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 /// Holds the extended (to the widest induction type) start index.
317 /// Maps scalars to widened vectors.
321 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
322 /// to what vectorization factor.
323 /// This class does not look at the profitability of vectorization, only the
324 /// legality. This class has two main kinds of checks:
325 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
326 /// will change the order of memory accesses in a way that will change the
327 /// correctness of the program.
328 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
329 /// checks for a number of different conditions, such as the availability of a
330 /// single induction variable, that all types are supported and vectorize-able,
331 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
332 /// This class is also used by InnerLoopVectorizer for identifying
333 /// induction variable and the different reduction variables.
334 class LoopVectorizationLegality {
336 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
337 DominatorTree *DT, TargetTransformInfo* TTI,
338 AliasAnalysis *AA, TargetLibraryInfo *TLI)
339 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
340 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false) {}
342 /// This enum represents the kinds of reductions that we support.
344 RK_NoReduction, ///< Not a reduction.
345 RK_IntegerAdd, ///< Sum of integers.
346 RK_IntegerMult, ///< Product of integers.
347 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
348 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
349 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
350 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
351 RK_FloatAdd, ///< Sum of floats.
352 RK_FloatMult, ///< Product of floats.
353 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
356 /// This enum represents the kinds of inductions that we support.
358 IK_NoInduction, ///< Not an induction variable.
359 IK_IntInduction, ///< Integer induction variable. Step = 1.
360 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
361 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
362 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
365 // This enum represents the kind of minmax reduction.
366 enum MinMaxReductionKind {
376 /// This POD struct holds information about reduction variables.
377 struct ReductionDescriptor {
378 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
379 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
381 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
382 MinMaxReductionKind MK)
383 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
385 // The starting value of the reduction.
386 // It does not have to be zero!
388 // The instruction who's value is used outside the loop.
389 Instruction *LoopExitInstr;
390 // The kind of the reduction.
392 // If this a min/max reduction the kind of reduction.
393 MinMaxReductionKind MinMaxKind;
396 /// This POD struct holds information about a potential reduction operation.
397 struct ReductionInstDesc {
398 ReductionInstDesc(bool IsRedux, Instruction *I) :
399 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
401 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
402 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
404 // Is this instruction a reduction candidate.
406 // The last instruction in a min/max pattern (select of the select(icmp())
407 // pattern), or the current reduction instruction otherwise.
408 Instruction *PatternLastInst;
409 // If this is a min/max pattern the comparison predicate.
410 MinMaxReductionKind MinMaxKind;
413 // This POD struct holds information about the memory runtime legality
414 // check that a group of pointers do not overlap.
415 struct RuntimePointerCheck {
416 RuntimePointerCheck() : Need(false) {}
418 /// Reset the state of the pointer runtime information.
426 /// Insert a pointer and calculate the start and end SCEVs.
427 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr);
429 /// This flag indicates if we need to add the runtime check.
431 /// Holds the pointers that we need to check.
432 SmallVector<Value*, 2> Pointers;
433 /// Holds the pointer value at the beginning of the loop.
434 SmallVector<const SCEV*, 2> Starts;
435 /// Holds the pointer value at the end of the loop.
436 SmallVector<const SCEV*, 2> Ends;
437 /// Holds the information if this pointer is used for writing to memory.
438 SmallVector<bool, 2> IsWritePtr;
441 /// A POD for saving information about induction variables.
442 struct InductionInfo {
443 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
444 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
451 /// ReductionList contains the reduction descriptors for all
452 /// of the reductions that were found in the loop.
453 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
455 /// InductionList saves induction variables and maps them to the
456 /// induction descriptor.
457 typedef MapVector<PHINode*, InductionInfo> InductionList;
459 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
460 /// respective Store/Load instruction(s) to calculate aliasing.
461 typedef MapVector<Value*, Instruction* > AliasMap;
462 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
464 /// Returns true if it is legal to vectorize this loop.
465 /// This does not mean that it is profitable to vectorize this
466 /// loop, only that it is legal to do so.
469 /// Returns the Induction variable.
470 PHINode *getInduction() { return Induction; }
472 /// Returns the reduction variables found in the loop.
473 ReductionList *getReductionVars() { return &Reductions; }
475 /// Returns the induction variables found in the loop.
476 InductionList *getInductionVars() { return &Inductions; }
478 /// Returns the widest induction type.
479 Type *getWidestInductionType() { return WidestIndTy; }
481 /// Returns True if V is an induction variable in this loop.
482 bool isInductionVariable(const Value *V);
484 /// Return true if the block BB needs to be predicated in order for the loop
485 /// to be vectorized.
486 bool blockNeedsPredication(BasicBlock *BB);
488 /// Check if this pointer is consecutive when vectorizing. This happens
489 /// when the last index of the GEP is the induction variable, or that the
490 /// pointer itself is an induction variable.
491 /// This check allows us to vectorize A[idx] into a wide load/store.
493 /// 0 - Stride is unknown or non consecutive.
494 /// 1 - Address is consecutive.
495 /// -1 - Address is consecutive, and decreasing.
496 int isConsecutivePtr(Value *Ptr);
498 /// Returns true if the value V is uniform within the loop.
499 bool isUniform(Value *V);
501 /// Returns true if this instruction will remain scalar after vectorization.
502 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
504 /// Returns the information that we collected about runtime memory check.
505 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
507 /// This function returns the identity element (or neutral element) for
509 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
511 /// Check if a single basic block loop is vectorizable.
512 /// At this point we know that this is a loop with a constant trip count
513 /// and we only need to check individual instructions.
514 bool canVectorizeInstrs();
516 /// When we vectorize loops we may change the order in which
517 /// we read and write from memory. This method checks if it is
518 /// legal to vectorize the code, considering only memory constrains.
519 /// Returns true if the loop is vectorizable
520 bool canVectorizeMemory();
522 /// Return true if we can vectorize this loop using the IF-conversion
524 bool canVectorizeWithIfConvert();
526 /// Collect the variables that need to stay uniform after vectorization.
527 void collectLoopUniforms();
529 /// Return true if all of the instructions in the block can be speculatively
531 bool blockCanBePredicated(BasicBlock *BB);
533 /// Returns True, if 'Phi' is the kind of reduction variable for type
534 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
535 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
536 /// Returns a struct describing if the instruction 'I' can be a reduction
537 /// variable of type 'Kind'. If the reduction is a min/max pattern of
538 /// select(icmp()) this function advances the instruction pointer 'I' from the
539 /// compare instruction to the select instruction and stores this pointer in
540 /// 'PatternLastInst' member of the returned struct.
541 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
542 ReductionInstDesc &Desc);
543 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
544 /// pattern corresponding to a min(X, Y) or max(X, Y).
545 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
546 ReductionInstDesc &Prev);
547 /// Returns the induction kind of Phi. This function may return NoInduction
548 /// if the PHI is not an induction variable.
549 InductionKind isInductionVariable(PHINode *Phi);
550 /// Return true if can compute the address bounds of Ptr within the loop.
551 bool hasComputableBounds(Value *Ptr);
552 /// Return true if there is the chance of write reorder.
553 bool hasPossibleGlobalWriteReorder(Value *Object,
555 AliasMultiMap &WriteObjects,
556 unsigned MaxByteWidth);
557 /// Return the AA location for a load or a store.
558 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
561 /// The loop that we evaluate.
565 /// DataLayout analysis.
570 TargetTransformInfo *TTI;
573 /// Target Library Info.
574 TargetLibraryInfo *TLI;
576 // --- vectorization state --- //
578 /// Holds the integer induction variable. This is the counter of the
581 /// Holds the reduction variables.
582 ReductionList Reductions;
583 /// Holds all of the induction variables that we found in the loop.
584 /// Notice that inductions don't need to start at zero and that induction
585 /// variables can be pointers.
586 InductionList Inductions;
587 /// Holds the widest induction type encountered.
590 /// Allowed outside users. This holds the reduction
591 /// vars which can be accessed from outside the loop.
592 SmallPtrSet<Value*, 4> AllowedExit;
593 /// This set holds the variables which are known to be uniform after
595 SmallPtrSet<Instruction*, 4> Uniforms;
596 /// We need to check that all of the pointers in this list are disjoint
598 RuntimePointerCheck PtrRtCheck;
599 /// Can we assume the absence of NaNs.
600 bool HasFunNoNaNAttr;
603 /// LoopVectorizationCostModel - estimates the expected speedups due to
605 /// In many cases vectorization is not profitable. This can happen because of
606 /// a number of reasons. In this class we mainly attempt to predict the
607 /// expected speedup/slowdowns due to the supported instruction set. We use the
608 /// TargetTransformInfo to query the different backends for the cost of
609 /// different operations.
610 class LoopVectorizationCostModel {
612 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
613 LoopVectorizationLegality *Legal,
614 const TargetTransformInfo &TTI,
615 DataLayout *DL, const TargetLibraryInfo *TLI)
616 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
618 /// Information about vectorization costs
619 struct VectorizationFactor {
620 unsigned Width; // Vector width with best cost
621 unsigned Cost; // Cost of the loop with that width
623 /// \return The most profitable vectorization factor and the cost of that VF.
624 /// This method checks every power of two up to VF. If UserVF is not ZERO
625 /// then this vectorization factor will be selected if vectorization is
627 VectorizationFactor selectVectorizationFactor(bool OptForSize,
630 /// \return The size (in bits) of the widest type in the code that
631 /// needs to be vectorized. We ignore values that remain scalar such as
632 /// 64 bit loop indices.
633 unsigned getWidestType();
635 /// \return The most profitable unroll factor.
636 /// If UserUF is non-zero then this method finds the best unroll-factor
637 /// based on register pressure and other parameters.
638 /// VF and LoopCost are the selected vectorization factor and the cost of the
640 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
643 /// \brief A struct that represents some properties of the register usage
645 struct RegisterUsage {
646 /// Holds the number of loop invariant values that are used in the loop.
647 unsigned LoopInvariantRegs;
648 /// Holds the maximum number of concurrent live intervals in the loop.
649 unsigned MaxLocalUsers;
650 /// Holds the number of instructions in the loop.
651 unsigned NumInstructions;
654 /// \return information about the register usage of the loop.
655 RegisterUsage calculateRegisterUsage();
658 /// Returns the expected execution cost. The unit of the cost does
659 /// not matter because we use the 'cost' units to compare different
660 /// vector widths. The cost that is returned is *not* normalized by
661 /// the factor width.
662 unsigned expectedCost(unsigned VF);
664 /// Returns the execution time cost of an instruction for a given vector
665 /// width. Vector width of one means scalar.
666 unsigned getInstructionCost(Instruction *I, unsigned VF);
668 /// A helper function for converting Scalar types to vector types.
669 /// If the incoming type is void, we return void. If the VF is 1, we return
671 static Type* ToVectorTy(Type *Scalar, unsigned VF);
673 /// Returns whether the instruction is a load or store and will be a emitted
674 /// as a vector operation.
675 bool isConsecutiveLoadOrStore(Instruction *I);
677 /// The loop that we evaluate.
681 /// Loop Info analysis.
683 /// Vectorization legality.
684 LoopVectorizationLegality *Legal;
685 /// Vector target information.
686 const TargetTransformInfo &TTI;
687 /// Target data layout information.
689 /// Target Library Info.
690 const TargetLibraryInfo *TLI;
693 /// The LoopVectorize Pass.
694 struct LoopVectorize : public LoopPass {
695 /// Pass identification, replacement for typeid
698 explicit LoopVectorize() : LoopPass(ID) {
699 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
705 TargetTransformInfo *TTI;
708 TargetLibraryInfo *TLI;
710 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
711 // We only vectorize innermost loops.
715 SE = &getAnalysis<ScalarEvolution>();
716 DL = getAnalysisIfAvailable<DataLayout>();
717 LI = &getAnalysis<LoopInfo>();
718 TTI = &getAnalysis<TargetTransformInfo>();
719 DT = &getAnalysis<DominatorTree>();
720 AA = getAnalysisIfAvailable<AliasAnalysis>();
721 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
724 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
728 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
729 L->getHeader()->getParent()->getName() << "\"\n");
731 // Check if it is legal to vectorize the loop.
732 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
733 if (!LVL.canVectorize()) {
734 DEBUG(dbgs() << "LV: Not vectorizing.\n");
738 // Use the cost model.
739 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
741 // Check the function attributes to find out if this function should be
742 // optimized for size.
743 Function *F = L->getHeader()->getParent();
744 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
745 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
746 unsigned FnIndex = AttributeSet::FunctionIndex;
747 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
748 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
751 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
752 "attribute is used.\n");
756 // Select the optimal vectorization factor.
757 LoopVectorizationCostModel::VectorizationFactor VF;
758 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
759 // Select the unroll factor.
760 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
764 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
768 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
769 F->getParent()->getModuleIdentifier()<<"\n");
770 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
772 // If we decided that it is *legal* to vectorize the loop then do it.
773 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
776 DEBUG(verifyFunction(*L->getHeader()->getParent()));
780 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
781 LoopPass::getAnalysisUsage(AU);
782 AU.addRequiredID(LoopSimplifyID);
783 AU.addRequiredID(LCSSAID);
784 AU.addRequired<DominatorTree>();
785 AU.addRequired<LoopInfo>();
786 AU.addRequired<ScalarEvolution>();
787 AU.addRequired<TargetTransformInfo>();
788 AU.addPreserved<LoopInfo>();
789 AU.addPreserved<DominatorTree>();
794 } // end anonymous namespace
796 //===----------------------------------------------------------------------===//
797 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
798 // LoopVectorizationCostModel.
799 //===----------------------------------------------------------------------===//
802 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
803 Loop *Lp, Value *Ptr,
805 const SCEV *Sc = SE->getSCEV(Ptr);
806 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
807 assert(AR && "Invalid addrec expression");
808 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
809 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
810 Pointers.push_back(Ptr);
811 Starts.push_back(AR->getStart());
812 Ends.push_back(ScEnd);
813 IsWritePtr.push_back(WritePtr);
816 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
817 // Save the current insertion location.
818 Instruction *Loc = Builder.GetInsertPoint();
820 // We need to place the broadcast of invariant variables outside the loop.
821 Instruction *Instr = dyn_cast<Instruction>(V);
822 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
823 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
825 // Place the code for broadcasting invariant variables in the new preheader.
827 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
829 // Broadcast the scalar into all locations in the vector.
830 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
832 // Restore the builder insertion point.
834 Builder.SetInsertPoint(Loc);
839 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
841 assert(Val->getType()->isVectorTy() && "Must be a vector");
842 assert(Val->getType()->getScalarType()->isIntegerTy() &&
843 "Elem must be an integer");
845 Type *ITy = Val->getType()->getScalarType();
846 VectorType *Ty = cast<VectorType>(Val->getType());
847 int VLen = Ty->getNumElements();
848 SmallVector<Constant*, 8> Indices;
850 // Create a vector of consecutive numbers from zero to VF.
851 for (int i = 0; i < VLen; ++i) {
852 int64_t Idx = Negate ? (-i) : i;
853 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
856 // Add the consecutive indices to the vector value.
857 Constant *Cv = ConstantVector::get(Indices);
858 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
859 return Builder.CreateAdd(Val, Cv, "induction");
862 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
863 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
864 // Make sure that the pointer does not point to structs.
865 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
868 // If this value is a pointer induction variable we know it is consecutive.
869 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
870 if (Phi && Inductions.count(Phi)) {
871 InductionInfo II = Inductions[Phi];
872 if (IK_PtrInduction == II.IK)
874 else if (IK_ReversePtrInduction == II.IK)
878 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
882 unsigned NumOperands = Gep->getNumOperands();
883 Value *LastIndex = Gep->getOperand(NumOperands - 1);
885 Value *GpPtr = Gep->getPointerOperand();
886 // If this GEP value is a consecutive pointer induction variable and all of
887 // the indices are constant then we know it is consecutive. We can
888 Phi = dyn_cast<PHINode>(GpPtr);
889 if (Phi && Inductions.count(Phi)) {
891 // Make sure that the pointer does not point to structs.
892 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
893 if (GepPtrType->getElementType()->isAggregateType())
896 // Make sure that all of the index operands are loop invariant.
897 for (unsigned i = 1; i < NumOperands; ++i)
898 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
901 InductionInfo II = Inductions[Phi];
902 if (IK_PtrInduction == II.IK)
904 else if (IK_ReversePtrInduction == II.IK)
908 // Check that all of the gep indices are uniform except for the last.
909 for (unsigned i = 0; i < NumOperands - 1; ++i)
910 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
913 // We can emit wide load/stores only if the last index is the induction
915 const SCEV *Last = SE->getSCEV(LastIndex);
916 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
917 const SCEV *Step = AR->getStepRecurrence(*SE);
919 // The memory is consecutive because the last index is consecutive
920 // and all other indices are loop invariant.
923 if (Step->isAllOnesValue())
930 bool LoopVectorizationLegality::isUniform(Value *V) {
931 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
934 InnerLoopVectorizer::VectorParts&
935 InnerLoopVectorizer::getVectorValue(Value *V) {
936 assert(V != Induction && "The new induction variable should not be used.");
937 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
939 // If we have this scalar in the map, return it.
941 return WidenMap.get(V);
943 // If this scalar is unknown, assume that it is a constant or that it is
944 // loop invariant. Broadcast V and save the value for future uses.
945 Value *B = getBroadcastInstrs(V);
946 return WidenMap.splat(V, B);
949 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
950 assert(Vec->getType()->isVectorTy() && "Invalid type");
951 SmallVector<Constant*, 8> ShuffleMask;
952 for (unsigned i = 0; i < VF; ++i)
953 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
955 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
956 ConstantVector::get(ShuffleMask),
961 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
962 LoopVectorizationLegality *Legal) {
963 // Attempt to issue a wide load.
964 LoadInst *LI = dyn_cast<LoadInst>(Instr);
965 StoreInst *SI = dyn_cast<StoreInst>(Instr);
967 assert((LI || SI) && "Invalid Load/Store instruction");
969 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
970 Type *DataTy = VectorType::get(ScalarDataTy, VF);
971 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
972 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
974 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
975 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
977 if (ScalarAllocatedSize != VectorElementSize)
978 return scalarizeInstruction(Instr);
980 // If the pointer is loop invariant or if it is non consecutive,
981 // scalarize the load.
982 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
983 bool Reverse = ConsecutiveStride < 0;
984 bool UniformLoad = LI && Legal->isUniform(Ptr);
985 if (!ConsecutiveStride || UniformLoad)
986 return scalarizeInstruction(Instr);
988 Constant *Zero = Builder.getInt32(0);
989 VectorParts &Entry = WidenMap.get(Instr);
991 // Handle consecutive loads/stores.
992 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
993 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
994 Value *PtrOperand = Gep->getPointerOperand();
995 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
996 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
998 // Create the new GEP with the new induction variable.
999 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1000 Gep2->setOperand(0, FirstBasePtr);
1001 Gep2->setName("gep.indvar.base");
1002 Ptr = Builder.Insert(Gep2);
1004 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1005 OrigLoop) && "Base ptr must be invariant");
1007 // The last index does not have to be the induction. It can be
1008 // consecutive and be a function of the index. For example A[I+1];
1009 unsigned NumOperands = Gep->getNumOperands();
1011 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1012 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1013 Value *LastIndex = GEPParts[0];
1014 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1016 // Create the new GEP with the new induction variable.
1017 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1018 Gep2->setOperand(NumOperands - 1, LastIndex);
1019 Gep2->setName("gep.indvar.idx");
1020 Ptr = Builder.Insert(Gep2);
1022 // Use the induction element ptr.
1023 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1024 VectorParts &PtrVal = getVectorValue(Ptr);
1025 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1030 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1031 "We do not allow storing to uniform addresses");
1033 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1034 for (unsigned Part = 0; Part < UF; ++Part) {
1035 // Calculate the pointer for the specific unroll-part.
1036 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1039 // If we store to reverse consecutive memory locations then we need
1040 // to reverse the order of elements in the stored value.
1041 StoredVal[Part] = reverseVector(StoredVal[Part]);
1042 // If the address is consecutive but reversed, then the
1043 // wide store needs to start at the last vector element.
1044 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1045 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1048 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1049 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1053 for (unsigned Part = 0; Part < UF; ++Part) {
1054 // Calculate the pointer for the specific unroll-part.
1055 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1058 // If the address is consecutive but reversed, then the
1059 // wide store needs to start at the last vector element.
1060 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1061 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1064 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1065 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1066 cast<LoadInst>(LI)->setAlignment(Alignment);
1067 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1071 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1072 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1073 // Holds vector parameters or scalars, in case of uniform vals.
1074 SmallVector<VectorParts, 4> Params;
1076 // Find all of the vectorized parameters.
1077 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1078 Value *SrcOp = Instr->getOperand(op);
1080 // If we are accessing the old induction variable, use the new one.
1081 if (SrcOp == OldInduction) {
1082 Params.push_back(getVectorValue(SrcOp));
1086 // Try using previously calculated values.
1087 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1089 // If the src is an instruction that appeared earlier in the basic block
1090 // then it should already be vectorized.
1091 if (SrcInst && OrigLoop->contains(SrcInst)) {
1092 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1093 // The parameter is a vector value from earlier.
1094 Params.push_back(WidenMap.get(SrcInst));
1096 // The parameter is a scalar from outside the loop. Maybe even a constant.
1097 VectorParts Scalars;
1098 Scalars.append(UF, SrcOp);
1099 Params.push_back(Scalars);
1103 assert(Params.size() == Instr->getNumOperands() &&
1104 "Invalid number of operands");
1106 // Does this instruction return a value ?
1107 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1109 Value *UndefVec = IsVoidRetTy ? 0 :
1110 UndefValue::get(VectorType::get(Instr->getType(), VF));
1111 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1112 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1114 // For each vector unroll 'part':
1115 for (unsigned Part = 0; Part < UF; ++Part) {
1116 // For each scalar that we create:
1117 for (unsigned Width = 0; Width < VF; ++Width) {
1118 Instruction *Cloned = Instr->clone();
1120 Cloned->setName(Instr->getName() + ".cloned");
1121 // Replace the operands of the cloned instrucions with extracted scalars.
1122 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1123 Value *Op = Params[op][Part];
1124 // Param is a vector. Need to extract the right lane.
1125 if (Op->getType()->isVectorTy())
1126 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1127 Cloned->setOperand(op, Op);
1130 // Place the cloned scalar in the new loop.
1131 Builder.Insert(Cloned);
1133 // If the original scalar returns a value we need to place it in a vector
1134 // so that future users will be able to use it.
1136 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1137 Builder.getInt32(Width));
1143 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1145 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1146 Legal->getRuntimePointerCheck();
1148 if (!PtrRtCheck->Need)
1151 Instruction *MemoryRuntimeCheck = 0;
1152 unsigned NumPointers = PtrRtCheck->Pointers.size();
1153 SmallVector<Value* , 2> Starts;
1154 SmallVector<Value* , 2> Ends;
1156 SCEVExpander Exp(*SE, "induction");
1158 // Use this type for pointer arithmetic.
1159 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1161 for (unsigned i = 0; i < NumPointers; ++i) {
1162 Value *Ptr = PtrRtCheck->Pointers[i];
1163 const SCEV *Sc = SE->getSCEV(Ptr);
1165 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1166 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1168 Starts.push_back(Ptr);
1169 Ends.push_back(Ptr);
1171 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1173 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1174 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1175 Starts.push_back(Start);
1176 Ends.push_back(End);
1180 IRBuilder<> ChkBuilder(Loc);
1182 for (unsigned i = 0; i < NumPointers; ++i) {
1183 for (unsigned j = i+1; j < NumPointers; ++j) {
1184 // No need to check if two readonly pointers intersect.
1185 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1188 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1189 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1190 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1191 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1193 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1194 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1195 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1196 if (MemoryRuntimeCheck)
1197 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1200 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1204 return MemoryRuntimeCheck;
1208 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1210 In this function we generate a new loop. The new loop will contain
1211 the vectorized instructions while the old loop will continue to run the
1214 [ ] <-- vector loop bypass (may consist of multiple blocks).
1217 | [ ] <-- vector pre header.
1221 | [ ]_| <-- vector loop.
1224 >[ ] <--- middle-block.
1227 | [ ] <--- new preheader.
1231 | [ ]_| <-- old scalar loop to handle remainder.
1234 >[ ] <-- exit block.
1238 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1239 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1240 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1241 assert(ExitBlock && "Must have an exit block");
1243 // Mark the old scalar loop with metadata that tells us not to vectorize this
1244 // loop again if we run into it.
1245 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None);
1246 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD);
1248 // Some loops have a single integer induction variable, while other loops
1249 // don't. One example is c++ iterators that often have multiple pointer
1250 // induction variables. In the code below we also support a case where we
1251 // don't have a single induction variable.
1252 OldInduction = Legal->getInduction();
1253 Type *IdxTy = Legal->getWidestInductionType();
1255 // Find the loop boundaries.
1256 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1257 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1259 // Get the total trip count from the count by adding 1.
1260 ExitCount = SE->getAddExpr(ExitCount,
1261 SE->getConstant(ExitCount->getType(), 1));
1263 // Expand the trip count and place the new instructions in the preheader.
1264 // Notice that the pre-header does not change, only the loop body.
1265 SCEVExpander Exp(*SE, "induction");
1267 // Count holds the overall loop count (N).
1268 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1269 BypassBlock->getTerminator());
1271 // The loop index does not have to start at Zero. Find the original start
1272 // value from the induction PHI node. If we don't have an induction variable
1273 // then we know that it starts at zero.
1274 Builder.SetInsertPoint(BypassBlock->getTerminator());
1275 Value *StartIdx = ExtendedIdx = OldInduction ?
1276 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1278 ConstantInt::get(IdxTy, 0);
1280 assert(BypassBlock && "Invalid loop structure");
1281 LoopBypassBlocks.push_back(BypassBlock);
1283 // Split the single block loop into the two loop structure described above.
1284 BasicBlock *VectorPH =
1285 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1286 BasicBlock *VecBody =
1287 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1288 BasicBlock *MiddleBlock =
1289 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1290 BasicBlock *ScalarPH =
1291 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1293 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1295 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1297 // Generate the induction variable.
1298 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1299 // The loop step is equal to the vectorization factor (num of SIMD elements)
1300 // times the unroll factor (num of SIMD instructions).
1301 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1303 // This is the IR builder that we use to add all of the logic for bypassing
1304 // the new vector loop.
1305 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1307 // We may need to extend the index in case there is a type mismatch.
1308 // We know that the count starts at zero and does not overflow.
1309 if (Count->getType() != IdxTy) {
1310 // The exit count can be of pointer type. Convert it to the correct
1312 if (ExitCount->getType()->isPointerTy())
1313 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1315 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1318 // Add the start index to the loop count to get the new end index.
1319 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1321 // Now we need to generate the expression for N - (N % VF), which is
1322 // the part that the vectorized body will execute.
1323 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1324 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1325 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1326 "end.idx.rnd.down");
1328 // Now, compare the new count to zero. If it is zero skip the vector loop and
1329 // jump to the scalar loop.
1330 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1333 BasicBlock *LastBypassBlock = BypassBlock;
1335 // Generate the code that checks in runtime if arrays overlap. We put the
1336 // checks into a separate block to make the more common case of few elements
1338 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1339 BypassBlock->getTerminator());
1340 if (MemRuntimeCheck) {
1341 // Create a new block containing the memory check.
1342 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1344 LoopBypassBlocks.push_back(CheckBlock);
1346 // Replace the branch into the memory check block with a conditional branch
1347 // for the "few elements case".
1348 Instruction *OldTerm = BypassBlock->getTerminator();
1349 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1350 OldTerm->eraseFromParent();
1352 Cmp = MemRuntimeCheck;
1353 LastBypassBlock = CheckBlock;
1356 LastBypassBlock->getTerminator()->eraseFromParent();
1357 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1360 // We are going to resume the execution of the scalar loop.
1361 // Go over all of the induction variables that we found and fix the
1362 // PHIs that are left in the scalar version of the loop.
1363 // The starting values of PHI nodes depend on the counter of the last
1364 // iteration in the vectorized loop.
1365 // If we come from a bypass edge then we need to start from the original
1368 // This variable saves the new starting index for the scalar loop.
1369 PHINode *ResumeIndex = 0;
1370 LoopVectorizationLegality::InductionList::iterator I, E;
1371 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1372 // Set builder to point to last bypass block.
1373 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1374 for (I = List->begin(), E = List->end(); I != E; ++I) {
1375 PHINode *OrigPhi = I->first;
1376 LoopVectorizationLegality::InductionInfo II = I->second;
1378 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1379 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1380 MiddleBlock->getTerminator());
1381 // We might have extended the type of the induction variable but we need a
1382 // truncated version for the scalar loop.
1383 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1384 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1385 MiddleBlock->getTerminator()) : 0;
1387 Value *EndValue = 0;
1389 case LoopVectorizationLegality::IK_NoInduction:
1390 llvm_unreachable("Unknown induction");
1391 case LoopVectorizationLegality::IK_IntInduction: {
1392 // Handle the integer induction counter:
1393 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1394 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1395 if (OrigPhi == OldInduction) {
1396 // Create a truncated version of the resume value for the scalar loop,
1397 // we might have promoted the type to a larger width.
1399 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1400 // The new PHI merges the original incoming value, in case of a bypass,
1401 // or the value at the end of the vectorized loop.
1402 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1403 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1404 TruncResumeVal->addIncoming(EndValue, VecBody);
1406 // We know what the end value is.
1407 EndValue = IdxEndRoundDown;
1408 // We also know which PHI node holds it.
1409 ResumeIndex = ResumeVal;
1412 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1413 // Convert the CountRoundDown variable to the PHI size.
1414 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1415 II.StartValue->getType(),
1417 // Handle reverse integer induction counter.
1418 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1421 case LoopVectorizationLegality::IK_PtrInduction: {
1422 // For pointer induction variables, calculate the offset using
1424 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1428 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1429 // The value at the end of the loop for the reverse pointer is calculated
1430 // by creating a GEP with a negative index starting from the start value.
1431 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1432 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1434 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1440 // The new PHI merges the original incoming value, in case of a bypass,
1441 // or the value at the end of the vectorized loop.
1442 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1443 if (OrigPhi == OldInduction)
1444 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1446 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1448 ResumeVal->addIncoming(EndValue, VecBody);
1450 // Fix the scalar body counter (PHI node).
1451 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1452 // The old inductions phi node in the scalar body needs the truncated value.
1453 if (OrigPhi == OldInduction)
1454 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1456 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1459 // If we are generating a new induction variable then we also need to
1460 // generate the code that calculates the exit value. This value is not
1461 // simply the end of the counter because we may skip the vectorized body
1462 // in case of a runtime check.
1464 assert(!ResumeIndex && "Unexpected resume value found");
1465 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1466 MiddleBlock->getTerminator());
1467 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1468 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1469 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1472 // Make sure that we found the index where scalar loop needs to continue.
1473 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1474 "Invalid resume Index");
1476 // Add a check in the middle block to see if we have completed
1477 // all of the iterations in the first vector loop.
1478 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1479 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1480 ResumeIndex, "cmp.n",
1481 MiddleBlock->getTerminator());
1483 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1484 // Remove the old terminator.
1485 MiddleBlock->getTerminator()->eraseFromParent();
1487 // Create i+1 and fill the PHINode.
1488 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1489 Induction->addIncoming(StartIdx, VectorPH);
1490 Induction->addIncoming(NextIdx, VecBody);
1491 // Create the compare.
1492 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1493 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1495 // Now we have two terminators. Remove the old one from the block.
1496 VecBody->getTerminator()->eraseFromParent();
1498 // Get ready to start creating new instructions into the vectorized body.
1499 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1501 // Create and register the new vector loop.
1502 Loop* Lp = new Loop();
1503 Loop *ParentLoop = OrigLoop->getParentLoop();
1505 // Insert the new loop into the loop nest and register the new basic blocks.
1507 ParentLoop->addChildLoop(Lp);
1508 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1509 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1510 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1511 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1512 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1514 LI->addTopLevelLoop(Lp);
1517 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1520 LoopVectorPreHeader = VectorPH;
1521 LoopScalarPreHeader = ScalarPH;
1522 LoopMiddleBlock = MiddleBlock;
1523 LoopExitBlock = ExitBlock;
1524 LoopVectorBody = VecBody;
1525 LoopScalarBody = OldBasicBlock;
1528 /// This function returns the identity element (or neutral element) for
1529 /// the operation K.
1531 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1536 // Adding, Xoring, Oring zero to a number does not change it.
1537 return ConstantInt::get(Tp, 0);
1538 case RK_IntegerMult:
1539 // Multiplying a number by 1 does not change it.
1540 return ConstantInt::get(Tp, 1);
1542 // AND-ing a number with an all-1 value does not change it.
1543 return ConstantInt::get(Tp, -1, true);
1545 // Multiplying a number by 1 does not change it.
1546 return ConstantFP::get(Tp, 1.0L);
1548 // Adding zero to a number does not change it.
1549 return ConstantFP::get(Tp, 0.0L);
1551 llvm_unreachable("Unknown reduction kind");
1555 static Intrinsic::ID
1556 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1557 // If we have an intrinsic call, check if it is trivially vectorizable.
1558 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1559 switch (II->getIntrinsicID()) {
1560 case Intrinsic::sqrt:
1561 case Intrinsic::sin:
1562 case Intrinsic::cos:
1563 case Intrinsic::exp:
1564 case Intrinsic::exp2:
1565 case Intrinsic::log:
1566 case Intrinsic::log10:
1567 case Intrinsic::log2:
1568 case Intrinsic::fabs:
1569 case Intrinsic::floor:
1570 case Intrinsic::ceil:
1571 case Intrinsic::trunc:
1572 case Intrinsic::rint:
1573 case Intrinsic::nearbyint:
1574 case Intrinsic::pow:
1575 case Intrinsic::fma:
1576 case Intrinsic::fmuladd:
1577 return II->getIntrinsicID();
1579 return Intrinsic::not_intrinsic;
1584 return Intrinsic::not_intrinsic;
1587 Function *F = CI->getCalledFunction();
1588 // We're going to make assumptions on the semantics of the functions, check
1589 // that the target knows that it's available in this environment.
1590 if (!F || !TLI->getLibFunc(F->getName(), Func))
1591 return Intrinsic::not_intrinsic;
1593 // Otherwise check if we have a call to a function that can be turned into a
1594 // vector intrinsic.
1601 return Intrinsic::sin;
1605 return Intrinsic::cos;
1609 return Intrinsic::exp;
1611 case LibFunc::exp2f:
1612 case LibFunc::exp2l:
1613 return Intrinsic::exp2;
1617 return Intrinsic::log;
1618 case LibFunc::log10:
1619 case LibFunc::log10f:
1620 case LibFunc::log10l:
1621 return Intrinsic::log10;
1623 case LibFunc::log2f:
1624 case LibFunc::log2l:
1625 return Intrinsic::log2;
1627 case LibFunc::fabsf:
1628 case LibFunc::fabsl:
1629 return Intrinsic::fabs;
1630 case LibFunc::floor:
1631 case LibFunc::floorf:
1632 case LibFunc::floorl:
1633 return Intrinsic::floor;
1635 case LibFunc::ceilf:
1636 case LibFunc::ceill:
1637 return Intrinsic::ceil;
1638 case LibFunc::trunc:
1639 case LibFunc::truncf:
1640 case LibFunc::truncl:
1641 return Intrinsic::trunc;
1643 case LibFunc::rintf:
1644 case LibFunc::rintl:
1645 return Intrinsic::rint;
1646 case LibFunc::nearbyint:
1647 case LibFunc::nearbyintf:
1648 case LibFunc::nearbyintl:
1649 return Intrinsic::nearbyint;
1653 return Intrinsic::pow;
1656 return Intrinsic::not_intrinsic;
1659 /// This function translates the reduction kind to an LLVM binary operator.
1661 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1663 case LoopVectorizationLegality::RK_IntegerAdd:
1664 return Instruction::Add;
1665 case LoopVectorizationLegality::RK_IntegerMult:
1666 return Instruction::Mul;
1667 case LoopVectorizationLegality::RK_IntegerOr:
1668 return Instruction::Or;
1669 case LoopVectorizationLegality::RK_IntegerAnd:
1670 return Instruction::And;
1671 case LoopVectorizationLegality::RK_IntegerXor:
1672 return Instruction::Xor;
1673 case LoopVectorizationLegality::RK_FloatMult:
1674 return Instruction::FMul;
1675 case LoopVectorizationLegality::RK_FloatAdd:
1676 return Instruction::FAdd;
1677 case LoopVectorizationLegality::RK_IntegerMinMax:
1678 return Instruction::ICmp;
1679 case LoopVectorizationLegality::RK_FloatMinMax:
1680 return Instruction::FCmp;
1682 llvm_unreachable("Unknown reduction operation");
1686 Value *createMinMaxOp(IRBuilder<> &Builder,
1687 LoopVectorizationLegality::MinMaxReductionKind RK,
1690 CmpInst::Predicate P = CmpInst::ICMP_NE;
1693 llvm_unreachable("Unknown min/max reduction kind");
1694 case LoopVectorizationLegality::MRK_UIntMin:
1695 P = CmpInst::ICMP_ULT;
1697 case LoopVectorizationLegality::MRK_UIntMax:
1698 P = CmpInst::ICMP_UGT;
1700 case LoopVectorizationLegality::MRK_SIntMin:
1701 P = CmpInst::ICMP_SLT;
1703 case LoopVectorizationLegality::MRK_SIntMax:
1704 P = CmpInst::ICMP_SGT;
1706 case LoopVectorizationLegality::MRK_FloatMin:
1707 P = CmpInst::FCMP_OLT;
1709 case LoopVectorizationLegality::MRK_FloatMax:
1710 P = CmpInst::FCMP_OGT;
1715 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1716 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1718 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1720 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1725 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1726 //===------------------------------------------------===//
1728 // Notice: any optimization or new instruction that go
1729 // into the code below should be also be implemented in
1732 //===------------------------------------------------===//
1733 Constant *Zero = Builder.getInt32(0);
1735 // In order to support reduction variables we need to be able to vectorize
1736 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1737 // stages. First, we create a new vector PHI node with no incoming edges.
1738 // We use this value when we vectorize all of the instructions that use the
1739 // PHI. Next, after all of the instructions in the block are complete we
1740 // add the new incoming edges to the PHI. At this point all of the
1741 // instructions in the basic block are vectorized, so we can use them to
1742 // construct the PHI.
1743 PhiVector RdxPHIsToFix;
1745 // Scan the loop in a topological order to ensure that defs are vectorized
1747 LoopBlocksDFS DFS(OrigLoop);
1750 // Vectorize all of the blocks in the original loop.
1751 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1752 be = DFS.endRPO(); bb != be; ++bb)
1753 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1755 // At this point every instruction in the original loop is widened to
1756 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1757 // that we vectorized. The PHI nodes are currently empty because we did
1758 // not want to introduce cycles. Notice that the remaining PHI nodes
1759 // that we need to fix are reduction variables.
1761 // Create the 'reduced' values for each of the induction vars.
1762 // The reduced values are the vector values that we scalarize and combine
1763 // after the loop is finished.
1764 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1766 PHINode *RdxPhi = *it;
1767 assert(RdxPhi && "Unable to recover vectorized PHI");
1769 // Find the reduction variable descriptor.
1770 assert(Legal->getReductionVars()->count(RdxPhi) &&
1771 "Unable to find the reduction variable");
1772 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1773 (*Legal->getReductionVars())[RdxPhi];
1775 // We need to generate a reduction vector from the incoming scalar.
1776 // To do so, we need to generate the 'identity' vector and overide
1777 // one of the elements with the incoming scalar reduction. We need
1778 // to do it in the vector-loop preheader.
1779 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1781 // This is the vector-clone of the value that leaves the loop.
1782 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1783 Type *VecTy = VectorExit[0]->getType();
1785 // Find the reduction identity variable. Zero for addition, or, xor,
1786 // one for multiplication, -1 for And.
1789 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
1790 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
1791 // MinMax reduction have the start value as their identify.
1792 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
1796 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1797 VecTy->getScalarType());
1798 Identity = ConstantVector::getSplat(VF, Iden);
1800 // This vector is the Identity vector where the first element is the
1801 // incoming scalar reduction.
1802 VectorStart = Builder.CreateInsertElement(Identity,
1803 RdxDesc.StartValue, Zero);
1806 // Fix the vector-loop phi.
1807 // We created the induction variable so we know that the
1808 // preheader is the first entry.
1809 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1811 // Reductions do not have to start at zero. They can start with
1812 // any loop invariant values.
1813 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1814 BasicBlock *Latch = OrigLoop->getLoopLatch();
1815 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1816 VectorParts &Val = getVectorValue(LoopVal);
1817 for (unsigned part = 0; part < UF; ++part) {
1818 // Make sure to add the reduction stat value only to the
1819 // first unroll part.
1820 Value *StartVal = (part == 0) ? VectorStart : Identity;
1821 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1822 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1825 // Before each round, move the insertion point right between
1826 // the PHIs and the values we are going to write.
1827 // This allows us to write both PHINodes and the extractelement
1829 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1831 VectorParts RdxParts;
1832 for (unsigned part = 0; part < UF; ++part) {
1833 // This PHINode contains the vectorized reduction variable, or
1834 // the initial value vector, if we bypass the vector loop.
1835 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1836 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1837 Value *StartVal = (part == 0) ? VectorStart : Identity;
1838 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1839 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1840 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1841 RdxParts.push_back(NewPhi);
1844 // Reduce all of the unrolled parts into a single vector.
1845 Value *ReducedPartRdx = RdxParts[0];
1846 unsigned Op = getReductionBinOp(RdxDesc.Kind);
1847 for (unsigned part = 1; part < UF; ++part) {
1848 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1849 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
1850 RdxParts[part], ReducedPartRdx,
1853 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
1854 ReducedPartRdx, RdxParts[part]);
1857 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1858 // and vector ops, reducing the set of values being computed by half each
1860 assert(isPowerOf2_32(VF) &&
1861 "Reduction emission only supported for pow2 vectors!");
1862 Value *TmpVec = ReducedPartRdx;
1863 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1864 for (unsigned i = VF; i != 1; i >>= 1) {
1865 // Move the upper half of the vector to the lower half.
1866 for (unsigned j = 0; j != i/2; ++j)
1867 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1869 // Fill the rest of the mask with undef.
1870 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1871 UndefValue::get(Builder.getInt32Ty()));
1874 Builder.CreateShuffleVector(TmpVec,
1875 UndefValue::get(TmpVec->getType()),
1876 ConstantVector::get(ShuffleMask),
1879 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1880 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1883 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
1886 // The result is in the first element of the vector.
1887 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1889 // Now, we need to fix the users of the reduction variable
1890 // inside and outside of the scalar remainder loop.
1891 // We know that the loop is in LCSSA form. We need to update the
1892 // PHI nodes in the exit blocks.
1893 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1894 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1895 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1896 if (!LCSSAPhi) continue;
1898 // All PHINodes need to have a single entry edge, or two if
1899 // we already fixed them.
1900 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1902 // We found our reduction value exit-PHI. Update it with the
1903 // incoming bypass edge.
1904 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1905 // Add an edge coming from the bypass.
1906 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1909 }// end of the LCSSA phi scan.
1911 // Fix the scalar loop reduction variable with the incoming reduction sum
1912 // from the vector body and from the backedge value.
1913 int IncomingEdgeBlockIdx =
1914 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1915 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1916 // Pick the other block.
1917 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1918 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1919 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1920 }// end of for each redux variable.
1922 // The Loop exit block may have single value PHI nodes where the incoming
1923 // value is 'undef'. While vectorizing we only handled real values that
1924 // were defined inside the loop. Here we handle the 'undef case'.
1926 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1927 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1928 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1929 if (!LCSSAPhi) continue;
1930 if (LCSSAPhi->getNumIncomingValues() == 1)
1931 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1936 InnerLoopVectorizer::VectorParts
1937 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1938 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1941 VectorParts SrcMask = createBlockInMask(Src);
1943 // The terminator has to be a branch inst!
1944 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1945 assert(BI && "Unexpected terminator found");
1947 if (BI->isConditional()) {
1948 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1950 if (BI->getSuccessor(0) != Dst)
1951 for (unsigned part = 0; part < UF; ++part)
1952 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1954 for (unsigned part = 0; part < UF; ++part)
1955 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1962 InnerLoopVectorizer::VectorParts
1963 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1964 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1966 // Loop incoming mask is all-one.
1967 if (OrigLoop->getHeader() == BB) {
1968 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1969 return getVectorValue(C);
1972 // This is the block mask. We OR all incoming edges, and with zero.
1973 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1974 VectorParts BlockMask = getVectorValue(Zero);
1977 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1978 VectorParts EM = createEdgeMask(*it, BB);
1979 for (unsigned part = 0; part < UF; ++part)
1980 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1987 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1988 BasicBlock *BB, PhiVector *PV) {
1989 // For each instruction in the old loop.
1990 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1991 VectorParts &Entry = WidenMap.get(it);
1992 switch (it->getOpcode()) {
1993 case Instruction::Br:
1994 // Nothing to do for PHIs and BR, since we already took care of the
1995 // loop control flow instructions.
1997 case Instruction::PHI:{
1998 PHINode* P = cast<PHINode>(it);
1999 // Handle reduction variables:
2000 if (Legal->getReductionVars()->count(P)) {
2001 for (unsigned part = 0; part < UF; ++part) {
2002 // This is phase one of vectorizing PHIs.
2003 Type *VecTy = VectorType::get(it->getType(), VF);
2004 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2005 LoopVectorBody-> getFirstInsertionPt());
2011 // Check for PHI nodes that are lowered to vector selects.
2012 if (P->getParent() != OrigLoop->getHeader()) {
2013 // We know that all PHIs in non header blocks are converted into
2014 // selects, so we don't have to worry about the insertion order and we
2015 // can just use the builder.
2016 // At this point we generate the predication tree. There may be
2017 // duplications since this is a simple recursive scan, but future
2018 // optimizations will clean it up.
2020 unsigned NumIncoming = P->getNumIncomingValues();
2021 assert(NumIncoming > 1 && "Invalid PHI");
2023 // Generate a sequence of selects of the form:
2024 // SELECT(Mask3, In3,
2025 // SELECT(Mask2, In2,
2027 for (unsigned In = 0; In < NumIncoming; In++) {
2028 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2030 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2032 for (unsigned part = 0; part < UF; ++part) {
2033 // We don't need to 'select' the first PHI operand because it is
2034 // the default value if all of the other masks don't match.
2036 Entry[part] = In0[part];
2038 // Select between the current value and the previous incoming edge
2039 // based on the incoming mask.
2040 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2041 Entry[part], "predphi");
2047 // This PHINode must be an induction variable.
2048 // Make sure that we know about it.
2049 assert(Legal->getInductionVars()->count(P) &&
2050 "Not an induction variable");
2052 LoopVectorizationLegality::InductionInfo II =
2053 Legal->getInductionVars()->lookup(P);
2056 case LoopVectorizationLegality::IK_NoInduction:
2057 llvm_unreachable("Unknown induction");
2058 case LoopVectorizationLegality::IK_IntInduction: {
2059 assert(P == OldInduction && "Unexpected PHI");
2060 // We might have had to extend the type.
2061 Value *Trunc = Builder.CreateTrunc(Induction, P->getType());
2062 Value *Broadcasted = getBroadcastInstrs(Trunc);
2063 // After broadcasting the induction variable we need to make the
2064 // vector consecutive by adding 0, 1, 2 ...
2065 for (unsigned part = 0; part < UF; ++part)
2066 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2069 case LoopVectorizationLegality::IK_ReverseIntInduction:
2070 case LoopVectorizationLegality::IK_PtrInduction:
2071 case LoopVectorizationLegality::IK_ReversePtrInduction:
2072 // Handle reverse integer and pointer inductions.
2073 Value *StartIdx = ExtendedIdx;
2074 // This is the normalized GEP that starts counting at zero.
2075 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2078 // Handle the reverse integer induction variable case.
2079 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2080 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2081 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2083 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2086 // This is a new value so do not hoist it out.
2087 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2088 // After broadcasting the induction variable we need to make the
2089 // vector consecutive by adding ... -3, -2, -1, 0.
2090 for (unsigned part = 0; part < UF; ++part)
2091 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2096 // Handle the pointer induction variable case.
2097 assert(P->getType()->isPointerTy() && "Unexpected type.");
2099 // Is this a reverse induction ptr or a consecutive induction ptr.
2100 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2103 // This is the vector of results. Notice that we don't generate
2104 // vector geps because scalar geps result in better code.
2105 for (unsigned part = 0; part < UF; ++part) {
2106 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2107 for (unsigned int i = 0; i < VF; ++i) {
2108 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2109 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2112 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2114 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2116 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2118 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2119 Builder.getInt32(i),
2122 Entry[part] = VecVal;
2129 case Instruction::Add:
2130 case Instruction::FAdd:
2131 case Instruction::Sub:
2132 case Instruction::FSub:
2133 case Instruction::Mul:
2134 case Instruction::FMul:
2135 case Instruction::UDiv:
2136 case Instruction::SDiv:
2137 case Instruction::FDiv:
2138 case Instruction::URem:
2139 case Instruction::SRem:
2140 case Instruction::FRem:
2141 case Instruction::Shl:
2142 case Instruction::LShr:
2143 case Instruction::AShr:
2144 case Instruction::And:
2145 case Instruction::Or:
2146 case Instruction::Xor: {
2147 // Just widen binops.
2148 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2149 VectorParts &A = getVectorValue(it->getOperand(0));
2150 VectorParts &B = getVectorValue(it->getOperand(1));
2152 // Use this vector value for all users of the original instruction.
2153 for (unsigned Part = 0; Part < UF; ++Part) {
2154 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2156 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2157 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2158 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2159 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2160 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2162 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2163 VecOp->setIsExact(BinOp->isExact());
2169 case Instruction::Select: {
2171 // If the selector is loop invariant we can create a select
2172 // instruction with a scalar condition. Otherwise, use vector-select.
2173 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2176 // The condition can be loop invariant but still defined inside the
2177 // loop. This means that we can't just use the original 'cond' value.
2178 // We have to take the 'vectorized' value and pick the first lane.
2179 // Instcombine will make this a no-op.
2180 VectorParts &Cond = getVectorValue(it->getOperand(0));
2181 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2182 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2183 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2184 Builder.getInt32(0));
2185 for (unsigned Part = 0; Part < UF; ++Part) {
2186 Entry[Part] = Builder.CreateSelect(
2187 InvariantCond ? ScalarCond : Cond[Part],
2194 case Instruction::ICmp:
2195 case Instruction::FCmp: {
2196 // Widen compares. Generate vector compares.
2197 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2198 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2199 VectorParts &A = getVectorValue(it->getOperand(0));
2200 VectorParts &B = getVectorValue(it->getOperand(1));
2201 for (unsigned Part = 0; Part < UF; ++Part) {
2204 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2206 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2212 case Instruction::Store:
2213 case Instruction::Load:
2214 vectorizeMemoryInstruction(it, Legal);
2216 case Instruction::ZExt:
2217 case Instruction::SExt:
2218 case Instruction::FPToUI:
2219 case Instruction::FPToSI:
2220 case Instruction::FPExt:
2221 case Instruction::PtrToInt:
2222 case Instruction::IntToPtr:
2223 case Instruction::SIToFP:
2224 case Instruction::UIToFP:
2225 case Instruction::Trunc:
2226 case Instruction::FPTrunc:
2227 case Instruction::BitCast: {
2228 CastInst *CI = dyn_cast<CastInst>(it);
2229 /// Optimize the special case where the source is the induction
2230 /// variable. Notice that we can only optimize the 'trunc' case
2231 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2232 /// c. other casts depend on pointer size.
2233 if (CI->getOperand(0) == OldInduction &&
2234 it->getOpcode() == Instruction::Trunc) {
2235 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2237 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2238 for (unsigned Part = 0; Part < UF; ++Part)
2239 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2242 /// Vectorize casts.
2243 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2245 VectorParts &A = getVectorValue(it->getOperand(0));
2246 for (unsigned Part = 0; Part < UF; ++Part)
2247 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2251 case Instruction::Call: {
2252 // Ignore dbg intrinsics.
2253 if (isa<DbgInfoIntrinsic>(it))
2256 Module *M = BB->getParent()->getParent();
2257 CallInst *CI = cast<CallInst>(it);
2258 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2259 assert(ID && "Not an intrinsic call!");
2260 for (unsigned Part = 0; Part < UF; ++Part) {
2261 SmallVector<Value*, 4> Args;
2262 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2263 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2264 Args.push_back(Arg[Part]);
2266 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2267 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2268 Entry[Part] = Builder.CreateCall(F, Args);
2274 // All other instructions are unsupported. Scalarize them.
2275 scalarizeInstruction(it);
2278 }// end of for_each instr.
2281 void InnerLoopVectorizer::updateAnalysis() {
2282 // Forget the original basic block.
2283 SE->forgetLoop(OrigLoop);
2285 // Update the dominator tree information.
2286 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2287 "Entry does not dominate exit.");
2289 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2290 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2291 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2292 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2293 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2294 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2295 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2296 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2298 DEBUG(DT->verifyAnalysis());
2301 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2302 if (!EnableIfConversion)
2305 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2306 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2308 // Collect the blocks that need predication.
2309 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2310 BasicBlock *BB = LoopBlocks[i];
2312 // We don't support switch statements inside loops.
2313 if (!isa<BranchInst>(BB->getTerminator()))
2316 // We must be able to predicate all blocks that need to be predicated.
2317 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2321 // We can if-convert this loop.
2325 bool LoopVectorizationLegality::canVectorize() {
2326 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2328 // We can only vectorize innermost loops.
2329 if (TheLoop->getSubLoopsVector().size())
2332 // We must have a single backedge.
2333 if (TheLoop->getNumBackEdges() != 1)
2336 // We must have a single exiting block.
2337 if (!TheLoop->getExitingBlock())
2340 unsigned NumBlocks = TheLoop->getNumBlocks();
2342 // Check if we can if-convert non single-bb loops.
2343 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2344 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2348 // We need to have a loop header.
2349 BasicBlock *Latch = TheLoop->getLoopLatch();
2350 DEBUG(dbgs() << "LV: Found a loop: " <<
2351 TheLoop->getHeader()->getName() << "\n");
2353 // ScalarEvolution needs to be able to find the exit count.
2354 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2355 if (ExitCount == SE->getCouldNotCompute()) {
2356 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2360 // Do not loop-vectorize loops with a tiny trip count.
2361 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2362 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2363 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2364 "This loop is not worth vectorizing.\n");
2368 // Check if we can vectorize the instructions and CFG in this loop.
2369 if (!canVectorizeInstrs()) {
2370 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2374 // Go over each instruction and look at memory deps.
2375 if (!canVectorizeMemory()) {
2376 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2380 // Collect all of the variables that remain uniform after vectorization.
2381 collectLoopUniforms();
2383 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2384 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2387 // Okay! We can vectorize. At this point we don't have any other mem analysis
2388 // which may limit our maximum vectorization factor, so just return true with
2393 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2394 if (Ty->isPointerTy())
2395 return DL.getIntPtrType(Ty->getContext());
2399 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2400 Ty0 = convertPointerToIntegerType(DL, Ty0);
2401 Ty1 = convertPointerToIntegerType(DL, Ty1);
2402 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2407 bool LoopVectorizationLegality::canVectorizeInstrs() {
2408 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2409 BasicBlock *Header = TheLoop->getHeader();
2411 // If we marked the scalar loop as "already vectorized" then no need
2412 // to vectorize it again.
2413 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) {
2414 DEBUG(dbgs() << "LV: This loop was vectorized before\n");
2418 // Look for the attribute signaling the absence of NaNs.
2419 Function &F = *Header->getParent();
2420 if (F.hasFnAttribute("no-nans-fp-math"))
2421 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2422 AttributeSet::FunctionIndex,
2423 "no-nans-fp-math").getValueAsString() == "true";
2425 // For each block in the loop.
2426 for (Loop::block_iterator bb = TheLoop->block_begin(),
2427 be = TheLoop->block_end(); bb != be; ++bb) {
2429 // Scan the instructions in the block and look for hazards.
2430 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2433 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2434 Type *PhiTy = Phi->getType();
2435 // Check that this PHI type is allowed.
2436 if (!PhiTy->isIntegerTy() &&
2437 !PhiTy->isFloatingPointTy() &&
2438 !PhiTy->isPointerTy()) {
2439 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2443 // If this PHINode is not in the header block, then we know that we
2444 // can convert it to select during if-conversion. No need to check if
2445 // the PHIs in this block are induction or reduction variables.
2449 // We only allow if-converted PHIs with more than two incoming values.
2450 if (Phi->getNumIncomingValues() != 2) {
2451 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2455 // This is the value coming from the preheader.
2456 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2457 // Check if this is an induction variable.
2458 InductionKind IK = isInductionVariable(Phi);
2460 if (IK_NoInduction != IK) {
2461 // Get the widest type.
2463 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2465 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2467 // Int inductions are special because we only allow one IV.
2468 if (IK == IK_IntInduction) {
2470 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2476 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2477 Inductions[Phi] = InductionInfo(StartValue, IK);
2481 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2482 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2485 if (AddReductionVar(Phi, RK_IntegerMult)) {
2486 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2489 if (AddReductionVar(Phi, RK_IntegerOr)) {
2490 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2493 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2494 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2497 if (AddReductionVar(Phi, RK_IntegerXor)) {
2498 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2501 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2502 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2505 if (AddReductionVar(Phi, RK_FloatMult)) {
2506 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2509 if (AddReductionVar(Phi, RK_FloatAdd)) {
2510 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2513 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2514 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2518 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2520 }// end of PHI handling
2522 // We still don't handle functions. However, we can ignore dbg intrinsic
2523 // calls and we do handle certain intrinsic and libm functions.
2524 CallInst *CI = dyn_cast<CallInst>(it);
2525 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2526 DEBUG(dbgs() << "LV: Found a call site.\n");
2530 // Check that the instruction return type is vectorizable.
2531 if (!VectorType::isValidElementType(it->getType()) &&
2532 !it->getType()->isVoidTy()) {
2533 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2537 // Check that the stored type is vectorizable.
2538 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2539 Type *T = ST->getValueOperand()->getType();
2540 if (!VectorType::isValidElementType(T))
2544 // Reduction instructions are allowed to have exit users.
2545 // All other instructions must not have external users.
2546 if (!AllowedExit.count(it))
2547 //Check that all of the users of the loop are inside the BB.
2548 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2550 Instruction *U = cast<Instruction>(*I);
2551 // This user may be a reduction exit value.
2552 if (!TheLoop->contains(U)) {
2553 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2562 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2563 if (Inductions.empty())
2570 void LoopVectorizationLegality::collectLoopUniforms() {
2571 // We now know that the loop is vectorizable!
2572 // Collect variables that will remain uniform after vectorization.
2573 std::vector<Value*> Worklist;
2574 BasicBlock *Latch = TheLoop->getLoopLatch();
2576 // Start with the conditional branch and walk up the block.
2577 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2579 while (Worklist.size()) {
2580 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2581 Worklist.pop_back();
2583 // Look at instructions inside this loop.
2584 // Stop when reaching PHI nodes.
2585 // TODO: we need to follow values all over the loop, not only in this block.
2586 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2589 // This is a known uniform.
2592 // Insert all operands.
2593 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2594 Worklist.push_back(I->getOperand(i));
2599 AliasAnalysis::Location
2600 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2601 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2602 return AA->getLocation(Store);
2603 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2604 return AA->getLocation(Load);
2606 llvm_unreachable("Should be either load or store instruction");
2610 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2613 AliasMultiMap& WriteObjects,
2614 unsigned MaxByteWidth) {
2616 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2618 std::vector<Instruction*>::iterator
2619 it = WriteObjects[Object].begin(),
2620 end = WriteObjects[Object].end();
2622 for (; it != end; ++it) {
2623 Instruction* I = *it;
2627 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2628 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2629 ThatLoc.getWithNewSize(MaxByteWidth)))
2635 bool LoopVectorizationLegality::canVectorizeMemory() {
2637 typedef SmallVector<Value*, 16> ValueVector;
2638 typedef SmallPtrSet<Value*, 16> ValueSet;
2639 // Holds the Load and Store *instructions*.
2642 PtrRtCheck.Pointers.clear();
2643 PtrRtCheck.Need = false;
2645 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2648 for (Loop::block_iterator bb = TheLoop->block_begin(),
2649 be = TheLoop->block_end(); bb != be; ++bb) {
2651 // Scan the BB and collect legal loads and stores.
2652 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2655 // If this is a load, save it. If this instruction can read from memory
2656 // but is not a load, then we quit. Notice that we don't handle function
2657 // calls that read or write.
2658 if (it->mayReadFromMemory()) {
2659 LoadInst *Ld = dyn_cast<LoadInst>(it);
2660 if (!Ld) return false;
2661 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2662 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2665 Loads.push_back(Ld);
2669 // Save 'store' instructions. Abort if other instructions write to memory.
2670 if (it->mayWriteToMemory()) {
2671 StoreInst *St = dyn_cast<StoreInst>(it);
2672 if (!St) return false;
2673 if (!St->isSimple() && !IsAnnotatedParallel) {
2674 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2677 Stores.push_back(St);
2682 // Now we have two lists that hold the loads and the stores.
2683 // Next, we find the pointers that they use.
2685 // Check if we see any stores. If there are no stores, then we don't
2686 // care if the pointers are *restrict*.
2687 if (!Stores.size()) {
2688 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2692 // Holds the read and read-write *pointers* that we find. These maps hold
2693 // unique values for pointers (so no need for multi-map).
2695 AliasMap ReadWrites;
2697 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2698 // multiple times on the same object. If the ptr is accessed twice, once
2699 // for read and once for write, it will only appear once (on the write
2700 // list). This is okay, since we are going to check for conflicts between
2701 // writes and between reads and writes, but not between reads and reads.
2704 ValueVector::iterator I, IE;
2705 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2706 StoreInst *ST = cast<StoreInst>(*I);
2707 Value* Ptr = ST->getPointerOperand();
2709 if (isUniform(Ptr)) {
2710 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2714 // If we did *not* see this pointer before, insert it to
2715 // the read-write list. At this phase it is only a 'write' list.
2716 if (Seen.insert(Ptr))
2717 ReadWrites.insert(std::make_pair(Ptr, ST));
2720 if (IsAnnotatedParallel) {
2722 << "LV: A loop annotated parallel, ignore memory dependency "
2727 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2728 LoadInst *LD = cast<LoadInst>(*I);
2729 Value* Ptr = LD->getPointerOperand();
2730 // If we did *not* see this pointer before, insert it to the
2731 // read list. If we *did* see it before, then it is already in
2732 // the read-write list. This allows us to vectorize expressions
2733 // such as A[i] += x; Because the address of A[i] is a read-write
2734 // pointer. This only works if the index of A[i] is consecutive.
2735 // If the address of i is unknown (for example A[B[i]]) then we may
2736 // read a few words, modify, and write a few words, and some of the
2737 // words may be written to the same address.
2738 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2739 Reads.insert(std::make_pair(Ptr, LD));
2742 // If we write (or read-write) to a single destination and there are no
2743 // other reads in this loop then is it safe to vectorize.
2744 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2745 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2749 unsigned NumReadPtrs = 0;
2750 unsigned NumWritePtrs = 0;
2752 // Find pointers with computable bounds. We are going to use this information
2753 // to place a runtime bound check.
2754 bool CanDoRT = true;
2755 AliasMap::iterator MI, ME;
2756 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2757 Value *V = (*MI).first;
2758 if (hasComputableBounds(V)) {
2759 PtrRtCheck.insert(SE, TheLoop, V, true);
2761 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2767 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2768 Value *V = (*MI).first;
2769 if (hasComputableBounds(V)) {
2770 PtrRtCheck.insert(SE, TheLoop, V, false);
2772 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2779 // Check that we did not collect too many pointers or found a
2780 // unsizeable pointer.
2781 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
2782 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
2783 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
2789 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2792 bool NeedRTCheck = false;
2794 // Biggest vectorized access possible, vector width * unroll factor.
2795 // TODO: We're being very pessimistic here, find a way to know the
2796 // real access width before getting here.
2797 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2798 TTI->getMaximumUnrollFactor();
2799 // Now that the pointers are in two lists (Reads and ReadWrites), we
2800 // can check that there are no conflicts between each of the writes and
2801 // between the writes to the reads.
2802 // Note that WriteObjects duplicates the stores (indexed now by underlying
2803 // objects) to avoid pointing to elements inside ReadWrites.
2804 // TODO: Maybe create a new type where they can interact without duplication.
2805 AliasMultiMap WriteObjects;
2806 ValueVector TempObjects;
2808 // Check that the read-writes do not conflict with other read-write
2810 bool AllWritesIdentified = true;
2811 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2812 Value *Val = (*MI).first;
2813 Instruction *Inst = (*MI).second;
2815 GetUnderlyingObjects(Val, TempObjects, DL);
2816 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2818 if (!isIdentifiedObject(*UI)) {
2819 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2821 AllWritesIdentified = false;
2824 // Never seen it before, can't alias.
2825 if (WriteObjects[*UI].empty()) {
2826 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2827 WriteObjects[*UI].push_back(Inst);
2830 // Direct alias found.
2831 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2832 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2836 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2838 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2839 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2841 // If global alias, make sure they do alias.
2842 if (hasPossibleGlobalWriteReorder(*UI,
2846 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
2851 // Didn't alias, insert into map for further reference.
2852 WriteObjects[*UI].push_back(Inst);
2854 TempObjects.clear();
2857 /// Check that the reads don't conflict with the read-writes.
2858 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2859 Value *Val = (*MI).first;
2860 GetUnderlyingObjects(Val, TempObjects, DL);
2861 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2863 // If all of the writes are identified then we don't care if the read
2864 // pointer is identified or not.
2865 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2866 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2870 // Never seen it before, can't alias.
2871 if (WriteObjects[*UI].empty())
2873 // Direct alias found.
2874 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2875 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2879 DEBUG(dbgs() << "LV: Found a global value: "
2881 Instruction *Inst = (*MI).second;
2882 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2883 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2885 // If global alias, make sure they do alias.
2886 if (hasPossibleGlobalWriteReorder(*UI,
2890 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
2895 TempObjects.clear();
2898 PtrRtCheck.Need = NeedRTCheck;
2899 if (NeedRTCheck && !CanDoRT) {
2900 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2901 "the array bounds.\n");
2906 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2907 " need a runtime memory check.\n");
2911 static bool hasMultipleUsesOf(Instruction *I,
2912 SmallPtrSet<Instruction *, 8> &Insts) {
2913 unsigned NumUses = 0;
2914 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
2915 if (Insts.count(dyn_cast<Instruction>(*Use)))
2924 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
2925 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
2926 if (!Set.count(dyn_cast<Instruction>(*Use)))
2931 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2932 ReductionKind Kind) {
2933 if (Phi->getNumIncomingValues() != 2)
2936 // Reduction variables are only found in the loop header block.
2937 if (Phi->getParent() != TheLoop->getHeader())
2940 // Obtain the reduction start value from the value that comes from the loop
2942 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2944 // ExitInstruction is the single value which is used outside the loop.
2945 // We only allow for a single reduction value to be used outside the loop.
2946 // This includes users of the reduction, variables (which form a cycle
2947 // which ends in the phi node).
2948 Instruction *ExitInstruction = 0;
2949 // Indicates that we found a reduction operation in our scan.
2950 bool FoundReduxOp = false;
2952 // We start with the PHI node and scan for all of the users of this
2953 // instruction. All users must be instructions that can be used as reduction
2954 // variables (such as ADD). We must have a single out-of-block user. The cycle
2955 // must include the original PHI.
2956 bool FoundStartPHI = false;
2958 // To recognize min/max patterns formed by a icmp select sequence, we store
2959 // the number of instruction we saw from the recognized min/max pattern,
2960 // to make sure we only see exactly the two instructions.
2961 unsigned NumCmpSelectPatternInst = 0;
2962 ReductionInstDesc ReduxDesc(false, 0);
2964 SmallPtrSet<Instruction *, 8> VisitedInsts;
2965 SmallVector<Instruction *, 8> Worklist;
2966 Worklist.push_back(Phi);
2967 VisitedInsts.insert(Phi);
2969 // A value in the reduction can be used:
2970 // - By the reduction:
2971 // - Reduction operation:
2972 // - One use of reduction value (safe).
2973 // - Multiple use of reduction value (not safe).
2975 // - All uses of the PHI must be the reduction (safe).
2976 // - Otherwise, not safe.
2977 // - By one instruction outside of the loop (safe).
2978 // - By further instructions outside of the loop (not safe).
2979 // - By an instruction that is not part of the reduction (not safe).
2981 // * An instruction type other than PHI or the reduction operation.
2982 // * A PHI in the header other than the initial PHI.
2983 while (!Worklist.empty()) {
2984 Instruction *Cur = Worklist.back();
2985 Worklist.pop_back();
2988 // If the instruction has no users then this is a broken chain and can't be
2989 // a reduction variable.
2990 if (Cur->use_empty())
2993 bool IsAPhi = isa<PHINode>(Cur);
2995 // A header PHI use other than the original PHI.
2996 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
2999 // Reductions of instructions such as Div, and Sub is only possible if the
3000 // LHS is the reduction variable.
3001 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3002 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3003 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3006 // Any reduction instruction must be of one of the allowed kinds.
3007 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3008 if (!ReduxDesc.IsReduction)
3011 // A reduction operation must only have one use of the reduction value.
3012 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3013 hasMultipleUsesOf(Cur, VisitedInsts))
3016 // All inputs to a PHI node must be a reduction value.
3017 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3020 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3021 isa<SelectInst>(Cur)))
3022 ++NumCmpSelectPatternInst;
3023 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3024 isa<SelectInst>(Cur)))
3025 ++NumCmpSelectPatternInst;
3027 // Check whether we found a reduction operator.
3028 FoundReduxOp |= !IsAPhi;
3030 // Process users of current instruction. Push non PHI nodes after PHI nodes
3031 // onto the stack. This way we are going to have seen all inputs to PHI
3032 // nodes once we get to them.
3033 SmallVector<Instruction *, 8> NonPHIs;
3034 SmallVector<Instruction *, 8> PHIs;
3035 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3037 Instruction *Usr = cast<Instruction>(*UI);
3039 // Check if we found the exit user.
3040 BasicBlock *Parent = Usr->getParent();
3041 if (!TheLoop->contains(Parent)) {
3042 // Exit if you find multiple outside users.
3043 if (ExitInstruction != 0)
3045 ExitInstruction = Cur;
3049 // Process instructions only once (termination).
3050 if (VisitedInsts.insert(Usr)) {
3051 if (isa<PHINode>(Usr))
3052 PHIs.push_back(Usr);
3054 NonPHIs.push_back(Usr);
3056 // Remember that we completed the cycle.
3058 FoundStartPHI = true;
3060 Worklist.append(PHIs.begin(), PHIs.end());
3061 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3064 // This means we have seen one but not the other instruction of the
3065 // pattern or more than just a select and cmp.
3066 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3067 NumCmpSelectPatternInst != 2)
3070 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3073 // We found a reduction var if we have reached the original phi node and we
3074 // only have a single instruction with out-of-loop users.
3076 // This instruction is allowed to have out-of-loop users.
3077 AllowedExit.insert(ExitInstruction);
3079 // Save the description of this reduction variable.
3080 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3081 ReduxDesc.MinMaxKind);
3082 Reductions[Phi] = RD;
3083 // We've ended the cycle. This is a reduction variable if we have an
3084 // outside user and it has a binary op.
3089 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3090 /// pattern corresponding to a min(X, Y) or max(X, Y).
3091 LoopVectorizationLegality::ReductionInstDesc
3092 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3093 ReductionInstDesc &Prev) {
3095 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3096 "Expect a select instruction");
3097 Instruction *Cmp = 0;
3098 SelectInst *Select = 0;
3100 // We must handle the select(cmp()) as a single instruction. Advance to the
3102 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3103 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3104 return ReductionInstDesc(false, I);
3105 return ReductionInstDesc(Select, Prev.MinMaxKind);
3108 // Only handle single use cases for now.
3109 if (!(Select = dyn_cast<SelectInst>(I)))
3110 return ReductionInstDesc(false, I);
3111 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3112 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3113 return ReductionInstDesc(false, I);
3114 if (!Cmp->hasOneUse())
3115 return ReductionInstDesc(false, I);
3120 // Look for a min/max pattern.
3121 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3122 return ReductionInstDesc(Select, MRK_UIntMin);
3123 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3124 return ReductionInstDesc(Select, MRK_UIntMax);
3125 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3126 return ReductionInstDesc(Select, MRK_SIntMax);
3127 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3128 return ReductionInstDesc(Select, MRK_SIntMin);
3129 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3130 return ReductionInstDesc(Select, MRK_FloatMin);
3131 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3132 return ReductionInstDesc(Select, MRK_FloatMax);
3133 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3134 return ReductionInstDesc(Select, MRK_FloatMin);
3135 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3136 return ReductionInstDesc(Select, MRK_FloatMax);
3138 return ReductionInstDesc(false, I);
3141 LoopVectorizationLegality::ReductionInstDesc
3142 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3144 ReductionInstDesc &Prev) {
3145 bool FP = I->getType()->isFloatingPointTy();
3146 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3147 switch (I->getOpcode()) {
3149 return ReductionInstDesc(false, I);
3150 case Instruction::PHI:
3151 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3152 Kind != RK_FloatMinMax))
3153 return ReductionInstDesc(false, I);
3154 return ReductionInstDesc(I, Prev.MinMaxKind);
3155 case Instruction::Sub:
3156 case Instruction::Add:
3157 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3158 case Instruction::Mul:
3159 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3160 case Instruction::And:
3161 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3162 case Instruction::Or:
3163 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3164 case Instruction::Xor:
3165 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3166 case Instruction::FMul:
3167 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3168 case Instruction::FAdd:
3169 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3170 case Instruction::FCmp:
3171 case Instruction::ICmp:
3172 case Instruction::Select:
3173 if (Kind != RK_IntegerMinMax &&
3174 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3175 return ReductionInstDesc(false, I);
3176 return isMinMaxSelectCmpPattern(I, Prev);
3180 LoopVectorizationLegality::InductionKind
3181 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3182 Type *PhiTy = Phi->getType();
3183 // We only handle integer and pointer inductions variables.
3184 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3185 return IK_NoInduction;
3187 // Check that the PHI is consecutive.
3188 const SCEV *PhiScev = SE->getSCEV(Phi);
3189 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3191 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3192 return IK_NoInduction;
3194 const SCEV *Step = AR->getStepRecurrence(*SE);
3196 // Integer inductions need to have a stride of one.
3197 if (PhiTy->isIntegerTy()) {
3199 return IK_IntInduction;
3200 if (Step->isAllOnesValue())
3201 return IK_ReverseIntInduction;
3202 return IK_NoInduction;
3205 // Calculate the pointer stride and check if it is consecutive.
3206 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3208 return IK_NoInduction;
3210 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3211 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3212 if (C->getValue()->equalsInt(Size))
3213 return IK_PtrInduction;
3214 else if (C->getValue()->equalsInt(0 - Size))
3215 return IK_ReversePtrInduction;
3217 return IK_NoInduction;
3220 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3221 Value *In0 = const_cast<Value*>(V);
3222 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3226 return Inductions.count(PN);
3229 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3230 assert(TheLoop->contains(BB) && "Unknown block used");
3232 // Blocks that do not dominate the latch need predication.
3233 BasicBlock* Latch = TheLoop->getLoopLatch();
3234 return !DT->dominates(BB, Latch);
3237 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3238 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3239 // We don't predicate loads/stores at the moment.
3240 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
3243 // The instructions below can trap.
3244 switch (it->getOpcode()) {
3246 case Instruction::UDiv:
3247 case Instruction::SDiv:
3248 case Instruction::URem:
3249 case Instruction::SRem:
3257 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3258 const SCEV *PhiScev = SE->getSCEV(Ptr);
3259 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3263 return AR->isAffine();
3266 LoopVectorizationCostModel::VectorizationFactor
3267 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3269 // Width 1 means no vectorize
3270 VectorizationFactor Factor = { 1U, 0U };
3271 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3272 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3276 // Find the trip count.
3277 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3278 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3280 unsigned WidestType = getWidestType();
3281 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3282 unsigned MaxVectorSize = WidestRegister / WidestType;
3283 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3284 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3286 if (MaxVectorSize == 0) {
3287 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3291 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3292 " into one vector!");
3294 unsigned VF = MaxVectorSize;
3296 // If we optimize the program for size, avoid creating the tail loop.
3298 // If we are unable to calculate the trip count then don't try to vectorize.
3300 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3304 // Find the maximum SIMD width that can fit within the trip count.
3305 VF = TC % MaxVectorSize;
3310 // If the trip count that we found modulo the vectorization factor is not
3311 // zero then we require a tail.
3313 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3319 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3320 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3322 Factor.Width = UserVF;
3326 float Cost = expectedCost(1);
3328 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3329 for (unsigned i=2; i <= VF; i*=2) {
3330 // Notice that the vector loop needs to be executed less times, so
3331 // we need to divide the cost of the vector loops by the width of
3332 // the vector elements.
3333 float VectorCost = expectedCost(i) / (float)i;
3334 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3335 (int)VectorCost << ".\n");
3336 if (VectorCost < Cost) {
3342 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3343 Factor.Width = Width;
3344 Factor.Cost = Width * Cost;
3348 unsigned LoopVectorizationCostModel::getWidestType() {
3349 unsigned MaxWidth = 8;
3352 for (Loop::block_iterator bb = TheLoop->block_begin(),
3353 be = TheLoop->block_end(); bb != be; ++bb) {
3354 BasicBlock *BB = *bb;
3356 // For each instruction in the loop.
3357 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3358 Type *T = it->getType();
3360 // Only examine Loads, Stores and PHINodes.
3361 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3364 // Examine PHI nodes that are reduction variables.
3365 if (PHINode *PN = dyn_cast<PHINode>(it))
3366 if (!Legal->getReductionVars()->count(PN))
3369 // Examine the stored values.
3370 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3371 T = ST->getValueOperand()->getType();
3373 // Ignore loaded pointer types and stored pointer types that are not
3374 // consecutive. However, we do want to take consecutive stores/loads of
3375 // pointer vectors into account.
3376 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3379 MaxWidth = std::max(MaxWidth,
3380 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3388 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3391 unsigned LoopCost) {
3393 // -- The unroll heuristics --
3394 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3395 // There are many micro-architectural considerations that we can't predict
3396 // at this level. For example frontend pressure (on decode or fetch) due to
3397 // code size, or the number and capabilities of the execution ports.
3399 // We use the following heuristics to select the unroll factor:
3400 // 1. If the code has reductions the we unroll in order to break the cross
3401 // iteration dependency.
3402 // 2. If the loop is really small then we unroll in order to reduce the loop
3404 // 3. We don't unroll if we think that we will spill registers to memory due
3405 // to the increased register pressure.
3407 // Use the user preference, unless 'auto' is selected.
3411 // When we optimize for size we don't unroll.
3415 // Do not unroll loops with a relatively small trip count.
3416 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3417 TheLoop->getLoopLatch());
3418 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3421 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3422 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3423 " vector registers\n");
3425 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3426 // We divide by these constants so assume that we have at least one
3427 // instruction that uses at least one register.
3428 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3429 R.NumInstructions = std::max(R.NumInstructions, 1U);
3431 // We calculate the unroll factor using the following formula.
3432 // Subtract the number of loop invariants from the number of available
3433 // registers. These registers are used by all of the unrolled instances.
3434 // Next, divide the remaining registers by the number of registers that is
3435 // required by the loop, in order to estimate how many parallel instances
3436 // fit without causing spills.
3437 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3439 // Clamp the unroll factor ranges to reasonable factors.
3440 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3442 // If we did not calculate the cost for VF (because the user selected the VF)
3443 // then we calculate the cost of VF here.
3445 LoopCost = expectedCost(VF);
3447 // Clamp the calculated UF to be between the 1 and the max unroll factor
3448 // that the target allows.
3449 if (UF > MaxUnrollSize)
3454 if (Legal->getReductionVars()->size()) {
3455 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3459 // We want to unroll tiny loops in order to reduce the loop overhead.
3460 // We assume that the cost overhead is 1 and we use the cost model
3461 // to estimate the cost of the loop and unroll until the cost of the
3462 // loop overhead is about 5% of the cost of the loop.
3463 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3464 if (LoopCost < 20) {
3465 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3466 unsigned NewUF = 20/LoopCost + 1;
3467 return std::min(NewUF, UF);
3470 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3474 LoopVectorizationCostModel::RegisterUsage
3475 LoopVectorizationCostModel::calculateRegisterUsage() {
3476 // This function calculates the register usage by measuring the highest number
3477 // of values that are alive at a single location. Obviously, this is a very
3478 // rough estimation. We scan the loop in a topological order in order and
3479 // assign a number to each instruction. We use RPO to ensure that defs are
3480 // met before their users. We assume that each instruction that has in-loop
3481 // users starts an interval. We record every time that an in-loop value is
3482 // used, so we have a list of the first and last occurrences of each
3483 // instruction. Next, we transpose this data structure into a multi map that
3484 // holds the list of intervals that *end* at a specific location. This multi
3485 // map allows us to perform a linear search. We scan the instructions linearly
3486 // and record each time that a new interval starts, by placing it in a set.
3487 // If we find this value in the multi-map then we remove it from the set.
3488 // The max register usage is the maximum size of the set.
3489 // We also search for instructions that are defined outside the loop, but are
3490 // used inside the loop. We need this number separately from the max-interval
3491 // usage number because when we unroll, loop-invariant values do not take
3493 LoopBlocksDFS DFS(TheLoop);
3497 R.NumInstructions = 0;
3499 // Each 'key' in the map opens a new interval. The values
3500 // of the map are the index of the 'last seen' usage of the
3501 // instruction that is the key.
3502 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3503 // Maps instruction to its index.
3504 DenseMap<unsigned, Instruction*> IdxToInstr;
3505 // Marks the end of each interval.
3506 IntervalMap EndPoint;
3507 // Saves the list of instruction indices that are used in the loop.
3508 SmallSet<Instruction*, 8> Ends;
3509 // Saves the list of values that are used in the loop but are
3510 // defined outside the loop, such as arguments and constants.
3511 SmallPtrSet<Value*, 8> LoopInvariants;
3514 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3515 be = DFS.endRPO(); bb != be; ++bb) {
3516 R.NumInstructions += (*bb)->size();
3517 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3519 Instruction *I = it;
3520 IdxToInstr[Index++] = I;
3522 // Save the end location of each USE.
3523 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3524 Value *U = I->getOperand(i);
3525 Instruction *Instr = dyn_cast<Instruction>(U);
3527 // Ignore non-instruction values such as arguments, constants, etc.
3528 if (!Instr) continue;
3530 // If this instruction is outside the loop then record it and continue.
3531 if (!TheLoop->contains(Instr)) {
3532 LoopInvariants.insert(Instr);
3536 // Overwrite previous end points.
3537 EndPoint[Instr] = Index;
3543 // Saves the list of intervals that end with the index in 'key'.
3544 typedef SmallVector<Instruction*, 2> InstrList;
3545 DenseMap<unsigned, InstrList> TransposeEnds;
3547 // Transpose the EndPoints to a list of values that end at each index.
3548 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3550 TransposeEnds[it->second].push_back(it->first);
3552 SmallSet<Instruction*, 8> OpenIntervals;
3553 unsigned MaxUsage = 0;
3556 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3557 for (unsigned int i = 0; i < Index; ++i) {
3558 Instruction *I = IdxToInstr[i];
3559 // Ignore instructions that are never used within the loop.
3560 if (!Ends.count(I)) continue;
3562 // Remove all of the instructions that end at this location.
3563 InstrList &List = TransposeEnds[i];
3564 for (unsigned int j=0, e = List.size(); j < e; ++j)
3565 OpenIntervals.erase(List[j]);
3567 // Count the number of live interals.
3568 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3570 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3571 OpenIntervals.size() <<"\n");
3573 // Add the current instruction to the list of open intervals.
3574 OpenIntervals.insert(I);
3577 unsigned Invariant = LoopInvariants.size();
3578 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3579 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3580 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3582 R.LoopInvariantRegs = Invariant;
3583 R.MaxLocalUsers = MaxUsage;
3587 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3591 for (Loop::block_iterator bb = TheLoop->block_begin(),
3592 be = TheLoop->block_end(); bb != be; ++bb) {
3593 unsigned BlockCost = 0;
3594 BasicBlock *BB = *bb;
3596 // For each instruction in the old loop.
3597 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3598 // Skip dbg intrinsics.
3599 if (isa<DbgInfoIntrinsic>(it))
3602 unsigned C = getInstructionCost(it, VF);
3604 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3605 VF << " For instruction: "<< *it << "\n");
3608 // We assume that if-converted blocks have a 50% chance of being executed.
3609 // When the code is scalar then some of the blocks are avoided due to CF.
3610 // When the code is vectorized we execute all code paths.
3611 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3621 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3622 // If we know that this instruction will remain uniform, check the cost of
3623 // the scalar version.
3624 if (Legal->isUniformAfterVectorization(I))
3627 Type *RetTy = I->getType();
3628 Type *VectorTy = ToVectorTy(RetTy, VF);
3630 // TODO: We need to estimate the cost of intrinsic calls.
3631 switch (I->getOpcode()) {
3632 case Instruction::GetElementPtr:
3633 // We mark this instruction as zero-cost because the cost of GEPs in
3634 // vectorized code depends on whether the corresponding memory instruction
3635 // is scalarized or not. Therefore, we handle GEPs with the memory
3636 // instruction cost.
3638 case Instruction::Br: {
3639 return TTI.getCFInstrCost(I->getOpcode());
3641 case Instruction::PHI:
3642 //TODO: IF-converted IFs become selects.
3644 case Instruction::Add:
3645 case Instruction::FAdd:
3646 case Instruction::Sub:
3647 case Instruction::FSub:
3648 case Instruction::Mul:
3649 case Instruction::FMul:
3650 case Instruction::UDiv:
3651 case Instruction::SDiv:
3652 case Instruction::FDiv:
3653 case Instruction::URem:
3654 case Instruction::SRem:
3655 case Instruction::FRem:
3656 case Instruction::Shl:
3657 case Instruction::LShr:
3658 case Instruction::AShr:
3659 case Instruction::And:
3660 case Instruction::Or:
3661 case Instruction::Xor: {
3662 // Certain instructions can be cheaper to vectorize if they have a constant
3663 // second vector operand. One example of this are shifts on x86.
3664 TargetTransformInfo::OperandValueKind Op1VK =
3665 TargetTransformInfo::OK_AnyValue;
3666 TargetTransformInfo::OperandValueKind Op2VK =
3667 TargetTransformInfo::OK_AnyValue;
3669 if (isa<ConstantInt>(I->getOperand(1)))
3670 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3672 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3674 case Instruction::Select: {
3675 SelectInst *SI = cast<SelectInst>(I);
3676 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3677 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3678 Type *CondTy = SI->getCondition()->getType();
3680 CondTy = VectorType::get(CondTy, VF);
3682 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3684 case Instruction::ICmp:
3685 case Instruction::FCmp: {
3686 Type *ValTy = I->getOperand(0)->getType();
3687 VectorTy = ToVectorTy(ValTy, VF);
3688 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3690 case Instruction::Store:
3691 case Instruction::Load: {
3692 StoreInst *SI = dyn_cast<StoreInst>(I);
3693 LoadInst *LI = dyn_cast<LoadInst>(I);
3694 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3696 VectorTy = ToVectorTy(ValTy, VF);
3698 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3699 unsigned AS = SI ? SI->getPointerAddressSpace() :
3700 LI->getPointerAddressSpace();
3701 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3702 // We add the cost of address computation here instead of with the gep
3703 // instruction because only here we know whether the operation is
3706 return TTI.getAddressComputationCost(VectorTy) +
3707 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3709 // Scalarized loads/stores.
3710 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3711 bool Reverse = ConsecutiveStride < 0;
3712 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3713 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3714 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3716 // The cost of extracting from the value vector and pointer vector.
3717 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3718 for (unsigned i = 0; i < VF; ++i) {
3719 // The cost of extracting the pointer operand.
3720 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3721 // In case of STORE, the cost of ExtractElement from the vector.
3722 // In case of LOAD, the cost of InsertElement into the returned
3724 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3725 Instruction::InsertElement,
3729 // The cost of the scalar loads/stores.
3730 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3731 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3736 // Wide load/stores.
3737 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3738 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3741 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3745 case Instruction::ZExt:
3746 case Instruction::SExt:
3747 case Instruction::FPToUI:
3748 case Instruction::FPToSI:
3749 case Instruction::FPExt:
3750 case Instruction::PtrToInt:
3751 case Instruction::IntToPtr:
3752 case Instruction::SIToFP:
3753 case Instruction::UIToFP:
3754 case Instruction::Trunc:
3755 case Instruction::FPTrunc:
3756 case Instruction::BitCast: {
3757 // We optimize the truncation of induction variable.
3758 // The cost of these is the same as the scalar operation.
3759 if (I->getOpcode() == Instruction::Trunc &&
3760 Legal->isInductionVariable(I->getOperand(0)))
3761 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3762 I->getOperand(0)->getType());
3764 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3765 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3767 case Instruction::Call: {
3768 CallInst *CI = cast<CallInst>(I);
3769 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3770 assert(ID && "Not an intrinsic call!");
3771 Type *RetTy = ToVectorTy(CI->getType(), VF);
3772 SmallVector<Type*, 4> Tys;
3773 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3774 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3775 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3778 // We are scalarizing the instruction. Return the cost of the scalar
3779 // instruction, plus the cost of insert and extract into vector
3780 // elements, times the vector width.
3783 if (!RetTy->isVoidTy() && VF != 1) {
3784 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3786 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3789 // The cost of inserting the results plus extracting each one of the
3791 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3794 // The cost of executing VF copies of the scalar instruction. This opcode
3795 // is unknown. Assume that it is the same as 'mul'.
3796 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3802 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3803 if (Scalar->isVoidTy() || VF == 1)
3805 return VectorType::get(Scalar, VF);
3808 char LoopVectorize::ID = 0;
3809 static const char lv_name[] = "Loop Vectorization";
3810 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3811 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3812 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3813 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3814 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3815 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3818 Pass *createLoopVectorizePass() {
3819 return new LoopVectorize();
3823 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3824 // Check for a store.
3825 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3826 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3828 // Check for a load.
3829 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3830 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;