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_alias_scope &&
539 Kind != LLVMContext::MD_noalias &&
540 Kind != LLVMContext::MD_fpmath)
543 To->setMetadata(Kind, M.second);
547 /// \brief Propagate known metadata from one instruction to a vector of others.
548 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
550 if (Instruction *I = dyn_cast<Instruction>(V))
551 propagateMetadata(I, From);
554 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
555 /// to what vectorization factor.
556 /// This class does not look at the profitability of vectorization, only the
557 /// legality. This class has two main kinds of checks:
558 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
559 /// will change the order of memory accesses in a way that will change the
560 /// correctness of the program.
561 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
562 /// checks for a number of different conditions, such as the availability of a
563 /// single induction variable, that all types are supported and vectorize-able,
564 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
565 /// This class is also used by InnerLoopVectorizer for identifying
566 /// induction variable and the different reduction variables.
567 class LoopVectorizationLegality {
571 unsigned NumPredStores;
573 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
574 DominatorTree *DT, TargetLibraryInfo *TLI,
575 AliasAnalysis *AA, Function *F)
576 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
577 DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
578 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
581 /// This enum represents the kinds of reductions that we support.
583 RK_NoReduction, ///< Not a reduction.
584 RK_IntegerAdd, ///< Sum of integers.
585 RK_IntegerMult, ///< Product of integers.
586 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
587 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
588 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
589 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
590 RK_FloatAdd, ///< Sum of floats.
591 RK_FloatMult, ///< Product of floats.
592 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
595 /// This enum represents the kinds of inductions that we support.
597 IK_NoInduction, ///< Not an induction variable.
598 IK_IntInduction, ///< Integer induction variable. Step = 1.
599 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
600 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
601 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
604 // This enum represents the kind of minmax reduction.
605 enum MinMaxReductionKind {
615 /// This struct holds information about reduction variables.
616 struct ReductionDescriptor {
617 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
618 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
620 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
621 MinMaxReductionKind MK)
622 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
624 // The starting value of the reduction.
625 // It does not have to be zero!
626 TrackingVH<Value> StartValue;
627 // The instruction who's value is used outside the loop.
628 Instruction *LoopExitInstr;
629 // The kind of the reduction.
631 // If this a min/max reduction the kind of reduction.
632 MinMaxReductionKind MinMaxKind;
635 /// This POD struct holds information about a potential reduction operation.
636 struct ReductionInstDesc {
637 ReductionInstDesc(bool IsRedux, Instruction *I) :
638 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
640 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
641 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
643 // Is this instruction a reduction candidate.
645 // The last instruction in a min/max pattern (select of the select(icmp())
646 // pattern), or the current reduction instruction otherwise.
647 Instruction *PatternLastInst;
648 // If this is a min/max pattern the comparison predicate.
649 MinMaxReductionKind MinMaxKind;
652 /// This struct holds information about the memory runtime legality
653 /// check that a group of pointers do not overlap.
654 struct RuntimePointerCheck {
655 RuntimePointerCheck() : Need(false) {}
657 /// Reset the state of the pointer runtime information.
664 DependencySetId.clear();
668 /// Insert a pointer and calculate the start and end SCEVs.
669 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
670 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
672 /// This flag indicates if we need to add the runtime check.
674 /// Holds the pointers that we need to check.
675 SmallVector<TrackingVH<Value>, 2> Pointers;
676 /// Holds the pointer value at the beginning of the loop.
677 SmallVector<const SCEV*, 2> Starts;
678 /// Holds the pointer value at the end of the loop.
679 SmallVector<const SCEV*, 2> Ends;
680 /// Holds the information if this pointer is used for writing to memory.
681 SmallVector<bool, 2> IsWritePtr;
682 /// Holds the id of the set of pointers that could be dependent because of a
683 /// shared underlying object.
684 SmallVector<unsigned, 2> DependencySetId;
685 /// Holds the id of the disjoint alias set to which this pointer belongs.
686 SmallVector<unsigned, 2> AliasSetId;
689 /// A struct for saving information about induction variables.
690 struct InductionInfo {
691 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
692 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
694 TrackingVH<Value> StartValue;
699 /// ReductionList contains the reduction descriptors for all
700 /// of the reductions that were found in the loop.
701 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
703 /// InductionList saves induction variables and maps them to the
704 /// induction descriptor.
705 typedef MapVector<PHINode*, InductionInfo> InductionList;
707 /// Returns true if it is legal to vectorize this loop.
708 /// This does not mean that it is profitable to vectorize this
709 /// loop, only that it is legal to do so.
712 /// Returns the Induction variable.
713 PHINode *getInduction() { return Induction; }
715 /// Returns the reduction variables found in the loop.
716 ReductionList *getReductionVars() { return &Reductions; }
718 /// Returns the induction variables found in the loop.
719 InductionList *getInductionVars() { return &Inductions; }
721 /// Returns the widest induction type.
722 Type *getWidestInductionType() { return WidestIndTy; }
724 /// Returns True if V is an induction variable in this loop.
725 bool isInductionVariable(const Value *V);
727 /// Return true if the block BB needs to be predicated in order for the loop
728 /// to be vectorized.
729 bool blockNeedsPredication(BasicBlock *BB);
731 /// Check if this pointer is consecutive when vectorizing. This happens
732 /// when the last index of the GEP is the induction variable, or that the
733 /// pointer itself is an induction variable.
734 /// This check allows us to vectorize A[idx] into a wide load/store.
736 /// 0 - Stride is unknown or non-consecutive.
737 /// 1 - Address is consecutive.
738 /// -1 - Address is consecutive, and decreasing.
739 int isConsecutivePtr(Value *Ptr);
741 /// Returns true if the value V is uniform within the loop.
742 bool isUniform(Value *V);
744 /// Returns true if this instruction will remain scalar after vectorization.
745 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
747 /// Returns the information that we collected about runtime memory check.
748 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
750 /// This function returns the identity element (or neutral element) for
752 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
754 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
756 bool hasStride(Value *V) { return StrideSet.count(V); }
757 bool mustCheckStrides() { return !StrideSet.empty(); }
758 SmallPtrSet<Value *, 8>::iterator strides_begin() {
759 return StrideSet.begin();
761 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
764 /// Check if a single basic block loop is vectorizable.
765 /// At this point we know that this is a loop with a constant trip count
766 /// and we only need to check individual instructions.
767 bool canVectorizeInstrs();
769 /// When we vectorize loops we may change the order in which
770 /// we read and write from memory. This method checks if it is
771 /// legal to vectorize the code, considering only memory constrains.
772 /// Returns true if the loop is vectorizable
773 bool canVectorizeMemory();
775 /// Return true if we can vectorize this loop using the IF-conversion
777 bool canVectorizeWithIfConvert();
779 /// Collect the variables that need to stay uniform after vectorization.
780 void collectLoopUniforms();
782 /// Return true if all of the instructions in the block can be speculatively
783 /// executed. \p SafePtrs is a list of addresses that are known to be legal
784 /// and we know that we can read from them without segfault.
785 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
787 /// Returns True, if 'Phi' is the kind of reduction variable for type
788 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
789 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
790 /// Returns a struct describing if the instruction 'I' can be a reduction
791 /// variable of type 'Kind'. If the reduction is a min/max pattern of
792 /// select(icmp()) this function advances the instruction pointer 'I' from the
793 /// compare instruction to the select instruction and stores this pointer in
794 /// 'PatternLastInst' member of the returned struct.
795 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
796 ReductionInstDesc &Desc);
797 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
798 /// pattern corresponding to a min(X, Y) or max(X, Y).
799 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
800 ReductionInstDesc &Prev);
801 /// Returns the induction kind of Phi. This function may return NoInduction
802 /// if the PHI is not an induction variable.
803 InductionKind isInductionVariable(PHINode *Phi);
805 /// \brief Collect memory access with loop invariant strides.
807 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
809 void collectStridedAcccess(Value *LoadOrStoreInst);
811 /// Report an analysis message to assist the user in diagnosing loops that are
813 void emitAnalysis(Report &Message) {
814 DebugLoc DL = TheLoop->getStartLoc();
815 if (Instruction *I = Message.getInstr())
816 DL = I->getDebugLoc();
817 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
818 *TheFunction, DL, Message.str());
821 /// The loop that we evaluate.
825 /// DataLayout analysis.
826 const DataLayout *DL;
829 /// Target Library Info.
830 TargetLibraryInfo *TLI;
834 Function *TheFunction;
836 // --- vectorization state --- //
838 /// Holds the integer induction variable. This is the counter of the
841 /// Holds the reduction variables.
842 ReductionList Reductions;
843 /// Holds all of the induction variables that we found in the loop.
844 /// Notice that inductions don't need to start at zero and that induction
845 /// variables can be pointers.
846 InductionList Inductions;
847 /// Holds the widest induction type encountered.
850 /// Allowed outside users. This holds the reduction
851 /// vars which can be accessed from outside the loop.
852 SmallPtrSet<Value*, 4> AllowedExit;
853 /// This set holds the variables which are known to be uniform after
855 SmallPtrSet<Instruction*, 4> Uniforms;
856 /// We need to check that all of the pointers in this list are disjoint
858 RuntimePointerCheck PtrRtCheck;
859 /// Can we assume the absence of NaNs.
860 bool HasFunNoNaNAttr;
862 unsigned MaxSafeDepDistBytes;
864 ValueToValueMap Strides;
865 SmallPtrSet<Value *, 8> StrideSet;
868 /// LoopVectorizationCostModel - estimates the expected speedups due to
870 /// In many cases vectorization is not profitable. This can happen because of
871 /// a number of reasons. In this class we mainly attempt to predict the
872 /// expected speedup/slowdowns due to the supported instruction set. We use the
873 /// TargetTransformInfo to query the different backends for the cost of
874 /// different operations.
875 class LoopVectorizationCostModel {
877 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
878 LoopVectorizationLegality *Legal,
879 const TargetTransformInfo &TTI,
880 const DataLayout *DL, const TargetLibraryInfo *TLI)
881 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
883 /// Information about vectorization costs
884 struct VectorizationFactor {
885 unsigned Width; // Vector width with best cost
886 unsigned Cost; // Cost of the loop with that width
888 /// \return The most profitable vectorization factor and the cost of that VF.
889 /// This method checks every power of two up to VF. If UserVF is not ZERO
890 /// then this vectorization factor will be selected if vectorization is
892 VectorizationFactor selectVectorizationFactor(bool OptForSize,
894 bool ForceVectorization);
896 /// \return The size (in bits) of the widest type in the code that
897 /// needs to be vectorized. We ignore values that remain scalar such as
898 /// 64 bit loop indices.
899 unsigned getWidestType();
901 /// \return The most profitable unroll factor.
902 /// If UserUF is non-zero then this method finds the best unroll-factor
903 /// based on register pressure and other parameters.
904 /// VF and LoopCost are the selected vectorization factor and the cost of the
906 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
909 /// \brief A struct that represents some properties of the register usage
911 struct RegisterUsage {
912 /// Holds the number of loop invariant values that are used in the loop.
913 unsigned LoopInvariantRegs;
914 /// Holds the maximum number of concurrent live intervals in the loop.
915 unsigned MaxLocalUsers;
916 /// Holds the number of instructions in the loop.
917 unsigned NumInstructions;
920 /// \return information about the register usage of the loop.
921 RegisterUsage calculateRegisterUsage();
924 /// Returns the expected execution cost. The unit of the cost does
925 /// not matter because we use the 'cost' units to compare different
926 /// vector widths. The cost that is returned is *not* normalized by
927 /// the factor width.
928 unsigned expectedCost(unsigned VF);
930 /// Returns the execution time cost of an instruction for a given vector
931 /// width. Vector width of one means scalar.
932 unsigned getInstructionCost(Instruction *I, unsigned VF);
934 /// A helper function for converting Scalar types to vector types.
935 /// If the incoming type is void, we return void. If the VF is 1, we return
937 static Type* ToVectorTy(Type *Scalar, unsigned VF);
939 /// Returns whether the instruction is a load or store and will be a emitted
940 /// as a vector operation.
941 bool isConsecutiveLoadOrStore(Instruction *I);
943 /// The loop that we evaluate.
947 /// Loop Info analysis.
949 /// Vectorization legality.
950 LoopVectorizationLegality *Legal;
951 /// Vector target information.
952 const TargetTransformInfo &TTI;
953 /// Target data layout information.
954 const DataLayout *DL;
955 /// Target Library Info.
956 const TargetLibraryInfo *TLI;
959 /// Utility class for getting and setting loop vectorizer hints in the form
960 /// of loop metadata.
961 class LoopVectorizeHints {
964 FK_Undefined = -1, ///< Not selected.
965 FK_Disabled = 0, ///< Forcing disabled.
966 FK_Enabled = 1, ///< Forcing enabled.
969 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
970 : Width(VectorizationFactor),
971 Unroll(DisableUnrolling),
973 LoopID(L->getLoopID()) {
975 // force-vector-unroll overrides DisableUnrolling.
976 if (VectorizationUnroll.getNumOccurrences() > 0)
977 Unroll = VectorizationUnroll;
979 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
980 << "LV: Unrolling disabled by the pass manager\n");
983 /// Return the loop metadata prefix.
984 static StringRef Prefix() { return "llvm.loop."; }
986 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
987 SmallVector<Value*, 2> Vals;
988 Vals.push_back(MDString::get(Context, Name));
989 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
990 return MDNode::get(Context, Vals);
993 /// Mark the loop L as already vectorized by setting the width to 1.
994 void setAlreadyVectorized(Loop *L) {
995 LLVMContext &Context = L->getHeader()->getContext();
999 // Create a new loop id with one more operand for the already_vectorized
1000 // hint. If the loop already has a loop id then copy the existing operands.
1001 SmallVector<Value*, 4> Vals(1);
1003 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
1004 Vals.push_back(LoopID->getOperand(i));
1007 createHint(Context, Twine(Prefix(), "vectorize.width").str(), Width));
1009 createHint(Context, Twine(Prefix(), "interleave.count").str(), 1));
1011 MDNode *NewLoopID = MDNode::get(Context, Vals);
1012 // Set operand 0 to refer to the loop id itself.
1013 NewLoopID->replaceOperandWith(0, NewLoopID);
1015 L->setLoopID(NewLoopID);
1017 LoopID->replaceAllUsesWith(NewLoopID);
1022 std::string emitRemark() const {
1024 R << "vectorization ";
1026 case LoopVectorizeHints::FK_Disabled:
1027 R << "is explicitly disabled";
1029 case LoopVectorizeHints::FK_Enabled:
1030 R << "is explicitly enabled";
1031 if (Width != 0 && Unroll != 0)
1032 R << " with width " << Width << " and interleave count " << Unroll;
1033 else if (Width != 0)
1034 R << " with width " << Width;
1035 else if (Unroll != 0)
1036 R << " with interleave count " << Unroll;
1038 case LoopVectorizeHints::FK_Undefined:
1039 R << "was not specified";
1045 unsigned getWidth() const { return Width; }
1046 unsigned getUnroll() const { return Unroll; }
1047 enum ForceKind getForce() const { return Force; }
1048 MDNode *getLoopID() const { return LoopID; }
1051 /// Find hints specified in the loop metadata.
1052 void getHints(const Loop *L) {
1056 // First operand should refer to the loop id itself.
1057 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1058 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1060 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1061 const MDString *S = nullptr;
1062 SmallVector<Value*, 4> Args;
1064 // The expected hint is either a MDString or a MDNode with the first
1065 // operand a MDString.
1066 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1067 if (!MD || MD->getNumOperands() == 0)
1069 S = dyn_cast<MDString>(MD->getOperand(0));
1070 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1071 Args.push_back(MD->getOperand(i));
1073 S = dyn_cast<MDString>(LoopID->getOperand(i));
1074 assert(Args.size() == 0 && "too many arguments for MDString");
1080 // Check if the hint starts with the loop metadata prefix.
1081 StringRef Hint = S->getString();
1082 if (!Hint.startswith(Prefix()))
1084 // Remove the prefix.
1085 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1087 if (Args.size() == 1)
1088 getHint(Hint, Args[0]);
1092 // Check string hint with one operand.
1093 void getHint(StringRef Hint, Value *Arg) {
1094 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1096 unsigned Val = C->getZExtValue();
1098 if (Hint == "vectorize.width") {
1099 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1102 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1103 } else if (Hint == "vectorize.enable") {
1104 if (C->getBitWidth() == 1)
1105 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1106 : LoopVectorizeHints::FK_Disabled;
1108 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1109 } else if (Hint == "interleave.count") {
1110 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1113 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1115 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1119 /// Vectorization width.
1121 /// Vectorization unroll factor.
1123 /// Vectorization forced
1124 enum ForceKind Force;
1129 static void emitMissedWarning(Function *F, Loop *L,
1130 const LoopVectorizeHints &LH) {
1131 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1132 L->getStartLoc(), LH.emitRemark());
1134 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1135 if (LH.getWidth() != 1)
1136 emitLoopVectorizeWarning(
1137 F->getContext(), *F, L->getStartLoc(),
1138 "failed explicitly specified loop vectorization");
1139 else if (LH.getUnroll() != 1)
1140 emitLoopInterleaveWarning(
1141 F->getContext(), *F, L->getStartLoc(),
1142 "failed explicitly specified loop interleaving");
1146 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1148 return V.push_back(&L);
1150 for (Loop *InnerL : L)
1151 addInnerLoop(*InnerL, V);
1154 /// The LoopVectorize Pass.
1155 struct LoopVectorize : public FunctionPass {
1156 /// Pass identification, replacement for typeid
1159 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1161 DisableUnrolling(NoUnrolling),
1162 AlwaysVectorize(AlwaysVectorize) {
1163 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1166 ScalarEvolution *SE;
1167 const DataLayout *DL;
1169 TargetTransformInfo *TTI;
1171 BlockFrequencyInfo *BFI;
1172 TargetLibraryInfo *TLI;
1174 bool DisableUnrolling;
1175 bool AlwaysVectorize;
1177 BlockFrequency ColdEntryFreq;
1179 bool runOnFunction(Function &F) override {
1180 SE = &getAnalysis<ScalarEvolution>();
1181 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1182 DL = DLP ? &DLP->getDataLayout() : nullptr;
1183 LI = &getAnalysis<LoopInfo>();
1184 TTI = &getAnalysis<TargetTransformInfo>();
1185 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1186 BFI = &getAnalysis<BlockFrequencyInfo>();
1187 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1188 AA = &getAnalysis<AliasAnalysis>();
1190 // Compute some weights outside of the loop over the loops. Compute this
1191 // using a BranchProbability to re-use its scaling math.
1192 const BranchProbability ColdProb(1, 5); // 20%
1193 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1195 // If the target claims to have no vector registers don't attempt
1197 if (!TTI->getNumberOfRegisters(true))
1201 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1202 << ": Missing data layout\n");
1206 // Build up a worklist of inner-loops to vectorize. This is necessary as
1207 // the act of vectorizing or partially unrolling a loop creates new loops
1208 // and can invalidate iterators across the loops.
1209 SmallVector<Loop *, 8> Worklist;
1212 addInnerLoop(*L, Worklist);
1214 LoopsAnalyzed += Worklist.size();
1216 // Now walk the identified inner loops.
1217 bool Changed = false;
1218 while (!Worklist.empty())
1219 Changed |= processLoop(Worklist.pop_back_val());
1221 // Process each loop nest in the function.
1225 bool processLoop(Loop *L) {
1226 assert(L->empty() && "Only process inner loops.");
1229 const std::string DebugLocStr = getDebugLocString(L);
1232 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1233 << L->getHeader()->getParent()->getName() << "\" from "
1234 << DebugLocStr << "\n");
1236 LoopVectorizeHints Hints(L, DisableUnrolling);
1238 DEBUG(dbgs() << "LV: Loop hints:"
1240 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1242 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1244 : "?")) << " width=" << Hints.getWidth()
1245 << " unroll=" << Hints.getUnroll() << "\n");
1247 // Function containing loop
1248 Function *F = L->getHeader()->getParent();
1250 // Looking at the diagnostic output is the only way to determine if a loop
1251 // was vectorized (other than looking at the IR or machine code), so it
1252 // is important to generate an optimization remark for each loop. Most of
1253 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1254 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1255 // less verbose reporting vectorized loops and unvectorized loops that may
1256 // benefit from vectorization, respectively.
1258 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1259 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1260 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1261 L->getStartLoc(), Hints.emitRemark());
1265 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1266 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1267 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1268 L->getStartLoc(), Hints.emitRemark());
1272 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1273 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1274 emitOptimizationRemarkAnalysis(
1275 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1276 "loop not vectorized: vector width and interleave count are "
1277 "explicitly set to 1");
1281 // Check the loop for a trip count threshold:
1282 // do not vectorize loops with a tiny trip count.
1283 BasicBlock *Latch = L->getLoopLatch();
1284 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1285 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1286 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1287 << "This loop is not worth vectorizing.");
1288 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1289 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1291 DEBUG(dbgs() << "\n");
1292 emitOptimizationRemarkAnalysis(
1293 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1294 "vectorization is not beneficial and is not explicitly forced");
1299 // Check if it is legal to vectorize the loop.
1300 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1301 if (!LVL.canVectorize()) {
1302 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1303 emitMissedWarning(F, L, Hints);
1307 // Use the cost model.
1308 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1310 // Check the function attributes to find out if this function should be
1311 // optimized for size.
1312 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1313 F->hasFnAttribute(Attribute::OptimizeForSize);
1315 // Compute the weighted frequency of this loop being executed and see if it
1316 // is less than 20% of the function entry baseline frequency. Note that we
1317 // always have a canonical loop here because we think we *can* vectoriez.
1318 // FIXME: This is hidden behind a flag due to pervasive problems with
1319 // exactly what block frequency models.
1320 if (LoopVectorizeWithBlockFrequency) {
1321 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1322 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1323 LoopEntryFreq < ColdEntryFreq)
1327 // Check the function attributes to see if implicit floats are allowed.a
1328 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1329 // an integer loop and the vector instructions selected are purely integer
1330 // vector instructions?
1331 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1332 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1333 "attribute is used.\n");
1334 emitOptimizationRemarkAnalysis(
1335 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1336 "loop not vectorized due to NoImplicitFloat attribute");
1337 emitMissedWarning(F, L, Hints);
1341 // Select the optimal vectorization factor.
1342 const LoopVectorizationCostModel::VectorizationFactor VF =
1343 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1345 LoopVectorizeHints::FK_Enabled);
1347 // Select the unroll factor.
1349 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1351 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1352 << DebugLocStr << '\n');
1353 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1355 if (VF.Width == 1) {
1356 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1359 emitOptimizationRemarkAnalysis(
1360 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1361 "not beneficial to vectorize and user disabled interleaving");
1364 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1366 // Report the unrolling decision.
1367 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1368 Twine("unrolled with interleaving factor " +
1370 " (vectorization not beneficial)"));
1372 // We decided not to vectorize, but we may want to unroll.
1374 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1375 Unroller.vectorize(&LVL);
1377 // If we decided that it is *legal* to vectorize the loop then do it.
1378 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1382 // Report the vectorization decision.
1383 emitOptimizationRemark(
1384 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1385 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1386 ", unrolling interleave factor: " + Twine(UF) + ")");
1389 // Mark the loop as already vectorized to avoid vectorizing again.
1390 Hints.setAlreadyVectorized(L);
1392 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1396 void getAnalysisUsage(AnalysisUsage &AU) const override {
1397 AU.addRequiredID(LoopSimplifyID);
1398 AU.addRequiredID(LCSSAID);
1399 AU.addRequired<BlockFrequencyInfo>();
1400 AU.addRequired<DominatorTreeWrapperPass>();
1401 AU.addRequired<LoopInfo>();
1402 AU.addRequired<ScalarEvolution>();
1403 AU.addRequired<TargetTransformInfo>();
1404 AU.addRequired<AliasAnalysis>();
1405 AU.addPreserved<LoopInfo>();
1406 AU.addPreserved<DominatorTreeWrapperPass>();
1407 AU.addPreserved<AliasAnalysis>();
1412 } // end anonymous namespace
1414 //===----------------------------------------------------------------------===//
1415 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1416 // LoopVectorizationCostModel.
1417 //===----------------------------------------------------------------------===//
1419 static Value *stripIntegerCast(Value *V) {
1420 if (CastInst *CI = dyn_cast<CastInst>(V))
1421 if (CI->getOperand(0)->getType()->isIntegerTy())
1422 return CI->getOperand(0);
1426 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1428 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1430 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1431 ValueToValueMap &PtrToStride,
1432 Value *Ptr, Value *OrigPtr = nullptr) {
1434 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1436 // If there is an entry in the map return the SCEV of the pointer with the
1437 // symbolic stride replaced by one.
1438 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1439 if (SI != PtrToStride.end()) {
1440 Value *StrideVal = SI->second;
1443 StrideVal = stripIntegerCast(StrideVal);
1445 // Replace symbolic stride by one.
1446 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1447 ValueToValueMap RewriteMap;
1448 RewriteMap[StrideVal] = One;
1451 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1452 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1457 // Otherwise, just return the SCEV of the original pointer.
1458 return SE->getSCEV(Ptr);
1461 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1462 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1463 unsigned ASId, ValueToValueMap &Strides) {
1464 // Get the stride replaced scev.
1465 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1466 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1467 assert(AR && "Invalid addrec expression");
1468 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1469 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1470 Pointers.push_back(Ptr);
1471 Starts.push_back(AR->getStart());
1472 Ends.push_back(ScEnd);
1473 IsWritePtr.push_back(WritePtr);
1474 DependencySetId.push_back(DepSetId);
1475 AliasSetId.push_back(ASId);
1478 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1479 // We need to place the broadcast of invariant variables outside the loop.
1480 Instruction *Instr = dyn_cast<Instruction>(V);
1482 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1483 Instr->getParent()) != LoopVectorBody.end());
1484 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1486 // Place the code for broadcasting invariant variables in the new preheader.
1487 IRBuilder<>::InsertPointGuard Guard(Builder);
1489 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1491 // Broadcast the scalar into all locations in the vector.
1492 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1497 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1499 assert(Val->getType()->isVectorTy() && "Must be a vector");
1500 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1501 "Elem must be an integer");
1502 // Create the types.
1503 Type *ITy = Val->getType()->getScalarType();
1504 VectorType *Ty = cast<VectorType>(Val->getType());
1505 int VLen = Ty->getNumElements();
1506 SmallVector<Constant*, 8> Indices;
1508 // Create a vector of consecutive numbers from zero to VF.
1509 for (int i = 0; i < VLen; ++i) {
1510 int64_t Idx = Negate ? (-i) : i;
1511 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1514 // Add the consecutive indices to the vector value.
1515 Constant *Cv = ConstantVector::get(Indices);
1516 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1517 return Builder.CreateAdd(Val, Cv, "induction");
1520 /// \brief Find the operand of the GEP that should be checked for consecutive
1521 /// stores. This ignores trailing indices that have no effect on the final
1523 static unsigned getGEPInductionOperand(const DataLayout *DL,
1524 const GetElementPtrInst *Gep) {
1525 unsigned LastOperand = Gep->getNumOperands() - 1;
1526 unsigned GEPAllocSize = DL->getTypeAllocSize(
1527 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1529 // Walk backwards and try to peel off zeros.
1530 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1531 // Find the type we're currently indexing into.
1532 gep_type_iterator GEPTI = gep_type_begin(Gep);
1533 std::advance(GEPTI, LastOperand - 1);
1535 // If it's a type with the same allocation size as the result of the GEP we
1536 // can peel off the zero index.
1537 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1545 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1546 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1547 // Make sure that the pointer does not point to structs.
1548 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1551 // If this value is a pointer induction variable we know it is consecutive.
1552 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1553 if (Phi && Inductions.count(Phi)) {
1554 InductionInfo II = Inductions[Phi];
1555 if (IK_PtrInduction == II.IK)
1557 else if (IK_ReversePtrInduction == II.IK)
1561 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1565 unsigned NumOperands = Gep->getNumOperands();
1566 Value *GpPtr = Gep->getPointerOperand();
1567 // If this GEP value is a consecutive pointer induction variable and all of
1568 // the indices are constant then we know it is consecutive. We can
1569 Phi = dyn_cast<PHINode>(GpPtr);
1570 if (Phi && Inductions.count(Phi)) {
1572 // Make sure that the pointer does not point to structs.
1573 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1574 if (GepPtrType->getElementType()->isAggregateType())
1577 // Make sure that all of the index operands are loop invariant.
1578 for (unsigned i = 1; i < NumOperands; ++i)
1579 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1582 InductionInfo II = Inductions[Phi];
1583 if (IK_PtrInduction == II.IK)
1585 else if (IK_ReversePtrInduction == II.IK)
1589 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1591 // Check that all of the gep indices are uniform except for our induction
1593 for (unsigned i = 0; i != NumOperands; ++i)
1594 if (i != InductionOperand &&
1595 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1598 // We can emit wide load/stores only if the last non-zero index is the
1599 // induction variable.
1600 const SCEV *Last = nullptr;
1601 if (!Strides.count(Gep))
1602 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1604 // Because of the multiplication by a stride we can have a s/zext cast.
1605 // We are going to replace this stride by 1 so the cast is safe to ignore.
1607 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1608 // %0 = trunc i64 %indvars.iv to i32
1609 // %mul = mul i32 %0, %Stride1
1610 // %idxprom = zext i32 %mul to i64 << Safe cast.
1611 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1613 Last = replaceSymbolicStrideSCEV(SE, Strides,
1614 Gep->getOperand(InductionOperand), Gep);
1615 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1617 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1621 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1622 const SCEV *Step = AR->getStepRecurrence(*SE);
1624 // The memory is consecutive because the last index is consecutive
1625 // and all other indices are loop invariant.
1628 if (Step->isAllOnesValue())
1635 bool LoopVectorizationLegality::isUniform(Value *V) {
1636 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1639 InnerLoopVectorizer::VectorParts&
1640 InnerLoopVectorizer::getVectorValue(Value *V) {
1641 assert(V != Induction && "The new induction variable should not be used.");
1642 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1644 // If we have a stride that is replaced by one, do it here.
1645 if (Legal->hasStride(V))
1646 V = ConstantInt::get(V->getType(), 1);
1648 // If we have this scalar in the map, return it.
1649 if (WidenMap.has(V))
1650 return WidenMap.get(V);
1652 // If this scalar is unknown, assume that it is a constant or that it is
1653 // loop invariant. Broadcast V and save the value for future uses.
1654 Value *B = getBroadcastInstrs(V);
1655 return WidenMap.splat(V, B);
1658 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1659 assert(Vec->getType()->isVectorTy() && "Invalid type");
1660 SmallVector<Constant*, 8> ShuffleMask;
1661 for (unsigned i = 0; i < VF; ++i)
1662 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1664 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1665 ConstantVector::get(ShuffleMask),
1669 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1670 // Attempt to issue a wide load.
1671 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1672 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1674 assert((LI || SI) && "Invalid Load/Store instruction");
1676 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1677 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1678 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1679 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1680 // An alignment of 0 means target abi alignment. We need to use the scalar's
1681 // target abi alignment in such a case.
1683 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1684 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1685 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1686 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1688 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1689 return scalarizeInstruction(Instr, true);
1691 if (ScalarAllocatedSize != VectorElementSize)
1692 return scalarizeInstruction(Instr);
1694 // If the pointer is loop invariant or if it is non-consecutive,
1695 // scalarize the load.
1696 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1697 bool Reverse = ConsecutiveStride < 0;
1698 bool UniformLoad = LI && Legal->isUniform(Ptr);
1699 if (!ConsecutiveStride || UniformLoad)
1700 return scalarizeInstruction(Instr);
1702 Constant *Zero = Builder.getInt32(0);
1703 VectorParts &Entry = WidenMap.get(Instr);
1705 // Handle consecutive loads/stores.
1706 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1707 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1708 setDebugLocFromInst(Builder, Gep);
1709 Value *PtrOperand = Gep->getPointerOperand();
1710 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1711 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1713 // Create the new GEP with the new induction variable.
1714 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1715 Gep2->setOperand(0, FirstBasePtr);
1716 Gep2->setName("gep.indvar.base");
1717 Ptr = Builder.Insert(Gep2);
1719 setDebugLocFromInst(Builder, Gep);
1720 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1721 OrigLoop) && "Base ptr must be invariant");
1723 // The last index does not have to be the induction. It can be
1724 // consecutive and be a function of the index. For example A[I+1];
1725 unsigned NumOperands = Gep->getNumOperands();
1726 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1727 // Create the new GEP with the new induction variable.
1728 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1730 for (unsigned i = 0; i < NumOperands; ++i) {
1731 Value *GepOperand = Gep->getOperand(i);
1732 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1734 // Update last index or loop invariant instruction anchored in loop.
1735 if (i == InductionOperand ||
1736 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1737 assert((i == InductionOperand ||
1738 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1739 "Must be last index or loop invariant");
1741 VectorParts &GEPParts = getVectorValue(GepOperand);
1742 Value *Index = GEPParts[0];
1743 Index = Builder.CreateExtractElement(Index, Zero);
1744 Gep2->setOperand(i, Index);
1745 Gep2->setName("gep.indvar.idx");
1748 Ptr = Builder.Insert(Gep2);
1750 // Use the induction element ptr.
1751 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1752 setDebugLocFromInst(Builder, Ptr);
1753 VectorParts &PtrVal = getVectorValue(Ptr);
1754 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1759 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1760 "We do not allow storing to uniform addresses");
1761 setDebugLocFromInst(Builder, SI);
1762 // We don't want to update the value in the map as it might be used in
1763 // another expression. So don't use a reference type for "StoredVal".
1764 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1766 for (unsigned Part = 0; Part < UF; ++Part) {
1767 // Calculate the pointer for the specific unroll-part.
1768 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1771 // If we store to reverse consecutive memory locations then we need
1772 // to reverse the order of elements in the stored value.
1773 StoredVal[Part] = reverseVector(StoredVal[Part]);
1774 // If the address is consecutive but reversed, then the
1775 // wide store needs to start at the last vector element.
1776 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1777 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1780 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1781 DataTy->getPointerTo(AddressSpace));
1783 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1784 propagateMetadata(NewSI, SI);
1790 assert(LI && "Must have a load instruction");
1791 setDebugLocFromInst(Builder, LI);
1792 for (unsigned Part = 0; Part < UF; ++Part) {
1793 // Calculate the pointer for the specific unroll-part.
1794 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1797 // If the address is consecutive but reversed, then the
1798 // wide store needs to start at the last vector element.
1799 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1800 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1803 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1804 DataTy->getPointerTo(AddressSpace));
1805 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1806 propagateMetadata(NewLI, LI);
1807 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1811 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1812 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1813 // Holds vector parameters or scalars, in case of uniform vals.
1814 SmallVector<VectorParts, 4> Params;
1816 setDebugLocFromInst(Builder, Instr);
1818 // Find all of the vectorized parameters.
1819 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1820 Value *SrcOp = Instr->getOperand(op);
1822 // If we are accessing the old induction variable, use the new one.
1823 if (SrcOp == OldInduction) {
1824 Params.push_back(getVectorValue(SrcOp));
1828 // Try using previously calculated values.
1829 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1831 // If the src is an instruction that appeared earlier in the basic block
1832 // then it should already be vectorized.
1833 if (SrcInst && OrigLoop->contains(SrcInst)) {
1834 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1835 // The parameter is a vector value from earlier.
1836 Params.push_back(WidenMap.get(SrcInst));
1838 // The parameter is a scalar from outside the loop. Maybe even a constant.
1839 VectorParts Scalars;
1840 Scalars.append(UF, SrcOp);
1841 Params.push_back(Scalars);
1845 assert(Params.size() == Instr->getNumOperands() &&
1846 "Invalid number of operands");
1848 // Does this instruction return a value ?
1849 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1851 Value *UndefVec = IsVoidRetTy ? nullptr :
1852 UndefValue::get(VectorType::get(Instr->getType(), VF));
1853 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1854 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1856 Instruction *InsertPt = Builder.GetInsertPoint();
1857 BasicBlock *IfBlock = Builder.GetInsertBlock();
1858 BasicBlock *CondBlock = nullptr;
1861 Loop *VectorLp = nullptr;
1862 if (IfPredicateStore) {
1863 assert(Instr->getParent()->getSinglePredecessor() &&
1864 "Only support single predecessor blocks");
1865 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1866 Instr->getParent());
1867 VectorLp = LI->getLoopFor(IfBlock);
1868 assert(VectorLp && "Must have a loop for this block");
1871 // For each vector unroll 'part':
1872 for (unsigned Part = 0; Part < UF; ++Part) {
1873 // For each scalar that we create:
1874 for (unsigned Width = 0; Width < VF; ++Width) {
1877 Value *Cmp = nullptr;
1878 if (IfPredicateStore) {
1879 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1880 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1881 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1882 LoopVectorBody.push_back(CondBlock);
1883 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1884 // Update Builder with newly created basic block.
1885 Builder.SetInsertPoint(InsertPt);
1888 Instruction *Cloned = Instr->clone();
1890 Cloned->setName(Instr->getName() + ".cloned");
1891 // Replace the operands of the cloned instructions with extracted scalars.
1892 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1893 Value *Op = Params[op][Part];
1894 // Param is a vector. Need to extract the right lane.
1895 if (Op->getType()->isVectorTy())
1896 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1897 Cloned->setOperand(op, Op);
1900 // Place the cloned scalar in the new loop.
1901 Builder.Insert(Cloned);
1903 // If the original scalar returns a value we need to place it in a vector
1904 // so that future users will be able to use it.
1906 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1907 Builder.getInt32(Width));
1909 if (IfPredicateStore) {
1910 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1911 LoopVectorBody.push_back(NewIfBlock);
1912 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1913 Builder.SetInsertPoint(InsertPt);
1914 Instruction *OldBr = IfBlock->getTerminator();
1915 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1916 OldBr->eraseFromParent();
1917 IfBlock = NewIfBlock;
1923 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1927 if (Instruction *I = dyn_cast<Instruction>(V))
1928 return I->getParent() == Loc->getParent() ? I : nullptr;
1932 std::pair<Instruction *, Instruction *>
1933 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1934 Instruction *tnullptr = nullptr;
1935 if (!Legal->mustCheckStrides())
1936 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1938 IRBuilder<> ChkBuilder(Loc);
1941 Value *Check = nullptr;
1942 Instruction *FirstInst = nullptr;
1943 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1944 SE = Legal->strides_end();
1946 Value *Ptr = stripIntegerCast(*SI);
1947 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1949 // Store the first instruction we create.
1950 FirstInst = getFirstInst(FirstInst, C, Loc);
1952 Check = ChkBuilder.CreateOr(Check, C);
1957 // We have to do this trickery because the IRBuilder might fold the check to a
1958 // constant expression in which case there is no Instruction anchored in a
1960 LLVMContext &Ctx = Loc->getContext();
1961 Instruction *TheCheck =
1962 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1963 ChkBuilder.Insert(TheCheck, "stride.not.one");
1964 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1966 return std::make_pair(FirstInst, TheCheck);
1969 std::pair<Instruction *, Instruction *>
1970 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1971 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1972 Legal->getRuntimePointerCheck();
1974 Instruction *tnullptr = nullptr;
1975 if (!PtrRtCheck->Need)
1976 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1978 unsigned NumPointers = PtrRtCheck->Pointers.size();
1979 SmallVector<TrackingVH<Value> , 2> Starts;
1980 SmallVector<TrackingVH<Value> , 2> Ends;
1982 LLVMContext &Ctx = Loc->getContext();
1983 SCEVExpander Exp(*SE, "induction");
1984 Instruction *FirstInst = nullptr;
1986 for (unsigned i = 0; i < NumPointers; ++i) {
1987 Value *Ptr = PtrRtCheck->Pointers[i];
1988 const SCEV *Sc = SE->getSCEV(Ptr);
1990 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1991 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1993 Starts.push_back(Ptr);
1994 Ends.push_back(Ptr);
1996 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1997 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1999 // Use this type for pointer arithmetic.
2000 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2002 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2003 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2004 Starts.push_back(Start);
2005 Ends.push_back(End);
2009 IRBuilder<> ChkBuilder(Loc);
2010 // Our instructions might fold to a constant.
2011 Value *MemoryRuntimeCheck = nullptr;
2012 for (unsigned i = 0; i < NumPointers; ++i) {
2013 for (unsigned j = i+1; j < NumPointers; ++j) {
2014 // No need to check if two readonly pointers intersect.
2015 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2018 // Only need to check pointers between two different dependency sets.
2019 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2021 // Only need to check pointers in the same alias set.
2022 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2025 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2026 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2028 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2029 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2030 "Trying to bounds check pointers with different address spaces");
2032 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2033 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2035 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2036 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2037 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2038 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2040 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2041 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2042 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2043 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2044 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2045 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2046 if (MemoryRuntimeCheck) {
2047 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2049 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2051 MemoryRuntimeCheck = IsConflict;
2055 // We have to do this trickery because the IRBuilder might fold the check to a
2056 // constant expression in which case there is no Instruction anchored in a
2058 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2059 ConstantInt::getTrue(Ctx));
2060 ChkBuilder.Insert(Check, "memcheck.conflict");
2061 FirstInst = getFirstInst(FirstInst, Check, Loc);
2062 return std::make_pair(FirstInst, Check);
2065 void InnerLoopVectorizer::createEmptyLoop() {
2067 In this function we generate a new loop. The new loop will contain
2068 the vectorized instructions while the old loop will continue to run the
2071 [ ] <-- Back-edge taken count overflow check.
2074 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2077 || [ ] <-- vector pre header.
2081 || [ ]_| <-- vector loop.
2084 | >[ ] <--- middle-block.
2087 -|- >[ ] <--- new preheader.
2091 | [ ]_| <-- old scalar loop to handle remainder.
2094 >[ ] <-- exit block.
2098 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2099 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2100 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2101 assert(BypassBlock && "Invalid loop structure");
2102 assert(ExitBlock && "Must have an exit block");
2104 // Some loops have a single integer induction variable, while other loops
2105 // don't. One example is c++ iterators that often have multiple pointer
2106 // induction variables. In the code below we also support a case where we
2107 // don't have a single induction variable.
2108 OldInduction = Legal->getInduction();
2109 Type *IdxTy = Legal->getWidestInductionType();
2111 // Find the loop boundaries.
2112 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2113 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2115 // The exit count might have the type of i64 while the phi is i32. This can
2116 // happen if we have an induction variable that is sign extended before the
2117 // compare. The only way that we get a backedge taken count is that the
2118 // induction variable was signed and as such will not overflow. In such a case
2119 // truncation is legal.
2120 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2121 IdxTy->getPrimitiveSizeInBits())
2122 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2124 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2125 // Get the total trip count from the count by adding 1.
2126 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2127 SE->getConstant(BackedgeTakeCount->getType(), 1));
2129 // Expand the trip count and place the new instructions in the preheader.
2130 // Notice that the pre-header does not change, only the loop body.
2131 SCEVExpander Exp(*SE, "induction");
2133 // We need to test whether the backedge-taken count is uint##_max. Adding one
2134 // to it will cause overflow and an incorrect loop trip count in the vector
2135 // body. In case of overflow we want to directly jump to the scalar remainder
2137 Value *BackedgeCount =
2138 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2139 BypassBlock->getTerminator());
2140 if (BackedgeCount->getType()->isPointerTy())
2141 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2142 "backedge.ptrcnt.to.int",
2143 BypassBlock->getTerminator());
2144 Instruction *CheckBCOverflow =
2145 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2146 Constant::getAllOnesValue(BackedgeCount->getType()),
2147 "backedge.overflow", BypassBlock->getTerminator());
2149 // The loop index does not have to start at Zero. Find the original start
2150 // value from the induction PHI node. If we don't have an induction variable
2151 // then we know that it starts at zero.
2152 Builder.SetInsertPoint(BypassBlock->getTerminator());
2153 Value *StartIdx = ExtendedIdx = OldInduction ?
2154 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2156 ConstantInt::get(IdxTy, 0);
2158 // We need an instruction to anchor the overflow check on. StartIdx needs to
2159 // be defined before the overflow check branch. Because the scalar preheader
2160 // is going to merge the start index and so the overflow branch block needs to
2161 // contain a definition of the start index.
2162 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2163 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2164 BypassBlock->getTerminator());
2166 // Count holds the overall loop count (N).
2167 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2168 BypassBlock->getTerminator());
2170 LoopBypassBlocks.push_back(BypassBlock);
2172 // Split the single block loop into the two loop structure described above.
2173 BasicBlock *VectorPH =
2174 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2175 BasicBlock *VecBody =
2176 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2177 BasicBlock *MiddleBlock =
2178 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2179 BasicBlock *ScalarPH =
2180 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2182 // Create and register the new vector loop.
2183 Loop* Lp = new Loop();
2184 Loop *ParentLoop = OrigLoop->getParentLoop();
2186 // Insert the new loop into the loop nest and register the new basic blocks
2187 // before calling any utilities such as SCEV that require valid LoopInfo.
2189 ParentLoop->addChildLoop(Lp);
2190 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2191 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2192 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2194 LI->addTopLevelLoop(Lp);
2196 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2198 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2200 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2202 // Generate the induction variable.
2203 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2204 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2205 // The loop step is equal to the vectorization factor (num of SIMD elements)
2206 // times the unroll factor (num of SIMD instructions).
2207 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2209 // This is the IR builder that we use to add all of the logic for bypassing
2210 // the new vector loop.
2211 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2212 setDebugLocFromInst(BypassBuilder,
2213 getDebugLocFromInstOrOperands(OldInduction));
2215 // We may need to extend the index in case there is a type mismatch.
2216 // We know that the count starts at zero and does not overflow.
2217 if (Count->getType() != IdxTy) {
2218 // The exit count can be of pointer type. Convert it to the correct
2220 if (ExitCount->getType()->isPointerTy())
2221 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2223 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2226 // Add the start index to the loop count to get the new end index.
2227 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2229 // Now we need to generate the expression for N - (N % VF), which is
2230 // the part that the vectorized body will execute.
2231 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2232 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2233 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2234 "end.idx.rnd.down");
2236 // Now, compare the new count to zero. If it is zero skip the vector loop and
2237 // jump to the scalar loop.
2239 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2241 BasicBlock *LastBypassBlock = BypassBlock;
2243 // Generate code to check that the loops trip count that we computed by adding
2244 // one to the backedge-taken count will not overflow.
2246 auto PastOverflowCheck =
2247 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2248 BasicBlock *CheckBlock =
2249 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2251 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2252 LoopBypassBlocks.push_back(CheckBlock);
2253 Instruction *OldTerm = LastBypassBlock->getTerminator();
2254 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2255 OldTerm->eraseFromParent();
2256 LastBypassBlock = CheckBlock;
2259 // Generate the code to check that the strides we assumed to be one are really
2260 // one. We want the new basic block to start at the first instruction in a
2261 // sequence of instructions that form a check.
2262 Instruction *StrideCheck;
2263 Instruction *FirstCheckInst;
2264 std::tie(FirstCheckInst, StrideCheck) =
2265 addStrideCheck(LastBypassBlock->getTerminator());
2267 // Create a new block containing the stride check.
2268 BasicBlock *CheckBlock =
2269 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2271 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2272 LoopBypassBlocks.push_back(CheckBlock);
2274 // Replace the branch into the memory check block with a conditional branch
2275 // for the "few elements case".
2276 Instruction *OldTerm = LastBypassBlock->getTerminator();
2277 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2278 OldTerm->eraseFromParent();
2281 LastBypassBlock = CheckBlock;
2284 // Generate the code that checks in runtime if arrays overlap. We put the
2285 // checks into a separate block to make the more common case of few elements
2287 Instruction *MemRuntimeCheck;
2288 std::tie(FirstCheckInst, MemRuntimeCheck) =
2289 addRuntimeCheck(LastBypassBlock->getTerminator());
2290 if (MemRuntimeCheck) {
2291 // Create a new block containing the memory check.
2292 BasicBlock *CheckBlock =
2293 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2295 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2296 LoopBypassBlocks.push_back(CheckBlock);
2298 // Replace the branch into the memory check block with a conditional branch
2299 // for the "few elements case".
2300 Instruction *OldTerm = LastBypassBlock->getTerminator();
2301 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2302 OldTerm->eraseFromParent();
2304 Cmp = MemRuntimeCheck;
2305 LastBypassBlock = CheckBlock;
2308 LastBypassBlock->getTerminator()->eraseFromParent();
2309 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2312 // We are going to resume the execution of the scalar loop.
2313 // Go over all of the induction variables that we found and fix the
2314 // PHIs that are left in the scalar version of the loop.
2315 // The starting values of PHI nodes depend on the counter of the last
2316 // iteration in the vectorized loop.
2317 // If we come from a bypass edge then we need to start from the original
2320 // This variable saves the new starting index for the scalar loop.
2321 PHINode *ResumeIndex = nullptr;
2322 LoopVectorizationLegality::InductionList::iterator I, E;
2323 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2324 // Set builder to point to last bypass block.
2325 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2326 for (I = List->begin(), E = List->end(); I != E; ++I) {
2327 PHINode *OrigPhi = I->first;
2328 LoopVectorizationLegality::InductionInfo II = I->second;
2330 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2331 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2332 MiddleBlock->getTerminator());
2333 // We might have extended the type of the induction variable but we need a
2334 // truncated version for the scalar loop.
2335 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2336 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2337 MiddleBlock->getTerminator()) : nullptr;
2339 // Create phi nodes to merge from the backedge-taken check block.
2340 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2341 ScalarPH->getTerminator());
2342 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2344 PHINode *BCTruncResumeVal = nullptr;
2345 if (OrigPhi == OldInduction) {
2347 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2348 ScalarPH->getTerminator());
2349 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2352 Value *EndValue = nullptr;
2354 case LoopVectorizationLegality::IK_NoInduction:
2355 llvm_unreachable("Unknown induction");
2356 case LoopVectorizationLegality::IK_IntInduction: {
2357 // Handle the integer induction counter.
2358 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2360 // We have the canonical induction variable.
2361 if (OrigPhi == OldInduction) {
2362 // Create a truncated version of the resume value for the scalar loop,
2363 // we might have promoted the type to a larger width.
2365 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2366 // The new PHI merges the original incoming value, in case of a bypass,
2367 // or the value at the end of the vectorized loop.
2368 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2369 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2370 TruncResumeVal->addIncoming(EndValue, VecBody);
2372 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2374 // We know what the end value is.
2375 EndValue = IdxEndRoundDown;
2376 // We also know which PHI node holds it.
2377 ResumeIndex = ResumeVal;
2381 // Not the canonical induction variable - add the vector loop count to the
2383 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2384 II.StartValue->getType(),
2386 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2389 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2390 // Convert the CountRoundDown variable to the PHI size.
2391 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2392 II.StartValue->getType(),
2394 // Handle reverse integer induction counter.
2395 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2398 case LoopVectorizationLegality::IK_PtrInduction: {
2399 // For pointer induction variables, calculate the offset using
2401 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2405 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2406 // The value at the end of the loop for the reverse pointer is calculated
2407 // by creating a GEP with a negative index starting from the start value.
2408 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2409 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2411 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2417 // The new PHI merges the original incoming value, in case of a bypass,
2418 // or the value at the end of the vectorized loop.
2419 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2420 if (OrigPhi == OldInduction)
2421 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2423 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2425 ResumeVal->addIncoming(EndValue, VecBody);
2427 // Fix the scalar body counter (PHI node).
2428 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2430 // The old induction's phi node in the scalar body needs the truncated
2432 if (OrigPhi == OldInduction) {
2433 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2434 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2436 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2437 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2441 // If we are generating a new induction variable then we also need to
2442 // generate the code that calculates the exit value. This value is not
2443 // simply the end of the counter because we may skip the vectorized body
2444 // in case of a runtime check.
2446 assert(!ResumeIndex && "Unexpected resume value found");
2447 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2448 MiddleBlock->getTerminator());
2449 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2450 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2451 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2454 // Make sure that we found the index where scalar loop needs to continue.
2455 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2456 "Invalid resume Index");
2458 // Add a check in the middle block to see if we have completed
2459 // all of the iterations in the first vector loop.
2460 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2461 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2462 ResumeIndex, "cmp.n",
2463 MiddleBlock->getTerminator());
2465 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2466 // Remove the old terminator.
2467 MiddleBlock->getTerminator()->eraseFromParent();
2469 // Create i+1 and fill the PHINode.
2470 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2471 Induction->addIncoming(StartIdx, VectorPH);
2472 Induction->addIncoming(NextIdx, VecBody);
2473 // Create the compare.
2474 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2475 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2477 // Now we have two terminators. Remove the old one from the block.
2478 VecBody->getTerminator()->eraseFromParent();
2480 // Get ready to start creating new instructions into the vectorized body.
2481 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2484 LoopVectorPreHeader = VectorPH;
2485 LoopScalarPreHeader = ScalarPH;
2486 LoopMiddleBlock = MiddleBlock;
2487 LoopExitBlock = ExitBlock;
2488 LoopVectorBody.push_back(VecBody);
2489 LoopScalarBody = OldBasicBlock;
2491 LoopVectorizeHints Hints(Lp, true);
2492 Hints.setAlreadyVectorized(Lp);
2495 /// This function returns the identity element (or neutral element) for
2496 /// the operation K.
2498 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2503 // Adding, Xoring, Oring zero to a number does not change it.
2504 return ConstantInt::get(Tp, 0);
2505 case RK_IntegerMult:
2506 // Multiplying a number by 1 does not change it.
2507 return ConstantInt::get(Tp, 1);
2509 // AND-ing a number with an all-1 value does not change it.
2510 return ConstantInt::get(Tp, -1, true);
2512 // Multiplying a number by 1 does not change it.
2513 return ConstantFP::get(Tp, 1.0L);
2515 // Adding zero to a number does not change it.
2516 return ConstantFP::get(Tp, 0.0L);
2518 llvm_unreachable("Unknown reduction kind");
2522 /// This function translates the reduction kind to an LLVM binary operator.
2524 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2526 case LoopVectorizationLegality::RK_IntegerAdd:
2527 return Instruction::Add;
2528 case LoopVectorizationLegality::RK_IntegerMult:
2529 return Instruction::Mul;
2530 case LoopVectorizationLegality::RK_IntegerOr:
2531 return Instruction::Or;
2532 case LoopVectorizationLegality::RK_IntegerAnd:
2533 return Instruction::And;
2534 case LoopVectorizationLegality::RK_IntegerXor:
2535 return Instruction::Xor;
2536 case LoopVectorizationLegality::RK_FloatMult:
2537 return Instruction::FMul;
2538 case LoopVectorizationLegality::RK_FloatAdd:
2539 return Instruction::FAdd;
2540 case LoopVectorizationLegality::RK_IntegerMinMax:
2541 return Instruction::ICmp;
2542 case LoopVectorizationLegality::RK_FloatMinMax:
2543 return Instruction::FCmp;
2545 llvm_unreachable("Unknown reduction operation");
2549 Value *createMinMaxOp(IRBuilder<> &Builder,
2550 LoopVectorizationLegality::MinMaxReductionKind RK,
2553 CmpInst::Predicate P = CmpInst::ICMP_NE;
2556 llvm_unreachable("Unknown min/max reduction kind");
2557 case LoopVectorizationLegality::MRK_UIntMin:
2558 P = CmpInst::ICMP_ULT;
2560 case LoopVectorizationLegality::MRK_UIntMax:
2561 P = CmpInst::ICMP_UGT;
2563 case LoopVectorizationLegality::MRK_SIntMin:
2564 P = CmpInst::ICMP_SLT;
2566 case LoopVectorizationLegality::MRK_SIntMax:
2567 P = CmpInst::ICMP_SGT;
2569 case LoopVectorizationLegality::MRK_FloatMin:
2570 P = CmpInst::FCMP_OLT;
2572 case LoopVectorizationLegality::MRK_FloatMax:
2573 P = CmpInst::FCMP_OGT;
2578 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2579 RK == LoopVectorizationLegality::MRK_FloatMax)
2580 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2582 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2584 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2589 struct CSEDenseMapInfo {
2590 static bool canHandle(Instruction *I) {
2591 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2592 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2594 static inline Instruction *getEmptyKey() {
2595 return DenseMapInfo<Instruction *>::getEmptyKey();
2597 static inline Instruction *getTombstoneKey() {
2598 return DenseMapInfo<Instruction *>::getTombstoneKey();
2600 static unsigned getHashValue(Instruction *I) {
2601 assert(canHandle(I) && "Unknown instruction!");
2602 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2603 I->value_op_end()));
2605 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2606 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2607 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2609 return LHS->isIdenticalTo(RHS);
2614 /// \brief Check whether this block is a predicated block.
2615 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2616 /// = ...; " blocks. We start with one vectorized basic block. For every
2617 /// conditional block we split this vectorized block. Therefore, every second
2618 /// block will be a predicated one.
2619 static bool isPredicatedBlock(unsigned BlockNum) {
2620 return BlockNum % 2;
2623 ///\brief Perform cse of induction variable instructions.
2624 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2625 // Perform simple cse.
2626 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2627 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2628 BasicBlock *BB = BBs[i];
2629 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2630 Instruction *In = I++;
2632 if (!CSEDenseMapInfo::canHandle(In))
2635 // Check if we can replace this instruction with any of the
2636 // visited instructions.
2637 if (Instruction *V = CSEMap.lookup(In)) {
2638 In->replaceAllUsesWith(V);
2639 In->eraseFromParent();
2642 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2643 // ...;" blocks for predicated stores. Every second block is a predicated
2645 if (isPredicatedBlock(i))
2653 /// \brief Adds a 'fast' flag to floating point operations.
2654 static Value *addFastMathFlag(Value *V) {
2655 if (isa<FPMathOperator>(V)){
2656 FastMathFlags Flags;
2657 Flags.setUnsafeAlgebra();
2658 cast<Instruction>(V)->setFastMathFlags(Flags);
2663 void InnerLoopVectorizer::vectorizeLoop() {
2664 //===------------------------------------------------===//
2666 // Notice: any optimization or new instruction that go
2667 // into the code below should be also be implemented in
2670 //===------------------------------------------------===//
2671 Constant *Zero = Builder.getInt32(0);
2673 // In order to support reduction variables we need to be able to vectorize
2674 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2675 // stages. First, we create a new vector PHI node with no incoming edges.
2676 // We use this value when we vectorize all of the instructions that use the
2677 // PHI. Next, after all of the instructions in the block are complete we
2678 // add the new incoming edges to the PHI. At this point all of the
2679 // instructions in the basic block are vectorized, so we can use them to
2680 // construct the PHI.
2681 PhiVector RdxPHIsToFix;
2683 // Scan the loop in a topological order to ensure that defs are vectorized
2685 LoopBlocksDFS DFS(OrigLoop);
2688 // Vectorize all of the blocks in the original loop.
2689 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2690 be = DFS.endRPO(); bb != be; ++bb)
2691 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2693 // At this point every instruction in the original loop is widened to
2694 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2695 // that we vectorized. The PHI nodes are currently empty because we did
2696 // not want to introduce cycles. Notice that the remaining PHI nodes
2697 // that we need to fix are reduction variables.
2699 // Create the 'reduced' values for each of the induction vars.
2700 // The reduced values are the vector values that we scalarize and combine
2701 // after the loop is finished.
2702 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2704 PHINode *RdxPhi = *it;
2705 assert(RdxPhi && "Unable to recover vectorized PHI");
2707 // Find the reduction variable descriptor.
2708 assert(Legal->getReductionVars()->count(RdxPhi) &&
2709 "Unable to find the reduction variable");
2710 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2711 (*Legal->getReductionVars())[RdxPhi];
2713 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2715 // We need to generate a reduction vector from the incoming scalar.
2716 // To do so, we need to generate the 'identity' vector and override
2717 // one of the elements with the incoming scalar reduction. We need
2718 // to do it in the vector-loop preheader.
2719 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2721 // This is the vector-clone of the value that leaves the loop.
2722 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2723 Type *VecTy = VectorExit[0]->getType();
2725 // Find the reduction identity variable. Zero for addition, or, xor,
2726 // one for multiplication, -1 for And.
2729 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2730 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2731 // MinMax reduction have the start value as their identify.
2733 VectorStart = Identity = RdxDesc.StartValue;
2735 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2740 // Handle other reduction kinds:
2742 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2743 VecTy->getScalarType());
2746 // This vector is the Identity vector where the first element is the
2747 // incoming scalar reduction.
2748 VectorStart = RdxDesc.StartValue;
2750 Identity = ConstantVector::getSplat(VF, Iden);
2752 // This vector is the Identity vector where the first element is the
2753 // incoming scalar reduction.
2754 VectorStart = Builder.CreateInsertElement(Identity,
2755 RdxDesc.StartValue, Zero);
2759 // Fix the vector-loop phi.
2760 // We created the induction variable so we know that the
2761 // preheader is the first entry.
2762 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2764 // Reductions do not have to start at zero. They can start with
2765 // any loop invariant values.
2766 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2767 BasicBlock *Latch = OrigLoop->getLoopLatch();
2768 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2769 VectorParts &Val = getVectorValue(LoopVal);
2770 for (unsigned part = 0; part < UF; ++part) {
2771 // Make sure to add the reduction stat value only to the
2772 // first unroll part.
2773 Value *StartVal = (part == 0) ? VectorStart : Identity;
2774 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2775 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2776 LoopVectorBody.back());
2779 // Before each round, move the insertion point right between
2780 // the PHIs and the values we are going to write.
2781 // This allows us to write both PHINodes and the extractelement
2783 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2785 VectorParts RdxParts;
2786 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2787 for (unsigned part = 0; part < UF; ++part) {
2788 // This PHINode contains the vectorized reduction variable, or
2789 // the initial value vector, if we bypass the vector loop.
2790 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2791 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2792 Value *StartVal = (part == 0) ? VectorStart : Identity;
2793 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2794 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2795 NewPhi->addIncoming(RdxExitVal[part],
2796 LoopVectorBody.back());
2797 RdxParts.push_back(NewPhi);
2800 // Reduce all of the unrolled parts into a single vector.
2801 Value *ReducedPartRdx = RdxParts[0];
2802 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2803 setDebugLocFromInst(Builder, ReducedPartRdx);
2804 for (unsigned part = 1; part < UF; ++part) {
2805 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2806 // Floating point operations had to be 'fast' to enable the reduction.
2807 ReducedPartRdx = addFastMathFlag(
2808 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2809 ReducedPartRdx, "bin.rdx"));
2811 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2812 ReducedPartRdx, RdxParts[part]);
2816 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2817 // and vector ops, reducing the set of values being computed by half each
2819 assert(isPowerOf2_32(VF) &&
2820 "Reduction emission only supported for pow2 vectors!");
2821 Value *TmpVec = ReducedPartRdx;
2822 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2823 for (unsigned i = VF; i != 1; i >>= 1) {
2824 // Move the upper half of the vector to the lower half.
2825 for (unsigned j = 0; j != i/2; ++j)
2826 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2828 // Fill the rest of the mask with undef.
2829 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2830 UndefValue::get(Builder.getInt32Ty()));
2833 Builder.CreateShuffleVector(TmpVec,
2834 UndefValue::get(TmpVec->getType()),
2835 ConstantVector::get(ShuffleMask),
2838 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2839 // Floating point operations had to be 'fast' to enable the reduction.
2840 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2841 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2843 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2846 // The result is in the first element of the vector.
2847 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2848 Builder.getInt32(0));
2851 // Create a phi node that merges control-flow from the backedge-taken check
2852 // block and the middle block.
2853 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2854 LoopScalarPreHeader->getTerminator());
2855 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2856 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2858 // Now, we need to fix the users of the reduction variable
2859 // inside and outside of the scalar remainder loop.
2860 // We know that the loop is in LCSSA form. We need to update the
2861 // PHI nodes in the exit blocks.
2862 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2863 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2864 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2865 if (!LCSSAPhi) break;
2867 // All PHINodes need to have a single entry edge, or two if
2868 // we already fixed them.
2869 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2871 // We found our reduction value exit-PHI. Update it with the
2872 // incoming bypass edge.
2873 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2874 // Add an edge coming from the bypass.
2875 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2878 }// end of the LCSSA phi scan.
2880 // Fix the scalar loop reduction variable with the incoming reduction sum
2881 // from the vector body and from the backedge value.
2882 int IncomingEdgeBlockIdx =
2883 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2884 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2885 // Pick the other block.
2886 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2887 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2888 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2889 }// end of for each redux variable.
2893 // Remove redundant induction instructions.
2894 cse(LoopVectorBody);
2897 void InnerLoopVectorizer::fixLCSSAPHIs() {
2898 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2899 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2900 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2901 if (!LCSSAPhi) break;
2902 if (LCSSAPhi->getNumIncomingValues() == 1)
2903 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2908 InnerLoopVectorizer::VectorParts
2909 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2910 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2913 // Look for cached value.
2914 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2915 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2916 if (ECEntryIt != MaskCache.end())
2917 return ECEntryIt->second;
2919 VectorParts SrcMask = createBlockInMask(Src);
2921 // The terminator has to be a branch inst!
2922 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2923 assert(BI && "Unexpected terminator found");
2925 if (BI->isConditional()) {
2926 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2928 if (BI->getSuccessor(0) != Dst)
2929 for (unsigned part = 0; part < UF; ++part)
2930 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2932 for (unsigned part = 0; part < UF; ++part)
2933 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2935 MaskCache[Edge] = EdgeMask;
2939 MaskCache[Edge] = SrcMask;
2943 InnerLoopVectorizer::VectorParts
2944 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2945 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2947 // Loop incoming mask is all-one.
2948 if (OrigLoop->getHeader() == BB) {
2949 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2950 return getVectorValue(C);
2953 // This is the block mask. We OR all incoming edges, and with zero.
2954 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2955 VectorParts BlockMask = getVectorValue(Zero);
2958 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2959 VectorParts EM = createEdgeMask(*it, BB);
2960 for (unsigned part = 0; part < UF; ++part)
2961 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2967 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2968 InnerLoopVectorizer::VectorParts &Entry,
2969 unsigned UF, unsigned VF, PhiVector *PV) {
2970 PHINode* P = cast<PHINode>(PN);
2971 // Handle reduction variables:
2972 if (Legal->getReductionVars()->count(P)) {
2973 for (unsigned part = 0; part < UF; ++part) {
2974 // This is phase one of vectorizing PHIs.
2975 Type *VecTy = (VF == 1) ? PN->getType() :
2976 VectorType::get(PN->getType(), VF);
2977 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2978 LoopVectorBody.back()-> getFirstInsertionPt());
2984 setDebugLocFromInst(Builder, P);
2985 // Check for PHI nodes that are lowered to vector selects.
2986 if (P->getParent() != OrigLoop->getHeader()) {
2987 // We know that all PHIs in non-header blocks are converted into
2988 // selects, so we don't have to worry about the insertion order and we
2989 // can just use the builder.
2990 // At this point we generate the predication tree. There may be
2991 // duplications since this is a simple recursive scan, but future
2992 // optimizations will clean it up.
2994 unsigned NumIncoming = P->getNumIncomingValues();
2996 // Generate a sequence of selects of the form:
2997 // SELECT(Mask3, In3,
2998 // SELECT(Mask2, In2,
3000 for (unsigned In = 0; In < NumIncoming; In++) {
3001 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3003 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3005 for (unsigned part = 0; part < UF; ++part) {
3006 // We might have single edge PHIs (blocks) - use an identity
3007 // 'select' for the first PHI operand.
3009 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3012 // Select between the current value and the previous incoming edge
3013 // based on the incoming mask.
3014 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3015 Entry[part], "predphi");
3021 // This PHINode must be an induction variable.
3022 // Make sure that we know about it.
3023 assert(Legal->getInductionVars()->count(P) &&
3024 "Not an induction variable");
3026 LoopVectorizationLegality::InductionInfo II =
3027 Legal->getInductionVars()->lookup(P);
3030 case LoopVectorizationLegality::IK_NoInduction:
3031 llvm_unreachable("Unknown induction");
3032 case LoopVectorizationLegality::IK_IntInduction: {
3033 assert(P->getType() == II.StartValue->getType() && "Types must match");
3034 Type *PhiTy = P->getType();
3036 if (P == OldInduction) {
3037 // Handle the canonical induction variable. We might have had to
3039 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3041 // Handle other induction variables that are now based on the
3043 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3045 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3046 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3049 Broadcasted = getBroadcastInstrs(Broadcasted);
3050 // After broadcasting the induction variable we need to make the vector
3051 // consecutive by adding 0, 1, 2, etc.
3052 for (unsigned part = 0; part < UF; ++part)
3053 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3056 case LoopVectorizationLegality::IK_ReverseIntInduction:
3057 case LoopVectorizationLegality::IK_PtrInduction:
3058 case LoopVectorizationLegality::IK_ReversePtrInduction:
3059 // Handle reverse integer and pointer inductions.
3060 Value *StartIdx = ExtendedIdx;
3061 // This is the normalized GEP that starts counting at zero.
3062 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3065 // Handle the reverse integer induction variable case.
3066 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3067 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3068 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3070 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3073 // This is a new value so do not hoist it out.
3074 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3075 // After broadcasting the induction variable we need to make the
3076 // vector consecutive by adding ... -3, -2, -1, 0.
3077 for (unsigned part = 0; part < UF; ++part)
3078 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3083 // Handle the pointer induction variable case.
3084 assert(P->getType()->isPointerTy() && "Unexpected type.");
3086 // Is this a reverse induction ptr or a consecutive induction ptr.
3087 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3090 // This is the vector of results. Notice that we don't generate
3091 // vector geps because scalar geps result in better code.
3092 for (unsigned part = 0; part < UF; ++part) {
3094 int EltIndex = (part) * (Reverse ? -1 : 1);
3095 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3098 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3100 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3102 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3104 Entry[part] = SclrGep;
3108 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3109 for (unsigned int i = 0; i < VF; ++i) {
3110 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3111 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3114 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3116 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3118 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3120 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3121 Builder.getInt32(i),
3124 Entry[part] = VecVal;
3130 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3131 // For each instruction in the old loop.
3132 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3133 VectorParts &Entry = WidenMap.get(it);
3134 switch (it->getOpcode()) {
3135 case Instruction::Br:
3136 // Nothing to do for PHIs and BR, since we already took care of the
3137 // loop control flow instructions.
3139 case Instruction::PHI:{
3140 // Vectorize PHINodes.
3141 widenPHIInstruction(it, Entry, UF, VF, PV);
3145 case Instruction::Add:
3146 case Instruction::FAdd:
3147 case Instruction::Sub:
3148 case Instruction::FSub:
3149 case Instruction::Mul:
3150 case Instruction::FMul:
3151 case Instruction::UDiv:
3152 case Instruction::SDiv:
3153 case Instruction::FDiv:
3154 case Instruction::URem:
3155 case Instruction::SRem:
3156 case Instruction::FRem:
3157 case Instruction::Shl:
3158 case Instruction::LShr:
3159 case Instruction::AShr:
3160 case Instruction::And:
3161 case Instruction::Or:
3162 case Instruction::Xor: {
3163 // Just widen binops.
3164 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3165 setDebugLocFromInst(Builder, BinOp);
3166 VectorParts &A = getVectorValue(it->getOperand(0));
3167 VectorParts &B = getVectorValue(it->getOperand(1));
3169 // Use this vector value for all users of the original instruction.
3170 for (unsigned Part = 0; Part < UF; ++Part) {
3171 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3173 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3174 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3175 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3176 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3177 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3179 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3180 VecOp->setIsExact(BinOp->isExact());
3182 // Copy the fast-math flags.
3183 if (VecOp && isa<FPMathOperator>(V))
3184 VecOp->setFastMathFlags(it->getFastMathFlags());
3189 propagateMetadata(Entry, it);
3192 case Instruction::Select: {
3194 // If the selector is loop invariant we can create a select
3195 // instruction with a scalar condition. Otherwise, use vector-select.
3196 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3198 setDebugLocFromInst(Builder, it);
3200 // The condition can be loop invariant but still defined inside the
3201 // loop. This means that we can't just use the original 'cond' value.
3202 // We have to take the 'vectorized' value and pick the first lane.
3203 // Instcombine will make this a no-op.
3204 VectorParts &Cond = getVectorValue(it->getOperand(0));
3205 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3206 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3208 Value *ScalarCond = (VF == 1) ? Cond[0] :
3209 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3211 for (unsigned Part = 0; Part < UF; ++Part) {
3212 Entry[Part] = Builder.CreateSelect(
3213 InvariantCond ? ScalarCond : Cond[Part],
3218 propagateMetadata(Entry, it);
3222 case Instruction::ICmp:
3223 case Instruction::FCmp: {
3224 // Widen compares. Generate vector compares.
3225 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3226 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3227 setDebugLocFromInst(Builder, it);
3228 VectorParts &A = getVectorValue(it->getOperand(0));
3229 VectorParts &B = getVectorValue(it->getOperand(1));
3230 for (unsigned Part = 0; Part < UF; ++Part) {
3233 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3235 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3239 propagateMetadata(Entry, it);
3243 case Instruction::Store:
3244 case Instruction::Load:
3245 vectorizeMemoryInstruction(it);
3247 case Instruction::ZExt:
3248 case Instruction::SExt:
3249 case Instruction::FPToUI:
3250 case Instruction::FPToSI:
3251 case Instruction::FPExt:
3252 case Instruction::PtrToInt:
3253 case Instruction::IntToPtr:
3254 case Instruction::SIToFP:
3255 case Instruction::UIToFP:
3256 case Instruction::Trunc:
3257 case Instruction::FPTrunc:
3258 case Instruction::BitCast: {
3259 CastInst *CI = dyn_cast<CastInst>(it);
3260 setDebugLocFromInst(Builder, it);
3261 /// Optimize the special case where the source is the induction
3262 /// variable. Notice that we can only optimize the 'trunc' case
3263 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3264 /// c. other casts depend on pointer size.
3265 if (CI->getOperand(0) == OldInduction &&
3266 it->getOpcode() == Instruction::Trunc) {
3267 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3269 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3270 for (unsigned Part = 0; Part < UF; ++Part)
3271 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3272 propagateMetadata(Entry, it);
3275 /// Vectorize casts.
3276 Type *DestTy = (VF == 1) ? CI->getType() :
3277 VectorType::get(CI->getType(), VF);
3279 VectorParts &A = getVectorValue(it->getOperand(0));
3280 for (unsigned Part = 0; Part < UF; ++Part)
3281 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3282 propagateMetadata(Entry, it);
3286 case Instruction::Call: {
3287 // Ignore dbg intrinsics.
3288 if (isa<DbgInfoIntrinsic>(it))
3290 setDebugLocFromInst(Builder, it);
3292 Module *M = BB->getParent()->getParent();
3293 CallInst *CI = cast<CallInst>(it);
3294 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3295 assert(ID && "Not an intrinsic call!");
3297 case Intrinsic::lifetime_end:
3298 case Intrinsic::lifetime_start:
3299 scalarizeInstruction(it);
3302 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3303 for (unsigned Part = 0; Part < UF; ++Part) {
3304 SmallVector<Value *, 4> Args;
3305 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3306 if (HasScalarOpd && i == 1) {
3307 Args.push_back(CI->getArgOperand(i));
3310 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3311 Args.push_back(Arg[Part]);
3313 Type *Tys[] = {CI->getType()};
3315 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3317 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3318 Entry[Part] = Builder.CreateCall(F, Args);
3321 propagateMetadata(Entry, it);
3328 // All other instructions are unsupported. Scalarize them.
3329 scalarizeInstruction(it);
3332 }// end of for_each instr.
3335 void InnerLoopVectorizer::updateAnalysis() {
3336 // Forget the original basic block.
3337 SE->forgetLoop(OrigLoop);
3339 // Update the dominator tree information.
3340 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3341 "Entry does not dominate exit.");
3343 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3344 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3345 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3347 // Due to if predication of stores we might create a sequence of "if(pred)
3348 // a[i] = ...; " blocks.
3349 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3351 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3352 else if (isPredicatedBlock(i)) {
3353 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3355 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3359 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3360 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3361 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3362 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3364 DEBUG(DT->verifyDomTree());
3367 /// \brief Check whether it is safe to if-convert this phi node.
3369 /// Phi nodes with constant expressions that can trap are not safe to if
3371 static bool canIfConvertPHINodes(BasicBlock *BB) {
3372 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3373 PHINode *Phi = dyn_cast<PHINode>(I);
3376 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3377 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3384 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3385 if (!EnableIfConversion) {
3386 emitAnalysis(Report() << "if-conversion is disabled");
3390 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3392 // A list of pointers that we can safely read and write to.
3393 SmallPtrSet<Value *, 8> SafePointes;
3395 // Collect safe addresses.
3396 for (Loop::block_iterator BI = TheLoop->block_begin(),
3397 BE = TheLoop->block_end(); BI != BE; ++BI) {
3398 BasicBlock *BB = *BI;
3400 if (blockNeedsPredication(BB))
3403 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3404 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3405 SafePointes.insert(LI->getPointerOperand());
3406 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3407 SafePointes.insert(SI->getPointerOperand());
3411 // Collect the blocks that need predication.
3412 BasicBlock *Header = TheLoop->getHeader();
3413 for (Loop::block_iterator BI = TheLoop->block_begin(),
3414 BE = TheLoop->block_end(); BI != BE; ++BI) {
3415 BasicBlock *BB = *BI;
3417 // We don't support switch statements inside loops.
3418 if (!isa<BranchInst>(BB->getTerminator())) {
3419 emitAnalysis(Report(BB->getTerminator())
3420 << "loop contains a switch statement");
3424 // We must be able to predicate all blocks that need to be predicated.
3425 if (blockNeedsPredication(BB)) {
3426 if (!blockCanBePredicated(BB, SafePointes)) {
3427 emitAnalysis(Report(BB->getTerminator())
3428 << "control flow cannot be substituted for a select");
3431 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3432 emitAnalysis(Report(BB->getTerminator())
3433 << "control flow cannot be substituted for a select");
3438 // We can if-convert this loop.
3442 bool LoopVectorizationLegality::canVectorize() {
3443 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3444 // be canonicalized.
3445 if (!TheLoop->getLoopPreheader()) {
3447 Report() << "loop control flow is not understood by vectorizer");
3451 // We can only vectorize innermost loops.
3452 if (TheLoop->getSubLoopsVector().size()) {
3453 emitAnalysis(Report() << "loop is not the innermost loop");
3457 // We must have a single backedge.
3458 if (TheLoop->getNumBackEdges() != 1) {
3460 Report() << "loop control flow is not understood by vectorizer");
3464 // We must have a single exiting block.
3465 if (!TheLoop->getExitingBlock()) {
3467 Report() << "loop control flow is not understood by vectorizer");
3471 // We need to have a loop header.
3472 DEBUG(dbgs() << "LV: Found a loop: " <<
3473 TheLoop->getHeader()->getName() << '\n');
3475 // Check if we can if-convert non-single-bb loops.
3476 unsigned NumBlocks = TheLoop->getNumBlocks();
3477 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3478 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3482 // ScalarEvolution needs to be able to find the exit count.
3483 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3484 if (ExitCount == SE->getCouldNotCompute()) {
3485 emitAnalysis(Report() << "could not determine number of loop iterations");
3486 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3490 // Check if we can vectorize the instructions and CFG in this loop.
3491 if (!canVectorizeInstrs()) {
3492 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3496 // Go over each instruction and look at memory deps.
3497 if (!canVectorizeMemory()) {
3498 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3502 // Collect all of the variables that remain uniform after vectorization.
3503 collectLoopUniforms();
3505 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3506 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3509 // Okay! We can vectorize. At this point we don't have any other mem analysis
3510 // which may limit our maximum vectorization factor, so just return true with
3515 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3516 if (Ty->isPointerTy())
3517 return DL.getIntPtrType(Ty);
3519 // It is possible that char's or short's overflow when we ask for the loop's
3520 // trip count, work around this by changing the type size.
3521 if (Ty->getScalarSizeInBits() < 32)
3522 return Type::getInt32Ty(Ty->getContext());
3527 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3528 Ty0 = convertPointerToIntegerType(DL, Ty0);
3529 Ty1 = convertPointerToIntegerType(DL, Ty1);
3530 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3535 /// \brief Check that the instruction has outside loop users and is not an
3536 /// identified reduction variable.
3537 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3538 SmallPtrSet<Value *, 4> &Reductions) {
3539 // Reduction instructions are allowed to have exit users. All other
3540 // instructions must not have external users.
3541 if (!Reductions.count(Inst))
3542 //Check that all of the users of the loop are inside the BB.
3543 for (User *U : Inst->users()) {
3544 Instruction *UI = cast<Instruction>(U);
3545 // This user may be a reduction exit value.
3546 if (!TheLoop->contains(UI)) {
3547 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3554 bool LoopVectorizationLegality::canVectorizeInstrs() {
3555 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3556 BasicBlock *Header = TheLoop->getHeader();
3558 // Look for the attribute signaling the absence of NaNs.
3559 Function &F = *Header->getParent();
3560 if (F.hasFnAttribute("no-nans-fp-math"))
3561 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3562 AttributeSet::FunctionIndex,
3563 "no-nans-fp-math").getValueAsString() == "true";
3565 // For each block in the loop.
3566 for (Loop::block_iterator bb = TheLoop->block_begin(),
3567 be = TheLoop->block_end(); bb != be; ++bb) {
3569 // Scan the instructions in the block and look for hazards.
3570 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3573 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3574 Type *PhiTy = Phi->getType();
3575 // Check that this PHI type is allowed.
3576 if (!PhiTy->isIntegerTy() &&
3577 !PhiTy->isFloatingPointTy() &&
3578 !PhiTy->isPointerTy()) {
3579 emitAnalysis(Report(it)
3580 << "loop control flow is not understood by vectorizer");
3581 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3585 // If this PHINode is not in the header block, then we know that we
3586 // can convert it to select during if-conversion. No need to check if
3587 // the PHIs in this block are induction or reduction variables.
3588 if (*bb != Header) {
3589 // Check that this instruction has no outside users or is an
3590 // identified reduction value with an outside user.
3591 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3593 emitAnalysis(Report(it) << "value that could not be identified as "
3594 "reduction is used outside the loop");
3598 // We only allow if-converted PHIs with more than two incoming values.
3599 if (Phi->getNumIncomingValues() != 2) {
3600 emitAnalysis(Report(it)
3601 << "control flow not understood by vectorizer");
3602 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3606 // This is the value coming from the preheader.
3607 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3608 // Check if this is an induction variable.
3609 InductionKind IK = isInductionVariable(Phi);
3611 if (IK_NoInduction != IK) {
3612 // Get the widest type.
3614 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3616 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3618 // Int inductions are special because we only allow one IV.
3619 if (IK == IK_IntInduction) {
3620 // Use the phi node with the widest type as induction. Use the last
3621 // one if there are multiple (no good reason for doing this other
3622 // than it is expedient).
3623 if (!Induction || PhiTy == WidestIndTy)
3627 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3628 Inductions[Phi] = InductionInfo(StartValue, IK);
3630 // Until we explicitly handle the case of an induction variable with
3631 // an outside loop user we have to give up vectorizing this loop.
3632 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3633 emitAnalysis(Report(it) << "use of induction value outside of the "
3634 "loop is not handled by vectorizer");
3641 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3642 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3645 if (AddReductionVar(Phi, RK_IntegerMult)) {
3646 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3649 if (AddReductionVar(Phi, RK_IntegerOr)) {
3650 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3653 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3654 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3657 if (AddReductionVar(Phi, RK_IntegerXor)) {
3658 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3661 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3662 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3665 if (AddReductionVar(Phi, RK_FloatMult)) {
3666 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3669 if (AddReductionVar(Phi, RK_FloatAdd)) {
3670 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3673 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3674 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3679 emitAnalysis(Report(it) << "unvectorizable operation");
3680 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3682 }// end of PHI handling
3684 // We still don't handle functions. However, we can ignore dbg intrinsic
3685 // calls and we do handle certain intrinsic and libm functions.
3686 CallInst *CI = dyn_cast<CallInst>(it);
3687 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3688 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3689 DEBUG(dbgs() << "LV: Found a call site.\n");
3693 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3694 // second argument is the same (i.e. loop invariant)
3696 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3697 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3698 emitAnalysis(Report(it)
3699 << "intrinsic instruction cannot be vectorized");
3700 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3705 // Check that the instruction return type is vectorizable.
3706 // Also, we can't vectorize extractelement instructions.
3707 if ((!VectorType::isValidElementType(it->getType()) &&
3708 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3709 emitAnalysis(Report(it)
3710 << "instruction return type cannot be vectorized");
3711 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3715 // Check that the stored type is vectorizable.
3716 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3717 Type *T = ST->getValueOperand()->getType();
3718 if (!VectorType::isValidElementType(T)) {
3719 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3722 if (EnableMemAccessVersioning)
3723 collectStridedAcccess(ST);
3726 if (EnableMemAccessVersioning)
3727 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3728 collectStridedAcccess(LI);
3730 // Reduction instructions are allowed to have exit users.
3731 // All other instructions must not have external users.
3732 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3733 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3742 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3743 if (Inductions.empty()) {
3744 emitAnalysis(Report()
3745 << "loop induction variable could not be identified");
3753 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3754 /// return the induction operand of the gep pointer.
3755 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3756 const DataLayout *DL, Loop *Lp) {
3757 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3761 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3763 // Check that all of the gep indices are uniform except for our induction
3765 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3766 if (i != InductionOperand &&
3767 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3769 return GEP->getOperand(InductionOperand);
3772 ///\brief Look for a cast use of the passed value.
3773 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3774 Value *UniqueCast = nullptr;
3775 for (User *U : Ptr->users()) {
3776 CastInst *CI = dyn_cast<CastInst>(U);
3777 if (CI && CI->getType() == Ty) {
3787 ///\brief Get the stride of a pointer access in a loop.
3788 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3789 /// pointer to the Value, or null otherwise.
3790 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3791 const DataLayout *DL, Loop *Lp) {
3792 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3793 if (!PtrTy || PtrTy->isAggregateType())
3796 // Try to remove a gep instruction to make the pointer (actually index at this
3797 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3798 // pointer, otherwise, we are analyzing the index.
3799 Value *OrigPtr = Ptr;
3801 // The size of the pointer access.
3802 int64_t PtrAccessSize = 1;
3804 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3805 const SCEV *V = SE->getSCEV(Ptr);
3809 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3810 V = C->getOperand();
3812 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3816 V = S->getStepRecurrence(*SE);
3820 // Strip off the size of access multiplication if we are still analyzing the
3822 if (OrigPtr == Ptr) {
3823 DL->getTypeAllocSize(PtrTy->getElementType());
3824 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3825 if (M->getOperand(0)->getSCEVType() != scConstant)
3828 const APInt &APStepVal =
3829 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3831 // Huge step value - give up.
3832 if (APStepVal.getBitWidth() > 64)
3835 int64_t StepVal = APStepVal.getSExtValue();
3836 if (PtrAccessSize != StepVal)
3838 V = M->getOperand(1);
3843 Type *StripedOffRecurrenceCast = nullptr;
3844 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3845 StripedOffRecurrenceCast = C->getType();
3846 V = C->getOperand();
3849 // Look for the loop invariant symbolic value.
3850 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3854 Value *Stride = U->getValue();
3855 if (!Lp->isLoopInvariant(Stride))
3858 // If we have stripped off the recurrence cast we have to make sure that we
3859 // return the value that is used in this loop so that we can replace it later.
3860 if (StripedOffRecurrenceCast)
3861 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3866 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3867 Value *Ptr = nullptr;
3868 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3869 Ptr = LI->getPointerOperand();
3870 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3871 Ptr = SI->getPointerOperand();
3875 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3879 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3880 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3881 Strides[Ptr] = Stride;
3882 StrideSet.insert(Stride);
3885 void LoopVectorizationLegality::collectLoopUniforms() {
3886 // We now know that the loop is vectorizable!
3887 // Collect variables that will remain uniform after vectorization.
3888 std::vector<Value*> Worklist;
3889 BasicBlock *Latch = TheLoop->getLoopLatch();
3891 // Start with the conditional branch and walk up the block.
3892 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3894 // Also add all consecutive pointer values; these values will be uniform
3895 // after vectorization (and subsequent cleanup) and, until revectorization is
3896 // supported, all dependencies must also be uniform.
3897 for (Loop::block_iterator B = TheLoop->block_begin(),
3898 BE = TheLoop->block_end(); B != BE; ++B)
3899 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3901 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3902 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3904 while (Worklist.size()) {
3905 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3906 Worklist.pop_back();
3908 // Look at instructions inside this loop.
3909 // Stop when reaching PHI nodes.
3910 // TODO: we need to follow values all over the loop, not only in this block.
3911 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3914 // This is a known uniform.
3917 // Insert all operands.
3918 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3923 /// \brief Analyses memory accesses in a loop.
3925 /// Checks whether run time pointer checks are needed and builds sets for data
3926 /// dependence checking.
3927 class AccessAnalysis {
3929 /// \brief Read or write access location.
3930 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3931 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3933 /// \brief Set of potential dependent memory accesses.
3934 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3936 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
3937 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
3939 /// \brief Register a load and whether it is only read from.
3940 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
3941 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3942 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3943 Accesses.insert(MemAccessInfo(Ptr, false));
3945 ReadOnlyPtr.insert(Ptr);
3948 /// \brief Register a store.
3949 void addStore(AliasAnalysis::Location &Loc) {
3950 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3951 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3952 Accesses.insert(MemAccessInfo(Ptr, true));
3955 /// \brief Check whether we can check the pointers at runtime for
3956 /// non-intersection.
3957 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3958 unsigned &NumComparisons, ScalarEvolution *SE,
3959 Loop *TheLoop, ValueToValueMap &Strides,
3960 bool ShouldCheckStride = false);
3962 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3963 /// and builds sets of dependent accesses.
3964 void buildDependenceSets() {
3965 processMemAccesses();
3968 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3970 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3971 void resetDepChecks() { CheckDeps.clear(); }
3973 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3976 typedef SetVector<MemAccessInfo> PtrAccessSet;
3978 /// \brief Go over all memory access and check whether runtime pointer checks
3979 /// are needed /// and build sets of dependency check candidates.
3980 void processMemAccesses();
3982 /// Set of all accesses.
3983 PtrAccessSet Accesses;
3985 /// Set of accesses that need a further dependence check.
3986 MemAccessInfoSet CheckDeps;
3988 /// Set of pointers that are read only.
3989 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3991 const DataLayout *DL;
3993 /// An alias set tracker to partition the access set by underlying object and
3994 //intrinsic property (such as TBAA metadata).
3995 AliasSetTracker AST;
3997 /// Sets of potentially dependent accesses - members of one set share an
3998 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3999 /// dependence check.
4000 DepCandidates &DepCands;
4002 bool IsRTCheckNeeded;
4005 } // end anonymous namespace
4007 /// \brief Check whether a pointer can participate in a runtime bounds check.
4008 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4010 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4011 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4015 return AR->isAffine();
4018 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4019 /// the address space.
4020 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4021 const Loop *Lp, ValueToValueMap &StridesMap);
4023 bool AccessAnalysis::canCheckPtrAtRT(
4024 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4025 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4026 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4027 // Find pointers with computable bounds. We are going to use this information
4028 // to place a runtime bound check.
4029 bool CanDoRT = true;
4031 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4034 // We assign a consecutive id to access from different alias sets.
4035 // Accesses between different groups doesn't need to be checked.
4037 for (auto &AS : AST) {
4038 unsigned NumReadPtrChecks = 0;
4039 unsigned NumWritePtrChecks = 0;
4041 // We assign consecutive id to access from different dependence sets.
4042 // Accesses within the same set don't need a runtime check.
4043 unsigned RunningDepId = 1;
4044 DenseMap<Value *, unsigned> DepSetId;
4047 Value *Ptr = A.getValue();
4048 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4049 MemAccessInfo Access(Ptr, IsWrite);
4052 ++NumWritePtrChecks;
4056 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4057 // When we run after a failing dependency check we have to make sure we
4058 // don't have wrapping pointers.
4059 (!ShouldCheckStride ||
4060 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4061 // The id of the dependence set.
4064 if (IsDepCheckNeeded) {
4065 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4066 unsigned &LeaderId = DepSetId[Leader];
4068 LeaderId = RunningDepId++;
4071 // Each access has its own dependence set.
4072 DepId = RunningDepId++;
4074 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4076 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4082 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4083 NumComparisons += 0; // Only one dependence set.
4085 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4086 NumWritePtrChecks - 1));
4092 // If the pointers that we would use for the bounds comparison have different
4093 // address spaces, assume the values aren't directly comparable, so we can't
4094 // use them for the runtime check. We also have to assume they could
4095 // overlap. In the future there should be metadata for whether address spaces
4097 unsigned NumPointers = RtCheck.Pointers.size();
4098 for (unsigned i = 0; i < NumPointers; ++i) {
4099 for (unsigned j = i + 1; j < NumPointers; ++j) {
4100 // Only need to check pointers between two different dependency sets.
4101 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4103 // Only need to check pointers in the same alias set.
4104 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4107 Value *PtrI = RtCheck.Pointers[i];
4108 Value *PtrJ = RtCheck.Pointers[j];
4110 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4111 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4113 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4114 " different address spaces\n");
4123 void AccessAnalysis::processMemAccesses() {
4124 // We process the set twice: first we process read-write pointers, last we
4125 // process read-only pointers. This allows us to skip dependence tests for
4126 // read-only pointers.
4128 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4129 DEBUG(dbgs() << " AST: "; AST.dump());
4130 DEBUG(dbgs() << "LV: Accesses:\n");
4132 for (auto A : Accesses)
4133 dbgs() << "\t" << *A.getPointer() << " (" <<
4134 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4135 "read-only" : "read")) << ")\n";
4138 // The AliasSetTracker has nicely partitioned our pointers by metadata
4139 // compatibility and potential for underlying-object overlap. As a result, we
4140 // only need to check for potential pointer dependencies within each alias
4142 for (auto &AS : AST) {
4143 // Note that both the alias-set tracker and the alias sets themselves used
4144 // linked lists internally and so the iteration order here is deterministic
4145 // (matching the original instruction order within each set).
4147 bool SetHasWrite = false;
4149 // Map of pointers to last access encountered.
4150 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4151 UnderlyingObjToAccessMap ObjToLastAccess;
4153 // Set of access to check after all writes have been processed.
4154 PtrAccessSet DeferredAccesses;
4156 // Iterate over each alias set twice, once to process read/write pointers,
4157 // and then to process read-only pointers.
4158 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4159 bool UseDeferred = SetIteration > 0;
4160 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4163 Value *Ptr = A.getValue();
4164 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4166 // If we're using the deferred access set, then it contains only reads.
4167 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4168 if (UseDeferred && !IsReadOnlyPtr)
4170 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4172 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4173 S.count(MemAccessInfo(Ptr, false))) &&
4174 "Alias-set pointer not in the access set?");
4176 MemAccessInfo Access(Ptr, IsWrite);
4177 DepCands.insert(Access);
4179 // Memorize read-only pointers for later processing and skip them in the
4180 // first round (they need to be checked after we have seen all write
4181 // pointers). Note: we also mark pointer that are not consecutive as
4182 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4183 // the second check for "!IsWrite".
4184 if (!UseDeferred && IsReadOnlyPtr) {
4185 DeferredAccesses.insert(Access);
4189 // If this is a write - check other reads and writes for conflicts. If
4190 // this is a read only check other writes for conflicts (but only if
4191 // there is no other write to the ptr - this is an optimization to
4192 // catch "a[i] = a[i] + " without having to do a dependence check).
4193 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4194 CheckDeps.insert(Access);
4195 IsRTCheckNeeded = true;
4201 // Create sets of pointers connected by a shared alias set and
4202 // underlying object.
4203 typedef SmallVector<Value*, 16> ValueVector;
4204 ValueVector TempObjects;
4205 GetUnderlyingObjects(Ptr, TempObjects, DL);
4206 for (Value *UnderlyingObj : TempObjects) {
4207 UnderlyingObjToAccessMap::iterator Prev =
4208 ObjToLastAccess.find(UnderlyingObj);
4209 if (Prev != ObjToLastAccess.end())
4210 DepCands.unionSets(Access, Prev->second);
4212 ObjToLastAccess[UnderlyingObj] = Access;
4220 /// \brief Checks memory dependences among accesses to the same underlying
4221 /// object to determine whether there vectorization is legal or not (and at
4222 /// which vectorization factor).
4224 /// This class works under the assumption that we already checked that memory
4225 /// locations with different underlying pointers are "must-not alias".
4226 /// We use the ScalarEvolution framework to symbolically evalutate access
4227 /// functions pairs. Since we currently don't restructure the loop we can rely
4228 /// on the program order of memory accesses to determine their safety.
4229 /// At the moment we will only deem accesses as safe for:
4230 /// * A negative constant distance assuming program order.
4232 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4233 /// a[i] = tmp; y = a[i];
4235 /// The latter case is safe because later checks guarantuee that there can't
4236 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4237 /// the same variable: a header phi can only be an induction or a reduction, a
4238 /// reduction can't have a memory sink, an induction can't have a memory
4239 /// source). This is important and must not be violated (or we have to
4240 /// resort to checking for cycles through memory).
4242 /// * A positive constant distance assuming program order that is bigger
4243 /// than the biggest memory access.
4245 /// tmp = a[i] OR b[i] = x
4246 /// a[i+2] = tmp y = b[i+2];
4248 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4250 /// * Zero distances and all accesses have the same size.
4252 class MemoryDepChecker {
4254 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4255 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4257 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4258 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4259 ShouldRetryWithRuntimeCheck(false) {}
4261 /// \brief Register the location (instructions are given increasing numbers)
4262 /// of a write access.
4263 void addAccess(StoreInst *SI) {
4264 Value *Ptr = SI->getPointerOperand();
4265 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4266 InstMap.push_back(SI);
4270 /// \brief Register the location (instructions are given increasing numbers)
4271 /// of a write access.
4272 void addAccess(LoadInst *LI) {
4273 Value *Ptr = LI->getPointerOperand();
4274 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4275 InstMap.push_back(LI);
4279 /// \brief Check whether the dependencies between the accesses are safe.
4281 /// Only checks sets with elements in \p CheckDeps.
4282 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4283 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4285 /// \brief The maximum number of bytes of a vector register we can vectorize
4286 /// the accesses safely with.
4287 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4289 /// \brief In same cases when the dependency check fails we can still
4290 /// vectorize the loop with a dynamic array access check.
4291 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4294 ScalarEvolution *SE;
4295 const DataLayout *DL;
4296 const Loop *InnermostLoop;
4298 /// \brief Maps access locations (ptr, read/write) to program order.
4299 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4301 /// \brief Memory access instructions in program order.
4302 SmallVector<Instruction *, 16> InstMap;
4304 /// \brief The program order index to be used for the next instruction.
4307 // We can access this many bytes in parallel safely.
4308 unsigned MaxSafeDepDistBytes;
4310 /// \brief If we see a non-constant dependence distance we can still try to
4311 /// vectorize this loop with runtime checks.
4312 bool ShouldRetryWithRuntimeCheck;
4314 /// \brief Check whether there is a plausible dependence between the two
4317 /// Access \p A must happen before \p B in program order. The two indices
4318 /// identify the index into the program order map.
4320 /// This function checks whether there is a plausible dependence (or the
4321 /// absence of such can't be proved) between the two accesses. If there is a
4322 /// plausible dependence but the dependence distance is bigger than one
4323 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4324 /// distance is smaller than any other distance encountered so far).
4325 /// Otherwise, this function returns true signaling a possible dependence.
4326 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4327 const MemAccessInfo &B, unsigned BIdx,
4328 ValueToValueMap &Strides);
4330 /// \brief Check whether the data dependence could prevent store-load
4332 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4335 } // end anonymous namespace
4337 static bool isInBoundsGep(Value *Ptr) {
4338 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4339 return GEP->isInBounds();
4343 /// \brief Check whether the access through \p Ptr has a constant stride.
4344 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4345 const Loop *Lp, ValueToValueMap &StridesMap) {
4346 const Type *Ty = Ptr->getType();
4347 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4349 // Make sure that the pointer does not point to aggregate types.
4350 const PointerType *PtrTy = cast<PointerType>(Ty);
4351 if (PtrTy->getElementType()->isAggregateType()) {
4352 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4357 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4359 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4361 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4362 << *Ptr << " SCEV: " << *PtrScev << "\n");
4366 // The accesss function must stride over the innermost loop.
4367 if (Lp != AR->getLoop()) {
4368 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4369 *Ptr << " SCEV: " << *PtrScev << "\n");
4372 // The address calculation must not wrap. Otherwise, a dependence could be
4374 // An inbounds getelementptr that is a AddRec with a unit stride
4375 // cannot wrap per definition. The unit stride requirement is checked later.
4376 // An getelementptr without an inbounds attribute and unit stride would have
4377 // to access the pointer value "0" which is undefined behavior in address
4378 // space 0, therefore we can also vectorize this case.
4379 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4380 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4381 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4382 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4383 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4384 << *Ptr << " SCEV: " << *PtrScev << "\n");
4388 // Check the step is constant.
4389 const SCEV *Step = AR->getStepRecurrence(*SE);
4391 // Calculate the pointer stride and check if it is consecutive.
4392 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4394 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4395 " SCEV: " << *PtrScev << "\n");
4399 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4400 const APInt &APStepVal = C->getValue()->getValue();
4402 // Huge step value - give up.
4403 if (APStepVal.getBitWidth() > 64)
4406 int64_t StepVal = APStepVal.getSExtValue();
4409 int64_t Stride = StepVal / Size;
4410 int64_t Rem = StepVal % Size;
4414 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4415 // know we can't "wrap around the address space". In case of address space
4416 // zero we know that this won't happen without triggering undefined behavior.
4417 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4418 Stride != 1 && Stride != -1)
4424 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4425 unsigned TypeByteSize) {
4426 // If loads occur at a distance that is not a multiple of a feasible vector
4427 // factor store-load forwarding does not take place.
4428 // Positive dependences might cause troubles because vectorizing them might
4429 // prevent store-load forwarding making vectorized code run a lot slower.
4430 // a[i] = a[i-3] ^ a[i-8];
4431 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4432 // hence on your typical architecture store-load forwarding does not take
4433 // place. Vectorizing in such cases does not make sense.
4434 // Store-load forwarding distance.
4435 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4436 // Maximum vector factor.
4437 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4438 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4439 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4441 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4443 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4444 MaxVFWithoutSLForwardIssues = (vf >>=1);
4449 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4450 DEBUG(dbgs() << "LV: Distance " << Distance <<
4451 " that could cause a store-load forwarding conflict\n");
4455 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4456 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4457 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4461 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4462 const MemAccessInfo &B, unsigned BIdx,
4463 ValueToValueMap &Strides) {
4464 assert (AIdx < BIdx && "Must pass arguments in program order");
4466 Value *APtr = A.getPointer();
4467 Value *BPtr = B.getPointer();
4468 bool AIsWrite = A.getInt();
4469 bool BIsWrite = B.getInt();
4471 // Two reads are independent.
4472 if (!AIsWrite && !BIsWrite)
4475 // We cannot check pointers in different address spaces.
4476 if (APtr->getType()->getPointerAddressSpace() !=
4477 BPtr->getType()->getPointerAddressSpace())
4480 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4481 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4483 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4484 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4486 const SCEV *Src = AScev;
4487 const SCEV *Sink = BScev;
4489 // If the induction step is negative we have to invert source and sink of the
4491 if (StrideAPtr < 0) {
4494 std::swap(APtr, BPtr);
4495 std::swap(Src, Sink);
4496 std::swap(AIsWrite, BIsWrite);
4497 std::swap(AIdx, BIdx);
4498 std::swap(StrideAPtr, StrideBPtr);
4501 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4503 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4504 << "(Induction step: " << StrideAPtr << ")\n");
4505 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4506 << *InstMap[BIdx] << ": " << *Dist << "\n");
4508 // Need consecutive accesses. We don't want to vectorize
4509 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4510 // the address space.
4511 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4512 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4516 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4518 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4519 ShouldRetryWithRuntimeCheck = true;
4523 Type *ATy = APtr->getType()->getPointerElementType();
4524 Type *BTy = BPtr->getType()->getPointerElementType();
4525 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4527 // Negative distances are not plausible dependencies.
4528 const APInt &Val = C->getValue()->getValue();
4529 if (Val.isNegative()) {
4530 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4531 if (IsTrueDataDependence &&
4532 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4536 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4540 // Write to the same location with the same size.
4541 // Could be improved to assert type sizes are the same (i32 == float, etc).
4545 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4549 assert(Val.isStrictlyPositive() && "Expect a positive value");
4551 // Positive distance bigger than max vectorization factor.
4554 "LV: ReadWrite-Write positive dependency with different types\n");
4558 unsigned Distance = (unsigned) Val.getZExtValue();
4560 // Bail out early if passed-in parameters make vectorization not feasible.
4561 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4562 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4564 // The distance must be bigger than the size needed for a vectorized version
4565 // of the operation and the size of the vectorized operation must not be
4566 // bigger than the currrent maximum size.
4567 if (Distance < 2*TypeByteSize ||
4568 2*TypeByteSize > MaxSafeDepDistBytes ||
4569 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4570 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4571 << Val.getSExtValue() << '\n');
4575 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4576 Distance : MaxSafeDepDistBytes;
4578 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4579 if (IsTrueDataDependence &&
4580 couldPreventStoreLoadForward(Distance, TypeByteSize))
4583 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4584 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4589 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4590 MemAccessInfoSet &CheckDeps,
4591 ValueToValueMap &Strides) {
4593 MaxSafeDepDistBytes = -1U;
4594 while (!CheckDeps.empty()) {
4595 MemAccessInfo CurAccess = *CheckDeps.begin();
4597 // Get the relevant memory access set.
4598 EquivalenceClasses<MemAccessInfo>::iterator I =
4599 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4601 // Check accesses within this set.
4602 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4603 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4605 // Check every access pair.
4607 CheckDeps.erase(*AI);
4608 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4610 // Check every accessing instruction pair in program order.
4611 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4612 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4613 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4614 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4615 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4617 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4628 bool LoopVectorizationLegality::canVectorizeMemory() {
4630 typedef SmallVector<Value*, 16> ValueVector;
4631 typedef SmallPtrSet<Value*, 16> ValueSet;
4633 // Holds the Load and Store *instructions*.
4637 // Holds all the different accesses in the loop.
4638 unsigned NumReads = 0;
4639 unsigned NumReadWrites = 0;
4641 PtrRtCheck.Pointers.clear();
4642 PtrRtCheck.Need = false;
4644 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4645 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4648 for (Loop::block_iterator bb = TheLoop->block_begin(),
4649 be = TheLoop->block_end(); bb != be; ++bb) {
4651 // Scan the BB and collect legal loads and stores.
4652 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4655 // If this is a load, save it. If this instruction can read from memory
4656 // but is not a load, then we quit. Notice that we don't handle function
4657 // calls that read or write.
4658 if (it->mayReadFromMemory()) {
4659 // Many math library functions read the rounding mode. We will only
4660 // vectorize a loop if it contains known function calls that don't set
4661 // the flag. Therefore, it is safe to ignore this read from memory.
4662 CallInst *Call = dyn_cast<CallInst>(it);
4663 if (Call && getIntrinsicIDForCall(Call, TLI))
4666 LoadInst *Ld = dyn_cast<LoadInst>(it);
4667 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4668 emitAnalysis(Report(Ld)
4669 << "read with atomic ordering or volatile read");
4670 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4674 Loads.push_back(Ld);
4675 DepChecker.addAccess(Ld);
4679 // Save 'store' instructions. Abort if other instructions write to memory.
4680 if (it->mayWriteToMemory()) {
4681 StoreInst *St = dyn_cast<StoreInst>(it);
4683 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4686 if (!St->isSimple() && !IsAnnotatedParallel) {
4687 emitAnalysis(Report(St)
4688 << "write with atomic ordering or volatile write");
4689 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4693 Stores.push_back(St);
4694 DepChecker.addAccess(St);
4699 // Now we have two lists that hold the loads and the stores.
4700 // Next, we find the pointers that they use.
4702 // Check if we see any stores. If there are no stores, then we don't
4703 // care if the pointers are *restrict*.
4704 if (!Stores.size()) {
4705 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4709 AccessAnalysis::DepCandidates DependentAccesses;
4710 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4712 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4713 // multiple times on the same object. If the ptr is accessed twice, once
4714 // for read and once for write, it will only appear once (on the write
4715 // list). This is okay, since we are going to check for conflicts between
4716 // writes and between reads and writes, but not between reads and reads.
4719 ValueVector::iterator I, IE;
4720 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4721 StoreInst *ST = cast<StoreInst>(*I);
4722 Value* Ptr = ST->getPointerOperand();
4724 if (isUniform(Ptr)) {
4727 << "write to a loop invariant address could not be vectorized");
4728 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4732 // If we did *not* see this pointer before, insert it to the read-write
4733 // list. At this phase it is only a 'write' list.
4734 if (Seen.insert(Ptr)) {
4737 AliasAnalysis::Location Loc = AA->getLocation(ST);
4738 // The TBAA metadata could have a control dependency on the predication
4739 // condition, so we cannot rely on it when determining whether or not we
4740 // need runtime pointer checks.
4741 if (blockNeedsPredication(ST->getParent()))
4742 Loc.AATags.TBAA = nullptr;
4744 Accesses.addStore(Loc);
4748 if (IsAnnotatedParallel) {
4750 << "LV: A loop annotated parallel, ignore memory dependency "
4755 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4756 LoadInst *LD = cast<LoadInst>(*I);
4757 Value* Ptr = LD->getPointerOperand();
4758 // If we did *not* see this pointer before, insert it to the
4759 // read list. If we *did* see it before, then it is already in
4760 // the read-write list. This allows us to vectorize expressions
4761 // such as A[i] += x; Because the address of A[i] is a read-write
4762 // pointer. This only works if the index of A[i] is consecutive.
4763 // If the address of i is unknown (for example A[B[i]]) then we may
4764 // read a few words, modify, and write a few words, and some of the
4765 // words may be written to the same address.
4766 bool IsReadOnlyPtr = false;
4767 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4769 IsReadOnlyPtr = true;
4772 AliasAnalysis::Location Loc = AA->getLocation(LD);
4773 // The TBAA metadata could have a control dependency on the predication
4774 // condition, so we cannot rely on it when determining whether or not we
4775 // need runtime pointer checks.
4776 if (blockNeedsPredication(LD->getParent()))
4777 Loc.AATags.TBAA = nullptr;
4779 Accesses.addLoad(Loc, IsReadOnlyPtr);
4782 // If we write (or read-write) to a single destination and there are no
4783 // other reads in this loop then is it safe to vectorize.
4784 if (NumReadWrites == 1 && NumReads == 0) {
4785 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4789 // Build dependence sets and check whether we need a runtime pointer bounds
4791 Accesses.buildDependenceSets();
4792 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4794 // Find pointers with computable bounds. We are going to use this information
4795 // to place a runtime bound check.
4796 unsigned NumComparisons = 0;
4797 bool CanDoRT = false;
4799 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4802 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4803 " pointer comparisons.\n");
4805 // If we only have one set of dependences to check pointers among we don't
4806 // need a runtime check.
4807 if (NumComparisons == 0 && NeedRTCheck)
4808 NeedRTCheck = false;
4810 // Check that we did not collect too many pointers or found an unsizeable
4812 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4818 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4821 if (NeedRTCheck && !CanDoRT) {
4822 emitAnalysis(Report() << "cannot identify array bounds");
4823 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4824 "the array bounds.\n");
4829 PtrRtCheck.Need = NeedRTCheck;
4831 bool CanVecMem = true;
4832 if (Accesses.isDependencyCheckNeeded()) {
4833 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4834 CanVecMem = DepChecker.areDepsSafe(
4835 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4836 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4838 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4839 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4842 // Clear the dependency checks. We assume they are not needed.
4843 Accesses.resetDepChecks();
4846 PtrRtCheck.Need = true;
4848 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4849 TheLoop, Strides, true);
4850 // Check that we did not collect too many pointers or found an unsizeable
4852 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4853 if (!CanDoRT && NumComparisons > 0)
4854 emitAnalysis(Report()
4855 << "cannot check memory dependencies at runtime");
4857 emitAnalysis(Report()
4858 << NumComparisons << " exceeds limit of "
4859 << RuntimeMemoryCheckThreshold
4860 << " dependent memory operations checked at runtime");
4861 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4871 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4873 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4874 " need a runtime memory check.\n");
4879 static bool hasMultipleUsesOf(Instruction *I,
4880 SmallPtrSet<Instruction *, 8> &Insts) {
4881 unsigned NumUses = 0;
4882 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4883 if (Insts.count(dyn_cast<Instruction>(*Use)))
4892 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4893 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4894 if (!Set.count(dyn_cast<Instruction>(*Use)))
4899 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4900 ReductionKind Kind) {
4901 if (Phi->getNumIncomingValues() != 2)
4904 // Reduction variables are only found in the loop header block.
4905 if (Phi->getParent() != TheLoop->getHeader())
4908 // Obtain the reduction start value from the value that comes from the loop
4910 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4912 // ExitInstruction is the single value which is used outside the loop.
4913 // We only allow for a single reduction value to be used outside the loop.
4914 // This includes users of the reduction, variables (which form a cycle
4915 // which ends in the phi node).
4916 Instruction *ExitInstruction = nullptr;
4917 // Indicates that we found a reduction operation in our scan.
4918 bool FoundReduxOp = false;
4920 // We start with the PHI node and scan for all of the users of this
4921 // instruction. All users must be instructions that can be used as reduction
4922 // variables (such as ADD). We must have a single out-of-block user. The cycle
4923 // must include the original PHI.
4924 bool FoundStartPHI = false;
4926 // To recognize min/max patterns formed by a icmp select sequence, we store
4927 // the number of instruction we saw from the recognized min/max pattern,
4928 // to make sure we only see exactly the two instructions.
4929 unsigned NumCmpSelectPatternInst = 0;
4930 ReductionInstDesc ReduxDesc(false, nullptr);
4932 SmallPtrSet<Instruction *, 8> VisitedInsts;
4933 SmallVector<Instruction *, 8> Worklist;
4934 Worklist.push_back(Phi);
4935 VisitedInsts.insert(Phi);
4937 // A value in the reduction can be used:
4938 // - By the reduction:
4939 // - Reduction operation:
4940 // - One use of reduction value (safe).
4941 // - Multiple use of reduction value (not safe).
4943 // - All uses of the PHI must be the reduction (safe).
4944 // - Otherwise, not safe.
4945 // - By one instruction outside of the loop (safe).
4946 // - By further instructions outside of the loop (not safe).
4947 // - By an instruction that is not part of the reduction (not safe).
4949 // * An instruction type other than PHI or the reduction operation.
4950 // * A PHI in the header other than the initial PHI.
4951 while (!Worklist.empty()) {
4952 Instruction *Cur = Worklist.back();
4953 Worklist.pop_back();
4956 // If the instruction has no users then this is a broken chain and can't be
4957 // a reduction variable.
4958 if (Cur->use_empty())
4961 bool IsAPhi = isa<PHINode>(Cur);
4963 // A header PHI use other than the original PHI.
4964 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4967 // Reductions of instructions such as Div, and Sub is only possible if the
4968 // LHS is the reduction variable.
4969 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4970 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4971 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4974 // Any reduction instruction must be of one of the allowed kinds.
4975 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4976 if (!ReduxDesc.IsReduction)
4979 // A reduction operation must only have one use of the reduction value.
4980 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4981 hasMultipleUsesOf(Cur, VisitedInsts))
4984 // All inputs to a PHI node must be a reduction value.
4985 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4988 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4989 isa<SelectInst>(Cur)))
4990 ++NumCmpSelectPatternInst;
4991 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4992 isa<SelectInst>(Cur)))
4993 ++NumCmpSelectPatternInst;
4995 // Check whether we found a reduction operator.
4996 FoundReduxOp |= !IsAPhi;
4998 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4999 // onto the stack. This way we are going to have seen all inputs to PHI
5000 // nodes once we get to them.
5001 SmallVector<Instruction *, 8> NonPHIs;
5002 SmallVector<Instruction *, 8> PHIs;
5003 for (User *U : Cur->users()) {
5004 Instruction *UI = cast<Instruction>(U);
5006 // Check if we found the exit user.
5007 BasicBlock *Parent = UI->getParent();
5008 if (!TheLoop->contains(Parent)) {
5009 // Exit if you find multiple outside users or if the header phi node is
5010 // being used. In this case the user uses the value of the previous
5011 // iteration, in which case we would loose "VF-1" iterations of the
5012 // reduction operation if we vectorize.
5013 if (ExitInstruction != nullptr || Cur == Phi)
5016 // The instruction used by an outside user must be the last instruction
5017 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5018 // operations on the value.
5019 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5022 ExitInstruction = Cur;
5026 // Process instructions only once (termination). Each reduction cycle
5027 // value must only be used once, except by phi nodes and min/max
5028 // reductions which are represented as a cmp followed by a select.
5029 ReductionInstDesc IgnoredVal(false, nullptr);
5030 if (VisitedInsts.insert(UI)) {
5031 if (isa<PHINode>(UI))
5034 NonPHIs.push_back(UI);
5035 } else if (!isa<PHINode>(UI) &&
5036 ((!isa<FCmpInst>(UI) &&
5037 !isa<ICmpInst>(UI) &&
5038 !isa<SelectInst>(UI)) ||
5039 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5042 // Remember that we completed the cycle.
5044 FoundStartPHI = true;
5046 Worklist.append(PHIs.begin(), PHIs.end());
5047 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5050 // This means we have seen one but not the other instruction of the
5051 // pattern or more than just a select and cmp.
5052 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5053 NumCmpSelectPatternInst != 2)
5056 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5059 // We found a reduction var if we have reached the original phi node and we
5060 // only have a single instruction with out-of-loop users.
5062 // This instruction is allowed to have out-of-loop users.
5063 AllowedExit.insert(ExitInstruction);
5065 // Save the description of this reduction variable.
5066 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5067 ReduxDesc.MinMaxKind);
5068 Reductions[Phi] = RD;
5069 // We've ended the cycle. This is a reduction variable if we have an
5070 // outside user and it has a binary op.
5075 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5076 /// pattern corresponding to a min(X, Y) or max(X, Y).
5077 LoopVectorizationLegality::ReductionInstDesc
5078 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5079 ReductionInstDesc &Prev) {
5081 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5082 "Expect a select instruction");
5083 Instruction *Cmp = nullptr;
5084 SelectInst *Select = nullptr;
5086 // We must handle the select(cmp()) as a single instruction. Advance to the
5088 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5089 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5090 return ReductionInstDesc(false, I);
5091 return ReductionInstDesc(Select, Prev.MinMaxKind);
5094 // Only handle single use cases for now.
5095 if (!(Select = dyn_cast<SelectInst>(I)))
5096 return ReductionInstDesc(false, I);
5097 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5098 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5099 return ReductionInstDesc(false, I);
5100 if (!Cmp->hasOneUse())
5101 return ReductionInstDesc(false, I);
5106 // Look for a min/max pattern.
5107 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5108 return ReductionInstDesc(Select, MRK_UIntMin);
5109 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5110 return ReductionInstDesc(Select, MRK_UIntMax);
5111 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5112 return ReductionInstDesc(Select, MRK_SIntMax);
5113 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5114 return ReductionInstDesc(Select, MRK_SIntMin);
5115 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5116 return ReductionInstDesc(Select, MRK_FloatMin);
5117 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5118 return ReductionInstDesc(Select, MRK_FloatMax);
5119 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5120 return ReductionInstDesc(Select, MRK_FloatMin);
5121 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5122 return ReductionInstDesc(Select, MRK_FloatMax);
5124 return ReductionInstDesc(false, I);
5127 LoopVectorizationLegality::ReductionInstDesc
5128 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5130 ReductionInstDesc &Prev) {
5131 bool FP = I->getType()->isFloatingPointTy();
5132 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
5133 switch (I->getOpcode()) {
5135 return ReductionInstDesc(false, I);
5136 case Instruction::PHI:
5137 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5138 Kind != RK_FloatMinMax))
5139 return ReductionInstDesc(false, I);
5140 return ReductionInstDesc(I, Prev.MinMaxKind);
5141 case Instruction::Sub:
5142 case Instruction::Add:
5143 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5144 case Instruction::Mul:
5145 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5146 case Instruction::And:
5147 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5148 case Instruction::Or:
5149 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5150 case Instruction::Xor:
5151 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5152 case Instruction::FMul:
5153 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5154 case Instruction::FAdd:
5155 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5156 case Instruction::FCmp:
5157 case Instruction::ICmp:
5158 case Instruction::Select:
5159 if (Kind != RK_IntegerMinMax &&
5160 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5161 return ReductionInstDesc(false, I);
5162 return isMinMaxSelectCmpPattern(I, Prev);
5166 LoopVectorizationLegality::InductionKind
5167 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5168 Type *PhiTy = Phi->getType();
5169 // We only handle integer and pointer inductions variables.
5170 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5171 return IK_NoInduction;
5173 // Check that the PHI is consecutive.
5174 const SCEV *PhiScev = SE->getSCEV(Phi);
5175 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5177 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5178 return IK_NoInduction;
5180 const SCEV *Step = AR->getStepRecurrence(*SE);
5182 // Integer inductions need to have a stride of one.
5183 if (PhiTy->isIntegerTy()) {
5185 return IK_IntInduction;
5186 if (Step->isAllOnesValue())
5187 return IK_ReverseIntInduction;
5188 return IK_NoInduction;
5191 // Calculate the pointer stride and check if it is consecutive.
5192 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5194 return IK_NoInduction;
5196 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5197 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5198 if (C->getValue()->equalsInt(Size))
5199 return IK_PtrInduction;
5200 else if (C->getValue()->equalsInt(0 - Size))
5201 return IK_ReversePtrInduction;
5203 return IK_NoInduction;
5206 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5207 Value *In0 = const_cast<Value*>(V);
5208 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5212 return Inductions.count(PN);
5215 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5216 assert(TheLoop->contains(BB) && "Unknown block used");
5218 // Blocks that do not dominate the latch need predication.
5219 BasicBlock* Latch = TheLoop->getLoopLatch();
5220 return !DT->dominates(BB, Latch);
5223 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5224 SmallPtrSet<Value *, 8>& SafePtrs) {
5225 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5226 // We might be able to hoist the load.
5227 if (it->mayReadFromMemory()) {
5228 LoadInst *LI = dyn_cast<LoadInst>(it);
5229 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5233 // We don't predicate stores at the moment.
5234 if (it->mayWriteToMemory()) {
5235 StoreInst *SI = dyn_cast<StoreInst>(it);
5236 // We only support predication of stores in basic blocks with one
5238 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5239 !SafePtrs.count(SI->getPointerOperand()) ||
5240 !SI->getParent()->getSinglePredecessor())
5246 // Check that we don't have a constant expression that can trap as operand.
5247 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5249 if (Constant *C = dyn_cast<Constant>(*OI))
5254 // The instructions below can trap.
5255 switch (it->getOpcode()) {
5257 case Instruction::UDiv:
5258 case Instruction::SDiv:
5259 case Instruction::URem:
5260 case Instruction::SRem:
5268 LoopVectorizationCostModel::VectorizationFactor
5269 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5271 bool ForceVectorization) {
5272 // Width 1 means no vectorize
5273 VectorizationFactor Factor = { 1U, 0U };
5274 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5275 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5279 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5280 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5284 // Find the trip count.
5285 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5286 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5288 unsigned WidestType = getWidestType();
5289 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5290 unsigned MaxSafeDepDist = -1U;
5291 if (Legal->getMaxSafeDepDistBytes() != -1U)
5292 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5293 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5294 WidestRegister : MaxSafeDepDist);
5295 unsigned MaxVectorSize = WidestRegister / WidestType;
5296 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5297 DEBUG(dbgs() << "LV: The Widest register is: "
5298 << WidestRegister << " bits.\n");
5300 if (MaxVectorSize == 0) {
5301 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5305 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5306 " into one vector!");
5308 unsigned VF = MaxVectorSize;
5310 // If we optimize the program for size, avoid creating the tail loop.
5312 // If we are unable to calculate the trip count then don't try to vectorize.
5314 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5318 // Find the maximum SIMD width that can fit within the trip count.
5319 VF = TC % MaxVectorSize;
5324 // If the trip count that we found modulo the vectorization factor is not
5325 // zero then we require a tail.
5327 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5333 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5334 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5336 Factor.Width = UserVF;
5340 float Cost = expectedCost(1);
5342 const float ScalarCost = Cost;
5345 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5347 // Ignore scalar width, because the user explicitly wants vectorization.
5348 if (ForceVectorization && VF > 1) {
5350 Cost = expectedCost(Width) / (float)Width;
5353 for (unsigned i=2; i <= VF; i*=2) {
5354 // Notice that the vector loop needs to be executed less times, so
5355 // we need to divide the cost of the vector loops by the width of
5356 // the vector elements.
5357 float VectorCost = expectedCost(i) / (float)i;
5358 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5359 (int)VectorCost << ".\n");
5360 if (VectorCost < Cost) {
5366 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5367 << "LV: Vectorization seems to be not beneficial, "
5368 << "but was forced by a user.\n");
5369 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5370 Factor.Width = Width;
5371 Factor.Cost = Width * Cost;
5375 unsigned LoopVectorizationCostModel::getWidestType() {
5376 unsigned MaxWidth = 8;
5379 for (Loop::block_iterator bb = TheLoop->block_begin(),
5380 be = TheLoop->block_end(); bb != be; ++bb) {
5381 BasicBlock *BB = *bb;
5383 // For each instruction in the loop.
5384 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5385 Type *T = it->getType();
5387 // Only examine Loads, Stores and PHINodes.
5388 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5391 // Examine PHI nodes that are reduction variables.
5392 if (PHINode *PN = dyn_cast<PHINode>(it))
5393 if (!Legal->getReductionVars()->count(PN))
5396 // Examine the stored values.
5397 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5398 T = ST->getValueOperand()->getType();
5400 // Ignore loaded pointer types and stored pointer types that are not
5401 // consecutive. However, we do want to take consecutive stores/loads of
5402 // pointer vectors into account.
5403 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5406 MaxWidth = std::max(MaxWidth,
5407 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5415 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5418 unsigned LoopCost) {
5420 // -- The unroll heuristics --
5421 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5422 // There are many micro-architectural considerations that we can't predict
5423 // at this level. For example frontend pressure (on decode or fetch) due to
5424 // code size, or the number and capabilities of the execution ports.
5426 // We use the following heuristics to select the unroll factor:
5427 // 1. If the code has reductions the we unroll in order to break the cross
5428 // iteration dependency.
5429 // 2. If the loop is really small then we unroll in order to reduce the loop
5431 // 3. We don't unroll if we think that we will spill registers to memory due
5432 // to the increased register pressure.
5434 // Use the user preference, unless 'auto' is selected.
5438 // When we optimize for size we don't unroll.
5442 // We used the distance for the unroll factor.
5443 if (Legal->getMaxSafeDepDistBytes() != -1U)
5446 // Do not unroll loops with a relatively small trip count.
5447 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5448 TheLoop->getLoopLatch());
5449 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5452 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5453 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5457 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5458 TargetNumRegisters = ForceTargetNumScalarRegs;
5460 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5461 TargetNumRegisters = ForceTargetNumVectorRegs;
5464 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5465 // We divide by these constants so assume that we have at least one
5466 // instruction that uses at least one register.
5467 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5468 R.NumInstructions = std::max(R.NumInstructions, 1U);
5470 // We calculate the unroll factor using the following formula.
5471 // Subtract the number of loop invariants from the number of available
5472 // registers. These registers are used by all of the unrolled instances.
5473 // Next, divide the remaining registers by the number of registers that is
5474 // required by the loop, in order to estimate how many parallel instances
5475 // fit without causing spills. All of this is rounded down if necessary to be
5476 // a power of two. We want power of two unroll factors to simplify any
5477 // addressing operations or alignment considerations.
5478 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5481 // Don't count the induction variable as unrolled.
5482 if (EnableIndVarRegisterHeur)
5483 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5484 std::max(1U, (R.MaxLocalUsers - 1)));
5486 // Clamp the unroll factor ranges to reasonable factors.
5487 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5489 // Check if the user has overridden the unroll max.
5491 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5492 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5494 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5495 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5498 // If we did not calculate the cost for VF (because the user selected the VF)
5499 // then we calculate the cost of VF here.
5501 LoopCost = expectedCost(VF);
5503 // Clamp the calculated UF to be between the 1 and the max unroll factor
5504 // that the target allows.
5505 if (UF > MaxUnrollSize)
5510 // Unroll if we vectorized this loop and there is a reduction that could
5511 // benefit from unrolling.
5512 if (VF > 1 && Legal->getReductionVars()->size()) {
5513 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5517 // Note that if we've already vectorized the loop we will have done the
5518 // runtime check and so unrolling won't require further checks.
5519 bool UnrollingRequiresRuntimePointerCheck =
5520 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5522 // We want to unroll small loops in order to reduce the loop overhead and
5523 // potentially expose ILP opportunities.
5524 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5525 if (!UnrollingRequiresRuntimePointerCheck &&
5526 LoopCost < SmallLoopCost) {
5527 // We assume that the cost overhead is 1 and we use the cost model
5528 // to estimate the cost of the loop and unroll until the cost of the
5529 // loop overhead is about 5% of the cost of the loop.
5530 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5532 // Unroll until store/load ports (estimated by max unroll factor) are
5534 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5535 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5537 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5538 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5539 return std::max(StoresUF, LoadsUF);
5542 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5546 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5550 LoopVectorizationCostModel::RegisterUsage
5551 LoopVectorizationCostModel::calculateRegisterUsage() {
5552 // This function calculates the register usage by measuring the highest number
5553 // of values that are alive at a single location. Obviously, this is a very
5554 // rough estimation. We scan the loop in a topological order in order and
5555 // assign a number to each instruction. We use RPO to ensure that defs are
5556 // met before their users. We assume that each instruction that has in-loop
5557 // users starts an interval. We record every time that an in-loop value is
5558 // used, so we have a list of the first and last occurrences of each
5559 // instruction. Next, we transpose this data structure into a multi map that
5560 // holds the list of intervals that *end* at a specific location. This multi
5561 // map allows us to perform a linear search. We scan the instructions linearly
5562 // and record each time that a new interval starts, by placing it in a set.
5563 // If we find this value in the multi-map then we remove it from the set.
5564 // The max register usage is the maximum size of the set.
5565 // We also search for instructions that are defined outside the loop, but are
5566 // used inside the loop. We need this number separately from the max-interval
5567 // usage number because when we unroll, loop-invariant values do not take
5569 LoopBlocksDFS DFS(TheLoop);
5573 R.NumInstructions = 0;
5575 // Each 'key' in the map opens a new interval. The values
5576 // of the map are the index of the 'last seen' usage of the
5577 // instruction that is the key.
5578 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5579 // Maps instruction to its index.
5580 DenseMap<unsigned, Instruction*> IdxToInstr;
5581 // Marks the end of each interval.
5582 IntervalMap EndPoint;
5583 // Saves the list of instruction indices that are used in the loop.
5584 SmallSet<Instruction*, 8> Ends;
5585 // Saves the list of values that are used in the loop but are
5586 // defined outside the loop, such as arguments and constants.
5587 SmallPtrSet<Value*, 8> LoopInvariants;
5590 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5591 be = DFS.endRPO(); bb != be; ++bb) {
5592 R.NumInstructions += (*bb)->size();
5593 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5595 Instruction *I = it;
5596 IdxToInstr[Index++] = I;
5598 // Save the end location of each USE.
5599 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5600 Value *U = I->getOperand(i);
5601 Instruction *Instr = dyn_cast<Instruction>(U);
5603 // Ignore non-instruction values such as arguments, constants, etc.
5604 if (!Instr) continue;
5606 // If this instruction is outside the loop then record it and continue.
5607 if (!TheLoop->contains(Instr)) {
5608 LoopInvariants.insert(Instr);
5612 // Overwrite previous end points.
5613 EndPoint[Instr] = Index;
5619 // Saves the list of intervals that end with the index in 'key'.
5620 typedef SmallVector<Instruction*, 2> InstrList;
5621 DenseMap<unsigned, InstrList> TransposeEnds;
5623 // Transpose the EndPoints to a list of values that end at each index.
5624 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5626 TransposeEnds[it->second].push_back(it->first);
5628 SmallSet<Instruction*, 8> OpenIntervals;
5629 unsigned MaxUsage = 0;
5632 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5633 for (unsigned int i = 0; i < Index; ++i) {
5634 Instruction *I = IdxToInstr[i];
5635 // Ignore instructions that are never used within the loop.
5636 if (!Ends.count(I)) continue;
5638 // Remove all of the instructions that end at this location.
5639 InstrList &List = TransposeEnds[i];
5640 for (unsigned int j=0, e = List.size(); j < e; ++j)
5641 OpenIntervals.erase(List[j]);
5643 // Count the number of live interals.
5644 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5646 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5647 OpenIntervals.size() << '\n');
5649 // Add the current instruction to the list of open intervals.
5650 OpenIntervals.insert(I);
5653 unsigned Invariant = LoopInvariants.size();
5654 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5655 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5656 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5658 R.LoopInvariantRegs = Invariant;
5659 R.MaxLocalUsers = MaxUsage;
5663 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5667 for (Loop::block_iterator bb = TheLoop->block_begin(),
5668 be = TheLoop->block_end(); bb != be; ++bb) {
5669 unsigned BlockCost = 0;
5670 BasicBlock *BB = *bb;
5672 // For each instruction in the old loop.
5673 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5674 // Skip dbg intrinsics.
5675 if (isa<DbgInfoIntrinsic>(it))
5678 unsigned C = getInstructionCost(it, VF);
5680 // Check if we should override the cost.
5681 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5682 C = ForceTargetInstructionCost;
5685 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5686 VF << " For instruction: " << *it << '\n');
5689 // We assume that if-converted blocks have a 50% chance of being executed.
5690 // When the code is scalar then some of the blocks are avoided due to CF.
5691 // When the code is vectorized we execute all code paths.
5692 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5701 /// \brief Check whether the address computation for a non-consecutive memory
5702 /// access looks like an unlikely candidate for being merged into the indexing
5705 /// We look for a GEP which has one index that is an induction variable and all
5706 /// other indices are loop invariant. If the stride of this access is also
5707 /// within a small bound we decide that this address computation can likely be
5708 /// merged into the addressing mode.
5709 /// In all other cases, we identify the address computation as complex.
5710 static bool isLikelyComplexAddressComputation(Value *Ptr,
5711 LoopVectorizationLegality *Legal,
5712 ScalarEvolution *SE,
5713 const Loop *TheLoop) {
5714 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5718 // We are looking for a gep with all loop invariant indices except for one
5719 // which should be an induction variable.
5720 unsigned NumOperands = Gep->getNumOperands();
5721 for (unsigned i = 1; i < NumOperands; ++i) {
5722 Value *Opd = Gep->getOperand(i);
5723 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5724 !Legal->isInductionVariable(Opd))
5728 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5729 // can likely be merged into the address computation.
5730 unsigned MaxMergeDistance = 64;
5732 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5736 // Check the step is constant.
5737 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5738 // Calculate the pointer stride and check if it is consecutive.
5739 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5743 const APInt &APStepVal = C->getValue()->getValue();
5745 // Huge step value - give up.
5746 if (APStepVal.getBitWidth() > 64)
5749 int64_t StepVal = APStepVal.getSExtValue();
5751 return StepVal > MaxMergeDistance;
5754 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5755 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5761 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5762 // If we know that this instruction will remain uniform, check the cost of
5763 // the scalar version.
5764 if (Legal->isUniformAfterVectorization(I))
5767 Type *RetTy = I->getType();
5768 Type *VectorTy = ToVectorTy(RetTy, VF);
5770 // TODO: We need to estimate the cost of intrinsic calls.
5771 switch (I->getOpcode()) {
5772 case Instruction::GetElementPtr:
5773 // We mark this instruction as zero-cost because the cost of GEPs in
5774 // vectorized code depends on whether the corresponding memory instruction
5775 // is scalarized or not. Therefore, we handle GEPs with the memory
5776 // instruction cost.
5778 case Instruction::Br: {
5779 return TTI.getCFInstrCost(I->getOpcode());
5781 case Instruction::PHI:
5782 //TODO: IF-converted IFs become selects.
5784 case Instruction::Add:
5785 case Instruction::FAdd:
5786 case Instruction::Sub:
5787 case Instruction::FSub:
5788 case Instruction::Mul:
5789 case Instruction::FMul:
5790 case Instruction::UDiv:
5791 case Instruction::SDiv:
5792 case Instruction::FDiv:
5793 case Instruction::URem:
5794 case Instruction::SRem:
5795 case Instruction::FRem:
5796 case Instruction::Shl:
5797 case Instruction::LShr:
5798 case Instruction::AShr:
5799 case Instruction::And:
5800 case Instruction::Or:
5801 case Instruction::Xor: {
5802 // Since we will replace the stride by 1 the multiplication should go away.
5803 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5805 // Certain instructions can be cheaper to vectorize if they have a constant
5806 // second vector operand. One example of this are shifts on x86.
5807 TargetTransformInfo::OperandValueKind Op1VK =
5808 TargetTransformInfo::OK_AnyValue;
5809 TargetTransformInfo::OperandValueKind Op2VK =
5810 TargetTransformInfo::OK_AnyValue;
5811 Value *Op2 = I->getOperand(1);
5813 // Check for a splat of a constant or for a non uniform vector of constants.
5814 if (isa<ConstantInt>(Op2))
5815 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5816 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5817 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5818 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5819 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5822 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5824 case Instruction::Select: {
5825 SelectInst *SI = cast<SelectInst>(I);
5826 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5827 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5828 Type *CondTy = SI->getCondition()->getType();
5830 CondTy = VectorType::get(CondTy, VF);
5832 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5834 case Instruction::ICmp:
5835 case Instruction::FCmp: {
5836 Type *ValTy = I->getOperand(0)->getType();
5837 VectorTy = ToVectorTy(ValTy, VF);
5838 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5840 case Instruction::Store:
5841 case Instruction::Load: {
5842 StoreInst *SI = dyn_cast<StoreInst>(I);
5843 LoadInst *LI = dyn_cast<LoadInst>(I);
5844 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5846 VectorTy = ToVectorTy(ValTy, VF);
5848 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5849 unsigned AS = SI ? SI->getPointerAddressSpace() :
5850 LI->getPointerAddressSpace();
5851 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5852 // We add the cost of address computation here instead of with the gep
5853 // instruction because only here we know whether the operation is
5856 return TTI.getAddressComputationCost(VectorTy) +
5857 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5859 // Scalarized loads/stores.
5860 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5861 bool Reverse = ConsecutiveStride < 0;
5862 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5863 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5864 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5865 bool IsComplexComputation =
5866 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5868 // The cost of extracting from the value vector and pointer vector.
5869 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5870 for (unsigned i = 0; i < VF; ++i) {
5871 // The cost of extracting the pointer operand.
5872 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5873 // In case of STORE, the cost of ExtractElement from the vector.
5874 // In case of LOAD, the cost of InsertElement into the returned
5876 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5877 Instruction::InsertElement,
5881 // The cost of the scalar loads/stores.
5882 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5883 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5888 // Wide load/stores.
5889 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5890 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5893 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5897 case Instruction::ZExt:
5898 case Instruction::SExt:
5899 case Instruction::FPToUI:
5900 case Instruction::FPToSI:
5901 case Instruction::FPExt:
5902 case Instruction::PtrToInt:
5903 case Instruction::IntToPtr:
5904 case Instruction::SIToFP:
5905 case Instruction::UIToFP:
5906 case Instruction::Trunc:
5907 case Instruction::FPTrunc:
5908 case Instruction::BitCast: {
5909 // We optimize the truncation of induction variable.
5910 // The cost of these is the same as the scalar operation.
5911 if (I->getOpcode() == Instruction::Trunc &&
5912 Legal->isInductionVariable(I->getOperand(0)))
5913 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5914 I->getOperand(0)->getType());
5916 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5917 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5919 case Instruction::Call: {
5920 CallInst *CI = cast<CallInst>(I);
5921 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5922 assert(ID && "Not an intrinsic call!");
5923 Type *RetTy = ToVectorTy(CI->getType(), VF);
5924 SmallVector<Type*, 4> Tys;
5925 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5926 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5927 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5930 // We are scalarizing the instruction. Return the cost of the scalar
5931 // instruction, plus the cost of insert and extract into vector
5932 // elements, times the vector width.
5935 if (!RetTy->isVoidTy() && VF != 1) {
5936 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5938 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5941 // The cost of inserting the results plus extracting each one of the
5943 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5946 // The cost of executing VF copies of the scalar instruction. This opcode
5947 // is unknown. Assume that it is the same as 'mul'.
5948 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5954 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5955 if (Scalar->isVoidTy() || VF == 1)
5957 return VectorType::get(Scalar, VF);
5960 char LoopVectorize::ID = 0;
5961 static const char lv_name[] = "Loop Vectorization";
5962 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5963 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5964 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5965 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5966 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5967 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5968 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5969 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5970 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5971 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5974 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5975 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5979 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5980 // Check for a store.
5981 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5982 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5984 // Check for a load.
5985 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5986 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5992 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5993 bool IfPredicateStore) {
5994 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5995 // Holds vector parameters or scalars, in case of uniform vals.
5996 SmallVector<VectorParts, 4> Params;
5998 setDebugLocFromInst(Builder, Instr);
6000 // Find all of the vectorized parameters.
6001 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6002 Value *SrcOp = Instr->getOperand(op);
6004 // If we are accessing the old induction variable, use the new one.
6005 if (SrcOp == OldInduction) {
6006 Params.push_back(getVectorValue(SrcOp));
6010 // Try using previously calculated values.
6011 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6013 // If the src is an instruction that appeared earlier in the basic block
6014 // then it should already be vectorized.
6015 if (SrcInst && OrigLoop->contains(SrcInst)) {
6016 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6017 // The parameter is a vector value from earlier.
6018 Params.push_back(WidenMap.get(SrcInst));
6020 // The parameter is a scalar from outside the loop. Maybe even a constant.
6021 VectorParts Scalars;
6022 Scalars.append(UF, SrcOp);
6023 Params.push_back(Scalars);
6027 assert(Params.size() == Instr->getNumOperands() &&
6028 "Invalid number of operands");
6030 // Does this instruction return a value ?
6031 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6033 Value *UndefVec = IsVoidRetTy ? nullptr :
6034 UndefValue::get(Instr->getType());
6035 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6036 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6038 Instruction *InsertPt = Builder.GetInsertPoint();
6039 BasicBlock *IfBlock = Builder.GetInsertBlock();
6040 BasicBlock *CondBlock = nullptr;
6043 Loop *VectorLp = nullptr;
6044 if (IfPredicateStore) {
6045 assert(Instr->getParent()->getSinglePredecessor() &&
6046 "Only support single predecessor blocks");
6047 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6048 Instr->getParent());
6049 VectorLp = LI->getLoopFor(IfBlock);
6050 assert(VectorLp && "Must have a loop for this block");
6053 // For each vector unroll 'part':
6054 for (unsigned Part = 0; Part < UF; ++Part) {
6055 // For each scalar that we create:
6057 // Start an "if (pred) a[i] = ..." block.
6058 Value *Cmp = nullptr;
6059 if (IfPredicateStore) {
6060 if (Cond[Part]->getType()->isVectorTy())
6062 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6063 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6064 ConstantInt::get(Cond[Part]->getType(), 1));
6065 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6066 LoopVectorBody.push_back(CondBlock);
6067 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6068 // Update Builder with newly created basic block.
6069 Builder.SetInsertPoint(InsertPt);
6072 Instruction *Cloned = Instr->clone();
6074 Cloned->setName(Instr->getName() + ".cloned");
6075 // Replace the operands of the cloned instructions with extracted scalars.
6076 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6077 Value *Op = Params[op][Part];
6078 Cloned->setOperand(op, Op);
6081 // Place the cloned scalar in the new loop.
6082 Builder.Insert(Cloned);
6084 // If the original scalar returns a value we need to place it in a vector
6085 // so that future users will be able to use it.
6087 VecResults[Part] = Cloned;
6090 if (IfPredicateStore) {
6091 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6092 LoopVectorBody.push_back(NewIfBlock);
6093 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6094 Builder.SetInsertPoint(InsertPt);
6095 Instruction *OldBr = IfBlock->getTerminator();
6096 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6097 OldBr->eraseFromParent();
6098 IfBlock = NewIfBlock;
6103 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6104 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6105 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6107 return scalarizeInstruction(Instr, IfPredicateStore);
6110 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6114 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6118 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6120 // When unrolling and the VF is 1, we only need to add a simple scalar.
6121 Type *ITy = Val->getType();
6122 assert(!ITy->isVectorTy() && "Val must be a scalar");
6123 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6124 return Builder.CreateAdd(Val, C, "induction");