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 scalarizeInstruction(Instruction *Instr);
358 virtual void vectorizeMemoryInstruction(Instruction *Instr,
359 LoopVectorizationLegality *Legal);
360 virtual Value *getBroadcastInstrs(Value *V);
361 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
362 virtual Value *reverseVector(Value *Vec);
365 /// \brief Look for a meaningful debug location on the instruction or it's
367 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
372 if (I->getDebugLoc() != Empty)
375 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
376 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
377 if (OpInst->getDebugLoc() != Empty)
384 /// \brief Set the debug location in the builder using the debug location in the
386 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
387 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
388 B.SetCurrentDebugLocation(Inst->getDebugLoc());
390 B.SetCurrentDebugLocation(DebugLoc());
393 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
394 /// to what vectorization factor.
395 /// This class does not look at the profitability of vectorization, only the
396 /// legality. This class has two main kinds of checks:
397 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
398 /// will change the order of memory accesses in a way that will change the
399 /// correctness of the program.
400 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
401 /// checks for a number of different conditions, such as the availability of a
402 /// single induction variable, that all types are supported and vectorize-able,
403 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
404 /// This class is also used by InnerLoopVectorizer for identifying
405 /// induction variable and the different reduction variables.
406 class LoopVectorizationLegality {
408 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
409 DominatorTree *DT, TargetLibraryInfo *TLI)
410 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
411 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
412 MaxSafeDepDistBytes(-1U) {}
414 /// This enum represents the kinds of reductions that we support.
416 RK_NoReduction, ///< Not a reduction.
417 RK_IntegerAdd, ///< Sum of integers.
418 RK_IntegerMult, ///< Product of integers.
419 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
420 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
421 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
422 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
423 RK_FloatAdd, ///< Sum of floats.
424 RK_FloatMult, ///< Product of floats.
425 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
428 /// This enum represents the kinds of inductions that we support.
430 IK_NoInduction, ///< Not an induction variable.
431 IK_IntInduction, ///< Integer induction variable. Step = 1.
432 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
433 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
434 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
437 // This enum represents the kind of minmax reduction.
438 enum MinMaxReductionKind {
448 /// This struct holds information about reduction variables.
449 struct ReductionDescriptor {
450 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
451 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
453 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
454 MinMaxReductionKind MK)
455 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
457 // The starting value of the reduction.
458 // It does not have to be zero!
459 TrackingVH<Value> StartValue;
460 // The instruction who's value is used outside the loop.
461 Instruction *LoopExitInstr;
462 // The kind of the reduction.
464 // If this a min/max reduction the kind of reduction.
465 MinMaxReductionKind MinMaxKind;
468 /// This POD struct holds information about a potential reduction operation.
469 struct ReductionInstDesc {
470 ReductionInstDesc(bool IsRedux, Instruction *I) :
471 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
473 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
474 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
476 // Is this instruction a reduction candidate.
478 // The last instruction in a min/max pattern (select of the select(icmp())
479 // pattern), or the current reduction instruction otherwise.
480 Instruction *PatternLastInst;
481 // If this is a min/max pattern the comparison predicate.
482 MinMaxReductionKind MinMaxKind;
485 /// This struct holds information about the memory runtime legality
486 /// check that a group of pointers do not overlap.
487 struct RuntimePointerCheck {
488 RuntimePointerCheck() : Need(false) {}
490 /// Reset the state of the pointer runtime information.
497 DependencySetId.clear();
500 /// Insert a pointer and calculate the start and end SCEVs.
501 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
504 /// This flag indicates if we need to add the runtime check.
506 /// Holds the pointers that we need to check.
507 SmallVector<TrackingVH<Value>, 2> Pointers;
508 /// Holds the pointer value at the beginning of the loop.
509 SmallVector<const SCEV*, 2> Starts;
510 /// Holds the pointer value at the end of the loop.
511 SmallVector<const SCEV*, 2> Ends;
512 /// Holds the information if this pointer is used for writing to memory.
513 SmallVector<bool, 2> IsWritePtr;
514 /// Holds the id of the set of pointers that could be dependent because of a
515 /// shared underlying object.
516 SmallVector<unsigned, 2> DependencySetId;
519 /// A struct for saving information about induction variables.
520 struct InductionInfo {
521 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
522 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
524 TrackingVH<Value> StartValue;
529 /// ReductionList contains the reduction descriptors for all
530 /// of the reductions that were found in the loop.
531 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
533 /// InductionList saves induction variables and maps them to the
534 /// induction descriptor.
535 typedef MapVector<PHINode*, InductionInfo> InductionList;
537 /// Returns true if it is legal to vectorize this loop.
538 /// This does not mean that it is profitable to vectorize this
539 /// loop, only that it is legal to do so.
542 /// Returns the Induction variable.
543 PHINode *getInduction() { return Induction; }
545 /// Returns the reduction variables found in the loop.
546 ReductionList *getReductionVars() { return &Reductions; }
548 /// Returns the induction variables found in the loop.
549 InductionList *getInductionVars() { return &Inductions; }
551 /// Returns the widest induction type.
552 Type *getWidestInductionType() { return WidestIndTy; }
554 /// Returns True if V is an induction variable in this loop.
555 bool isInductionVariable(const Value *V);
557 /// Return true if the block BB needs to be predicated in order for the loop
558 /// to be vectorized.
559 bool blockNeedsPredication(BasicBlock *BB);
561 /// Check if this pointer is consecutive when vectorizing. This happens
562 /// when the last index of the GEP is the induction variable, or that the
563 /// pointer itself is an induction variable.
564 /// This check allows us to vectorize A[idx] into a wide load/store.
566 /// 0 - Stride is unknown or non consecutive.
567 /// 1 - Address is consecutive.
568 /// -1 - Address is consecutive, and decreasing.
569 int isConsecutivePtr(Value *Ptr);
571 /// Returns true if the value V is uniform within the loop.
572 bool isUniform(Value *V);
574 /// Returns true if this instruction will remain scalar after vectorization.
575 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
577 /// Returns the information that we collected about runtime memory check.
578 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
580 /// This function returns the identity element (or neutral element) for
582 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
584 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
587 /// Check if a single basic block loop is vectorizable.
588 /// At this point we know that this is a loop with a constant trip count
589 /// and we only need to check individual instructions.
590 bool canVectorizeInstrs();
592 /// When we vectorize loops we may change the order in which
593 /// we read and write from memory. This method checks if it is
594 /// legal to vectorize the code, considering only memory constrains.
595 /// Returns true if the loop is vectorizable
596 bool canVectorizeMemory();
598 /// Return true if we can vectorize this loop using the IF-conversion
600 bool canVectorizeWithIfConvert();
602 /// Collect the variables that need to stay uniform after vectorization.
603 void collectLoopUniforms();
605 /// Return true if all of the instructions in the block can be speculatively
606 /// executed. \p SafePtrs is a list of addresses that are known to be legal
607 /// and we know that we can read from them without segfault.
608 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
610 /// Returns True, if 'Phi' is the kind of reduction variable for type
611 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
612 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
613 /// Returns a struct describing if the instruction 'I' can be a reduction
614 /// variable of type 'Kind'. If the reduction is a min/max pattern of
615 /// select(icmp()) this function advances the instruction pointer 'I' from the
616 /// compare instruction to the select instruction and stores this pointer in
617 /// 'PatternLastInst' member of the returned struct.
618 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
619 ReductionInstDesc &Desc);
620 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
621 /// pattern corresponding to a min(X, Y) or max(X, Y).
622 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
623 ReductionInstDesc &Prev);
624 /// Returns the induction kind of Phi. This function may return NoInduction
625 /// if the PHI is not an induction variable.
626 InductionKind isInductionVariable(PHINode *Phi);
628 /// The loop that we evaluate.
632 /// DataLayout analysis.
636 /// Target Library Info.
637 TargetLibraryInfo *TLI;
639 // --- vectorization state --- //
641 /// Holds the integer induction variable. This is the counter of the
644 /// Holds the reduction variables.
645 ReductionList Reductions;
646 /// Holds all of the induction variables that we found in the loop.
647 /// Notice that inductions don't need to start at zero and that induction
648 /// variables can be pointers.
649 InductionList Inductions;
650 /// Holds the widest induction type encountered.
653 /// Allowed outside users. This holds the reduction
654 /// vars which can be accessed from outside the loop.
655 SmallPtrSet<Value*, 4> AllowedExit;
656 /// This set holds the variables which are known to be uniform after
658 SmallPtrSet<Instruction*, 4> Uniforms;
659 /// We need to check that all of the pointers in this list are disjoint
661 RuntimePointerCheck PtrRtCheck;
662 /// Can we assume the absence of NaNs.
663 bool HasFunNoNaNAttr;
665 unsigned MaxSafeDepDistBytes;
668 /// LoopVectorizationCostModel - estimates the expected speedups due to
670 /// In many cases vectorization is not profitable. This can happen because of
671 /// a number of reasons. In this class we mainly attempt to predict the
672 /// expected speedup/slowdowns due to the supported instruction set. We use the
673 /// TargetTransformInfo to query the different backends for the cost of
674 /// different operations.
675 class LoopVectorizationCostModel {
677 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
678 LoopVectorizationLegality *Legal,
679 const TargetTransformInfo &TTI,
680 DataLayout *DL, const TargetLibraryInfo *TLI)
681 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
683 /// Information about vectorization costs
684 struct VectorizationFactor {
685 unsigned Width; // Vector width with best cost
686 unsigned Cost; // Cost of the loop with that width
688 /// \return The most profitable vectorization factor and the cost of that VF.
689 /// This method checks every power of two up to VF. If UserVF is not ZERO
690 /// then this vectorization factor will be selected if vectorization is
692 VectorizationFactor selectVectorizationFactor(bool OptForSize,
695 /// \return The size (in bits) of the widest type in the code that
696 /// needs to be vectorized. We ignore values that remain scalar such as
697 /// 64 bit loop indices.
698 unsigned getWidestType();
700 /// \return The most profitable unroll factor.
701 /// If UserUF is non-zero then this method finds the best unroll-factor
702 /// based on register pressure and other parameters.
703 /// VF and LoopCost are the selected vectorization factor and the cost of the
705 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
708 /// \brief A struct that represents some properties of the register usage
710 struct RegisterUsage {
711 /// Holds the number of loop invariant values that are used in the loop.
712 unsigned LoopInvariantRegs;
713 /// Holds the maximum number of concurrent live intervals in the loop.
714 unsigned MaxLocalUsers;
715 /// Holds the number of instructions in the loop.
716 unsigned NumInstructions;
719 /// \return information about the register usage of the loop.
720 RegisterUsage calculateRegisterUsage();
723 /// Returns the expected execution cost. The unit of the cost does
724 /// not matter because we use the 'cost' units to compare different
725 /// vector widths. The cost that is returned is *not* normalized by
726 /// the factor width.
727 unsigned expectedCost(unsigned VF);
729 /// Returns the execution time cost of an instruction for a given vector
730 /// width. Vector width of one means scalar.
731 unsigned getInstructionCost(Instruction *I, unsigned VF);
733 /// A helper function for converting Scalar types to vector types.
734 /// If the incoming type is void, we return void. If the VF is 1, we return
736 static Type* ToVectorTy(Type *Scalar, unsigned VF);
738 /// Returns whether the instruction is a load or store and will be a emitted
739 /// as a vector operation.
740 bool isConsecutiveLoadOrStore(Instruction *I);
742 /// The loop that we evaluate.
746 /// Loop Info analysis.
748 /// Vectorization legality.
749 LoopVectorizationLegality *Legal;
750 /// Vector target information.
751 const TargetTransformInfo &TTI;
752 /// Target data layout information.
754 /// Target Library Info.
755 const TargetLibraryInfo *TLI;
758 /// Utility class for getting and setting loop vectorizer hints in the form
759 /// of loop metadata.
760 struct LoopVectorizeHints {
761 /// Vectorization width.
763 /// Vectorization unroll factor.
766 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
767 : Width(VectorizationFactor)
768 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
769 , LoopID(L->getLoopID()) {
771 // The command line options override any loop metadata except for when
772 // width == 1 which is used to indicate the loop is already vectorized.
773 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
774 Width = VectorizationFactor;
775 if (VectorizationUnroll.getNumOccurrences() > 0)
776 Unroll = VectorizationUnroll;
778 DEBUG(if (DisableUnrolling && Unroll == 1)
779 dbgs() << "LV: Unrolling disabled by the pass manager\n");
782 /// Return the loop vectorizer metadata prefix.
783 static StringRef Prefix() { return "llvm.vectorizer."; }
785 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
786 SmallVector<Value*, 2> Vals;
787 Vals.push_back(MDString::get(Context, Name));
788 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
789 return MDNode::get(Context, Vals);
792 /// Mark the loop L as already vectorized by setting the width to 1.
793 void setAlreadyVectorized(Loop *L) {
794 LLVMContext &Context = L->getHeader()->getContext();
798 // Create a new loop id with one more operand for the already_vectorized
799 // hint. If the loop already has a loop id then copy the existing operands.
800 SmallVector<Value*, 4> Vals(1);
802 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
803 Vals.push_back(LoopID->getOperand(i));
805 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
806 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
808 MDNode *NewLoopID = MDNode::get(Context, Vals);
809 // Set operand 0 to refer to the loop id itself.
810 NewLoopID->replaceOperandWith(0, NewLoopID);
812 L->setLoopID(NewLoopID);
814 LoopID->replaceAllUsesWith(NewLoopID);
822 /// Find hints specified in the loop metadata.
823 void getHints(const Loop *L) {
827 // First operand should refer to the loop id itself.
828 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
829 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
831 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
832 const MDString *S = 0;
833 SmallVector<Value*, 4> Args;
835 // The expected hint is either a MDString or a MDNode with the first
836 // operand a MDString.
837 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
838 if (!MD || MD->getNumOperands() == 0)
840 S = dyn_cast<MDString>(MD->getOperand(0));
841 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
842 Args.push_back(MD->getOperand(i));
844 S = dyn_cast<MDString>(LoopID->getOperand(i));
845 assert(Args.size() == 0 && "too many arguments for MDString");
851 // Check if the hint starts with the vectorizer prefix.
852 StringRef Hint = S->getString();
853 if (!Hint.startswith(Prefix()))
855 // Remove the prefix.
856 Hint = Hint.substr(Prefix().size(), StringRef::npos);
858 if (Args.size() == 1)
859 getHint(Hint, Args[0]);
863 // Check string hint with one operand.
864 void getHint(StringRef Hint, Value *Arg) {
865 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
867 unsigned Val = C->getZExtValue();
869 if (Hint == "width") {
870 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
873 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
874 } else if (Hint == "unroll") {
875 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
878 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
880 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
885 /// The LoopVectorize Pass.
886 struct LoopVectorize : public LoopPass {
887 /// Pass identification, replacement for typeid
890 explicit LoopVectorize(bool NoUnrolling = false)
891 : LoopPass(ID), DisableUnrolling(NoUnrolling) {
892 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
898 TargetTransformInfo *TTI;
900 TargetLibraryInfo *TLI;
901 bool DisableUnrolling;
903 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
904 // We only vectorize innermost loops.
908 SE = &getAnalysis<ScalarEvolution>();
909 DL = getAnalysisIfAvailable<DataLayout>();
910 LI = &getAnalysis<LoopInfo>();
911 TTI = &getAnalysis<TargetTransformInfo>();
912 DT = &getAnalysis<DominatorTree>();
913 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
915 // If the target claims to have no vector registers don't attempt
917 if (!TTI->getNumberOfRegisters(true))
921 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout\n");
925 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
926 L->getHeader()->getParent()->getName() << "\"\n");
928 LoopVectorizeHints Hints(L, DisableUnrolling);
930 if (Hints.Width == 1 && Hints.Unroll == 1) {
931 DEBUG(dbgs() << "LV: Not vectorizing.\n");
935 // Check if it is legal to vectorize the loop.
936 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
937 if (!LVL.canVectorize()) {
938 DEBUG(dbgs() << "LV: Not vectorizing.\n");
942 // Use the cost model.
943 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
945 // Check the function attributes to find out if this function should be
946 // optimized for size.
947 Function *F = L->getHeader()->getParent();
948 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
949 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
950 unsigned FnIndex = AttributeSet::FunctionIndex;
951 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
952 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
955 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
956 "attribute is used.\n");
960 // Select the optimal vectorization factor.
961 LoopVectorizationCostModel::VectorizationFactor VF;
962 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
963 // Select the unroll factor.
964 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
968 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
971 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
972 F->getParent()->getModuleIdentifier() << '\n');
973 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
978 // We decided not to vectorize, but we may want to unroll.
979 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
980 Unroller.vectorize(&LVL);
982 // If we decided that it is *legal* to vectorize the loop then do it.
983 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
987 // Mark the loop as already vectorized to avoid vectorizing again.
988 Hints.setAlreadyVectorized(L);
990 DEBUG(verifyFunction(*L->getHeader()->getParent()));
994 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
995 LoopPass::getAnalysisUsage(AU);
996 AU.addRequiredID(LoopSimplifyID);
997 AU.addRequiredID(LCSSAID);
998 AU.addRequired<DominatorTree>();
999 AU.addRequired<LoopInfo>();
1000 AU.addRequired<ScalarEvolution>();
1001 AU.addRequired<TargetTransformInfo>();
1002 AU.addPreserved<LoopInfo>();
1003 AU.addPreserved<DominatorTree>();
1008 } // end anonymous namespace
1010 //===----------------------------------------------------------------------===//
1011 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1012 // LoopVectorizationCostModel.
1013 //===----------------------------------------------------------------------===//
1016 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1017 Loop *Lp, Value *Ptr,
1019 unsigned DepSetId) {
1020 const SCEV *Sc = SE->getSCEV(Ptr);
1021 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1022 assert(AR && "Invalid addrec expression");
1023 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1024 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1025 Pointers.push_back(Ptr);
1026 Starts.push_back(AR->getStart());
1027 Ends.push_back(ScEnd);
1028 IsWritePtr.push_back(WritePtr);
1029 DependencySetId.push_back(DepSetId);
1032 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1033 // We need to place the broadcast of invariant variables outside the loop.
1034 Instruction *Instr = dyn_cast<Instruction>(V);
1035 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1036 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1038 // Place the code for broadcasting invariant variables in the new preheader.
1039 IRBuilder<>::InsertPointGuard Guard(Builder);
1041 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1043 // Broadcast the scalar into all locations in the vector.
1044 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1049 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1051 assert(Val->getType()->isVectorTy() && "Must be a vector");
1052 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1053 "Elem must be an integer");
1054 // Create the types.
1055 Type *ITy = Val->getType()->getScalarType();
1056 VectorType *Ty = cast<VectorType>(Val->getType());
1057 int VLen = Ty->getNumElements();
1058 SmallVector<Constant*, 8> Indices;
1060 // Create a vector of consecutive numbers from zero to VF.
1061 for (int i = 0; i < VLen; ++i) {
1062 int64_t Idx = Negate ? (-i) : i;
1063 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1066 // Add the consecutive indices to the vector value.
1067 Constant *Cv = ConstantVector::get(Indices);
1068 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1069 return Builder.CreateAdd(Val, Cv, "induction");
1072 /// \brief Find the operand of the GEP that should be checked for consecutive
1073 /// stores. This ignores trailing indices that have no effect on the final
1075 static unsigned getGEPInductionOperand(DataLayout *DL,
1076 const GetElementPtrInst *Gep) {
1077 unsigned LastOperand = Gep->getNumOperands() - 1;
1078 unsigned GEPAllocSize = DL->getTypeAllocSize(
1079 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1081 // Walk backwards and try to peel off zeros.
1082 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1083 // Find the type we're currently indexing into.
1084 gep_type_iterator GEPTI = gep_type_begin(Gep);
1085 std::advance(GEPTI, LastOperand - 1);
1087 // If it's a type with the same allocation size as the result of the GEP we
1088 // can peel off the zero index.
1089 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1097 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1098 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1099 // Make sure that the pointer does not point to structs.
1100 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1103 // If this value is a pointer induction variable we know it is consecutive.
1104 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1105 if (Phi && Inductions.count(Phi)) {
1106 InductionInfo II = Inductions[Phi];
1107 if (IK_PtrInduction == II.IK)
1109 else if (IK_ReversePtrInduction == II.IK)
1113 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1117 unsigned NumOperands = Gep->getNumOperands();
1118 Value *GpPtr = Gep->getPointerOperand();
1119 // If this GEP value is a consecutive pointer induction variable and all of
1120 // the indices are constant then we know it is consecutive. We can
1121 Phi = dyn_cast<PHINode>(GpPtr);
1122 if (Phi && Inductions.count(Phi)) {
1124 // Make sure that the pointer does not point to structs.
1125 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1126 if (GepPtrType->getElementType()->isAggregateType())
1129 // Make sure that all of the index operands are loop invariant.
1130 for (unsigned i = 1; i < NumOperands; ++i)
1131 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1134 InductionInfo II = Inductions[Phi];
1135 if (IK_PtrInduction == II.IK)
1137 else if (IK_ReversePtrInduction == II.IK)
1141 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1143 // Check that all of the gep indices are uniform except for our induction
1145 for (unsigned i = 0; i != NumOperands; ++i)
1146 if (i != InductionOperand &&
1147 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1150 // We can emit wide load/stores only if the last non-zero index is the
1151 // induction variable.
1152 const SCEV *Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1153 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1154 const SCEV *Step = AR->getStepRecurrence(*SE);
1156 // The memory is consecutive because the last index is consecutive
1157 // and all other indices are loop invariant.
1160 if (Step->isAllOnesValue())
1167 bool LoopVectorizationLegality::isUniform(Value *V) {
1168 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1171 InnerLoopVectorizer::VectorParts&
1172 InnerLoopVectorizer::getVectorValue(Value *V) {
1173 assert(V != Induction && "The new induction variable should not be used.");
1174 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1176 // If we have this scalar in the map, return it.
1177 if (WidenMap.has(V))
1178 return WidenMap.get(V);
1180 // If this scalar is unknown, assume that it is a constant or that it is
1181 // loop invariant. Broadcast V and save the value for future uses.
1182 Value *B = getBroadcastInstrs(V);
1183 return WidenMap.splat(V, B);
1186 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1187 assert(Vec->getType()->isVectorTy() && "Invalid type");
1188 SmallVector<Constant*, 8> ShuffleMask;
1189 for (unsigned i = 0; i < VF; ++i)
1190 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1192 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1193 ConstantVector::get(ShuffleMask),
1198 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1199 LoopVectorizationLegality *Legal) {
1200 // Attempt to issue a wide load.
1201 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1202 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1204 assert((LI || SI) && "Invalid Load/Store instruction");
1206 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1207 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1208 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1209 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1210 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1211 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1212 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1214 if (ScalarAllocatedSize != VectorElementSize)
1215 return scalarizeInstruction(Instr);
1217 // If the pointer is loop invariant or if it is non consecutive,
1218 // scalarize the load.
1219 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1220 bool Reverse = ConsecutiveStride < 0;
1221 bool UniformLoad = LI && Legal->isUniform(Ptr);
1222 if (!ConsecutiveStride || UniformLoad)
1223 return scalarizeInstruction(Instr);
1225 Constant *Zero = Builder.getInt32(0);
1226 VectorParts &Entry = WidenMap.get(Instr);
1228 // Handle consecutive loads/stores.
1229 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1230 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1231 setDebugLocFromInst(Builder, Gep);
1232 Value *PtrOperand = Gep->getPointerOperand();
1233 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1234 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1236 // Create the new GEP with the new induction variable.
1237 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1238 Gep2->setOperand(0, FirstBasePtr);
1239 Gep2->setName("gep.indvar.base");
1240 Ptr = Builder.Insert(Gep2);
1242 setDebugLocFromInst(Builder, Gep);
1243 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1244 OrigLoop) && "Base ptr must be invariant");
1246 // The last index does not have to be the induction. It can be
1247 // consecutive and be a function of the index. For example A[I+1];
1248 unsigned NumOperands = Gep->getNumOperands();
1249 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1250 // Create the new GEP with the new induction variable.
1251 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1253 for (unsigned i = 0; i < NumOperands; ++i) {
1254 Value *GepOperand = Gep->getOperand(i);
1255 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1257 // Update last index or loop invariant instruction anchored in loop.
1258 if (i == InductionOperand ||
1259 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1260 assert((i == InductionOperand ||
1261 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1262 "Must be last index or loop invariant");
1264 VectorParts &GEPParts = getVectorValue(GepOperand);
1265 Value *Index = GEPParts[0];
1266 Index = Builder.CreateExtractElement(Index, Zero);
1267 Gep2->setOperand(i, Index);
1268 Gep2->setName("gep.indvar.idx");
1271 Ptr = Builder.Insert(Gep2);
1273 // Use the induction element ptr.
1274 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1275 setDebugLocFromInst(Builder, Ptr);
1276 VectorParts &PtrVal = getVectorValue(Ptr);
1277 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1282 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1283 "We do not allow storing to uniform addresses");
1284 setDebugLocFromInst(Builder, SI);
1285 // We don't want to update the value in the map as it might be used in
1286 // another expression. So don't use a reference type for "StoredVal".
1287 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1289 for (unsigned Part = 0; Part < UF; ++Part) {
1290 // Calculate the pointer for the specific unroll-part.
1291 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1294 // If we store to reverse consecutive memory locations then we need
1295 // to reverse the order of elements in the stored value.
1296 StoredVal[Part] = reverseVector(StoredVal[Part]);
1297 // If the address is consecutive but reversed, then the
1298 // wide store needs to start at the last vector element.
1299 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1300 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1303 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1304 DataTy->getPointerTo(AddressSpace));
1305 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1311 assert(LI && "Must have a load instruction");
1312 setDebugLocFromInst(Builder, LI);
1313 for (unsigned Part = 0; Part < UF; ++Part) {
1314 // Calculate the pointer for the specific unroll-part.
1315 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1318 // If the address is consecutive but reversed, then the
1319 // wide store needs to start at the last vector element.
1320 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1321 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1324 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1325 DataTy->getPointerTo(AddressSpace));
1326 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1327 cast<LoadInst>(LI)->setAlignment(Alignment);
1328 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1332 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1333 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1334 // Holds vector parameters or scalars, in case of uniform vals.
1335 SmallVector<VectorParts, 4> Params;
1337 setDebugLocFromInst(Builder, Instr);
1339 // Find all of the vectorized parameters.
1340 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1341 Value *SrcOp = Instr->getOperand(op);
1343 // If we are accessing the old induction variable, use the new one.
1344 if (SrcOp == OldInduction) {
1345 Params.push_back(getVectorValue(SrcOp));
1349 // Try using previously calculated values.
1350 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1352 // If the src is an instruction that appeared earlier in the basic block
1353 // then it should already be vectorized.
1354 if (SrcInst && OrigLoop->contains(SrcInst)) {
1355 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1356 // The parameter is a vector value from earlier.
1357 Params.push_back(WidenMap.get(SrcInst));
1359 // The parameter is a scalar from outside the loop. Maybe even a constant.
1360 VectorParts Scalars;
1361 Scalars.append(UF, SrcOp);
1362 Params.push_back(Scalars);
1366 assert(Params.size() == Instr->getNumOperands() &&
1367 "Invalid number of operands");
1369 // Does this instruction return a value ?
1370 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1372 Value *UndefVec = IsVoidRetTy ? 0 :
1373 UndefValue::get(VectorType::get(Instr->getType(), VF));
1374 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1375 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1377 // For each vector unroll 'part':
1378 for (unsigned Part = 0; Part < UF; ++Part) {
1379 // For each scalar that we create:
1380 for (unsigned Width = 0; Width < VF; ++Width) {
1381 Instruction *Cloned = Instr->clone();
1383 Cloned->setName(Instr->getName() + ".cloned");
1384 // Replace the operands of the cloned instructions with extracted scalars.
1385 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1386 Value *Op = Params[op][Part];
1387 // Param is a vector. Need to extract the right lane.
1388 if (Op->getType()->isVectorTy())
1389 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1390 Cloned->setOperand(op, Op);
1393 // Place the cloned scalar in the new loop.
1394 Builder.Insert(Cloned);
1396 // If the original scalar returns a value we need to place it in a vector
1397 // so that future users will be able to use it.
1399 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1400 Builder.getInt32(Width));
1406 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1408 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1409 Legal->getRuntimePointerCheck();
1411 if (!PtrRtCheck->Need)
1414 unsigned NumPointers = PtrRtCheck->Pointers.size();
1415 SmallVector<TrackingVH<Value> , 2> Starts;
1416 SmallVector<TrackingVH<Value> , 2> Ends;
1418 LLVMContext &Ctx = Loc->getContext();
1419 SCEVExpander Exp(*SE, "induction");
1421 for (unsigned i = 0; i < NumPointers; ++i) {
1422 Value *Ptr = PtrRtCheck->Pointers[i];
1423 const SCEV *Sc = SE->getSCEV(Ptr);
1425 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1426 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1428 Starts.push_back(Ptr);
1429 Ends.push_back(Ptr);
1431 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1432 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1434 // Use this type for pointer arithmetic.
1435 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1437 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1438 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1439 Starts.push_back(Start);
1440 Ends.push_back(End);
1444 IRBuilder<> ChkBuilder(Loc);
1445 // Our instructions might fold to a constant.
1446 Value *MemoryRuntimeCheck = 0;
1447 for (unsigned i = 0; i < NumPointers; ++i) {
1448 for (unsigned j = i+1; j < NumPointers; ++j) {
1449 // No need to check if two readonly pointers intersect.
1450 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1453 // Only need to check pointers between two different dependency sets.
1454 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1457 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1458 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1460 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1461 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1462 "Trying to bounds check pointers with different address spaces");
1464 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1465 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1467 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1468 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1469 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1470 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1472 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1473 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1474 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1475 if (MemoryRuntimeCheck)
1476 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1478 MemoryRuntimeCheck = IsConflict;
1482 // We have to do this trickery because the IRBuilder might fold the check to a
1483 // constant expression in which case there is no Instruction anchored in a
1485 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1486 ConstantInt::getTrue(Ctx));
1487 ChkBuilder.Insert(Check, "memcheck.conflict");
1492 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1494 In this function we generate a new loop. The new loop will contain
1495 the vectorized instructions while the old loop will continue to run the
1498 [ ] <-- vector loop bypass (may consist of multiple blocks).
1501 | [ ] <-- vector pre header.
1505 | [ ]_| <-- vector loop.
1508 >[ ] <--- middle-block.
1511 | [ ] <--- new preheader.
1515 | [ ]_| <-- old scalar loop to handle remainder.
1518 >[ ] <-- exit block.
1522 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1523 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1524 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1525 assert(ExitBlock && "Must have an exit block");
1527 // Some loops have a single integer induction variable, while other loops
1528 // don't. One example is c++ iterators that often have multiple pointer
1529 // induction variables. In the code below we also support a case where we
1530 // don't have a single induction variable.
1531 OldInduction = Legal->getInduction();
1532 Type *IdxTy = Legal->getWidestInductionType();
1534 // Find the loop boundaries.
1535 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1536 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1538 // Get the total trip count from the count by adding 1.
1539 ExitCount = SE->getAddExpr(ExitCount,
1540 SE->getConstant(ExitCount->getType(), 1));
1542 // Expand the trip count and place the new instructions in the preheader.
1543 // Notice that the pre-header does not change, only the loop body.
1544 SCEVExpander Exp(*SE, "induction");
1546 // Count holds the overall loop count (N).
1547 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1548 BypassBlock->getTerminator());
1550 // The loop index does not have to start at Zero. Find the original start
1551 // value from the induction PHI node. If we don't have an induction variable
1552 // then we know that it starts at zero.
1553 Builder.SetInsertPoint(BypassBlock->getTerminator());
1554 Value *StartIdx = ExtendedIdx = OldInduction ?
1555 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1557 ConstantInt::get(IdxTy, 0);
1559 assert(BypassBlock && "Invalid loop structure");
1560 LoopBypassBlocks.push_back(BypassBlock);
1562 // Split the single block loop into the two loop structure described above.
1563 BasicBlock *VectorPH =
1564 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1565 BasicBlock *VecBody =
1566 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1567 BasicBlock *MiddleBlock =
1568 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1569 BasicBlock *ScalarPH =
1570 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1572 // Create and register the new vector loop.
1573 Loop* Lp = new Loop();
1574 Loop *ParentLoop = OrigLoop->getParentLoop();
1576 // Insert the new loop into the loop nest and register the new basic blocks
1577 // before calling any utilities such as SCEV that require valid LoopInfo.
1579 ParentLoop->addChildLoop(Lp);
1580 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1581 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1582 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1584 LI->addTopLevelLoop(Lp);
1586 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1588 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1590 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1592 // Generate the induction variable.
1593 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1594 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1595 // The loop step is equal to the vectorization factor (num of SIMD elements)
1596 // times the unroll factor (num of SIMD instructions).
1597 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1599 // This is the IR builder that we use to add all of the logic for bypassing
1600 // the new vector loop.
1601 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1602 setDebugLocFromInst(BypassBuilder,
1603 getDebugLocFromInstOrOperands(OldInduction));
1605 // We may need to extend the index in case there is a type mismatch.
1606 // We know that the count starts at zero and does not overflow.
1607 if (Count->getType() != IdxTy) {
1608 // The exit count can be of pointer type. Convert it to the correct
1610 if (ExitCount->getType()->isPointerTy())
1611 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1613 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1616 // Add the start index to the loop count to get the new end index.
1617 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1619 // Now we need to generate the expression for N - (N % VF), which is
1620 // the part that the vectorized body will execute.
1621 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1622 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1623 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1624 "end.idx.rnd.down");
1626 // Now, compare the new count to zero. If it is zero skip the vector loop and
1627 // jump to the scalar loop.
1628 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1631 BasicBlock *LastBypassBlock = BypassBlock;
1633 // Generate the code that checks in runtime if arrays overlap. We put the
1634 // checks into a separate block to make the more common case of few elements
1636 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1637 BypassBlock->getTerminator());
1638 if (MemRuntimeCheck) {
1639 // Create a new block containing the memory check.
1640 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1643 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1644 LoopBypassBlocks.push_back(CheckBlock);
1646 // Replace the branch into the memory check block with a conditional branch
1647 // for the "few elements case".
1648 Instruction *OldTerm = BypassBlock->getTerminator();
1649 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1650 OldTerm->eraseFromParent();
1652 Cmp = MemRuntimeCheck;
1653 LastBypassBlock = CheckBlock;
1656 LastBypassBlock->getTerminator()->eraseFromParent();
1657 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1660 // We are going to resume the execution of the scalar loop.
1661 // Go over all of the induction variables that we found and fix the
1662 // PHIs that are left in the scalar version of the loop.
1663 // The starting values of PHI nodes depend on the counter of the last
1664 // iteration in the vectorized loop.
1665 // If we come from a bypass edge then we need to start from the original
1668 // This variable saves the new starting index for the scalar loop.
1669 PHINode *ResumeIndex = 0;
1670 LoopVectorizationLegality::InductionList::iterator I, E;
1671 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1672 // Set builder to point to last bypass block.
1673 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1674 for (I = List->begin(), E = List->end(); I != E; ++I) {
1675 PHINode *OrigPhi = I->first;
1676 LoopVectorizationLegality::InductionInfo II = I->second;
1678 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1679 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1680 MiddleBlock->getTerminator());
1681 // We might have extended the type of the induction variable but we need a
1682 // truncated version for the scalar loop.
1683 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1684 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1685 MiddleBlock->getTerminator()) : 0;
1687 Value *EndValue = 0;
1689 case LoopVectorizationLegality::IK_NoInduction:
1690 llvm_unreachable("Unknown induction");
1691 case LoopVectorizationLegality::IK_IntInduction: {
1692 // Handle the integer induction counter.
1693 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1695 // We have the canonical induction variable.
1696 if (OrigPhi == OldInduction) {
1697 // Create a truncated version of the resume value for the scalar loop,
1698 // we might have promoted the type to a larger width.
1700 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1701 // The new PHI merges the original incoming value, in case of a bypass,
1702 // or the value at the end of the vectorized loop.
1703 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1704 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1705 TruncResumeVal->addIncoming(EndValue, VecBody);
1707 // We know what the end value is.
1708 EndValue = IdxEndRoundDown;
1709 // We also know which PHI node holds it.
1710 ResumeIndex = ResumeVal;
1714 // Not the canonical induction variable - add the vector loop count to the
1716 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1717 II.StartValue->getType(),
1719 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1722 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1723 // Convert the CountRoundDown variable to the PHI size.
1724 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1725 II.StartValue->getType(),
1727 // Handle reverse integer induction counter.
1728 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1731 case LoopVectorizationLegality::IK_PtrInduction: {
1732 // For pointer induction variables, calculate the offset using
1734 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1738 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1739 // The value at the end of the loop for the reverse pointer is calculated
1740 // by creating a GEP with a negative index starting from the start value.
1741 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1742 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1744 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1750 // The new PHI merges the original incoming value, in case of a bypass,
1751 // or the value at the end of the vectorized loop.
1752 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1753 if (OrigPhi == OldInduction)
1754 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1756 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1758 ResumeVal->addIncoming(EndValue, VecBody);
1760 // Fix the scalar body counter (PHI node).
1761 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1762 // The old inductions phi node in the scalar body needs the truncated value.
1763 if (OrigPhi == OldInduction)
1764 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1766 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1769 // If we are generating a new induction variable then we also need to
1770 // generate the code that calculates the exit value. This value is not
1771 // simply the end of the counter because we may skip the vectorized body
1772 // in case of a runtime check.
1774 assert(!ResumeIndex && "Unexpected resume value found");
1775 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1776 MiddleBlock->getTerminator());
1777 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1778 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1779 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1782 // Make sure that we found the index where scalar loop needs to continue.
1783 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1784 "Invalid resume Index");
1786 // Add a check in the middle block to see if we have completed
1787 // all of the iterations in the first vector loop.
1788 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1789 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1790 ResumeIndex, "cmp.n",
1791 MiddleBlock->getTerminator());
1793 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1794 // Remove the old terminator.
1795 MiddleBlock->getTerminator()->eraseFromParent();
1797 // Create i+1 and fill the PHINode.
1798 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1799 Induction->addIncoming(StartIdx, VectorPH);
1800 Induction->addIncoming(NextIdx, VecBody);
1801 // Create the compare.
1802 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1803 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1805 // Now we have two terminators. Remove the old one from the block.
1806 VecBody->getTerminator()->eraseFromParent();
1808 // Get ready to start creating new instructions into the vectorized body.
1809 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1812 LoopVectorPreHeader = VectorPH;
1813 LoopScalarPreHeader = ScalarPH;
1814 LoopMiddleBlock = MiddleBlock;
1815 LoopExitBlock = ExitBlock;
1816 LoopVectorBody = VecBody;
1817 LoopScalarBody = OldBasicBlock;
1819 LoopVectorizeHints Hints(Lp, true);
1820 Hints.setAlreadyVectorized(Lp);
1823 /// This function returns the identity element (or neutral element) for
1824 /// the operation K.
1826 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1831 // Adding, Xoring, Oring zero to a number does not change it.
1832 return ConstantInt::get(Tp, 0);
1833 case RK_IntegerMult:
1834 // Multiplying a number by 1 does not change it.
1835 return ConstantInt::get(Tp, 1);
1837 // AND-ing a number with an all-1 value does not change it.
1838 return ConstantInt::get(Tp, -1, true);
1840 // Multiplying a number by 1 does not change it.
1841 return ConstantFP::get(Tp, 1.0L);
1843 // Adding zero to a number does not change it.
1844 return ConstantFP::get(Tp, 0.0L);
1846 llvm_unreachable("Unknown reduction kind");
1850 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
1851 Intrinsic::ID ValidIntrinsicID) {
1852 if (I.getNumArgOperands() != 1 ||
1853 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1854 I.getType() != I.getArgOperand(0)->getType() ||
1855 !I.onlyReadsMemory())
1856 return Intrinsic::not_intrinsic;
1858 return ValidIntrinsicID;
1861 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
1862 Intrinsic::ID ValidIntrinsicID) {
1863 if (I.getNumArgOperands() != 2 ||
1864 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1865 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
1866 I.getType() != I.getArgOperand(0)->getType() ||
1867 I.getType() != I.getArgOperand(1)->getType() ||
1868 !I.onlyReadsMemory())
1869 return Intrinsic::not_intrinsic;
1871 return ValidIntrinsicID;
1875 static Intrinsic::ID
1876 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1877 // If we have an intrinsic call, check if it is trivially vectorizable.
1878 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1879 switch (II->getIntrinsicID()) {
1880 case Intrinsic::sqrt:
1881 case Intrinsic::sin:
1882 case Intrinsic::cos:
1883 case Intrinsic::exp:
1884 case Intrinsic::exp2:
1885 case Intrinsic::log:
1886 case Intrinsic::log10:
1887 case Intrinsic::log2:
1888 case Intrinsic::fabs:
1889 case Intrinsic::copysign:
1890 case Intrinsic::floor:
1891 case Intrinsic::ceil:
1892 case Intrinsic::trunc:
1893 case Intrinsic::rint:
1894 case Intrinsic::nearbyint:
1895 case Intrinsic::round:
1896 case Intrinsic::pow:
1897 case Intrinsic::fma:
1898 case Intrinsic::fmuladd:
1899 case Intrinsic::lifetime_start:
1900 case Intrinsic::lifetime_end:
1901 return II->getIntrinsicID();
1903 return Intrinsic::not_intrinsic;
1908 return Intrinsic::not_intrinsic;
1911 Function *F = CI->getCalledFunction();
1912 // We're going to make assumptions on the semantics of the functions, check
1913 // that the target knows that it's available in this environment and it does
1914 // not have local linkage.
1915 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
1916 return Intrinsic::not_intrinsic;
1918 // Otherwise check if we have a call to a function that can be turned into a
1919 // vector intrinsic.
1926 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
1930 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
1934 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
1936 case LibFunc::exp2f:
1937 case LibFunc::exp2l:
1938 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
1942 return checkUnaryFloatSignature(*CI, Intrinsic::log);
1943 case LibFunc::log10:
1944 case LibFunc::log10f:
1945 case LibFunc::log10l:
1946 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
1948 case LibFunc::log2f:
1949 case LibFunc::log2l:
1950 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
1952 case LibFunc::fabsf:
1953 case LibFunc::fabsl:
1954 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
1955 case LibFunc::copysign:
1956 case LibFunc::copysignf:
1957 case LibFunc::copysignl:
1958 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
1959 case LibFunc::floor:
1960 case LibFunc::floorf:
1961 case LibFunc::floorl:
1962 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
1964 case LibFunc::ceilf:
1965 case LibFunc::ceill:
1966 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
1967 case LibFunc::trunc:
1968 case LibFunc::truncf:
1969 case LibFunc::truncl:
1970 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
1972 case LibFunc::rintf:
1973 case LibFunc::rintl:
1974 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
1975 case LibFunc::nearbyint:
1976 case LibFunc::nearbyintf:
1977 case LibFunc::nearbyintl:
1978 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
1979 case LibFunc::round:
1980 case LibFunc::roundf:
1981 case LibFunc::roundl:
1982 return checkUnaryFloatSignature(*CI, Intrinsic::round);
1986 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
1989 return Intrinsic::not_intrinsic;
1992 /// This function translates the reduction kind to an LLVM binary operator.
1994 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1996 case LoopVectorizationLegality::RK_IntegerAdd:
1997 return Instruction::Add;
1998 case LoopVectorizationLegality::RK_IntegerMult:
1999 return Instruction::Mul;
2000 case LoopVectorizationLegality::RK_IntegerOr:
2001 return Instruction::Or;
2002 case LoopVectorizationLegality::RK_IntegerAnd:
2003 return Instruction::And;
2004 case LoopVectorizationLegality::RK_IntegerXor:
2005 return Instruction::Xor;
2006 case LoopVectorizationLegality::RK_FloatMult:
2007 return Instruction::FMul;
2008 case LoopVectorizationLegality::RK_FloatAdd:
2009 return Instruction::FAdd;
2010 case LoopVectorizationLegality::RK_IntegerMinMax:
2011 return Instruction::ICmp;
2012 case LoopVectorizationLegality::RK_FloatMinMax:
2013 return Instruction::FCmp;
2015 llvm_unreachable("Unknown reduction operation");
2019 Value *createMinMaxOp(IRBuilder<> &Builder,
2020 LoopVectorizationLegality::MinMaxReductionKind RK,
2023 CmpInst::Predicate P = CmpInst::ICMP_NE;
2026 llvm_unreachable("Unknown min/max reduction kind");
2027 case LoopVectorizationLegality::MRK_UIntMin:
2028 P = CmpInst::ICMP_ULT;
2030 case LoopVectorizationLegality::MRK_UIntMax:
2031 P = CmpInst::ICMP_UGT;
2033 case LoopVectorizationLegality::MRK_SIntMin:
2034 P = CmpInst::ICMP_SLT;
2036 case LoopVectorizationLegality::MRK_SIntMax:
2037 P = CmpInst::ICMP_SGT;
2039 case LoopVectorizationLegality::MRK_FloatMin:
2040 P = CmpInst::FCMP_OLT;
2042 case LoopVectorizationLegality::MRK_FloatMax:
2043 P = CmpInst::FCMP_OGT;
2048 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2049 RK == LoopVectorizationLegality::MRK_FloatMax)
2050 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2052 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2054 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2059 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2060 //===------------------------------------------------===//
2062 // Notice: any optimization or new instruction that go
2063 // into the code below should be also be implemented in
2066 //===------------------------------------------------===//
2067 Constant *Zero = Builder.getInt32(0);
2069 // In order to support reduction variables we need to be able to vectorize
2070 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2071 // stages. First, we create a new vector PHI node with no incoming edges.
2072 // We use this value when we vectorize all of the instructions that use the
2073 // PHI. Next, after all of the instructions in the block are complete we
2074 // add the new incoming edges to the PHI. At this point all of the
2075 // instructions in the basic block are vectorized, so we can use them to
2076 // construct the PHI.
2077 PhiVector RdxPHIsToFix;
2079 // Scan the loop in a topological order to ensure that defs are vectorized
2081 LoopBlocksDFS DFS(OrigLoop);
2084 // Vectorize all of the blocks in the original loop.
2085 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2086 be = DFS.endRPO(); bb != be; ++bb)
2087 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2089 // At this point every instruction in the original loop is widened to
2090 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2091 // that we vectorized. The PHI nodes are currently empty because we did
2092 // not want to introduce cycles. Notice that the remaining PHI nodes
2093 // that we need to fix are reduction variables.
2095 // Create the 'reduced' values for each of the induction vars.
2096 // The reduced values are the vector values that we scalarize and combine
2097 // after the loop is finished.
2098 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2100 PHINode *RdxPhi = *it;
2101 assert(RdxPhi && "Unable to recover vectorized PHI");
2103 // Find the reduction variable descriptor.
2104 assert(Legal->getReductionVars()->count(RdxPhi) &&
2105 "Unable to find the reduction variable");
2106 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2107 (*Legal->getReductionVars())[RdxPhi];
2109 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2111 // We need to generate a reduction vector from the incoming scalar.
2112 // To do so, we need to generate the 'identity' vector and overide
2113 // one of the elements with the incoming scalar reduction. We need
2114 // to do it in the vector-loop preheader.
2115 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2117 // This is the vector-clone of the value that leaves the loop.
2118 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2119 Type *VecTy = VectorExit[0]->getType();
2121 // Find the reduction identity variable. Zero for addition, or, xor,
2122 // one for multiplication, -1 for And.
2125 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2126 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2127 // MinMax reduction have the start value as their identify.
2129 VectorStart = Identity = RdxDesc.StartValue;
2131 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2136 // Handle other reduction kinds:
2138 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2139 VecTy->getScalarType());
2142 // This vector is the Identity vector where the first element is the
2143 // incoming scalar reduction.
2144 VectorStart = RdxDesc.StartValue;
2146 Identity = ConstantVector::getSplat(VF, Iden);
2148 // This vector is the Identity vector where the first element is the
2149 // incoming scalar reduction.
2150 VectorStart = Builder.CreateInsertElement(Identity,
2151 RdxDesc.StartValue, Zero);
2155 // Fix the vector-loop phi.
2156 // We created the induction variable so we know that the
2157 // preheader is the first entry.
2158 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2160 // Reductions do not have to start at zero. They can start with
2161 // any loop invariant values.
2162 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2163 BasicBlock *Latch = OrigLoop->getLoopLatch();
2164 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2165 VectorParts &Val = getVectorValue(LoopVal);
2166 for (unsigned part = 0; part < UF; ++part) {
2167 // Make sure to add the reduction stat value only to the
2168 // first unroll part.
2169 Value *StartVal = (part == 0) ? VectorStart : Identity;
2170 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2171 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2174 // Before each round, move the insertion point right between
2175 // the PHIs and the values we are going to write.
2176 // This allows us to write both PHINodes and the extractelement
2178 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2180 VectorParts RdxParts;
2181 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2182 for (unsigned part = 0; part < UF; ++part) {
2183 // This PHINode contains the vectorized reduction variable, or
2184 // the initial value vector, if we bypass the vector loop.
2185 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2186 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2187 Value *StartVal = (part == 0) ? VectorStart : Identity;
2188 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2189 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2190 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2191 RdxParts.push_back(NewPhi);
2194 // Reduce all of the unrolled parts into a single vector.
2195 Value *ReducedPartRdx = RdxParts[0];
2196 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2197 setDebugLocFromInst(Builder, ReducedPartRdx);
2198 for (unsigned part = 1; part < UF; ++part) {
2199 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2200 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2201 RdxParts[part], ReducedPartRdx,
2204 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2205 ReducedPartRdx, RdxParts[part]);
2209 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2210 // and vector ops, reducing the set of values being computed by half each
2212 assert(isPowerOf2_32(VF) &&
2213 "Reduction emission only supported for pow2 vectors!");
2214 Value *TmpVec = ReducedPartRdx;
2215 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2216 for (unsigned i = VF; i != 1; i >>= 1) {
2217 // Move the upper half of the vector to the lower half.
2218 for (unsigned j = 0; j != i/2; ++j)
2219 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2221 // Fill the rest of the mask with undef.
2222 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2223 UndefValue::get(Builder.getInt32Ty()));
2226 Builder.CreateShuffleVector(TmpVec,
2227 UndefValue::get(TmpVec->getType()),
2228 ConstantVector::get(ShuffleMask),
2231 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2232 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2235 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2238 // The result is in the first element of the vector.
2239 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2240 Builder.getInt32(0));
2243 // Now, we need to fix the users of the reduction variable
2244 // inside and outside of the scalar remainder loop.
2245 // We know that the loop is in LCSSA form. We need to update the
2246 // PHI nodes in the exit blocks.
2247 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2248 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2249 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2250 if (!LCSSAPhi) break;
2252 // All PHINodes need to have a single entry edge, or two if
2253 // we already fixed them.
2254 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2256 // We found our reduction value exit-PHI. Update it with the
2257 // incoming bypass edge.
2258 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2259 // Add an edge coming from the bypass.
2260 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2263 }// end of the LCSSA phi scan.
2265 // Fix the scalar loop reduction variable with the incoming reduction sum
2266 // from the vector body and from the backedge value.
2267 int IncomingEdgeBlockIdx =
2268 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2269 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2270 // Pick the other block.
2271 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2272 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2273 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2274 }// end of for each redux variable.
2278 // Perform simple cse.
2279 SmallPtrSet<Instruction*, 16> Visited;
2280 SmallVector<Instruction*, 16> ToRemove;
2281 for (BasicBlock::iterator I = LoopVectorBody->begin(),
2282 E = LoopVectorBody->end(); I != E; ++I) {
2283 Instruction *In = I;
2285 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In) &&
2286 !isa<ShuffleVectorInst>(In) && !isa<GetElementPtrInst>(In))
2289 // Check if we can replace this instruction with any of the
2290 // visited instructions.
2291 for (SmallPtrSet<Instruction*, 16>::iterator v = Visited.begin(),
2292 ve = Visited.end(); v != ve; ++v) {
2293 if (In->isIdenticalTo(*v)) {
2294 In->replaceAllUsesWith(*v);
2295 ToRemove.push_back(In);
2304 // Erase all of the instructions that we RAUWed.
2305 for (SmallVectorImpl<Instruction *>::iterator v = ToRemove.begin(),
2306 ve = ToRemove.end(); v != ve; ++v) {
2307 assert((*v)->getNumUses() == 0 && "Can't remove instructions with uses");
2308 (*v)->eraseFromParent();
2312 void InnerLoopVectorizer::fixLCSSAPHIs() {
2313 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2314 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2315 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2316 if (!LCSSAPhi) break;
2317 if (LCSSAPhi->getNumIncomingValues() == 1)
2318 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2323 InnerLoopVectorizer::VectorParts
2324 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2325 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2328 // Look for cached value.
2329 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2330 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2331 if (ECEntryIt != MaskCache.end())
2332 return ECEntryIt->second;
2334 VectorParts SrcMask = createBlockInMask(Src);
2336 // The terminator has to be a branch inst!
2337 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2338 assert(BI && "Unexpected terminator found");
2340 if (BI->isConditional()) {
2341 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2343 if (BI->getSuccessor(0) != Dst)
2344 for (unsigned part = 0; part < UF; ++part)
2345 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2347 for (unsigned part = 0; part < UF; ++part)
2348 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2350 MaskCache[Edge] = EdgeMask;
2354 MaskCache[Edge] = SrcMask;
2358 InnerLoopVectorizer::VectorParts
2359 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2360 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2362 // Loop incoming mask is all-one.
2363 if (OrigLoop->getHeader() == BB) {
2364 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2365 return getVectorValue(C);
2368 // This is the block mask. We OR all incoming edges, and with zero.
2369 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2370 VectorParts BlockMask = getVectorValue(Zero);
2373 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2374 VectorParts EM = createEdgeMask(*it, BB);
2375 for (unsigned part = 0; part < UF; ++part)
2376 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2382 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2383 InnerLoopVectorizer::VectorParts &Entry,
2384 LoopVectorizationLegality *Legal,
2385 unsigned UF, unsigned VF, PhiVector *PV) {
2386 PHINode* P = cast<PHINode>(PN);
2387 // Handle reduction variables:
2388 if (Legal->getReductionVars()->count(P)) {
2389 for (unsigned part = 0; part < UF; ++part) {
2390 // This is phase one of vectorizing PHIs.
2391 Type *VecTy = (VF == 1) ? PN->getType() :
2392 VectorType::get(PN->getType(), VF);
2393 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2394 LoopVectorBody-> getFirstInsertionPt());
2400 setDebugLocFromInst(Builder, P);
2401 // Check for PHI nodes that are lowered to vector selects.
2402 if (P->getParent() != OrigLoop->getHeader()) {
2403 // We know that all PHIs in non header blocks are converted into
2404 // selects, so we don't have to worry about the insertion order and we
2405 // can just use the builder.
2406 // At this point we generate the predication tree. There may be
2407 // duplications since this is a simple recursive scan, but future
2408 // optimizations will clean it up.
2410 unsigned NumIncoming = P->getNumIncomingValues();
2412 // Generate a sequence of selects of the form:
2413 // SELECT(Mask3, In3,
2414 // SELECT(Mask2, In2,
2416 for (unsigned In = 0; In < NumIncoming; In++) {
2417 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2419 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2421 for (unsigned part = 0; part < UF; ++part) {
2422 // We might have single edge PHIs (blocks) - use an identity
2423 // 'select' for the first PHI operand.
2425 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2428 // Select between the current value and the previous incoming edge
2429 // based on the incoming mask.
2430 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2431 Entry[part], "predphi");
2437 // This PHINode must be an induction variable.
2438 // Make sure that we know about it.
2439 assert(Legal->getInductionVars()->count(P) &&
2440 "Not an induction variable");
2442 LoopVectorizationLegality::InductionInfo II =
2443 Legal->getInductionVars()->lookup(P);
2446 case LoopVectorizationLegality::IK_NoInduction:
2447 llvm_unreachable("Unknown induction");
2448 case LoopVectorizationLegality::IK_IntInduction: {
2449 assert(P->getType() == II.StartValue->getType() && "Types must match");
2450 Type *PhiTy = P->getType();
2452 if (P == OldInduction) {
2453 // Handle the canonical induction variable. We might have had to
2455 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2457 // Handle other induction variables that are now based on the
2459 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2461 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2462 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2465 Broadcasted = getBroadcastInstrs(Broadcasted);
2466 // After broadcasting the induction variable we need to make the vector
2467 // consecutive by adding 0, 1, 2, etc.
2468 for (unsigned part = 0; part < UF; ++part)
2469 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2472 case LoopVectorizationLegality::IK_ReverseIntInduction:
2473 case LoopVectorizationLegality::IK_PtrInduction:
2474 case LoopVectorizationLegality::IK_ReversePtrInduction:
2475 // Handle reverse integer and pointer inductions.
2476 Value *StartIdx = ExtendedIdx;
2477 // This is the normalized GEP that starts counting at zero.
2478 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2481 // Handle the reverse integer induction variable case.
2482 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2483 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2484 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2486 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2489 // This is a new value so do not hoist it out.
2490 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2491 // After broadcasting the induction variable we need to make the
2492 // vector consecutive by adding ... -3, -2, -1, 0.
2493 for (unsigned part = 0; part < UF; ++part)
2494 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2499 // Handle the pointer induction variable case.
2500 assert(P->getType()->isPointerTy() && "Unexpected type.");
2502 // Is this a reverse induction ptr or a consecutive induction ptr.
2503 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2506 // This is the vector of results. Notice that we don't generate
2507 // vector geps because scalar geps result in better code.
2508 for (unsigned part = 0; part < UF; ++part) {
2510 int EltIndex = (part) * (Reverse ? -1 : 1);
2511 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2514 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2516 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2518 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2520 Entry[part] = SclrGep;
2524 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2525 for (unsigned int i = 0; i < VF; ++i) {
2526 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2527 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2530 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2532 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2534 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2536 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2537 Builder.getInt32(i),
2540 Entry[part] = VecVal;
2547 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2548 BasicBlock *BB, PhiVector *PV) {
2549 // For each instruction in the old loop.
2550 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2551 VectorParts &Entry = WidenMap.get(it);
2552 switch (it->getOpcode()) {
2553 case Instruction::Br:
2554 // Nothing to do for PHIs and BR, since we already took care of the
2555 // loop control flow instructions.
2557 case Instruction::PHI:{
2558 // Vectorize PHINodes.
2559 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2563 case Instruction::Add:
2564 case Instruction::FAdd:
2565 case Instruction::Sub:
2566 case Instruction::FSub:
2567 case Instruction::Mul:
2568 case Instruction::FMul:
2569 case Instruction::UDiv:
2570 case Instruction::SDiv:
2571 case Instruction::FDiv:
2572 case Instruction::URem:
2573 case Instruction::SRem:
2574 case Instruction::FRem:
2575 case Instruction::Shl:
2576 case Instruction::LShr:
2577 case Instruction::AShr:
2578 case Instruction::And:
2579 case Instruction::Or:
2580 case Instruction::Xor: {
2581 // Just widen binops.
2582 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2583 setDebugLocFromInst(Builder, BinOp);
2584 VectorParts &A = getVectorValue(it->getOperand(0));
2585 VectorParts &B = getVectorValue(it->getOperand(1));
2587 // Use this vector value for all users of the original instruction.
2588 for (unsigned Part = 0; Part < UF; ++Part) {
2589 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2591 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2592 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2593 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2594 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2595 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2597 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2598 VecOp->setIsExact(BinOp->isExact());
2604 case Instruction::Select: {
2606 // If the selector is loop invariant we can create a select
2607 // instruction with a scalar condition. Otherwise, use vector-select.
2608 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2610 setDebugLocFromInst(Builder, it);
2612 // The condition can be loop invariant but still defined inside the
2613 // loop. This means that we can't just use the original 'cond' value.
2614 // We have to take the 'vectorized' value and pick the first lane.
2615 // Instcombine will make this a no-op.
2616 VectorParts &Cond = getVectorValue(it->getOperand(0));
2617 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2618 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2620 Value *ScalarCond = (VF == 1) ? Cond[0] :
2621 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2623 for (unsigned Part = 0; Part < UF; ++Part) {
2624 Entry[Part] = Builder.CreateSelect(
2625 InvariantCond ? ScalarCond : Cond[Part],
2632 case Instruction::ICmp:
2633 case Instruction::FCmp: {
2634 // Widen compares. Generate vector compares.
2635 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2636 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2637 setDebugLocFromInst(Builder, it);
2638 VectorParts &A = getVectorValue(it->getOperand(0));
2639 VectorParts &B = getVectorValue(it->getOperand(1));
2640 for (unsigned Part = 0; Part < UF; ++Part) {
2643 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2645 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2651 case Instruction::Store:
2652 case Instruction::Load:
2653 vectorizeMemoryInstruction(it, Legal);
2655 case Instruction::ZExt:
2656 case Instruction::SExt:
2657 case Instruction::FPToUI:
2658 case Instruction::FPToSI:
2659 case Instruction::FPExt:
2660 case Instruction::PtrToInt:
2661 case Instruction::IntToPtr:
2662 case Instruction::SIToFP:
2663 case Instruction::UIToFP:
2664 case Instruction::Trunc:
2665 case Instruction::FPTrunc:
2666 case Instruction::BitCast: {
2667 CastInst *CI = dyn_cast<CastInst>(it);
2668 setDebugLocFromInst(Builder, it);
2669 /// Optimize the special case where the source is the induction
2670 /// variable. Notice that we can only optimize the 'trunc' case
2671 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2672 /// c. other casts depend on pointer size.
2673 if (CI->getOperand(0) == OldInduction &&
2674 it->getOpcode() == Instruction::Trunc) {
2675 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2677 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2678 for (unsigned Part = 0; Part < UF; ++Part)
2679 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2682 /// Vectorize casts.
2683 Type *DestTy = (VF == 1) ? CI->getType() :
2684 VectorType::get(CI->getType(), VF);
2686 VectorParts &A = getVectorValue(it->getOperand(0));
2687 for (unsigned Part = 0; Part < UF; ++Part)
2688 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2692 case Instruction::Call: {
2693 // Ignore dbg intrinsics.
2694 if (isa<DbgInfoIntrinsic>(it))
2696 setDebugLocFromInst(Builder, it);
2698 Module *M = BB->getParent()->getParent();
2699 CallInst *CI = cast<CallInst>(it);
2700 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2701 assert(ID && "Not an intrinsic call!");
2703 case Intrinsic::lifetime_end:
2704 case Intrinsic::lifetime_start:
2705 scalarizeInstruction(it);
2708 for (unsigned Part = 0; Part < UF; ++Part) {
2709 SmallVector<Value *, 4> Args;
2710 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2711 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2712 Args.push_back(Arg[Part]);
2714 Type *Tys[] = {CI->getType()};
2716 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2718 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2719 Entry[Part] = Builder.CreateCall(F, Args);
2727 // All other instructions are unsupported. Scalarize them.
2728 scalarizeInstruction(it);
2731 }// end of for_each instr.
2734 void InnerLoopVectorizer::updateAnalysis() {
2735 // Forget the original basic block.
2736 SE->forgetLoop(OrigLoop);
2738 // Update the dominator tree information.
2739 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2740 "Entry does not dominate exit.");
2742 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2743 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2744 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2745 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2746 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2747 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2748 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2749 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2751 DEBUG(DT->verifyAnalysis());
2754 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2755 if (!EnableIfConversion)
2758 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2760 // A list of pointers that we can safely read and write to.
2761 SmallPtrSet<Value *, 8> SafePointes;
2763 // Collect safe addresses.
2764 for (Loop::block_iterator BI = TheLoop->block_begin(),
2765 BE = TheLoop->block_end(); BI != BE; ++BI) {
2766 BasicBlock *BB = *BI;
2768 if (blockNeedsPredication(BB))
2771 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2772 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2773 SafePointes.insert(LI->getPointerOperand());
2774 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2775 SafePointes.insert(SI->getPointerOperand());
2779 // Collect the blocks that need predication.
2780 for (Loop::block_iterator BI = TheLoop->block_begin(),
2781 BE = TheLoop->block_end(); BI != BE; ++BI) {
2782 BasicBlock *BB = *BI;
2784 // We don't support switch statements inside loops.
2785 if (!isa<BranchInst>(BB->getTerminator()))
2788 // We must be able to predicate all blocks that need to be predicated.
2789 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2793 // We can if-convert this loop.
2797 bool LoopVectorizationLegality::canVectorize() {
2798 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2799 // be canonicalized.
2800 if (!TheLoop->getLoopPreheader())
2803 // We can only vectorize innermost loops.
2804 if (TheLoop->getSubLoopsVector().size())
2807 // We must have a single backedge.
2808 if (TheLoop->getNumBackEdges() != 1)
2811 // We must have a single exiting block.
2812 if (!TheLoop->getExitingBlock())
2815 // We need to have a loop header.
2816 DEBUG(dbgs() << "LV: Found a loop: " <<
2817 TheLoop->getHeader()->getName() << '\n');
2819 // Check if we can if-convert non single-bb loops.
2820 unsigned NumBlocks = TheLoop->getNumBlocks();
2821 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2822 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2826 // ScalarEvolution needs to be able to find the exit count.
2827 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2828 if (ExitCount == SE->getCouldNotCompute()) {
2829 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2833 // Do not loop-vectorize loops with a tiny trip count.
2834 BasicBlock *Latch = TheLoop->getLoopLatch();
2835 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2836 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2837 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2838 "This loop is not worth vectorizing.\n");
2842 // Check if we can vectorize the instructions and CFG in this loop.
2843 if (!canVectorizeInstrs()) {
2844 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2848 // Go over each instruction and look at memory deps.
2849 if (!canVectorizeMemory()) {
2850 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2854 // Collect all of the variables that remain uniform after vectorization.
2855 collectLoopUniforms();
2857 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2858 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2861 // Okay! We can vectorize. At this point we don't have any other mem analysis
2862 // which may limit our maximum vectorization factor, so just return true with
2867 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2868 if (Ty->isPointerTy())
2869 return DL.getIntPtrType(Ty);
2874 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2875 Ty0 = convertPointerToIntegerType(DL, Ty0);
2876 Ty1 = convertPointerToIntegerType(DL, Ty1);
2877 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2882 /// \brief Check that the instruction has outside loop users and is not an
2883 /// identified reduction variable.
2884 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2885 SmallPtrSet<Value *, 4> &Reductions) {
2886 // Reduction instructions are allowed to have exit users. All other
2887 // instructions must not have external users.
2888 if (!Reductions.count(Inst))
2889 //Check that all of the users of the loop are inside the BB.
2890 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2892 Instruction *U = cast<Instruction>(*I);
2893 // This user may be a reduction exit value.
2894 if (!TheLoop->contains(U)) {
2895 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
2902 bool LoopVectorizationLegality::canVectorizeInstrs() {
2903 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2904 BasicBlock *Header = TheLoop->getHeader();
2906 // Look for the attribute signaling the absence of NaNs.
2907 Function &F = *Header->getParent();
2908 if (F.hasFnAttribute("no-nans-fp-math"))
2909 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2910 AttributeSet::FunctionIndex,
2911 "no-nans-fp-math").getValueAsString() == "true";
2913 // For each block in the loop.
2914 for (Loop::block_iterator bb = TheLoop->block_begin(),
2915 be = TheLoop->block_end(); bb != be; ++bb) {
2917 // Scan the instructions in the block and look for hazards.
2918 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2921 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2922 Type *PhiTy = Phi->getType();
2923 // Check that this PHI type is allowed.
2924 if (!PhiTy->isIntegerTy() &&
2925 !PhiTy->isFloatingPointTy() &&
2926 !PhiTy->isPointerTy()) {
2927 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2931 // If this PHINode is not in the header block, then we know that we
2932 // can convert it to select during if-conversion. No need to check if
2933 // the PHIs in this block are induction or reduction variables.
2934 if (*bb != Header) {
2935 // Check that this instruction has no outside users or is an
2936 // identified reduction value with an outside user.
2937 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2942 // We only allow if-converted PHIs with more than two incoming values.
2943 if (Phi->getNumIncomingValues() != 2) {
2944 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2948 // This is the value coming from the preheader.
2949 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2950 // Check if this is an induction variable.
2951 InductionKind IK = isInductionVariable(Phi);
2953 if (IK_NoInduction != IK) {
2954 // Get the widest type.
2956 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2958 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2960 // Int inductions are special because we only allow one IV.
2961 if (IK == IK_IntInduction) {
2962 // Use the phi node with the widest type as induction. Use the last
2963 // one if there are multiple (no good reason for doing this other
2964 // than it is expedient).
2965 if (!Induction || PhiTy == WidestIndTy)
2969 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2970 Inductions[Phi] = InductionInfo(StartValue, IK);
2972 // Until we explicitly handle the case of an induction variable with
2973 // an outside loop user we have to give up vectorizing this loop.
2974 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2980 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2981 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2984 if (AddReductionVar(Phi, RK_IntegerMult)) {
2985 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2988 if (AddReductionVar(Phi, RK_IntegerOr)) {
2989 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2992 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2993 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2996 if (AddReductionVar(Phi, RK_IntegerXor)) {
2997 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3000 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3001 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3004 if (AddReductionVar(Phi, RK_FloatMult)) {
3005 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3008 if (AddReductionVar(Phi, RK_FloatAdd)) {
3009 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3012 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3013 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3018 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3020 }// end of PHI handling
3022 // We still don't handle functions. However, we can ignore dbg intrinsic
3023 // calls and we do handle certain intrinsic and libm functions.
3024 CallInst *CI = dyn_cast<CallInst>(it);
3025 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3026 DEBUG(dbgs() << "LV: Found a call site.\n");
3030 // Check that the instruction return type is vectorizable.
3031 // Also, we can't vectorize extractelement instructions.
3032 if ((!VectorType::isValidElementType(it->getType()) &&
3033 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3034 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3038 // Check that the stored type is vectorizable.
3039 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3040 Type *T = ST->getValueOperand()->getType();
3041 if (!VectorType::isValidElementType(T))
3045 // Reduction instructions are allowed to have exit users.
3046 // All other instructions must not have external users.
3047 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3055 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3056 if (Inductions.empty())
3063 void LoopVectorizationLegality::collectLoopUniforms() {
3064 // We now know that the loop is vectorizable!
3065 // Collect variables that will remain uniform after vectorization.
3066 std::vector<Value*> Worklist;
3067 BasicBlock *Latch = TheLoop->getLoopLatch();
3069 // Start with the conditional branch and walk up the block.
3070 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3072 while (Worklist.size()) {
3073 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3074 Worklist.pop_back();
3076 // Look at instructions inside this loop.
3077 // Stop when reaching PHI nodes.
3078 // TODO: we need to follow values all over the loop, not only in this block.
3079 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3082 // This is a known uniform.
3085 // Insert all operands.
3086 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3091 /// \brief Analyses memory accesses in a loop.
3093 /// Checks whether run time pointer checks are needed and builds sets for data
3094 /// dependence checking.
3095 class AccessAnalysis {
3097 /// \brief Read or write access location.
3098 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3099 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3101 /// \brief Set of potential dependent memory accesses.
3102 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3104 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3105 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3106 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3108 /// \brief Register a load and whether it is only read from.
3109 void addLoad(Value *Ptr, bool IsReadOnly) {
3110 Accesses.insert(MemAccessInfo(Ptr, false));
3112 ReadOnlyPtr.insert(Ptr);
3115 /// \brief Register a store.
3116 void addStore(Value *Ptr) {
3117 Accesses.insert(MemAccessInfo(Ptr, true));
3120 /// \brief Check whether we can check the pointers at runtime for
3121 /// non-intersection.
3122 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3123 unsigned &NumComparisons, ScalarEvolution *SE,
3124 Loop *TheLoop, bool ShouldCheckStride = false);
3126 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3127 /// and builds sets of dependent accesses.
3128 void buildDependenceSets() {
3129 // Process read-write pointers first.
3130 processMemAccesses(false);
3131 // Next, process read pointers.
3132 processMemAccesses(true);
3135 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3137 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3138 void resetDepChecks() { CheckDeps.clear(); }
3140 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3143 typedef SetVector<MemAccessInfo> PtrAccessSet;
3144 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3146 /// \brief Go over all memory access or only the deferred ones if
3147 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3148 /// and build sets of dependency check candidates.
3149 void processMemAccesses(bool UseDeferred);
3151 /// Set of all accesses.
3152 PtrAccessSet Accesses;
3154 /// Set of access to check after all writes have been processed.
3155 PtrAccessSet DeferredAccesses;
3157 /// Map of pointers to last access encountered.
3158 UnderlyingObjToAccessMap ObjToLastAccess;
3160 /// Set of accesses that need a further dependence check.
3161 MemAccessInfoSet CheckDeps;
3163 /// Set of pointers that are read only.
3164 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3166 /// Set of underlying objects already written to.
3167 SmallPtrSet<Value*, 16> WriteObjects;
3171 /// Sets of potentially dependent accesses - members of one set share an
3172 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3173 /// dependence check.
3174 DepCandidates &DepCands;
3176 bool AreAllWritesIdentified;
3177 bool AreAllReadsIdentified;
3178 bool IsRTCheckNeeded;
3181 } // end anonymous namespace
3183 /// \brief Check whether a pointer can participate in a runtime bounds check.
3184 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3185 const SCEV *PtrScev = SE->getSCEV(Ptr);
3186 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3190 return AR->isAffine();
3193 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3194 /// the address space.
3195 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3198 bool AccessAnalysis::canCheckPtrAtRT(
3199 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3200 unsigned &NumComparisons, ScalarEvolution *SE,
3201 Loop *TheLoop, bool ShouldCheckStride) {
3202 // Find pointers with computable bounds. We are going to use this information
3203 // to place a runtime bound check.
3204 unsigned NumReadPtrChecks = 0;
3205 unsigned NumWritePtrChecks = 0;
3206 bool CanDoRT = true;
3208 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3209 // We assign consecutive id to access from different dependence sets.
3210 // Accesses within the same set don't need a runtime check.
3211 unsigned RunningDepId = 1;
3212 DenseMap<Value *, unsigned> DepSetId;
3214 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3216 const MemAccessInfo &Access = *AI;
3217 Value *Ptr = Access.getPointer();
3218 bool IsWrite = Access.getInt();
3220 // Just add write checks if we have both.
3221 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3225 ++NumWritePtrChecks;
3229 if (hasComputableBounds(SE, Ptr) &&
3230 // When we run after a failing dependency check we have to make sure we
3231 // don't have wrapping pointers.
3232 (!ShouldCheckStride || isStridedPtr(SE, DL, Ptr, TheLoop) == 1)) {
3233 // The id of the dependence set.
3236 if (IsDepCheckNeeded) {
3237 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3238 unsigned &LeaderId = DepSetId[Leader];
3240 LeaderId = RunningDepId++;
3243 // Each access has its own dependence set.
3244 DepId = RunningDepId++;
3246 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3248 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3254 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3255 NumComparisons = 0; // Only one dependence set.
3257 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3258 NumWritePtrChecks - 1));
3261 // If the pointers that we would use for the bounds comparison have different
3262 // address spaces, assume the values aren't directly comparable, so we can't
3263 // use them for the runtime check. We also have to assume they could
3264 // overlap. In the future there should be metadata for whether address spaces
3266 unsigned NumPointers = RtCheck.Pointers.size();
3267 for (unsigned i = 0; i < NumPointers; ++i) {
3268 for (unsigned j = i + 1; j < NumPointers; ++j) {
3269 // Only need to check pointers between two different dependency sets.
3270 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3273 Value *PtrI = RtCheck.Pointers[i];
3274 Value *PtrJ = RtCheck.Pointers[j];
3276 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3277 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3279 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3280 " different address spaces\n");
3289 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3290 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3293 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3294 // We process the set twice: first we process read-write pointers, last we
3295 // process read-only pointers. This allows us to skip dependence tests for
3296 // read-only pointers.
3298 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3299 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3300 const MemAccessInfo &Access = *AI;
3301 Value *Ptr = Access.getPointer();
3302 bool IsWrite = Access.getInt();
3304 DepCands.insert(Access);
3306 // Memorize read-only pointers for later processing and skip them in the
3307 // first round (they need to be checked after we have seen all write
3308 // pointers). Note: we also mark pointer that are not consecutive as
3309 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3310 // second check for "!IsWrite".
3311 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3312 if (!UseDeferred && IsReadOnlyPtr) {
3313 DeferredAccesses.insert(Access);
3317 bool NeedDepCheck = false;
3318 // Check whether there is the possiblity of dependency because of underlying
3319 // objects being the same.
3320 typedef SmallVector<Value*, 16> ValueVector;
3321 ValueVector TempObjects;
3322 GetUnderlyingObjects(Ptr, TempObjects, DL);
3323 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3325 Value *UnderlyingObj = *UI;
3327 // If this is a write then it needs to be an identified object. If this a
3328 // read and all writes (so far) are identified function scope objects we
3329 // don't need an identified underlying object but only an Argument (the
3330 // next write is going to invalidate this assumption if it is
3332 // This is a micro-optimization for the case where all writes are
3333 // identified and we have one argument pointer.
3334 // Otherwise, we do need a runtime check.
3335 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3336 (!IsWrite && (!AreAllWritesIdentified ||
3337 !isa<Argument>(UnderlyingObj)) &&
3338 !isIdentifiedObject(UnderlyingObj))) {
3339 DEBUG(dbgs() << "LV: Found an unidentified " <<
3340 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3342 IsRTCheckNeeded = (IsRTCheckNeeded ||
3343 !isIdentifiedObject(UnderlyingObj) ||
3344 !AreAllReadsIdentified);
3347 AreAllWritesIdentified = false;
3349 AreAllReadsIdentified = false;
3352 // If this is a write - check other reads and writes for conflicts. If
3353 // this is a read only check other writes for conflicts (but only if there
3354 // is no other write to the ptr - this is an optimization to catch "a[i] =
3355 // a[i] + " without having to do a dependence check).
3356 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3357 NeedDepCheck = true;
3360 WriteObjects.insert(UnderlyingObj);
3362 // Create sets of pointers connected by shared underlying objects.
3363 UnderlyingObjToAccessMap::iterator Prev =
3364 ObjToLastAccess.find(UnderlyingObj);
3365 if (Prev != ObjToLastAccess.end())
3366 DepCands.unionSets(Access, Prev->second);
3368 ObjToLastAccess[UnderlyingObj] = Access;
3372 CheckDeps.insert(Access);
3377 /// \brief Checks memory dependences among accesses to the same underlying
3378 /// object to determine whether there vectorization is legal or not (and at
3379 /// which vectorization factor).
3381 /// This class works under the assumption that we already checked that memory
3382 /// locations with different underlying pointers are "must-not alias".
3383 /// We use the ScalarEvolution framework to symbolically evalutate access
3384 /// functions pairs. Since we currently don't restructure the loop we can rely
3385 /// on the program order of memory accesses to determine their safety.
3386 /// At the moment we will only deem accesses as safe for:
3387 /// * A negative constant distance assuming program order.
3389 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3390 /// a[i] = tmp; y = a[i];
3392 /// The latter case is safe because later checks guarantuee that there can't
3393 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3394 /// the same variable: a header phi can only be an induction or a reduction, a
3395 /// reduction can't have a memory sink, an induction can't have a memory
3396 /// source). This is important and must not be violated (or we have to
3397 /// resort to checking for cycles through memory).
3399 /// * A positive constant distance assuming program order that is bigger
3400 /// than the biggest memory access.
3402 /// tmp = a[i] OR b[i] = x
3403 /// a[i+2] = tmp y = b[i+2];
3405 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3407 /// * Zero distances and all accesses have the same size.
3409 class MemoryDepChecker {
3411 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3412 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3414 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3415 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3416 ShouldRetryWithRuntimeCheck(false) {}
3418 /// \brief Register the location (instructions are given increasing numbers)
3419 /// of a write access.
3420 void addAccess(StoreInst *SI) {
3421 Value *Ptr = SI->getPointerOperand();
3422 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3423 InstMap.push_back(SI);
3427 /// \brief Register the location (instructions are given increasing numbers)
3428 /// of a write access.
3429 void addAccess(LoadInst *LI) {
3430 Value *Ptr = LI->getPointerOperand();
3431 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3432 InstMap.push_back(LI);
3436 /// \brief Check whether the dependencies between the accesses are safe.
3438 /// Only checks sets with elements in \p CheckDeps.
3439 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3440 MemAccessInfoSet &CheckDeps);
3442 /// \brief The maximum number of bytes of a vector register we can vectorize
3443 /// the accesses safely with.
3444 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3446 /// \brief In same cases when the dependency check fails we can still
3447 /// vectorize the loop with a dynamic array access check.
3448 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3451 ScalarEvolution *SE;
3453 const Loop *InnermostLoop;
3455 /// \brief Maps access locations (ptr, read/write) to program order.
3456 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3458 /// \brief Memory access instructions in program order.
3459 SmallVector<Instruction *, 16> InstMap;
3461 /// \brief The program order index to be used for the next instruction.
3464 // We can access this many bytes in parallel safely.
3465 unsigned MaxSafeDepDistBytes;
3467 /// \brief If we see a non constant dependence distance we can still try to
3468 /// vectorize this loop with runtime checks.
3469 bool ShouldRetryWithRuntimeCheck;
3471 /// \brief Check whether there is a plausible dependence between the two
3474 /// Access \p A must happen before \p B in program order. The two indices
3475 /// identify the index into the program order map.
3477 /// This function checks whether there is a plausible dependence (or the
3478 /// absence of such can't be proved) between the two accesses. If there is a
3479 /// plausible dependence but the dependence distance is bigger than one
3480 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3481 /// distance is smaller than any other distance encountered so far).
3482 /// Otherwise, this function returns true signaling a possible dependence.
3483 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3484 const MemAccessInfo &B, unsigned BIdx);
3486 /// \brief Check whether the data dependence could prevent store-load
3488 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3491 } // end anonymous namespace
3493 static bool isInBoundsGep(Value *Ptr) {
3494 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3495 return GEP->isInBounds();
3499 /// \brief Check whether the access through \p Ptr has a constant stride.
3500 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3502 const Type *Ty = Ptr->getType();
3503 assert(Ty->isPointerTy() && "Unexpected non ptr");
3505 // Make sure that the pointer does not point to aggregate types.
3506 const PointerType *PtrTy = cast<PointerType>(Ty);
3507 if (PtrTy->getElementType()->isAggregateType()) {
3508 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3513 const SCEV *PtrScev = SE->getSCEV(Ptr);
3514 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3516 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3517 << *Ptr << " SCEV: " << *PtrScev << "\n");
3521 // The accesss function must stride over the innermost loop.
3522 if (Lp != AR->getLoop()) {
3523 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3524 *Ptr << " SCEV: " << *PtrScev << "\n");
3527 // The address calculation must not wrap. Otherwise, a dependence could be
3529 // An inbounds getelementptr that is a AddRec with a unit stride
3530 // cannot wrap per definition. The unit stride requirement is checked later.
3531 // An getelementptr without an inbounds attribute and unit stride would have
3532 // to access the pointer value "0" which is undefined behavior in address
3533 // space 0, therefore we can also vectorize this case.
3534 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3535 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3536 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3537 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3538 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3539 << *Ptr << " SCEV: " << *PtrScev << "\n");
3543 // Check the step is constant.
3544 const SCEV *Step = AR->getStepRecurrence(*SE);
3546 // Calculate the pointer stride and check if it is consecutive.
3547 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3549 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3550 " SCEV: " << *PtrScev << "\n");
3554 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3555 const APInt &APStepVal = C->getValue()->getValue();
3557 // Huge step value - give up.
3558 if (APStepVal.getBitWidth() > 64)
3561 int64_t StepVal = APStepVal.getSExtValue();
3564 int64_t Stride = StepVal / Size;
3565 int64_t Rem = StepVal % Size;
3569 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3570 // know we can't "wrap around the address space". In case of address space
3571 // zero we know that this won't happen without triggering undefined behavior.
3572 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3573 Stride != 1 && Stride != -1)
3579 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3580 unsigned TypeByteSize) {
3581 // If loads occur at a distance that is not a multiple of a feasible vector
3582 // factor store-load forwarding does not take place.
3583 // Positive dependences might cause troubles because vectorizing them might
3584 // prevent store-load forwarding making vectorized code run a lot slower.
3585 // a[i] = a[i-3] ^ a[i-8];
3586 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3587 // hence on your typical architecture store-load forwarding does not take
3588 // place. Vectorizing in such cases does not make sense.
3589 // Store-load forwarding distance.
3590 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3591 // Maximum vector factor.
3592 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3593 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3594 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3596 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3598 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3599 MaxVFWithoutSLForwardIssues = (vf >>=1);
3604 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3605 DEBUG(dbgs() << "LV: Distance " << Distance <<
3606 " that could cause a store-load forwarding conflict\n");
3610 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3611 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3612 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3616 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3617 const MemAccessInfo &B, unsigned BIdx) {
3618 assert (AIdx < BIdx && "Must pass arguments in program order");
3620 Value *APtr = A.getPointer();
3621 Value *BPtr = B.getPointer();
3622 bool AIsWrite = A.getInt();
3623 bool BIsWrite = B.getInt();
3625 // Two reads are independent.
3626 if (!AIsWrite && !BIsWrite)
3629 const SCEV *AScev = SE->getSCEV(APtr);
3630 const SCEV *BScev = SE->getSCEV(BPtr);
3632 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3633 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3635 const SCEV *Src = AScev;
3636 const SCEV *Sink = BScev;
3638 // If the induction step is negative we have to invert source and sink of the
3640 if (StrideAPtr < 0) {
3643 std::swap(APtr, BPtr);
3644 std::swap(Src, Sink);
3645 std::swap(AIsWrite, BIsWrite);
3646 std::swap(AIdx, BIdx);
3647 std::swap(StrideAPtr, StrideBPtr);
3650 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3652 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3653 << "(Induction step: " << StrideAPtr << ")\n");
3654 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3655 << *InstMap[BIdx] << ": " << *Dist << "\n");
3657 // Need consecutive accesses. We don't want to vectorize
3658 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3659 // the address space.
3660 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3661 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3665 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3667 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3668 ShouldRetryWithRuntimeCheck = true;
3672 Type *ATy = APtr->getType()->getPointerElementType();
3673 Type *BTy = BPtr->getType()->getPointerElementType();
3674 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3676 // Negative distances are not plausible dependencies.
3677 const APInt &Val = C->getValue()->getValue();
3678 if (Val.isNegative()) {
3679 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3680 if (IsTrueDataDependence &&
3681 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3685 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3689 // Write to the same location with the same size.
3690 // Could be improved to assert type sizes are the same (i32 == float, etc).
3694 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
3698 assert(Val.isStrictlyPositive() && "Expect a positive value");
3700 // Positive distance bigger than max vectorization factor.
3703 "LV: ReadWrite-Write positive dependency with different types\n");
3707 unsigned Distance = (unsigned) Val.getZExtValue();
3709 // Bail out early if passed-in parameters make vectorization not feasible.
3710 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3711 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3713 // The distance must be bigger than the size needed for a vectorized version
3714 // of the operation and the size of the vectorized operation must not be
3715 // bigger than the currrent maximum size.
3716 if (Distance < 2*TypeByteSize ||
3717 2*TypeByteSize > MaxSafeDepDistBytes ||
3718 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3719 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3720 << Val.getSExtValue() << '\n');
3724 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3725 Distance : MaxSafeDepDistBytes;
3727 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3728 if (IsTrueDataDependence &&
3729 couldPreventStoreLoadForward(Distance, TypeByteSize))
3732 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3733 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
3739 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3740 MemAccessInfoSet &CheckDeps) {
3742 MaxSafeDepDistBytes = -1U;
3743 while (!CheckDeps.empty()) {
3744 MemAccessInfo CurAccess = *CheckDeps.begin();
3746 // Get the relevant memory access set.
3747 EquivalenceClasses<MemAccessInfo>::iterator I =
3748 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3750 // Check accesses within this set.
3751 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3752 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3754 // Check every access pair.
3756 CheckDeps.erase(*AI);
3757 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3759 // Check every accessing instruction pair in program order.
3760 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3761 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3762 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3763 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3764 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3766 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3777 bool LoopVectorizationLegality::canVectorizeMemory() {
3779 typedef SmallVector<Value*, 16> ValueVector;
3780 typedef SmallPtrSet<Value*, 16> ValueSet;
3782 // Holds the Load and Store *instructions*.
3786 // Holds all the different accesses in the loop.
3787 unsigned NumReads = 0;
3788 unsigned NumReadWrites = 0;
3790 PtrRtCheck.Pointers.clear();
3791 PtrRtCheck.Need = false;
3793 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3794 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3797 for (Loop::block_iterator bb = TheLoop->block_begin(),
3798 be = TheLoop->block_end(); bb != be; ++bb) {
3800 // Scan the BB and collect legal loads and stores.
3801 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3804 // If this is a load, save it. If this instruction can read from memory
3805 // but is not a load, then we quit. Notice that we don't handle function
3806 // calls that read or write.
3807 if (it->mayReadFromMemory()) {
3808 // Many math library functions read the rounding mode. We will only
3809 // vectorize a loop if it contains known function calls that don't set
3810 // the flag. Therefore, it is safe to ignore this read from memory.
3811 CallInst *Call = dyn_cast<CallInst>(it);
3812 if (Call && getIntrinsicIDForCall(Call, TLI))
3815 LoadInst *Ld = dyn_cast<LoadInst>(it);
3816 if (!Ld) return false;
3817 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3818 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3821 Loads.push_back(Ld);
3822 DepChecker.addAccess(Ld);
3826 // Save 'store' instructions. Abort if other instructions write to memory.
3827 if (it->mayWriteToMemory()) {
3828 StoreInst *St = dyn_cast<StoreInst>(it);
3829 if (!St) return false;
3830 if (!St->isSimple() && !IsAnnotatedParallel) {
3831 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3834 Stores.push_back(St);
3835 DepChecker.addAccess(St);
3840 // Now we have two lists that hold the loads and the stores.
3841 // Next, we find the pointers that they use.
3843 // Check if we see any stores. If there are no stores, then we don't
3844 // care if the pointers are *restrict*.
3845 if (!Stores.size()) {
3846 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3850 AccessAnalysis::DepCandidates DependentAccesses;
3851 AccessAnalysis Accesses(DL, DependentAccesses);
3853 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3854 // multiple times on the same object. If the ptr is accessed twice, once
3855 // for read and once for write, it will only appear once (on the write
3856 // list). This is okay, since we are going to check for conflicts between
3857 // writes and between reads and writes, but not between reads and reads.
3860 ValueVector::iterator I, IE;
3861 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3862 StoreInst *ST = cast<StoreInst>(*I);
3863 Value* Ptr = ST->getPointerOperand();
3865 if (isUniform(Ptr)) {
3866 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3870 // If we did *not* see this pointer before, insert it to the read-write
3871 // list. At this phase it is only a 'write' list.
3872 if (Seen.insert(Ptr)) {
3874 Accesses.addStore(Ptr);
3878 if (IsAnnotatedParallel) {
3880 << "LV: A loop annotated parallel, ignore memory dependency "
3885 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3886 LoadInst *LD = cast<LoadInst>(*I);
3887 Value* Ptr = LD->getPointerOperand();
3888 // If we did *not* see this pointer before, insert it to the
3889 // read list. If we *did* see it before, then it is already in
3890 // the read-write list. This allows us to vectorize expressions
3891 // such as A[i] += x; Because the address of A[i] is a read-write
3892 // pointer. This only works if the index of A[i] is consecutive.
3893 // If the address of i is unknown (for example A[B[i]]) then we may
3894 // read a few words, modify, and write a few words, and some of the
3895 // words may be written to the same address.
3896 bool IsReadOnlyPtr = false;
3897 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3899 IsReadOnlyPtr = true;
3901 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3904 // If we write (or read-write) to a single destination and there are no
3905 // other reads in this loop then is it safe to vectorize.
3906 if (NumReadWrites == 1 && NumReads == 0) {
3907 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3911 // Build dependence sets and check whether we need a runtime pointer bounds
3913 Accesses.buildDependenceSets();
3914 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3916 // Find pointers with computable bounds. We are going to use this information
3917 // to place a runtime bound check.
3918 unsigned NumComparisons = 0;
3919 bool CanDoRT = false;
3921 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3924 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3925 " pointer comparisons.\n");
3927 // If we only have one set of dependences to check pointers among we don't
3928 // need a runtime check.
3929 if (NumComparisons == 0 && NeedRTCheck)
3930 NeedRTCheck = false;
3932 // Check that we did not collect too many pointers or found an unsizeable
3934 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3940 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3943 if (NeedRTCheck && !CanDoRT) {
3944 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3945 "the array bounds.\n");
3950 PtrRtCheck.Need = NeedRTCheck;
3952 bool CanVecMem = true;
3953 if (Accesses.isDependencyCheckNeeded()) {
3954 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3955 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3956 Accesses.getDependenciesToCheck());
3957 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3959 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
3960 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
3963 // Clear the dependency checks. We assume they are not needed.
3964 Accesses.resetDepChecks();
3967 PtrRtCheck.Need = true;
3969 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
3971 // Check that we did not collect too many pointers or found an unsizeable
3973 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3974 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
3983 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
3984 " need a runtime memory check.\n");
3989 static bool hasMultipleUsesOf(Instruction *I,
3990 SmallPtrSet<Instruction *, 8> &Insts) {
3991 unsigned NumUses = 0;
3992 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3993 if (Insts.count(dyn_cast<Instruction>(*Use)))
4002 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4003 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4004 if (!Set.count(dyn_cast<Instruction>(*Use)))
4009 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4010 ReductionKind Kind) {
4011 if (Phi->getNumIncomingValues() != 2)
4014 // Reduction variables are only found in the loop header block.
4015 if (Phi->getParent() != TheLoop->getHeader())
4018 // Obtain the reduction start value from the value that comes from the loop
4020 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4022 // ExitInstruction is the single value which is used outside the loop.
4023 // We only allow for a single reduction value to be used outside the loop.
4024 // This includes users of the reduction, variables (which form a cycle
4025 // which ends in the phi node).
4026 Instruction *ExitInstruction = 0;
4027 // Indicates that we found a reduction operation in our scan.
4028 bool FoundReduxOp = false;
4030 // We start with the PHI node and scan for all of the users of this
4031 // instruction. All users must be instructions that can be used as reduction
4032 // variables (such as ADD). We must have a single out-of-block user. The cycle
4033 // must include the original PHI.
4034 bool FoundStartPHI = false;
4036 // To recognize min/max patterns formed by a icmp select sequence, we store
4037 // the number of instruction we saw from the recognized min/max pattern,
4038 // to make sure we only see exactly the two instructions.
4039 unsigned NumCmpSelectPatternInst = 0;
4040 ReductionInstDesc ReduxDesc(false, 0);
4042 SmallPtrSet<Instruction *, 8> VisitedInsts;
4043 SmallVector<Instruction *, 8> Worklist;
4044 Worklist.push_back(Phi);
4045 VisitedInsts.insert(Phi);
4047 // A value in the reduction can be used:
4048 // - By the reduction:
4049 // - Reduction operation:
4050 // - One use of reduction value (safe).
4051 // - Multiple use of reduction value (not safe).
4053 // - All uses of the PHI must be the reduction (safe).
4054 // - Otherwise, not safe.
4055 // - By one instruction outside of the loop (safe).
4056 // - By further instructions outside of the loop (not safe).
4057 // - By an instruction that is not part of the reduction (not safe).
4059 // * An instruction type other than PHI or the reduction operation.
4060 // * A PHI in the header other than the initial PHI.
4061 while (!Worklist.empty()) {
4062 Instruction *Cur = Worklist.back();
4063 Worklist.pop_back();
4066 // If the instruction has no users then this is a broken chain and can't be
4067 // a reduction variable.
4068 if (Cur->use_empty())
4071 bool IsAPhi = isa<PHINode>(Cur);
4073 // A header PHI use other than the original PHI.
4074 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4077 // Reductions of instructions such as Div, and Sub is only possible if the
4078 // LHS is the reduction variable.
4079 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4080 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4081 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4084 // Any reduction instruction must be of one of the allowed kinds.
4085 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4086 if (!ReduxDesc.IsReduction)
4089 // A reduction operation must only have one use of the reduction value.
4090 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4091 hasMultipleUsesOf(Cur, VisitedInsts))
4094 // All inputs to a PHI node must be a reduction value.
4095 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4098 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4099 isa<SelectInst>(Cur)))
4100 ++NumCmpSelectPatternInst;
4101 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4102 isa<SelectInst>(Cur)))
4103 ++NumCmpSelectPatternInst;
4105 // Check whether we found a reduction operator.
4106 FoundReduxOp |= !IsAPhi;
4108 // Process users of current instruction. Push non PHI nodes after PHI nodes
4109 // onto the stack. This way we are going to have seen all inputs to PHI
4110 // nodes once we get to them.
4111 SmallVector<Instruction *, 8> NonPHIs;
4112 SmallVector<Instruction *, 8> PHIs;
4113 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4115 Instruction *Usr = cast<Instruction>(*UI);
4117 // Check if we found the exit user.
4118 BasicBlock *Parent = Usr->getParent();
4119 if (!TheLoop->contains(Parent)) {
4120 // Exit if you find multiple outside users or if the header phi node is
4121 // being used. In this case the user uses the value of the previous
4122 // iteration, in which case we would loose "VF-1" iterations of the
4123 // reduction operation if we vectorize.
4124 if (ExitInstruction != 0 || Cur == Phi)
4127 // The instruction used by an outside user must be the last instruction
4128 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4129 // operations on the value.
4130 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4133 ExitInstruction = Cur;
4137 // Process instructions only once (termination).
4138 if (VisitedInsts.insert(Usr)) {
4139 if (isa<PHINode>(Usr))
4140 PHIs.push_back(Usr);
4142 NonPHIs.push_back(Usr);
4144 // Remember that we completed the cycle.
4146 FoundStartPHI = true;
4148 Worklist.append(PHIs.begin(), PHIs.end());
4149 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4152 // This means we have seen one but not the other instruction of the
4153 // pattern or more than just a select and cmp.
4154 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4155 NumCmpSelectPatternInst != 2)
4158 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4161 // We found a reduction var if we have reached the original phi node and we
4162 // only have a single instruction with out-of-loop users.
4164 // This instruction is allowed to have out-of-loop users.
4165 AllowedExit.insert(ExitInstruction);
4167 // Save the description of this reduction variable.
4168 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4169 ReduxDesc.MinMaxKind);
4170 Reductions[Phi] = RD;
4171 // We've ended the cycle. This is a reduction variable if we have an
4172 // outside user and it has a binary op.
4177 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4178 /// pattern corresponding to a min(X, Y) or max(X, Y).
4179 LoopVectorizationLegality::ReductionInstDesc
4180 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4181 ReductionInstDesc &Prev) {
4183 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4184 "Expect a select instruction");
4185 Instruction *Cmp = 0;
4186 SelectInst *Select = 0;
4188 // We must handle the select(cmp()) as a single instruction. Advance to the
4190 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4191 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4192 return ReductionInstDesc(false, I);
4193 return ReductionInstDesc(Select, Prev.MinMaxKind);
4196 // Only handle single use cases for now.
4197 if (!(Select = dyn_cast<SelectInst>(I)))
4198 return ReductionInstDesc(false, I);
4199 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4200 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4201 return ReductionInstDesc(false, I);
4202 if (!Cmp->hasOneUse())
4203 return ReductionInstDesc(false, I);
4208 // Look for a min/max pattern.
4209 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4210 return ReductionInstDesc(Select, MRK_UIntMin);
4211 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4212 return ReductionInstDesc(Select, MRK_UIntMax);
4213 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4214 return ReductionInstDesc(Select, MRK_SIntMax);
4215 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4216 return ReductionInstDesc(Select, MRK_SIntMin);
4217 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4218 return ReductionInstDesc(Select, MRK_FloatMin);
4219 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4220 return ReductionInstDesc(Select, MRK_FloatMax);
4221 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4222 return ReductionInstDesc(Select, MRK_FloatMin);
4223 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4224 return ReductionInstDesc(Select, MRK_FloatMax);
4226 return ReductionInstDesc(false, I);
4229 LoopVectorizationLegality::ReductionInstDesc
4230 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4232 ReductionInstDesc &Prev) {
4233 bool FP = I->getType()->isFloatingPointTy();
4234 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4235 switch (I->getOpcode()) {
4237 return ReductionInstDesc(false, I);
4238 case Instruction::PHI:
4239 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4240 Kind != RK_FloatMinMax))
4241 return ReductionInstDesc(false, I);
4242 return ReductionInstDesc(I, Prev.MinMaxKind);
4243 case Instruction::Sub:
4244 case Instruction::Add:
4245 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4246 case Instruction::Mul:
4247 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4248 case Instruction::And:
4249 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4250 case Instruction::Or:
4251 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4252 case Instruction::Xor:
4253 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4254 case Instruction::FMul:
4255 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4256 case Instruction::FAdd:
4257 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4258 case Instruction::FCmp:
4259 case Instruction::ICmp:
4260 case Instruction::Select:
4261 if (Kind != RK_IntegerMinMax &&
4262 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4263 return ReductionInstDesc(false, I);
4264 return isMinMaxSelectCmpPattern(I, Prev);
4268 LoopVectorizationLegality::InductionKind
4269 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4270 Type *PhiTy = Phi->getType();
4271 // We only handle integer and pointer inductions variables.
4272 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4273 return IK_NoInduction;
4275 // Check that the PHI is consecutive.
4276 const SCEV *PhiScev = SE->getSCEV(Phi);
4277 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4279 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4280 return IK_NoInduction;
4282 const SCEV *Step = AR->getStepRecurrence(*SE);
4284 // Integer inductions need to have a stride of one.
4285 if (PhiTy->isIntegerTy()) {
4287 return IK_IntInduction;
4288 if (Step->isAllOnesValue())
4289 return IK_ReverseIntInduction;
4290 return IK_NoInduction;
4293 // Calculate the pointer stride and check if it is consecutive.
4294 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4296 return IK_NoInduction;
4298 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4299 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4300 if (C->getValue()->equalsInt(Size))
4301 return IK_PtrInduction;
4302 else if (C->getValue()->equalsInt(0 - Size))
4303 return IK_ReversePtrInduction;
4305 return IK_NoInduction;
4308 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4309 Value *In0 = const_cast<Value*>(V);
4310 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4314 return Inductions.count(PN);
4317 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4318 assert(TheLoop->contains(BB) && "Unknown block used");
4320 // Blocks that do not dominate the latch need predication.
4321 BasicBlock* Latch = TheLoop->getLoopLatch();
4322 return !DT->dominates(BB, Latch);
4325 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4326 SmallPtrSet<Value *, 8>& SafePtrs) {
4327 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4328 // We might be able to hoist the load.
4329 if (it->mayReadFromMemory()) {
4330 LoadInst *LI = dyn_cast<LoadInst>(it);
4331 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4335 // We don't predicate stores at the moment.
4336 if (it->mayWriteToMemory() || it->mayThrow())
4339 // The instructions below can trap.
4340 switch (it->getOpcode()) {
4342 case Instruction::UDiv:
4343 case Instruction::SDiv:
4344 case Instruction::URem:
4345 case Instruction::SRem:
4353 LoopVectorizationCostModel::VectorizationFactor
4354 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4356 // Width 1 means no vectorize
4357 VectorizationFactor Factor = { 1U, 0U };
4358 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4359 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4363 // Find the trip count.
4364 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4365 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4367 unsigned WidestType = getWidestType();
4368 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4369 unsigned MaxSafeDepDist = -1U;
4370 if (Legal->getMaxSafeDepDistBytes() != -1U)
4371 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4372 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4373 WidestRegister : MaxSafeDepDist);
4374 unsigned MaxVectorSize = WidestRegister / WidestType;
4375 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4376 DEBUG(dbgs() << "LV: The Widest register is: "
4377 << WidestRegister << " bits.\n");
4379 if (MaxVectorSize == 0) {
4380 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4384 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4385 " into one vector!");
4387 unsigned VF = MaxVectorSize;
4389 // If we optimize the program for size, avoid creating the tail loop.
4391 // If we are unable to calculate the trip count then don't try to vectorize.
4393 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4397 // Find the maximum SIMD width that can fit within the trip count.
4398 VF = TC % MaxVectorSize;
4403 // If the trip count that we found modulo the vectorization factor is not
4404 // zero then we require a tail.
4406 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4412 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4413 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4415 Factor.Width = UserVF;
4419 float Cost = expectedCost(1);
4421 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
4422 for (unsigned i=2; i <= VF; i*=2) {
4423 // Notice that the vector loop needs to be executed less times, so
4424 // we need to divide the cost of the vector loops by the width of
4425 // the vector elements.
4426 float VectorCost = expectedCost(i) / (float)i;
4427 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4428 (int)VectorCost << ".\n");
4429 if (VectorCost < Cost) {
4435 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4436 Factor.Width = Width;
4437 Factor.Cost = Width * Cost;
4441 unsigned LoopVectorizationCostModel::getWidestType() {
4442 unsigned MaxWidth = 8;
4445 for (Loop::block_iterator bb = TheLoop->block_begin(),
4446 be = TheLoop->block_end(); bb != be; ++bb) {
4447 BasicBlock *BB = *bb;
4449 // For each instruction in the loop.
4450 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4451 Type *T = it->getType();
4453 // Only examine Loads, Stores and PHINodes.
4454 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4457 // Examine PHI nodes that are reduction variables.
4458 if (PHINode *PN = dyn_cast<PHINode>(it))
4459 if (!Legal->getReductionVars()->count(PN))
4462 // Examine the stored values.
4463 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4464 T = ST->getValueOperand()->getType();
4466 // Ignore loaded pointer types and stored pointer types that are not
4467 // consecutive. However, we do want to take consecutive stores/loads of
4468 // pointer vectors into account.
4469 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4472 MaxWidth = std::max(MaxWidth,
4473 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4481 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4484 unsigned LoopCost) {
4486 // -- The unroll heuristics --
4487 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4488 // There are many micro-architectural considerations that we can't predict
4489 // at this level. For example frontend pressure (on decode or fetch) due to
4490 // code size, or the number and capabilities of the execution ports.
4492 // We use the following heuristics to select the unroll factor:
4493 // 1. If the code has reductions the we unroll in order to break the cross
4494 // iteration dependency.
4495 // 2. If the loop is really small then we unroll in order to reduce the loop
4497 // 3. We don't unroll if we think that we will spill registers to memory due
4498 // to the increased register pressure.
4500 // Use the user preference, unless 'auto' is selected.
4504 // When we optimize for size we don't unroll.
4508 // We used the distance for the unroll factor.
4509 if (Legal->getMaxSafeDepDistBytes() != -1U)
4512 // Do not unroll loops with a relatively small trip count.
4513 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4514 TheLoop->getLoopLatch());
4515 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4518 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4519 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4520 " vector registers\n");
4522 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4523 // We divide by these constants so assume that we have at least one
4524 // instruction that uses at least one register.
4525 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4526 R.NumInstructions = std::max(R.NumInstructions, 1U);
4528 // We calculate the unroll factor using the following formula.
4529 // Subtract the number of loop invariants from the number of available
4530 // registers. These registers are used by all of the unrolled instances.
4531 // Next, divide the remaining registers by the number of registers that is
4532 // required by the loop, in order to estimate how many parallel instances
4533 // fit without causing spills.
4534 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4536 // Clamp the unroll factor ranges to reasonable factors.
4537 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4539 // If we did not calculate the cost for VF (because the user selected the VF)
4540 // then we calculate the cost of VF here.
4542 LoopCost = expectedCost(VF);
4544 // Clamp the calculated UF to be between the 1 and the max unroll factor
4545 // that the target allows.
4546 if (UF > MaxUnrollSize)
4551 bool HasReductions = Legal->getReductionVars()->size();
4553 // Decide if we want to unroll if we decided that it is legal to vectorize
4554 // but not profitable.
4556 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4557 LoopCost > SmallLoopCost)
4563 if (HasReductions) {
4564 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4568 // We want to unroll tiny loops in order to reduce the loop overhead.
4569 // We assume that the cost overhead is 1 and we use the cost model
4570 // to estimate the cost of the loop and unroll until the cost of the
4571 // loop overhead is about 5% of the cost of the loop.
4572 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4573 if (LoopCost < SmallLoopCost) {
4574 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4575 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4576 return std::min(NewUF, UF);
4579 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4583 LoopVectorizationCostModel::RegisterUsage
4584 LoopVectorizationCostModel::calculateRegisterUsage() {
4585 // This function calculates the register usage by measuring the highest number
4586 // of values that are alive at a single location. Obviously, this is a very
4587 // rough estimation. We scan the loop in a topological order in order and
4588 // assign a number to each instruction. We use RPO to ensure that defs are
4589 // met before their users. We assume that each instruction that has in-loop
4590 // users starts an interval. We record every time that an in-loop value is
4591 // used, so we have a list of the first and last occurrences of each
4592 // instruction. Next, we transpose this data structure into a multi map that
4593 // holds the list of intervals that *end* at a specific location. This multi
4594 // map allows us to perform a linear search. We scan the instructions linearly
4595 // and record each time that a new interval starts, by placing it in a set.
4596 // If we find this value in the multi-map then we remove it from the set.
4597 // The max register usage is the maximum size of the set.
4598 // We also search for instructions that are defined outside the loop, but are
4599 // used inside the loop. We need this number separately from the max-interval
4600 // usage number because when we unroll, loop-invariant values do not take
4602 LoopBlocksDFS DFS(TheLoop);
4606 R.NumInstructions = 0;
4608 // Each 'key' in the map opens a new interval. The values
4609 // of the map are the index of the 'last seen' usage of the
4610 // instruction that is the key.
4611 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4612 // Maps instruction to its index.
4613 DenseMap<unsigned, Instruction*> IdxToInstr;
4614 // Marks the end of each interval.
4615 IntervalMap EndPoint;
4616 // Saves the list of instruction indices that are used in the loop.
4617 SmallSet<Instruction*, 8> Ends;
4618 // Saves the list of values that are used in the loop but are
4619 // defined outside the loop, such as arguments and constants.
4620 SmallPtrSet<Value*, 8> LoopInvariants;
4623 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4624 be = DFS.endRPO(); bb != be; ++bb) {
4625 R.NumInstructions += (*bb)->size();
4626 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4628 Instruction *I = it;
4629 IdxToInstr[Index++] = I;
4631 // Save the end location of each USE.
4632 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4633 Value *U = I->getOperand(i);
4634 Instruction *Instr = dyn_cast<Instruction>(U);
4636 // Ignore non-instruction values such as arguments, constants, etc.
4637 if (!Instr) continue;
4639 // If this instruction is outside the loop then record it and continue.
4640 if (!TheLoop->contains(Instr)) {
4641 LoopInvariants.insert(Instr);
4645 // Overwrite previous end points.
4646 EndPoint[Instr] = Index;
4652 // Saves the list of intervals that end with the index in 'key'.
4653 typedef SmallVector<Instruction*, 2> InstrList;
4654 DenseMap<unsigned, InstrList> TransposeEnds;
4656 // Transpose the EndPoints to a list of values that end at each index.
4657 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4659 TransposeEnds[it->second].push_back(it->first);
4661 SmallSet<Instruction*, 8> OpenIntervals;
4662 unsigned MaxUsage = 0;
4665 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4666 for (unsigned int i = 0; i < Index; ++i) {
4667 Instruction *I = IdxToInstr[i];
4668 // Ignore instructions that are never used within the loop.
4669 if (!Ends.count(I)) continue;
4671 // Remove all of the instructions that end at this location.
4672 InstrList &List = TransposeEnds[i];
4673 for (unsigned int j=0, e = List.size(); j < e; ++j)
4674 OpenIntervals.erase(List[j]);
4676 // Count the number of live interals.
4677 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4679 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4680 OpenIntervals.size() << '\n');
4682 // Add the current instruction to the list of open intervals.
4683 OpenIntervals.insert(I);
4686 unsigned Invariant = LoopInvariants.size();
4687 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4688 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4689 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4691 R.LoopInvariantRegs = Invariant;
4692 R.MaxLocalUsers = MaxUsage;
4696 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4700 for (Loop::block_iterator bb = TheLoop->block_begin(),
4701 be = TheLoop->block_end(); bb != be; ++bb) {
4702 unsigned BlockCost = 0;
4703 BasicBlock *BB = *bb;
4705 // For each instruction in the old loop.
4706 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4707 // Skip dbg intrinsics.
4708 if (isa<DbgInfoIntrinsic>(it))
4711 unsigned C = getInstructionCost(it, VF);
4713 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4714 VF << " For instruction: " << *it << '\n');
4717 // We assume that if-converted blocks have a 50% chance of being executed.
4718 // When the code is scalar then some of the blocks are avoided due to CF.
4719 // When the code is vectorized we execute all code paths.
4720 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4729 /// \brief Check whether the address computation for a non-consecutive memory
4730 /// access looks like an unlikely candidate for being merged into the indexing
4733 /// We look for a GEP which has one index that is an induction variable and all
4734 /// other indices are loop invariant. If the stride of this access is also
4735 /// within a small bound we decide that this address computation can likely be
4736 /// merged into the addressing mode.
4737 /// In all other cases, we identify the address computation as complex.
4738 static bool isLikelyComplexAddressComputation(Value *Ptr,
4739 LoopVectorizationLegality *Legal,
4740 ScalarEvolution *SE,
4741 const Loop *TheLoop) {
4742 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4746 // We are looking for a gep with all loop invariant indices except for one
4747 // which should be an induction variable.
4748 unsigned NumOperands = Gep->getNumOperands();
4749 for (unsigned i = 1; i < NumOperands; ++i) {
4750 Value *Opd = Gep->getOperand(i);
4751 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4752 !Legal->isInductionVariable(Opd))
4756 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4757 // can likely be merged into the address computation.
4758 unsigned MaxMergeDistance = 64;
4760 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4764 // Check the step is constant.
4765 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4766 // Calculate the pointer stride and check if it is consecutive.
4767 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4771 const APInt &APStepVal = C->getValue()->getValue();
4773 // Huge step value - give up.
4774 if (APStepVal.getBitWidth() > 64)
4777 int64_t StepVal = APStepVal.getSExtValue();
4779 return StepVal > MaxMergeDistance;
4783 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4784 // If we know that this instruction will remain uniform, check the cost of
4785 // the scalar version.
4786 if (Legal->isUniformAfterVectorization(I))
4789 Type *RetTy = I->getType();
4790 Type *VectorTy = ToVectorTy(RetTy, VF);
4792 // TODO: We need to estimate the cost of intrinsic calls.
4793 switch (I->getOpcode()) {
4794 case Instruction::GetElementPtr:
4795 // We mark this instruction as zero-cost because the cost of GEPs in
4796 // vectorized code depends on whether the corresponding memory instruction
4797 // is scalarized or not. Therefore, we handle GEPs with the memory
4798 // instruction cost.
4800 case Instruction::Br: {
4801 return TTI.getCFInstrCost(I->getOpcode());
4803 case Instruction::PHI:
4804 //TODO: IF-converted IFs become selects.
4806 case Instruction::Add:
4807 case Instruction::FAdd:
4808 case Instruction::Sub:
4809 case Instruction::FSub:
4810 case Instruction::Mul:
4811 case Instruction::FMul:
4812 case Instruction::UDiv:
4813 case Instruction::SDiv:
4814 case Instruction::FDiv:
4815 case Instruction::URem:
4816 case Instruction::SRem:
4817 case Instruction::FRem:
4818 case Instruction::Shl:
4819 case Instruction::LShr:
4820 case Instruction::AShr:
4821 case Instruction::And:
4822 case Instruction::Or:
4823 case Instruction::Xor: {
4824 // Certain instructions can be cheaper to vectorize if they have a constant
4825 // second vector operand. One example of this are shifts on x86.
4826 TargetTransformInfo::OperandValueKind Op1VK =
4827 TargetTransformInfo::OK_AnyValue;
4828 TargetTransformInfo::OperandValueKind Op2VK =
4829 TargetTransformInfo::OK_AnyValue;
4831 if (isa<ConstantInt>(I->getOperand(1)))
4832 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4834 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4836 case Instruction::Select: {
4837 SelectInst *SI = cast<SelectInst>(I);
4838 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4839 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4840 Type *CondTy = SI->getCondition()->getType();
4842 CondTy = VectorType::get(CondTy, VF);
4844 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4846 case Instruction::ICmp:
4847 case Instruction::FCmp: {
4848 Type *ValTy = I->getOperand(0)->getType();
4849 VectorTy = ToVectorTy(ValTy, VF);
4850 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4852 case Instruction::Store:
4853 case Instruction::Load: {
4854 StoreInst *SI = dyn_cast<StoreInst>(I);
4855 LoadInst *LI = dyn_cast<LoadInst>(I);
4856 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4858 VectorTy = ToVectorTy(ValTy, VF);
4860 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4861 unsigned AS = SI ? SI->getPointerAddressSpace() :
4862 LI->getPointerAddressSpace();
4863 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4864 // We add the cost of address computation here instead of with the gep
4865 // instruction because only here we know whether the operation is
4868 return TTI.getAddressComputationCost(VectorTy) +
4869 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4871 // Scalarized loads/stores.
4872 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4873 bool Reverse = ConsecutiveStride < 0;
4874 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4875 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4876 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4877 bool IsComplexComputation =
4878 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4880 // The cost of extracting from the value vector and pointer vector.
4881 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4882 for (unsigned i = 0; i < VF; ++i) {
4883 // The cost of extracting the pointer operand.
4884 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4885 // In case of STORE, the cost of ExtractElement from the vector.
4886 // In case of LOAD, the cost of InsertElement into the returned
4888 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4889 Instruction::InsertElement,
4893 // The cost of the scalar loads/stores.
4894 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4895 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4900 // Wide load/stores.
4901 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4902 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4905 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4909 case Instruction::ZExt:
4910 case Instruction::SExt:
4911 case Instruction::FPToUI:
4912 case Instruction::FPToSI:
4913 case Instruction::FPExt:
4914 case Instruction::PtrToInt:
4915 case Instruction::IntToPtr:
4916 case Instruction::SIToFP:
4917 case Instruction::UIToFP:
4918 case Instruction::Trunc:
4919 case Instruction::FPTrunc:
4920 case Instruction::BitCast: {
4921 // We optimize the truncation of induction variable.
4922 // The cost of these is the same as the scalar operation.
4923 if (I->getOpcode() == Instruction::Trunc &&
4924 Legal->isInductionVariable(I->getOperand(0)))
4925 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4926 I->getOperand(0)->getType());
4928 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4929 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4931 case Instruction::Call: {
4932 CallInst *CI = cast<CallInst>(I);
4933 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4934 assert(ID && "Not an intrinsic call!");
4935 Type *RetTy = ToVectorTy(CI->getType(), VF);
4936 SmallVector<Type*, 4> Tys;
4937 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4938 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4939 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4942 // We are scalarizing the instruction. Return the cost of the scalar
4943 // instruction, plus the cost of insert and extract into vector
4944 // elements, times the vector width.
4947 if (!RetTy->isVoidTy() && VF != 1) {
4948 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4950 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4953 // The cost of inserting the results plus extracting each one of the
4955 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4958 // The cost of executing VF copies of the scalar instruction. This opcode
4959 // is unknown. Assume that it is the same as 'mul'.
4960 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4966 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4967 if (Scalar->isVoidTy() || VF == 1)
4969 return VectorType::get(Scalar, VF);
4972 char LoopVectorize::ID = 0;
4973 static const char lv_name[] = "Loop Vectorization";
4974 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4975 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4976 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4977 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4978 INITIALIZE_PASS_DEPENDENCY(LCSSA)
4979 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4980 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4981 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4984 Pass *createLoopVectorizePass(bool NoUnrolling) {
4985 return new LoopVectorize(NoUnrolling);
4989 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4990 // Check for a store.
4991 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4992 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4994 // Check for a load.
4995 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4996 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5002 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
5003 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5004 // Holds vector parameters or scalars, in case of uniform vals.
5005 SmallVector<VectorParts, 4> Params;
5007 setDebugLocFromInst(Builder, Instr);
5009 // Find all of the vectorized parameters.
5010 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5011 Value *SrcOp = Instr->getOperand(op);
5013 // If we are accessing the old induction variable, use the new one.
5014 if (SrcOp == OldInduction) {
5015 Params.push_back(getVectorValue(SrcOp));
5019 // Try using previously calculated values.
5020 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5022 // If the src is an instruction that appeared earlier in the basic block
5023 // then it should already be vectorized.
5024 if (SrcInst && OrigLoop->contains(SrcInst)) {
5025 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5026 // The parameter is a vector value from earlier.
5027 Params.push_back(WidenMap.get(SrcInst));
5029 // The parameter is a scalar from outside the loop. Maybe even a constant.
5030 VectorParts Scalars;
5031 Scalars.append(UF, SrcOp);
5032 Params.push_back(Scalars);
5036 assert(Params.size() == Instr->getNumOperands() &&
5037 "Invalid number of operands");
5039 // Does this instruction return a value ?
5040 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5042 Value *UndefVec = IsVoidRetTy ? 0 :
5043 UndefValue::get(Instr->getType());
5044 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5045 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5047 // For each vector unroll 'part':
5048 for (unsigned Part = 0; Part < UF; ++Part) {
5049 // For each scalar that we create:
5051 Instruction *Cloned = Instr->clone();
5053 Cloned->setName(Instr->getName() + ".cloned");
5054 // Replace the operands of the cloned instructions with extracted scalars.
5055 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5056 Value *Op = Params[op][Part];
5057 Cloned->setOperand(op, Op);
5060 // Place the cloned scalar in the new loop.
5061 Builder.Insert(Cloned);
5063 // If the original scalar returns a value we need to place it in a vector
5064 // so that future users will be able to use it.
5066 VecResults[Part] = Cloned;
5071 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
5072 LoopVectorizationLegality*) {
5073 return scalarizeInstruction(Instr);
5076 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5080 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5084 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5086 // When unrolling and the VF is 1, we only need to add a simple scalar.
5087 Type *ITy = Val->getType();
5088 assert(!ITy->isVectorTy() && "Val must be a scalar");
5089 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5090 return Builder.CreateAdd(Val, C, "induction");