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/Target/TargetLibraryInfo.h"
84 #include "llvm/Transforms/Scalar.h"
85 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
86 #include "llvm/Transforms/Utils/Local.h"
91 using namespace llvm::PatternMatch;
93 static cl::opt<unsigned>
94 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
95 cl::desc("Sets the SIMD width. Zero is autoselect."));
97 static cl::opt<unsigned>
98 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
99 cl::desc("Sets the vectorization unroll count. "
100 "Zero is autoselect."));
103 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
104 cl::desc("Enable if-conversion during vectorization."));
106 /// We don't vectorize loops with a known constant trip count below this number.
107 static cl::opt<unsigned>
108 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
110 cl::desc("Don't vectorize loops with a constant "
111 "trip count that is smaller than this "
114 /// We don't unroll loops with a known constant trip count below this number.
115 static const unsigned TinyTripCountUnrollThreshold = 128;
117 /// When performing memory disambiguation checks at runtime do not make more
118 /// than this number of comparisons.
119 static const unsigned RuntimeMemoryCheckThreshold = 8;
121 /// We use a metadata with this name to indicate that a scalar loop was
122 /// vectorized and that we don't need to re-vectorize it if we run into it
125 AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized";
129 // Forward declarations.
130 class LoopVectorizationLegality;
131 class LoopVectorizationCostModel;
133 /// InnerLoopVectorizer vectorizes loops which contain only one basic
134 /// block to a specified vectorization factor (VF).
135 /// This class performs the widening of scalars into vectors, or multiple
136 /// scalars. This class also implements the following features:
137 /// * It inserts an epilogue loop for handling loops that don't have iteration
138 /// counts that are known to be a multiple of the vectorization factor.
139 /// * It handles the code generation for reduction variables.
140 /// * Scalarization (implementation using scalars) of un-vectorizable
142 /// InnerLoopVectorizer does not perform any vectorization-legality
143 /// checks, and relies on the caller to check for the different legality
144 /// aspects. The InnerLoopVectorizer relies on the
145 /// LoopVectorizationLegality class to provide information about the induction
146 /// and reduction variables that were found to a given vectorization factor.
147 class InnerLoopVectorizer {
149 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
150 DominatorTree *DT, DataLayout *DL,
151 const TargetLibraryInfo *TLI, unsigned VecWidth,
152 unsigned UnrollFactor)
153 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
154 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
155 OldInduction(0), WidenMap(UnrollFactor) {}
157 // Perform the actual loop widening (vectorization).
158 void vectorize(LoopVectorizationLegality *Legal) {
159 // Create a new empty loop. Unlink the old loop and connect the new one.
160 createEmptyLoop(Legal);
161 // Widen each instruction in the old loop to a new one in the new loop.
162 // Use the Legality module to find the induction and reduction variables.
163 vectorizeLoop(Legal);
164 // Register the new loop and update the analysis passes.
169 /// A small list of PHINodes.
170 typedef SmallVector<PHINode*, 4> PhiVector;
171 /// When we unroll loops we have multiple vector values for each scalar.
172 /// This data structure holds the unrolled and vectorized values that
173 /// originated from one scalar instruction.
174 typedef SmallVector<Value*, 2> VectorParts;
176 /// Add code that checks at runtime if the accessed arrays overlap.
177 /// Returns the comparator value or NULL if no check is needed.
178 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
180 /// Create an empty loop, based on the loop ranges of the old loop.
181 void createEmptyLoop(LoopVectorizationLegality *Legal);
182 /// Copy and widen the instructions from the old loop.
183 void vectorizeLoop(LoopVectorizationLegality *Legal);
185 /// A helper function that computes the predicate of the block BB, assuming
186 /// that the header block of the loop is set to True. It returns the *entry*
187 /// mask for the block BB.
188 VectorParts createBlockInMask(BasicBlock *BB);
189 /// A helper function that computes the predicate of the edge between SRC
191 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
193 /// A helper function to vectorize a single BB within the innermost loop.
194 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
197 /// Insert the new loop to the loop hierarchy and pass manager
198 /// and update the analysis passes.
199 void updateAnalysis();
201 /// This instruction is un-vectorizable. Implement it as a sequence
203 void scalarizeInstruction(Instruction *Instr);
205 /// Vectorize Load and Store instructions,
206 void vectorizeMemoryInstruction(Instruction *Instr,
207 LoopVectorizationLegality *Legal);
209 /// Create a broadcast instruction. This method generates a broadcast
210 /// instruction (shuffle) for loop invariant values and for the induction
211 /// value. If this is the induction variable then we extend it to N, N+1, ...
212 /// this is needed because each iteration in the loop corresponds to a SIMD
214 Value *getBroadcastInstrs(Value *V);
216 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
217 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
218 /// The sequence starts at StartIndex.
219 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
221 /// When we go over instructions in the basic block we rely on previous
222 /// values within the current basic block or on loop invariant values.
223 /// When we widen (vectorize) values we place them in the map. If the values
224 /// are not within the map, they have to be loop invariant, so we simply
225 /// broadcast them into a vector.
226 VectorParts &getVectorValue(Value *V);
228 /// Generate a shuffle sequence that will reverse the vector Vec.
229 Value *reverseVector(Value *Vec);
231 /// This is a helper class that holds the vectorizer state. It maps scalar
232 /// instructions to vector instructions. When the code is 'unrolled' then
233 /// then a single scalar value is mapped to multiple vector parts. The parts
234 /// are stored in the VectorPart type.
236 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
238 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
240 /// \return True if 'Key' is saved in the Value Map.
241 bool has(Value *Key) const { return MapStorage.count(Key); }
243 /// Initializes a new entry in the map. Sets all of the vector parts to the
244 /// save value in 'Val'.
245 /// \return A reference to a vector with splat values.
246 VectorParts &splat(Value *Key, Value *Val) {
247 VectorParts &Entry = MapStorage[Key];
248 Entry.assign(UF, Val);
252 ///\return A reference to the value that is stored at 'Key'.
253 VectorParts &get(Value *Key) {
254 VectorParts &Entry = MapStorage[Key];
257 assert(Entry.size() == UF);
262 /// The unroll factor. Each entry in the map stores this number of vector
266 /// Map storage. We use std::map and not DenseMap because insertions to a
267 /// dense map invalidates its iterators.
268 std::map<Value *, VectorParts> MapStorage;
271 /// The original loop.
273 /// Scev analysis to use.
281 /// Target Library Info.
282 const TargetLibraryInfo *TLI;
284 /// The vectorization SIMD factor to use. Each vector will have this many
287 /// The vectorization unroll factor to use. Each scalar is vectorized to this
288 /// many different vector instructions.
291 /// The builder that we use
294 // --- Vectorization state ---
296 /// The vector-loop preheader.
297 BasicBlock *LoopVectorPreHeader;
298 /// The scalar-loop preheader.
299 BasicBlock *LoopScalarPreHeader;
300 /// Middle Block between the vector and the scalar.
301 BasicBlock *LoopMiddleBlock;
302 ///The ExitBlock of the scalar loop.
303 BasicBlock *LoopExitBlock;
304 ///The vector loop body.
305 BasicBlock *LoopVectorBody;
306 ///The scalar loop body.
307 BasicBlock *LoopScalarBody;
308 /// A list of all bypass blocks. The first block is the entry of the loop.
309 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
311 /// The new Induction variable which was added to the new block.
313 /// The induction variable of the old basic block.
314 PHINode *OldInduction;
315 /// Maps scalars to widened vectors.
319 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
320 /// to what vectorization factor.
321 /// This class does not look at the profitability of vectorization, only the
322 /// legality. This class has two main kinds of checks:
323 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
324 /// will change the order of memory accesses in a way that will change the
325 /// correctness of the program.
326 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
327 /// checks for a number of different conditions, such as the availability of a
328 /// single induction variable, that all types are supported and vectorize-able,
329 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
330 /// This class is also used by InnerLoopVectorizer for identifying
331 /// induction variable and the different reduction variables.
332 class LoopVectorizationLegality {
334 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
335 DominatorTree *DT, TargetTransformInfo* TTI,
336 AliasAnalysis *AA, TargetLibraryInfo *TLI)
337 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
338 Induction(0), HasFunNoNaNAttr(false) {}
340 /// This enum represents the kinds of reductions that we support.
342 RK_NoReduction, ///< Not a reduction.
343 RK_IntegerAdd, ///< Sum of integers.
344 RK_IntegerMult, ///< Product of integers.
345 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
346 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
347 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
348 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
349 RK_FloatAdd, ///< Sum of floats.
350 RK_FloatMult, ///< Product of floats.
351 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
354 /// This enum represents the kinds of inductions that we support.
356 IK_NoInduction, ///< Not an induction variable.
357 IK_IntInduction, ///< Integer induction variable. Step = 1.
358 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
359 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
360 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
363 // This enum represents the kind of minmax reduction.
364 enum MinMaxReductionKind {
374 /// This POD struct holds information about reduction variables.
375 struct ReductionDescriptor {
376 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
377 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
379 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
380 MinMaxReductionKind MK)
381 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
383 // The starting value of the reduction.
384 // It does not have to be zero!
386 // The instruction who's value is used outside the loop.
387 Instruction *LoopExitInstr;
388 // The kind of the reduction.
390 // If this a min/max reduction the kind of reduction.
391 MinMaxReductionKind MinMaxKind;
394 /// This POD struct holds information about a potential reduction operation.
395 struct ReductionInstDesc {
396 ReductionInstDesc(bool IsRedux, Instruction *I) :
397 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
399 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
400 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
402 // Is this instruction a reduction candidate.
404 // The last instruction in a min/max pattern (select of the select(icmp())
405 // pattern), or the current reduction instruction otherwise.
406 Instruction *PatternLastInst;
407 // If this is a min/max pattern the comparison predicate.
408 MinMaxReductionKind MinMaxKind;
411 // This POD struct holds information about the memory runtime legality
412 // check that a group of pointers do not overlap.
413 struct RuntimePointerCheck {
414 RuntimePointerCheck() : Need(false) {}
416 /// Reset the state of the pointer runtime information.
424 /// Insert a pointer and calculate the start and end SCEVs.
425 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr);
427 /// This flag indicates if we need to add the runtime check.
429 /// Holds the pointers that we need to check.
430 SmallVector<Value*, 2> Pointers;
431 /// Holds the pointer value at the beginning of the loop.
432 SmallVector<const SCEV*, 2> Starts;
433 /// Holds the pointer value at the end of the loop.
434 SmallVector<const SCEV*, 2> Ends;
435 /// Holds the information if this pointer is used for writing to memory.
436 SmallVector<bool, 2> IsWritePtr;
439 /// A POD for saving information about induction variables.
440 struct InductionInfo {
441 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
442 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
449 /// ReductionList contains the reduction descriptors for all
450 /// of the reductions that were found in the loop.
451 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
453 /// InductionList saves induction variables and maps them to the
454 /// induction descriptor.
455 typedef MapVector<PHINode*, InductionInfo> InductionList;
457 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
458 /// respective Store/Load instruction(s) to calculate aliasing.
459 typedef MapVector<Value*, Instruction* > AliasMap;
460 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
462 /// Returns true if it is legal to vectorize this loop.
463 /// This does not mean that it is profitable to vectorize this
464 /// loop, only that it is legal to do so.
467 /// Returns the Induction variable.
468 PHINode *getInduction() { return Induction; }
470 /// Returns the reduction variables found in the loop.
471 ReductionList *getReductionVars() { return &Reductions; }
473 /// Returns the induction variables found in the loop.
474 InductionList *getInductionVars() { return &Inductions; }
476 /// Returns True if V is an induction variable in this loop.
477 bool isInductionVariable(const Value *V);
479 /// Return true if the block BB needs to be predicated in order for the loop
480 /// to be vectorized.
481 bool blockNeedsPredication(BasicBlock *BB);
483 /// Check if this pointer is consecutive when vectorizing. This happens
484 /// when the last index of the GEP is the induction variable, or that the
485 /// pointer itself is an induction variable.
486 /// This check allows us to vectorize A[idx] into a wide load/store.
488 /// 0 - Stride is unknown or non consecutive.
489 /// 1 - Address is consecutive.
490 /// -1 - Address is consecutive, and decreasing.
491 int isConsecutivePtr(Value *Ptr);
493 /// Returns true if the value V is uniform within the loop.
494 bool isUniform(Value *V);
496 /// Returns true if this instruction will remain scalar after vectorization.
497 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
499 /// Returns the information that we collected about runtime memory check.
500 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
502 /// This function returns the identity element (or neutral element) for
504 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
506 /// Check if a single basic block loop is vectorizable.
507 /// At this point we know that this is a loop with a constant trip count
508 /// and we only need to check individual instructions.
509 bool canVectorizeInstrs();
511 /// When we vectorize loops we may change the order in which
512 /// we read and write from memory. This method checks if it is
513 /// legal to vectorize the code, considering only memory constrains.
514 /// Returns true if the loop is vectorizable
515 bool canVectorizeMemory();
517 /// Return true if we can vectorize this loop using the IF-conversion
519 bool canVectorizeWithIfConvert();
521 /// Collect the variables that need to stay uniform after vectorization.
522 void collectLoopUniforms();
524 /// Return true if all of the instructions in the block can be speculatively
526 bool blockCanBePredicated(BasicBlock *BB);
528 /// Returns True, if 'Phi' is the kind of reduction variable for type
529 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
530 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
531 /// Returns a struct describing if the instruction 'I' can be a reduction
532 /// variable of type 'Kind'. If the reduction is a min/max pattern of
533 /// select(icmp()) this function advances the instruction pointer 'I' from the
534 /// compare instruction to the select instruction and stores this pointer in
535 /// 'PatternLastInst' member of the returned struct.
536 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
537 ReductionInstDesc &Desc);
538 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
539 /// pattern corresponding to a min(X, Y) or max(X, Y).
540 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
541 ReductionInstDesc &Prev);
542 /// Returns the induction kind of Phi. This function may return NoInduction
543 /// if the PHI is not an induction variable.
544 InductionKind isInductionVariable(PHINode *Phi);
545 /// Return true if can compute the address bounds of Ptr within the loop.
546 bool hasComputableBounds(Value *Ptr);
547 /// Return true if there is the chance of write reorder.
548 bool hasPossibleGlobalWriteReorder(Value *Object,
550 AliasMultiMap &WriteObjects,
551 unsigned MaxByteWidth);
552 /// Return the AA location for a load or a store.
553 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
556 /// The loop that we evaluate.
560 /// DataLayout analysis.
565 TargetTransformInfo *TTI;
568 /// Target Library Info.
569 TargetLibraryInfo *TLI;
571 // --- vectorization state --- //
573 /// Holds the integer induction variable. This is the counter of the
576 /// Holds the reduction variables.
577 ReductionList Reductions;
578 /// Holds all of the induction variables that we found in the loop.
579 /// Notice that inductions don't need to start at zero and that induction
580 /// variables can be pointers.
581 InductionList Inductions;
583 /// Allowed outside users. This holds the reduction
584 /// vars which can be accessed from outside the loop.
585 SmallPtrSet<Value*, 4> AllowedExit;
586 /// This set holds the variables which are known to be uniform after
588 SmallPtrSet<Instruction*, 4> Uniforms;
589 /// We need to check that all of the pointers in this list are disjoint
591 RuntimePointerCheck PtrRtCheck;
592 /// Can we assume the absence of NaNs.
593 bool HasFunNoNaNAttr;
596 /// LoopVectorizationCostModel - estimates the expected speedups due to
598 /// In many cases vectorization is not profitable. This can happen because of
599 /// a number of reasons. In this class we mainly attempt to predict the
600 /// expected speedup/slowdowns due to the supported instruction set. We use the
601 /// TargetTransformInfo to query the different backends for the cost of
602 /// different operations.
603 class LoopVectorizationCostModel {
605 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
606 LoopVectorizationLegality *Legal,
607 const TargetTransformInfo &TTI,
608 DataLayout *DL, const TargetLibraryInfo *TLI)
609 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
611 /// Information about vectorization costs
612 struct VectorizationFactor {
613 unsigned Width; // Vector width with best cost
614 unsigned Cost; // Cost of the loop with that width
616 /// \return The most profitable vectorization factor and the cost of that VF.
617 /// This method checks every power of two up to VF. If UserVF is not ZERO
618 /// then this vectorization factor will be selected if vectorization is
620 VectorizationFactor selectVectorizationFactor(bool OptForSize,
623 /// \return The size (in bits) of the widest type in the code that
624 /// needs to be vectorized. We ignore values that remain scalar such as
625 /// 64 bit loop indices.
626 unsigned getWidestType();
628 /// \return The most profitable unroll factor.
629 /// If UserUF is non-zero then this method finds the best unroll-factor
630 /// based on register pressure and other parameters.
631 /// VF and LoopCost are the selected vectorization factor and the cost of the
633 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
636 /// \brief A struct that represents some properties of the register usage
638 struct RegisterUsage {
639 /// Holds the number of loop invariant values that are used in the loop.
640 unsigned LoopInvariantRegs;
641 /// Holds the maximum number of concurrent live intervals in the loop.
642 unsigned MaxLocalUsers;
643 /// Holds the number of instructions in the loop.
644 unsigned NumInstructions;
647 /// \return information about the register usage of the loop.
648 RegisterUsage calculateRegisterUsage();
651 /// Returns the expected execution cost. The unit of the cost does
652 /// not matter because we use the 'cost' units to compare different
653 /// vector widths. The cost that is returned is *not* normalized by
654 /// the factor width.
655 unsigned expectedCost(unsigned VF);
657 /// Returns the execution time cost of an instruction for a given vector
658 /// width. Vector width of one means scalar.
659 unsigned getInstructionCost(Instruction *I, unsigned VF);
661 /// A helper function for converting Scalar types to vector types.
662 /// If the incoming type is void, we return void. If the VF is 1, we return
664 static Type* ToVectorTy(Type *Scalar, unsigned VF);
666 /// Returns whether the instruction is a load or store and will be a emitted
667 /// as a vector operation.
668 bool isConsecutiveLoadOrStore(Instruction *I);
670 /// The loop that we evaluate.
674 /// Loop Info analysis.
676 /// Vectorization legality.
677 LoopVectorizationLegality *Legal;
678 /// Vector target information.
679 const TargetTransformInfo &TTI;
680 /// Target data layout information.
682 /// Target Library Info.
683 const TargetLibraryInfo *TLI;
686 /// The LoopVectorize Pass.
687 struct LoopVectorize : public LoopPass {
688 /// Pass identification, replacement for typeid
691 explicit LoopVectorize() : LoopPass(ID) {
692 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
698 TargetTransformInfo *TTI;
701 TargetLibraryInfo *TLI;
703 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
704 // We only vectorize innermost loops.
708 SE = &getAnalysis<ScalarEvolution>();
709 DL = getAnalysisIfAvailable<DataLayout>();
710 LI = &getAnalysis<LoopInfo>();
711 TTI = &getAnalysis<TargetTransformInfo>();
712 DT = &getAnalysis<DominatorTree>();
713 AA = getAnalysisIfAvailable<AliasAnalysis>();
714 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
717 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
721 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
722 L->getHeader()->getParent()->getName() << "\"\n");
724 // Check if it is legal to vectorize the loop.
725 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
726 if (!LVL.canVectorize()) {
727 DEBUG(dbgs() << "LV: Not vectorizing.\n");
731 // Use the cost model.
732 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
734 // Check the function attributes to find out if this function should be
735 // optimized for size.
736 Function *F = L->getHeader()->getParent();
737 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
738 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
739 unsigned FnIndex = AttributeSet::FunctionIndex;
740 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
741 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
744 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
745 "attribute is used.\n");
749 // Select the optimal vectorization factor.
750 LoopVectorizationCostModel::VectorizationFactor VF;
751 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
752 // Select the unroll factor.
753 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
757 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
761 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
762 F->getParent()->getModuleIdentifier()<<"\n");
763 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
765 // If we decided that it is *legal* to vectorize the loop then do it.
766 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
769 DEBUG(verifyFunction(*L->getHeader()->getParent()));
773 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
774 LoopPass::getAnalysisUsage(AU);
775 AU.addRequiredID(LoopSimplifyID);
776 AU.addRequiredID(LCSSAID);
777 AU.addRequired<DominatorTree>();
778 AU.addRequired<LoopInfo>();
779 AU.addRequired<ScalarEvolution>();
780 AU.addRequired<TargetTransformInfo>();
781 AU.addPreserved<LoopInfo>();
782 AU.addPreserved<DominatorTree>();
787 } // end anonymous namespace
789 //===----------------------------------------------------------------------===//
790 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
791 // LoopVectorizationCostModel.
792 //===----------------------------------------------------------------------===//
795 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
796 Loop *Lp, Value *Ptr,
798 const SCEV *Sc = SE->getSCEV(Ptr);
799 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
800 assert(AR && "Invalid addrec expression");
801 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
802 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
803 Pointers.push_back(Ptr);
804 Starts.push_back(AR->getStart());
805 Ends.push_back(ScEnd);
806 IsWritePtr.push_back(WritePtr);
809 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
810 // Save the current insertion location.
811 Instruction *Loc = Builder.GetInsertPoint();
813 // We need to place the broadcast of invariant variables outside the loop.
814 Instruction *Instr = dyn_cast<Instruction>(V);
815 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
816 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
818 // Place the code for broadcasting invariant variables in the new preheader.
820 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
822 // Broadcast the scalar into all locations in the vector.
823 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
825 // Restore the builder insertion point.
827 Builder.SetInsertPoint(Loc);
832 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
834 assert(Val->getType()->isVectorTy() && "Must be a vector");
835 assert(Val->getType()->getScalarType()->isIntegerTy() &&
836 "Elem must be an integer");
838 Type *ITy = Val->getType()->getScalarType();
839 VectorType *Ty = cast<VectorType>(Val->getType());
840 int VLen = Ty->getNumElements();
841 SmallVector<Constant*, 8> Indices;
843 // Create a vector of consecutive numbers from zero to VF.
844 for (int i = 0; i < VLen; ++i) {
845 int64_t Idx = Negate ? (-i) : i;
846 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
849 // Add the consecutive indices to the vector value.
850 Constant *Cv = ConstantVector::get(Indices);
851 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
852 return Builder.CreateAdd(Val, Cv, "induction");
855 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
856 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
857 // Make sure that the pointer does not point to structs.
858 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
861 // If this value is a pointer induction variable we know it is consecutive.
862 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
863 if (Phi && Inductions.count(Phi)) {
864 InductionInfo II = Inductions[Phi];
865 if (IK_PtrInduction == II.IK)
867 else if (IK_ReversePtrInduction == II.IK)
871 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
875 unsigned NumOperands = Gep->getNumOperands();
876 Value *LastIndex = Gep->getOperand(NumOperands - 1);
878 Value *GpPtr = Gep->getPointerOperand();
879 // If this GEP value is a consecutive pointer induction variable and all of
880 // the indices are constant then we know it is consecutive. We can
881 Phi = dyn_cast<PHINode>(GpPtr);
882 if (Phi && Inductions.count(Phi)) {
884 // Make sure that the pointer does not point to structs.
885 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
886 if (GepPtrType->getElementType()->isAggregateType())
889 // Make sure that all of the index operands are loop invariant.
890 for (unsigned i = 1; i < NumOperands; ++i)
891 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
894 InductionInfo II = Inductions[Phi];
895 if (IK_PtrInduction == II.IK)
897 else if (IK_ReversePtrInduction == II.IK)
901 // Check that all of the gep indices are uniform except for the last.
902 for (unsigned i = 0; i < NumOperands - 1; ++i)
903 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
906 // We can emit wide load/stores only if the last index is the induction
908 const SCEV *Last = SE->getSCEV(LastIndex);
909 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
910 const SCEV *Step = AR->getStepRecurrence(*SE);
912 // The memory is consecutive because the last index is consecutive
913 // and all other indices are loop invariant.
916 if (Step->isAllOnesValue())
923 bool LoopVectorizationLegality::isUniform(Value *V) {
924 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
927 InnerLoopVectorizer::VectorParts&
928 InnerLoopVectorizer::getVectorValue(Value *V) {
929 assert(V != Induction && "The new induction variable should not be used.");
930 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
932 // If we have this scalar in the map, return it.
934 return WidenMap.get(V);
936 // If this scalar is unknown, assume that it is a constant or that it is
937 // loop invariant. Broadcast V and save the value for future uses.
938 Value *B = getBroadcastInstrs(V);
939 return WidenMap.splat(V, B);
942 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
943 assert(Vec->getType()->isVectorTy() && "Invalid type");
944 SmallVector<Constant*, 8> ShuffleMask;
945 for (unsigned i = 0; i < VF; ++i)
946 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
948 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
949 ConstantVector::get(ShuffleMask),
954 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
955 LoopVectorizationLegality *Legal) {
956 // Attempt to issue a wide load.
957 LoadInst *LI = dyn_cast<LoadInst>(Instr);
958 StoreInst *SI = dyn_cast<StoreInst>(Instr);
960 assert((LI || SI) && "Invalid Load/Store instruction");
962 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
963 Type *DataTy = VectorType::get(ScalarDataTy, VF);
964 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
965 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
967 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
968 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
970 if (ScalarAllocatedSize != VectorElementSize)
971 return scalarizeInstruction(Instr);
973 // If the pointer is loop invariant or if it is non consecutive,
974 // scalarize the load.
975 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
976 bool Reverse = ConsecutiveStride < 0;
977 bool UniformLoad = LI && Legal->isUniform(Ptr);
978 if (!ConsecutiveStride || UniformLoad)
979 return scalarizeInstruction(Instr);
981 Constant *Zero = Builder.getInt32(0);
982 VectorParts &Entry = WidenMap.get(Instr);
984 // Handle consecutive loads/stores.
985 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
986 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
987 Value *PtrOperand = Gep->getPointerOperand();
988 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
989 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
991 // Create the new GEP with the new induction variable.
992 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
993 Gep2->setOperand(0, FirstBasePtr);
994 Gep2->setName("gep.indvar.base");
995 Ptr = Builder.Insert(Gep2);
997 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
998 OrigLoop) && "Base ptr must be invariant");
1000 // The last index does not have to be the induction. It can be
1001 // consecutive and be a function of the index. For example A[I+1];
1002 unsigned NumOperands = Gep->getNumOperands();
1004 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1005 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1006 Value *LastIndex = GEPParts[0];
1007 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1009 // Create the new GEP with the new induction variable.
1010 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1011 Gep2->setOperand(NumOperands - 1, LastIndex);
1012 Gep2->setName("gep.indvar.idx");
1013 Ptr = Builder.Insert(Gep2);
1015 // Use the induction element ptr.
1016 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1017 VectorParts &PtrVal = getVectorValue(Ptr);
1018 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1023 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1024 "We do not allow storing to uniform addresses");
1026 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1027 for (unsigned Part = 0; Part < UF; ++Part) {
1028 // Calculate the pointer for the specific unroll-part.
1029 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1032 // If we store to reverse consecutive memory locations then we need
1033 // to reverse the order of elements in the stored value.
1034 StoredVal[Part] = reverseVector(StoredVal[Part]);
1035 // If the address is consecutive but reversed, then the
1036 // wide store needs to start at the last vector element.
1037 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1038 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1041 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1042 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1046 for (unsigned Part = 0; Part < UF; ++Part) {
1047 // Calculate the pointer for the specific unroll-part.
1048 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1051 // If the address is consecutive but reversed, then the
1052 // wide store needs to start at the last vector element.
1053 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1054 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1057 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1058 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1059 cast<LoadInst>(LI)->setAlignment(Alignment);
1060 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1064 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1065 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1066 // Holds vector parameters or scalars, in case of uniform vals.
1067 SmallVector<VectorParts, 4> Params;
1069 // Find all of the vectorized parameters.
1070 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1071 Value *SrcOp = Instr->getOperand(op);
1073 // If we are accessing the old induction variable, use the new one.
1074 if (SrcOp == OldInduction) {
1075 Params.push_back(getVectorValue(SrcOp));
1079 // Try using previously calculated values.
1080 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1082 // If the src is an instruction that appeared earlier in the basic block
1083 // then it should already be vectorized.
1084 if (SrcInst && OrigLoop->contains(SrcInst)) {
1085 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1086 // The parameter is a vector value from earlier.
1087 Params.push_back(WidenMap.get(SrcInst));
1089 // The parameter is a scalar from outside the loop. Maybe even a constant.
1090 VectorParts Scalars;
1091 Scalars.append(UF, SrcOp);
1092 Params.push_back(Scalars);
1096 assert(Params.size() == Instr->getNumOperands() &&
1097 "Invalid number of operands");
1099 // Does this instruction return a value ?
1100 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1102 Value *UndefVec = IsVoidRetTy ? 0 :
1103 UndefValue::get(VectorType::get(Instr->getType(), VF));
1104 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1105 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1107 // For each vector unroll 'part':
1108 for (unsigned Part = 0; Part < UF; ++Part) {
1109 // For each scalar that we create:
1110 for (unsigned Width = 0; Width < VF; ++Width) {
1111 Instruction *Cloned = Instr->clone();
1113 Cloned->setName(Instr->getName() + ".cloned");
1114 // Replace the operands of the cloned instrucions with extracted scalars.
1115 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1116 Value *Op = Params[op][Part];
1117 // Param is a vector. Need to extract the right lane.
1118 if (Op->getType()->isVectorTy())
1119 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1120 Cloned->setOperand(op, Op);
1123 // Place the cloned scalar in the new loop.
1124 Builder.Insert(Cloned);
1126 // If the original scalar returns a value we need to place it in a vector
1127 // so that future users will be able to use it.
1129 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1130 Builder.getInt32(Width));
1136 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1138 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1139 Legal->getRuntimePointerCheck();
1141 if (!PtrRtCheck->Need)
1144 Instruction *MemoryRuntimeCheck = 0;
1145 unsigned NumPointers = PtrRtCheck->Pointers.size();
1146 SmallVector<Value* , 2> Starts;
1147 SmallVector<Value* , 2> Ends;
1149 SCEVExpander Exp(*SE, "induction");
1151 // Use this type for pointer arithmetic.
1152 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1154 for (unsigned i = 0; i < NumPointers; ++i) {
1155 Value *Ptr = PtrRtCheck->Pointers[i];
1156 const SCEV *Sc = SE->getSCEV(Ptr);
1158 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1159 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1161 Starts.push_back(Ptr);
1162 Ends.push_back(Ptr);
1164 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1166 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1167 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1168 Starts.push_back(Start);
1169 Ends.push_back(End);
1173 IRBuilder<> ChkBuilder(Loc);
1175 for (unsigned i = 0; i < NumPointers; ++i) {
1176 for (unsigned j = i+1; j < NumPointers; ++j) {
1177 // No need to check if two readonly pointers intersect.
1178 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1181 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1182 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1183 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1184 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1186 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1187 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1188 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1189 if (MemoryRuntimeCheck)
1190 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1193 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1197 return MemoryRuntimeCheck;
1201 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1203 In this function we generate a new loop. The new loop will contain
1204 the vectorized instructions while the old loop will continue to run the
1207 [ ] <-- vector loop bypass (may consist of multiple blocks).
1210 | [ ] <-- vector pre header.
1214 | [ ]_| <-- vector loop.
1217 >[ ] <--- middle-block.
1220 | [ ] <--- new preheader.
1224 | [ ]_| <-- old scalar loop to handle remainder.
1227 >[ ] <-- exit block.
1231 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1232 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1233 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1234 assert(ExitBlock && "Must have an exit block");
1236 // Mark the old scalar loop with metadata that tells us not to vectorize this
1237 // loop again if we run into it.
1238 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None);
1239 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD);
1241 // Some loops have a single integer induction variable, while other loops
1242 // don't. One example is c++ iterators that often have multiple pointer
1243 // induction variables. In the code below we also support a case where we
1244 // don't have a single induction variable.
1245 OldInduction = Legal->getInduction();
1246 Type *IdxTy = OldInduction ? OldInduction->getType() :
1247 DL->getIntPtrType(SE->getContext());
1249 // Find the loop boundaries.
1250 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1251 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1253 // Get the total trip count from the count by adding 1.
1254 ExitCount = SE->getAddExpr(ExitCount,
1255 SE->getConstant(ExitCount->getType(), 1));
1257 // Expand the trip count and place the new instructions in the preheader.
1258 // Notice that the pre-header does not change, only the loop body.
1259 SCEVExpander Exp(*SE, "induction");
1261 // Count holds the overall loop count (N).
1262 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1263 BypassBlock->getTerminator());
1265 // The loop index does not have to start at Zero. Find the original start
1266 // value from the induction PHI node. If we don't have an induction variable
1267 // then we know that it starts at zero.
1268 Value *StartIdx = OldInduction ?
1269 OldInduction->getIncomingValueForBlock(BypassBlock):
1270 ConstantInt::get(IdxTy, 0);
1272 assert(BypassBlock && "Invalid loop structure");
1273 LoopBypassBlocks.push_back(BypassBlock);
1275 // Split the single block loop into the two loop structure described above.
1276 BasicBlock *VectorPH =
1277 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1278 BasicBlock *VecBody =
1279 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1280 BasicBlock *MiddleBlock =
1281 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1282 BasicBlock *ScalarPH =
1283 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1285 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1287 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1289 // Generate the induction variable.
1290 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1291 // The loop step is equal to the vectorization factor (num of SIMD elements)
1292 // times the unroll factor (num of SIMD instructions).
1293 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1295 // This is the IR builder that we use to add all of the logic for bypassing
1296 // the new vector loop.
1297 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1299 // We may need to extend the index in case there is a type mismatch.
1300 // We know that the count starts at zero and does not overflow.
1301 if (Count->getType() != IdxTy) {
1302 // The exit count can be of pointer type. Convert it to the correct
1304 if (ExitCount->getType()->isPointerTy())
1305 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1307 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1310 // Add the start index to the loop count to get the new end index.
1311 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1313 // Now we need to generate the expression for N - (N % VF), which is
1314 // the part that the vectorized body will execute.
1315 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1316 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1317 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1318 "end.idx.rnd.down");
1320 // Now, compare the new count to zero. If it is zero skip the vector loop and
1321 // jump to the scalar loop.
1322 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1325 BasicBlock *LastBypassBlock = BypassBlock;
1327 // Generate the code that checks in runtime if arrays overlap. We put the
1328 // checks into a separate block to make the more common case of few elements
1330 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1331 BypassBlock->getTerminator());
1332 if (MemRuntimeCheck) {
1333 // Create a new block containing the memory check.
1334 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1336 LoopBypassBlocks.push_back(CheckBlock);
1338 // Replace the branch into the memory check block with a conditional branch
1339 // for the "few elements case".
1340 Instruction *OldTerm = BypassBlock->getTerminator();
1341 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1342 OldTerm->eraseFromParent();
1344 Cmp = MemRuntimeCheck;
1345 LastBypassBlock = CheckBlock;
1348 LastBypassBlock->getTerminator()->eraseFromParent();
1349 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1352 // We are going to resume the execution of the scalar loop.
1353 // Go over all of the induction variables that we found and fix the
1354 // PHIs that are left in the scalar version of the loop.
1355 // The starting values of PHI nodes depend on the counter of the last
1356 // iteration in the vectorized loop.
1357 // If we come from a bypass edge then we need to start from the original
1360 // This variable saves the new starting index for the scalar loop.
1361 PHINode *ResumeIndex = 0;
1362 LoopVectorizationLegality::InductionList::iterator I, E;
1363 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1364 // Set builder to point to last bypass block.
1365 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1366 for (I = List->begin(), E = List->end(); I != E; ++I) {
1367 PHINode *OrigPhi = I->first;
1368 LoopVectorizationLegality::InductionInfo II = I->second;
1369 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1370 MiddleBlock->getTerminator());
1371 Value *EndValue = 0;
1373 case LoopVectorizationLegality::IK_NoInduction:
1374 llvm_unreachable("Unknown induction");
1375 case LoopVectorizationLegality::IK_IntInduction: {
1376 // Handle the integer induction counter:
1377 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1378 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1379 // We know what the end value is.
1380 EndValue = IdxEndRoundDown;
1381 // We also know which PHI node holds it.
1382 ResumeIndex = ResumeVal;
1385 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1386 // Convert the CountRoundDown variable to the PHI size.
1387 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1388 II.StartValue->getType(),
1390 // Handle reverse integer induction counter.
1391 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1394 case LoopVectorizationLegality::IK_PtrInduction: {
1395 // For pointer induction variables, calculate the offset using
1397 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1401 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1402 // The value at the end of the loop for the reverse pointer is calculated
1403 // by creating a GEP with a negative index starting from the start value.
1404 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1405 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1407 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1413 // The new PHI merges the original incoming value, in case of a bypass,
1414 // or the value at the end of the vectorized loop.
1415 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1416 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1417 ResumeVal->addIncoming(EndValue, VecBody);
1419 // Fix the scalar body counter (PHI node).
1420 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1421 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1424 // If we are generating a new induction variable then we also need to
1425 // generate the code that calculates the exit value. This value is not
1426 // simply the end of the counter because we may skip the vectorized body
1427 // in case of a runtime check.
1429 assert(!ResumeIndex && "Unexpected resume value found");
1430 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1431 MiddleBlock->getTerminator());
1432 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1433 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1434 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1437 // Make sure that we found the index where scalar loop needs to continue.
1438 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1439 "Invalid resume Index");
1441 // Add a check in the middle block to see if we have completed
1442 // all of the iterations in the first vector loop.
1443 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1444 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1445 ResumeIndex, "cmp.n",
1446 MiddleBlock->getTerminator());
1448 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1449 // Remove the old terminator.
1450 MiddleBlock->getTerminator()->eraseFromParent();
1452 // Create i+1 and fill the PHINode.
1453 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1454 Induction->addIncoming(StartIdx, VectorPH);
1455 Induction->addIncoming(NextIdx, VecBody);
1456 // Create the compare.
1457 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1458 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1460 // Now we have two terminators. Remove the old one from the block.
1461 VecBody->getTerminator()->eraseFromParent();
1463 // Get ready to start creating new instructions into the vectorized body.
1464 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1466 // Create and register the new vector loop.
1467 Loop* Lp = new Loop();
1468 Loop *ParentLoop = OrigLoop->getParentLoop();
1470 // Insert the new loop into the loop nest and register the new basic blocks.
1472 ParentLoop->addChildLoop(Lp);
1473 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1474 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1475 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1476 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1477 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1479 LI->addTopLevelLoop(Lp);
1482 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1485 LoopVectorPreHeader = VectorPH;
1486 LoopScalarPreHeader = ScalarPH;
1487 LoopMiddleBlock = MiddleBlock;
1488 LoopExitBlock = ExitBlock;
1489 LoopVectorBody = VecBody;
1490 LoopScalarBody = OldBasicBlock;
1493 /// This function returns the identity element (or neutral element) for
1494 /// the operation K.
1496 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1501 // Adding, Xoring, Oring zero to a number does not change it.
1502 return ConstantInt::get(Tp, 0);
1503 case RK_IntegerMult:
1504 // Multiplying a number by 1 does not change it.
1505 return ConstantInt::get(Tp, 1);
1507 // AND-ing a number with an all-1 value does not change it.
1508 return ConstantInt::get(Tp, -1, true);
1510 // Multiplying a number by 1 does not change it.
1511 return ConstantFP::get(Tp, 1.0L);
1513 // Adding zero to a number does not change it.
1514 return ConstantFP::get(Tp, 0.0L);
1516 llvm_unreachable("Unknown reduction kind");
1520 static Intrinsic::ID
1521 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1522 // If we have an intrinsic call, check if it is trivially vectorizable.
1523 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1524 switch (II->getIntrinsicID()) {
1525 case Intrinsic::sqrt:
1526 case Intrinsic::sin:
1527 case Intrinsic::cos:
1528 case Intrinsic::exp:
1529 case Intrinsic::exp2:
1530 case Intrinsic::log:
1531 case Intrinsic::log10:
1532 case Intrinsic::log2:
1533 case Intrinsic::fabs:
1534 case Intrinsic::floor:
1535 case Intrinsic::ceil:
1536 case Intrinsic::trunc:
1537 case Intrinsic::rint:
1538 case Intrinsic::nearbyint:
1539 case Intrinsic::pow:
1540 case Intrinsic::fma:
1541 case Intrinsic::fmuladd:
1542 return II->getIntrinsicID();
1544 return Intrinsic::not_intrinsic;
1549 return Intrinsic::not_intrinsic;
1552 Function *F = CI->getCalledFunction();
1553 // We're going to make assumptions on the semantics of the functions, check
1554 // that the target knows that it's available in this environment.
1555 if (!F || !TLI->getLibFunc(F->getName(), Func))
1556 return Intrinsic::not_intrinsic;
1558 // Otherwise check if we have a call to a function that can be turned into a
1559 // vector intrinsic.
1566 return Intrinsic::sin;
1570 return Intrinsic::cos;
1574 return Intrinsic::exp;
1576 case LibFunc::exp2f:
1577 case LibFunc::exp2l:
1578 return Intrinsic::exp2;
1582 return Intrinsic::log;
1583 case LibFunc::log10:
1584 case LibFunc::log10f:
1585 case LibFunc::log10l:
1586 return Intrinsic::log10;
1588 case LibFunc::log2f:
1589 case LibFunc::log2l:
1590 return Intrinsic::log2;
1592 case LibFunc::fabsf:
1593 case LibFunc::fabsl:
1594 return Intrinsic::fabs;
1595 case LibFunc::floor:
1596 case LibFunc::floorf:
1597 case LibFunc::floorl:
1598 return Intrinsic::floor;
1600 case LibFunc::ceilf:
1601 case LibFunc::ceill:
1602 return Intrinsic::ceil;
1603 case LibFunc::trunc:
1604 case LibFunc::truncf:
1605 case LibFunc::truncl:
1606 return Intrinsic::trunc;
1608 case LibFunc::rintf:
1609 case LibFunc::rintl:
1610 return Intrinsic::rint;
1611 case LibFunc::nearbyint:
1612 case LibFunc::nearbyintf:
1613 case LibFunc::nearbyintl:
1614 return Intrinsic::nearbyint;
1618 return Intrinsic::pow;
1621 return Intrinsic::not_intrinsic;
1624 /// This function translates the reduction kind to an LLVM binary operator.
1626 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1628 case LoopVectorizationLegality::RK_IntegerAdd:
1629 return Instruction::Add;
1630 case LoopVectorizationLegality::RK_IntegerMult:
1631 return Instruction::Mul;
1632 case LoopVectorizationLegality::RK_IntegerOr:
1633 return Instruction::Or;
1634 case LoopVectorizationLegality::RK_IntegerAnd:
1635 return Instruction::And;
1636 case LoopVectorizationLegality::RK_IntegerXor:
1637 return Instruction::Xor;
1638 case LoopVectorizationLegality::RK_FloatMult:
1639 return Instruction::FMul;
1640 case LoopVectorizationLegality::RK_FloatAdd:
1641 return Instruction::FAdd;
1642 case LoopVectorizationLegality::RK_IntegerMinMax:
1643 return Instruction::ICmp;
1644 case LoopVectorizationLegality::RK_FloatMinMax:
1645 return Instruction::FCmp;
1647 llvm_unreachable("Unknown reduction operation");
1651 Value *createMinMaxOp(IRBuilder<> &Builder,
1652 LoopVectorizationLegality::MinMaxReductionKind RK,
1655 CmpInst::Predicate P = CmpInst::ICMP_NE;
1658 llvm_unreachable("Unknown min/max reduction kind");
1659 case LoopVectorizationLegality::MRK_UIntMin:
1660 P = CmpInst::ICMP_ULT;
1662 case LoopVectorizationLegality::MRK_UIntMax:
1663 P = CmpInst::ICMP_UGT;
1665 case LoopVectorizationLegality::MRK_SIntMin:
1666 P = CmpInst::ICMP_SLT;
1668 case LoopVectorizationLegality::MRK_SIntMax:
1669 P = CmpInst::ICMP_SGT;
1671 case LoopVectorizationLegality::MRK_FloatMin:
1672 P = CmpInst::FCMP_OLT;
1674 case LoopVectorizationLegality::MRK_FloatMax:
1675 P = CmpInst::FCMP_OGT;
1680 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1681 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1683 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1685 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1690 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1691 //===------------------------------------------------===//
1693 // Notice: any optimization or new instruction that go
1694 // into the code below should be also be implemented in
1697 //===------------------------------------------------===//
1698 Constant *Zero = Builder.getInt32(0);
1700 // In order to support reduction variables we need to be able to vectorize
1701 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1702 // stages. First, we create a new vector PHI node with no incoming edges.
1703 // We use this value when we vectorize all of the instructions that use the
1704 // PHI. Next, after all of the instructions in the block are complete we
1705 // add the new incoming edges to the PHI. At this point all of the
1706 // instructions in the basic block are vectorized, so we can use them to
1707 // construct the PHI.
1708 PhiVector RdxPHIsToFix;
1710 // Scan the loop in a topological order to ensure that defs are vectorized
1712 LoopBlocksDFS DFS(OrigLoop);
1715 // Vectorize all of the blocks in the original loop.
1716 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1717 be = DFS.endRPO(); bb != be; ++bb)
1718 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1720 // At this point every instruction in the original loop is widened to
1721 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1722 // that we vectorized. The PHI nodes are currently empty because we did
1723 // not want to introduce cycles. Notice that the remaining PHI nodes
1724 // that we need to fix are reduction variables.
1726 // Create the 'reduced' values for each of the induction vars.
1727 // The reduced values are the vector values that we scalarize and combine
1728 // after the loop is finished.
1729 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1731 PHINode *RdxPhi = *it;
1732 assert(RdxPhi && "Unable to recover vectorized PHI");
1734 // Find the reduction variable descriptor.
1735 assert(Legal->getReductionVars()->count(RdxPhi) &&
1736 "Unable to find the reduction variable");
1737 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1738 (*Legal->getReductionVars())[RdxPhi];
1740 // We need to generate a reduction vector from the incoming scalar.
1741 // To do so, we need to generate the 'identity' vector and overide
1742 // one of the elements with the incoming scalar reduction. We need
1743 // to do it in the vector-loop preheader.
1744 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1746 // This is the vector-clone of the value that leaves the loop.
1747 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1748 Type *VecTy = VectorExit[0]->getType();
1750 // Find the reduction identity variable. Zero for addition, or, xor,
1751 // one for multiplication, -1 for And.
1754 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
1755 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
1756 // MinMax reduction have the start value as their identify.
1757 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
1761 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1762 VecTy->getScalarType());
1763 Identity = ConstantVector::getSplat(VF, Iden);
1765 // This vector is the Identity vector where the first element is the
1766 // incoming scalar reduction.
1767 VectorStart = Builder.CreateInsertElement(Identity,
1768 RdxDesc.StartValue, Zero);
1771 // Fix the vector-loop phi.
1772 // We created the induction variable so we know that the
1773 // preheader is the first entry.
1774 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1776 // Reductions do not have to start at zero. They can start with
1777 // any loop invariant values.
1778 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1779 BasicBlock *Latch = OrigLoop->getLoopLatch();
1780 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1781 VectorParts &Val = getVectorValue(LoopVal);
1782 for (unsigned part = 0; part < UF; ++part) {
1783 // Make sure to add the reduction stat value only to the
1784 // first unroll part.
1785 Value *StartVal = (part == 0) ? VectorStart : Identity;
1786 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1787 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1790 // Before each round, move the insertion point right between
1791 // the PHIs and the values we are going to write.
1792 // This allows us to write both PHINodes and the extractelement
1794 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1796 VectorParts RdxParts;
1797 for (unsigned part = 0; part < UF; ++part) {
1798 // This PHINode contains the vectorized reduction variable, or
1799 // the initial value vector, if we bypass the vector loop.
1800 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1801 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1802 Value *StartVal = (part == 0) ? VectorStart : Identity;
1803 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1804 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1805 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1806 RdxParts.push_back(NewPhi);
1809 // Reduce all of the unrolled parts into a single vector.
1810 Value *ReducedPartRdx = RdxParts[0];
1811 unsigned Op = getReductionBinOp(RdxDesc.Kind);
1812 for (unsigned part = 1; part < UF; ++part) {
1813 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1814 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
1815 RdxParts[part], ReducedPartRdx,
1818 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
1819 ReducedPartRdx, RdxParts[part]);
1822 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1823 // and vector ops, reducing the set of values being computed by half each
1825 assert(isPowerOf2_32(VF) &&
1826 "Reduction emission only supported for pow2 vectors!");
1827 Value *TmpVec = ReducedPartRdx;
1828 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1829 for (unsigned i = VF; i != 1; i >>= 1) {
1830 // Move the upper half of the vector to the lower half.
1831 for (unsigned j = 0; j != i/2; ++j)
1832 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1834 // Fill the rest of the mask with undef.
1835 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1836 UndefValue::get(Builder.getInt32Ty()));
1839 Builder.CreateShuffleVector(TmpVec,
1840 UndefValue::get(TmpVec->getType()),
1841 ConstantVector::get(ShuffleMask),
1844 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1845 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1848 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
1851 // The result is in the first element of the vector.
1852 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1854 // Now, we need to fix the users of the reduction variable
1855 // inside and outside of the scalar remainder loop.
1856 // We know that the loop is in LCSSA form. We need to update the
1857 // PHI nodes in the exit blocks.
1858 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1859 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1860 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1861 if (!LCSSAPhi) continue;
1863 // All PHINodes need to have a single entry edge, or two if
1864 // we already fixed them.
1865 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1867 // We found our reduction value exit-PHI. Update it with the
1868 // incoming bypass edge.
1869 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1870 // Add an edge coming from the bypass.
1871 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1874 }// end of the LCSSA phi scan.
1876 // Fix the scalar loop reduction variable with the incoming reduction sum
1877 // from the vector body and from the backedge value.
1878 int IncomingEdgeBlockIdx =
1879 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1880 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1881 // Pick the other block.
1882 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1883 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1884 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1885 }// end of for each redux variable.
1887 // The Loop exit block may have single value PHI nodes where the incoming
1888 // value is 'undef'. While vectorizing we only handled real values that
1889 // were defined inside the loop. Here we handle the 'undef case'.
1891 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1892 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1893 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1894 if (!LCSSAPhi) continue;
1895 if (LCSSAPhi->getNumIncomingValues() == 1)
1896 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1901 InnerLoopVectorizer::VectorParts
1902 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1903 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1906 VectorParts SrcMask = createBlockInMask(Src);
1908 // The terminator has to be a branch inst!
1909 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1910 assert(BI && "Unexpected terminator found");
1912 if (BI->isConditional()) {
1913 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1915 if (BI->getSuccessor(0) != Dst)
1916 for (unsigned part = 0; part < UF; ++part)
1917 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1919 for (unsigned part = 0; part < UF; ++part)
1920 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1927 InnerLoopVectorizer::VectorParts
1928 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1929 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1931 // Loop incoming mask is all-one.
1932 if (OrigLoop->getHeader() == BB) {
1933 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1934 return getVectorValue(C);
1937 // This is the block mask. We OR all incoming edges, and with zero.
1938 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1939 VectorParts BlockMask = getVectorValue(Zero);
1942 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1943 VectorParts EM = createEdgeMask(*it, BB);
1944 for (unsigned part = 0; part < UF; ++part)
1945 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1952 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1953 BasicBlock *BB, PhiVector *PV) {
1954 // For each instruction in the old loop.
1955 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1956 VectorParts &Entry = WidenMap.get(it);
1957 switch (it->getOpcode()) {
1958 case Instruction::Br:
1959 // Nothing to do for PHIs and BR, since we already took care of the
1960 // loop control flow instructions.
1962 case Instruction::PHI:{
1963 PHINode* P = cast<PHINode>(it);
1964 // Handle reduction variables:
1965 if (Legal->getReductionVars()->count(P)) {
1966 for (unsigned part = 0; part < UF; ++part) {
1967 // This is phase one of vectorizing PHIs.
1968 Type *VecTy = VectorType::get(it->getType(), VF);
1969 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1970 LoopVectorBody-> getFirstInsertionPt());
1976 // Check for PHI nodes that are lowered to vector selects.
1977 if (P->getParent() != OrigLoop->getHeader()) {
1978 // We know that all PHIs in non header blocks are converted into
1979 // selects, so we don't have to worry about the insertion order and we
1980 // can just use the builder.
1981 // At this point we generate the predication tree. There may be
1982 // duplications since this is a simple recursive scan, but future
1983 // optimizations will clean it up.
1985 unsigned NumIncoming = P->getNumIncomingValues();
1986 assert(NumIncoming > 1 && "Invalid PHI");
1988 // Generate a sequence of selects of the form:
1989 // SELECT(Mask3, In3,
1990 // SELECT(Mask2, In2,
1992 for (unsigned In = 0; In < NumIncoming; In++) {
1993 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
1995 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
1997 for (unsigned part = 0; part < UF; ++part) {
1998 // We don't need to 'select' the first PHI operand because it is
1999 // the default value if all of the other masks don't match.
2001 Entry[part] = In0[part];
2003 // Select between the current value and the previous incoming edge
2004 // based on the incoming mask.
2005 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2006 Entry[part], "predphi");
2012 // This PHINode must be an induction variable.
2013 // Make sure that we know about it.
2014 assert(Legal->getInductionVars()->count(P) &&
2015 "Not an induction variable");
2017 LoopVectorizationLegality::InductionInfo II =
2018 Legal->getInductionVars()->lookup(P);
2021 case LoopVectorizationLegality::IK_NoInduction:
2022 llvm_unreachable("Unknown induction");
2023 case LoopVectorizationLegality::IK_IntInduction: {
2024 assert(P == OldInduction && "Unexpected PHI");
2025 Value *Broadcasted = getBroadcastInstrs(Induction);
2026 // After broadcasting the induction variable we need to make the
2027 // vector consecutive by adding 0, 1, 2 ...
2028 for (unsigned part = 0; part < UF; ++part)
2029 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2032 case LoopVectorizationLegality::IK_ReverseIntInduction:
2033 case LoopVectorizationLegality::IK_PtrInduction:
2034 case LoopVectorizationLegality::IK_ReversePtrInduction:
2035 // Handle reverse integer and pointer inductions.
2036 Value *StartIdx = 0;
2037 // If we have a single integer induction variable then use it.
2038 // Otherwise, start counting at zero.
2040 LoopVectorizationLegality::InductionInfo OldII =
2041 Legal->getInductionVars()->lookup(OldInduction);
2042 StartIdx = OldII.StartValue;
2044 StartIdx = ConstantInt::get(Induction->getType(), 0);
2046 // This is the normalized GEP that starts counting at zero.
2047 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2050 // Handle the reverse integer induction variable case.
2051 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2052 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2053 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2055 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2058 // This is a new value so do not hoist it out.
2059 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2060 // After broadcasting the induction variable we need to make the
2061 // vector consecutive by adding ... -3, -2, -1, 0.
2062 for (unsigned part = 0; part < UF; ++part)
2063 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2068 // Handle the pointer induction variable case.
2069 assert(P->getType()->isPointerTy() && "Unexpected type.");
2071 // Is this a reverse induction ptr or a consecutive induction ptr.
2072 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2075 // This is the vector of results. Notice that we don't generate
2076 // vector geps because scalar geps result in better code.
2077 for (unsigned part = 0; part < UF; ++part) {
2078 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2079 for (unsigned int i = 0; i < VF; ++i) {
2080 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2081 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2084 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2086 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2088 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2090 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2091 Builder.getInt32(i),
2094 Entry[part] = VecVal;
2101 case Instruction::Add:
2102 case Instruction::FAdd:
2103 case Instruction::Sub:
2104 case Instruction::FSub:
2105 case Instruction::Mul:
2106 case Instruction::FMul:
2107 case Instruction::UDiv:
2108 case Instruction::SDiv:
2109 case Instruction::FDiv:
2110 case Instruction::URem:
2111 case Instruction::SRem:
2112 case Instruction::FRem:
2113 case Instruction::Shl:
2114 case Instruction::LShr:
2115 case Instruction::AShr:
2116 case Instruction::And:
2117 case Instruction::Or:
2118 case Instruction::Xor: {
2119 // Just widen binops.
2120 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2121 VectorParts &A = getVectorValue(it->getOperand(0));
2122 VectorParts &B = getVectorValue(it->getOperand(1));
2124 // Use this vector value for all users of the original instruction.
2125 for (unsigned Part = 0; Part < UF; ++Part) {
2126 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2128 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2129 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2130 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2131 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2132 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2134 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2135 VecOp->setIsExact(BinOp->isExact());
2141 case Instruction::Select: {
2143 // If the selector is loop invariant we can create a select
2144 // instruction with a scalar condition. Otherwise, use vector-select.
2145 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2148 // The condition can be loop invariant but still defined inside the
2149 // loop. This means that we can't just use the original 'cond' value.
2150 // We have to take the 'vectorized' value and pick the first lane.
2151 // Instcombine will make this a no-op.
2152 VectorParts &Cond = getVectorValue(it->getOperand(0));
2153 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2154 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2155 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2156 Builder.getInt32(0));
2157 for (unsigned Part = 0; Part < UF; ++Part) {
2158 Entry[Part] = Builder.CreateSelect(
2159 InvariantCond ? ScalarCond : Cond[Part],
2166 case Instruction::ICmp:
2167 case Instruction::FCmp: {
2168 // Widen compares. Generate vector compares.
2169 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2170 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2171 VectorParts &A = getVectorValue(it->getOperand(0));
2172 VectorParts &B = getVectorValue(it->getOperand(1));
2173 for (unsigned Part = 0; Part < UF; ++Part) {
2176 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2178 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2184 case Instruction::Store:
2185 case Instruction::Load:
2186 vectorizeMemoryInstruction(it, Legal);
2188 case Instruction::ZExt:
2189 case Instruction::SExt:
2190 case Instruction::FPToUI:
2191 case Instruction::FPToSI:
2192 case Instruction::FPExt:
2193 case Instruction::PtrToInt:
2194 case Instruction::IntToPtr:
2195 case Instruction::SIToFP:
2196 case Instruction::UIToFP:
2197 case Instruction::Trunc:
2198 case Instruction::FPTrunc:
2199 case Instruction::BitCast: {
2200 CastInst *CI = dyn_cast<CastInst>(it);
2201 /// Optimize the special case where the source is the induction
2202 /// variable. Notice that we can only optimize the 'trunc' case
2203 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2204 /// c. other casts depend on pointer size.
2205 if (CI->getOperand(0) == OldInduction &&
2206 it->getOpcode() == Instruction::Trunc) {
2207 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2209 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2210 for (unsigned Part = 0; Part < UF; ++Part)
2211 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2214 /// Vectorize casts.
2215 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2217 VectorParts &A = getVectorValue(it->getOperand(0));
2218 for (unsigned Part = 0; Part < UF; ++Part)
2219 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2223 case Instruction::Call: {
2224 // Ignore dbg intrinsics.
2225 if (isa<DbgInfoIntrinsic>(it))
2228 Module *M = BB->getParent()->getParent();
2229 CallInst *CI = cast<CallInst>(it);
2230 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2231 assert(ID && "Not an intrinsic call!");
2232 for (unsigned Part = 0; Part < UF; ++Part) {
2233 SmallVector<Value*, 4> Args;
2234 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2235 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2236 Args.push_back(Arg[Part]);
2238 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2239 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2240 Entry[Part] = Builder.CreateCall(F, Args);
2246 // All other instructions are unsupported. Scalarize them.
2247 scalarizeInstruction(it);
2250 }// end of for_each instr.
2253 void InnerLoopVectorizer::updateAnalysis() {
2254 // Forget the original basic block.
2255 SE->forgetLoop(OrigLoop);
2257 // Update the dominator tree information.
2258 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2259 "Entry does not dominate exit.");
2261 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2262 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2263 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2264 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2265 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2266 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2267 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2268 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2270 DEBUG(DT->verifyAnalysis());
2273 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2274 if (!EnableIfConversion)
2277 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2278 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2280 // Collect the blocks that need predication.
2281 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2282 BasicBlock *BB = LoopBlocks[i];
2284 // We don't support switch statements inside loops.
2285 if (!isa<BranchInst>(BB->getTerminator()))
2288 // We must be able to predicate all blocks that need to be predicated.
2289 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2293 // We can if-convert this loop.
2297 bool LoopVectorizationLegality::canVectorize() {
2298 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2300 // We can only vectorize innermost loops.
2301 if (TheLoop->getSubLoopsVector().size())
2304 // We must have a single backedge.
2305 if (TheLoop->getNumBackEdges() != 1)
2308 // We must have a single exiting block.
2309 if (!TheLoop->getExitingBlock())
2312 unsigned NumBlocks = TheLoop->getNumBlocks();
2314 // Check if we can if-convert non single-bb loops.
2315 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2316 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2320 // We need to have a loop header.
2321 BasicBlock *Latch = TheLoop->getLoopLatch();
2322 DEBUG(dbgs() << "LV: Found a loop: " <<
2323 TheLoop->getHeader()->getName() << "\n");
2325 // ScalarEvolution needs to be able to find the exit count.
2326 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2327 if (ExitCount == SE->getCouldNotCompute()) {
2328 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2332 // Do not loop-vectorize loops with a tiny trip count.
2333 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2334 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2335 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2336 "This loop is not worth vectorizing.\n");
2340 // Check if we can vectorize the instructions and CFG in this loop.
2341 if (!canVectorizeInstrs()) {
2342 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2346 // Go over each instruction and look at memory deps.
2347 if (!canVectorizeMemory()) {
2348 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2352 // Collect all of the variables that remain uniform after vectorization.
2353 collectLoopUniforms();
2355 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2356 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2359 // Okay! We can vectorize. At this point we don't have any other mem analysis
2360 // which may limit our maximum vectorization factor, so just return true with
2365 bool LoopVectorizationLegality::canVectorizeInstrs() {
2366 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2367 BasicBlock *Header = TheLoop->getHeader();
2369 // If we marked the scalar loop as "already vectorized" then no need
2370 // to vectorize it again.
2371 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) {
2372 DEBUG(dbgs() << "LV: This loop was vectorized before\n");
2376 // Look for the attribute signaling the absence of NaNs.
2377 Function &F = *Header->getParent();
2378 if (F.hasFnAttribute("no-nans-fp-math"))
2379 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2380 AttributeSet::FunctionIndex,
2381 "no-nans-fp-math").getValueAsString() == "true";
2383 // For each block in the loop.
2384 for (Loop::block_iterator bb = TheLoop->block_begin(),
2385 be = TheLoop->block_end(); bb != be; ++bb) {
2387 // Scan the instructions in the block and look for hazards.
2388 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2391 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2392 Type *PhiTy = Phi->getType();
2393 // Check that this PHI type is allowed.
2394 if (!PhiTy->isIntegerTy() &&
2395 !PhiTy->isFloatingPointTy() &&
2396 !PhiTy->isPointerTy()) {
2397 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2401 // If this PHINode is not in the header block, then we know that we
2402 // can convert it to select during if-conversion. No need to check if
2403 // the PHIs in this block are induction or reduction variables.
2407 // We only allow if-converted PHIs with more than two incoming values.
2408 if (Phi->getNumIncomingValues() != 2) {
2409 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2413 // This is the value coming from the preheader.
2414 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2415 // Check if this is an induction variable.
2416 InductionKind IK = isInductionVariable(Phi);
2418 if (IK_NoInduction != IK) {
2419 // Int inductions are special because we only allow one IV.
2420 if (IK == IK_IntInduction) {
2422 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2428 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2429 Inductions[Phi] = InductionInfo(StartValue, IK);
2433 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2434 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2437 if (AddReductionVar(Phi, RK_IntegerMult)) {
2438 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2441 if (AddReductionVar(Phi, RK_IntegerOr)) {
2442 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2445 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2446 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2449 if (AddReductionVar(Phi, RK_IntegerXor)) {
2450 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2453 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2454 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2457 if (AddReductionVar(Phi, RK_FloatMult)) {
2458 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2461 if (AddReductionVar(Phi, RK_FloatAdd)) {
2462 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2465 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2466 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2470 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2472 }// end of PHI handling
2474 // We still don't handle functions. However, we can ignore dbg intrinsic
2475 // calls and we do handle certain intrinsic and libm functions.
2476 CallInst *CI = dyn_cast<CallInst>(it);
2477 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2478 DEBUG(dbgs() << "LV: Found a call site.\n");
2482 // Check that the instruction return type is vectorizable.
2483 if (!VectorType::isValidElementType(it->getType()) &&
2484 !it->getType()->isVoidTy()) {
2485 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2489 // Check that the stored type is vectorizable.
2490 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2491 Type *T = ST->getValueOperand()->getType();
2492 if (!VectorType::isValidElementType(T))
2496 // Reduction instructions are allowed to have exit users.
2497 // All other instructions must not have external users.
2498 if (!AllowedExit.count(it))
2499 //Check that all of the users of the loop are inside the BB.
2500 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2502 Instruction *U = cast<Instruction>(*I);
2503 // This user may be a reduction exit value.
2504 if (!TheLoop->contains(U)) {
2505 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2514 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2515 if (Inductions.empty())
2522 void LoopVectorizationLegality::collectLoopUniforms() {
2523 // We now know that the loop is vectorizable!
2524 // Collect variables that will remain uniform after vectorization.
2525 std::vector<Value*> Worklist;
2526 BasicBlock *Latch = TheLoop->getLoopLatch();
2528 // Start with the conditional branch and walk up the block.
2529 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2531 while (Worklist.size()) {
2532 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2533 Worklist.pop_back();
2535 // Look at instructions inside this loop.
2536 // Stop when reaching PHI nodes.
2537 // TODO: we need to follow values all over the loop, not only in this block.
2538 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2541 // This is a known uniform.
2544 // Insert all operands.
2545 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2546 Worklist.push_back(I->getOperand(i));
2551 AliasAnalysis::Location
2552 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2553 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2554 return AA->getLocation(Store);
2555 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2556 return AA->getLocation(Load);
2558 llvm_unreachable("Should be either load or store instruction");
2562 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2565 AliasMultiMap& WriteObjects,
2566 unsigned MaxByteWidth) {
2568 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2570 std::vector<Instruction*>::iterator
2571 it = WriteObjects[Object].begin(),
2572 end = WriteObjects[Object].end();
2574 for (; it != end; ++it) {
2575 Instruction* I = *it;
2579 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2580 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2581 ThatLoc.getWithNewSize(MaxByteWidth)))
2587 bool LoopVectorizationLegality::canVectorizeMemory() {
2589 typedef SmallVector<Value*, 16> ValueVector;
2590 typedef SmallPtrSet<Value*, 16> ValueSet;
2591 // Holds the Load and Store *instructions*.
2594 PtrRtCheck.Pointers.clear();
2595 PtrRtCheck.Need = false;
2597 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2600 for (Loop::block_iterator bb = TheLoop->block_begin(),
2601 be = TheLoop->block_end(); bb != be; ++bb) {
2603 // Scan the BB and collect legal loads and stores.
2604 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2607 // If this is a load, save it. If this instruction can read from memory
2608 // but is not a load, then we quit. Notice that we don't handle function
2609 // calls that read or write.
2610 if (it->mayReadFromMemory()) {
2611 LoadInst *Ld = dyn_cast<LoadInst>(it);
2612 if (!Ld) return false;
2613 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2614 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2617 Loads.push_back(Ld);
2621 // Save 'store' instructions. Abort if other instructions write to memory.
2622 if (it->mayWriteToMemory()) {
2623 StoreInst *St = dyn_cast<StoreInst>(it);
2624 if (!St) return false;
2625 if (!St->isSimple() && !IsAnnotatedParallel) {
2626 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2629 Stores.push_back(St);
2634 // Now we have two lists that hold the loads and the stores.
2635 // Next, we find the pointers that they use.
2637 // Check if we see any stores. If there are no stores, then we don't
2638 // care if the pointers are *restrict*.
2639 if (!Stores.size()) {
2640 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2644 // Holds the read and read-write *pointers* that we find. These maps hold
2645 // unique values for pointers (so no need for multi-map).
2647 AliasMap ReadWrites;
2649 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2650 // multiple times on the same object. If the ptr is accessed twice, once
2651 // for read and once for write, it will only appear once (on the write
2652 // list). This is okay, since we are going to check for conflicts between
2653 // writes and between reads and writes, but not between reads and reads.
2656 ValueVector::iterator I, IE;
2657 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2658 StoreInst *ST = cast<StoreInst>(*I);
2659 Value* Ptr = ST->getPointerOperand();
2661 if (isUniform(Ptr)) {
2662 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2666 // If we did *not* see this pointer before, insert it to
2667 // the read-write list. At this phase it is only a 'write' list.
2668 if (Seen.insert(Ptr))
2669 ReadWrites.insert(std::make_pair(Ptr, ST));
2672 if (IsAnnotatedParallel) {
2674 << "LV: A loop annotated parallel, ignore memory dependency "
2679 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2680 LoadInst *LD = cast<LoadInst>(*I);
2681 Value* Ptr = LD->getPointerOperand();
2682 // If we did *not* see this pointer before, insert it to the
2683 // read list. If we *did* see it before, then it is already in
2684 // the read-write list. This allows us to vectorize expressions
2685 // such as A[i] += x; Because the address of A[i] is a read-write
2686 // pointer. This only works if the index of A[i] is consecutive.
2687 // If the address of i is unknown (for example A[B[i]]) then we may
2688 // read a few words, modify, and write a few words, and some of the
2689 // words may be written to the same address.
2690 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2691 Reads.insert(std::make_pair(Ptr, LD));
2694 // If we write (or read-write) to a single destination and there are no
2695 // other reads in this loop then is it safe to vectorize.
2696 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2697 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2701 unsigned NumReadPtrs = 0;
2702 unsigned NumWritePtrs = 0;
2704 // Find pointers with computable bounds. We are going to use this information
2705 // to place a runtime bound check.
2706 bool CanDoRT = true;
2707 AliasMap::iterator MI, ME;
2708 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2709 Value *V = (*MI).first;
2710 if (hasComputableBounds(V)) {
2711 PtrRtCheck.insert(SE, TheLoop, V, true);
2713 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2719 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2720 Value *V = (*MI).first;
2721 if (hasComputableBounds(V)) {
2722 PtrRtCheck.insert(SE, TheLoop, V, false);
2724 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2731 // Check that we did not collect too many pointers or found a
2732 // unsizeable pointer.
2733 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
2734 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
2735 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
2741 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2744 bool NeedRTCheck = false;
2746 // Biggest vectorized access possible, vector width * unroll factor.
2747 // TODO: We're being very pessimistic here, find a way to know the
2748 // real access width before getting here.
2749 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2750 TTI->getMaximumUnrollFactor();
2751 // Now that the pointers are in two lists (Reads and ReadWrites), we
2752 // can check that there are no conflicts between each of the writes and
2753 // between the writes to the reads.
2754 // Note that WriteObjects duplicates the stores (indexed now by underlying
2755 // objects) to avoid pointing to elements inside ReadWrites.
2756 // TODO: Maybe create a new type where they can interact without duplication.
2757 AliasMultiMap WriteObjects;
2758 ValueVector TempObjects;
2760 // Check that the read-writes do not conflict with other read-write
2762 bool AllWritesIdentified = true;
2763 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2764 Value *Val = (*MI).first;
2765 Instruction *Inst = (*MI).second;
2767 GetUnderlyingObjects(Val, TempObjects, DL);
2768 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2770 if (!isIdentifiedObject(*UI)) {
2771 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2773 AllWritesIdentified = false;
2776 // Never seen it before, can't alias.
2777 if (WriteObjects[*UI].empty()) {
2778 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2779 WriteObjects[*UI].push_back(Inst);
2782 // Direct alias found.
2783 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2784 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2788 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2790 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2791 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2793 // If global alias, make sure they do alias.
2794 if (hasPossibleGlobalWriteReorder(*UI,
2798 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
2803 // Didn't alias, insert into map for further reference.
2804 WriteObjects[*UI].push_back(Inst);
2806 TempObjects.clear();
2809 /// Check that the reads don't conflict with the read-writes.
2810 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2811 Value *Val = (*MI).first;
2812 GetUnderlyingObjects(Val, TempObjects, DL);
2813 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2815 // If all of the writes are identified then we don't care if the read
2816 // pointer is identified or not.
2817 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2818 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2822 // Never seen it before, can't alias.
2823 if (WriteObjects[*UI].empty())
2825 // Direct alias found.
2826 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2827 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2831 DEBUG(dbgs() << "LV: Found a global value: "
2833 Instruction *Inst = (*MI).second;
2834 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2835 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2837 // If global alias, make sure they do alias.
2838 if (hasPossibleGlobalWriteReorder(*UI,
2842 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
2847 TempObjects.clear();
2850 PtrRtCheck.Need = NeedRTCheck;
2851 if (NeedRTCheck && !CanDoRT) {
2852 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2853 "the array bounds.\n");
2858 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2859 " need a runtime memory check.\n");
2863 static bool hasMultipleUsesOf(Instruction *I,
2864 SmallPtrSet<Instruction *, 8> &Insts) {
2865 unsigned NumUses = 0;
2866 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
2867 if (Insts.count(dyn_cast<Instruction>(*Use)))
2876 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
2877 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
2878 if (!Set.count(dyn_cast<Instruction>(*Use)))
2883 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2884 ReductionKind Kind) {
2885 if (Phi->getNumIncomingValues() != 2)
2888 // Reduction variables are only found in the loop header block.
2889 if (Phi->getParent() != TheLoop->getHeader())
2892 // Obtain the reduction start value from the value that comes from the loop
2894 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2896 // ExitInstruction is the single value which is used outside the loop.
2897 // We only allow for a single reduction value to be used outside the loop.
2898 // This includes users of the reduction, variables (which form a cycle
2899 // which ends in the phi node).
2900 Instruction *ExitInstruction = 0;
2901 // Indicates that we found a reduction operation in our scan.
2902 bool FoundReduxOp = false;
2904 // We start with the PHI node and scan for all of the users of this
2905 // instruction. All users must be instructions that can be used as reduction
2906 // variables (such as ADD). We must have a single out-of-block user. The cycle
2907 // must include the original PHI.
2908 bool FoundStartPHI = false;
2910 // To recognize min/max patterns formed by a icmp select sequence, we store
2911 // the number of instruction we saw from the recognized min/max pattern,
2912 // to make sure we only see exactly the two instructions.
2913 unsigned NumCmpSelectPatternInst = 0;
2914 ReductionInstDesc ReduxDesc(false, 0);
2916 SmallPtrSet<Instruction *, 8> VisitedInsts;
2917 SmallVector<Instruction *, 8> Worklist;
2918 Worklist.push_back(Phi);
2919 VisitedInsts.insert(Phi);
2921 // A value in the reduction can be used:
2922 // - By the reduction:
2923 // - Reduction operation:
2924 // - One use of reduction value (safe).
2925 // - Multiple use of reduction value (not safe).
2927 // - All uses of the PHI must be the reduction (safe).
2928 // - Otherwise, not safe.
2929 // - By one instruction outside of the loop (safe).
2930 // - By further instructions outside of the loop (not safe).
2931 // - By an instruction that is not part of the reduction (not safe).
2933 // * An instruction type other than PHI or the reduction operation.
2934 // * A PHI in the header other than the initial PHI.
2935 while (!Worklist.empty()) {
2936 Instruction *Cur = Worklist.back();
2937 Worklist.pop_back();
2940 // If the instruction has no users then this is a broken chain and can't be
2941 // a reduction variable.
2942 if (Cur->use_empty())
2945 bool IsAPhi = isa<PHINode>(Cur);
2947 // A header PHI use other than the original PHI.
2948 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
2951 // Reductions of instructions such as Div, and Sub is only possible if the
2952 // LHS is the reduction variable.
2953 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
2954 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
2955 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
2958 // Any reduction instruction must be of one of the allowed kinds.
2959 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
2960 if (!ReduxDesc.IsReduction)
2963 // A reduction operation must only have one use of the reduction value.
2964 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
2965 hasMultipleUsesOf(Cur, VisitedInsts))
2968 // All inputs to a PHI node must be a reduction value.
2969 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
2972 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
2973 isa<SelectInst>(Cur)))
2974 ++NumCmpSelectPatternInst;
2975 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
2976 isa<SelectInst>(Cur)))
2977 ++NumCmpSelectPatternInst;
2979 // Check whether we found a reduction operator.
2980 FoundReduxOp |= !IsAPhi;
2982 // Process users of current instruction. Push non PHI nodes after PHI nodes
2983 // onto the stack. This way we are going to have seen all inputs to PHI
2984 // nodes once we get to them.
2985 SmallVector<Instruction *, 8> NonPHIs;
2986 SmallVector<Instruction *, 8> PHIs;
2987 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
2989 Instruction *Usr = cast<Instruction>(*UI);
2991 // Check if we found the exit user.
2992 BasicBlock *Parent = Usr->getParent();
2993 if (!TheLoop->contains(Parent)) {
2994 // Exit if you find multiple outside users.
2995 if (ExitInstruction != 0)
2997 ExitInstruction = Cur;
3001 // Process instructions only once (termination).
3002 if (VisitedInsts.insert(Usr)) {
3003 if (isa<PHINode>(Usr))
3004 PHIs.push_back(Usr);
3006 NonPHIs.push_back(Usr);
3008 // Remember that we completed the cycle.
3010 FoundStartPHI = true;
3012 Worklist.append(PHIs.begin(), PHIs.end());
3013 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3016 // This means we have seen one but not the other instruction of the
3017 // pattern or more than just a select and cmp.
3018 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3019 NumCmpSelectPatternInst != 2)
3022 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3025 // We found a reduction var if we have reached the original phi node and we
3026 // only have a single instruction with out-of-loop users.
3028 // This instruction is allowed to have out-of-loop users.
3029 AllowedExit.insert(ExitInstruction);
3031 // Save the description of this reduction variable.
3032 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3033 ReduxDesc.MinMaxKind);
3034 Reductions[Phi] = RD;
3035 // We've ended the cycle. This is a reduction variable if we have an
3036 // outside user and it has a binary op.
3041 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3042 /// pattern corresponding to a min(X, Y) or max(X, Y).
3043 LoopVectorizationLegality::ReductionInstDesc
3044 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3045 ReductionInstDesc &Prev) {
3047 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3048 "Expect a select instruction");
3049 Instruction *Cmp = 0;
3050 SelectInst *Select = 0;
3052 // We must handle the select(cmp()) as a single instruction. Advance to the
3054 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3055 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3056 return ReductionInstDesc(false, I);
3057 return ReductionInstDesc(Select, Prev.MinMaxKind);
3060 // Only handle single use cases for now.
3061 if (!(Select = dyn_cast<SelectInst>(I)))
3062 return ReductionInstDesc(false, I);
3063 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3064 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3065 return ReductionInstDesc(false, I);
3066 if (!Cmp->hasOneUse())
3067 return ReductionInstDesc(false, I);
3072 // Look for a min/max pattern.
3073 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3074 return ReductionInstDesc(Select, MRK_UIntMin);
3075 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3076 return ReductionInstDesc(Select, MRK_UIntMax);
3077 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3078 return ReductionInstDesc(Select, MRK_SIntMax);
3079 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3080 return ReductionInstDesc(Select, MRK_SIntMin);
3081 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3082 return ReductionInstDesc(Select, MRK_FloatMin);
3083 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3084 return ReductionInstDesc(Select, MRK_FloatMax);
3085 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3086 return ReductionInstDesc(Select, MRK_FloatMin);
3087 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3088 return ReductionInstDesc(Select, MRK_FloatMax);
3090 return ReductionInstDesc(false, I);
3093 LoopVectorizationLegality::ReductionInstDesc
3094 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3096 ReductionInstDesc &Prev) {
3097 bool FP = I->getType()->isFloatingPointTy();
3098 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3099 switch (I->getOpcode()) {
3101 return ReductionInstDesc(false, I);
3102 case Instruction::PHI:
3103 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3104 Kind != RK_FloatMinMax))
3105 return ReductionInstDesc(false, I);
3106 return ReductionInstDesc(I, Prev.MinMaxKind);
3107 case Instruction::Sub:
3108 case Instruction::Add:
3109 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3110 case Instruction::Mul:
3111 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3112 case Instruction::And:
3113 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3114 case Instruction::Or:
3115 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3116 case Instruction::Xor:
3117 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3118 case Instruction::FMul:
3119 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3120 case Instruction::FAdd:
3121 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3122 case Instruction::FCmp:
3123 case Instruction::ICmp:
3124 case Instruction::Select:
3125 if (Kind != RK_IntegerMinMax &&
3126 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3127 return ReductionInstDesc(false, I);
3128 return isMinMaxSelectCmpPattern(I, Prev);
3132 LoopVectorizationLegality::InductionKind
3133 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3134 Type *PhiTy = Phi->getType();
3135 // We only handle integer and pointer inductions variables.
3136 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3137 return IK_NoInduction;
3139 // Check that the PHI is consecutive.
3140 const SCEV *PhiScev = SE->getSCEV(Phi);
3141 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3143 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3144 return IK_NoInduction;
3146 const SCEV *Step = AR->getStepRecurrence(*SE);
3148 // Integer inductions need to have a stride of one.
3149 if (PhiTy->isIntegerTy()) {
3151 return IK_IntInduction;
3152 if (Step->isAllOnesValue())
3153 return IK_ReverseIntInduction;
3154 return IK_NoInduction;
3157 // Calculate the pointer stride and check if it is consecutive.
3158 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3160 return IK_NoInduction;
3162 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3163 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3164 if (C->getValue()->equalsInt(Size))
3165 return IK_PtrInduction;
3166 else if (C->getValue()->equalsInt(0 - Size))
3167 return IK_ReversePtrInduction;
3169 return IK_NoInduction;
3172 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3173 Value *In0 = const_cast<Value*>(V);
3174 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3178 return Inductions.count(PN);
3181 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3182 assert(TheLoop->contains(BB) && "Unknown block used");
3184 // Blocks that do not dominate the latch need predication.
3185 BasicBlock* Latch = TheLoop->getLoopLatch();
3186 return !DT->dominates(BB, Latch);
3189 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3190 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3191 // We don't predicate loads/stores at the moment.
3192 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
3195 // The instructions below can trap.
3196 switch (it->getOpcode()) {
3198 case Instruction::UDiv:
3199 case Instruction::SDiv:
3200 case Instruction::URem:
3201 case Instruction::SRem:
3209 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3210 const SCEV *PhiScev = SE->getSCEV(Ptr);
3211 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3215 return AR->isAffine();
3218 LoopVectorizationCostModel::VectorizationFactor
3219 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3221 // Width 1 means no vectorize
3222 VectorizationFactor Factor = { 1U, 0U };
3223 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3224 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3228 // Find the trip count.
3229 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3230 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3232 unsigned WidestType = getWidestType();
3233 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3234 unsigned MaxVectorSize = WidestRegister / WidestType;
3235 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3236 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3238 if (MaxVectorSize == 0) {
3239 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3243 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3244 " into one vector!");
3246 unsigned VF = MaxVectorSize;
3248 // If we optimize the program for size, avoid creating the tail loop.
3250 // If we are unable to calculate the trip count then don't try to vectorize.
3252 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3256 // Find the maximum SIMD width that can fit within the trip count.
3257 VF = TC % MaxVectorSize;
3262 // If the trip count that we found modulo the vectorization factor is not
3263 // zero then we require a tail.
3265 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3271 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3272 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3274 Factor.Width = UserVF;
3278 float Cost = expectedCost(1);
3280 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3281 for (unsigned i=2; i <= VF; i*=2) {
3282 // Notice that the vector loop needs to be executed less times, so
3283 // we need to divide the cost of the vector loops by the width of
3284 // the vector elements.
3285 float VectorCost = expectedCost(i) / (float)i;
3286 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3287 (int)VectorCost << ".\n");
3288 if (VectorCost < Cost) {
3294 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3295 Factor.Width = Width;
3296 Factor.Cost = Width * Cost;
3300 unsigned LoopVectorizationCostModel::getWidestType() {
3301 unsigned MaxWidth = 8;
3304 for (Loop::block_iterator bb = TheLoop->block_begin(),
3305 be = TheLoop->block_end(); bb != be; ++bb) {
3306 BasicBlock *BB = *bb;
3308 // For each instruction in the loop.
3309 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3310 Type *T = it->getType();
3312 // Only examine Loads, Stores and PHINodes.
3313 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3316 // Examine PHI nodes that are reduction variables.
3317 if (PHINode *PN = dyn_cast<PHINode>(it))
3318 if (!Legal->getReductionVars()->count(PN))
3321 // Examine the stored values.
3322 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3323 T = ST->getValueOperand()->getType();
3325 // Ignore loaded pointer types and stored pointer types that are not
3326 // consecutive. However, we do want to take consecutive stores/loads of
3327 // pointer vectors into account.
3328 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3331 MaxWidth = std::max(MaxWidth,
3332 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3340 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3343 unsigned LoopCost) {
3345 // -- The unroll heuristics --
3346 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3347 // There are many micro-architectural considerations that we can't predict
3348 // at this level. For example frontend pressure (on decode or fetch) due to
3349 // code size, or the number and capabilities of the execution ports.
3351 // We use the following heuristics to select the unroll factor:
3352 // 1. If the code has reductions the we unroll in order to break the cross
3353 // iteration dependency.
3354 // 2. If the loop is really small then we unroll in order to reduce the loop
3356 // 3. We don't unroll if we think that we will spill registers to memory due
3357 // to the increased register pressure.
3359 // Use the user preference, unless 'auto' is selected.
3363 // When we optimize for size we don't unroll.
3367 // Do not unroll loops with a relatively small trip count.
3368 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3369 TheLoop->getLoopLatch());
3370 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3373 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3374 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3375 " vector registers\n");
3377 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3378 // We divide by these constants so assume that we have at least one
3379 // instruction that uses at least one register.
3380 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3381 R.NumInstructions = std::max(R.NumInstructions, 1U);
3383 // We calculate the unroll factor using the following formula.
3384 // Subtract the number of loop invariants from the number of available
3385 // registers. These registers are used by all of the unrolled instances.
3386 // Next, divide the remaining registers by the number of registers that is
3387 // required by the loop, in order to estimate how many parallel instances
3388 // fit without causing spills.
3389 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3391 // Clamp the unroll factor ranges to reasonable factors.
3392 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3394 // If we did not calculate the cost for VF (because the user selected the VF)
3395 // then we calculate the cost of VF here.
3397 LoopCost = expectedCost(VF);
3399 // Clamp the calculated UF to be between the 1 and the max unroll factor
3400 // that the target allows.
3401 if (UF > MaxUnrollSize)
3406 if (Legal->getReductionVars()->size()) {
3407 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3411 // We want to unroll tiny loops in order to reduce the loop overhead.
3412 // We assume that the cost overhead is 1 and we use the cost model
3413 // to estimate the cost of the loop and unroll until the cost of the
3414 // loop overhead is about 5% of the cost of the loop.
3415 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3416 if (LoopCost < 20) {
3417 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3418 unsigned NewUF = 20/LoopCost + 1;
3419 return std::min(NewUF, UF);
3422 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3426 LoopVectorizationCostModel::RegisterUsage
3427 LoopVectorizationCostModel::calculateRegisterUsage() {
3428 // This function calculates the register usage by measuring the highest number
3429 // of values that are alive at a single location. Obviously, this is a very
3430 // rough estimation. We scan the loop in a topological order in order and
3431 // assign a number to each instruction. We use RPO to ensure that defs are
3432 // met before their users. We assume that each instruction that has in-loop
3433 // users starts an interval. We record every time that an in-loop value is
3434 // used, so we have a list of the first and last occurrences of each
3435 // instruction. Next, we transpose this data structure into a multi map that
3436 // holds the list of intervals that *end* at a specific location. This multi
3437 // map allows us to perform a linear search. We scan the instructions linearly
3438 // and record each time that a new interval starts, by placing it in a set.
3439 // If we find this value in the multi-map then we remove it from the set.
3440 // The max register usage is the maximum size of the set.
3441 // We also search for instructions that are defined outside the loop, but are
3442 // used inside the loop. We need this number separately from the max-interval
3443 // usage number because when we unroll, loop-invariant values do not take
3445 LoopBlocksDFS DFS(TheLoop);
3449 R.NumInstructions = 0;
3451 // Each 'key' in the map opens a new interval. The values
3452 // of the map are the index of the 'last seen' usage of the
3453 // instruction that is the key.
3454 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3455 // Maps instruction to its index.
3456 DenseMap<unsigned, Instruction*> IdxToInstr;
3457 // Marks the end of each interval.
3458 IntervalMap EndPoint;
3459 // Saves the list of instruction indices that are used in the loop.
3460 SmallSet<Instruction*, 8> Ends;
3461 // Saves the list of values that are used in the loop but are
3462 // defined outside the loop, such as arguments and constants.
3463 SmallPtrSet<Value*, 8> LoopInvariants;
3466 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3467 be = DFS.endRPO(); bb != be; ++bb) {
3468 R.NumInstructions += (*bb)->size();
3469 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3471 Instruction *I = it;
3472 IdxToInstr[Index++] = I;
3474 // Save the end location of each USE.
3475 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3476 Value *U = I->getOperand(i);
3477 Instruction *Instr = dyn_cast<Instruction>(U);
3479 // Ignore non-instruction values such as arguments, constants, etc.
3480 if (!Instr) continue;
3482 // If this instruction is outside the loop then record it and continue.
3483 if (!TheLoop->contains(Instr)) {
3484 LoopInvariants.insert(Instr);
3488 // Overwrite previous end points.
3489 EndPoint[Instr] = Index;
3495 // Saves the list of intervals that end with the index in 'key'.
3496 typedef SmallVector<Instruction*, 2> InstrList;
3497 DenseMap<unsigned, InstrList> TransposeEnds;
3499 // Transpose the EndPoints to a list of values that end at each index.
3500 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3502 TransposeEnds[it->second].push_back(it->first);
3504 SmallSet<Instruction*, 8> OpenIntervals;
3505 unsigned MaxUsage = 0;
3508 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3509 for (unsigned int i = 0; i < Index; ++i) {
3510 Instruction *I = IdxToInstr[i];
3511 // Ignore instructions that are never used within the loop.
3512 if (!Ends.count(I)) continue;
3514 // Remove all of the instructions that end at this location.
3515 InstrList &List = TransposeEnds[i];
3516 for (unsigned int j=0, e = List.size(); j < e; ++j)
3517 OpenIntervals.erase(List[j]);
3519 // Count the number of live interals.
3520 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3522 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3523 OpenIntervals.size() <<"\n");
3525 // Add the current instruction to the list of open intervals.
3526 OpenIntervals.insert(I);
3529 unsigned Invariant = LoopInvariants.size();
3530 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3531 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3532 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3534 R.LoopInvariantRegs = Invariant;
3535 R.MaxLocalUsers = MaxUsage;
3539 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3543 for (Loop::block_iterator bb = TheLoop->block_begin(),
3544 be = TheLoop->block_end(); bb != be; ++bb) {
3545 unsigned BlockCost = 0;
3546 BasicBlock *BB = *bb;
3548 // For each instruction in the old loop.
3549 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3550 // Skip dbg intrinsics.
3551 if (isa<DbgInfoIntrinsic>(it))
3554 unsigned C = getInstructionCost(it, VF);
3556 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3557 VF << " For instruction: "<< *it << "\n");
3560 // We assume that if-converted blocks have a 50% chance of being executed.
3561 // When the code is scalar then some of the blocks are avoided due to CF.
3562 // When the code is vectorized we execute all code paths.
3563 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3573 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3574 // If we know that this instruction will remain uniform, check the cost of
3575 // the scalar version.
3576 if (Legal->isUniformAfterVectorization(I))
3579 Type *RetTy = I->getType();
3580 Type *VectorTy = ToVectorTy(RetTy, VF);
3582 // TODO: We need to estimate the cost of intrinsic calls.
3583 switch (I->getOpcode()) {
3584 case Instruction::GetElementPtr:
3585 // We mark this instruction as zero-cost because the cost of GEPs in
3586 // vectorized code depends on whether the corresponding memory instruction
3587 // is scalarized or not. Therefore, we handle GEPs with the memory
3588 // instruction cost.
3590 case Instruction::Br: {
3591 return TTI.getCFInstrCost(I->getOpcode());
3593 case Instruction::PHI:
3594 //TODO: IF-converted IFs become selects.
3596 case Instruction::Add:
3597 case Instruction::FAdd:
3598 case Instruction::Sub:
3599 case Instruction::FSub:
3600 case Instruction::Mul:
3601 case Instruction::FMul:
3602 case Instruction::UDiv:
3603 case Instruction::SDiv:
3604 case Instruction::FDiv:
3605 case Instruction::URem:
3606 case Instruction::SRem:
3607 case Instruction::FRem:
3608 case Instruction::Shl:
3609 case Instruction::LShr:
3610 case Instruction::AShr:
3611 case Instruction::And:
3612 case Instruction::Or:
3613 case Instruction::Xor: {
3614 // Certain instructions can be cheaper to vectorize if they have a constant
3615 // second vector operand. One example of this are shifts on x86.
3616 TargetTransformInfo::OperandValueKind Op1VK =
3617 TargetTransformInfo::OK_AnyValue;
3618 TargetTransformInfo::OperandValueKind Op2VK =
3619 TargetTransformInfo::OK_AnyValue;
3621 if (isa<ConstantInt>(I->getOperand(1)))
3622 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3624 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3626 case Instruction::Select: {
3627 SelectInst *SI = cast<SelectInst>(I);
3628 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3629 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3630 Type *CondTy = SI->getCondition()->getType();
3632 CondTy = VectorType::get(CondTy, VF);
3634 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3636 case Instruction::ICmp:
3637 case Instruction::FCmp: {
3638 Type *ValTy = I->getOperand(0)->getType();
3639 VectorTy = ToVectorTy(ValTy, VF);
3640 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3642 case Instruction::Store:
3643 case Instruction::Load: {
3644 StoreInst *SI = dyn_cast<StoreInst>(I);
3645 LoadInst *LI = dyn_cast<LoadInst>(I);
3646 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3648 VectorTy = ToVectorTy(ValTy, VF);
3650 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3651 unsigned AS = SI ? SI->getPointerAddressSpace() :
3652 LI->getPointerAddressSpace();
3653 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3654 // We add the cost of address computation here instead of with the gep
3655 // instruction because only here we know whether the operation is
3658 return TTI.getAddressComputationCost(VectorTy) +
3659 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3661 // Scalarized loads/stores.
3662 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3663 bool Reverse = ConsecutiveStride < 0;
3664 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3665 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3666 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3668 // The cost of extracting from the value vector and pointer vector.
3669 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3670 for (unsigned i = 0; i < VF; ++i) {
3671 // The cost of extracting the pointer operand.
3672 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3673 // In case of STORE, the cost of ExtractElement from the vector.
3674 // In case of LOAD, the cost of InsertElement into the returned
3676 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3677 Instruction::InsertElement,
3681 // The cost of the scalar loads/stores.
3682 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3683 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3688 // Wide load/stores.
3689 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3690 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3693 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3697 case Instruction::ZExt:
3698 case Instruction::SExt:
3699 case Instruction::FPToUI:
3700 case Instruction::FPToSI:
3701 case Instruction::FPExt:
3702 case Instruction::PtrToInt:
3703 case Instruction::IntToPtr:
3704 case Instruction::SIToFP:
3705 case Instruction::UIToFP:
3706 case Instruction::Trunc:
3707 case Instruction::FPTrunc:
3708 case Instruction::BitCast: {
3709 // We optimize the truncation of induction variable.
3710 // The cost of these is the same as the scalar operation.
3711 if (I->getOpcode() == Instruction::Trunc &&
3712 Legal->isInductionVariable(I->getOperand(0)))
3713 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3714 I->getOperand(0)->getType());
3716 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3717 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3719 case Instruction::Call: {
3720 CallInst *CI = cast<CallInst>(I);
3721 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3722 assert(ID && "Not an intrinsic call!");
3723 Type *RetTy = ToVectorTy(CI->getType(), VF);
3724 SmallVector<Type*, 4> Tys;
3725 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3726 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3727 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3730 // We are scalarizing the instruction. Return the cost of the scalar
3731 // instruction, plus the cost of insert and extract into vector
3732 // elements, times the vector width.
3735 if (!RetTy->isVoidTy() && VF != 1) {
3736 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3738 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3741 // The cost of inserting the results plus extracting each one of the
3743 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3746 // The cost of executing VF copies of the scalar instruction. This opcode
3747 // is unknown. Assume that it is the same as 'mul'.
3748 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3754 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3755 if (Scalar->isVoidTy() || VF == 1)
3757 return VectorType::get(Scalar, VF);
3760 char LoopVectorize::ID = 0;
3761 static const char lv_name[] = "Loop Vectorization";
3762 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3763 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3764 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3765 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3766 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3767 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3770 Pass *createLoopVectorizePass() {
3771 return new LoopVectorize();
3775 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3776 // Check for a store.
3777 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3778 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3780 // Check for a load.
3781 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3782 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;