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/BlockFrequencyInfo.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopIterator.h"
60 #include "llvm/Analysis/LoopPass.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpander.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/TargetTransformInfo.h"
65 #include "llvm/Analysis/ValueTracking.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DataLayout.h"
68 #include "llvm/IR/DebugInfo.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/DiagnosticInfo.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/PatternMatch.h"
79 #include "llvm/IR/Type.h"
80 #include "llvm/IR/Value.h"
81 #include "llvm/IR/ValueHandle.h"
82 #include "llvm/IR/Verifier.h"
83 #include "llvm/Pass.h"
84 #include "llvm/Support/BranchProbability.h"
85 #include "llvm/Support/CommandLine.h"
86 #include "llvm/Support/Debug.h"
87 #include "llvm/Support/raw_ostream.h"
88 #include "llvm/Transforms/Scalar.h"
89 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
90 #include "llvm/Transforms/Utils/Local.h"
91 #include "llvm/Transforms/Utils/VectorUtils.h"
97 using namespace llvm::PatternMatch;
99 #define LV_NAME "loop-vectorize"
100 #define DEBUG_TYPE LV_NAME
102 STATISTIC(LoopsVectorized, "Number of loops vectorized");
103 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
105 static cl::opt<unsigned>
106 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
107 cl::desc("Sets the SIMD width. Zero is autoselect."));
109 static cl::opt<unsigned>
110 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
111 cl::desc("Sets the vectorization unroll count. "
112 "Zero is autoselect."));
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 /// We don't unroll loops with a known constant trip count below this number.
142 static const unsigned TinyTripCountUnrollThreshold = 128;
144 /// When performing memory disambiguation checks at runtime do not make more
145 /// than this number of comparisons.
146 static const unsigned RuntimeMemoryCheckThreshold = 8;
148 /// Maximum simd width.
149 static const unsigned MaxVectorWidth = 64;
151 static cl::opt<unsigned> ForceTargetNumScalarRegs(
152 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
153 cl::desc("A flag that overrides the target's number of scalar registers."));
155 static cl::opt<unsigned> ForceTargetNumVectorRegs(
156 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of vector registers."));
159 /// Maximum vectorization unroll count.
160 static const unsigned MaxUnrollFactor = 16;
162 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
163 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
164 cl::desc("A flag that overrides the target's max unroll factor for scalar "
167 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
168 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
169 cl::desc("A flag that overrides the target's max unroll factor for "
170 "vectorized loops."));
172 static cl::opt<unsigned> ForceTargetInstructionCost(
173 "force-target-instruction-cost", cl::init(0), cl::Hidden,
174 cl::desc("A flag that overrides the target's expected cost for "
175 "an instruction to a single constant value. Mostly "
176 "useful for getting consistent testing."));
178 static cl::opt<unsigned> SmallLoopCost(
179 "small-loop-cost", cl::init(20), cl::Hidden,
180 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
182 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
183 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
184 cl::desc("Enable the use of the block frequency analysis to access PGO "
185 "heuristics minimizing code growth in cold regions and being more "
186 "aggressive in hot regions."));
188 // Runtime unroll loops for load/store throughput.
189 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
190 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
191 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
193 /// The number of stores in a loop that are allowed to need predication.
194 static cl::opt<unsigned> NumberOfStoresToPredicate(
195 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
196 cl::desc("Max number of stores to be predicated behind an if."));
198 static cl::opt<bool> EnableIndVarRegisterHeur(
199 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
200 cl::desc("Count the induction variable only once when unrolling"));
202 static cl::opt<bool> EnableCondStoresVectorization(
203 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
204 cl::desc("Enable if predication of stores during vectorization."));
208 // Forward declarations.
209 class LoopVectorizationLegality;
210 class LoopVectorizationCostModel;
212 /// Optimization analysis message produced during vectorization. Messages inform
213 /// the user why vectorization did not occur.
215 string_ostream Message;
219 Report(Instruction *I = nullptr) : Instr(I) {
220 Message << "loop not vectorized: ";
223 template <typename A> Report &operator<<(const A &Value) {
228 Instruction *getInstr() { return Instr; }
230 StringRef str() { return Message.str(); }
231 operator Twine() { return Message.str(); }
234 /// InnerLoopVectorizer vectorizes loops which contain only one basic
235 /// block to a specified vectorization factor (VF).
236 /// This class performs the widening of scalars into vectors, or multiple
237 /// scalars. This class also implements the following features:
238 /// * It inserts an epilogue loop for handling loops that don't have iteration
239 /// counts that are known to be a multiple of the vectorization factor.
240 /// * It handles the code generation for reduction variables.
241 /// * Scalarization (implementation using scalars) of un-vectorizable
243 /// InnerLoopVectorizer does not perform any vectorization-legality
244 /// checks, and relies on the caller to check for the different legality
245 /// aspects. The InnerLoopVectorizer relies on the
246 /// LoopVectorizationLegality class to provide information about the induction
247 /// and reduction variables that were found to a given vectorization factor.
248 class InnerLoopVectorizer {
250 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
251 DominatorTree *DT, const DataLayout *DL,
252 const TargetLibraryInfo *TLI, unsigned VecWidth,
253 unsigned UnrollFactor)
254 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
255 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
256 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
259 // Perform the actual loop widening (vectorization).
260 void vectorize(LoopVectorizationLegality *L) {
262 // Create a new empty loop. Unlink the old loop and connect the new one.
264 // Widen each instruction in the old loop to a new one in the new loop.
265 // Use the Legality module to find the induction and reduction variables.
267 // Register the new loop and update the analysis passes.
271 virtual ~InnerLoopVectorizer() {}
274 /// A small list of PHINodes.
275 typedef SmallVector<PHINode*, 4> PhiVector;
276 /// When we unroll loops we have multiple vector values for each scalar.
277 /// This data structure holds the unrolled and vectorized values that
278 /// originated from one scalar instruction.
279 typedef SmallVector<Value*, 2> VectorParts;
281 // When we if-convert we need create edge masks. We have to cache values so
282 // that we don't end up with exponential recursion/IR.
283 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
284 VectorParts> EdgeMaskCache;
286 /// \brief Add code that checks at runtime if the accessed arrays overlap.
288 /// Returns a pair of instructions where the first element is the first
289 /// instruction generated in possibly a sequence of instructions and the
290 /// second value is the final comparator value or NULL if no check is needed.
291 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
293 /// \brief Add checks for strides that where assumed to be 1.
295 /// Returns the last check instruction and the first check instruction in the
296 /// pair as (first, last).
297 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
299 /// Create an empty loop, based on the loop ranges of the old loop.
300 void createEmptyLoop();
301 /// Copy and widen the instructions from the old loop.
302 virtual void vectorizeLoop();
304 /// \brief The Loop exit block may have single value PHI nodes where the
305 /// incoming value is 'Undef'. While vectorizing we only handled real values
306 /// that were defined inside the loop. Here we fix the 'undef case'.
310 /// A helper function that computes the predicate of the block BB, assuming
311 /// that the header block of the loop is set to True. It returns the *entry*
312 /// mask for the block BB.
313 VectorParts createBlockInMask(BasicBlock *BB);
314 /// A helper function that computes the predicate of the edge between SRC
316 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
318 /// A helper function to vectorize a single BB within the innermost loop.
319 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
321 /// Vectorize a single PHINode in a block. This method handles the induction
322 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
323 /// arbitrary length vectors.
324 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
325 unsigned UF, unsigned VF, PhiVector *PV);
327 /// Insert the new loop to the loop hierarchy and pass manager
328 /// and update the analysis passes.
329 void updateAnalysis();
331 /// This instruction is un-vectorizable. Implement it as a sequence
332 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
333 /// scalarized instruction behind an if block predicated on the control
334 /// dependence of the instruction.
335 virtual void scalarizeInstruction(Instruction *Instr,
336 bool IfPredicateStore=false);
338 /// Vectorize Load and Store instructions,
339 virtual void vectorizeMemoryInstruction(Instruction *Instr);
341 /// Create a broadcast instruction. This method generates a broadcast
342 /// instruction (shuffle) for loop invariant values and for the induction
343 /// value. If this is the induction variable then we extend it to N, N+1, ...
344 /// this is needed because each iteration in the loop corresponds to a SIMD
346 virtual Value *getBroadcastInstrs(Value *V);
348 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
349 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
350 /// The sequence starts at StartIndex.
351 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
353 /// When we go over instructions in the basic block we rely on previous
354 /// values within the current basic block or on loop invariant values.
355 /// When we widen (vectorize) values we place them in the map. If the values
356 /// are not within the map, they have to be loop invariant, so we simply
357 /// broadcast them into a vector.
358 VectorParts &getVectorValue(Value *V);
360 /// Generate a shuffle sequence that will reverse the vector Vec.
361 virtual Value *reverseVector(Value *Vec);
363 /// This is a helper class that holds the vectorizer state. It maps scalar
364 /// instructions to vector instructions. When the code is 'unrolled' then
365 /// then a single scalar value is mapped to multiple vector parts. The parts
366 /// are stored in the VectorPart type.
368 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
370 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
372 /// \return True if 'Key' is saved in the Value Map.
373 bool has(Value *Key) const { return MapStorage.count(Key); }
375 /// Initializes a new entry in the map. Sets all of the vector parts to the
376 /// save value in 'Val'.
377 /// \return A reference to a vector with splat values.
378 VectorParts &splat(Value *Key, Value *Val) {
379 VectorParts &Entry = MapStorage[Key];
380 Entry.assign(UF, Val);
384 ///\return A reference to the value that is stored at 'Key'.
385 VectorParts &get(Value *Key) {
386 VectorParts &Entry = MapStorage[Key];
389 assert(Entry.size() == UF);
394 /// The unroll factor. Each entry in the map stores this number of vector
398 /// Map storage. We use std::map and not DenseMap because insertions to a
399 /// dense map invalidates its iterators.
400 std::map<Value *, VectorParts> MapStorage;
403 /// The original loop.
405 /// Scev analysis to use.
412 const DataLayout *DL;
413 /// Target Library Info.
414 const TargetLibraryInfo *TLI;
416 /// The vectorization SIMD factor to use. Each vector will have this many
421 /// The vectorization unroll factor to use. Each scalar is vectorized to this
422 /// many different vector instructions.
425 /// The builder that we use
428 // --- Vectorization state ---
430 /// The vector-loop preheader.
431 BasicBlock *LoopVectorPreHeader;
432 /// The scalar-loop preheader.
433 BasicBlock *LoopScalarPreHeader;
434 /// Middle Block between the vector and the scalar.
435 BasicBlock *LoopMiddleBlock;
436 ///The ExitBlock of the scalar loop.
437 BasicBlock *LoopExitBlock;
438 ///The vector loop body.
439 SmallVector<BasicBlock *, 4> LoopVectorBody;
440 ///The scalar loop body.
441 BasicBlock *LoopScalarBody;
442 /// A list of all bypass blocks. The first block is the entry of the loop.
443 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
445 /// The new Induction variable which was added to the new block.
447 /// The induction variable of the old basic block.
448 PHINode *OldInduction;
449 /// Holds the extended (to the widest induction type) start index.
451 /// Maps scalars to widened vectors.
453 EdgeMaskCache MaskCache;
455 LoopVectorizationLegality *Legal;
458 class InnerLoopUnroller : public InnerLoopVectorizer {
460 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
461 DominatorTree *DT, const DataLayout *DL,
462 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
463 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
466 void scalarizeInstruction(Instruction *Instr,
467 bool IfPredicateStore = false) override;
468 void vectorizeMemoryInstruction(Instruction *Instr) override;
469 Value *getBroadcastInstrs(Value *V) override;
470 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
471 Value *reverseVector(Value *Vec) override;
474 /// \brief Look for a meaningful debug location on the instruction or it's
476 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
481 if (I->getDebugLoc() != Empty)
484 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
485 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
486 if (OpInst->getDebugLoc() != Empty)
493 /// \brief Set the debug location in the builder using the debug location in the
495 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
496 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
497 B.SetCurrentDebugLocation(Inst->getDebugLoc());
499 B.SetCurrentDebugLocation(DebugLoc());
503 /// \return string containing a file name and a line # for the given loop.
504 static std::string getDebugLocString(const Loop *L) {
506 return std::string();
509 const DebugLoc LoopDbgLoc = L->getStartLoc();
510 if (!LoopDbgLoc.isUnknown())
511 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
513 // Just print the module name.
514 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
519 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
520 /// to what vectorization factor.
521 /// This class does not look at the profitability of vectorization, only the
522 /// legality. This class has two main kinds of checks:
523 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
524 /// will change the order of memory accesses in a way that will change the
525 /// correctness of the program.
526 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
527 /// checks for a number of different conditions, such as the availability of a
528 /// single induction variable, that all types are supported and vectorize-able,
529 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
530 /// This class is also used by InnerLoopVectorizer for identifying
531 /// induction variable and the different reduction variables.
532 class LoopVectorizationLegality {
536 unsigned NumPredStores;
538 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
539 DominatorTree *DT, TargetLibraryInfo *TLI,
541 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
542 DT(DT), TLI(TLI), TheFunction(F), Induction(nullptr),
543 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
546 /// This enum represents the kinds of reductions that we support.
548 RK_NoReduction, ///< Not a reduction.
549 RK_IntegerAdd, ///< Sum of integers.
550 RK_IntegerMult, ///< Product of integers.
551 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
552 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
553 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
554 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
555 RK_FloatAdd, ///< Sum of floats.
556 RK_FloatMult, ///< Product of floats.
557 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
560 /// This enum represents the kinds of inductions that we support.
562 IK_NoInduction, ///< Not an induction variable.
563 IK_IntInduction, ///< Integer induction variable. Step = 1.
564 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
565 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
566 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
569 // This enum represents the kind of minmax reduction.
570 enum MinMaxReductionKind {
580 /// This struct holds information about reduction variables.
581 struct ReductionDescriptor {
582 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
583 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
585 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
586 MinMaxReductionKind MK)
587 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
589 // The starting value of the reduction.
590 // It does not have to be zero!
591 TrackingVH<Value> StartValue;
592 // The instruction who's value is used outside the loop.
593 Instruction *LoopExitInstr;
594 // The kind of the reduction.
596 // If this a min/max reduction the kind of reduction.
597 MinMaxReductionKind MinMaxKind;
600 /// This POD struct holds information about a potential reduction operation.
601 struct ReductionInstDesc {
602 ReductionInstDesc(bool IsRedux, Instruction *I) :
603 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
605 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
606 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
608 // Is this instruction a reduction candidate.
610 // The last instruction in a min/max pattern (select of the select(icmp())
611 // pattern), or the current reduction instruction otherwise.
612 Instruction *PatternLastInst;
613 // If this is a min/max pattern the comparison predicate.
614 MinMaxReductionKind MinMaxKind;
617 /// This struct holds information about the memory runtime legality
618 /// check that a group of pointers do not overlap.
619 struct RuntimePointerCheck {
620 RuntimePointerCheck() : Need(false) {}
622 /// Reset the state of the pointer runtime information.
629 DependencySetId.clear();
632 /// Insert a pointer and calculate the start and end SCEVs.
633 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
634 unsigned DepSetId, ValueToValueMap &Strides);
636 /// This flag indicates if we need to add the runtime check.
638 /// Holds the pointers that we need to check.
639 SmallVector<TrackingVH<Value>, 2> Pointers;
640 /// Holds the pointer value at the beginning of the loop.
641 SmallVector<const SCEV*, 2> Starts;
642 /// Holds the pointer value at the end of the loop.
643 SmallVector<const SCEV*, 2> Ends;
644 /// Holds the information if this pointer is used for writing to memory.
645 SmallVector<bool, 2> IsWritePtr;
646 /// Holds the id of the set of pointers that could be dependent because of a
647 /// shared underlying object.
648 SmallVector<unsigned, 2> DependencySetId;
651 /// A struct for saving information about induction variables.
652 struct InductionInfo {
653 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
654 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
656 TrackingVH<Value> StartValue;
661 /// ReductionList contains the reduction descriptors for all
662 /// of the reductions that were found in the loop.
663 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
665 /// InductionList saves induction variables and maps them to the
666 /// induction descriptor.
667 typedef MapVector<PHINode*, InductionInfo> InductionList;
669 /// Returns true if it is legal to vectorize this loop.
670 /// This does not mean that it is profitable to vectorize this
671 /// loop, only that it is legal to do so.
674 /// Returns the Induction variable.
675 PHINode *getInduction() { return Induction; }
677 /// Returns the reduction variables found in the loop.
678 ReductionList *getReductionVars() { return &Reductions; }
680 /// Returns the induction variables found in the loop.
681 InductionList *getInductionVars() { return &Inductions; }
683 /// Returns the widest induction type.
684 Type *getWidestInductionType() { return WidestIndTy; }
686 /// Returns True if V is an induction variable in this loop.
687 bool isInductionVariable(const Value *V);
689 /// Return true if the block BB needs to be predicated in order for the loop
690 /// to be vectorized.
691 bool blockNeedsPredication(BasicBlock *BB);
693 /// Check if this pointer is consecutive when vectorizing. This happens
694 /// when the last index of the GEP is the induction variable, or that the
695 /// pointer itself is an induction variable.
696 /// This check allows us to vectorize A[idx] into a wide load/store.
698 /// 0 - Stride is unknown or non-consecutive.
699 /// 1 - Address is consecutive.
700 /// -1 - Address is consecutive, and decreasing.
701 int isConsecutivePtr(Value *Ptr);
703 /// Returns true if the value V is uniform within the loop.
704 bool isUniform(Value *V);
706 /// Returns true if this instruction will remain scalar after vectorization.
707 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
709 /// Returns the information that we collected about runtime memory check.
710 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
712 /// This function returns the identity element (or neutral element) for
714 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
716 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
718 bool hasStride(Value *V) { return StrideSet.count(V); }
719 bool mustCheckStrides() { return !StrideSet.empty(); }
720 SmallPtrSet<Value *, 8>::iterator strides_begin() {
721 return StrideSet.begin();
723 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
726 /// Check if a single basic block loop is vectorizable.
727 /// At this point we know that this is a loop with a constant trip count
728 /// and we only need to check individual instructions.
729 bool canVectorizeInstrs();
731 /// When we vectorize loops we may change the order in which
732 /// we read and write from memory. This method checks if it is
733 /// legal to vectorize the code, considering only memory constrains.
734 /// Returns true if the loop is vectorizable
735 bool canVectorizeMemory();
737 /// Return true if we can vectorize this loop using the IF-conversion
739 bool canVectorizeWithIfConvert();
741 /// Collect the variables that need to stay uniform after vectorization.
742 void collectLoopUniforms();
744 /// Return true if all of the instructions in the block can be speculatively
745 /// executed. \p SafePtrs is a list of addresses that are known to be legal
746 /// and we know that we can read from them without segfault.
747 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
749 /// Returns True, if 'Phi' is the kind of reduction variable for type
750 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
751 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
752 /// Returns a struct describing if the instruction 'I' can be a reduction
753 /// variable of type 'Kind'. If the reduction is a min/max pattern of
754 /// select(icmp()) this function advances the instruction pointer 'I' from the
755 /// compare instruction to the select instruction and stores this pointer in
756 /// 'PatternLastInst' member of the returned struct.
757 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
758 ReductionInstDesc &Desc);
759 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
760 /// pattern corresponding to a min(X, Y) or max(X, Y).
761 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
762 ReductionInstDesc &Prev);
763 /// Returns the induction kind of Phi. This function may return NoInduction
764 /// if the PHI is not an induction variable.
765 InductionKind isInductionVariable(PHINode *Phi);
767 /// \brief Collect memory access with loop invariant strides.
769 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
771 void collectStridedAcccess(Value *LoadOrStoreInst);
773 /// Report an analysis message to assist the user in diagnosing loops that are
775 void emitAnalysis(Report &Message) {
776 DebugLoc DL = TheLoop->getStartLoc();
777 if (Instruction *I = Message.getInstr())
778 DL = I->getDebugLoc();
779 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
780 *TheFunction, DL, Message.str());
783 /// The loop that we evaluate.
787 /// DataLayout analysis.
788 const DataLayout *DL;
791 /// Target Library Info.
792 TargetLibraryInfo *TLI;
794 Function *TheFunction;
796 // --- vectorization state --- //
798 /// Holds the integer induction variable. This is the counter of the
801 /// Holds the reduction variables.
802 ReductionList Reductions;
803 /// Holds all of the induction variables that we found in the loop.
804 /// Notice that inductions don't need to start at zero and that induction
805 /// variables can be pointers.
806 InductionList Inductions;
807 /// Holds the widest induction type encountered.
810 /// Allowed outside users. This holds the reduction
811 /// vars which can be accessed from outside the loop.
812 SmallPtrSet<Value*, 4> AllowedExit;
813 /// This set holds the variables which are known to be uniform after
815 SmallPtrSet<Instruction*, 4> Uniforms;
816 /// We need to check that all of the pointers in this list are disjoint
818 RuntimePointerCheck PtrRtCheck;
819 /// Can we assume the absence of NaNs.
820 bool HasFunNoNaNAttr;
822 unsigned MaxSafeDepDistBytes;
824 ValueToValueMap Strides;
825 SmallPtrSet<Value *, 8> StrideSet;
828 /// LoopVectorizationCostModel - estimates the expected speedups due to
830 /// In many cases vectorization is not profitable. This can happen because of
831 /// a number of reasons. In this class we mainly attempt to predict the
832 /// expected speedup/slowdowns due to the supported instruction set. We use the
833 /// TargetTransformInfo to query the different backends for the cost of
834 /// different operations.
835 class LoopVectorizationCostModel {
837 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
838 LoopVectorizationLegality *Legal,
839 const TargetTransformInfo &TTI,
840 const DataLayout *DL, const TargetLibraryInfo *TLI)
841 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
843 /// Information about vectorization costs
844 struct VectorizationFactor {
845 unsigned Width; // Vector width with best cost
846 unsigned Cost; // Cost of the loop with that width
848 /// \return The most profitable vectorization factor and the cost of that VF.
849 /// This method checks every power of two up to VF. If UserVF is not ZERO
850 /// then this vectorization factor will be selected if vectorization is
852 VectorizationFactor selectVectorizationFactor(bool OptForSize,
854 bool ForceVectorization);
856 /// \return The size (in bits) of the widest type in the code that
857 /// needs to be vectorized. We ignore values that remain scalar such as
858 /// 64 bit loop indices.
859 unsigned getWidestType();
861 /// \return The most profitable unroll factor.
862 /// If UserUF is non-zero then this method finds the best unroll-factor
863 /// based on register pressure and other parameters.
864 /// VF and LoopCost are the selected vectorization factor and the cost of the
866 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
869 /// \brief A struct that represents some properties of the register usage
871 struct RegisterUsage {
872 /// Holds the number of loop invariant values that are used in the loop.
873 unsigned LoopInvariantRegs;
874 /// Holds the maximum number of concurrent live intervals in the loop.
875 unsigned MaxLocalUsers;
876 /// Holds the number of instructions in the loop.
877 unsigned NumInstructions;
880 /// \return information about the register usage of the loop.
881 RegisterUsage calculateRegisterUsage();
884 /// Returns the expected execution cost. The unit of the cost does
885 /// not matter because we use the 'cost' units to compare different
886 /// vector widths. The cost that is returned is *not* normalized by
887 /// the factor width.
888 unsigned expectedCost(unsigned VF);
890 /// Returns the execution time cost of an instruction for a given vector
891 /// width. Vector width of one means scalar.
892 unsigned getInstructionCost(Instruction *I, unsigned VF);
894 /// A helper function for converting Scalar types to vector types.
895 /// If the incoming type is void, we return void. If the VF is 1, we return
897 static Type* ToVectorTy(Type *Scalar, unsigned VF);
899 /// Returns whether the instruction is a load or store and will be a emitted
900 /// as a vector operation.
901 bool isConsecutiveLoadOrStore(Instruction *I);
903 /// The loop that we evaluate.
907 /// Loop Info analysis.
909 /// Vectorization legality.
910 LoopVectorizationLegality *Legal;
911 /// Vector target information.
912 const TargetTransformInfo &TTI;
913 /// Target data layout information.
914 const DataLayout *DL;
915 /// Target Library Info.
916 const TargetLibraryInfo *TLI;
919 /// Utility class for getting and setting loop vectorizer hints in the form
920 /// of loop metadata.
921 class LoopVectorizeHints {
924 FK_Undefined = -1, ///< Not selected.
925 FK_Disabled = 0, ///< Forcing disabled.
926 FK_Enabled = 1, ///< Forcing enabled.
929 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
930 : Width(VectorizationFactor),
931 Unroll(DisableUnrolling),
933 LoopID(L->getLoopID()) {
935 // force-vector-unroll overrides DisableUnrolling.
936 if (VectorizationUnroll.getNumOccurrences() > 0)
937 Unroll = VectorizationUnroll;
939 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
940 << "LV: Unrolling disabled by the pass manager\n");
943 /// Return the loop vectorizer metadata prefix.
944 static StringRef Prefix() { return "llvm.loop.vectorize."; }
946 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
947 SmallVector<Value*, 2> Vals;
948 Vals.push_back(MDString::get(Context, Name));
949 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
950 return MDNode::get(Context, Vals);
953 /// Mark the loop L as already vectorized by setting the width to 1.
954 void setAlreadyVectorized(Loop *L) {
955 LLVMContext &Context = L->getHeader()->getContext();
959 // Create a new loop id with one more operand for the already_vectorized
960 // hint. If the loop already has a loop id then copy the existing operands.
961 SmallVector<Value*, 4> Vals(1);
963 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
964 Vals.push_back(LoopID->getOperand(i));
966 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
967 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
969 MDNode *NewLoopID = MDNode::get(Context, Vals);
970 // Set operand 0 to refer to the loop id itself.
971 NewLoopID->replaceOperandWith(0, NewLoopID);
973 L->setLoopID(NewLoopID);
975 LoopID->replaceAllUsesWith(NewLoopID);
980 std::string emitRemark() const {
982 R << "vectorization ";
984 case LoopVectorizeHints::FK_Disabled:
985 R << "is explicitly disabled";
987 case LoopVectorizeHints::FK_Enabled:
988 R << "is explicitly enabled";
989 if (Width != 0 && Unroll != 0)
990 R << " with width " << Width << " and interleave count " << Unroll;
992 R << " with width " << Width;
993 else if (Unroll != 0)
994 R << " with interleave count " << Unroll;
996 case LoopVectorizeHints::FK_Undefined:
997 R << "was not specified";
1003 unsigned getWidth() const { return Width; }
1004 unsigned getUnroll() const { return Unroll; }
1005 enum ForceKind getForce() const { return Force; }
1006 MDNode *getLoopID() const { return LoopID; }
1009 /// Find hints specified in the loop metadata.
1010 void getHints(const Loop *L) {
1014 // First operand should refer to the loop id itself.
1015 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1016 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1018 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1019 const MDString *S = nullptr;
1020 SmallVector<Value*, 4> Args;
1022 // The expected hint is either a MDString or a MDNode with the first
1023 // operand a MDString.
1024 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1025 if (!MD || MD->getNumOperands() == 0)
1027 S = dyn_cast<MDString>(MD->getOperand(0));
1028 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1029 Args.push_back(MD->getOperand(i));
1031 S = dyn_cast<MDString>(LoopID->getOperand(i));
1032 assert(Args.size() == 0 && "too many arguments for MDString");
1038 // Check if the hint starts with the vectorizer prefix.
1039 StringRef Hint = S->getString();
1040 if (!Hint.startswith(Prefix()))
1042 // Remove the prefix.
1043 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1045 if (Args.size() == 1)
1046 getHint(Hint, Args[0]);
1050 // Check string hint with one operand.
1051 void getHint(StringRef Hint, Value *Arg) {
1052 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1054 unsigned Val = C->getZExtValue();
1056 if (Hint == "width") {
1057 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1060 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1061 } else if (Hint == "unroll") {
1062 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1065 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1066 } else if (Hint == "enable") {
1067 if (C->getBitWidth() == 1)
1068 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1069 : LoopVectorizeHints::FK_Disabled;
1071 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1073 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1077 /// Vectorization width.
1079 /// Vectorization unroll factor.
1081 /// Vectorization forced
1082 enum ForceKind Force;
1087 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1089 return V.push_back(&L);
1091 for (Loop *InnerL : L)
1092 addInnerLoop(*InnerL, V);
1095 /// The LoopVectorize Pass.
1096 struct LoopVectorize : public FunctionPass {
1097 /// Pass identification, replacement for typeid
1100 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1102 DisableUnrolling(NoUnrolling),
1103 AlwaysVectorize(AlwaysVectorize) {
1104 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1107 ScalarEvolution *SE;
1108 const DataLayout *DL;
1110 TargetTransformInfo *TTI;
1112 BlockFrequencyInfo *BFI;
1113 TargetLibraryInfo *TLI;
1114 bool DisableUnrolling;
1115 bool AlwaysVectorize;
1117 BlockFrequency ColdEntryFreq;
1119 bool runOnFunction(Function &F) override {
1120 SE = &getAnalysis<ScalarEvolution>();
1121 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1122 DL = DLP ? &DLP->getDataLayout() : nullptr;
1123 LI = &getAnalysis<LoopInfo>();
1124 TTI = &getAnalysis<TargetTransformInfo>();
1125 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1126 BFI = &getAnalysis<BlockFrequencyInfo>();
1127 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1129 // Compute some weights outside of the loop over the loops. Compute this
1130 // using a BranchProbability to re-use its scaling math.
1131 const BranchProbability ColdProb(1, 5); // 20%
1132 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1134 // If the target claims to have no vector registers don't attempt
1136 if (!TTI->getNumberOfRegisters(true))
1140 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1141 << ": Missing data layout\n");
1145 // Build up a worklist of inner-loops to vectorize. This is necessary as
1146 // the act of vectorizing or partially unrolling a loop creates new loops
1147 // and can invalidate iterators across the loops.
1148 SmallVector<Loop *, 8> Worklist;
1151 addInnerLoop(*L, Worklist);
1153 LoopsAnalyzed += Worklist.size();
1155 // Now walk the identified inner loops.
1156 bool Changed = false;
1157 while (!Worklist.empty())
1158 Changed |= processLoop(Worklist.pop_back_val());
1160 // Process each loop nest in the function.
1164 bool processLoop(Loop *L) {
1165 assert(L->empty() && "Only process inner loops.");
1168 const std::string DebugLocStr = getDebugLocString(L);
1171 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1172 << L->getHeader()->getParent()->getName() << "\" from "
1173 << DebugLocStr << "\n");
1175 LoopVectorizeHints Hints(L, DisableUnrolling);
1177 DEBUG(dbgs() << "LV: Loop hints:"
1179 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1181 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1183 : "?")) << " width=" << Hints.getWidth()
1184 << " unroll=" << Hints.getUnroll() << "\n");
1186 // Function containing loop
1187 Function *F = L->getHeader()->getParent();
1189 // Looking at the diagnostic output is the only way to determine if a loop
1190 // was vectorized (other than looking at the IR or machine code), so it
1191 // is important to generate an optimization remark for each loop. Most of
1192 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1193 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1194 // less verbose reporting vectorized loops and unvectorized loops that may
1195 // benefit from vectorization, respectively.
1197 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1198 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1199 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1200 L->getStartLoc(), Hints.emitRemark());
1204 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1205 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1206 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1207 L->getStartLoc(), Hints.emitRemark());
1211 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1212 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1213 emitOptimizationRemarkAnalysis(
1214 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1215 "loop not vectorized: vector width and interleave count are "
1216 "explicitly set to 1");
1220 // Check the loop for a trip count threshold:
1221 // do not vectorize loops with a tiny trip count.
1222 BasicBlock *Latch = L->getLoopLatch();
1223 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1224 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1225 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1226 << "This loop is not worth vectorizing.");
1227 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1228 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1230 DEBUG(dbgs() << "\n");
1231 emitOptimizationRemarkAnalysis(
1232 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1233 "vectorization is not beneficial and is not explicitly forced");
1238 // Check if it is legal to vectorize the loop.
1239 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, F);
1240 if (!LVL.canVectorize()) {
1241 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1242 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1243 L->getStartLoc(), Hints.emitRemark());
1247 // Use the cost model.
1248 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1250 // Check the function attributes to find out if this function should be
1251 // optimized for size.
1252 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1253 F->hasFnAttribute(Attribute::OptimizeForSize);
1255 // Compute the weighted frequency of this loop being executed and see if it
1256 // is less than 20% of the function entry baseline frequency. Note that we
1257 // always have a canonical loop here because we think we *can* vectoriez.
1258 // FIXME: This is hidden behind a flag due to pervasive problems with
1259 // exactly what block frequency models.
1260 if (LoopVectorizeWithBlockFrequency) {
1261 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1262 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1263 LoopEntryFreq < ColdEntryFreq)
1267 // Check the function attributes to see if implicit floats are allowed.a
1268 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1269 // an integer loop and the vector instructions selected are purely integer
1270 // vector instructions?
1271 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1272 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1273 "attribute is used.\n");
1274 emitOptimizationRemarkAnalysis(
1275 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1276 "loop not vectorized due to NoImplicitFloat attribute");
1277 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1278 L->getStartLoc(), Hints.emitRemark());
1282 // Select the optimal vectorization factor.
1283 const LoopVectorizationCostModel::VectorizationFactor VF =
1284 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1286 LoopVectorizeHints::FK_Enabled);
1288 // Select the unroll factor.
1290 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1292 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1293 << DebugLocStr << '\n');
1294 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1296 if (VF.Width == 1) {
1297 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1300 emitOptimizationRemarkAnalysis(
1301 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1302 "not beneficial to vectorize and user disabled interleaving");
1305 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1307 // Report the unrolling decision.
1308 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1309 Twine("unrolled with interleaving factor " +
1311 " (vectorization not beneficial)"));
1313 // We decided not to vectorize, but we may want to unroll.
1315 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1316 Unroller.vectorize(&LVL);
1318 // If we decided that it is *legal* to vectorize the loop then do it.
1319 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1323 // Report the vectorization decision.
1324 emitOptimizationRemark(
1325 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1326 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1327 ", unrolling interleave factor: " + Twine(UF) + ")");
1330 // Mark the loop as already vectorized to avoid vectorizing again.
1331 Hints.setAlreadyVectorized(L);
1333 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1337 void getAnalysisUsage(AnalysisUsage &AU) const override {
1338 AU.addRequiredID(LoopSimplifyID);
1339 AU.addRequiredID(LCSSAID);
1340 AU.addRequired<BlockFrequencyInfo>();
1341 AU.addRequired<DominatorTreeWrapperPass>();
1342 AU.addRequired<LoopInfo>();
1343 AU.addRequired<ScalarEvolution>();
1344 AU.addRequired<TargetTransformInfo>();
1345 AU.addPreserved<LoopInfo>();
1346 AU.addPreserved<DominatorTreeWrapperPass>();
1351 } // end anonymous namespace
1353 //===----------------------------------------------------------------------===//
1354 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1355 // LoopVectorizationCostModel.
1356 //===----------------------------------------------------------------------===//
1358 static Value *stripIntegerCast(Value *V) {
1359 if (CastInst *CI = dyn_cast<CastInst>(V))
1360 if (CI->getOperand(0)->getType()->isIntegerTy())
1361 return CI->getOperand(0);
1365 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1367 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1369 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1370 ValueToValueMap &PtrToStride,
1371 Value *Ptr, Value *OrigPtr = nullptr) {
1373 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1375 // If there is an entry in the map return the SCEV of the pointer with the
1376 // symbolic stride replaced by one.
1377 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1378 if (SI != PtrToStride.end()) {
1379 Value *StrideVal = SI->second;
1382 StrideVal = stripIntegerCast(StrideVal);
1384 // Replace symbolic stride by one.
1385 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1386 ValueToValueMap RewriteMap;
1387 RewriteMap[StrideVal] = One;
1390 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1391 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1396 // Otherwise, just return the SCEV of the original pointer.
1397 return SE->getSCEV(Ptr);
1400 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1401 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1402 ValueToValueMap &Strides) {
1403 // Get the stride replaced scev.
1404 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1405 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1406 assert(AR && "Invalid addrec expression");
1407 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1408 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1409 Pointers.push_back(Ptr);
1410 Starts.push_back(AR->getStart());
1411 Ends.push_back(ScEnd);
1412 IsWritePtr.push_back(WritePtr);
1413 DependencySetId.push_back(DepSetId);
1416 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1417 // We need to place the broadcast of invariant variables outside the loop.
1418 Instruction *Instr = dyn_cast<Instruction>(V);
1420 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1421 Instr->getParent()) != LoopVectorBody.end());
1422 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1424 // Place the code for broadcasting invariant variables in the new preheader.
1425 IRBuilder<>::InsertPointGuard Guard(Builder);
1427 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1429 // Broadcast the scalar into all locations in the vector.
1430 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1435 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1437 assert(Val->getType()->isVectorTy() && "Must be a vector");
1438 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1439 "Elem must be an integer");
1440 // Create the types.
1441 Type *ITy = Val->getType()->getScalarType();
1442 VectorType *Ty = cast<VectorType>(Val->getType());
1443 int VLen = Ty->getNumElements();
1444 SmallVector<Constant*, 8> Indices;
1446 // Create a vector of consecutive numbers from zero to VF.
1447 for (int i = 0; i < VLen; ++i) {
1448 int64_t Idx = Negate ? (-i) : i;
1449 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1452 // Add the consecutive indices to the vector value.
1453 Constant *Cv = ConstantVector::get(Indices);
1454 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1455 return Builder.CreateAdd(Val, Cv, "induction");
1458 /// \brief Find the operand of the GEP that should be checked for consecutive
1459 /// stores. This ignores trailing indices that have no effect on the final
1461 static unsigned getGEPInductionOperand(const DataLayout *DL,
1462 const GetElementPtrInst *Gep) {
1463 unsigned LastOperand = Gep->getNumOperands() - 1;
1464 unsigned GEPAllocSize = DL->getTypeAllocSize(
1465 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1467 // Walk backwards and try to peel off zeros.
1468 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1469 // Find the type we're currently indexing into.
1470 gep_type_iterator GEPTI = gep_type_begin(Gep);
1471 std::advance(GEPTI, LastOperand - 1);
1473 // If it's a type with the same allocation size as the result of the GEP we
1474 // can peel off the zero index.
1475 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1483 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1484 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1485 // Make sure that the pointer does not point to structs.
1486 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1489 // If this value is a pointer induction variable we know it is consecutive.
1490 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1491 if (Phi && Inductions.count(Phi)) {
1492 InductionInfo II = Inductions[Phi];
1493 if (IK_PtrInduction == II.IK)
1495 else if (IK_ReversePtrInduction == II.IK)
1499 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1503 unsigned NumOperands = Gep->getNumOperands();
1504 Value *GpPtr = Gep->getPointerOperand();
1505 // If this GEP value is a consecutive pointer induction variable and all of
1506 // the indices are constant then we know it is consecutive. We can
1507 Phi = dyn_cast<PHINode>(GpPtr);
1508 if (Phi && Inductions.count(Phi)) {
1510 // Make sure that the pointer does not point to structs.
1511 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1512 if (GepPtrType->getElementType()->isAggregateType())
1515 // Make sure that all of the index operands are loop invariant.
1516 for (unsigned i = 1; i < NumOperands; ++i)
1517 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1520 InductionInfo II = Inductions[Phi];
1521 if (IK_PtrInduction == II.IK)
1523 else if (IK_ReversePtrInduction == II.IK)
1527 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1529 // Check that all of the gep indices are uniform except for our induction
1531 for (unsigned i = 0; i != NumOperands; ++i)
1532 if (i != InductionOperand &&
1533 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1536 // We can emit wide load/stores only if the last non-zero index is the
1537 // induction variable.
1538 const SCEV *Last = nullptr;
1539 if (!Strides.count(Gep))
1540 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1542 // Because of the multiplication by a stride we can have a s/zext cast.
1543 // We are going to replace this stride by 1 so the cast is safe to ignore.
1545 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1546 // %0 = trunc i64 %indvars.iv to i32
1547 // %mul = mul i32 %0, %Stride1
1548 // %idxprom = zext i32 %mul to i64 << Safe cast.
1549 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1551 Last = replaceSymbolicStrideSCEV(SE, Strides,
1552 Gep->getOperand(InductionOperand), Gep);
1553 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1555 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1559 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1560 const SCEV *Step = AR->getStepRecurrence(*SE);
1562 // The memory is consecutive because the last index is consecutive
1563 // and all other indices are loop invariant.
1566 if (Step->isAllOnesValue())
1573 bool LoopVectorizationLegality::isUniform(Value *V) {
1574 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1577 InnerLoopVectorizer::VectorParts&
1578 InnerLoopVectorizer::getVectorValue(Value *V) {
1579 assert(V != Induction && "The new induction variable should not be used.");
1580 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1582 // If we have a stride that is replaced by one, do it here.
1583 if (Legal->hasStride(V))
1584 V = ConstantInt::get(V->getType(), 1);
1586 // If we have this scalar in the map, return it.
1587 if (WidenMap.has(V))
1588 return WidenMap.get(V);
1590 // If this scalar is unknown, assume that it is a constant or that it is
1591 // loop invariant. Broadcast V and save the value for future uses.
1592 Value *B = getBroadcastInstrs(V);
1593 return WidenMap.splat(V, B);
1596 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1597 assert(Vec->getType()->isVectorTy() && "Invalid type");
1598 SmallVector<Constant*, 8> ShuffleMask;
1599 for (unsigned i = 0; i < VF; ++i)
1600 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1602 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1603 ConstantVector::get(ShuffleMask),
1607 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1608 // Attempt to issue a wide load.
1609 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1610 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1612 assert((LI || SI) && "Invalid Load/Store instruction");
1614 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1615 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1616 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1617 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1618 // An alignment of 0 means target abi alignment. We need to use the scalar's
1619 // target abi alignment in such a case.
1621 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1622 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1623 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1624 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1626 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1627 return scalarizeInstruction(Instr, true);
1629 if (ScalarAllocatedSize != VectorElementSize)
1630 return scalarizeInstruction(Instr);
1632 // If the pointer is loop invariant or if it is non-consecutive,
1633 // scalarize the load.
1634 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1635 bool Reverse = ConsecutiveStride < 0;
1636 bool UniformLoad = LI && Legal->isUniform(Ptr);
1637 if (!ConsecutiveStride || UniformLoad)
1638 return scalarizeInstruction(Instr);
1640 Constant *Zero = Builder.getInt32(0);
1641 VectorParts &Entry = WidenMap.get(Instr);
1643 // Handle consecutive loads/stores.
1644 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1645 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1646 setDebugLocFromInst(Builder, Gep);
1647 Value *PtrOperand = Gep->getPointerOperand();
1648 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1649 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1651 // Create the new GEP with the new induction variable.
1652 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1653 Gep2->setOperand(0, FirstBasePtr);
1654 Gep2->setName("gep.indvar.base");
1655 Ptr = Builder.Insert(Gep2);
1657 setDebugLocFromInst(Builder, Gep);
1658 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1659 OrigLoop) && "Base ptr must be invariant");
1661 // The last index does not have to be the induction. It can be
1662 // consecutive and be a function of the index. For example A[I+1];
1663 unsigned NumOperands = Gep->getNumOperands();
1664 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1665 // Create the new GEP with the new induction variable.
1666 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1668 for (unsigned i = 0; i < NumOperands; ++i) {
1669 Value *GepOperand = Gep->getOperand(i);
1670 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1672 // Update last index or loop invariant instruction anchored in loop.
1673 if (i == InductionOperand ||
1674 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1675 assert((i == InductionOperand ||
1676 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1677 "Must be last index or loop invariant");
1679 VectorParts &GEPParts = getVectorValue(GepOperand);
1680 Value *Index = GEPParts[0];
1681 Index = Builder.CreateExtractElement(Index, Zero);
1682 Gep2->setOperand(i, Index);
1683 Gep2->setName("gep.indvar.idx");
1686 Ptr = Builder.Insert(Gep2);
1688 // Use the induction element ptr.
1689 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1690 setDebugLocFromInst(Builder, Ptr);
1691 VectorParts &PtrVal = getVectorValue(Ptr);
1692 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1697 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1698 "We do not allow storing to uniform addresses");
1699 setDebugLocFromInst(Builder, SI);
1700 // We don't want to update the value in the map as it might be used in
1701 // another expression. So don't use a reference type for "StoredVal".
1702 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1704 for (unsigned Part = 0; Part < UF; ++Part) {
1705 // Calculate the pointer for the specific unroll-part.
1706 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1709 // If we store to reverse consecutive memory locations then we need
1710 // to reverse the order of elements in the stored value.
1711 StoredVal[Part] = reverseVector(StoredVal[Part]);
1712 // If the address is consecutive but reversed, then the
1713 // wide store needs to start at the last vector element.
1714 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1715 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1718 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1719 DataTy->getPointerTo(AddressSpace));
1720 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1726 assert(LI && "Must have a load instruction");
1727 setDebugLocFromInst(Builder, LI);
1728 for (unsigned Part = 0; Part < UF; ++Part) {
1729 // Calculate the pointer for the specific unroll-part.
1730 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1733 // If the address is consecutive but reversed, then the
1734 // wide store needs to start at the last vector element.
1735 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1736 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1739 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1740 DataTy->getPointerTo(AddressSpace));
1741 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1742 cast<LoadInst>(LI)->setAlignment(Alignment);
1743 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1747 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1748 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1749 // Holds vector parameters or scalars, in case of uniform vals.
1750 SmallVector<VectorParts, 4> Params;
1752 setDebugLocFromInst(Builder, Instr);
1754 // Find all of the vectorized parameters.
1755 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1756 Value *SrcOp = Instr->getOperand(op);
1758 // If we are accessing the old induction variable, use the new one.
1759 if (SrcOp == OldInduction) {
1760 Params.push_back(getVectorValue(SrcOp));
1764 // Try using previously calculated values.
1765 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1767 // If the src is an instruction that appeared earlier in the basic block
1768 // then it should already be vectorized.
1769 if (SrcInst && OrigLoop->contains(SrcInst)) {
1770 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1771 // The parameter is a vector value from earlier.
1772 Params.push_back(WidenMap.get(SrcInst));
1774 // The parameter is a scalar from outside the loop. Maybe even a constant.
1775 VectorParts Scalars;
1776 Scalars.append(UF, SrcOp);
1777 Params.push_back(Scalars);
1781 assert(Params.size() == Instr->getNumOperands() &&
1782 "Invalid number of operands");
1784 // Does this instruction return a value ?
1785 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1787 Value *UndefVec = IsVoidRetTy ? nullptr :
1788 UndefValue::get(VectorType::get(Instr->getType(), VF));
1789 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1790 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1792 Instruction *InsertPt = Builder.GetInsertPoint();
1793 BasicBlock *IfBlock = Builder.GetInsertBlock();
1794 BasicBlock *CondBlock = nullptr;
1797 Loop *VectorLp = nullptr;
1798 if (IfPredicateStore) {
1799 assert(Instr->getParent()->getSinglePredecessor() &&
1800 "Only support single predecessor blocks");
1801 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1802 Instr->getParent());
1803 VectorLp = LI->getLoopFor(IfBlock);
1804 assert(VectorLp && "Must have a loop for this block");
1807 // For each vector unroll 'part':
1808 for (unsigned Part = 0; Part < UF; ++Part) {
1809 // For each scalar that we create:
1810 for (unsigned Width = 0; Width < VF; ++Width) {
1813 Value *Cmp = nullptr;
1814 if (IfPredicateStore) {
1815 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1816 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1817 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1818 LoopVectorBody.push_back(CondBlock);
1819 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1820 // Update Builder with newly created basic block.
1821 Builder.SetInsertPoint(InsertPt);
1824 Instruction *Cloned = Instr->clone();
1826 Cloned->setName(Instr->getName() + ".cloned");
1827 // Replace the operands of the cloned instructions with extracted scalars.
1828 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1829 Value *Op = Params[op][Part];
1830 // Param is a vector. Need to extract the right lane.
1831 if (Op->getType()->isVectorTy())
1832 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1833 Cloned->setOperand(op, Op);
1836 // Place the cloned scalar in the new loop.
1837 Builder.Insert(Cloned);
1839 // If the original scalar returns a value we need to place it in a vector
1840 // so that future users will be able to use it.
1842 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1843 Builder.getInt32(Width));
1845 if (IfPredicateStore) {
1846 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1847 LoopVectorBody.push_back(NewIfBlock);
1848 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1849 Builder.SetInsertPoint(InsertPt);
1850 Instruction *OldBr = IfBlock->getTerminator();
1851 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1852 OldBr->eraseFromParent();
1853 IfBlock = NewIfBlock;
1859 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1863 if (Instruction *I = dyn_cast<Instruction>(V))
1864 return I->getParent() == Loc->getParent() ? I : nullptr;
1868 std::pair<Instruction *, Instruction *>
1869 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1870 Instruction *tnullptr = nullptr;
1871 if (!Legal->mustCheckStrides())
1872 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1874 IRBuilder<> ChkBuilder(Loc);
1877 Value *Check = nullptr;
1878 Instruction *FirstInst = nullptr;
1879 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1880 SE = Legal->strides_end();
1882 Value *Ptr = stripIntegerCast(*SI);
1883 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1885 // Store the first instruction we create.
1886 FirstInst = getFirstInst(FirstInst, C, Loc);
1888 Check = ChkBuilder.CreateOr(Check, C);
1893 // We have to do this trickery because the IRBuilder might fold the check to a
1894 // constant expression in which case there is no Instruction anchored in a
1896 LLVMContext &Ctx = Loc->getContext();
1897 Instruction *TheCheck =
1898 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1899 ChkBuilder.Insert(TheCheck, "stride.not.one");
1900 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1902 return std::make_pair(FirstInst, TheCheck);
1905 std::pair<Instruction *, Instruction *>
1906 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1907 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1908 Legal->getRuntimePointerCheck();
1910 Instruction *tnullptr = nullptr;
1911 if (!PtrRtCheck->Need)
1912 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1914 unsigned NumPointers = PtrRtCheck->Pointers.size();
1915 SmallVector<TrackingVH<Value> , 2> Starts;
1916 SmallVector<TrackingVH<Value> , 2> Ends;
1918 LLVMContext &Ctx = Loc->getContext();
1919 SCEVExpander Exp(*SE, "induction");
1920 Instruction *FirstInst = nullptr;
1922 for (unsigned i = 0; i < NumPointers; ++i) {
1923 Value *Ptr = PtrRtCheck->Pointers[i];
1924 const SCEV *Sc = SE->getSCEV(Ptr);
1926 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1927 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1929 Starts.push_back(Ptr);
1930 Ends.push_back(Ptr);
1932 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1933 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1935 // Use this type for pointer arithmetic.
1936 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1938 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1939 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1940 Starts.push_back(Start);
1941 Ends.push_back(End);
1945 IRBuilder<> ChkBuilder(Loc);
1946 // Our instructions might fold to a constant.
1947 Value *MemoryRuntimeCheck = nullptr;
1948 for (unsigned i = 0; i < NumPointers; ++i) {
1949 for (unsigned j = i+1; j < NumPointers; ++j) {
1950 // No need to check if two readonly pointers intersect.
1951 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1954 // Only need to check pointers between two different dependency sets.
1955 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1958 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1959 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1961 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1962 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1963 "Trying to bounds check pointers with different address spaces");
1965 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1966 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1968 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1969 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1970 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1971 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1973 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1974 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1975 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1976 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1977 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1978 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1979 if (MemoryRuntimeCheck) {
1980 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1982 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1984 MemoryRuntimeCheck = IsConflict;
1988 // We have to do this trickery because the IRBuilder might fold the check to a
1989 // constant expression in which case there is no Instruction anchored in a
1991 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1992 ConstantInt::getTrue(Ctx));
1993 ChkBuilder.Insert(Check, "memcheck.conflict");
1994 FirstInst = getFirstInst(FirstInst, Check, Loc);
1995 return std::make_pair(FirstInst, Check);
1998 void InnerLoopVectorizer::createEmptyLoop() {
2000 In this function we generate a new loop. The new loop will contain
2001 the vectorized instructions while the old loop will continue to run the
2004 [ ] <-- Back-edge taken count overflow check.
2007 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2010 || [ ] <-- vector pre header.
2014 || [ ]_| <-- vector loop.
2017 | >[ ] <--- middle-block.
2020 -|- >[ ] <--- new preheader.
2024 | [ ]_| <-- old scalar loop to handle remainder.
2027 >[ ] <-- exit block.
2031 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2032 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2033 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2034 assert(BypassBlock && "Invalid loop structure");
2035 assert(ExitBlock && "Must have an exit block");
2037 // Some loops have a single integer induction variable, while other loops
2038 // don't. One example is c++ iterators that often have multiple pointer
2039 // induction variables. In the code below we also support a case where we
2040 // don't have a single induction variable.
2041 OldInduction = Legal->getInduction();
2042 Type *IdxTy = Legal->getWidestInductionType();
2044 // Find the loop boundaries.
2045 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2046 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2048 // The exit count might have the type of i64 while the phi is i32. This can
2049 // happen if we have an induction variable that is sign extended before the
2050 // compare. The only way that we get a backedge taken count is that the
2051 // induction variable was signed and as such will not overflow. In such a case
2052 // truncation is legal.
2053 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2054 IdxTy->getPrimitiveSizeInBits())
2055 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2057 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2058 // Get the total trip count from the count by adding 1.
2059 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2060 SE->getConstant(BackedgeTakeCount->getType(), 1));
2062 // Expand the trip count and place the new instructions in the preheader.
2063 // Notice that the pre-header does not change, only the loop body.
2064 SCEVExpander Exp(*SE, "induction");
2066 // We need to test whether the backedge-taken count is uint##_max. Adding one
2067 // to it will cause overflow and an incorrect loop trip count in the vector
2068 // body. In case of overflow we want to directly jump to the scalar remainder
2070 Value *BackedgeCount =
2071 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2072 BypassBlock->getTerminator());
2073 if (BackedgeCount->getType()->isPointerTy())
2074 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2075 "backedge.ptrcnt.to.int",
2076 BypassBlock->getTerminator());
2077 Instruction *CheckBCOverflow =
2078 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2079 Constant::getAllOnesValue(BackedgeCount->getType()),
2080 "backedge.overflow", BypassBlock->getTerminator());
2082 // The loop index does not have to start at Zero. Find the original start
2083 // value from the induction PHI node. If we don't have an induction variable
2084 // then we know that it starts at zero.
2085 Builder.SetInsertPoint(BypassBlock->getTerminator());
2086 Value *StartIdx = ExtendedIdx = OldInduction ?
2087 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2089 ConstantInt::get(IdxTy, 0);
2091 // We need an instruction to anchor the overflow check on. StartIdx needs to
2092 // be defined before the overflow check branch. Because the scalar preheader
2093 // is going to merge the start index and so the overflow branch block needs to
2094 // contain a definition of the start index.
2095 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2096 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2097 BypassBlock->getTerminator());
2099 // Count holds the overall loop count (N).
2100 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2101 BypassBlock->getTerminator());
2103 LoopBypassBlocks.push_back(BypassBlock);
2105 // Split the single block loop into the two loop structure described above.
2106 BasicBlock *VectorPH =
2107 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2108 BasicBlock *VecBody =
2109 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2110 BasicBlock *MiddleBlock =
2111 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2112 BasicBlock *ScalarPH =
2113 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2115 // Create and register the new vector loop.
2116 Loop* Lp = new Loop();
2117 Loop *ParentLoop = OrigLoop->getParentLoop();
2119 // Insert the new loop into the loop nest and register the new basic blocks
2120 // before calling any utilities such as SCEV that require valid LoopInfo.
2122 ParentLoop->addChildLoop(Lp);
2123 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2124 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2125 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2127 LI->addTopLevelLoop(Lp);
2129 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2131 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2133 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2135 // Generate the induction variable.
2136 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2137 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2138 // The loop step is equal to the vectorization factor (num of SIMD elements)
2139 // times the unroll factor (num of SIMD instructions).
2140 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2142 // This is the IR builder that we use to add all of the logic for bypassing
2143 // the new vector loop.
2144 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2145 setDebugLocFromInst(BypassBuilder,
2146 getDebugLocFromInstOrOperands(OldInduction));
2148 // We may need to extend the index in case there is a type mismatch.
2149 // We know that the count starts at zero and does not overflow.
2150 if (Count->getType() != IdxTy) {
2151 // The exit count can be of pointer type. Convert it to the correct
2153 if (ExitCount->getType()->isPointerTy())
2154 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2156 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2159 // Add the start index to the loop count to get the new end index.
2160 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2162 // Now we need to generate the expression for N - (N % VF), which is
2163 // the part that the vectorized body will execute.
2164 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2165 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2166 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2167 "end.idx.rnd.down");
2169 // Now, compare the new count to zero. If it is zero skip the vector loop and
2170 // jump to the scalar loop.
2172 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2174 BasicBlock *LastBypassBlock = BypassBlock;
2176 // Generate code to check that the loops trip count that we computed by adding
2177 // one to the backedge-taken count will not overflow.
2179 auto PastOverflowCheck =
2180 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2181 BasicBlock *CheckBlock =
2182 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2184 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2185 LoopBypassBlocks.push_back(CheckBlock);
2186 Instruction *OldTerm = LastBypassBlock->getTerminator();
2187 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2188 OldTerm->eraseFromParent();
2189 LastBypassBlock = CheckBlock;
2192 // Generate the code to check that the strides we assumed to be one are really
2193 // one. We want the new basic block to start at the first instruction in a
2194 // sequence of instructions that form a check.
2195 Instruction *StrideCheck;
2196 Instruction *FirstCheckInst;
2197 std::tie(FirstCheckInst, StrideCheck) =
2198 addStrideCheck(LastBypassBlock->getTerminator());
2200 // Create a new block containing the stride check.
2201 BasicBlock *CheckBlock =
2202 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2204 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2205 LoopBypassBlocks.push_back(CheckBlock);
2207 // Replace the branch into the memory check block with a conditional branch
2208 // for the "few elements case".
2209 Instruction *OldTerm = LastBypassBlock->getTerminator();
2210 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2211 OldTerm->eraseFromParent();
2214 LastBypassBlock = CheckBlock;
2217 // Generate the code that checks in runtime if arrays overlap. We put the
2218 // checks into a separate block to make the more common case of few elements
2220 Instruction *MemRuntimeCheck;
2221 std::tie(FirstCheckInst, MemRuntimeCheck) =
2222 addRuntimeCheck(LastBypassBlock->getTerminator());
2223 if (MemRuntimeCheck) {
2224 // Create a new block containing the memory check.
2225 BasicBlock *CheckBlock =
2226 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2228 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2229 LoopBypassBlocks.push_back(CheckBlock);
2231 // Replace the branch into the memory check block with a conditional branch
2232 // for the "few elements case".
2233 Instruction *OldTerm = LastBypassBlock->getTerminator();
2234 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2235 OldTerm->eraseFromParent();
2237 Cmp = MemRuntimeCheck;
2238 LastBypassBlock = CheckBlock;
2241 LastBypassBlock->getTerminator()->eraseFromParent();
2242 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2245 // We are going to resume the execution of the scalar loop.
2246 // Go over all of the induction variables that we found and fix the
2247 // PHIs that are left in the scalar version of the loop.
2248 // The starting values of PHI nodes depend on the counter of the last
2249 // iteration in the vectorized loop.
2250 // If we come from a bypass edge then we need to start from the original
2253 // This variable saves the new starting index for the scalar loop.
2254 PHINode *ResumeIndex = nullptr;
2255 LoopVectorizationLegality::InductionList::iterator I, E;
2256 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2257 // Set builder to point to last bypass block.
2258 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2259 for (I = List->begin(), E = List->end(); I != E; ++I) {
2260 PHINode *OrigPhi = I->first;
2261 LoopVectorizationLegality::InductionInfo II = I->second;
2263 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2264 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2265 MiddleBlock->getTerminator());
2266 // We might have extended the type of the induction variable but we need a
2267 // truncated version for the scalar loop.
2268 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2269 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2270 MiddleBlock->getTerminator()) : nullptr;
2272 // Create phi nodes to merge from the backedge-taken check block.
2273 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2274 ScalarPH->getTerminator());
2275 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2277 PHINode *BCTruncResumeVal = nullptr;
2278 if (OrigPhi == OldInduction) {
2280 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2281 ScalarPH->getTerminator());
2282 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2285 Value *EndValue = nullptr;
2287 case LoopVectorizationLegality::IK_NoInduction:
2288 llvm_unreachable("Unknown induction");
2289 case LoopVectorizationLegality::IK_IntInduction: {
2290 // Handle the integer induction counter.
2291 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2293 // We have the canonical induction variable.
2294 if (OrigPhi == OldInduction) {
2295 // Create a truncated version of the resume value for the scalar loop,
2296 // we might have promoted the type to a larger width.
2298 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2299 // The new PHI merges the original incoming value, in case of a bypass,
2300 // or the value at the end of the vectorized loop.
2301 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2302 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2303 TruncResumeVal->addIncoming(EndValue, VecBody);
2305 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2307 // We know what the end value is.
2308 EndValue = IdxEndRoundDown;
2309 // We also know which PHI node holds it.
2310 ResumeIndex = ResumeVal;
2314 // Not the canonical induction variable - add the vector loop count to the
2316 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2317 II.StartValue->getType(),
2319 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2322 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2323 // Convert the CountRoundDown variable to the PHI size.
2324 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2325 II.StartValue->getType(),
2327 // Handle reverse integer induction counter.
2328 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2331 case LoopVectorizationLegality::IK_PtrInduction: {
2332 // For pointer induction variables, calculate the offset using
2334 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2338 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2339 // The value at the end of the loop for the reverse pointer is calculated
2340 // by creating a GEP with a negative index starting from the start value.
2341 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2342 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2344 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2350 // The new PHI merges the original incoming value, in case of a bypass,
2351 // or the value at the end of the vectorized loop.
2352 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2353 if (OrigPhi == OldInduction)
2354 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2356 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2358 ResumeVal->addIncoming(EndValue, VecBody);
2360 // Fix the scalar body counter (PHI node).
2361 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2363 // The old induction's phi node in the scalar body needs the truncated
2365 if (OrigPhi == OldInduction) {
2366 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2367 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2369 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2370 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2374 // If we are generating a new induction variable then we also need to
2375 // generate the code that calculates the exit value. This value is not
2376 // simply the end of the counter because we may skip the vectorized body
2377 // in case of a runtime check.
2379 assert(!ResumeIndex && "Unexpected resume value found");
2380 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2381 MiddleBlock->getTerminator());
2382 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2383 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2384 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2387 // Make sure that we found the index where scalar loop needs to continue.
2388 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2389 "Invalid resume Index");
2391 // Add a check in the middle block to see if we have completed
2392 // all of the iterations in the first vector loop.
2393 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2394 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2395 ResumeIndex, "cmp.n",
2396 MiddleBlock->getTerminator());
2398 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2399 // Remove the old terminator.
2400 MiddleBlock->getTerminator()->eraseFromParent();
2402 // Create i+1 and fill the PHINode.
2403 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2404 Induction->addIncoming(StartIdx, VectorPH);
2405 Induction->addIncoming(NextIdx, VecBody);
2406 // Create the compare.
2407 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2408 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2410 // Now we have two terminators. Remove the old one from the block.
2411 VecBody->getTerminator()->eraseFromParent();
2413 // Get ready to start creating new instructions into the vectorized body.
2414 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2417 LoopVectorPreHeader = VectorPH;
2418 LoopScalarPreHeader = ScalarPH;
2419 LoopMiddleBlock = MiddleBlock;
2420 LoopExitBlock = ExitBlock;
2421 LoopVectorBody.push_back(VecBody);
2422 LoopScalarBody = OldBasicBlock;
2424 LoopVectorizeHints Hints(Lp, true);
2425 Hints.setAlreadyVectorized(Lp);
2428 /// This function returns the identity element (or neutral element) for
2429 /// the operation K.
2431 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2436 // Adding, Xoring, Oring zero to a number does not change it.
2437 return ConstantInt::get(Tp, 0);
2438 case RK_IntegerMult:
2439 // Multiplying a number by 1 does not change it.
2440 return ConstantInt::get(Tp, 1);
2442 // AND-ing a number with an all-1 value does not change it.
2443 return ConstantInt::get(Tp, -1, true);
2445 // Multiplying a number by 1 does not change it.
2446 return ConstantFP::get(Tp, 1.0L);
2448 // Adding zero to a number does not change it.
2449 return ConstantFP::get(Tp, 0.0L);
2451 llvm_unreachable("Unknown reduction kind");
2455 /// This function translates the reduction kind to an LLVM binary operator.
2457 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2459 case LoopVectorizationLegality::RK_IntegerAdd:
2460 return Instruction::Add;
2461 case LoopVectorizationLegality::RK_IntegerMult:
2462 return Instruction::Mul;
2463 case LoopVectorizationLegality::RK_IntegerOr:
2464 return Instruction::Or;
2465 case LoopVectorizationLegality::RK_IntegerAnd:
2466 return Instruction::And;
2467 case LoopVectorizationLegality::RK_IntegerXor:
2468 return Instruction::Xor;
2469 case LoopVectorizationLegality::RK_FloatMult:
2470 return Instruction::FMul;
2471 case LoopVectorizationLegality::RK_FloatAdd:
2472 return Instruction::FAdd;
2473 case LoopVectorizationLegality::RK_IntegerMinMax:
2474 return Instruction::ICmp;
2475 case LoopVectorizationLegality::RK_FloatMinMax:
2476 return Instruction::FCmp;
2478 llvm_unreachable("Unknown reduction operation");
2482 Value *createMinMaxOp(IRBuilder<> &Builder,
2483 LoopVectorizationLegality::MinMaxReductionKind RK,
2486 CmpInst::Predicate P = CmpInst::ICMP_NE;
2489 llvm_unreachable("Unknown min/max reduction kind");
2490 case LoopVectorizationLegality::MRK_UIntMin:
2491 P = CmpInst::ICMP_ULT;
2493 case LoopVectorizationLegality::MRK_UIntMax:
2494 P = CmpInst::ICMP_UGT;
2496 case LoopVectorizationLegality::MRK_SIntMin:
2497 P = CmpInst::ICMP_SLT;
2499 case LoopVectorizationLegality::MRK_SIntMax:
2500 P = CmpInst::ICMP_SGT;
2502 case LoopVectorizationLegality::MRK_FloatMin:
2503 P = CmpInst::FCMP_OLT;
2505 case LoopVectorizationLegality::MRK_FloatMax:
2506 P = CmpInst::FCMP_OGT;
2511 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2512 RK == LoopVectorizationLegality::MRK_FloatMax)
2513 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2515 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2517 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2522 struct CSEDenseMapInfo {
2523 static bool canHandle(Instruction *I) {
2524 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2525 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2527 static inline Instruction *getEmptyKey() {
2528 return DenseMapInfo<Instruction *>::getEmptyKey();
2530 static inline Instruction *getTombstoneKey() {
2531 return DenseMapInfo<Instruction *>::getTombstoneKey();
2533 static unsigned getHashValue(Instruction *I) {
2534 assert(canHandle(I) && "Unknown instruction!");
2535 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2536 I->value_op_end()));
2538 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2539 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2540 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2542 return LHS->isIdenticalTo(RHS);
2547 /// \brief Check whether this block is a predicated block.
2548 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2549 /// = ...; " blocks. We start with one vectorized basic block. For every
2550 /// conditional block we split this vectorized block. Therefore, every second
2551 /// block will be a predicated one.
2552 static bool isPredicatedBlock(unsigned BlockNum) {
2553 return BlockNum % 2;
2556 ///\brief Perform cse of induction variable instructions.
2557 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2558 // Perform simple cse.
2559 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2560 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2561 BasicBlock *BB = BBs[i];
2562 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2563 Instruction *In = I++;
2565 if (!CSEDenseMapInfo::canHandle(In))
2568 // Check if we can replace this instruction with any of the
2569 // visited instructions.
2570 if (Instruction *V = CSEMap.lookup(In)) {
2571 In->replaceAllUsesWith(V);
2572 In->eraseFromParent();
2575 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2576 // ...;" blocks for predicated stores. Every second block is a predicated
2578 if (isPredicatedBlock(i))
2586 /// \brief Adds a 'fast' flag to floating point operations.
2587 static Value *addFastMathFlag(Value *V) {
2588 if (isa<FPMathOperator>(V)){
2589 FastMathFlags Flags;
2590 Flags.setUnsafeAlgebra();
2591 cast<Instruction>(V)->setFastMathFlags(Flags);
2596 void InnerLoopVectorizer::vectorizeLoop() {
2597 //===------------------------------------------------===//
2599 // Notice: any optimization or new instruction that go
2600 // into the code below should be also be implemented in
2603 //===------------------------------------------------===//
2604 Constant *Zero = Builder.getInt32(0);
2606 // In order to support reduction variables we need to be able to vectorize
2607 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2608 // stages. First, we create a new vector PHI node with no incoming edges.
2609 // We use this value when we vectorize all of the instructions that use the
2610 // PHI. Next, after all of the instructions in the block are complete we
2611 // add the new incoming edges to the PHI. At this point all of the
2612 // instructions in the basic block are vectorized, so we can use them to
2613 // construct the PHI.
2614 PhiVector RdxPHIsToFix;
2616 // Scan the loop in a topological order to ensure that defs are vectorized
2618 LoopBlocksDFS DFS(OrigLoop);
2621 // Vectorize all of the blocks in the original loop.
2622 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2623 be = DFS.endRPO(); bb != be; ++bb)
2624 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2626 // At this point every instruction in the original loop is widened to
2627 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2628 // that we vectorized. The PHI nodes are currently empty because we did
2629 // not want to introduce cycles. Notice that the remaining PHI nodes
2630 // that we need to fix are reduction variables.
2632 // Create the 'reduced' values for each of the induction vars.
2633 // The reduced values are the vector values that we scalarize and combine
2634 // after the loop is finished.
2635 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2637 PHINode *RdxPhi = *it;
2638 assert(RdxPhi && "Unable to recover vectorized PHI");
2640 // Find the reduction variable descriptor.
2641 assert(Legal->getReductionVars()->count(RdxPhi) &&
2642 "Unable to find the reduction variable");
2643 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2644 (*Legal->getReductionVars())[RdxPhi];
2646 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2648 // We need to generate a reduction vector from the incoming scalar.
2649 // To do so, we need to generate the 'identity' vector and override
2650 // one of the elements with the incoming scalar reduction. We need
2651 // to do it in the vector-loop preheader.
2652 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2654 // This is the vector-clone of the value that leaves the loop.
2655 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2656 Type *VecTy = VectorExit[0]->getType();
2658 // Find the reduction identity variable. Zero for addition, or, xor,
2659 // one for multiplication, -1 for And.
2662 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2663 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2664 // MinMax reduction have the start value as their identify.
2666 VectorStart = Identity = RdxDesc.StartValue;
2668 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2673 // Handle other reduction kinds:
2675 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2676 VecTy->getScalarType());
2679 // This vector is the Identity vector where the first element is the
2680 // incoming scalar reduction.
2681 VectorStart = RdxDesc.StartValue;
2683 Identity = ConstantVector::getSplat(VF, Iden);
2685 // This vector is the Identity vector where the first element is the
2686 // incoming scalar reduction.
2687 VectorStart = Builder.CreateInsertElement(Identity,
2688 RdxDesc.StartValue, Zero);
2692 // Fix the vector-loop phi.
2693 // We created the induction variable so we know that the
2694 // preheader is the first entry.
2695 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2697 // Reductions do not have to start at zero. They can start with
2698 // any loop invariant values.
2699 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2700 BasicBlock *Latch = OrigLoop->getLoopLatch();
2701 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2702 VectorParts &Val = getVectorValue(LoopVal);
2703 for (unsigned part = 0; part < UF; ++part) {
2704 // Make sure to add the reduction stat value only to the
2705 // first unroll part.
2706 Value *StartVal = (part == 0) ? VectorStart : Identity;
2707 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2708 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2709 LoopVectorBody.back());
2712 // Before each round, move the insertion point right between
2713 // the PHIs and the values we are going to write.
2714 // This allows us to write both PHINodes and the extractelement
2716 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2718 VectorParts RdxParts;
2719 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2720 for (unsigned part = 0; part < UF; ++part) {
2721 // This PHINode contains the vectorized reduction variable, or
2722 // the initial value vector, if we bypass the vector loop.
2723 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2724 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2725 Value *StartVal = (part == 0) ? VectorStart : Identity;
2726 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2727 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2728 NewPhi->addIncoming(RdxExitVal[part],
2729 LoopVectorBody.back());
2730 RdxParts.push_back(NewPhi);
2733 // Reduce all of the unrolled parts into a single vector.
2734 Value *ReducedPartRdx = RdxParts[0];
2735 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2736 setDebugLocFromInst(Builder, ReducedPartRdx);
2737 for (unsigned part = 1; part < UF; ++part) {
2738 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2739 // Floating point operations had to be 'fast' to enable the reduction.
2740 ReducedPartRdx = addFastMathFlag(
2741 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2742 ReducedPartRdx, "bin.rdx"));
2744 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2745 ReducedPartRdx, RdxParts[part]);
2749 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2750 // and vector ops, reducing the set of values being computed by half each
2752 assert(isPowerOf2_32(VF) &&
2753 "Reduction emission only supported for pow2 vectors!");
2754 Value *TmpVec = ReducedPartRdx;
2755 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2756 for (unsigned i = VF; i != 1; i >>= 1) {
2757 // Move the upper half of the vector to the lower half.
2758 for (unsigned j = 0; j != i/2; ++j)
2759 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2761 // Fill the rest of the mask with undef.
2762 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2763 UndefValue::get(Builder.getInt32Ty()));
2766 Builder.CreateShuffleVector(TmpVec,
2767 UndefValue::get(TmpVec->getType()),
2768 ConstantVector::get(ShuffleMask),
2771 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2772 // Floating point operations had to be 'fast' to enable the reduction.
2773 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2774 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2776 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2779 // The result is in the first element of the vector.
2780 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2781 Builder.getInt32(0));
2784 // Create a phi node that merges control-flow from the backedge-taken check
2785 // block and the middle block.
2786 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2787 LoopScalarPreHeader->getTerminator());
2788 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2789 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2791 // Now, we need to fix the users of the reduction variable
2792 // inside and outside of the scalar remainder loop.
2793 // We know that the loop is in LCSSA form. We need to update the
2794 // PHI nodes in the exit blocks.
2795 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2796 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2797 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2798 if (!LCSSAPhi) break;
2800 // All PHINodes need to have a single entry edge, or two if
2801 // we already fixed them.
2802 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2804 // We found our reduction value exit-PHI. Update it with the
2805 // incoming bypass edge.
2806 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2807 // Add an edge coming from the bypass.
2808 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2811 }// end of the LCSSA phi scan.
2813 // Fix the scalar loop reduction variable with the incoming reduction sum
2814 // from the vector body and from the backedge value.
2815 int IncomingEdgeBlockIdx =
2816 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2817 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2818 // Pick the other block.
2819 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2820 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2821 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2822 }// end of for each redux variable.
2826 // Remove redundant induction instructions.
2827 cse(LoopVectorBody);
2830 void InnerLoopVectorizer::fixLCSSAPHIs() {
2831 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2832 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2833 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2834 if (!LCSSAPhi) break;
2835 if (LCSSAPhi->getNumIncomingValues() == 1)
2836 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2841 InnerLoopVectorizer::VectorParts
2842 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2843 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2846 // Look for cached value.
2847 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2848 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2849 if (ECEntryIt != MaskCache.end())
2850 return ECEntryIt->second;
2852 VectorParts SrcMask = createBlockInMask(Src);
2854 // The terminator has to be a branch inst!
2855 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2856 assert(BI && "Unexpected terminator found");
2858 if (BI->isConditional()) {
2859 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2861 if (BI->getSuccessor(0) != Dst)
2862 for (unsigned part = 0; part < UF; ++part)
2863 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2865 for (unsigned part = 0; part < UF; ++part)
2866 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2868 MaskCache[Edge] = EdgeMask;
2872 MaskCache[Edge] = SrcMask;
2876 InnerLoopVectorizer::VectorParts
2877 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2878 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2880 // Loop incoming mask is all-one.
2881 if (OrigLoop->getHeader() == BB) {
2882 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2883 return getVectorValue(C);
2886 // This is the block mask. We OR all incoming edges, and with zero.
2887 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2888 VectorParts BlockMask = getVectorValue(Zero);
2891 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2892 VectorParts EM = createEdgeMask(*it, BB);
2893 for (unsigned part = 0; part < UF; ++part)
2894 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2900 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2901 InnerLoopVectorizer::VectorParts &Entry,
2902 unsigned UF, unsigned VF, PhiVector *PV) {
2903 PHINode* P = cast<PHINode>(PN);
2904 // Handle reduction variables:
2905 if (Legal->getReductionVars()->count(P)) {
2906 for (unsigned part = 0; part < UF; ++part) {
2907 // This is phase one of vectorizing PHIs.
2908 Type *VecTy = (VF == 1) ? PN->getType() :
2909 VectorType::get(PN->getType(), VF);
2910 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2911 LoopVectorBody.back()-> getFirstInsertionPt());
2917 setDebugLocFromInst(Builder, P);
2918 // Check for PHI nodes that are lowered to vector selects.
2919 if (P->getParent() != OrigLoop->getHeader()) {
2920 // We know that all PHIs in non-header blocks are converted into
2921 // selects, so we don't have to worry about the insertion order and we
2922 // can just use the builder.
2923 // At this point we generate the predication tree. There may be
2924 // duplications since this is a simple recursive scan, but future
2925 // optimizations will clean it up.
2927 unsigned NumIncoming = P->getNumIncomingValues();
2929 // Generate a sequence of selects of the form:
2930 // SELECT(Mask3, In3,
2931 // SELECT(Mask2, In2,
2933 for (unsigned In = 0; In < NumIncoming; In++) {
2934 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2936 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2938 for (unsigned part = 0; part < UF; ++part) {
2939 // We might have single edge PHIs (blocks) - use an identity
2940 // 'select' for the first PHI operand.
2942 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2945 // Select between the current value and the previous incoming edge
2946 // based on the incoming mask.
2947 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2948 Entry[part], "predphi");
2954 // This PHINode must be an induction variable.
2955 // Make sure that we know about it.
2956 assert(Legal->getInductionVars()->count(P) &&
2957 "Not an induction variable");
2959 LoopVectorizationLegality::InductionInfo II =
2960 Legal->getInductionVars()->lookup(P);
2963 case LoopVectorizationLegality::IK_NoInduction:
2964 llvm_unreachable("Unknown induction");
2965 case LoopVectorizationLegality::IK_IntInduction: {
2966 assert(P->getType() == II.StartValue->getType() && "Types must match");
2967 Type *PhiTy = P->getType();
2969 if (P == OldInduction) {
2970 // Handle the canonical induction variable. We might have had to
2972 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2974 // Handle other induction variables that are now based on the
2976 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2978 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2979 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2982 Broadcasted = getBroadcastInstrs(Broadcasted);
2983 // After broadcasting the induction variable we need to make the vector
2984 // consecutive by adding 0, 1, 2, etc.
2985 for (unsigned part = 0; part < UF; ++part)
2986 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2989 case LoopVectorizationLegality::IK_ReverseIntInduction:
2990 case LoopVectorizationLegality::IK_PtrInduction:
2991 case LoopVectorizationLegality::IK_ReversePtrInduction:
2992 // Handle reverse integer and pointer inductions.
2993 Value *StartIdx = ExtendedIdx;
2994 // This is the normalized GEP that starts counting at zero.
2995 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2998 // Handle the reverse integer induction variable case.
2999 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3000 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3001 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3003 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3006 // This is a new value so do not hoist it out.
3007 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3008 // After broadcasting the induction variable we need to make the
3009 // vector consecutive by adding ... -3, -2, -1, 0.
3010 for (unsigned part = 0; part < UF; ++part)
3011 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3016 // Handle the pointer induction variable case.
3017 assert(P->getType()->isPointerTy() && "Unexpected type.");
3019 // Is this a reverse induction ptr or a consecutive induction ptr.
3020 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3023 // This is the vector of results. Notice that we don't generate
3024 // vector geps because scalar geps result in better code.
3025 for (unsigned part = 0; part < UF; ++part) {
3027 int EltIndex = (part) * (Reverse ? -1 : 1);
3028 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3031 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3033 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3035 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3037 Entry[part] = SclrGep;
3041 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3042 for (unsigned int i = 0; i < VF; ++i) {
3043 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3044 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3047 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3049 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3051 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3053 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3054 Builder.getInt32(i),
3057 Entry[part] = VecVal;
3063 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3064 // For each instruction in the old loop.
3065 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3066 VectorParts &Entry = WidenMap.get(it);
3067 switch (it->getOpcode()) {
3068 case Instruction::Br:
3069 // Nothing to do for PHIs and BR, since we already took care of the
3070 // loop control flow instructions.
3072 case Instruction::PHI:{
3073 // Vectorize PHINodes.
3074 widenPHIInstruction(it, Entry, UF, VF, PV);
3078 case Instruction::Add:
3079 case Instruction::FAdd:
3080 case Instruction::Sub:
3081 case Instruction::FSub:
3082 case Instruction::Mul:
3083 case Instruction::FMul:
3084 case Instruction::UDiv:
3085 case Instruction::SDiv:
3086 case Instruction::FDiv:
3087 case Instruction::URem:
3088 case Instruction::SRem:
3089 case Instruction::FRem:
3090 case Instruction::Shl:
3091 case Instruction::LShr:
3092 case Instruction::AShr:
3093 case Instruction::And:
3094 case Instruction::Or:
3095 case Instruction::Xor: {
3096 // Just widen binops.
3097 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3098 setDebugLocFromInst(Builder, BinOp);
3099 VectorParts &A = getVectorValue(it->getOperand(0));
3100 VectorParts &B = getVectorValue(it->getOperand(1));
3102 // Use this vector value for all users of the original instruction.
3103 for (unsigned Part = 0; Part < UF; ++Part) {
3104 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3106 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3107 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3108 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3109 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3110 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3112 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3113 VecOp->setIsExact(BinOp->isExact());
3115 // Copy the fast-math flags.
3116 if (VecOp && isa<FPMathOperator>(V))
3117 VecOp->setFastMathFlags(it->getFastMathFlags());
3123 case Instruction::Select: {
3125 // If the selector is loop invariant we can create a select
3126 // instruction with a scalar condition. Otherwise, use vector-select.
3127 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3129 setDebugLocFromInst(Builder, it);
3131 // The condition can be loop invariant but still defined inside the
3132 // loop. This means that we can't just use the original 'cond' value.
3133 // We have to take the 'vectorized' value and pick the first lane.
3134 // Instcombine will make this a no-op.
3135 VectorParts &Cond = getVectorValue(it->getOperand(0));
3136 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3137 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3139 Value *ScalarCond = (VF == 1) ? Cond[0] :
3140 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3142 for (unsigned Part = 0; Part < UF; ++Part) {
3143 Entry[Part] = Builder.CreateSelect(
3144 InvariantCond ? ScalarCond : Cond[Part],
3151 case Instruction::ICmp:
3152 case Instruction::FCmp: {
3153 // Widen compares. Generate vector compares.
3154 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3155 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3156 setDebugLocFromInst(Builder, it);
3157 VectorParts &A = getVectorValue(it->getOperand(0));
3158 VectorParts &B = getVectorValue(it->getOperand(1));
3159 for (unsigned Part = 0; Part < UF; ++Part) {
3162 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3164 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3170 case Instruction::Store:
3171 case Instruction::Load:
3172 vectorizeMemoryInstruction(it);
3174 case Instruction::ZExt:
3175 case Instruction::SExt:
3176 case Instruction::FPToUI:
3177 case Instruction::FPToSI:
3178 case Instruction::FPExt:
3179 case Instruction::PtrToInt:
3180 case Instruction::IntToPtr:
3181 case Instruction::SIToFP:
3182 case Instruction::UIToFP:
3183 case Instruction::Trunc:
3184 case Instruction::FPTrunc:
3185 case Instruction::BitCast: {
3186 CastInst *CI = dyn_cast<CastInst>(it);
3187 setDebugLocFromInst(Builder, it);
3188 /// Optimize the special case where the source is the induction
3189 /// variable. Notice that we can only optimize the 'trunc' case
3190 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3191 /// c. other casts depend on pointer size.
3192 if (CI->getOperand(0) == OldInduction &&
3193 it->getOpcode() == Instruction::Trunc) {
3194 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3196 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3197 for (unsigned Part = 0; Part < UF; ++Part)
3198 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3201 /// Vectorize casts.
3202 Type *DestTy = (VF == 1) ? CI->getType() :
3203 VectorType::get(CI->getType(), VF);
3205 VectorParts &A = getVectorValue(it->getOperand(0));
3206 for (unsigned Part = 0; Part < UF; ++Part)
3207 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3211 case Instruction::Call: {
3212 // Ignore dbg intrinsics.
3213 if (isa<DbgInfoIntrinsic>(it))
3215 setDebugLocFromInst(Builder, it);
3217 Module *M = BB->getParent()->getParent();
3218 CallInst *CI = cast<CallInst>(it);
3219 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3220 assert(ID && "Not an intrinsic call!");
3222 case Intrinsic::lifetime_end:
3223 case Intrinsic::lifetime_start:
3224 scalarizeInstruction(it);
3227 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3228 for (unsigned Part = 0; Part < UF; ++Part) {
3229 SmallVector<Value *, 4> Args;
3230 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3231 if (HasScalarOpd && i == 1) {
3232 Args.push_back(CI->getArgOperand(i));
3235 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3236 Args.push_back(Arg[Part]);
3238 Type *Tys[] = {CI->getType()};
3240 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3242 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3243 Entry[Part] = Builder.CreateCall(F, Args);
3251 // All other instructions are unsupported. Scalarize them.
3252 scalarizeInstruction(it);
3255 }// end of for_each instr.
3258 void InnerLoopVectorizer::updateAnalysis() {
3259 // Forget the original basic block.
3260 SE->forgetLoop(OrigLoop);
3262 // Update the dominator tree information.
3263 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3264 "Entry does not dominate exit.");
3266 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3267 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3268 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3270 // Due to if predication of stores we might create a sequence of "if(pred)
3271 // a[i] = ...; " blocks.
3272 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3274 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3275 else if (isPredicatedBlock(i)) {
3276 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3278 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3282 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3283 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3284 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3285 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3287 DEBUG(DT->verifyDomTree());
3290 /// \brief Check whether it is safe to if-convert this phi node.
3292 /// Phi nodes with constant expressions that can trap are not safe to if
3294 static bool canIfConvertPHINodes(BasicBlock *BB) {
3295 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3296 PHINode *Phi = dyn_cast<PHINode>(I);
3299 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3300 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3307 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3308 if (!EnableIfConversion) {
3309 emitAnalysis(Report() << "if-conversion is disabled");
3313 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3315 // A list of pointers that we can safely read and write to.
3316 SmallPtrSet<Value *, 8> SafePointes;
3318 // Collect safe addresses.
3319 for (Loop::block_iterator BI = TheLoop->block_begin(),
3320 BE = TheLoop->block_end(); BI != BE; ++BI) {
3321 BasicBlock *BB = *BI;
3323 if (blockNeedsPredication(BB))
3326 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3327 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3328 SafePointes.insert(LI->getPointerOperand());
3329 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3330 SafePointes.insert(SI->getPointerOperand());
3334 // Collect the blocks that need predication.
3335 BasicBlock *Header = TheLoop->getHeader();
3336 for (Loop::block_iterator BI = TheLoop->block_begin(),
3337 BE = TheLoop->block_end(); BI != BE; ++BI) {
3338 BasicBlock *BB = *BI;
3340 // We don't support switch statements inside loops.
3341 if (!isa<BranchInst>(BB->getTerminator())) {
3342 emitAnalysis(Report(BB->getTerminator())
3343 << "loop contains a switch statement");
3347 // We must be able to predicate all blocks that need to be predicated.
3348 if (blockNeedsPredication(BB)) {
3349 if (!blockCanBePredicated(BB, SafePointes)) {
3350 emitAnalysis(Report(BB->getTerminator())
3351 << "control flow cannot be substituted for a select");
3354 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3355 emitAnalysis(Report(BB->getTerminator())
3356 << "control flow cannot be substituted for a select");
3361 // We can if-convert this loop.
3365 bool LoopVectorizationLegality::canVectorize() {
3366 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3367 // be canonicalized.
3368 if (!TheLoop->getLoopPreheader()) {
3370 Report() << "loop control flow is not understood by vectorizer");
3374 // We can only vectorize innermost loops.
3375 if (TheLoop->getSubLoopsVector().size()) {
3376 emitAnalysis(Report() << "loop is not the innermost loop");
3380 // We must have a single backedge.
3381 if (TheLoop->getNumBackEdges() != 1) {
3383 Report() << "loop control flow is not understood by vectorizer");
3387 // We must have a single exiting block.
3388 if (!TheLoop->getExitingBlock()) {
3390 Report() << "loop control flow is not understood by vectorizer");
3394 // We need to have a loop header.
3395 DEBUG(dbgs() << "LV: Found a loop: " <<
3396 TheLoop->getHeader()->getName() << '\n');
3398 // Check if we can if-convert non-single-bb loops.
3399 unsigned NumBlocks = TheLoop->getNumBlocks();
3400 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3401 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3405 // ScalarEvolution needs to be able to find the exit count.
3406 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3407 if (ExitCount == SE->getCouldNotCompute()) {
3408 emitAnalysis(Report() << "could not determine number of loop iterations");
3409 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3413 // Check if we can vectorize the instructions and CFG in this loop.
3414 if (!canVectorizeInstrs()) {
3415 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3419 // Go over each instruction and look at memory deps.
3420 if (!canVectorizeMemory()) {
3421 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3425 // Collect all of the variables that remain uniform after vectorization.
3426 collectLoopUniforms();
3428 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3429 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3432 // Okay! We can vectorize. At this point we don't have any other mem analysis
3433 // which may limit our maximum vectorization factor, so just return true with
3438 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3439 if (Ty->isPointerTy())
3440 return DL.getIntPtrType(Ty);
3442 // It is possible that char's or short's overflow when we ask for the loop's
3443 // trip count, work around this by changing the type size.
3444 if (Ty->getScalarSizeInBits() < 32)
3445 return Type::getInt32Ty(Ty->getContext());
3450 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3451 Ty0 = convertPointerToIntegerType(DL, Ty0);
3452 Ty1 = convertPointerToIntegerType(DL, Ty1);
3453 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3458 /// \brief Check that the instruction has outside loop users and is not an
3459 /// identified reduction variable.
3460 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3461 SmallPtrSet<Value *, 4> &Reductions) {
3462 // Reduction instructions are allowed to have exit users. All other
3463 // instructions must not have external users.
3464 if (!Reductions.count(Inst))
3465 //Check that all of the users of the loop are inside the BB.
3466 for (User *U : Inst->users()) {
3467 Instruction *UI = cast<Instruction>(U);
3468 // This user may be a reduction exit value.
3469 if (!TheLoop->contains(UI)) {
3470 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3477 bool LoopVectorizationLegality::canVectorizeInstrs() {
3478 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3479 BasicBlock *Header = TheLoop->getHeader();
3481 // Look for the attribute signaling the absence of NaNs.
3482 Function &F = *Header->getParent();
3483 if (F.hasFnAttribute("no-nans-fp-math"))
3484 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3485 AttributeSet::FunctionIndex,
3486 "no-nans-fp-math").getValueAsString() == "true";
3488 // For each block in the loop.
3489 for (Loop::block_iterator bb = TheLoop->block_begin(),
3490 be = TheLoop->block_end(); bb != be; ++bb) {
3492 // Scan the instructions in the block and look for hazards.
3493 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3496 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3497 Type *PhiTy = Phi->getType();
3498 // Check that this PHI type is allowed.
3499 if (!PhiTy->isIntegerTy() &&
3500 !PhiTy->isFloatingPointTy() &&
3501 !PhiTy->isPointerTy()) {
3502 emitAnalysis(Report(it)
3503 << "loop control flow is not understood by vectorizer");
3504 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3508 // If this PHINode is not in the header block, then we know that we
3509 // can convert it to select during if-conversion. No need to check if
3510 // the PHIs in this block are induction or reduction variables.
3511 if (*bb != Header) {
3512 // Check that this instruction has no outside users or is an
3513 // identified reduction value with an outside user.
3514 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3516 emitAnalysis(Report(it) << "value that could not be identified as "
3517 "reduction is used outside the loop");
3521 // We only allow if-converted PHIs with more than two incoming values.
3522 if (Phi->getNumIncomingValues() != 2) {
3523 emitAnalysis(Report(it)
3524 << "control flow not understood by vectorizer");
3525 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3529 // This is the value coming from the preheader.
3530 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3531 // Check if this is an induction variable.
3532 InductionKind IK = isInductionVariable(Phi);
3534 if (IK_NoInduction != IK) {
3535 // Get the widest type.
3537 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3539 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3541 // Int inductions are special because we only allow one IV.
3542 if (IK == IK_IntInduction) {
3543 // Use the phi node with the widest type as induction. Use the last
3544 // one if there are multiple (no good reason for doing this other
3545 // than it is expedient).
3546 if (!Induction || PhiTy == WidestIndTy)
3550 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3551 Inductions[Phi] = InductionInfo(StartValue, IK);
3553 // Until we explicitly handle the case of an induction variable with
3554 // an outside loop user we have to give up vectorizing this loop.
3555 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3556 emitAnalysis(Report(it) << "use of induction value outside of the "
3557 "loop is not handled by vectorizer");
3564 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3565 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3568 if (AddReductionVar(Phi, RK_IntegerMult)) {
3569 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3572 if (AddReductionVar(Phi, RK_IntegerOr)) {
3573 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3576 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3577 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3580 if (AddReductionVar(Phi, RK_IntegerXor)) {
3581 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3584 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3585 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3588 if (AddReductionVar(Phi, RK_FloatMult)) {
3589 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3592 if (AddReductionVar(Phi, RK_FloatAdd)) {
3593 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3596 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3597 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3602 emitAnalysis(Report(it) << "unvectorizable operation");
3603 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3605 }// end of PHI handling
3607 // We still don't handle functions. However, we can ignore dbg intrinsic
3608 // calls and we do handle certain intrinsic and libm functions.
3609 CallInst *CI = dyn_cast<CallInst>(it);
3610 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3611 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3612 DEBUG(dbgs() << "LV: Found a call site.\n");
3616 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3617 // second argument is the same (i.e. loop invariant)
3619 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3620 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3621 emitAnalysis(Report(it)
3622 << "intrinsic instruction cannot be vectorized");
3623 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3628 // Check that the instruction return type is vectorizable.
3629 // Also, we can't vectorize extractelement instructions.
3630 if ((!VectorType::isValidElementType(it->getType()) &&
3631 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3632 emitAnalysis(Report(it)
3633 << "instruction return type cannot be vectorized");
3634 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3638 // Check that the stored type is vectorizable.
3639 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3640 Type *T = ST->getValueOperand()->getType();
3641 if (!VectorType::isValidElementType(T)) {
3642 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3645 if (EnableMemAccessVersioning)
3646 collectStridedAcccess(ST);
3649 if (EnableMemAccessVersioning)
3650 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3651 collectStridedAcccess(LI);
3653 // Reduction instructions are allowed to have exit users.
3654 // All other instructions must not have external users.
3655 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3656 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3665 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3666 if (Inductions.empty()) {
3667 emitAnalysis(Report()
3668 << "loop induction variable could not be identified");
3676 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3677 /// return the induction operand of the gep pointer.
3678 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3679 const DataLayout *DL, Loop *Lp) {
3680 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3684 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3686 // Check that all of the gep indices are uniform except for our induction
3688 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3689 if (i != InductionOperand &&
3690 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3692 return GEP->getOperand(InductionOperand);
3695 ///\brief Look for a cast use of the passed value.
3696 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3697 Value *UniqueCast = nullptr;
3698 for (User *U : Ptr->users()) {
3699 CastInst *CI = dyn_cast<CastInst>(U);
3700 if (CI && CI->getType() == Ty) {
3710 ///\brief Get the stride of a pointer access in a loop.
3711 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3712 /// pointer to the Value, or null otherwise.
3713 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3714 const DataLayout *DL, Loop *Lp) {
3715 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3716 if (!PtrTy || PtrTy->isAggregateType())
3719 // Try to remove a gep instruction to make the pointer (actually index at this
3720 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3721 // pointer, otherwise, we are analyzing the index.
3722 Value *OrigPtr = Ptr;
3724 // The size of the pointer access.
3725 int64_t PtrAccessSize = 1;
3727 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3728 const SCEV *V = SE->getSCEV(Ptr);
3732 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3733 V = C->getOperand();
3735 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3739 V = S->getStepRecurrence(*SE);
3743 // Strip off the size of access multiplication if we are still analyzing the
3745 if (OrigPtr == Ptr) {
3746 DL->getTypeAllocSize(PtrTy->getElementType());
3747 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3748 if (M->getOperand(0)->getSCEVType() != scConstant)
3751 const APInt &APStepVal =
3752 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3754 // Huge step value - give up.
3755 if (APStepVal.getBitWidth() > 64)
3758 int64_t StepVal = APStepVal.getSExtValue();
3759 if (PtrAccessSize != StepVal)
3761 V = M->getOperand(1);
3766 Type *StripedOffRecurrenceCast = nullptr;
3767 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3768 StripedOffRecurrenceCast = C->getType();
3769 V = C->getOperand();
3772 // Look for the loop invariant symbolic value.
3773 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3777 Value *Stride = U->getValue();
3778 if (!Lp->isLoopInvariant(Stride))
3781 // If we have stripped off the recurrence cast we have to make sure that we
3782 // return the value that is used in this loop so that we can replace it later.
3783 if (StripedOffRecurrenceCast)
3784 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3789 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3790 Value *Ptr = nullptr;
3791 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3792 Ptr = LI->getPointerOperand();
3793 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3794 Ptr = SI->getPointerOperand();
3798 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3802 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3803 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3804 Strides[Ptr] = Stride;
3805 StrideSet.insert(Stride);
3808 void LoopVectorizationLegality::collectLoopUniforms() {
3809 // We now know that the loop is vectorizable!
3810 // Collect variables that will remain uniform after vectorization.
3811 std::vector<Value*> Worklist;
3812 BasicBlock *Latch = TheLoop->getLoopLatch();
3814 // Start with the conditional branch and walk up the block.
3815 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3817 // Also add all consecutive pointer values; these values will be uniform
3818 // after vectorization (and subsequent cleanup) and, until revectorization is
3819 // supported, all dependencies must also be uniform.
3820 for (Loop::block_iterator B = TheLoop->block_begin(),
3821 BE = TheLoop->block_end(); B != BE; ++B)
3822 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3824 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3825 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3827 while (Worklist.size()) {
3828 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3829 Worklist.pop_back();
3831 // Look at instructions inside this loop.
3832 // Stop when reaching PHI nodes.
3833 // TODO: we need to follow values all over the loop, not only in this block.
3834 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3837 // This is a known uniform.
3840 // Insert all operands.
3841 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3846 /// \brief Analyses memory accesses in a loop.
3848 /// Checks whether run time pointer checks are needed and builds sets for data
3849 /// dependence checking.
3850 class AccessAnalysis {
3852 /// \brief Read or write access location.
3853 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3854 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3856 /// \brief Set of potential dependent memory accesses.
3857 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3859 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3860 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3861 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3863 /// \brief Register a load and whether it is only read from.
3864 void addLoad(Value *Ptr, bool IsReadOnly) {
3865 Accesses.insert(MemAccessInfo(Ptr, false));
3867 ReadOnlyPtr.insert(Ptr);
3870 /// \brief Register a store.
3871 void addStore(Value *Ptr) {
3872 Accesses.insert(MemAccessInfo(Ptr, true));
3875 /// \brief Check whether we can check the pointers at runtime for
3876 /// non-intersection.
3877 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3878 unsigned &NumComparisons, ScalarEvolution *SE,
3879 Loop *TheLoop, ValueToValueMap &Strides,
3880 bool ShouldCheckStride = false);
3882 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3883 /// and builds sets of dependent accesses.
3884 void buildDependenceSets() {
3885 // Process read-write pointers first.
3886 processMemAccesses(false);
3887 // Next, process read pointers.
3888 processMemAccesses(true);
3891 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3893 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3894 void resetDepChecks() { CheckDeps.clear(); }
3896 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3899 typedef SetVector<MemAccessInfo> PtrAccessSet;
3900 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3902 /// \brief Go over all memory access or only the deferred ones if
3903 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3904 /// and build sets of dependency check candidates.
3905 void processMemAccesses(bool UseDeferred);
3907 /// Set of all accesses.
3908 PtrAccessSet Accesses;
3910 /// Set of access to check after all writes have been processed.
3911 PtrAccessSet DeferredAccesses;
3913 /// Map of pointers to last access encountered.
3914 UnderlyingObjToAccessMap ObjToLastAccess;
3916 /// Set of accesses that need a further dependence check.
3917 MemAccessInfoSet CheckDeps;
3919 /// Set of pointers that are read only.
3920 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3922 /// Set of underlying objects already written to.
3923 SmallPtrSet<Value*, 16> WriteObjects;
3925 const DataLayout *DL;
3927 /// Sets of potentially dependent accesses - members of one set share an
3928 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3929 /// dependence check.
3930 DepCandidates &DepCands;
3932 bool AreAllWritesIdentified;
3933 bool AreAllReadsIdentified;
3934 bool IsRTCheckNeeded;
3937 } // end anonymous namespace
3939 /// \brief Check whether a pointer can participate in a runtime bounds check.
3940 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3942 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3943 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3947 return AR->isAffine();
3950 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3951 /// the address space.
3952 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3953 const Loop *Lp, ValueToValueMap &StridesMap);
3955 bool AccessAnalysis::canCheckPtrAtRT(
3956 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3957 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3958 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3959 // Find pointers with computable bounds. We are going to use this information
3960 // to place a runtime bound check.
3961 unsigned NumReadPtrChecks = 0;
3962 unsigned NumWritePtrChecks = 0;
3963 bool CanDoRT = true;
3965 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3966 // We assign consecutive id to access from different dependence sets.
3967 // Accesses within the same set don't need a runtime check.
3968 unsigned RunningDepId = 1;
3969 DenseMap<Value *, unsigned> DepSetId;
3971 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3973 const MemAccessInfo &Access = *AI;
3974 Value *Ptr = Access.getPointer();
3975 bool IsWrite = Access.getInt();
3977 // Just add write checks if we have both.
3978 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3982 ++NumWritePtrChecks;
3986 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3987 // When we run after a failing dependency check we have to make sure we
3988 // don't have wrapping pointers.
3989 (!ShouldCheckStride ||
3990 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3991 // The id of the dependence set.
3994 if (IsDepCheckNeeded) {
3995 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3996 unsigned &LeaderId = DepSetId[Leader];
3998 LeaderId = RunningDepId++;
4001 // Each access has its own dependence set.
4002 DepId = RunningDepId++;
4004 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
4006 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4012 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4013 NumComparisons = 0; // Only one dependence set.
4015 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
4016 NumWritePtrChecks - 1));
4019 // If the pointers that we would use for the bounds comparison have different
4020 // address spaces, assume the values aren't directly comparable, so we can't
4021 // use them for the runtime check. We also have to assume they could
4022 // overlap. In the future there should be metadata for whether address spaces
4024 unsigned NumPointers = RtCheck.Pointers.size();
4025 for (unsigned i = 0; i < NumPointers; ++i) {
4026 for (unsigned j = i + 1; j < NumPointers; ++j) {
4027 // Only need to check pointers between two different dependency sets.
4028 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4031 Value *PtrI = RtCheck.Pointers[i];
4032 Value *PtrJ = RtCheck.Pointers[j];
4034 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4035 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4037 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4038 " different address spaces\n");
4047 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
4048 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
4051 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
4052 // We process the set twice: first we process read-write pointers, last we
4053 // process read-only pointers. This allows us to skip dependence tests for
4054 // read-only pointers.
4056 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4057 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
4058 const MemAccessInfo &Access = *AI;
4059 Value *Ptr = Access.getPointer();
4060 bool IsWrite = Access.getInt();
4062 DepCands.insert(Access);
4064 // Memorize read-only pointers for later processing and skip them in the
4065 // first round (they need to be checked after we have seen all write
4066 // pointers). Note: we also mark pointer that are not consecutive as
4067 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
4068 // second check for "!IsWrite".
4069 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4070 if (!UseDeferred && IsReadOnlyPtr) {
4071 DeferredAccesses.insert(Access);
4075 bool NeedDepCheck = false;
4076 // Check whether there is the possibility of dependency because of
4077 // underlying objects being the same.
4078 typedef SmallVector<Value*, 16> ValueVector;
4079 ValueVector TempObjects;
4080 GetUnderlyingObjects(Ptr, TempObjects, DL);
4081 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
4083 Value *UnderlyingObj = *UI;
4085 // If this is a write then it needs to be an identified object. If this a
4086 // read and all writes (so far) are identified function scope objects we
4087 // don't need an identified underlying object but only an Argument (the
4088 // next write is going to invalidate this assumption if it is
4090 // This is a micro-optimization for the case where all writes are
4091 // identified and we have one argument pointer.
4092 // Otherwise, we do need a runtime check.
4093 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
4094 (!IsWrite && (!AreAllWritesIdentified ||
4095 !isa<Argument>(UnderlyingObj)) &&
4096 !isIdentifiedObject(UnderlyingObj))) {
4097 DEBUG(dbgs() << "LV: Found an unidentified " <<
4098 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
4100 IsRTCheckNeeded = (IsRTCheckNeeded ||
4101 !isIdentifiedObject(UnderlyingObj) ||
4102 !AreAllReadsIdentified);
4105 AreAllWritesIdentified = false;
4107 AreAllReadsIdentified = false;
4110 // If this is a write - check other reads and writes for conflicts. If
4111 // this is a read only check other writes for conflicts (but only if there
4112 // is no other write to the ptr - this is an optimization to catch "a[i] =
4113 // a[i] + " without having to do a dependence check).
4114 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
4115 NeedDepCheck = true;
4118 WriteObjects.insert(UnderlyingObj);
4120 // Create sets of pointers connected by shared underlying objects.
4121 UnderlyingObjToAccessMap::iterator Prev =
4122 ObjToLastAccess.find(UnderlyingObj);
4123 if (Prev != ObjToLastAccess.end())
4124 DepCands.unionSets(Access, Prev->second);
4126 ObjToLastAccess[UnderlyingObj] = Access;
4130 CheckDeps.insert(Access);
4135 /// \brief Checks memory dependences among accesses to the same underlying
4136 /// object to determine whether there vectorization is legal or not (and at
4137 /// which vectorization factor).
4139 /// This class works under the assumption that we already checked that memory
4140 /// locations with different underlying pointers are "must-not alias".
4141 /// We use the ScalarEvolution framework to symbolically evalutate access
4142 /// functions pairs. Since we currently don't restructure the loop we can rely
4143 /// on the program order of memory accesses to determine their safety.
4144 /// At the moment we will only deem accesses as safe for:
4145 /// * A negative constant distance assuming program order.
4147 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4148 /// a[i] = tmp; y = a[i];
4150 /// The latter case is safe because later checks guarantuee that there can't
4151 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4152 /// the same variable: a header phi can only be an induction or a reduction, a
4153 /// reduction can't have a memory sink, an induction can't have a memory
4154 /// source). This is important and must not be violated (or we have to
4155 /// resort to checking for cycles through memory).
4157 /// * A positive constant distance assuming program order that is bigger
4158 /// than the biggest memory access.
4160 /// tmp = a[i] OR b[i] = x
4161 /// a[i+2] = tmp y = b[i+2];
4163 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4165 /// * Zero distances and all accesses have the same size.
4167 class MemoryDepChecker {
4169 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4170 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4172 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4173 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4174 ShouldRetryWithRuntimeCheck(false) {}
4176 /// \brief Register the location (instructions are given increasing numbers)
4177 /// of a write access.
4178 void addAccess(StoreInst *SI) {
4179 Value *Ptr = SI->getPointerOperand();
4180 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4181 InstMap.push_back(SI);
4185 /// \brief Register the location (instructions are given increasing numbers)
4186 /// of a write access.
4187 void addAccess(LoadInst *LI) {
4188 Value *Ptr = LI->getPointerOperand();
4189 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4190 InstMap.push_back(LI);
4194 /// \brief Check whether the dependencies between the accesses are safe.
4196 /// Only checks sets with elements in \p CheckDeps.
4197 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4198 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4200 /// \brief The maximum number of bytes of a vector register we can vectorize
4201 /// the accesses safely with.
4202 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4204 /// \brief In same cases when the dependency check fails we can still
4205 /// vectorize the loop with a dynamic array access check.
4206 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4209 ScalarEvolution *SE;
4210 const DataLayout *DL;
4211 const Loop *InnermostLoop;
4213 /// \brief Maps access locations (ptr, read/write) to program order.
4214 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4216 /// \brief Memory access instructions in program order.
4217 SmallVector<Instruction *, 16> InstMap;
4219 /// \brief The program order index to be used for the next instruction.
4222 // We can access this many bytes in parallel safely.
4223 unsigned MaxSafeDepDistBytes;
4225 /// \brief If we see a non-constant dependence distance we can still try to
4226 /// vectorize this loop with runtime checks.
4227 bool ShouldRetryWithRuntimeCheck;
4229 /// \brief Check whether there is a plausible dependence between the two
4232 /// Access \p A must happen before \p B in program order. The two indices
4233 /// identify the index into the program order map.
4235 /// This function checks whether there is a plausible dependence (or the
4236 /// absence of such can't be proved) between the two accesses. If there is a
4237 /// plausible dependence but the dependence distance is bigger than one
4238 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4239 /// distance is smaller than any other distance encountered so far).
4240 /// Otherwise, this function returns true signaling a possible dependence.
4241 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4242 const MemAccessInfo &B, unsigned BIdx,
4243 ValueToValueMap &Strides);
4245 /// \brief Check whether the data dependence could prevent store-load
4247 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4250 } // end anonymous namespace
4252 static bool isInBoundsGep(Value *Ptr) {
4253 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4254 return GEP->isInBounds();
4258 /// \brief Check whether the access through \p Ptr has a constant stride.
4259 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4260 const Loop *Lp, ValueToValueMap &StridesMap) {
4261 const Type *Ty = Ptr->getType();
4262 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4264 // Make sure that the pointer does not point to aggregate types.
4265 const PointerType *PtrTy = cast<PointerType>(Ty);
4266 if (PtrTy->getElementType()->isAggregateType()) {
4267 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4272 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4274 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4276 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4277 << *Ptr << " SCEV: " << *PtrScev << "\n");
4281 // The accesss function must stride over the innermost loop.
4282 if (Lp != AR->getLoop()) {
4283 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4284 *Ptr << " SCEV: " << *PtrScev << "\n");
4287 // The address calculation must not wrap. Otherwise, a dependence could be
4289 // An inbounds getelementptr that is a AddRec with a unit stride
4290 // cannot wrap per definition. The unit stride requirement is checked later.
4291 // An getelementptr without an inbounds attribute and unit stride would have
4292 // to access the pointer value "0" which is undefined behavior in address
4293 // space 0, therefore we can also vectorize this case.
4294 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4295 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4296 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4297 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4298 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4299 << *Ptr << " SCEV: " << *PtrScev << "\n");
4303 // Check the step is constant.
4304 const SCEV *Step = AR->getStepRecurrence(*SE);
4306 // Calculate the pointer stride and check if it is consecutive.
4307 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4309 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4310 " SCEV: " << *PtrScev << "\n");
4314 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4315 const APInt &APStepVal = C->getValue()->getValue();
4317 // Huge step value - give up.
4318 if (APStepVal.getBitWidth() > 64)
4321 int64_t StepVal = APStepVal.getSExtValue();
4324 int64_t Stride = StepVal / Size;
4325 int64_t Rem = StepVal % Size;
4329 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4330 // know we can't "wrap around the address space". In case of address space
4331 // zero we know that this won't happen without triggering undefined behavior.
4332 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4333 Stride != 1 && Stride != -1)
4339 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4340 unsigned TypeByteSize) {
4341 // If loads occur at a distance that is not a multiple of a feasible vector
4342 // factor store-load forwarding does not take place.
4343 // Positive dependences might cause troubles because vectorizing them might
4344 // prevent store-load forwarding making vectorized code run a lot slower.
4345 // a[i] = a[i-3] ^ a[i-8];
4346 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4347 // hence on your typical architecture store-load forwarding does not take
4348 // place. Vectorizing in such cases does not make sense.
4349 // Store-load forwarding distance.
4350 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4351 // Maximum vector factor.
4352 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4353 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4354 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4356 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4358 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4359 MaxVFWithoutSLForwardIssues = (vf >>=1);
4364 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4365 DEBUG(dbgs() << "LV: Distance " << Distance <<
4366 " that could cause a store-load forwarding conflict\n");
4370 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4371 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4372 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4376 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4377 const MemAccessInfo &B, unsigned BIdx,
4378 ValueToValueMap &Strides) {
4379 assert (AIdx < BIdx && "Must pass arguments in program order");
4381 Value *APtr = A.getPointer();
4382 Value *BPtr = B.getPointer();
4383 bool AIsWrite = A.getInt();
4384 bool BIsWrite = B.getInt();
4386 // Two reads are independent.
4387 if (!AIsWrite && !BIsWrite)
4390 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4391 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4393 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4394 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4396 const SCEV *Src = AScev;
4397 const SCEV *Sink = BScev;
4399 // If the induction step is negative we have to invert source and sink of the
4401 if (StrideAPtr < 0) {
4404 std::swap(APtr, BPtr);
4405 std::swap(Src, Sink);
4406 std::swap(AIsWrite, BIsWrite);
4407 std::swap(AIdx, BIdx);
4408 std::swap(StrideAPtr, StrideBPtr);
4411 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4413 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4414 << "(Induction step: " << StrideAPtr << ")\n");
4415 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4416 << *InstMap[BIdx] << ": " << *Dist << "\n");
4418 // Need consecutive accesses. We don't want to vectorize
4419 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4420 // the address space.
4421 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4422 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4426 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4428 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4429 ShouldRetryWithRuntimeCheck = true;
4433 Type *ATy = APtr->getType()->getPointerElementType();
4434 Type *BTy = BPtr->getType()->getPointerElementType();
4435 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4437 // Negative distances are not plausible dependencies.
4438 const APInt &Val = C->getValue()->getValue();
4439 if (Val.isNegative()) {
4440 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4441 if (IsTrueDataDependence &&
4442 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4446 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4450 // Write to the same location with the same size.
4451 // Could be improved to assert type sizes are the same (i32 == float, etc).
4455 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4459 assert(Val.isStrictlyPositive() && "Expect a positive value");
4461 // Positive distance bigger than max vectorization factor.
4464 "LV: ReadWrite-Write positive dependency with different types\n");
4468 unsigned Distance = (unsigned) Val.getZExtValue();
4470 // Bail out early if passed-in parameters make vectorization not feasible.
4471 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4472 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4474 // The distance must be bigger than the size needed for a vectorized version
4475 // of the operation and the size of the vectorized operation must not be
4476 // bigger than the currrent maximum size.
4477 if (Distance < 2*TypeByteSize ||
4478 2*TypeByteSize > MaxSafeDepDistBytes ||
4479 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4480 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4481 << Val.getSExtValue() << '\n');
4485 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4486 Distance : MaxSafeDepDistBytes;
4488 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4489 if (IsTrueDataDependence &&
4490 couldPreventStoreLoadForward(Distance, TypeByteSize))
4493 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4494 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4499 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4500 MemAccessInfoSet &CheckDeps,
4501 ValueToValueMap &Strides) {
4503 MaxSafeDepDistBytes = -1U;
4504 while (!CheckDeps.empty()) {
4505 MemAccessInfo CurAccess = *CheckDeps.begin();
4507 // Get the relevant memory access set.
4508 EquivalenceClasses<MemAccessInfo>::iterator I =
4509 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4511 // Check accesses within this set.
4512 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4513 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4515 // Check every access pair.
4517 CheckDeps.erase(*AI);
4518 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4520 // Check every accessing instruction pair in program order.
4521 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4522 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4523 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4524 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4525 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4527 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4538 bool LoopVectorizationLegality::canVectorizeMemory() {
4540 typedef SmallVector<Value*, 16> ValueVector;
4541 typedef SmallPtrSet<Value*, 16> ValueSet;
4543 // Holds the Load and Store *instructions*.
4547 // Holds all the different accesses in the loop.
4548 unsigned NumReads = 0;
4549 unsigned NumReadWrites = 0;
4551 PtrRtCheck.Pointers.clear();
4552 PtrRtCheck.Need = false;
4554 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4555 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4558 for (Loop::block_iterator bb = TheLoop->block_begin(),
4559 be = TheLoop->block_end(); bb != be; ++bb) {
4561 // Scan the BB and collect legal loads and stores.
4562 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4565 // If this is a load, save it. If this instruction can read from memory
4566 // but is not a load, then we quit. Notice that we don't handle function
4567 // calls that read or write.
4568 if (it->mayReadFromMemory()) {
4569 // Many math library functions read the rounding mode. We will only
4570 // vectorize a loop if it contains known function calls that don't set
4571 // the flag. Therefore, it is safe to ignore this read from memory.
4572 CallInst *Call = dyn_cast<CallInst>(it);
4573 if (Call && getIntrinsicIDForCall(Call, TLI))
4576 LoadInst *Ld = dyn_cast<LoadInst>(it);
4577 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4578 emitAnalysis(Report(Ld)
4579 << "read with atomic ordering or volatile read");
4580 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4584 Loads.push_back(Ld);
4585 DepChecker.addAccess(Ld);
4589 // Save 'store' instructions. Abort if other instructions write to memory.
4590 if (it->mayWriteToMemory()) {
4591 StoreInst *St = dyn_cast<StoreInst>(it);
4593 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4596 if (!St->isSimple() && !IsAnnotatedParallel) {
4597 emitAnalysis(Report(St)
4598 << "write with atomic ordering or volatile write");
4599 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4603 Stores.push_back(St);
4604 DepChecker.addAccess(St);
4609 // Now we have two lists that hold the loads and the stores.
4610 // Next, we find the pointers that they use.
4612 // Check if we see any stores. If there are no stores, then we don't
4613 // care if the pointers are *restrict*.
4614 if (!Stores.size()) {
4615 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4619 AccessAnalysis::DepCandidates DependentAccesses;
4620 AccessAnalysis Accesses(DL, DependentAccesses);
4622 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4623 // multiple times on the same object. If the ptr is accessed twice, once
4624 // for read and once for write, it will only appear once (on the write
4625 // list). This is okay, since we are going to check for conflicts between
4626 // writes and between reads and writes, but not between reads and reads.
4629 ValueVector::iterator I, IE;
4630 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4631 StoreInst *ST = cast<StoreInst>(*I);
4632 Value* Ptr = ST->getPointerOperand();
4634 if (isUniform(Ptr)) {
4637 << "write to a loop invariant address could not be vectorized");
4638 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4642 // If we did *not* see this pointer before, insert it to the read-write
4643 // list. At this phase it is only a 'write' list.
4644 if (Seen.insert(Ptr)) {
4646 Accesses.addStore(Ptr);
4650 if (IsAnnotatedParallel) {
4652 << "LV: A loop annotated parallel, ignore memory dependency "
4657 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4658 LoadInst *LD = cast<LoadInst>(*I);
4659 Value* Ptr = LD->getPointerOperand();
4660 // If we did *not* see this pointer before, insert it to the
4661 // read list. If we *did* see it before, then it is already in
4662 // the read-write list. This allows us to vectorize expressions
4663 // such as A[i] += x; Because the address of A[i] is a read-write
4664 // pointer. This only works if the index of A[i] is consecutive.
4665 // If the address of i is unknown (for example A[B[i]]) then we may
4666 // read a few words, modify, and write a few words, and some of the
4667 // words may be written to the same address.
4668 bool IsReadOnlyPtr = false;
4669 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4671 IsReadOnlyPtr = true;
4673 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4676 // If we write (or read-write) to a single destination and there are no
4677 // other reads in this loop then is it safe to vectorize.
4678 if (NumReadWrites == 1 && NumReads == 0) {
4679 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4683 // Build dependence sets and check whether we need a runtime pointer bounds
4685 Accesses.buildDependenceSets();
4686 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4688 // Find pointers with computable bounds. We are going to use this information
4689 // to place a runtime bound check.
4690 unsigned NumComparisons = 0;
4691 bool CanDoRT = false;
4693 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4696 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4697 " pointer comparisons.\n");
4699 // If we only have one set of dependences to check pointers among we don't
4700 // need a runtime check.
4701 if (NumComparisons == 0 && NeedRTCheck)
4702 NeedRTCheck = false;
4704 // Check that we did not collect too many pointers or found an unsizeable
4706 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4712 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4715 if (NeedRTCheck && !CanDoRT) {
4716 emitAnalysis(Report() << "cannot identify array bounds");
4717 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4718 "the array bounds.\n");
4723 PtrRtCheck.Need = NeedRTCheck;
4725 bool CanVecMem = true;
4726 if (Accesses.isDependencyCheckNeeded()) {
4727 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4728 CanVecMem = DepChecker.areDepsSafe(
4729 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4730 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4732 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4733 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4736 // Clear the dependency checks. We assume they are not needed.
4737 Accesses.resetDepChecks();
4740 PtrRtCheck.Need = true;
4742 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4743 TheLoop, Strides, true);
4744 // Check that we did not collect too many pointers or found an unsizeable
4746 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4747 if (!CanDoRT && NumComparisons > 0)
4748 emitAnalysis(Report()
4749 << "cannot check memory dependencies at runtime");
4751 emitAnalysis(Report()
4752 << NumComparisons << " exceeds limit of "
4753 << RuntimeMemoryCheckThreshold
4754 << " dependent memory operations checked at runtime");
4755 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4765 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4767 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4768 " need a runtime memory check.\n");
4773 static bool hasMultipleUsesOf(Instruction *I,
4774 SmallPtrSet<Instruction *, 8> &Insts) {
4775 unsigned NumUses = 0;
4776 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4777 if (Insts.count(dyn_cast<Instruction>(*Use)))
4786 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4787 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4788 if (!Set.count(dyn_cast<Instruction>(*Use)))
4793 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4794 ReductionKind Kind) {
4795 if (Phi->getNumIncomingValues() != 2)
4798 // Reduction variables are only found in the loop header block.
4799 if (Phi->getParent() != TheLoop->getHeader())
4802 // Obtain the reduction start value from the value that comes from the loop
4804 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4806 // ExitInstruction is the single value which is used outside the loop.
4807 // We only allow for a single reduction value to be used outside the loop.
4808 // This includes users of the reduction, variables (which form a cycle
4809 // which ends in the phi node).
4810 Instruction *ExitInstruction = nullptr;
4811 // Indicates that we found a reduction operation in our scan.
4812 bool FoundReduxOp = false;
4814 // We start with the PHI node and scan for all of the users of this
4815 // instruction. All users must be instructions that can be used as reduction
4816 // variables (such as ADD). We must have a single out-of-block user. The cycle
4817 // must include the original PHI.
4818 bool FoundStartPHI = false;
4820 // To recognize min/max patterns formed by a icmp select sequence, we store
4821 // the number of instruction we saw from the recognized min/max pattern,
4822 // to make sure we only see exactly the two instructions.
4823 unsigned NumCmpSelectPatternInst = 0;
4824 ReductionInstDesc ReduxDesc(false, nullptr);
4826 SmallPtrSet<Instruction *, 8> VisitedInsts;
4827 SmallVector<Instruction *, 8> Worklist;
4828 Worklist.push_back(Phi);
4829 VisitedInsts.insert(Phi);
4831 // A value in the reduction can be used:
4832 // - By the reduction:
4833 // - Reduction operation:
4834 // - One use of reduction value (safe).
4835 // - Multiple use of reduction value (not safe).
4837 // - All uses of the PHI must be the reduction (safe).
4838 // - Otherwise, not safe.
4839 // - By one instruction outside of the loop (safe).
4840 // - By further instructions outside of the loop (not safe).
4841 // - By an instruction that is not part of the reduction (not safe).
4843 // * An instruction type other than PHI or the reduction operation.
4844 // * A PHI in the header other than the initial PHI.
4845 while (!Worklist.empty()) {
4846 Instruction *Cur = Worklist.back();
4847 Worklist.pop_back();
4850 // If the instruction has no users then this is a broken chain and can't be
4851 // a reduction variable.
4852 if (Cur->use_empty())
4855 bool IsAPhi = isa<PHINode>(Cur);
4857 // A header PHI use other than the original PHI.
4858 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4861 // Reductions of instructions such as Div, and Sub is only possible if the
4862 // LHS is the reduction variable.
4863 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4864 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4865 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4868 // Any reduction instruction must be of one of the allowed kinds.
4869 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4870 if (!ReduxDesc.IsReduction)
4873 // A reduction operation must only have one use of the reduction value.
4874 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4875 hasMultipleUsesOf(Cur, VisitedInsts))
4878 // All inputs to a PHI node must be a reduction value.
4879 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4882 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4883 isa<SelectInst>(Cur)))
4884 ++NumCmpSelectPatternInst;
4885 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4886 isa<SelectInst>(Cur)))
4887 ++NumCmpSelectPatternInst;
4889 // Check whether we found a reduction operator.
4890 FoundReduxOp |= !IsAPhi;
4892 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4893 // onto the stack. This way we are going to have seen all inputs to PHI
4894 // nodes once we get to them.
4895 SmallVector<Instruction *, 8> NonPHIs;
4896 SmallVector<Instruction *, 8> PHIs;
4897 for (User *U : Cur->users()) {
4898 Instruction *UI = cast<Instruction>(U);
4900 // Check if we found the exit user.
4901 BasicBlock *Parent = UI->getParent();
4902 if (!TheLoop->contains(Parent)) {
4903 // Exit if you find multiple outside users or if the header phi node is
4904 // being used. In this case the user uses the value of the previous
4905 // iteration, in which case we would loose "VF-1" iterations of the
4906 // reduction operation if we vectorize.
4907 if (ExitInstruction != nullptr || Cur == Phi)
4910 // The instruction used by an outside user must be the last instruction
4911 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4912 // operations on the value.
4913 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4916 ExitInstruction = Cur;
4920 // Process instructions only once (termination). Each reduction cycle
4921 // value must only be used once, except by phi nodes and min/max
4922 // reductions which are represented as a cmp followed by a select.
4923 ReductionInstDesc IgnoredVal(false, nullptr);
4924 if (VisitedInsts.insert(UI)) {
4925 if (isa<PHINode>(UI))
4928 NonPHIs.push_back(UI);
4929 } else if (!isa<PHINode>(UI) &&
4930 ((!isa<FCmpInst>(UI) &&
4931 !isa<ICmpInst>(UI) &&
4932 !isa<SelectInst>(UI)) ||
4933 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4936 // Remember that we completed the cycle.
4938 FoundStartPHI = true;
4940 Worklist.append(PHIs.begin(), PHIs.end());
4941 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4944 // This means we have seen one but not the other instruction of the
4945 // pattern or more than just a select and cmp.
4946 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4947 NumCmpSelectPatternInst != 2)
4950 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4953 // We found a reduction var if we have reached the original phi node and we
4954 // only have a single instruction with out-of-loop users.
4956 // This instruction is allowed to have out-of-loop users.
4957 AllowedExit.insert(ExitInstruction);
4959 // Save the description of this reduction variable.
4960 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4961 ReduxDesc.MinMaxKind);
4962 Reductions[Phi] = RD;
4963 // We've ended the cycle. This is a reduction variable if we have an
4964 // outside user and it has a binary op.
4969 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4970 /// pattern corresponding to a min(X, Y) or max(X, Y).
4971 LoopVectorizationLegality::ReductionInstDesc
4972 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4973 ReductionInstDesc &Prev) {
4975 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4976 "Expect a select instruction");
4977 Instruction *Cmp = nullptr;
4978 SelectInst *Select = nullptr;
4980 // We must handle the select(cmp()) as a single instruction. Advance to the
4982 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4983 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4984 return ReductionInstDesc(false, I);
4985 return ReductionInstDesc(Select, Prev.MinMaxKind);
4988 // Only handle single use cases for now.
4989 if (!(Select = dyn_cast<SelectInst>(I)))
4990 return ReductionInstDesc(false, I);
4991 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4992 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4993 return ReductionInstDesc(false, I);
4994 if (!Cmp->hasOneUse())
4995 return ReductionInstDesc(false, I);
5000 // Look for a min/max pattern.
5001 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5002 return ReductionInstDesc(Select, MRK_UIntMin);
5003 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5004 return ReductionInstDesc(Select, MRK_UIntMax);
5005 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5006 return ReductionInstDesc(Select, MRK_SIntMax);
5007 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5008 return ReductionInstDesc(Select, MRK_SIntMin);
5009 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5010 return ReductionInstDesc(Select, MRK_FloatMin);
5011 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5012 return ReductionInstDesc(Select, MRK_FloatMax);
5013 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5014 return ReductionInstDesc(Select, MRK_FloatMin);
5015 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5016 return ReductionInstDesc(Select, MRK_FloatMax);
5018 return ReductionInstDesc(false, I);
5021 LoopVectorizationLegality::ReductionInstDesc
5022 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5024 ReductionInstDesc &Prev) {
5025 bool FP = I->getType()->isFloatingPointTy();
5026 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
5027 switch (I->getOpcode()) {
5029 return ReductionInstDesc(false, I);
5030 case Instruction::PHI:
5031 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5032 Kind != RK_FloatMinMax))
5033 return ReductionInstDesc(false, I);
5034 return ReductionInstDesc(I, Prev.MinMaxKind);
5035 case Instruction::Sub:
5036 case Instruction::Add:
5037 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5038 case Instruction::Mul:
5039 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5040 case Instruction::And:
5041 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5042 case Instruction::Or:
5043 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5044 case Instruction::Xor:
5045 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5046 case Instruction::FMul:
5047 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5048 case Instruction::FAdd:
5049 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5050 case Instruction::FCmp:
5051 case Instruction::ICmp:
5052 case Instruction::Select:
5053 if (Kind != RK_IntegerMinMax &&
5054 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5055 return ReductionInstDesc(false, I);
5056 return isMinMaxSelectCmpPattern(I, Prev);
5060 LoopVectorizationLegality::InductionKind
5061 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5062 Type *PhiTy = Phi->getType();
5063 // We only handle integer and pointer inductions variables.
5064 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5065 return IK_NoInduction;
5067 // Check that the PHI is consecutive.
5068 const SCEV *PhiScev = SE->getSCEV(Phi);
5069 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5071 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5072 return IK_NoInduction;
5074 const SCEV *Step = AR->getStepRecurrence(*SE);
5076 // Integer inductions need to have a stride of one.
5077 if (PhiTy->isIntegerTy()) {
5079 return IK_IntInduction;
5080 if (Step->isAllOnesValue())
5081 return IK_ReverseIntInduction;
5082 return IK_NoInduction;
5085 // Calculate the pointer stride and check if it is consecutive.
5086 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5088 return IK_NoInduction;
5090 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5091 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5092 if (C->getValue()->equalsInt(Size))
5093 return IK_PtrInduction;
5094 else if (C->getValue()->equalsInt(0 - Size))
5095 return IK_ReversePtrInduction;
5097 return IK_NoInduction;
5100 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5101 Value *In0 = const_cast<Value*>(V);
5102 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5106 return Inductions.count(PN);
5109 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5110 assert(TheLoop->contains(BB) && "Unknown block used");
5112 // Blocks that do not dominate the latch need predication.
5113 BasicBlock* Latch = TheLoop->getLoopLatch();
5114 return !DT->dominates(BB, Latch);
5117 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5118 SmallPtrSet<Value *, 8>& SafePtrs) {
5119 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5120 // We might be able to hoist the load.
5121 if (it->mayReadFromMemory()) {
5122 LoadInst *LI = dyn_cast<LoadInst>(it);
5123 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5127 // We don't predicate stores at the moment.
5128 if (it->mayWriteToMemory()) {
5129 StoreInst *SI = dyn_cast<StoreInst>(it);
5130 // We only support predication of stores in basic blocks with one
5132 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5133 !SafePtrs.count(SI->getPointerOperand()) ||
5134 !SI->getParent()->getSinglePredecessor())
5140 // Check that we don't have a constant expression that can trap as operand.
5141 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5143 if (Constant *C = dyn_cast<Constant>(*OI))
5148 // The instructions below can trap.
5149 switch (it->getOpcode()) {
5151 case Instruction::UDiv:
5152 case Instruction::SDiv:
5153 case Instruction::URem:
5154 case Instruction::SRem:
5162 LoopVectorizationCostModel::VectorizationFactor
5163 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5165 bool ForceVectorization) {
5166 // Width 1 means no vectorize
5167 VectorizationFactor Factor = { 1U, 0U };
5168 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5169 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5173 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5174 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5178 // Find the trip count.
5179 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5180 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5182 unsigned WidestType = getWidestType();
5183 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5184 unsigned MaxSafeDepDist = -1U;
5185 if (Legal->getMaxSafeDepDistBytes() != -1U)
5186 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5187 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5188 WidestRegister : MaxSafeDepDist);
5189 unsigned MaxVectorSize = WidestRegister / WidestType;
5190 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5191 DEBUG(dbgs() << "LV: The Widest register is: "
5192 << WidestRegister << " bits.\n");
5194 if (MaxVectorSize == 0) {
5195 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5199 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5200 " into one vector!");
5202 unsigned VF = MaxVectorSize;
5204 // If we optimize the program for size, avoid creating the tail loop.
5206 // If we are unable to calculate the trip count then don't try to vectorize.
5208 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5212 // Find the maximum SIMD width that can fit within the trip count.
5213 VF = TC % MaxVectorSize;
5218 // If the trip count that we found modulo the vectorization factor is not
5219 // zero then we require a tail.
5221 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5227 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5228 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5230 Factor.Width = UserVF;
5234 float Cost = expectedCost(1);
5236 const float ScalarCost = Cost;
5239 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5241 // Ignore scalar width, because the user explicitly wants vectorization.
5242 if (ForceVectorization && VF > 1) {
5244 Cost = expectedCost(Width) / (float)Width;
5247 for (unsigned i=2; i <= VF; i*=2) {
5248 // Notice that the vector loop needs to be executed less times, so
5249 // we need to divide the cost of the vector loops by the width of
5250 // the vector elements.
5251 float VectorCost = expectedCost(i) / (float)i;
5252 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5253 (int)VectorCost << ".\n");
5254 if (VectorCost < Cost) {
5260 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5261 << "LV: Vectorization seems to be not beneficial, "
5262 << "but was forced by a user.\n");
5263 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5264 Factor.Width = Width;
5265 Factor.Cost = Width * Cost;
5269 unsigned LoopVectorizationCostModel::getWidestType() {
5270 unsigned MaxWidth = 8;
5273 for (Loop::block_iterator bb = TheLoop->block_begin(),
5274 be = TheLoop->block_end(); bb != be; ++bb) {
5275 BasicBlock *BB = *bb;
5277 // For each instruction in the loop.
5278 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5279 Type *T = it->getType();
5281 // Only examine Loads, Stores and PHINodes.
5282 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5285 // Examine PHI nodes that are reduction variables.
5286 if (PHINode *PN = dyn_cast<PHINode>(it))
5287 if (!Legal->getReductionVars()->count(PN))
5290 // Examine the stored values.
5291 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5292 T = ST->getValueOperand()->getType();
5294 // Ignore loaded pointer types and stored pointer types that are not
5295 // consecutive. However, we do want to take consecutive stores/loads of
5296 // pointer vectors into account.
5297 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5300 MaxWidth = std::max(MaxWidth,
5301 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5309 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5312 unsigned LoopCost) {
5314 // -- The unroll heuristics --
5315 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5316 // There are many micro-architectural considerations that we can't predict
5317 // at this level. For example frontend pressure (on decode or fetch) due to
5318 // code size, or the number and capabilities of the execution ports.
5320 // We use the following heuristics to select the unroll factor:
5321 // 1. If the code has reductions the we unroll in order to break the cross
5322 // iteration dependency.
5323 // 2. If the loop is really small then we unroll in order to reduce the loop
5325 // 3. We don't unroll if we think that we will spill registers to memory due
5326 // to the increased register pressure.
5328 // Use the user preference, unless 'auto' is selected.
5332 // When we optimize for size we don't unroll.
5336 // We used the distance for the unroll factor.
5337 if (Legal->getMaxSafeDepDistBytes() != -1U)
5340 // Do not unroll loops with a relatively small trip count.
5341 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5342 TheLoop->getLoopLatch());
5343 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5346 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5347 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5351 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5352 TargetNumRegisters = ForceTargetNumScalarRegs;
5354 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5355 TargetNumRegisters = ForceTargetNumVectorRegs;
5358 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5359 // We divide by these constants so assume that we have at least one
5360 // instruction that uses at least one register.
5361 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5362 R.NumInstructions = std::max(R.NumInstructions, 1U);
5364 // We calculate the unroll factor using the following formula.
5365 // Subtract the number of loop invariants from the number of available
5366 // registers. These registers are used by all of the unrolled instances.
5367 // Next, divide the remaining registers by the number of registers that is
5368 // required by the loop, in order to estimate how many parallel instances
5369 // fit without causing spills. All of this is rounded down if necessary to be
5370 // a power of two. We want power of two unroll factors to simplify any
5371 // addressing operations or alignment considerations.
5372 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5375 // Don't count the induction variable as unrolled.
5376 if (EnableIndVarRegisterHeur)
5377 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5378 std::max(1U, (R.MaxLocalUsers - 1)));
5380 // Clamp the unroll factor ranges to reasonable factors.
5381 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5383 // Check if the user has overridden the unroll max.
5385 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5386 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5388 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5389 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5392 // If we did not calculate the cost for VF (because the user selected the VF)
5393 // then we calculate the cost of VF here.
5395 LoopCost = expectedCost(VF);
5397 // Clamp the calculated UF to be between the 1 and the max unroll factor
5398 // that the target allows.
5399 if (UF > MaxUnrollSize)
5404 // Unroll if we vectorized this loop and there is a reduction that could
5405 // benefit from unrolling.
5406 if (VF > 1 && Legal->getReductionVars()->size()) {
5407 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5411 // Note that if we've already vectorized the loop we will have done the
5412 // runtime check and so unrolling won't require further checks.
5413 bool UnrollingRequiresRuntimePointerCheck =
5414 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5416 // We want to unroll small loops in order to reduce the loop overhead and
5417 // potentially expose ILP opportunities.
5418 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5419 if (!UnrollingRequiresRuntimePointerCheck &&
5420 LoopCost < SmallLoopCost) {
5421 // We assume that the cost overhead is 1 and we use the cost model
5422 // to estimate the cost of the loop and unroll until the cost of the
5423 // loop overhead is about 5% of the cost of the loop.
5424 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5426 // Unroll until store/load ports (estimated by max unroll factor) are
5428 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5429 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5431 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5432 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5433 return std::max(StoresUF, LoadsUF);
5436 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5440 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5444 LoopVectorizationCostModel::RegisterUsage
5445 LoopVectorizationCostModel::calculateRegisterUsage() {
5446 // This function calculates the register usage by measuring the highest number
5447 // of values that are alive at a single location. Obviously, this is a very
5448 // rough estimation. We scan the loop in a topological order in order and
5449 // assign a number to each instruction. We use RPO to ensure that defs are
5450 // met before their users. We assume that each instruction that has in-loop
5451 // users starts an interval. We record every time that an in-loop value is
5452 // used, so we have a list of the first and last occurrences of each
5453 // instruction. Next, we transpose this data structure into a multi map that
5454 // holds the list of intervals that *end* at a specific location. This multi
5455 // map allows us to perform a linear search. We scan the instructions linearly
5456 // and record each time that a new interval starts, by placing it in a set.
5457 // If we find this value in the multi-map then we remove it from the set.
5458 // The max register usage is the maximum size of the set.
5459 // We also search for instructions that are defined outside the loop, but are
5460 // used inside the loop. We need this number separately from the max-interval
5461 // usage number because when we unroll, loop-invariant values do not take
5463 LoopBlocksDFS DFS(TheLoop);
5467 R.NumInstructions = 0;
5469 // Each 'key' in the map opens a new interval. The values
5470 // of the map are the index of the 'last seen' usage of the
5471 // instruction that is the key.
5472 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5473 // Maps instruction to its index.
5474 DenseMap<unsigned, Instruction*> IdxToInstr;
5475 // Marks the end of each interval.
5476 IntervalMap EndPoint;
5477 // Saves the list of instruction indices that are used in the loop.
5478 SmallSet<Instruction*, 8> Ends;
5479 // Saves the list of values that are used in the loop but are
5480 // defined outside the loop, such as arguments and constants.
5481 SmallPtrSet<Value*, 8> LoopInvariants;
5484 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5485 be = DFS.endRPO(); bb != be; ++bb) {
5486 R.NumInstructions += (*bb)->size();
5487 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5489 Instruction *I = it;
5490 IdxToInstr[Index++] = I;
5492 // Save the end location of each USE.
5493 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5494 Value *U = I->getOperand(i);
5495 Instruction *Instr = dyn_cast<Instruction>(U);
5497 // Ignore non-instruction values such as arguments, constants, etc.
5498 if (!Instr) continue;
5500 // If this instruction is outside the loop then record it and continue.
5501 if (!TheLoop->contains(Instr)) {
5502 LoopInvariants.insert(Instr);
5506 // Overwrite previous end points.
5507 EndPoint[Instr] = Index;
5513 // Saves the list of intervals that end with the index in 'key'.
5514 typedef SmallVector<Instruction*, 2> InstrList;
5515 DenseMap<unsigned, InstrList> TransposeEnds;
5517 // Transpose the EndPoints to a list of values that end at each index.
5518 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5520 TransposeEnds[it->second].push_back(it->first);
5522 SmallSet<Instruction*, 8> OpenIntervals;
5523 unsigned MaxUsage = 0;
5526 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5527 for (unsigned int i = 0; i < Index; ++i) {
5528 Instruction *I = IdxToInstr[i];
5529 // Ignore instructions that are never used within the loop.
5530 if (!Ends.count(I)) continue;
5532 // Remove all of the instructions that end at this location.
5533 InstrList &List = TransposeEnds[i];
5534 for (unsigned int j=0, e = List.size(); j < e; ++j)
5535 OpenIntervals.erase(List[j]);
5537 // Count the number of live interals.
5538 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5540 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5541 OpenIntervals.size() << '\n');
5543 // Add the current instruction to the list of open intervals.
5544 OpenIntervals.insert(I);
5547 unsigned Invariant = LoopInvariants.size();
5548 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5549 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5550 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5552 R.LoopInvariantRegs = Invariant;
5553 R.MaxLocalUsers = MaxUsage;
5557 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5561 for (Loop::block_iterator bb = TheLoop->block_begin(),
5562 be = TheLoop->block_end(); bb != be; ++bb) {
5563 unsigned BlockCost = 0;
5564 BasicBlock *BB = *bb;
5566 // For each instruction in the old loop.
5567 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5568 // Skip dbg intrinsics.
5569 if (isa<DbgInfoIntrinsic>(it))
5572 unsigned C = getInstructionCost(it, VF);
5574 // Check if we should override the cost.
5575 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5576 C = ForceTargetInstructionCost;
5579 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5580 VF << " For instruction: " << *it << '\n');
5583 // We assume that if-converted blocks have a 50% chance of being executed.
5584 // When the code is scalar then some of the blocks are avoided due to CF.
5585 // When the code is vectorized we execute all code paths.
5586 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5595 /// \brief Check whether the address computation for a non-consecutive memory
5596 /// access looks like an unlikely candidate for being merged into the indexing
5599 /// We look for a GEP which has one index that is an induction variable and all
5600 /// other indices are loop invariant. If the stride of this access is also
5601 /// within a small bound we decide that this address computation can likely be
5602 /// merged into the addressing mode.
5603 /// In all other cases, we identify the address computation as complex.
5604 static bool isLikelyComplexAddressComputation(Value *Ptr,
5605 LoopVectorizationLegality *Legal,
5606 ScalarEvolution *SE,
5607 const Loop *TheLoop) {
5608 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5612 // We are looking for a gep with all loop invariant indices except for one
5613 // which should be an induction variable.
5614 unsigned NumOperands = Gep->getNumOperands();
5615 for (unsigned i = 1; i < NumOperands; ++i) {
5616 Value *Opd = Gep->getOperand(i);
5617 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5618 !Legal->isInductionVariable(Opd))
5622 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5623 // can likely be merged into the address computation.
5624 unsigned MaxMergeDistance = 64;
5626 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5630 // Check the step is constant.
5631 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5632 // Calculate the pointer stride and check if it is consecutive.
5633 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5637 const APInt &APStepVal = C->getValue()->getValue();
5639 // Huge step value - give up.
5640 if (APStepVal.getBitWidth() > 64)
5643 int64_t StepVal = APStepVal.getSExtValue();
5645 return StepVal > MaxMergeDistance;
5648 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5649 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5655 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5656 // If we know that this instruction will remain uniform, check the cost of
5657 // the scalar version.
5658 if (Legal->isUniformAfterVectorization(I))
5661 Type *RetTy = I->getType();
5662 Type *VectorTy = ToVectorTy(RetTy, VF);
5664 // TODO: We need to estimate the cost of intrinsic calls.
5665 switch (I->getOpcode()) {
5666 case Instruction::GetElementPtr:
5667 // We mark this instruction as zero-cost because the cost of GEPs in
5668 // vectorized code depends on whether the corresponding memory instruction
5669 // is scalarized or not. Therefore, we handle GEPs with the memory
5670 // instruction cost.
5672 case Instruction::Br: {
5673 return TTI.getCFInstrCost(I->getOpcode());
5675 case Instruction::PHI:
5676 //TODO: IF-converted IFs become selects.
5678 case Instruction::Add:
5679 case Instruction::FAdd:
5680 case Instruction::Sub:
5681 case Instruction::FSub:
5682 case Instruction::Mul:
5683 case Instruction::FMul:
5684 case Instruction::UDiv:
5685 case Instruction::SDiv:
5686 case Instruction::FDiv:
5687 case Instruction::URem:
5688 case Instruction::SRem:
5689 case Instruction::FRem:
5690 case Instruction::Shl:
5691 case Instruction::LShr:
5692 case Instruction::AShr:
5693 case Instruction::And:
5694 case Instruction::Or:
5695 case Instruction::Xor: {
5696 // Since we will replace the stride by 1 the multiplication should go away.
5697 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5699 // Certain instructions can be cheaper to vectorize if they have a constant
5700 // second vector operand. One example of this are shifts on x86.
5701 TargetTransformInfo::OperandValueKind Op1VK =
5702 TargetTransformInfo::OK_AnyValue;
5703 TargetTransformInfo::OperandValueKind Op2VK =
5704 TargetTransformInfo::OK_AnyValue;
5705 Value *Op2 = I->getOperand(1);
5707 // Check for a splat of a constant or for a non uniform vector of constants.
5708 if (isa<ConstantInt>(Op2))
5709 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5710 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5711 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5712 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5713 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5716 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5718 case Instruction::Select: {
5719 SelectInst *SI = cast<SelectInst>(I);
5720 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5721 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5722 Type *CondTy = SI->getCondition()->getType();
5724 CondTy = VectorType::get(CondTy, VF);
5726 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5728 case Instruction::ICmp:
5729 case Instruction::FCmp: {
5730 Type *ValTy = I->getOperand(0)->getType();
5731 VectorTy = ToVectorTy(ValTy, VF);
5732 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5734 case Instruction::Store:
5735 case Instruction::Load: {
5736 StoreInst *SI = dyn_cast<StoreInst>(I);
5737 LoadInst *LI = dyn_cast<LoadInst>(I);
5738 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5740 VectorTy = ToVectorTy(ValTy, VF);
5742 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5743 unsigned AS = SI ? SI->getPointerAddressSpace() :
5744 LI->getPointerAddressSpace();
5745 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5746 // We add the cost of address computation here instead of with the gep
5747 // instruction because only here we know whether the operation is
5750 return TTI.getAddressComputationCost(VectorTy) +
5751 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5753 // Scalarized loads/stores.
5754 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5755 bool Reverse = ConsecutiveStride < 0;
5756 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5757 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5758 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5759 bool IsComplexComputation =
5760 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5762 // The cost of extracting from the value vector and pointer vector.
5763 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5764 for (unsigned i = 0; i < VF; ++i) {
5765 // The cost of extracting the pointer operand.
5766 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5767 // In case of STORE, the cost of ExtractElement from the vector.
5768 // In case of LOAD, the cost of InsertElement into the returned
5770 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5771 Instruction::InsertElement,
5775 // The cost of the scalar loads/stores.
5776 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5777 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5782 // Wide load/stores.
5783 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5784 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5787 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5791 case Instruction::ZExt:
5792 case Instruction::SExt:
5793 case Instruction::FPToUI:
5794 case Instruction::FPToSI:
5795 case Instruction::FPExt:
5796 case Instruction::PtrToInt:
5797 case Instruction::IntToPtr:
5798 case Instruction::SIToFP:
5799 case Instruction::UIToFP:
5800 case Instruction::Trunc:
5801 case Instruction::FPTrunc:
5802 case Instruction::BitCast: {
5803 // We optimize the truncation of induction variable.
5804 // The cost of these is the same as the scalar operation.
5805 if (I->getOpcode() == Instruction::Trunc &&
5806 Legal->isInductionVariable(I->getOperand(0)))
5807 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5808 I->getOperand(0)->getType());
5810 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5811 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5813 case Instruction::Call: {
5814 CallInst *CI = cast<CallInst>(I);
5815 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5816 assert(ID && "Not an intrinsic call!");
5817 Type *RetTy = ToVectorTy(CI->getType(), VF);
5818 SmallVector<Type*, 4> Tys;
5819 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5820 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5821 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5824 // We are scalarizing the instruction. Return the cost of the scalar
5825 // instruction, plus the cost of insert and extract into vector
5826 // elements, times the vector width.
5829 if (!RetTy->isVoidTy() && VF != 1) {
5830 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5832 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5835 // The cost of inserting the results plus extracting each one of the
5837 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5840 // The cost of executing VF copies of the scalar instruction. This opcode
5841 // is unknown. Assume that it is the same as 'mul'.
5842 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5848 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5849 if (Scalar->isVoidTy() || VF == 1)
5851 return VectorType::get(Scalar, VF);
5854 char LoopVectorize::ID = 0;
5855 static const char lv_name[] = "Loop Vectorization";
5856 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5857 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5858 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5859 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5860 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5861 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5862 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5863 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5864 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5867 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5868 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5872 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5873 // Check for a store.
5874 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5875 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5877 // Check for a load.
5878 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5879 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5885 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5886 bool IfPredicateStore) {
5887 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5888 // Holds vector parameters or scalars, in case of uniform vals.
5889 SmallVector<VectorParts, 4> Params;
5891 setDebugLocFromInst(Builder, Instr);
5893 // Find all of the vectorized parameters.
5894 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5895 Value *SrcOp = Instr->getOperand(op);
5897 // If we are accessing the old induction variable, use the new one.
5898 if (SrcOp == OldInduction) {
5899 Params.push_back(getVectorValue(SrcOp));
5903 // Try using previously calculated values.
5904 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5906 // If the src is an instruction that appeared earlier in the basic block
5907 // then it should already be vectorized.
5908 if (SrcInst && OrigLoop->contains(SrcInst)) {
5909 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5910 // The parameter is a vector value from earlier.
5911 Params.push_back(WidenMap.get(SrcInst));
5913 // The parameter is a scalar from outside the loop. Maybe even a constant.
5914 VectorParts Scalars;
5915 Scalars.append(UF, SrcOp);
5916 Params.push_back(Scalars);
5920 assert(Params.size() == Instr->getNumOperands() &&
5921 "Invalid number of operands");
5923 // Does this instruction return a value ?
5924 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5926 Value *UndefVec = IsVoidRetTy ? nullptr :
5927 UndefValue::get(Instr->getType());
5928 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5929 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5931 Instruction *InsertPt = Builder.GetInsertPoint();
5932 BasicBlock *IfBlock = Builder.GetInsertBlock();
5933 BasicBlock *CondBlock = nullptr;
5936 Loop *VectorLp = nullptr;
5937 if (IfPredicateStore) {
5938 assert(Instr->getParent()->getSinglePredecessor() &&
5939 "Only support single predecessor blocks");
5940 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5941 Instr->getParent());
5942 VectorLp = LI->getLoopFor(IfBlock);
5943 assert(VectorLp && "Must have a loop for this block");
5946 // For each vector unroll 'part':
5947 for (unsigned Part = 0; Part < UF; ++Part) {
5948 // For each scalar that we create:
5950 // Start an "if (pred) a[i] = ..." block.
5951 Value *Cmp = nullptr;
5952 if (IfPredicateStore) {
5953 if (Cond[Part]->getType()->isVectorTy())
5955 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5956 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5957 ConstantInt::get(Cond[Part]->getType(), 1));
5958 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5959 LoopVectorBody.push_back(CondBlock);
5960 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5961 // Update Builder with newly created basic block.
5962 Builder.SetInsertPoint(InsertPt);
5965 Instruction *Cloned = Instr->clone();
5967 Cloned->setName(Instr->getName() + ".cloned");
5968 // Replace the operands of the cloned instructions with extracted scalars.
5969 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5970 Value *Op = Params[op][Part];
5971 Cloned->setOperand(op, Op);
5974 // Place the cloned scalar in the new loop.
5975 Builder.Insert(Cloned);
5977 // If the original scalar returns a value we need to place it in a vector
5978 // so that future users will be able to use it.
5980 VecResults[Part] = Cloned;
5983 if (IfPredicateStore) {
5984 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5985 LoopVectorBody.push_back(NewIfBlock);
5986 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5987 Builder.SetInsertPoint(InsertPt);
5988 Instruction *OldBr = IfBlock->getTerminator();
5989 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5990 OldBr->eraseFromParent();
5991 IfBlock = NewIfBlock;
5996 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5997 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5998 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6000 return scalarizeInstruction(Instr, IfPredicateStore);
6003 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6007 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6011 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6013 // When unrolling and the VF is 1, we only need to add a simple scalar.
6014 Type *ITy = Val->getType();
6015 assert(!ITy->isVectorTy() && "Val must be a scalar");
6016 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6017 return Builder.CreateAdd(Val, C, "induction");