1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/EquivalenceClasses.h"
51 #include "llvm/ADT/MapVector.h"
52 #include "llvm/ADT/SetVector.h"
53 #include "llvm/ADT/SmallPtrSet.h"
54 #include "llvm/ADT/SmallSet.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/StringExtras.h"
57 #include "llvm/Analysis/AliasAnalysis.h"
58 #include "llvm/Analysis/Dominators.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/Analysis/Verifier.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/LLVMContext.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/IR/Type.h"
78 #include "llvm/IR/Value.h"
79 #include "llvm/Pass.h"
80 #include "llvm/Support/CommandLine.h"
81 #include "llvm/Support/Debug.h"
82 #include "llvm/Support/PatternMatch.h"
83 #include "llvm/Support/raw_ostream.h"
84 #include "llvm/Support/ValueHandle.h"
85 #include "llvm/Target/TargetLibraryInfo.h"
86 #include "llvm/Transforms/Scalar.h"
87 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
88 #include "llvm/Transforms/Utils/Local.h"
93 using namespace llvm::PatternMatch;
95 static cl::opt<unsigned>
96 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
97 cl::desc("Sets the SIMD width. Zero is autoselect."));
99 static cl::opt<unsigned>
100 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
101 cl::desc("Sets the vectorization unroll count. "
102 "Zero is autoselect."));
105 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
106 cl::desc("Enable if-conversion during vectorization."));
108 /// We don't vectorize loops with a known constant trip count below this number.
109 static cl::opt<unsigned>
110 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
112 cl::desc("Don't vectorize loops with a constant "
113 "trip count that is smaller than this "
116 /// We don't unroll loops with a known constant trip count below this number.
117 static const unsigned TinyTripCountUnrollThreshold = 128;
119 /// When performing memory disambiguation checks at runtime do not make more
120 /// than this number of comparisons.
121 static const unsigned RuntimeMemoryCheckThreshold = 8;
123 /// Maximum simd width.
124 static const unsigned MaxVectorWidth = 64;
126 /// Maximum vectorization unroll count.
127 static const unsigned MaxUnrollFactor = 16;
129 /// The cost of a loop that is considered 'small' by the unroller.
130 static const unsigned SmallLoopCost = 20;
134 // Forward declarations.
135 class LoopVectorizationLegality;
136 class LoopVectorizationCostModel;
138 /// InnerLoopVectorizer vectorizes loops which contain only one basic
139 /// block to a specified vectorization factor (VF).
140 /// This class performs the widening of scalars into vectors, or multiple
141 /// scalars. This class also implements the following features:
142 /// * It inserts an epilogue loop for handling loops that don't have iteration
143 /// counts that are known to be a multiple of the vectorization factor.
144 /// * It handles the code generation for reduction variables.
145 /// * Scalarization (implementation using scalars) of un-vectorizable
147 /// InnerLoopVectorizer does not perform any vectorization-legality
148 /// checks, and relies on the caller to check for the different legality
149 /// aspects. The InnerLoopVectorizer relies on the
150 /// LoopVectorizationLegality class to provide information about the induction
151 /// and reduction variables that were found to a given vectorization factor.
152 class InnerLoopVectorizer {
154 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
155 DominatorTree *DT, DataLayout *DL,
156 const TargetLibraryInfo *TLI, unsigned VecWidth,
157 unsigned UnrollFactor)
158 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
159 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
160 OldInduction(0), WidenMap(UnrollFactor) {}
162 // Perform the actual loop widening (vectorization).
163 void vectorize(LoopVectorizationLegality *Legal) {
164 // Create a new empty loop. Unlink the old loop and connect the new one.
165 createEmptyLoop(Legal);
166 // Widen each instruction in the old loop to a new one in the new loop.
167 // Use the Legality module to find the induction and reduction variables.
168 vectorizeLoop(Legal);
169 // Register the new loop and update the analysis passes.
173 virtual ~InnerLoopVectorizer() {}
176 /// A small list of PHINodes.
177 typedef SmallVector<PHINode*, 4> PhiVector;
178 /// When we unroll loops we have multiple vector values for each scalar.
179 /// This data structure holds the unrolled and vectorized values that
180 /// originated from one scalar instruction.
181 typedef SmallVector<Value*, 2> VectorParts;
183 // When we if-convert we need create edge masks. We have to cache values so
184 // that we don't end up with exponential recursion/IR.
185 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
186 VectorParts> EdgeMaskCache;
188 /// Add code that checks at runtime if the accessed arrays overlap.
189 /// Returns the comparator value or NULL if no check is needed.
190 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
192 /// Create an empty loop, based on the loop ranges of the old loop.
193 void createEmptyLoop(LoopVectorizationLegality *Legal);
194 /// Copy and widen the instructions from the old loop.
195 virtual void vectorizeLoop(LoopVectorizationLegality *Legal);
197 /// \brief The Loop exit block may have single value PHI nodes where the
198 /// incoming value is 'Undef'. While vectorizing we only handled real values
199 /// that were defined inside the loop. Here we fix the 'undef case'.
203 /// A helper function that computes the predicate of the block BB, assuming
204 /// that the header block of the loop is set to True. It returns the *entry*
205 /// mask for the block BB.
206 VectorParts createBlockInMask(BasicBlock *BB);
207 /// A helper function that computes the predicate of the edge between SRC
209 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
211 /// A helper function to vectorize a single BB within the innermost loop.
212 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
215 /// Vectorize a single PHINode in a block. This method handles the induction
216 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
217 /// arbitrary length vectors.
218 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
219 LoopVectorizationLegality *Legal,
220 unsigned UF, unsigned VF, PhiVector *PV);
222 /// Insert the new loop to the loop hierarchy and pass manager
223 /// and update the analysis passes.
224 void updateAnalysis();
226 /// This instruction is un-vectorizable. Implement it as a sequence
228 virtual void scalarizeInstruction(Instruction *Instr);
230 /// Vectorize Load and Store instructions,
231 virtual void vectorizeMemoryInstruction(Instruction *Instr,
232 LoopVectorizationLegality *Legal);
234 /// Create a broadcast instruction. This method generates a broadcast
235 /// instruction (shuffle) for loop invariant values and for the induction
236 /// value. If this is the induction variable then we extend it to N, N+1, ...
237 /// this is needed because each iteration in the loop corresponds to a SIMD
239 virtual Value *getBroadcastInstrs(Value *V);
241 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
242 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
243 /// The sequence starts at StartIndex.
244 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
246 /// When we go over instructions in the basic block we rely on previous
247 /// values within the current basic block or on loop invariant values.
248 /// When we widen (vectorize) values we place them in the map. If the values
249 /// are not within the map, they have to be loop invariant, so we simply
250 /// broadcast them into a vector.
251 VectorParts &getVectorValue(Value *V);
253 /// Generate a shuffle sequence that will reverse the vector Vec.
254 virtual Value *reverseVector(Value *Vec);
256 /// This is a helper class that holds the vectorizer state. It maps scalar
257 /// instructions to vector instructions. When the code is 'unrolled' then
258 /// then a single scalar value is mapped to multiple vector parts. The parts
259 /// are stored in the VectorPart type.
261 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
263 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
265 /// \return True if 'Key' is saved in the Value Map.
266 bool has(Value *Key) const { return MapStorage.count(Key); }
268 /// Initializes a new entry in the map. Sets all of the vector parts to the
269 /// save value in 'Val'.
270 /// \return A reference to a vector with splat values.
271 VectorParts &splat(Value *Key, Value *Val) {
272 VectorParts &Entry = MapStorage[Key];
273 Entry.assign(UF, Val);
277 ///\return A reference to the value that is stored at 'Key'.
278 VectorParts &get(Value *Key) {
279 VectorParts &Entry = MapStorage[Key];
282 assert(Entry.size() == UF);
287 /// The unroll factor. Each entry in the map stores this number of vector
291 /// Map storage. We use std::map and not DenseMap because insertions to a
292 /// dense map invalidates its iterators.
293 std::map<Value *, VectorParts> MapStorage;
296 /// The original loop.
298 /// Scev analysis to use.
306 /// Target Library Info.
307 const TargetLibraryInfo *TLI;
309 /// The vectorization SIMD factor to use. Each vector will have this many
314 /// The vectorization unroll factor to use. Each scalar is vectorized to this
315 /// many different vector instructions.
318 /// The builder that we use
321 // --- Vectorization state ---
323 /// The vector-loop preheader.
324 BasicBlock *LoopVectorPreHeader;
325 /// The scalar-loop preheader.
326 BasicBlock *LoopScalarPreHeader;
327 /// Middle Block between the vector and the scalar.
328 BasicBlock *LoopMiddleBlock;
329 ///The ExitBlock of the scalar loop.
330 BasicBlock *LoopExitBlock;
331 ///The vector loop body.
332 BasicBlock *LoopVectorBody;
333 ///The scalar loop body.
334 BasicBlock *LoopScalarBody;
335 /// A list of all bypass blocks. The first block is the entry of the loop.
336 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
338 /// The new Induction variable which was added to the new block.
340 /// The induction variable of the old basic block.
341 PHINode *OldInduction;
342 /// Holds the extended (to the widest induction type) start index.
344 /// Maps scalars to widened vectors.
346 EdgeMaskCache MaskCache;
349 class InnerLoopUnroller : public InnerLoopVectorizer {
351 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
352 DominatorTree *DT, DataLayout *DL,
353 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
354 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
357 virtual void vectorizeLoop(LoopVectorizationLegality *Legal);
358 virtual void scalarizeInstruction(Instruction *Instr);
359 virtual void vectorizeMemoryInstruction(Instruction *Instr,
360 LoopVectorizationLegality *Legal);
361 virtual Value *getBroadcastInstrs(Value *V);
362 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
363 virtual Value *reverseVector(Value *Vec);
366 /// \brief Look for a meaningful debug location on the instruction or it's
368 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
373 if (I->getDebugLoc() != Empty)
376 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
377 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
378 if (OpInst->getDebugLoc() != Empty)
385 /// \brief Set the debug location in the builder using the debug location in the
387 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
388 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
389 B.SetCurrentDebugLocation(Inst->getDebugLoc());
391 B.SetCurrentDebugLocation(DebugLoc());
394 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
395 /// to what vectorization factor.
396 /// This class does not look at the profitability of vectorization, only the
397 /// legality. This class has two main kinds of checks:
398 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
399 /// will change the order of memory accesses in a way that will change the
400 /// correctness of the program.
401 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
402 /// checks for a number of different conditions, such as the availability of a
403 /// single induction variable, that all types are supported and vectorize-able,
404 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
405 /// This class is also used by InnerLoopVectorizer for identifying
406 /// induction variable and the different reduction variables.
407 class LoopVectorizationLegality {
409 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
410 DominatorTree *DT, TargetLibraryInfo *TLI)
411 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
412 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
413 MaxSafeDepDistBytes(-1U) {}
415 /// This enum represents the kinds of reductions that we support.
417 RK_NoReduction, ///< Not a reduction.
418 RK_IntegerAdd, ///< Sum of integers.
419 RK_IntegerMult, ///< Product of integers.
420 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
421 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
422 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
423 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
424 RK_FloatAdd, ///< Sum of floats.
425 RK_FloatMult, ///< Product of floats.
426 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
429 /// This enum represents the kinds of inductions that we support.
431 IK_NoInduction, ///< Not an induction variable.
432 IK_IntInduction, ///< Integer induction variable. Step = 1.
433 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
434 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
435 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
438 // This enum represents the kind of minmax reduction.
439 enum MinMaxReductionKind {
449 /// This POD struct holds information about reduction variables.
450 struct ReductionDescriptor {
451 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
452 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
454 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
455 MinMaxReductionKind MK)
456 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
458 // The starting value of the reduction.
459 // It does not have to be zero!
460 TrackingVH<Value> StartValue;
461 // The instruction who's value is used outside the loop.
462 Instruction *LoopExitInstr;
463 // The kind of the reduction.
465 // If this a min/max reduction the kind of reduction.
466 MinMaxReductionKind MinMaxKind;
469 /// This POD struct holds information about a potential reduction operation.
470 struct ReductionInstDesc {
471 ReductionInstDesc(bool IsRedux, Instruction *I) :
472 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
474 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
475 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
477 // Is this instruction a reduction candidate.
479 // The last instruction in a min/max pattern (select of the select(icmp())
480 // pattern), or the current reduction instruction otherwise.
481 Instruction *PatternLastInst;
482 // If this is a min/max pattern the comparison predicate.
483 MinMaxReductionKind MinMaxKind;
486 // This POD struct holds information about the memory runtime legality
487 // check that a group of pointers do not overlap.
488 struct RuntimePointerCheck {
489 RuntimePointerCheck() : Need(false) {}
491 /// Reset the state of the pointer runtime information.
499 /// Insert a pointer and calculate the start and end SCEVs.
500 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
503 /// This flag indicates if we need to add the runtime check.
505 /// Holds the pointers that we need to check.
506 SmallVector<TrackingVH<Value>, 2> Pointers;
507 /// Holds the pointer value at the beginning of the loop.
508 SmallVector<const SCEV*, 2> Starts;
509 /// Holds the pointer value at the end of the loop.
510 SmallVector<const SCEV*, 2> Ends;
511 /// Holds the information if this pointer is used for writing to memory.
512 SmallVector<bool, 2> IsWritePtr;
513 /// Holds the id of the set of pointers that could be dependent because of a
514 /// shared underlying object.
515 SmallVector<unsigned, 2> DependencySetId;
518 /// A POD for saving information about induction variables.
519 struct InductionInfo {
520 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
521 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
523 TrackingVH<Value> StartValue;
528 /// ReductionList contains the reduction descriptors for all
529 /// of the reductions that were found in the loop.
530 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
532 /// InductionList saves induction variables and maps them to the
533 /// induction descriptor.
534 typedef MapVector<PHINode*, InductionInfo> InductionList;
536 /// Returns true if it is legal to vectorize this loop.
537 /// This does not mean that it is profitable to vectorize this
538 /// loop, only that it is legal to do so.
541 /// Returns the Induction variable.
542 PHINode *getInduction() { return Induction; }
544 /// Returns the reduction variables found in the loop.
545 ReductionList *getReductionVars() { return &Reductions; }
547 /// Returns the induction variables found in the loop.
548 InductionList *getInductionVars() { return &Inductions; }
550 /// Returns the widest induction type.
551 Type *getWidestInductionType() { return WidestIndTy; }
553 /// Returns True if V is an induction variable in this loop.
554 bool isInductionVariable(const Value *V);
556 /// Return true if the block BB needs to be predicated in order for the loop
557 /// to be vectorized.
558 bool blockNeedsPredication(BasicBlock *BB);
560 /// Check if this pointer is consecutive when vectorizing. This happens
561 /// when the last index of the GEP is the induction variable, or that the
562 /// pointer itself is an induction variable.
563 /// This check allows us to vectorize A[idx] into a wide load/store.
565 /// 0 - Stride is unknown or non consecutive.
566 /// 1 - Address is consecutive.
567 /// -1 - Address is consecutive, and decreasing.
568 int isConsecutivePtr(Value *Ptr);
570 /// Returns true if the value V is uniform within the loop.
571 bool isUniform(Value *V);
573 /// Returns true if this instruction will remain scalar after vectorization.
574 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
576 /// Returns the information that we collected about runtime memory check.
577 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
579 /// This function returns the identity element (or neutral element) for
581 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
583 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
586 /// Check if a single basic block loop is vectorizable.
587 /// At this point we know that this is a loop with a constant trip count
588 /// and we only need to check individual instructions.
589 bool canVectorizeInstrs();
591 /// When we vectorize loops we may change the order in which
592 /// we read and write from memory. This method checks if it is
593 /// legal to vectorize the code, considering only memory constrains.
594 /// Returns true if the loop is vectorizable
595 bool canVectorizeMemory();
597 /// Return true if we can vectorize this loop using the IF-conversion
599 bool canVectorizeWithIfConvert();
601 /// Collect the variables that need to stay uniform after vectorization.
602 void collectLoopUniforms();
604 /// Return true if all of the instructions in the block can be speculatively
605 /// executed. \p SafePtrs is a list of addresses that are known to be legal
606 /// and we know that we can read from them without segfault.
607 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
609 /// Returns True, if 'Phi' is the kind of reduction variable for type
610 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
611 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
612 /// Returns a struct describing if the instruction 'I' can be a reduction
613 /// variable of type 'Kind'. If the reduction is a min/max pattern of
614 /// select(icmp()) this function advances the instruction pointer 'I' from the
615 /// compare instruction to the select instruction and stores this pointer in
616 /// 'PatternLastInst' member of the returned struct.
617 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
618 ReductionInstDesc &Desc);
619 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
620 /// pattern corresponding to a min(X, Y) or max(X, Y).
621 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
622 ReductionInstDesc &Prev);
623 /// Returns the induction kind of Phi. This function may return NoInduction
624 /// if the PHI is not an induction variable.
625 InductionKind isInductionVariable(PHINode *Phi);
627 /// The loop that we evaluate.
631 /// DataLayout analysis.
635 /// Target Library Info.
636 TargetLibraryInfo *TLI;
638 // --- vectorization state --- //
640 /// Holds the integer induction variable. This is the counter of the
643 /// Holds the reduction variables.
644 ReductionList Reductions;
645 /// Holds all of the induction variables that we found in the loop.
646 /// Notice that inductions don't need to start at zero and that induction
647 /// variables can be pointers.
648 InductionList Inductions;
649 /// Holds the widest induction type encountered.
652 /// Allowed outside users. This holds the reduction
653 /// vars which can be accessed from outside the loop.
654 SmallPtrSet<Value*, 4> AllowedExit;
655 /// This set holds the variables which are known to be uniform after
657 SmallPtrSet<Instruction*, 4> Uniforms;
658 /// We need to check that all of the pointers in this list are disjoint
660 RuntimePointerCheck PtrRtCheck;
661 /// Can we assume the absence of NaNs.
662 bool HasFunNoNaNAttr;
664 unsigned MaxSafeDepDistBytes;
667 /// LoopVectorizationCostModel - estimates the expected speedups due to
669 /// In many cases vectorization is not profitable. This can happen because of
670 /// a number of reasons. In this class we mainly attempt to predict the
671 /// expected speedup/slowdowns due to the supported instruction set. We use the
672 /// TargetTransformInfo to query the different backends for the cost of
673 /// different operations.
674 class LoopVectorizationCostModel {
676 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
677 LoopVectorizationLegality *Legal,
678 const TargetTransformInfo &TTI,
679 DataLayout *DL, const TargetLibraryInfo *TLI)
680 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
682 /// Information about vectorization costs
683 struct VectorizationFactor {
684 unsigned Width; // Vector width with best cost
685 unsigned Cost; // Cost of the loop with that width
687 /// \return The most profitable vectorization factor and the cost of that VF.
688 /// This method checks every power of two up to VF. If UserVF is not ZERO
689 /// then this vectorization factor will be selected if vectorization is
691 VectorizationFactor selectVectorizationFactor(bool OptForSize,
694 /// \return The size (in bits) of the widest type in the code that
695 /// needs to be vectorized. We ignore values that remain scalar such as
696 /// 64 bit loop indices.
697 unsigned getWidestType();
699 /// \return The most profitable unroll factor.
700 /// If UserUF is non-zero then this method finds the best unroll-factor
701 /// based on register pressure and other parameters.
702 /// VF and LoopCost are the selected vectorization factor and the cost of the
704 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
707 /// \brief A struct that represents some properties of the register usage
709 struct RegisterUsage {
710 /// Holds the number of loop invariant values that are used in the loop.
711 unsigned LoopInvariantRegs;
712 /// Holds the maximum number of concurrent live intervals in the loop.
713 unsigned MaxLocalUsers;
714 /// Holds the number of instructions in the loop.
715 unsigned NumInstructions;
718 /// \return information about the register usage of the loop.
719 RegisterUsage calculateRegisterUsage();
722 /// Returns the expected execution cost. The unit of the cost does
723 /// not matter because we use the 'cost' units to compare different
724 /// vector widths. The cost that is returned is *not* normalized by
725 /// the factor width.
726 unsigned expectedCost(unsigned VF);
728 /// Returns the execution time cost of an instruction for a given vector
729 /// width. Vector width of one means scalar.
730 unsigned getInstructionCost(Instruction *I, unsigned VF);
732 /// A helper function for converting Scalar types to vector types.
733 /// If the incoming type is void, we return void. If the VF is 1, we return
735 static Type* ToVectorTy(Type *Scalar, unsigned VF);
737 /// Returns whether the instruction is a load or store and will be a emitted
738 /// as a vector operation.
739 bool isConsecutiveLoadOrStore(Instruction *I);
741 /// The loop that we evaluate.
745 /// Loop Info analysis.
747 /// Vectorization legality.
748 LoopVectorizationLegality *Legal;
749 /// Vector target information.
750 const TargetTransformInfo &TTI;
751 /// Target data layout information.
753 /// Target Library Info.
754 const TargetLibraryInfo *TLI;
757 /// Utility class for getting and setting loop vectorizer hints in the form
758 /// of loop metadata.
759 struct LoopVectorizeHints {
760 /// Vectorization width.
762 /// Vectorization unroll factor.
765 LoopVectorizeHints(const Loop *L)
766 : Width(VectorizationFactor)
767 , Unroll(VectorizationUnroll)
768 , LoopID(L->getLoopID()) {
770 // The command line options override any loop metadata except for when
771 // width == 1 which is used to indicate the loop is already vectorized.
772 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
773 Width = VectorizationFactor;
774 if (VectorizationUnroll.getNumOccurrences() > 0)
775 Unroll = VectorizationUnroll;
778 /// Return the loop vectorizer metadata prefix.
779 static StringRef Prefix() { return "llvm.vectorizer."; }
781 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
782 SmallVector<Value*, 2> Vals;
783 Vals.push_back(MDString::get(Context, Name));
784 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
785 return MDNode::get(Context, Vals);
788 /// Mark the loop L as already vectorized by setting the width to 1.
789 void setAlreadyVectorized(Loop *L) {
790 LLVMContext &Context = L->getHeader()->getContext();
794 // Create a new loop id with one more operand for the already_vectorized
795 // hint. If the loop already has a loop id then copy the existing operands.
796 SmallVector<Value*, 4> Vals(1);
798 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
799 Vals.push_back(LoopID->getOperand(i));
801 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
803 MDNode *NewLoopID = MDNode::get(Context, Vals);
804 // Set operand 0 to refer to the loop id itself.
805 NewLoopID->replaceOperandWith(0, NewLoopID);
807 L->setLoopID(NewLoopID);
809 LoopID->replaceAllUsesWith(NewLoopID);
817 /// Find hints specified in the loop metadata.
818 void getHints(const Loop *L) {
822 // First operand should refer to the loop id itself.
823 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
824 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
826 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
827 const MDString *S = 0;
828 SmallVector<Value*, 4> Args;
830 // The expected hint is either a MDString or a MDNode with the first
831 // operand a MDString.
832 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
833 if (!MD || MD->getNumOperands() == 0)
835 S = dyn_cast<MDString>(MD->getOperand(0));
836 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
837 Args.push_back(MD->getOperand(i));
839 S = dyn_cast<MDString>(LoopID->getOperand(i));
840 assert(Args.size() == 0 && "too many arguments for MDString");
846 // Check if the hint starts with the vectorizer prefix.
847 StringRef Hint = S->getString();
848 if (!Hint.startswith(Prefix()))
850 // Remove the prefix.
851 Hint = Hint.substr(Prefix().size(), StringRef::npos);
853 if (Args.size() == 1)
854 getHint(Hint, Args[0]);
858 // Check string hint with one operand.
859 void getHint(StringRef Hint, Value *Arg) {
860 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
862 unsigned Val = C->getZExtValue();
864 if (Hint == "width") {
865 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
866 "Invalid width metadata");
868 } else if (Hint == "unroll") {
869 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
870 "Invalid unroll metadata");
873 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
877 /// The LoopVectorize Pass.
878 struct LoopVectorize : public LoopPass {
879 /// Pass identification, replacement for typeid
882 explicit LoopVectorize() : LoopPass(ID) {
883 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
889 TargetTransformInfo *TTI;
891 TargetLibraryInfo *TLI;
893 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
894 // We only vectorize innermost loops.
898 SE = &getAnalysis<ScalarEvolution>();
899 DL = getAnalysisIfAvailable<DataLayout>();
900 LI = &getAnalysis<LoopInfo>();
901 TTI = &getAnalysis<TargetTransformInfo>();
902 DT = &getAnalysis<DominatorTree>();
903 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
906 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
910 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
911 L->getHeader()->getParent()->getName() << "\"\n");
913 LoopVectorizeHints Hints(L);
915 if (Hints.Width == 1 && Hints.Unroll == 1) {
916 DEBUG(dbgs() << "LV: Not vectorizing.\n");
920 // Check if it is legal to vectorize the loop.
921 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
922 if (!LVL.canVectorize()) {
923 DEBUG(dbgs() << "LV: Not vectorizing.\n");
927 // Use the cost model.
928 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
930 // Check the function attributes to find out if this function should be
931 // optimized for size.
932 Function *F = L->getHeader()->getParent();
933 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
934 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
935 unsigned FnIndex = AttributeSet::FunctionIndex;
936 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
937 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
940 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
941 "attribute is used.\n");
945 // Select the optimal vectorization factor.
946 LoopVectorizationCostModel::VectorizationFactor VF;
947 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
948 // Select the unroll factor.
949 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
953 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
956 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
957 F->getParent()->getModuleIdentifier()<<"\n");
958 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
963 // We decided not to vectorize, but we may want to unroll.
964 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
965 Unroller.vectorize(&LVL);
967 // If we decided that it is *legal* to vectorize the loop then do it.
968 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
972 // Mark the loop as already vectorized to avoid vectorizing again.
973 Hints.setAlreadyVectorized(L);
975 DEBUG(verifyFunction(*L->getHeader()->getParent()));
979 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
980 LoopPass::getAnalysisUsage(AU);
981 AU.addRequiredID(LoopSimplifyID);
982 AU.addRequiredID(LCSSAID);
983 AU.addRequired<DominatorTree>();
984 AU.addRequired<LoopInfo>();
985 AU.addRequired<ScalarEvolution>();
986 AU.addRequired<TargetTransformInfo>();
987 AU.addPreserved<LoopInfo>();
988 AU.addPreserved<DominatorTree>();
993 } // end anonymous namespace
995 //===----------------------------------------------------------------------===//
996 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
997 // LoopVectorizationCostModel.
998 //===----------------------------------------------------------------------===//
1001 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1002 Loop *Lp, Value *Ptr,
1004 unsigned DepSetId) {
1005 const SCEV *Sc = SE->getSCEV(Ptr);
1006 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1007 assert(AR && "Invalid addrec expression");
1008 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1009 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1010 Pointers.push_back(Ptr);
1011 Starts.push_back(AR->getStart());
1012 Ends.push_back(ScEnd);
1013 IsWritePtr.push_back(WritePtr);
1014 DependencySetId.push_back(DepSetId);
1017 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1018 // Save the current insertion location.
1019 Instruction *Loc = Builder.GetInsertPoint();
1021 // We need to place the broadcast of invariant variables outside the loop.
1022 Instruction *Instr = dyn_cast<Instruction>(V);
1023 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1024 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1026 // Place the code for broadcasting invariant variables in the new preheader.
1028 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1030 // Broadcast the scalar into all locations in the vector.
1031 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1033 // Restore the builder insertion point.
1035 Builder.SetInsertPoint(Loc);
1040 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1042 assert(Val->getType()->isVectorTy() && "Must be a vector");
1043 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1044 "Elem must be an integer");
1045 // Create the types.
1046 Type *ITy = Val->getType()->getScalarType();
1047 VectorType *Ty = cast<VectorType>(Val->getType());
1048 int VLen = Ty->getNumElements();
1049 SmallVector<Constant*, 8> Indices;
1051 // Create a vector of consecutive numbers from zero to VF.
1052 for (int i = 0; i < VLen; ++i) {
1053 int64_t Idx = Negate ? (-i) : i;
1054 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1057 // Add the consecutive indices to the vector value.
1058 Constant *Cv = ConstantVector::get(Indices);
1059 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1060 return Builder.CreateAdd(Val, Cv, "induction");
1063 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1064 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1065 // Make sure that the pointer does not point to structs.
1066 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1069 // If this value is a pointer induction variable we know it is consecutive.
1070 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1071 if (Phi && Inductions.count(Phi)) {
1072 InductionInfo II = Inductions[Phi];
1073 if (IK_PtrInduction == II.IK)
1075 else if (IK_ReversePtrInduction == II.IK)
1079 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1083 unsigned NumOperands = Gep->getNumOperands();
1084 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1086 Value *GpPtr = Gep->getPointerOperand();
1087 // If this GEP value is a consecutive pointer induction variable and all of
1088 // the indices are constant then we know it is consecutive. We can
1089 Phi = dyn_cast<PHINode>(GpPtr);
1090 if (Phi && Inductions.count(Phi)) {
1092 // Make sure that the pointer does not point to structs.
1093 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1094 if (GepPtrType->getElementType()->isAggregateType())
1097 // Make sure that all of the index operands are loop invariant.
1098 for (unsigned i = 1; i < NumOperands; ++i)
1099 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1102 InductionInfo II = Inductions[Phi];
1103 if (IK_PtrInduction == II.IK)
1105 else if (IK_ReversePtrInduction == II.IK)
1109 // Check that all of the gep indices are uniform except for the last.
1110 for (unsigned i = 0; i < NumOperands - 1; ++i)
1111 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1114 // We can emit wide load/stores only if the last index is the induction
1116 const SCEV *Last = SE->getSCEV(LastIndex);
1117 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1118 const SCEV *Step = AR->getStepRecurrence(*SE);
1120 // The memory is consecutive because the last index is consecutive
1121 // and all other indices are loop invariant.
1124 if (Step->isAllOnesValue())
1131 bool LoopVectorizationLegality::isUniform(Value *V) {
1132 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1135 InnerLoopVectorizer::VectorParts&
1136 InnerLoopVectorizer::getVectorValue(Value *V) {
1137 assert(V != Induction && "The new induction variable should not be used.");
1138 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1140 // If we have this scalar in the map, return it.
1141 if (WidenMap.has(V))
1142 return WidenMap.get(V);
1144 // If this scalar is unknown, assume that it is a constant or that it is
1145 // loop invariant. Broadcast V and save the value for future uses.
1146 Value *B = getBroadcastInstrs(V);
1147 return WidenMap.splat(V, B);
1150 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1151 assert(Vec->getType()->isVectorTy() && "Invalid type");
1152 SmallVector<Constant*, 8> ShuffleMask;
1153 for (unsigned i = 0; i < VF; ++i)
1154 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1156 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1157 ConstantVector::get(ShuffleMask),
1162 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1163 LoopVectorizationLegality *Legal) {
1164 // Attempt to issue a wide load.
1165 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1166 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1168 assert((LI || SI) && "Invalid Load/Store instruction");
1170 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1171 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1172 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1173 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1174 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1175 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1176 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1178 if (ScalarAllocatedSize != VectorElementSize)
1179 return scalarizeInstruction(Instr);
1181 // If the pointer is loop invariant or if it is non consecutive,
1182 // scalarize the load.
1183 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1184 bool Reverse = ConsecutiveStride < 0;
1185 bool UniformLoad = LI && Legal->isUniform(Ptr);
1186 if (!ConsecutiveStride || UniformLoad)
1187 return scalarizeInstruction(Instr);
1189 Constant *Zero = Builder.getInt32(0);
1190 VectorParts &Entry = WidenMap.get(Instr);
1192 // Handle consecutive loads/stores.
1193 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1194 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1195 setDebugLocFromInst(Builder, Gep);
1196 Value *PtrOperand = Gep->getPointerOperand();
1197 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1198 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1200 // Create the new GEP with the new induction variable.
1201 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1202 Gep2->setOperand(0, FirstBasePtr);
1203 Gep2->setName("gep.indvar.base");
1204 Ptr = Builder.Insert(Gep2);
1206 setDebugLocFromInst(Builder, Gep);
1207 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1208 OrigLoop) && "Base ptr must be invariant");
1210 // The last index does not have to be the induction. It can be
1211 // consecutive and be a function of the index. For example A[I+1];
1212 unsigned NumOperands = Gep->getNumOperands();
1213 unsigned LastOperand = NumOperands - 1;
1214 // Create the new GEP with the new induction variable.
1215 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1217 for (unsigned i = 0; i < NumOperands; ++i) {
1218 Value *GepOperand = Gep->getOperand(i);
1219 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1221 // Update last index or loop invariant instruction anchored in loop.
1222 if (i == LastOperand ||
1223 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1224 assert((i == LastOperand ||
1225 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1226 "Must be last index or loop invariant");
1228 VectorParts &GEPParts = getVectorValue(GepOperand);
1229 Value *Index = GEPParts[0];
1230 Index = Builder.CreateExtractElement(Index, Zero);
1231 Gep2->setOperand(i, Index);
1232 Gep2->setName("gep.indvar.idx");
1235 Ptr = Builder.Insert(Gep2);
1237 // Use the induction element ptr.
1238 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1239 setDebugLocFromInst(Builder, Ptr);
1240 VectorParts &PtrVal = getVectorValue(Ptr);
1241 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1246 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1247 "We do not allow storing to uniform addresses");
1248 setDebugLocFromInst(Builder, SI);
1249 // We don't want to update the value in the map as it might be used in
1250 // another expression. So don't use a reference type for "StoredVal".
1251 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1253 for (unsigned Part = 0; Part < UF; ++Part) {
1254 // Calculate the pointer for the specific unroll-part.
1255 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1258 // If we store to reverse consecutive memory locations then we need
1259 // to reverse the order of elements in the stored value.
1260 StoredVal[Part] = reverseVector(StoredVal[Part]);
1261 // If the address is consecutive but reversed, then the
1262 // wide store needs to start at the last vector element.
1263 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1264 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1267 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1268 DataTy->getPointerTo(AddressSpace));
1269 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1275 assert(LI && "Must have a load instruction");
1276 setDebugLocFromInst(Builder, LI);
1277 for (unsigned Part = 0; Part < UF; ++Part) {
1278 // Calculate the pointer for the specific unroll-part.
1279 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1282 // If the address is consecutive but reversed, then the
1283 // wide store needs to start at the last vector element.
1284 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1285 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1288 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1289 DataTy->getPointerTo(AddressSpace));
1290 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1291 cast<LoadInst>(LI)->setAlignment(Alignment);
1292 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1296 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1297 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1298 // Holds vector parameters or scalars, in case of uniform vals.
1299 SmallVector<VectorParts, 4> Params;
1301 setDebugLocFromInst(Builder, Instr);
1303 // Find all of the vectorized parameters.
1304 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1305 Value *SrcOp = Instr->getOperand(op);
1307 // If we are accessing the old induction variable, use the new one.
1308 if (SrcOp == OldInduction) {
1309 Params.push_back(getVectorValue(SrcOp));
1313 // Try using previously calculated values.
1314 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1316 // If the src is an instruction that appeared earlier in the basic block
1317 // then it should already be vectorized.
1318 if (SrcInst && OrigLoop->contains(SrcInst)) {
1319 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1320 // The parameter is a vector value from earlier.
1321 Params.push_back(WidenMap.get(SrcInst));
1323 // The parameter is a scalar from outside the loop. Maybe even a constant.
1324 VectorParts Scalars;
1325 Scalars.append(UF, SrcOp);
1326 Params.push_back(Scalars);
1330 assert(Params.size() == Instr->getNumOperands() &&
1331 "Invalid number of operands");
1333 // Does this instruction return a value ?
1334 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1336 Value *UndefVec = IsVoidRetTy ? 0 :
1337 UndefValue::get(VectorType::get(Instr->getType(), VF));
1338 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1339 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1341 // For each vector unroll 'part':
1342 for (unsigned Part = 0; Part < UF; ++Part) {
1343 // For each scalar that we create:
1344 for (unsigned Width = 0; Width < VF; ++Width) {
1345 Instruction *Cloned = Instr->clone();
1347 Cloned->setName(Instr->getName() + ".cloned");
1348 // Replace the operands of the cloned instrucions with extracted scalars.
1349 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1350 Value *Op = Params[op][Part];
1351 // Param is a vector. Need to extract the right lane.
1352 if (Op->getType()->isVectorTy())
1353 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1354 Cloned->setOperand(op, Op);
1357 // Place the cloned scalar in the new loop.
1358 Builder.Insert(Cloned);
1360 // If the original scalar returns a value we need to place it in a vector
1361 // so that future users will be able to use it.
1363 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1364 Builder.getInt32(Width));
1370 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1372 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1373 Legal->getRuntimePointerCheck();
1375 if (!PtrRtCheck->Need)
1378 unsigned NumPointers = PtrRtCheck->Pointers.size();
1379 SmallVector<TrackingVH<Value> , 2> Starts;
1380 SmallVector<TrackingVH<Value> , 2> Ends;
1382 SCEVExpander Exp(*SE, "induction");
1384 // Use this type for pointer arithmetic.
1385 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1387 for (unsigned i = 0; i < NumPointers; ++i) {
1388 Value *Ptr = PtrRtCheck->Pointers[i];
1389 const SCEV *Sc = SE->getSCEV(Ptr);
1391 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1392 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1394 Starts.push_back(Ptr);
1395 Ends.push_back(Ptr);
1397 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1399 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1400 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1401 Starts.push_back(Start);
1402 Ends.push_back(End);
1406 IRBuilder<> ChkBuilder(Loc);
1407 // Our instructions might fold to a constant.
1408 Value *MemoryRuntimeCheck = 0;
1409 for (unsigned i = 0; i < NumPointers; ++i) {
1410 for (unsigned j = i+1; j < NumPointers; ++j) {
1411 // No need to check if two readonly pointers intersect.
1412 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1415 // Only need to check pointers between two different dependency sets.
1416 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1419 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1420 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1421 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1422 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1424 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1425 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1426 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1427 if (MemoryRuntimeCheck)
1428 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1430 MemoryRuntimeCheck = IsConflict;
1434 // We have to do this trickery because the IRBuilder might fold the check to a
1435 // constant expression in which case there is no Instruction anchored in a
1437 LLVMContext &Ctx = Loc->getContext();
1438 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1439 ConstantInt::getTrue(Ctx));
1440 ChkBuilder.Insert(Check, "memcheck.conflict");
1445 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1447 In this function we generate a new loop. The new loop will contain
1448 the vectorized instructions while the old loop will continue to run the
1451 [ ] <-- vector loop bypass (may consist of multiple blocks).
1454 | [ ] <-- vector pre header.
1458 | [ ]_| <-- vector loop.
1461 >[ ] <--- middle-block.
1464 | [ ] <--- new preheader.
1468 | [ ]_| <-- old scalar loop to handle remainder.
1471 >[ ] <-- exit block.
1475 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1476 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1477 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1478 assert(ExitBlock && "Must have an exit block");
1480 // Some loops have a single integer induction variable, while other loops
1481 // don't. One example is c++ iterators that often have multiple pointer
1482 // induction variables. In the code below we also support a case where we
1483 // don't have a single induction variable.
1484 OldInduction = Legal->getInduction();
1485 Type *IdxTy = Legal->getWidestInductionType();
1487 // Find the loop boundaries.
1488 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1489 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1491 // Get the total trip count from the count by adding 1.
1492 ExitCount = SE->getAddExpr(ExitCount,
1493 SE->getConstant(ExitCount->getType(), 1));
1495 // Expand the trip count and place the new instructions in the preheader.
1496 // Notice that the pre-header does not change, only the loop body.
1497 SCEVExpander Exp(*SE, "induction");
1499 // Count holds the overall loop count (N).
1500 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1501 BypassBlock->getTerminator());
1503 // The loop index does not have to start at Zero. Find the original start
1504 // value from the induction PHI node. If we don't have an induction variable
1505 // then we know that it starts at zero.
1506 Builder.SetInsertPoint(BypassBlock->getTerminator());
1507 Value *StartIdx = ExtendedIdx = OldInduction ?
1508 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1510 ConstantInt::get(IdxTy, 0);
1512 assert(BypassBlock && "Invalid loop structure");
1513 LoopBypassBlocks.push_back(BypassBlock);
1515 // Split the single block loop into the two loop structure described above.
1516 BasicBlock *VectorPH =
1517 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1518 BasicBlock *VecBody =
1519 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1520 BasicBlock *MiddleBlock =
1521 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1522 BasicBlock *ScalarPH =
1523 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1525 // Create and register the new vector loop.
1526 Loop* Lp = new Loop();
1527 Loop *ParentLoop = OrigLoop->getParentLoop();
1529 // Insert the new loop into the loop nest and register the new basic blocks
1530 // before calling any utilities such as SCEV that require valid LoopInfo.
1532 ParentLoop->addChildLoop(Lp);
1533 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1534 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1535 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1537 LI->addTopLevelLoop(Lp);
1539 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1541 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1543 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1545 // Generate the induction variable.
1546 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1547 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1548 // The loop step is equal to the vectorization factor (num of SIMD elements)
1549 // times the unroll factor (num of SIMD instructions).
1550 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1552 // This is the IR builder that we use to add all of the logic for bypassing
1553 // the new vector loop.
1554 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1555 setDebugLocFromInst(BypassBuilder,
1556 getDebugLocFromInstOrOperands(OldInduction));
1558 // We may need to extend the index in case there is a type mismatch.
1559 // We know that the count starts at zero and does not overflow.
1560 if (Count->getType() != IdxTy) {
1561 // The exit count can be of pointer type. Convert it to the correct
1563 if (ExitCount->getType()->isPointerTy())
1564 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1566 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1569 // Add the start index to the loop count to get the new end index.
1570 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1572 // Now we need to generate the expression for N - (N % VF), which is
1573 // the part that the vectorized body will execute.
1574 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1575 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1576 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1577 "end.idx.rnd.down");
1579 // Now, compare the new count to zero. If it is zero skip the vector loop and
1580 // jump to the scalar loop.
1581 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1584 BasicBlock *LastBypassBlock = BypassBlock;
1586 // Generate the code that checks in runtime if arrays overlap. We put the
1587 // checks into a separate block to make the more common case of few elements
1589 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1590 BypassBlock->getTerminator());
1591 if (MemRuntimeCheck) {
1592 // Create a new block containing the memory check.
1593 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1596 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1597 LoopBypassBlocks.push_back(CheckBlock);
1599 // Replace the branch into the memory check block with a conditional branch
1600 // for the "few elements case".
1601 Instruction *OldTerm = BypassBlock->getTerminator();
1602 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1603 OldTerm->eraseFromParent();
1605 Cmp = MemRuntimeCheck;
1606 LastBypassBlock = CheckBlock;
1609 LastBypassBlock->getTerminator()->eraseFromParent();
1610 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1613 // We are going to resume the execution of the scalar loop.
1614 // Go over all of the induction variables that we found and fix the
1615 // PHIs that are left in the scalar version of the loop.
1616 // The starting values of PHI nodes depend on the counter of the last
1617 // iteration in the vectorized loop.
1618 // If we come from a bypass edge then we need to start from the original
1621 // This variable saves the new starting index for the scalar loop.
1622 PHINode *ResumeIndex = 0;
1623 LoopVectorizationLegality::InductionList::iterator I, E;
1624 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1625 // Set builder to point to last bypass block.
1626 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1627 for (I = List->begin(), E = List->end(); I != E; ++I) {
1628 PHINode *OrigPhi = I->first;
1629 LoopVectorizationLegality::InductionInfo II = I->second;
1631 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1632 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1633 MiddleBlock->getTerminator());
1634 // We might have extended the type of the induction variable but we need a
1635 // truncated version for the scalar loop.
1636 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1637 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1638 MiddleBlock->getTerminator()) : 0;
1640 Value *EndValue = 0;
1642 case LoopVectorizationLegality::IK_NoInduction:
1643 llvm_unreachable("Unknown induction");
1644 case LoopVectorizationLegality::IK_IntInduction: {
1645 // Handle the integer induction counter.
1646 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1648 // We have the canonical induction variable.
1649 if (OrigPhi == OldInduction) {
1650 // Create a truncated version of the resume value for the scalar loop,
1651 // we might have promoted the type to a larger width.
1653 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1654 // The new PHI merges the original incoming value, in case of a bypass,
1655 // or the value at the end of the vectorized loop.
1656 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1657 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1658 TruncResumeVal->addIncoming(EndValue, VecBody);
1660 // We know what the end value is.
1661 EndValue = IdxEndRoundDown;
1662 // We also know which PHI node holds it.
1663 ResumeIndex = ResumeVal;
1667 // Not the canonical induction variable - add the vector loop count to the
1669 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1670 II.StartValue->getType(),
1672 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1675 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1676 // Convert the CountRoundDown variable to the PHI size.
1677 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1678 II.StartValue->getType(),
1680 // Handle reverse integer induction counter.
1681 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1684 case LoopVectorizationLegality::IK_PtrInduction: {
1685 // For pointer induction variables, calculate the offset using
1687 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1691 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1692 // The value at the end of the loop for the reverse pointer is calculated
1693 // by creating a GEP with a negative index starting from the start value.
1694 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1695 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1697 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1703 // The new PHI merges the original incoming value, in case of a bypass,
1704 // or the value at the end of the vectorized loop.
1705 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1706 if (OrigPhi == OldInduction)
1707 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1709 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1711 ResumeVal->addIncoming(EndValue, VecBody);
1713 // Fix the scalar body counter (PHI node).
1714 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1715 // The old inductions phi node in the scalar body needs the truncated value.
1716 if (OrigPhi == OldInduction)
1717 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1719 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1722 // If we are generating a new induction variable then we also need to
1723 // generate the code that calculates the exit value. This value is not
1724 // simply the end of the counter because we may skip the vectorized body
1725 // in case of a runtime check.
1727 assert(!ResumeIndex && "Unexpected resume value found");
1728 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1729 MiddleBlock->getTerminator());
1730 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1731 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1732 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1735 // Make sure that we found the index where scalar loop needs to continue.
1736 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1737 "Invalid resume Index");
1739 // Add a check in the middle block to see if we have completed
1740 // all of the iterations in the first vector loop.
1741 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1742 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1743 ResumeIndex, "cmp.n",
1744 MiddleBlock->getTerminator());
1746 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1747 // Remove the old terminator.
1748 MiddleBlock->getTerminator()->eraseFromParent();
1750 // Create i+1 and fill the PHINode.
1751 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1752 Induction->addIncoming(StartIdx, VectorPH);
1753 Induction->addIncoming(NextIdx, VecBody);
1754 // Create the compare.
1755 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1756 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1758 // Now we have two terminators. Remove the old one from the block.
1759 VecBody->getTerminator()->eraseFromParent();
1761 // Get ready to start creating new instructions into the vectorized body.
1762 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1765 LoopVectorPreHeader = VectorPH;
1766 LoopScalarPreHeader = ScalarPH;
1767 LoopMiddleBlock = MiddleBlock;
1768 LoopExitBlock = ExitBlock;
1769 LoopVectorBody = VecBody;
1770 LoopScalarBody = OldBasicBlock;
1773 /// This function returns the identity element (or neutral element) for
1774 /// the operation K.
1776 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1781 // Adding, Xoring, Oring zero to a number does not change it.
1782 return ConstantInt::get(Tp, 0);
1783 case RK_IntegerMult:
1784 // Multiplying a number by 1 does not change it.
1785 return ConstantInt::get(Tp, 1);
1787 // AND-ing a number with an all-1 value does not change it.
1788 return ConstantInt::get(Tp, -1, true);
1790 // Multiplying a number by 1 does not change it.
1791 return ConstantFP::get(Tp, 1.0L);
1793 // Adding zero to a number does not change it.
1794 return ConstantFP::get(Tp, 0.0L);
1796 llvm_unreachable("Unknown reduction kind");
1800 static Intrinsic::ID
1801 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1802 // If we have an intrinsic call, check if it is trivially vectorizable.
1803 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1804 switch (II->getIntrinsicID()) {
1805 case Intrinsic::sqrt:
1806 case Intrinsic::sin:
1807 case Intrinsic::cos:
1808 case Intrinsic::exp:
1809 case Intrinsic::exp2:
1810 case Intrinsic::log:
1811 case Intrinsic::log10:
1812 case Intrinsic::log2:
1813 case Intrinsic::fabs:
1814 case Intrinsic::copysign:
1815 case Intrinsic::floor:
1816 case Intrinsic::ceil:
1817 case Intrinsic::trunc:
1818 case Intrinsic::rint:
1819 case Intrinsic::nearbyint:
1820 case Intrinsic::round:
1821 case Intrinsic::pow:
1822 case Intrinsic::fma:
1823 case Intrinsic::fmuladd:
1824 case Intrinsic::lifetime_start:
1825 case Intrinsic::lifetime_end:
1826 return II->getIntrinsicID();
1828 return Intrinsic::not_intrinsic;
1833 return Intrinsic::not_intrinsic;
1836 Function *F = CI->getCalledFunction();
1837 // We're going to make assumptions on the semantics of the functions, check
1838 // that the target knows that it's available in this environment.
1839 if (!F || !TLI->getLibFunc(F->getName(), Func))
1840 return Intrinsic::not_intrinsic;
1842 // Otherwise check if we have a call to a function that can be turned into a
1843 // vector intrinsic.
1850 return Intrinsic::sin;
1854 return Intrinsic::cos;
1858 return Intrinsic::exp;
1860 case LibFunc::exp2f:
1861 case LibFunc::exp2l:
1862 return Intrinsic::exp2;
1866 return Intrinsic::log;
1867 case LibFunc::log10:
1868 case LibFunc::log10f:
1869 case LibFunc::log10l:
1870 return Intrinsic::log10;
1872 case LibFunc::log2f:
1873 case LibFunc::log2l:
1874 return Intrinsic::log2;
1876 case LibFunc::fabsf:
1877 case LibFunc::fabsl:
1878 return Intrinsic::fabs;
1879 case LibFunc::copysign:
1880 case LibFunc::copysignf:
1881 case LibFunc::copysignl:
1882 return Intrinsic::copysign;
1883 case LibFunc::floor:
1884 case LibFunc::floorf:
1885 case LibFunc::floorl:
1886 return Intrinsic::floor;
1888 case LibFunc::ceilf:
1889 case LibFunc::ceill:
1890 return Intrinsic::ceil;
1891 case LibFunc::trunc:
1892 case LibFunc::truncf:
1893 case LibFunc::truncl:
1894 return Intrinsic::trunc;
1896 case LibFunc::rintf:
1897 case LibFunc::rintl:
1898 return Intrinsic::rint;
1899 case LibFunc::nearbyint:
1900 case LibFunc::nearbyintf:
1901 case LibFunc::nearbyintl:
1902 return Intrinsic::nearbyint;
1903 case LibFunc::round:
1904 case LibFunc::roundf:
1905 case LibFunc::roundl:
1906 return Intrinsic::round;
1910 return Intrinsic::pow;
1913 return Intrinsic::not_intrinsic;
1916 /// This function translates the reduction kind to an LLVM binary operator.
1918 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1920 case LoopVectorizationLegality::RK_IntegerAdd:
1921 return Instruction::Add;
1922 case LoopVectorizationLegality::RK_IntegerMult:
1923 return Instruction::Mul;
1924 case LoopVectorizationLegality::RK_IntegerOr:
1925 return Instruction::Or;
1926 case LoopVectorizationLegality::RK_IntegerAnd:
1927 return Instruction::And;
1928 case LoopVectorizationLegality::RK_IntegerXor:
1929 return Instruction::Xor;
1930 case LoopVectorizationLegality::RK_FloatMult:
1931 return Instruction::FMul;
1932 case LoopVectorizationLegality::RK_FloatAdd:
1933 return Instruction::FAdd;
1934 case LoopVectorizationLegality::RK_IntegerMinMax:
1935 return Instruction::ICmp;
1936 case LoopVectorizationLegality::RK_FloatMinMax:
1937 return Instruction::FCmp;
1939 llvm_unreachable("Unknown reduction operation");
1943 Value *createMinMaxOp(IRBuilder<> &Builder,
1944 LoopVectorizationLegality::MinMaxReductionKind RK,
1947 CmpInst::Predicate P = CmpInst::ICMP_NE;
1950 llvm_unreachable("Unknown min/max reduction kind");
1951 case LoopVectorizationLegality::MRK_UIntMin:
1952 P = CmpInst::ICMP_ULT;
1954 case LoopVectorizationLegality::MRK_UIntMax:
1955 P = CmpInst::ICMP_UGT;
1957 case LoopVectorizationLegality::MRK_SIntMin:
1958 P = CmpInst::ICMP_SLT;
1960 case LoopVectorizationLegality::MRK_SIntMax:
1961 P = CmpInst::ICMP_SGT;
1963 case LoopVectorizationLegality::MRK_FloatMin:
1964 P = CmpInst::FCMP_OLT;
1966 case LoopVectorizationLegality::MRK_FloatMax:
1967 P = CmpInst::FCMP_OGT;
1972 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
1973 RK == LoopVectorizationLegality::MRK_FloatMax)
1974 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1976 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1978 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1983 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1984 //===------------------------------------------------===//
1986 // Notice: any optimization or new instruction that go
1987 // into the code below should be also be implemented in
1990 //===------------------------------------------------===//
1991 Constant *Zero = Builder.getInt32(0);
1993 // In order to support reduction variables we need to be able to vectorize
1994 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1995 // stages. First, we create a new vector PHI node with no incoming edges.
1996 // We use this value when we vectorize all of the instructions that use the
1997 // PHI. Next, after all of the instructions in the block are complete we
1998 // add the new incoming edges to the PHI. At this point all of the
1999 // instructions in the basic block are vectorized, so we can use them to
2000 // construct the PHI.
2001 PhiVector RdxPHIsToFix;
2003 // Scan the loop in a topological order to ensure that defs are vectorized
2005 LoopBlocksDFS DFS(OrigLoop);
2008 // Vectorize all of the blocks in the original loop.
2009 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2010 be = DFS.endRPO(); bb != be; ++bb)
2011 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2013 // At this point every instruction in the original loop is widened to
2014 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2015 // that we vectorized. The PHI nodes are currently empty because we did
2016 // not want to introduce cycles. Notice that the remaining PHI nodes
2017 // that we need to fix are reduction variables.
2019 // Create the 'reduced' values for each of the induction vars.
2020 // The reduced values are the vector values that we scalarize and combine
2021 // after the loop is finished.
2022 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2024 PHINode *RdxPhi = *it;
2025 assert(RdxPhi && "Unable to recover vectorized PHI");
2027 // Find the reduction variable descriptor.
2028 assert(Legal->getReductionVars()->count(RdxPhi) &&
2029 "Unable to find the reduction variable");
2030 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2031 (*Legal->getReductionVars())[RdxPhi];
2033 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2035 // We need to generate a reduction vector from the incoming scalar.
2036 // To do so, we need to generate the 'identity' vector and overide
2037 // one of the elements with the incoming scalar reduction. We need
2038 // to do it in the vector-loop preheader.
2039 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2041 // This is the vector-clone of the value that leaves the loop.
2042 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2043 Type *VecTy = VectorExit[0]->getType();
2045 // Find the reduction identity variable. Zero for addition, or, xor,
2046 // one for multiplication, -1 for And.
2049 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2050 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2051 // MinMax reduction have the start value as their identify.
2052 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2056 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2057 VecTy->getScalarType());
2058 Identity = ConstantVector::getSplat(VF, Iden);
2060 // This vector is the Identity vector where the first element is the
2061 // incoming scalar reduction.
2062 VectorStart = Builder.CreateInsertElement(Identity,
2063 RdxDesc.StartValue, Zero);
2066 // Fix the vector-loop phi.
2067 // We created the induction variable so we know that the
2068 // preheader is the first entry.
2069 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2071 // Reductions do not have to start at zero. They can start with
2072 // any loop invariant values.
2073 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2074 BasicBlock *Latch = OrigLoop->getLoopLatch();
2075 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2076 VectorParts &Val = getVectorValue(LoopVal);
2077 for (unsigned part = 0; part < UF; ++part) {
2078 // Make sure to add the reduction stat value only to the
2079 // first unroll part.
2080 Value *StartVal = (part == 0) ? VectorStart : Identity;
2081 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2082 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2085 // Before each round, move the insertion point right between
2086 // the PHIs and the values we are going to write.
2087 // This allows us to write both PHINodes and the extractelement
2089 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2091 VectorParts RdxParts;
2092 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2093 for (unsigned part = 0; part < UF; ++part) {
2094 // This PHINode contains the vectorized reduction variable, or
2095 // the initial value vector, if we bypass the vector loop.
2096 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2097 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2098 Value *StartVal = (part == 0) ? VectorStart : Identity;
2099 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2100 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2101 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2102 RdxParts.push_back(NewPhi);
2105 // Reduce all of the unrolled parts into a single vector.
2106 Value *ReducedPartRdx = RdxParts[0];
2107 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2108 setDebugLocFromInst(Builder, ReducedPartRdx);
2109 for (unsigned part = 1; part < UF; ++part) {
2110 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2111 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2112 RdxParts[part], ReducedPartRdx,
2115 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2116 ReducedPartRdx, RdxParts[part]);
2119 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2120 // and vector ops, reducing the set of values being computed by half each
2122 assert(isPowerOf2_32(VF) &&
2123 "Reduction emission only supported for pow2 vectors!");
2124 Value *TmpVec = ReducedPartRdx;
2125 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2126 for (unsigned i = VF; i != 1; i >>= 1) {
2127 // Move the upper half of the vector to the lower half.
2128 for (unsigned j = 0; j != i/2; ++j)
2129 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2131 // Fill the rest of the mask with undef.
2132 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2133 UndefValue::get(Builder.getInt32Ty()));
2136 Builder.CreateShuffleVector(TmpVec,
2137 UndefValue::get(TmpVec->getType()),
2138 ConstantVector::get(ShuffleMask),
2141 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2142 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2145 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2148 // The result is in the first element of the vector.
2149 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2151 // Now, we need to fix the users of the reduction variable
2152 // inside and outside of the scalar remainder loop.
2153 // We know that the loop is in LCSSA form. We need to update the
2154 // PHI nodes in the exit blocks.
2155 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2156 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2157 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2158 if (!LCSSAPhi) continue;
2160 // All PHINodes need to have a single entry edge, or two if
2161 // we already fixed them.
2162 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2164 // We found our reduction value exit-PHI. Update it with the
2165 // incoming bypass edge.
2166 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2167 // Add an edge coming from the bypass.
2168 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2171 }// end of the LCSSA phi scan.
2173 // Fix the scalar loop reduction variable with the incoming reduction sum
2174 // from the vector body and from the backedge value.
2175 int IncomingEdgeBlockIdx =
2176 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2177 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2178 // Pick the other block.
2179 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2180 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2181 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2182 }// end of for each redux variable.
2187 void InnerLoopVectorizer::fixLCSSAPHIs() {
2188 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2189 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2190 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2191 if (!LCSSAPhi) continue;
2192 if (LCSSAPhi->getNumIncomingValues() == 1)
2193 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2198 InnerLoopVectorizer::VectorParts
2199 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2200 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2203 // Look for cached value.
2204 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2205 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2206 if (ECEntryIt != MaskCache.end())
2207 return ECEntryIt->second;
2209 VectorParts SrcMask = createBlockInMask(Src);
2211 // The terminator has to be a branch inst!
2212 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2213 assert(BI && "Unexpected terminator found");
2215 if (BI->isConditional()) {
2216 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2218 if (BI->getSuccessor(0) != Dst)
2219 for (unsigned part = 0; part < UF; ++part)
2220 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2222 for (unsigned part = 0; part < UF; ++part)
2223 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2225 MaskCache[Edge] = EdgeMask;
2229 MaskCache[Edge] = SrcMask;
2233 InnerLoopVectorizer::VectorParts
2234 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2235 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2237 // Loop incoming mask is all-one.
2238 if (OrigLoop->getHeader() == BB) {
2239 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2240 return getVectorValue(C);
2243 // This is the block mask. We OR all incoming edges, and with zero.
2244 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2245 VectorParts BlockMask = getVectorValue(Zero);
2248 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2249 VectorParts EM = createEdgeMask(*it, BB);
2250 for (unsigned part = 0; part < UF; ++part)
2251 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2257 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2258 InnerLoopVectorizer::VectorParts &Entry,
2259 LoopVectorizationLegality *Legal,
2260 unsigned UF, unsigned VF, PhiVector *PV) {
2261 PHINode* P = cast<PHINode>(PN);
2262 // Handle reduction variables:
2263 if (Legal->getReductionVars()->count(P)) {
2264 for (unsigned part = 0; part < UF; ++part) {
2265 // This is phase one of vectorizing PHIs.
2266 Type *VecTy = (VF == 1) ? PN->getType() :
2267 VectorType::get(PN->getType(), VF);
2268 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2269 LoopVectorBody-> getFirstInsertionPt());
2275 setDebugLocFromInst(Builder, P);
2276 // Check for PHI nodes that are lowered to vector selects.
2277 if (P->getParent() != OrigLoop->getHeader()) {
2278 // We know that all PHIs in non header blocks are converted into
2279 // selects, so we don't have to worry about the insertion order and we
2280 // can just use the builder.
2281 // At this point we generate the predication tree. There may be
2282 // duplications since this is a simple recursive scan, but future
2283 // optimizations will clean it up.
2285 unsigned NumIncoming = P->getNumIncomingValues();
2287 // Generate a sequence of selects of the form:
2288 // SELECT(Mask3, In3,
2289 // SELECT(Mask2, In2,
2291 for (unsigned In = 0; In < NumIncoming; In++) {
2292 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2294 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2296 for (unsigned part = 0; part < UF; ++part) {
2297 // We might have single edge PHIs (blocks) - use an identity
2298 // 'select' for the first PHI operand.
2300 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2303 // Select between the current value and the previous incoming edge
2304 // based on the incoming mask.
2305 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2306 Entry[part], "predphi");
2312 // This PHINode must be an induction variable.
2313 // Make sure that we know about it.
2314 assert(Legal->getInductionVars()->count(P) &&
2315 "Not an induction variable");
2317 LoopVectorizationLegality::InductionInfo II =
2318 Legal->getInductionVars()->lookup(P);
2321 case LoopVectorizationLegality::IK_NoInduction:
2322 llvm_unreachable("Unknown induction");
2323 case LoopVectorizationLegality::IK_IntInduction: {
2324 assert(P->getType() == II.StartValue->getType() && "Types must match");
2325 Type *PhiTy = P->getType();
2327 if (P == OldInduction) {
2328 // Handle the canonical induction variable. We might have had to
2330 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2332 // Handle other induction variables that are now based on the
2334 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2336 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2337 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2340 Broadcasted = getBroadcastInstrs(Broadcasted);
2341 // After broadcasting the induction variable we need to make the vector
2342 // consecutive by adding 0, 1, 2, etc.
2343 for (unsigned part = 0; part < UF; ++part)
2344 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2347 case LoopVectorizationLegality::IK_ReverseIntInduction:
2348 case LoopVectorizationLegality::IK_PtrInduction:
2349 case LoopVectorizationLegality::IK_ReversePtrInduction:
2350 // Handle reverse integer and pointer inductions.
2351 Value *StartIdx = ExtendedIdx;
2352 // This is the normalized GEP that starts counting at zero.
2353 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2356 // Handle the reverse integer induction variable case.
2357 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2358 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2359 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2361 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2364 // This is a new value so do not hoist it out.
2365 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2366 // After broadcasting the induction variable we need to make the
2367 // vector consecutive by adding ... -3, -2, -1, 0.
2368 for (unsigned part = 0; part < UF; ++part)
2369 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2374 // Handle the pointer induction variable case.
2375 assert(P->getType()->isPointerTy() && "Unexpected type.");
2377 // Is this a reverse induction ptr or a consecutive induction ptr.
2378 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2381 // This is the vector of results. Notice that we don't generate
2382 // vector geps because scalar geps result in better code.
2383 for (unsigned part = 0; part < UF; ++part) {
2385 int EltIndex = (part) * (Reverse ? -1 : 1);
2386 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2389 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2391 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2393 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2395 Entry[part] = SclrGep;
2399 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2400 for (unsigned int i = 0; i < VF; ++i) {
2401 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2402 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2405 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2407 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2409 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2411 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2412 Builder.getInt32(i),
2415 Entry[part] = VecVal;
2422 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2423 BasicBlock *BB, PhiVector *PV) {
2424 // For each instruction in the old loop.
2425 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2426 VectorParts &Entry = WidenMap.get(it);
2427 switch (it->getOpcode()) {
2428 case Instruction::Br:
2429 // Nothing to do for PHIs and BR, since we already took care of the
2430 // loop control flow instructions.
2432 case Instruction::PHI:{
2433 // Vectorize PHINodes.
2434 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2438 case Instruction::Add:
2439 case Instruction::FAdd:
2440 case Instruction::Sub:
2441 case Instruction::FSub:
2442 case Instruction::Mul:
2443 case Instruction::FMul:
2444 case Instruction::UDiv:
2445 case Instruction::SDiv:
2446 case Instruction::FDiv:
2447 case Instruction::URem:
2448 case Instruction::SRem:
2449 case Instruction::FRem:
2450 case Instruction::Shl:
2451 case Instruction::LShr:
2452 case Instruction::AShr:
2453 case Instruction::And:
2454 case Instruction::Or:
2455 case Instruction::Xor: {
2456 // Just widen binops.
2457 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2458 setDebugLocFromInst(Builder, BinOp);
2459 VectorParts &A = getVectorValue(it->getOperand(0));
2460 VectorParts &B = getVectorValue(it->getOperand(1));
2462 // Use this vector value for all users of the original instruction.
2463 for (unsigned Part = 0; Part < UF; ++Part) {
2464 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2466 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2467 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2468 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2469 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2470 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2472 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2473 VecOp->setIsExact(BinOp->isExact());
2479 case Instruction::Select: {
2481 // If the selector is loop invariant we can create a select
2482 // instruction with a scalar condition. Otherwise, use vector-select.
2483 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2485 setDebugLocFromInst(Builder, it);
2487 // The condition can be loop invariant but still defined inside the
2488 // loop. This means that we can't just use the original 'cond' value.
2489 // We have to take the 'vectorized' value and pick the first lane.
2490 // Instcombine will make this a no-op.
2491 VectorParts &Cond = getVectorValue(it->getOperand(0));
2492 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2493 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2495 Value *ScalarCond = (VF == 1) ? Cond[0] :
2496 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2498 for (unsigned Part = 0; Part < UF; ++Part) {
2499 Entry[Part] = Builder.CreateSelect(
2500 InvariantCond ? ScalarCond : Cond[Part],
2507 case Instruction::ICmp:
2508 case Instruction::FCmp: {
2509 // Widen compares. Generate vector compares.
2510 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2511 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2512 setDebugLocFromInst(Builder, it);
2513 VectorParts &A = getVectorValue(it->getOperand(0));
2514 VectorParts &B = getVectorValue(it->getOperand(1));
2515 for (unsigned Part = 0; Part < UF; ++Part) {
2518 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2520 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2526 case Instruction::Store:
2527 case Instruction::Load:
2528 vectorizeMemoryInstruction(it, Legal);
2530 case Instruction::ZExt:
2531 case Instruction::SExt:
2532 case Instruction::FPToUI:
2533 case Instruction::FPToSI:
2534 case Instruction::FPExt:
2535 case Instruction::PtrToInt:
2536 case Instruction::IntToPtr:
2537 case Instruction::SIToFP:
2538 case Instruction::UIToFP:
2539 case Instruction::Trunc:
2540 case Instruction::FPTrunc:
2541 case Instruction::BitCast: {
2542 CastInst *CI = dyn_cast<CastInst>(it);
2543 setDebugLocFromInst(Builder, it);
2544 /// Optimize the special case where the source is the induction
2545 /// variable. Notice that we can only optimize the 'trunc' case
2546 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2547 /// c. other casts depend on pointer size.
2548 if (CI->getOperand(0) == OldInduction &&
2549 it->getOpcode() == Instruction::Trunc) {
2550 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2552 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2553 for (unsigned Part = 0; Part < UF; ++Part)
2554 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2557 /// Vectorize casts.
2558 Type *DestTy = (VF == 1) ? CI->getType() :
2559 VectorType::get(CI->getType(), VF);
2561 VectorParts &A = getVectorValue(it->getOperand(0));
2562 for (unsigned Part = 0; Part < UF; ++Part)
2563 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2567 case Instruction::Call: {
2568 // Ignore dbg intrinsics.
2569 if (isa<DbgInfoIntrinsic>(it))
2571 setDebugLocFromInst(Builder, it);
2573 Module *M = BB->getParent()->getParent();
2574 CallInst *CI = cast<CallInst>(it);
2575 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2576 assert(ID && "Not an intrinsic call!");
2578 case Intrinsic::lifetime_end:
2579 case Intrinsic::lifetime_start:
2580 scalarizeInstruction(it);
2583 for (unsigned Part = 0; Part < UF; ++Part) {
2584 SmallVector<Value *, 4> Args;
2585 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2586 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2587 Args.push_back(Arg[Part]);
2589 Type *Tys[] = {CI->getType()};
2591 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2593 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2594 Entry[Part] = Builder.CreateCall(F, Args);
2602 // All other instructions are unsupported. Scalarize them.
2603 scalarizeInstruction(it);
2606 }// end of for_each instr.
2609 void InnerLoopVectorizer::updateAnalysis() {
2610 // Forget the original basic block.
2611 SE->forgetLoop(OrigLoop);
2613 // Update the dominator tree information.
2614 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2615 "Entry does not dominate exit.");
2617 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2618 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2619 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2620 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2621 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2622 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2623 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2624 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2626 DEBUG(DT->verifyAnalysis());
2629 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2630 if (!EnableIfConversion)
2633 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2634 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2636 // A list of pointers that we can safely read and write to.
2637 SmallPtrSet<Value *, 8> SafePointes;
2639 // Collect safe addresses.
2640 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2641 BasicBlock *BB = LoopBlocks[i];
2643 if (blockNeedsPredication(BB))
2646 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2647 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2648 SafePointes.insert(LI->getPointerOperand());
2649 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2650 SafePointes.insert(SI->getPointerOperand());
2654 // Collect the blocks that need predication.
2655 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2656 BasicBlock *BB = LoopBlocks[i];
2658 // We don't support switch statements inside loops.
2659 if (!isa<BranchInst>(BB->getTerminator()))
2662 // We must be able to predicate all blocks that need to be predicated.
2663 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2667 // We can if-convert this loop.
2671 bool LoopVectorizationLegality::canVectorize() {
2672 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2673 // be canonicalized.
2674 if (!TheLoop->getLoopPreheader())
2677 // We can only vectorize innermost loops.
2678 if (TheLoop->getSubLoopsVector().size())
2681 // We must have a single backedge.
2682 if (TheLoop->getNumBackEdges() != 1)
2685 // We must have a single exiting block.
2686 if (!TheLoop->getExitingBlock())
2689 unsigned NumBlocks = TheLoop->getNumBlocks();
2691 // Check if we can if-convert non single-bb loops.
2692 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2693 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2697 // We need to have a loop header.
2698 BasicBlock *Latch = TheLoop->getLoopLatch();
2699 DEBUG(dbgs() << "LV: Found a loop: " <<
2700 TheLoop->getHeader()->getName() << "\n");
2702 // ScalarEvolution needs to be able to find the exit count.
2703 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2704 if (ExitCount == SE->getCouldNotCompute()) {
2705 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2709 // Do not loop-vectorize loops with a tiny trip count.
2710 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2711 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2712 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2713 "This loop is not worth vectorizing.\n");
2717 // Check if we can vectorize the instructions and CFG in this loop.
2718 if (!canVectorizeInstrs()) {
2719 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2723 // Go over each instruction and look at memory deps.
2724 if (!canVectorizeMemory()) {
2725 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2729 // Collect all of the variables that remain uniform after vectorization.
2730 collectLoopUniforms();
2732 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2733 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2736 // Okay! We can vectorize. At this point we don't have any other mem analysis
2737 // which may limit our maximum vectorization factor, so just return true with
2742 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2743 if (Ty->isPointerTy())
2744 return DL.getIntPtrType(Ty);
2749 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2750 Ty0 = convertPointerToIntegerType(DL, Ty0);
2751 Ty1 = convertPointerToIntegerType(DL, Ty1);
2752 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2757 /// \brief Check that the instruction has outside loop users and is not an
2758 /// identified reduction variable.
2759 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2760 SmallPtrSet<Value *, 4> &Reductions) {
2761 // Reduction instructions are allowed to have exit users. All other
2762 // instructions must not have external users.
2763 if (!Reductions.count(Inst))
2764 //Check that all of the users of the loop are inside the BB.
2765 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2767 Instruction *U = cast<Instruction>(*I);
2768 // This user may be a reduction exit value.
2769 if (!TheLoop->contains(U)) {
2770 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2777 bool LoopVectorizationLegality::canVectorizeInstrs() {
2778 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2779 BasicBlock *Header = TheLoop->getHeader();
2781 // Look for the attribute signaling the absence of NaNs.
2782 Function &F = *Header->getParent();
2783 if (F.hasFnAttribute("no-nans-fp-math"))
2784 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2785 AttributeSet::FunctionIndex,
2786 "no-nans-fp-math").getValueAsString() == "true";
2788 // For each block in the loop.
2789 for (Loop::block_iterator bb = TheLoop->block_begin(),
2790 be = TheLoop->block_end(); bb != be; ++bb) {
2792 // Scan the instructions in the block and look for hazards.
2793 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2796 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2797 Type *PhiTy = Phi->getType();
2798 // Check that this PHI type is allowed.
2799 if (!PhiTy->isIntegerTy() &&
2800 !PhiTy->isFloatingPointTy() &&
2801 !PhiTy->isPointerTy()) {
2802 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2806 // If this PHINode is not in the header block, then we know that we
2807 // can convert it to select during if-conversion. No need to check if
2808 // the PHIs in this block are induction or reduction variables.
2809 if (*bb != Header) {
2810 // Check that this instruction has no outside users or is an
2811 // identified reduction value with an outside user.
2812 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2817 // We only allow if-converted PHIs with more than two incoming values.
2818 if (Phi->getNumIncomingValues() != 2) {
2819 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2823 // This is the value coming from the preheader.
2824 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2825 // Check if this is an induction variable.
2826 InductionKind IK = isInductionVariable(Phi);
2828 if (IK_NoInduction != IK) {
2829 // Get the widest type.
2831 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2833 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2835 // Int inductions are special because we only allow one IV.
2836 if (IK == IK_IntInduction) {
2837 // Use the phi node with the widest type as induction. Use the last
2838 // one if there are multiple (no good reason for doing this other
2839 // than it is expedient).
2840 if (!Induction || PhiTy == WidestIndTy)
2844 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2845 Inductions[Phi] = InductionInfo(StartValue, IK);
2849 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2850 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2853 if (AddReductionVar(Phi, RK_IntegerMult)) {
2854 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2857 if (AddReductionVar(Phi, RK_IntegerOr)) {
2858 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2861 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2862 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2865 if (AddReductionVar(Phi, RK_IntegerXor)) {
2866 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2869 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2870 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2873 if (AddReductionVar(Phi, RK_FloatMult)) {
2874 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2877 if (AddReductionVar(Phi, RK_FloatAdd)) {
2878 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2881 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2882 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
2887 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2889 }// end of PHI handling
2891 // We still don't handle functions. However, we can ignore dbg intrinsic
2892 // calls and we do handle certain intrinsic and libm functions.
2893 CallInst *CI = dyn_cast<CallInst>(it);
2894 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2895 DEBUG(dbgs() << "LV: Found a call site.\n");
2899 // Check that the instruction return type is vectorizable.
2900 if (!VectorType::isValidElementType(it->getType()) &&
2901 !it->getType()->isVoidTy()) {
2902 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2906 // Check that the stored type is vectorizable.
2907 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2908 Type *T = ST->getValueOperand()->getType();
2909 if (!VectorType::isValidElementType(T))
2913 // Reduction instructions are allowed to have exit users.
2914 // All other instructions must not have external users.
2915 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2923 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2924 if (Inductions.empty())
2931 void LoopVectorizationLegality::collectLoopUniforms() {
2932 // We now know that the loop is vectorizable!
2933 // Collect variables that will remain uniform after vectorization.
2934 std::vector<Value*> Worklist;
2935 BasicBlock *Latch = TheLoop->getLoopLatch();
2937 // Start with the conditional branch and walk up the block.
2938 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2940 while (Worklist.size()) {
2941 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2942 Worklist.pop_back();
2944 // Look at instructions inside this loop.
2945 // Stop when reaching PHI nodes.
2946 // TODO: we need to follow values all over the loop, not only in this block.
2947 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2950 // This is a known uniform.
2953 // Insert all operands.
2954 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2959 /// \brief Analyses memory accesses in a loop.
2961 /// Checks whether run time pointer checks are needed and builds sets for data
2962 /// dependence checking.
2963 class AccessAnalysis {
2965 /// \brief Read or write access location.
2966 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
2967 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
2969 /// \brief Set of potential dependent memory accesses.
2970 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2972 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2973 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2974 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2976 /// \brief Register a load and whether it is only read from.
2977 void addLoad(Value *Ptr, bool IsReadOnly) {
2978 Accesses.insert(MemAccessInfo(Ptr, false));
2980 ReadOnlyPtr.insert(Ptr);
2983 /// \brief Register a store.
2984 void addStore(Value *Ptr) {
2985 Accesses.insert(MemAccessInfo(Ptr, true));
2988 /// \brief Check whether we can check the pointers at runtime for
2989 /// non-intersection.
2990 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2991 unsigned &NumComparisons, ScalarEvolution *SE,
2994 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2995 /// and builds sets of dependent accesses.
2996 void buildDependenceSets() {
2997 // Process read-write pointers first.
2998 processMemAccesses(false);
2999 // Next, process read pointers.
3000 processMemAccesses(true);
3003 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3005 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3007 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3010 typedef SetVector<MemAccessInfo> PtrAccessSet;
3011 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3013 /// \brief Go over all memory access or only the deferred ones if
3014 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3015 /// and build sets of dependency check candidates.
3016 void processMemAccesses(bool UseDeferred);
3018 /// Set of all accesses.
3019 PtrAccessSet Accesses;
3021 /// Set of access to check after all writes have been processed.
3022 PtrAccessSet DeferredAccesses;
3024 /// Map of pointers to last access encountered.
3025 UnderlyingObjToAccessMap ObjToLastAccess;
3027 /// Set of accesses that need a further dependence check.
3028 MemAccessInfoSet CheckDeps;
3030 /// Set of pointers that are read only.
3031 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3033 /// Set of underlying objects already written to.
3034 SmallPtrSet<Value*, 16> WriteObjects;
3038 /// Sets of potentially dependent accesses - members of one set share an
3039 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3040 /// dependence check.
3041 DepCandidates &DepCands;
3043 bool AreAllWritesIdentified;
3044 bool AreAllReadsIdentified;
3045 bool IsRTCheckNeeded;
3048 } // end anonymous namespace
3050 /// \brief Check whether a pointer can participate in a runtime bounds check.
3051 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3052 const SCEV *PtrScev = SE->getSCEV(Ptr);
3053 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3057 return AR->isAffine();
3060 bool AccessAnalysis::canCheckPtrAtRT(
3061 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3062 unsigned &NumComparisons, ScalarEvolution *SE,
3064 // Find pointers with computable bounds. We are going to use this information
3065 // to place a runtime bound check.
3066 unsigned NumReadPtrChecks = 0;
3067 unsigned NumWritePtrChecks = 0;
3068 bool CanDoRT = true;
3070 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3071 // We assign consecutive id to access from different dependence sets.
3072 // Accesses within the same set don't need a runtime check.
3073 unsigned RunningDepId = 1;
3074 DenseMap<Value *, unsigned> DepSetId;
3076 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3078 const MemAccessInfo &Access = *AI;
3079 Value *Ptr = Access.getPointer();
3080 bool IsWrite = Access.getInt();
3082 // Just add write checks if we have both.
3083 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3087 ++NumWritePtrChecks;
3091 if (hasComputableBounds(SE, Ptr)) {
3092 // The id of the dependence set.
3095 if (IsDepCheckNeeded) {
3096 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3097 unsigned &LeaderId = DepSetId[Leader];
3099 LeaderId = RunningDepId++;
3102 // Each access has its own dependence set.
3103 DepId = RunningDepId++;
3105 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3107 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3113 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3114 NumComparisons = 0; // Only one dependence set.
3116 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3117 NumWritePtrChecks - 1));
3121 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3122 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3125 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3126 // We process the set twice: first we process read-write pointers, last we
3127 // process read-only pointers. This allows us to skip dependence tests for
3128 // read-only pointers.
3130 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3131 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3132 const MemAccessInfo &Access = *AI;
3133 Value *Ptr = Access.getPointer();
3134 bool IsWrite = Access.getInt();
3136 DepCands.insert(Access);
3138 // Memorize read-only pointers for later processing and skip them in the
3139 // first round (they need to be checked after we have seen all write
3140 // pointers). Note: we also mark pointer that are not consecutive as
3141 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3142 // second check for "!IsWrite".
3143 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3144 if (!UseDeferred && IsReadOnlyPtr) {
3145 DeferredAccesses.insert(Access);
3149 bool NeedDepCheck = false;
3150 // Check whether there is the possiblity of dependency because of underlying
3151 // objects being the same.
3152 typedef SmallVector<Value*, 16> ValueVector;
3153 ValueVector TempObjects;
3154 GetUnderlyingObjects(Ptr, TempObjects, DL);
3155 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3157 Value *UnderlyingObj = *UI;
3159 // If this is a write then it needs to be an identified object. If this a
3160 // read and all writes (so far) are identified function scope objects we
3161 // don't need an identified underlying object but only an Argument (the
3162 // next write is going to invalidate this assumption if it is
3164 // This is a micro-optimization for the case where all writes are
3165 // identified and we have one argument pointer.
3166 // Otherwise, we do need a runtime check.
3167 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3168 (!IsWrite && (!AreAllWritesIdentified ||
3169 !isa<Argument>(UnderlyingObj)) &&
3170 !isIdentifiedObject(UnderlyingObj))) {
3171 DEBUG(dbgs() << "LV: Found an unidentified " <<
3172 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3174 IsRTCheckNeeded = (IsRTCheckNeeded ||
3175 !isIdentifiedObject(UnderlyingObj) ||
3176 !AreAllReadsIdentified);
3179 AreAllWritesIdentified = false;
3181 AreAllReadsIdentified = false;
3184 // If this is a write - check other reads and writes for conflicts. If
3185 // this is a read only check other writes for conflicts (but only if there
3186 // is no other write to the ptr - this is an optimization to catch "a[i] =
3187 // a[i] + " without having to do a dependence check).
3188 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3189 NeedDepCheck = true;
3192 WriteObjects.insert(UnderlyingObj);
3194 // Create sets of pointers connected by shared underlying objects.
3195 UnderlyingObjToAccessMap::iterator Prev =
3196 ObjToLastAccess.find(UnderlyingObj);
3197 if (Prev != ObjToLastAccess.end())
3198 DepCands.unionSets(Access, Prev->second);
3200 ObjToLastAccess[UnderlyingObj] = Access;
3204 CheckDeps.insert(Access);
3209 /// \brief Checks memory dependences among accesses to the same underlying
3210 /// object to determine whether there vectorization is legal or not (and at
3211 /// which vectorization factor).
3213 /// This class works under the assumption that we already checked that memory
3214 /// locations with different underlying pointers are "must-not alias".
3215 /// We use the ScalarEvolution framework to symbolically evalutate access
3216 /// functions pairs. Since we currently don't restructure the loop we can rely
3217 /// on the program order of memory accesses to determine their safety.
3218 /// At the moment we will only deem accesses as safe for:
3219 /// * A negative constant distance assuming program order.
3221 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3222 /// a[i] = tmp; y = a[i];
3224 /// The latter case is safe because later checks guarantuee that there can't
3225 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3226 /// the same variable: a header phi can only be an induction or a reduction, a
3227 /// reduction can't have a memory sink, an induction can't have a memory
3228 /// source). This is important and must not be violated (or we have to
3229 /// resort to checking for cycles through memory).
3231 /// * A positive constant distance assuming program order that is bigger
3232 /// than the biggest memory access.
3234 /// tmp = a[i] OR b[i] = x
3235 /// a[i+2] = tmp y = b[i+2];
3237 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3239 /// * Zero distances and all accesses have the same size.
3241 class MemoryDepChecker {
3243 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3244 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3246 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3247 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3249 /// \brief Register the location (instructions are given increasing numbers)
3250 /// of a write access.
3251 void addAccess(StoreInst *SI) {
3252 Value *Ptr = SI->getPointerOperand();
3253 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3254 InstMap.push_back(SI);
3258 /// \brief Register the location (instructions are given increasing numbers)
3259 /// of a write access.
3260 void addAccess(LoadInst *LI) {
3261 Value *Ptr = LI->getPointerOperand();
3262 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3263 InstMap.push_back(LI);
3267 /// \brief Check whether the dependencies between the accesses are safe.
3269 /// Only checks sets with elements in \p CheckDeps.
3270 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3271 MemAccessInfoSet &CheckDeps);
3273 /// \brief The maximum number of bytes of a vector register we can vectorize
3274 /// the accesses safely with.
3275 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3278 ScalarEvolution *SE;
3280 const Loop *InnermostLoop;
3282 /// \brief Maps access locations (ptr, read/write) to program order.
3283 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3285 /// \brief Memory access instructions in program order.
3286 SmallVector<Instruction *, 16> InstMap;
3288 /// \brief The program order index to be used for the next instruction.
3291 // We can access this many bytes in parallel safely.
3292 unsigned MaxSafeDepDistBytes;
3294 /// \brief Check whether there is a plausible dependence between the two
3297 /// Access \p A must happen before \p B in program order. The two indices
3298 /// identify the index into the program order map.
3300 /// This function checks whether there is a plausible dependence (or the
3301 /// absence of such can't be proved) between the two accesses. If there is a
3302 /// plausible dependence but the dependence distance is bigger than one
3303 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3304 /// distance is smaller than any other distance encountered so far).
3305 /// Otherwise, this function returns true signaling a possible dependence.
3306 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3307 const MemAccessInfo &B, unsigned BIdx);
3309 /// \brief Check whether the data dependence could prevent store-load
3311 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3314 } // end anonymous namespace
3316 static bool isInBoundsGep(Value *Ptr) {
3317 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3318 return GEP->isInBounds();
3322 /// \brief Check whether the access through \p Ptr has a constant stride.
3323 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3325 const Type *Ty = Ptr->getType();
3326 assert(Ty->isPointerTy() && "Unexpected non ptr");
3328 // Make sure that the pointer does not point to aggregate types.
3329 const PointerType *PtrTy = cast<PointerType>(Ty);
3330 if (PtrTy->getElementType()->isAggregateType()) {
3331 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3336 const SCEV *PtrScev = SE->getSCEV(Ptr);
3337 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3339 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3340 << *Ptr << " SCEV: " << *PtrScev << "\n");
3344 // The accesss function must stride over the innermost loop.
3345 if (Lp != AR->getLoop()) {
3346 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3347 *Ptr << " SCEV: " << *PtrScev << "\n");
3350 // The address calculation must not wrap. Otherwise, a dependence could be
3352 // An inbounds getelementptr that is a AddRec with a unit stride
3353 // cannot wrap per definition. The unit stride requirement is checked later.
3354 // An getelementptr without an inbounds attribute and unit stride would have
3355 // to access the pointer value "0" which is undefined behavior in address
3356 // space 0, therefore we can also vectorize this case.
3357 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3358 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3359 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3360 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3361 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3362 << *Ptr << " SCEV: " << *PtrScev << "\n");
3366 // Check the step is constant.
3367 const SCEV *Step = AR->getStepRecurrence(*SE);
3369 // Calculate the pointer stride and check if it is consecutive.
3370 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3372 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3373 " SCEV: " << *PtrScev << "\n");
3377 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3378 const APInt &APStepVal = C->getValue()->getValue();
3380 // Huge step value - give up.
3381 if (APStepVal.getBitWidth() > 64)
3384 int64_t StepVal = APStepVal.getSExtValue();
3387 int64_t Stride = StepVal / Size;
3388 int64_t Rem = StepVal % Size;
3392 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3393 // know we can't "wrap around the address space". In case of address space
3394 // zero we know that this won't happen without triggering undefined behavior.
3395 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3396 Stride != 1 && Stride != -1)
3402 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3403 unsigned TypeByteSize) {
3404 // If loads occur at a distance that is not a multiple of a feasible vector
3405 // factor store-load forwarding does not take place.
3406 // Positive dependences might cause troubles because vectorizing them might
3407 // prevent store-load forwarding making vectorized code run a lot slower.
3408 // a[i] = a[i-3] ^ a[i-8];
3409 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3410 // hence on your typical architecture store-load forwarding does not take
3411 // place. Vectorizing in such cases does not make sense.
3412 // Store-load forwarding distance.
3413 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3414 // Maximum vector factor.
3415 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3416 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3417 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3419 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3421 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3422 MaxVFWithoutSLForwardIssues = (vf >>=1);
3427 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3428 DEBUG(dbgs() << "LV: Distance " << Distance <<
3429 " that could cause a store-load forwarding conflict\n");
3433 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3434 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3435 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3439 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3440 const MemAccessInfo &B, unsigned BIdx) {
3441 assert (AIdx < BIdx && "Must pass arguments in program order");
3443 Value *APtr = A.getPointer();
3444 Value *BPtr = B.getPointer();
3445 bool AIsWrite = A.getInt();
3446 bool BIsWrite = B.getInt();
3448 // Two reads are independent.
3449 if (!AIsWrite && !BIsWrite)
3452 const SCEV *AScev = SE->getSCEV(APtr);
3453 const SCEV *BScev = SE->getSCEV(BPtr);
3455 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3456 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3458 const SCEV *Src = AScev;
3459 const SCEV *Sink = BScev;
3461 // If the induction step is negative we have to invert source and sink of the
3463 if (StrideAPtr < 0) {
3466 std::swap(APtr, BPtr);
3467 std::swap(Src, Sink);
3468 std::swap(AIsWrite, BIsWrite);
3469 std::swap(AIdx, BIdx);
3470 std::swap(StrideAPtr, StrideBPtr);
3473 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3475 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3476 << "(Induction step: " << StrideAPtr << ")\n");
3477 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3478 << *InstMap[BIdx] << ": " << *Dist << "\n");
3480 // Need consecutive accesses. We don't want to vectorize
3481 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3482 // the address space.
3483 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3484 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3488 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3490 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3494 Type *ATy = APtr->getType()->getPointerElementType();
3495 Type *BTy = BPtr->getType()->getPointerElementType();
3496 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3498 // Negative distances are not plausible dependencies.
3499 const APInt &Val = C->getValue()->getValue();
3500 if (Val.isNegative()) {
3501 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3502 if (IsTrueDataDependence &&
3503 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3507 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3511 // Write to the same location with the same size.
3512 // Could be improved to assert type sizes are the same (i32 == float, etc).
3516 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3520 assert(Val.isStrictlyPositive() && "Expect a positive value");
3522 // Positive distance bigger than max vectorization factor.
3525 "LV: ReadWrite-Write positive dependency with different types");
3529 unsigned Distance = (unsigned) Val.getZExtValue();
3531 // Bail out early if passed-in parameters make vectorization not feasible.
3532 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3533 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3535 // The distance must be bigger than the size needed for a vectorized version
3536 // of the operation and the size of the vectorized operation must not be
3537 // bigger than the currrent maximum size.
3538 if (Distance < 2*TypeByteSize ||
3539 2*TypeByteSize > MaxSafeDepDistBytes ||
3540 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3541 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3542 << Val.getSExtValue() << "\n");
3546 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3547 Distance : MaxSafeDepDistBytes;
3549 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3550 if (IsTrueDataDependence &&
3551 couldPreventStoreLoadForward(Distance, TypeByteSize))
3554 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3555 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3561 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3562 MemAccessInfoSet &CheckDeps) {
3564 MaxSafeDepDistBytes = -1U;
3565 while (!CheckDeps.empty()) {
3566 MemAccessInfo CurAccess = *CheckDeps.begin();
3568 // Get the relevant memory access set.
3569 EquivalenceClasses<MemAccessInfo>::iterator I =
3570 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3572 // Check accesses within this set.
3573 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3574 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3576 // Check every access pair.
3578 CheckDeps.erase(*AI);
3579 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3581 // Check every accessing instruction pair in program order.
3582 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3583 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3584 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3585 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3586 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3588 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3599 bool LoopVectorizationLegality::canVectorizeMemory() {
3601 typedef SmallVector<Value*, 16> ValueVector;
3602 typedef SmallPtrSet<Value*, 16> ValueSet;
3604 // Holds the Load and Store *instructions*.
3608 // Holds all the different accesses in the loop.
3609 unsigned NumReads = 0;
3610 unsigned NumReadWrites = 0;
3612 PtrRtCheck.Pointers.clear();
3613 PtrRtCheck.Need = false;
3615 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3616 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3619 for (Loop::block_iterator bb = TheLoop->block_begin(),
3620 be = TheLoop->block_end(); bb != be; ++bb) {
3622 // Scan the BB and collect legal loads and stores.
3623 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3626 // If this is a load, save it. If this instruction can read from memory
3627 // but is not a load, then we quit. Notice that we don't handle function
3628 // calls that read or write.
3629 if (it->mayReadFromMemory()) {
3630 // Many math library functions read the rounding mode. We will only
3631 // vectorize a loop if it contains known function calls that don't set
3632 // the flag. Therefore, it is safe to ignore this read from memory.
3633 CallInst *Call = dyn_cast<CallInst>(it);
3634 if (Call && getIntrinsicIDForCall(Call, TLI))
3637 LoadInst *Ld = dyn_cast<LoadInst>(it);
3638 if (!Ld) return false;
3639 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3640 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3643 Loads.push_back(Ld);
3644 DepChecker.addAccess(Ld);
3648 // Save 'store' instructions. Abort if other instructions write to memory.
3649 if (it->mayWriteToMemory()) {
3650 StoreInst *St = dyn_cast<StoreInst>(it);
3651 if (!St) return false;
3652 if (!St->isSimple() && !IsAnnotatedParallel) {
3653 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3656 Stores.push_back(St);
3657 DepChecker.addAccess(St);
3662 // Now we have two lists that hold the loads and the stores.
3663 // Next, we find the pointers that they use.
3665 // Check if we see any stores. If there are no stores, then we don't
3666 // care if the pointers are *restrict*.
3667 if (!Stores.size()) {
3668 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3672 AccessAnalysis::DepCandidates DependentAccesses;
3673 AccessAnalysis Accesses(DL, DependentAccesses);
3675 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3676 // multiple times on the same object. If the ptr is accessed twice, once
3677 // for read and once for write, it will only appear once (on the write
3678 // list). This is okay, since we are going to check for conflicts between
3679 // writes and between reads and writes, but not between reads and reads.
3682 ValueVector::iterator I, IE;
3683 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3684 StoreInst *ST = cast<StoreInst>(*I);
3685 Value* Ptr = ST->getPointerOperand();
3687 if (isUniform(Ptr)) {
3688 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3692 // If we did *not* see this pointer before, insert it to the read-write
3693 // list. At this phase it is only a 'write' list.
3694 if (Seen.insert(Ptr)) {
3696 Accesses.addStore(Ptr);
3700 if (IsAnnotatedParallel) {
3702 << "LV: A loop annotated parallel, ignore memory dependency "
3707 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3708 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3709 LoadInst *LD = cast<LoadInst>(*I);
3710 Value* Ptr = LD->getPointerOperand();
3711 // If we did *not* see this pointer before, insert it to the
3712 // read list. If we *did* see it before, then it is already in
3713 // the read-write list. This allows us to vectorize expressions
3714 // such as A[i] += x; Because the address of A[i] is a read-write
3715 // pointer. This only works if the index of A[i] is consecutive.
3716 // If the address of i is unknown (for example A[B[i]]) then we may
3717 // read a few words, modify, and write a few words, and some of the
3718 // words may be written to the same address.
3719 bool IsReadOnlyPtr = false;
3720 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3722 IsReadOnlyPtr = true;
3724 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3727 // If we write (or read-write) to a single destination and there are no
3728 // other reads in this loop then is it safe to vectorize.
3729 if (NumReadWrites == 1 && NumReads == 0) {
3730 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3734 // Build dependence sets and check whether we need a runtime pointer bounds
3736 Accesses.buildDependenceSets();
3737 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3739 // Find pointers with computable bounds. We are going to use this information
3740 // to place a runtime bound check.
3741 unsigned NumComparisons = 0;
3742 bool CanDoRT = false;
3744 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3747 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3748 " pointer comparisons.\n");
3750 // If we only have one set of dependences to check pointers among we don't
3751 // need a runtime check.
3752 if (NumComparisons == 0 && NeedRTCheck)
3753 NeedRTCheck = false;
3755 // Check that we did not collect too many pointers or found a unsizeable
3757 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3763 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3766 if (NeedRTCheck && !CanDoRT) {
3767 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3768 "the array bounds.\n");
3773 PtrRtCheck.Need = NeedRTCheck;
3775 bool CanVecMem = true;
3776 if (Accesses.isDependencyCheckNeeded()) {
3777 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3778 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3779 Accesses.getDependenciesToCheck());
3780 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3783 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3784 " need a runtime memory check.\n");
3789 static bool hasMultipleUsesOf(Instruction *I,
3790 SmallPtrSet<Instruction *, 8> &Insts) {
3791 unsigned NumUses = 0;
3792 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3793 if (Insts.count(dyn_cast<Instruction>(*Use)))
3802 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3803 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3804 if (!Set.count(dyn_cast<Instruction>(*Use)))
3809 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3810 ReductionKind Kind) {
3811 if (Phi->getNumIncomingValues() != 2)
3814 // Reduction variables are only found in the loop header block.
3815 if (Phi->getParent() != TheLoop->getHeader())
3818 // Obtain the reduction start value from the value that comes from the loop
3820 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3822 // ExitInstruction is the single value which is used outside the loop.
3823 // We only allow for a single reduction value to be used outside the loop.
3824 // This includes users of the reduction, variables (which form a cycle
3825 // which ends in the phi node).
3826 Instruction *ExitInstruction = 0;
3827 // Indicates that we found a reduction operation in our scan.
3828 bool FoundReduxOp = false;
3830 // We start with the PHI node and scan for all of the users of this
3831 // instruction. All users must be instructions that can be used as reduction
3832 // variables (such as ADD). We must have a single out-of-block user. The cycle
3833 // must include the original PHI.
3834 bool FoundStartPHI = false;
3836 // To recognize min/max patterns formed by a icmp select sequence, we store
3837 // the number of instruction we saw from the recognized min/max pattern,
3838 // to make sure we only see exactly the two instructions.
3839 unsigned NumCmpSelectPatternInst = 0;
3840 ReductionInstDesc ReduxDesc(false, 0);
3842 SmallPtrSet<Instruction *, 8> VisitedInsts;
3843 SmallVector<Instruction *, 8> Worklist;
3844 Worklist.push_back(Phi);
3845 VisitedInsts.insert(Phi);
3847 // A value in the reduction can be used:
3848 // - By the reduction:
3849 // - Reduction operation:
3850 // - One use of reduction value (safe).
3851 // - Multiple use of reduction value (not safe).
3853 // - All uses of the PHI must be the reduction (safe).
3854 // - Otherwise, not safe.
3855 // - By one instruction outside of the loop (safe).
3856 // - By further instructions outside of the loop (not safe).
3857 // - By an instruction that is not part of the reduction (not safe).
3859 // * An instruction type other than PHI or the reduction operation.
3860 // * A PHI in the header other than the initial PHI.
3861 while (!Worklist.empty()) {
3862 Instruction *Cur = Worklist.back();
3863 Worklist.pop_back();
3866 // If the instruction has no users then this is a broken chain and can't be
3867 // a reduction variable.
3868 if (Cur->use_empty())
3871 bool IsAPhi = isa<PHINode>(Cur);
3873 // A header PHI use other than the original PHI.
3874 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3877 // Reductions of instructions such as Div, and Sub is only possible if the
3878 // LHS is the reduction variable.
3879 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3880 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3881 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3884 // Any reduction instruction must be of one of the allowed kinds.
3885 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3886 if (!ReduxDesc.IsReduction)
3889 // A reduction operation must only have one use of the reduction value.
3890 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3891 hasMultipleUsesOf(Cur, VisitedInsts))
3894 // All inputs to a PHI node must be a reduction value.
3895 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3898 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3899 isa<SelectInst>(Cur)))
3900 ++NumCmpSelectPatternInst;
3901 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3902 isa<SelectInst>(Cur)))
3903 ++NumCmpSelectPatternInst;
3905 // Check whether we found a reduction operator.
3906 FoundReduxOp |= !IsAPhi;
3908 // Process users of current instruction. Push non PHI nodes after PHI nodes
3909 // onto the stack. This way we are going to have seen all inputs to PHI
3910 // nodes once we get to them.
3911 SmallVector<Instruction *, 8> NonPHIs;
3912 SmallVector<Instruction *, 8> PHIs;
3913 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3915 Instruction *Usr = cast<Instruction>(*UI);
3917 // Check if we found the exit user.
3918 BasicBlock *Parent = Usr->getParent();
3919 if (!TheLoop->contains(Parent)) {
3920 // Exit if you find multiple outside users or if the header phi node is
3921 // being used. In this case the user uses the value of the previous
3922 // iteration, in which case we would loose "VF-1" iterations of the
3923 // reduction operation if we vectorize.
3924 if (ExitInstruction != 0 || Cur == Phi)
3927 ExitInstruction = Cur;
3931 // Process instructions only once (termination).
3932 if (VisitedInsts.insert(Usr)) {
3933 if (isa<PHINode>(Usr))
3934 PHIs.push_back(Usr);
3936 NonPHIs.push_back(Usr);
3938 // Remember that we completed the cycle.
3940 FoundStartPHI = true;
3942 Worklist.append(PHIs.begin(), PHIs.end());
3943 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3946 // This means we have seen one but not the other instruction of the
3947 // pattern or more than just a select and cmp.
3948 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3949 NumCmpSelectPatternInst != 2)
3952 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3955 // We found a reduction var if we have reached the original phi node and we
3956 // only have a single instruction with out-of-loop users.
3958 // This instruction is allowed to have out-of-loop users.
3959 AllowedExit.insert(ExitInstruction);
3961 // Save the description of this reduction variable.
3962 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3963 ReduxDesc.MinMaxKind);
3964 Reductions[Phi] = RD;
3965 // We've ended the cycle. This is a reduction variable if we have an
3966 // outside user and it has a binary op.
3971 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3972 /// pattern corresponding to a min(X, Y) or max(X, Y).
3973 LoopVectorizationLegality::ReductionInstDesc
3974 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3975 ReductionInstDesc &Prev) {
3977 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3978 "Expect a select instruction");
3979 Instruction *Cmp = 0;
3980 SelectInst *Select = 0;
3982 // We must handle the select(cmp()) as a single instruction. Advance to the
3984 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3985 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3986 return ReductionInstDesc(false, I);
3987 return ReductionInstDesc(Select, Prev.MinMaxKind);
3990 // Only handle single use cases for now.
3991 if (!(Select = dyn_cast<SelectInst>(I)))
3992 return ReductionInstDesc(false, I);
3993 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3994 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3995 return ReductionInstDesc(false, I);
3996 if (!Cmp->hasOneUse())
3997 return ReductionInstDesc(false, I);
4002 // Look for a min/max pattern.
4003 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4004 return ReductionInstDesc(Select, MRK_UIntMin);
4005 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4006 return ReductionInstDesc(Select, MRK_UIntMax);
4007 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4008 return ReductionInstDesc(Select, MRK_SIntMax);
4009 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4010 return ReductionInstDesc(Select, MRK_SIntMin);
4011 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4012 return ReductionInstDesc(Select, MRK_FloatMin);
4013 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4014 return ReductionInstDesc(Select, MRK_FloatMax);
4015 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4016 return ReductionInstDesc(Select, MRK_FloatMin);
4017 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4018 return ReductionInstDesc(Select, MRK_FloatMax);
4020 return ReductionInstDesc(false, I);
4023 LoopVectorizationLegality::ReductionInstDesc
4024 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4026 ReductionInstDesc &Prev) {
4027 bool FP = I->getType()->isFloatingPointTy();
4028 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4029 switch (I->getOpcode()) {
4031 return ReductionInstDesc(false, I);
4032 case Instruction::PHI:
4033 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4034 Kind != RK_FloatMinMax))
4035 return ReductionInstDesc(false, I);
4036 return ReductionInstDesc(I, Prev.MinMaxKind);
4037 case Instruction::Sub:
4038 case Instruction::Add:
4039 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4040 case Instruction::Mul:
4041 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4042 case Instruction::And:
4043 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4044 case Instruction::Or:
4045 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4046 case Instruction::Xor:
4047 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4048 case Instruction::FMul:
4049 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4050 case Instruction::FAdd:
4051 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4052 case Instruction::FCmp:
4053 case Instruction::ICmp:
4054 case Instruction::Select:
4055 if (Kind != RK_IntegerMinMax &&
4056 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4057 return ReductionInstDesc(false, I);
4058 return isMinMaxSelectCmpPattern(I, Prev);
4062 LoopVectorizationLegality::InductionKind
4063 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4064 Type *PhiTy = Phi->getType();
4065 // We only handle integer and pointer inductions variables.
4066 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4067 return IK_NoInduction;
4069 // Check that the PHI is consecutive.
4070 const SCEV *PhiScev = SE->getSCEV(Phi);
4071 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4073 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4074 return IK_NoInduction;
4076 const SCEV *Step = AR->getStepRecurrence(*SE);
4078 // Integer inductions need to have a stride of one.
4079 if (PhiTy->isIntegerTy()) {
4081 return IK_IntInduction;
4082 if (Step->isAllOnesValue())
4083 return IK_ReverseIntInduction;
4084 return IK_NoInduction;
4087 // Calculate the pointer stride and check if it is consecutive.
4088 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4090 return IK_NoInduction;
4092 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4093 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4094 if (C->getValue()->equalsInt(Size))
4095 return IK_PtrInduction;
4096 else if (C->getValue()->equalsInt(0 - Size))
4097 return IK_ReversePtrInduction;
4099 return IK_NoInduction;
4102 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4103 Value *In0 = const_cast<Value*>(V);
4104 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4108 return Inductions.count(PN);
4111 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4112 assert(TheLoop->contains(BB) && "Unknown block used");
4114 // Blocks that do not dominate the latch need predication.
4115 BasicBlock* Latch = TheLoop->getLoopLatch();
4116 return !DT->dominates(BB, Latch);
4119 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4120 SmallPtrSet<Value *, 8>& SafePtrs) {
4121 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4122 // We might be able to hoist the load.
4123 if (it->mayReadFromMemory()) {
4124 LoadInst *LI = dyn_cast<LoadInst>(it);
4125 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4129 // We don't predicate stores at the moment.
4130 if (it->mayWriteToMemory() || it->mayThrow())
4133 // The instructions below can trap.
4134 switch (it->getOpcode()) {
4136 case Instruction::UDiv:
4137 case Instruction::SDiv:
4138 case Instruction::URem:
4139 case Instruction::SRem:
4147 LoopVectorizationCostModel::VectorizationFactor
4148 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4150 // Width 1 means no vectorize
4151 VectorizationFactor Factor = { 1U, 0U };
4152 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4153 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4157 // Find the trip count.
4158 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4159 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4161 unsigned WidestType = getWidestType();
4162 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4163 unsigned MaxSafeDepDist = -1U;
4164 if (Legal->getMaxSafeDepDistBytes() != -1U)
4165 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4166 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4167 WidestRegister : MaxSafeDepDist);
4168 unsigned MaxVectorSize = WidestRegister / WidestType;
4169 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4170 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4172 if (MaxVectorSize == 0) {
4173 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4177 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4178 " into one vector!");
4180 unsigned VF = MaxVectorSize;
4182 // If we optimize the program for size, avoid creating the tail loop.
4184 // If we are unable to calculate the trip count then don't try to vectorize.
4186 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4190 // Find the maximum SIMD width that can fit within the trip count.
4191 VF = TC % MaxVectorSize;
4196 // If the trip count that we found modulo the vectorization factor is not
4197 // zero then we require a tail.
4199 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4205 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4206 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4208 Factor.Width = UserVF;
4212 float Cost = expectedCost(1);
4214 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4215 for (unsigned i=2; i <= VF; i*=2) {
4216 // Notice that the vector loop needs to be executed less times, so
4217 // we need to divide the cost of the vector loops by the width of
4218 // the vector elements.
4219 float VectorCost = expectedCost(i) / (float)i;
4220 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4221 (int)VectorCost << ".\n");
4222 if (VectorCost < Cost) {
4228 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4229 Factor.Width = Width;
4230 Factor.Cost = Width * Cost;
4234 unsigned LoopVectorizationCostModel::getWidestType() {
4235 unsigned MaxWidth = 8;
4238 for (Loop::block_iterator bb = TheLoop->block_begin(),
4239 be = TheLoop->block_end(); bb != be; ++bb) {
4240 BasicBlock *BB = *bb;
4242 // For each instruction in the loop.
4243 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4244 Type *T = it->getType();
4246 // Only examine Loads, Stores and PHINodes.
4247 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4250 // Examine PHI nodes that are reduction variables.
4251 if (PHINode *PN = dyn_cast<PHINode>(it))
4252 if (!Legal->getReductionVars()->count(PN))
4255 // Examine the stored values.
4256 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4257 T = ST->getValueOperand()->getType();
4259 // Ignore loaded pointer types and stored pointer types that are not
4260 // consecutive. However, we do want to take consecutive stores/loads of
4261 // pointer vectors into account.
4262 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4265 MaxWidth = std::max(MaxWidth,
4266 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4274 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4277 unsigned LoopCost) {
4279 // -- The unroll heuristics --
4280 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4281 // There are many micro-architectural considerations that we can't predict
4282 // at this level. For example frontend pressure (on decode or fetch) due to
4283 // code size, or the number and capabilities of the execution ports.
4285 // We use the following heuristics to select the unroll factor:
4286 // 1. If the code has reductions the we unroll in order to break the cross
4287 // iteration dependency.
4288 // 2. If the loop is really small then we unroll in order to reduce the loop
4290 // 3. We don't unroll if we think that we will spill registers to memory due
4291 // to the increased register pressure.
4293 // Use the user preference, unless 'auto' is selected.
4297 // When we optimize for size we don't unroll.
4301 // We used the distance for the unroll factor.
4302 if (Legal->getMaxSafeDepDistBytes() != -1U)
4305 // Do not unroll loops with a relatively small trip count.
4306 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4307 TheLoop->getLoopLatch());
4308 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4311 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4312 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4313 " vector registers\n");
4315 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4316 // We divide by these constants so assume that we have at least one
4317 // instruction that uses at least one register.
4318 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4319 R.NumInstructions = std::max(R.NumInstructions, 1U);
4321 // We calculate the unroll factor using the following formula.
4322 // Subtract the number of loop invariants from the number of available
4323 // registers. These registers are used by all of the unrolled instances.
4324 // Next, divide the remaining registers by the number of registers that is
4325 // required by the loop, in order to estimate how many parallel instances
4326 // fit without causing spills.
4327 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4329 // Clamp the unroll factor ranges to reasonable factors.
4330 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4332 // If we did not calculate the cost for VF (because the user selected the VF)
4333 // then we calculate the cost of VF here.
4335 LoopCost = expectedCost(VF);
4337 // Clamp the calculated UF to be between the 1 and the max unroll factor
4338 // that the target allows.
4339 if (UF > MaxUnrollSize)
4344 bool HasReductions = Legal->getReductionVars()->size();
4346 // Decide if we want to unroll if we decided that it is legal to vectorize
4347 // but not profitable.
4349 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4350 LoopCost > SmallLoopCost)
4356 if (HasReductions) {
4357 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4361 // We want to unroll tiny loops in order to reduce the loop overhead.
4362 // We assume that the cost overhead is 1 and we use the cost model
4363 // to estimate the cost of the loop and unroll until the cost of the
4364 // loop overhead is about 5% of the cost of the loop.
4365 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4366 if (LoopCost < SmallLoopCost) {
4367 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4368 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4369 return std::min(NewUF, UF);
4372 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4376 LoopVectorizationCostModel::RegisterUsage
4377 LoopVectorizationCostModel::calculateRegisterUsage() {
4378 // This function calculates the register usage by measuring the highest number
4379 // of values that are alive at a single location. Obviously, this is a very
4380 // rough estimation. We scan the loop in a topological order in order and
4381 // assign a number to each instruction. We use RPO to ensure that defs are
4382 // met before their users. We assume that each instruction that has in-loop
4383 // users starts an interval. We record every time that an in-loop value is
4384 // used, so we have a list of the first and last occurrences of each
4385 // instruction. Next, we transpose this data structure into a multi map that
4386 // holds the list of intervals that *end* at a specific location. This multi
4387 // map allows us to perform a linear search. We scan the instructions linearly
4388 // and record each time that a new interval starts, by placing it in a set.
4389 // If we find this value in the multi-map then we remove it from the set.
4390 // The max register usage is the maximum size of the set.
4391 // We also search for instructions that are defined outside the loop, but are
4392 // used inside the loop. We need this number separately from the max-interval
4393 // usage number because when we unroll, loop-invariant values do not take
4395 LoopBlocksDFS DFS(TheLoop);
4399 R.NumInstructions = 0;
4401 // Each 'key' in the map opens a new interval. The values
4402 // of the map are the index of the 'last seen' usage of the
4403 // instruction that is the key.
4404 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4405 // Maps instruction to its index.
4406 DenseMap<unsigned, Instruction*> IdxToInstr;
4407 // Marks the end of each interval.
4408 IntervalMap EndPoint;
4409 // Saves the list of instruction indices that are used in the loop.
4410 SmallSet<Instruction*, 8> Ends;
4411 // Saves the list of values that are used in the loop but are
4412 // defined outside the loop, such as arguments and constants.
4413 SmallPtrSet<Value*, 8> LoopInvariants;
4416 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4417 be = DFS.endRPO(); bb != be; ++bb) {
4418 R.NumInstructions += (*bb)->size();
4419 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4421 Instruction *I = it;
4422 IdxToInstr[Index++] = I;
4424 // Save the end location of each USE.
4425 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4426 Value *U = I->getOperand(i);
4427 Instruction *Instr = dyn_cast<Instruction>(U);
4429 // Ignore non-instruction values such as arguments, constants, etc.
4430 if (!Instr) continue;
4432 // If this instruction is outside the loop then record it and continue.
4433 if (!TheLoop->contains(Instr)) {
4434 LoopInvariants.insert(Instr);
4438 // Overwrite previous end points.
4439 EndPoint[Instr] = Index;
4445 // Saves the list of intervals that end with the index in 'key'.
4446 typedef SmallVector<Instruction*, 2> InstrList;
4447 DenseMap<unsigned, InstrList> TransposeEnds;
4449 // Transpose the EndPoints to a list of values that end at each index.
4450 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4452 TransposeEnds[it->second].push_back(it->first);
4454 SmallSet<Instruction*, 8> OpenIntervals;
4455 unsigned MaxUsage = 0;
4458 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4459 for (unsigned int i = 0; i < Index; ++i) {
4460 Instruction *I = IdxToInstr[i];
4461 // Ignore instructions that are never used within the loop.
4462 if (!Ends.count(I)) continue;
4464 // Remove all of the instructions that end at this location.
4465 InstrList &List = TransposeEnds[i];
4466 for (unsigned int j=0, e = List.size(); j < e; ++j)
4467 OpenIntervals.erase(List[j]);
4469 // Count the number of live interals.
4470 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4472 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4473 OpenIntervals.size() <<"\n");
4475 // Add the current instruction to the list of open intervals.
4476 OpenIntervals.insert(I);
4479 unsigned Invariant = LoopInvariants.size();
4480 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4481 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4482 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4484 R.LoopInvariantRegs = Invariant;
4485 R.MaxLocalUsers = MaxUsage;
4489 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4493 for (Loop::block_iterator bb = TheLoop->block_begin(),
4494 be = TheLoop->block_end(); bb != be; ++bb) {
4495 unsigned BlockCost = 0;
4496 BasicBlock *BB = *bb;
4498 // For each instruction in the old loop.
4499 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4500 // Skip dbg intrinsics.
4501 if (isa<DbgInfoIntrinsic>(it))
4504 unsigned C = getInstructionCost(it, VF);
4506 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4507 VF << " For instruction: "<< *it << "\n");
4510 // We assume that if-converted blocks have a 50% chance of being executed.
4511 // When the code is scalar then some of the blocks are avoided due to CF.
4512 // When the code is vectorized we execute all code paths.
4513 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4522 /// \brief Check whether the address computation for a non-consecutive memory
4523 /// access looks like an unlikely candidate for being merged into the indexing
4526 /// We look for a GEP which has one index that is an induction variable and all
4527 /// other indices are loop invariant. If the stride of this access is also
4528 /// within a small bound we decide that this address computation can likely be
4529 /// merged into the addressing mode.
4530 /// In all other cases, we identify the address computation as complex.
4531 static bool isLikelyComplexAddressComputation(Value *Ptr,
4532 LoopVectorizationLegality *Legal,
4533 ScalarEvolution *SE,
4534 const Loop *TheLoop) {
4535 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4539 // We are looking for a gep with all loop invariant indices except for one
4540 // which should be an induction variable.
4541 unsigned NumOperands = Gep->getNumOperands();
4542 for (unsigned i = 1; i < NumOperands; ++i) {
4543 Value *Opd = Gep->getOperand(i);
4544 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4545 !Legal->isInductionVariable(Opd))
4549 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4550 // can likely be merged into the address computation.
4551 unsigned MaxMergeDistance = 64;
4553 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4557 // Check the step is constant.
4558 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4559 // Calculate the pointer stride and check if it is consecutive.
4560 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4564 const APInt &APStepVal = C->getValue()->getValue();
4566 // Huge step value - give up.
4567 if (APStepVal.getBitWidth() > 64)
4570 int64_t StepVal = APStepVal.getSExtValue();
4572 return StepVal > MaxMergeDistance;
4576 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4577 // If we know that this instruction will remain uniform, check the cost of
4578 // the scalar version.
4579 if (Legal->isUniformAfterVectorization(I))
4582 Type *RetTy = I->getType();
4583 Type *VectorTy = ToVectorTy(RetTy, VF);
4585 // TODO: We need to estimate the cost of intrinsic calls.
4586 switch (I->getOpcode()) {
4587 case Instruction::GetElementPtr:
4588 // We mark this instruction as zero-cost because the cost of GEPs in
4589 // vectorized code depends on whether the corresponding memory instruction
4590 // is scalarized or not. Therefore, we handle GEPs with the memory
4591 // instruction cost.
4593 case Instruction::Br: {
4594 return TTI.getCFInstrCost(I->getOpcode());
4596 case Instruction::PHI:
4597 //TODO: IF-converted IFs become selects.
4599 case Instruction::Add:
4600 case Instruction::FAdd:
4601 case Instruction::Sub:
4602 case Instruction::FSub:
4603 case Instruction::Mul:
4604 case Instruction::FMul:
4605 case Instruction::UDiv:
4606 case Instruction::SDiv:
4607 case Instruction::FDiv:
4608 case Instruction::URem:
4609 case Instruction::SRem:
4610 case Instruction::FRem:
4611 case Instruction::Shl:
4612 case Instruction::LShr:
4613 case Instruction::AShr:
4614 case Instruction::And:
4615 case Instruction::Or:
4616 case Instruction::Xor: {
4617 // Certain instructions can be cheaper to vectorize if they have a constant
4618 // second vector operand. One example of this are shifts on x86.
4619 TargetTransformInfo::OperandValueKind Op1VK =
4620 TargetTransformInfo::OK_AnyValue;
4621 TargetTransformInfo::OperandValueKind Op2VK =
4622 TargetTransformInfo::OK_AnyValue;
4624 if (isa<ConstantInt>(I->getOperand(1)))
4625 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4627 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4629 case Instruction::Select: {
4630 SelectInst *SI = cast<SelectInst>(I);
4631 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4632 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4633 Type *CondTy = SI->getCondition()->getType();
4635 CondTy = VectorType::get(CondTy, VF);
4637 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4639 case Instruction::ICmp:
4640 case Instruction::FCmp: {
4641 Type *ValTy = I->getOperand(0)->getType();
4642 VectorTy = ToVectorTy(ValTy, VF);
4643 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4645 case Instruction::Store:
4646 case Instruction::Load: {
4647 StoreInst *SI = dyn_cast<StoreInst>(I);
4648 LoadInst *LI = dyn_cast<LoadInst>(I);
4649 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4651 VectorTy = ToVectorTy(ValTy, VF);
4653 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4654 unsigned AS = SI ? SI->getPointerAddressSpace() :
4655 LI->getPointerAddressSpace();
4656 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4657 // We add the cost of address computation here instead of with the gep
4658 // instruction because only here we know whether the operation is
4661 return TTI.getAddressComputationCost(VectorTy) +
4662 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4664 // Scalarized loads/stores.
4665 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4666 bool Reverse = ConsecutiveStride < 0;
4667 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4668 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4669 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4670 bool IsComplexComputation =
4671 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4673 // The cost of extracting from the value vector and pointer vector.
4674 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4675 for (unsigned i = 0; i < VF; ++i) {
4676 // The cost of extracting the pointer operand.
4677 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4678 // In case of STORE, the cost of ExtractElement from the vector.
4679 // In case of LOAD, the cost of InsertElement into the returned
4681 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4682 Instruction::InsertElement,
4686 // The cost of the scalar loads/stores.
4687 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4688 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4693 // Wide load/stores.
4694 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4695 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4698 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4702 case Instruction::ZExt:
4703 case Instruction::SExt:
4704 case Instruction::FPToUI:
4705 case Instruction::FPToSI:
4706 case Instruction::FPExt:
4707 case Instruction::PtrToInt:
4708 case Instruction::IntToPtr:
4709 case Instruction::SIToFP:
4710 case Instruction::UIToFP:
4711 case Instruction::Trunc:
4712 case Instruction::FPTrunc:
4713 case Instruction::BitCast: {
4714 // We optimize the truncation of induction variable.
4715 // The cost of these is the same as the scalar operation.
4716 if (I->getOpcode() == Instruction::Trunc &&
4717 Legal->isInductionVariable(I->getOperand(0)))
4718 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4719 I->getOperand(0)->getType());
4721 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4722 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4724 case Instruction::Call: {
4725 CallInst *CI = cast<CallInst>(I);
4726 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4727 assert(ID && "Not an intrinsic call!");
4728 Type *RetTy = ToVectorTy(CI->getType(), VF);
4729 SmallVector<Type*, 4> Tys;
4730 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4731 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4732 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4735 // We are scalarizing the instruction. Return the cost of the scalar
4736 // instruction, plus the cost of insert and extract into vector
4737 // elements, times the vector width.
4740 if (!RetTy->isVoidTy() && VF != 1) {
4741 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4743 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4746 // The cost of inserting the results plus extracting each one of the
4748 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4751 // The cost of executing VF copies of the scalar instruction. This opcode
4752 // is unknown. Assume that it is the same as 'mul'.
4753 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4759 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4760 if (Scalar->isVoidTy() || VF == 1)
4762 return VectorType::get(Scalar, VF);
4765 char LoopVectorize::ID = 0;
4766 static const char lv_name[] = "Loop Vectorization";
4767 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4768 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4769 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4770 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4771 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4774 Pass *createLoopVectorizePass() {
4775 return new LoopVectorize();
4779 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4780 // Check for a store.
4781 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4782 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4784 // Check for a load.
4785 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4786 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
4792 InnerLoopUnroller::vectorizeLoop(LoopVectorizationLegality *Legal) {
4793 // In order to support reduction variables we need to be able to unroll
4794 // Phi nodes. Phi nodes have cycles, so we need to unroll them in two
4795 // stages. See InnerLoopVectorizer::vectorizeLoop for more details.
4796 PhiVector RdxPHIsToFix;
4798 // Scan the loop in a topological order to ensure that defs are vectorized
4800 LoopBlocksDFS DFS(OrigLoop);
4803 // Unroll all of the blocks in the original loop.
4804 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(), be = DFS.endRPO();
4806 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
4808 // Create the 'reduced' values for each of the induction vars.
4809 // The reduced values are the vector values that we scalarize and combine
4810 // after the loop is finished.
4811 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
4813 PHINode *RdxPhi = *it;
4814 assert(RdxPhi && "Unable to recover vectorized PHI");
4816 // Find the reduction variable descriptor.
4817 assert(Legal->getReductionVars()->count(RdxPhi) &&
4818 "Unable to find the reduction variable");
4819 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
4820 (*Legal->getReductionVars())[RdxPhi];
4822 setDebugLocFromInst(Builder, RdxDesc.StartValue);
4824 // We need to generate a reduction vector from the incoming scalar.
4825 // To do so, we need to generate the 'identity' vector and overide
4826 // one of the elements with the incoming scalar reduction. We need
4827 // to do it in the vector-loop preheader.
4828 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
4830 // This is the vector-clone of the value that leaves the loop.
4831 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
4832 Type *VecTy = VectorExit[0]->getType();
4834 // Find the reduction identity variable. Zero for addition, or, xor,
4835 // one for multiplication, -1 for And.
4838 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
4839 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
4840 // MinMax reduction have the start value as their identify.
4841 VectorStart = Identity = RdxDesc.StartValue;
4844 Identity = LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
4845 VecTy->getScalarType());
4847 // This vector is the Identity vector where the first element is the
4848 // incoming scalar reduction.
4849 VectorStart = RdxDesc.StartValue;
4852 // Fix the vector-loop phi.
4853 // We created the induction variable so we know that the
4854 // preheader is the first entry.
4855 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
4857 // Reductions do not have to start at zero. They can start with
4858 // any loop invariant values.
4859 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
4860 BasicBlock *Latch = OrigLoop->getLoopLatch();
4861 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
4862 VectorParts &Val = getVectorValue(LoopVal);
4863 for (unsigned part = 0; part < UF; ++part) {
4864 // Make sure to add the reduction stat value only to the
4865 // first unroll part.
4866 Value *StartVal = (part == 0) ? VectorStart : Identity;
4867 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
4868 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
4871 // Before each round, move the insertion point right between
4872 // the PHIs and the values we are going to write.
4873 // This allows us to write both PHINodes and the extractelement
4875 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
4877 VectorParts RdxParts;
4878 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
4879 for (unsigned part = 0; part < UF; ++part) {
4880 // This PHINode contains the vectorized reduction variable, or
4881 // the initial value vector, if we bypass the vector loop.
4882 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
4883 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
4884 Value *StartVal = (part == 0) ? VectorStart : Identity;
4885 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4886 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
4887 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
4888 RdxParts.push_back(NewPhi);
4891 // Reduce all of the unrolled parts into a single vector.
4892 Value *ReducedPartRdx = RdxParts[0];
4893 unsigned Op = getReductionBinOp(RdxDesc.Kind);
4894 setDebugLocFromInst(Builder, ReducedPartRdx);
4895 for (unsigned part = 1; part < UF; ++part) {
4896 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4897 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
4898 RdxParts[part], ReducedPartRdx,
4901 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
4902 ReducedPartRdx, RdxParts[part]);
4905 // Now, we need to fix the users of the reduction variable
4906 // inside and outside of the scalar remainder loop.
4907 // We know that the loop is in LCSSA form. We need to update the
4908 // PHI nodes in the exit blocks.
4909 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4910 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
4911 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4912 if (!LCSSAPhi) continue;
4914 // All PHINodes need to have a single entry edge, or two if
4915 // we already fixed them.
4916 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
4918 // We found our reduction value exit-PHI. Update it with the
4919 // incoming bypass edge.
4920 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
4921 // Add an edge coming from the bypass.
4922 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4925 }// end of the LCSSA phi scan.
4927 // Fix the scalar loop reduction variable with the incoming reduction sum
4928 // from the vector body and from the backedge value.
4929 int IncomingEdgeBlockIdx =
4930 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
4931 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
4932 // Pick the other block.
4933 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4934 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
4935 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
4936 }// end of for each redux variable.
4941 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
4942 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
4943 // Holds vector parameters or scalars, in case of uniform vals.
4944 SmallVector<VectorParts, 4> Params;
4946 setDebugLocFromInst(Builder, Instr);
4948 // Find all of the vectorized parameters.
4949 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
4950 Value *SrcOp = Instr->getOperand(op);
4952 // If we are accessing the old induction variable, use the new one.
4953 if (SrcOp == OldInduction) {
4954 Params.push_back(getVectorValue(SrcOp));
4958 // Try using previously calculated values.
4959 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
4961 // If the src is an instruction that appeared earlier in the basic block
4962 // then it should already be vectorized.
4963 if (SrcInst && OrigLoop->contains(SrcInst)) {
4964 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
4965 // The parameter is a vector value from earlier.
4966 Params.push_back(WidenMap.get(SrcInst));
4968 // The parameter is a scalar from outside the loop. Maybe even a constant.
4969 VectorParts Scalars;
4970 Scalars.append(UF, SrcOp);
4971 Params.push_back(Scalars);
4975 assert(Params.size() == Instr->getNumOperands() &&
4976 "Invalid number of operands");
4978 // Does this instruction return a value ?
4979 bool IsVoidRetTy = Instr->getType()->isVoidTy();
4981 Value *UndefVec = IsVoidRetTy ? 0 :
4982 UndefValue::get(Instr->getType());
4983 // Create a new entry in the WidenMap and initialize it to Undef or Null.
4984 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
4986 // For each vector unroll 'part':
4987 for (unsigned Part = 0; Part < UF; ++Part) {
4988 // For each scalar that we create:
4990 Instruction *Cloned = Instr->clone();
4992 Cloned->setName(Instr->getName() + ".cloned");
4993 // Replace the operands of the cloned instrucions with extracted scalars.
4994 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
4995 Value *Op = Params[op][Part];
4996 Cloned->setOperand(op, Op);
4999 // Place the cloned scalar in the new loop.
5000 Builder.Insert(Cloned);
5002 // If the original scalar returns a value we need to place it in a vector
5003 // so that future users will be able to use it.
5005 VecResults[Part] = Cloned;
5010 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
5011 LoopVectorizationLegality*) {
5012 return scalarizeInstruction(Instr);
5015 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5019 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5023 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5025 // When unrolling and the VF is 1, we only need to add a simple scalar.
5026 Type *ITy = Val->getType();
5027 assert(!ITy->isVectorTy() && "Val must be a scalar");
5028 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5029 return Builder.CreateAdd(Val, C, "induction");