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 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/EquivalenceClasses.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/StringExtras.h"
58 #include "llvm/Analysis/AliasAnalysis.h"
59 #include "llvm/Analysis/Dominators.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/Analysis/Verifier.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DerivedTypes.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.h"
80 #include "llvm/Pass.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/Debug.h"
83 #include "llvm/Support/PatternMatch.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/Support/ValueHandle.h"
86 #include "llvm/Target/TargetLibraryInfo.h"
87 #include "llvm/Transforms/Scalar.h"
88 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
89 #include "llvm/Transforms/Utils/Local.h"
94 using namespace llvm::PatternMatch;
96 static cl::opt<unsigned>
97 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
98 cl::desc("Sets the SIMD width. Zero is autoselect."));
100 static cl::opt<unsigned>
101 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
102 cl::desc("Sets the vectorization unroll count. "
103 "Zero is autoselect."));
106 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
107 cl::desc("Enable if-conversion during vectorization."));
109 /// We don't vectorize loops with a known constant trip count below this number.
110 static cl::opt<unsigned>
111 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
113 cl::desc("Don't vectorize loops with a constant "
114 "trip count that is smaller than this "
117 /// We don't unroll loops with a known constant trip count below this number.
118 static const unsigned TinyTripCountUnrollThreshold = 128;
120 /// When performing memory disambiguation checks at runtime do not make more
121 /// than this number of comparisons.
122 static const unsigned RuntimeMemoryCheckThreshold = 8;
124 /// Maximum simd width.
125 static const unsigned MaxVectorWidth = 64;
127 /// Maximum vectorization unroll count.
128 static const unsigned MaxUnrollFactor = 16;
130 /// The cost of a loop that is considered 'small' by the unroller.
131 static const unsigned SmallLoopCost = 20;
135 // Forward declarations.
136 class LoopVectorizationLegality;
137 class LoopVectorizationCostModel;
139 /// InnerLoopVectorizer vectorizes loops which contain only one basic
140 /// block to a specified vectorization factor (VF).
141 /// This class performs the widening of scalars into vectors, or multiple
142 /// scalars. This class also implements the following features:
143 /// * It inserts an epilogue loop for handling loops that don't have iteration
144 /// counts that are known to be a multiple of the vectorization factor.
145 /// * It handles the code generation for reduction variables.
146 /// * Scalarization (implementation using scalars) of un-vectorizable
148 /// InnerLoopVectorizer does not perform any vectorization-legality
149 /// checks, and relies on the caller to check for the different legality
150 /// aspects. The InnerLoopVectorizer relies on the
151 /// LoopVectorizationLegality class to provide information about the induction
152 /// and reduction variables that were found to a given vectorization factor.
153 class InnerLoopVectorizer {
155 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
156 DominatorTree *DT, DataLayout *DL,
157 const TargetLibraryInfo *TLI, unsigned VecWidth,
158 unsigned UnrollFactor)
159 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
160 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
161 OldInduction(0), WidenMap(UnrollFactor) {}
163 // Perform the actual loop widening (vectorization).
164 void vectorize(LoopVectorizationLegality *Legal) {
165 // Create a new empty loop. Unlink the old loop and connect the new one.
166 createEmptyLoop(Legal);
167 // Widen each instruction in the old loop to a new one in the new loop.
168 // Use the Legality module to find the induction and reduction variables.
169 vectorizeLoop(Legal);
170 // Register the new loop and update the analysis passes.
174 virtual ~InnerLoopVectorizer() {}
177 /// A small list of PHINodes.
178 typedef SmallVector<PHINode*, 4> PhiVector;
179 /// When we unroll loops we have multiple vector values for each scalar.
180 /// This data structure holds the unrolled and vectorized values that
181 /// originated from one scalar instruction.
182 typedef SmallVector<Value*, 2> VectorParts;
184 // When we if-convert we need create edge masks. We have to cache values so
185 // that we don't end up with exponential recursion/IR.
186 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
187 VectorParts> EdgeMaskCache;
189 /// Add code that checks at runtime if the accessed arrays overlap.
190 /// Returns the comparator value or NULL if no check is needed.
191 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
193 /// Create an empty loop, based on the loop ranges of the old loop.
194 void createEmptyLoop(LoopVectorizationLegality *Legal);
195 /// Copy and widen the instructions from the old loop.
196 virtual void vectorizeLoop(LoopVectorizationLegality *Legal);
198 /// \brief The Loop exit block may have single value PHI nodes where the
199 /// incoming value is 'Undef'. While vectorizing we only handled real values
200 /// that were defined inside the loop. Here we fix the 'undef case'.
204 /// A helper function that computes the predicate of the block BB, assuming
205 /// that the header block of the loop is set to True. It returns the *entry*
206 /// mask for the block BB.
207 VectorParts createBlockInMask(BasicBlock *BB);
208 /// A helper function that computes the predicate of the edge between SRC
210 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
212 /// A helper function to vectorize a single BB within the innermost loop.
213 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
216 /// Vectorize a single PHINode in a block. This method handles the induction
217 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
218 /// arbitrary length vectors.
219 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
220 LoopVectorizationLegality *Legal,
221 unsigned UF, unsigned VF, PhiVector *PV);
223 /// Insert the new loop to the loop hierarchy and pass manager
224 /// and update the analysis passes.
225 void updateAnalysis();
227 /// This instruction is un-vectorizable. Implement it as a sequence
229 virtual void scalarizeInstruction(Instruction *Instr);
231 /// Vectorize Load and Store instructions,
232 virtual void vectorizeMemoryInstruction(Instruction *Instr,
233 LoopVectorizationLegality *Legal);
235 /// Create a broadcast instruction. This method generates a broadcast
236 /// instruction (shuffle) for loop invariant values and for the induction
237 /// value. If this is the induction variable then we extend it to N, N+1, ...
238 /// this is needed because each iteration in the loop corresponds to a SIMD
240 virtual Value *getBroadcastInstrs(Value *V);
242 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
243 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
244 /// The sequence starts at StartIndex.
245 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
247 /// When we go over instructions in the basic block we rely on previous
248 /// values within the current basic block or on loop invariant values.
249 /// When we widen (vectorize) values we place them in the map. If the values
250 /// are not within the map, they have to be loop invariant, so we simply
251 /// broadcast them into a vector.
252 VectorParts &getVectorValue(Value *V);
254 /// Generate a shuffle sequence that will reverse the vector Vec.
255 virtual Value *reverseVector(Value *Vec);
257 /// This is a helper class that holds the vectorizer state. It maps scalar
258 /// instructions to vector instructions. When the code is 'unrolled' then
259 /// then a single scalar value is mapped to multiple vector parts. The parts
260 /// are stored in the VectorPart type.
262 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
264 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
266 /// \return True if 'Key' is saved in the Value Map.
267 bool has(Value *Key) const { return MapStorage.count(Key); }
269 /// Initializes a new entry in the map. Sets all of the vector parts to the
270 /// save value in 'Val'.
271 /// \return A reference to a vector with splat values.
272 VectorParts &splat(Value *Key, Value *Val) {
273 VectorParts &Entry = MapStorage[Key];
274 Entry.assign(UF, Val);
278 ///\return A reference to the value that is stored at 'Key'.
279 VectorParts &get(Value *Key) {
280 VectorParts &Entry = MapStorage[Key];
283 assert(Entry.size() == UF);
288 /// The unroll factor. Each entry in the map stores this number of vector
292 /// Map storage. We use std::map and not DenseMap because insertions to a
293 /// dense map invalidates its iterators.
294 std::map<Value *, VectorParts> MapStorage;
297 /// The original loop.
299 /// Scev analysis to use.
307 /// Target Library Info.
308 const TargetLibraryInfo *TLI;
310 /// The vectorization SIMD factor to use. Each vector will have this many
315 /// The vectorization unroll factor to use. Each scalar is vectorized to this
316 /// many different vector instructions.
319 /// The builder that we use
322 // --- Vectorization state ---
324 /// The vector-loop preheader.
325 BasicBlock *LoopVectorPreHeader;
326 /// The scalar-loop preheader.
327 BasicBlock *LoopScalarPreHeader;
328 /// Middle Block between the vector and the scalar.
329 BasicBlock *LoopMiddleBlock;
330 ///The ExitBlock of the scalar loop.
331 BasicBlock *LoopExitBlock;
332 ///The vector loop body.
333 BasicBlock *LoopVectorBody;
334 ///The scalar loop body.
335 BasicBlock *LoopScalarBody;
336 /// A list of all bypass blocks. The first block is the entry of the loop.
337 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
339 /// The new Induction variable which was added to the new block.
341 /// The induction variable of the old basic block.
342 PHINode *OldInduction;
343 /// Holds the extended (to the widest induction type) start index.
345 /// Maps scalars to widened vectors.
347 EdgeMaskCache MaskCache;
350 class InnerLoopUnroller : public InnerLoopVectorizer {
352 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
353 DominatorTree *DT, DataLayout *DL,
354 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
355 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
358 virtual void scalarizeInstruction(Instruction *Instr);
359 virtual void vectorizeMemoryInstruction(Instruction *Instr,
360 LoopVectorizationLegality *Legal);
361 virtual Value *getBroadcastInstrs(Value *V);
362 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
363 virtual Value *reverseVector(Value *Vec);
366 /// \brief Look for a meaningful debug location on the instruction or it's
368 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
373 if (I->getDebugLoc() != Empty)
376 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
377 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
378 if (OpInst->getDebugLoc() != Empty)
385 /// \brief Set the debug location in the builder using the debug location in the
387 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
388 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
389 B.SetCurrentDebugLocation(Inst->getDebugLoc());
391 B.SetCurrentDebugLocation(DebugLoc());
394 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
395 /// to what vectorization factor.
396 /// This class does not look at the profitability of vectorization, only the
397 /// legality. This class has two main kinds of checks:
398 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
399 /// will change the order of memory accesses in a way that will change the
400 /// correctness of the program.
401 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
402 /// checks for a number of different conditions, such as the availability of a
403 /// single induction variable, that all types are supported and vectorize-able,
404 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
405 /// This class is also used by InnerLoopVectorizer for identifying
406 /// induction variable and the different reduction variables.
407 class LoopVectorizationLegality {
409 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
410 DominatorTree *DT, TargetLibraryInfo *TLI)
411 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
412 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
413 MaxSafeDepDistBytes(-1U) {}
415 /// This enum represents the kinds of reductions that we support.
417 RK_NoReduction, ///< Not a reduction.
418 RK_IntegerAdd, ///< Sum of integers.
419 RK_IntegerMult, ///< Product of integers.
420 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
421 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
422 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
423 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
424 RK_FloatAdd, ///< Sum of floats.
425 RK_FloatMult, ///< Product of floats.
426 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
429 /// This enum represents the kinds of inductions that we support.
431 IK_NoInduction, ///< Not an induction variable.
432 IK_IntInduction, ///< Integer induction variable. Step = 1.
433 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
434 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
435 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
438 // This enum represents the kind of minmax reduction.
439 enum MinMaxReductionKind {
449 /// This struct holds information about reduction variables.
450 struct ReductionDescriptor {
451 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
452 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
454 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
455 MinMaxReductionKind MK)
456 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
458 // The starting value of the reduction.
459 // It does not have to be zero!
460 TrackingVH<Value> StartValue;
461 // The instruction who's value is used outside the loop.
462 Instruction *LoopExitInstr;
463 // The kind of the reduction.
465 // If this a min/max reduction the kind of reduction.
466 MinMaxReductionKind MinMaxKind;
469 /// This POD struct holds information about a potential reduction operation.
470 struct ReductionInstDesc {
471 ReductionInstDesc(bool IsRedux, Instruction *I) :
472 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
474 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
475 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
477 // Is this instruction a reduction candidate.
479 // The last instruction in a min/max pattern (select of the select(icmp())
480 // pattern), or the current reduction instruction otherwise.
481 Instruction *PatternLastInst;
482 // If this is a min/max pattern the comparison predicate.
483 MinMaxReductionKind MinMaxKind;
486 /// This struct holds information about the memory runtime legality
487 /// check that a group of pointers do not overlap.
488 struct RuntimePointerCheck {
489 RuntimePointerCheck() : Need(false) {}
491 /// Reset the state of the pointer runtime information.
498 DependencySetId.clear();
501 /// Insert a pointer and calculate the start and end SCEVs.
502 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
505 /// This flag indicates if we need to add the runtime check.
507 /// Holds the pointers that we need to check.
508 SmallVector<TrackingVH<Value>, 2> Pointers;
509 /// Holds the pointer value at the beginning of the loop.
510 SmallVector<const SCEV*, 2> Starts;
511 /// Holds the pointer value at the end of the loop.
512 SmallVector<const SCEV*, 2> Ends;
513 /// Holds the information if this pointer is used for writing to memory.
514 SmallVector<bool, 2> IsWritePtr;
515 /// Holds the id of the set of pointers that could be dependent because of a
516 /// shared underlying object.
517 SmallVector<unsigned, 2> DependencySetId;
520 /// A struct for saving information about induction variables.
521 struct InductionInfo {
522 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
523 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
525 TrackingVH<Value> StartValue;
530 /// ReductionList contains the reduction descriptors for all
531 /// of the reductions that were found in the loop.
532 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
534 /// InductionList saves induction variables and maps them to the
535 /// induction descriptor.
536 typedef MapVector<PHINode*, InductionInfo> InductionList;
538 /// Returns true if it is legal to vectorize this loop.
539 /// This does not mean that it is profitable to vectorize this
540 /// loop, only that it is legal to do so.
543 /// Returns the Induction variable.
544 PHINode *getInduction() { return Induction; }
546 /// Returns the reduction variables found in the loop.
547 ReductionList *getReductionVars() { return &Reductions; }
549 /// Returns the induction variables found in the loop.
550 InductionList *getInductionVars() { return &Inductions; }
552 /// Returns the widest induction type.
553 Type *getWidestInductionType() { return WidestIndTy; }
555 /// Returns True if V is an induction variable in this loop.
556 bool isInductionVariable(const Value *V);
558 /// Return true if the block BB needs to be predicated in order for the loop
559 /// to be vectorized.
560 bool blockNeedsPredication(BasicBlock *BB);
562 /// Check if this pointer is consecutive when vectorizing. This happens
563 /// when the last index of the GEP is the induction variable, or that the
564 /// pointer itself is an induction variable.
565 /// This check allows us to vectorize A[idx] into a wide load/store.
567 /// 0 - Stride is unknown or non consecutive.
568 /// 1 - Address is consecutive.
569 /// -1 - Address is consecutive, and decreasing.
570 int isConsecutivePtr(Value *Ptr);
572 /// Returns true if the value V is uniform within the loop.
573 bool isUniform(Value *V);
575 /// Returns true if this instruction will remain scalar after vectorization.
576 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
578 /// Returns the information that we collected about runtime memory check.
579 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
581 /// This function returns the identity element (or neutral element) for
583 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
585 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
588 /// Check if a single basic block loop is vectorizable.
589 /// At this point we know that this is a loop with a constant trip count
590 /// and we only need to check individual instructions.
591 bool canVectorizeInstrs();
593 /// When we vectorize loops we may change the order in which
594 /// we read and write from memory. This method checks if it is
595 /// legal to vectorize the code, considering only memory constrains.
596 /// Returns true if the loop is vectorizable
597 bool canVectorizeMemory();
599 /// Return true if we can vectorize this loop using the IF-conversion
601 bool canVectorizeWithIfConvert();
603 /// Collect the variables that need to stay uniform after vectorization.
604 void collectLoopUniforms();
606 /// Return true if all of the instructions in the block can be speculatively
607 /// executed. \p SafePtrs is a list of addresses that are known to be legal
608 /// and we know that we can read from them without segfault.
609 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
611 /// Returns True, if 'Phi' is the kind of reduction variable for type
612 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
613 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
614 /// Returns a struct describing if the instruction 'I' can be a reduction
615 /// variable of type 'Kind'. If the reduction is a min/max pattern of
616 /// select(icmp()) this function advances the instruction pointer 'I' from the
617 /// compare instruction to the select instruction and stores this pointer in
618 /// 'PatternLastInst' member of the returned struct.
619 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
620 ReductionInstDesc &Desc);
621 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
622 /// pattern corresponding to a min(X, Y) or max(X, Y).
623 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
624 ReductionInstDesc &Prev);
625 /// Returns the induction kind of Phi. This function may return NoInduction
626 /// if the PHI is not an induction variable.
627 InductionKind isInductionVariable(PHINode *Phi);
629 /// The loop that we evaluate.
633 /// DataLayout analysis.
637 /// Target Library Info.
638 TargetLibraryInfo *TLI;
640 // --- vectorization state --- //
642 /// Holds the integer induction variable. This is the counter of the
645 /// Holds the reduction variables.
646 ReductionList Reductions;
647 /// Holds all of the induction variables that we found in the loop.
648 /// Notice that inductions don't need to start at zero and that induction
649 /// variables can be pointers.
650 InductionList Inductions;
651 /// Holds the widest induction type encountered.
654 /// Allowed outside users. This holds the reduction
655 /// vars which can be accessed from outside the loop.
656 SmallPtrSet<Value*, 4> AllowedExit;
657 /// This set holds the variables which are known to be uniform after
659 SmallPtrSet<Instruction*, 4> Uniforms;
660 /// We need to check that all of the pointers in this list are disjoint
662 RuntimePointerCheck PtrRtCheck;
663 /// Can we assume the absence of NaNs.
664 bool HasFunNoNaNAttr;
666 unsigned MaxSafeDepDistBytes;
669 /// LoopVectorizationCostModel - estimates the expected speedups due to
671 /// In many cases vectorization is not profitable. This can happen because of
672 /// a number of reasons. In this class we mainly attempt to predict the
673 /// expected speedup/slowdowns due to the supported instruction set. We use the
674 /// TargetTransformInfo to query the different backends for the cost of
675 /// different operations.
676 class LoopVectorizationCostModel {
678 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
679 LoopVectorizationLegality *Legal,
680 const TargetTransformInfo &TTI,
681 DataLayout *DL, const TargetLibraryInfo *TLI)
682 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
684 /// Information about vectorization costs
685 struct VectorizationFactor {
686 unsigned Width; // Vector width with best cost
687 unsigned Cost; // Cost of the loop with that width
689 /// \return The most profitable vectorization factor and the cost of that VF.
690 /// This method checks every power of two up to VF. If UserVF is not ZERO
691 /// then this vectorization factor will be selected if vectorization is
693 VectorizationFactor selectVectorizationFactor(bool OptForSize,
696 /// \return The size (in bits) of the widest type in the code that
697 /// needs to be vectorized. We ignore values that remain scalar such as
698 /// 64 bit loop indices.
699 unsigned getWidestType();
701 /// \return The most profitable unroll factor.
702 /// If UserUF is non-zero then this method finds the best unroll-factor
703 /// based on register pressure and other parameters.
704 /// VF and LoopCost are the selected vectorization factor and the cost of the
706 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
709 /// \brief A struct that represents some properties of the register usage
711 struct RegisterUsage {
712 /// Holds the number of loop invariant values that are used in the loop.
713 unsigned LoopInvariantRegs;
714 /// Holds the maximum number of concurrent live intervals in the loop.
715 unsigned MaxLocalUsers;
716 /// Holds the number of instructions in the loop.
717 unsigned NumInstructions;
720 /// \return information about the register usage of the loop.
721 RegisterUsage calculateRegisterUsage();
724 /// Returns the expected execution cost. The unit of the cost does
725 /// not matter because we use the 'cost' units to compare different
726 /// vector widths. The cost that is returned is *not* normalized by
727 /// the factor width.
728 unsigned expectedCost(unsigned VF);
730 /// Returns the execution time cost of an instruction for a given vector
731 /// width. Vector width of one means scalar.
732 unsigned getInstructionCost(Instruction *I, unsigned VF);
734 /// A helper function for converting Scalar types to vector types.
735 /// If the incoming type is void, we return void. If the VF is 1, we return
737 static Type* ToVectorTy(Type *Scalar, unsigned VF);
739 /// Returns whether the instruction is a load or store and will be a emitted
740 /// as a vector operation.
741 bool isConsecutiveLoadOrStore(Instruction *I);
743 /// The loop that we evaluate.
747 /// Loop Info analysis.
749 /// Vectorization legality.
750 LoopVectorizationLegality *Legal;
751 /// Vector target information.
752 const TargetTransformInfo &TTI;
753 /// Target data layout information.
755 /// Target Library Info.
756 const TargetLibraryInfo *TLI;
759 /// Utility class for getting and setting loop vectorizer hints in the form
760 /// of loop metadata.
761 struct LoopVectorizeHints {
762 /// Vectorization width.
764 /// Vectorization unroll factor.
767 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
768 : Width(VectorizationFactor)
769 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
770 , LoopID(L->getLoopID()) {
772 // The command line options override any loop metadata except for when
773 // width == 1 which is used to indicate the loop is already vectorized.
774 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
775 Width = VectorizationFactor;
776 if (VectorizationUnroll.getNumOccurrences() > 0)
777 Unroll = VectorizationUnroll;
779 DEBUG(if (DisableUnrolling && Unroll == 1)
780 dbgs() << "LV: Unrolling disabled by the pass manager\n");
783 /// Return the loop vectorizer metadata prefix.
784 static StringRef Prefix() { return "llvm.vectorizer."; }
786 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
787 SmallVector<Value*, 2> Vals;
788 Vals.push_back(MDString::get(Context, Name));
789 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
790 return MDNode::get(Context, Vals);
793 /// Mark the loop L as already vectorized by setting the width to 1.
794 void setAlreadyVectorized(Loop *L) {
795 LLVMContext &Context = L->getHeader()->getContext();
799 // Create a new loop id with one more operand for the already_vectorized
800 // hint. If the loop already has a loop id then copy the existing operands.
801 SmallVector<Value*, 4> Vals(1);
803 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
804 Vals.push_back(LoopID->getOperand(i));
806 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
807 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
809 MDNode *NewLoopID = MDNode::get(Context, Vals);
810 // Set operand 0 to refer to the loop id itself.
811 NewLoopID->replaceOperandWith(0, NewLoopID);
813 L->setLoopID(NewLoopID);
815 LoopID->replaceAllUsesWith(NewLoopID);
823 /// Find hints specified in the loop metadata.
824 void getHints(const Loop *L) {
828 // First operand should refer to the loop id itself.
829 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
830 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
832 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
833 const MDString *S = 0;
834 SmallVector<Value*, 4> Args;
836 // The expected hint is either a MDString or a MDNode with the first
837 // operand a MDString.
838 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
839 if (!MD || MD->getNumOperands() == 0)
841 S = dyn_cast<MDString>(MD->getOperand(0));
842 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
843 Args.push_back(MD->getOperand(i));
845 S = dyn_cast<MDString>(LoopID->getOperand(i));
846 assert(Args.size() == 0 && "too many arguments for MDString");
852 // Check if the hint starts with the vectorizer prefix.
853 StringRef Hint = S->getString();
854 if (!Hint.startswith(Prefix()))
856 // Remove the prefix.
857 Hint = Hint.substr(Prefix().size(), StringRef::npos);
859 if (Args.size() == 1)
860 getHint(Hint, Args[0]);
864 // Check string hint with one operand.
865 void getHint(StringRef Hint, Value *Arg) {
866 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
868 unsigned Val = C->getZExtValue();
870 if (Hint == "width") {
871 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
874 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
875 } else if (Hint == "unroll") {
876 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
879 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
881 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
886 /// The LoopVectorize Pass.
887 struct LoopVectorize : public LoopPass {
888 /// Pass identification, replacement for typeid
891 explicit LoopVectorize(bool NoUnrolling = false)
892 : LoopPass(ID), DisableUnrolling(NoUnrolling) {
893 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
899 TargetTransformInfo *TTI;
901 TargetLibraryInfo *TLI;
902 bool DisableUnrolling;
904 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
905 // We only vectorize innermost loops.
909 SE = &getAnalysis<ScalarEvolution>();
910 DL = getAnalysisIfAvailable<DataLayout>();
911 LI = &getAnalysis<LoopInfo>();
912 TTI = &getAnalysis<TargetTransformInfo>();
913 DT = &getAnalysis<DominatorTree>();
914 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
916 // If the target claims to have no vector registers don't attempt
918 if (!TTI->getNumberOfRegisters(true))
922 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout\n");
926 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
927 L->getHeader()->getParent()->getName() << "\"\n");
929 LoopVectorizeHints Hints(L, DisableUnrolling);
931 if (Hints.Width == 1 && Hints.Unroll == 1) {
932 DEBUG(dbgs() << "LV: Not vectorizing.\n");
936 // Check if it is legal to vectorize the loop.
937 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
938 if (!LVL.canVectorize()) {
939 DEBUG(dbgs() << "LV: Not vectorizing.\n");
943 // Use the cost model.
944 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
946 // Check the function attributes to find out if this function should be
947 // optimized for size.
948 Function *F = L->getHeader()->getParent();
949 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
950 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
951 unsigned FnIndex = AttributeSet::FunctionIndex;
952 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
953 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
956 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
957 "attribute is used.\n");
961 // Select the optimal vectorization factor.
962 LoopVectorizationCostModel::VectorizationFactor VF;
963 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
964 // Select the unroll factor.
965 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
968 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
969 F->getParent()->getModuleIdentifier() << '\n');
970 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
973 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
976 // We decided not to vectorize, but we may want to unroll.
977 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
978 Unroller.vectorize(&LVL);
980 // If we decided that it is *legal* to vectorize the loop then do it.
981 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
985 // Mark the loop as already vectorized to avoid vectorizing again.
986 Hints.setAlreadyVectorized(L);
988 DEBUG(verifyFunction(*L->getHeader()->getParent()));
992 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
993 LoopPass::getAnalysisUsage(AU);
994 AU.addRequiredID(LoopSimplifyID);
995 AU.addRequiredID(LCSSAID);
996 AU.addRequired<DominatorTree>();
997 AU.addRequired<LoopInfo>();
998 AU.addRequired<ScalarEvolution>();
999 AU.addRequired<TargetTransformInfo>();
1000 AU.addPreserved<LoopInfo>();
1001 AU.addPreserved<DominatorTree>();
1006 } // end anonymous namespace
1008 //===----------------------------------------------------------------------===//
1009 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1010 // LoopVectorizationCostModel.
1011 //===----------------------------------------------------------------------===//
1014 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1015 Loop *Lp, Value *Ptr,
1017 unsigned DepSetId) {
1018 const SCEV *Sc = SE->getSCEV(Ptr);
1019 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1020 assert(AR && "Invalid addrec expression");
1021 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1022 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1023 Pointers.push_back(Ptr);
1024 Starts.push_back(AR->getStart());
1025 Ends.push_back(ScEnd);
1026 IsWritePtr.push_back(WritePtr);
1027 DependencySetId.push_back(DepSetId);
1030 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1031 // We need to place the broadcast of invariant variables outside the loop.
1032 Instruction *Instr = dyn_cast<Instruction>(V);
1033 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1034 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1036 // Place the code for broadcasting invariant variables in the new preheader.
1037 IRBuilder<>::InsertPointGuard Guard(Builder);
1039 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1041 // Broadcast the scalar into all locations in the vector.
1042 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1047 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1049 assert(Val->getType()->isVectorTy() && "Must be a vector");
1050 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1051 "Elem must be an integer");
1052 // Create the types.
1053 Type *ITy = Val->getType()->getScalarType();
1054 VectorType *Ty = cast<VectorType>(Val->getType());
1055 int VLen = Ty->getNumElements();
1056 SmallVector<Constant*, 8> Indices;
1058 // Create a vector of consecutive numbers from zero to VF.
1059 for (int i = 0; i < VLen; ++i) {
1060 int64_t Idx = Negate ? (-i) : i;
1061 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1064 // Add the consecutive indices to the vector value.
1065 Constant *Cv = ConstantVector::get(Indices);
1066 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1067 return Builder.CreateAdd(Val, Cv, "induction");
1070 /// \brief Find the operand of the GEP that should be checked for consecutive
1071 /// stores. This ignores trailing indices that have no effect on the final
1073 static unsigned getGEPInductionOperand(DataLayout *DL,
1074 const GetElementPtrInst *Gep) {
1075 unsigned LastOperand = Gep->getNumOperands() - 1;
1076 unsigned GEPAllocSize = DL->getTypeAllocSize(
1077 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1079 // Walk backwards and try to peel off zeros.
1080 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1081 // Find the type we're currently indexing into.
1082 gep_type_iterator GEPTI = gep_type_begin(Gep);
1083 std::advance(GEPTI, LastOperand - 1);
1085 // If it's a type with the same allocation size as the result of the GEP we
1086 // can peel off the zero index.
1087 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1095 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1096 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1097 // Make sure that the pointer does not point to structs.
1098 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1101 // If this value is a pointer induction variable we know it is consecutive.
1102 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1103 if (Phi && Inductions.count(Phi)) {
1104 InductionInfo II = Inductions[Phi];
1105 if (IK_PtrInduction == II.IK)
1107 else if (IK_ReversePtrInduction == II.IK)
1111 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1115 unsigned NumOperands = Gep->getNumOperands();
1116 Value *GpPtr = Gep->getPointerOperand();
1117 // If this GEP value is a consecutive pointer induction variable and all of
1118 // the indices are constant then we know it is consecutive. We can
1119 Phi = dyn_cast<PHINode>(GpPtr);
1120 if (Phi && Inductions.count(Phi)) {
1122 // Make sure that the pointer does not point to structs.
1123 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1124 if (GepPtrType->getElementType()->isAggregateType())
1127 // Make sure that all of the index operands are loop invariant.
1128 for (unsigned i = 1; i < NumOperands; ++i)
1129 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1132 InductionInfo II = Inductions[Phi];
1133 if (IK_PtrInduction == II.IK)
1135 else if (IK_ReversePtrInduction == II.IK)
1139 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1141 // Check that all of the gep indices are uniform except for our induction
1143 for (unsigned i = 0; i != NumOperands; ++i)
1144 if (i != InductionOperand &&
1145 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1148 // We can emit wide load/stores only if the last non-zero index is the
1149 // induction variable.
1150 const SCEV *Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1151 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1152 const SCEV *Step = AR->getStepRecurrence(*SE);
1154 // The memory is consecutive because the last index is consecutive
1155 // and all other indices are loop invariant.
1158 if (Step->isAllOnesValue())
1165 bool LoopVectorizationLegality::isUniform(Value *V) {
1166 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1169 InnerLoopVectorizer::VectorParts&
1170 InnerLoopVectorizer::getVectorValue(Value *V) {
1171 assert(V != Induction && "The new induction variable should not be used.");
1172 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1174 // If we have this scalar in the map, return it.
1175 if (WidenMap.has(V))
1176 return WidenMap.get(V);
1178 // If this scalar is unknown, assume that it is a constant or that it is
1179 // loop invariant. Broadcast V and save the value for future uses.
1180 Value *B = getBroadcastInstrs(V);
1181 return WidenMap.splat(V, B);
1184 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1185 assert(Vec->getType()->isVectorTy() && "Invalid type");
1186 SmallVector<Constant*, 8> ShuffleMask;
1187 for (unsigned i = 0; i < VF; ++i)
1188 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1190 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1191 ConstantVector::get(ShuffleMask),
1196 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1197 LoopVectorizationLegality *Legal) {
1198 // Attempt to issue a wide load.
1199 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1200 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1202 assert((LI || SI) && "Invalid Load/Store instruction");
1204 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1205 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1206 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1207 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1208 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1209 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1210 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1212 if (ScalarAllocatedSize != VectorElementSize)
1213 return scalarizeInstruction(Instr);
1215 // If the pointer is loop invariant or if it is non consecutive,
1216 // scalarize the load.
1217 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1218 bool Reverse = ConsecutiveStride < 0;
1219 bool UniformLoad = LI && Legal->isUniform(Ptr);
1220 if (!ConsecutiveStride || UniformLoad)
1221 return scalarizeInstruction(Instr);
1223 Constant *Zero = Builder.getInt32(0);
1224 VectorParts &Entry = WidenMap.get(Instr);
1226 // Handle consecutive loads/stores.
1227 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1228 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1229 setDebugLocFromInst(Builder, Gep);
1230 Value *PtrOperand = Gep->getPointerOperand();
1231 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1232 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1234 // Create the new GEP with the new induction variable.
1235 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1236 Gep2->setOperand(0, FirstBasePtr);
1237 Gep2->setName("gep.indvar.base");
1238 Ptr = Builder.Insert(Gep2);
1240 setDebugLocFromInst(Builder, Gep);
1241 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1242 OrigLoop) && "Base ptr must be invariant");
1244 // The last index does not have to be the induction. It can be
1245 // consecutive and be a function of the index. For example A[I+1];
1246 unsigned NumOperands = Gep->getNumOperands();
1247 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1248 // Create the new GEP with the new induction variable.
1249 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1251 for (unsigned i = 0; i < NumOperands; ++i) {
1252 Value *GepOperand = Gep->getOperand(i);
1253 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1255 // Update last index or loop invariant instruction anchored in loop.
1256 if (i == InductionOperand ||
1257 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1258 assert((i == InductionOperand ||
1259 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1260 "Must be last index or loop invariant");
1262 VectorParts &GEPParts = getVectorValue(GepOperand);
1263 Value *Index = GEPParts[0];
1264 Index = Builder.CreateExtractElement(Index, Zero);
1265 Gep2->setOperand(i, Index);
1266 Gep2->setName("gep.indvar.idx");
1269 Ptr = Builder.Insert(Gep2);
1271 // Use the induction element ptr.
1272 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1273 setDebugLocFromInst(Builder, Ptr);
1274 VectorParts &PtrVal = getVectorValue(Ptr);
1275 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1280 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1281 "We do not allow storing to uniform addresses");
1282 setDebugLocFromInst(Builder, SI);
1283 // We don't want to update the value in the map as it might be used in
1284 // another expression. So don't use a reference type for "StoredVal".
1285 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1287 for (unsigned Part = 0; Part < UF; ++Part) {
1288 // Calculate the pointer for the specific unroll-part.
1289 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1292 // If we store to reverse consecutive memory locations then we need
1293 // to reverse the order of elements in the stored value.
1294 StoredVal[Part] = reverseVector(StoredVal[Part]);
1295 // If the address is consecutive but reversed, then the
1296 // wide store needs to start at the last vector element.
1297 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1298 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1301 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1302 DataTy->getPointerTo(AddressSpace));
1303 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1309 assert(LI && "Must have a load instruction");
1310 setDebugLocFromInst(Builder, LI);
1311 for (unsigned Part = 0; Part < UF; ++Part) {
1312 // Calculate the pointer for the specific unroll-part.
1313 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1316 // If the address is consecutive but reversed, then the
1317 // wide store needs to start at the last vector element.
1318 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1319 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1322 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1323 DataTy->getPointerTo(AddressSpace));
1324 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1325 cast<LoadInst>(LI)->setAlignment(Alignment);
1326 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1330 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1331 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1332 // Holds vector parameters or scalars, in case of uniform vals.
1333 SmallVector<VectorParts, 4> Params;
1335 setDebugLocFromInst(Builder, Instr);
1337 // Find all of the vectorized parameters.
1338 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1339 Value *SrcOp = Instr->getOperand(op);
1341 // If we are accessing the old induction variable, use the new one.
1342 if (SrcOp == OldInduction) {
1343 Params.push_back(getVectorValue(SrcOp));
1347 // Try using previously calculated values.
1348 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1350 // If the src is an instruction that appeared earlier in the basic block
1351 // then it should already be vectorized.
1352 if (SrcInst && OrigLoop->contains(SrcInst)) {
1353 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1354 // The parameter is a vector value from earlier.
1355 Params.push_back(WidenMap.get(SrcInst));
1357 // The parameter is a scalar from outside the loop. Maybe even a constant.
1358 VectorParts Scalars;
1359 Scalars.append(UF, SrcOp);
1360 Params.push_back(Scalars);
1364 assert(Params.size() == Instr->getNumOperands() &&
1365 "Invalid number of operands");
1367 // Does this instruction return a value ?
1368 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1370 Value *UndefVec = IsVoidRetTy ? 0 :
1371 UndefValue::get(VectorType::get(Instr->getType(), VF));
1372 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1373 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1375 // For each vector unroll 'part':
1376 for (unsigned Part = 0; Part < UF; ++Part) {
1377 // For each scalar that we create:
1378 for (unsigned Width = 0; Width < VF; ++Width) {
1379 Instruction *Cloned = Instr->clone();
1381 Cloned->setName(Instr->getName() + ".cloned");
1382 // Replace the operands of the cloned instructions with extracted scalars.
1383 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1384 Value *Op = Params[op][Part];
1385 // Param is a vector. Need to extract the right lane.
1386 if (Op->getType()->isVectorTy())
1387 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1388 Cloned->setOperand(op, Op);
1391 // Place the cloned scalar in the new loop.
1392 Builder.Insert(Cloned);
1394 // If the original scalar returns a value we need to place it in a vector
1395 // so that future users will be able to use it.
1397 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1398 Builder.getInt32(Width));
1404 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1406 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1407 Legal->getRuntimePointerCheck();
1409 if (!PtrRtCheck->Need)
1412 unsigned NumPointers = PtrRtCheck->Pointers.size();
1413 SmallVector<TrackingVH<Value> , 2> Starts;
1414 SmallVector<TrackingVH<Value> , 2> Ends;
1416 LLVMContext &Ctx = Loc->getContext();
1417 SCEVExpander Exp(*SE, "induction");
1419 for (unsigned i = 0; i < NumPointers; ++i) {
1420 Value *Ptr = PtrRtCheck->Pointers[i];
1421 const SCEV *Sc = SE->getSCEV(Ptr);
1423 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1424 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1426 Starts.push_back(Ptr);
1427 Ends.push_back(Ptr);
1429 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1430 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1432 // Use this type for pointer arithmetic.
1433 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1435 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1436 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1437 Starts.push_back(Start);
1438 Ends.push_back(End);
1442 IRBuilder<> ChkBuilder(Loc);
1443 // Our instructions might fold to a constant.
1444 Value *MemoryRuntimeCheck = 0;
1445 for (unsigned i = 0; i < NumPointers; ++i) {
1446 for (unsigned j = i+1; j < NumPointers; ++j) {
1447 // No need to check if two readonly pointers intersect.
1448 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1451 // Only need to check pointers between two different dependency sets.
1452 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1455 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1456 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1458 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1459 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1460 "Trying to bounds check pointers with different address spaces");
1462 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1463 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1465 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1466 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1467 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1468 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1470 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1471 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1472 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1473 if (MemoryRuntimeCheck)
1474 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1476 MemoryRuntimeCheck = IsConflict;
1480 // We have to do this trickery because the IRBuilder might fold the check to a
1481 // constant expression in which case there is no Instruction anchored in a
1483 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1484 ConstantInt::getTrue(Ctx));
1485 ChkBuilder.Insert(Check, "memcheck.conflict");
1490 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1492 In this function we generate a new loop. The new loop will contain
1493 the vectorized instructions while the old loop will continue to run the
1496 [ ] <-- vector loop bypass (may consist of multiple blocks).
1499 | [ ] <-- vector pre header.
1503 | [ ]_| <-- vector loop.
1506 >[ ] <--- middle-block.
1509 | [ ] <--- new preheader.
1513 | [ ]_| <-- old scalar loop to handle remainder.
1516 >[ ] <-- exit block.
1520 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1521 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1522 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1523 assert(ExitBlock && "Must have an exit block");
1525 // Some loops have a single integer induction variable, while other loops
1526 // don't. One example is c++ iterators that often have multiple pointer
1527 // induction variables. In the code below we also support a case where we
1528 // don't have a single induction variable.
1529 OldInduction = Legal->getInduction();
1530 Type *IdxTy = Legal->getWidestInductionType();
1532 // Find the loop boundaries.
1533 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1534 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1536 // Get the total trip count from the count by adding 1.
1537 ExitCount = SE->getAddExpr(ExitCount,
1538 SE->getConstant(ExitCount->getType(), 1));
1540 // Expand the trip count and place the new instructions in the preheader.
1541 // Notice that the pre-header does not change, only the loop body.
1542 SCEVExpander Exp(*SE, "induction");
1544 // Count holds the overall loop count (N).
1545 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1546 BypassBlock->getTerminator());
1548 // The loop index does not have to start at Zero. Find the original start
1549 // value from the induction PHI node. If we don't have an induction variable
1550 // then we know that it starts at zero.
1551 Builder.SetInsertPoint(BypassBlock->getTerminator());
1552 Value *StartIdx = ExtendedIdx = OldInduction ?
1553 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1555 ConstantInt::get(IdxTy, 0);
1557 assert(BypassBlock && "Invalid loop structure");
1558 LoopBypassBlocks.push_back(BypassBlock);
1560 // Split the single block loop into the two loop structure described above.
1561 BasicBlock *VectorPH =
1562 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1563 BasicBlock *VecBody =
1564 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1565 BasicBlock *MiddleBlock =
1566 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1567 BasicBlock *ScalarPH =
1568 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1570 // Create and register the new vector loop.
1571 Loop* Lp = new Loop();
1572 Loop *ParentLoop = OrigLoop->getParentLoop();
1574 // Insert the new loop into the loop nest and register the new basic blocks
1575 // before calling any utilities such as SCEV that require valid LoopInfo.
1577 ParentLoop->addChildLoop(Lp);
1578 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1579 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1580 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1582 LI->addTopLevelLoop(Lp);
1584 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1586 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1588 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1590 // Generate the induction variable.
1591 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1592 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1593 // The loop step is equal to the vectorization factor (num of SIMD elements)
1594 // times the unroll factor (num of SIMD instructions).
1595 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1597 // This is the IR builder that we use to add all of the logic for bypassing
1598 // the new vector loop.
1599 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1600 setDebugLocFromInst(BypassBuilder,
1601 getDebugLocFromInstOrOperands(OldInduction));
1603 // We may need to extend the index in case there is a type mismatch.
1604 // We know that the count starts at zero and does not overflow.
1605 if (Count->getType() != IdxTy) {
1606 // The exit count can be of pointer type. Convert it to the correct
1608 if (ExitCount->getType()->isPointerTy())
1609 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1611 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1614 // Add the start index to the loop count to get the new end index.
1615 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1617 // Now we need to generate the expression for N - (N % VF), which is
1618 // the part that the vectorized body will execute.
1619 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1620 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1621 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1622 "end.idx.rnd.down");
1624 // Now, compare the new count to zero. If it is zero skip the vector loop and
1625 // jump to the scalar loop.
1626 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1629 BasicBlock *LastBypassBlock = BypassBlock;
1631 // Generate the code that checks in runtime if arrays overlap. We put the
1632 // checks into a separate block to make the more common case of few elements
1634 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1635 BypassBlock->getTerminator());
1636 if (MemRuntimeCheck) {
1637 // Create a new block containing the memory check.
1638 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1641 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1642 LoopBypassBlocks.push_back(CheckBlock);
1644 // Replace the branch into the memory check block with a conditional branch
1645 // for the "few elements case".
1646 Instruction *OldTerm = BypassBlock->getTerminator();
1647 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1648 OldTerm->eraseFromParent();
1650 Cmp = MemRuntimeCheck;
1651 LastBypassBlock = CheckBlock;
1654 LastBypassBlock->getTerminator()->eraseFromParent();
1655 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1658 // We are going to resume the execution of the scalar loop.
1659 // Go over all of the induction variables that we found and fix the
1660 // PHIs that are left in the scalar version of the loop.
1661 // The starting values of PHI nodes depend on the counter of the last
1662 // iteration in the vectorized loop.
1663 // If we come from a bypass edge then we need to start from the original
1666 // This variable saves the new starting index for the scalar loop.
1667 PHINode *ResumeIndex = 0;
1668 LoopVectorizationLegality::InductionList::iterator I, E;
1669 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1670 // Set builder to point to last bypass block.
1671 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1672 for (I = List->begin(), E = List->end(); I != E; ++I) {
1673 PHINode *OrigPhi = I->first;
1674 LoopVectorizationLegality::InductionInfo II = I->second;
1676 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1677 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1678 MiddleBlock->getTerminator());
1679 // We might have extended the type of the induction variable but we need a
1680 // truncated version for the scalar loop.
1681 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1682 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1683 MiddleBlock->getTerminator()) : 0;
1685 Value *EndValue = 0;
1687 case LoopVectorizationLegality::IK_NoInduction:
1688 llvm_unreachable("Unknown induction");
1689 case LoopVectorizationLegality::IK_IntInduction: {
1690 // Handle the integer induction counter.
1691 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1693 // We have the canonical induction variable.
1694 if (OrigPhi == OldInduction) {
1695 // Create a truncated version of the resume value for the scalar loop,
1696 // we might have promoted the type to a larger width.
1698 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1699 // The new PHI merges the original incoming value, in case of a bypass,
1700 // or the value at the end of the vectorized loop.
1701 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1702 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1703 TruncResumeVal->addIncoming(EndValue, VecBody);
1705 // We know what the end value is.
1706 EndValue = IdxEndRoundDown;
1707 // We also know which PHI node holds it.
1708 ResumeIndex = ResumeVal;
1712 // Not the canonical induction variable - add the vector loop count to the
1714 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1715 II.StartValue->getType(),
1717 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1720 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1721 // Convert the CountRoundDown variable to the PHI size.
1722 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1723 II.StartValue->getType(),
1725 // Handle reverse integer induction counter.
1726 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1729 case LoopVectorizationLegality::IK_PtrInduction: {
1730 // For pointer induction variables, calculate the offset using
1732 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1736 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1737 // The value at the end of the loop for the reverse pointer is calculated
1738 // by creating a GEP with a negative index starting from the start value.
1739 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1740 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1742 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1748 // The new PHI merges the original incoming value, in case of a bypass,
1749 // or the value at the end of the vectorized loop.
1750 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1751 if (OrigPhi == OldInduction)
1752 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1754 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1756 ResumeVal->addIncoming(EndValue, VecBody);
1758 // Fix the scalar body counter (PHI node).
1759 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1760 // The old inductions phi node in the scalar body needs the truncated value.
1761 if (OrigPhi == OldInduction)
1762 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1764 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1767 // If we are generating a new induction variable then we also need to
1768 // generate the code that calculates the exit value. This value is not
1769 // simply the end of the counter because we may skip the vectorized body
1770 // in case of a runtime check.
1772 assert(!ResumeIndex && "Unexpected resume value found");
1773 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1774 MiddleBlock->getTerminator());
1775 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1776 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1777 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1780 // Make sure that we found the index where scalar loop needs to continue.
1781 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1782 "Invalid resume Index");
1784 // Add a check in the middle block to see if we have completed
1785 // all of the iterations in the first vector loop.
1786 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1787 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1788 ResumeIndex, "cmp.n",
1789 MiddleBlock->getTerminator());
1791 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1792 // Remove the old terminator.
1793 MiddleBlock->getTerminator()->eraseFromParent();
1795 // Create i+1 and fill the PHINode.
1796 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1797 Induction->addIncoming(StartIdx, VectorPH);
1798 Induction->addIncoming(NextIdx, VecBody);
1799 // Create the compare.
1800 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1801 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1803 // Now we have two terminators. Remove the old one from the block.
1804 VecBody->getTerminator()->eraseFromParent();
1806 // Get ready to start creating new instructions into the vectorized body.
1807 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1810 LoopVectorPreHeader = VectorPH;
1811 LoopScalarPreHeader = ScalarPH;
1812 LoopMiddleBlock = MiddleBlock;
1813 LoopExitBlock = ExitBlock;
1814 LoopVectorBody = VecBody;
1815 LoopScalarBody = OldBasicBlock;
1817 LoopVectorizeHints Hints(Lp, true);
1818 Hints.setAlreadyVectorized(Lp);
1821 /// This function returns the identity element (or neutral element) for
1822 /// the operation K.
1824 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1829 // Adding, Xoring, Oring zero to a number does not change it.
1830 return ConstantInt::get(Tp, 0);
1831 case RK_IntegerMult:
1832 // Multiplying a number by 1 does not change it.
1833 return ConstantInt::get(Tp, 1);
1835 // AND-ing a number with an all-1 value does not change it.
1836 return ConstantInt::get(Tp, -1, true);
1838 // Multiplying a number by 1 does not change it.
1839 return ConstantFP::get(Tp, 1.0L);
1841 // Adding zero to a number does not change it.
1842 return ConstantFP::get(Tp, 0.0L);
1844 llvm_unreachable("Unknown reduction kind");
1848 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
1849 Intrinsic::ID ValidIntrinsicID) {
1850 if (I.getNumArgOperands() != 1 ||
1851 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1852 I.getType() != I.getArgOperand(0)->getType() ||
1853 !I.onlyReadsMemory())
1854 return Intrinsic::not_intrinsic;
1856 return ValidIntrinsicID;
1859 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
1860 Intrinsic::ID ValidIntrinsicID) {
1861 if (I.getNumArgOperands() != 2 ||
1862 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1863 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
1864 I.getType() != I.getArgOperand(0)->getType() ||
1865 I.getType() != I.getArgOperand(1)->getType() ||
1866 !I.onlyReadsMemory())
1867 return Intrinsic::not_intrinsic;
1869 return ValidIntrinsicID;
1873 static Intrinsic::ID
1874 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1875 // If we have an intrinsic call, check if it is trivially vectorizable.
1876 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1877 switch (II->getIntrinsicID()) {
1878 case Intrinsic::sqrt:
1879 case Intrinsic::sin:
1880 case Intrinsic::cos:
1881 case Intrinsic::exp:
1882 case Intrinsic::exp2:
1883 case Intrinsic::log:
1884 case Intrinsic::log10:
1885 case Intrinsic::log2:
1886 case Intrinsic::fabs:
1887 case Intrinsic::copysign:
1888 case Intrinsic::floor:
1889 case Intrinsic::ceil:
1890 case Intrinsic::trunc:
1891 case Intrinsic::rint:
1892 case Intrinsic::nearbyint:
1893 case Intrinsic::round:
1894 case Intrinsic::pow:
1895 case Intrinsic::fma:
1896 case Intrinsic::fmuladd:
1897 case Intrinsic::lifetime_start:
1898 case Intrinsic::lifetime_end:
1899 return II->getIntrinsicID();
1901 return Intrinsic::not_intrinsic;
1906 return Intrinsic::not_intrinsic;
1909 Function *F = CI->getCalledFunction();
1910 // We're going to make assumptions on the semantics of the functions, check
1911 // that the target knows that it's available in this environment and it does
1912 // not have local linkage.
1913 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
1914 return Intrinsic::not_intrinsic;
1916 // Otherwise check if we have a call to a function that can be turned into a
1917 // vector intrinsic.
1924 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
1928 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
1932 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
1934 case LibFunc::exp2f:
1935 case LibFunc::exp2l:
1936 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
1940 return checkUnaryFloatSignature(*CI, Intrinsic::log);
1941 case LibFunc::log10:
1942 case LibFunc::log10f:
1943 case LibFunc::log10l:
1944 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
1946 case LibFunc::log2f:
1947 case LibFunc::log2l:
1948 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
1950 case LibFunc::fabsf:
1951 case LibFunc::fabsl:
1952 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
1953 case LibFunc::copysign:
1954 case LibFunc::copysignf:
1955 case LibFunc::copysignl:
1956 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
1957 case LibFunc::floor:
1958 case LibFunc::floorf:
1959 case LibFunc::floorl:
1960 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
1962 case LibFunc::ceilf:
1963 case LibFunc::ceill:
1964 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
1965 case LibFunc::trunc:
1966 case LibFunc::truncf:
1967 case LibFunc::truncl:
1968 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
1970 case LibFunc::rintf:
1971 case LibFunc::rintl:
1972 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
1973 case LibFunc::nearbyint:
1974 case LibFunc::nearbyintf:
1975 case LibFunc::nearbyintl:
1976 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
1977 case LibFunc::round:
1978 case LibFunc::roundf:
1979 case LibFunc::roundl:
1980 return checkUnaryFloatSignature(*CI, Intrinsic::round);
1984 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
1987 return Intrinsic::not_intrinsic;
1990 /// This function translates the reduction kind to an LLVM binary operator.
1992 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1994 case LoopVectorizationLegality::RK_IntegerAdd:
1995 return Instruction::Add;
1996 case LoopVectorizationLegality::RK_IntegerMult:
1997 return Instruction::Mul;
1998 case LoopVectorizationLegality::RK_IntegerOr:
1999 return Instruction::Or;
2000 case LoopVectorizationLegality::RK_IntegerAnd:
2001 return Instruction::And;
2002 case LoopVectorizationLegality::RK_IntegerXor:
2003 return Instruction::Xor;
2004 case LoopVectorizationLegality::RK_FloatMult:
2005 return Instruction::FMul;
2006 case LoopVectorizationLegality::RK_FloatAdd:
2007 return Instruction::FAdd;
2008 case LoopVectorizationLegality::RK_IntegerMinMax:
2009 return Instruction::ICmp;
2010 case LoopVectorizationLegality::RK_FloatMinMax:
2011 return Instruction::FCmp;
2013 llvm_unreachable("Unknown reduction operation");
2017 Value *createMinMaxOp(IRBuilder<> &Builder,
2018 LoopVectorizationLegality::MinMaxReductionKind RK,
2021 CmpInst::Predicate P = CmpInst::ICMP_NE;
2024 llvm_unreachable("Unknown min/max reduction kind");
2025 case LoopVectorizationLegality::MRK_UIntMin:
2026 P = CmpInst::ICMP_ULT;
2028 case LoopVectorizationLegality::MRK_UIntMax:
2029 P = CmpInst::ICMP_UGT;
2031 case LoopVectorizationLegality::MRK_SIntMin:
2032 P = CmpInst::ICMP_SLT;
2034 case LoopVectorizationLegality::MRK_SIntMax:
2035 P = CmpInst::ICMP_SGT;
2037 case LoopVectorizationLegality::MRK_FloatMin:
2038 P = CmpInst::FCMP_OLT;
2040 case LoopVectorizationLegality::MRK_FloatMax:
2041 P = CmpInst::FCMP_OGT;
2046 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2047 RK == LoopVectorizationLegality::MRK_FloatMax)
2048 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2050 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2052 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2057 struct CSEDenseMapInfo {
2058 static bool canHandle(Instruction *I) {
2059 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2060 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2062 static inline Instruction *getEmptyKey() {
2063 return DenseMapInfo<Instruction *>::getEmptyKey();
2065 static inline Instruction *getTombstoneKey() {
2066 return DenseMapInfo<Instruction *>::getTombstoneKey();
2068 static unsigned getHashValue(Instruction *I) {
2069 assert(canHandle(I) && "Unknown instruction!");
2070 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2071 I->value_op_end()));
2073 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2074 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2075 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2077 return LHS->isIdenticalTo(RHS);
2082 ///\brief Perform cse of induction variable instructions.
2083 static void cse(BasicBlock *BB) {
2084 // Perform simple cse.
2085 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2086 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2087 Instruction *In = I++;
2089 if (!CSEDenseMapInfo::canHandle(In))
2092 // Check if we can replace this instruction with any of the
2093 // visited instructions.
2094 if (Instruction *V = CSEMap.lookup(In)) {
2095 In->replaceAllUsesWith(V);
2096 In->eraseFromParent();
2105 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2106 //===------------------------------------------------===//
2108 // Notice: any optimization or new instruction that go
2109 // into the code below should be also be implemented in
2112 //===------------------------------------------------===//
2113 Constant *Zero = Builder.getInt32(0);
2115 // In order to support reduction variables we need to be able to vectorize
2116 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2117 // stages. First, we create a new vector PHI node with no incoming edges.
2118 // We use this value when we vectorize all of the instructions that use the
2119 // PHI. Next, after all of the instructions in the block are complete we
2120 // add the new incoming edges to the PHI. At this point all of the
2121 // instructions in the basic block are vectorized, so we can use them to
2122 // construct the PHI.
2123 PhiVector RdxPHIsToFix;
2125 // Scan the loop in a topological order to ensure that defs are vectorized
2127 LoopBlocksDFS DFS(OrigLoop);
2130 // Vectorize all of the blocks in the original loop.
2131 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2132 be = DFS.endRPO(); bb != be; ++bb)
2133 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2135 // At this point every instruction in the original loop is widened to
2136 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2137 // that we vectorized. The PHI nodes are currently empty because we did
2138 // not want to introduce cycles. Notice that the remaining PHI nodes
2139 // that we need to fix are reduction variables.
2141 // Create the 'reduced' values for each of the induction vars.
2142 // The reduced values are the vector values that we scalarize and combine
2143 // after the loop is finished.
2144 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2146 PHINode *RdxPhi = *it;
2147 assert(RdxPhi && "Unable to recover vectorized PHI");
2149 // Find the reduction variable descriptor.
2150 assert(Legal->getReductionVars()->count(RdxPhi) &&
2151 "Unable to find the reduction variable");
2152 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2153 (*Legal->getReductionVars())[RdxPhi];
2155 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2157 // We need to generate a reduction vector from the incoming scalar.
2158 // To do so, we need to generate the 'identity' vector and overide
2159 // one of the elements with the incoming scalar reduction. We need
2160 // to do it in the vector-loop preheader.
2161 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2163 // This is the vector-clone of the value that leaves the loop.
2164 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2165 Type *VecTy = VectorExit[0]->getType();
2167 // Find the reduction identity variable. Zero for addition, or, xor,
2168 // one for multiplication, -1 for And.
2171 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2172 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2173 // MinMax reduction have the start value as their identify.
2175 VectorStart = Identity = RdxDesc.StartValue;
2177 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2182 // Handle other reduction kinds:
2184 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2185 VecTy->getScalarType());
2188 // This vector is the Identity vector where the first element is the
2189 // incoming scalar reduction.
2190 VectorStart = RdxDesc.StartValue;
2192 Identity = ConstantVector::getSplat(VF, Iden);
2194 // This vector is the Identity vector where the first element is the
2195 // incoming scalar reduction.
2196 VectorStart = Builder.CreateInsertElement(Identity,
2197 RdxDesc.StartValue, Zero);
2201 // Fix the vector-loop phi.
2202 // We created the induction variable so we know that the
2203 // preheader is the first entry.
2204 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2206 // Reductions do not have to start at zero. They can start with
2207 // any loop invariant values.
2208 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2209 BasicBlock *Latch = OrigLoop->getLoopLatch();
2210 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2211 VectorParts &Val = getVectorValue(LoopVal);
2212 for (unsigned part = 0; part < UF; ++part) {
2213 // Make sure to add the reduction stat value only to the
2214 // first unroll part.
2215 Value *StartVal = (part == 0) ? VectorStart : Identity;
2216 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2217 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2220 // Before each round, move the insertion point right between
2221 // the PHIs and the values we are going to write.
2222 // This allows us to write both PHINodes and the extractelement
2224 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2226 VectorParts RdxParts;
2227 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2228 for (unsigned part = 0; part < UF; ++part) {
2229 // This PHINode contains the vectorized reduction variable, or
2230 // the initial value vector, if we bypass the vector loop.
2231 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2232 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2233 Value *StartVal = (part == 0) ? VectorStart : Identity;
2234 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2235 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2236 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2237 RdxParts.push_back(NewPhi);
2240 // Reduce all of the unrolled parts into a single vector.
2241 Value *ReducedPartRdx = RdxParts[0];
2242 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2243 setDebugLocFromInst(Builder, ReducedPartRdx);
2244 for (unsigned part = 1; part < UF; ++part) {
2245 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2246 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2247 RdxParts[part], ReducedPartRdx,
2250 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2251 ReducedPartRdx, RdxParts[part]);
2255 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2256 // and vector ops, reducing the set of values being computed by half each
2258 assert(isPowerOf2_32(VF) &&
2259 "Reduction emission only supported for pow2 vectors!");
2260 Value *TmpVec = ReducedPartRdx;
2261 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2262 for (unsigned i = VF; i != 1; i >>= 1) {
2263 // Move the upper half of the vector to the lower half.
2264 for (unsigned j = 0; j != i/2; ++j)
2265 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2267 // Fill the rest of the mask with undef.
2268 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2269 UndefValue::get(Builder.getInt32Ty()));
2272 Builder.CreateShuffleVector(TmpVec,
2273 UndefValue::get(TmpVec->getType()),
2274 ConstantVector::get(ShuffleMask),
2277 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2278 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2281 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2284 // The result is in the first element of the vector.
2285 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2286 Builder.getInt32(0));
2289 // Now, we need to fix the users of the reduction variable
2290 // inside and outside of the scalar remainder loop.
2291 // We know that the loop is in LCSSA form. We need to update the
2292 // PHI nodes in the exit blocks.
2293 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2294 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2295 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2296 if (!LCSSAPhi) break;
2298 // All PHINodes need to have a single entry edge, or two if
2299 // we already fixed them.
2300 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2302 // We found our reduction value exit-PHI. Update it with the
2303 // incoming bypass edge.
2304 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2305 // Add an edge coming from the bypass.
2306 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2309 }// end of the LCSSA phi scan.
2311 // Fix the scalar loop reduction variable with the incoming reduction sum
2312 // from the vector body and from the backedge value.
2313 int IncomingEdgeBlockIdx =
2314 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2315 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2316 // Pick the other block.
2317 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2318 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2319 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2320 }// end of for each redux variable.
2324 // Remove redundant induction instructions.
2325 cse(LoopVectorBody);
2328 void InnerLoopVectorizer::fixLCSSAPHIs() {
2329 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2330 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2331 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2332 if (!LCSSAPhi) break;
2333 if (LCSSAPhi->getNumIncomingValues() == 1)
2334 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2339 InnerLoopVectorizer::VectorParts
2340 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2341 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2344 // Look for cached value.
2345 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2346 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2347 if (ECEntryIt != MaskCache.end())
2348 return ECEntryIt->second;
2350 VectorParts SrcMask = createBlockInMask(Src);
2352 // The terminator has to be a branch inst!
2353 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2354 assert(BI && "Unexpected terminator found");
2356 if (BI->isConditional()) {
2357 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2359 if (BI->getSuccessor(0) != Dst)
2360 for (unsigned part = 0; part < UF; ++part)
2361 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2363 for (unsigned part = 0; part < UF; ++part)
2364 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2366 MaskCache[Edge] = EdgeMask;
2370 MaskCache[Edge] = SrcMask;
2374 InnerLoopVectorizer::VectorParts
2375 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2376 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2378 // Loop incoming mask is all-one.
2379 if (OrigLoop->getHeader() == BB) {
2380 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2381 return getVectorValue(C);
2384 // This is the block mask. We OR all incoming edges, and with zero.
2385 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2386 VectorParts BlockMask = getVectorValue(Zero);
2389 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2390 VectorParts EM = createEdgeMask(*it, BB);
2391 for (unsigned part = 0; part < UF; ++part)
2392 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2398 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2399 InnerLoopVectorizer::VectorParts &Entry,
2400 LoopVectorizationLegality *Legal,
2401 unsigned UF, unsigned VF, PhiVector *PV) {
2402 PHINode* P = cast<PHINode>(PN);
2403 // Handle reduction variables:
2404 if (Legal->getReductionVars()->count(P)) {
2405 for (unsigned part = 0; part < UF; ++part) {
2406 // This is phase one of vectorizing PHIs.
2407 Type *VecTy = (VF == 1) ? PN->getType() :
2408 VectorType::get(PN->getType(), VF);
2409 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2410 LoopVectorBody-> getFirstInsertionPt());
2416 setDebugLocFromInst(Builder, P);
2417 // Check for PHI nodes that are lowered to vector selects.
2418 if (P->getParent() != OrigLoop->getHeader()) {
2419 // We know that all PHIs in non header blocks are converted into
2420 // selects, so we don't have to worry about the insertion order and we
2421 // can just use the builder.
2422 // At this point we generate the predication tree. There may be
2423 // duplications since this is a simple recursive scan, but future
2424 // optimizations will clean it up.
2426 unsigned NumIncoming = P->getNumIncomingValues();
2428 // Generate a sequence of selects of the form:
2429 // SELECT(Mask3, In3,
2430 // SELECT(Mask2, In2,
2432 for (unsigned In = 0; In < NumIncoming; In++) {
2433 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2435 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2437 for (unsigned part = 0; part < UF; ++part) {
2438 // We might have single edge PHIs (blocks) - use an identity
2439 // 'select' for the first PHI operand.
2441 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2444 // Select between the current value and the previous incoming edge
2445 // based on the incoming mask.
2446 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2447 Entry[part], "predphi");
2453 // This PHINode must be an induction variable.
2454 // Make sure that we know about it.
2455 assert(Legal->getInductionVars()->count(P) &&
2456 "Not an induction variable");
2458 LoopVectorizationLegality::InductionInfo II =
2459 Legal->getInductionVars()->lookup(P);
2462 case LoopVectorizationLegality::IK_NoInduction:
2463 llvm_unreachable("Unknown induction");
2464 case LoopVectorizationLegality::IK_IntInduction: {
2465 assert(P->getType() == II.StartValue->getType() && "Types must match");
2466 Type *PhiTy = P->getType();
2468 if (P == OldInduction) {
2469 // Handle the canonical induction variable. We might have had to
2471 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2473 // Handle other induction variables that are now based on the
2475 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2477 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2478 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2481 Broadcasted = getBroadcastInstrs(Broadcasted);
2482 // After broadcasting the induction variable we need to make the vector
2483 // consecutive by adding 0, 1, 2, etc.
2484 for (unsigned part = 0; part < UF; ++part)
2485 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2488 case LoopVectorizationLegality::IK_ReverseIntInduction:
2489 case LoopVectorizationLegality::IK_PtrInduction:
2490 case LoopVectorizationLegality::IK_ReversePtrInduction:
2491 // Handle reverse integer and pointer inductions.
2492 Value *StartIdx = ExtendedIdx;
2493 // This is the normalized GEP that starts counting at zero.
2494 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2497 // Handle the reverse integer induction variable case.
2498 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2499 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2500 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2502 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2505 // This is a new value so do not hoist it out.
2506 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2507 // After broadcasting the induction variable we need to make the
2508 // vector consecutive by adding ... -3, -2, -1, 0.
2509 for (unsigned part = 0; part < UF; ++part)
2510 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2515 // Handle the pointer induction variable case.
2516 assert(P->getType()->isPointerTy() && "Unexpected type.");
2518 // Is this a reverse induction ptr or a consecutive induction ptr.
2519 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2522 // This is the vector of results. Notice that we don't generate
2523 // vector geps because scalar geps result in better code.
2524 for (unsigned part = 0; part < UF; ++part) {
2526 int EltIndex = (part) * (Reverse ? -1 : 1);
2527 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2530 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2532 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2534 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2536 Entry[part] = SclrGep;
2540 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2541 for (unsigned int i = 0; i < VF; ++i) {
2542 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2543 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2546 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2548 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2550 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2552 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2553 Builder.getInt32(i),
2556 Entry[part] = VecVal;
2563 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2564 BasicBlock *BB, PhiVector *PV) {
2565 // For each instruction in the old loop.
2566 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2567 VectorParts &Entry = WidenMap.get(it);
2568 switch (it->getOpcode()) {
2569 case Instruction::Br:
2570 // Nothing to do for PHIs and BR, since we already took care of the
2571 // loop control flow instructions.
2573 case Instruction::PHI:{
2574 // Vectorize PHINodes.
2575 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2579 case Instruction::Add:
2580 case Instruction::FAdd:
2581 case Instruction::Sub:
2582 case Instruction::FSub:
2583 case Instruction::Mul:
2584 case Instruction::FMul:
2585 case Instruction::UDiv:
2586 case Instruction::SDiv:
2587 case Instruction::FDiv:
2588 case Instruction::URem:
2589 case Instruction::SRem:
2590 case Instruction::FRem:
2591 case Instruction::Shl:
2592 case Instruction::LShr:
2593 case Instruction::AShr:
2594 case Instruction::And:
2595 case Instruction::Or:
2596 case Instruction::Xor: {
2597 // Just widen binops.
2598 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2599 setDebugLocFromInst(Builder, BinOp);
2600 VectorParts &A = getVectorValue(it->getOperand(0));
2601 VectorParts &B = getVectorValue(it->getOperand(1));
2603 // Use this vector value for all users of the original instruction.
2604 for (unsigned Part = 0; Part < UF; ++Part) {
2605 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2607 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2608 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2609 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2610 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2611 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2613 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2614 VecOp->setIsExact(BinOp->isExact());
2620 case Instruction::Select: {
2622 // If the selector is loop invariant we can create a select
2623 // instruction with a scalar condition. Otherwise, use vector-select.
2624 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2626 setDebugLocFromInst(Builder, it);
2628 // The condition can be loop invariant but still defined inside the
2629 // loop. This means that we can't just use the original 'cond' value.
2630 // We have to take the 'vectorized' value and pick the first lane.
2631 // Instcombine will make this a no-op.
2632 VectorParts &Cond = getVectorValue(it->getOperand(0));
2633 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2634 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2636 Value *ScalarCond = (VF == 1) ? Cond[0] :
2637 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2639 for (unsigned Part = 0; Part < UF; ++Part) {
2640 Entry[Part] = Builder.CreateSelect(
2641 InvariantCond ? ScalarCond : Cond[Part],
2648 case Instruction::ICmp:
2649 case Instruction::FCmp: {
2650 // Widen compares. Generate vector compares.
2651 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2652 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2653 setDebugLocFromInst(Builder, it);
2654 VectorParts &A = getVectorValue(it->getOperand(0));
2655 VectorParts &B = getVectorValue(it->getOperand(1));
2656 for (unsigned Part = 0; Part < UF; ++Part) {
2659 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2661 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2667 case Instruction::Store:
2668 case Instruction::Load:
2669 vectorizeMemoryInstruction(it, Legal);
2671 case Instruction::ZExt:
2672 case Instruction::SExt:
2673 case Instruction::FPToUI:
2674 case Instruction::FPToSI:
2675 case Instruction::FPExt:
2676 case Instruction::PtrToInt:
2677 case Instruction::IntToPtr:
2678 case Instruction::SIToFP:
2679 case Instruction::UIToFP:
2680 case Instruction::Trunc:
2681 case Instruction::FPTrunc:
2682 case Instruction::BitCast: {
2683 CastInst *CI = dyn_cast<CastInst>(it);
2684 setDebugLocFromInst(Builder, it);
2685 /// Optimize the special case where the source is the induction
2686 /// variable. Notice that we can only optimize the 'trunc' case
2687 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2688 /// c. other casts depend on pointer size.
2689 if (CI->getOperand(0) == OldInduction &&
2690 it->getOpcode() == Instruction::Trunc) {
2691 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2693 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2694 for (unsigned Part = 0; Part < UF; ++Part)
2695 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2698 /// Vectorize casts.
2699 Type *DestTy = (VF == 1) ? CI->getType() :
2700 VectorType::get(CI->getType(), VF);
2702 VectorParts &A = getVectorValue(it->getOperand(0));
2703 for (unsigned Part = 0; Part < UF; ++Part)
2704 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2708 case Instruction::Call: {
2709 // Ignore dbg intrinsics.
2710 if (isa<DbgInfoIntrinsic>(it))
2712 setDebugLocFromInst(Builder, it);
2714 Module *M = BB->getParent()->getParent();
2715 CallInst *CI = cast<CallInst>(it);
2716 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2717 assert(ID && "Not an intrinsic call!");
2719 case Intrinsic::lifetime_end:
2720 case Intrinsic::lifetime_start:
2721 scalarizeInstruction(it);
2724 for (unsigned Part = 0; Part < UF; ++Part) {
2725 SmallVector<Value *, 4> Args;
2726 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2727 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2728 Args.push_back(Arg[Part]);
2730 Type *Tys[] = {CI->getType()};
2732 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2734 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2735 Entry[Part] = Builder.CreateCall(F, Args);
2743 // All other instructions are unsupported. Scalarize them.
2744 scalarizeInstruction(it);
2747 }// end of for_each instr.
2750 void InnerLoopVectorizer::updateAnalysis() {
2751 // Forget the original basic block.
2752 SE->forgetLoop(OrigLoop);
2754 // Update the dominator tree information.
2755 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2756 "Entry does not dominate exit.");
2758 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2759 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2760 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2761 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2762 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2763 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2764 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2765 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2767 DEBUG(DT->verifyAnalysis());
2770 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2771 if (!EnableIfConversion)
2774 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2776 // A list of pointers that we can safely read and write to.
2777 SmallPtrSet<Value *, 8> SafePointes;
2779 // Collect safe addresses.
2780 for (Loop::block_iterator BI = TheLoop->block_begin(),
2781 BE = TheLoop->block_end(); BI != BE; ++BI) {
2782 BasicBlock *BB = *BI;
2784 if (blockNeedsPredication(BB))
2787 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2788 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2789 SafePointes.insert(LI->getPointerOperand());
2790 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2791 SafePointes.insert(SI->getPointerOperand());
2795 // Collect the blocks that need predication.
2796 for (Loop::block_iterator BI = TheLoop->block_begin(),
2797 BE = TheLoop->block_end(); BI != BE; ++BI) {
2798 BasicBlock *BB = *BI;
2800 // We don't support switch statements inside loops.
2801 if (!isa<BranchInst>(BB->getTerminator()))
2804 // We must be able to predicate all blocks that need to be predicated.
2805 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2809 // We can if-convert this loop.
2813 bool LoopVectorizationLegality::canVectorize() {
2814 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2815 // be canonicalized.
2816 if (!TheLoop->getLoopPreheader())
2819 // We can only vectorize innermost loops.
2820 if (TheLoop->getSubLoopsVector().size())
2823 // We must have a single backedge.
2824 if (TheLoop->getNumBackEdges() != 1)
2827 // We must have a single exiting block.
2828 if (!TheLoop->getExitingBlock())
2831 // We need to have a loop header.
2832 DEBUG(dbgs() << "LV: Found a loop: " <<
2833 TheLoop->getHeader()->getName() << '\n');
2835 // Check if we can if-convert non single-bb loops.
2836 unsigned NumBlocks = TheLoop->getNumBlocks();
2837 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2838 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2842 // ScalarEvolution needs to be able to find the exit count.
2843 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2844 if (ExitCount == SE->getCouldNotCompute()) {
2845 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2849 // Do not loop-vectorize loops with a tiny trip count.
2850 BasicBlock *Latch = TheLoop->getLoopLatch();
2851 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2852 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2853 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2854 "This loop is not worth vectorizing.\n");
2858 // Check if we can vectorize the instructions and CFG in this loop.
2859 if (!canVectorizeInstrs()) {
2860 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2864 // Go over each instruction and look at memory deps.
2865 if (!canVectorizeMemory()) {
2866 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2870 // Collect all of the variables that remain uniform after vectorization.
2871 collectLoopUniforms();
2873 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2874 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2877 // Okay! We can vectorize. At this point we don't have any other mem analysis
2878 // which may limit our maximum vectorization factor, so just return true with
2883 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2884 if (Ty->isPointerTy())
2885 return DL.getIntPtrType(Ty);
2890 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2891 Ty0 = convertPointerToIntegerType(DL, Ty0);
2892 Ty1 = convertPointerToIntegerType(DL, Ty1);
2893 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2898 /// \brief Check that the instruction has outside loop users and is not an
2899 /// identified reduction variable.
2900 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2901 SmallPtrSet<Value *, 4> &Reductions) {
2902 // Reduction instructions are allowed to have exit users. All other
2903 // instructions must not have external users.
2904 if (!Reductions.count(Inst))
2905 //Check that all of the users of the loop are inside the BB.
2906 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2908 Instruction *U = cast<Instruction>(*I);
2909 // This user may be a reduction exit value.
2910 if (!TheLoop->contains(U)) {
2911 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
2918 bool LoopVectorizationLegality::canVectorizeInstrs() {
2919 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2920 BasicBlock *Header = TheLoop->getHeader();
2922 // Look for the attribute signaling the absence of NaNs.
2923 Function &F = *Header->getParent();
2924 if (F.hasFnAttribute("no-nans-fp-math"))
2925 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2926 AttributeSet::FunctionIndex,
2927 "no-nans-fp-math").getValueAsString() == "true";
2929 // For each block in the loop.
2930 for (Loop::block_iterator bb = TheLoop->block_begin(),
2931 be = TheLoop->block_end(); bb != be; ++bb) {
2933 // Scan the instructions in the block and look for hazards.
2934 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2937 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2938 Type *PhiTy = Phi->getType();
2939 // Check that this PHI type is allowed.
2940 if (!PhiTy->isIntegerTy() &&
2941 !PhiTy->isFloatingPointTy() &&
2942 !PhiTy->isPointerTy()) {
2943 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2947 // If this PHINode is not in the header block, then we know that we
2948 // can convert it to select during if-conversion. No need to check if
2949 // the PHIs in this block are induction or reduction variables.
2950 if (*bb != Header) {
2951 // Check that this instruction has no outside users or is an
2952 // identified reduction value with an outside user.
2953 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2958 // We only allow if-converted PHIs with more than two incoming values.
2959 if (Phi->getNumIncomingValues() != 2) {
2960 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2964 // This is the value coming from the preheader.
2965 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2966 // Check if this is an induction variable.
2967 InductionKind IK = isInductionVariable(Phi);
2969 if (IK_NoInduction != IK) {
2970 // Get the widest type.
2972 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2974 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2976 // Int inductions are special because we only allow one IV.
2977 if (IK == IK_IntInduction) {
2978 // Use the phi node with the widest type as induction. Use the last
2979 // one if there are multiple (no good reason for doing this other
2980 // than it is expedient).
2981 if (!Induction || PhiTy == WidestIndTy)
2985 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2986 Inductions[Phi] = InductionInfo(StartValue, IK);
2988 // Until we explicitly handle the case of an induction variable with
2989 // an outside loop user we have to give up vectorizing this loop.
2990 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2996 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2997 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3000 if (AddReductionVar(Phi, RK_IntegerMult)) {
3001 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3004 if (AddReductionVar(Phi, RK_IntegerOr)) {
3005 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3008 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3009 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3012 if (AddReductionVar(Phi, RK_IntegerXor)) {
3013 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3016 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3017 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3020 if (AddReductionVar(Phi, RK_FloatMult)) {
3021 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3024 if (AddReductionVar(Phi, RK_FloatAdd)) {
3025 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3028 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3029 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3034 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3036 }// end of PHI handling
3038 // We still don't handle functions. However, we can ignore dbg intrinsic
3039 // calls and we do handle certain intrinsic and libm functions.
3040 CallInst *CI = dyn_cast<CallInst>(it);
3041 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3042 DEBUG(dbgs() << "LV: Found a call site.\n");
3046 // Check that the instruction return type is vectorizable.
3047 // Also, we can't vectorize extractelement instructions.
3048 if ((!VectorType::isValidElementType(it->getType()) &&
3049 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3050 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3054 // Check that the stored type is vectorizable.
3055 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3056 Type *T = ST->getValueOperand()->getType();
3057 if (!VectorType::isValidElementType(T))
3061 // Reduction instructions are allowed to have exit users.
3062 // All other instructions must not have external users.
3063 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3071 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3072 if (Inductions.empty())
3079 void LoopVectorizationLegality::collectLoopUniforms() {
3080 // We now know that the loop is vectorizable!
3081 // Collect variables that will remain uniform after vectorization.
3082 std::vector<Value*> Worklist;
3083 BasicBlock *Latch = TheLoop->getLoopLatch();
3085 // Start with the conditional branch and walk up the block.
3086 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3088 while (Worklist.size()) {
3089 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3090 Worklist.pop_back();
3092 // Look at instructions inside this loop.
3093 // Stop when reaching PHI nodes.
3094 // TODO: we need to follow values all over the loop, not only in this block.
3095 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3098 // This is a known uniform.
3101 // Insert all operands.
3102 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3107 /// \brief Analyses memory accesses in a loop.
3109 /// Checks whether run time pointer checks are needed and builds sets for data
3110 /// dependence checking.
3111 class AccessAnalysis {
3113 /// \brief Read or write access location.
3114 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3115 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3117 /// \brief Set of potential dependent memory accesses.
3118 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3120 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3121 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3122 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3124 /// \brief Register a load and whether it is only read from.
3125 void addLoad(Value *Ptr, bool IsReadOnly) {
3126 Accesses.insert(MemAccessInfo(Ptr, false));
3128 ReadOnlyPtr.insert(Ptr);
3131 /// \brief Register a store.
3132 void addStore(Value *Ptr) {
3133 Accesses.insert(MemAccessInfo(Ptr, true));
3136 /// \brief Check whether we can check the pointers at runtime for
3137 /// non-intersection.
3138 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3139 unsigned &NumComparisons, ScalarEvolution *SE,
3140 Loop *TheLoop, bool ShouldCheckStride = false);
3142 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3143 /// and builds sets of dependent accesses.
3144 void buildDependenceSets() {
3145 // Process read-write pointers first.
3146 processMemAccesses(false);
3147 // Next, process read pointers.
3148 processMemAccesses(true);
3151 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3153 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3154 void resetDepChecks() { CheckDeps.clear(); }
3156 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3159 typedef SetVector<MemAccessInfo> PtrAccessSet;
3160 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3162 /// \brief Go over all memory access or only the deferred ones if
3163 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3164 /// and build sets of dependency check candidates.
3165 void processMemAccesses(bool UseDeferred);
3167 /// Set of all accesses.
3168 PtrAccessSet Accesses;
3170 /// Set of access to check after all writes have been processed.
3171 PtrAccessSet DeferredAccesses;
3173 /// Map of pointers to last access encountered.
3174 UnderlyingObjToAccessMap ObjToLastAccess;
3176 /// Set of accesses that need a further dependence check.
3177 MemAccessInfoSet CheckDeps;
3179 /// Set of pointers that are read only.
3180 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3182 /// Set of underlying objects already written to.
3183 SmallPtrSet<Value*, 16> WriteObjects;
3187 /// Sets of potentially dependent accesses - members of one set share an
3188 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3189 /// dependence check.
3190 DepCandidates &DepCands;
3192 bool AreAllWritesIdentified;
3193 bool AreAllReadsIdentified;
3194 bool IsRTCheckNeeded;
3197 } // end anonymous namespace
3199 /// \brief Check whether a pointer can participate in a runtime bounds check.
3200 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3201 const SCEV *PtrScev = SE->getSCEV(Ptr);
3202 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3206 return AR->isAffine();
3209 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3210 /// the address space.
3211 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3214 bool AccessAnalysis::canCheckPtrAtRT(
3215 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3216 unsigned &NumComparisons, ScalarEvolution *SE,
3217 Loop *TheLoop, bool ShouldCheckStride) {
3218 // Find pointers with computable bounds. We are going to use this information
3219 // to place a runtime bound check.
3220 unsigned NumReadPtrChecks = 0;
3221 unsigned NumWritePtrChecks = 0;
3222 bool CanDoRT = true;
3224 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3225 // We assign consecutive id to access from different dependence sets.
3226 // Accesses within the same set don't need a runtime check.
3227 unsigned RunningDepId = 1;
3228 DenseMap<Value *, unsigned> DepSetId;
3230 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3232 const MemAccessInfo &Access = *AI;
3233 Value *Ptr = Access.getPointer();
3234 bool IsWrite = Access.getInt();
3236 // Just add write checks if we have both.
3237 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3241 ++NumWritePtrChecks;
3245 if (hasComputableBounds(SE, Ptr) &&
3246 // When we run after a failing dependency check we have to make sure we
3247 // don't have wrapping pointers.
3248 (!ShouldCheckStride || isStridedPtr(SE, DL, Ptr, TheLoop) == 1)) {
3249 // The id of the dependence set.
3252 if (IsDepCheckNeeded) {
3253 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3254 unsigned &LeaderId = DepSetId[Leader];
3256 LeaderId = RunningDepId++;
3259 // Each access has its own dependence set.
3260 DepId = RunningDepId++;
3262 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3264 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3270 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3271 NumComparisons = 0; // Only one dependence set.
3273 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3274 NumWritePtrChecks - 1));
3277 // If the pointers that we would use for the bounds comparison have different
3278 // address spaces, assume the values aren't directly comparable, so we can't
3279 // use them for the runtime check. We also have to assume they could
3280 // overlap. In the future there should be metadata for whether address spaces
3282 unsigned NumPointers = RtCheck.Pointers.size();
3283 for (unsigned i = 0; i < NumPointers; ++i) {
3284 for (unsigned j = i + 1; j < NumPointers; ++j) {
3285 // Only need to check pointers between two different dependency sets.
3286 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3289 Value *PtrI = RtCheck.Pointers[i];
3290 Value *PtrJ = RtCheck.Pointers[j];
3292 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3293 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3295 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3296 " different address spaces\n");
3305 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3306 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3309 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3310 // We process the set twice: first we process read-write pointers, last we
3311 // process read-only pointers. This allows us to skip dependence tests for
3312 // read-only pointers.
3314 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3315 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3316 const MemAccessInfo &Access = *AI;
3317 Value *Ptr = Access.getPointer();
3318 bool IsWrite = Access.getInt();
3320 DepCands.insert(Access);
3322 // Memorize read-only pointers for later processing and skip them in the
3323 // first round (they need to be checked after we have seen all write
3324 // pointers). Note: we also mark pointer that are not consecutive as
3325 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3326 // second check for "!IsWrite".
3327 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3328 if (!UseDeferred && IsReadOnlyPtr) {
3329 DeferredAccesses.insert(Access);
3333 bool NeedDepCheck = false;
3334 // Check whether there is the possiblity of dependency because of underlying
3335 // objects being the same.
3336 typedef SmallVector<Value*, 16> ValueVector;
3337 ValueVector TempObjects;
3338 GetUnderlyingObjects(Ptr, TempObjects, DL);
3339 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3341 Value *UnderlyingObj = *UI;
3343 // If this is a write then it needs to be an identified object. If this a
3344 // read and all writes (so far) are identified function scope objects we
3345 // don't need an identified underlying object but only an Argument (the
3346 // next write is going to invalidate this assumption if it is
3348 // This is a micro-optimization for the case where all writes are
3349 // identified and we have one argument pointer.
3350 // Otherwise, we do need a runtime check.
3351 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3352 (!IsWrite && (!AreAllWritesIdentified ||
3353 !isa<Argument>(UnderlyingObj)) &&
3354 !isIdentifiedObject(UnderlyingObj))) {
3355 DEBUG(dbgs() << "LV: Found an unidentified " <<
3356 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3358 IsRTCheckNeeded = (IsRTCheckNeeded ||
3359 !isIdentifiedObject(UnderlyingObj) ||
3360 !AreAllReadsIdentified);
3363 AreAllWritesIdentified = false;
3365 AreAllReadsIdentified = false;
3368 // If this is a write - check other reads and writes for conflicts. If
3369 // this is a read only check other writes for conflicts (but only if there
3370 // is no other write to the ptr - this is an optimization to catch "a[i] =
3371 // a[i] + " without having to do a dependence check).
3372 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3373 NeedDepCheck = true;
3376 WriteObjects.insert(UnderlyingObj);
3378 // Create sets of pointers connected by shared underlying objects.
3379 UnderlyingObjToAccessMap::iterator Prev =
3380 ObjToLastAccess.find(UnderlyingObj);
3381 if (Prev != ObjToLastAccess.end())
3382 DepCands.unionSets(Access, Prev->second);
3384 ObjToLastAccess[UnderlyingObj] = Access;
3388 CheckDeps.insert(Access);
3393 /// \brief Checks memory dependences among accesses to the same underlying
3394 /// object to determine whether there vectorization is legal or not (and at
3395 /// which vectorization factor).
3397 /// This class works under the assumption that we already checked that memory
3398 /// locations with different underlying pointers are "must-not alias".
3399 /// We use the ScalarEvolution framework to symbolically evalutate access
3400 /// functions pairs. Since we currently don't restructure the loop we can rely
3401 /// on the program order of memory accesses to determine their safety.
3402 /// At the moment we will only deem accesses as safe for:
3403 /// * A negative constant distance assuming program order.
3405 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3406 /// a[i] = tmp; y = a[i];
3408 /// The latter case is safe because later checks guarantuee that there can't
3409 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3410 /// the same variable: a header phi can only be an induction or a reduction, a
3411 /// reduction can't have a memory sink, an induction can't have a memory
3412 /// source). This is important and must not be violated (or we have to
3413 /// resort to checking for cycles through memory).
3415 /// * A positive constant distance assuming program order that is bigger
3416 /// than the biggest memory access.
3418 /// tmp = a[i] OR b[i] = x
3419 /// a[i+2] = tmp y = b[i+2];
3421 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3423 /// * Zero distances and all accesses have the same size.
3425 class MemoryDepChecker {
3427 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3428 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3430 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3431 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3432 ShouldRetryWithRuntimeCheck(false) {}
3434 /// \brief Register the location (instructions are given increasing numbers)
3435 /// of a write access.
3436 void addAccess(StoreInst *SI) {
3437 Value *Ptr = SI->getPointerOperand();
3438 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3439 InstMap.push_back(SI);
3443 /// \brief Register the location (instructions are given increasing numbers)
3444 /// of a write access.
3445 void addAccess(LoadInst *LI) {
3446 Value *Ptr = LI->getPointerOperand();
3447 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3448 InstMap.push_back(LI);
3452 /// \brief Check whether the dependencies between the accesses are safe.
3454 /// Only checks sets with elements in \p CheckDeps.
3455 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3456 MemAccessInfoSet &CheckDeps);
3458 /// \brief The maximum number of bytes of a vector register we can vectorize
3459 /// the accesses safely with.
3460 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3462 /// \brief In same cases when the dependency check fails we can still
3463 /// vectorize the loop with a dynamic array access check.
3464 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3467 ScalarEvolution *SE;
3469 const Loop *InnermostLoop;
3471 /// \brief Maps access locations (ptr, read/write) to program order.
3472 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3474 /// \brief Memory access instructions in program order.
3475 SmallVector<Instruction *, 16> InstMap;
3477 /// \brief The program order index to be used for the next instruction.
3480 // We can access this many bytes in parallel safely.
3481 unsigned MaxSafeDepDistBytes;
3483 /// \brief If we see a non constant dependence distance we can still try to
3484 /// vectorize this loop with runtime checks.
3485 bool ShouldRetryWithRuntimeCheck;
3487 /// \brief Check whether there is a plausible dependence between the two
3490 /// Access \p A must happen before \p B in program order. The two indices
3491 /// identify the index into the program order map.
3493 /// This function checks whether there is a plausible dependence (or the
3494 /// absence of such can't be proved) between the two accesses. If there is a
3495 /// plausible dependence but the dependence distance is bigger than one
3496 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3497 /// distance is smaller than any other distance encountered so far).
3498 /// Otherwise, this function returns true signaling a possible dependence.
3499 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3500 const MemAccessInfo &B, unsigned BIdx);
3502 /// \brief Check whether the data dependence could prevent store-load
3504 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3507 } // end anonymous namespace
3509 static bool isInBoundsGep(Value *Ptr) {
3510 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3511 return GEP->isInBounds();
3515 /// \brief Check whether the access through \p Ptr has a constant stride.
3516 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3518 const Type *Ty = Ptr->getType();
3519 assert(Ty->isPointerTy() && "Unexpected non ptr");
3521 // Make sure that the pointer does not point to aggregate types.
3522 const PointerType *PtrTy = cast<PointerType>(Ty);
3523 if (PtrTy->getElementType()->isAggregateType()) {
3524 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3529 const SCEV *PtrScev = SE->getSCEV(Ptr);
3530 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3532 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3533 << *Ptr << " SCEV: " << *PtrScev << "\n");
3537 // The accesss function must stride over the innermost loop.
3538 if (Lp != AR->getLoop()) {
3539 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3540 *Ptr << " SCEV: " << *PtrScev << "\n");
3543 // The address calculation must not wrap. Otherwise, a dependence could be
3545 // An inbounds getelementptr that is a AddRec with a unit stride
3546 // cannot wrap per definition. The unit stride requirement is checked later.
3547 // An getelementptr without an inbounds attribute and unit stride would have
3548 // to access the pointer value "0" which is undefined behavior in address
3549 // space 0, therefore we can also vectorize this case.
3550 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3551 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3552 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3553 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3554 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3555 << *Ptr << " SCEV: " << *PtrScev << "\n");
3559 // Check the step is constant.
3560 const SCEV *Step = AR->getStepRecurrence(*SE);
3562 // Calculate the pointer stride and check if it is consecutive.
3563 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3565 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3566 " SCEV: " << *PtrScev << "\n");
3570 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3571 const APInt &APStepVal = C->getValue()->getValue();
3573 // Huge step value - give up.
3574 if (APStepVal.getBitWidth() > 64)
3577 int64_t StepVal = APStepVal.getSExtValue();
3580 int64_t Stride = StepVal / Size;
3581 int64_t Rem = StepVal % Size;
3585 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3586 // know we can't "wrap around the address space". In case of address space
3587 // zero we know that this won't happen without triggering undefined behavior.
3588 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3589 Stride != 1 && Stride != -1)
3595 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3596 unsigned TypeByteSize) {
3597 // If loads occur at a distance that is not a multiple of a feasible vector
3598 // factor store-load forwarding does not take place.
3599 // Positive dependences might cause troubles because vectorizing them might
3600 // prevent store-load forwarding making vectorized code run a lot slower.
3601 // a[i] = a[i-3] ^ a[i-8];
3602 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3603 // hence on your typical architecture store-load forwarding does not take
3604 // place. Vectorizing in such cases does not make sense.
3605 // Store-load forwarding distance.
3606 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3607 // Maximum vector factor.
3608 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3609 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3610 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3612 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3614 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3615 MaxVFWithoutSLForwardIssues = (vf >>=1);
3620 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3621 DEBUG(dbgs() << "LV: Distance " << Distance <<
3622 " that could cause a store-load forwarding conflict\n");
3626 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3627 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3628 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3632 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3633 const MemAccessInfo &B, unsigned BIdx) {
3634 assert (AIdx < BIdx && "Must pass arguments in program order");
3636 Value *APtr = A.getPointer();
3637 Value *BPtr = B.getPointer();
3638 bool AIsWrite = A.getInt();
3639 bool BIsWrite = B.getInt();
3641 // Two reads are independent.
3642 if (!AIsWrite && !BIsWrite)
3645 const SCEV *AScev = SE->getSCEV(APtr);
3646 const SCEV *BScev = SE->getSCEV(BPtr);
3648 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3649 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3651 const SCEV *Src = AScev;
3652 const SCEV *Sink = BScev;
3654 // If the induction step is negative we have to invert source and sink of the
3656 if (StrideAPtr < 0) {
3659 std::swap(APtr, BPtr);
3660 std::swap(Src, Sink);
3661 std::swap(AIsWrite, BIsWrite);
3662 std::swap(AIdx, BIdx);
3663 std::swap(StrideAPtr, StrideBPtr);
3666 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3668 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3669 << "(Induction step: " << StrideAPtr << ")\n");
3670 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3671 << *InstMap[BIdx] << ": " << *Dist << "\n");
3673 // Need consecutive accesses. We don't want to vectorize
3674 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3675 // the address space.
3676 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3677 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3681 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3683 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3684 ShouldRetryWithRuntimeCheck = true;
3688 Type *ATy = APtr->getType()->getPointerElementType();
3689 Type *BTy = BPtr->getType()->getPointerElementType();
3690 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3692 // Negative distances are not plausible dependencies.
3693 const APInt &Val = C->getValue()->getValue();
3694 if (Val.isNegative()) {
3695 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3696 if (IsTrueDataDependence &&
3697 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3701 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3705 // Write to the same location with the same size.
3706 // Could be improved to assert type sizes are the same (i32 == float, etc).
3710 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
3714 assert(Val.isStrictlyPositive() && "Expect a positive value");
3716 // Positive distance bigger than max vectorization factor.
3719 "LV: ReadWrite-Write positive dependency with different types\n");
3723 unsigned Distance = (unsigned) Val.getZExtValue();
3725 // Bail out early if passed-in parameters make vectorization not feasible.
3726 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3727 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3729 // The distance must be bigger than the size needed for a vectorized version
3730 // of the operation and the size of the vectorized operation must not be
3731 // bigger than the currrent maximum size.
3732 if (Distance < 2*TypeByteSize ||
3733 2*TypeByteSize > MaxSafeDepDistBytes ||
3734 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3735 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3736 << Val.getSExtValue() << '\n');
3740 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3741 Distance : MaxSafeDepDistBytes;
3743 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3744 if (IsTrueDataDependence &&
3745 couldPreventStoreLoadForward(Distance, TypeByteSize))
3748 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3749 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
3755 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3756 MemAccessInfoSet &CheckDeps) {
3758 MaxSafeDepDistBytes = -1U;
3759 while (!CheckDeps.empty()) {
3760 MemAccessInfo CurAccess = *CheckDeps.begin();
3762 // Get the relevant memory access set.
3763 EquivalenceClasses<MemAccessInfo>::iterator I =
3764 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3766 // Check accesses within this set.
3767 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3768 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3770 // Check every access pair.
3772 CheckDeps.erase(*AI);
3773 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3775 // Check every accessing instruction pair in program order.
3776 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3777 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3778 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3779 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3780 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3782 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3793 bool LoopVectorizationLegality::canVectorizeMemory() {
3795 typedef SmallVector<Value*, 16> ValueVector;
3796 typedef SmallPtrSet<Value*, 16> ValueSet;
3798 // Holds the Load and Store *instructions*.
3802 // Holds all the different accesses in the loop.
3803 unsigned NumReads = 0;
3804 unsigned NumReadWrites = 0;
3806 PtrRtCheck.Pointers.clear();
3807 PtrRtCheck.Need = false;
3809 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3810 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3813 for (Loop::block_iterator bb = TheLoop->block_begin(),
3814 be = TheLoop->block_end(); bb != be; ++bb) {
3816 // Scan the BB and collect legal loads and stores.
3817 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3820 // If this is a load, save it. If this instruction can read from memory
3821 // but is not a load, then we quit. Notice that we don't handle function
3822 // calls that read or write.
3823 if (it->mayReadFromMemory()) {
3824 // Many math library functions read the rounding mode. We will only
3825 // vectorize a loop if it contains known function calls that don't set
3826 // the flag. Therefore, it is safe to ignore this read from memory.
3827 CallInst *Call = dyn_cast<CallInst>(it);
3828 if (Call && getIntrinsicIDForCall(Call, TLI))
3831 LoadInst *Ld = dyn_cast<LoadInst>(it);
3832 if (!Ld) return false;
3833 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3834 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3837 Loads.push_back(Ld);
3838 DepChecker.addAccess(Ld);
3842 // Save 'store' instructions. Abort if other instructions write to memory.
3843 if (it->mayWriteToMemory()) {
3844 StoreInst *St = dyn_cast<StoreInst>(it);
3845 if (!St) return false;
3846 if (!St->isSimple() && !IsAnnotatedParallel) {
3847 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3850 Stores.push_back(St);
3851 DepChecker.addAccess(St);
3856 // Now we have two lists that hold the loads and the stores.
3857 // Next, we find the pointers that they use.
3859 // Check if we see any stores. If there are no stores, then we don't
3860 // care if the pointers are *restrict*.
3861 if (!Stores.size()) {
3862 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3866 AccessAnalysis::DepCandidates DependentAccesses;
3867 AccessAnalysis Accesses(DL, DependentAccesses);
3869 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3870 // multiple times on the same object. If the ptr is accessed twice, once
3871 // for read and once for write, it will only appear once (on the write
3872 // list). This is okay, since we are going to check for conflicts between
3873 // writes and between reads and writes, but not between reads and reads.
3876 ValueVector::iterator I, IE;
3877 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3878 StoreInst *ST = cast<StoreInst>(*I);
3879 Value* Ptr = ST->getPointerOperand();
3881 if (isUniform(Ptr)) {
3882 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3886 // If we did *not* see this pointer before, insert it to the read-write
3887 // list. At this phase it is only a 'write' list.
3888 if (Seen.insert(Ptr)) {
3890 Accesses.addStore(Ptr);
3894 if (IsAnnotatedParallel) {
3896 << "LV: A loop annotated parallel, ignore memory dependency "
3901 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3902 LoadInst *LD = cast<LoadInst>(*I);
3903 Value* Ptr = LD->getPointerOperand();
3904 // If we did *not* see this pointer before, insert it to the
3905 // read list. If we *did* see it before, then it is already in
3906 // the read-write list. This allows us to vectorize expressions
3907 // such as A[i] += x; Because the address of A[i] is a read-write
3908 // pointer. This only works if the index of A[i] is consecutive.
3909 // If the address of i is unknown (for example A[B[i]]) then we may
3910 // read a few words, modify, and write a few words, and some of the
3911 // words may be written to the same address.
3912 bool IsReadOnlyPtr = false;
3913 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3915 IsReadOnlyPtr = true;
3917 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3920 // If we write (or read-write) to a single destination and there are no
3921 // other reads in this loop then is it safe to vectorize.
3922 if (NumReadWrites == 1 && NumReads == 0) {
3923 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3927 // Build dependence sets and check whether we need a runtime pointer bounds
3929 Accesses.buildDependenceSets();
3930 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3932 // Find pointers with computable bounds. We are going to use this information
3933 // to place a runtime bound check.
3934 unsigned NumComparisons = 0;
3935 bool CanDoRT = false;
3937 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3940 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3941 " pointer comparisons.\n");
3943 // If we only have one set of dependences to check pointers among we don't
3944 // need a runtime check.
3945 if (NumComparisons == 0 && NeedRTCheck)
3946 NeedRTCheck = false;
3948 // Check that we did not collect too many pointers or found an unsizeable
3950 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3956 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3959 if (NeedRTCheck && !CanDoRT) {
3960 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3961 "the array bounds.\n");
3966 PtrRtCheck.Need = NeedRTCheck;
3968 bool CanVecMem = true;
3969 if (Accesses.isDependencyCheckNeeded()) {
3970 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3971 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3972 Accesses.getDependenciesToCheck());
3973 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3975 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
3976 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
3979 // Clear the dependency checks. We assume they are not needed.
3980 Accesses.resetDepChecks();
3983 PtrRtCheck.Need = true;
3985 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
3987 // Check that we did not collect too many pointers or found an unsizeable
3989 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3990 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
3999 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4000 " need a runtime memory check.\n");
4005 static bool hasMultipleUsesOf(Instruction *I,
4006 SmallPtrSet<Instruction *, 8> &Insts) {
4007 unsigned NumUses = 0;
4008 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4009 if (Insts.count(dyn_cast<Instruction>(*Use)))
4018 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4019 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4020 if (!Set.count(dyn_cast<Instruction>(*Use)))
4025 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4026 ReductionKind Kind) {
4027 if (Phi->getNumIncomingValues() != 2)
4030 // Reduction variables are only found in the loop header block.
4031 if (Phi->getParent() != TheLoop->getHeader())
4034 // Obtain the reduction start value from the value that comes from the loop
4036 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4038 // ExitInstruction is the single value which is used outside the loop.
4039 // We only allow for a single reduction value to be used outside the loop.
4040 // This includes users of the reduction, variables (which form a cycle
4041 // which ends in the phi node).
4042 Instruction *ExitInstruction = 0;
4043 // Indicates that we found a reduction operation in our scan.
4044 bool FoundReduxOp = false;
4046 // We start with the PHI node and scan for all of the users of this
4047 // instruction. All users must be instructions that can be used as reduction
4048 // variables (such as ADD). We must have a single out-of-block user. The cycle
4049 // must include the original PHI.
4050 bool FoundStartPHI = false;
4052 // To recognize min/max patterns formed by a icmp select sequence, we store
4053 // the number of instruction we saw from the recognized min/max pattern,
4054 // to make sure we only see exactly the two instructions.
4055 unsigned NumCmpSelectPatternInst = 0;
4056 ReductionInstDesc ReduxDesc(false, 0);
4058 SmallPtrSet<Instruction *, 8> VisitedInsts;
4059 SmallVector<Instruction *, 8> Worklist;
4060 Worklist.push_back(Phi);
4061 VisitedInsts.insert(Phi);
4063 // A value in the reduction can be used:
4064 // - By the reduction:
4065 // - Reduction operation:
4066 // - One use of reduction value (safe).
4067 // - Multiple use of reduction value (not safe).
4069 // - All uses of the PHI must be the reduction (safe).
4070 // - Otherwise, not safe.
4071 // - By one instruction outside of the loop (safe).
4072 // - By further instructions outside of the loop (not safe).
4073 // - By an instruction that is not part of the reduction (not safe).
4075 // * An instruction type other than PHI or the reduction operation.
4076 // * A PHI in the header other than the initial PHI.
4077 while (!Worklist.empty()) {
4078 Instruction *Cur = Worklist.back();
4079 Worklist.pop_back();
4082 // If the instruction has no users then this is a broken chain and can't be
4083 // a reduction variable.
4084 if (Cur->use_empty())
4087 bool IsAPhi = isa<PHINode>(Cur);
4089 // A header PHI use other than the original PHI.
4090 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4093 // Reductions of instructions such as Div, and Sub is only possible if the
4094 // LHS is the reduction variable.
4095 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4096 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4097 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4100 // Any reduction instruction must be of one of the allowed kinds.
4101 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4102 if (!ReduxDesc.IsReduction)
4105 // A reduction operation must only have one use of the reduction value.
4106 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4107 hasMultipleUsesOf(Cur, VisitedInsts))
4110 // All inputs to a PHI node must be a reduction value.
4111 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4114 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4115 isa<SelectInst>(Cur)))
4116 ++NumCmpSelectPatternInst;
4117 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4118 isa<SelectInst>(Cur)))
4119 ++NumCmpSelectPatternInst;
4121 // Check whether we found a reduction operator.
4122 FoundReduxOp |= !IsAPhi;
4124 // Process users of current instruction. Push non PHI nodes after PHI nodes
4125 // onto the stack. This way we are going to have seen all inputs to PHI
4126 // nodes once we get to them.
4127 SmallVector<Instruction *, 8> NonPHIs;
4128 SmallVector<Instruction *, 8> PHIs;
4129 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4131 Instruction *Usr = cast<Instruction>(*UI);
4133 // Check if we found the exit user.
4134 BasicBlock *Parent = Usr->getParent();
4135 if (!TheLoop->contains(Parent)) {
4136 // Exit if you find multiple outside users or if the header phi node is
4137 // being used. In this case the user uses the value of the previous
4138 // iteration, in which case we would loose "VF-1" iterations of the
4139 // reduction operation if we vectorize.
4140 if (ExitInstruction != 0 || Cur == Phi)
4143 // The instruction used by an outside user must be the last instruction
4144 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4145 // operations on the value.
4146 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4149 ExitInstruction = Cur;
4153 // Process instructions only once (termination).
4154 if (VisitedInsts.insert(Usr)) {
4155 if (isa<PHINode>(Usr))
4156 PHIs.push_back(Usr);
4158 NonPHIs.push_back(Usr);
4160 // Remember that we completed the cycle.
4162 FoundStartPHI = true;
4164 Worklist.append(PHIs.begin(), PHIs.end());
4165 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4168 // This means we have seen one but not the other instruction of the
4169 // pattern or more than just a select and cmp.
4170 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4171 NumCmpSelectPatternInst != 2)
4174 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4177 // We found a reduction var if we have reached the original phi node and we
4178 // only have a single instruction with out-of-loop users.
4180 // This instruction is allowed to have out-of-loop users.
4181 AllowedExit.insert(ExitInstruction);
4183 // Save the description of this reduction variable.
4184 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4185 ReduxDesc.MinMaxKind);
4186 Reductions[Phi] = RD;
4187 // We've ended the cycle. This is a reduction variable if we have an
4188 // outside user and it has a binary op.
4193 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4194 /// pattern corresponding to a min(X, Y) or max(X, Y).
4195 LoopVectorizationLegality::ReductionInstDesc
4196 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4197 ReductionInstDesc &Prev) {
4199 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4200 "Expect a select instruction");
4201 Instruction *Cmp = 0;
4202 SelectInst *Select = 0;
4204 // We must handle the select(cmp()) as a single instruction. Advance to the
4206 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4207 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4208 return ReductionInstDesc(false, I);
4209 return ReductionInstDesc(Select, Prev.MinMaxKind);
4212 // Only handle single use cases for now.
4213 if (!(Select = dyn_cast<SelectInst>(I)))
4214 return ReductionInstDesc(false, I);
4215 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4216 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4217 return ReductionInstDesc(false, I);
4218 if (!Cmp->hasOneUse())
4219 return ReductionInstDesc(false, I);
4224 // Look for a min/max pattern.
4225 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4226 return ReductionInstDesc(Select, MRK_UIntMin);
4227 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4228 return ReductionInstDesc(Select, MRK_UIntMax);
4229 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4230 return ReductionInstDesc(Select, MRK_SIntMax);
4231 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4232 return ReductionInstDesc(Select, MRK_SIntMin);
4233 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4234 return ReductionInstDesc(Select, MRK_FloatMin);
4235 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4236 return ReductionInstDesc(Select, MRK_FloatMax);
4237 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4238 return ReductionInstDesc(Select, MRK_FloatMin);
4239 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4240 return ReductionInstDesc(Select, MRK_FloatMax);
4242 return ReductionInstDesc(false, I);
4245 LoopVectorizationLegality::ReductionInstDesc
4246 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4248 ReductionInstDesc &Prev) {
4249 bool FP = I->getType()->isFloatingPointTy();
4250 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4251 switch (I->getOpcode()) {
4253 return ReductionInstDesc(false, I);
4254 case Instruction::PHI:
4255 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4256 Kind != RK_FloatMinMax))
4257 return ReductionInstDesc(false, I);
4258 return ReductionInstDesc(I, Prev.MinMaxKind);
4259 case Instruction::Sub:
4260 case Instruction::Add:
4261 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4262 case Instruction::Mul:
4263 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4264 case Instruction::And:
4265 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4266 case Instruction::Or:
4267 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4268 case Instruction::Xor:
4269 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4270 case Instruction::FMul:
4271 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4272 case Instruction::FAdd:
4273 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4274 case Instruction::FCmp:
4275 case Instruction::ICmp:
4276 case Instruction::Select:
4277 if (Kind != RK_IntegerMinMax &&
4278 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4279 return ReductionInstDesc(false, I);
4280 return isMinMaxSelectCmpPattern(I, Prev);
4284 LoopVectorizationLegality::InductionKind
4285 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4286 Type *PhiTy = Phi->getType();
4287 // We only handle integer and pointer inductions variables.
4288 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4289 return IK_NoInduction;
4291 // Check that the PHI is consecutive.
4292 const SCEV *PhiScev = SE->getSCEV(Phi);
4293 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4295 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4296 return IK_NoInduction;
4298 const SCEV *Step = AR->getStepRecurrence(*SE);
4300 // Integer inductions need to have a stride of one.
4301 if (PhiTy->isIntegerTy()) {
4303 return IK_IntInduction;
4304 if (Step->isAllOnesValue())
4305 return IK_ReverseIntInduction;
4306 return IK_NoInduction;
4309 // Calculate the pointer stride and check if it is consecutive.
4310 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4312 return IK_NoInduction;
4314 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4315 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4316 if (C->getValue()->equalsInt(Size))
4317 return IK_PtrInduction;
4318 else if (C->getValue()->equalsInt(0 - Size))
4319 return IK_ReversePtrInduction;
4321 return IK_NoInduction;
4324 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4325 Value *In0 = const_cast<Value*>(V);
4326 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4330 return Inductions.count(PN);
4333 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4334 assert(TheLoop->contains(BB) && "Unknown block used");
4336 // Blocks that do not dominate the latch need predication.
4337 BasicBlock* Latch = TheLoop->getLoopLatch();
4338 return !DT->dominates(BB, Latch);
4341 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4342 SmallPtrSet<Value *, 8>& SafePtrs) {
4343 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4344 // We might be able to hoist the load.
4345 if (it->mayReadFromMemory()) {
4346 LoadInst *LI = dyn_cast<LoadInst>(it);
4347 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4351 // We don't predicate stores at the moment.
4352 if (it->mayWriteToMemory() || it->mayThrow())
4355 // The instructions below can trap.
4356 switch (it->getOpcode()) {
4358 case Instruction::UDiv:
4359 case Instruction::SDiv:
4360 case Instruction::URem:
4361 case Instruction::SRem:
4369 LoopVectorizationCostModel::VectorizationFactor
4370 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4372 // Width 1 means no vectorize
4373 VectorizationFactor Factor = { 1U, 0U };
4374 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4375 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4379 // Find the trip count.
4380 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4381 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4383 unsigned WidestType = getWidestType();
4384 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4385 unsigned MaxSafeDepDist = -1U;
4386 if (Legal->getMaxSafeDepDistBytes() != -1U)
4387 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4388 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4389 WidestRegister : MaxSafeDepDist);
4390 unsigned MaxVectorSize = WidestRegister / WidestType;
4391 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4392 DEBUG(dbgs() << "LV: The Widest register is: "
4393 << WidestRegister << " bits.\n");
4395 if (MaxVectorSize == 0) {
4396 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4400 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4401 " into one vector!");
4403 unsigned VF = MaxVectorSize;
4405 // If we optimize the program for size, avoid creating the tail loop.
4407 // If we are unable to calculate the trip count then don't try to vectorize.
4409 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4413 // Find the maximum SIMD width that can fit within the trip count.
4414 VF = TC % MaxVectorSize;
4419 // If the trip count that we found modulo the vectorization factor is not
4420 // zero then we require a tail.
4422 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4428 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4429 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4431 Factor.Width = UserVF;
4435 float Cost = expectedCost(1);
4437 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
4438 for (unsigned i=2; i <= VF; i*=2) {
4439 // Notice that the vector loop needs to be executed less times, so
4440 // we need to divide the cost of the vector loops by the width of
4441 // the vector elements.
4442 float VectorCost = expectedCost(i) / (float)i;
4443 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4444 (int)VectorCost << ".\n");
4445 if (VectorCost < Cost) {
4451 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4452 Factor.Width = Width;
4453 Factor.Cost = Width * Cost;
4457 unsigned LoopVectorizationCostModel::getWidestType() {
4458 unsigned MaxWidth = 8;
4461 for (Loop::block_iterator bb = TheLoop->block_begin(),
4462 be = TheLoop->block_end(); bb != be; ++bb) {
4463 BasicBlock *BB = *bb;
4465 // For each instruction in the loop.
4466 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4467 Type *T = it->getType();
4469 // Only examine Loads, Stores and PHINodes.
4470 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4473 // Examine PHI nodes that are reduction variables.
4474 if (PHINode *PN = dyn_cast<PHINode>(it))
4475 if (!Legal->getReductionVars()->count(PN))
4478 // Examine the stored values.
4479 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4480 T = ST->getValueOperand()->getType();
4482 // Ignore loaded pointer types and stored pointer types that are not
4483 // consecutive. However, we do want to take consecutive stores/loads of
4484 // pointer vectors into account.
4485 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4488 MaxWidth = std::max(MaxWidth,
4489 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4497 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4500 unsigned LoopCost) {
4502 // -- The unroll heuristics --
4503 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4504 // There are many micro-architectural considerations that we can't predict
4505 // at this level. For example frontend pressure (on decode or fetch) due to
4506 // code size, or the number and capabilities of the execution ports.
4508 // We use the following heuristics to select the unroll factor:
4509 // 1. If the code has reductions the we unroll in order to break the cross
4510 // iteration dependency.
4511 // 2. If the loop is really small then we unroll in order to reduce the loop
4513 // 3. We don't unroll if we think that we will spill registers to memory due
4514 // to the increased register pressure.
4516 // Use the user preference, unless 'auto' is selected.
4520 // When we optimize for size we don't unroll.
4524 // We used the distance for the unroll factor.
4525 if (Legal->getMaxSafeDepDistBytes() != -1U)
4528 // Do not unroll loops with a relatively small trip count.
4529 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4530 TheLoop->getLoopLatch());
4531 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4534 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4535 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4536 " vector registers\n");
4538 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4539 // We divide by these constants so assume that we have at least one
4540 // instruction that uses at least one register.
4541 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4542 R.NumInstructions = std::max(R.NumInstructions, 1U);
4544 // We calculate the unroll factor using the following formula.
4545 // Subtract the number of loop invariants from the number of available
4546 // registers. These registers are used by all of the unrolled instances.
4547 // Next, divide the remaining registers by the number of registers that is
4548 // required by the loop, in order to estimate how many parallel instances
4549 // fit without causing spills.
4550 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4552 // Clamp the unroll factor ranges to reasonable factors.
4553 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4555 // If we did not calculate the cost for VF (because the user selected the VF)
4556 // then we calculate the cost of VF here.
4558 LoopCost = expectedCost(VF);
4560 // Clamp the calculated UF to be between the 1 and the max unroll factor
4561 // that the target allows.
4562 if (UF > MaxUnrollSize)
4567 bool HasReductions = Legal->getReductionVars()->size();
4569 // Decide if we want to unroll if we decided that it is legal to vectorize
4570 // but not profitable.
4572 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4573 LoopCost > SmallLoopCost)
4579 if (HasReductions) {
4580 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4584 // We want to unroll tiny loops in order to reduce the loop overhead.
4585 // We assume that the cost overhead is 1 and we use the cost model
4586 // to estimate the cost of the loop and unroll until the cost of the
4587 // loop overhead is about 5% of the cost of the loop.
4588 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4589 if (LoopCost < SmallLoopCost) {
4590 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4591 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4592 return std::min(NewUF, UF);
4595 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4599 LoopVectorizationCostModel::RegisterUsage
4600 LoopVectorizationCostModel::calculateRegisterUsage() {
4601 // This function calculates the register usage by measuring the highest number
4602 // of values that are alive at a single location. Obviously, this is a very
4603 // rough estimation. We scan the loop in a topological order in order and
4604 // assign a number to each instruction. We use RPO to ensure that defs are
4605 // met before their users. We assume that each instruction that has in-loop
4606 // users starts an interval. We record every time that an in-loop value is
4607 // used, so we have a list of the first and last occurrences of each
4608 // instruction. Next, we transpose this data structure into a multi map that
4609 // holds the list of intervals that *end* at a specific location. This multi
4610 // map allows us to perform a linear search. We scan the instructions linearly
4611 // and record each time that a new interval starts, by placing it in a set.
4612 // If we find this value in the multi-map then we remove it from the set.
4613 // The max register usage is the maximum size of the set.
4614 // We also search for instructions that are defined outside the loop, but are
4615 // used inside the loop. We need this number separately from the max-interval
4616 // usage number because when we unroll, loop-invariant values do not take
4618 LoopBlocksDFS DFS(TheLoop);
4622 R.NumInstructions = 0;
4624 // Each 'key' in the map opens a new interval. The values
4625 // of the map are the index of the 'last seen' usage of the
4626 // instruction that is the key.
4627 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4628 // Maps instruction to its index.
4629 DenseMap<unsigned, Instruction*> IdxToInstr;
4630 // Marks the end of each interval.
4631 IntervalMap EndPoint;
4632 // Saves the list of instruction indices that are used in the loop.
4633 SmallSet<Instruction*, 8> Ends;
4634 // Saves the list of values that are used in the loop but are
4635 // defined outside the loop, such as arguments and constants.
4636 SmallPtrSet<Value*, 8> LoopInvariants;
4639 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4640 be = DFS.endRPO(); bb != be; ++bb) {
4641 R.NumInstructions += (*bb)->size();
4642 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4644 Instruction *I = it;
4645 IdxToInstr[Index++] = I;
4647 // Save the end location of each USE.
4648 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4649 Value *U = I->getOperand(i);
4650 Instruction *Instr = dyn_cast<Instruction>(U);
4652 // Ignore non-instruction values such as arguments, constants, etc.
4653 if (!Instr) continue;
4655 // If this instruction is outside the loop then record it and continue.
4656 if (!TheLoop->contains(Instr)) {
4657 LoopInvariants.insert(Instr);
4661 // Overwrite previous end points.
4662 EndPoint[Instr] = Index;
4668 // Saves the list of intervals that end with the index in 'key'.
4669 typedef SmallVector<Instruction*, 2> InstrList;
4670 DenseMap<unsigned, InstrList> TransposeEnds;
4672 // Transpose the EndPoints to a list of values that end at each index.
4673 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4675 TransposeEnds[it->second].push_back(it->first);
4677 SmallSet<Instruction*, 8> OpenIntervals;
4678 unsigned MaxUsage = 0;
4681 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4682 for (unsigned int i = 0; i < Index; ++i) {
4683 Instruction *I = IdxToInstr[i];
4684 // Ignore instructions that are never used within the loop.
4685 if (!Ends.count(I)) continue;
4687 // Remove all of the instructions that end at this location.
4688 InstrList &List = TransposeEnds[i];
4689 for (unsigned int j=0, e = List.size(); j < e; ++j)
4690 OpenIntervals.erase(List[j]);
4692 // Count the number of live interals.
4693 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4695 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4696 OpenIntervals.size() << '\n');
4698 // Add the current instruction to the list of open intervals.
4699 OpenIntervals.insert(I);
4702 unsigned Invariant = LoopInvariants.size();
4703 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4704 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4705 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4707 R.LoopInvariantRegs = Invariant;
4708 R.MaxLocalUsers = MaxUsage;
4712 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4716 for (Loop::block_iterator bb = TheLoop->block_begin(),
4717 be = TheLoop->block_end(); bb != be; ++bb) {
4718 unsigned BlockCost = 0;
4719 BasicBlock *BB = *bb;
4721 // For each instruction in the old loop.
4722 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4723 // Skip dbg intrinsics.
4724 if (isa<DbgInfoIntrinsic>(it))
4727 unsigned C = getInstructionCost(it, VF);
4729 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4730 VF << " For instruction: " << *it << '\n');
4733 // We assume that if-converted blocks have a 50% chance of being executed.
4734 // When the code is scalar then some of the blocks are avoided due to CF.
4735 // When the code is vectorized we execute all code paths.
4736 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4745 /// \brief Check whether the address computation for a non-consecutive memory
4746 /// access looks like an unlikely candidate for being merged into the indexing
4749 /// We look for a GEP which has one index that is an induction variable and all
4750 /// other indices are loop invariant. If the stride of this access is also
4751 /// within a small bound we decide that this address computation can likely be
4752 /// merged into the addressing mode.
4753 /// In all other cases, we identify the address computation as complex.
4754 static bool isLikelyComplexAddressComputation(Value *Ptr,
4755 LoopVectorizationLegality *Legal,
4756 ScalarEvolution *SE,
4757 const Loop *TheLoop) {
4758 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4762 // We are looking for a gep with all loop invariant indices except for one
4763 // which should be an induction variable.
4764 unsigned NumOperands = Gep->getNumOperands();
4765 for (unsigned i = 1; i < NumOperands; ++i) {
4766 Value *Opd = Gep->getOperand(i);
4767 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4768 !Legal->isInductionVariable(Opd))
4772 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4773 // can likely be merged into the address computation.
4774 unsigned MaxMergeDistance = 64;
4776 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4780 // Check the step is constant.
4781 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4782 // Calculate the pointer stride and check if it is consecutive.
4783 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4787 const APInt &APStepVal = C->getValue()->getValue();
4789 // Huge step value - give up.
4790 if (APStepVal.getBitWidth() > 64)
4793 int64_t StepVal = APStepVal.getSExtValue();
4795 return StepVal > MaxMergeDistance;
4799 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4800 // If we know that this instruction will remain uniform, check the cost of
4801 // the scalar version.
4802 if (Legal->isUniformAfterVectorization(I))
4805 Type *RetTy = I->getType();
4806 Type *VectorTy = ToVectorTy(RetTy, VF);
4808 // TODO: We need to estimate the cost of intrinsic calls.
4809 switch (I->getOpcode()) {
4810 case Instruction::GetElementPtr:
4811 // We mark this instruction as zero-cost because the cost of GEPs in
4812 // vectorized code depends on whether the corresponding memory instruction
4813 // is scalarized or not. Therefore, we handle GEPs with the memory
4814 // instruction cost.
4816 case Instruction::Br: {
4817 return TTI.getCFInstrCost(I->getOpcode());
4819 case Instruction::PHI:
4820 //TODO: IF-converted IFs become selects.
4822 case Instruction::Add:
4823 case Instruction::FAdd:
4824 case Instruction::Sub:
4825 case Instruction::FSub:
4826 case Instruction::Mul:
4827 case Instruction::FMul:
4828 case Instruction::UDiv:
4829 case Instruction::SDiv:
4830 case Instruction::FDiv:
4831 case Instruction::URem:
4832 case Instruction::SRem:
4833 case Instruction::FRem:
4834 case Instruction::Shl:
4835 case Instruction::LShr:
4836 case Instruction::AShr:
4837 case Instruction::And:
4838 case Instruction::Or:
4839 case Instruction::Xor: {
4840 // Certain instructions can be cheaper to vectorize if they have a constant
4841 // second vector operand. One example of this are shifts on x86.
4842 TargetTransformInfo::OperandValueKind Op1VK =
4843 TargetTransformInfo::OK_AnyValue;
4844 TargetTransformInfo::OperandValueKind Op2VK =
4845 TargetTransformInfo::OK_AnyValue;
4847 if (isa<ConstantInt>(I->getOperand(1)))
4848 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4850 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4852 case Instruction::Select: {
4853 SelectInst *SI = cast<SelectInst>(I);
4854 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4855 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4856 Type *CondTy = SI->getCondition()->getType();
4858 CondTy = VectorType::get(CondTy, VF);
4860 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4862 case Instruction::ICmp:
4863 case Instruction::FCmp: {
4864 Type *ValTy = I->getOperand(0)->getType();
4865 VectorTy = ToVectorTy(ValTy, VF);
4866 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4868 case Instruction::Store:
4869 case Instruction::Load: {
4870 StoreInst *SI = dyn_cast<StoreInst>(I);
4871 LoadInst *LI = dyn_cast<LoadInst>(I);
4872 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4874 VectorTy = ToVectorTy(ValTy, VF);
4876 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4877 unsigned AS = SI ? SI->getPointerAddressSpace() :
4878 LI->getPointerAddressSpace();
4879 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4880 // We add the cost of address computation here instead of with the gep
4881 // instruction because only here we know whether the operation is
4884 return TTI.getAddressComputationCost(VectorTy) +
4885 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4887 // Scalarized loads/stores.
4888 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4889 bool Reverse = ConsecutiveStride < 0;
4890 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4891 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4892 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4893 bool IsComplexComputation =
4894 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4896 // The cost of extracting from the value vector and pointer vector.
4897 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4898 for (unsigned i = 0; i < VF; ++i) {
4899 // The cost of extracting the pointer operand.
4900 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4901 // In case of STORE, the cost of ExtractElement from the vector.
4902 // In case of LOAD, the cost of InsertElement into the returned
4904 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4905 Instruction::InsertElement,
4909 // The cost of the scalar loads/stores.
4910 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4911 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4916 // Wide load/stores.
4917 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4918 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4921 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4925 case Instruction::ZExt:
4926 case Instruction::SExt:
4927 case Instruction::FPToUI:
4928 case Instruction::FPToSI:
4929 case Instruction::FPExt:
4930 case Instruction::PtrToInt:
4931 case Instruction::IntToPtr:
4932 case Instruction::SIToFP:
4933 case Instruction::UIToFP:
4934 case Instruction::Trunc:
4935 case Instruction::FPTrunc:
4936 case Instruction::BitCast: {
4937 // We optimize the truncation of induction variable.
4938 // The cost of these is the same as the scalar operation.
4939 if (I->getOpcode() == Instruction::Trunc &&
4940 Legal->isInductionVariable(I->getOperand(0)))
4941 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4942 I->getOperand(0)->getType());
4944 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4945 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4947 case Instruction::Call: {
4948 CallInst *CI = cast<CallInst>(I);
4949 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4950 assert(ID && "Not an intrinsic call!");
4951 Type *RetTy = ToVectorTy(CI->getType(), VF);
4952 SmallVector<Type*, 4> Tys;
4953 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4954 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4955 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4958 // We are scalarizing the instruction. Return the cost of the scalar
4959 // instruction, plus the cost of insert and extract into vector
4960 // elements, times the vector width.
4963 if (!RetTy->isVoidTy() && VF != 1) {
4964 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4966 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4969 // The cost of inserting the results plus extracting each one of the
4971 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4974 // The cost of executing VF copies of the scalar instruction. This opcode
4975 // is unknown. Assume that it is the same as 'mul'.
4976 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4982 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4983 if (Scalar->isVoidTy() || VF == 1)
4985 return VectorType::get(Scalar, VF);
4988 char LoopVectorize::ID = 0;
4989 static const char lv_name[] = "Loop Vectorization";
4990 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4991 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4992 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4993 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4994 INITIALIZE_PASS_DEPENDENCY(LCSSA)
4995 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4996 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4997 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5000 Pass *createLoopVectorizePass(bool NoUnrolling) {
5001 return new LoopVectorize(NoUnrolling);
5005 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5006 // Check for a store.
5007 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5008 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5010 // Check for a load.
5011 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5012 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5018 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
5019 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5020 // Holds vector parameters or scalars, in case of uniform vals.
5021 SmallVector<VectorParts, 4> Params;
5023 setDebugLocFromInst(Builder, Instr);
5025 // Find all of the vectorized parameters.
5026 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5027 Value *SrcOp = Instr->getOperand(op);
5029 // If we are accessing the old induction variable, use the new one.
5030 if (SrcOp == OldInduction) {
5031 Params.push_back(getVectorValue(SrcOp));
5035 // Try using previously calculated values.
5036 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5038 // If the src is an instruction that appeared earlier in the basic block
5039 // then it should already be vectorized.
5040 if (SrcInst && OrigLoop->contains(SrcInst)) {
5041 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5042 // The parameter is a vector value from earlier.
5043 Params.push_back(WidenMap.get(SrcInst));
5045 // The parameter is a scalar from outside the loop. Maybe even a constant.
5046 VectorParts Scalars;
5047 Scalars.append(UF, SrcOp);
5048 Params.push_back(Scalars);
5052 assert(Params.size() == Instr->getNumOperands() &&
5053 "Invalid number of operands");
5055 // Does this instruction return a value ?
5056 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5058 Value *UndefVec = IsVoidRetTy ? 0 :
5059 UndefValue::get(Instr->getType());
5060 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5061 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5063 // For each vector unroll 'part':
5064 for (unsigned Part = 0; Part < UF; ++Part) {
5065 // For each scalar that we create:
5067 Instruction *Cloned = Instr->clone();
5069 Cloned->setName(Instr->getName() + ".cloned");
5070 // Replace the operands of the cloned instructions with extracted scalars.
5071 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5072 Value *Op = Params[op][Part];
5073 Cloned->setOperand(op, Op);
5076 // Place the cloned scalar in the new loop.
5077 Builder.Insert(Cloned);
5079 // If the original scalar returns a value we need to place it in a vector
5080 // so that future users will be able to use it.
5082 VecResults[Part] = Cloned;
5087 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
5088 LoopVectorizationLegality*) {
5089 return scalarizeInstruction(Instr);
5092 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5096 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5100 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5102 // When unrolling and the VF is 1, we only need to add a simple scalar.
5103 Type *ITy = Val->getType();
5104 assert(!ITy->isVectorTy() && "Val must be a scalar");
5105 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5106 return Builder.CreateAdd(Val, C, "induction");