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 vectorizer metadata prefix.
982 static StringRef Prefix() { return "llvm.loop.vectorize."; }
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));
1004 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
1005 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
1007 MDNode *NewLoopID = MDNode::get(Context, Vals);
1008 // Set operand 0 to refer to the loop id itself.
1009 NewLoopID->replaceOperandWith(0, NewLoopID);
1011 L->setLoopID(NewLoopID);
1013 LoopID->replaceAllUsesWith(NewLoopID);
1018 std::string emitRemark() const {
1020 R << "vectorization ";
1022 case LoopVectorizeHints::FK_Disabled:
1023 R << "is explicitly disabled";
1025 case LoopVectorizeHints::FK_Enabled:
1026 R << "is explicitly enabled";
1027 if (Width != 0 && Unroll != 0)
1028 R << " with width " << Width << " and interleave count " << Unroll;
1029 else if (Width != 0)
1030 R << " with width " << Width;
1031 else if (Unroll != 0)
1032 R << " with interleave count " << Unroll;
1034 case LoopVectorizeHints::FK_Undefined:
1035 R << "was not specified";
1041 unsigned getWidth() const { return Width; }
1042 unsigned getUnroll() const { return Unroll; }
1043 enum ForceKind getForce() const { return Force; }
1044 MDNode *getLoopID() const { return LoopID; }
1047 /// Find hints specified in the loop metadata.
1048 void getHints(const Loop *L) {
1052 // First operand should refer to the loop id itself.
1053 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1054 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1056 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1057 const MDString *S = nullptr;
1058 SmallVector<Value*, 4> Args;
1060 // The expected hint is either a MDString or a MDNode with the first
1061 // operand a MDString.
1062 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1063 if (!MD || MD->getNumOperands() == 0)
1065 S = dyn_cast<MDString>(MD->getOperand(0));
1066 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1067 Args.push_back(MD->getOperand(i));
1069 S = dyn_cast<MDString>(LoopID->getOperand(i));
1070 assert(Args.size() == 0 && "too many arguments for MDString");
1076 // Check if the hint starts with the vectorizer prefix.
1077 StringRef Hint = S->getString();
1078 if (!Hint.startswith(Prefix()))
1080 // Remove the prefix.
1081 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1083 if (Args.size() == 1)
1084 getHint(Hint, Args[0]);
1088 // Check string hint with one operand.
1089 void getHint(StringRef Hint, Value *Arg) {
1090 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1092 unsigned Val = C->getZExtValue();
1094 if (Hint == "width") {
1095 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1098 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1099 } else if (Hint == "unroll") {
1100 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1103 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1104 } else if (Hint == "enable") {
1105 if (C->getBitWidth() == 1)
1106 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1107 : LoopVectorizeHints::FK_Disabled;
1109 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1111 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1115 /// Vectorization width.
1117 /// Vectorization unroll factor.
1119 /// Vectorization forced
1120 enum ForceKind Force;
1125 static void emitMissedWarning(Function *F, Loop *L,
1126 const LoopVectorizeHints &LH) {
1127 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1128 L->getStartLoc(), LH.emitRemark());
1130 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1131 if (LH.getWidth() != 1)
1132 emitLoopVectorizeWarning(
1133 F->getContext(), *F, L->getStartLoc(),
1134 "failed explicitly specified loop vectorization");
1135 else if (LH.getUnroll() != 1)
1136 emitLoopInterleaveWarning(
1137 F->getContext(), *F, L->getStartLoc(),
1138 "failed explicitly specified loop interleaving");
1142 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1144 return V.push_back(&L);
1146 for (Loop *InnerL : L)
1147 addInnerLoop(*InnerL, V);
1150 /// The LoopVectorize Pass.
1151 struct LoopVectorize : public FunctionPass {
1152 /// Pass identification, replacement for typeid
1155 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1157 DisableUnrolling(NoUnrolling),
1158 AlwaysVectorize(AlwaysVectorize) {
1159 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1162 ScalarEvolution *SE;
1163 const DataLayout *DL;
1165 TargetTransformInfo *TTI;
1167 BlockFrequencyInfo *BFI;
1168 TargetLibraryInfo *TLI;
1170 bool DisableUnrolling;
1171 bool AlwaysVectorize;
1173 BlockFrequency ColdEntryFreq;
1175 bool runOnFunction(Function &F) override {
1176 SE = &getAnalysis<ScalarEvolution>();
1177 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1178 DL = DLP ? &DLP->getDataLayout() : nullptr;
1179 LI = &getAnalysis<LoopInfo>();
1180 TTI = &getAnalysis<TargetTransformInfo>();
1181 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1182 BFI = &getAnalysis<BlockFrequencyInfo>();
1183 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1184 AA = &getAnalysis<AliasAnalysis>();
1186 // Compute some weights outside of the loop over the loops. Compute this
1187 // using a BranchProbability to re-use its scaling math.
1188 const BranchProbability ColdProb(1, 5); // 20%
1189 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1191 // If the target claims to have no vector registers don't attempt
1193 if (!TTI->getNumberOfRegisters(true))
1197 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1198 << ": Missing data layout\n");
1202 // Build up a worklist of inner-loops to vectorize. This is necessary as
1203 // the act of vectorizing or partially unrolling a loop creates new loops
1204 // and can invalidate iterators across the loops.
1205 SmallVector<Loop *, 8> Worklist;
1208 addInnerLoop(*L, Worklist);
1210 LoopsAnalyzed += Worklist.size();
1212 // Now walk the identified inner loops.
1213 bool Changed = false;
1214 while (!Worklist.empty())
1215 Changed |= processLoop(Worklist.pop_back_val());
1217 // Process each loop nest in the function.
1221 bool processLoop(Loop *L) {
1222 assert(L->empty() && "Only process inner loops.");
1225 const std::string DebugLocStr = getDebugLocString(L);
1228 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1229 << L->getHeader()->getParent()->getName() << "\" from "
1230 << DebugLocStr << "\n");
1232 LoopVectorizeHints Hints(L, DisableUnrolling);
1234 DEBUG(dbgs() << "LV: Loop hints:"
1236 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1238 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1240 : "?")) << " width=" << Hints.getWidth()
1241 << " unroll=" << Hints.getUnroll() << "\n");
1243 // Function containing loop
1244 Function *F = L->getHeader()->getParent();
1246 // Looking at the diagnostic output is the only way to determine if a loop
1247 // was vectorized (other than looking at the IR or machine code), so it
1248 // is important to generate an optimization remark for each loop. Most of
1249 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1250 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1251 // less verbose reporting vectorized loops and unvectorized loops that may
1252 // benefit from vectorization, respectively.
1254 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1255 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1256 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1257 L->getStartLoc(), Hints.emitRemark());
1261 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1262 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1263 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1264 L->getStartLoc(), Hints.emitRemark());
1268 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1269 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1270 emitOptimizationRemarkAnalysis(
1271 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1272 "loop not vectorized: vector width and interleave count are "
1273 "explicitly set to 1");
1277 // Check the loop for a trip count threshold:
1278 // do not vectorize loops with a tiny trip count.
1279 BasicBlock *Latch = L->getLoopLatch();
1280 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1281 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1282 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1283 << "This loop is not worth vectorizing.");
1284 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1285 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1287 DEBUG(dbgs() << "\n");
1288 emitOptimizationRemarkAnalysis(
1289 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1290 "vectorization is not beneficial and is not explicitly forced");
1295 // Check if it is legal to vectorize the loop.
1296 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1297 if (!LVL.canVectorize()) {
1298 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1299 emitMissedWarning(F, L, Hints);
1303 // Use the cost model.
1304 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1306 // Check the function attributes to find out if this function should be
1307 // optimized for size.
1308 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1309 F->hasFnAttribute(Attribute::OptimizeForSize);
1311 // Compute the weighted frequency of this loop being executed and see if it
1312 // is less than 20% of the function entry baseline frequency. Note that we
1313 // always have a canonical loop here because we think we *can* vectoriez.
1314 // FIXME: This is hidden behind a flag due to pervasive problems with
1315 // exactly what block frequency models.
1316 if (LoopVectorizeWithBlockFrequency) {
1317 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1318 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1319 LoopEntryFreq < ColdEntryFreq)
1323 // Check the function attributes to see if implicit floats are allowed.a
1324 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1325 // an integer loop and the vector instructions selected are purely integer
1326 // vector instructions?
1327 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1328 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1329 "attribute is used.\n");
1330 emitOptimizationRemarkAnalysis(
1331 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1332 "loop not vectorized due to NoImplicitFloat attribute");
1333 emitMissedWarning(F, L, Hints);
1337 // Select the optimal vectorization factor.
1338 const LoopVectorizationCostModel::VectorizationFactor VF =
1339 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1341 LoopVectorizeHints::FK_Enabled);
1343 // Select the unroll factor.
1345 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1347 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1348 << DebugLocStr << '\n');
1349 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1351 if (VF.Width == 1) {
1352 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1355 emitOptimizationRemarkAnalysis(
1356 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1357 "not beneficial to vectorize and user disabled interleaving");
1360 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1362 // Report the unrolling decision.
1363 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1364 Twine("unrolled with interleaving factor " +
1366 " (vectorization not beneficial)"));
1368 // We decided not to vectorize, but we may want to unroll.
1370 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1371 Unroller.vectorize(&LVL);
1373 // If we decided that it is *legal* to vectorize the loop then do it.
1374 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1378 // Report the vectorization decision.
1379 emitOptimizationRemark(
1380 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1381 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1382 ", unrolling interleave factor: " + Twine(UF) + ")");
1385 // Mark the loop as already vectorized to avoid vectorizing again.
1386 Hints.setAlreadyVectorized(L);
1388 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1392 void getAnalysisUsage(AnalysisUsage &AU) const override {
1393 AU.addRequiredID(LoopSimplifyID);
1394 AU.addRequiredID(LCSSAID);
1395 AU.addRequired<BlockFrequencyInfo>();
1396 AU.addRequired<DominatorTreeWrapperPass>();
1397 AU.addRequired<LoopInfo>();
1398 AU.addRequired<ScalarEvolution>();
1399 AU.addRequired<TargetTransformInfo>();
1400 AU.addRequired<AliasAnalysis>();
1401 AU.addPreserved<LoopInfo>();
1402 AU.addPreserved<DominatorTreeWrapperPass>();
1403 AU.addPreserved<AliasAnalysis>();
1408 } // end anonymous namespace
1410 //===----------------------------------------------------------------------===//
1411 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1412 // LoopVectorizationCostModel.
1413 //===----------------------------------------------------------------------===//
1415 static Value *stripIntegerCast(Value *V) {
1416 if (CastInst *CI = dyn_cast<CastInst>(V))
1417 if (CI->getOperand(0)->getType()->isIntegerTy())
1418 return CI->getOperand(0);
1422 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1424 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1426 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1427 ValueToValueMap &PtrToStride,
1428 Value *Ptr, Value *OrigPtr = nullptr) {
1430 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1432 // If there is an entry in the map return the SCEV of the pointer with the
1433 // symbolic stride replaced by one.
1434 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1435 if (SI != PtrToStride.end()) {
1436 Value *StrideVal = SI->second;
1439 StrideVal = stripIntegerCast(StrideVal);
1441 // Replace symbolic stride by one.
1442 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1443 ValueToValueMap RewriteMap;
1444 RewriteMap[StrideVal] = One;
1447 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1448 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1453 // Otherwise, just return the SCEV of the original pointer.
1454 return SE->getSCEV(Ptr);
1457 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1458 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1459 unsigned ASId, ValueToValueMap &Strides) {
1460 // Get the stride replaced scev.
1461 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1462 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1463 assert(AR && "Invalid addrec expression");
1464 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1465 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1466 Pointers.push_back(Ptr);
1467 Starts.push_back(AR->getStart());
1468 Ends.push_back(ScEnd);
1469 IsWritePtr.push_back(WritePtr);
1470 DependencySetId.push_back(DepSetId);
1471 AliasSetId.push_back(ASId);
1474 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1475 // We need to place the broadcast of invariant variables outside the loop.
1476 Instruction *Instr = dyn_cast<Instruction>(V);
1478 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1479 Instr->getParent()) != LoopVectorBody.end());
1480 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1482 // Place the code for broadcasting invariant variables in the new preheader.
1483 IRBuilder<>::InsertPointGuard Guard(Builder);
1485 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1487 // Broadcast the scalar into all locations in the vector.
1488 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1493 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1495 assert(Val->getType()->isVectorTy() && "Must be a vector");
1496 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1497 "Elem must be an integer");
1498 // Create the types.
1499 Type *ITy = Val->getType()->getScalarType();
1500 VectorType *Ty = cast<VectorType>(Val->getType());
1501 int VLen = Ty->getNumElements();
1502 SmallVector<Constant*, 8> Indices;
1504 // Create a vector of consecutive numbers from zero to VF.
1505 for (int i = 0; i < VLen; ++i) {
1506 int64_t Idx = Negate ? (-i) : i;
1507 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1510 // Add the consecutive indices to the vector value.
1511 Constant *Cv = ConstantVector::get(Indices);
1512 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1513 return Builder.CreateAdd(Val, Cv, "induction");
1516 /// \brief Find the operand of the GEP that should be checked for consecutive
1517 /// stores. This ignores trailing indices that have no effect on the final
1519 static unsigned getGEPInductionOperand(const DataLayout *DL,
1520 const GetElementPtrInst *Gep) {
1521 unsigned LastOperand = Gep->getNumOperands() - 1;
1522 unsigned GEPAllocSize = DL->getTypeAllocSize(
1523 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1525 // Walk backwards and try to peel off zeros.
1526 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1527 // Find the type we're currently indexing into.
1528 gep_type_iterator GEPTI = gep_type_begin(Gep);
1529 std::advance(GEPTI, LastOperand - 1);
1531 // If it's a type with the same allocation size as the result of the GEP we
1532 // can peel off the zero index.
1533 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1541 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1542 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1543 // Make sure that the pointer does not point to structs.
1544 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1547 // If this value is a pointer induction variable we know it is consecutive.
1548 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1549 if (Phi && Inductions.count(Phi)) {
1550 InductionInfo II = Inductions[Phi];
1551 if (IK_PtrInduction == II.IK)
1553 else if (IK_ReversePtrInduction == II.IK)
1557 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1561 unsigned NumOperands = Gep->getNumOperands();
1562 Value *GpPtr = Gep->getPointerOperand();
1563 // If this GEP value is a consecutive pointer induction variable and all of
1564 // the indices are constant then we know it is consecutive. We can
1565 Phi = dyn_cast<PHINode>(GpPtr);
1566 if (Phi && Inductions.count(Phi)) {
1568 // Make sure that the pointer does not point to structs.
1569 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1570 if (GepPtrType->getElementType()->isAggregateType())
1573 // Make sure that all of the index operands are loop invariant.
1574 for (unsigned i = 1; i < NumOperands; ++i)
1575 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1578 InductionInfo II = Inductions[Phi];
1579 if (IK_PtrInduction == II.IK)
1581 else if (IK_ReversePtrInduction == II.IK)
1585 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1587 // Check that all of the gep indices are uniform except for our induction
1589 for (unsigned i = 0; i != NumOperands; ++i)
1590 if (i != InductionOperand &&
1591 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1594 // We can emit wide load/stores only if the last non-zero index is the
1595 // induction variable.
1596 const SCEV *Last = nullptr;
1597 if (!Strides.count(Gep))
1598 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1600 // Because of the multiplication by a stride we can have a s/zext cast.
1601 // We are going to replace this stride by 1 so the cast is safe to ignore.
1603 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1604 // %0 = trunc i64 %indvars.iv to i32
1605 // %mul = mul i32 %0, %Stride1
1606 // %idxprom = zext i32 %mul to i64 << Safe cast.
1607 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1609 Last = replaceSymbolicStrideSCEV(SE, Strides,
1610 Gep->getOperand(InductionOperand), Gep);
1611 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1613 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1617 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1618 const SCEV *Step = AR->getStepRecurrence(*SE);
1620 // The memory is consecutive because the last index is consecutive
1621 // and all other indices are loop invariant.
1624 if (Step->isAllOnesValue())
1631 bool LoopVectorizationLegality::isUniform(Value *V) {
1632 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1635 InnerLoopVectorizer::VectorParts&
1636 InnerLoopVectorizer::getVectorValue(Value *V) {
1637 assert(V != Induction && "The new induction variable should not be used.");
1638 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1640 // If we have a stride that is replaced by one, do it here.
1641 if (Legal->hasStride(V))
1642 V = ConstantInt::get(V->getType(), 1);
1644 // If we have this scalar in the map, return it.
1645 if (WidenMap.has(V))
1646 return WidenMap.get(V);
1648 // If this scalar is unknown, assume that it is a constant or that it is
1649 // loop invariant. Broadcast V and save the value for future uses.
1650 Value *B = getBroadcastInstrs(V);
1651 return WidenMap.splat(V, B);
1654 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1655 assert(Vec->getType()->isVectorTy() && "Invalid type");
1656 SmallVector<Constant*, 8> ShuffleMask;
1657 for (unsigned i = 0; i < VF; ++i)
1658 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1660 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1661 ConstantVector::get(ShuffleMask),
1665 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1666 // Attempt to issue a wide load.
1667 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1668 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1670 assert((LI || SI) && "Invalid Load/Store instruction");
1672 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1673 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1674 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1675 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1676 // An alignment of 0 means target abi alignment. We need to use the scalar's
1677 // target abi alignment in such a case.
1679 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1680 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1681 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1682 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1684 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1685 return scalarizeInstruction(Instr, true);
1687 if (ScalarAllocatedSize != VectorElementSize)
1688 return scalarizeInstruction(Instr);
1690 // If the pointer is loop invariant or if it is non-consecutive,
1691 // scalarize the load.
1692 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1693 bool Reverse = ConsecutiveStride < 0;
1694 bool UniformLoad = LI && Legal->isUniform(Ptr);
1695 if (!ConsecutiveStride || UniformLoad)
1696 return scalarizeInstruction(Instr);
1698 Constant *Zero = Builder.getInt32(0);
1699 VectorParts &Entry = WidenMap.get(Instr);
1701 // Handle consecutive loads/stores.
1702 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1703 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1704 setDebugLocFromInst(Builder, Gep);
1705 Value *PtrOperand = Gep->getPointerOperand();
1706 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1707 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1709 // Create the new GEP with the new induction variable.
1710 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1711 Gep2->setOperand(0, FirstBasePtr);
1712 Gep2->setName("gep.indvar.base");
1713 Ptr = Builder.Insert(Gep2);
1715 setDebugLocFromInst(Builder, Gep);
1716 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1717 OrigLoop) && "Base ptr must be invariant");
1719 // The last index does not have to be the induction. It can be
1720 // consecutive and be a function of the index. For example A[I+1];
1721 unsigned NumOperands = Gep->getNumOperands();
1722 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1723 // Create the new GEP with the new induction variable.
1724 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1726 for (unsigned i = 0; i < NumOperands; ++i) {
1727 Value *GepOperand = Gep->getOperand(i);
1728 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1730 // Update last index or loop invariant instruction anchored in loop.
1731 if (i == InductionOperand ||
1732 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1733 assert((i == InductionOperand ||
1734 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1735 "Must be last index or loop invariant");
1737 VectorParts &GEPParts = getVectorValue(GepOperand);
1738 Value *Index = GEPParts[0];
1739 Index = Builder.CreateExtractElement(Index, Zero);
1740 Gep2->setOperand(i, Index);
1741 Gep2->setName("gep.indvar.idx");
1744 Ptr = Builder.Insert(Gep2);
1746 // Use the induction element ptr.
1747 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1748 setDebugLocFromInst(Builder, Ptr);
1749 VectorParts &PtrVal = getVectorValue(Ptr);
1750 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1755 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1756 "We do not allow storing to uniform addresses");
1757 setDebugLocFromInst(Builder, SI);
1758 // We don't want to update the value in the map as it might be used in
1759 // another expression. So don't use a reference type for "StoredVal".
1760 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1762 for (unsigned Part = 0; Part < UF; ++Part) {
1763 // Calculate the pointer for the specific unroll-part.
1764 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1767 // If we store to reverse consecutive memory locations then we need
1768 // to reverse the order of elements in the stored value.
1769 StoredVal[Part] = reverseVector(StoredVal[Part]);
1770 // If the address is consecutive but reversed, then the
1771 // wide store needs to start at the last vector element.
1772 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1773 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1776 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1777 DataTy->getPointerTo(AddressSpace));
1779 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1780 propagateMetadata(NewSI, SI);
1786 assert(LI && "Must have a load instruction");
1787 setDebugLocFromInst(Builder, LI);
1788 for (unsigned Part = 0; Part < UF; ++Part) {
1789 // Calculate the pointer for the specific unroll-part.
1790 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1793 // If the address is consecutive but reversed, then the
1794 // wide store needs to start at the last vector element.
1795 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1796 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1799 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1800 DataTy->getPointerTo(AddressSpace));
1801 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1802 propagateMetadata(NewLI, LI);
1803 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1807 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1808 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1809 // Holds vector parameters or scalars, in case of uniform vals.
1810 SmallVector<VectorParts, 4> Params;
1812 setDebugLocFromInst(Builder, Instr);
1814 // Find all of the vectorized parameters.
1815 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1816 Value *SrcOp = Instr->getOperand(op);
1818 // If we are accessing the old induction variable, use the new one.
1819 if (SrcOp == OldInduction) {
1820 Params.push_back(getVectorValue(SrcOp));
1824 // Try using previously calculated values.
1825 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1827 // If the src is an instruction that appeared earlier in the basic block
1828 // then it should already be vectorized.
1829 if (SrcInst && OrigLoop->contains(SrcInst)) {
1830 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1831 // The parameter is a vector value from earlier.
1832 Params.push_back(WidenMap.get(SrcInst));
1834 // The parameter is a scalar from outside the loop. Maybe even a constant.
1835 VectorParts Scalars;
1836 Scalars.append(UF, SrcOp);
1837 Params.push_back(Scalars);
1841 assert(Params.size() == Instr->getNumOperands() &&
1842 "Invalid number of operands");
1844 // Does this instruction return a value ?
1845 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1847 Value *UndefVec = IsVoidRetTy ? nullptr :
1848 UndefValue::get(VectorType::get(Instr->getType(), VF));
1849 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1850 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1852 Instruction *InsertPt = Builder.GetInsertPoint();
1853 BasicBlock *IfBlock = Builder.GetInsertBlock();
1854 BasicBlock *CondBlock = nullptr;
1857 Loop *VectorLp = nullptr;
1858 if (IfPredicateStore) {
1859 assert(Instr->getParent()->getSinglePredecessor() &&
1860 "Only support single predecessor blocks");
1861 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1862 Instr->getParent());
1863 VectorLp = LI->getLoopFor(IfBlock);
1864 assert(VectorLp && "Must have a loop for this block");
1867 // For each vector unroll 'part':
1868 for (unsigned Part = 0; Part < UF; ++Part) {
1869 // For each scalar that we create:
1870 for (unsigned Width = 0; Width < VF; ++Width) {
1873 Value *Cmp = nullptr;
1874 if (IfPredicateStore) {
1875 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1876 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1877 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1878 LoopVectorBody.push_back(CondBlock);
1879 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1880 // Update Builder with newly created basic block.
1881 Builder.SetInsertPoint(InsertPt);
1884 Instruction *Cloned = Instr->clone();
1886 Cloned->setName(Instr->getName() + ".cloned");
1887 // Replace the operands of the cloned instructions with extracted scalars.
1888 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1889 Value *Op = Params[op][Part];
1890 // Param is a vector. Need to extract the right lane.
1891 if (Op->getType()->isVectorTy())
1892 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1893 Cloned->setOperand(op, Op);
1896 // Place the cloned scalar in the new loop.
1897 Builder.Insert(Cloned);
1899 // If the original scalar returns a value we need to place it in a vector
1900 // so that future users will be able to use it.
1902 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1903 Builder.getInt32(Width));
1905 if (IfPredicateStore) {
1906 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1907 LoopVectorBody.push_back(NewIfBlock);
1908 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1909 Builder.SetInsertPoint(InsertPt);
1910 Instruction *OldBr = IfBlock->getTerminator();
1911 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1912 OldBr->eraseFromParent();
1913 IfBlock = NewIfBlock;
1919 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1923 if (Instruction *I = dyn_cast<Instruction>(V))
1924 return I->getParent() == Loc->getParent() ? I : nullptr;
1928 std::pair<Instruction *, Instruction *>
1929 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1930 Instruction *tnullptr = nullptr;
1931 if (!Legal->mustCheckStrides())
1932 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1934 IRBuilder<> ChkBuilder(Loc);
1937 Value *Check = nullptr;
1938 Instruction *FirstInst = nullptr;
1939 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1940 SE = Legal->strides_end();
1942 Value *Ptr = stripIntegerCast(*SI);
1943 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1945 // Store the first instruction we create.
1946 FirstInst = getFirstInst(FirstInst, C, Loc);
1948 Check = ChkBuilder.CreateOr(Check, C);
1953 // We have to do this trickery because the IRBuilder might fold the check to a
1954 // constant expression in which case there is no Instruction anchored in a
1956 LLVMContext &Ctx = Loc->getContext();
1957 Instruction *TheCheck =
1958 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1959 ChkBuilder.Insert(TheCheck, "stride.not.one");
1960 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1962 return std::make_pair(FirstInst, TheCheck);
1965 std::pair<Instruction *, Instruction *>
1966 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1967 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1968 Legal->getRuntimePointerCheck();
1970 Instruction *tnullptr = nullptr;
1971 if (!PtrRtCheck->Need)
1972 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1974 unsigned NumPointers = PtrRtCheck->Pointers.size();
1975 SmallVector<TrackingVH<Value> , 2> Starts;
1976 SmallVector<TrackingVH<Value> , 2> Ends;
1978 LLVMContext &Ctx = Loc->getContext();
1979 SCEVExpander Exp(*SE, "induction");
1980 Instruction *FirstInst = nullptr;
1982 for (unsigned i = 0; i < NumPointers; ++i) {
1983 Value *Ptr = PtrRtCheck->Pointers[i];
1984 const SCEV *Sc = SE->getSCEV(Ptr);
1986 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1987 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1989 Starts.push_back(Ptr);
1990 Ends.push_back(Ptr);
1992 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1993 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1995 // Use this type for pointer arithmetic.
1996 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1998 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1999 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2000 Starts.push_back(Start);
2001 Ends.push_back(End);
2005 IRBuilder<> ChkBuilder(Loc);
2006 // Our instructions might fold to a constant.
2007 Value *MemoryRuntimeCheck = nullptr;
2008 for (unsigned i = 0; i < NumPointers; ++i) {
2009 for (unsigned j = i+1; j < NumPointers; ++j) {
2010 // No need to check if two readonly pointers intersect.
2011 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2014 // Only need to check pointers between two different dependency sets.
2015 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2017 // Only need to check pointers in the same alias set.
2018 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2021 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2022 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2024 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2025 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2026 "Trying to bounds check pointers with different address spaces");
2028 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2029 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2031 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2032 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2033 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2034 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2036 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2037 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2038 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2039 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2040 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2041 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2042 if (MemoryRuntimeCheck) {
2043 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2045 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2047 MemoryRuntimeCheck = IsConflict;
2051 // We have to do this trickery because the IRBuilder might fold the check to a
2052 // constant expression in which case there is no Instruction anchored in a
2054 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2055 ConstantInt::getTrue(Ctx));
2056 ChkBuilder.Insert(Check, "memcheck.conflict");
2057 FirstInst = getFirstInst(FirstInst, Check, Loc);
2058 return std::make_pair(FirstInst, Check);
2061 void InnerLoopVectorizer::createEmptyLoop() {
2063 In this function we generate a new loop. The new loop will contain
2064 the vectorized instructions while the old loop will continue to run the
2067 [ ] <-- Back-edge taken count overflow check.
2070 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2073 || [ ] <-- vector pre header.
2077 || [ ]_| <-- vector loop.
2080 | >[ ] <--- middle-block.
2083 -|- >[ ] <--- new preheader.
2087 | [ ]_| <-- old scalar loop to handle remainder.
2090 >[ ] <-- exit block.
2094 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2095 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2096 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2097 assert(BypassBlock && "Invalid loop structure");
2098 assert(ExitBlock && "Must have an exit block");
2100 // Some loops have a single integer induction variable, while other loops
2101 // don't. One example is c++ iterators that often have multiple pointer
2102 // induction variables. In the code below we also support a case where we
2103 // don't have a single induction variable.
2104 OldInduction = Legal->getInduction();
2105 Type *IdxTy = Legal->getWidestInductionType();
2107 // Find the loop boundaries.
2108 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2109 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2111 // The exit count might have the type of i64 while the phi is i32. This can
2112 // happen if we have an induction variable that is sign extended before the
2113 // compare. The only way that we get a backedge taken count is that the
2114 // induction variable was signed and as such will not overflow. In such a case
2115 // truncation is legal.
2116 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2117 IdxTy->getPrimitiveSizeInBits())
2118 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2120 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2121 // Get the total trip count from the count by adding 1.
2122 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2123 SE->getConstant(BackedgeTakeCount->getType(), 1));
2125 // Expand the trip count and place the new instructions in the preheader.
2126 // Notice that the pre-header does not change, only the loop body.
2127 SCEVExpander Exp(*SE, "induction");
2129 // We need to test whether the backedge-taken count is uint##_max. Adding one
2130 // to it will cause overflow and an incorrect loop trip count in the vector
2131 // body. In case of overflow we want to directly jump to the scalar remainder
2133 Value *BackedgeCount =
2134 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2135 BypassBlock->getTerminator());
2136 if (BackedgeCount->getType()->isPointerTy())
2137 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2138 "backedge.ptrcnt.to.int",
2139 BypassBlock->getTerminator());
2140 Instruction *CheckBCOverflow =
2141 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2142 Constant::getAllOnesValue(BackedgeCount->getType()),
2143 "backedge.overflow", BypassBlock->getTerminator());
2145 // The loop index does not have to start at Zero. Find the original start
2146 // value from the induction PHI node. If we don't have an induction variable
2147 // then we know that it starts at zero.
2148 Builder.SetInsertPoint(BypassBlock->getTerminator());
2149 Value *StartIdx = ExtendedIdx = OldInduction ?
2150 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2152 ConstantInt::get(IdxTy, 0);
2154 // We need an instruction to anchor the overflow check on. StartIdx needs to
2155 // be defined before the overflow check branch. Because the scalar preheader
2156 // is going to merge the start index and so the overflow branch block needs to
2157 // contain a definition of the start index.
2158 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2159 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2160 BypassBlock->getTerminator());
2162 // Count holds the overall loop count (N).
2163 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2164 BypassBlock->getTerminator());
2166 LoopBypassBlocks.push_back(BypassBlock);
2168 // Split the single block loop into the two loop structure described above.
2169 BasicBlock *VectorPH =
2170 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2171 BasicBlock *VecBody =
2172 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2173 BasicBlock *MiddleBlock =
2174 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2175 BasicBlock *ScalarPH =
2176 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2178 // Create and register the new vector loop.
2179 Loop* Lp = new Loop();
2180 Loop *ParentLoop = OrigLoop->getParentLoop();
2182 // Insert the new loop into the loop nest and register the new basic blocks
2183 // before calling any utilities such as SCEV that require valid LoopInfo.
2185 ParentLoop->addChildLoop(Lp);
2186 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2187 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2188 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2190 LI->addTopLevelLoop(Lp);
2192 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2194 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2196 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2198 // Generate the induction variable.
2199 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2200 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2201 // The loop step is equal to the vectorization factor (num of SIMD elements)
2202 // times the unroll factor (num of SIMD instructions).
2203 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2205 // This is the IR builder that we use to add all of the logic for bypassing
2206 // the new vector loop.
2207 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2208 setDebugLocFromInst(BypassBuilder,
2209 getDebugLocFromInstOrOperands(OldInduction));
2211 // We may need to extend the index in case there is a type mismatch.
2212 // We know that the count starts at zero and does not overflow.
2213 if (Count->getType() != IdxTy) {
2214 // The exit count can be of pointer type. Convert it to the correct
2216 if (ExitCount->getType()->isPointerTy())
2217 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2219 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2222 // Add the start index to the loop count to get the new end index.
2223 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2225 // Now we need to generate the expression for N - (N % VF), which is
2226 // the part that the vectorized body will execute.
2227 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2228 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2229 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2230 "end.idx.rnd.down");
2232 // Now, compare the new count to zero. If it is zero skip the vector loop and
2233 // jump to the scalar loop.
2235 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2237 BasicBlock *LastBypassBlock = BypassBlock;
2239 // Generate code to check that the loops trip count that we computed by adding
2240 // one to the backedge-taken count will not overflow.
2242 auto PastOverflowCheck =
2243 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2244 BasicBlock *CheckBlock =
2245 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2247 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2248 LoopBypassBlocks.push_back(CheckBlock);
2249 Instruction *OldTerm = LastBypassBlock->getTerminator();
2250 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2251 OldTerm->eraseFromParent();
2252 LastBypassBlock = CheckBlock;
2255 // Generate the code to check that the strides we assumed to be one are really
2256 // one. We want the new basic block to start at the first instruction in a
2257 // sequence of instructions that form a check.
2258 Instruction *StrideCheck;
2259 Instruction *FirstCheckInst;
2260 std::tie(FirstCheckInst, StrideCheck) =
2261 addStrideCheck(LastBypassBlock->getTerminator());
2263 // Create a new block containing the stride check.
2264 BasicBlock *CheckBlock =
2265 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2267 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2268 LoopBypassBlocks.push_back(CheckBlock);
2270 // Replace the branch into the memory check block with a conditional branch
2271 // for the "few elements case".
2272 Instruction *OldTerm = LastBypassBlock->getTerminator();
2273 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2274 OldTerm->eraseFromParent();
2277 LastBypassBlock = CheckBlock;
2280 // Generate the code that checks in runtime if arrays overlap. We put the
2281 // checks into a separate block to make the more common case of few elements
2283 Instruction *MemRuntimeCheck;
2284 std::tie(FirstCheckInst, MemRuntimeCheck) =
2285 addRuntimeCheck(LastBypassBlock->getTerminator());
2286 if (MemRuntimeCheck) {
2287 // Create a new block containing the memory check.
2288 BasicBlock *CheckBlock =
2289 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2291 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2292 LoopBypassBlocks.push_back(CheckBlock);
2294 // Replace the branch into the memory check block with a conditional branch
2295 // for the "few elements case".
2296 Instruction *OldTerm = LastBypassBlock->getTerminator();
2297 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2298 OldTerm->eraseFromParent();
2300 Cmp = MemRuntimeCheck;
2301 LastBypassBlock = CheckBlock;
2304 LastBypassBlock->getTerminator()->eraseFromParent();
2305 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2308 // We are going to resume the execution of the scalar loop.
2309 // Go over all of the induction variables that we found and fix the
2310 // PHIs that are left in the scalar version of the loop.
2311 // The starting values of PHI nodes depend on the counter of the last
2312 // iteration in the vectorized loop.
2313 // If we come from a bypass edge then we need to start from the original
2316 // This variable saves the new starting index for the scalar loop.
2317 PHINode *ResumeIndex = nullptr;
2318 LoopVectorizationLegality::InductionList::iterator I, E;
2319 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2320 // Set builder to point to last bypass block.
2321 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2322 for (I = List->begin(), E = List->end(); I != E; ++I) {
2323 PHINode *OrigPhi = I->first;
2324 LoopVectorizationLegality::InductionInfo II = I->second;
2326 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2327 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2328 MiddleBlock->getTerminator());
2329 // We might have extended the type of the induction variable but we need a
2330 // truncated version for the scalar loop.
2331 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2332 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2333 MiddleBlock->getTerminator()) : nullptr;
2335 // Create phi nodes to merge from the backedge-taken check block.
2336 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2337 ScalarPH->getTerminator());
2338 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2340 PHINode *BCTruncResumeVal = nullptr;
2341 if (OrigPhi == OldInduction) {
2343 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2344 ScalarPH->getTerminator());
2345 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2348 Value *EndValue = nullptr;
2350 case LoopVectorizationLegality::IK_NoInduction:
2351 llvm_unreachable("Unknown induction");
2352 case LoopVectorizationLegality::IK_IntInduction: {
2353 // Handle the integer induction counter.
2354 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2356 // We have the canonical induction variable.
2357 if (OrigPhi == OldInduction) {
2358 // Create a truncated version of the resume value for the scalar loop,
2359 // we might have promoted the type to a larger width.
2361 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2362 // The new PHI merges the original incoming value, in case of a bypass,
2363 // or the value at the end of the vectorized loop.
2364 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2365 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2366 TruncResumeVal->addIncoming(EndValue, VecBody);
2368 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2370 // We know what the end value is.
2371 EndValue = IdxEndRoundDown;
2372 // We also know which PHI node holds it.
2373 ResumeIndex = ResumeVal;
2377 // Not the canonical induction variable - add the vector loop count to the
2379 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2380 II.StartValue->getType(),
2382 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2385 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2386 // Convert the CountRoundDown variable to the PHI size.
2387 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2388 II.StartValue->getType(),
2390 // Handle reverse integer induction counter.
2391 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2394 case LoopVectorizationLegality::IK_PtrInduction: {
2395 // For pointer induction variables, calculate the offset using
2397 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2401 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2402 // The value at the end of the loop for the reverse pointer is calculated
2403 // by creating a GEP with a negative index starting from the start value.
2404 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2405 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2407 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2413 // The new PHI merges the original incoming value, in case of a bypass,
2414 // or the value at the end of the vectorized loop.
2415 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2416 if (OrigPhi == OldInduction)
2417 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2419 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2421 ResumeVal->addIncoming(EndValue, VecBody);
2423 // Fix the scalar body counter (PHI node).
2424 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2426 // The old induction's phi node in the scalar body needs the truncated
2428 if (OrigPhi == OldInduction) {
2429 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2430 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2432 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2433 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2437 // If we are generating a new induction variable then we also need to
2438 // generate the code that calculates the exit value. This value is not
2439 // simply the end of the counter because we may skip the vectorized body
2440 // in case of a runtime check.
2442 assert(!ResumeIndex && "Unexpected resume value found");
2443 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2444 MiddleBlock->getTerminator());
2445 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2446 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2447 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2450 // Make sure that we found the index where scalar loop needs to continue.
2451 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2452 "Invalid resume Index");
2454 // Add a check in the middle block to see if we have completed
2455 // all of the iterations in the first vector loop.
2456 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2457 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2458 ResumeIndex, "cmp.n",
2459 MiddleBlock->getTerminator());
2461 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2462 // Remove the old terminator.
2463 MiddleBlock->getTerminator()->eraseFromParent();
2465 // Create i+1 and fill the PHINode.
2466 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2467 Induction->addIncoming(StartIdx, VectorPH);
2468 Induction->addIncoming(NextIdx, VecBody);
2469 // Create the compare.
2470 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2471 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2473 // Now we have two terminators. Remove the old one from the block.
2474 VecBody->getTerminator()->eraseFromParent();
2476 // Get ready to start creating new instructions into the vectorized body.
2477 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2480 LoopVectorPreHeader = VectorPH;
2481 LoopScalarPreHeader = ScalarPH;
2482 LoopMiddleBlock = MiddleBlock;
2483 LoopExitBlock = ExitBlock;
2484 LoopVectorBody.push_back(VecBody);
2485 LoopScalarBody = OldBasicBlock;
2487 LoopVectorizeHints Hints(Lp, true);
2488 Hints.setAlreadyVectorized(Lp);
2491 /// This function returns the identity element (or neutral element) for
2492 /// the operation K.
2494 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2499 // Adding, Xoring, Oring zero to a number does not change it.
2500 return ConstantInt::get(Tp, 0);
2501 case RK_IntegerMult:
2502 // Multiplying a number by 1 does not change it.
2503 return ConstantInt::get(Tp, 1);
2505 // AND-ing a number with an all-1 value does not change it.
2506 return ConstantInt::get(Tp, -1, true);
2508 // Multiplying a number by 1 does not change it.
2509 return ConstantFP::get(Tp, 1.0L);
2511 // Adding zero to a number does not change it.
2512 return ConstantFP::get(Tp, 0.0L);
2514 llvm_unreachable("Unknown reduction kind");
2518 /// This function translates the reduction kind to an LLVM binary operator.
2520 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2522 case LoopVectorizationLegality::RK_IntegerAdd:
2523 return Instruction::Add;
2524 case LoopVectorizationLegality::RK_IntegerMult:
2525 return Instruction::Mul;
2526 case LoopVectorizationLegality::RK_IntegerOr:
2527 return Instruction::Or;
2528 case LoopVectorizationLegality::RK_IntegerAnd:
2529 return Instruction::And;
2530 case LoopVectorizationLegality::RK_IntegerXor:
2531 return Instruction::Xor;
2532 case LoopVectorizationLegality::RK_FloatMult:
2533 return Instruction::FMul;
2534 case LoopVectorizationLegality::RK_FloatAdd:
2535 return Instruction::FAdd;
2536 case LoopVectorizationLegality::RK_IntegerMinMax:
2537 return Instruction::ICmp;
2538 case LoopVectorizationLegality::RK_FloatMinMax:
2539 return Instruction::FCmp;
2541 llvm_unreachable("Unknown reduction operation");
2545 Value *createMinMaxOp(IRBuilder<> &Builder,
2546 LoopVectorizationLegality::MinMaxReductionKind RK,
2549 CmpInst::Predicate P = CmpInst::ICMP_NE;
2552 llvm_unreachable("Unknown min/max reduction kind");
2553 case LoopVectorizationLegality::MRK_UIntMin:
2554 P = CmpInst::ICMP_ULT;
2556 case LoopVectorizationLegality::MRK_UIntMax:
2557 P = CmpInst::ICMP_UGT;
2559 case LoopVectorizationLegality::MRK_SIntMin:
2560 P = CmpInst::ICMP_SLT;
2562 case LoopVectorizationLegality::MRK_SIntMax:
2563 P = CmpInst::ICMP_SGT;
2565 case LoopVectorizationLegality::MRK_FloatMin:
2566 P = CmpInst::FCMP_OLT;
2568 case LoopVectorizationLegality::MRK_FloatMax:
2569 P = CmpInst::FCMP_OGT;
2574 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2575 RK == LoopVectorizationLegality::MRK_FloatMax)
2576 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2578 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2580 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2585 struct CSEDenseMapInfo {
2586 static bool canHandle(Instruction *I) {
2587 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2588 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2590 static inline Instruction *getEmptyKey() {
2591 return DenseMapInfo<Instruction *>::getEmptyKey();
2593 static inline Instruction *getTombstoneKey() {
2594 return DenseMapInfo<Instruction *>::getTombstoneKey();
2596 static unsigned getHashValue(Instruction *I) {
2597 assert(canHandle(I) && "Unknown instruction!");
2598 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2599 I->value_op_end()));
2601 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2602 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2603 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2605 return LHS->isIdenticalTo(RHS);
2610 /// \brief Check whether this block is a predicated block.
2611 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2612 /// = ...; " blocks. We start with one vectorized basic block. For every
2613 /// conditional block we split this vectorized block. Therefore, every second
2614 /// block will be a predicated one.
2615 static bool isPredicatedBlock(unsigned BlockNum) {
2616 return BlockNum % 2;
2619 ///\brief Perform cse of induction variable instructions.
2620 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2621 // Perform simple cse.
2622 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2623 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2624 BasicBlock *BB = BBs[i];
2625 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2626 Instruction *In = I++;
2628 if (!CSEDenseMapInfo::canHandle(In))
2631 // Check if we can replace this instruction with any of the
2632 // visited instructions.
2633 if (Instruction *V = CSEMap.lookup(In)) {
2634 In->replaceAllUsesWith(V);
2635 In->eraseFromParent();
2638 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2639 // ...;" blocks for predicated stores. Every second block is a predicated
2641 if (isPredicatedBlock(i))
2649 /// \brief Adds a 'fast' flag to floating point operations.
2650 static Value *addFastMathFlag(Value *V) {
2651 if (isa<FPMathOperator>(V)){
2652 FastMathFlags Flags;
2653 Flags.setUnsafeAlgebra();
2654 cast<Instruction>(V)->setFastMathFlags(Flags);
2659 void InnerLoopVectorizer::vectorizeLoop() {
2660 //===------------------------------------------------===//
2662 // Notice: any optimization or new instruction that go
2663 // into the code below should be also be implemented in
2666 //===------------------------------------------------===//
2667 Constant *Zero = Builder.getInt32(0);
2669 // In order to support reduction variables we need to be able to vectorize
2670 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2671 // stages. First, we create a new vector PHI node with no incoming edges.
2672 // We use this value when we vectorize all of the instructions that use the
2673 // PHI. Next, after all of the instructions in the block are complete we
2674 // add the new incoming edges to the PHI. At this point all of the
2675 // instructions in the basic block are vectorized, so we can use them to
2676 // construct the PHI.
2677 PhiVector RdxPHIsToFix;
2679 // Scan the loop in a topological order to ensure that defs are vectorized
2681 LoopBlocksDFS DFS(OrigLoop);
2684 // Vectorize all of the blocks in the original loop.
2685 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2686 be = DFS.endRPO(); bb != be; ++bb)
2687 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2689 // At this point every instruction in the original loop is widened to
2690 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2691 // that we vectorized. The PHI nodes are currently empty because we did
2692 // not want to introduce cycles. Notice that the remaining PHI nodes
2693 // that we need to fix are reduction variables.
2695 // Create the 'reduced' values for each of the induction vars.
2696 // The reduced values are the vector values that we scalarize and combine
2697 // after the loop is finished.
2698 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2700 PHINode *RdxPhi = *it;
2701 assert(RdxPhi && "Unable to recover vectorized PHI");
2703 // Find the reduction variable descriptor.
2704 assert(Legal->getReductionVars()->count(RdxPhi) &&
2705 "Unable to find the reduction variable");
2706 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2707 (*Legal->getReductionVars())[RdxPhi];
2709 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2711 // We need to generate a reduction vector from the incoming scalar.
2712 // To do so, we need to generate the 'identity' vector and override
2713 // one of the elements with the incoming scalar reduction. We need
2714 // to do it in the vector-loop preheader.
2715 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2717 // This is the vector-clone of the value that leaves the loop.
2718 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2719 Type *VecTy = VectorExit[0]->getType();
2721 // Find the reduction identity variable. Zero for addition, or, xor,
2722 // one for multiplication, -1 for And.
2725 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2726 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2727 // MinMax reduction have the start value as their identify.
2729 VectorStart = Identity = RdxDesc.StartValue;
2731 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2736 // Handle other reduction kinds:
2738 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2739 VecTy->getScalarType());
2742 // This vector is the Identity vector where the first element is the
2743 // incoming scalar reduction.
2744 VectorStart = RdxDesc.StartValue;
2746 Identity = ConstantVector::getSplat(VF, Iden);
2748 // This vector is the Identity vector where the first element is the
2749 // incoming scalar reduction.
2750 VectorStart = Builder.CreateInsertElement(Identity,
2751 RdxDesc.StartValue, Zero);
2755 // Fix the vector-loop phi.
2756 // We created the induction variable so we know that the
2757 // preheader is the first entry.
2758 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2760 // Reductions do not have to start at zero. They can start with
2761 // any loop invariant values.
2762 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2763 BasicBlock *Latch = OrigLoop->getLoopLatch();
2764 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2765 VectorParts &Val = getVectorValue(LoopVal);
2766 for (unsigned part = 0; part < UF; ++part) {
2767 // Make sure to add the reduction stat value only to the
2768 // first unroll part.
2769 Value *StartVal = (part == 0) ? VectorStart : Identity;
2770 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2771 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2772 LoopVectorBody.back());
2775 // Before each round, move the insertion point right between
2776 // the PHIs and the values we are going to write.
2777 // This allows us to write both PHINodes and the extractelement
2779 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2781 VectorParts RdxParts;
2782 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2783 for (unsigned part = 0; part < UF; ++part) {
2784 // This PHINode contains the vectorized reduction variable, or
2785 // the initial value vector, if we bypass the vector loop.
2786 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2787 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2788 Value *StartVal = (part == 0) ? VectorStart : Identity;
2789 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2790 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2791 NewPhi->addIncoming(RdxExitVal[part],
2792 LoopVectorBody.back());
2793 RdxParts.push_back(NewPhi);
2796 // Reduce all of the unrolled parts into a single vector.
2797 Value *ReducedPartRdx = RdxParts[0];
2798 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2799 setDebugLocFromInst(Builder, ReducedPartRdx);
2800 for (unsigned part = 1; part < UF; ++part) {
2801 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2802 // Floating point operations had to be 'fast' to enable the reduction.
2803 ReducedPartRdx = addFastMathFlag(
2804 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2805 ReducedPartRdx, "bin.rdx"));
2807 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2808 ReducedPartRdx, RdxParts[part]);
2812 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2813 // and vector ops, reducing the set of values being computed by half each
2815 assert(isPowerOf2_32(VF) &&
2816 "Reduction emission only supported for pow2 vectors!");
2817 Value *TmpVec = ReducedPartRdx;
2818 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2819 for (unsigned i = VF; i != 1; i >>= 1) {
2820 // Move the upper half of the vector to the lower half.
2821 for (unsigned j = 0; j != i/2; ++j)
2822 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2824 // Fill the rest of the mask with undef.
2825 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2826 UndefValue::get(Builder.getInt32Ty()));
2829 Builder.CreateShuffleVector(TmpVec,
2830 UndefValue::get(TmpVec->getType()),
2831 ConstantVector::get(ShuffleMask),
2834 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2835 // Floating point operations had to be 'fast' to enable the reduction.
2836 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2837 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2839 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2842 // The result is in the first element of the vector.
2843 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2844 Builder.getInt32(0));
2847 // Create a phi node that merges control-flow from the backedge-taken check
2848 // block and the middle block.
2849 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2850 LoopScalarPreHeader->getTerminator());
2851 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2852 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2854 // Now, we need to fix the users of the reduction variable
2855 // inside and outside of the scalar remainder loop.
2856 // We know that the loop is in LCSSA form. We need to update the
2857 // PHI nodes in the exit blocks.
2858 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2859 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2860 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2861 if (!LCSSAPhi) break;
2863 // All PHINodes need to have a single entry edge, or two if
2864 // we already fixed them.
2865 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2867 // We found our reduction value exit-PHI. Update it with the
2868 // incoming bypass edge.
2869 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2870 // Add an edge coming from the bypass.
2871 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2874 }// end of the LCSSA phi scan.
2876 // Fix the scalar loop reduction variable with the incoming reduction sum
2877 // from the vector body and from the backedge value.
2878 int IncomingEdgeBlockIdx =
2879 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2880 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2881 // Pick the other block.
2882 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2883 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2884 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2885 }// end of for each redux variable.
2889 // Remove redundant induction instructions.
2890 cse(LoopVectorBody);
2893 void InnerLoopVectorizer::fixLCSSAPHIs() {
2894 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2895 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2896 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2897 if (!LCSSAPhi) break;
2898 if (LCSSAPhi->getNumIncomingValues() == 1)
2899 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2904 InnerLoopVectorizer::VectorParts
2905 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2906 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2909 // Look for cached value.
2910 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2911 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2912 if (ECEntryIt != MaskCache.end())
2913 return ECEntryIt->second;
2915 VectorParts SrcMask = createBlockInMask(Src);
2917 // The terminator has to be a branch inst!
2918 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2919 assert(BI && "Unexpected terminator found");
2921 if (BI->isConditional()) {
2922 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2924 if (BI->getSuccessor(0) != Dst)
2925 for (unsigned part = 0; part < UF; ++part)
2926 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2928 for (unsigned part = 0; part < UF; ++part)
2929 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2931 MaskCache[Edge] = EdgeMask;
2935 MaskCache[Edge] = SrcMask;
2939 InnerLoopVectorizer::VectorParts
2940 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2941 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2943 // Loop incoming mask is all-one.
2944 if (OrigLoop->getHeader() == BB) {
2945 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2946 return getVectorValue(C);
2949 // This is the block mask. We OR all incoming edges, and with zero.
2950 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2951 VectorParts BlockMask = getVectorValue(Zero);
2953 for (BasicBlock *Pred : predecessors(BB)) {
2954 VectorParts EM = createEdgeMask(Pred, BB);
2955 for (unsigned part = 0; part < UF; ++part)
2956 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2962 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2963 InnerLoopVectorizer::VectorParts &Entry,
2964 unsigned UF, unsigned VF, PhiVector *PV) {
2965 PHINode* P = cast<PHINode>(PN);
2966 // Handle reduction variables:
2967 if (Legal->getReductionVars()->count(P)) {
2968 for (unsigned part = 0; part < UF; ++part) {
2969 // This is phase one of vectorizing PHIs.
2970 Type *VecTy = (VF == 1) ? PN->getType() :
2971 VectorType::get(PN->getType(), VF);
2972 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2973 LoopVectorBody.back()-> getFirstInsertionPt());
2979 setDebugLocFromInst(Builder, P);
2980 // Check for PHI nodes that are lowered to vector selects.
2981 if (P->getParent() != OrigLoop->getHeader()) {
2982 // We know that all PHIs in non-header blocks are converted into
2983 // selects, so we don't have to worry about the insertion order and we
2984 // can just use the builder.
2985 // At this point we generate the predication tree. There may be
2986 // duplications since this is a simple recursive scan, but future
2987 // optimizations will clean it up.
2989 unsigned NumIncoming = P->getNumIncomingValues();
2991 // Generate a sequence of selects of the form:
2992 // SELECT(Mask3, In3,
2993 // SELECT(Mask2, In2,
2995 for (unsigned In = 0; In < NumIncoming; In++) {
2996 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2998 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3000 for (unsigned part = 0; part < UF; ++part) {
3001 // We might have single edge PHIs (blocks) - use an identity
3002 // 'select' for the first PHI operand.
3004 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3007 // Select between the current value and the previous incoming edge
3008 // based on the incoming mask.
3009 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3010 Entry[part], "predphi");
3016 // This PHINode must be an induction variable.
3017 // Make sure that we know about it.
3018 assert(Legal->getInductionVars()->count(P) &&
3019 "Not an induction variable");
3021 LoopVectorizationLegality::InductionInfo II =
3022 Legal->getInductionVars()->lookup(P);
3025 case LoopVectorizationLegality::IK_NoInduction:
3026 llvm_unreachable("Unknown induction");
3027 case LoopVectorizationLegality::IK_IntInduction: {
3028 assert(P->getType() == II.StartValue->getType() && "Types must match");
3029 Type *PhiTy = P->getType();
3031 if (P == OldInduction) {
3032 // Handle the canonical induction variable. We might have had to
3034 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3036 // Handle other induction variables that are now based on the
3038 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3040 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3041 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3044 Broadcasted = getBroadcastInstrs(Broadcasted);
3045 // After broadcasting the induction variable we need to make the vector
3046 // consecutive by adding 0, 1, 2, etc.
3047 for (unsigned part = 0; part < UF; ++part)
3048 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3051 case LoopVectorizationLegality::IK_ReverseIntInduction:
3052 case LoopVectorizationLegality::IK_PtrInduction:
3053 case LoopVectorizationLegality::IK_ReversePtrInduction:
3054 // Handle reverse integer and pointer inductions.
3055 Value *StartIdx = ExtendedIdx;
3056 // This is the normalized GEP that starts counting at zero.
3057 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3060 // Handle the reverse integer induction variable case.
3061 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3062 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3063 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3065 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3068 // This is a new value so do not hoist it out.
3069 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3070 // After broadcasting the induction variable we need to make the
3071 // vector consecutive by adding ... -3, -2, -1, 0.
3072 for (unsigned part = 0; part < UF; ++part)
3073 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3078 // Handle the pointer induction variable case.
3079 assert(P->getType()->isPointerTy() && "Unexpected type.");
3081 // Is this a reverse induction ptr or a consecutive induction ptr.
3082 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3085 // This is the vector of results. Notice that we don't generate
3086 // vector geps because scalar geps result in better code.
3087 for (unsigned part = 0; part < UF; ++part) {
3089 int EltIndex = (part) * (Reverse ? -1 : 1);
3090 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3093 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3095 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3097 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3099 Entry[part] = SclrGep;
3103 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3104 for (unsigned int i = 0; i < VF; ++i) {
3105 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3106 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3109 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3111 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3113 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3115 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3116 Builder.getInt32(i),
3119 Entry[part] = VecVal;
3125 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3126 // For each instruction in the old loop.
3127 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3128 VectorParts &Entry = WidenMap.get(it);
3129 switch (it->getOpcode()) {
3130 case Instruction::Br:
3131 // Nothing to do for PHIs and BR, since we already took care of the
3132 // loop control flow instructions.
3134 case Instruction::PHI:{
3135 // Vectorize PHINodes.
3136 widenPHIInstruction(it, Entry, UF, VF, PV);
3140 case Instruction::Add:
3141 case Instruction::FAdd:
3142 case Instruction::Sub:
3143 case Instruction::FSub:
3144 case Instruction::Mul:
3145 case Instruction::FMul:
3146 case Instruction::UDiv:
3147 case Instruction::SDiv:
3148 case Instruction::FDiv:
3149 case Instruction::URem:
3150 case Instruction::SRem:
3151 case Instruction::FRem:
3152 case Instruction::Shl:
3153 case Instruction::LShr:
3154 case Instruction::AShr:
3155 case Instruction::And:
3156 case Instruction::Or:
3157 case Instruction::Xor: {
3158 // Just widen binops.
3159 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3160 setDebugLocFromInst(Builder, BinOp);
3161 VectorParts &A = getVectorValue(it->getOperand(0));
3162 VectorParts &B = getVectorValue(it->getOperand(1));
3164 // Use this vector value for all users of the original instruction.
3165 for (unsigned Part = 0; Part < UF; ++Part) {
3166 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3168 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3169 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3170 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3171 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3172 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3174 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3175 VecOp->setIsExact(BinOp->isExact());
3177 // Copy the fast-math flags.
3178 if (VecOp && isa<FPMathOperator>(V))
3179 VecOp->setFastMathFlags(it->getFastMathFlags());
3184 propagateMetadata(Entry, it);
3187 case Instruction::Select: {
3189 // If the selector is loop invariant we can create a select
3190 // instruction with a scalar condition. Otherwise, use vector-select.
3191 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3193 setDebugLocFromInst(Builder, it);
3195 // The condition can be loop invariant but still defined inside the
3196 // loop. This means that we can't just use the original 'cond' value.
3197 // We have to take the 'vectorized' value and pick the first lane.
3198 // Instcombine will make this a no-op.
3199 VectorParts &Cond = getVectorValue(it->getOperand(0));
3200 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3201 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3203 Value *ScalarCond = (VF == 1) ? Cond[0] :
3204 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3206 for (unsigned Part = 0; Part < UF; ++Part) {
3207 Entry[Part] = Builder.CreateSelect(
3208 InvariantCond ? ScalarCond : Cond[Part],
3213 propagateMetadata(Entry, it);
3217 case Instruction::ICmp:
3218 case Instruction::FCmp: {
3219 // Widen compares. Generate vector compares.
3220 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3221 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3222 setDebugLocFromInst(Builder, it);
3223 VectorParts &A = getVectorValue(it->getOperand(0));
3224 VectorParts &B = getVectorValue(it->getOperand(1));
3225 for (unsigned Part = 0; Part < UF; ++Part) {
3228 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3230 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3234 propagateMetadata(Entry, it);
3238 case Instruction::Store:
3239 case Instruction::Load:
3240 vectorizeMemoryInstruction(it);
3242 case Instruction::ZExt:
3243 case Instruction::SExt:
3244 case Instruction::FPToUI:
3245 case Instruction::FPToSI:
3246 case Instruction::FPExt:
3247 case Instruction::PtrToInt:
3248 case Instruction::IntToPtr:
3249 case Instruction::SIToFP:
3250 case Instruction::UIToFP:
3251 case Instruction::Trunc:
3252 case Instruction::FPTrunc:
3253 case Instruction::BitCast: {
3254 CastInst *CI = dyn_cast<CastInst>(it);
3255 setDebugLocFromInst(Builder, it);
3256 /// Optimize the special case where the source is the induction
3257 /// variable. Notice that we can only optimize the 'trunc' case
3258 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3259 /// c. other casts depend on pointer size.
3260 if (CI->getOperand(0) == OldInduction &&
3261 it->getOpcode() == Instruction::Trunc) {
3262 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3264 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3265 for (unsigned Part = 0; Part < UF; ++Part)
3266 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3267 propagateMetadata(Entry, it);
3270 /// Vectorize casts.
3271 Type *DestTy = (VF == 1) ? CI->getType() :
3272 VectorType::get(CI->getType(), VF);
3274 VectorParts &A = getVectorValue(it->getOperand(0));
3275 for (unsigned Part = 0; Part < UF; ++Part)
3276 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3277 propagateMetadata(Entry, it);
3281 case Instruction::Call: {
3282 // Ignore dbg intrinsics.
3283 if (isa<DbgInfoIntrinsic>(it))
3285 setDebugLocFromInst(Builder, it);
3287 Module *M = BB->getParent()->getParent();
3288 CallInst *CI = cast<CallInst>(it);
3289 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3290 assert(ID && "Not an intrinsic call!");
3292 case Intrinsic::lifetime_end:
3293 case Intrinsic::lifetime_start:
3294 scalarizeInstruction(it);
3297 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3298 for (unsigned Part = 0; Part < UF; ++Part) {
3299 SmallVector<Value *, 4> Args;
3300 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3301 if (HasScalarOpd && i == 1) {
3302 Args.push_back(CI->getArgOperand(i));
3305 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3306 Args.push_back(Arg[Part]);
3308 Type *Tys[] = {CI->getType()};
3310 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3312 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3313 Entry[Part] = Builder.CreateCall(F, Args);
3316 propagateMetadata(Entry, it);
3323 // All other instructions are unsupported. Scalarize them.
3324 scalarizeInstruction(it);
3327 }// end of for_each instr.
3330 void InnerLoopVectorizer::updateAnalysis() {
3331 // Forget the original basic block.
3332 SE->forgetLoop(OrigLoop);
3334 // Update the dominator tree information.
3335 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3336 "Entry does not dominate exit.");
3338 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3339 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3340 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3342 // Due to if predication of stores we might create a sequence of "if(pred)
3343 // a[i] = ...; " blocks.
3344 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3346 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3347 else if (isPredicatedBlock(i)) {
3348 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3350 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3354 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3355 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3356 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3357 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3359 DEBUG(DT->verifyDomTree());
3362 /// \brief Check whether it is safe to if-convert this phi node.
3364 /// Phi nodes with constant expressions that can trap are not safe to if
3366 static bool canIfConvertPHINodes(BasicBlock *BB) {
3367 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3368 PHINode *Phi = dyn_cast<PHINode>(I);
3371 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3372 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3379 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3380 if (!EnableIfConversion) {
3381 emitAnalysis(Report() << "if-conversion is disabled");
3385 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3387 // A list of pointers that we can safely read and write to.
3388 SmallPtrSet<Value *, 8> SafePointes;
3390 // Collect safe addresses.
3391 for (Loop::block_iterator BI = TheLoop->block_begin(),
3392 BE = TheLoop->block_end(); BI != BE; ++BI) {
3393 BasicBlock *BB = *BI;
3395 if (blockNeedsPredication(BB))
3398 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3399 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3400 SafePointes.insert(LI->getPointerOperand());
3401 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3402 SafePointes.insert(SI->getPointerOperand());
3406 // Collect the blocks that need predication.
3407 BasicBlock *Header = TheLoop->getHeader();
3408 for (Loop::block_iterator BI = TheLoop->block_begin(),
3409 BE = TheLoop->block_end(); BI != BE; ++BI) {
3410 BasicBlock *BB = *BI;
3412 // We don't support switch statements inside loops.
3413 if (!isa<BranchInst>(BB->getTerminator())) {
3414 emitAnalysis(Report(BB->getTerminator())
3415 << "loop contains a switch statement");
3419 // We must be able to predicate all blocks that need to be predicated.
3420 if (blockNeedsPredication(BB)) {
3421 if (!blockCanBePredicated(BB, SafePointes)) {
3422 emitAnalysis(Report(BB->getTerminator())
3423 << "control flow cannot be substituted for a select");
3426 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3427 emitAnalysis(Report(BB->getTerminator())
3428 << "control flow cannot be substituted for a select");
3433 // We can if-convert this loop.
3437 bool LoopVectorizationLegality::canVectorize() {
3438 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3439 // be canonicalized.
3440 if (!TheLoop->getLoopPreheader()) {
3442 Report() << "loop control flow is not understood by vectorizer");
3446 // We can only vectorize innermost loops.
3447 if (TheLoop->getSubLoopsVector().size()) {
3448 emitAnalysis(Report() << "loop is not the innermost loop");
3452 // We must have a single backedge.
3453 if (TheLoop->getNumBackEdges() != 1) {
3455 Report() << "loop control flow is not understood by vectorizer");
3459 // We must have a single exiting block.
3460 if (!TheLoop->getExitingBlock()) {
3462 Report() << "loop control flow is not understood by vectorizer");
3466 // We need to have a loop header.
3467 DEBUG(dbgs() << "LV: Found a loop: " <<
3468 TheLoop->getHeader()->getName() << '\n');
3470 // Check if we can if-convert non-single-bb loops.
3471 unsigned NumBlocks = TheLoop->getNumBlocks();
3472 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3473 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3477 // ScalarEvolution needs to be able to find the exit count.
3478 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3479 if (ExitCount == SE->getCouldNotCompute()) {
3480 emitAnalysis(Report() << "could not determine number of loop iterations");
3481 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3485 // Check if we can vectorize the instructions and CFG in this loop.
3486 if (!canVectorizeInstrs()) {
3487 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3491 // Go over each instruction and look at memory deps.
3492 if (!canVectorizeMemory()) {
3493 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3497 // Collect all of the variables that remain uniform after vectorization.
3498 collectLoopUniforms();
3500 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3501 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3504 // Okay! We can vectorize. At this point we don't have any other mem analysis
3505 // which may limit our maximum vectorization factor, so just return true with
3510 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3511 if (Ty->isPointerTy())
3512 return DL.getIntPtrType(Ty);
3514 // It is possible that char's or short's overflow when we ask for the loop's
3515 // trip count, work around this by changing the type size.
3516 if (Ty->getScalarSizeInBits() < 32)
3517 return Type::getInt32Ty(Ty->getContext());
3522 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3523 Ty0 = convertPointerToIntegerType(DL, Ty0);
3524 Ty1 = convertPointerToIntegerType(DL, Ty1);
3525 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3530 /// \brief Check that the instruction has outside loop users and is not an
3531 /// identified reduction variable.
3532 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3533 SmallPtrSet<Value *, 4> &Reductions) {
3534 // Reduction instructions are allowed to have exit users. All other
3535 // instructions must not have external users.
3536 if (!Reductions.count(Inst))
3537 //Check that all of the users of the loop are inside the BB.
3538 for (User *U : Inst->users()) {
3539 Instruction *UI = cast<Instruction>(U);
3540 // This user may be a reduction exit value.
3541 if (!TheLoop->contains(UI)) {
3542 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3549 bool LoopVectorizationLegality::canVectorizeInstrs() {
3550 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3551 BasicBlock *Header = TheLoop->getHeader();
3553 // Look for the attribute signaling the absence of NaNs.
3554 Function &F = *Header->getParent();
3555 if (F.hasFnAttribute("no-nans-fp-math"))
3556 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3557 AttributeSet::FunctionIndex,
3558 "no-nans-fp-math").getValueAsString() == "true";
3560 // For each block in the loop.
3561 for (Loop::block_iterator bb = TheLoop->block_begin(),
3562 be = TheLoop->block_end(); bb != be; ++bb) {
3564 // Scan the instructions in the block and look for hazards.
3565 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3568 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3569 Type *PhiTy = Phi->getType();
3570 // Check that this PHI type is allowed.
3571 if (!PhiTy->isIntegerTy() &&
3572 !PhiTy->isFloatingPointTy() &&
3573 !PhiTy->isPointerTy()) {
3574 emitAnalysis(Report(it)
3575 << "loop control flow is not understood by vectorizer");
3576 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3580 // If this PHINode is not in the header block, then we know that we
3581 // can convert it to select during if-conversion. No need to check if
3582 // the PHIs in this block are induction or reduction variables.
3583 if (*bb != Header) {
3584 // Check that this instruction has no outside users or is an
3585 // identified reduction value with an outside user.
3586 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3588 emitAnalysis(Report(it) << "value that could not be identified as "
3589 "reduction is used outside the loop");
3593 // We only allow if-converted PHIs with more than two incoming values.
3594 if (Phi->getNumIncomingValues() != 2) {
3595 emitAnalysis(Report(it)
3596 << "control flow not understood by vectorizer");
3597 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3601 // This is the value coming from the preheader.
3602 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3603 // Check if this is an induction variable.
3604 InductionKind IK = isInductionVariable(Phi);
3606 if (IK_NoInduction != IK) {
3607 // Get the widest type.
3609 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3611 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3613 // Int inductions are special because we only allow one IV.
3614 if (IK == IK_IntInduction) {
3615 // Use the phi node with the widest type as induction. Use the last
3616 // one if there are multiple (no good reason for doing this other
3617 // than it is expedient).
3618 if (!Induction || PhiTy == WidestIndTy)
3622 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3623 Inductions[Phi] = InductionInfo(StartValue, IK);
3625 // Until we explicitly handle the case of an induction variable with
3626 // an outside loop user we have to give up vectorizing this loop.
3627 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3628 emitAnalysis(Report(it) << "use of induction value outside of the "
3629 "loop is not handled by vectorizer");
3636 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3637 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3640 if (AddReductionVar(Phi, RK_IntegerMult)) {
3641 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3644 if (AddReductionVar(Phi, RK_IntegerOr)) {
3645 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3648 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3649 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3652 if (AddReductionVar(Phi, RK_IntegerXor)) {
3653 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3656 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3657 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3660 if (AddReductionVar(Phi, RK_FloatMult)) {
3661 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3664 if (AddReductionVar(Phi, RK_FloatAdd)) {
3665 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3668 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3669 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3674 emitAnalysis(Report(it) << "unvectorizable operation");
3675 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3677 }// end of PHI handling
3679 // We still don't handle functions. However, we can ignore dbg intrinsic
3680 // calls and we do handle certain intrinsic and libm functions.
3681 CallInst *CI = dyn_cast<CallInst>(it);
3682 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3683 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3684 DEBUG(dbgs() << "LV: Found a call site.\n");
3688 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3689 // second argument is the same (i.e. loop invariant)
3691 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3692 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3693 emitAnalysis(Report(it)
3694 << "intrinsic instruction cannot be vectorized");
3695 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3700 // Check that the instruction return type is vectorizable.
3701 // Also, we can't vectorize extractelement instructions.
3702 if ((!VectorType::isValidElementType(it->getType()) &&
3703 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3704 emitAnalysis(Report(it)
3705 << "instruction return type cannot be vectorized");
3706 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3710 // Check that the stored type is vectorizable.
3711 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3712 Type *T = ST->getValueOperand()->getType();
3713 if (!VectorType::isValidElementType(T)) {
3714 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3717 if (EnableMemAccessVersioning)
3718 collectStridedAcccess(ST);
3721 if (EnableMemAccessVersioning)
3722 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3723 collectStridedAcccess(LI);
3725 // Reduction instructions are allowed to have exit users.
3726 // All other instructions must not have external users.
3727 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3728 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3737 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3738 if (Inductions.empty()) {
3739 emitAnalysis(Report()
3740 << "loop induction variable could not be identified");
3748 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3749 /// return the induction operand of the gep pointer.
3750 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3751 const DataLayout *DL, Loop *Lp) {
3752 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3756 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3758 // Check that all of the gep indices are uniform except for our induction
3760 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3761 if (i != InductionOperand &&
3762 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3764 return GEP->getOperand(InductionOperand);
3767 ///\brief Look for a cast use of the passed value.
3768 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3769 Value *UniqueCast = nullptr;
3770 for (User *U : Ptr->users()) {
3771 CastInst *CI = dyn_cast<CastInst>(U);
3772 if (CI && CI->getType() == Ty) {
3782 ///\brief Get the stride of a pointer access in a loop.
3783 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3784 /// pointer to the Value, or null otherwise.
3785 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3786 const DataLayout *DL, Loop *Lp) {
3787 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3788 if (!PtrTy || PtrTy->isAggregateType())
3791 // Try to remove a gep instruction to make the pointer (actually index at this
3792 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3793 // pointer, otherwise, we are analyzing the index.
3794 Value *OrigPtr = Ptr;
3796 // The size of the pointer access.
3797 int64_t PtrAccessSize = 1;
3799 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3800 const SCEV *V = SE->getSCEV(Ptr);
3804 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3805 V = C->getOperand();
3807 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3811 V = S->getStepRecurrence(*SE);
3815 // Strip off the size of access multiplication if we are still analyzing the
3817 if (OrigPtr == Ptr) {
3818 DL->getTypeAllocSize(PtrTy->getElementType());
3819 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3820 if (M->getOperand(0)->getSCEVType() != scConstant)
3823 const APInt &APStepVal =
3824 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3826 // Huge step value - give up.
3827 if (APStepVal.getBitWidth() > 64)
3830 int64_t StepVal = APStepVal.getSExtValue();
3831 if (PtrAccessSize != StepVal)
3833 V = M->getOperand(1);
3838 Type *StripedOffRecurrenceCast = nullptr;
3839 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3840 StripedOffRecurrenceCast = C->getType();
3841 V = C->getOperand();
3844 // Look for the loop invariant symbolic value.
3845 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3849 Value *Stride = U->getValue();
3850 if (!Lp->isLoopInvariant(Stride))
3853 // If we have stripped off the recurrence cast we have to make sure that we
3854 // return the value that is used in this loop so that we can replace it later.
3855 if (StripedOffRecurrenceCast)
3856 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3861 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3862 Value *Ptr = nullptr;
3863 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3864 Ptr = LI->getPointerOperand();
3865 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3866 Ptr = SI->getPointerOperand();
3870 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3874 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3875 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3876 Strides[Ptr] = Stride;
3877 StrideSet.insert(Stride);
3880 void LoopVectorizationLegality::collectLoopUniforms() {
3881 // We now know that the loop is vectorizable!
3882 // Collect variables that will remain uniform after vectorization.
3883 std::vector<Value*> Worklist;
3884 BasicBlock *Latch = TheLoop->getLoopLatch();
3886 // Start with the conditional branch and walk up the block.
3887 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3889 // Also add all consecutive pointer values; these values will be uniform
3890 // after vectorization (and subsequent cleanup) and, until revectorization is
3891 // supported, all dependencies must also be uniform.
3892 for (Loop::block_iterator B = TheLoop->block_begin(),
3893 BE = TheLoop->block_end(); B != BE; ++B)
3894 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3896 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3897 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3899 while (Worklist.size()) {
3900 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3901 Worklist.pop_back();
3903 // Look at instructions inside this loop.
3904 // Stop when reaching PHI nodes.
3905 // TODO: we need to follow values all over the loop, not only in this block.
3906 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3909 // This is a known uniform.
3912 // Insert all operands.
3913 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3918 /// \brief Analyses memory accesses in a loop.
3920 /// Checks whether run time pointer checks are needed and builds sets for data
3921 /// dependence checking.
3922 class AccessAnalysis {
3924 /// \brief Read or write access location.
3925 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3926 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3928 /// \brief Set of potential dependent memory accesses.
3929 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3931 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
3932 DL(Dl), AA(AA), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
3934 /// \brief Register a load and whether it is only read from.
3935 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
3936 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3937 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.TBAATag);
3938 Accesses.insert(MemAccessInfo(Ptr, false));
3940 ReadOnlyPtr.insert(Ptr);
3943 /// \brief Register a store.
3944 void addStore(AliasAnalysis::Location &Loc) {
3945 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3946 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.TBAATag);
3947 Accesses.insert(MemAccessInfo(Ptr, true));
3950 /// \brief Check whether we can check the pointers at runtime for
3951 /// non-intersection.
3952 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3953 unsigned &NumComparisons, ScalarEvolution *SE,
3954 Loop *TheLoop, ValueToValueMap &Strides,
3955 bool ShouldCheckStride = false);
3957 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3958 /// and builds sets of dependent accesses.
3959 void buildDependenceSets() {
3960 processMemAccesses();
3963 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3965 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3966 void resetDepChecks() { CheckDeps.clear(); }
3968 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3971 typedef SetVector<MemAccessInfo> PtrAccessSet;
3973 /// \brief Go over all memory access and check whether runtime pointer checks
3974 /// are needed /// and build sets of dependency check candidates.
3975 void processMemAccesses();
3977 /// Set of all accesses.
3978 PtrAccessSet Accesses;
3980 /// Set of accesses that need a further dependence check.
3981 MemAccessInfoSet CheckDeps;
3983 /// Set of pointers that are read only.
3984 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3986 const DataLayout *DL;
3989 /// An alias set tracker to partition the access set by underlying object and
3990 //intrinsic property (such as TBAA metadata).
3991 AliasSetTracker AST;
3993 /// Sets of potentially dependent accesses - members of one set share an
3994 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3995 /// dependence check.
3996 DepCandidates &DepCands;
3998 bool IsRTCheckNeeded;
4001 } // end anonymous namespace
4003 /// \brief Check whether a pointer can participate in a runtime bounds check.
4004 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4006 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4007 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4011 return AR->isAffine();
4014 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4015 /// the address space.
4016 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4017 const Loop *Lp, ValueToValueMap &StridesMap);
4019 bool AccessAnalysis::canCheckPtrAtRT(
4020 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4021 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4022 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4023 // Find pointers with computable bounds. We are going to use this information
4024 // to place a runtime bound check.
4025 bool CanDoRT = true;
4027 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4030 // We assign a consecutive id to access from different alias sets.
4031 // Accesses between different groups doesn't need to be checked.
4033 for (auto &AS : AST) {
4034 unsigned NumReadPtrChecks = 0;
4035 unsigned NumWritePtrChecks = 0;
4037 // We assign consecutive id to access from different dependence sets.
4038 // Accesses within the same set don't need a runtime check.
4039 unsigned RunningDepId = 1;
4040 DenseMap<Value *, unsigned> DepSetId;
4043 Value *Ptr = A.getValue();
4044 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4045 MemAccessInfo Access(Ptr, IsWrite);
4048 ++NumWritePtrChecks;
4052 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4053 // When we run after a failing dependency check we have to make sure we
4054 // don't have wrapping pointers.
4055 (!ShouldCheckStride ||
4056 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4057 // The id of the dependence set.
4060 if (IsDepCheckNeeded) {
4061 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4062 unsigned &LeaderId = DepSetId[Leader];
4064 LeaderId = RunningDepId++;
4067 // Each access has its own dependence set.
4068 DepId = RunningDepId++;
4070 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4072 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4078 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4079 NumComparisons += 0; // Only one dependence set.
4081 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4082 NumWritePtrChecks - 1));
4088 // If the pointers that we would use for the bounds comparison have different
4089 // address spaces, assume the values aren't directly comparable, so we can't
4090 // use them for the runtime check. We also have to assume they could
4091 // overlap. In the future there should be metadata for whether address spaces
4093 unsigned NumPointers = RtCheck.Pointers.size();
4094 for (unsigned i = 0; i < NumPointers; ++i) {
4095 for (unsigned j = i + 1; j < NumPointers; ++j) {
4096 // Only need to check pointers between two different dependency sets.
4097 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4099 // Only need to check pointers in the same alias set.
4100 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4103 Value *PtrI = RtCheck.Pointers[i];
4104 Value *PtrJ = RtCheck.Pointers[j];
4106 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4107 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4109 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4110 " different address spaces\n");
4119 void AccessAnalysis::processMemAccesses() {
4120 // We process the set twice: first we process read-write pointers, last we
4121 // process read-only pointers. This allows us to skip dependence tests for
4122 // read-only pointers.
4124 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4125 DEBUG(dbgs() << " AST: "; AST.dump());
4126 DEBUG(dbgs() << "LV: Accesses:\n");
4128 for (auto A : Accesses)
4129 dbgs() << "\t" << *A.getPointer() << " (" <<
4130 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4131 "read-only" : "read")) << ")\n";
4134 // The AliasSetTracker has nicely partitioned our pointers by metadata
4135 // compatibility and potential for underlying-object overlap. As a result, we
4136 // only need to check for potential pointer dependencies within each alias
4138 for (auto &AS : AST) {
4139 // Note that both the alias-set tracker and the alias sets themselves used
4140 // linked lists internally and so the iteration order here is deterministic
4141 // (matching the original instruction order within each set).
4143 bool SetHasWrite = false;
4145 // Map of pointers to last access encountered.
4146 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4147 UnderlyingObjToAccessMap ObjToLastAccess;
4149 // Set of access to check after all writes have been processed.
4150 PtrAccessSet DeferredAccesses;
4152 // Iterate over each alias set twice, once to process read/write pointers,
4153 // and then to process read-only pointers.
4154 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4155 bool UseDeferred = SetIteration > 0;
4156 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4159 Value *Ptr = A.getValue();
4160 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4162 // If we're using the deferred access set, then it contains only reads.
4163 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4164 if (UseDeferred && !IsReadOnlyPtr)
4166 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4168 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4169 S.count(MemAccessInfo(Ptr, false))) &&
4170 "Alias-set pointer not in the access set?");
4172 MemAccessInfo Access(Ptr, IsWrite);
4173 DepCands.insert(Access);
4175 // Memorize read-only pointers for later processing and skip them in the
4176 // first round (they need to be checked after we have seen all write
4177 // pointers). Note: we also mark pointer that are not consecutive as
4178 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4179 // the second check for "!IsWrite".
4180 if (!UseDeferred && IsReadOnlyPtr) {
4181 DeferredAccesses.insert(Access);
4185 // If this is a write - check other reads and writes for conflicts. If
4186 // this is a read only check other writes for conflicts (but only if
4187 // there is no other write to the ptr - this is an optimization to
4188 // catch "a[i] = a[i] + " without having to do a dependence check).
4189 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4190 CheckDeps.insert(Access);
4191 IsRTCheckNeeded = true;
4197 // Create sets of pointers connected by a shared alias set and
4198 // underlying object.
4199 typedef SmallVector<Value*, 16> ValueVector;
4200 ValueVector TempObjects;
4201 GetUnderlyingObjects(Ptr, TempObjects, DL);
4202 for (Value *UnderlyingObj : TempObjects) {
4203 UnderlyingObjToAccessMap::iterator Prev =
4204 ObjToLastAccess.find(UnderlyingObj);
4205 if (Prev != ObjToLastAccess.end())
4206 DepCands.unionSets(Access, Prev->second);
4208 ObjToLastAccess[UnderlyingObj] = Access;
4216 /// \brief Checks memory dependences among accesses to the same underlying
4217 /// object to determine whether there vectorization is legal or not (and at
4218 /// which vectorization factor).
4220 /// This class works under the assumption that we already checked that memory
4221 /// locations with different underlying pointers are "must-not alias".
4222 /// We use the ScalarEvolution framework to symbolically evalutate access
4223 /// functions pairs. Since we currently don't restructure the loop we can rely
4224 /// on the program order of memory accesses to determine their safety.
4225 /// At the moment we will only deem accesses as safe for:
4226 /// * A negative constant distance assuming program order.
4228 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4229 /// a[i] = tmp; y = a[i];
4231 /// The latter case is safe because later checks guarantuee that there can't
4232 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4233 /// the same variable: a header phi can only be an induction or a reduction, a
4234 /// reduction can't have a memory sink, an induction can't have a memory
4235 /// source). This is important and must not be violated (or we have to
4236 /// resort to checking for cycles through memory).
4238 /// * A positive constant distance assuming program order that is bigger
4239 /// than the biggest memory access.
4241 /// tmp = a[i] OR b[i] = x
4242 /// a[i+2] = tmp y = b[i+2];
4244 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4246 /// * Zero distances and all accesses have the same size.
4248 class MemoryDepChecker {
4250 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4251 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4253 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4254 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4255 ShouldRetryWithRuntimeCheck(false) {}
4257 /// \brief Register the location (instructions are given increasing numbers)
4258 /// of a write access.
4259 void addAccess(StoreInst *SI) {
4260 Value *Ptr = SI->getPointerOperand();
4261 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4262 InstMap.push_back(SI);
4266 /// \brief Register the location (instructions are given increasing numbers)
4267 /// of a write access.
4268 void addAccess(LoadInst *LI) {
4269 Value *Ptr = LI->getPointerOperand();
4270 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4271 InstMap.push_back(LI);
4275 /// \brief Check whether the dependencies between the accesses are safe.
4277 /// Only checks sets with elements in \p CheckDeps.
4278 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4279 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4281 /// \brief The maximum number of bytes of a vector register we can vectorize
4282 /// the accesses safely with.
4283 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4285 /// \brief In same cases when the dependency check fails we can still
4286 /// vectorize the loop with a dynamic array access check.
4287 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4290 ScalarEvolution *SE;
4291 const DataLayout *DL;
4292 const Loop *InnermostLoop;
4294 /// \brief Maps access locations (ptr, read/write) to program order.
4295 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4297 /// \brief Memory access instructions in program order.
4298 SmallVector<Instruction *, 16> InstMap;
4300 /// \brief The program order index to be used for the next instruction.
4303 // We can access this many bytes in parallel safely.
4304 unsigned MaxSafeDepDistBytes;
4306 /// \brief If we see a non-constant dependence distance we can still try to
4307 /// vectorize this loop with runtime checks.
4308 bool ShouldRetryWithRuntimeCheck;
4310 /// \brief Check whether there is a plausible dependence between the two
4313 /// Access \p A must happen before \p B in program order. The two indices
4314 /// identify the index into the program order map.
4316 /// This function checks whether there is a plausible dependence (or the
4317 /// absence of such can't be proved) between the two accesses. If there is a
4318 /// plausible dependence but the dependence distance is bigger than one
4319 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4320 /// distance is smaller than any other distance encountered so far).
4321 /// Otherwise, this function returns true signaling a possible dependence.
4322 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4323 const MemAccessInfo &B, unsigned BIdx,
4324 ValueToValueMap &Strides);
4326 /// \brief Check whether the data dependence could prevent store-load
4328 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4331 } // end anonymous namespace
4333 static bool isInBoundsGep(Value *Ptr) {
4334 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4335 return GEP->isInBounds();
4339 /// \brief Check whether the access through \p Ptr has a constant stride.
4340 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4341 const Loop *Lp, ValueToValueMap &StridesMap) {
4342 const Type *Ty = Ptr->getType();
4343 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4345 // Make sure that the pointer does not point to aggregate types.
4346 const PointerType *PtrTy = cast<PointerType>(Ty);
4347 if (PtrTy->getElementType()->isAggregateType()) {
4348 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4353 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4355 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4357 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4358 << *Ptr << " SCEV: " << *PtrScev << "\n");
4362 // The accesss function must stride over the innermost loop.
4363 if (Lp != AR->getLoop()) {
4364 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4365 *Ptr << " SCEV: " << *PtrScev << "\n");
4368 // The address calculation must not wrap. Otherwise, a dependence could be
4370 // An inbounds getelementptr that is a AddRec with a unit stride
4371 // cannot wrap per definition. The unit stride requirement is checked later.
4372 // An getelementptr without an inbounds attribute and unit stride would have
4373 // to access the pointer value "0" which is undefined behavior in address
4374 // space 0, therefore we can also vectorize this case.
4375 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4376 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4377 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4378 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4379 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4380 << *Ptr << " SCEV: " << *PtrScev << "\n");
4384 // Check the step is constant.
4385 const SCEV *Step = AR->getStepRecurrence(*SE);
4387 // Calculate the pointer stride and check if it is consecutive.
4388 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4390 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4391 " SCEV: " << *PtrScev << "\n");
4395 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4396 const APInt &APStepVal = C->getValue()->getValue();
4398 // Huge step value - give up.
4399 if (APStepVal.getBitWidth() > 64)
4402 int64_t StepVal = APStepVal.getSExtValue();
4405 int64_t Stride = StepVal / Size;
4406 int64_t Rem = StepVal % Size;
4410 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4411 // know we can't "wrap around the address space". In case of address space
4412 // zero we know that this won't happen without triggering undefined behavior.
4413 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4414 Stride != 1 && Stride != -1)
4420 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4421 unsigned TypeByteSize) {
4422 // If loads occur at a distance that is not a multiple of a feasible vector
4423 // factor store-load forwarding does not take place.
4424 // Positive dependences might cause troubles because vectorizing them might
4425 // prevent store-load forwarding making vectorized code run a lot slower.
4426 // a[i] = a[i-3] ^ a[i-8];
4427 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4428 // hence on your typical architecture store-load forwarding does not take
4429 // place. Vectorizing in such cases does not make sense.
4430 // Store-load forwarding distance.
4431 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4432 // Maximum vector factor.
4433 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4434 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4435 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4437 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4439 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4440 MaxVFWithoutSLForwardIssues = (vf >>=1);
4445 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4446 DEBUG(dbgs() << "LV: Distance " << Distance <<
4447 " that could cause a store-load forwarding conflict\n");
4451 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4452 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4453 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4457 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4458 const MemAccessInfo &B, unsigned BIdx,
4459 ValueToValueMap &Strides) {
4460 assert (AIdx < BIdx && "Must pass arguments in program order");
4462 Value *APtr = A.getPointer();
4463 Value *BPtr = B.getPointer();
4464 bool AIsWrite = A.getInt();
4465 bool BIsWrite = B.getInt();
4467 // Two reads are independent.
4468 if (!AIsWrite && !BIsWrite)
4471 // We cannot check pointers in different address spaces.
4472 if (APtr->getType()->getPointerAddressSpace() !=
4473 BPtr->getType()->getPointerAddressSpace())
4476 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4477 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4479 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4480 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4482 const SCEV *Src = AScev;
4483 const SCEV *Sink = BScev;
4485 // If the induction step is negative we have to invert source and sink of the
4487 if (StrideAPtr < 0) {
4490 std::swap(APtr, BPtr);
4491 std::swap(Src, Sink);
4492 std::swap(AIsWrite, BIsWrite);
4493 std::swap(AIdx, BIdx);
4494 std::swap(StrideAPtr, StrideBPtr);
4497 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4499 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4500 << "(Induction step: " << StrideAPtr << ")\n");
4501 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4502 << *InstMap[BIdx] << ": " << *Dist << "\n");
4504 // Need consecutive accesses. We don't want to vectorize
4505 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4506 // the address space.
4507 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4508 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4512 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4514 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4515 ShouldRetryWithRuntimeCheck = true;
4519 Type *ATy = APtr->getType()->getPointerElementType();
4520 Type *BTy = BPtr->getType()->getPointerElementType();
4521 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4523 // Negative distances are not plausible dependencies.
4524 const APInt &Val = C->getValue()->getValue();
4525 if (Val.isNegative()) {
4526 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4527 if (IsTrueDataDependence &&
4528 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4532 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4536 // Write to the same location with the same size.
4537 // Could be improved to assert type sizes are the same (i32 == float, etc).
4541 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4545 assert(Val.isStrictlyPositive() && "Expect a positive value");
4547 // Positive distance bigger than max vectorization factor.
4550 "LV: ReadWrite-Write positive dependency with different types\n");
4554 unsigned Distance = (unsigned) Val.getZExtValue();
4556 // Bail out early if passed-in parameters make vectorization not feasible.
4557 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4558 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4560 // The distance must be bigger than the size needed for a vectorized version
4561 // of the operation and the size of the vectorized operation must not be
4562 // bigger than the currrent maximum size.
4563 if (Distance < 2*TypeByteSize ||
4564 2*TypeByteSize > MaxSafeDepDistBytes ||
4565 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4566 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4567 << Val.getSExtValue() << '\n');
4571 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4572 Distance : MaxSafeDepDistBytes;
4574 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4575 if (IsTrueDataDependence &&
4576 couldPreventStoreLoadForward(Distance, TypeByteSize))
4579 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4580 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4585 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4586 MemAccessInfoSet &CheckDeps,
4587 ValueToValueMap &Strides) {
4589 MaxSafeDepDistBytes = -1U;
4590 while (!CheckDeps.empty()) {
4591 MemAccessInfo CurAccess = *CheckDeps.begin();
4593 // Get the relevant memory access set.
4594 EquivalenceClasses<MemAccessInfo>::iterator I =
4595 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4597 // Check accesses within this set.
4598 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4599 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4601 // Check every access pair.
4603 CheckDeps.erase(*AI);
4604 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4606 // Check every accessing instruction pair in program order.
4607 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4608 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4609 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4610 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4611 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4613 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4624 bool LoopVectorizationLegality::canVectorizeMemory() {
4626 typedef SmallVector<Value*, 16> ValueVector;
4627 typedef SmallPtrSet<Value*, 16> ValueSet;
4629 // Holds the Load and Store *instructions*.
4633 // Holds all the different accesses in the loop.
4634 unsigned NumReads = 0;
4635 unsigned NumReadWrites = 0;
4637 PtrRtCheck.Pointers.clear();
4638 PtrRtCheck.Need = false;
4640 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4641 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4644 for (Loop::block_iterator bb = TheLoop->block_begin(),
4645 be = TheLoop->block_end(); bb != be; ++bb) {
4647 // Scan the BB and collect legal loads and stores.
4648 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4651 // If this is a load, save it. If this instruction can read from memory
4652 // but is not a load, then we quit. Notice that we don't handle function
4653 // calls that read or write.
4654 if (it->mayReadFromMemory()) {
4655 // Many math library functions read the rounding mode. We will only
4656 // vectorize a loop if it contains known function calls that don't set
4657 // the flag. Therefore, it is safe to ignore this read from memory.
4658 CallInst *Call = dyn_cast<CallInst>(it);
4659 if (Call && getIntrinsicIDForCall(Call, TLI))
4662 LoadInst *Ld = dyn_cast<LoadInst>(it);
4663 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4664 emitAnalysis(Report(Ld)
4665 << "read with atomic ordering or volatile read");
4666 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4670 Loads.push_back(Ld);
4671 DepChecker.addAccess(Ld);
4675 // Save 'store' instructions. Abort if other instructions write to memory.
4676 if (it->mayWriteToMemory()) {
4677 StoreInst *St = dyn_cast<StoreInst>(it);
4679 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4682 if (!St->isSimple() && !IsAnnotatedParallel) {
4683 emitAnalysis(Report(St)
4684 << "write with atomic ordering or volatile write");
4685 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4689 Stores.push_back(St);
4690 DepChecker.addAccess(St);
4695 // Now we have two lists that hold the loads and the stores.
4696 // Next, we find the pointers that they use.
4698 // Check if we see any stores. If there are no stores, then we don't
4699 // care if the pointers are *restrict*.
4700 if (!Stores.size()) {
4701 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4705 AccessAnalysis::DepCandidates DependentAccesses;
4706 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4708 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4709 // multiple times on the same object. If the ptr is accessed twice, once
4710 // for read and once for write, it will only appear once (on the write
4711 // list). This is okay, since we are going to check for conflicts between
4712 // writes and between reads and writes, but not between reads and reads.
4715 ValueVector::iterator I, IE;
4716 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4717 StoreInst *ST = cast<StoreInst>(*I);
4718 Value* Ptr = ST->getPointerOperand();
4720 if (isUniform(Ptr)) {
4723 << "write to a loop invariant address could not be vectorized");
4724 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4728 // If we did *not* see this pointer before, insert it to the read-write
4729 // list. At this phase it is only a 'write' list.
4730 if (Seen.insert(Ptr)) {
4733 AliasAnalysis::Location Loc = AA->getLocation(ST);
4734 // The TBAA metadata could have a control dependency on the predication
4735 // condition, so we cannot rely on it when determining whether or not we
4736 // need runtime pointer checks.
4737 if (blockNeedsPredication(ST->getParent()))
4738 Loc.TBAATag = nullptr;
4740 Accesses.addStore(Loc);
4744 if (IsAnnotatedParallel) {
4746 << "LV: A loop annotated parallel, ignore memory dependency "
4751 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4752 LoadInst *LD = cast<LoadInst>(*I);
4753 Value* Ptr = LD->getPointerOperand();
4754 // If we did *not* see this pointer before, insert it to the
4755 // read list. If we *did* see it before, then it is already in
4756 // the read-write list. This allows us to vectorize expressions
4757 // such as A[i] += x; Because the address of A[i] is a read-write
4758 // pointer. This only works if the index of A[i] is consecutive.
4759 // If the address of i is unknown (for example A[B[i]]) then we may
4760 // read a few words, modify, and write a few words, and some of the
4761 // words may be written to the same address.
4762 bool IsReadOnlyPtr = false;
4763 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4765 IsReadOnlyPtr = true;
4768 AliasAnalysis::Location Loc = AA->getLocation(LD);
4769 // The TBAA metadata could have a control dependency on the predication
4770 // condition, so we cannot rely on it when determining whether or not we
4771 // need runtime pointer checks.
4772 if (blockNeedsPredication(LD->getParent()))
4773 Loc.TBAATag = nullptr;
4775 Accesses.addLoad(Loc, IsReadOnlyPtr);
4778 // If we write (or read-write) to a single destination and there are no
4779 // other reads in this loop then is it safe to vectorize.
4780 if (NumReadWrites == 1 && NumReads == 0) {
4781 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4785 // Build dependence sets and check whether we need a runtime pointer bounds
4787 Accesses.buildDependenceSets();
4788 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4790 // Find pointers with computable bounds. We are going to use this information
4791 // to place a runtime bound check.
4792 unsigned NumComparisons = 0;
4793 bool CanDoRT = false;
4795 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4798 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4799 " pointer comparisons.\n");
4801 // If we only have one set of dependences to check pointers among we don't
4802 // need a runtime check.
4803 if (NumComparisons == 0 && NeedRTCheck)
4804 NeedRTCheck = false;
4806 // Check that we did not collect too many pointers or found an unsizeable
4808 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4814 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4817 if (NeedRTCheck && !CanDoRT) {
4818 emitAnalysis(Report() << "cannot identify array bounds");
4819 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4820 "the array bounds.\n");
4825 PtrRtCheck.Need = NeedRTCheck;
4827 bool CanVecMem = true;
4828 if (Accesses.isDependencyCheckNeeded()) {
4829 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4830 CanVecMem = DepChecker.areDepsSafe(
4831 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4832 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4834 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4835 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4838 // Clear the dependency checks. We assume they are not needed.
4839 Accesses.resetDepChecks();
4842 PtrRtCheck.Need = true;
4844 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4845 TheLoop, Strides, true);
4846 // Check that we did not collect too many pointers or found an unsizeable
4848 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4849 if (!CanDoRT && NumComparisons > 0)
4850 emitAnalysis(Report()
4851 << "cannot check memory dependencies at runtime");
4853 emitAnalysis(Report()
4854 << NumComparisons << " exceeds limit of "
4855 << RuntimeMemoryCheckThreshold
4856 << " dependent memory operations checked at runtime");
4857 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4867 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4869 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4870 " need a runtime memory check.\n");
4875 static bool hasMultipleUsesOf(Instruction *I,
4876 SmallPtrSet<Instruction *, 8> &Insts) {
4877 unsigned NumUses = 0;
4878 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4879 if (Insts.count(dyn_cast<Instruction>(*Use)))
4888 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4889 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4890 if (!Set.count(dyn_cast<Instruction>(*Use)))
4895 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4896 ReductionKind Kind) {
4897 if (Phi->getNumIncomingValues() != 2)
4900 // Reduction variables are only found in the loop header block.
4901 if (Phi->getParent() != TheLoop->getHeader())
4904 // Obtain the reduction start value from the value that comes from the loop
4906 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4908 // ExitInstruction is the single value which is used outside the loop.
4909 // We only allow for a single reduction value to be used outside the loop.
4910 // This includes users of the reduction, variables (which form a cycle
4911 // which ends in the phi node).
4912 Instruction *ExitInstruction = nullptr;
4913 // Indicates that we found a reduction operation in our scan.
4914 bool FoundReduxOp = false;
4916 // We start with the PHI node and scan for all of the users of this
4917 // instruction. All users must be instructions that can be used as reduction
4918 // variables (such as ADD). We must have a single out-of-block user. The cycle
4919 // must include the original PHI.
4920 bool FoundStartPHI = false;
4922 // To recognize min/max patterns formed by a icmp select sequence, we store
4923 // the number of instruction we saw from the recognized min/max pattern,
4924 // to make sure we only see exactly the two instructions.
4925 unsigned NumCmpSelectPatternInst = 0;
4926 ReductionInstDesc ReduxDesc(false, nullptr);
4928 SmallPtrSet<Instruction *, 8> VisitedInsts;
4929 SmallVector<Instruction *, 8> Worklist;
4930 Worklist.push_back(Phi);
4931 VisitedInsts.insert(Phi);
4933 // A value in the reduction can be used:
4934 // - By the reduction:
4935 // - Reduction operation:
4936 // - One use of reduction value (safe).
4937 // - Multiple use of reduction value (not safe).
4939 // - All uses of the PHI must be the reduction (safe).
4940 // - Otherwise, not safe.
4941 // - By one instruction outside of the loop (safe).
4942 // - By further instructions outside of the loop (not safe).
4943 // - By an instruction that is not part of the reduction (not safe).
4945 // * An instruction type other than PHI or the reduction operation.
4946 // * A PHI in the header other than the initial PHI.
4947 while (!Worklist.empty()) {
4948 Instruction *Cur = Worklist.back();
4949 Worklist.pop_back();
4952 // If the instruction has no users then this is a broken chain and can't be
4953 // a reduction variable.
4954 if (Cur->use_empty())
4957 bool IsAPhi = isa<PHINode>(Cur);
4959 // A header PHI use other than the original PHI.
4960 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4963 // Reductions of instructions such as Div, and Sub is only possible if the
4964 // LHS is the reduction variable.
4965 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4966 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4967 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4970 // Any reduction instruction must be of one of the allowed kinds.
4971 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4972 if (!ReduxDesc.IsReduction)
4975 // A reduction operation must only have one use of the reduction value.
4976 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4977 hasMultipleUsesOf(Cur, VisitedInsts))
4980 // All inputs to a PHI node must be a reduction value.
4981 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4984 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4985 isa<SelectInst>(Cur)))
4986 ++NumCmpSelectPatternInst;
4987 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4988 isa<SelectInst>(Cur)))
4989 ++NumCmpSelectPatternInst;
4991 // Check whether we found a reduction operator.
4992 FoundReduxOp |= !IsAPhi;
4994 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4995 // onto the stack. This way we are going to have seen all inputs to PHI
4996 // nodes once we get to them.
4997 SmallVector<Instruction *, 8> NonPHIs;
4998 SmallVector<Instruction *, 8> PHIs;
4999 for (User *U : Cur->users()) {
5000 Instruction *UI = cast<Instruction>(U);
5002 // Check if we found the exit user.
5003 BasicBlock *Parent = UI->getParent();
5004 if (!TheLoop->contains(Parent)) {
5005 // Exit if you find multiple outside users or if the header phi node is
5006 // being used. In this case the user uses the value of the previous
5007 // iteration, in which case we would loose "VF-1" iterations of the
5008 // reduction operation if we vectorize.
5009 if (ExitInstruction != nullptr || Cur == Phi)
5012 // The instruction used by an outside user must be the last instruction
5013 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5014 // operations on the value.
5015 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5018 ExitInstruction = Cur;
5022 // Process instructions only once (termination). Each reduction cycle
5023 // value must only be used once, except by phi nodes and min/max
5024 // reductions which are represented as a cmp followed by a select.
5025 ReductionInstDesc IgnoredVal(false, nullptr);
5026 if (VisitedInsts.insert(UI)) {
5027 if (isa<PHINode>(UI))
5030 NonPHIs.push_back(UI);
5031 } else if (!isa<PHINode>(UI) &&
5032 ((!isa<FCmpInst>(UI) &&
5033 !isa<ICmpInst>(UI) &&
5034 !isa<SelectInst>(UI)) ||
5035 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5038 // Remember that we completed the cycle.
5040 FoundStartPHI = true;
5042 Worklist.append(PHIs.begin(), PHIs.end());
5043 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5046 // This means we have seen one but not the other instruction of the
5047 // pattern or more than just a select and cmp.
5048 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5049 NumCmpSelectPatternInst != 2)
5052 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5055 // We found a reduction var if we have reached the original phi node and we
5056 // only have a single instruction with out-of-loop users.
5058 // This instruction is allowed to have out-of-loop users.
5059 AllowedExit.insert(ExitInstruction);
5061 // Save the description of this reduction variable.
5062 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5063 ReduxDesc.MinMaxKind);
5064 Reductions[Phi] = RD;
5065 // We've ended the cycle. This is a reduction variable if we have an
5066 // outside user and it has a binary op.
5071 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5072 /// pattern corresponding to a min(X, Y) or max(X, Y).
5073 LoopVectorizationLegality::ReductionInstDesc
5074 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5075 ReductionInstDesc &Prev) {
5077 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5078 "Expect a select instruction");
5079 Instruction *Cmp = nullptr;
5080 SelectInst *Select = nullptr;
5082 // We must handle the select(cmp()) as a single instruction. Advance to the
5084 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5085 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5086 return ReductionInstDesc(false, I);
5087 return ReductionInstDesc(Select, Prev.MinMaxKind);
5090 // Only handle single use cases for now.
5091 if (!(Select = dyn_cast<SelectInst>(I)))
5092 return ReductionInstDesc(false, I);
5093 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5094 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5095 return ReductionInstDesc(false, I);
5096 if (!Cmp->hasOneUse())
5097 return ReductionInstDesc(false, I);
5102 // Look for a min/max pattern.
5103 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5104 return ReductionInstDesc(Select, MRK_UIntMin);
5105 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5106 return ReductionInstDesc(Select, MRK_UIntMax);
5107 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5108 return ReductionInstDesc(Select, MRK_SIntMax);
5109 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5110 return ReductionInstDesc(Select, MRK_SIntMin);
5111 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5112 return ReductionInstDesc(Select, MRK_FloatMin);
5113 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5114 return ReductionInstDesc(Select, MRK_FloatMax);
5115 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5116 return ReductionInstDesc(Select, MRK_FloatMin);
5117 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5118 return ReductionInstDesc(Select, MRK_FloatMax);
5120 return ReductionInstDesc(false, I);
5123 LoopVectorizationLegality::ReductionInstDesc
5124 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5126 ReductionInstDesc &Prev) {
5127 bool FP = I->getType()->isFloatingPointTy();
5128 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
5129 switch (I->getOpcode()) {
5131 return ReductionInstDesc(false, I);
5132 case Instruction::PHI:
5133 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5134 Kind != RK_FloatMinMax))
5135 return ReductionInstDesc(false, I);
5136 return ReductionInstDesc(I, Prev.MinMaxKind);
5137 case Instruction::Sub:
5138 case Instruction::Add:
5139 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5140 case Instruction::Mul:
5141 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5142 case Instruction::And:
5143 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5144 case Instruction::Or:
5145 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5146 case Instruction::Xor:
5147 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5148 case Instruction::FMul:
5149 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5150 case Instruction::FAdd:
5151 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5152 case Instruction::FCmp:
5153 case Instruction::ICmp:
5154 case Instruction::Select:
5155 if (Kind != RK_IntegerMinMax &&
5156 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5157 return ReductionInstDesc(false, I);
5158 return isMinMaxSelectCmpPattern(I, Prev);
5162 LoopVectorizationLegality::InductionKind
5163 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5164 Type *PhiTy = Phi->getType();
5165 // We only handle integer and pointer inductions variables.
5166 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5167 return IK_NoInduction;
5169 // Check that the PHI is consecutive.
5170 const SCEV *PhiScev = SE->getSCEV(Phi);
5171 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5173 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5174 return IK_NoInduction;
5176 const SCEV *Step = AR->getStepRecurrence(*SE);
5178 // Integer inductions need to have a stride of one.
5179 if (PhiTy->isIntegerTy()) {
5181 return IK_IntInduction;
5182 if (Step->isAllOnesValue())
5183 return IK_ReverseIntInduction;
5184 return IK_NoInduction;
5187 // Calculate the pointer stride and check if it is consecutive.
5188 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5190 return IK_NoInduction;
5192 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5193 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5194 if (C->getValue()->equalsInt(Size))
5195 return IK_PtrInduction;
5196 else if (C->getValue()->equalsInt(0 - Size))
5197 return IK_ReversePtrInduction;
5199 return IK_NoInduction;
5202 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5203 Value *In0 = const_cast<Value*>(V);
5204 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5208 return Inductions.count(PN);
5211 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5212 assert(TheLoop->contains(BB) && "Unknown block used");
5214 // Blocks that do not dominate the latch need predication.
5215 BasicBlock* Latch = TheLoop->getLoopLatch();
5216 return !DT->dominates(BB, Latch);
5219 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5220 SmallPtrSet<Value *, 8>& SafePtrs) {
5221 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5222 // We might be able to hoist the load.
5223 if (it->mayReadFromMemory()) {
5224 LoadInst *LI = dyn_cast<LoadInst>(it);
5225 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5229 // We don't predicate stores at the moment.
5230 if (it->mayWriteToMemory()) {
5231 StoreInst *SI = dyn_cast<StoreInst>(it);
5232 // We only support predication of stores in basic blocks with one
5234 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5235 !SafePtrs.count(SI->getPointerOperand()) ||
5236 !SI->getParent()->getSinglePredecessor())
5242 // Check that we don't have a constant expression that can trap as operand.
5243 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5245 if (Constant *C = dyn_cast<Constant>(*OI))
5250 // The instructions below can trap.
5251 switch (it->getOpcode()) {
5253 case Instruction::UDiv:
5254 case Instruction::SDiv:
5255 case Instruction::URem:
5256 case Instruction::SRem:
5264 LoopVectorizationCostModel::VectorizationFactor
5265 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5267 bool ForceVectorization) {
5268 // Width 1 means no vectorize
5269 VectorizationFactor Factor = { 1U, 0U };
5270 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5271 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5275 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5276 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5280 // Find the trip count.
5281 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5282 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5284 unsigned WidestType = getWidestType();
5285 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5286 unsigned MaxSafeDepDist = -1U;
5287 if (Legal->getMaxSafeDepDistBytes() != -1U)
5288 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5289 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5290 WidestRegister : MaxSafeDepDist);
5291 unsigned MaxVectorSize = WidestRegister / WidestType;
5292 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5293 DEBUG(dbgs() << "LV: The Widest register is: "
5294 << WidestRegister << " bits.\n");
5296 if (MaxVectorSize == 0) {
5297 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5301 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5302 " into one vector!");
5304 unsigned VF = MaxVectorSize;
5306 // If we optimize the program for size, avoid creating the tail loop.
5308 // If we are unable to calculate the trip count then don't try to vectorize.
5310 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5314 // Find the maximum SIMD width that can fit within the trip count.
5315 VF = TC % MaxVectorSize;
5320 // If the trip count that we found modulo the vectorization factor is not
5321 // zero then we require a tail.
5323 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5329 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5330 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5332 Factor.Width = UserVF;
5336 float Cost = expectedCost(1);
5338 const float ScalarCost = Cost;
5341 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5343 // Ignore scalar width, because the user explicitly wants vectorization.
5344 if (ForceVectorization && VF > 1) {
5346 Cost = expectedCost(Width) / (float)Width;
5349 for (unsigned i=2; i <= VF; i*=2) {
5350 // Notice that the vector loop needs to be executed less times, so
5351 // we need to divide the cost of the vector loops by the width of
5352 // the vector elements.
5353 float VectorCost = expectedCost(i) / (float)i;
5354 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5355 (int)VectorCost << ".\n");
5356 if (VectorCost < Cost) {
5362 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5363 << "LV: Vectorization seems to be not beneficial, "
5364 << "but was forced by a user.\n");
5365 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5366 Factor.Width = Width;
5367 Factor.Cost = Width * Cost;
5371 unsigned LoopVectorizationCostModel::getWidestType() {
5372 unsigned MaxWidth = 8;
5375 for (Loop::block_iterator bb = TheLoop->block_begin(),
5376 be = TheLoop->block_end(); bb != be; ++bb) {
5377 BasicBlock *BB = *bb;
5379 // For each instruction in the loop.
5380 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5381 Type *T = it->getType();
5383 // Only examine Loads, Stores and PHINodes.
5384 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5387 // Examine PHI nodes that are reduction variables.
5388 if (PHINode *PN = dyn_cast<PHINode>(it))
5389 if (!Legal->getReductionVars()->count(PN))
5392 // Examine the stored values.
5393 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5394 T = ST->getValueOperand()->getType();
5396 // Ignore loaded pointer types and stored pointer types that are not
5397 // consecutive. However, we do want to take consecutive stores/loads of
5398 // pointer vectors into account.
5399 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5402 MaxWidth = std::max(MaxWidth,
5403 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5411 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5414 unsigned LoopCost) {
5416 // -- The unroll heuristics --
5417 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5418 // There are many micro-architectural considerations that we can't predict
5419 // at this level. For example frontend pressure (on decode or fetch) due to
5420 // code size, or the number and capabilities of the execution ports.
5422 // We use the following heuristics to select the unroll factor:
5423 // 1. If the code has reductions the we unroll in order to break the cross
5424 // iteration dependency.
5425 // 2. If the loop is really small then we unroll in order to reduce the loop
5427 // 3. We don't unroll if we think that we will spill registers to memory due
5428 // to the increased register pressure.
5430 // Use the user preference, unless 'auto' is selected.
5434 // When we optimize for size we don't unroll.
5438 // We used the distance for the unroll factor.
5439 if (Legal->getMaxSafeDepDistBytes() != -1U)
5442 // Do not unroll loops with a relatively small trip count.
5443 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5444 TheLoop->getLoopLatch());
5445 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5448 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5449 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5453 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5454 TargetNumRegisters = ForceTargetNumScalarRegs;
5456 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5457 TargetNumRegisters = ForceTargetNumVectorRegs;
5460 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5461 // We divide by these constants so assume that we have at least one
5462 // instruction that uses at least one register.
5463 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5464 R.NumInstructions = std::max(R.NumInstructions, 1U);
5466 // We calculate the unroll factor using the following formula.
5467 // Subtract the number of loop invariants from the number of available
5468 // registers. These registers are used by all of the unrolled instances.
5469 // Next, divide the remaining registers by the number of registers that is
5470 // required by the loop, in order to estimate how many parallel instances
5471 // fit without causing spills. All of this is rounded down if necessary to be
5472 // a power of two. We want power of two unroll factors to simplify any
5473 // addressing operations or alignment considerations.
5474 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5477 // Don't count the induction variable as unrolled.
5478 if (EnableIndVarRegisterHeur)
5479 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5480 std::max(1U, (R.MaxLocalUsers - 1)));
5482 // Clamp the unroll factor ranges to reasonable factors.
5483 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5485 // Check if the user has overridden the unroll max.
5487 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5488 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5490 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5491 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5494 // If we did not calculate the cost for VF (because the user selected the VF)
5495 // then we calculate the cost of VF here.
5497 LoopCost = expectedCost(VF);
5499 // Clamp the calculated UF to be between the 1 and the max unroll factor
5500 // that the target allows.
5501 if (UF > MaxUnrollSize)
5506 // Unroll if we vectorized this loop and there is a reduction that could
5507 // benefit from unrolling.
5508 if (VF > 1 && Legal->getReductionVars()->size()) {
5509 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5513 // Note that if we've already vectorized the loop we will have done the
5514 // runtime check and so unrolling won't require further checks.
5515 bool UnrollingRequiresRuntimePointerCheck =
5516 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5518 // We want to unroll small loops in order to reduce the loop overhead and
5519 // potentially expose ILP opportunities.
5520 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5521 if (!UnrollingRequiresRuntimePointerCheck &&
5522 LoopCost < SmallLoopCost) {
5523 // We assume that the cost overhead is 1 and we use the cost model
5524 // to estimate the cost of the loop and unroll until the cost of the
5525 // loop overhead is about 5% of the cost of the loop.
5526 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5528 // Unroll until store/load ports (estimated by max unroll factor) are
5530 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5531 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5533 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5534 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5535 return std::max(StoresUF, LoadsUF);
5538 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5542 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5546 LoopVectorizationCostModel::RegisterUsage
5547 LoopVectorizationCostModel::calculateRegisterUsage() {
5548 // This function calculates the register usage by measuring the highest number
5549 // of values that are alive at a single location. Obviously, this is a very
5550 // rough estimation. We scan the loop in a topological order in order and
5551 // assign a number to each instruction. We use RPO to ensure that defs are
5552 // met before their users. We assume that each instruction that has in-loop
5553 // users starts an interval. We record every time that an in-loop value is
5554 // used, so we have a list of the first and last occurrences of each
5555 // instruction. Next, we transpose this data structure into a multi map that
5556 // holds the list of intervals that *end* at a specific location. This multi
5557 // map allows us to perform a linear search. We scan the instructions linearly
5558 // and record each time that a new interval starts, by placing it in a set.
5559 // If we find this value in the multi-map then we remove it from the set.
5560 // The max register usage is the maximum size of the set.
5561 // We also search for instructions that are defined outside the loop, but are
5562 // used inside the loop. We need this number separately from the max-interval
5563 // usage number because when we unroll, loop-invariant values do not take
5565 LoopBlocksDFS DFS(TheLoop);
5569 R.NumInstructions = 0;
5571 // Each 'key' in the map opens a new interval. The values
5572 // of the map are the index of the 'last seen' usage of the
5573 // instruction that is the key.
5574 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5575 // Maps instruction to its index.
5576 DenseMap<unsigned, Instruction*> IdxToInstr;
5577 // Marks the end of each interval.
5578 IntervalMap EndPoint;
5579 // Saves the list of instruction indices that are used in the loop.
5580 SmallSet<Instruction*, 8> Ends;
5581 // Saves the list of values that are used in the loop but are
5582 // defined outside the loop, such as arguments and constants.
5583 SmallPtrSet<Value*, 8> LoopInvariants;
5586 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5587 be = DFS.endRPO(); bb != be; ++bb) {
5588 R.NumInstructions += (*bb)->size();
5589 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5591 Instruction *I = it;
5592 IdxToInstr[Index++] = I;
5594 // Save the end location of each USE.
5595 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5596 Value *U = I->getOperand(i);
5597 Instruction *Instr = dyn_cast<Instruction>(U);
5599 // Ignore non-instruction values such as arguments, constants, etc.
5600 if (!Instr) continue;
5602 // If this instruction is outside the loop then record it and continue.
5603 if (!TheLoop->contains(Instr)) {
5604 LoopInvariants.insert(Instr);
5608 // Overwrite previous end points.
5609 EndPoint[Instr] = Index;
5615 // Saves the list of intervals that end with the index in 'key'.
5616 typedef SmallVector<Instruction*, 2> InstrList;
5617 DenseMap<unsigned, InstrList> TransposeEnds;
5619 // Transpose the EndPoints to a list of values that end at each index.
5620 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5622 TransposeEnds[it->second].push_back(it->first);
5624 SmallSet<Instruction*, 8> OpenIntervals;
5625 unsigned MaxUsage = 0;
5628 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5629 for (unsigned int i = 0; i < Index; ++i) {
5630 Instruction *I = IdxToInstr[i];
5631 // Ignore instructions that are never used within the loop.
5632 if (!Ends.count(I)) continue;
5634 // Remove all of the instructions that end at this location.
5635 InstrList &List = TransposeEnds[i];
5636 for (unsigned int j=0, e = List.size(); j < e; ++j)
5637 OpenIntervals.erase(List[j]);
5639 // Count the number of live interals.
5640 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5642 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5643 OpenIntervals.size() << '\n');
5645 // Add the current instruction to the list of open intervals.
5646 OpenIntervals.insert(I);
5649 unsigned Invariant = LoopInvariants.size();
5650 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5651 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5652 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5654 R.LoopInvariantRegs = Invariant;
5655 R.MaxLocalUsers = MaxUsage;
5659 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5663 for (Loop::block_iterator bb = TheLoop->block_begin(),
5664 be = TheLoop->block_end(); bb != be; ++bb) {
5665 unsigned BlockCost = 0;
5666 BasicBlock *BB = *bb;
5668 // For each instruction in the old loop.
5669 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5670 // Skip dbg intrinsics.
5671 if (isa<DbgInfoIntrinsic>(it))
5674 unsigned C = getInstructionCost(it, VF);
5676 // Check if we should override the cost.
5677 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5678 C = ForceTargetInstructionCost;
5681 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5682 VF << " For instruction: " << *it << '\n');
5685 // We assume that if-converted blocks have a 50% chance of being executed.
5686 // When the code is scalar then some of the blocks are avoided due to CF.
5687 // When the code is vectorized we execute all code paths.
5688 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5697 /// \brief Check whether the address computation for a non-consecutive memory
5698 /// access looks like an unlikely candidate for being merged into the indexing
5701 /// We look for a GEP which has one index that is an induction variable and all
5702 /// other indices are loop invariant. If the stride of this access is also
5703 /// within a small bound we decide that this address computation can likely be
5704 /// merged into the addressing mode.
5705 /// In all other cases, we identify the address computation as complex.
5706 static bool isLikelyComplexAddressComputation(Value *Ptr,
5707 LoopVectorizationLegality *Legal,
5708 ScalarEvolution *SE,
5709 const Loop *TheLoop) {
5710 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5714 // We are looking for a gep with all loop invariant indices except for one
5715 // which should be an induction variable.
5716 unsigned NumOperands = Gep->getNumOperands();
5717 for (unsigned i = 1; i < NumOperands; ++i) {
5718 Value *Opd = Gep->getOperand(i);
5719 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5720 !Legal->isInductionVariable(Opd))
5724 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5725 // can likely be merged into the address computation.
5726 unsigned MaxMergeDistance = 64;
5728 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5732 // Check the step is constant.
5733 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5734 // Calculate the pointer stride and check if it is consecutive.
5735 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5739 const APInt &APStepVal = C->getValue()->getValue();
5741 // Huge step value - give up.
5742 if (APStepVal.getBitWidth() > 64)
5745 int64_t StepVal = APStepVal.getSExtValue();
5747 return StepVal > MaxMergeDistance;
5750 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5751 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5757 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5758 // If we know that this instruction will remain uniform, check the cost of
5759 // the scalar version.
5760 if (Legal->isUniformAfterVectorization(I))
5763 Type *RetTy = I->getType();
5764 Type *VectorTy = ToVectorTy(RetTy, VF);
5766 // TODO: We need to estimate the cost of intrinsic calls.
5767 switch (I->getOpcode()) {
5768 case Instruction::GetElementPtr:
5769 // We mark this instruction as zero-cost because the cost of GEPs in
5770 // vectorized code depends on whether the corresponding memory instruction
5771 // is scalarized or not. Therefore, we handle GEPs with the memory
5772 // instruction cost.
5774 case Instruction::Br: {
5775 return TTI.getCFInstrCost(I->getOpcode());
5777 case Instruction::PHI:
5778 //TODO: IF-converted IFs become selects.
5780 case Instruction::Add:
5781 case Instruction::FAdd:
5782 case Instruction::Sub:
5783 case Instruction::FSub:
5784 case Instruction::Mul:
5785 case Instruction::FMul:
5786 case Instruction::UDiv:
5787 case Instruction::SDiv:
5788 case Instruction::FDiv:
5789 case Instruction::URem:
5790 case Instruction::SRem:
5791 case Instruction::FRem:
5792 case Instruction::Shl:
5793 case Instruction::LShr:
5794 case Instruction::AShr:
5795 case Instruction::And:
5796 case Instruction::Or:
5797 case Instruction::Xor: {
5798 // Since we will replace the stride by 1 the multiplication should go away.
5799 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5801 // Certain instructions can be cheaper to vectorize if they have a constant
5802 // second vector operand. One example of this are shifts on x86.
5803 TargetTransformInfo::OperandValueKind Op1VK =
5804 TargetTransformInfo::OK_AnyValue;
5805 TargetTransformInfo::OperandValueKind Op2VK =
5806 TargetTransformInfo::OK_AnyValue;
5807 Value *Op2 = I->getOperand(1);
5809 // Check for a splat of a constant or for a non uniform vector of constants.
5810 if (isa<ConstantInt>(Op2))
5811 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5812 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5813 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5814 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5815 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5818 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5820 case Instruction::Select: {
5821 SelectInst *SI = cast<SelectInst>(I);
5822 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5823 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5824 Type *CondTy = SI->getCondition()->getType();
5826 CondTy = VectorType::get(CondTy, VF);
5828 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5830 case Instruction::ICmp:
5831 case Instruction::FCmp: {
5832 Type *ValTy = I->getOperand(0)->getType();
5833 VectorTy = ToVectorTy(ValTy, VF);
5834 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5836 case Instruction::Store:
5837 case Instruction::Load: {
5838 StoreInst *SI = dyn_cast<StoreInst>(I);
5839 LoadInst *LI = dyn_cast<LoadInst>(I);
5840 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5842 VectorTy = ToVectorTy(ValTy, VF);
5844 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5845 unsigned AS = SI ? SI->getPointerAddressSpace() :
5846 LI->getPointerAddressSpace();
5847 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5848 // We add the cost of address computation here instead of with the gep
5849 // instruction because only here we know whether the operation is
5852 return TTI.getAddressComputationCost(VectorTy) +
5853 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5855 // Scalarized loads/stores.
5856 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5857 bool Reverse = ConsecutiveStride < 0;
5858 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5859 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5860 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5861 bool IsComplexComputation =
5862 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5864 // The cost of extracting from the value vector and pointer vector.
5865 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5866 for (unsigned i = 0; i < VF; ++i) {
5867 // The cost of extracting the pointer operand.
5868 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5869 // In case of STORE, the cost of ExtractElement from the vector.
5870 // In case of LOAD, the cost of InsertElement into the returned
5872 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5873 Instruction::InsertElement,
5877 // The cost of the scalar loads/stores.
5878 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5879 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5884 // Wide load/stores.
5885 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5886 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5889 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5893 case Instruction::ZExt:
5894 case Instruction::SExt:
5895 case Instruction::FPToUI:
5896 case Instruction::FPToSI:
5897 case Instruction::FPExt:
5898 case Instruction::PtrToInt:
5899 case Instruction::IntToPtr:
5900 case Instruction::SIToFP:
5901 case Instruction::UIToFP:
5902 case Instruction::Trunc:
5903 case Instruction::FPTrunc:
5904 case Instruction::BitCast: {
5905 // We optimize the truncation of induction variable.
5906 // The cost of these is the same as the scalar operation.
5907 if (I->getOpcode() == Instruction::Trunc &&
5908 Legal->isInductionVariable(I->getOperand(0)))
5909 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5910 I->getOperand(0)->getType());
5912 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5913 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5915 case Instruction::Call: {
5916 CallInst *CI = cast<CallInst>(I);
5917 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5918 assert(ID && "Not an intrinsic call!");
5919 Type *RetTy = ToVectorTy(CI->getType(), VF);
5920 SmallVector<Type*, 4> Tys;
5921 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5922 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5923 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5926 // We are scalarizing the instruction. Return the cost of the scalar
5927 // instruction, plus the cost of insert and extract into vector
5928 // elements, times the vector width.
5931 if (!RetTy->isVoidTy() && VF != 1) {
5932 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5934 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5937 // The cost of inserting the results plus extracting each one of the
5939 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5942 // The cost of executing VF copies of the scalar instruction. This opcode
5943 // is unknown. Assume that it is the same as 'mul'.
5944 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5950 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5951 if (Scalar->isVoidTy() || VF == 1)
5953 return VectorType::get(Scalar, VF);
5956 char LoopVectorize::ID = 0;
5957 static const char lv_name[] = "Loop Vectorization";
5958 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5959 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5960 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5961 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5962 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5963 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5964 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5965 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5966 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5967 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5970 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5971 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5975 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5976 // Check for a store.
5977 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5978 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5980 // Check for a load.
5981 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5982 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5988 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5989 bool IfPredicateStore) {
5990 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5991 // Holds vector parameters or scalars, in case of uniform vals.
5992 SmallVector<VectorParts, 4> Params;
5994 setDebugLocFromInst(Builder, Instr);
5996 // Find all of the vectorized parameters.
5997 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5998 Value *SrcOp = Instr->getOperand(op);
6000 // If we are accessing the old induction variable, use the new one.
6001 if (SrcOp == OldInduction) {
6002 Params.push_back(getVectorValue(SrcOp));
6006 // Try using previously calculated values.
6007 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6009 // If the src is an instruction that appeared earlier in the basic block
6010 // then it should already be vectorized.
6011 if (SrcInst && OrigLoop->contains(SrcInst)) {
6012 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6013 // The parameter is a vector value from earlier.
6014 Params.push_back(WidenMap.get(SrcInst));
6016 // The parameter is a scalar from outside the loop. Maybe even a constant.
6017 VectorParts Scalars;
6018 Scalars.append(UF, SrcOp);
6019 Params.push_back(Scalars);
6023 assert(Params.size() == Instr->getNumOperands() &&
6024 "Invalid number of operands");
6026 // Does this instruction return a value ?
6027 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6029 Value *UndefVec = IsVoidRetTy ? nullptr :
6030 UndefValue::get(Instr->getType());
6031 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6032 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6034 Instruction *InsertPt = Builder.GetInsertPoint();
6035 BasicBlock *IfBlock = Builder.GetInsertBlock();
6036 BasicBlock *CondBlock = nullptr;
6039 Loop *VectorLp = nullptr;
6040 if (IfPredicateStore) {
6041 assert(Instr->getParent()->getSinglePredecessor() &&
6042 "Only support single predecessor blocks");
6043 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6044 Instr->getParent());
6045 VectorLp = LI->getLoopFor(IfBlock);
6046 assert(VectorLp && "Must have a loop for this block");
6049 // For each vector unroll 'part':
6050 for (unsigned Part = 0; Part < UF; ++Part) {
6051 // For each scalar that we create:
6053 // Start an "if (pred) a[i] = ..." block.
6054 Value *Cmp = nullptr;
6055 if (IfPredicateStore) {
6056 if (Cond[Part]->getType()->isVectorTy())
6058 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6059 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6060 ConstantInt::get(Cond[Part]->getType(), 1));
6061 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6062 LoopVectorBody.push_back(CondBlock);
6063 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6064 // Update Builder with newly created basic block.
6065 Builder.SetInsertPoint(InsertPt);
6068 Instruction *Cloned = Instr->clone();
6070 Cloned->setName(Instr->getName() + ".cloned");
6071 // Replace the operands of the cloned instructions with extracted scalars.
6072 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6073 Value *Op = Params[op][Part];
6074 Cloned->setOperand(op, Op);
6077 // Place the cloned scalar in the new loop.
6078 Builder.Insert(Cloned);
6080 // If the original scalar returns a value we need to place it in a vector
6081 // so that future users will be able to use it.
6083 VecResults[Part] = Cloned;
6086 if (IfPredicateStore) {
6087 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6088 LoopVectorBody.push_back(NewIfBlock);
6089 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6090 Builder.SetInsertPoint(InsertPt);
6091 Instruction *OldBr = IfBlock->getTerminator();
6092 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6093 OldBr->eraseFromParent();
6094 IfBlock = NewIfBlock;
6099 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6100 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6101 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6103 return scalarizeInstruction(Instr, IfPredicateStore);
6106 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6110 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6114 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6116 // When unrolling and the VF is 1, we only need to add a simple scalar.
6117 Type *ITy = Val->getType();
6118 assert(!ITy->isVectorTy() && "Val must be a scalar");
6119 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6120 return Builder.CreateAdd(Val, C, "induction");