions.
4927 unsigned NumCmpSelectPatternInst = 0;
4928 ReductionInstDesc ReduxDesc(false, nullptr);
4930 SmallPtrSet<Instruction *, 8> VisitedInsts;
4931 SmallVector<Instruction *, 8> Worklist;
4932 Worklist.push_back(Phi);
4933 VisitedInsts.insert(Phi);
4935 // A value in the reduction can be used:
4936 // - By the reduction:
4937 // - Reduction operation:
4938 // - One use of reduction value (safe).
4939 // - Multiple use of reduction value (not safe).
4941 // - All uses of the PHI must be the reduction (safe).
4942 // - Otherwise, not safe.
4943 // - By one instruction outside of the loop (safe).
4944 // - By further instructions outside of the loop (not safe).
4945 // - By an instruction that is not part of the reduction (not safe).
4947 // * An instruction type other than PHI or the reduction operation.
4948 // * A PHI in the header other than the initial PHI.
4949 while (!Worklist.empty()) {
4950 Instruction *Cur = Worklist.back();
4951 Worklist.pop_back();
4954 // If the instruction has no users then this is a broken chain and can't be
4955 // a reduction variable.
4956 if (Cur->use_empty())
4959 bool IsAPhi = isa<PHINode>(Cur);
4961 // A header PHI use other than the original PHI.
4962 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4965 // Reductions of instructions such as Div, and Sub is only possible if the
4966 // LHS is the reduction variable.
4967 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4968 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4969 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4972 // Any reduction instruction must be of one of the allowed kinds.
4973 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4974 if (!ReduxDesc.IsReduction)
4977 // A reduction operation must only have one use of the reduction value.
4978 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4979 hasMultipleUsesOf(Cur, VisitedInsts))
4982 // All inputs to a PHI node must be a reduction value.
4983 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4986 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4987 isa<SelectInst>(Cur)))
4988 ++NumCmpSelectPatternInst;
4989 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4990 isa<SelectInst>(Cur)))
4991 ++NumCmpSelectPatternInst;
4993 // Check whether we found a reduction operator.
4994 FoundReduxOp |= !IsAPhi;
4996 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4997 // onto the stack. This way we are going to have seen all inputs to PHI
4998 // nodes once we get to them.
4999 SmallVector<Instruction *, 8> NonPHIs;
5000 SmallVector<Instruction *, 8> PHIs;
5001 for (User *U : Cur->users()) {
5002 Instruction *UI = cast<Instruction>(U);
5004 // Check if we found the exit user.
5005 BasicBlock *Parent = UI->getParent();
5006 if (!TheLoop->contains(Parent)) {
5007 // Exit if you find multiple outside users or if the header phi node is
5008 // being used. In this case the user uses the value of the previous
5009 // iteration, in which case we would loose "VF-1" iterations of the
5010 // reduction operation if we vectorize.
5011 if (ExitInstruction != nullptr || Cur == Phi)
5014 // The instruction used by an outside user must be the last instruction
5015 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5016 // operations on the value.
5017 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5020 ExitInstruction = Cur;
5024 // Process instructions only once (termination). Each reduction cycle
5025 // value must only be used once, except by phi nodes and min/max
5026 // reductions which are represented as a cmp followed by a select.
5027 ReductionInstDesc IgnoredVal(false, nullptr);
5028 if (VisitedInsts.insert(UI)) {
5029 if (isa<PHINode>(UI))
5032 NonPHIs.push_back(UI);
5033 } else if (!isa<PHINode>(UI) &&
5034 ((!isa<FCmpInst>(UI) &&
5035 !isa<ICmpInst>(UI) &&
5036 !isa<SelectInst>(UI)) ||
5037 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5040 // Remember that we completed the cycle.
5042 FoundStartPHI = true;
5044 Worklist.append(PHIs.begin(), PHIs.end());
5045 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5048 // This means we have seen one but not the other instruction of the
5049 // pattern or more than just a select and cmp.
5050 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5051 NumCmpSelectPatternInst != 2)
5054 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5057 // We found a reduction var if we have reached the original phi node and we
5058 // only have a single instruction with out-of-loop users.
5060 // This instruction is allowed to have out-of-loop users.
5061 AllowedExit.insert(ExitInstruction);
5063 // Save the description of this reduction variable.
5064 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5065 ReduxDesc.MinMaxKind);
5066 Reductions[Phi] = RD;
5067 // We've ended the cycle. This is a reduction variable if we have an
5068 // outside user and it has a binary op.
5073 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5074 /// pattern corresponding to a min(X, Y) or max(X, Y).
5075 LoopVectorizationLegality::ReductionInstDesc
5076 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5077 ReductionInstDesc &Prev) {
5079 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5080 "Expect a select instruction");
5081 Instruction *Cmp = nullptr;
5082 SelectInst *Select = nullptr;
5084 // We must handle the select(cmp()) as a single instruction. Advance to the
5086 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5087 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5088 return ReductionInstDesc(false, I);
5089 return ReductionInstDesc(Select, Prev.MinMaxKind);
5092 // Only handle single use cases for now.
5093 if (!(Select = dyn_cast<SelectInst>(I)))
5094 return ReductionInstDesc(false, I);
5095 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5096 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5097 return ReductionInstDesc(false, I);
5098 if (!Cmp->hasOneUse())
5099 return ReductionInstDesc(false, I);
5104 // Look for a min/max pattern.
5105 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5106 return ReductionInstDesc(Select, MRK_UIntMin);
5107 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5108 return ReductionInstDesc(Select, MRK_UIntMax);
5109 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5110 return ReductionInstDesc(Select, MRK_SIntMax);
5111 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5112 return ReductionInstDesc(Select, MRK_SIntMin);
5113 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5114 return ReductionInstDesc(Select, MRK_FloatMin);
5115 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5116 return ReductionInstDesc(Select, MRK_FloatMax);
5117 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5118 return ReductionInstDesc(Select, MRK_FloatMin);
5119 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5120 return ReductionInstDesc(Select, MRK_FloatMax);
5122 return ReductionInstDesc(false, I);
5125 LoopVectorizationLegality::ReductionInstDesc
5126 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5128 ReductionInstDesc &Prev) {
5129 bool FP = I->getType()->isFloatingPointTy();
5130 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
5131 switch (I->getOpcode()) {
5133 return ReductionInstDesc(false, I);
5134 case Instruction::PHI:
5135 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5136 Kind != RK_FloatMinMax))
5137 return ReductionInstDesc(false, I);
5138 return ReductionInstDesc(I, Prev.MinMaxKind);
5139 case Instruction::Sub:
5140 case Instruction::Add:
5141 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5142 case Instruction::Mul:
5143 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5144 case Instruction::And:
5145 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5146 case Instruction::Or:
5147 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5148 case Instruction::Xor:
5149 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5150 case Instruction::FMul:
5151 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5152 case Instruction::FAdd:
5153 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5154 case Instruction::FCmp:
5155 case Instruction::ICmp:
5156 case Instruction::Select:
5157 if (Kind != RK_IntegerMinMax &&
5158 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5159 return ReductionInstDesc(false, I);
5160 return isMinMaxSelectCmpPattern(I, Prev);
5164 LoopVectorizationLegality::InductionKind
5165 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5166 Type *PhiTy = Phi->getType();
5167 // We only handle integer and pointer inductions variables.
5168 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5169 return IK_NoInduction;
5171 // Check that the PHI is consecutive.
5172 const SCEV *PhiScev = SE->getSCEV(Phi);
5173 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5175 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5176 return IK_NoInduction;
5178 const SCEV *Step = AR->getStepRecurrence(*SE);
5180 // Integer inductions need to have a stride of one.
5181 if (PhiTy->isIntegerTy()) {
5183 return IK_IntInduction;
5184 if (Step->isAllOnesValue())
5185 return IK_ReverseIntInduction;
5186 return IK_NoInduction;
5189 // Calculate the pointer stride and check if it is consecutive.
5190 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5192 return IK_NoInduction;
5194 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5195 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5196 if (C->getValue()->equalsInt(Size))
5197 return IK_PtrInduction;
5198 else if (C->getValue()->equalsInt(0 - Size))
5199 return IK_ReversePtrInduction;
5201 return IK_NoInduction;
5204 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5205 Value *In0 = const_cast<Value*>(V);
5206 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5210 return Inductions.count(PN);
5213 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5214 assert(TheLoop->contains(BB) && "Unknown block used");
5216 // Blocks that do not dominate the latch need predication.
5217 BasicBlock* Latch = TheLoop->getLoopLatch();
5218 return !DT->dominates(BB, Latch);
5221 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5222 SmallPtrSet<Value *, 8>& SafePtrs) {
5223 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5224 // We might be able to hoist the load.
5225 if (it->mayReadFromMemory()) {
5226 LoadInst *LI = dyn_cast<LoadInst>(it);
5227 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5231 // We don't predicate stores at the moment.
5232 if (it->mayWriteToMemory()) {
5233 StoreInst *SI = dyn_cast<StoreInst>(it);
5234 // We only support predication of stores in basic blocks with one
5236 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5237 !SafePtrs.count(SI->getPointerOperand()) ||
5238 !SI->getParent()->getSinglePredecessor())
5244 // Check that we don't have a constant expression that can trap as operand.
5245 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5247 if (Constant *C = dyn_cast<Constant>(*OI))
5252 // The instructions below can trap.
5253 switch (it->getOpcode()) {
5255 case Instruction::UDiv:
5256 case Instruction::SDiv:
5257 case Instruction::URem:
5258 case Instruction::SRem:
5266 LoopVectorizationCostModel::VectorizationFactor
5267 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5269 bool ForceVectorization) {
5270 // Width 1 means no vectorize
5271 VectorizationFactor Factor = { 1U, 0U };
5272 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5273 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5277 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5278 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5282 // Find the trip count.
5283 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5284 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5286 unsigned WidestType = getWidestType();
5287 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5288 unsigned MaxSafeDepDist = -1U;
5289 if (Legal->getMaxSafeDepDistBytes() != -1U)
5290 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5291 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5292 WidestRegister : MaxSafeDepDist);
5293 unsigned MaxVectorSize = WidestRegister / WidestType;
5294 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5295 DEBUG(dbgs() << "LV: The Widest register is: "
5296 << WidestRegister << " bits.\n");
5298 if (MaxVectorSize == 0) {
5299 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5303 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5304 " into one vector!");
5306 unsigned VF = MaxVectorSize;
5308 // If we optimize the program for size, avoid creating the tail loop.
5310 // If we are unable to calculate the trip count then don't try to vectorize.
5312 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5316 // Find the maximum SIMD width that can fit within the trip count.
5317 VF = TC % MaxVectorSize;
5322 // If the trip count that we found modulo the vectorization factor is not
5323 // zero then we require a tail.
5325 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5331 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5332 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5334 Factor.Width = UserVF;
5338 float Cost = expectedCost(1);
5340 const float ScalarCost = Cost;
5343 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5345 // Ignore scalar width, because the user explicitly wants vectorization.
5346 if (ForceVectorization && VF > 1) {
5348 Cost = expectedCost(Width) / (float)Width;
5351 for (unsigned i=2; i <= VF; i*=2) {
5352 // Notice that the vector loop needs to be executed less times, so
5353 // we need to divide the cost of the vector loops by the width of
5354 // the vector elements.
5355 float VectorCost = expectedCost(i) / (float)i;
5356 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5357 (int)VectorCost << ".\n");
5358 if (VectorCost < Cost) {
5364 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5365 << "LV: Vectorization seems to be not beneficial, "
5366 << "but was forced by a user.\n");
5367 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5368 Factor.Width = Width;
5369 Factor.Cost = Width * Cost;
5373 unsigned LoopVectorizationCostModel::getWidestType() {
5374 unsigned MaxWidth = 8;
5377 for (Loop::block_iterator bb = TheLoop->block_begin(),
5378 be = TheLoop->block_end(); bb != be; ++bb) {
5379 BasicBlock *BB = *bb;
5381 // For each instruction in the loop.
5382 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5383 Type *T = it->getType();
5385 // Only examine Loads, Stores and PHINodes.
5386 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5389 // Examine PHI nodes that are reduction variables.
5390 if (PHINode *PN = dyn_cast<PHINode>(it))
5391 if (!Legal->getReductionVars()->count(PN))
5394 // Examine the stored values.
5395 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5396 T = ST->getValueOperand()->getType();
5398 // Ignore loaded pointer types and stored pointer types that are not
5399 // consecutive. However, we do want to take consecutive stores/loads of
5400 // pointer vectors into account.
5401 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5404 MaxWidth = std::max(MaxWidth,
5405 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5413 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5416 unsigned LoopCost) {
5418 // -- The unroll heuristics --
5419 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5420 // There are many micro-architectural considerations that we can't predict
5421 // at this level. For example frontend pressure (on decode or fetch) due to
5422 // code size, or the number and capabilities of the execution ports.
5424 // We use the following heuristics to select the unroll factor:
5425 // 1. If the code has reductions the we unroll in order to break the cross
5426 // iteration dependency.
5427 // 2. If the loop is really small then we unroll in order to reduce the loop
5429 // 3. We don't unroll if we think that we will spill registers to memory due
5430 // to the increased register pressure.
5432 // Use the user preference, unless 'auto' is selected.
5436 // When we optimize for size we don't unroll.
5440 // We used the distance for the unroll factor.
5441 if (Legal->getMaxSafeDepDistBytes() != -1U)
5444 // Do not unroll loops with a relatively small trip count.
5445 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5446 TheLoop->getLoopLatch());
5447 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5450 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5451 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5455 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5456 TargetNumRegisters = ForceTargetNumScalarRegs;
5458 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5459 TargetNumRegisters = ForceTargetNumVectorRegs;
5462 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5463 // We divide by these constants so assume that we have at least one
5464 // instruction that uses at least one register.
5465 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5466 R.NumInstructions = std::max(R.NumInstructions, 1U);
5468 // We calculate the unroll factor using the following formula.
5469 // Subtract the number of loop invariants from the number of available
5470 // registers. These registers are used by all of the unrolled instances.
5471 // Next, divide the remaining registers by the number of registers that is
5472 // required by the loop, in order to estimate how many parallel instances
5473 // fit without causing spills. All of this is rounded down if necessary to be
5474 // a power of two. We want power of two unroll factors to simplify any
5475 // addressing operations or alignment considerations.
5476 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5479 // Don't count the induction variable as unrolled.
5480 if (EnableIndVarRegisterHeur)
5481 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5482 std::max(1U, (R.MaxLocalUsers - 1)));
5484 // Clamp the unroll factor ranges to reasonable factors.
5485 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5487 // Check if the user has overridden the unroll max.
5489 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5490 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5492 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5493 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5496 // If we did not calculate the cost for VF (because the user selected the VF)
5497 // then we calculate the cost of VF here.
5499 LoopCost = expectedCost(VF);
5501 // Clamp the calculated UF to be between the 1 and the max unroll factor
5502 // that the target allows.
5503 if (UF > MaxUnrollSize)
5508 // Unroll if we vectorized this loop and there is a reduction that could
5509 // benefit from unrolling.
5510 if (VF > 1 && Legal->getReductionVars()->size()) {
5511 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5515 // Note that if we've already vectorized the loop we will have done the
5516 // runtime check and so unrolling won't require further checks.
5517 bool UnrollingRequiresRuntimePointerCheck =
5518 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5520 // We want to unroll small loops in order to reduce the loop overhead and
5521 // potentially expose ILP opportunities.
5522 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5523 if (!UnrollingRequiresRuntimePointerCheck &&
5524 LoopCost < SmallLoopCost) {
5525 // We assume that the cost overhead is 1 and we use the cost model
5526 // to estimate the cost of the loop and unroll until the cost of the
5527 // loop overhead is about 5% of the cost of the loop.
5528 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5530 // Unroll until store/load ports (estimated by max unroll factor) are
5532 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5533 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5535 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5536 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5537 return std::max(StoresUF, LoadsUF);
5540 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5544 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5548 LoopVectorizationCostModel::RegisterUsage
5549 LoopVectorizationCostModel::calculateRegisterUsage() {
5550 // This function calculates the register usage by measuring the highest number
5551 // of values that are alive at a single location. Obviously, this is a very
5552 // rough estimation. We scan the loop in a topological order in order and
5553 // assign a number to each instruction. We use RPO to ensure that defs are
5554 // met before their users. We assume that each instruction that has in-loop
5555 // users starts an interval. We record every time that an in-loop value is
5556 // used, so we have a list of the first and last occurrences of each
5557 // instruction. Next, we transpose this data structure into a multi map that
5558 // holds the list of intervals that *end* at a specific location. This multi
5559 // map allows us to perform a linear search. We scan the instructions linearly
5560 // and record each time that a new interval starts, by placing it in a set.
5561 // If we find this value in the multi-map then we remove it from the set.
5562 // The max register usage is the maximum size of the set.
5563 // We also search for instructions that are defined outside the loop, but are
5564 // used inside the loop. We need this number separately from the max-interval
5565 // usage number because when we unroll, loop-invariant values do not take
5567 LoopBlocksDFS DFS(TheLoop);
5571 R.NumInstructions = 0;
5573 // Each 'key' in the map opens a new interval. The values
5574 // of the map are the index of the 'last seen' usage of the
5575 // instruction that is the key.
5576 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5577 // Maps instruction to its index.
5578 DenseMap<unsigned, Instruction*> IdxToInstr;
5579 // Marks the end of each interval.
5580 IntervalMap EndPoint;
5581 // Saves the list of instruction indices that are used in the loop.
5582 SmallSet<Instruction*, 8> Ends;
5583 // Saves the list of values that are used in the loop but are
5584 // defined outside the loop, such as arguments and constants.
5585 SmallPtrSet<Value*, 8> LoopInvariants;
5588 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5589 be = DFS.endRPO(); bb != be; ++bb) {
5590 R.NumInstructions += (*bb)->size();
5591 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5593 Instruction *I = it;
5594 IdxToInstr[Index++] = I;
5596 // Save the end location of each USE.
5597 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5598 Value *U = I->getOperand(i);
5599 Instruction *Instr = dyn_cast<Instruction>(U);
5601 // Ignore non-instruction values such as arguments, constants, etc.
5602 if (!Instr) continue;
5604 // If this instruction is outside the loop then record it and continue.
5605 if (!TheLoop->contains(Instr)) {
5606 LoopInvariants.insert(Instr);
5610 // Overwrite previous end points.
5611 EndPoint[Instr] = Index;
5617 // Saves the list of intervals that end with the index in 'key'.
5618 typedef SmallVector<Instruction*, 2> InstrList;
5619 DenseMap<unsigned, InstrList> TransposeEnds;
5621 // Transpose the EndPoints to a list of values that end at each index.
5622 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5624 TransposeEnds[it->second].push_back(it->first);
5626 SmallSet<Instruction*, 8> OpenIntervals;
5627 unsigned MaxUsage = 0;
5630 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5631 for (unsigned int i = 0; i < Index; ++i) {
5632 Instruction *I = IdxToInstr[i];
5633 // Ignore instructions that are never used within the loop.
5634 if (!Ends.count(I)) continue;
5636 // Remove all of the instructions that end at this location.
5637 InstrList &List = TransposeEnds[i];
5638 for (unsigned int j=0, e = List.size(); j < e; ++j)
5639 OpenIntervals.erase(List[j]);
5641 // Count the number of live interals.
5642 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5644 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5645 OpenIntervals.size() << '\n');
5647 // Add the current instruction to the list of open intervals.
5648 OpenIntervals.insert(I);
5651 unsigned Invariant = LoopInvariants.size();
5652 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5653 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5654 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5656 R.LoopInvariantRegs = Invariant;
5657 R.MaxLocalUsers = MaxUsage;
5661 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5665 for (Loop::block_iterator bb = TheLoop->block_begin(),
5666 be = TheLoop->block_end(); bb != be; ++bb) {
5667 unsigned BlockCost = 0;
5668 BasicBlock *BB = *bb;
5670 // For each instruction in the old loop.
5671 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5672 // Skip dbg intrinsics.
5673 if (isa<DbgInfoIntrinsic>(it))
5676 unsigned C = getInstructionCost(it, VF);
5678 // Check if we should override the cost.
5679 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5680 C = ForceTargetInstructionCost;
5683 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5684 VF << " For instruction: " << *it << '\n');
5687 // We assume that if-converted blocks have a 50% chance of being executed.
5688 // When the code is scalar then some of the blocks are avoided due to CF.
5689 // When the code is vectorized we execute all code paths.
5690 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5699 /// \brief Check whether the address computation for a non-consecutive memory
5700 /// access looks like an unlikely candidate for being merged into the indexing
5703 /// We look for a GEP which has one index that is an induction variable and all
5704 /// other indices are loop invariant. If the stride of this access is also
5705 /// within a small bound we decide that this address computation can likely be
5706 /// merged into the addressing mode.
5707 /// In all other cases, we identify the address computation as complex.
5708 static bool isLikelyComplexAddressComputation(Value *Ptr,
5709 LoopVectorizationLegality *Legal,
5710 ScalarEvolution *SE,
5711 const Loop *TheLoop) {
5712 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5716 // We are looking for a gep with all loop invariant indices except for one
5717 // which should be an induction variable.
5718 unsigned NumOperands = Gep->getNumOperands();
5719 for (unsigned i = 1; i < NumOperands; ++i) {
5720 Value *Opd = Gep->getOperand(i);
5721 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5722 !Legal->isInductionVariable(Opd))
5726 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5727 // can likely be merged into the address computation.
5728 unsigned MaxMergeDistance = 64;
5730 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5734 // Check the step is constant.
5735 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5736 // Calculate the pointer stride and check if it is consecutive.
5737 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5741 const APInt &APStepVal = C->getValue()->getValue();
5743 // Huge step value - give up.
5744 if (APStepVal.getBitWidth() > 64)
5747 int64_t StepVal = APStepVal.getSExtValue();
5749 return StepVal > MaxMergeDistance;
5752 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5753 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5759 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5760 // If we know that this instruction will remain uniform, check the cost of
5761 // the scalar version.
5762 if (Legal->isUniformAfterVectorization(I))
5765 Type *RetTy = I->getType();
5766 Type *VectorTy = ToVectorTy(RetTy, VF);
5768 // TODO: We need to estimate the cost of intrinsic calls.
5769 switch (I->getOpcode()) {
5770 case Instruction::GetElementPtr:
5771 // We mark this instruction as zero-cost because the cost of GEPs in
5772 // vectorized code depends on whether the corresponding memory instruction
5773 // is scalarized or not. Therefore, we handle GEPs with the memory
5774 // instruction cost.
5776 case Instruction::Br: {
5777 return TTI.getCFInstrCost(I->getOpcode());
5779 case Instruction::PHI:
5780 //TODO: IF-converted IFs become selects.
5782 case Instruction::Add:
5783 case Instruction::FAdd:
5784 case Instruction::Sub:
5785 case Instruction::FSub:
5786 case Instruction::Mul:
5787 case Instruction::FMul:
5788 case Instruction::UDiv:
5789 case Instruction::SDiv:
5790 case Instruction::FDiv:
5791 case Instruction::URem:
5792 case Instruction::SRem:
5793 case Instruction::FRem:
5794 case Instruction::Shl:
5795 case Instruction::LShr:
5796 case Instruction::AShr:
5797 case Instruction::And:
5798 case Instruction::Or:
5799 case Instruction::Xor: {
5800 // Since we will replace the stride by 1 the multiplication should go away.
5801 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5803 // Certain instructions can be cheaper to vectorize if they have a constant
5804 // second vector operand. One example of this are shifts on x86.
5805 TargetTransformInfo::OperandValueKind Op1VK =
5806 TargetTransformInfo::OK_AnyValue;
5807 TargetTransformInfo::OperandValueKind Op2VK =
5808 TargetTransformInfo::OK_AnyValue;
5809 Value *Op2 = I->getOperand(1);
5811 // Check for a splat of a constant or for a non uniform vector of constants.
5812 if (isa<ConstantInt>(Op2))
5813 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5814 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5815 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5816 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5817 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5820 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5822 case Instruction::Select: {
5823 SelectInst *SI = cast<SelectInst>(I);
5824 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5825 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5826 Type *CondTy = SI->getCondition()->getType();
5828 CondTy = VectorType::get(CondTy, VF);
5830 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5832 case Instruction::ICmp:
5833 case Instruction::FCmp: {
5834 Type *ValTy = I->getOperand(0)->getType();
5835 VectorTy = ToVectorTy(ValTy, VF);
5836 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5838 case Instruction::Store:
5839 case Instruction::Load: {
5840 StoreInst *SI = dyn_cast<StoreInst>(I);
5841 LoadInst *LI = dyn_cast<LoadInst>(I);
5842 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5844 VectorTy = ToVectorTy(ValTy, VF);
5846 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5847 unsigned AS = SI ? SI->getPointerAddressSpace() :
5848 LI->getPointerAddressSpace();
5849 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5850 // We add the cost of address computation here instead of with the gep
5851 // instruction because only here we know whether the operation is
5854 return TTI.getAddressComputationCost(VectorTy) +
5855 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5857 // Scalarized loads/stores.
5858 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5859 bool Reverse = ConsecutiveStride < 0;
5860 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5861 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5862 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5863 bool IsComplexComputation =
5864 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5866 // The cost of extracting from the value vector and pointer vector.
5867 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5868 for (unsigned i = 0; i < VF; ++i) {
5869 // The cost of extracting the pointer operand.
5870 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5871 // In case of STORE, the cost of ExtractElement from the vector.
5872 // In case of LOAD, the cost of InsertElement into the returned
5874 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5875 Instruction::InsertElement,
5879 // The cost of the scalar loads/stores.
5880 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5881 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5886 // Wide load/stores.
5887 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5888 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5891 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5895 case Instruction::ZExt:
5896 case Instruction::SExt:
5897 case Instruction::FPToUI:
5898 case Instruction::FPToSI:
5899 case Instruction::FPExt:
5900 case Instruction::PtrToInt:
5901 case Instruction::IntToPtr:
5902 case Instruction::SIToFP:
5903 case Instruction::UIToFP:
5904 case Instruction::Trunc:
5905 case Instruction::FPTrunc:
5906 case Instruction::BitCast: {
5907 // We optimize the truncation of induction variable.
5908 // The cost of these is the same as the scalar operation.
5909 if (I->getOpcode() == Instruction::Trunc &&
5910 Legal->isInductionVariable(I->getOperand(0)))
5911 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5912 I->getOperand(0)->getType());
5914 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5915 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5917 case Instruction::Call: {
5918 CallInst *CI = cast<CallInst>(I);
5919 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5920 assert(ID && "Not an intrinsic call!");
5921 Type *RetTy = ToVectorTy(CI->getType(), VF);
5922 SmallVector<Type*, 4> Tys;
5923 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5924 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5925 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5928 // We are scalarizing the instruction. Return the cost of the scalar
5929 // instruction, plus the cost of insert and extract into vector
5930 // elements, times the vector width.
5933 if (!RetTy->isVoidTy() && VF != 1) {
5934 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5936 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5939 // The cost of inserting the results plus extracting each one of the
5941 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5944 // The cost of executing VF copies of the scalar instruction. This opcode
5945 // is unknown. Assume that it is the same as 'mul'.
5946 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5952 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5953 if (Scalar->isVoidTy() || VF == 1)
5955 return VectorType::get(Scalar, VF);
5958 char LoopVectorize::ID = 0;
5959 static const char lv_name[] = "Loop Vectorization";
5960 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5961 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5962 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5963 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5964 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5965 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5966 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5967 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5968 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5969 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5972 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5973 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5977 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5978 // Check for a store.
5979 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5980 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5982 // Check for a load.
5983 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5984 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5990 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5991 bool IfPredicateStore) {
5992 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5993 // Holds vector parameters or scalars, in case of uniform vals.
5994 SmallVector<VectorParts, 4> Params;
5996 setDebugLocFromInst(Builder, Instr);
5998 // Find all of the vectorized parameters.
5999 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6000 Value *SrcOp = Instr->getOperand(op);
6002 // If we are accessing the old induction variable, use the new one.
6003 if (SrcOp == OldInduction) {
6004 Params.push_back(getVectorValue(SrcOp));
6008 // Try using previously calculated values.
6009 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6011 // If the src is an instruction that appeared earlier in the basic block
6012 // then it should already be vectorized.
6013 if (SrcInst && OrigLoop->contains(SrcInst)) {
6014 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6015 // The parameter is a vector value from earlier.
6016 Params.push_back(WidenMap.get(SrcInst));
6018 // The parameter is a scalar from outside the loop. Maybe even a constant.
6019 VectorParts Scalars;
6020 Scalars.append(UF, SrcOp);
6021 Params.push_back(Scalars);
6025 assert(Params.size() == Instr->getNumOperands() &&
6026 "Invalid number of operands");
6028 // Does this instruction return a value ?
6029 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6031 Value *UndefVec = IsVoidRetTy ? nullptr :
6032 UndefValue::get(Instr->getType());
6033 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6034 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6036 Instruction *InsertPt = Builder.GetInsertPoint();
6037 BasicBlock *IfBlock = Builder.GetInsertBlock();
6038 BasicBlock *CondBlock = nullptr;
6041 Loop *VectorLp = nullptr;
6042 if (IfPredicateStore) {
6043 assert(Instr->getParent()->getSinglePredecessor() &&
6044 "Only support single predecessor blocks");
6045 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6046 Instr->getParent());
6047 VectorLp = LI->getLoopFor(IfBlock);
6048 assert(VectorLp && "Must have a loop for this block");
6051 // For each vector unroll 'part':
6052 for (unsigned Part = 0; Part < UF; ++Part) {
6053 // For each scalar that we create:
6055 // Start an "if (pred) a[i] = ..." block.
6056 Value *Cmp = nullptr;
6057 if (IfPredicateStore) {
6058 if (Cond[Part]->getType()->isVectorTy())
6060 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6061 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6062 ConstantInt::get(Cond[Part]->getType(), 1));
6063 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6064 LoopVectorBody.push_back(CondBlock);
6065 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6066 // Update Builder with newly created basic block.
6067 Builder.SetInsertPoint(InsertPt);
6070 Instruction *Cloned = Instr->clone();
6072 Cloned->setName(Instr->getName() + ".cloned");
6073 // Replace the operands of the cloned instructions with extracted scalars.
6074 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6075 Value *Op = Params[op][Part];
6076 Cloned->setOperand(op, Op);
6079 // Place the cloned scalar in the new loop.
6080 Builder.Insert(Cloned);
6082 // If the original scalar returns a value we need to place it in a vector
6083 // so that future users will be able to use it.
6085 VecResults[Part] = Cloned;
6088 if (IfPredicateStore) {
6089 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6090 LoopVectorBody.push_back(NewIfBlock);
6091 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6092 Builder.SetInsertPoint(InsertPt);
6093 Instruction *OldBr = IfBlock->getTerminator();
6094 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6095 OldBr->eraseFromParent();
6096 IfBlock = NewIfBlock;
6101 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6102 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6103 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6105 return scalarizeInstruction(Instr, IfPredicateStore);
6108 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6112 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6116 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6118 // When unrolling and the VF is 1, we only need to add a simple scalar.
6119 Type *ITy = Val->getType();
6120 assert(!ITy->isVectorTy() && "Val must be a scalar");
6121 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6122 return Builder.CreateAdd(Val, C, "induction");