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/Dominators.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/LLVMContext.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/IR/PatternMatch.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.h"
80 #include "llvm/IR/ValueHandle.h"
81 #include "llvm/IR/Verifier.h"
82 #include "llvm/Pass.h"
83 #include "llvm/Support/BranchProbability.h"
84 #include "llvm/Support/CommandLine.h"
85 #include "llvm/Support/Debug.h"
86 #include "llvm/Support/raw_ostream.h"
87 #include "llvm/Transforms/Scalar.h"
88 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
89 #include "llvm/Transforms/Utils/Local.h"
90 #include "llvm/Transforms/Utils/VectorUtils.h"
96 using namespace llvm::PatternMatch;
98 #define LV_NAME "loop-vectorize"
99 #define DEBUG_TYPE LV_NAME
101 STATISTIC(LoopsVectorized, "Number of loops vectorized");
102 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
104 static cl::opt<unsigned>
105 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
106 cl::desc("Sets the SIMD width. Zero is autoselect."));
108 static cl::opt<unsigned>
109 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
110 cl::desc("Sets the vectorization unroll count. "
111 "Zero is autoselect."));
114 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
115 cl::desc("Enable if-conversion during vectorization."));
117 /// We don't vectorize loops with a known constant trip count below this number.
118 static cl::opt<unsigned>
119 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
121 cl::desc("Don't vectorize loops with a constant "
122 "trip count that is smaller than this "
125 /// This enables versioning on the strides of symbolically striding memory
126 /// accesses in code like the following.
127 /// for (i = 0; i < N; ++i)
128 /// A[i * Stride1] += B[i * Stride2] ...
130 /// Will be roughly translated to
131 /// if (Stride1 == 1 && Stride2 == 1) {
132 /// for (i = 0; i < N; i+=4)
136 static cl::opt<bool> EnableMemAccessVersioning(
137 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
138 cl::desc("Enable symblic stride memory access versioning"));
140 /// We don't unroll loops with a known constant trip count below this number.
141 static const unsigned TinyTripCountUnrollThreshold = 128;
143 /// When performing memory disambiguation checks at runtime do not make more
144 /// than this number of comparisons.
145 static const unsigned RuntimeMemoryCheckThreshold = 8;
147 /// Maximum simd width.
148 static const unsigned MaxVectorWidth = 64;
150 static cl::opt<unsigned> ForceTargetNumScalarRegs(
151 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
152 cl::desc("A flag that overrides the target's number of scalar registers."));
154 static cl::opt<unsigned> ForceTargetNumVectorRegs(
155 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's number of vector registers."));
158 /// Maximum vectorization unroll count.
159 static const unsigned MaxUnrollFactor = 16;
161 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
162 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
163 cl::desc("A flag that overrides the target's max unroll factor for scalar "
166 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
167 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max unroll factor for "
169 "vectorized loops."));
171 static cl::opt<unsigned> ForceTargetInstructionCost(
172 "force-target-instruction-cost", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's expected cost for "
174 "an instruction to a single constant value. Mostly "
175 "useful for getting consistent testing."));
177 static cl::opt<unsigned> SmallLoopCost(
178 "small-loop-cost", cl::init(20), cl::Hidden,
179 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
181 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
182 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
183 cl::desc("Enable the use of the block frequency analysis to access PGO "
184 "heuristics minimizing code growth in cold regions and being more "
185 "aggressive in hot regions."));
187 // Runtime unroll loops for load/store throughput.
188 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
189 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
190 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
192 /// The number of stores in a loop that are allowed to need predication.
193 static cl::opt<unsigned> NumberOfStoresToPredicate(
194 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
195 cl::desc("Max number of stores to be predicated behind an if."));
197 static cl::opt<bool> EnableIndVarRegisterHeur(
198 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
199 cl::desc("Count the induction variable only once when unrolling"));
201 static cl::opt<bool> EnableCondStoresVectorization(
202 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
203 cl::desc("Enable if predication of stores during vectorization."));
207 // Forward declarations.
208 class LoopVectorizationLegality;
209 class LoopVectorizationCostModel;
211 /// InnerLoopVectorizer vectorizes loops which contain only one basic
212 /// block to a specified vectorization factor (VF).
213 /// This class performs the widening of scalars into vectors, or multiple
214 /// scalars. This class also implements the following features:
215 /// * It inserts an epilogue loop for handling loops that don't have iteration
216 /// counts that are known to be a multiple of the vectorization factor.
217 /// * It handles the code generation for reduction variables.
218 /// * Scalarization (implementation using scalars) of un-vectorizable
220 /// InnerLoopVectorizer does not perform any vectorization-legality
221 /// checks, and relies on the caller to check for the different legality
222 /// aspects. The InnerLoopVectorizer relies on the
223 /// LoopVectorizationLegality class to provide information about the induction
224 /// and reduction variables that were found to a given vectorization factor.
225 class InnerLoopVectorizer {
227 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
228 DominatorTree *DT, const DataLayout *DL,
229 const TargetLibraryInfo *TLI, unsigned VecWidth,
230 unsigned UnrollFactor)
231 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
232 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
233 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
236 // Perform the actual loop widening (vectorization).
237 void vectorize(LoopVectorizationLegality *L) {
239 // Create a new empty loop. Unlink the old loop and connect the new one.
241 // Widen each instruction in the old loop to a new one in the new loop.
242 // Use the Legality module to find the induction and reduction variables.
244 // Register the new loop and update the analysis passes.
248 virtual ~InnerLoopVectorizer() {}
251 /// A small list of PHINodes.
252 typedef SmallVector<PHINode*, 4> PhiVector;
253 /// When we unroll loops we have multiple vector values for each scalar.
254 /// This data structure holds the unrolled and vectorized values that
255 /// originated from one scalar instruction.
256 typedef SmallVector<Value*, 2> VectorParts;
258 // When we if-convert we need create edge masks. We have to cache values so
259 // that we don't end up with exponential recursion/IR.
260 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
261 VectorParts> EdgeMaskCache;
263 /// \brief Add code that checks at runtime if the accessed arrays overlap.
265 /// Returns a pair of instructions where the first element is the first
266 /// instruction generated in possibly a sequence of instructions and the
267 /// second value is the final comparator value or NULL if no check is needed.
268 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
270 /// \brief Add checks for strides that where assumed to be 1.
272 /// Returns the last check instruction and the first check instruction in the
273 /// pair as (first, last).
274 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
276 /// Create an empty loop, based on the loop ranges of the old loop.
277 void createEmptyLoop();
278 /// Copy and widen the instructions from the old loop.
279 virtual void vectorizeLoop();
281 /// \brief The Loop exit block may have single value PHI nodes where the
282 /// incoming value is 'Undef'. While vectorizing we only handled real values
283 /// that were defined inside the loop. Here we fix the 'undef case'.
287 /// A helper function that computes the predicate of the block BB, assuming
288 /// that the header block of the loop is set to True. It returns the *entry*
289 /// mask for the block BB.
290 VectorParts createBlockInMask(BasicBlock *BB);
291 /// A helper function that computes the predicate of the edge between SRC
293 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
295 /// A helper function to vectorize a single BB within the innermost loop.
296 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
298 /// Vectorize a single PHINode in a block. This method handles the induction
299 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
300 /// arbitrary length vectors.
301 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
302 unsigned UF, unsigned VF, PhiVector *PV);
304 /// Insert the new loop to the loop hierarchy and pass manager
305 /// and update the analysis passes.
306 void updateAnalysis();
308 /// This instruction is un-vectorizable. Implement it as a sequence
309 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
310 /// scalarized instruction behind an if block predicated on the control
311 /// dependence of the instruction.
312 virtual void scalarizeInstruction(Instruction *Instr,
313 bool IfPredicateStore=false);
315 /// Vectorize Load and Store instructions,
316 virtual void vectorizeMemoryInstruction(Instruction *Instr);
318 /// Create a broadcast instruction. This method generates a broadcast
319 /// instruction (shuffle) for loop invariant values and for the induction
320 /// value. If this is the induction variable then we extend it to N, N+1, ...
321 /// this is needed because each iteration in the loop corresponds to a SIMD
323 virtual Value *getBroadcastInstrs(Value *V);
325 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
326 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
327 /// The sequence starts at StartIndex.
328 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
330 /// When we go over instructions in the basic block we rely on previous
331 /// values within the current basic block or on loop invariant values.
332 /// When we widen (vectorize) values we place them in the map. If the values
333 /// are not within the map, they have to be loop invariant, so we simply
334 /// broadcast them into a vector.
335 VectorParts &getVectorValue(Value *V);
337 /// Generate a shuffle sequence that will reverse the vector Vec.
338 virtual Value *reverseVector(Value *Vec);
340 /// This is a helper class that holds the vectorizer state. It maps scalar
341 /// instructions to vector instructions. When the code is 'unrolled' then
342 /// then a single scalar value is mapped to multiple vector parts. The parts
343 /// are stored in the VectorPart type.
345 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
347 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
349 /// \return True if 'Key' is saved in the Value Map.
350 bool has(Value *Key) const { return MapStorage.count(Key); }
352 /// Initializes a new entry in the map. Sets all of the vector parts to the
353 /// save value in 'Val'.
354 /// \return A reference to a vector with splat values.
355 VectorParts &splat(Value *Key, Value *Val) {
356 VectorParts &Entry = MapStorage[Key];
357 Entry.assign(UF, Val);
361 ///\return A reference to the value that is stored at 'Key'.
362 VectorParts &get(Value *Key) {
363 VectorParts &Entry = MapStorage[Key];
366 assert(Entry.size() == UF);
371 /// The unroll factor. Each entry in the map stores this number of vector
375 /// Map storage. We use std::map and not DenseMap because insertions to a
376 /// dense map invalidates its iterators.
377 std::map<Value *, VectorParts> MapStorage;
380 /// The original loop.
382 /// Scev analysis to use.
389 const DataLayout *DL;
390 /// Target Library Info.
391 const TargetLibraryInfo *TLI;
393 /// The vectorization SIMD factor to use. Each vector will have this many
398 /// The vectorization unroll factor to use. Each scalar is vectorized to this
399 /// many different vector instructions.
402 /// The builder that we use
405 // --- Vectorization state ---
407 /// The vector-loop preheader.
408 BasicBlock *LoopVectorPreHeader;
409 /// The scalar-loop preheader.
410 BasicBlock *LoopScalarPreHeader;
411 /// Middle Block between the vector and the scalar.
412 BasicBlock *LoopMiddleBlock;
413 ///The ExitBlock of the scalar loop.
414 BasicBlock *LoopExitBlock;
415 ///The vector loop body.
416 SmallVector<BasicBlock *, 4> LoopVectorBody;
417 ///The scalar loop body.
418 BasicBlock *LoopScalarBody;
419 /// A list of all bypass blocks. The first block is the entry of the loop.
420 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
422 /// The new Induction variable which was added to the new block.
424 /// The induction variable of the old basic block.
425 PHINode *OldInduction;
426 /// Holds the extended (to the widest induction type) start index.
428 /// Maps scalars to widened vectors.
430 EdgeMaskCache MaskCache;
432 LoopVectorizationLegality *Legal;
435 class InnerLoopUnroller : public InnerLoopVectorizer {
437 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
438 DominatorTree *DT, const DataLayout *DL,
439 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
440 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
443 void scalarizeInstruction(Instruction *Instr,
444 bool IfPredicateStore = false) override;
445 void vectorizeMemoryInstruction(Instruction *Instr) override;
446 Value *getBroadcastInstrs(Value *V) override;
447 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
448 Value *reverseVector(Value *Vec) override;
451 /// \brief Look for a meaningful debug location on the instruction or it's
453 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
458 if (I->getDebugLoc() != Empty)
461 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
462 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
463 if (OpInst->getDebugLoc() != Empty)
470 /// \brief Set the debug location in the builder using the debug location in the
472 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
473 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
474 B.SetCurrentDebugLocation(Inst->getDebugLoc());
476 B.SetCurrentDebugLocation(DebugLoc());
480 /// \return string containing a file name and a line # for the given
482 static std::string getDebugLocString(const Instruction *I) {
485 raw_string_ostream OS(Result);
486 const DebugLoc &InstrDebugLoc = I->getDebugLoc();
487 if (!InstrDebugLoc.isUnknown())
488 InstrDebugLoc.print(I->getContext(), OS);
490 // Just print the module name.
491 OS << I->getParent()->getParent()->getParent()->getModuleIdentifier();
498 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
499 /// to what vectorization factor.
500 /// This class does not look at the profitability of vectorization, only the
501 /// legality. This class has two main kinds of checks:
502 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
503 /// will change the order of memory accesses in a way that will change the
504 /// correctness of the program.
505 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
506 /// checks for a number of different conditions, such as the availability of a
507 /// single induction variable, that all types are supported and vectorize-able,
508 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
509 /// This class is also used by InnerLoopVectorizer for identifying
510 /// induction variable and the different reduction variables.
511 class LoopVectorizationLegality {
515 unsigned NumPredStores;
517 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
518 DominatorTree *DT, TargetLibraryInfo *TLI)
519 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
520 DT(DT), TLI(TLI), Induction(nullptr), WidestIndTy(nullptr),
521 HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {}
523 /// This enum represents the kinds of reductions that we support.
525 RK_NoReduction, ///< Not a reduction.
526 RK_IntegerAdd, ///< Sum of integers.
527 RK_IntegerMult, ///< Product of integers.
528 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
529 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
530 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
531 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
532 RK_FloatAdd, ///< Sum of floats.
533 RK_FloatMult, ///< Product of floats.
534 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
537 /// This enum represents the kinds of inductions that we support.
539 IK_NoInduction, ///< Not an induction variable.
540 IK_IntInduction, ///< Integer induction variable. Step = 1.
541 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
542 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
543 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
546 // This enum represents the kind of minmax reduction.
547 enum MinMaxReductionKind {
557 /// This struct holds information about reduction variables.
558 struct ReductionDescriptor {
559 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
560 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
562 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
563 MinMaxReductionKind MK)
564 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
566 // The starting value of the reduction.
567 // It does not have to be zero!
568 TrackingVH<Value> StartValue;
569 // The instruction who's value is used outside the loop.
570 Instruction *LoopExitInstr;
571 // The kind of the reduction.
573 // If this a min/max reduction the kind of reduction.
574 MinMaxReductionKind MinMaxKind;
577 /// This POD struct holds information about a potential reduction operation.
578 struct ReductionInstDesc {
579 ReductionInstDesc(bool IsRedux, Instruction *I) :
580 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
582 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
583 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
585 // Is this instruction a reduction candidate.
587 // The last instruction in a min/max pattern (select of the select(icmp())
588 // pattern), or the current reduction instruction otherwise.
589 Instruction *PatternLastInst;
590 // If this is a min/max pattern the comparison predicate.
591 MinMaxReductionKind MinMaxKind;
594 /// This struct holds information about the memory runtime legality
595 /// check that a group of pointers do not overlap.
596 struct RuntimePointerCheck {
597 RuntimePointerCheck() : Need(false) {}
599 /// Reset the state of the pointer runtime information.
606 DependencySetId.clear();
609 /// Insert a pointer and calculate the start and end SCEVs.
610 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
611 unsigned DepSetId, ValueToValueMap &Strides);
613 /// This flag indicates if we need to add the runtime check.
615 /// Holds the pointers that we need to check.
616 SmallVector<TrackingVH<Value>, 2> Pointers;
617 /// Holds the pointer value at the beginning of the loop.
618 SmallVector<const SCEV*, 2> Starts;
619 /// Holds the pointer value at the end of the loop.
620 SmallVector<const SCEV*, 2> Ends;
621 /// Holds the information if this pointer is used for writing to memory.
622 SmallVector<bool, 2> IsWritePtr;
623 /// Holds the id of the set of pointers that could be dependent because of a
624 /// shared underlying object.
625 SmallVector<unsigned, 2> DependencySetId;
628 /// A struct for saving information about induction variables.
629 struct InductionInfo {
630 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
631 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
633 TrackingVH<Value> StartValue;
638 /// ReductionList contains the reduction descriptors for all
639 /// of the reductions that were found in the loop.
640 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
642 /// InductionList saves induction variables and maps them to the
643 /// induction descriptor.
644 typedef MapVector<PHINode*, InductionInfo> InductionList;
646 /// Returns true if it is legal to vectorize this loop.
647 /// This does not mean that it is profitable to vectorize this
648 /// loop, only that it is legal to do so.
651 /// Returns the Induction variable.
652 PHINode *getInduction() { return Induction; }
654 /// Returns the reduction variables found in the loop.
655 ReductionList *getReductionVars() { return &Reductions; }
657 /// Returns the induction variables found in the loop.
658 InductionList *getInductionVars() { return &Inductions; }
660 /// Returns the widest induction type.
661 Type *getWidestInductionType() { return WidestIndTy; }
663 /// Returns True if V is an induction variable in this loop.
664 bool isInductionVariable(const Value *V);
666 /// Return true if the block BB needs to be predicated in order for the loop
667 /// to be vectorized.
668 bool blockNeedsPredication(BasicBlock *BB);
670 /// Check if this pointer is consecutive when vectorizing. This happens
671 /// when the last index of the GEP is the induction variable, or that the
672 /// pointer itself is an induction variable.
673 /// This check allows us to vectorize A[idx] into a wide load/store.
675 /// 0 - Stride is unknown or non-consecutive.
676 /// 1 - Address is consecutive.
677 /// -1 - Address is consecutive, and decreasing.
678 int isConsecutivePtr(Value *Ptr);
680 /// Returns true if the value V is uniform within the loop.
681 bool isUniform(Value *V);
683 /// Returns true if this instruction will remain scalar after vectorization.
684 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
686 /// Returns the information that we collected about runtime memory check.
687 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
689 /// This function returns the identity element (or neutral element) for
691 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
693 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
695 bool hasStride(Value *V) { return StrideSet.count(V); }
696 bool mustCheckStrides() { return !StrideSet.empty(); }
697 SmallPtrSet<Value *, 8>::iterator strides_begin() {
698 return StrideSet.begin();
700 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
703 /// Check if a single basic block loop is vectorizable.
704 /// At this point we know that this is a loop with a constant trip count
705 /// and we only need to check individual instructions.
706 bool canVectorizeInstrs();
708 /// When we vectorize loops we may change the order in which
709 /// we read and write from memory. This method checks if it is
710 /// legal to vectorize the code, considering only memory constrains.
711 /// Returns true if the loop is vectorizable
712 bool canVectorizeMemory();
714 /// Return true if we can vectorize this loop using the IF-conversion
716 bool canVectorizeWithIfConvert();
718 /// Collect the variables that need to stay uniform after vectorization.
719 void collectLoopUniforms();
721 /// Return true if all of the instructions in the block can be speculatively
722 /// executed. \p SafePtrs is a list of addresses that are known to be legal
723 /// and we know that we can read from them without segfault.
724 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
726 /// Returns True, if 'Phi' is the kind of reduction variable for type
727 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
728 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
729 /// Returns a struct describing if the instruction 'I' can be a reduction
730 /// variable of type 'Kind'. If the reduction is a min/max pattern of
731 /// select(icmp()) this function advances the instruction pointer 'I' from the
732 /// compare instruction to the select instruction and stores this pointer in
733 /// 'PatternLastInst' member of the returned struct.
734 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
735 ReductionInstDesc &Desc);
736 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
737 /// pattern corresponding to a min(X, Y) or max(X, Y).
738 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
739 ReductionInstDesc &Prev);
740 /// Returns the induction kind of Phi. This function may return NoInduction
741 /// if the PHI is not an induction variable.
742 InductionKind isInductionVariable(PHINode *Phi);
744 /// \brief Collect memory access with loop invariant strides.
746 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
748 void collectStridedAcccess(Value *LoadOrStoreInst);
750 /// The loop that we evaluate.
754 /// DataLayout analysis.
755 const DataLayout *DL;
758 /// Target Library Info.
759 TargetLibraryInfo *TLI;
761 // --- vectorization state --- //
763 /// Holds the integer induction variable. This is the counter of the
766 /// Holds the reduction variables.
767 ReductionList Reductions;
768 /// Holds all of the induction variables that we found in the loop.
769 /// Notice that inductions don't need to start at zero and that induction
770 /// variables can be pointers.
771 InductionList Inductions;
772 /// Holds the widest induction type encountered.
775 /// Allowed outside users. This holds the reduction
776 /// vars which can be accessed from outside the loop.
777 SmallPtrSet<Value*, 4> AllowedExit;
778 /// This set holds the variables which are known to be uniform after
780 SmallPtrSet<Instruction*, 4> Uniforms;
781 /// We need to check that all of the pointers in this list are disjoint
783 RuntimePointerCheck PtrRtCheck;
784 /// Can we assume the absence of NaNs.
785 bool HasFunNoNaNAttr;
787 unsigned MaxSafeDepDistBytes;
789 ValueToValueMap Strides;
790 SmallPtrSet<Value *, 8> StrideSet;
793 /// LoopVectorizationCostModel - estimates the expected speedups due to
795 /// In many cases vectorization is not profitable. This can happen because of
796 /// a number of reasons. In this class we mainly attempt to predict the
797 /// expected speedup/slowdowns due to the supported instruction set. We use the
798 /// TargetTransformInfo to query the different backends for the cost of
799 /// different operations.
800 class LoopVectorizationCostModel {
802 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
803 LoopVectorizationLegality *Legal,
804 const TargetTransformInfo &TTI,
805 const DataLayout *DL, const TargetLibraryInfo *TLI)
806 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
808 /// Information about vectorization costs
809 struct VectorizationFactor {
810 unsigned Width; // Vector width with best cost
811 unsigned Cost; // Cost of the loop with that width
813 /// \return The most profitable vectorization factor and the cost of that VF.
814 /// This method checks every power of two up to VF. If UserVF is not ZERO
815 /// then this vectorization factor will be selected if vectorization is
817 VectorizationFactor selectVectorizationFactor(bool OptForSize,
819 bool ForceVectorization);
821 /// \return The size (in bits) of the widest type in the code that
822 /// needs to be vectorized. We ignore values that remain scalar such as
823 /// 64 bit loop indices.
824 unsigned getWidestType();
826 /// \return The most profitable unroll factor.
827 /// If UserUF is non-zero then this method finds the best unroll-factor
828 /// based on register pressure and other parameters.
829 /// VF and LoopCost are the selected vectorization factor and the cost of the
831 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
834 /// \brief A struct that represents some properties of the register usage
836 struct RegisterUsage {
837 /// Holds the number of loop invariant values that are used in the loop.
838 unsigned LoopInvariantRegs;
839 /// Holds the maximum number of concurrent live intervals in the loop.
840 unsigned MaxLocalUsers;
841 /// Holds the number of instructions in the loop.
842 unsigned NumInstructions;
845 /// \return information about the register usage of the loop.
846 RegisterUsage calculateRegisterUsage();
849 /// Returns the expected execution cost. The unit of the cost does
850 /// not matter because we use the 'cost' units to compare different
851 /// vector widths. The cost that is returned is *not* normalized by
852 /// the factor width.
853 unsigned expectedCost(unsigned VF);
855 /// Returns the execution time cost of an instruction for a given vector
856 /// width. Vector width of one means scalar.
857 unsigned getInstructionCost(Instruction *I, unsigned VF);
859 /// A helper function for converting Scalar types to vector types.
860 /// If the incoming type is void, we return void. If the VF is 1, we return
862 static Type* ToVectorTy(Type *Scalar, unsigned VF);
864 /// Returns whether the instruction is a load or store and will be a emitted
865 /// as a vector operation.
866 bool isConsecutiveLoadOrStore(Instruction *I);
868 /// The loop that we evaluate.
872 /// Loop Info analysis.
874 /// Vectorization legality.
875 LoopVectorizationLegality *Legal;
876 /// Vector target information.
877 const TargetTransformInfo &TTI;
878 /// Target data layout information.
879 const DataLayout *DL;
880 /// Target Library Info.
881 const TargetLibraryInfo *TLI;
884 /// Utility class for getting and setting loop vectorizer hints in the form
885 /// of loop metadata.
886 class LoopVectorizeHints {
889 FK_Undefined = -1, ///< Not selected.
890 FK_Disabled = 0, ///< Forcing disabled.
891 FK_Enabled = 1, ///< Forcing enabled.
894 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
895 : Width(VectorizationFactor),
896 Unroll(DisableUnrolling),
898 LoopID(L->getLoopID()) {
900 // force-vector-unroll overrides DisableUnrolling.
901 if (VectorizationUnroll.getNumOccurrences() > 0)
902 Unroll = VectorizationUnroll;
904 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
905 << "LV: Unrolling disabled by the pass manager\n");
908 /// Return the loop vectorizer metadata prefix.
909 static StringRef Prefix() { return "llvm.vectorizer."; }
911 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
912 SmallVector<Value*, 2> Vals;
913 Vals.push_back(MDString::get(Context, Name));
914 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
915 return MDNode::get(Context, Vals);
918 /// Mark the loop L as already vectorized by setting the width to 1.
919 void setAlreadyVectorized(Loop *L) {
920 LLVMContext &Context = L->getHeader()->getContext();
924 // Create a new loop id with one more operand for the already_vectorized
925 // hint. If the loop already has a loop id then copy the existing operands.
926 SmallVector<Value*, 4> Vals(1);
928 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
929 Vals.push_back(LoopID->getOperand(i));
931 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
932 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
934 MDNode *NewLoopID = MDNode::get(Context, Vals);
935 // Set operand 0 to refer to the loop id itself.
936 NewLoopID->replaceOperandWith(0, NewLoopID);
938 L->setLoopID(NewLoopID);
940 LoopID->replaceAllUsesWith(NewLoopID);
945 unsigned getWidth() const { return Width; }
946 unsigned getUnroll() const { return Unroll; }
947 enum ForceKind getForce() const { return Force; }
948 MDNode *getLoopID() const { return LoopID; }
951 /// Find hints specified in the loop metadata.
952 void getHints(const Loop *L) {
956 // First operand should refer to the loop id itself.
957 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
958 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
960 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
961 const MDString *S = nullptr;
962 SmallVector<Value*, 4> Args;
964 // The expected hint is either a MDString or a MDNode with the first
965 // operand a MDString.
966 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
967 if (!MD || MD->getNumOperands() == 0)
969 S = dyn_cast<MDString>(MD->getOperand(0));
970 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
971 Args.push_back(MD->getOperand(i));
973 S = dyn_cast<MDString>(LoopID->getOperand(i));
974 assert(Args.size() == 0 && "too many arguments for MDString");
980 // Check if the hint starts with the vectorizer prefix.
981 StringRef Hint = S->getString();
982 if (!Hint.startswith(Prefix()))
984 // Remove the prefix.
985 Hint = Hint.substr(Prefix().size(), StringRef::npos);
987 if (Args.size() == 1)
988 getHint(Hint, Args[0]);
992 // Check string hint with one operand.
993 void getHint(StringRef Hint, Value *Arg) {
994 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
996 unsigned Val = C->getZExtValue();
998 if (Hint == "width") {
999 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1002 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1003 } else if (Hint == "unroll") {
1004 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1007 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1008 } else if (Hint == "enable") {
1009 if (C->getBitWidth() == 1)
1010 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1011 : LoopVectorizeHints::FK_Disabled;
1013 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1015 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1019 /// Vectorization width.
1021 /// Vectorization unroll factor.
1023 /// Vectorization forced
1024 enum ForceKind Force;
1029 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1031 return V.push_back(&L);
1033 for (Loop *InnerL : L)
1034 addInnerLoop(*InnerL, V);
1037 /// The LoopVectorize Pass.
1038 struct LoopVectorize : public FunctionPass {
1039 /// Pass identification, replacement for typeid
1042 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1044 DisableUnrolling(NoUnrolling),
1045 AlwaysVectorize(AlwaysVectorize) {
1046 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1049 ScalarEvolution *SE;
1050 const DataLayout *DL;
1052 TargetTransformInfo *TTI;
1054 BlockFrequencyInfo *BFI;
1055 TargetLibraryInfo *TLI;
1056 bool DisableUnrolling;
1057 bool AlwaysVectorize;
1059 BlockFrequency ColdEntryFreq;
1061 bool runOnFunction(Function &F) override {
1062 SE = &getAnalysis<ScalarEvolution>();
1063 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1064 DL = DLP ? &DLP->getDataLayout() : nullptr;
1065 LI = &getAnalysis<LoopInfo>();
1066 TTI = &getAnalysis<TargetTransformInfo>();
1067 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1068 BFI = &getAnalysis<BlockFrequencyInfo>();
1069 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1071 // Compute some weights outside of the loop over the loops. Compute this
1072 // using a BranchProbability to re-use its scaling math.
1073 const BranchProbability ColdProb(1, 5); // 20%
1074 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1076 // If the target claims to have no vector registers don't attempt
1078 if (!TTI->getNumberOfRegisters(true))
1082 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1083 << ": Missing data layout\n");
1087 // Build up a worklist of inner-loops to vectorize. This is necessary as
1088 // the act of vectorizing or partially unrolling a loop creates new loops
1089 // and can invalidate iterators across the loops.
1090 SmallVector<Loop *, 8> Worklist;
1093 addInnerLoop(*L, Worklist);
1095 LoopsAnalyzed += Worklist.size();
1097 // Now walk the identified inner loops.
1098 bool Changed = false;
1099 while (!Worklist.empty())
1100 Changed |= processLoop(Worklist.pop_back_val());
1102 // Process each loop nest in the function.
1106 bool processLoop(Loop *L) {
1107 assert(L->empty() && "Only process inner loops.");
1110 const std::string DebugLocStr =
1111 getDebugLocString(L->getHeader()->getFirstNonPHIOrDbgOrLifetime());
1114 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1115 << L->getHeader()->getParent()->getName() << "\" from "
1116 << DebugLocStr << "\n");
1118 LoopVectorizeHints Hints(L, DisableUnrolling);
1120 DEBUG(dbgs() << "LV: Loop hints:"
1122 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1124 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1126 : "?")) << " width=" << Hints.getWidth()
1127 << " unroll=" << Hints.getUnroll() << "\n");
1129 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1130 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1134 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1135 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1139 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1140 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1144 // Check the loop for a trip count threshold:
1145 // do not vectorize loops with a tiny trip count.
1146 BasicBlock *Latch = L->getLoopLatch();
1147 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1148 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1149 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1150 << "This loop is not worth vectorizing.");
1151 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1152 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1154 DEBUG(dbgs() << "\n");
1159 // Check if it is legal to vectorize the loop.
1160 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1161 if (!LVL.canVectorize()) {
1162 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1166 // Use the cost model.
1167 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1169 // Check the function attributes to find out if this function should be
1170 // optimized for size.
1171 Function *F = L->getHeader()->getParent();
1172 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1173 F->hasFnAttribute(Attribute::OptimizeForSize);
1175 // Compute the weighted frequency of this loop being executed and see if it
1176 // is less than 20% of the function entry baseline frequency. Note that we
1177 // always have a canonical loop here because we think we *can* vectoriez.
1178 // FIXME: This is hidden behind a flag due to pervasive problems with
1179 // exactly what block frequency models.
1180 if (LoopVectorizeWithBlockFrequency) {
1181 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1182 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1183 LoopEntryFreq < ColdEntryFreq)
1187 // Check the function attributes to see if implicit floats are allowed.a
1188 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1189 // an integer loop and the vector instructions selected are purely integer
1190 // vector instructions?
1191 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1192 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1193 "attribute is used.\n");
1197 // Select the optimal vectorization factor.
1198 const LoopVectorizationCostModel::VectorizationFactor VF =
1199 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1201 LoopVectorizeHints::FK_Enabled);
1203 // Select the unroll factor.
1205 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1207 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1208 << DebugLocStr << '\n');
1209 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1211 if (VF.Width == 1) {
1212 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1215 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1217 // Report the unrolling decision.
1218 F->getContext().emitOptimizationRemark(
1219 DEBUG_TYPE, *F, L->getStartLoc(),
1220 Twine("unrolled with interleaving factor " + Twine(UF) +
1221 " (vectorization not beneficial)"));
1223 // We decided not to vectorize, but we may want to unroll.
1224 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1225 Unroller.vectorize(&LVL);
1227 // If we decided that it is *legal* to vectorize the loop then do it.
1228 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1232 // Report the vectorization decision.
1233 F->getContext().emitOptimizationRemark(
1234 DEBUG_TYPE, *F, L->getStartLoc(),
1235 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1236 ", unrolling interleave factor: " + Twine(UF) + ")");
1239 // Mark the loop as already vectorized to avoid vectorizing again.
1240 Hints.setAlreadyVectorized(L);
1242 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1246 void getAnalysisUsage(AnalysisUsage &AU) const override {
1247 AU.addRequiredID(LoopSimplifyID);
1248 AU.addRequiredID(LCSSAID);
1249 AU.addRequired<BlockFrequencyInfo>();
1250 AU.addRequired<DominatorTreeWrapperPass>();
1251 AU.addRequired<LoopInfo>();
1252 AU.addRequired<ScalarEvolution>();
1253 AU.addRequired<TargetTransformInfo>();
1254 AU.addPreserved<LoopInfo>();
1255 AU.addPreserved<DominatorTreeWrapperPass>();
1260 } // end anonymous namespace
1262 //===----------------------------------------------------------------------===//
1263 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1264 // LoopVectorizationCostModel.
1265 //===----------------------------------------------------------------------===//
1267 static Value *stripIntegerCast(Value *V) {
1268 if (CastInst *CI = dyn_cast<CastInst>(V))
1269 if (CI->getOperand(0)->getType()->isIntegerTy())
1270 return CI->getOperand(0);
1274 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1276 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1278 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1279 ValueToValueMap &PtrToStride,
1280 Value *Ptr, Value *OrigPtr = nullptr) {
1282 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1284 // If there is an entry in the map return the SCEV of the pointer with the
1285 // symbolic stride replaced by one.
1286 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1287 if (SI != PtrToStride.end()) {
1288 Value *StrideVal = SI->second;
1291 StrideVal = stripIntegerCast(StrideVal);
1293 // Replace symbolic stride by one.
1294 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1295 ValueToValueMap RewriteMap;
1296 RewriteMap[StrideVal] = One;
1299 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1300 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1305 // Otherwise, just return the SCEV of the original pointer.
1306 return SE->getSCEV(Ptr);
1309 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1310 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1311 ValueToValueMap &Strides) {
1312 // Get the stride replaced scev.
1313 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1314 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1315 assert(AR && "Invalid addrec expression");
1316 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1317 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1318 Pointers.push_back(Ptr);
1319 Starts.push_back(AR->getStart());
1320 Ends.push_back(ScEnd);
1321 IsWritePtr.push_back(WritePtr);
1322 DependencySetId.push_back(DepSetId);
1325 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1326 // We need to place the broadcast of invariant variables outside the loop.
1327 Instruction *Instr = dyn_cast<Instruction>(V);
1329 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1330 Instr->getParent()) != LoopVectorBody.end());
1331 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1333 // Place the code for broadcasting invariant variables in the new preheader.
1334 IRBuilder<>::InsertPointGuard Guard(Builder);
1336 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1338 // Broadcast the scalar into all locations in the vector.
1339 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1344 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1346 assert(Val->getType()->isVectorTy() && "Must be a vector");
1347 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1348 "Elem must be an integer");
1349 // Create the types.
1350 Type *ITy = Val->getType()->getScalarType();
1351 VectorType *Ty = cast<VectorType>(Val->getType());
1352 int VLen = Ty->getNumElements();
1353 SmallVector<Constant*, 8> Indices;
1355 // Create a vector of consecutive numbers from zero to VF.
1356 for (int i = 0; i < VLen; ++i) {
1357 int64_t Idx = Negate ? (-i) : i;
1358 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1361 // Add the consecutive indices to the vector value.
1362 Constant *Cv = ConstantVector::get(Indices);
1363 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1364 return Builder.CreateAdd(Val, Cv, "induction");
1367 /// \brief Find the operand of the GEP that should be checked for consecutive
1368 /// stores. This ignores trailing indices that have no effect on the final
1370 static unsigned getGEPInductionOperand(const DataLayout *DL,
1371 const GetElementPtrInst *Gep) {
1372 unsigned LastOperand = Gep->getNumOperands() - 1;
1373 unsigned GEPAllocSize = DL->getTypeAllocSize(
1374 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1376 // Walk backwards and try to peel off zeros.
1377 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1378 // Find the type we're currently indexing into.
1379 gep_type_iterator GEPTI = gep_type_begin(Gep);
1380 std::advance(GEPTI, LastOperand - 1);
1382 // If it's a type with the same allocation size as the result of the GEP we
1383 // can peel off the zero index.
1384 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1392 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1393 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1394 // Make sure that the pointer does not point to structs.
1395 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1398 // If this value is a pointer induction variable we know it is consecutive.
1399 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1400 if (Phi && Inductions.count(Phi)) {
1401 InductionInfo II = Inductions[Phi];
1402 if (IK_PtrInduction == II.IK)
1404 else if (IK_ReversePtrInduction == II.IK)
1408 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1412 unsigned NumOperands = Gep->getNumOperands();
1413 Value *GpPtr = Gep->getPointerOperand();
1414 // If this GEP value is a consecutive pointer induction variable and all of
1415 // the indices are constant then we know it is consecutive. We can
1416 Phi = dyn_cast<PHINode>(GpPtr);
1417 if (Phi && Inductions.count(Phi)) {
1419 // Make sure that the pointer does not point to structs.
1420 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1421 if (GepPtrType->getElementType()->isAggregateType())
1424 // Make sure that all of the index operands are loop invariant.
1425 for (unsigned i = 1; i < NumOperands; ++i)
1426 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1429 InductionInfo II = Inductions[Phi];
1430 if (IK_PtrInduction == II.IK)
1432 else if (IK_ReversePtrInduction == II.IK)
1436 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1438 // Check that all of the gep indices are uniform except for our induction
1440 for (unsigned i = 0; i != NumOperands; ++i)
1441 if (i != InductionOperand &&
1442 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1445 // We can emit wide load/stores only if the last non-zero index is the
1446 // induction variable.
1447 const SCEV *Last = nullptr;
1448 if (!Strides.count(Gep))
1449 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1451 // Because of the multiplication by a stride we can have a s/zext cast.
1452 // We are going to replace this stride by 1 so the cast is safe to ignore.
1454 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1455 // %0 = trunc i64 %indvars.iv to i32
1456 // %mul = mul i32 %0, %Stride1
1457 // %idxprom = zext i32 %mul to i64 << Safe cast.
1458 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1460 Last = replaceSymbolicStrideSCEV(SE, Strides,
1461 Gep->getOperand(InductionOperand), Gep);
1462 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1464 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1468 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1469 const SCEV *Step = AR->getStepRecurrence(*SE);
1471 // The memory is consecutive because the last index is consecutive
1472 // and all other indices are loop invariant.
1475 if (Step->isAllOnesValue())
1482 bool LoopVectorizationLegality::isUniform(Value *V) {
1483 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1486 InnerLoopVectorizer::VectorParts&
1487 InnerLoopVectorizer::getVectorValue(Value *V) {
1488 assert(V != Induction && "The new induction variable should not be used.");
1489 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1491 // If we have a stride that is replaced by one, do it here.
1492 if (Legal->hasStride(V))
1493 V = ConstantInt::get(V->getType(), 1);
1495 // If we have this scalar in the map, return it.
1496 if (WidenMap.has(V))
1497 return WidenMap.get(V);
1499 // If this scalar is unknown, assume that it is a constant or that it is
1500 // loop invariant. Broadcast V and save the value for future uses.
1501 Value *B = getBroadcastInstrs(V);
1502 return WidenMap.splat(V, B);
1505 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1506 assert(Vec->getType()->isVectorTy() && "Invalid type");
1507 SmallVector<Constant*, 8> ShuffleMask;
1508 for (unsigned i = 0; i < VF; ++i)
1509 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1511 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1512 ConstantVector::get(ShuffleMask),
1516 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1517 // Attempt to issue a wide load.
1518 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1519 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1521 assert((LI || SI) && "Invalid Load/Store instruction");
1523 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1524 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1525 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1526 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1527 // An alignment of 0 means target abi alignment. We need to use the scalar's
1528 // target abi alignment in such a case.
1530 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1531 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1532 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1533 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1535 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1536 return scalarizeInstruction(Instr, true);
1538 if (ScalarAllocatedSize != VectorElementSize)
1539 return scalarizeInstruction(Instr);
1541 // If the pointer is loop invariant or if it is non-consecutive,
1542 // scalarize the load.
1543 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1544 bool Reverse = ConsecutiveStride < 0;
1545 bool UniformLoad = LI && Legal->isUniform(Ptr);
1546 if (!ConsecutiveStride || UniformLoad)
1547 return scalarizeInstruction(Instr);
1549 Constant *Zero = Builder.getInt32(0);
1550 VectorParts &Entry = WidenMap.get(Instr);
1552 // Handle consecutive loads/stores.
1553 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1554 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1555 setDebugLocFromInst(Builder, Gep);
1556 Value *PtrOperand = Gep->getPointerOperand();
1557 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1558 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1560 // Create the new GEP with the new induction variable.
1561 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1562 Gep2->setOperand(0, FirstBasePtr);
1563 Gep2->setName("gep.indvar.base");
1564 Ptr = Builder.Insert(Gep2);
1566 setDebugLocFromInst(Builder, Gep);
1567 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1568 OrigLoop) && "Base ptr must be invariant");
1570 // The last index does not have to be the induction. It can be
1571 // consecutive and be a function of the index. For example A[I+1];
1572 unsigned NumOperands = Gep->getNumOperands();
1573 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1574 // Create the new GEP with the new induction variable.
1575 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1577 for (unsigned i = 0; i < NumOperands; ++i) {
1578 Value *GepOperand = Gep->getOperand(i);
1579 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1581 // Update last index or loop invariant instruction anchored in loop.
1582 if (i == InductionOperand ||
1583 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1584 assert((i == InductionOperand ||
1585 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1586 "Must be last index or loop invariant");
1588 VectorParts &GEPParts = getVectorValue(GepOperand);
1589 Value *Index = GEPParts[0];
1590 Index = Builder.CreateExtractElement(Index, Zero);
1591 Gep2->setOperand(i, Index);
1592 Gep2->setName("gep.indvar.idx");
1595 Ptr = Builder.Insert(Gep2);
1597 // Use the induction element ptr.
1598 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1599 setDebugLocFromInst(Builder, Ptr);
1600 VectorParts &PtrVal = getVectorValue(Ptr);
1601 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1606 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1607 "We do not allow storing to uniform addresses");
1608 setDebugLocFromInst(Builder, SI);
1609 // We don't want to update the value in the map as it might be used in
1610 // another expression. So don't use a reference type for "StoredVal".
1611 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1613 for (unsigned Part = 0; Part < UF; ++Part) {
1614 // Calculate the pointer for the specific unroll-part.
1615 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1618 // If we store to reverse consecutive memory locations then we need
1619 // to reverse the order of elements in the stored value.
1620 StoredVal[Part] = reverseVector(StoredVal[Part]);
1621 // If the address is consecutive but reversed, then the
1622 // wide store needs to start at the last vector element.
1623 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1624 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1627 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1628 DataTy->getPointerTo(AddressSpace));
1629 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1635 assert(LI && "Must have a load instruction");
1636 setDebugLocFromInst(Builder, LI);
1637 for (unsigned Part = 0; Part < UF; ++Part) {
1638 // Calculate the pointer for the specific unroll-part.
1639 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1642 // If the address is consecutive but reversed, then the
1643 // wide store needs to start at the last vector element.
1644 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1645 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1648 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1649 DataTy->getPointerTo(AddressSpace));
1650 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1651 cast<LoadInst>(LI)->setAlignment(Alignment);
1652 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1656 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1657 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1658 // Holds vector parameters or scalars, in case of uniform vals.
1659 SmallVector<VectorParts, 4> Params;
1661 setDebugLocFromInst(Builder, Instr);
1663 // Find all of the vectorized parameters.
1664 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1665 Value *SrcOp = Instr->getOperand(op);
1667 // If we are accessing the old induction variable, use the new one.
1668 if (SrcOp == OldInduction) {
1669 Params.push_back(getVectorValue(SrcOp));
1673 // Try using previously calculated values.
1674 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1676 // If the src is an instruction that appeared earlier in the basic block
1677 // then it should already be vectorized.
1678 if (SrcInst && OrigLoop->contains(SrcInst)) {
1679 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1680 // The parameter is a vector value from earlier.
1681 Params.push_back(WidenMap.get(SrcInst));
1683 // The parameter is a scalar from outside the loop. Maybe even a constant.
1684 VectorParts Scalars;
1685 Scalars.append(UF, SrcOp);
1686 Params.push_back(Scalars);
1690 assert(Params.size() == Instr->getNumOperands() &&
1691 "Invalid number of operands");
1693 // Does this instruction return a value ?
1694 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1696 Value *UndefVec = IsVoidRetTy ? nullptr :
1697 UndefValue::get(VectorType::get(Instr->getType(), VF));
1698 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1699 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1701 Instruction *InsertPt = Builder.GetInsertPoint();
1702 BasicBlock *IfBlock = Builder.GetInsertBlock();
1703 BasicBlock *CondBlock = nullptr;
1706 Loop *VectorLp = nullptr;
1707 if (IfPredicateStore) {
1708 assert(Instr->getParent()->getSinglePredecessor() &&
1709 "Only support single predecessor blocks");
1710 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1711 Instr->getParent());
1712 VectorLp = LI->getLoopFor(IfBlock);
1713 assert(VectorLp && "Must have a loop for this block");
1716 // For each vector unroll 'part':
1717 for (unsigned Part = 0; Part < UF; ++Part) {
1718 // For each scalar that we create:
1719 for (unsigned Width = 0; Width < VF; ++Width) {
1722 Value *Cmp = nullptr;
1723 if (IfPredicateStore) {
1724 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1725 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1726 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1727 LoopVectorBody.push_back(CondBlock);
1728 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1729 // Update Builder with newly created basic block.
1730 Builder.SetInsertPoint(InsertPt);
1733 Instruction *Cloned = Instr->clone();
1735 Cloned->setName(Instr->getName() + ".cloned");
1736 // Replace the operands of the cloned instructions with extracted scalars.
1737 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1738 Value *Op = Params[op][Part];
1739 // Param is a vector. Need to extract the right lane.
1740 if (Op->getType()->isVectorTy())
1741 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1742 Cloned->setOperand(op, Op);
1745 // Place the cloned scalar in the new loop.
1746 Builder.Insert(Cloned);
1748 // If the original scalar returns a value we need to place it in a vector
1749 // so that future users will be able to use it.
1751 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1752 Builder.getInt32(Width));
1754 if (IfPredicateStore) {
1755 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1756 LoopVectorBody.push_back(NewIfBlock);
1757 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1758 Builder.SetInsertPoint(InsertPt);
1759 Instruction *OldBr = IfBlock->getTerminator();
1760 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1761 OldBr->eraseFromParent();
1762 IfBlock = NewIfBlock;
1768 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1772 if (Instruction *I = dyn_cast<Instruction>(V))
1773 return I->getParent() == Loc->getParent() ? I : nullptr;
1777 std::pair<Instruction *, Instruction *>
1778 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1779 Instruction *tnullptr = nullptr;
1780 if (!Legal->mustCheckStrides())
1781 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1783 IRBuilder<> ChkBuilder(Loc);
1786 Value *Check = nullptr;
1787 Instruction *FirstInst = nullptr;
1788 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1789 SE = Legal->strides_end();
1791 Value *Ptr = stripIntegerCast(*SI);
1792 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1794 // Store the first instruction we create.
1795 FirstInst = getFirstInst(FirstInst, C, Loc);
1797 Check = ChkBuilder.CreateOr(Check, C);
1802 // We have to do this trickery because the IRBuilder might fold the check to a
1803 // constant expression in which case there is no Instruction anchored in a
1805 LLVMContext &Ctx = Loc->getContext();
1806 Instruction *TheCheck =
1807 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1808 ChkBuilder.Insert(TheCheck, "stride.not.one");
1809 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1811 return std::make_pair(FirstInst, TheCheck);
1814 std::pair<Instruction *, Instruction *>
1815 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1816 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1817 Legal->getRuntimePointerCheck();
1819 Instruction *tnullptr = nullptr;
1820 if (!PtrRtCheck->Need)
1821 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1823 unsigned NumPointers = PtrRtCheck->Pointers.size();
1824 SmallVector<TrackingVH<Value> , 2> Starts;
1825 SmallVector<TrackingVH<Value> , 2> Ends;
1827 LLVMContext &Ctx = Loc->getContext();
1828 SCEVExpander Exp(*SE, "induction");
1829 Instruction *FirstInst = nullptr;
1831 for (unsigned i = 0; i < NumPointers; ++i) {
1832 Value *Ptr = PtrRtCheck->Pointers[i];
1833 const SCEV *Sc = SE->getSCEV(Ptr);
1835 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1836 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1838 Starts.push_back(Ptr);
1839 Ends.push_back(Ptr);
1841 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1842 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1844 // Use this type for pointer arithmetic.
1845 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1847 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1848 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1849 Starts.push_back(Start);
1850 Ends.push_back(End);
1854 IRBuilder<> ChkBuilder(Loc);
1855 // Our instructions might fold to a constant.
1856 Value *MemoryRuntimeCheck = nullptr;
1857 for (unsigned i = 0; i < NumPointers; ++i) {
1858 for (unsigned j = i+1; j < NumPointers; ++j) {
1859 // No need to check if two readonly pointers intersect.
1860 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1863 // Only need to check pointers between two different dependency sets.
1864 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1867 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1868 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1870 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1871 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1872 "Trying to bounds check pointers with different address spaces");
1874 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1875 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1877 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1878 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1879 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1880 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1882 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1883 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1884 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1885 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1886 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1887 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1888 if (MemoryRuntimeCheck) {
1889 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1891 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1893 MemoryRuntimeCheck = IsConflict;
1897 // We have to do this trickery because the IRBuilder might fold the check to a
1898 // constant expression in which case there is no Instruction anchored in a
1900 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1901 ConstantInt::getTrue(Ctx));
1902 ChkBuilder.Insert(Check, "memcheck.conflict");
1903 FirstInst = getFirstInst(FirstInst, Check, Loc);
1904 return std::make_pair(FirstInst, Check);
1907 void InnerLoopVectorizer::createEmptyLoop() {
1909 In this function we generate a new loop. The new loop will contain
1910 the vectorized instructions while the old loop will continue to run the
1913 [ ] <-- vector loop bypass (may consist of multiple blocks).
1916 | [ ] <-- vector pre header.
1920 | [ ]_| <-- vector loop.
1923 >[ ] <--- middle-block.
1926 | [ ] <--- new preheader.
1930 | [ ]_| <-- old scalar loop to handle remainder.
1933 >[ ] <-- exit block.
1937 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1938 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1939 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1940 assert(ExitBlock && "Must have an exit block");
1942 // Some loops have a single integer induction variable, while other loops
1943 // don't. One example is c++ iterators that often have multiple pointer
1944 // induction variables. In the code below we also support a case where we
1945 // don't have a single induction variable.
1946 OldInduction = Legal->getInduction();
1947 Type *IdxTy = Legal->getWidestInductionType();
1949 // Find the loop boundaries.
1950 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1951 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1953 // The exit count might have the type of i64 while the phi is i32. This can
1954 // happen if we have an induction variable that is sign extended before the
1955 // compare. The only way that we get a backedge taken count is that the
1956 // induction variable was signed and as such will not overflow. In such a case
1957 // truncation is legal.
1958 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1959 IdxTy->getPrimitiveSizeInBits())
1960 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1962 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1963 // Get the total trip count from the count by adding 1.
1964 ExitCount = SE->getAddExpr(ExitCount,
1965 SE->getConstant(ExitCount->getType(), 1));
1967 // Expand the trip count and place the new instructions in the preheader.
1968 // Notice that the pre-header does not change, only the loop body.
1969 SCEVExpander Exp(*SE, "induction");
1971 // Count holds the overall loop count (N).
1972 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1973 BypassBlock->getTerminator());
1975 // The loop index does not have to start at Zero. Find the original start
1976 // value from the induction PHI node. If we don't have an induction variable
1977 // then we know that it starts at zero.
1978 Builder.SetInsertPoint(BypassBlock->getTerminator());
1979 Value *StartIdx = ExtendedIdx = OldInduction ?
1980 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1982 ConstantInt::get(IdxTy, 0);
1984 assert(BypassBlock && "Invalid loop structure");
1985 LoopBypassBlocks.push_back(BypassBlock);
1987 // Split the single block loop into the two loop structure described above.
1988 BasicBlock *VectorPH =
1989 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1990 BasicBlock *VecBody =
1991 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1992 BasicBlock *MiddleBlock =
1993 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1994 BasicBlock *ScalarPH =
1995 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1997 // Create and register the new vector loop.
1998 Loop* Lp = new Loop();
1999 Loop *ParentLoop = OrigLoop->getParentLoop();
2001 // Insert the new loop into the loop nest and register the new basic blocks
2002 // before calling any utilities such as SCEV that require valid LoopInfo.
2004 ParentLoop->addChildLoop(Lp);
2005 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2006 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2007 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2009 LI->addTopLevelLoop(Lp);
2011 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2013 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2015 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2017 // Generate the induction variable.
2018 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2019 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2020 // The loop step is equal to the vectorization factor (num of SIMD elements)
2021 // times the unroll factor (num of SIMD instructions).
2022 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2024 // This is the IR builder that we use to add all of the logic for bypassing
2025 // the new vector loop.
2026 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2027 setDebugLocFromInst(BypassBuilder,
2028 getDebugLocFromInstOrOperands(OldInduction));
2030 // We may need to extend the index in case there is a type mismatch.
2031 // We know that the count starts at zero and does not overflow.
2032 if (Count->getType() != IdxTy) {
2033 // The exit count can be of pointer type. Convert it to the correct
2035 if (ExitCount->getType()->isPointerTy())
2036 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2038 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2041 // Add the start index to the loop count to get the new end index.
2042 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2044 // Now we need to generate the expression for N - (N % VF), which is
2045 // the part that the vectorized body will execute.
2046 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2047 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2048 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2049 "end.idx.rnd.down");
2051 // Now, compare the new count to zero. If it is zero skip the vector loop and
2052 // jump to the scalar loop.
2053 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
2056 BasicBlock *LastBypassBlock = BypassBlock;
2058 // Generate the code to check that the strides we assumed to be one are really
2059 // one. We want the new basic block to start at the first instruction in a
2060 // sequence of instructions that form a check.
2061 Instruction *StrideCheck;
2062 Instruction *FirstCheckInst;
2063 std::tie(FirstCheckInst, StrideCheck) =
2064 addStrideCheck(BypassBlock->getTerminator());
2066 // Create a new block containing the stride check.
2067 BasicBlock *CheckBlock =
2068 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2070 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2071 LoopBypassBlocks.push_back(CheckBlock);
2073 // Replace the branch into the memory check block with a conditional branch
2074 // for the "few elements case".
2075 Instruction *OldTerm = BypassBlock->getTerminator();
2076 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2077 OldTerm->eraseFromParent();
2080 LastBypassBlock = CheckBlock;
2083 // Generate the code that checks in runtime if arrays overlap. We put the
2084 // checks into a separate block to make the more common case of few elements
2086 Instruction *MemRuntimeCheck;
2087 std::tie(FirstCheckInst, MemRuntimeCheck) =
2088 addRuntimeCheck(LastBypassBlock->getTerminator());
2089 if (MemRuntimeCheck) {
2090 // Create a new block containing the memory check.
2091 BasicBlock *CheckBlock =
2092 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2094 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2095 LoopBypassBlocks.push_back(CheckBlock);
2097 // Replace the branch into the memory check block with a conditional branch
2098 // for the "few elements case".
2099 Instruction *OldTerm = LastBypassBlock->getTerminator();
2100 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2101 OldTerm->eraseFromParent();
2103 Cmp = MemRuntimeCheck;
2104 LastBypassBlock = CheckBlock;
2107 LastBypassBlock->getTerminator()->eraseFromParent();
2108 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2111 // We are going to resume the execution of the scalar loop.
2112 // Go over all of the induction variables that we found and fix the
2113 // PHIs that are left in the scalar version of the loop.
2114 // The starting values of PHI nodes depend on the counter of the last
2115 // iteration in the vectorized loop.
2116 // If we come from a bypass edge then we need to start from the original
2119 // This variable saves the new starting index for the scalar loop.
2120 PHINode *ResumeIndex = nullptr;
2121 LoopVectorizationLegality::InductionList::iterator I, E;
2122 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2123 // Set builder to point to last bypass block.
2124 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2125 for (I = List->begin(), E = List->end(); I != E; ++I) {
2126 PHINode *OrigPhi = I->first;
2127 LoopVectorizationLegality::InductionInfo II = I->second;
2129 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2130 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2131 MiddleBlock->getTerminator());
2132 // We might have extended the type of the induction variable but we need a
2133 // truncated version for the scalar loop.
2134 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2135 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2136 MiddleBlock->getTerminator()) : nullptr;
2138 Value *EndValue = nullptr;
2140 case LoopVectorizationLegality::IK_NoInduction:
2141 llvm_unreachable("Unknown induction");
2142 case LoopVectorizationLegality::IK_IntInduction: {
2143 // Handle the integer induction counter.
2144 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2146 // We have the canonical induction variable.
2147 if (OrigPhi == OldInduction) {
2148 // Create a truncated version of the resume value for the scalar loop,
2149 // we might have promoted the type to a larger width.
2151 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2152 // The new PHI merges the original incoming value, in case of a bypass,
2153 // or the value at the end of the vectorized loop.
2154 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2155 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2156 TruncResumeVal->addIncoming(EndValue, VecBody);
2158 // We know what the end value is.
2159 EndValue = IdxEndRoundDown;
2160 // We also know which PHI node holds it.
2161 ResumeIndex = ResumeVal;
2165 // Not the canonical induction variable - add the vector loop count to the
2167 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2168 II.StartValue->getType(),
2170 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2173 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2174 // Convert the CountRoundDown variable to the PHI size.
2175 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2176 II.StartValue->getType(),
2178 // Handle reverse integer induction counter.
2179 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2182 case LoopVectorizationLegality::IK_PtrInduction: {
2183 // For pointer induction variables, calculate the offset using
2185 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2189 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2190 // The value at the end of the loop for the reverse pointer is calculated
2191 // by creating a GEP with a negative index starting from the start value.
2192 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2193 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2195 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2201 // The new PHI merges the original incoming value, in case of a bypass,
2202 // or the value at the end of the vectorized loop.
2203 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2204 if (OrigPhi == OldInduction)
2205 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2207 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2209 ResumeVal->addIncoming(EndValue, VecBody);
2211 // Fix the scalar body counter (PHI node).
2212 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2213 // The old inductions phi node in the scalar body needs the truncated value.
2214 if (OrigPhi == OldInduction)
2215 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2217 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2220 // If we are generating a new induction variable then we also need to
2221 // generate the code that calculates the exit value. This value is not
2222 // simply the end of the counter because we may skip the vectorized body
2223 // in case of a runtime check.
2225 assert(!ResumeIndex && "Unexpected resume value found");
2226 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2227 MiddleBlock->getTerminator());
2228 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2229 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2230 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2233 // Make sure that we found the index where scalar loop needs to continue.
2234 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2235 "Invalid resume Index");
2237 // Add a check in the middle block to see if we have completed
2238 // all of the iterations in the first vector loop.
2239 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2240 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2241 ResumeIndex, "cmp.n",
2242 MiddleBlock->getTerminator());
2244 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2245 // Remove the old terminator.
2246 MiddleBlock->getTerminator()->eraseFromParent();
2248 // Create i+1 and fill the PHINode.
2249 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2250 Induction->addIncoming(StartIdx, VectorPH);
2251 Induction->addIncoming(NextIdx, VecBody);
2252 // Create the compare.
2253 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2254 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2256 // Now we have two terminators. Remove the old one from the block.
2257 VecBody->getTerminator()->eraseFromParent();
2259 // Get ready to start creating new instructions into the vectorized body.
2260 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2263 LoopVectorPreHeader = VectorPH;
2264 LoopScalarPreHeader = ScalarPH;
2265 LoopMiddleBlock = MiddleBlock;
2266 LoopExitBlock = ExitBlock;
2267 LoopVectorBody.push_back(VecBody);
2268 LoopScalarBody = OldBasicBlock;
2270 LoopVectorizeHints Hints(Lp, true);
2271 Hints.setAlreadyVectorized(Lp);
2274 /// This function returns the identity element (or neutral element) for
2275 /// the operation K.
2277 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2282 // Adding, Xoring, Oring zero to a number does not change it.
2283 return ConstantInt::get(Tp, 0);
2284 case RK_IntegerMult:
2285 // Multiplying a number by 1 does not change it.
2286 return ConstantInt::get(Tp, 1);
2288 // AND-ing a number with an all-1 value does not change it.
2289 return ConstantInt::get(Tp, -1, true);
2291 // Multiplying a number by 1 does not change it.
2292 return ConstantFP::get(Tp, 1.0L);
2294 // Adding zero to a number does not change it.
2295 return ConstantFP::get(Tp, 0.0L);
2297 llvm_unreachable("Unknown reduction kind");
2301 /// This function translates the reduction kind to an LLVM binary operator.
2303 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2305 case LoopVectorizationLegality::RK_IntegerAdd:
2306 return Instruction::Add;
2307 case LoopVectorizationLegality::RK_IntegerMult:
2308 return Instruction::Mul;
2309 case LoopVectorizationLegality::RK_IntegerOr:
2310 return Instruction::Or;
2311 case LoopVectorizationLegality::RK_IntegerAnd:
2312 return Instruction::And;
2313 case LoopVectorizationLegality::RK_IntegerXor:
2314 return Instruction::Xor;
2315 case LoopVectorizationLegality::RK_FloatMult:
2316 return Instruction::FMul;
2317 case LoopVectorizationLegality::RK_FloatAdd:
2318 return Instruction::FAdd;
2319 case LoopVectorizationLegality::RK_IntegerMinMax:
2320 return Instruction::ICmp;
2321 case LoopVectorizationLegality::RK_FloatMinMax:
2322 return Instruction::FCmp;
2324 llvm_unreachable("Unknown reduction operation");
2328 Value *createMinMaxOp(IRBuilder<> &Builder,
2329 LoopVectorizationLegality::MinMaxReductionKind RK,
2332 CmpInst::Predicate P = CmpInst::ICMP_NE;
2335 llvm_unreachable("Unknown min/max reduction kind");
2336 case LoopVectorizationLegality::MRK_UIntMin:
2337 P = CmpInst::ICMP_ULT;
2339 case LoopVectorizationLegality::MRK_UIntMax:
2340 P = CmpInst::ICMP_UGT;
2342 case LoopVectorizationLegality::MRK_SIntMin:
2343 P = CmpInst::ICMP_SLT;
2345 case LoopVectorizationLegality::MRK_SIntMax:
2346 P = CmpInst::ICMP_SGT;
2348 case LoopVectorizationLegality::MRK_FloatMin:
2349 P = CmpInst::FCMP_OLT;
2351 case LoopVectorizationLegality::MRK_FloatMax:
2352 P = CmpInst::FCMP_OGT;
2357 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2358 RK == LoopVectorizationLegality::MRK_FloatMax)
2359 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2361 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2363 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2368 struct CSEDenseMapInfo {
2369 static bool canHandle(Instruction *I) {
2370 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2371 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2373 static inline Instruction *getEmptyKey() {
2374 return DenseMapInfo<Instruction *>::getEmptyKey();
2376 static inline Instruction *getTombstoneKey() {
2377 return DenseMapInfo<Instruction *>::getTombstoneKey();
2379 static unsigned getHashValue(Instruction *I) {
2380 assert(canHandle(I) && "Unknown instruction!");
2381 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2382 I->value_op_end()));
2384 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2385 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2386 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2388 return LHS->isIdenticalTo(RHS);
2393 /// \brief Check whether this block is a predicated block.
2394 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2395 /// = ...; " blocks. We start with one vectorized basic block. For every
2396 /// conditional block we split this vectorized block. Therefore, every second
2397 /// block will be a predicated one.
2398 static bool isPredicatedBlock(unsigned BlockNum) {
2399 return BlockNum % 2;
2402 ///\brief Perform cse of induction variable instructions.
2403 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2404 // Perform simple cse.
2405 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2406 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2407 BasicBlock *BB = BBs[i];
2408 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2409 Instruction *In = I++;
2411 if (!CSEDenseMapInfo::canHandle(In))
2414 // Check if we can replace this instruction with any of the
2415 // visited instructions.
2416 if (Instruction *V = CSEMap.lookup(In)) {
2417 In->replaceAllUsesWith(V);
2418 In->eraseFromParent();
2421 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2422 // ...;" blocks for predicated stores. Every second block is a predicated
2424 if (isPredicatedBlock(i))
2432 /// \brief Adds a 'fast' flag to floating point operations.
2433 static Value *addFastMathFlag(Value *V) {
2434 if (isa<FPMathOperator>(V)){
2435 FastMathFlags Flags;
2436 Flags.setUnsafeAlgebra();
2437 cast<Instruction>(V)->setFastMathFlags(Flags);
2442 void InnerLoopVectorizer::vectorizeLoop() {
2443 //===------------------------------------------------===//
2445 // Notice: any optimization or new instruction that go
2446 // into the code below should be also be implemented in
2449 //===------------------------------------------------===//
2450 Constant *Zero = Builder.getInt32(0);
2452 // In order to support reduction variables we need to be able to vectorize
2453 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2454 // stages. First, we create a new vector PHI node with no incoming edges.
2455 // We use this value when we vectorize all of the instructions that use the
2456 // PHI. Next, after all of the instructions in the block are complete we
2457 // add the new incoming edges to the PHI. At this point all of the
2458 // instructions in the basic block are vectorized, so we can use them to
2459 // construct the PHI.
2460 PhiVector RdxPHIsToFix;
2462 // Scan the loop in a topological order to ensure that defs are vectorized
2464 LoopBlocksDFS DFS(OrigLoop);
2467 // Vectorize all of the blocks in the original loop.
2468 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2469 be = DFS.endRPO(); bb != be; ++bb)
2470 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2472 // At this point every instruction in the original loop is widened to
2473 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2474 // that we vectorized. The PHI nodes are currently empty because we did
2475 // not want to introduce cycles. Notice that the remaining PHI nodes
2476 // that we need to fix are reduction variables.
2478 // Create the 'reduced' values for each of the induction vars.
2479 // The reduced values are the vector values that we scalarize and combine
2480 // after the loop is finished.
2481 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2483 PHINode *RdxPhi = *it;
2484 assert(RdxPhi && "Unable to recover vectorized PHI");
2486 // Find the reduction variable descriptor.
2487 assert(Legal->getReductionVars()->count(RdxPhi) &&
2488 "Unable to find the reduction variable");
2489 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2490 (*Legal->getReductionVars())[RdxPhi];
2492 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2494 // We need to generate a reduction vector from the incoming scalar.
2495 // To do so, we need to generate the 'identity' vector and override
2496 // one of the elements with the incoming scalar reduction. We need
2497 // to do it in the vector-loop preheader.
2498 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2500 // This is the vector-clone of the value that leaves the loop.
2501 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2502 Type *VecTy = VectorExit[0]->getType();
2504 // Find the reduction identity variable. Zero for addition, or, xor,
2505 // one for multiplication, -1 for And.
2508 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2509 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2510 // MinMax reduction have the start value as their identify.
2512 VectorStart = Identity = RdxDesc.StartValue;
2514 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2519 // Handle other reduction kinds:
2521 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2522 VecTy->getScalarType());
2525 // This vector is the Identity vector where the first element is the
2526 // incoming scalar reduction.
2527 VectorStart = RdxDesc.StartValue;
2529 Identity = ConstantVector::getSplat(VF, Iden);
2531 // This vector is the Identity vector where the first element is the
2532 // incoming scalar reduction.
2533 VectorStart = Builder.CreateInsertElement(Identity,
2534 RdxDesc.StartValue, Zero);
2538 // Fix the vector-loop phi.
2539 // We created the induction variable so we know that the
2540 // preheader is the first entry.
2541 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2543 // Reductions do not have to start at zero. They can start with
2544 // any loop invariant values.
2545 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2546 BasicBlock *Latch = OrigLoop->getLoopLatch();
2547 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2548 VectorParts &Val = getVectorValue(LoopVal);
2549 for (unsigned part = 0; part < UF; ++part) {
2550 // Make sure to add the reduction stat value only to the
2551 // first unroll part.
2552 Value *StartVal = (part == 0) ? VectorStart : Identity;
2553 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2554 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2555 LoopVectorBody.back());
2558 // Before each round, move the insertion point right between
2559 // the PHIs and the values we are going to write.
2560 // This allows us to write both PHINodes and the extractelement
2562 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2564 VectorParts RdxParts;
2565 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2566 for (unsigned part = 0; part < UF; ++part) {
2567 // This PHINode contains the vectorized reduction variable, or
2568 // the initial value vector, if we bypass the vector loop.
2569 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2570 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2571 Value *StartVal = (part == 0) ? VectorStart : Identity;
2572 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2573 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2574 NewPhi->addIncoming(RdxExitVal[part],
2575 LoopVectorBody.back());
2576 RdxParts.push_back(NewPhi);
2579 // Reduce all of the unrolled parts into a single vector.
2580 Value *ReducedPartRdx = RdxParts[0];
2581 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2582 setDebugLocFromInst(Builder, ReducedPartRdx);
2583 for (unsigned part = 1; part < UF; ++part) {
2584 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2585 // Floating point operations had to be 'fast' to enable the reduction.
2586 ReducedPartRdx = addFastMathFlag(
2587 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2588 ReducedPartRdx, "bin.rdx"));
2590 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2591 ReducedPartRdx, RdxParts[part]);
2595 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2596 // and vector ops, reducing the set of values being computed by half each
2598 assert(isPowerOf2_32(VF) &&
2599 "Reduction emission only supported for pow2 vectors!");
2600 Value *TmpVec = ReducedPartRdx;
2601 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2602 for (unsigned i = VF; i != 1; i >>= 1) {
2603 // Move the upper half of the vector to the lower half.
2604 for (unsigned j = 0; j != i/2; ++j)
2605 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2607 // Fill the rest of the mask with undef.
2608 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2609 UndefValue::get(Builder.getInt32Ty()));
2612 Builder.CreateShuffleVector(TmpVec,
2613 UndefValue::get(TmpVec->getType()),
2614 ConstantVector::get(ShuffleMask),
2617 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2618 // Floating point operations had to be 'fast' to enable the reduction.
2619 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2620 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2622 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2625 // The result is in the first element of the vector.
2626 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2627 Builder.getInt32(0));
2630 // Now, we need to fix the users of the reduction variable
2631 // inside and outside of the scalar remainder loop.
2632 // We know that the loop is in LCSSA form. We need to update the
2633 // PHI nodes in the exit blocks.
2634 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2635 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2636 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2637 if (!LCSSAPhi) break;
2639 // All PHINodes need to have a single entry edge, or two if
2640 // we already fixed them.
2641 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2643 // We found our reduction value exit-PHI. Update it with the
2644 // incoming bypass edge.
2645 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2646 // Add an edge coming from the bypass.
2647 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2650 }// end of the LCSSA phi scan.
2652 // Fix the scalar loop reduction variable with the incoming reduction sum
2653 // from the vector body and from the backedge value.
2654 int IncomingEdgeBlockIdx =
2655 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2656 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2657 // Pick the other block.
2658 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2659 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2660 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2661 }// end of for each redux variable.
2665 // Remove redundant induction instructions.
2666 cse(LoopVectorBody);
2669 void InnerLoopVectorizer::fixLCSSAPHIs() {
2670 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2671 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2672 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2673 if (!LCSSAPhi) break;
2674 if (LCSSAPhi->getNumIncomingValues() == 1)
2675 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2680 InnerLoopVectorizer::VectorParts
2681 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2682 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2685 // Look for cached value.
2686 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2687 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2688 if (ECEntryIt != MaskCache.end())
2689 return ECEntryIt->second;
2691 VectorParts SrcMask = createBlockInMask(Src);
2693 // The terminator has to be a branch inst!
2694 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2695 assert(BI && "Unexpected terminator found");
2697 if (BI->isConditional()) {
2698 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2700 if (BI->getSuccessor(0) != Dst)
2701 for (unsigned part = 0; part < UF; ++part)
2702 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2704 for (unsigned part = 0; part < UF; ++part)
2705 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2707 MaskCache[Edge] = EdgeMask;
2711 MaskCache[Edge] = SrcMask;
2715 InnerLoopVectorizer::VectorParts
2716 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2717 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2719 // Loop incoming mask is all-one.
2720 if (OrigLoop->getHeader() == BB) {
2721 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2722 return getVectorValue(C);
2725 // This is the block mask. We OR all incoming edges, and with zero.
2726 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2727 VectorParts BlockMask = getVectorValue(Zero);
2730 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2731 VectorParts EM = createEdgeMask(*it, BB);
2732 for (unsigned part = 0; part < UF; ++part)
2733 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2739 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2740 InnerLoopVectorizer::VectorParts &Entry,
2741 unsigned UF, unsigned VF, PhiVector *PV) {
2742 PHINode* P = cast<PHINode>(PN);
2743 // Handle reduction variables:
2744 if (Legal->getReductionVars()->count(P)) {
2745 for (unsigned part = 0; part < UF; ++part) {
2746 // This is phase one of vectorizing PHIs.
2747 Type *VecTy = (VF == 1) ? PN->getType() :
2748 VectorType::get(PN->getType(), VF);
2749 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2750 LoopVectorBody.back()-> getFirstInsertionPt());
2756 setDebugLocFromInst(Builder, P);
2757 // Check for PHI nodes that are lowered to vector selects.
2758 if (P->getParent() != OrigLoop->getHeader()) {
2759 // We know that all PHIs in non-header blocks are converted into
2760 // selects, so we don't have to worry about the insertion order and we
2761 // can just use the builder.
2762 // At this point we generate the predication tree. There may be
2763 // duplications since this is a simple recursive scan, but future
2764 // optimizations will clean it up.
2766 unsigned NumIncoming = P->getNumIncomingValues();
2768 // Generate a sequence of selects of the form:
2769 // SELECT(Mask3, In3,
2770 // SELECT(Mask2, In2,
2772 for (unsigned In = 0; In < NumIncoming; In++) {
2773 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2775 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2777 for (unsigned part = 0; part < UF; ++part) {
2778 // We might have single edge PHIs (blocks) - use an identity
2779 // 'select' for the first PHI operand.
2781 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2784 // Select between the current value and the previous incoming edge
2785 // based on the incoming mask.
2786 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2787 Entry[part], "predphi");
2793 // This PHINode must be an induction variable.
2794 // Make sure that we know about it.
2795 assert(Legal->getInductionVars()->count(P) &&
2796 "Not an induction variable");
2798 LoopVectorizationLegality::InductionInfo II =
2799 Legal->getInductionVars()->lookup(P);
2802 case LoopVectorizationLegality::IK_NoInduction:
2803 llvm_unreachable("Unknown induction");
2804 case LoopVectorizationLegality::IK_IntInduction: {
2805 assert(P->getType() == II.StartValue->getType() && "Types must match");
2806 Type *PhiTy = P->getType();
2808 if (P == OldInduction) {
2809 // Handle the canonical induction variable. We might have had to
2811 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2813 // Handle other induction variables that are now based on the
2815 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2817 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2818 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2821 Broadcasted = getBroadcastInstrs(Broadcasted);
2822 // After broadcasting the induction variable we need to make the vector
2823 // consecutive by adding 0, 1, 2, etc.
2824 for (unsigned part = 0; part < UF; ++part)
2825 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2828 case LoopVectorizationLegality::IK_ReverseIntInduction:
2829 case LoopVectorizationLegality::IK_PtrInduction:
2830 case LoopVectorizationLegality::IK_ReversePtrInduction:
2831 // Handle reverse integer and pointer inductions.
2832 Value *StartIdx = ExtendedIdx;
2833 // This is the normalized GEP that starts counting at zero.
2834 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2837 // Handle the reverse integer induction variable case.
2838 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2839 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2840 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2842 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2845 // This is a new value so do not hoist it out.
2846 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2847 // After broadcasting the induction variable we need to make the
2848 // vector consecutive by adding ... -3, -2, -1, 0.
2849 for (unsigned part = 0; part < UF; ++part)
2850 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2855 // Handle the pointer induction variable case.
2856 assert(P->getType()->isPointerTy() && "Unexpected type.");
2858 // Is this a reverse induction ptr or a consecutive induction ptr.
2859 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2862 // This is the vector of results. Notice that we don't generate
2863 // vector geps because scalar geps result in better code.
2864 for (unsigned part = 0; part < UF; ++part) {
2866 int EltIndex = (part) * (Reverse ? -1 : 1);
2867 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2870 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2872 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2874 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2876 Entry[part] = SclrGep;
2880 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2881 for (unsigned int i = 0; i < VF; ++i) {
2882 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2883 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2886 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2888 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2890 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2892 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2893 Builder.getInt32(i),
2896 Entry[part] = VecVal;
2902 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2903 // For each instruction in the old loop.
2904 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2905 VectorParts &Entry = WidenMap.get(it);
2906 switch (it->getOpcode()) {
2907 case Instruction::Br:
2908 // Nothing to do for PHIs and BR, since we already took care of the
2909 // loop control flow instructions.
2911 case Instruction::PHI:{
2912 // Vectorize PHINodes.
2913 widenPHIInstruction(it, Entry, UF, VF, PV);
2917 case Instruction::Add:
2918 case Instruction::FAdd:
2919 case Instruction::Sub:
2920 case Instruction::FSub:
2921 case Instruction::Mul:
2922 case Instruction::FMul:
2923 case Instruction::UDiv:
2924 case Instruction::SDiv:
2925 case Instruction::FDiv:
2926 case Instruction::URem:
2927 case Instruction::SRem:
2928 case Instruction::FRem:
2929 case Instruction::Shl:
2930 case Instruction::LShr:
2931 case Instruction::AShr:
2932 case Instruction::And:
2933 case Instruction::Or:
2934 case Instruction::Xor: {
2935 // Just widen binops.
2936 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2937 setDebugLocFromInst(Builder, BinOp);
2938 VectorParts &A = getVectorValue(it->getOperand(0));
2939 VectorParts &B = getVectorValue(it->getOperand(1));
2941 // Use this vector value for all users of the original instruction.
2942 for (unsigned Part = 0; Part < UF; ++Part) {
2943 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2945 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2946 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2947 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2948 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2949 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2951 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2952 VecOp->setIsExact(BinOp->isExact());
2954 // Copy the fast-math flags.
2955 if (VecOp && isa<FPMathOperator>(V))
2956 VecOp->setFastMathFlags(it->getFastMathFlags());
2962 case Instruction::Select: {
2964 // If the selector is loop invariant we can create a select
2965 // instruction with a scalar condition. Otherwise, use vector-select.
2966 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2968 setDebugLocFromInst(Builder, it);
2970 // The condition can be loop invariant but still defined inside the
2971 // loop. This means that we can't just use the original 'cond' value.
2972 // We have to take the 'vectorized' value and pick the first lane.
2973 // Instcombine will make this a no-op.
2974 VectorParts &Cond = getVectorValue(it->getOperand(0));
2975 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2976 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2978 Value *ScalarCond = (VF == 1) ? Cond[0] :
2979 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2981 for (unsigned Part = 0; Part < UF; ++Part) {
2982 Entry[Part] = Builder.CreateSelect(
2983 InvariantCond ? ScalarCond : Cond[Part],
2990 case Instruction::ICmp:
2991 case Instruction::FCmp: {
2992 // Widen compares. Generate vector compares.
2993 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2994 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2995 setDebugLocFromInst(Builder, it);
2996 VectorParts &A = getVectorValue(it->getOperand(0));
2997 VectorParts &B = getVectorValue(it->getOperand(1));
2998 for (unsigned Part = 0; Part < UF; ++Part) {
3001 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3003 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3009 case Instruction::Store:
3010 case Instruction::Load:
3011 vectorizeMemoryInstruction(it);
3013 case Instruction::ZExt:
3014 case Instruction::SExt:
3015 case Instruction::FPToUI:
3016 case Instruction::FPToSI:
3017 case Instruction::FPExt:
3018 case Instruction::PtrToInt:
3019 case Instruction::IntToPtr:
3020 case Instruction::SIToFP:
3021 case Instruction::UIToFP:
3022 case Instruction::Trunc:
3023 case Instruction::FPTrunc:
3024 case Instruction::BitCast: {
3025 CastInst *CI = dyn_cast<CastInst>(it);
3026 setDebugLocFromInst(Builder, it);
3027 /// Optimize the special case where the source is the induction
3028 /// variable. Notice that we can only optimize the 'trunc' case
3029 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3030 /// c. other casts depend on pointer size.
3031 if (CI->getOperand(0) == OldInduction &&
3032 it->getOpcode() == Instruction::Trunc) {
3033 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3035 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3036 for (unsigned Part = 0; Part < UF; ++Part)
3037 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3040 /// Vectorize casts.
3041 Type *DestTy = (VF == 1) ? CI->getType() :
3042 VectorType::get(CI->getType(), VF);
3044 VectorParts &A = getVectorValue(it->getOperand(0));
3045 for (unsigned Part = 0; Part < UF; ++Part)
3046 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3050 case Instruction::Call: {
3051 // Ignore dbg intrinsics.
3052 if (isa<DbgInfoIntrinsic>(it))
3054 setDebugLocFromInst(Builder, it);
3056 Module *M = BB->getParent()->getParent();
3057 CallInst *CI = cast<CallInst>(it);
3058 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3059 assert(ID && "Not an intrinsic call!");
3061 case Intrinsic::lifetime_end:
3062 case Intrinsic::lifetime_start:
3063 scalarizeInstruction(it);
3066 for (unsigned Part = 0; Part < UF; ++Part) {
3067 SmallVector<Value *, 4> Args;
3068 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3069 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3070 Args.push_back(Arg[Part]);
3072 Type *Tys[] = {CI->getType()};
3074 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3076 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3077 Entry[Part] = Builder.CreateCall(F, Args);
3085 // All other instructions are unsupported. Scalarize them.
3086 scalarizeInstruction(it);
3089 }// end of for_each instr.
3092 void InnerLoopVectorizer::updateAnalysis() {
3093 // Forget the original basic block.
3094 SE->forgetLoop(OrigLoop);
3096 // Update the dominator tree information.
3097 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3098 "Entry does not dominate exit.");
3100 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3101 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3102 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3104 // Due to if predication of stores we might create a sequence of "if(pred)
3105 // a[i] = ...; " blocks.
3106 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3108 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3109 else if (isPredicatedBlock(i)) {
3110 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3112 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3116 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3117 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3118 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3119 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3121 DEBUG(DT->verifyDomTree());
3124 /// \brief Check whether it is safe to if-convert this phi node.
3126 /// Phi nodes with constant expressions that can trap are not safe to if
3128 static bool canIfConvertPHINodes(BasicBlock *BB) {
3129 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3130 PHINode *Phi = dyn_cast<PHINode>(I);
3133 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3134 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3141 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3142 if (!EnableIfConversion)
3145 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3147 // A list of pointers that we can safely read and write to.
3148 SmallPtrSet<Value *, 8> SafePointes;
3150 // Collect safe addresses.
3151 for (Loop::block_iterator BI = TheLoop->block_begin(),
3152 BE = TheLoop->block_end(); BI != BE; ++BI) {
3153 BasicBlock *BB = *BI;
3155 if (blockNeedsPredication(BB))
3158 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3159 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3160 SafePointes.insert(LI->getPointerOperand());
3161 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3162 SafePointes.insert(SI->getPointerOperand());
3166 // Collect the blocks that need predication.
3167 BasicBlock *Header = TheLoop->getHeader();
3168 for (Loop::block_iterator BI = TheLoop->block_begin(),
3169 BE = TheLoop->block_end(); BI != BE; ++BI) {
3170 BasicBlock *BB = *BI;
3172 // We don't support switch statements inside loops.
3173 if (!isa<BranchInst>(BB->getTerminator()))
3176 // We must be able to predicate all blocks that need to be predicated.
3177 if (blockNeedsPredication(BB)) {
3178 if (!blockCanBePredicated(BB, SafePointes))
3180 } else if (BB != Header && !canIfConvertPHINodes(BB))
3185 // We can if-convert this loop.
3189 bool LoopVectorizationLegality::canVectorize() {
3190 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3191 // be canonicalized.
3192 if (!TheLoop->getLoopPreheader())
3195 // We can only vectorize innermost loops.
3196 if (TheLoop->getSubLoopsVector().size())
3199 // We must have a single backedge.
3200 if (TheLoop->getNumBackEdges() != 1)
3203 // We must have a single exiting block.
3204 if (!TheLoop->getExitingBlock())
3207 // We need to have a loop header.
3208 DEBUG(dbgs() << "LV: Found a loop: " <<
3209 TheLoop->getHeader()->getName() << '\n');
3211 // Check if we can if-convert non-single-bb loops.
3212 unsigned NumBlocks = TheLoop->getNumBlocks();
3213 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3214 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3218 // ScalarEvolution needs to be able to find the exit count.
3219 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3220 if (ExitCount == SE->getCouldNotCompute()) {
3221 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3225 // Check if we can vectorize the instructions and CFG in this loop.
3226 if (!canVectorizeInstrs()) {
3227 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3231 // Go over each instruction and look at memory deps.
3232 if (!canVectorizeMemory()) {
3233 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3237 // Collect all of the variables that remain uniform after vectorization.
3238 collectLoopUniforms();
3240 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3241 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3244 // Okay! We can vectorize. At this point we don't have any other mem analysis
3245 // which may limit our maximum vectorization factor, so just return true with
3250 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3251 if (Ty->isPointerTy())
3252 return DL.getIntPtrType(Ty);
3254 // It is possible that char's or short's overflow when we ask for the loop's
3255 // trip count, work around this by changing the type size.
3256 if (Ty->getScalarSizeInBits() < 32)
3257 return Type::getInt32Ty(Ty->getContext());
3262 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3263 Ty0 = convertPointerToIntegerType(DL, Ty0);
3264 Ty1 = convertPointerToIntegerType(DL, Ty1);
3265 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3270 /// \brief Check that the instruction has outside loop users and is not an
3271 /// identified reduction variable.
3272 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3273 SmallPtrSet<Value *, 4> &Reductions) {
3274 // Reduction instructions are allowed to have exit users. All other
3275 // instructions must not have external users.
3276 if (!Reductions.count(Inst))
3277 //Check that all of the users of the loop are inside the BB.
3278 for (User *U : Inst->users()) {
3279 Instruction *UI = cast<Instruction>(U);
3280 // This user may be a reduction exit value.
3281 if (!TheLoop->contains(UI)) {
3282 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3289 bool LoopVectorizationLegality::canVectorizeInstrs() {
3290 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3291 BasicBlock *Header = TheLoop->getHeader();
3293 // Look for the attribute signaling the absence of NaNs.
3294 Function &F = *Header->getParent();
3295 if (F.hasFnAttribute("no-nans-fp-math"))
3296 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3297 AttributeSet::FunctionIndex,
3298 "no-nans-fp-math").getValueAsString() == "true";
3300 // For each block in the loop.
3301 for (Loop::block_iterator bb = TheLoop->block_begin(),
3302 be = TheLoop->block_end(); bb != be; ++bb) {
3304 // Scan the instructions in the block and look for hazards.
3305 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3308 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3309 Type *PhiTy = Phi->getType();
3310 // Check that this PHI type is allowed.
3311 if (!PhiTy->isIntegerTy() &&
3312 !PhiTy->isFloatingPointTy() &&
3313 !PhiTy->isPointerTy()) {
3314 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3318 // If this PHINode is not in the header block, then we know that we
3319 // can convert it to select during if-conversion. No need to check if
3320 // the PHIs in this block are induction or reduction variables.
3321 if (*bb != Header) {
3322 // Check that this instruction has no outside users or is an
3323 // identified reduction value with an outside user.
3324 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3329 // We only allow if-converted PHIs with more than two incoming values.
3330 if (Phi->getNumIncomingValues() != 2) {
3331 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3335 // This is the value coming from the preheader.
3336 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3337 // Check if this is an induction variable.
3338 InductionKind IK = isInductionVariable(Phi);
3340 if (IK_NoInduction != IK) {
3341 // Get the widest type.
3343 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3345 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3347 // Int inductions are special because we only allow one IV.
3348 if (IK == IK_IntInduction) {
3349 // Use the phi node with the widest type as induction. Use the last
3350 // one if there are multiple (no good reason for doing this other
3351 // than it is expedient).
3352 if (!Induction || PhiTy == WidestIndTy)
3356 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3357 Inductions[Phi] = InductionInfo(StartValue, IK);
3359 // Until we explicitly handle the case of an induction variable with
3360 // an outside loop user we have to give up vectorizing this loop.
3361 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3367 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3368 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3371 if (AddReductionVar(Phi, RK_IntegerMult)) {
3372 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3375 if (AddReductionVar(Phi, RK_IntegerOr)) {
3376 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3379 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3380 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3383 if (AddReductionVar(Phi, RK_IntegerXor)) {
3384 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3387 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3388 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3391 if (AddReductionVar(Phi, RK_FloatMult)) {
3392 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3395 if (AddReductionVar(Phi, RK_FloatAdd)) {
3396 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3399 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3400 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3405 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3407 }// end of PHI handling
3409 // We still don't handle functions. However, we can ignore dbg intrinsic
3410 // calls and we do handle certain intrinsic and libm functions.
3411 CallInst *CI = dyn_cast<CallInst>(it);
3412 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3413 DEBUG(dbgs() << "LV: Found a call site.\n");
3417 // Check that the instruction return type is vectorizable.
3418 // Also, we can't vectorize extractelement instructions.
3419 if ((!VectorType::isValidElementType(it->getType()) &&
3420 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3421 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3425 // Check that the stored type is vectorizable.
3426 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3427 Type *T = ST->getValueOperand()->getType();
3428 if (!VectorType::isValidElementType(T))
3430 if (EnableMemAccessVersioning)
3431 collectStridedAcccess(ST);
3434 if (EnableMemAccessVersioning)
3435 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3436 collectStridedAcccess(LI);
3438 // Reduction instructions are allowed to have exit users.
3439 // All other instructions must not have external users.
3440 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3448 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3449 if (Inductions.empty())
3456 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3457 /// return the induction operand of the gep pointer.
3458 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3459 const DataLayout *DL, Loop *Lp) {
3460 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3464 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3466 // Check that all of the gep indices are uniform except for our induction
3468 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3469 if (i != InductionOperand &&
3470 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3472 return GEP->getOperand(InductionOperand);
3475 ///\brief Look for a cast use of the passed value.
3476 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3477 Value *UniqueCast = nullptr;
3478 for (User *U : Ptr->users()) {
3479 CastInst *CI = dyn_cast<CastInst>(U);
3480 if (CI && CI->getType() == Ty) {
3490 ///\brief Get the stride of a pointer access in a loop.
3491 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3492 /// pointer to the Value, or null otherwise.
3493 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3494 const DataLayout *DL, Loop *Lp) {
3495 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3496 if (!PtrTy || PtrTy->isAggregateType())
3499 // Try to remove a gep instruction to make the pointer (actually index at this
3500 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3501 // pointer, otherwise, we are analyzing the index.
3502 Value *OrigPtr = Ptr;
3504 // The size of the pointer access.
3505 int64_t PtrAccessSize = 1;
3507 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3508 const SCEV *V = SE->getSCEV(Ptr);
3512 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3513 V = C->getOperand();
3515 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3519 V = S->getStepRecurrence(*SE);
3523 // Strip off the size of access multiplication if we are still analyzing the
3525 if (OrigPtr == Ptr) {
3526 DL->getTypeAllocSize(PtrTy->getElementType());
3527 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3528 if (M->getOperand(0)->getSCEVType() != scConstant)
3531 const APInt &APStepVal =
3532 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3534 // Huge step value - give up.
3535 if (APStepVal.getBitWidth() > 64)
3538 int64_t StepVal = APStepVal.getSExtValue();
3539 if (PtrAccessSize != StepVal)
3541 V = M->getOperand(1);
3546 Type *StripedOffRecurrenceCast = nullptr;
3547 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3548 StripedOffRecurrenceCast = C->getType();
3549 V = C->getOperand();
3552 // Look for the loop invariant symbolic value.
3553 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3557 Value *Stride = U->getValue();
3558 if (!Lp->isLoopInvariant(Stride))
3561 // If we have stripped off the recurrence cast we have to make sure that we
3562 // return the value that is used in this loop so that we can replace it later.
3563 if (StripedOffRecurrenceCast)
3564 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3569 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3570 Value *Ptr = nullptr;
3571 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3572 Ptr = LI->getPointerOperand();
3573 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3574 Ptr = SI->getPointerOperand();
3578 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3582 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3583 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3584 Strides[Ptr] = Stride;
3585 StrideSet.insert(Stride);
3588 void LoopVectorizationLegality::collectLoopUniforms() {
3589 // We now know that the loop is vectorizable!
3590 // Collect variables that will remain uniform after vectorization.
3591 std::vector<Value*> Worklist;
3592 BasicBlock *Latch = TheLoop->getLoopLatch();
3594 // Start with the conditional branch and walk up the block.
3595 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3597 // Also add all consecutive pointer values; these values will be uniform
3598 // after vectorization (and subsequent cleanup) and, until revectorization is
3599 // supported, all dependencies must also be uniform.
3600 for (Loop::block_iterator B = TheLoop->block_begin(),
3601 BE = TheLoop->block_end(); B != BE; ++B)
3602 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3604 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3605 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3607 while (Worklist.size()) {
3608 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3609 Worklist.pop_back();
3611 // Look at instructions inside this loop.
3612 // Stop when reaching PHI nodes.
3613 // TODO: we need to follow values all over the loop, not only in this block.
3614 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3617 // This is a known uniform.
3620 // Insert all operands.
3621 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3626 /// \brief Analyses memory accesses in a loop.
3628 /// Checks whether run time pointer checks are needed and builds sets for data
3629 /// dependence checking.
3630 class AccessAnalysis {
3632 /// \brief Read or write access location.
3633 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3634 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3636 /// \brief Set of potential dependent memory accesses.
3637 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3639 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3640 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3641 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3643 /// \brief Register a load and whether it is only read from.
3644 void addLoad(Value *Ptr, bool IsReadOnly) {
3645 Accesses.insert(MemAccessInfo(Ptr, false));
3647 ReadOnlyPtr.insert(Ptr);
3650 /// \brief Register a store.
3651 void addStore(Value *Ptr) {
3652 Accesses.insert(MemAccessInfo(Ptr, true));
3655 /// \brief Check whether we can check the pointers at runtime for
3656 /// non-intersection.
3657 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3658 unsigned &NumComparisons, ScalarEvolution *SE,
3659 Loop *TheLoop, ValueToValueMap &Strides,
3660 bool ShouldCheckStride = false);
3662 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3663 /// and builds sets of dependent accesses.
3664 void buildDependenceSets() {
3665 // Process read-write pointers first.
3666 processMemAccesses(false);
3667 // Next, process read pointers.
3668 processMemAccesses(true);
3671 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3673 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3674 void resetDepChecks() { CheckDeps.clear(); }
3676 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3679 typedef SetVector<MemAccessInfo> PtrAccessSet;
3680 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3682 /// \brief Go over all memory access or only the deferred ones if
3683 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3684 /// and build sets of dependency check candidates.
3685 void processMemAccesses(bool UseDeferred);
3687 /// Set of all accesses.
3688 PtrAccessSet Accesses;
3690 /// Set of access to check after all writes have been processed.
3691 PtrAccessSet DeferredAccesses;
3693 /// Map of pointers to last access encountered.
3694 UnderlyingObjToAccessMap ObjToLastAccess;
3696 /// Set of accesses that need a further dependence check.
3697 MemAccessInfoSet CheckDeps;
3699 /// Set of pointers that are read only.
3700 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3702 /// Set of underlying objects already written to.
3703 SmallPtrSet<Value*, 16> WriteObjects;
3705 const DataLayout *DL;
3707 /// Sets of potentially dependent accesses - members of one set share an
3708 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3709 /// dependence check.
3710 DepCandidates &DepCands;
3712 bool AreAllWritesIdentified;
3713 bool AreAllReadsIdentified;
3714 bool IsRTCheckNeeded;
3717 } // end anonymous namespace
3719 /// \brief Check whether a pointer can participate in a runtime bounds check.
3720 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3722 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3723 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3727 return AR->isAffine();
3730 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3731 /// the address space.
3732 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3733 const Loop *Lp, ValueToValueMap &StridesMap);
3735 bool AccessAnalysis::canCheckPtrAtRT(
3736 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3737 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3738 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3739 // Find pointers with computable bounds. We are going to use this information
3740 // to place a runtime bound check.
3741 unsigned NumReadPtrChecks = 0;
3742 unsigned NumWritePtrChecks = 0;
3743 bool CanDoRT = true;
3745 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3746 // We assign consecutive id to access from different dependence sets.
3747 // Accesses within the same set don't need a runtime check.
3748 unsigned RunningDepId = 1;
3749 DenseMap<Value *, unsigned> DepSetId;
3751 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3753 const MemAccessInfo &Access = *AI;
3754 Value *Ptr = Access.getPointer();
3755 bool IsWrite = Access.getInt();
3757 // Just add write checks if we have both.
3758 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3762 ++NumWritePtrChecks;
3766 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3767 // When we run after a failing dependency check we have to make sure we
3768 // don't have wrapping pointers.
3769 (!ShouldCheckStride ||
3770 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3771 // The id of the dependence set.
3774 if (IsDepCheckNeeded) {
3775 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3776 unsigned &LeaderId = DepSetId[Leader];
3778 LeaderId = RunningDepId++;
3781 // Each access has its own dependence set.
3782 DepId = RunningDepId++;
3784 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3786 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3792 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3793 NumComparisons = 0; // Only one dependence set.
3795 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3796 NumWritePtrChecks - 1));
3799 // If the pointers that we would use for the bounds comparison have different
3800 // address spaces, assume the values aren't directly comparable, so we can't
3801 // use them for the runtime check. We also have to assume they could
3802 // overlap. In the future there should be metadata for whether address spaces
3804 unsigned NumPointers = RtCheck.Pointers.size();
3805 for (unsigned i = 0; i < NumPointers; ++i) {
3806 for (unsigned j = i + 1; j < NumPointers; ++j) {
3807 // Only need to check pointers between two different dependency sets.
3808 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3811 Value *PtrI = RtCheck.Pointers[i];
3812 Value *PtrJ = RtCheck.Pointers[j];
3814 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3815 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3817 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3818 " different address spaces\n");
3827 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3828 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3831 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3832 // We process the set twice: first we process read-write pointers, last we
3833 // process read-only pointers. This allows us to skip dependence tests for
3834 // read-only pointers.
3836 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3837 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3838 const MemAccessInfo &Access = *AI;
3839 Value *Ptr = Access.getPointer();
3840 bool IsWrite = Access.getInt();
3842 DepCands.insert(Access);
3844 // Memorize read-only pointers for later processing and skip them in the
3845 // first round (they need to be checked after we have seen all write
3846 // pointers). Note: we also mark pointer that are not consecutive as
3847 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3848 // second check for "!IsWrite".
3849 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3850 if (!UseDeferred && IsReadOnlyPtr) {
3851 DeferredAccesses.insert(Access);
3855 bool NeedDepCheck = false;
3856 // Check whether there is the possibility of dependency because of
3857 // underlying objects being the same.
3858 typedef SmallVector<Value*, 16> ValueVector;
3859 ValueVector TempObjects;
3860 GetUnderlyingObjects(Ptr, TempObjects, DL);
3861 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3863 Value *UnderlyingObj = *UI;
3865 // If this is a write then it needs to be an identified object. If this a
3866 // read and all writes (so far) are identified function scope objects we
3867 // don't need an identified underlying object but only an Argument (the
3868 // next write is going to invalidate this assumption if it is
3870 // This is a micro-optimization for the case where all writes are
3871 // identified and we have one argument pointer.
3872 // Otherwise, we do need a runtime check.
3873 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3874 (!IsWrite && (!AreAllWritesIdentified ||
3875 !isa<Argument>(UnderlyingObj)) &&
3876 !isIdentifiedObject(UnderlyingObj))) {
3877 DEBUG(dbgs() << "LV: Found an unidentified " <<
3878 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3880 IsRTCheckNeeded = (IsRTCheckNeeded ||
3881 !isIdentifiedObject(UnderlyingObj) ||
3882 !AreAllReadsIdentified);
3885 AreAllWritesIdentified = false;
3887 AreAllReadsIdentified = false;
3890 // If this is a write - check other reads and writes for conflicts. If
3891 // this is a read only check other writes for conflicts (but only if there
3892 // is no other write to the ptr - this is an optimization to catch "a[i] =
3893 // a[i] + " without having to do a dependence check).
3894 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3895 NeedDepCheck = true;
3898 WriteObjects.insert(UnderlyingObj);
3900 // Create sets of pointers connected by shared underlying objects.
3901 UnderlyingObjToAccessMap::iterator Prev =
3902 ObjToLastAccess.find(UnderlyingObj);
3903 if (Prev != ObjToLastAccess.end())
3904 DepCands.unionSets(Access, Prev->second);
3906 ObjToLastAccess[UnderlyingObj] = Access;
3910 CheckDeps.insert(Access);
3915 /// \brief Checks memory dependences among accesses to the same underlying
3916 /// object to determine whether there vectorization is legal or not (and at
3917 /// which vectorization factor).
3919 /// This class works under the assumption that we already checked that memory
3920 /// locations with different underlying pointers are "must-not alias".
3921 /// We use the ScalarEvolution framework to symbolically evalutate access
3922 /// functions pairs. Since we currently don't restructure the loop we can rely
3923 /// on the program order of memory accesses to determine their safety.
3924 /// At the moment we will only deem accesses as safe for:
3925 /// * A negative constant distance assuming program order.
3927 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3928 /// a[i] = tmp; y = a[i];
3930 /// The latter case is safe because later checks guarantuee that there can't
3931 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3932 /// the same variable: a header phi can only be an induction or a reduction, a
3933 /// reduction can't have a memory sink, an induction can't have a memory
3934 /// source). This is important and must not be violated (or we have to
3935 /// resort to checking for cycles through memory).
3937 /// * A positive constant distance assuming program order that is bigger
3938 /// than the biggest memory access.
3940 /// tmp = a[i] OR b[i] = x
3941 /// a[i+2] = tmp y = b[i+2];
3943 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3945 /// * Zero distances and all accesses have the same size.
3947 class MemoryDepChecker {
3949 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3950 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3952 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
3953 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3954 ShouldRetryWithRuntimeCheck(false) {}
3956 /// \brief Register the location (instructions are given increasing numbers)
3957 /// of a write access.
3958 void addAccess(StoreInst *SI) {
3959 Value *Ptr = SI->getPointerOperand();
3960 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3961 InstMap.push_back(SI);
3965 /// \brief Register the location (instructions are given increasing numbers)
3966 /// of a write access.
3967 void addAccess(LoadInst *LI) {
3968 Value *Ptr = LI->getPointerOperand();
3969 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3970 InstMap.push_back(LI);
3974 /// \brief Check whether the dependencies between the accesses are safe.
3976 /// Only checks sets with elements in \p CheckDeps.
3977 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3978 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
3980 /// \brief The maximum number of bytes of a vector register we can vectorize
3981 /// the accesses safely with.
3982 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3984 /// \brief In same cases when the dependency check fails we can still
3985 /// vectorize the loop with a dynamic array access check.
3986 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3989 ScalarEvolution *SE;
3990 const DataLayout *DL;
3991 const Loop *InnermostLoop;
3993 /// \brief Maps access locations (ptr, read/write) to program order.
3994 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3996 /// \brief Memory access instructions in program order.
3997 SmallVector<Instruction *, 16> InstMap;
3999 /// \brief The program order index to be used for the next instruction.
4002 // We can access this many bytes in parallel safely.
4003 unsigned MaxSafeDepDistBytes;
4005 /// \brief If we see a non-constant dependence distance we can still try to
4006 /// vectorize this loop with runtime checks.
4007 bool ShouldRetryWithRuntimeCheck;
4009 /// \brief Check whether there is a plausible dependence between the two
4012 /// Access \p A must happen before \p B in program order. The two indices
4013 /// identify the index into the program order map.
4015 /// This function checks whether there is a plausible dependence (or the
4016 /// absence of such can't be proved) between the two accesses. If there is a
4017 /// plausible dependence but the dependence distance is bigger than one
4018 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4019 /// distance is smaller than any other distance encountered so far).
4020 /// Otherwise, this function returns true signaling a possible dependence.
4021 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4022 const MemAccessInfo &B, unsigned BIdx,
4023 ValueToValueMap &Strides);
4025 /// \brief Check whether the data dependence could prevent store-load
4027 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4030 } // end anonymous namespace
4032 static bool isInBoundsGep(Value *Ptr) {
4033 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4034 return GEP->isInBounds();
4038 /// \brief Check whether the access through \p Ptr has a constant stride.
4039 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4040 const Loop *Lp, ValueToValueMap &StridesMap) {
4041 const Type *Ty = Ptr->getType();
4042 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4044 // Make sure that the pointer does not point to aggregate types.
4045 const PointerType *PtrTy = cast<PointerType>(Ty);
4046 if (PtrTy->getElementType()->isAggregateType()) {
4047 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4052 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4054 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4056 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4057 << *Ptr << " SCEV: " << *PtrScev << "\n");
4061 // The accesss function must stride over the innermost loop.
4062 if (Lp != AR->getLoop()) {
4063 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4064 *Ptr << " SCEV: " << *PtrScev << "\n");
4067 // The address calculation must not wrap. Otherwise, a dependence could be
4069 // An inbounds getelementptr that is a AddRec with a unit stride
4070 // cannot wrap per definition. The unit stride requirement is checked later.
4071 // An getelementptr without an inbounds attribute and unit stride would have
4072 // to access the pointer value "0" which is undefined behavior in address
4073 // space 0, therefore we can also vectorize this case.
4074 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4075 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4076 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4077 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4078 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4079 << *Ptr << " SCEV: " << *PtrScev << "\n");
4083 // Check the step is constant.
4084 const SCEV *Step = AR->getStepRecurrence(*SE);
4086 // Calculate the pointer stride and check if it is consecutive.
4087 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4089 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4090 " SCEV: " << *PtrScev << "\n");
4094 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4095 const APInt &APStepVal = C->getValue()->getValue();
4097 // Huge step value - give up.
4098 if (APStepVal.getBitWidth() > 64)
4101 int64_t StepVal = APStepVal.getSExtValue();
4104 int64_t Stride = StepVal / Size;
4105 int64_t Rem = StepVal % Size;
4109 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4110 // know we can't "wrap around the address space". In case of address space
4111 // zero we know that this won't happen without triggering undefined behavior.
4112 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4113 Stride != 1 && Stride != -1)
4119 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4120 unsigned TypeByteSize) {
4121 // If loads occur at a distance that is not a multiple of a feasible vector
4122 // factor store-load forwarding does not take place.
4123 // Positive dependences might cause troubles because vectorizing them might
4124 // prevent store-load forwarding making vectorized code run a lot slower.
4125 // a[i] = a[i-3] ^ a[i-8];
4126 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4127 // hence on your typical architecture store-load forwarding does not take
4128 // place. Vectorizing in such cases does not make sense.
4129 // Store-load forwarding distance.
4130 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4131 // Maximum vector factor.
4132 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4133 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4134 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4136 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4138 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4139 MaxVFWithoutSLForwardIssues = (vf >>=1);
4144 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4145 DEBUG(dbgs() << "LV: Distance " << Distance <<
4146 " that could cause a store-load forwarding conflict\n");
4150 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4151 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4152 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4156 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4157 const MemAccessInfo &B, unsigned BIdx,
4158 ValueToValueMap &Strides) {
4159 assert (AIdx < BIdx && "Must pass arguments in program order");
4161 Value *APtr = A.getPointer();
4162 Value *BPtr = B.getPointer();
4163 bool AIsWrite = A.getInt();
4164 bool BIsWrite = B.getInt();
4166 // Two reads are independent.
4167 if (!AIsWrite && !BIsWrite)
4170 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4171 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4173 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4174 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4176 const SCEV *Src = AScev;
4177 const SCEV *Sink = BScev;
4179 // If the induction step is negative we have to invert source and sink of the
4181 if (StrideAPtr < 0) {
4184 std::swap(APtr, BPtr);
4185 std::swap(Src, Sink);
4186 std::swap(AIsWrite, BIsWrite);
4187 std::swap(AIdx, BIdx);
4188 std::swap(StrideAPtr, StrideBPtr);
4191 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4193 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4194 << "(Induction step: " << StrideAPtr << ")\n");
4195 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4196 << *InstMap[BIdx] << ": " << *Dist << "\n");
4198 // Need consecutive accesses. We don't want to vectorize
4199 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4200 // the address space.
4201 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4202 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4206 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4208 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4209 ShouldRetryWithRuntimeCheck = true;
4213 Type *ATy = APtr->getType()->getPointerElementType();
4214 Type *BTy = BPtr->getType()->getPointerElementType();
4215 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4217 // Negative distances are not plausible dependencies.
4218 const APInt &Val = C->getValue()->getValue();
4219 if (Val.isNegative()) {
4220 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4221 if (IsTrueDataDependence &&
4222 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4226 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4230 // Write to the same location with the same size.
4231 // Could be improved to assert type sizes are the same (i32 == float, etc).
4235 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4239 assert(Val.isStrictlyPositive() && "Expect a positive value");
4241 // Positive distance bigger than max vectorization factor.
4244 "LV: ReadWrite-Write positive dependency with different types\n");
4248 unsigned Distance = (unsigned) Val.getZExtValue();
4250 // Bail out early if passed-in parameters make vectorization not feasible.
4251 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4252 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4254 // The distance must be bigger than the size needed for a vectorized version
4255 // of the operation and the size of the vectorized operation must not be
4256 // bigger than the currrent maximum size.
4257 if (Distance < 2*TypeByteSize ||
4258 2*TypeByteSize > MaxSafeDepDistBytes ||
4259 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4260 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4261 << Val.getSExtValue() << '\n');
4265 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4266 Distance : MaxSafeDepDistBytes;
4268 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4269 if (IsTrueDataDependence &&
4270 couldPreventStoreLoadForward(Distance, TypeByteSize))
4273 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4274 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4279 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4280 MemAccessInfoSet &CheckDeps,
4281 ValueToValueMap &Strides) {
4283 MaxSafeDepDistBytes = -1U;
4284 while (!CheckDeps.empty()) {
4285 MemAccessInfo CurAccess = *CheckDeps.begin();
4287 // Get the relevant memory access set.
4288 EquivalenceClasses<MemAccessInfo>::iterator I =
4289 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4291 // Check accesses within this set.
4292 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4293 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4295 // Check every access pair.
4297 CheckDeps.erase(*AI);
4298 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4300 // Check every accessing instruction pair in program order.
4301 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4302 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4303 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4304 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4305 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4307 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4318 bool LoopVectorizationLegality::canVectorizeMemory() {
4320 typedef SmallVector<Value*, 16> ValueVector;
4321 typedef SmallPtrSet<Value*, 16> ValueSet;
4323 // Holds the Load and Store *instructions*.
4327 // Holds all the different accesses in the loop.
4328 unsigned NumReads = 0;
4329 unsigned NumReadWrites = 0;
4331 PtrRtCheck.Pointers.clear();
4332 PtrRtCheck.Need = false;
4334 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4335 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4338 for (Loop::block_iterator bb = TheLoop->block_begin(),
4339 be = TheLoop->block_end(); bb != be; ++bb) {
4341 // Scan the BB and collect legal loads and stores.
4342 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4345 // If this is a load, save it. If this instruction can read from memory
4346 // but is not a load, then we quit. Notice that we don't handle function
4347 // calls that read or write.
4348 if (it->mayReadFromMemory()) {
4349 // Many math library functions read the rounding mode. We will only
4350 // vectorize a loop if it contains known function calls that don't set
4351 // the flag. Therefore, it is safe to ignore this read from memory.
4352 CallInst *Call = dyn_cast<CallInst>(it);
4353 if (Call && getIntrinsicIDForCall(Call, TLI))
4356 LoadInst *Ld = dyn_cast<LoadInst>(it);
4357 if (!Ld) return false;
4358 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4359 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4363 Loads.push_back(Ld);
4364 DepChecker.addAccess(Ld);
4368 // Save 'store' instructions. Abort if other instructions write to memory.
4369 if (it->mayWriteToMemory()) {
4370 StoreInst *St = dyn_cast<StoreInst>(it);
4371 if (!St) return false;
4372 if (!St->isSimple() && !IsAnnotatedParallel) {
4373 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4377 Stores.push_back(St);
4378 DepChecker.addAccess(St);
4383 // Now we have two lists that hold the loads and the stores.
4384 // Next, we find the pointers that they use.
4386 // Check if we see any stores. If there are no stores, then we don't
4387 // care if the pointers are *restrict*.
4388 if (!Stores.size()) {
4389 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4393 AccessAnalysis::DepCandidates DependentAccesses;
4394 AccessAnalysis Accesses(DL, DependentAccesses);
4396 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4397 // multiple times on the same object. If the ptr is accessed twice, once
4398 // for read and once for write, it will only appear once (on the write
4399 // list). This is okay, since we are going to check for conflicts between
4400 // writes and between reads and writes, but not between reads and reads.
4403 ValueVector::iterator I, IE;
4404 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4405 StoreInst *ST = cast<StoreInst>(*I);
4406 Value* Ptr = ST->getPointerOperand();
4408 if (isUniform(Ptr)) {
4409 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4413 // If we did *not* see this pointer before, insert it to the read-write
4414 // list. At this phase it is only a 'write' list.
4415 if (Seen.insert(Ptr)) {
4417 Accesses.addStore(Ptr);
4421 if (IsAnnotatedParallel) {
4423 << "LV: A loop annotated parallel, ignore memory dependency "
4428 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4429 LoadInst *LD = cast<LoadInst>(*I);
4430 Value* Ptr = LD->getPointerOperand();
4431 // If we did *not* see this pointer before, insert it to the
4432 // read list. If we *did* see it before, then it is already in
4433 // the read-write list. This allows us to vectorize expressions
4434 // such as A[i] += x; Because the address of A[i] is a read-write
4435 // pointer. This only works if the index of A[i] is consecutive.
4436 // If the address of i is unknown (for example A[B[i]]) then we may
4437 // read a few words, modify, and write a few words, and some of the
4438 // words may be written to the same address.
4439 bool IsReadOnlyPtr = false;
4440 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4442 IsReadOnlyPtr = true;
4444 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4447 // If we write (or read-write) to a single destination and there are no
4448 // other reads in this loop then is it safe to vectorize.
4449 if (NumReadWrites == 1 && NumReads == 0) {
4450 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4454 // Build dependence sets and check whether we need a runtime pointer bounds
4456 Accesses.buildDependenceSets();
4457 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4459 // Find pointers with computable bounds. We are going to use this information
4460 // to place a runtime bound check.
4461 unsigned NumComparisons = 0;
4462 bool CanDoRT = false;
4464 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4467 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4468 " pointer comparisons.\n");
4470 // If we only have one set of dependences to check pointers among we don't
4471 // need a runtime check.
4472 if (NumComparisons == 0 && NeedRTCheck)
4473 NeedRTCheck = false;
4475 // Check that we did not collect too many pointers or found an unsizeable
4477 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4483 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4486 if (NeedRTCheck && !CanDoRT) {
4487 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4488 "the array bounds.\n");
4493 PtrRtCheck.Need = NeedRTCheck;
4495 bool CanVecMem = true;
4496 if (Accesses.isDependencyCheckNeeded()) {
4497 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4498 CanVecMem = DepChecker.areDepsSafe(
4499 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4500 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4502 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4503 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4506 // Clear the dependency checks. We assume they are not needed.
4507 Accesses.resetDepChecks();
4510 PtrRtCheck.Need = true;
4512 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4513 TheLoop, Strides, true);
4514 // Check that we did not collect too many pointers or found an unsizeable
4516 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4517 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4526 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4527 " need a runtime memory check.\n");
4532 static bool hasMultipleUsesOf(Instruction *I,
4533 SmallPtrSet<Instruction *, 8> &Insts) {
4534 unsigned NumUses = 0;
4535 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4536 if (Insts.count(dyn_cast<Instruction>(*Use)))
4545 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4546 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4547 if (!Set.count(dyn_cast<Instruction>(*Use)))
4552 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4553 ReductionKind Kind) {
4554 if (Phi->getNumIncomingValues() != 2)
4557 // Reduction variables are only found in the loop header block.
4558 if (Phi->getParent() != TheLoop->getHeader())
4561 // Obtain the reduction start value from the value that comes from the loop
4563 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4565 // ExitInstruction is the single value which is used outside the loop.
4566 // We only allow for a single reduction value to be used outside the loop.
4567 // This includes users of the reduction, variables (which form a cycle
4568 // which ends in the phi node).
4569 Instruction *ExitInstruction = nullptr;
4570 // Indicates that we found a reduction operation in our scan.
4571 bool FoundReduxOp = false;
4573 // We start with the PHI node and scan for all of the users of this
4574 // instruction. All users must be instructions that can be used as reduction
4575 // variables (such as ADD). We must have a single out-of-block user. The cycle
4576 // must include the original PHI.
4577 bool FoundStartPHI = false;
4579 // To recognize min/max patterns formed by a icmp select sequence, we store
4580 // the number of instruction we saw from the recognized min/max pattern,
4581 // to make sure we only see exactly the two instructions.
4582 unsigned NumCmpSelectPatternInst = 0;
4583 ReductionInstDesc ReduxDesc(false, nullptr);
4585 SmallPtrSet<Instruction *, 8> VisitedInsts;
4586 SmallVector<Instruction *, 8> Worklist;
4587 Worklist.push_back(Phi);
4588 VisitedInsts.insert(Phi);
4590 // A value in the reduction can be used:
4591 // - By the reduction:
4592 // - Reduction operation:
4593 // - One use of reduction value (safe).
4594 // - Multiple use of reduction value (not safe).
4596 // - All uses of the PHI must be the reduction (safe).
4597 // - Otherwise, not safe.
4598 // - By one instruction outside of the loop (safe).
4599 // - By further instructions outside of the loop (not safe).
4600 // - By an instruction that is not part of the reduction (not safe).
4602 // * An instruction type other than PHI or the reduction operation.
4603 // * A PHI in the header other than the initial PHI.
4604 while (!Worklist.empty()) {
4605 Instruction *Cur = Worklist.back();
4606 Worklist.pop_back();
4609 // If the instruction has no users then this is a broken chain and can't be
4610 // a reduction variable.
4611 if (Cur->use_empty())
4614 bool IsAPhi = isa<PHINode>(Cur);
4616 // A header PHI use other than the original PHI.
4617 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4620 // Reductions of instructions such as Div, and Sub is only possible if the
4621 // LHS is the reduction variable.
4622 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4623 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4624 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4627 // Any reduction instruction must be of one of the allowed kinds.
4628 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4629 if (!ReduxDesc.IsReduction)
4632 // A reduction operation must only have one use of the reduction value.
4633 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4634 hasMultipleUsesOf(Cur, VisitedInsts))
4637 // All inputs to a PHI node must be a reduction value.
4638 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4641 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4642 isa<SelectInst>(Cur)))
4643 ++NumCmpSelectPatternInst;
4644 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4645 isa<SelectInst>(Cur)))
4646 ++NumCmpSelectPatternInst;
4648 // Check whether we found a reduction operator.
4649 FoundReduxOp |= !IsAPhi;
4651 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4652 // onto the stack. This way we are going to have seen all inputs to PHI
4653 // nodes once we get to them.
4654 SmallVector<Instruction *, 8> NonPHIs;
4655 SmallVector<Instruction *, 8> PHIs;
4656 for (User *U : Cur->users()) {
4657 Instruction *UI = cast<Instruction>(U);
4659 // Check if we found the exit user.
4660 BasicBlock *Parent = UI->getParent();
4661 if (!TheLoop->contains(Parent)) {
4662 // Exit if you find multiple outside users or if the header phi node is
4663 // being used. In this case the user uses the value of the previous
4664 // iteration, in which case we would loose "VF-1" iterations of the
4665 // reduction operation if we vectorize.
4666 if (ExitInstruction != nullptr || Cur == Phi)
4669 // The instruction used by an outside user must be the last instruction
4670 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4671 // operations on the value.
4672 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4675 ExitInstruction = Cur;
4679 // Process instructions only once (termination). Each reduction cycle
4680 // value must only be used once, except by phi nodes and min/max
4681 // reductions which are represented as a cmp followed by a select.
4682 ReductionInstDesc IgnoredVal(false, nullptr);
4683 if (VisitedInsts.insert(UI)) {
4684 if (isa<PHINode>(UI))
4687 NonPHIs.push_back(UI);
4688 } else if (!isa<PHINode>(UI) &&
4689 ((!isa<FCmpInst>(UI) &&
4690 !isa<ICmpInst>(UI) &&
4691 !isa<SelectInst>(UI)) ||
4692 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4695 // Remember that we completed the cycle.
4697 FoundStartPHI = true;
4699 Worklist.append(PHIs.begin(), PHIs.end());
4700 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4703 // This means we have seen one but not the other instruction of the
4704 // pattern or more than just a select and cmp.
4705 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4706 NumCmpSelectPatternInst != 2)
4709 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4712 // We found a reduction var if we have reached the original phi node and we
4713 // only have a single instruction with out-of-loop users.
4715 // This instruction is allowed to have out-of-loop users.
4716 AllowedExit.insert(ExitInstruction);
4718 // Save the description of this reduction variable.
4719 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4720 ReduxDesc.MinMaxKind);
4721 Reductions[Phi] = RD;
4722 // We've ended the cycle. This is a reduction variable if we have an
4723 // outside user and it has a binary op.
4728 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4729 /// pattern corresponding to a min(X, Y) or max(X, Y).
4730 LoopVectorizationLegality::ReductionInstDesc
4731 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4732 ReductionInstDesc &Prev) {
4734 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4735 "Expect a select instruction");
4736 Instruction *Cmp = nullptr;
4737 SelectInst *Select = nullptr;
4739 // We must handle the select(cmp()) as a single instruction. Advance to the
4741 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4742 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4743 return ReductionInstDesc(false, I);
4744 return ReductionInstDesc(Select, Prev.MinMaxKind);
4747 // Only handle single use cases for now.
4748 if (!(Select = dyn_cast<SelectInst>(I)))
4749 return ReductionInstDesc(false, I);
4750 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4751 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4752 return ReductionInstDesc(false, I);
4753 if (!Cmp->hasOneUse())
4754 return ReductionInstDesc(false, I);
4759 // Look for a min/max pattern.
4760 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4761 return ReductionInstDesc(Select, MRK_UIntMin);
4762 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4763 return ReductionInstDesc(Select, MRK_UIntMax);
4764 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4765 return ReductionInstDesc(Select, MRK_SIntMax);
4766 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4767 return ReductionInstDesc(Select, MRK_SIntMin);
4768 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4769 return ReductionInstDesc(Select, MRK_FloatMin);
4770 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4771 return ReductionInstDesc(Select, MRK_FloatMax);
4772 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4773 return ReductionInstDesc(Select, MRK_FloatMin);
4774 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4775 return ReductionInstDesc(Select, MRK_FloatMax);
4777 return ReductionInstDesc(false, I);
4780 LoopVectorizationLegality::ReductionInstDesc
4781 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4783 ReductionInstDesc &Prev) {
4784 bool FP = I->getType()->isFloatingPointTy();
4785 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4786 switch (I->getOpcode()) {
4788 return ReductionInstDesc(false, I);
4789 case Instruction::PHI:
4790 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4791 Kind != RK_FloatMinMax))
4792 return ReductionInstDesc(false, I);
4793 return ReductionInstDesc(I, Prev.MinMaxKind);
4794 case Instruction::Sub:
4795 case Instruction::Add:
4796 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4797 case Instruction::Mul:
4798 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4799 case Instruction::And:
4800 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4801 case Instruction::Or:
4802 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4803 case Instruction::Xor:
4804 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4805 case Instruction::FMul:
4806 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4807 case Instruction::FAdd:
4808 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4809 case Instruction::FCmp:
4810 case Instruction::ICmp:
4811 case Instruction::Select:
4812 if (Kind != RK_IntegerMinMax &&
4813 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4814 return ReductionInstDesc(false, I);
4815 return isMinMaxSelectCmpPattern(I, Prev);
4819 LoopVectorizationLegality::InductionKind
4820 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4821 Type *PhiTy = Phi->getType();
4822 // We only handle integer and pointer inductions variables.
4823 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4824 return IK_NoInduction;
4826 // Check that the PHI is consecutive.
4827 const SCEV *PhiScev = SE->getSCEV(Phi);
4828 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4830 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4831 return IK_NoInduction;
4833 const SCEV *Step = AR->getStepRecurrence(*SE);
4835 // Integer inductions need to have a stride of one.
4836 if (PhiTy->isIntegerTy()) {
4838 return IK_IntInduction;
4839 if (Step->isAllOnesValue())
4840 return IK_ReverseIntInduction;
4841 return IK_NoInduction;
4844 // Calculate the pointer stride and check if it is consecutive.
4845 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4847 return IK_NoInduction;
4849 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4850 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4851 if (C->getValue()->equalsInt(Size))
4852 return IK_PtrInduction;
4853 else if (C->getValue()->equalsInt(0 - Size))
4854 return IK_ReversePtrInduction;
4856 return IK_NoInduction;
4859 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4860 Value *In0 = const_cast<Value*>(V);
4861 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4865 return Inductions.count(PN);
4868 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4869 assert(TheLoop->contains(BB) && "Unknown block used");
4871 // Blocks that do not dominate the latch need predication.
4872 BasicBlock* Latch = TheLoop->getLoopLatch();
4873 return !DT->dominates(BB, Latch);
4876 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4877 SmallPtrSet<Value *, 8>& SafePtrs) {
4878 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4879 // We might be able to hoist the load.
4880 if (it->mayReadFromMemory()) {
4881 LoadInst *LI = dyn_cast<LoadInst>(it);
4882 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4886 // We don't predicate stores at the moment.
4887 if (it->mayWriteToMemory()) {
4888 StoreInst *SI = dyn_cast<StoreInst>(it);
4889 // We only support predication of stores in basic blocks with one
4891 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
4892 !SafePtrs.count(SI->getPointerOperand()) ||
4893 !SI->getParent()->getSinglePredecessor())
4899 // Check that we don't have a constant expression that can trap as operand.
4900 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4902 if (Constant *C = dyn_cast<Constant>(*OI))
4907 // The instructions below can trap.
4908 switch (it->getOpcode()) {
4910 case Instruction::UDiv:
4911 case Instruction::SDiv:
4912 case Instruction::URem:
4913 case Instruction::SRem:
4921 LoopVectorizationCostModel::VectorizationFactor
4922 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4924 bool ForceVectorization) {
4925 // Width 1 means no vectorize
4926 VectorizationFactor Factor = { 1U, 0U };
4927 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4928 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4932 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
4933 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4937 // Find the trip count.
4938 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4939 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4941 unsigned WidestType = getWidestType();
4942 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4943 unsigned MaxSafeDepDist = -1U;
4944 if (Legal->getMaxSafeDepDistBytes() != -1U)
4945 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4946 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4947 WidestRegister : MaxSafeDepDist);
4948 unsigned MaxVectorSize = WidestRegister / WidestType;
4949 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4950 DEBUG(dbgs() << "LV: The Widest register is: "
4951 << WidestRegister << " bits.\n");
4953 if (MaxVectorSize == 0) {
4954 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4958 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4959 " into one vector!");
4961 unsigned VF = MaxVectorSize;
4963 // If we optimize the program for size, avoid creating the tail loop.
4965 // If we are unable to calculate the trip count then don't try to vectorize.
4967 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4971 // Find the maximum SIMD width that can fit within the trip count.
4972 VF = TC % MaxVectorSize;
4977 // If the trip count that we found modulo the vectorization factor is not
4978 // zero then we require a tail.
4980 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4986 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4987 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4989 Factor.Width = UserVF;
4993 float Cost = expectedCost(1);
4995 const float ScalarCost = Cost;
4998 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5000 // Ignore scalar width, because the user explicitly wants vectorization.
5001 if (ForceVectorization && VF > 1) {
5003 Cost = expectedCost(Width) / (float)Width;
5006 for (unsigned i=2; i <= VF; i*=2) {
5007 // Notice that the vector loop needs to be executed less times, so
5008 // we need to divide the cost of the vector loops by the width of
5009 // the vector elements.
5010 float VectorCost = expectedCost(i) / (float)i;
5011 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5012 (int)VectorCost << ".\n");
5013 if (VectorCost < Cost) {
5019 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5020 << "LV: Vectorization seems to be not beneficial, "
5021 << "but was forced by a user.\n");
5022 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5023 Factor.Width = Width;
5024 Factor.Cost = Width * Cost;
5028 unsigned LoopVectorizationCostModel::getWidestType() {
5029 unsigned MaxWidth = 8;
5032 for (Loop::block_iterator bb = TheLoop->block_begin(),
5033 be = TheLoop->block_end(); bb != be; ++bb) {
5034 BasicBlock *BB = *bb;
5036 // For each instruction in the loop.
5037 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5038 Type *T = it->getType();
5040 // Only examine Loads, Stores and PHINodes.
5041 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5044 // Examine PHI nodes that are reduction variables.
5045 if (PHINode *PN = dyn_cast<PHINode>(it))
5046 if (!Legal->getReductionVars()->count(PN))
5049 // Examine the stored values.
5050 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5051 T = ST->getValueOperand()->getType();
5053 // Ignore loaded pointer types and stored pointer types that are not
5054 // consecutive. However, we do want to take consecutive stores/loads of
5055 // pointer vectors into account.
5056 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5059 MaxWidth = std::max(MaxWidth,
5060 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5068 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5071 unsigned LoopCost) {
5073 // -- The unroll heuristics --
5074 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5075 // There are many micro-architectural considerations that we can't predict
5076 // at this level. For example frontend pressure (on decode or fetch) due to
5077 // code size, or the number and capabilities of the execution ports.
5079 // We use the following heuristics to select the unroll factor:
5080 // 1. If the code has reductions the we unroll in order to break the cross
5081 // iteration dependency.
5082 // 2. If the loop is really small then we unroll in order to reduce the loop
5084 // 3. We don't unroll if we think that we will spill registers to memory due
5085 // to the increased register pressure.
5087 // Use the user preference, unless 'auto' is selected.
5091 // When we optimize for size we don't unroll.
5095 // We used the distance for the unroll factor.
5096 if (Legal->getMaxSafeDepDistBytes() != -1U)
5099 // Do not unroll loops with a relatively small trip count.
5100 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5101 TheLoop->getLoopLatch());
5102 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5105 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5106 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5110 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5111 TargetNumRegisters = ForceTargetNumScalarRegs;
5113 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5114 TargetNumRegisters = ForceTargetNumVectorRegs;
5117 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5118 // We divide by these constants so assume that we have at least one
5119 // instruction that uses at least one register.
5120 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5121 R.NumInstructions = std::max(R.NumInstructions, 1U);
5123 // We calculate the unroll factor using the following formula.
5124 // Subtract the number of loop invariants from the number of available
5125 // registers. These registers are used by all of the unrolled instances.
5126 // Next, divide the remaining registers by the number of registers that is
5127 // required by the loop, in order to estimate how many parallel instances
5128 // fit without causing spills. All of this is rounded down if necessary to be
5129 // a power of two. We want power of two unroll factors to simplify any
5130 // addressing operations or alignment considerations.
5131 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5134 // Don't count the induction variable as unrolled.
5135 if (EnableIndVarRegisterHeur)
5136 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5137 std::max(1U, (R.MaxLocalUsers - 1)));
5139 // Clamp the unroll factor ranges to reasonable factors.
5140 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5142 // Check if the user has overridden the unroll max.
5144 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5145 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5147 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5148 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5151 // If we did not calculate the cost for VF (because the user selected the VF)
5152 // then we calculate the cost of VF here.
5154 LoopCost = expectedCost(VF);
5156 // Clamp the calculated UF to be between the 1 and the max unroll factor
5157 // that the target allows.
5158 if (UF > MaxUnrollSize)
5163 // Unroll if we vectorized this loop and there is a reduction that could
5164 // benefit from unrolling.
5165 if (VF > 1 && Legal->getReductionVars()->size()) {
5166 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5170 // Note that if we've already vectorized the loop we will have done the
5171 // runtime check and so unrolling won't require further checks.
5172 bool UnrollingRequiresRuntimePointerCheck =
5173 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5175 // We want to unroll small loops in order to reduce the loop overhead and
5176 // potentially expose ILP opportunities.
5177 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5178 if (!UnrollingRequiresRuntimePointerCheck &&
5179 LoopCost < SmallLoopCost) {
5180 // We assume that the cost overhead is 1 and we use the cost model
5181 // to estimate the cost of the loop and unroll until the cost of the
5182 // loop overhead is about 5% of the cost of the loop.
5183 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5185 // Unroll until store/load ports (estimated by max unroll factor) are
5187 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5188 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5190 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5191 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5192 return std::max(StoresUF, LoadsUF);
5195 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5199 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5203 LoopVectorizationCostModel::RegisterUsage
5204 LoopVectorizationCostModel::calculateRegisterUsage() {
5205 // This function calculates the register usage by measuring the highest number
5206 // of values that are alive at a single location. Obviously, this is a very
5207 // rough estimation. We scan the loop in a topological order in order and
5208 // assign a number to each instruction. We use RPO to ensure that defs are
5209 // met before their users. We assume that each instruction that has in-loop
5210 // users starts an interval. We record every time that an in-loop value is
5211 // used, so we have a list of the first and last occurrences of each
5212 // instruction. Next, we transpose this data structure into a multi map that
5213 // holds the list of intervals that *end* at a specific location. This multi
5214 // map allows us to perform a linear search. We scan the instructions linearly
5215 // and record each time that a new interval starts, by placing it in a set.
5216 // If we find this value in the multi-map then we remove it from the set.
5217 // The max register usage is the maximum size of the set.
5218 // We also search for instructions that are defined outside the loop, but are
5219 // used inside the loop. We need this number separately from the max-interval
5220 // usage number because when we unroll, loop-invariant values do not take
5222 LoopBlocksDFS DFS(TheLoop);
5226 R.NumInstructions = 0;
5228 // Each 'key' in the map opens a new interval. The values
5229 // of the map are the index of the 'last seen' usage of the
5230 // instruction that is the key.
5231 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5232 // Maps instruction to its index.
5233 DenseMap<unsigned, Instruction*> IdxToInstr;
5234 // Marks the end of each interval.
5235 IntervalMap EndPoint;
5236 // Saves the list of instruction indices that are used in the loop.
5237 SmallSet<Instruction*, 8> Ends;
5238 // Saves the list of values that are used in the loop but are
5239 // defined outside the loop, such as arguments and constants.
5240 SmallPtrSet<Value*, 8> LoopInvariants;
5243 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5244 be = DFS.endRPO(); bb != be; ++bb) {
5245 R.NumInstructions += (*bb)->size();
5246 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5248 Instruction *I = it;
5249 IdxToInstr[Index++] = I;
5251 // Save the end location of each USE.
5252 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5253 Value *U = I->getOperand(i);
5254 Instruction *Instr = dyn_cast<Instruction>(U);
5256 // Ignore non-instruction values such as arguments, constants, etc.
5257 if (!Instr) continue;
5259 // If this instruction is outside the loop then record it and continue.
5260 if (!TheLoop->contains(Instr)) {
5261 LoopInvariants.insert(Instr);
5265 // Overwrite previous end points.
5266 EndPoint[Instr] = Index;
5272 // Saves the list of intervals that end with the index in 'key'.
5273 typedef SmallVector<Instruction*, 2> InstrList;
5274 DenseMap<unsigned, InstrList> TransposeEnds;
5276 // Transpose the EndPoints to a list of values that end at each index.
5277 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5279 TransposeEnds[it->second].push_back(it->first);
5281 SmallSet<Instruction*, 8> OpenIntervals;
5282 unsigned MaxUsage = 0;
5285 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5286 for (unsigned int i = 0; i < Index; ++i) {
5287 Instruction *I = IdxToInstr[i];
5288 // Ignore instructions that are never used within the loop.
5289 if (!Ends.count(I)) continue;
5291 // Remove all of the instructions that end at this location.
5292 InstrList &List = TransposeEnds[i];
5293 for (unsigned int j=0, e = List.size(); j < e; ++j)
5294 OpenIntervals.erase(List[j]);
5296 // Count the number of live interals.
5297 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5299 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5300 OpenIntervals.size() << '\n');
5302 // Add the current instruction to the list of open intervals.
5303 OpenIntervals.insert(I);
5306 unsigned Invariant = LoopInvariants.size();
5307 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5308 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5309 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5311 R.LoopInvariantRegs = Invariant;
5312 R.MaxLocalUsers = MaxUsage;
5316 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5320 for (Loop::block_iterator bb = TheLoop->block_begin(),
5321 be = TheLoop->block_end(); bb != be; ++bb) {
5322 unsigned BlockCost = 0;
5323 BasicBlock *BB = *bb;
5325 // For each instruction in the old loop.
5326 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5327 // Skip dbg intrinsics.
5328 if (isa<DbgInfoIntrinsic>(it))
5331 unsigned C = getInstructionCost(it, VF);
5333 // Check if we should override the cost.
5334 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5335 C = ForceTargetInstructionCost;
5338 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5339 VF << " For instruction: " << *it << '\n');
5342 // We assume that if-converted blocks have a 50% chance of being executed.
5343 // When the code is scalar then some of the blocks are avoided due to CF.
5344 // When the code is vectorized we execute all code paths.
5345 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5354 /// \brief Check whether the address computation for a non-consecutive memory
5355 /// access looks like an unlikely candidate for being merged into the indexing
5358 /// We look for a GEP which has one index that is an induction variable and all
5359 /// other indices are loop invariant. If the stride of this access is also
5360 /// within a small bound we decide that this address computation can likely be
5361 /// merged into the addressing mode.
5362 /// In all other cases, we identify the address computation as complex.
5363 static bool isLikelyComplexAddressComputation(Value *Ptr,
5364 LoopVectorizationLegality *Legal,
5365 ScalarEvolution *SE,
5366 const Loop *TheLoop) {
5367 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5371 // We are looking for a gep with all loop invariant indices except for one
5372 // which should be an induction variable.
5373 unsigned NumOperands = Gep->getNumOperands();
5374 for (unsigned i = 1; i < NumOperands; ++i) {
5375 Value *Opd = Gep->getOperand(i);
5376 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5377 !Legal->isInductionVariable(Opd))
5381 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5382 // can likely be merged into the address computation.
5383 unsigned MaxMergeDistance = 64;
5385 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5389 // Check the step is constant.
5390 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5391 // Calculate the pointer stride and check if it is consecutive.
5392 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5396 const APInt &APStepVal = C->getValue()->getValue();
5398 // Huge step value - give up.
5399 if (APStepVal.getBitWidth() > 64)
5402 int64_t StepVal = APStepVal.getSExtValue();
5404 return StepVal > MaxMergeDistance;
5407 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5408 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5414 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5415 // If we know that this instruction will remain uniform, check the cost of
5416 // the scalar version.
5417 if (Legal->isUniformAfterVectorization(I))
5420 Type *RetTy = I->getType();
5421 Type *VectorTy = ToVectorTy(RetTy, VF);
5423 // TODO: We need to estimate the cost of intrinsic calls.
5424 switch (I->getOpcode()) {
5425 case Instruction::GetElementPtr:
5426 // We mark this instruction as zero-cost because the cost of GEPs in
5427 // vectorized code depends on whether the corresponding memory instruction
5428 // is scalarized or not. Therefore, we handle GEPs with the memory
5429 // instruction cost.
5431 case Instruction::Br: {
5432 return TTI.getCFInstrCost(I->getOpcode());
5434 case Instruction::PHI:
5435 //TODO: IF-converted IFs become selects.
5437 case Instruction::Add:
5438 case Instruction::FAdd:
5439 case Instruction::Sub:
5440 case Instruction::FSub:
5441 case Instruction::Mul:
5442 case Instruction::FMul:
5443 case Instruction::UDiv:
5444 case Instruction::SDiv:
5445 case Instruction::FDiv:
5446 case Instruction::URem:
5447 case Instruction::SRem:
5448 case Instruction::FRem:
5449 case Instruction::Shl:
5450 case Instruction::LShr:
5451 case Instruction::AShr:
5452 case Instruction::And:
5453 case Instruction::Or:
5454 case Instruction::Xor: {
5455 // Since we will replace the stride by 1 the multiplication should go away.
5456 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5458 // Certain instructions can be cheaper to vectorize if they have a constant
5459 // second vector operand. One example of this are shifts on x86.
5460 TargetTransformInfo::OperandValueKind Op1VK =
5461 TargetTransformInfo::OK_AnyValue;
5462 TargetTransformInfo::OperandValueKind Op2VK =
5463 TargetTransformInfo::OK_AnyValue;
5464 Value *Op2 = I->getOperand(1);
5466 // Check for a splat of a constant or for a non uniform vector of constants.
5467 if (isa<ConstantInt>(Op2))
5468 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5469 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5470 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5471 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5472 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5475 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5477 case Instruction::Select: {
5478 SelectInst *SI = cast<SelectInst>(I);
5479 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5480 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5481 Type *CondTy = SI->getCondition()->getType();
5483 CondTy = VectorType::get(CondTy, VF);
5485 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5487 case Instruction::ICmp:
5488 case Instruction::FCmp: {
5489 Type *ValTy = I->getOperand(0)->getType();
5490 VectorTy = ToVectorTy(ValTy, VF);
5491 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5493 case Instruction::Store:
5494 case Instruction::Load: {
5495 StoreInst *SI = dyn_cast<StoreInst>(I);
5496 LoadInst *LI = dyn_cast<LoadInst>(I);
5497 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5499 VectorTy = ToVectorTy(ValTy, VF);
5501 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5502 unsigned AS = SI ? SI->getPointerAddressSpace() :
5503 LI->getPointerAddressSpace();
5504 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5505 // We add the cost of address computation here instead of with the gep
5506 // instruction because only here we know whether the operation is
5509 return TTI.getAddressComputationCost(VectorTy) +
5510 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5512 // Scalarized loads/stores.
5513 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5514 bool Reverse = ConsecutiveStride < 0;
5515 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5516 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5517 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5518 bool IsComplexComputation =
5519 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5521 // The cost of extracting from the value vector and pointer vector.
5522 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5523 for (unsigned i = 0; i < VF; ++i) {
5524 // The cost of extracting the pointer operand.
5525 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5526 // In case of STORE, the cost of ExtractElement from the vector.
5527 // In case of LOAD, the cost of InsertElement into the returned
5529 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5530 Instruction::InsertElement,
5534 // The cost of the scalar loads/stores.
5535 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5536 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5541 // Wide load/stores.
5542 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5543 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5546 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5550 case Instruction::ZExt:
5551 case Instruction::SExt:
5552 case Instruction::FPToUI:
5553 case Instruction::FPToSI:
5554 case Instruction::FPExt:
5555 case Instruction::PtrToInt:
5556 case Instruction::IntToPtr:
5557 case Instruction::SIToFP:
5558 case Instruction::UIToFP:
5559 case Instruction::Trunc:
5560 case Instruction::FPTrunc:
5561 case Instruction::BitCast: {
5562 // We optimize the truncation of induction variable.
5563 // The cost of these is the same as the scalar operation.
5564 if (I->getOpcode() == Instruction::Trunc &&
5565 Legal->isInductionVariable(I->getOperand(0)))
5566 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5567 I->getOperand(0)->getType());
5569 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5570 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5572 case Instruction::Call: {
5573 CallInst *CI = cast<CallInst>(I);
5574 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5575 assert(ID && "Not an intrinsic call!");
5576 Type *RetTy = ToVectorTy(CI->getType(), VF);
5577 SmallVector<Type*, 4> Tys;
5578 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5579 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5580 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5583 // We are scalarizing the instruction. Return the cost of the scalar
5584 // instruction, plus the cost of insert and extract into vector
5585 // elements, times the vector width.
5588 if (!RetTy->isVoidTy() && VF != 1) {
5589 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5591 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5594 // The cost of inserting the results plus extracting each one of the
5596 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5599 // The cost of executing VF copies of the scalar instruction. This opcode
5600 // is unknown. Assume that it is the same as 'mul'.
5601 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5607 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5608 if (Scalar->isVoidTy() || VF == 1)
5610 return VectorType::get(Scalar, VF);
5613 char LoopVectorize::ID = 0;
5614 static const char lv_name[] = "Loop Vectorization";
5615 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5616 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5617 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5618 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5619 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5620 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5621 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5622 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5623 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5626 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5627 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5631 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5632 // Check for a store.
5633 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5634 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5636 // Check for a load.
5637 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5638 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5644 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5645 bool IfPredicateStore) {
5646 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5647 // Holds vector parameters or scalars, in case of uniform vals.
5648 SmallVector<VectorParts, 4> Params;
5650 setDebugLocFromInst(Builder, Instr);
5652 // Find all of the vectorized parameters.
5653 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5654 Value *SrcOp = Instr->getOperand(op);
5656 // If we are accessing the old induction variable, use the new one.
5657 if (SrcOp == OldInduction) {
5658 Params.push_back(getVectorValue(SrcOp));
5662 // Try using previously calculated values.
5663 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5665 // If the src is an instruction that appeared earlier in the basic block
5666 // then it should already be vectorized.
5667 if (SrcInst && OrigLoop->contains(SrcInst)) {
5668 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5669 // The parameter is a vector value from earlier.
5670 Params.push_back(WidenMap.get(SrcInst));
5672 // The parameter is a scalar from outside the loop. Maybe even a constant.
5673 VectorParts Scalars;
5674 Scalars.append(UF, SrcOp);
5675 Params.push_back(Scalars);
5679 assert(Params.size() == Instr->getNumOperands() &&
5680 "Invalid number of operands");
5682 // Does this instruction return a value ?
5683 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5685 Value *UndefVec = IsVoidRetTy ? nullptr :
5686 UndefValue::get(Instr->getType());
5687 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5688 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5690 Instruction *InsertPt = Builder.GetInsertPoint();
5691 BasicBlock *IfBlock = Builder.GetInsertBlock();
5692 BasicBlock *CondBlock = nullptr;
5695 Loop *VectorLp = nullptr;
5696 if (IfPredicateStore) {
5697 assert(Instr->getParent()->getSinglePredecessor() &&
5698 "Only support single predecessor blocks");
5699 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5700 Instr->getParent());
5701 VectorLp = LI->getLoopFor(IfBlock);
5702 assert(VectorLp && "Must have a loop for this block");
5705 // For each vector unroll 'part':
5706 for (unsigned Part = 0; Part < UF; ++Part) {
5707 // For each scalar that we create:
5709 // Start an "if (pred) a[i] = ..." block.
5710 Value *Cmp = nullptr;
5711 if (IfPredicateStore) {
5712 if (Cond[Part]->getType()->isVectorTy())
5714 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5715 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5716 ConstantInt::get(Cond[Part]->getType(), 1));
5717 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5718 LoopVectorBody.push_back(CondBlock);
5719 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5720 // Update Builder with newly created basic block.
5721 Builder.SetInsertPoint(InsertPt);
5724 Instruction *Cloned = Instr->clone();
5726 Cloned->setName(Instr->getName() + ".cloned");
5727 // Replace the operands of the cloned instructions with extracted scalars.
5728 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5729 Value *Op = Params[op][Part];
5730 Cloned->setOperand(op, Op);
5733 // Place the cloned scalar in the new loop.
5734 Builder.Insert(Cloned);
5736 // If the original scalar returns a value we need to place it in a vector
5737 // so that future users will be able to use it.
5739 VecResults[Part] = Cloned;
5742 if (IfPredicateStore) {
5743 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5744 LoopVectorBody.push_back(NewIfBlock);
5745 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5746 Builder.SetInsertPoint(InsertPt);
5747 Instruction *OldBr = IfBlock->getTerminator();
5748 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5749 OldBr->eraseFromParent();
5750 IfBlock = NewIfBlock;
5755 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5756 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5757 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5759 return scalarizeInstruction(Instr, IfPredicateStore);
5762 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5766 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5770 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5772 // When unrolling and the VF is 1, we only need to add a simple scalar.
5773 Type *ITy = Val->getType();
5774 assert(!ITy->isVectorTy() && "Val must be a scalar");
5775 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5776 return Builder.CreateAdd(Val, C, "induction");