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/MapVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/StringExtras.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/AliasSetTracker.h"
57 #include "llvm/Analysis/Dominators.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopIterator.h"
60 #include "llvm/Analysis/LoopPass.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpander.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/TargetTransformInfo.h"
65 #include "llvm/Analysis/ValueTracking.h"
66 #include "llvm/Analysis/Verifier.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Function.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/LLVMContext.h"
75 #include "llvm/IR/Module.h"
76 #include "llvm/IR/Type.h"
77 #include "llvm/IR/Value.h"
78 #include "llvm/Pass.h"
79 #include "llvm/Support/CommandLine.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/PatternMatch.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/Support/ValueHandle.h"
84 #include "llvm/Target/TargetLibraryInfo.h"
85 #include "llvm/Transforms/Scalar.h"
86 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
87 #include "llvm/Transforms/Utils/Local.h"
92 using namespace llvm::PatternMatch;
94 static cl::opt<unsigned>
95 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
96 cl::desc("Sets the SIMD width. Zero is autoselect."));
98 static cl::opt<unsigned>
99 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
100 cl::desc("Sets the vectorization unroll count. "
101 "Zero is autoselect."));
104 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
105 cl::desc("Enable if-conversion during vectorization."));
107 /// We don't vectorize loops with a known constant trip count below this number.
108 static cl::opt<unsigned>
109 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
111 cl::desc("Don't vectorize loops with a constant "
112 "trip count that is smaller than this "
115 /// We don't unroll loops with a known constant trip count below this number.
116 static const unsigned TinyTripCountUnrollThreshold = 128;
118 /// When performing memory disambiguation checks at runtime do not make more
119 /// than this number of comparisons.
120 static const unsigned RuntimeMemoryCheckThreshold = 8;
122 /// We use a metadata with this name to indicate that a scalar loop was
123 /// vectorized and that we don't need to re-vectorize it if we run into it
126 AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized";
130 // Forward declarations.
131 class LoopVectorizationLegality;
132 class LoopVectorizationCostModel;
134 /// InnerLoopVectorizer vectorizes loops which contain only one basic
135 /// block to a specified vectorization factor (VF).
136 /// This class performs the widening of scalars into vectors, or multiple
137 /// scalars. This class also implements the following features:
138 /// * It inserts an epilogue loop for handling loops that don't have iteration
139 /// counts that are known to be a multiple of the vectorization factor.
140 /// * It handles the code generation for reduction variables.
141 /// * Scalarization (implementation using scalars) of un-vectorizable
143 /// InnerLoopVectorizer does not perform any vectorization-legality
144 /// checks, and relies on the caller to check for the different legality
145 /// aspects. The InnerLoopVectorizer relies on the
146 /// LoopVectorizationLegality class to provide information about the induction
147 /// and reduction variables that were found to a given vectorization factor.
148 class InnerLoopVectorizer {
150 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
151 DominatorTree *DT, DataLayout *DL,
152 const TargetLibraryInfo *TLI, unsigned VecWidth,
153 unsigned UnrollFactor)
154 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
155 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
156 OldInduction(0), WidenMap(UnrollFactor) {}
158 // Perform the actual loop widening (vectorization).
159 void vectorize(LoopVectorizationLegality *Legal) {
160 // Create a new empty loop. Unlink the old loop and connect the new one.
161 createEmptyLoop(Legal);
162 // Widen each instruction in the old loop to a new one in the new loop.
163 // Use the Legality module to find the induction and reduction variables.
164 vectorizeLoop(Legal);
165 // Register the new loop and update the analysis passes.
170 /// A small list of PHINodes.
171 typedef SmallVector<PHINode*, 4> PhiVector;
172 /// When we unroll loops we have multiple vector values for each scalar.
173 /// This data structure holds the unrolled and vectorized values that
174 /// originated from one scalar instruction.
175 typedef SmallVector<Value*, 2> VectorParts;
177 /// Add code that checks at runtime if the accessed arrays overlap.
178 /// Returns the comparator value or NULL if no check is needed.
179 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
181 /// Create an empty loop, based on the loop ranges of the old loop.
182 void createEmptyLoop(LoopVectorizationLegality *Legal);
183 /// Copy and widen the instructions from the old loop.
184 void vectorizeLoop(LoopVectorizationLegality *Legal);
186 /// A helper function that computes the predicate of the block BB, assuming
187 /// that the header block of the loop is set to True. It returns the *entry*
188 /// mask for the block BB.
189 VectorParts createBlockInMask(BasicBlock *BB);
190 /// A helper function that computes the predicate of the edge between SRC
192 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
194 /// A helper function to vectorize a single BB within the innermost loop.
195 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
198 /// Insert the new loop to the loop hierarchy and pass manager
199 /// and update the analysis passes.
200 void updateAnalysis();
202 /// This instruction is un-vectorizable. Implement it as a sequence
204 void scalarizeInstruction(Instruction *Instr);
206 /// Vectorize Load and Store instructions,
207 void vectorizeMemoryInstruction(Instruction *Instr,
208 LoopVectorizationLegality *Legal);
210 /// Create a broadcast instruction. This method generates a broadcast
211 /// instruction (shuffle) for loop invariant values and for the induction
212 /// value. If this is the induction variable then we extend it to N, N+1, ...
213 /// this is needed because each iteration in the loop corresponds to a SIMD
215 Value *getBroadcastInstrs(Value *V);
217 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
218 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
219 /// The sequence starts at StartIndex.
220 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
222 /// When we go over instructions in the basic block we rely on previous
223 /// values within the current basic block or on loop invariant values.
224 /// When we widen (vectorize) values we place them in the map. If the values
225 /// are not within the map, they have to be loop invariant, so we simply
226 /// broadcast them into a vector.
227 VectorParts &getVectorValue(Value *V);
229 /// Generate a shuffle sequence that will reverse the vector Vec.
230 Value *reverseVector(Value *Vec);
232 /// This is a helper class that holds the vectorizer state. It maps scalar
233 /// instructions to vector instructions. When the code is 'unrolled' then
234 /// then a single scalar value is mapped to multiple vector parts. The parts
235 /// are stored in the VectorPart type.
237 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
239 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
241 /// \return True if 'Key' is saved in the Value Map.
242 bool has(Value *Key) const { return MapStorage.count(Key); }
244 /// Initializes a new entry in the map. Sets all of the vector parts to the
245 /// save value in 'Val'.
246 /// \return A reference to a vector with splat values.
247 VectorParts &splat(Value *Key, Value *Val) {
248 VectorParts &Entry = MapStorage[Key];
249 Entry.assign(UF, Val);
253 ///\return A reference to the value that is stored at 'Key'.
254 VectorParts &get(Value *Key) {
255 VectorParts &Entry = MapStorage[Key];
258 assert(Entry.size() == UF);
263 /// The unroll factor. Each entry in the map stores this number of vector
267 /// Map storage. We use std::map and not DenseMap because insertions to a
268 /// dense map invalidates its iterators.
269 std::map<Value *, VectorParts> MapStorage;
272 /// The original loop.
274 /// Scev analysis to use.
282 /// Target Library Info.
283 const TargetLibraryInfo *TLI;
285 /// The vectorization SIMD factor to use. Each vector will have this many
288 /// The vectorization unroll factor to use. Each scalar is vectorized to this
289 /// many different vector instructions.
292 /// The builder that we use
295 // --- Vectorization state ---
297 /// The vector-loop preheader.
298 BasicBlock *LoopVectorPreHeader;
299 /// The scalar-loop preheader.
300 BasicBlock *LoopScalarPreHeader;
301 /// Middle Block between the vector and the scalar.
302 BasicBlock *LoopMiddleBlock;
303 ///The ExitBlock of the scalar loop.
304 BasicBlock *LoopExitBlock;
305 ///The vector loop body.
306 BasicBlock *LoopVectorBody;
307 ///The scalar loop body.
308 BasicBlock *LoopScalarBody;
309 /// A list of all bypass blocks. The first block is the entry of the loop.
310 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
312 /// The new Induction variable which was added to the new block.
314 /// The induction variable of the old basic block.
315 PHINode *OldInduction;
316 /// Maps scalars to widened vectors.
320 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
321 /// to what vectorization factor.
322 /// This class does not look at the profitability of vectorization, only the
323 /// legality. This class has two main kinds of checks:
324 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
325 /// will change the order of memory accesses in a way that will change the
326 /// correctness of the program.
327 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
328 /// checks for a number of different conditions, such as the availability of a
329 /// single induction variable, that all types are supported and vectorize-able,
330 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
331 /// This class is also used by InnerLoopVectorizer for identifying
332 /// induction variable and the different reduction variables.
333 class LoopVectorizationLegality {
335 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
336 DominatorTree *DT, TargetTransformInfo* TTI,
337 AliasAnalysis *AA, TargetLibraryInfo *TLI)
338 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
339 Induction(0), HasFunNoNaNAttr(false) {}
341 /// This enum represents the kinds of reductions that we support.
343 RK_NoReduction, ///< Not a reduction.
344 RK_IntegerAdd, ///< Sum of integers.
345 RK_IntegerMult, ///< Product of integers.
346 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
347 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
348 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
349 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
350 RK_FloatAdd, ///< Sum of floats.
351 RK_FloatMult, ///< Product of floats.
352 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
355 /// This enum represents the kinds of inductions that we support.
357 IK_NoInduction, ///< Not an induction variable.
358 IK_IntInduction, ///< Integer induction variable. Step = 1.
359 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
360 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
361 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
364 // This enum represents the kind of minmax reduction.
365 enum MinMaxReductionKind {
375 /// This POD struct holds information about reduction variables.
376 struct ReductionDescriptor {
377 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
378 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
380 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
381 MinMaxReductionKind MK)
382 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
384 // The starting value of the reduction.
385 // It does not have to be zero!
386 TrackingVH<Value> StartValue;
387 // The instruction who's value is used outside the loop.
388 Instruction *LoopExitInstr;
389 // The kind of the reduction.
391 // If this a min/max reduction the kind of reduction.
392 MinMaxReductionKind MinMaxKind;
395 /// This POD struct holds information about a potential reduction operation.
396 struct ReductionInstDesc {
397 ReductionInstDesc(bool IsRedux, Instruction *I) :
398 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
400 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
401 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
403 // Is this instruction a reduction candidate.
405 // The last instruction in a min/max pattern (select of the select(icmp())
406 // pattern), or the current reduction instruction otherwise.
407 Instruction *PatternLastInst;
408 // If this is a min/max pattern the comparison predicate.
409 MinMaxReductionKind MinMaxKind;
412 // This POD struct holds information about the memory runtime legality
413 // check that a group of pointers do not overlap.
414 struct RuntimePointerCheck {
415 RuntimePointerCheck() : Need(false) {}
417 /// Reset the state of the pointer runtime information.
425 /// Insert a pointer and calculate the start and end SCEVs.
426 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr);
428 /// This flag indicates if we need to add the runtime check.
430 /// Holds the pointers that we need to check.
431 SmallVector<TrackingVH<Value>, 2> Pointers;
432 /// Holds the pointer value at the beginning of the loop.
433 SmallVector<const SCEV*, 2> Starts;
434 /// Holds the pointer value at the end of the loop.
435 SmallVector<const SCEV*, 2> Ends;
436 /// Holds the information if this pointer is used for writing to memory.
437 SmallVector<bool, 2> IsWritePtr;
440 /// A POD for saving information about induction variables.
441 struct InductionInfo {
442 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
443 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
445 TrackingVH<Value> StartValue;
450 /// ReductionList contains the reduction descriptors for all
451 /// of the reductions that were found in the loop.
452 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
454 /// InductionList saves induction variables and maps them to the
455 /// induction descriptor.
456 typedef MapVector<PHINode*, InductionInfo> InductionList;
458 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
459 /// respective Store/Load instruction(s) to calculate aliasing.
460 typedef MapVector<Value*, Instruction* > AliasMap;
461 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
463 /// Returns true if it is legal to vectorize this loop.
464 /// This does not mean that it is profitable to vectorize this
465 /// loop, only that it is legal to do so.
468 /// Returns the Induction variable.
469 PHINode *getInduction() { return Induction; }
471 /// Returns the reduction variables found in the loop.
472 ReductionList *getReductionVars() { return &Reductions; }
474 /// Returns the induction variables found in the loop.
475 InductionList *getInductionVars() { return &Inductions; }
477 /// Returns True if V is an induction variable in this loop.
478 bool isInductionVariable(const Value *V);
480 /// Return true if the block BB needs to be predicated in order for the loop
481 /// to be vectorized.
482 bool blockNeedsPredication(BasicBlock *BB);
484 /// Check if this pointer is consecutive when vectorizing. This happens
485 /// when the last index of the GEP is the induction variable, or that the
486 /// pointer itself is an induction variable.
487 /// This check allows us to vectorize A[idx] into a wide load/store.
489 /// 0 - Stride is unknown or non consecutive.
490 /// 1 - Address is consecutive.
491 /// -1 - Address is consecutive, and decreasing.
492 int isConsecutivePtr(Value *Ptr);
494 /// Returns true if the value V is uniform within the loop.
495 bool isUniform(Value *V);
497 /// Returns true if this instruction will remain scalar after vectorization.
498 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
500 /// Returns the information that we collected about runtime memory check.
501 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
503 /// This function returns the identity element (or neutral element) for
505 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
507 /// Check if a single basic block loop is vectorizable.
508 /// At this point we know that this is a loop with a constant trip count
509 /// and we only need to check individual instructions.
510 bool canVectorizeInstrs();
512 /// When we vectorize loops we may change the order in which
513 /// we read and write from memory. This method checks if it is
514 /// legal to vectorize the code, considering only memory constrains.
515 /// Returns true if the loop is vectorizable
516 bool canVectorizeMemory();
518 /// Return true if we can vectorize this loop using the IF-conversion
520 bool canVectorizeWithIfConvert();
522 /// Collect the variables that need to stay uniform after vectorization.
523 void collectLoopUniforms();
525 /// Return true if all of the instructions in the block can be speculatively
527 bool blockCanBePredicated(BasicBlock *BB);
529 /// Returns True, if 'Phi' is the kind of reduction variable for type
530 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
531 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
532 /// Returns a struct describing if the instruction 'I' can be a reduction
533 /// variable of type 'Kind'. If the reduction is a min/max pattern of
534 /// select(icmp()) this function advances the instruction pointer 'I' from the
535 /// compare instruction to the select instruction and stores this pointer in
536 /// 'PatternLastInst' member of the returned struct.
537 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
538 ReductionInstDesc &Desc);
539 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
540 /// pattern corresponding to a min(X, Y) or max(X, Y).
541 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
542 ReductionInstDesc &Prev);
543 /// Returns the induction kind of Phi. This function may return NoInduction
544 /// if the PHI is not an induction variable.
545 InductionKind isInductionVariable(PHINode *Phi);
546 /// Return true if can compute the address bounds of Ptr within the loop.
547 bool hasComputableBounds(Value *Ptr);
548 /// Return true if there is the chance of write reorder.
549 bool hasPossibleGlobalWriteReorder(Value *Object,
551 AliasMultiMap &WriteObjects,
552 unsigned MaxByteWidth);
553 /// Return the AA location for a load or a store.
554 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
557 /// The loop that we evaluate.
561 /// DataLayout analysis.
566 TargetTransformInfo *TTI;
569 /// Target Library Info.
570 TargetLibraryInfo *TLI;
572 // --- vectorization state --- //
574 /// Holds the integer induction variable. This is the counter of the
577 /// Holds the reduction variables.
578 ReductionList Reductions;
579 /// Holds all of the induction variables that we found in the loop.
580 /// Notice that inductions don't need to start at zero and that induction
581 /// variables can be pointers.
582 InductionList Inductions;
584 /// Allowed outside users. This holds the reduction
585 /// vars which can be accessed from outside the loop.
586 SmallPtrSet<Value*, 4> AllowedExit;
587 /// This set holds the variables which are known to be uniform after
589 SmallPtrSet<Instruction*, 4> Uniforms;
590 /// We need to check that all of the pointers in this list are disjoint
592 RuntimePointerCheck PtrRtCheck;
593 /// Can we assume the absence of NaNs.
594 bool HasFunNoNaNAttr;
597 /// LoopVectorizationCostModel - estimates the expected speedups due to
599 /// In many cases vectorization is not profitable. This can happen because of
600 /// a number of reasons. In this class we mainly attempt to predict the
601 /// expected speedup/slowdowns due to the supported instruction set. We use the
602 /// TargetTransformInfo to query the different backends for the cost of
603 /// different operations.
604 class LoopVectorizationCostModel {
606 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
607 LoopVectorizationLegality *Legal,
608 const TargetTransformInfo &TTI,
609 DataLayout *DL, const TargetLibraryInfo *TLI)
610 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
612 /// Information about vectorization costs
613 struct VectorizationFactor {
614 unsigned Width; // Vector width with best cost
615 unsigned Cost; // Cost of the loop with that width
617 /// \return The most profitable vectorization factor and the cost of that VF.
618 /// This method checks every power of two up to VF. If UserVF is not ZERO
619 /// then this vectorization factor will be selected if vectorization is
621 VectorizationFactor selectVectorizationFactor(bool OptForSize,
624 /// \return The size (in bits) of the widest type in the code that
625 /// needs to be vectorized. We ignore values that remain scalar such as
626 /// 64 bit loop indices.
627 unsigned getWidestType();
629 /// \return The most profitable unroll factor.
630 /// If UserUF is non-zero then this method finds the best unroll-factor
631 /// based on register pressure and other parameters.
632 /// VF and LoopCost are the selected vectorization factor and the cost of the
634 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
637 /// \brief A struct that represents some properties of the register usage
639 struct RegisterUsage {
640 /// Holds the number of loop invariant values that are used in the loop.
641 unsigned LoopInvariantRegs;
642 /// Holds the maximum number of concurrent live intervals in the loop.
643 unsigned MaxLocalUsers;
644 /// Holds the number of instructions in the loop.
645 unsigned NumInstructions;
648 /// \return information about the register usage of the loop.
649 RegisterUsage calculateRegisterUsage();
652 /// Returns the expected execution cost. The unit of the cost does
653 /// not matter because we use the 'cost' units to compare different
654 /// vector widths. The cost that is returned is *not* normalized by
655 /// the factor width.
656 unsigned expectedCost(unsigned VF);
658 /// Returns the execution time cost of an instruction for a given vector
659 /// width. Vector width of one means scalar.
660 unsigned getInstructionCost(Instruction *I, unsigned VF);
662 /// A helper function for converting Scalar types to vector types.
663 /// If the incoming type is void, we return void. If the VF is 1, we return
665 static Type* ToVectorTy(Type *Scalar, unsigned VF);
667 /// Returns whether the instruction is a load or store and will be a emitted
668 /// as a vector operation.
669 bool isConsecutiveLoadOrStore(Instruction *I);
671 /// The loop that we evaluate.
675 /// Loop Info analysis.
677 /// Vectorization legality.
678 LoopVectorizationLegality *Legal;
679 /// Vector target information.
680 const TargetTransformInfo &TTI;
681 /// Target data layout information.
683 /// Target Library Info.
684 const TargetLibraryInfo *TLI;
687 /// The LoopVectorize Pass.
688 struct LoopVectorize : public LoopPass {
689 /// Pass identification, replacement for typeid
692 explicit LoopVectorize() : LoopPass(ID) {
693 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
699 TargetTransformInfo *TTI;
702 TargetLibraryInfo *TLI;
704 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
705 // We only vectorize innermost loops.
709 SE = &getAnalysis<ScalarEvolution>();
710 DL = getAnalysisIfAvailable<DataLayout>();
711 LI = &getAnalysis<LoopInfo>();
712 TTI = &getAnalysis<TargetTransformInfo>();
713 DT = &getAnalysis<DominatorTree>();
714 AA = getAnalysisIfAvailable<AliasAnalysis>();
715 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
718 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
722 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
723 L->getHeader()->getParent()->getName() << "\"\n");
725 // Check if it is legal to vectorize the loop.
726 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
727 if (!LVL.canVectorize()) {
728 DEBUG(dbgs() << "LV: Not vectorizing.\n");
732 // Use the cost model.
733 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
735 // Check the function attributes to find out if this function should be
736 // optimized for size.
737 Function *F = L->getHeader()->getParent();
738 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
739 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
740 unsigned FnIndex = AttributeSet::FunctionIndex;
741 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
742 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
745 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
746 "attribute is used.\n");
750 // Select the optimal vectorization factor.
751 LoopVectorizationCostModel::VectorizationFactor VF;
752 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
753 // Select the unroll factor.
754 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
758 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
762 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
763 F->getParent()->getModuleIdentifier()<<"\n");
764 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
766 // If we decided that it is *legal* to vectorize the loop then do it.
767 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
770 DEBUG(verifyFunction(*L->getHeader()->getParent()));
774 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
775 LoopPass::getAnalysisUsage(AU);
776 AU.addRequiredID(LoopSimplifyID);
777 AU.addRequiredID(LCSSAID);
778 AU.addRequired<DominatorTree>();
779 AU.addRequired<LoopInfo>();
780 AU.addRequired<ScalarEvolution>();
781 AU.addRequired<TargetTransformInfo>();
782 AU.addPreserved<LoopInfo>();
783 AU.addPreserved<DominatorTree>();
788 } // end anonymous namespace
790 //===----------------------------------------------------------------------===//
791 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
792 // LoopVectorizationCostModel.
793 //===----------------------------------------------------------------------===//
796 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
797 Loop *Lp, Value *Ptr,
799 const SCEV *Sc = SE->getSCEV(Ptr);
800 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
801 assert(AR && "Invalid addrec expression");
802 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
803 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
804 Pointers.push_back(Ptr);
805 Starts.push_back(AR->getStart());
806 Ends.push_back(ScEnd);
807 IsWritePtr.push_back(WritePtr);
810 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
811 // Save the current insertion location.
812 Instruction *Loc = Builder.GetInsertPoint();
814 // We need to place the broadcast of invariant variables outside the loop.
815 Instruction *Instr = dyn_cast<Instruction>(V);
816 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
817 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
819 // Place the code for broadcasting invariant variables in the new preheader.
821 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
823 // Broadcast the scalar into all locations in the vector.
824 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
826 // Restore the builder insertion point.
828 Builder.SetInsertPoint(Loc);
833 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
835 assert(Val->getType()->isVectorTy() && "Must be a vector");
836 assert(Val->getType()->getScalarType()->isIntegerTy() &&
837 "Elem must be an integer");
839 Type *ITy = Val->getType()->getScalarType();
840 VectorType *Ty = cast<VectorType>(Val->getType());
841 int VLen = Ty->getNumElements();
842 SmallVector<Constant*, 8> Indices;
844 // Create a vector of consecutive numbers from zero to VF.
845 for (int i = 0; i < VLen; ++i) {
846 int64_t Idx = Negate ? (-i) : i;
847 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
850 // Add the consecutive indices to the vector value.
851 Constant *Cv = ConstantVector::get(Indices);
852 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
853 return Builder.CreateAdd(Val, Cv, "induction");
856 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
857 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
858 // Make sure that the pointer does not point to structs.
859 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
862 // If this value is a pointer induction variable we know it is consecutive.
863 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
864 if (Phi && Inductions.count(Phi)) {
865 InductionInfo II = Inductions[Phi];
866 if (IK_PtrInduction == II.IK)
868 else if (IK_ReversePtrInduction == II.IK)
872 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
876 unsigned NumOperands = Gep->getNumOperands();
877 Value *LastIndex = Gep->getOperand(NumOperands - 1);
879 Value *GpPtr = Gep->getPointerOperand();
880 // If this GEP value is a consecutive pointer induction variable and all of
881 // the indices are constant then we know it is consecutive. We can
882 Phi = dyn_cast<PHINode>(GpPtr);
883 if (Phi && Inductions.count(Phi)) {
885 // Make sure that the pointer does not point to structs.
886 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
887 if (GepPtrType->getElementType()->isAggregateType())
890 // Make sure that all of the index operands are loop invariant.
891 for (unsigned i = 1; i < NumOperands; ++i)
892 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
895 InductionInfo II = Inductions[Phi];
896 if (IK_PtrInduction == II.IK)
898 else if (IK_ReversePtrInduction == II.IK)
902 // Check that all of the gep indices are uniform except for the last.
903 for (unsigned i = 0; i < NumOperands - 1; ++i)
904 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
907 // We can emit wide load/stores only if the last index is the induction
909 const SCEV *Last = SE->getSCEV(LastIndex);
910 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
911 const SCEV *Step = AR->getStepRecurrence(*SE);
913 // The memory is consecutive because the last index is consecutive
914 // and all other indices are loop invariant.
917 if (Step->isAllOnesValue())
924 bool LoopVectorizationLegality::isUniform(Value *V) {
925 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
928 InnerLoopVectorizer::VectorParts&
929 InnerLoopVectorizer::getVectorValue(Value *V) {
930 assert(V != Induction && "The new induction variable should not be used.");
931 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
933 // If we have this scalar in the map, return it.
935 return WidenMap.get(V);
937 // If this scalar is unknown, assume that it is a constant or that it is
938 // loop invariant. Broadcast V and save the value for future uses.
939 Value *B = getBroadcastInstrs(V);
940 return WidenMap.splat(V, B);
943 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
944 assert(Vec->getType()->isVectorTy() && "Invalid type");
945 SmallVector<Constant*, 8> ShuffleMask;
946 for (unsigned i = 0; i < VF; ++i)
947 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
949 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
950 ConstantVector::get(ShuffleMask),
955 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
956 LoopVectorizationLegality *Legal) {
957 // Attempt to issue a wide load.
958 LoadInst *LI = dyn_cast<LoadInst>(Instr);
959 StoreInst *SI = dyn_cast<StoreInst>(Instr);
961 assert((LI || SI) && "Invalid Load/Store instruction");
963 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
964 Type *DataTy = VectorType::get(ScalarDataTy, VF);
965 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
966 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
968 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
969 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
971 if (ScalarAllocatedSize != VectorElementSize)
972 return scalarizeInstruction(Instr);
974 // If the pointer is loop invariant or if it is non consecutive,
975 // scalarize the load.
976 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
977 bool Reverse = ConsecutiveStride < 0;
978 bool UniformLoad = LI && Legal->isUniform(Ptr);
979 if (!ConsecutiveStride || UniformLoad)
980 return scalarizeInstruction(Instr);
982 Constant *Zero = Builder.getInt32(0);
983 VectorParts &Entry = WidenMap.get(Instr);
985 // Handle consecutive loads/stores.
986 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
987 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
988 Value *PtrOperand = Gep->getPointerOperand();
989 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
990 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
992 // Create the new GEP with the new induction variable.
993 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
994 Gep2->setOperand(0, FirstBasePtr);
995 Gep2->setName("gep.indvar.base");
996 Ptr = Builder.Insert(Gep2);
998 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
999 OrigLoop) && "Base ptr must be invariant");
1001 // The last index does not have to be the induction. It can be
1002 // consecutive and be a function of the index. For example A[I+1];
1003 unsigned NumOperands = Gep->getNumOperands();
1005 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1006 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1007 Value *LastIndex = GEPParts[0];
1008 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1010 // Create the new GEP with the new induction variable.
1011 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1012 Gep2->setOperand(NumOperands - 1, LastIndex);
1013 Gep2->setName("gep.indvar.idx");
1014 Ptr = Builder.Insert(Gep2);
1016 // Use the induction element ptr.
1017 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1018 VectorParts &PtrVal = getVectorValue(Ptr);
1019 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1024 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1025 "We do not allow storing to uniform addresses");
1027 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1028 for (unsigned Part = 0; Part < UF; ++Part) {
1029 // Calculate the pointer for the specific unroll-part.
1030 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1033 // If we store to reverse consecutive memory locations then we need
1034 // to reverse the order of elements in the stored value.
1035 StoredVal[Part] = reverseVector(StoredVal[Part]);
1036 // If the address is consecutive but reversed, then the
1037 // wide store needs to start at the last vector element.
1038 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1039 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1042 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1043 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1047 for (unsigned Part = 0; Part < UF; ++Part) {
1048 // Calculate the pointer for the specific unroll-part.
1049 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1052 // If the address is consecutive but reversed, then the
1053 // wide store needs to start at the last vector element.
1054 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1055 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1058 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1059 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1060 cast<LoadInst>(LI)->setAlignment(Alignment);
1061 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1065 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1066 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1067 // Holds vector parameters or scalars, in case of uniform vals.
1068 SmallVector<VectorParts, 4> Params;
1070 // Find all of the vectorized parameters.
1071 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1072 Value *SrcOp = Instr->getOperand(op);
1074 // If we are accessing the old induction variable, use the new one.
1075 if (SrcOp == OldInduction) {
1076 Params.push_back(getVectorValue(SrcOp));
1080 // Try using previously calculated values.
1081 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1083 // If the src is an instruction that appeared earlier in the basic block
1084 // then it should already be vectorized.
1085 if (SrcInst && OrigLoop->contains(SrcInst)) {
1086 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1087 // The parameter is a vector value from earlier.
1088 Params.push_back(WidenMap.get(SrcInst));
1090 // The parameter is a scalar from outside the loop. Maybe even a constant.
1091 VectorParts Scalars;
1092 Scalars.append(UF, SrcOp);
1093 Params.push_back(Scalars);
1097 assert(Params.size() == Instr->getNumOperands() &&
1098 "Invalid number of operands");
1100 // Does this instruction return a value ?
1101 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1103 Value *UndefVec = IsVoidRetTy ? 0 :
1104 UndefValue::get(VectorType::get(Instr->getType(), VF));
1105 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1106 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1108 // For each vector unroll 'part':
1109 for (unsigned Part = 0; Part < UF; ++Part) {
1110 // For each scalar that we create:
1111 for (unsigned Width = 0; Width < VF; ++Width) {
1112 Instruction *Cloned = Instr->clone();
1114 Cloned->setName(Instr->getName() + ".cloned");
1115 // Replace the operands of the cloned instrucions with extracted scalars.
1116 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1117 Value *Op = Params[op][Part];
1118 // Param is a vector. Need to extract the right lane.
1119 if (Op->getType()->isVectorTy())
1120 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1121 Cloned->setOperand(op, Op);
1124 // Place the cloned scalar in the new loop.
1125 Builder.Insert(Cloned);
1127 // If the original scalar returns a value we need to place it in a vector
1128 // so that future users will be able to use it.
1130 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1131 Builder.getInt32(Width));
1137 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1139 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1140 Legal->getRuntimePointerCheck();
1142 if (!PtrRtCheck->Need)
1145 Instruction *MemoryRuntimeCheck = 0;
1146 unsigned NumPointers = PtrRtCheck->Pointers.size();
1147 SmallVector<Value* , 2> Starts;
1148 SmallVector<Value* , 2> Ends;
1150 SCEVExpander Exp(*SE, "induction");
1152 // Use this type for pointer arithmetic.
1153 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1155 for (unsigned i = 0; i < NumPointers; ++i) {
1156 Value *Ptr = PtrRtCheck->Pointers[i];
1157 const SCEV *Sc = SE->getSCEV(Ptr);
1159 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1160 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1162 Starts.push_back(Ptr);
1163 Ends.push_back(Ptr);
1165 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1167 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1168 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1169 Starts.push_back(Start);
1170 Ends.push_back(End);
1174 IRBuilder<> ChkBuilder(Loc);
1176 for (unsigned i = 0; i < NumPointers; ++i) {
1177 for (unsigned j = i+1; j < NumPointers; ++j) {
1178 // No need to check if two readonly pointers intersect.
1179 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1182 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1183 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1184 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1185 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1187 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1188 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1189 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1190 if (MemoryRuntimeCheck)
1191 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1194 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1198 return MemoryRuntimeCheck;
1202 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1204 In this function we generate a new loop. The new loop will contain
1205 the vectorized instructions while the old loop will continue to run the
1208 [ ] <-- vector loop bypass (may consist of multiple blocks).
1211 | [ ] <-- vector pre header.
1215 | [ ]_| <-- vector loop.
1218 >[ ] <--- middle-block.
1221 | [ ] <--- new preheader.
1225 | [ ]_| <-- old scalar loop to handle remainder.
1228 >[ ] <-- exit block.
1232 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1233 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1234 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1235 assert(ExitBlock && "Must have an exit block");
1237 // Mark the old scalar loop with metadata that tells us not to vectorize this
1238 // loop again if we run into it.
1239 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None);
1240 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD);
1242 // Some loops have a single integer induction variable, while other loops
1243 // don't. One example is c++ iterators that often have multiple pointer
1244 // induction variables. In the code below we also support a case where we
1245 // don't have a single induction variable.
1246 OldInduction = Legal->getInduction();
1247 Type *IdxTy = OldInduction ? OldInduction->getType() :
1248 DL->getIntPtrType(SE->getContext());
1250 // Find the loop boundaries.
1251 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1252 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1254 // Get the total trip count from the count by adding 1.
1255 ExitCount = SE->getAddExpr(ExitCount,
1256 SE->getConstant(ExitCount->getType(), 1));
1258 // Expand the trip count and place the new instructions in the preheader.
1259 // Notice that the pre-header does not change, only the loop body.
1260 SCEVExpander Exp(*SE, "induction");
1262 // Count holds the overall loop count (N).
1263 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1264 BypassBlock->getTerminator());
1266 // The loop index does not have to start at Zero. Find the original start
1267 // value from the induction PHI node. If we don't have an induction variable
1268 // then we know that it starts at zero.
1269 Value *StartIdx = OldInduction ?
1270 OldInduction->getIncomingValueForBlock(BypassBlock):
1271 ConstantInt::get(IdxTy, 0);
1273 assert(BypassBlock && "Invalid loop structure");
1274 LoopBypassBlocks.push_back(BypassBlock);
1276 // Split the single block loop into the two loop structure described above.
1277 BasicBlock *VectorPH =
1278 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1279 BasicBlock *VecBody =
1280 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1281 BasicBlock *MiddleBlock =
1282 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1283 BasicBlock *ScalarPH =
1284 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1286 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1288 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1290 // Generate the induction variable.
1291 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1292 // The loop step is equal to the vectorization factor (num of SIMD elements)
1293 // times the unroll factor (num of SIMD instructions).
1294 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1296 // This is the IR builder that we use to add all of the logic for bypassing
1297 // the new vector loop.
1298 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1300 // We may need to extend the index in case there is a type mismatch.
1301 // We know that the count starts at zero and does not overflow.
1302 if (Count->getType() != IdxTy) {
1303 // The exit count can be of pointer type. Convert it to the correct
1305 if (ExitCount->getType()->isPointerTy())
1306 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1308 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1311 // Add the start index to the loop count to get the new end index.
1312 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1314 // Now we need to generate the expression for N - (N % VF), which is
1315 // the part that the vectorized body will execute.
1316 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1317 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1318 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1319 "end.idx.rnd.down");
1321 // Now, compare the new count to zero. If it is zero skip the vector loop and
1322 // jump to the scalar loop.
1323 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1326 BasicBlock *LastBypassBlock = BypassBlock;
1328 // Generate the code that checks in runtime if arrays overlap. We put the
1329 // checks into a separate block to make the more common case of few elements
1331 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1332 BypassBlock->getTerminator());
1333 if (MemRuntimeCheck) {
1334 // Create a new block containing the memory check.
1335 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1337 LoopBypassBlocks.push_back(CheckBlock);
1339 // Replace the branch into the memory check block with a conditional branch
1340 // for the "few elements case".
1341 Instruction *OldTerm = BypassBlock->getTerminator();
1342 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1343 OldTerm->eraseFromParent();
1345 Cmp = MemRuntimeCheck;
1346 LastBypassBlock = CheckBlock;
1349 LastBypassBlock->getTerminator()->eraseFromParent();
1350 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1353 // We are going to resume the execution of the scalar loop.
1354 // Go over all of the induction variables that we found and fix the
1355 // PHIs that are left in the scalar version of the loop.
1356 // The starting values of PHI nodes depend on the counter of the last
1357 // iteration in the vectorized loop.
1358 // If we come from a bypass edge then we need to start from the original
1361 // This variable saves the new starting index for the scalar loop.
1362 PHINode *ResumeIndex = 0;
1363 LoopVectorizationLegality::InductionList::iterator I, E;
1364 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1365 for (I = List->begin(), E = List->end(); I != E; ++I) {
1366 PHINode *OrigPhi = I->first;
1367 LoopVectorizationLegality::InductionInfo II = I->second;
1368 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1369 MiddleBlock->getTerminator());
1370 Value *EndValue = 0;
1372 case LoopVectorizationLegality::IK_NoInduction:
1373 llvm_unreachable("Unknown induction");
1374 case LoopVectorizationLegality::IK_IntInduction: {
1375 // Handle the integer induction counter:
1376 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1377 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1378 // We know what the end value is.
1379 EndValue = IdxEndRoundDown;
1380 // We also know which PHI node holds it.
1381 ResumeIndex = ResumeVal;
1384 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1385 // Convert the CountRoundDown variable to the PHI size.
1386 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1387 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1388 Value *CRD = CountRoundDown;
1389 if (CRDSize > IISize)
1390 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1391 II.StartValue->getType(), "tr.crd",
1392 LoopBypassBlocks.back()->getTerminator());
1393 else if (CRDSize < IISize)
1394 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1395 II.StartValue->getType(),
1397 LoopBypassBlocks.back()->getTerminator());
1398 // Handle reverse integer induction counter:
1400 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1401 LoopBypassBlocks.back()->getTerminator());
1404 case LoopVectorizationLegality::IK_PtrInduction: {
1405 // For pointer induction variables, calculate the offset using
1408 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end",
1409 LoopBypassBlocks.back()->getTerminator());
1412 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1413 // The value at the end of the loop for the reverse pointer is calculated
1414 // by creating a GEP with a negative index starting from the start value.
1415 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1416 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown,
1418 LoopBypassBlocks.back()->getTerminator());
1419 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx,
1421 LoopBypassBlocks.back()->getTerminator());
1426 // The new PHI merges the original incoming value, in case of a bypass,
1427 // or the value at the end of the vectorized loop.
1428 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1429 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1430 ResumeVal->addIncoming(EndValue, VecBody);
1432 // Fix the scalar body counter (PHI node).
1433 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1434 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1437 // If we are generating a new induction variable then we also need to
1438 // generate the code that calculates the exit value. This value is not
1439 // simply the end of the counter because we may skip the vectorized body
1440 // in case of a runtime check.
1442 assert(!ResumeIndex && "Unexpected resume value found");
1443 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1444 MiddleBlock->getTerminator());
1445 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1446 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1447 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1450 // Make sure that we found the index where scalar loop needs to continue.
1451 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1452 "Invalid resume Index");
1454 // Add a check in the middle block to see if we have completed
1455 // all of the iterations in the first vector loop.
1456 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1457 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1458 ResumeIndex, "cmp.n",
1459 MiddleBlock->getTerminator());
1461 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1462 // Remove the old terminator.
1463 MiddleBlock->getTerminator()->eraseFromParent();
1465 // Create i+1 and fill the PHINode.
1466 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1467 Induction->addIncoming(StartIdx, VectorPH);
1468 Induction->addIncoming(NextIdx, VecBody);
1469 // Create the compare.
1470 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1471 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1473 // Now we have two terminators. Remove the old one from the block.
1474 VecBody->getTerminator()->eraseFromParent();
1476 // Get ready to start creating new instructions into the vectorized body.
1477 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1479 // Create and register the new vector loop.
1480 Loop* Lp = new Loop();
1481 Loop *ParentLoop = OrigLoop->getParentLoop();
1483 // Insert the new loop into the loop nest and register the new basic blocks.
1485 ParentLoop->addChildLoop(Lp);
1486 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1487 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1488 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1489 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1490 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1492 LI->addTopLevelLoop(Lp);
1495 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1498 LoopVectorPreHeader = VectorPH;
1499 LoopScalarPreHeader = ScalarPH;
1500 LoopMiddleBlock = MiddleBlock;
1501 LoopExitBlock = ExitBlock;
1502 LoopVectorBody = VecBody;
1503 LoopScalarBody = OldBasicBlock;
1506 /// This function returns the identity element (or neutral element) for
1507 /// the operation K.
1509 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1514 // Adding, Xoring, Oring zero to a number does not change it.
1515 return ConstantInt::get(Tp, 0);
1516 case RK_IntegerMult:
1517 // Multiplying a number by 1 does not change it.
1518 return ConstantInt::get(Tp, 1);
1520 // AND-ing a number with an all-1 value does not change it.
1521 return ConstantInt::get(Tp, -1, true);
1523 // Multiplying a number by 1 does not change it.
1524 return ConstantFP::get(Tp, 1.0L);
1526 // Adding zero to a number does not change it.
1527 return ConstantFP::get(Tp, 0.0L);
1529 llvm_unreachable("Unknown reduction kind");
1533 static Intrinsic::ID
1534 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1535 // If we have an intrinsic call, check if it is trivially vectorizable.
1536 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1537 switch (II->getIntrinsicID()) {
1538 case Intrinsic::sqrt:
1539 case Intrinsic::sin:
1540 case Intrinsic::cos:
1541 case Intrinsic::exp:
1542 case Intrinsic::exp2:
1543 case Intrinsic::log:
1544 case Intrinsic::log10:
1545 case Intrinsic::log2:
1546 case Intrinsic::fabs:
1547 case Intrinsic::floor:
1548 case Intrinsic::ceil:
1549 case Intrinsic::trunc:
1550 case Intrinsic::rint:
1551 case Intrinsic::nearbyint:
1552 case Intrinsic::pow:
1553 case Intrinsic::fma:
1554 case Intrinsic::fmuladd:
1555 return II->getIntrinsicID();
1557 return Intrinsic::not_intrinsic;
1562 return Intrinsic::not_intrinsic;
1565 Function *F = CI->getCalledFunction();
1566 // We're going to make assumptions on the semantics of the functions, check
1567 // that the target knows that it's available in this environment.
1568 if (!F || !TLI->getLibFunc(F->getName(), Func))
1569 return Intrinsic::not_intrinsic;
1571 // Otherwise check if we have a call to a function that can be turned into a
1572 // vector intrinsic.
1579 return Intrinsic::sin;
1583 return Intrinsic::cos;
1587 return Intrinsic::exp;
1589 case LibFunc::exp2f:
1590 case LibFunc::exp2l:
1591 return Intrinsic::exp2;
1595 return Intrinsic::log;
1596 case LibFunc::log10:
1597 case LibFunc::log10f:
1598 case LibFunc::log10l:
1599 return Intrinsic::log10;
1601 case LibFunc::log2f:
1602 case LibFunc::log2l:
1603 return Intrinsic::log2;
1605 case LibFunc::fabsf:
1606 case LibFunc::fabsl:
1607 return Intrinsic::fabs;
1608 case LibFunc::floor:
1609 case LibFunc::floorf:
1610 case LibFunc::floorl:
1611 return Intrinsic::floor;
1613 case LibFunc::ceilf:
1614 case LibFunc::ceill:
1615 return Intrinsic::ceil;
1616 case LibFunc::trunc:
1617 case LibFunc::truncf:
1618 case LibFunc::truncl:
1619 return Intrinsic::trunc;
1621 case LibFunc::rintf:
1622 case LibFunc::rintl:
1623 return Intrinsic::rint;
1624 case LibFunc::nearbyint:
1625 case LibFunc::nearbyintf:
1626 case LibFunc::nearbyintl:
1627 return Intrinsic::nearbyint;
1631 return Intrinsic::pow;
1634 return Intrinsic::not_intrinsic;
1637 /// This function translates the reduction kind to an LLVM binary operator.
1639 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1641 case LoopVectorizationLegality::RK_IntegerAdd:
1642 return Instruction::Add;
1643 case LoopVectorizationLegality::RK_IntegerMult:
1644 return Instruction::Mul;
1645 case LoopVectorizationLegality::RK_IntegerOr:
1646 return Instruction::Or;
1647 case LoopVectorizationLegality::RK_IntegerAnd:
1648 return Instruction::And;
1649 case LoopVectorizationLegality::RK_IntegerXor:
1650 return Instruction::Xor;
1651 case LoopVectorizationLegality::RK_FloatMult:
1652 return Instruction::FMul;
1653 case LoopVectorizationLegality::RK_FloatAdd:
1654 return Instruction::FAdd;
1655 case LoopVectorizationLegality::RK_IntegerMinMax:
1656 return Instruction::ICmp;
1657 case LoopVectorizationLegality::RK_FloatMinMax:
1658 return Instruction::FCmp;
1660 llvm_unreachable("Unknown reduction operation");
1664 Value *createMinMaxOp(IRBuilder<> &Builder,
1665 LoopVectorizationLegality::MinMaxReductionKind RK,
1668 CmpInst::Predicate P = CmpInst::ICMP_NE;
1671 llvm_unreachable("Unknown min/max reduction kind");
1672 case LoopVectorizationLegality::MRK_UIntMin:
1673 P = CmpInst::ICMP_ULT;
1675 case LoopVectorizationLegality::MRK_UIntMax:
1676 P = CmpInst::ICMP_UGT;
1678 case LoopVectorizationLegality::MRK_SIntMin:
1679 P = CmpInst::ICMP_SLT;
1681 case LoopVectorizationLegality::MRK_SIntMax:
1682 P = CmpInst::ICMP_SGT;
1684 case LoopVectorizationLegality::MRK_FloatMin:
1685 P = CmpInst::FCMP_OLT;
1687 case LoopVectorizationLegality::MRK_FloatMax:
1688 P = CmpInst::FCMP_OGT;
1693 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1694 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1696 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1698 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1703 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1704 //===------------------------------------------------===//
1706 // Notice: any optimization or new instruction that go
1707 // into the code below should be also be implemented in
1710 //===------------------------------------------------===//
1711 Constant *Zero = Builder.getInt32(0);
1713 // In order to support reduction variables we need to be able to vectorize
1714 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1715 // stages. First, we create a new vector PHI node with no incoming edges.
1716 // We use this value when we vectorize all of the instructions that use the
1717 // PHI. Next, after all of the instructions in the block are complete we
1718 // add the new incoming edges to the PHI. At this point all of the
1719 // instructions in the basic block are vectorized, so we can use them to
1720 // construct the PHI.
1721 PhiVector RdxPHIsToFix;
1723 // Scan the loop in a topological order to ensure that defs are vectorized
1725 LoopBlocksDFS DFS(OrigLoop);
1728 // Vectorize all of the blocks in the original loop.
1729 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1730 be = DFS.endRPO(); bb != be; ++bb)
1731 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1733 // At this point every instruction in the original loop is widened to
1734 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1735 // that we vectorized. The PHI nodes are currently empty because we did
1736 // not want to introduce cycles. Notice that the remaining PHI nodes
1737 // that we need to fix are reduction variables.
1739 // Create the 'reduced' values for each of the induction vars.
1740 // The reduced values are the vector values that we scalarize and combine
1741 // after the loop is finished.
1742 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1744 PHINode *RdxPhi = *it;
1745 assert(RdxPhi && "Unable to recover vectorized PHI");
1747 // Find the reduction variable descriptor.
1748 assert(Legal->getReductionVars()->count(RdxPhi) &&
1749 "Unable to find the reduction variable");
1750 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1751 (*Legal->getReductionVars())[RdxPhi];
1753 // We need to generate a reduction vector from the incoming scalar.
1754 // To do so, we need to generate the 'identity' vector and overide
1755 // one of the elements with the incoming scalar reduction. We need
1756 // to do it in the vector-loop preheader.
1757 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1759 // This is the vector-clone of the value that leaves the loop.
1760 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1761 Type *VecTy = VectorExit[0]->getType();
1763 // Find the reduction identity variable. Zero for addition, or, xor,
1764 // one for multiplication, -1 for And.
1767 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
1768 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
1769 // MinMax reduction have the start value as their identify.
1770 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
1774 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1775 VecTy->getScalarType());
1776 Identity = ConstantVector::getSplat(VF, Iden);
1778 // This vector is the Identity vector where the first element is the
1779 // incoming scalar reduction.
1780 VectorStart = Builder.CreateInsertElement(Identity,
1781 RdxDesc.StartValue, Zero);
1784 // Fix the vector-loop phi.
1785 // We created the induction variable so we know that the
1786 // preheader is the first entry.
1787 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1789 // Reductions do not have to start at zero. They can start with
1790 // any loop invariant values.
1791 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1792 BasicBlock *Latch = OrigLoop->getLoopLatch();
1793 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1794 VectorParts &Val = getVectorValue(LoopVal);
1795 for (unsigned part = 0; part < UF; ++part) {
1796 // Make sure to add the reduction stat value only to the
1797 // first unroll part.
1798 Value *StartVal = (part == 0) ? VectorStart : Identity;
1799 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1800 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1803 // Before each round, move the insertion point right between
1804 // the PHIs and the values we are going to write.
1805 // This allows us to write both PHINodes and the extractelement
1807 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1809 VectorParts RdxParts;
1810 for (unsigned part = 0; part < UF; ++part) {
1811 // This PHINode contains the vectorized reduction variable, or
1812 // the initial value vector, if we bypass the vector loop.
1813 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1814 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1815 Value *StartVal = (part == 0) ? VectorStart : Identity;
1816 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1817 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1818 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1819 RdxParts.push_back(NewPhi);
1822 // Reduce all of the unrolled parts into a single vector.
1823 Value *ReducedPartRdx = RdxParts[0];
1824 unsigned Op = getReductionBinOp(RdxDesc.Kind);
1825 for (unsigned part = 1; part < UF; ++part) {
1826 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1827 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
1828 RdxParts[part], ReducedPartRdx,
1831 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
1832 ReducedPartRdx, RdxParts[part]);
1835 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1836 // and vector ops, reducing the set of values being computed by half each
1838 assert(isPowerOf2_32(VF) &&
1839 "Reduction emission only supported for pow2 vectors!");
1840 Value *TmpVec = ReducedPartRdx;
1841 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1842 for (unsigned i = VF; i != 1; i >>= 1) {
1843 // Move the upper half of the vector to the lower half.
1844 for (unsigned j = 0; j != i/2; ++j)
1845 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1847 // Fill the rest of the mask with undef.
1848 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1849 UndefValue::get(Builder.getInt32Ty()));
1852 Builder.CreateShuffleVector(TmpVec,
1853 UndefValue::get(TmpVec->getType()),
1854 ConstantVector::get(ShuffleMask),
1857 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1858 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1861 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
1864 // The result is in the first element of the vector.
1865 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1867 // Now, we need to fix the users of the reduction variable
1868 // inside and outside of the scalar remainder loop.
1869 // We know that the loop is in LCSSA form. We need to update the
1870 // PHI nodes in the exit blocks.
1871 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1872 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1873 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1874 if (!LCSSAPhi) continue;
1876 // All PHINodes need to have a single entry edge, or two if
1877 // we already fixed them.
1878 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1880 // We found our reduction value exit-PHI. Update it with the
1881 // incoming bypass edge.
1882 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1883 // Add an edge coming from the bypass.
1884 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1887 }// end of the LCSSA phi scan.
1889 // Fix the scalar loop reduction variable with the incoming reduction sum
1890 // from the vector body and from the backedge value.
1891 int IncomingEdgeBlockIdx =
1892 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1893 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1894 // Pick the other block.
1895 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1896 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1897 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1898 }// end of for each redux variable.
1900 // The Loop exit block may have single value PHI nodes where the incoming
1901 // value is 'undef'. While vectorizing we only handled real values that
1902 // were defined inside the loop. Here we handle the 'undef case'.
1904 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1905 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1906 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1907 if (!LCSSAPhi) continue;
1908 if (LCSSAPhi->getNumIncomingValues() == 1)
1909 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1914 InnerLoopVectorizer::VectorParts
1915 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1916 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1919 VectorParts SrcMask = createBlockInMask(Src);
1921 // The terminator has to be a branch inst!
1922 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1923 assert(BI && "Unexpected terminator found");
1925 if (BI->isConditional()) {
1926 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1928 if (BI->getSuccessor(0) != Dst)
1929 for (unsigned part = 0; part < UF; ++part)
1930 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1932 for (unsigned part = 0; part < UF; ++part)
1933 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1940 InnerLoopVectorizer::VectorParts
1941 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1942 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1944 // Loop incoming mask is all-one.
1945 if (OrigLoop->getHeader() == BB) {
1946 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1947 return getVectorValue(C);
1950 // This is the block mask. We OR all incoming edges, and with zero.
1951 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1952 VectorParts BlockMask = getVectorValue(Zero);
1955 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1956 VectorParts EM = createEdgeMask(*it, BB);
1957 for (unsigned part = 0; part < UF; ++part)
1958 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1965 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1966 BasicBlock *BB, PhiVector *PV) {
1967 // For each instruction in the old loop.
1968 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1969 VectorParts &Entry = WidenMap.get(it);
1970 switch (it->getOpcode()) {
1971 case Instruction::Br:
1972 // Nothing to do for PHIs and BR, since we already took care of the
1973 // loop control flow instructions.
1975 case Instruction::PHI:{
1976 PHINode* P = cast<PHINode>(it);
1977 // Handle reduction variables:
1978 if (Legal->getReductionVars()->count(P)) {
1979 for (unsigned part = 0; part < UF; ++part) {
1980 // This is phase one of vectorizing PHIs.
1981 Type *VecTy = VectorType::get(it->getType(), VF);
1982 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1983 LoopVectorBody-> getFirstInsertionPt());
1989 // Check for PHI nodes that are lowered to vector selects.
1990 if (P->getParent() != OrigLoop->getHeader()) {
1991 // We know that all PHIs in non header blocks are converted into
1992 // selects, so we don't have to worry about the insertion order and we
1993 // can just use the builder.
1994 // At this point we generate the predication tree. There may be
1995 // duplications since this is a simple recursive scan, but future
1996 // optimizations will clean it up.
1998 unsigned NumIncoming = P->getNumIncomingValues();
1999 assert(NumIncoming > 1 && "Invalid PHI");
2001 // Generate a sequence of selects of the form:
2002 // SELECT(Mask3, In3,
2003 // SELECT(Mask2, In2,
2005 for (unsigned In = 0; In < NumIncoming; In++) {
2006 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2008 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2010 for (unsigned part = 0; part < UF; ++part) {
2011 // We don't need to 'select' the first PHI operand because it is
2012 // the default value if all of the other masks don't match.
2014 Entry[part] = In0[part];
2016 // Select between the current value and the previous incoming edge
2017 // based on the incoming mask.
2018 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2019 Entry[part], "predphi");
2025 // This PHINode must be an induction variable.
2026 // Make sure that we know about it.
2027 assert(Legal->getInductionVars()->count(P) &&
2028 "Not an induction variable");
2030 LoopVectorizationLegality::InductionInfo II =
2031 Legal->getInductionVars()->lookup(P);
2034 case LoopVectorizationLegality::IK_NoInduction:
2035 llvm_unreachable("Unknown induction");
2036 case LoopVectorizationLegality::IK_IntInduction: {
2037 assert(P == OldInduction && "Unexpected PHI");
2038 Value *Broadcasted = getBroadcastInstrs(Induction);
2039 // After broadcasting the induction variable we need to make the
2040 // vector consecutive by adding 0, 1, 2 ...
2041 for (unsigned part = 0; part < UF; ++part)
2042 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2045 case LoopVectorizationLegality::IK_ReverseIntInduction:
2046 case LoopVectorizationLegality::IK_PtrInduction:
2047 case LoopVectorizationLegality::IK_ReversePtrInduction:
2048 // Handle reverse integer and pointer inductions.
2049 Value *StartIdx = 0;
2050 // If we have a single integer induction variable then use it.
2051 // Otherwise, start counting at zero.
2053 LoopVectorizationLegality::InductionInfo OldII =
2054 Legal->getInductionVars()->lookup(OldInduction);
2055 StartIdx = OldII.StartValue;
2057 StartIdx = ConstantInt::get(Induction->getType(), 0);
2059 // This is the normalized GEP that starts counting at zero.
2060 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2063 // Handle the reverse integer induction variable case.
2064 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2065 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2066 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2068 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2071 // This is a new value so do not hoist it out.
2072 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2073 // After broadcasting the induction variable we need to make the
2074 // vector consecutive by adding ... -3, -2, -1, 0.
2075 for (unsigned part = 0; part < UF; ++part)
2076 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2081 // Handle the pointer induction variable case.
2082 assert(P->getType()->isPointerTy() && "Unexpected type.");
2084 // Is this a reverse induction ptr or a consecutive induction ptr.
2085 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2088 // This is the vector of results. Notice that we don't generate
2089 // vector geps because scalar geps result in better code.
2090 for (unsigned part = 0; part < UF; ++part) {
2091 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2092 for (unsigned int i = 0; i < VF; ++i) {
2093 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2094 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2097 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2099 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2101 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2103 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2104 Builder.getInt32(i),
2107 Entry[part] = VecVal;
2114 case Instruction::Add:
2115 case Instruction::FAdd:
2116 case Instruction::Sub:
2117 case Instruction::FSub:
2118 case Instruction::Mul:
2119 case Instruction::FMul:
2120 case Instruction::UDiv:
2121 case Instruction::SDiv:
2122 case Instruction::FDiv:
2123 case Instruction::URem:
2124 case Instruction::SRem:
2125 case Instruction::FRem:
2126 case Instruction::Shl:
2127 case Instruction::LShr:
2128 case Instruction::AShr:
2129 case Instruction::And:
2130 case Instruction::Or:
2131 case Instruction::Xor: {
2132 // Just widen binops.
2133 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2134 VectorParts &A = getVectorValue(it->getOperand(0));
2135 VectorParts &B = getVectorValue(it->getOperand(1));
2137 // Use this vector value for all users of the original instruction.
2138 for (unsigned Part = 0; Part < UF; ++Part) {
2139 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2141 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2142 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2143 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2144 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2145 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2147 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2148 VecOp->setIsExact(BinOp->isExact());
2154 case Instruction::Select: {
2156 // If the selector is loop invariant we can create a select
2157 // instruction with a scalar condition. Otherwise, use vector-select.
2158 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2161 // The condition can be loop invariant but still defined inside the
2162 // loop. This means that we can't just use the original 'cond' value.
2163 // We have to take the 'vectorized' value and pick the first lane.
2164 // Instcombine will make this a no-op.
2165 VectorParts &Cond = getVectorValue(it->getOperand(0));
2166 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2167 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2168 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2169 Builder.getInt32(0));
2170 for (unsigned Part = 0; Part < UF; ++Part) {
2171 Entry[Part] = Builder.CreateSelect(
2172 InvariantCond ? ScalarCond : Cond[Part],
2179 case Instruction::ICmp:
2180 case Instruction::FCmp: {
2181 // Widen compares. Generate vector compares.
2182 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2183 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2184 VectorParts &A = getVectorValue(it->getOperand(0));
2185 VectorParts &B = getVectorValue(it->getOperand(1));
2186 for (unsigned Part = 0; Part < UF; ++Part) {
2189 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2191 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2197 case Instruction::Store:
2198 case Instruction::Load:
2199 vectorizeMemoryInstruction(it, Legal);
2201 case Instruction::ZExt:
2202 case Instruction::SExt:
2203 case Instruction::FPToUI:
2204 case Instruction::FPToSI:
2205 case Instruction::FPExt:
2206 case Instruction::PtrToInt:
2207 case Instruction::IntToPtr:
2208 case Instruction::SIToFP:
2209 case Instruction::UIToFP:
2210 case Instruction::Trunc:
2211 case Instruction::FPTrunc:
2212 case Instruction::BitCast: {
2213 CastInst *CI = dyn_cast<CastInst>(it);
2214 /// Optimize the special case where the source is the induction
2215 /// variable. Notice that we can only optimize the 'trunc' case
2216 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2217 /// c. other casts depend on pointer size.
2218 if (CI->getOperand(0) == OldInduction &&
2219 it->getOpcode() == Instruction::Trunc) {
2220 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2222 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2223 for (unsigned Part = 0; Part < UF; ++Part)
2224 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2227 /// Vectorize casts.
2228 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2230 VectorParts &A = getVectorValue(it->getOperand(0));
2231 for (unsigned Part = 0; Part < UF; ++Part)
2232 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2236 case Instruction::Call: {
2237 // Ignore dbg intrinsics.
2238 if (isa<DbgInfoIntrinsic>(it))
2241 Module *M = BB->getParent()->getParent();
2242 CallInst *CI = cast<CallInst>(it);
2243 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2244 assert(ID && "Not an intrinsic call!");
2245 for (unsigned Part = 0; Part < UF; ++Part) {
2246 SmallVector<Value*, 4> Args;
2247 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2248 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2249 Args.push_back(Arg[Part]);
2251 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2252 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2253 Entry[Part] = Builder.CreateCall(F, Args);
2259 // All other instructions are unsupported. Scalarize them.
2260 scalarizeInstruction(it);
2263 }// end of for_each instr.
2266 void InnerLoopVectorizer::updateAnalysis() {
2267 // Forget the original basic block.
2268 SE->forgetLoop(OrigLoop);
2270 // Update the dominator tree information.
2271 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2272 "Entry does not dominate exit.");
2274 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2275 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2276 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2277 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2278 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2279 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2280 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2281 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2283 DEBUG(DT->verifyAnalysis());
2286 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2287 if (!EnableIfConversion)
2290 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2291 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2293 // Collect the blocks that need predication.
2294 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2295 BasicBlock *BB = LoopBlocks[i];
2297 // We don't support switch statements inside loops.
2298 if (!isa<BranchInst>(BB->getTerminator()))
2301 // We must be able to predicate all blocks that need to be predicated.
2302 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2306 // We can if-convert this loop.
2310 bool LoopVectorizationLegality::canVectorize() {
2311 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2312 // be canonicalized.
2313 if (!TheLoop->getLoopPreheader())
2316 // We can only vectorize innermost loops.
2317 if (TheLoop->getSubLoopsVector().size())
2320 // We must have a single backedge.
2321 if (TheLoop->getNumBackEdges() != 1)
2324 // We must have a single exiting block.
2325 if (!TheLoop->getExitingBlock())
2328 unsigned NumBlocks = TheLoop->getNumBlocks();
2330 // Check if we can if-convert non single-bb loops.
2331 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2332 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2336 // We need to have a loop header.
2337 BasicBlock *Latch = TheLoop->getLoopLatch();
2338 DEBUG(dbgs() << "LV: Found a loop: " <<
2339 TheLoop->getHeader()->getName() << "\n");
2341 // ScalarEvolution needs to be able to find the exit count.
2342 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2343 if (ExitCount == SE->getCouldNotCompute()) {
2344 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2348 // Do not loop-vectorize loops with a tiny trip count.
2349 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2350 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2351 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2352 "This loop is not worth vectorizing.\n");
2356 // Check if we can vectorize the instructions and CFG in this loop.
2357 if (!canVectorizeInstrs()) {
2358 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2362 // Go over each instruction and look at memory deps.
2363 if (!canVectorizeMemory()) {
2364 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2368 // Collect all of the variables that remain uniform after vectorization.
2369 collectLoopUniforms();
2371 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2372 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2375 // Okay! We can vectorize. At this point we don't have any other mem analysis
2376 // which may limit our maximum vectorization factor, so just return true with
2381 /// \brief Check that the instruction has outside loop users and is not an
2382 /// identified reduction variable.
2383 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2384 SmallPtrSet<Value *, 4> &Reductions) {
2385 // Reduction instructions are allowed to have exit users. All other
2386 // instructions must not have external users.
2387 if (!Reductions.count(Inst))
2388 //Check that all of the users of the loop are inside the BB.
2389 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2391 Instruction *U = cast<Instruction>(*I);
2392 // This user may be a reduction exit value.
2393 if (!TheLoop->contains(U)) {
2394 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2401 bool LoopVectorizationLegality::canVectorizeInstrs() {
2402 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2403 BasicBlock *Header = TheLoop->getHeader();
2405 // If we marked the scalar loop as "already vectorized" then no need
2406 // to vectorize it again.
2407 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) {
2408 DEBUG(dbgs() << "LV: This loop was vectorized before\n");
2412 // Look for the attribute signaling the absence of NaNs.
2413 Function &F = *Header->getParent();
2414 if (F.hasFnAttribute("no-nans-fp-math"))
2415 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2416 AttributeSet::FunctionIndex,
2417 "no-nans-fp-math").getValueAsString() == "true";
2419 // For each block in the loop.
2420 for (Loop::block_iterator bb = TheLoop->block_begin(),
2421 be = TheLoop->block_end(); bb != be; ++bb) {
2423 // Scan the instructions in the block and look for hazards.
2424 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2427 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2428 // Check that this PHI type is allowed.
2429 if (!Phi->getType()->isIntegerTy() &&
2430 !Phi->getType()->isFloatingPointTy() &&
2431 !Phi->getType()->isPointerTy()) {
2432 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2436 // If this PHINode is not in the header block, then we know that we
2437 // can convert it to select during if-conversion. No need to check if
2438 // the PHIs in this block are induction or reduction variables.
2439 if (*bb != Header) {
2440 // Check that this instruction has no outside users or is an
2441 // identified reduction value with an outside user.
2442 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2447 // We only allow if-converted PHIs with more than two incoming values.
2448 if (Phi->getNumIncomingValues() != 2) {
2449 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2453 // This is the value coming from the preheader.
2454 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2455 // Check if this is an induction variable.
2456 InductionKind IK = isInductionVariable(Phi);
2458 if (IK_NoInduction != IK) {
2459 // Int inductions are special because we only allow one IV.
2460 if (IK == IK_IntInduction) {
2462 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2468 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2469 Inductions[Phi] = InductionInfo(StartValue, IK);
2473 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2474 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2477 if (AddReductionVar(Phi, RK_IntegerMult)) {
2478 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2481 if (AddReductionVar(Phi, RK_IntegerOr)) {
2482 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2485 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2486 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2489 if (AddReductionVar(Phi, RK_IntegerXor)) {
2490 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2493 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2494 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2497 if (AddReductionVar(Phi, RK_FloatMult)) {
2498 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2501 if (AddReductionVar(Phi, RK_FloatAdd)) {
2502 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2505 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2506 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2510 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2512 }// end of PHI handling
2514 // We still don't handle functions. However, we can ignore dbg intrinsic
2515 // calls and we do handle certain intrinsic and libm functions.
2516 CallInst *CI = dyn_cast<CallInst>(it);
2517 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2518 DEBUG(dbgs() << "LV: Found a call site.\n");
2522 // Check that the instruction return type is vectorizable.
2523 if (!VectorType::isValidElementType(it->getType()) &&
2524 !it->getType()->isVoidTy()) {
2525 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2529 // Check that the stored type is vectorizable.
2530 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2531 Type *T = ST->getValueOperand()->getType();
2532 if (!VectorType::isValidElementType(T))
2536 // Reduction instructions are allowed to have exit users.
2537 // All other instructions must not have external users.
2538 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2546 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2547 assert(getInductionVars()->size() && "No induction variables");
2553 void LoopVectorizationLegality::collectLoopUniforms() {
2554 // We now know that the loop is vectorizable!
2555 // Collect variables that will remain uniform after vectorization.
2556 std::vector<Value*> Worklist;
2557 BasicBlock *Latch = TheLoop->getLoopLatch();
2559 // Start with the conditional branch and walk up the block.
2560 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2562 while (Worklist.size()) {
2563 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2564 Worklist.pop_back();
2566 // Look at instructions inside this loop.
2567 // Stop when reaching PHI nodes.
2568 // TODO: we need to follow values all over the loop, not only in this block.
2569 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2572 // This is a known uniform.
2575 // Insert all operands.
2576 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2577 Worklist.push_back(I->getOperand(i));
2582 AliasAnalysis::Location
2583 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2584 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2585 return AA->getLocation(Store);
2586 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2587 return AA->getLocation(Load);
2589 llvm_unreachable("Should be either load or store instruction");
2593 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2596 AliasMultiMap& WriteObjects,
2597 unsigned MaxByteWidth) {
2599 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2601 std::vector<Instruction*>::iterator
2602 it = WriteObjects[Object].begin(),
2603 end = WriteObjects[Object].end();
2605 for (; it != end; ++it) {
2606 Instruction* I = *it;
2610 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2611 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2612 ThatLoc.getWithNewSize(MaxByteWidth)))
2618 bool LoopVectorizationLegality::canVectorizeMemory() {
2620 typedef SmallVector<Value*, 16> ValueVector;
2621 typedef SmallPtrSet<Value*, 16> ValueSet;
2622 // Holds the Load and Store *instructions*.
2625 PtrRtCheck.Pointers.clear();
2626 PtrRtCheck.Need = false;
2628 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2631 for (Loop::block_iterator bb = TheLoop->block_begin(),
2632 be = TheLoop->block_end(); bb != be; ++bb) {
2634 // Scan the BB and collect legal loads and stores.
2635 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2638 // If this is a load, save it. If this instruction can read from memory
2639 // but is not a load, then we quit. Notice that we don't handle function
2640 // calls that read or write.
2641 if (it->mayReadFromMemory()) {
2642 LoadInst *Ld = dyn_cast<LoadInst>(it);
2643 if (!Ld) return false;
2644 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2645 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2648 Loads.push_back(Ld);
2652 // Save 'store' instructions. Abort if other instructions write to memory.
2653 if (it->mayWriteToMemory()) {
2654 StoreInst *St = dyn_cast<StoreInst>(it);
2655 if (!St) return false;
2656 if (!St->isSimple() && !IsAnnotatedParallel) {
2657 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2660 Stores.push_back(St);
2665 // Now we have two lists that hold the loads and the stores.
2666 // Next, we find the pointers that they use.
2668 // Check if we see any stores. If there are no stores, then we don't
2669 // care if the pointers are *restrict*.
2670 if (!Stores.size()) {
2671 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2675 // Holds the read and read-write *pointers* that we find. These maps hold
2676 // unique values for pointers (so no need for multi-map).
2678 AliasMap ReadWrites;
2680 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2681 // multiple times on the same object. If the ptr is accessed twice, once
2682 // for read and once for write, it will only appear once (on the write
2683 // list). This is okay, since we are going to check for conflicts between
2684 // writes and between reads and writes, but not between reads and reads.
2687 ValueVector::iterator I, IE;
2688 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2689 StoreInst *ST = cast<StoreInst>(*I);
2690 Value* Ptr = ST->getPointerOperand();
2692 if (isUniform(Ptr)) {
2693 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2697 // If we did *not* see this pointer before, insert it to
2698 // the read-write list. At this phase it is only a 'write' list.
2699 if (Seen.insert(Ptr))
2700 ReadWrites.insert(std::make_pair(Ptr, ST));
2703 if (IsAnnotatedParallel) {
2705 << "LV: A loop annotated parallel, ignore memory dependency "
2710 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2711 LoadInst *LD = cast<LoadInst>(*I);
2712 Value* Ptr = LD->getPointerOperand();
2713 // If we did *not* see this pointer before, insert it to the
2714 // read list. If we *did* see it before, then it is already in
2715 // the read-write list. This allows us to vectorize expressions
2716 // such as A[i] += x; Because the address of A[i] is a read-write
2717 // pointer. This only works if the index of A[i] is consecutive.
2718 // If the address of i is unknown (for example A[B[i]]) then we may
2719 // read a few words, modify, and write a few words, and some of the
2720 // words may be written to the same address.
2721 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2722 Reads.insert(std::make_pair(Ptr, LD));
2725 // If we write (or read-write) to a single destination and there are no
2726 // other reads in this loop then is it safe to vectorize.
2727 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2728 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2732 unsigned NumReadPtrs = 0;
2733 unsigned NumWritePtrs = 0;
2735 // Find pointers with computable bounds. We are going to use this information
2736 // to place a runtime bound check.
2737 bool CanDoRT = true;
2738 AliasMap::iterator MI, ME;
2739 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2740 Value *V = (*MI).first;
2741 if (hasComputableBounds(V)) {
2742 PtrRtCheck.insert(SE, TheLoop, V, true);
2744 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2750 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2751 Value *V = (*MI).first;
2752 if (hasComputableBounds(V)) {
2753 PtrRtCheck.insert(SE, TheLoop, V, false);
2755 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2762 // Check that we did not collect too many pointers or found a
2763 // unsizeable pointer.
2764 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
2765 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
2766 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
2772 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2775 bool NeedRTCheck = false;
2777 // Biggest vectorized access possible, vector width * unroll factor.
2778 // TODO: We're being very pessimistic here, find a way to know the
2779 // real access width before getting here.
2780 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2781 TTI->getMaximumUnrollFactor();
2782 // Now that the pointers are in two lists (Reads and ReadWrites), we
2783 // can check that there are no conflicts between each of the writes and
2784 // between the writes to the reads.
2785 // Note that WriteObjects duplicates the stores (indexed now by underlying
2786 // objects) to avoid pointing to elements inside ReadWrites.
2787 // TODO: Maybe create a new type where they can interact without duplication.
2788 AliasMultiMap WriteObjects;
2789 ValueVector TempObjects;
2791 // Check that the read-writes do not conflict with other read-write
2793 bool AllWritesIdentified = true;
2794 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2795 Value *Val = (*MI).first;
2796 Instruction *Inst = (*MI).second;
2798 GetUnderlyingObjects(Val, TempObjects, DL);
2799 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2801 if (!isIdentifiedObject(*UI)) {
2802 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2804 AllWritesIdentified = false;
2807 // Never seen it before, can't alias.
2808 if (WriteObjects[*UI].empty()) {
2809 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2810 WriteObjects[*UI].push_back(Inst);
2813 // Direct alias found.
2814 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2815 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2819 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2821 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2822 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2824 // If global alias, make sure they do alias.
2825 if (hasPossibleGlobalWriteReorder(*UI,
2829 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
2834 // Didn't alias, insert into map for further reference.
2835 WriteObjects[*UI].push_back(Inst);
2837 TempObjects.clear();
2840 /// Check that the reads don't conflict with the read-writes.
2841 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2842 Value *Val = (*MI).first;
2843 GetUnderlyingObjects(Val, TempObjects, DL);
2844 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2846 // If all of the writes are identified then we don't care if the read
2847 // pointer is identified or not.
2848 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2849 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2853 // Never seen it before, can't alias.
2854 if (WriteObjects[*UI].empty())
2856 // Direct alias found.
2857 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2858 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2862 DEBUG(dbgs() << "LV: Found a global value: "
2864 Instruction *Inst = (*MI).second;
2865 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2866 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2868 // If global alias, make sure they do alias.
2869 if (hasPossibleGlobalWriteReorder(*UI,
2873 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
2878 TempObjects.clear();
2881 PtrRtCheck.Need = NeedRTCheck;
2882 if (NeedRTCheck && !CanDoRT) {
2883 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2884 "the array bounds.\n");
2889 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2890 " need a runtime memory check.\n");
2894 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2895 ReductionKind Kind) {
2896 if (Phi->getNumIncomingValues() != 2)
2899 // Reduction variables are only found in the loop header block.
2900 if (Phi->getParent() != TheLoop->getHeader())
2903 // Obtain the reduction start value from the value that comes from the loop
2905 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2907 // ExitInstruction is the single value which is used outside the loop.
2908 // We only allow for a single reduction value to be used outside the loop.
2909 // This includes users of the reduction, variables (which form a cycle
2910 // which ends in the phi node).
2911 Instruction *ExitInstruction = 0;
2912 // Indicates that we found a binary operation in our scan.
2913 bool FoundBinOp = false;
2915 // Iter is our iterator. We start with the PHI node and scan for all of the
2916 // users of this instruction. All users must be instructions that can be
2917 // used as reduction variables (such as ADD). We may have a single
2918 // out-of-block user. The cycle must end with the original PHI.
2919 Instruction *Iter = Phi;
2921 // To recognize min/max patterns formed by a icmp select sequence, we store
2922 // the number of instruction we saw from the recognized min/max pattern,
2923 // such that we don't stop when we see the phi has two uses (one by the select
2924 // and one by the icmp) and to make sure we only see exactly the two
2926 unsigned NumCmpSelectPatternInst = 0;
2927 ReductionInstDesc ReduxDesc(false, 0);
2929 // Avoid cycles in the chain.
2930 SmallPtrSet<Instruction *, 8> VisitedInsts;
2931 while (VisitedInsts.insert(Iter)) {
2932 // If the instruction has no users then this is a broken
2933 // chain and can't be a reduction variable.
2934 if (Iter->use_empty())
2937 // Did we find a user inside this loop already ?
2938 bool FoundInBlockUser = false;
2939 // Did we reach the initial PHI node already ?
2940 bool FoundStartPHI = false;
2942 // Is this a bin op ?
2943 FoundBinOp |= !isa<PHINode>(Iter);
2945 // For each of the *users* of iter.
2946 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2948 Instruction *U = cast<Instruction>(*it);
2949 // We already know that the PHI is a user.
2951 FoundStartPHI = true;
2955 // Check if we found the exit user.
2956 BasicBlock *Parent = U->getParent();
2957 if (!TheLoop->contains(Parent)) {
2958 // Exit if you find multiple outside users.
2959 if (ExitInstruction != 0)
2961 ExitInstruction = Iter;
2964 // We allow in-loop PHINodes which are not the original reduction PHI
2965 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2966 // structure) then don't skip this PHI.
2967 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2968 U->getParent() != TheLoop->getHeader() &&
2969 TheLoop->contains(U) &&
2970 Iter->hasNUsesOrMore(2))
2973 // We can't have multiple inside users except for a combination of
2974 // icmp/select both using the phi.
2975 if (FoundInBlockUser && !NumCmpSelectPatternInst)
2977 FoundInBlockUser = true;
2979 // Any reduction instr must be of one of the allowed kinds.
2980 ReduxDesc = isReductionInstr(U, Kind, ReduxDesc);
2981 if (!ReduxDesc.IsReduction)
2984 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(U) || isa<SelectInst>(U)))
2985 ++NumCmpSelectPatternInst;
2986 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(U) || isa<SelectInst>(U)))
2987 ++NumCmpSelectPatternInst;
2989 // Reductions of instructions such as Div, and Sub is only
2990 // possible if the LHS is the reduction variable.
2991 if (!U->isCommutative() && !isa<PHINode>(U) && !isa<SelectInst>(U) &&
2992 !isa<ICmpInst>(U) && !isa<FCmpInst>(U) && U->getOperand(0) != Iter)
2995 Iter = ReduxDesc.PatternLastInst;
2998 // This means we have seen one but not the other instruction of the
2999 // pattern or more than just a select and cmp.
3000 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3001 NumCmpSelectPatternInst != 2)
3004 // We found a reduction var if we have reached the original
3005 // phi node and we only have a single instruction with out-of-loop
3007 if (FoundStartPHI) {
3008 // This instruction is allowed to have out-of-loop users.
3009 AllowedExit.insert(ExitInstruction);
3011 // Save the description of this reduction variable.
3012 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3013 ReduxDesc.MinMaxKind);
3014 Reductions[Phi] = RD;
3015 // We've ended the cycle. This is a reduction variable if we have an
3016 // outside user and it has a binary op.
3017 return FoundBinOp && ExitInstruction;
3024 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3025 /// pattern corresponding to a min(X, Y) or max(X, Y).
3026 LoopVectorizationLegality::ReductionInstDesc
3027 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3028 ReductionInstDesc &Prev) {
3030 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3031 "Expect a select instruction");
3032 Instruction *Cmp = 0;
3033 SelectInst *Select = 0;
3035 // We must handle the select(cmp()) as a single instruction. Advance to the
3037 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3038 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3039 return ReductionInstDesc(false, I);
3040 return ReductionInstDesc(Select, Prev.MinMaxKind);
3043 // Only handle single use cases for now.
3044 if (!(Select = dyn_cast<SelectInst>(I)))
3045 return ReductionInstDesc(false, I);
3046 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3047 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3048 return ReductionInstDesc(false, I);
3049 if (!Cmp->hasOneUse())
3050 return ReductionInstDesc(false, I);
3055 // Look for a min/max pattern.
3056 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3057 return ReductionInstDesc(Select, MRK_UIntMin);
3058 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3059 return ReductionInstDesc(Select, MRK_UIntMax);
3060 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3061 return ReductionInstDesc(Select, MRK_SIntMax);
3062 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3063 return ReductionInstDesc(Select, MRK_SIntMin);
3064 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3065 return ReductionInstDesc(Select, MRK_FloatMin);
3066 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3067 return ReductionInstDesc(Select, MRK_FloatMax);
3068 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3069 return ReductionInstDesc(Select, MRK_FloatMin);
3070 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3071 return ReductionInstDesc(Select, MRK_FloatMax);
3073 return ReductionInstDesc(false, I);
3076 LoopVectorizationLegality::ReductionInstDesc
3077 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3079 ReductionInstDesc &Prev) {
3080 bool FP = I->getType()->isFloatingPointTy();
3081 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3082 switch (I->getOpcode()) {
3084 return ReductionInstDesc(false, I);
3085 case Instruction::PHI:
3086 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3087 Kind != RK_FloatMinMax))
3088 return ReductionInstDesc(false, I);
3089 return ReductionInstDesc(I, Prev.MinMaxKind);
3090 case Instruction::Sub:
3091 case Instruction::Add:
3092 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3093 case Instruction::Mul:
3094 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3095 case Instruction::And:
3096 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3097 case Instruction::Or:
3098 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3099 case Instruction::Xor:
3100 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3101 case Instruction::FMul:
3102 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3103 case Instruction::FAdd:
3104 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3105 case Instruction::FCmp:
3106 case Instruction::ICmp:
3107 case Instruction::Select:
3108 if (Kind != RK_IntegerMinMax &&
3109 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3110 return ReductionInstDesc(false, I);
3111 return isMinMaxSelectCmpPattern(I, Prev);
3115 LoopVectorizationLegality::InductionKind
3116 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3117 Type *PhiTy = Phi->getType();
3118 // We only handle integer and pointer inductions variables.
3119 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3120 return IK_NoInduction;
3122 // Check that the PHI is consecutive.
3123 const SCEV *PhiScev = SE->getSCEV(Phi);
3124 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3126 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3127 return IK_NoInduction;
3129 const SCEV *Step = AR->getStepRecurrence(*SE);
3131 // Integer inductions need to have a stride of one.
3132 if (PhiTy->isIntegerTy()) {
3134 return IK_IntInduction;
3135 if (Step->isAllOnesValue())
3136 return IK_ReverseIntInduction;
3137 return IK_NoInduction;
3140 // Calculate the pointer stride and check if it is consecutive.
3141 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3143 return IK_NoInduction;
3145 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3146 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3147 if (C->getValue()->equalsInt(Size))
3148 return IK_PtrInduction;
3149 else if (C->getValue()->equalsInt(0 - Size))
3150 return IK_ReversePtrInduction;
3152 return IK_NoInduction;
3155 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3156 Value *In0 = const_cast<Value*>(V);
3157 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3161 return Inductions.count(PN);
3164 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3165 assert(TheLoop->contains(BB) && "Unknown block used");
3167 // Blocks that do not dominate the latch need predication.
3168 BasicBlock* Latch = TheLoop->getLoopLatch();
3169 return !DT->dominates(BB, Latch);
3172 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3173 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3174 // We don't predicate loads/stores at the moment.
3175 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
3178 // The instructions below can trap.
3179 switch (it->getOpcode()) {
3181 case Instruction::UDiv:
3182 case Instruction::SDiv:
3183 case Instruction::URem:
3184 case Instruction::SRem:
3192 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3193 const SCEV *PhiScev = SE->getSCEV(Ptr);
3194 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3198 return AR->isAffine();
3201 LoopVectorizationCostModel::VectorizationFactor
3202 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3204 // Width 1 means no vectorize
3205 VectorizationFactor Factor = { 1U, 0U };
3206 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3207 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3211 // Find the trip count.
3212 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3213 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3215 unsigned WidestType = getWidestType();
3216 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3217 unsigned MaxVectorSize = WidestRegister / WidestType;
3218 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3219 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3221 if (MaxVectorSize == 0) {
3222 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3226 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3227 " into one vector!");
3229 unsigned VF = MaxVectorSize;
3231 // If we optimize the program for size, avoid creating the tail loop.
3233 // If we are unable to calculate the trip count then don't try to vectorize.
3235 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3239 // Find the maximum SIMD width that can fit within the trip count.
3240 VF = TC % MaxVectorSize;
3245 // If the trip count that we found modulo the vectorization factor is not
3246 // zero then we require a tail.
3248 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3254 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3255 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3257 Factor.Width = UserVF;
3261 float Cost = expectedCost(1);
3263 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3264 for (unsigned i=2; i <= VF; i*=2) {
3265 // Notice that the vector loop needs to be executed less times, so
3266 // we need to divide the cost of the vector loops by the width of
3267 // the vector elements.
3268 float VectorCost = expectedCost(i) / (float)i;
3269 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3270 (int)VectorCost << ".\n");
3271 if (VectorCost < Cost) {
3277 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3278 Factor.Width = Width;
3279 Factor.Cost = Width * Cost;
3283 unsigned LoopVectorizationCostModel::getWidestType() {
3284 unsigned MaxWidth = 8;
3287 for (Loop::block_iterator bb = TheLoop->block_begin(),
3288 be = TheLoop->block_end(); bb != be; ++bb) {
3289 BasicBlock *BB = *bb;
3291 // For each instruction in the loop.
3292 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3293 Type *T = it->getType();
3295 // Only examine Loads, Stores and PHINodes.
3296 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3299 // Examine PHI nodes that are reduction variables.
3300 if (PHINode *PN = dyn_cast<PHINode>(it))
3301 if (!Legal->getReductionVars()->count(PN))
3304 // Examine the stored values.
3305 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3306 T = ST->getValueOperand()->getType();
3308 // Ignore loaded pointer types and stored pointer types that are not
3309 // consecutive. However, we do want to take consecutive stores/loads of
3310 // pointer vectors into account.
3311 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3314 MaxWidth = std::max(MaxWidth,
3315 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3323 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3326 unsigned LoopCost) {
3328 // -- The unroll heuristics --
3329 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3330 // There are many micro-architectural considerations that we can't predict
3331 // at this level. For example frontend pressure (on decode or fetch) due to
3332 // code size, or the number and capabilities of the execution ports.
3334 // We use the following heuristics to select the unroll factor:
3335 // 1. If the code has reductions the we unroll in order to break the cross
3336 // iteration dependency.
3337 // 2. If the loop is really small then we unroll in order to reduce the loop
3339 // 3. We don't unroll if we think that we will spill registers to memory due
3340 // to the increased register pressure.
3342 // Use the user preference, unless 'auto' is selected.
3346 // When we optimize for size we don't unroll.
3350 // Do not unroll loops with a relatively small trip count.
3351 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3352 TheLoop->getLoopLatch());
3353 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3356 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3357 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3358 " vector registers\n");
3360 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3361 // We divide by these constants so assume that we have at least one
3362 // instruction that uses at least one register.
3363 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3364 R.NumInstructions = std::max(R.NumInstructions, 1U);
3366 // We calculate the unroll factor using the following formula.
3367 // Subtract the number of loop invariants from the number of available
3368 // registers. These registers are used by all of the unrolled instances.
3369 // Next, divide the remaining registers by the number of registers that is
3370 // required by the loop, in order to estimate how many parallel instances
3371 // fit without causing spills.
3372 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3374 // Clamp the unroll factor ranges to reasonable factors.
3375 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3377 // If we did not calculate the cost for VF (because the user selected the VF)
3378 // then we calculate the cost of VF here.
3380 LoopCost = expectedCost(VF);
3382 // Clamp the calculated UF to be between the 1 and the max unroll factor
3383 // that the target allows.
3384 if (UF > MaxUnrollSize)
3389 if (Legal->getReductionVars()->size()) {
3390 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3394 // We want to unroll tiny loops in order to reduce the loop overhead.
3395 // We assume that the cost overhead is 1 and we use the cost model
3396 // to estimate the cost of the loop and unroll until the cost of the
3397 // loop overhead is about 5% of the cost of the loop.
3398 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3399 if (LoopCost < 20) {
3400 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3401 unsigned NewUF = 20/LoopCost + 1;
3402 return std::min(NewUF, UF);
3405 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3409 LoopVectorizationCostModel::RegisterUsage
3410 LoopVectorizationCostModel::calculateRegisterUsage() {
3411 // This function calculates the register usage by measuring the highest number
3412 // of values that are alive at a single location. Obviously, this is a very
3413 // rough estimation. We scan the loop in a topological order in order and
3414 // assign a number to each instruction. We use RPO to ensure that defs are
3415 // met before their users. We assume that each instruction that has in-loop
3416 // users starts an interval. We record every time that an in-loop value is
3417 // used, so we have a list of the first and last occurrences of each
3418 // instruction. Next, we transpose this data structure into a multi map that
3419 // holds the list of intervals that *end* at a specific location. This multi
3420 // map allows us to perform a linear search. We scan the instructions linearly
3421 // and record each time that a new interval starts, by placing it in a set.
3422 // If we find this value in the multi-map then we remove it from the set.
3423 // The max register usage is the maximum size of the set.
3424 // We also search for instructions that are defined outside the loop, but are
3425 // used inside the loop. We need this number separately from the max-interval
3426 // usage number because when we unroll, loop-invariant values do not take
3428 LoopBlocksDFS DFS(TheLoop);
3432 R.NumInstructions = 0;
3434 // Each 'key' in the map opens a new interval. The values
3435 // of the map are the index of the 'last seen' usage of the
3436 // instruction that is the key.
3437 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3438 // Maps instruction to its index.
3439 DenseMap<unsigned, Instruction*> IdxToInstr;
3440 // Marks the end of each interval.
3441 IntervalMap EndPoint;
3442 // Saves the list of instruction indices that are used in the loop.
3443 SmallSet<Instruction*, 8> Ends;
3444 // Saves the list of values that are used in the loop but are
3445 // defined outside the loop, such as arguments and constants.
3446 SmallPtrSet<Value*, 8> LoopInvariants;
3449 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3450 be = DFS.endRPO(); bb != be; ++bb) {
3451 R.NumInstructions += (*bb)->size();
3452 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3454 Instruction *I = it;
3455 IdxToInstr[Index++] = I;
3457 // Save the end location of each USE.
3458 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3459 Value *U = I->getOperand(i);
3460 Instruction *Instr = dyn_cast<Instruction>(U);
3462 // Ignore non-instruction values such as arguments, constants, etc.
3463 if (!Instr) continue;
3465 // If this instruction is outside the loop then record it and continue.
3466 if (!TheLoop->contains(Instr)) {
3467 LoopInvariants.insert(Instr);
3471 // Overwrite previous end points.
3472 EndPoint[Instr] = Index;
3478 // Saves the list of intervals that end with the index in 'key'.
3479 typedef SmallVector<Instruction*, 2> InstrList;
3480 DenseMap<unsigned, InstrList> TransposeEnds;
3482 // Transpose the EndPoints to a list of values that end at each index.
3483 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3485 TransposeEnds[it->second].push_back(it->first);
3487 SmallSet<Instruction*, 8> OpenIntervals;
3488 unsigned MaxUsage = 0;
3491 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3492 for (unsigned int i = 0; i < Index; ++i) {
3493 Instruction *I = IdxToInstr[i];
3494 // Ignore instructions that are never used within the loop.
3495 if (!Ends.count(I)) continue;
3497 // Remove all of the instructions that end at this location.
3498 InstrList &List = TransposeEnds[i];
3499 for (unsigned int j=0, e = List.size(); j < e; ++j)
3500 OpenIntervals.erase(List[j]);
3502 // Count the number of live interals.
3503 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3505 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3506 OpenIntervals.size() <<"\n");
3508 // Add the current instruction to the list of open intervals.
3509 OpenIntervals.insert(I);
3512 unsigned Invariant = LoopInvariants.size();
3513 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3514 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3515 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3517 R.LoopInvariantRegs = Invariant;
3518 R.MaxLocalUsers = MaxUsage;
3522 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3526 for (Loop::block_iterator bb = TheLoop->block_begin(),
3527 be = TheLoop->block_end(); bb != be; ++bb) {
3528 unsigned BlockCost = 0;
3529 BasicBlock *BB = *bb;
3531 // For each instruction in the old loop.
3532 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3533 // Skip dbg intrinsics.
3534 if (isa<DbgInfoIntrinsic>(it))
3537 unsigned C = getInstructionCost(it, VF);
3539 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3540 VF << " For instruction: "<< *it << "\n");
3543 // We assume that if-converted blocks have a 50% chance of being executed.
3544 // When the code is scalar then some of the blocks are avoided due to CF.
3545 // When the code is vectorized we execute all code paths.
3546 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3556 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3557 // If we know that this instruction will remain uniform, check the cost of
3558 // the scalar version.
3559 if (Legal->isUniformAfterVectorization(I))
3562 Type *RetTy = I->getType();
3563 Type *VectorTy = ToVectorTy(RetTy, VF);
3565 // TODO: We need to estimate the cost of intrinsic calls.
3566 switch (I->getOpcode()) {
3567 case Instruction::GetElementPtr:
3568 // We mark this instruction as zero-cost because the cost of GEPs in
3569 // vectorized code depends on whether the corresponding memory instruction
3570 // is scalarized or not. Therefore, we handle GEPs with the memory
3571 // instruction cost.
3573 case Instruction::Br: {
3574 return TTI.getCFInstrCost(I->getOpcode());
3576 case Instruction::PHI:
3577 //TODO: IF-converted IFs become selects.
3579 case Instruction::Add:
3580 case Instruction::FAdd:
3581 case Instruction::Sub:
3582 case Instruction::FSub:
3583 case Instruction::Mul:
3584 case Instruction::FMul:
3585 case Instruction::UDiv:
3586 case Instruction::SDiv:
3587 case Instruction::FDiv:
3588 case Instruction::URem:
3589 case Instruction::SRem:
3590 case Instruction::FRem:
3591 case Instruction::Shl:
3592 case Instruction::LShr:
3593 case Instruction::AShr:
3594 case Instruction::And:
3595 case Instruction::Or:
3596 case Instruction::Xor: {
3597 // Certain instructions can be cheaper to vectorize if they have a constant
3598 // second vector operand. One example of this are shifts on x86.
3599 TargetTransformInfo::OperandValueKind Op1VK =
3600 TargetTransformInfo::OK_AnyValue;
3601 TargetTransformInfo::OperandValueKind Op2VK =
3602 TargetTransformInfo::OK_AnyValue;
3604 if (isa<ConstantInt>(I->getOperand(1)))
3605 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3607 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3609 case Instruction::Select: {
3610 SelectInst *SI = cast<SelectInst>(I);
3611 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3612 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3613 Type *CondTy = SI->getCondition()->getType();
3615 CondTy = VectorType::get(CondTy, VF);
3617 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3619 case Instruction::ICmp:
3620 case Instruction::FCmp: {
3621 Type *ValTy = I->getOperand(0)->getType();
3622 VectorTy = ToVectorTy(ValTy, VF);
3623 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3625 case Instruction::Store:
3626 case Instruction::Load: {
3627 StoreInst *SI = dyn_cast<StoreInst>(I);
3628 LoadInst *LI = dyn_cast<LoadInst>(I);
3629 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3631 VectorTy = ToVectorTy(ValTy, VF);
3633 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3634 unsigned AS = SI ? SI->getPointerAddressSpace() :
3635 LI->getPointerAddressSpace();
3636 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3637 // We add the cost of address computation here instead of with the gep
3638 // instruction because only here we know whether the operation is
3641 return TTI.getAddressComputationCost(VectorTy) +
3642 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3644 // Scalarized loads/stores.
3645 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3646 bool Reverse = ConsecutiveStride < 0;
3647 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3648 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3649 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3651 // The cost of extracting from the value vector and pointer vector.
3652 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3653 for (unsigned i = 0; i < VF; ++i) {
3654 // The cost of extracting the pointer operand.
3655 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3656 // In case of STORE, the cost of ExtractElement from the vector.
3657 // In case of LOAD, the cost of InsertElement into the returned
3659 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3660 Instruction::InsertElement,
3664 // The cost of the scalar loads/stores.
3665 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3666 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3671 // Wide load/stores.
3672 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3673 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3676 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3680 case Instruction::ZExt:
3681 case Instruction::SExt:
3682 case Instruction::FPToUI:
3683 case Instruction::FPToSI:
3684 case Instruction::FPExt:
3685 case Instruction::PtrToInt:
3686 case Instruction::IntToPtr:
3687 case Instruction::SIToFP:
3688 case Instruction::UIToFP:
3689 case Instruction::Trunc:
3690 case Instruction::FPTrunc:
3691 case Instruction::BitCast: {
3692 // We optimize the truncation of induction variable.
3693 // The cost of these is the same as the scalar operation.
3694 if (I->getOpcode() == Instruction::Trunc &&
3695 Legal->isInductionVariable(I->getOperand(0)))
3696 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3697 I->getOperand(0)->getType());
3699 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3700 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3702 case Instruction::Call: {
3703 CallInst *CI = cast<CallInst>(I);
3704 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3705 assert(ID && "Not an intrinsic call!");
3706 Type *RetTy = ToVectorTy(CI->getType(), VF);
3707 SmallVector<Type*, 4> Tys;
3708 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3709 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3710 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3713 // We are scalarizing the instruction. Return the cost of the scalar
3714 // instruction, plus the cost of insert and extract into vector
3715 // elements, times the vector width.
3718 if (!RetTy->isVoidTy() && VF != 1) {
3719 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3721 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3724 // The cost of inserting the results plus extracting each one of the
3726 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3729 // The cost of executing VF copies of the scalar instruction. This opcode
3730 // is unknown. Assume that it is the same as 'mul'.
3731 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3737 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3738 if (Scalar->isVoidTy() || VF == 1)
3740 return VectorType::get(Scalar, VF);
3743 char LoopVectorize::ID = 0;
3744 static const char lv_name[] = "Loop Vectorization";
3745 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3746 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3747 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3748 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3749 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3750 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3753 Pass *createLoopVectorizePass() {
3754 return new LoopVectorize();
3758 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3759 // Check for a store.
3760 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3761 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3763 // Check for a load.
3764 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3765 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;