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 POD 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 POD 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.
498 /// Insert a pointer and calculate the start and end SCEVs.
499 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
502 /// This flag indicates if we need to add the runtime check.
504 /// Holds the pointers that we need to check.
505 SmallVector<TrackingVH<Value>, 2> Pointers;
506 /// Holds the pointer value at the beginning of the loop.
507 SmallVector<const SCEV*, 2> Starts;
508 /// Holds the pointer value at the end of the loop.
509 SmallVector<const SCEV*, 2> Ends;
510 /// Holds the information if this pointer is used for writing to memory.
511 SmallVector<bool, 2> IsWritePtr;
512 /// Holds the id of the set of pointers that could be dependent because of a
513 /// shared underlying object.
514 SmallVector<unsigned, 2> DependencySetId;
517 /// A POD for saving information about induction variables.
518 struct InductionInfo {
519 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
520 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
522 TrackingVH<Value> StartValue;
527 /// ReductionList contains the reduction descriptors for all
528 /// of the reductions that were found in the loop.
529 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
531 /// InductionList saves induction variables and maps them to the
532 /// induction descriptor.
533 typedef MapVector<PHINode*, InductionInfo> InductionList;
535 /// Returns true if it is legal to vectorize this loop.
536 /// This does not mean that it is profitable to vectorize this
537 /// loop, only that it is legal to do so.
540 /// Returns the Induction variable.
541 PHINode *getInduction() { return Induction; }
543 /// Returns the reduction variables found in the loop.
544 ReductionList *getReductionVars() { return &Reductions; }
546 /// Returns the induction variables found in the loop.
547 InductionList *getInductionVars() { return &Inductions; }
549 /// Returns the widest induction type.
550 Type *getWidestInductionType() { return WidestIndTy; }
552 /// Returns True if V is an induction variable in this loop.
553 bool isInductionVariable(const Value *V);
555 /// Return true if the block BB needs to be predicated in order for the loop
556 /// to be vectorized.
557 bool blockNeedsPredication(BasicBlock *BB);
559 /// Check if this pointer is consecutive when vectorizing. This happens
560 /// when the last index of the GEP is the induction variable, or that the
561 /// pointer itself is an induction variable.
562 /// This check allows us to vectorize A[idx] into a wide load/store.
564 /// 0 - Stride is unknown or non consecutive.
565 /// 1 - Address is consecutive.
566 /// -1 - Address is consecutive, and decreasing.
567 int isConsecutivePtr(Value *Ptr);
569 /// Returns true if the value V is uniform within the loop.
570 bool isUniform(Value *V);
572 /// Returns true if this instruction will remain scalar after vectorization.
573 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
575 /// Returns the information that we collected about runtime memory check.
576 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
578 /// This function returns the identity element (or neutral element) for
580 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
582 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
585 /// Check if a single basic block loop is vectorizable.
586 /// At this point we know that this is a loop with a constant trip count
587 /// and we only need to check individual instructions.
588 bool canVectorizeInstrs();
590 /// When we vectorize loops we may change the order in which
591 /// we read and write from memory. This method checks if it is
592 /// legal to vectorize the code, considering only memory constrains.
593 /// Returns true if the loop is vectorizable
594 bool canVectorizeMemory();
596 /// Return true if we can vectorize this loop using the IF-conversion
598 bool canVectorizeWithIfConvert();
600 /// Collect the variables that need to stay uniform after vectorization.
601 void collectLoopUniforms();
603 /// Return true if all of the instructions in the block can be speculatively
604 /// executed. \p SafePtrs is a list of addresses that are known to be legal
605 /// and we know that we can read from them without segfault.
606 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
608 /// Returns True, if 'Phi' is the kind of reduction variable for type
609 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
610 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
611 /// Returns a struct describing if the instruction 'I' can be a reduction
612 /// variable of type 'Kind'. If the reduction is a min/max pattern of
613 /// select(icmp()) this function advances the instruction pointer 'I' from the
614 /// compare instruction to the select instruction and stores this pointer in
615 /// 'PatternLastInst' member of the returned struct.
616 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
617 ReductionInstDesc &Desc);
618 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
619 /// pattern corresponding to a min(X, Y) or max(X, Y).
620 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
621 ReductionInstDesc &Prev);
622 /// Returns the induction kind of Phi. This function may return NoInduction
623 /// if the PHI is not an induction variable.
624 InductionKind isInductionVariable(PHINode *Phi);
626 /// The loop that we evaluate.
630 /// DataLayout analysis.
634 /// Target Library Info.
635 TargetLibraryInfo *TLI;
637 // --- vectorization state --- //
639 /// Holds the integer induction variable. This is the counter of the
642 /// Holds the reduction variables.
643 ReductionList Reductions;
644 /// Holds all of the induction variables that we found in the loop.
645 /// Notice that inductions don't need to start at zero and that induction
646 /// variables can be pointers.
647 InductionList Inductions;
648 /// Holds the widest induction type encountered.
651 /// Allowed outside users. This holds the reduction
652 /// vars which can be accessed from outside the loop.
653 SmallPtrSet<Value*, 4> AllowedExit;
654 /// This set holds the variables which are known to be uniform after
656 SmallPtrSet<Instruction*, 4> Uniforms;
657 /// We need to check that all of the pointers in this list are disjoint
659 RuntimePointerCheck PtrRtCheck;
660 /// Can we assume the absence of NaNs.
661 bool HasFunNoNaNAttr;
663 unsigned MaxSafeDepDistBytes;
666 /// LoopVectorizationCostModel - estimates the expected speedups due to
668 /// In many cases vectorization is not profitable. This can happen because of
669 /// a number of reasons. In this class we mainly attempt to predict the
670 /// expected speedup/slowdowns due to the supported instruction set. We use the
671 /// TargetTransformInfo to query the different backends for the cost of
672 /// different operations.
673 class LoopVectorizationCostModel {
675 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
676 LoopVectorizationLegality *Legal,
677 const TargetTransformInfo &TTI,
678 DataLayout *DL, const TargetLibraryInfo *TLI)
679 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
681 /// Information about vectorization costs
682 struct VectorizationFactor {
683 unsigned Width; // Vector width with best cost
684 unsigned Cost; // Cost of the loop with that width
686 /// \return The most profitable vectorization factor and the cost of that VF.
687 /// This method checks every power of two up to VF. If UserVF is not ZERO
688 /// then this vectorization factor will be selected if vectorization is
690 VectorizationFactor selectVectorizationFactor(bool OptForSize,
693 /// \return The size (in bits) of the widest type in the code that
694 /// needs to be vectorized. We ignore values that remain scalar such as
695 /// 64 bit loop indices.
696 unsigned getWidestType();
698 /// \return The most profitable unroll factor.
699 /// If UserUF is non-zero then this method finds the best unroll-factor
700 /// based on register pressure and other parameters.
701 /// VF and LoopCost are the selected vectorization factor and the cost of the
703 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
706 /// \brief A struct that represents some properties of the register usage
708 struct RegisterUsage {
709 /// Holds the number of loop invariant values that are used in the loop.
710 unsigned LoopInvariantRegs;
711 /// Holds the maximum number of concurrent live intervals in the loop.
712 unsigned MaxLocalUsers;
713 /// Holds the number of instructions in the loop.
714 unsigned NumInstructions;
717 /// \return information about the register usage of the loop.
718 RegisterUsage calculateRegisterUsage();
721 /// Returns the expected execution cost. The unit of the cost does
722 /// not matter because we use the 'cost' units to compare different
723 /// vector widths. The cost that is returned is *not* normalized by
724 /// the factor width.
725 unsigned expectedCost(unsigned VF);
727 /// Returns the execution time cost of an instruction for a given vector
728 /// width. Vector width of one means scalar.
729 unsigned getInstructionCost(Instruction *I, unsigned VF);
731 /// A helper function for converting Scalar types to vector types.
732 /// If the incoming type is void, we return void. If the VF is 1, we return
734 static Type* ToVectorTy(Type *Scalar, unsigned VF);
736 /// Returns whether the instruction is a load or store and will be a emitted
737 /// as a vector operation.
738 bool isConsecutiveLoadOrStore(Instruction *I);
740 /// The loop that we evaluate.
744 /// Loop Info analysis.
746 /// Vectorization legality.
747 LoopVectorizationLegality *Legal;
748 /// Vector target information.
749 const TargetTransformInfo &TTI;
750 /// Target data layout information.
752 /// Target Library Info.
753 const TargetLibraryInfo *TLI;
756 /// Utility class for getting and setting loop vectorizer hints in the form
757 /// of loop metadata.
758 struct LoopVectorizeHints {
759 /// Vectorization width.
761 /// Vectorization unroll factor.
764 LoopVectorizeHints(const Loop *L)
765 : Width(VectorizationFactor)
766 , Unroll(VectorizationUnroll)
767 , LoopID(L->getLoopID()) {
769 // The command line options override any loop metadata except for when
770 // width == 1 which is used to indicate the loop is already vectorized.
771 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
772 Width = VectorizationFactor;
773 if (VectorizationUnroll.getNumOccurrences() > 0)
774 Unroll = VectorizationUnroll;
777 /// Return the loop vectorizer metadata prefix.
778 static StringRef Prefix() { return "llvm.vectorizer."; }
780 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
781 SmallVector<Value*, 2> Vals;
782 Vals.push_back(MDString::get(Context, Name));
783 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
784 return MDNode::get(Context, Vals);
787 /// Mark the loop L as already vectorized by setting the width to 1.
788 void setAlreadyVectorized(Loop *L) {
789 LLVMContext &Context = L->getHeader()->getContext();
793 // Create a new loop id with one more operand for the already_vectorized
794 // hint. If the loop already has a loop id then copy the existing operands.
795 SmallVector<Value*, 4> Vals(1);
797 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
798 Vals.push_back(LoopID->getOperand(i));
800 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
802 MDNode *NewLoopID = MDNode::get(Context, Vals);
803 // Set operand 0 to refer to the loop id itself.
804 NewLoopID->replaceOperandWith(0, NewLoopID);
806 L->setLoopID(NewLoopID);
808 LoopID->replaceAllUsesWith(NewLoopID);
816 /// Find hints specified in the loop metadata.
817 void getHints(const Loop *L) {
821 // First operand should refer to the loop id itself.
822 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
823 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
825 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
826 const MDString *S = 0;
827 SmallVector<Value*, 4> Args;
829 // The expected hint is either a MDString or a MDNode with the first
830 // operand a MDString.
831 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
832 if (!MD || MD->getNumOperands() == 0)
834 S = dyn_cast<MDString>(MD->getOperand(0));
835 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
836 Args.push_back(MD->getOperand(i));
838 S = dyn_cast<MDString>(LoopID->getOperand(i));
839 assert(Args.size() == 0 && "too many arguments for MDString");
845 // Check if the hint starts with the vectorizer prefix.
846 StringRef Hint = S->getString();
847 if (!Hint.startswith(Prefix()))
849 // Remove the prefix.
850 Hint = Hint.substr(Prefix().size(), StringRef::npos);
852 if (Args.size() == 1)
853 getHint(Hint, Args[0]);
857 // Check string hint with one operand.
858 void getHint(StringRef Hint, Value *Arg) {
859 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
861 unsigned Val = C->getZExtValue();
863 if (Hint == "width") {
864 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
865 "Invalid width metadata");
867 } else if (Hint == "unroll") {
868 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
869 "Invalid unroll metadata");
872 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
876 /// The LoopVectorize Pass.
877 struct LoopVectorize : public LoopPass {
878 /// Pass identification, replacement for typeid
881 explicit LoopVectorize() : LoopPass(ID) {
882 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
888 TargetTransformInfo *TTI;
890 TargetLibraryInfo *TLI;
892 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
893 // We only vectorize innermost loops.
897 SE = &getAnalysis<ScalarEvolution>();
898 DL = getAnalysisIfAvailable<DataLayout>();
899 LI = &getAnalysis<LoopInfo>();
900 TTI = &getAnalysis<TargetTransformInfo>();
901 DT = &getAnalysis<DominatorTree>();
902 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
905 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
909 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
910 L->getHeader()->getParent()->getName() << "\"\n");
912 LoopVectorizeHints Hints(L);
914 if (Hints.Width == 1 && Hints.Unroll == 1) {
915 DEBUG(dbgs() << "LV: Not vectorizing.\n");
919 // Check if it is legal to vectorize the loop.
920 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
921 if (!LVL.canVectorize()) {
922 DEBUG(dbgs() << "LV: Not vectorizing.\n");
926 // Use the cost model.
927 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
929 // Check the function attributes to find out if this function should be
930 // optimized for size.
931 Function *F = L->getHeader()->getParent();
932 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
933 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
934 unsigned FnIndex = AttributeSet::FunctionIndex;
935 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
936 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
939 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
940 "attribute is used.\n");
944 // Select the optimal vectorization factor.
945 LoopVectorizationCostModel::VectorizationFactor VF;
946 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
947 // Select the unroll factor.
948 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
952 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
955 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
956 F->getParent()->getModuleIdentifier()<<"\n");
957 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
962 // We decided not to vectorize, but we may want to unroll.
963 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
964 Unroller.vectorize(&LVL);
966 // If we decided that it is *legal* to vectorize the loop then do it.
967 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
971 // Mark the loop as already vectorized to avoid vectorizing again.
972 Hints.setAlreadyVectorized(L);
974 DEBUG(verifyFunction(*L->getHeader()->getParent()));
978 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
979 LoopPass::getAnalysisUsage(AU);
980 AU.addRequiredID(LoopSimplifyID);
981 AU.addRequiredID(LCSSAID);
982 AU.addRequired<DominatorTree>();
983 AU.addRequired<LoopInfo>();
984 AU.addRequired<ScalarEvolution>();
985 AU.addRequired<TargetTransformInfo>();
986 AU.addPreserved<LoopInfo>();
987 AU.addPreserved<DominatorTree>();
992 } // end anonymous namespace
994 //===----------------------------------------------------------------------===//
995 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
996 // LoopVectorizationCostModel.
997 //===----------------------------------------------------------------------===//
1000 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1001 Loop *Lp, Value *Ptr,
1003 unsigned DepSetId) {
1004 const SCEV *Sc = SE->getSCEV(Ptr);
1005 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1006 assert(AR && "Invalid addrec expression");
1007 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1008 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1009 Pointers.push_back(Ptr);
1010 Starts.push_back(AR->getStart());
1011 Ends.push_back(ScEnd);
1012 IsWritePtr.push_back(WritePtr);
1013 DependencySetId.push_back(DepSetId);
1016 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1017 // Save the current insertion location.
1018 Instruction *Loc = Builder.GetInsertPoint();
1020 // We need to place the broadcast of invariant variables outside the loop.
1021 Instruction *Instr = dyn_cast<Instruction>(V);
1022 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1023 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1025 // Place the code for broadcasting invariant variables in the new preheader.
1027 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1029 // Broadcast the scalar into all locations in the vector.
1030 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1032 // Restore the builder insertion point.
1034 Builder.SetInsertPoint(Loc);
1039 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1041 assert(Val->getType()->isVectorTy() && "Must be a vector");
1042 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1043 "Elem must be an integer");
1044 // Create the types.
1045 Type *ITy = Val->getType()->getScalarType();
1046 VectorType *Ty = cast<VectorType>(Val->getType());
1047 int VLen = Ty->getNumElements();
1048 SmallVector<Constant*, 8> Indices;
1050 // Create a vector of consecutive numbers from zero to VF.
1051 for (int i = 0; i < VLen; ++i) {
1052 int64_t Idx = Negate ? (-i) : i;
1053 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1056 // Add the consecutive indices to the vector value.
1057 Constant *Cv = ConstantVector::get(Indices);
1058 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1059 return Builder.CreateAdd(Val, Cv, "induction");
1062 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1063 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1064 // Make sure that the pointer does not point to structs.
1065 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1068 // If this value is a pointer induction variable we know it is consecutive.
1069 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1070 if (Phi && Inductions.count(Phi)) {
1071 InductionInfo II = Inductions[Phi];
1072 if (IK_PtrInduction == II.IK)
1074 else if (IK_ReversePtrInduction == II.IK)
1078 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1082 unsigned NumOperands = Gep->getNumOperands();
1083 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1085 Value *GpPtr = Gep->getPointerOperand();
1086 // If this GEP value is a consecutive pointer induction variable and all of
1087 // the indices are constant then we know it is consecutive. We can
1088 Phi = dyn_cast<PHINode>(GpPtr);
1089 if (Phi && Inductions.count(Phi)) {
1091 // Make sure that the pointer does not point to structs.
1092 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1093 if (GepPtrType->getElementType()->isAggregateType())
1096 // Make sure that all of the index operands are loop invariant.
1097 for (unsigned i = 1; i < NumOperands; ++i)
1098 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1101 InductionInfo II = Inductions[Phi];
1102 if (IK_PtrInduction == II.IK)
1104 else if (IK_ReversePtrInduction == II.IK)
1108 // Check that all of the gep indices are uniform except for the last.
1109 for (unsigned i = 0; i < NumOperands - 1; ++i)
1110 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1113 // We can emit wide load/stores only if the last index is the induction
1115 const SCEV *Last = SE->getSCEV(LastIndex);
1116 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1117 const SCEV *Step = AR->getStepRecurrence(*SE);
1119 // The memory is consecutive because the last index is consecutive
1120 // and all other indices are loop invariant.
1123 if (Step->isAllOnesValue())
1130 bool LoopVectorizationLegality::isUniform(Value *V) {
1131 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1134 InnerLoopVectorizer::VectorParts&
1135 InnerLoopVectorizer::getVectorValue(Value *V) {
1136 assert(V != Induction && "The new induction variable should not be used.");
1137 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1139 // If we have this scalar in the map, return it.
1140 if (WidenMap.has(V))
1141 return WidenMap.get(V);
1143 // If this scalar is unknown, assume that it is a constant or that it is
1144 // loop invariant. Broadcast V and save the value for future uses.
1145 Value *B = getBroadcastInstrs(V);
1146 return WidenMap.splat(V, B);
1149 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1150 assert(Vec->getType()->isVectorTy() && "Invalid type");
1151 SmallVector<Constant*, 8> ShuffleMask;
1152 for (unsigned i = 0; i < VF; ++i)
1153 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1155 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1156 ConstantVector::get(ShuffleMask),
1161 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1162 LoopVectorizationLegality *Legal) {
1163 // Attempt to issue a wide load.
1164 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1165 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1167 assert((LI || SI) && "Invalid Load/Store instruction");
1169 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1170 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1171 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1172 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1173 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1174 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1175 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1177 if (ScalarAllocatedSize != VectorElementSize)
1178 return scalarizeInstruction(Instr);
1180 // If the pointer is loop invariant or if it is non consecutive,
1181 // scalarize the load.
1182 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1183 bool Reverse = ConsecutiveStride < 0;
1184 bool UniformLoad = LI && Legal->isUniform(Ptr);
1185 if (!ConsecutiveStride || UniformLoad)
1186 return scalarizeInstruction(Instr);
1188 Constant *Zero = Builder.getInt32(0);
1189 VectorParts &Entry = WidenMap.get(Instr);
1191 // Handle consecutive loads/stores.
1192 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1193 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1194 setDebugLocFromInst(Builder, Gep);
1195 Value *PtrOperand = Gep->getPointerOperand();
1196 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1197 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1199 // Create the new GEP with the new induction variable.
1200 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1201 Gep2->setOperand(0, FirstBasePtr);
1202 Gep2->setName("gep.indvar.base");
1203 Ptr = Builder.Insert(Gep2);
1205 setDebugLocFromInst(Builder, Gep);
1206 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1207 OrigLoop) && "Base ptr must be invariant");
1209 // The last index does not have to be the induction. It can be
1210 // consecutive and be a function of the index. For example A[I+1];
1211 unsigned NumOperands = Gep->getNumOperands();
1212 unsigned LastOperand = NumOperands - 1;
1213 // Create the new GEP with the new induction variable.
1214 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1216 for (unsigned i = 0; i < NumOperands; ++i) {
1217 Value *GepOperand = Gep->getOperand(i);
1218 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1220 // Update last index or loop invariant instruction anchored in loop.
1221 if (i == LastOperand ||
1222 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1223 assert((i == LastOperand ||
1224 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1225 "Must be last index or loop invariant");
1227 VectorParts &GEPParts = getVectorValue(GepOperand);
1228 Value *Index = GEPParts[0];
1229 Index = Builder.CreateExtractElement(Index, Zero);
1230 Gep2->setOperand(i, Index);
1231 Gep2->setName("gep.indvar.idx");
1234 Ptr = Builder.Insert(Gep2);
1236 // Use the induction element ptr.
1237 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1238 setDebugLocFromInst(Builder, Ptr);
1239 VectorParts &PtrVal = getVectorValue(Ptr);
1240 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1245 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1246 "We do not allow storing to uniform addresses");
1247 setDebugLocFromInst(Builder, SI);
1248 // We don't want to update the value in the map as it might be used in
1249 // another expression. So don't use a reference type for "StoredVal".
1250 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1252 for (unsigned Part = 0; Part < UF; ++Part) {
1253 // Calculate the pointer for the specific unroll-part.
1254 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1257 // If we store to reverse consecutive memory locations then we need
1258 // to reverse the order of elements in the stored value.
1259 StoredVal[Part] = reverseVector(StoredVal[Part]);
1260 // If the address is consecutive but reversed, then the
1261 // wide store needs to start at the last vector element.
1262 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1263 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1266 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1267 DataTy->getPointerTo(AddressSpace));
1268 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1274 assert(LI && "Must have a load instruction");
1275 setDebugLocFromInst(Builder, LI);
1276 for (unsigned Part = 0; Part < UF; ++Part) {
1277 // Calculate the pointer for the specific unroll-part.
1278 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1281 // If the address is consecutive but reversed, then the
1282 // wide store needs to start at the last vector element.
1283 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1284 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1287 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1288 DataTy->getPointerTo(AddressSpace));
1289 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1290 cast<LoadInst>(LI)->setAlignment(Alignment);
1291 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1295 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1296 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1297 // Holds vector parameters or scalars, in case of uniform vals.
1298 SmallVector<VectorParts, 4> Params;
1300 setDebugLocFromInst(Builder, Instr);
1302 // Find all of the vectorized parameters.
1303 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1304 Value *SrcOp = Instr->getOperand(op);
1306 // If we are accessing the old induction variable, use the new one.
1307 if (SrcOp == OldInduction) {
1308 Params.push_back(getVectorValue(SrcOp));
1312 // Try using previously calculated values.
1313 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1315 // If the src is an instruction that appeared earlier in the basic block
1316 // then it should already be vectorized.
1317 if (SrcInst && OrigLoop->contains(SrcInst)) {
1318 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1319 // The parameter is a vector value from earlier.
1320 Params.push_back(WidenMap.get(SrcInst));
1322 // The parameter is a scalar from outside the loop. Maybe even a constant.
1323 VectorParts Scalars;
1324 Scalars.append(UF, SrcOp);
1325 Params.push_back(Scalars);
1329 assert(Params.size() == Instr->getNumOperands() &&
1330 "Invalid number of operands");
1332 // Does this instruction return a value ?
1333 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1335 Value *UndefVec = IsVoidRetTy ? 0 :
1336 UndefValue::get(VectorType::get(Instr->getType(), VF));
1337 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1338 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1340 // For each vector unroll 'part':
1341 for (unsigned Part = 0; Part < UF; ++Part) {
1342 // For each scalar that we create:
1343 for (unsigned Width = 0; Width < VF; ++Width) {
1344 Instruction *Cloned = Instr->clone();
1346 Cloned->setName(Instr->getName() + ".cloned");
1347 // Replace the operands of the cloned instrucions with extracted scalars.
1348 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1349 Value *Op = Params[op][Part];
1350 // Param is a vector. Need to extract the right lane.
1351 if (Op->getType()->isVectorTy())
1352 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1353 Cloned->setOperand(op, Op);
1356 // Place the cloned scalar in the new loop.
1357 Builder.Insert(Cloned);
1359 // If the original scalar returns a value we need to place it in a vector
1360 // so that future users will be able to use it.
1362 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1363 Builder.getInt32(Width));
1369 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1371 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1372 Legal->getRuntimePointerCheck();
1374 if (!PtrRtCheck->Need)
1377 unsigned NumPointers = PtrRtCheck->Pointers.size();
1378 SmallVector<TrackingVH<Value> , 2> Starts;
1379 SmallVector<TrackingVH<Value> , 2> Ends;
1381 SCEVExpander Exp(*SE, "induction");
1383 // Use this type for pointer arithmetic.
1384 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1386 for (unsigned i = 0; i < NumPointers; ++i) {
1387 Value *Ptr = PtrRtCheck->Pointers[i];
1388 const SCEV *Sc = SE->getSCEV(Ptr);
1390 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1391 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1393 Starts.push_back(Ptr);
1394 Ends.push_back(Ptr);
1396 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1398 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1399 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1400 Starts.push_back(Start);
1401 Ends.push_back(End);
1405 IRBuilder<> ChkBuilder(Loc);
1406 // Our instructions might fold to a constant.
1407 Value *MemoryRuntimeCheck = 0;
1408 for (unsigned i = 0; i < NumPointers; ++i) {
1409 for (unsigned j = i+1; j < NumPointers; ++j) {
1410 // No need to check if two readonly pointers intersect.
1411 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1414 // Only need to check pointers between two different dependency sets.
1415 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1418 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1419 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1420 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1421 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1423 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1424 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1425 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1426 if (MemoryRuntimeCheck)
1427 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1429 MemoryRuntimeCheck = IsConflict;
1433 // We have to do this trickery because the IRBuilder might fold the check to a
1434 // constant expression in which case there is no Instruction anchored in a
1436 LLVMContext &Ctx = Loc->getContext();
1437 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1438 ConstantInt::getTrue(Ctx));
1439 ChkBuilder.Insert(Check, "memcheck.conflict");
1444 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1446 In this function we generate a new loop. The new loop will contain
1447 the vectorized instructions while the old loop will continue to run the
1450 [ ] <-- vector loop bypass (may consist of multiple blocks).
1453 | [ ] <-- vector pre header.
1457 | [ ]_| <-- vector loop.
1460 >[ ] <--- middle-block.
1463 | [ ] <--- new preheader.
1467 | [ ]_| <-- old scalar loop to handle remainder.
1470 >[ ] <-- exit block.
1474 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1475 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1476 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1477 assert(ExitBlock && "Must have an exit block");
1479 // Some loops have a single integer induction variable, while other loops
1480 // don't. One example is c++ iterators that often have multiple pointer
1481 // induction variables. In the code below we also support a case where we
1482 // don't have a single induction variable.
1483 OldInduction = Legal->getInduction();
1484 Type *IdxTy = Legal->getWidestInductionType();
1486 // Find the loop boundaries.
1487 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1488 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1490 // Get the total trip count from the count by adding 1.
1491 ExitCount = SE->getAddExpr(ExitCount,
1492 SE->getConstant(ExitCount->getType(), 1));
1494 // Expand the trip count and place the new instructions in the preheader.
1495 // Notice that the pre-header does not change, only the loop body.
1496 SCEVExpander Exp(*SE, "induction");
1498 // Count holds the overall loop count (N).
1499 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1500 BypassBlock->getTerminator());
1502 // The loop index does not have to start at Zero. Find the original start
1503 // value from the induction PHI node. If we don't have an induction variable
1504 // then we know that it starts at zero.
1505 Builder.SetInsertPoint(BypassBlock->getTerminator());
1506 Value *StartIdx = ExtendedIdx = OldInduction ?
1507 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1509 ConstantInt::get(IdxTy, 0);
1511 assert(BypassBlock && "Invalid loop structure");
1512 LoopBypassBlocks.push_back(BypassBlock);
1514 // Split the single block loop into the two loop structure described above.
1515 BasicBlock *VectorPH =
1516 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1517 BasicBlock *VecBody =
1518 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1519 BasicBlock *MiddleBlock =
1520 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1521 BasicBlock *ScalarPH =
1522 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1524 // Create and register the new vector loop.
1525 Loop* Lp = new Loop();
1526 Loop *ParentLoop = OrigLoop->getParentLoop();
1528 // Insert the new loop into the loop nest and register the new basic blocks
1529 // before calling any utilities such as SCEV that require valid LoopInfo.
1531 ParentLoop->addChildLoop(Lp);
1532 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1533 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1534 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1536 LI->addTopLevelLoop(Lp);
1538 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1540 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1542 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1544 // Generate the induction variable.
1545 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1546 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1547 // The loop step is equal to the vectorization factor (num of SIMD elements)
1548 // times the unroll factor (num of SIMD instructions).
1549 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1551 // This is the IR builder that we use to add all of the logic for bypassing
1552 // the new vector loop.
1553 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1554 setDebugLocFromInst(BypassBuilder,
1555 getDebugLocFromInstOrOperands(OldInduction));
1557 // We may need to extend the index in case there is a type mismatch.
1558 // We know that the count starts at zero and does not overflow.
1559 if (Count->getType() != IdxTy) {
1560 // The exit count can be of pointer type. Convert it to the correct
1562 if (ExitCount->getType()->isPointerTy())
1563 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1565 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1568 // Add the start index to the loop count to get the new end index.
1569 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1571 // Now we need to generate the expression for N - (N % VF), which is
1572 // the part that the vectorized body will execute.
1573 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1574 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1575 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1576 "end.idx.rnd.down");
1578 // Now, compare the new count to zero. If it is zero skip the vector loop and
1579 // jump to the scalar loop.
1580 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1583 BasicBlock *LastBypassBlock = BypassBlock;
1585 // Generate the code that checks in runtime if arrays overlap. We put the
1586 // checks into a separate block to make the more common case of few elements
1588 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1589 BypassBlock->getTerminator());
1590 if (MemRuntimeCheck) {
1591 // Create a new block containing the memory check.
1592 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1595 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1596 LoopBypassBlocks.push_back(CheckBlock);
1598 // Replace the branch into the memory check block with a conditional branch
1599 // for the "few elements case".
1600 Instruction *OldTerm = BypassBlock->getTerminator();
1601 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1602 OldTerm->eraseFromParent();
1604 Cmp = MemRuntimeCheck;
1605 LastBypassBlock = CheckBlock;
1608 LastBypassBlock->getTerminator()->eraseFromParent();
1609 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1612 // We are going to resume the execution of the scalar loop.
1613 // Go over all of the induction variables that we found and fix the
1614 // PHIs that are left in the scalar version of the loop.
1615 // The starting values of PHI nodes depend on the counter of the last
1616 // iteration in the vectorized loop.
1617 // If we come from a bypass edge then we need to start from the original
1620 // This variable saves the new starting index for the scalar loop.
1621 PHINode *ResumeIndex = 0;
1622 LoopVectorizationLegality::InductionList::iterator I, E;
1623 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1624 // Set builder to point to last bypass block.
1625 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1626 for (I = List->begin(), E = List->end(); I != E; ++I) {
1627 PHINode *OrigPhi = I->first;
1628 LoopVectorizationLegality::InductionInfo II = I->second;
1630 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1631 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1632 MiddleBlock->getTerminator());
1633 // We might have extended the type of the induction variable but we need a
1634 // truncated version for the scalar loop.
1635 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1636 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1637 MiddleBlock->getTerminator()) : 0;
1639 Value *EndValue = 0;
1641 case LoopVectorizationLegality::IK_NoInduction:
1642 llvm_unreachable("Unknown induction");
1643 case LoopVectorizationLegality::IK_IntInduction: {
1644 // Handle the integer induction counter.
1645 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1647 // We have the canonical induction variable.
1648 if (OrigPhi == OldInduction) {
1649 // Create a truncated version of the resume value for the scalar loop,
1650 // we might have promoted the type to a larger width.
1652 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1653 // The new PHI merges the original incoming value, in case of a bypass,
1654 // or the value at the end of the vectorized loop.
1655 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1656 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1657 TruncResumeVal->addIncoming(EndValue, VecBody);
1659 // We know what the end value is.
1660 EndValue = IdxEndRoundDown;
1661 // We also know which PHI node holds it.
1662 ResumeIndex = ResumeVal;
1666 // Not the canonical induction variable - add the vector loop count to the
1668 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1669 II.StartValue->getType(),
1671 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1674 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1675 // Convert the CountRoundDown variable to the PHI size.
1676 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1677 II.StartValue->getType(),
1679 // Handle reverse integer induction counter.
1680 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1683 case LoopVectorizationLegality::IK_PtrInduction: {
1684 // For pointer induction variables, calculate the offset using
1686 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1690 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1691 // The value at the end of the loop for the reverse pointer is calculated
1692 // by creating a GEP with a negative index starting from the start value.
1693 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1694 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1696 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1702 // The new PHI merges the original incoming value, in case of a bypass,
1703 // or the value at the end of the vectorized loop.
1704 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1705 if (OrigPhi == OldInduction)
1706 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1708 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1710 ResumeVal->addIncoming(EndValue, VecBody);
1712 // Fix the scalar body counter (PHI node).
1713 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1714 // The old inductions phi node in the scalar body needs the truncated value.
1715 if (OrigPhi == OldInduction)
1716 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1718 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1721 // If we are generating a new induction variable then we also need to
1722 // generate the code that calculates the exit value. This value is not
1723 // simply the end of the counter because we may skip the vectorized body
1724 // in case of a runtime check.
1726 assert(!ResumeIndex && "Unexpected resume value found");
1727 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1728 MiddleBlock->getTerminator());
1729 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1730 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1731 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1734 // Make sure that we found the index where scalar loop needs to continue.
1735 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1736 "Invalid resume Index");
1738 // Add a check in the middle block to see if we have completed
1739 // all of the iterations in the first vector loop.
1740 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1741 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1742 ResumeIndex, "cmp.n",
1743 MiddleBlock->getTerminator());
1745 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1746 // Remove the old terminator.
1747 MiddleBlock->getTerminator()->eraseFromParent();
1749 // Create i+1 and fill the PHINode.
1750 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1751 Induction->addIncoming(StartIdx, VectorPH);
1752 Induction->addIncoming(NextIdx, VecBody);
1753 // Create the compare.
1754 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1755 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1757 // Now we have two terminators. Remove the old one from the block.
1758 VecBody->getTerminator()->eraseFromParent();
1760 // Get ready to start creating new instructions into the vectorized body.
1761 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1764 LoopVectorPreHeader = VectorPH;
1765 LoopScalarPreHeader = ScalarPH;
1766 LoopMiddleBlock = MiddleBlock;
1767 LoopExitBlock = ExitBlock;
1768 LoopVectorBody = VecBody;
1769 LoopScalarBody = OldBasicBlock;
1772 /// This function returns the identity element (or neutral element) for
1773 /// the operation K.
1775 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1780 // Adding, Xoring, Oring zero to a number does not change it.
1781 return ConstantInt::get(Tp, 0);
1782 case RK_IntegerMult:
1783 // Multiplying a number by 1 does not change it.
1784 return ConstantInt::get(Tp, 1);
1786 // AND-ing a number with an all-1 value does not change it.
1787 return ConstantInt::get(Tp, -1, true);
1789 // Multiplying a number by 1 does not change it.
1790 return ConstantFP::get(Tp, 1.0L);
1792 // Adding zero to a number does not change it.
1793 return ConstantFP::get(Tp, 0.0L);
1795 llvm_unreachable("Unknown reduction kind");
1799 static Intrinsic::ID
1800 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1801 // If we have an intrinsic call, check if it is trivially vectorizable.
1802 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1803 switch (II->getIntrinsicID()) {
1804 case Intrinsic::sqrt:
1805 case Intrinsic::sin:
1806 case Intrinsic::cos:
1807 case Intrinsic::exp:
1808 case Intrinsic::exp2:
1809 case Intrinsic::log:
1810 case Intrinsic::log10:
1811 case Intrinsic::log2:
1812 case Intrinsic::fabs:
1813 case Intrinsic::copysign:
1814 case Intrinsic::floor:
1815 case Intrinsic::ceil:
1816 case Intrinsic::trunc:
1817 case Intrinsic::rint:
1818 case Intrinsic::nearbyint:
1819 case Intrinsic::round:
1820 case Intrinsic::pow:
1821 case Intrinsic::fma:
1822 case Intrinsic::fmuladd:
1823 case Intrinsic::lifetime_start:
1824 case Intrinsic::lifetime_end:
1825 return II->getIntrinsicID();
1827 return Intrinsic::not_intrinsic;
1832 return Intrinsic::not_intrinsic;
1835 Function *F = CI->getCalledFunction();
1836 // We're going to make assumptions on the semantics of the functions, check
1837 // that the target knows that it's available in this environment.
1838 if (!F || !TLI->getLibFunc(F->getName(), Func))
1839 return Intrinsic::not_intrinsic;
1841 // Otherwise check if we have a call to a function that can be turned into a
1842 // vector intrinsic.
1849 return Intrinsic::sin;
1853 return Intrinsic::cos;
1857 return Intrinsic::exp;
1859 case LibFunc::exp2f:
1860 case LibFunc::exp2l:
1861 return Intrinsic::exp2;
1865 return Intrinsic::log;
1866 case LibFunc::log10:
1867 case LibFunc::log10f:
1868 case LibFunc::log10l:
1869 return Intrinsic::log10;
1871 case LibFunc::log2f:
1872 case LibFunc::log2l:
1873 return Intrinsic::log2;
1875 case LibFunc::fabsf:
1876 case LibFunc::fabsl:
1877 return Intrinsic::fabs;
1878 case LibFunc::copysign:
1879 case LibFunc::copysignf:
1880 case LibFunc::copysignl:
1881 return Intrinsic::copysign;
1882 case LibFunc::floor:
1883 case LibFunc::floorf:
1884 case LibFunc::floorl:
1885 return Intrinsic::floor;
1887 case LibFunc::ceilf:
1888 case LibFunc::ceill:
1889 return Intrinsic::ceil;
1890 case LibFunc::trunc:
1891 case LibFunc::truncf:
1892 case LibFunc::truncl:
1893 return Intrinsic::trunc;
1895 case LibFunc::rintf:
1896 case LibFunc::rintl:
1897 return Intrinsic::rint;
1898 case LibFunc::nearbyint:
1899 case LibFunc::nearbyintf:
1900 case LibFunc::nearbyintl:
1901 return Intrinsic::nearbyint;
1902 case LibFunc::round:
1903 case LibFunc::roundf:
1904 case LibFunc::roundl:
1905 return Intrinsic::round;
1909 return Intrinsic::pow;
1912 return Intrinsic::not_intrinsic;
1915 /// This function translates the reduction kind to an LLVM binary operator.
1917 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1919 case LoopVectorizationLegality::RK_IntegerAdd:
1920 return Instruction::Add;
1921 case LoopVectorizationLegality::RK_IntegerMult:
1922 return Instruction::Mul;
1923 case LoopVectorizationLegality::RK_IntegerOr:
1924 return Instruction::Or;
1925 case LoopVectorizationLegality::RK_IntegerAnd:
1926 return Instruction::And;
1927 case LoopVectorizationLegality::RK_IntegerXor:
1928 return Instruction::Xor;
1929 case LoopVectorizationLegality::RK_FloatMult:
1930 return Instruction::FMul;
1931 case LoopVectorizationLegality::RK_FloatAdd:
1932 return Instruction::FAdd;
1933 case LoopVectorizationLegality::RK_IntegerMinMax:
1934 return Instruction::ICmp;
1935 case LoopVectorizationLegality::RK_FloatMinMax:
1936 return Instruction::FCmp;
1938 llvm_unreachable("Unknown reduction operation");
1942 Value *createMinMaxOp(IRBuilder<> &Builder,
1943 LoopVectorizationLegality::MinMaxReductionKind RK,
1946 CmpInst::Predicate P = CmpInst::ICMP_NE;
1949 llvm_unreachable("Unknown min/max reduction kind");
1950 case LoopVectorizationLegality::MRK_UIntMin:
1951 P = CmpInst::ICMP_ULT;
1953 case LoopVectorizationLegality::MRK_UIntMax:
1954 P = CmpInst::ICMP_UGT;
1956 case LoopVectorizationLegality::MRK_SIntMin:
1957 P = CmpInst::ICMP_SLT;
1959 case LoopVectorizationLegality::MRK_SIntMax:
1960 P = CmpInst::ICMP_SGT;
1962 case LoopVectorizationLegality::MRK_FloatMin:
1963 P = CmpInst::FCMP_OLT;
1965 case LoopVectorizationLegality::MRK_FloatMax:
1966 P = CmpInst::FCMP_OGT;
1971 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
1972 RK == LoopVectorizationLegality::MRK_FloatMax)
1973 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1975 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1977 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1982 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1983 //===------------------------------------------------===//
1985 // Notice: any optimization or new instruction that go
1986 // into the code below should be also be implemented in
1989 //===------------------------------------------------===//
1990 Constant *Zero = Builder.getInt32(0);
1992 // In order to support reduction variables we need to be able to vectorize
1993 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1994 // stages. First, we create a new vector PHI node with no incoming edges.
1995 // We use this value when we vectorize all of the instructions that use the
1996 // PHI. Next, after all of the instructions in the block are complete we
1997 // add the new incoming edges to the PHI. At this point all of the
1998 // instructions in the basic block are vectorized, so we can use them to
1999 // construct the PHI.
2000 PhiVector RdxPHIsToFix;
2002 // Scan the loop in a topological order to ensure that defs are vectorized
2004 LoopBlocksDFS DFS(OrigLoop);
2007 // Vectorize all of the blocks in the original loop.
2008 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2009 be = DFS.endRPO(); bb != be; ++bb)
2010 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2012 // At this point every instruction in the original loop is widened to
2013 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2014 // that we vectorized. The PHI nodes are currently empty because we did
2015 // not want to introduce cycles. Notice that the remaining PHI nodes
2016 // that we need to fix are reduction variables.
2018 // Create the 'reduced' values for each of the induction vars.
2019 // The reduced values are the vector values that we scalarize and combine
2020 // after the loop is finished.
2021 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2023 PHINode *RdxPhi = *it;
2024 assert(RdxPhi && "Unable to recover vectorized PHI");
2026 // Find the reduction variable descriptor.
2027 assert(Legal->getReductionVars()->count(RdxPhi) &&
2028 "Unable to find the reduction variable");
2029 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2030 (*Legal->getReductionVars())[RdxPhi];
2032 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2034 // We need to generate a reduction vector from the incoming scalar.
2035 // To do so, we need to generate the 'identity' vector and overide
2036 // one of the elements with the incoming scalar reduction. We need
2037 // to do it in the vector-loop preheader.
2038 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2040 // This is the vector-clone of the value that leaves the loop.
2041 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2042 Type *VecTy = VectorExit[0]->getType();
2044 // Find the reduction identity variable. Zero for addition, or, xor,
2045 // one for multiplication, -1 for And.
2048 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2049 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2050 // MinMax reduction have the start value as their identify.
2052 VectorStart = Identity = RdxDesc.StartValue;
2054 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2059 // Handle other reduction kinds:
2061 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2062 VecTy->getScalarType());
2065 // This vector is the Identity vector where the first element is the
2066 // incoming scalar reduction.
2067 VectorStart = RdxDesc.StartValue;
2069 Identity = ConstantVector::getSplat(VF, Iden);
2071 // This vector is the Identity vector where the first element is the
2072 // incoming scalar reduction.
2073 VectorStart = Builder.CreateInsertElement(Identity,
2074 RdxDesc.StartValue, Zero);
2078 // Fix the vector-loop phi.
2079 // We created the induction variable so we know that the
2080 // preheader is the first entry.
2081 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2083 // Reductions do not have to start at zero. They can start with
2084 // any loop invariant values.
2085 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2086 BasicBlock *Latch = OrigLoop->getLoopLatch();
2087 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2088 VectorParts &Val = getVectorValue(LoopVal);
2089 for (unsigned part = 0; part < UF; ++part) {
2090 // Make sure to add the reduction stat value only to the
2091 // first unroll part.
2092 Value *StartVal = (part == 0) ? VectorStart : Identity;
2093 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2094 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2097 // Before each round, move the insertion point right between
2098 // the PHIs and the values we are going to write.
2099 // This allows us to write both PHINodes and the extractelement
2101 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2103 VectorParts RdxParts;
2104 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2105 for (unsigned part = 0; part < UF; ++part) {
2106 // This PHINode contains the vectorized reduction variable, or
2107 // the initial value vector, if we bypass the vector loop.
2108 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2109 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2110 Value *StartVal = (part == 0) ? VectorStart : Identity;
2111 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2112 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2113 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2114 RdxParts.push_back(NewPhi);
2117 // Reduce all of the unrolled parts into a single vector.
2118 Value *ReducedPartRdx = RdxParts[0];
2119 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2120 setDebugLocFromInst(Builder, ReducedPartRdx);
2121 for (unsigned part = 1; part < UF; ++part) {
2122 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2123 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2124 RdxParts[part], ReducedPartRdx,
2127 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2128 ReducedPartRdx, RdxParts[part]);
2132 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2133 // and vector ops, reducing the set of values being computed by half each
2135 assert(isPowerOf2_32(VF) &&
2136 "Reduction emission only supported for pow2 vectors!");
2137 Value *TmpVec = ReducedPartRdx;
2138 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2139 for (unsigned i = VF; i != 1; i >>= 1) {
2140 // Move the upper half of the vector to the lower half.
2141 for (unsigned j = 0; j != i/2; ++j)
2142 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2144 // Fill the rest of the mask with undef.
2145 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2146 UndefValue::get(Builder.getInt32Ty()));
2149 Builder.CreateShuffleVector(TmpVec,
2150 UndefValue::get(TmpVec->getType()),
2151 ConstantVector::get(ShuffleMask),
2154 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2155 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2158 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2161 // The result is in the first element of the vector.
2162 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2163 Builder.getInt32(0));
2166 // Now, we need to fix the users of the reduction variable
2167 // inside and outside of the scalar remainder loop.
2168 // We know that the loop is in LCSSA form. We need to update the
2169 // PHI nodes in the exit blocks.
2170 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2171 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2172 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2173 if (!LCSSAPhi) continue;
2175 // All PHINodes need to have a single entry edge, or two if
2176 // we already fixed them.
2177 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2179 // We found our reduction value exit-PHI. Update it with the
2180 // incoming bypass edge.
2181 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2182 // Add an edge coming from the bypass.
2183 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2186 }// end of the LCSSA phi scan.
2188 // Fix the scalar loop reduction variable with the incoming reduction sum
2189 // from the vector body and from the backedge value.
2190 int IncomingEdgeBlockIdx =
2191 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2192 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2193 // Pick the other block.
2194 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2195 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2196 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2197 }// end of for each redux variable.
2202 void InnerLoopVectorizer::fixLCSSAPHIs() {
2203 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2204 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2205 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2206 if (!LCSSAPhi) continue;
2207 if (LCSSAPhi->getNumIncomingValues() == 1)
2208 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2213 InnerLoopVectorizer::VectorParts
2214 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2215 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2218 // Look for cached value.
2219 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2220 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2221 if (ECEntryIt != MaskCache.end())
2222 return ECEntryIt->second;
2224 VectorParts SrcMask = createBlockInMask(Src);
2226 // The terminator has to be a branch inst!
2227 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2228 assert(BI && "Unexpected terminator found");
2230 if (BI->isConditional()) {
2231 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2233 if (BI->getSuccessor(0) != Dst)
2234 for (unsigned part = 0; part < UF; ++part)
2235 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2237 for (unsigned part = 0; part < UF; ++part)
2238 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2240 MaskCache[Edge] = EdgeMask;
2244 MaskCache[Edge] = SrcMask;
2248 InnerLoopVectorizer::VectorParts
2249 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2250 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2252 // Loop incoming mask is all-one.
2253 if (OrigLoop->getHeader() == BB) {
2254 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2255 return getVectorValue(C);
2258 // This is the block mask. We OR all incoming edges, and with zero.
2259 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2260 VectorParts BlockMask = getVectorValue(Zero);
2263 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2264 VectorParts EM = createEdgeMask(*it, BB);
2265 for (unsigned part = 0; part < UF; ++part)
2266 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2272 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2273 InnerLoopVectorizer::VectorParts &Entry,
2274 LoopVectorizationLegality *Legal,
2275 unsigned UF, unsigned VF, PhiVector *PV) {
2276 PHINode* P = cast<PHINode>(PN);
2277 // Handle reduction variables:
2278 if (Legal->getReductionVars()->count(P)) {
2279 for (unsigned part = 0; part < UF; ++part) {
2280 // This is phase one of vectorizing PHIs.
2281 Type *VecTy = (VF == 1) ? PN->getType() :
2282 VectorType::get(PN->getType(), VF);
2283 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2284 LoopVectorBody-> getFirstInsertionPt());
2290 setDebugLocFromInst(Builder, P);
2291 // Check for PHI nodes that are lowered to vector selects.
2292 if (P->getParent() != OrigLoop->getHeader()) {
2293 // We know that all PHIs in non header blocks are converted into
2294 // selects, so we don't have to worry about the insertion order and we
2295 // can just use the builder.
2296 // At this point we generate the predication tree. There may be
2297 // duplications since this is a simple recursive scan, but future
2298 // optimizations will clean it up.
2300 unsigned NumIncoming = P->getNumIncomingValues();
2302 // Generate a sequence of selects of the form:
2303 // SELECT(Mask3, In3,
2304 // SELECT(Mask2, In2,
2306 for (unsigned In = 0; In < NumIncoming; In++) {
2307 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2309 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2311 for (unsigned part = 0; part < UF; ++part) {
2312 // We might have single edge PHIs (blocks) - use an identity
2313 // 'select' for the first PHI operand.
2315 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2318 // Select between the current value and the previous incoming edge
2319 // based on the incoming mask.
2320 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2321 Entry[part], "predphi");
2327 // This PHINode must be an induction variable.
2328 // Make sure that we know about it.
2329 assert(Legal->getInductionVars()->count(P) &&
2330 "Not an induction variable");
2332 LoopVectorizationLegality::InductionInfo II =
2333 Legal->getInductionVars()->lookup(P);
2336 case LoopVectorizationLegality::IK_NoInduction:
2337 llvm_unreachable("Unknown induction");
2338 case LoopVectorizationLegality::IK_IntInduction: {
2339 assert(P->getType() == II.StartValue->getType() && "Types must match");
2340 Type *PhiTy = P->getType();
2342 if (P == OldInduction) {
2343 // Handle the canonical induction variable. We might have had to
2345 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2347 // Handle other induction variables that are now based on the
2349 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2351 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2352 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2355 Broadcasted = getBroadcastInstrs(Broadcasted);
2356 // After broadcasting the induction variable we need to make the vector
2357 // consecutive by adding 0, 1, 2, etc.
2358 for (unsigned part = 0; part < UF; ++part)
2359 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2362 case LoopVectorizationLegality::IK_ReverseIntInduction:
2363 case LoopVectorizationLegality::IK_PtrInduction:
2364 case LoopVectorizationLegality::IK_ReversePtrInduction:
2365 // Handle reverse integer and pointer inductions.
2366 Value *StartIdx = ExtendedIdx;
2367 // This is the normalized GEP that starts counting at zero.
2368 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2371 // Handle the reverse integer induction variable case.
2372 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2373 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2374 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2376 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2379 // This is a new value so do not hoist it out.
2380 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2381 // After broadcasting the induction variable we need to make the
2382 // vector consecutive by adding ... -3, -2, -1, 0.
2383 for (unsigned part = 0; part < UF; ++part)
2384 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2389 // Handle the pointer induction variable case.
2390 assert(P->getType()->isPointerTy() && "Unexpected type.");
2392 // Is this a reverse induction ptr or a consecutive induction ptr.
2393 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2396 // This is the vector of results. Notice that we don't generate
2397 // vector geps because scalar geps result in better code.
2398 for (unsigned part = 0; part < UF; ++part) {
2400 int EltIndex = (part) * (Reverse ? -1 : 1);
2401 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2404 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2406 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2408 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2410 Entry[part] = SclrGep;
2414 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2415 for (unsigned int i = 0; i < VF; ++i) {
2416 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2417 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2420 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2422 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2424 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2426 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2427 Builder.getInt32(i),
2430 Entry[part] = VecVal;
2437 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2438 BasicBlock *BB, PhiVector *PV) {
2439 // For each instruction in the old loop.
2440 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2441 VectorParts &Entry = WidenMap.get(it);
2442 switch (it->getOpcode()) {
2443 case Instruction::Br:
2444 // Nothing to do for PHIs and BR, since we already took care of the
2445 // loop control flow instructions.
2447 case Instruction::PHI:{
2448 // Vectorize PHINodes.
2449 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2453 case Instruction::Add:
2454 case Instruction::FAdd:
2455 case Instruction::Sub:
2456 case Instruction::FSub:
2457 case Instruction::Mul:
2458 case Instruction::FMul:
2459 case Instruction::UDiv:
2460 case Instruction::SDiv:
2461 case Instruction::FDiv:
2462 case Instruction::URem:
2463 case Instruction::SRem:
2464 case Instruction::FRem:
2465 case Instruction::Shl:
2466 case Instruction::LShr:
2467 case Instruction::AShr:
2468 case Instruction::And:
2469 case Instruction::Or:
2470 case Instruction::Xor: {
2471 // Just widen binops.
2472 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2473 setDebugLocFromInst(Builder, BinOp);
2474 VectorParts &A = getVectorValue(it->getOperand(0));
2475 VectorParts &B = getVectorValue(it->getOperand(1));
2477 // Use this vector value for all users of the original instruction.
2478 for (unsigned Part = 0; Part < UF; ++Part) {
2479 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2481 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2482 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2483 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2484 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2485 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2487 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2488 VecOp->setIsExact(BinOp->isExact());
2494 case Instruction::Select: {
2496 // If the selector is loop invariant we can create a select
2497 // instruction with a scalar condition. Otherwise, use vector-select.
2498 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2500 setDebugLocFromInst(Builder, it);
2502 // The condition can be loop invariant but still defined inside the
2503 // loop. This means that we can't just use the original 'cond' value.
2504 // We have to take the 'vectorized' value and pick the first lane.
2505 // Instcombine will make this a no-op.
2506 VectorParts &Cond = getVectorValue(it->getOperand(0));
2507 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2508 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2510 Value *ScalarCond = (VF == 1) ? Cond[0] :
2511 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2513 for (unsigned Part = 0; Part < UF; ++Part) {
2514 Entry[Part] = Builder.CreateSelect(
2515 InvariantCond ? ScalarCond : Cond[Part],
2522 case Instruction::ICmp:
2523 case Instruction::FCmp: {
2524 // Widen compares. Generate vector compares.
2525 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2526 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2527 setDebugLocFromInst(Builder, it);
2528 VectorParts &A = getVectorValue(it->getOperand(0));
2529 VectorParts &B = getVectorValue(it->getOperand(1));
2530 for (unsigned Part = 0; Part < UF; ++Part) {
2533 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2535 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2541 case Instruction::Store:
2542 case Instruction::Load:
2543 vectorizeMemoryInstruction(it, Legal);
2545 case Instruction::ZExt:
2546 case Instruction::SExt:
2547 case Instruction::FPToUI:
2548 case Instruction::FPToSI:
2549 case Instruction::FPExt:
2550 case Instruction::PtrToInt:
2551 case Instruction::IntToPtr:
2552 case Instruction::SIToFP:
2553 case Instruction::UIToFP:
2554 case Instruction::Trunc:
2555 case Instruction::FPTrunc:
2556 case Instruction::BitCast: {
2557 CastInst *CI = dyn_cast<CastInst>(it);
2558 setDebugLocFromInst(Builder, it);
2559 /// Optimize the special case where the source is the induction
2560 /// variable. Notice that we can only optimize the 'trunc' case
2561 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2562 /// c. other casts depend on pointer size.
2563 if (CI->getOperand(0) == OldInduction &&
2564 it->getOpcode() == Instruction::Trunc) {
2565 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2567 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2568 for (unsigned Part = 0; Part < UF; ++Part)
2569 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2572 /// Vectorize casts.
2573 Type *DestTy = (VF == 1) ? CI->getType() :
2574 VectorType::get(CI->getType(), VF);
2576 VectorParts &A = getVectorValue(it->getOperand(0));
2577 for (unsigned Part = 0; Part < UF; ++Part)
2578 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2582 case Instruction::Call: {
2583 // Ignore dbg intrinsics.
2584 if (isa<DbgInfoIntrinsic>(it))
2586 setDebugLocFromInst(Builder, it);
2588 Module *M = BB->getParent()->getParent();
2589 CallInst *CI = cast<CallInst>(it);
2590 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2591 assert(ID && "Not an intrinsic call!");
2593 case Intrinsic::lifetime_end:
2594 case Intrinsic::lifetime_start:
2595 scalarizeInstruction(it);
2598 for (unsigned Part = 0; Part < UF; ++Part) {
2599 SmallVector<Value *, 4> Args;
2600 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2601 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2602 Args.push_back(Arg[Part]);
2604 Type *Tys[] = {CI->getType()};
2606 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2608 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2609 Entry[Part] = Builder.CreateCall(F, Args);
2617 // All other instructions are unsupported. Scalarize them.
2618 scalarizeInstruction(it);
2621 }// end of for_each instr.
2624 void InnerLoopVectorizer::updateAnalysis() {
2625 // Forget the original basic block.
2626 SE->forgetLoop(OrigLoop);
2628 // Update the dominator tree information.
2629 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2630 "Entry does not dominate exit.");
2632 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2633 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2634 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2635 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2636 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2637 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2638 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2639 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2641 DEBUG(DT->verifyAnalysis());
2644 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2645 if (!EnableIfConversion)
2648 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2649 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2651 // A list of pointers that we can safely read and write to.
2652 SmallPtrSet<Value *, 8> SafePointes;
2654 // Collect safe addresses.
2655 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2656 BasicBlock *BB = LoopBlocks[i];
2658 if (blockNeedsPredication(BB))
2661 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2662 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2663 SafePointes.insert(LI->getPointerOperand());
2664 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2665 SafePointes.insert(SI->getPointerOperand());
2669 // Collect the blocks that need predication.
2670 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2671 BasicBlock *BB = LoopBlocks[i];
2673 // We don't support switch statements inside loops.
2674 if (!isa<BranchInst>(BB->getTerminator()))
2677 // We must be able to predicate all blocks that need to be predicated.
2678 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2682 // We can if-convert this loop.
2686 bool LoopVectorizationLegality::canVectorize() {
2687 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2688 // be canonicalized.
2689 if (!TheLoop->getLoopPreheader())
2692 // We can only vectorize innermost loops.
2693 if (TheLoop->getSubLoopsVector().size())
2696 // We must have a single backedge.
2697 if (TheLoop->getNumBackEdges() != 1)
2700 // We must have a single exiting block.
2701 if (!TheLoop->getExitingBlock())
2704 unsigned NumBlocks = TheLoop->getNumBlocks();
2706 // Check if we can if-convert non single-bb loops.
2707 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2708 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2712 // We need to have a loop header.
2713 BasicBlock *Latch = TheLoop->getLoopLatch();
2714 DEBUG(dbgs() << "LV: Found a loop: " <<
2715 TheLoop->getHeader()->getName() << "\n");
2717 // ScalarEvolution needs to be able to find the exit count.
2718 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2719 if (ExitCount == SE->getCouldNotCompute()) {
2720 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2724 // Do not loop-vectorize loops with a tiny trip count.
2725 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2726 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2727 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2728 "This loop is not worth vectorizing.\n");
2732 // Check if we can vectorize the instructions and CFG in this loop.
2733 if (!canVectorizeInstrs()) {
2734 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2738 // Go over each instruction and look at memory deps.
2739 if (!canVectorizeMemory()) {
2740 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2744 // Collect all of the variables that remain uniform after vectorization.
2745 collectLoopUniforms();
2747 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2748 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2751 // Okay! We can vectorize. At this point we don't have any other mem analysis
2752 // which may limit our maximum vectorization factor, so just return true with
2757 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2758 if (Ty->isPointerTy())
2759 return DL.getIntPtrType(Ty);
2764 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2765 Ty0 = convertPointerToIntegerType(DL, Ty0);
2766 Ty1 = convertPointerToIntegerType(DL, Ty1);
2767 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2772 /// \brief Check that the instruction has outside loop users and is not an
2773 /// identified reduction variable.
2774 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2775 SmallPtrSet<Value *, 4> &Reductions) {
2776 // Reduction instructions are allowed to have exit users. All other
2777 // instructions must not have external users.
2778 if (!Reductions.count(Inst))
2779 //Check that all of the users of the loop are inside the BB.
2780 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2782 Instruction *U = cast<Instruction>(*I);
2783 // This user may be a reduction exit value.
2784 if (!TheLoop->contains(U)) {
2785 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2792 bool LoopVectorizationLegality::canVectorizeInstrs() {
2793 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2794 BasicBlock *Header = TheLoop->getHeader();
2796 // Look for the attribute signaling the absence of NaNs.
2797 Function &F = *Header->getParent();
2798 if (F.hasFnAttribute("no-nans-fp-math"))
2799 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2800 AttributeSet::FunctionIndex,
2801 "no-nans-fp-math").getValueAsString() == "true";
2803 // For each block in the loop.
2804 for (Loop::block_iterator bb = TheLoop->block_begin(),
2805 be = TheLoop->block_end(); bb != be; ++bb) {
2807 // Scan the instructions in the block and look for hazards.
2808 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2811 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2812 Type *PhiTy = Phi->getType();
2813 // Check that this PHI type is allowed.
2814 if (!PhiTy->isIntegerTy() &&
2815 !PhiTy->isFloatingPointTy() &&
2816 !PhiTy->isPointerTy()) {
2817 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2821 // If this PHINode is not in the header block, then we know that we
2822 // can convert it to select during if-conversion. No need to check if
2823 // the PHIs in this block are induction or reduction variables.
2824 if (*bb != Header) {
2825 // Check that this instruction has no outside users or is an
2826 // identified reduction value with an outside user.
2827 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2832 // We only allow if-converted PHIs with more than two incoming values.
2833 if (Phi->getNumIncomingValues() != 2) {
2834 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2838 // This is the value coming from the preheader.
2839 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2840 // Check if this is an induction variable.
2841 InductionKind IK = isInductionVariable(Phi);
2843 if (IK_NoInduction != IK) {
2844 // Get the widest type.
2846 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2848 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2850 // Int inductions are special because we only allow one IV.
2851 if (IK == IK_IntInduction) {
2852 // Use the phi node with the widest type as induction. Use the last
2853 // one if there are multiple (no good reason for doing this other
2854 // than it is expedient).
2855 if (!Induction || PhiTy == WidestIndTy)
2859 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2860 Inductions[Phi] = InductionInfo(StartValue, IK);
2864 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2865 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2868 if (AddReductionVar(Phi, RK_IntegerMult)) {
2869 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2872 if (AddReductionVar(Phi, RK_IntegerOr)) {
2873 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2876 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2877 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2880 if (AddReductionVar(Phi, RK_IntegerXor)) {
2881 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2884 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2885 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2888 if (AddReductionVar(Phi, RK_FloatMult)) {
2889 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2892 if (AddReductionVar(Phi, RK_FloatAdd)) {
2893 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2896 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2897 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
2902 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2904 }// end of PHI handling
2906 // We still don't handle functions. However, we can ignore dbg intrinsic
2907 // calls and we do handle certain intrinsic and libm functions.
2908 CallInst *CI = dyn_cast<CallInst>(it);
2909 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2910 DEBUG(dbgs() << "LV: Found a call site.\n");
2914 // Check that the instruction return type is vectorizable.
2915 if (!VectorType::isValidElementType(it->getType()) &&
2916 !it->getType()->isVoidTy()) {
2917 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2921 // Check that the stored type is vectorizable.
2922 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2923 Type *T = ST->getValueOperand()->getType();
2924 if (!VectorType::isValidElementType(T))
2928 // Reduction instructions are allowed to have exit users.
2929 // All other instructions must not have external users.
2930 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2938 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2939 if (Inductions.empty())
2946 void LoopVectorizationLegality::collectLoopUniforms() {
2947 // We now know that the loop is vectorizable!
2948 // Collect variables that will remain uniform after vectorization.
2949 std::vector<Value*> Worklist;
2950 BasicBlock *Latch = TheLoop->getLoopLatch();
2952 // Start with the conditional branch and walk up the block.
2953 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2955 while (Worklist.size()) {
2956 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2957 Worklist.pop_back();
2959 // Look at instructions inside this loop.
2960 // Stop when reaching PHI nodes.
2961 // TODO: we need to follow values all over the loop, not only in this block.
2962 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2965 // This is a known uniform.
2968 // Insert all operands.
2969 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2974 /// \brief Analyses memory accesses in a loop.
2976 /// Checks whether run time pointer checks are needed and builds sets for data
2977 /// dependence checking.
2978 class AccessAnalysis {
2980 /// \brief Read or write access location.
2981 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
2982 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
2984 /// \brief Set of potential dependent memory accesses.
2985 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2987 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2988 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2989 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2991 /// \brief Register a load and whether it is only read from.
2992 void addLoad(Value *Ptr, bool IsReadOnly) {
2993 Accesses.insert(MemAccessInfo(Ptr, false));
2995 ReadOnlyPtr.insert(Ptr);
2998 /// \brief Register a store.
2999 void addStore(Value *Ptr) {
3000 Accesses.insert(MemAccessInfo(Ptr, true));
3003 /// \brief Check whether we can check the pointers at runtime for
3004 /// non-intersection.
3005 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3006 unsigned &NumComparisons, ScalarEvolution *SE,
3009 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3010 /// and builds sets of dependent accesses.
3011 void buildDependenceSets() {
3012 // Process read-write pointers first.
3013 processMemAccesses(false);
3014 // Next, process read pointers.
3015 processMemAccesses(true);
3018 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3020 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3022 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3025 typedef SetVector<MemAccessInfo> PtrAccessSet;
3026 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3028 /// \brief Go over all memory access or only the deferred ones if
3029 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3030 /// and build sets of dependency check candidates.
3031 void processMemAccesses(bool UseDeferred);
3033 /// Set of all accesses.
3034 PtrAccessSet Accesses;
3036 /// Set of access to check after all writes have been processed.
3037 PtrAccessSet DeferredAccesses;
3039 /// Map of pointers to last access encountered.
3040 UnderlyingObjToAccessMap ObjToLastAccess;
3042 /// Set of accesses that need a further dependence check.
3043 MemAccessInfoSet CheckDeps;
3045 /// Set of pointers that are read only.
3046 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3048 /// Set of underlying objects already written to.
3049 SmallPtrSet<Value*, 16> WriteObjects;
3053 /// Sets of potentially dependent accesses - members of one set share an
3054 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3055 /// dependence check.
3056 DepCandidates &DepCands;
3058 bool AreAllWritesIdentified;
3059 bool AreAllReadsIdentified;
3060 bool IsRTCheckNeeded;
3063 } // end anonymous namespace
3065 /// \brief Check whether a pointer can participate in a runtime bounds check.
3066 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3067 const SCEV *PtrScev = SE->getSCEV(Ptr);
3068 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3072 return AR->isAffine();
3075 bool AccessAnalysis::canCheckPtrAtRT(
3076 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3077 unsigned &NumComparisons, ScalarEvolution *SE,
3079 // Find pointers with computable bounds. We are going to use this information
3080 // to place a runtime bound check.
3081 unsigned NumReadPtrChecks = 0;
3082 unsigned NumWritePtrChecks = 0;
3083 bool CanDoRT = true;
3085 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3086 // We assign consecutive id to access from different dependence sets.
3087 // Accesses within the same set don't need a runtime check.
3088 unsigned RunningDepId = 1;
3089 DenseMap<Value *, unsigned> DepSetId;
3091 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3093 const MemAccessInfo &Access = *AI;
3094 Value *Ptr = Access.getPointer();
3095 bool IsWrite = Access.getInt();
3097 // Just add write checks if we have both.
3098 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3102 ++NumWritePtrChecks;
3106 if (hasComputableBounds(SE, Ptr)) {
3107 // The id of the dependence set.
3110 if (IsDepCheckNeeded) {
3111 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3112 unsigned &LeaderId = DepSetId[Leader];
3114 LeaderId = RunningDepId++;
3117 // Each access has its own dependence set.
3118 DepId = RunningDepId++;
3120 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3122 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3128 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3129 NumComparisons = 0; // Only one dependence set.
3131 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3132 NumWritePtrChecks - 1));
3136 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3137 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3140 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3141 // We process the set twice: first we process read-write pointers, last we
3142 // process read-only pointers. This allows us to skip dependence tests for
3143 // read-only pointers.
3145 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3146 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3147 const MemAccessInfo &Access = *AI;
3148 Value *Ptr = Access.getPointer();
3149 bool IsWrite = Access.getInt();
3151 DepCands.insert(Access);
3153 // Memorize read-only pointers for later processing and skip them in the
3154 // first round (they need to be checked after we have seen all write
3155 // pointers). Note: we also mark pointer that are not consecutive as
3156 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3157 // second check for "!IsWrite".
3158 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3159 if (!UseDeferred && IsReadOnlyPtr) {
3160 DeferredAccesses.insert(Access);
3164 bool NeedDepCheck = false;
3165 // Check whether there is the possiblity of dependency because of underlying
3166 // objects being the same.
3167 typedef SmallVector<Value*, 16> ValueVector;
3168 ValueVector TempObjects;
3169 GetUnderlyingObjects(Ptr, TempObjects, DL);
3170 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3172 Value *UnderlyingObj = *UI;
3174 // If this is a write then it needs to be an identified object. If this a
3175 // read and all writes (so far) are identified function scope objects we
3176 // don't need an identified underlying object but only an Argument (the
3177 // next write is going to invalidate this assumption if it is
3179 // This is a micro-optimization for the case where all writes are
3180 // identified and we have one argument pointer.
3181 // Otherwise, we do need a runtime check.
3182 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3183 (!IsWrite && (!AreAllWritesIdentified ||
3184 !isa<Argument>(UnderlyingObj)) &&
3185 !isIdentifiedObject(UnderlyingObj))) {
3186 DEBUG(dbgs() << "LV: Found an unidentified " <<
3187 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3189 IsRTCheckNeeded = (IsRTCheckNeeded ||
3190 !isIdentifiedObject(UnderlyingObj) ||
3191 !AreAllReadsIdentified);
3194 AreAllWritesIdentified = false;
3196 AreAllReadsIdentified = false;
3199 // If this is a write - check other reads and writes for conflicts. If
3200 // this is a read only check other writes for conflicts (but only if there
3201 // is no other write to the ptr - this is an optimization to catch "a[i] =
3202 // a[i] + " without having to do a dependence check).
3203 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3204 NeedDepCheck = true;
3207 WriteObjects.insert(UnderlyingObj);
3209 // Create sets of pointers connected by shared underlying objects.
3210 UnderlyingObjToAccessMap::iterator Prev =
3211 ObjToLastAccess.find(UnderlyingObj);
3212 if (Prev != ObjToLastAccess.end())
3213 DepCands.unionSets(Access, Prev->second);
3215 ObjToLastAccess[UnderlyingObj] = Access;
3219 CheckDeps.insert(Access);
3224 /// \brief Checks memory dependences among accesses to the same underlying
3225 /// object to determine whether there vectorization is legal or not (and at
3226 /// which vectorization factor).
3228 /// This class works under the assumption that we already checked that memory
3229 /// locations with different underlying pointers are "must-not alias".
3230 /// We use the ScalarEvolution framework to symbolically evalutate access
3231 /// functions pairs. Since we currently don't restructure the loop we can rely
3232 /// on the program order of memory accesses to determine their safety.
3233 /// At the moment we will only deem accesses as safe for:
3234 /// * A negative constant distance assuming program order.
3236 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3237 /// a[i] = tmp; y = a[i];
3239 /// The latter case is safe because later checks guarantuee that there can't
3240 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3241 /// the same variable: a header phi can only be an induction or a reduction, a
3242 /// reduction can't have a memory sink, an induction can't have a memory
3243 /// source). This is important and must not be violated (or we have to
3244 /// resort to checking for cycles through memory).
3246 /// * A positive constant distance assuming program order that is bigger
3247 /// than the biggest memory access.
3249 /// tmp = a[i] OR b[i] = x
3250 /// a[i+2] = tmp y = b[i+2];
3252 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3254 /// * Zero distances and all accesses have the same size.
3256 class MemoryDepChecker {
3258 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3259 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3261 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3262 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3264 /// \brief Register the location (instructions are given increasing numbers)
3265 /// of a write access.
3266 void addAccess(StoreInst *SI) {
3267 Value *Ptr = SI->getPointerOperand();
3268 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3269 InstMap.push_back(SI);
3273 /// \brief Register the location (instructions are given increasing numbers)
3274 /// of a write access.
3275 void addAccess(LoadInst *LI) {
3276 Value *Ptr = LI->getPointerOperand();
3277 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3278 InstMap.push_back(LI);
3282 /// \brief Check whether the dependencies between the accesses are safe.
3284 /// Only checks sets with elements in \p CheckDeps.
3285 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3286 MemAccessInfoSet &CheckDeps);
3288 /// \brief The maximum number of bytes of a vector register we can vectorize
3289 /// the accesses safely with.
3290 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3293 ScalarEvolution *SE;
3295 const Loop *InnermostLoop;
3297 /// \brief Maps access locations (ptr, read/write) to program order.
3298 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3300 /// \brief Memory access instructions in program order.
3301 SmallVector<Instruction *, 16> InstMap;
3303 /// \brief The program order index to be used for the next instruction.
3306 // We can access this many bytes in parallel safely.
3307 unsigned MaxSafeDepDistBytes;
3309 /// \brief Check whether there is a plausible dependence between the two
3312 /// Access \p A must happen before \p B in program order. The two indices
3313 /// identify the index into the program order map.
3315 /// This function checks whether there is a plausible dependence (or the
3316 /// absence of such can't be proved) between the two accesses. If there is a
3317 /// plausible dependence but the dependence distance is bigger than one
3318 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3319 /// distance is smaller than any other distance encountered so far).
3320 /// Otherwise, this function returns true signaling a possible dependence.
3321 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3322 const MemAccessInfo &B, unsigned BIdx);
3324 /// \brief Check whether the data dependence could prevent store-load
3326 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3329 } // end anonymous namespace
3331 static bool isInBoundsGep(Value *Ptr) {
3332 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3333 return GEP->isInBounds();
3337 /// \brief Check whether the access through \p Ptr has a constant stride.
3338 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3340 const Type *Ty = Ptr->getType();
3341 assert(Ty->isPointerTy() && "Unexpected non ptr");
3343 // Make sure that the pointer does not point to aggregate types.
3344 const PointerType *PtrTy = cast<PointerType>(Ty);
3345 if (PtrTy->getElementType()->isAggregateType()) {
3346 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3351 const SCEV *PtrScev = SE->getSCEV(Ptr);
3352 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3354 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3355 << *Ptr << " SCEV: " << *PtrScev << "\n");
3359 // The accesss function must stride over the innermost loop.
3360 if (Lp != AR->getLoop()) {
3361 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3362 *Ptr << " SCEV: " << *PtrScev << "\n");
3365 // The address calculation must not wrap. Otherwise, a dependence could be
3367 // An inbounds getelementptr that is a AddRec with a unit stride
3368 // cannot wrap per definition. The unit stride requirement is checked later.
3369 // An getelementptr without an inbounds attribute and unit stride would have
3370 // to access the pointer value "0" which is undefined behavior in address
3371 // space 0, therefore we can also vectorize this case.
3372 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3373 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3374 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3375 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3376 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3377 << *Ptr << " SCEV: " << *PtrScev << "\n");
3381 // Check the step is constant.
3382 const SCEV *Step = AR->getStepRecurrence(*SE);
3384 // Calculate the pointer stride and check if it is consecutive.
3385 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3387 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3388 " SCEV: " << *PtrScev << "\n");
3392 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3393 const APInt &APStepVal = C->getValue()->getValue();
3395 // Huge step value - give up.
3396 if (APStepVal.getBitWidth() > 64)
3399 int64_t StepVal = APStepVal.getSExtValue();
3402 int64_t Stride = StepVal / Size;
3403 int64_t Rem = StepVal % Size;
3407 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3408 // know we can't "wrap around the address space". In case of address space
3409 // zero we know that this won't happen without triggering undefined behavior.
3410 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3411 Stride != 1 && Stride != -1)
3417 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3418 unsigned TypeByteSize) {
3419 // If loads occur at a distance that is not a multiple of a feasible vector
3420 // factor store-load forwarding does not take place.
3421 // Positive dependences might cause troubles because vectorizing them might
3422 // prevent store-load forwarding making vectorized code run a lot slower.
3423 // a[i] = a[i-3] ^ a[i-8];
3424 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3425 // hence on your typical architecture store-load forwarding does not take
3426 // place. Vectorizing in such cases does not make sense.
3427 // Store-load forwarding distance.
3428 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3429 // Maximum vector factor.
3430 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3431 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3432 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3434 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3436 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3437 MaxVFWithoutSLForwardIssues = (vf >>=1);
3442 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3443 DEBUG(dbgs() << "LV: Distance " << Distance <<
3444 " that could cause a store-load forwarding conflict\n");
3448 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3449 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3450 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3454 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3455 const MemAccessInfo &B, unsigned BIdx) {
3456 assert (AIdx < BIdx && "Must pass arguments in program order");
3458 Value *APtr = A.getPointer();
3459 Value *BPtr = B.getPointer();
3460 bool AIsWrite = A.getInt();
3461 bool BIsWrite = B.getInt();
3463 // Two reads are independent.
3464 if (!AIsWrite && !BIsWrite)
3467 const SCEV *AScev = SE->getSCEV(APtr);
3468 const SCEV *BScev = SE->getSCEV(BPtr);
3470 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3471 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3473 const SCEV *Src = AScev;
3474 const SCEV *Sink = BScev;
3476 // If the induction step is negative we have to invert source and sink of the
3478 if (StrideAPtr < 0) {
3481 std::swap(APtr, BPtr);
3482 std::swap(Src, Sink);
3483 std::swap(AIsWrite, BIsWrite);
3484 std::swap(AIdx, BIdx);
3485 std::swap(StrideAPtr, StrideBPtr);
3488 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3490 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3491 << "(Induction step: " << StrideAPtr << ")\n");
3492 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3493 << *InstMap[BIdx] << ": " << *Dist << "\n");
3495 // Need consecutive accesses. We don't want to vectorize
3496 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3497 // the address space.
3498 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3499 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3503 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3505 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3509 Type *ATy = APtr->getType()->getPointerElementType();
3510 Type *BTy = BPtr->getType()->getPointerElementType();
3511 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3513 // Negative distances are not plausible dependencies.
3514 const APInt &Val = C->getValue()->getValue();
3515 if (Val.isNegative()) {
3516 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3517 if (IsTrueDataDependence &&
3518 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3522 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3526 // Write to the same location with the same size.
3527 // Could be improved to assert type sizes are the same (i32 == float, etc).
3531 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3535 assert(Val.isStrictlyPositive() && "Expect a positive value");
3537 // Positive distance bigger than max vectorization factor.
3540 "LV: ReadWrite-Write positive dependency with different types");
3544 unsigned Distance = (unsigned) Val.getZExtValue();
3546 // Bail out early if passed-in parameters make vectorization not feasible.
3547 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3548 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3550 // The distance must be bigger than the size needed for a vectorized version
3551 // of the operation and the size of the vectorized operation must not be
3552 // bigger than the currrent maximum size.
3553 if (Distance < 2*TypeByteSize ||
3554 2*TypeByteSize > MaxSafeDepDistBytes ||
3555 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3556 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3557 << Val.getSExtValue() << "\n");
3561 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3562 Distance : MaxSafeDepDistBytes;
3564 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3565 if (IsTrueDataDependence &&
3566 couldPreventStoreLoadForward(Distance, TypeByteSize))
3569 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3570 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3576 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3577 MemAccessInfoSet &CheckDeps) {
3579 MaxSafeDepDistBytes = -1U;
3580 while (!CheckDeps.empty()) {
3581 MemAccessInfo CurAccess = *CheckDeps.begin();
3583 // Get the relevant memory access set.
3584 EquivalenceClasses<MemAccessInfo>::iterator I =
3585 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3587 // Check accesses within this set.
3588 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3589 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3591 // Check every access pair.
3593 CheckDeps.erase(*AI);
3594 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3596 // Check every accessing instruction pair in program order.
3597 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3598 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3599 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3600 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3601 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3603 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3614 bool LoopVectorizationLegality::canVectorizeMemory() {
3616 typedef SmallVector<Value*, 16> ValueVector;
3617 typedef SmallPtrSet<Value*, 16> ValueSet;
3619 // Holds the Load and Store *instructions*.
3623 // Holds all the different accesses in the loop.
3624 unsigned NumReads = 0;
3625 unsigned NumReadWrites = 0;
3627 PtrRtCheck.Pointers.clear();
3628 PtrRtCheck.Need = false;
3630 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3631 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3634 for (Loop::block_iterator bb = TheLoop->block_begin(),
3635 be = TheLoop->block_end(); bb != be; ++bb) {
3637 // Scan the BB and collect legal loads and stores.
3638 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3641 // If this is a load, save it. If this instruction can read from memory
3642 // but is not a load, then we quit. Notice that we don't handle function
3643 // calls that read or write.
3644 if (it->mayReadFromMemory()) {
3645 // Many math library functions read the rounding mode. We will only
3646 // vectorize a loop if it contains known function calls that don't set
3647 // the flag. Therefore, it is safe to ignore this read from memory.
3648 CallInst *Call = dyn_cast<CallInst>(it);
3649 if (Call && getIntrinsicIDForCall(Call, TLI))
3652 LoadInst *Ld = dyn_cast<LoadInst>(it);
3653 if (!Ld) return false;
3654 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3655 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3658 Loads.push_back(Ld);
3659 DepChecker.addAccess(Ld);
3663 // Save 'store' instructions. Abort if other instructions write to memory.
3664 if (it->mayWriteToMemory()) {
3665 StoreInst *St = dyn_cast<StoreInst>(it);
3666 if (!St) return false;
3667 if (!St->isSimple() && !IsAnnotatedParallel) {
3668 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3671 Stores.push_back(St);
3672 DepChecker.addAccess(St);
3677 // Now we have two lists that hold the loads and the stores.
3678 // Next, we find the pointers that they use.
3680 // Check if we see any stores. If there are no stores, then we don't
3681 // care if the pointers are *restrict*.
3682 if (!Stores.size()) {
3683 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3687 AccessAnalysis::DepCandidates DependentAccesses;
3688 AccessAnalysis Accesses(DL, DependentAccesses);
3690 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3691 // multiple times on the same object. If the ptr is accessed twice, once
3692 // for read and once for write, it will only appear once (on the write
3693 // list). This is okay, since we are going to check for conflicts between
3694 // writes and between reads and writes, but not between reads and reads.
3697 ValueVector::iterator I, IE;
3698 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3699 StoreInst *ST = cast<StoreInst>(*I);
3700 Value* Ptr = ST->getPointerOperand();
3702 if (isUniform(Ptr)) {
3703 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3707 // If we did *not* see this pointer before, insert it to the read-write
3708 // list. At this phase it is only a 'write' list.
3709 if (Seen.insert(Ptr)) {
3711 Accesses.addStore(Ptr);
3715 if (IsAnnotatedParallel) {
3717 << "LV: A loop annotated parallel, ignore memory dependency "
3722 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3723 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3724 LoadInst *LD = cast<LoadInst>(*I);
3725 Value* Ptr = LD->getPointerOperand();
3726 // If we did *not* see this pointer before, insert it to the
3727 // read list. If we *did* see it before, then it is already in
3728 // the read-write list. This allows us to vectorize expressions
3729 // such as A[i] += x; Because the address of A[i] is a read-write
3730 // pointer. This only works if the index of A[i] is consecutive.
3731 // If the address of i is unknown (for example A[B[i]]) then we may
3732 // read a few words, modify, and write a few words, and some of the
3733 // words may be written to the same address.
3734 bool IsReadOnlyPtr = false;
3735 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3737 IsReadOnlyPtr = true;
3739 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3742 // If we write (or read-write) to a single destination and there are no
3743 // other reads in this loop then is it safe to vectorize.
3744 if (NumReadWrites == 1 && NumReads == 0) {
3745 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3749 // Build dependence sets and check whether we need a runtime pointer bounds
3751 Accesses.buildDependenceSets();
3752 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3754 // Find pointers with computable bounds. We are going to use this information
3755 // to place a runtime bound check.
3756 unsigned NumComparisons = 0;
3757 bool CanDoRT = false;
3759 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3762 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3763 " pointer comparisons.\n");
3765 // If we only have one set of dependences to check pointers among we don't
3766 // need a runtime check.
3767 if (NumComparisons == 0 && NeedRTCheck)
3768 NeedRTCheck = false;
3770 // Check that we did not collect too many pointers or found a unsizeable
3772 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3778 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3781 if (NeedRTCheck && !CanDoRT) {
3782 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3783 "the array bounds.\n");
3788 PtrRtCheck.Need = NeedRTCheck;
3790 bool CanVecMem = true;
3791 if (Accesses.isDependencyCheckNeeded()) {
3792 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3793 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3794 Accesses.getDependenciesToCheck());
3795 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3798 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3799 " need a runtime memory check.\n");
3804 static bool hasMultipleUsesOf(Instruction *I,
3805 SmallPtrSet<Instruction *, 8> &Insts) {
3806 unsigned NumUses = 0;
3807 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3808 if (Insts.count(dyn_cast<Instruction>(*Use)))
3817 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3818 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3819 if (!Set.count(dyn_cast<Instruction>(*Use)))
3824 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3825 ReductionKind Kind) {
3826 if (Phi->getNumIncomingValues() != 2)
3829 // Reduction variables are only found in the loop header block.
3830 if (Phi->getParent() != TheLoop->getHeader())
3833 // Obtain the reduction start value from the value that comes from the loop
3835 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3837 // ExitInstruction is the single value which is used outside the loop.
3838 // We only allow for a single reduction value to be used outside the loop.
3839 // This includes users of the reduction, variables (which form a cycle
3840 // which ends in the phi node).
3841 Instruction *ExitInstruction = 0;
3842 // Indicates that we found a reduction operation in our scan.
3843 bool FoundReduxOp = false;
3845 // We start with the PHI node and scan for all of the users of this
3846 // instruction. All users must be instructions that can be used as reduction
3847 // variables (such as ADD). We must have a single out-of-block user. The cycle
3848 // must include the original PHI.
3849 bool FoundStartPHI = false;
3851 // To recognize min/max patterns formed by a icmp select sequence, we store
3852 // the number of instruction we saw from the recognized min/max pattern,
3853 // to make sure we only see exactly the two instructions.
3854 unsigned NumCmpSelectPatternInst = 0;
3855 ReductionInstDesc ReduxDesc(false, 0);
3857 SmallPtrSet<Instruction *, 8> VisitedInsts;
3858 SmallVector<Instruction *, 8> Worklist;
3859 Worklist.push_back(Phi);
3860 VisitedInsts.insert(Phi);
3862 // A value in the reduction can be used:
3863 // - By the reduction:
3864 // - Reduction operation:
3865 // - One use of reduction value (safe).
3866 // - Multiple use of reduction value (not safe).
3868 // - All uses of the PHI must be the reduction (safe).
3869 // - Otherwise, not safe.
3870 // - By one instruction outside of the loop (safe).
3871 // - By further instructions outside of the loop (not safe).
3872 // - By an instruction that is not part of the reduction (not safe).
3874 // * An instruction type other than PHI or the reduction operation.
3875 // * A PHI in the header other than the initial PHI.
3876 while (!Worklist.empty()) {
3877 Instruction *Cur = Worklist.back();
3878 Worklist.pop_back();
3881 // If the instruction has no users then this is a broken chain and can't be
3882 // a reduction variable.
3883 if (Cur->use_empty())
3886 bool IsAPhi = isa<PHINode>(Cur);
3888 // A header PHI use other than the original PHI.
3889 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3892 // Reductions of instructions such as Div, and Sub is only possible if the
3893 // LHS is the reduction variable.
3894 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3895 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3896 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3899 // Any reduction instruction must be of one of the allowed kinds.
3900 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3901 if (!ReduxDesc.IsReduction)
3904 // A reduction operation must only have one use of the reduction value.
3905 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3906 hasMultipleUsesOf(Cur, VisitedInsts))
3909 // All inputs to a PHI node must be a reduction value.
3910 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3913 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3914 isa<SelectInst>(Cur)))
3915 ++NumCmpSelectPatternInst;
3916 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3917 isa<SelectInst>(Cur)))
3918 ++NumCmpSelectPatternInst;
3920 // Check whether we found a reduction operator.
3921 FoundReduxOp |= !IsAPhi;
3923 // Process users of current instruction. Push non PHI nodes after PHI nodes
3924 // onto the stack. This way we are going to have seen all inputs to PHI
3925 // nodes once we get to them.
3926 SmallVector<Instruction *, 8> NonPHIs;
3927 SmallVector<Instruction *, 8> PHIs;
3928 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3930 Instruction *Usr = cast<Instruction>(*UI);
3932 // Check if we found the exit user.
3933 BasicBlock *Parent = Usr->getParent();
3934 if (!TheLoop->contains(Parent)) {
3935 // Exit if you find multiple outside users or if the header phi node is
3936 // being used. In this case the user uses the value of the previous
3937 // iteration, in which case we would loose "VF-1" iterations of the
3938 // reduction operation if we vectorize.
3939 if (ExitInstruction != 0 || Cur == Phi)
3942 ExitInstruction = Cur;
3946 // Process instructions only once (termination).
3947 if (VisitedInsts.insert(Usr)) {
3948 if (isa<PHINode>(Usr))
3949 PHIs.push_back(Usr);
3951 NonPHIs.push_back(Usr);
3953 // Remember that we completed the cycle.
3955 FoundStartPHI = true;
3957 Worklist.append(PHIs.begin(), PHIs.end());
3958 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3961 // This means we have seen one but not the other instruction of the
3962 // pattern or more than just a select and cmp.
3963 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3964 NumCmpSelectPatternInst != 2)
3967 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3970 // We found a reduction var if we have reached the original phi node and we
3971 // only have a single instruction with out-of-loop users.
3973 // This instruction is allowed to have out-of-loop users.
3974 AllowedExit.insert(ExitInstruction);
3976 // Save the description of this reduction variable.
3977 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3978 ReduxDesc.MinMaxKind);
3979 Reductions[Phi] = RD;
3980 // We've ended the cycle. This is a reduction variable if we have an
3981 // outside user and it has a binary op.
3986 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3987 /// pattern corresponding to a min(X, Y) or max(X, Y).
3988 LoopVectorizationLegality::ReductionInstDesc
3989 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3990 ReductionInstDesc &Prev) {
3992 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3993 "Expect a select instruction");
3994 Instruction *Cmp = 0;
3995 SelectInst *Select = 0;
3997 // We must handle the select(cmp()) as a single instruction. Advance to the
3999 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4000 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4001 return ReductionInstDesc(false, I);
4002 return ReductionInstDesc(Select, Prev.MinMaxKind);
4005 // Only handle single use cases for now.
4006 if (!(Select = dyn_cast<SelectInst>(I)))
4007 return ReductionInstDesc(false, I);
4008 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4009 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4010 return ReductionInstDesc(false, I);
4011 if (!Cmp->hasOneUse())
4012 return ReductionInstDesc(false, I);
4017 // Look for a min/max pattern.
4018 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4019 return ReductionInstDesc(Select, MRK_UIntMin);
4020 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4021 return ReductionInstDesc(Select, MRK_UIntMax);
4022 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4023 return ReductionInstDesc(Select, MRK_SIntMax);
4024 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4025 return ReductionInstDesc(Select, MRK_SIntMin);
4026 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4027 return ReductionInstDesc(Select, MRK_FloatMin);
4028 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4029 return ReductionInstDesc(Select, MRK_FloatMax);
4030 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4031 return ReductionInstDesc(Select, MRK_FloatMin);
4032 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4033 return ReductionInstDesc(Select, MRK_FloatMax);
4035 return ReductionInstDesc(false, I);
4038 LoopVectorizationLegality::ReductionInstDesc
4039 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4041 ReductionInstDesc &Prev) {
4042 bool FP = I->getType()->isFloatingPointTy();
4043 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4044 switch (I->getOpcode()) {
4046 return ReductionInstDesc(false, I);
4047 case Instruction::PHI:
4048 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4049 Kind != RK_FloatMinMax))
4050 return ReductionInstDesc(false, I);
4051 return ReductionInstDesc(I, Prev.MinMaxKind);
4052 case Instruction::Sub:
4053 case Instruction::Add:
4054 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4055 case Instruction::Mul:
4056 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4057 case Instruction::And:
4058 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4059 case Instruction::Or:
4060 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4061 case Instruction::Xor:
4062 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4063 case Instruction::FMul:
4064 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4065 case Instruction::FAdd:
4066 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4067 case Instruction::FCmp:
4068 case Instruction::ICmp:
4069 case Instruction::Select:
4070 if (Kind != RK_IntegerMinMax &&
4071 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4072 return ReductionInstDesc(false, I);
4073 return isMinMaxSelectCmpPattern(I, Prev);
4077 LoopVectorizationLegality::InductionKind
4078 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4079 Type *PhiTy = Phi->getType();
4080 // We only handle integer and pointer inductions variables.
4081 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4082 return IK_NoInduction;
4084 // Check that the PHI is consecutive.
4085 const SCEV *PhiScev = SE->getSCEV(Phi);
4086 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4088 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4089 return IK_NoInduction;
4091 const SCEV *Step = AR->getStepRecurrence(*SE);
4093 // Integer inductions need to have a stride of one.
4094 if (PhiTy->isIntegerTy()) {
4096 return IK_IntInduction;
4097 if (Step->isAllOnesValue())
4098 return IK_ReverseIntInduction;
4099 return IK_NoInduction;
4102 // Calculate the pointer stride and check if it is consecutive.
4103 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4105 return IK_NoInduction;
4107 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4108 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4109 if (C->getValue()->equalsInt(Size))
4110 return IK_PtrInduction;
4111 else if (C->getValue()->equalsInt(0 - Size))
4112 return IK_ReversePtrInduction;
4114 return IK_NoInduction;
4117 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4118 Value *In0 = const_cast<Value*>(V);
4119 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4123 return Inductions.count(PN);
4126 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4127 assert(TheLoop->contains(BB) && "Unknown block used");
4129 // Blocks that do not dominate the latch need predication.
4130 BasicBlock* Latch = TheLoop->getLoopLatch();
4131 return !DT->dominates(BB, Latch);
4134 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4135 SmallPtrSet<Value *, 8>& SafePtrs) {
4136 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4137 // We might be able to hoist the load.
4138 if (it->mayReadFromMemory()) {
4139 LoadInst *LI = dyn_cast<LoadInst>(it);
4140 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4144 // We don't predicate stores at the moment.
4145 if (it->mayWriteToMemory() || it->mayThrow())
4148 // The instructions below can trap.
4149 switch (it->getOpcode()) {
4151 case Instruction::UDiv:
4152 case Instruction::SDiv:
4153 case Instruction::URem:
4154 case Instruction::SRem:
4162 LoopVectorizationCostModel::VectorizationFactor
4163 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4165 // Width 1 means no vectorize
4166 VectorizationFactor Factor = { 1U, 0U };
4167 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4168 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4172 // Find the trip count.
4173 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4174 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4176 unsigned WidestType = getWidestType();
4177 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4178 unsigned MaxSafeDepDist = -1U;
4179 if (Legal->getMaxSafeDepDistBytes() != -1U)
4180 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4181 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4182 WidestRegister : MaxSafeDepDist);
4183 unsigned MaxVectorSize = WidestRegister / WidestType;
4184 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4185 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4187 if (MaxVectorSize == 0) {
4188 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4192 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4193 " into one vector!");
4195 unsigned VF = MaxVectorSize;
4197 // If we optimize the program for size, avoid creating the tail loop.
4199 // If we are unable to calculate the trip count then don't try to vectorize.
4201 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4205 // Find the maximum SIMD width that can fit within the trip count.
4206 VF = TC % MaxVectorSize;
4211 // If the trip count that we found modulo the vectorization factor is not
4212 // zero then we require a tail.
4214 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4220 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4221 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4223 Factor.Width = UserVF;
4227 float Cost = expectedCost(1);
4229 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4230 for (unsigned i=2; i <= VF; i*=2) {
4231 // Notice that the vector loop needs to be executed less times, so
4232 // we need to divide the cost of the vector loops by the width of
4233 // the vector elements.
4234 float VectorCost = expectedCost(i) / (float)i;
4235 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4236 (int)VectorCost << ".\n");
4237 if (VectorCost < Cost) {
4243 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4244 Factor.Width = Width;
4245 Factor.Cost = Width * Cost;
4249 unsigned LoopVectorizationCostModel::getWidestType() {
4250 unsigned MaxWidth = 8;
4253 for (Loop::block_iterator bb = TheLoop->block_begin(),
4254 be = TheLoop->block_end(); bb != be; ++bb) {
4255 BasicBlock *BB = *bb;
4257 // For each instruction in the loop.
4258 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4259 Type *T = it->getType();
4261 // Only examine Loads, Stores and PHINodes.
4262 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4265 // Examine PHI nodes that are reduction variables.
4266 if (PHINode *PN = dyn_cast<PHINode>(it))
4267 if (!Legal->getReductionVars()->count(PN))
4270 // Examine the stored values.
4271 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4272 T = ST->getValueOperand()->getType();
4274 // Ignore loaded pointer types and stored pointer types that are not
4275 // consecutive. However, we do want to take consecutive stores/loads of
4276 // pointer vectors into account.
4277 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4280 MaxWidth = std::max(MaxWidth,
4281 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4289 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4292 unsigned LoopCost) {
4294 // -- The unroll heuristics --
4295 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4296 // There are many micro-architectural considerations that we can't predict
4297 // at this level. For example frontend pressure (on decode or fetch) due to
4298 // code size, or the number and capabilities of the execution ports.
4300 // We use the following heuristics to select the unroll factor:
4301 // 1. If the code has reductions the we unroll in order to break the cross
4302 // iteration dependency.
4303 // 2. If the loop is really small then we unroll in order to reduce the loop
4305 // 3. We don't unroll if we think that we will spill registers to memory due
4306 // to the increased register pressure.
4308 // Use the user preference, unless 'auto' is selected.
4312 // When we optimize for size we don't unroll.
4316 // We used the distance for the unroll factor.
4317 if (Legal->getMaxSafeDepDistBytes() != -1U)
4320 // Do not unroll loops with a relatively small trip count.
4321 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4322 TheLoop->getLoopLatch());
4323 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4326 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4327 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4328 " vector registers\n");
4330 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4331 // We divide by these constants so assume that we have at least one
4332 // instruction that uses at least one register.
4333 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4334 R.NumInstructions = std::max(R.NumInstructions, 1U);
4336 // We calculate the unroll factor using the following formula.
4337 // Subtract the number of loop invariants from the number of available
4338 // registers. These registers are used by all of the unrolled instances.
4339 // Next, divide the remaining registers by the number of registers that is
4340 // required by the loop, in order to estimate how many parallel instances
4341 // fit without causing spills.
4342 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4344 // Clamp the unroll factor ranges to reasonable factors.
4345 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4347 // If we did not calculate the cost for VF (because the user selected the VF)
4348 // then we calculate the cost of VF here.
4350 LoopCost = expectedCost(VF);
4352 // Clamp the calculated UF to be between the 1 and the max unroll factor
4353 // that the target allows.
4354 if (UF > MaxUnrollSize)
4359 bool HasReductions = Legal->getReductionVars()->size();
4361 // Decide if we want to unroll if we decided that it is legal to vectorize
4362 // but not profitable.
4364 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4365 LoopCost > SmallLoopCost)
4371 if (HasReductions) {
4372 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4376 // We want to unroll tiny loops in order to reduce the loop overhead.
4377 // We assume that the cost overhead is 1 and we use the cost model
4378 // to estimate the cost of the loop and unroll until the cost of the
4379 // loop overhead is about 5% of the cost of the loop.
4380 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4381 if (LoopCost < SmallLoopCost) {
4382 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4383 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4384 return std::min(NewUF, UF);
4387 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4391 LoopVectorizationCostModel::RegisterUsage
4392 LoopVectorizationCostModel::calculateRegisterUsage() {
4393 // This function calculates the register usage by measuring the highest number
4394 // of values that are alive at a single location. Obviously, this is a very
4395 // rough estimation. We scan the loop in a topological order in order and
4396 // assign a number to each instruction. We use RPO to ensure that defs are
4397 // met before their users. We assume that each instruction that has in-loop
4398 // users starts an interval. We record every time that an in-loop value is
4399 // used, so we have a list of the first and last occurrences of each
4400 // instruction. Next, we transpose this data structure into a multi map that
4401 // holds the list of intervals that *end* at a specific location. This multi
4402 // map allows us to perform a linear search. We scan the instructions linearly
4403 // and record each time that a new interval starts, by placing it in a set.
4404 // If we find this value in the multi-map then we remove it from the set.
4405 // The max register usage is the maximum size of the set.
4406 // We also search for instructions that are defined outside the loop, but are
4407 // used inside the loop. We need this number separately from the max-interval
4408 // usage number because when we unroll, loop-invariant values do not take
4410 LoopBlocksDFS DFS(TheLoop);
4414 R.NumInstructions = 0;
4416 // Each 'key' in the map opens a new interval. The values
4417 // of the map are the index of the 'last seen' usage of the
4418 // instruction that is the key.
4419 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4420 // Maps instruction to its index.
4421 DenseMap<unsigned, Instruction*> IdxToInstr;
4422 // Marks the end of each interval.
4423 IntervalMap EndPoint;
4424 // Saves the list of instruction indices that are used in the loop.
4425 SmallSet<Instruction*, 8> Ends;
4426 // Saves the list of values that are used in the loop but are
4427 // defined outside the loop, such as arguments and constants.
4428 SmallPtrSet<Value*, 8> LoopInvariants;
4431 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4432 be = DFS.endRPO(); bb != be; ++bb) {
4433 R.NumInstructions += (*bb)->size();
4434 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4436 Instruction *I = it;
4437 IdxToInstr[Index++] = I;
4439 // Save the end location of each USE.
4440 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4441 Value *U = I->getOperand(i);
4442 Instruction *Instr = dyn_cast<Instruction>(U);
4444 // Ignore non-instruction values such as arguments, constants, etc.
4445 if (!Instr) continue;
4447 // If this instruction is outside the loop then record it and continue.
4448 if (!TheLoop->contains(Instr)) {
4449 LoopInvariants.insert(Instr);
4453 // Overwrite previous end points.
4454 EndPoint[Instr] = Index;
4460 // Saves the list of intervals that end with the index in 'key'.
4461 typedef SmallVector<Instruction*, 2> InstrList;
4462 DenseMap<unsigned, InstrList> TransposeEnds;
4464 // Transpose the EndPoints to a list of values that end at each index.
4465 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4467 TransposeEnds[it->second].push_back(it->first);
4469 SmallSet<Instruction*, 8> OpenIntervals;
4470 unsigned MaxUsage = 0;
4473 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4474 for (unsigned int i = 0; i < Index; ++i) {
4475 Instruction *I = IdxToInstr[i];
4476 // Ignore instructions that are never used within the loop.
4477 if (!Ends.count(I)) continue;
4479 // Remove all of the instructions that end at this location.
4480 InstrList &List = TransposeEnds[i];
4481 for (unsigned int j=0, e = List.size(); j < e; ++j)
4482 OpenIntervals.erase(List[j]);
4484 // Count the number of live interals.
4485 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4487 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4488 OpenIntervals.size() <<"\n");
4490 // Add the current instruction to the list of open intervals.
4491 OpenIntervals.insert(I);
4494 unsigned Invariant = LoopInvariants.size();
4495 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4496 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4497 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4499 R.LoopInvariantRegs = Invariant;
4500 R.MaxLocalUsers = MaxUsage;
4504 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4508 for (Loop::block_iterator bb = TheLoop->block_begin(),
4509 be = TheLoop->block_end(); bb != be; ++bb) {
4510 unsigned BlockCost = 0;
4511 BasicBlock *BB = *bb;
4513 // For each instruction in the old loop.
4514 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4515 // Skip dbg intrinsics.
4516 if (isa<DbgInfoIntrinsic>(it))
4519 unsigned C = getInstructionCost(it, VF);
4521 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4522 VF << " For instruction: "<< *it << "\n");
4525 // We assume that if-converted blocks have a 50% chance of being executed.
4526 // When the code is scalar then some of the blocks are avoided due to CF.
4527 // When the code is vectorized we execute all code paths.
4528 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4537 /// \brief Check whether the address computation for a non-consecutive memory
4538 /// access looks like an unlikely candidate for being merged into the indexing
4541 /// We look for a GEP which has one index that is an induction variable and all
4542 /// other indices are loop invariant. If the stride of this access is also
4543 /// within a small bound we decide that this address computation can likely be
4544 /// merged into the addressing mode.
4545 /// In all other cases, we identify the address computation as complex.
4546 static bool isLikelyComplexAddressComputation(Value *Ptr,
4547 LoopVectorizationLegality *Legal,
4548 ScalarEvolution *SE,
4549 const Loop *TheLoop) {
4550 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4554 // We are looking for a gep with all loop invariant indices except for one
4555 // which should be an induction variable.
4556 unsigned NumOperands = Gep->getNumOperands();
4557 for (unsigned i = 1; i < NumOperands; ++i) {
4558 Value *Opd = Gep->getOperand(i);
4559 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4560 !Legal->isInductionVariable(Opd))
4564 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4565 // can likely be merged into the address computation.
4566 unsigned MaxMergeDistance = 64;
4568 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4572 // Check the step is constant.
4573 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4574 // Calculate the pointer stride and check if it is consecutive.
4575 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4579 const APInt &APStepVal = C->getValue()->getValue();
4581 // Huge step value - give up.
4582 if (APStepVal.getBitWidth() > 64)
4585 int64_t StepVal = APStepVal.getSExtValue();
4587 return StepVal > MaxMergeDistance;
4591 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4592 // If we know that this instruction will remain uniform, check the cost of
4593 // the scalar version.
4594 if (Legal->isUniformAfterVectorization(I))
4597 Type *RetTy = I->getType();
4598 Type *VectorTy = ToVectorTy(RetTy, VF);
4600 // TODO: We need to estimate the cost of intrinsic calls.
4601 switch (I->getOpcode()) {
4602 case Instruction::GetElementPtr:
4603 // We mark this instruction as zero-cost because the cost of GEPs in
4604 // vectorized code depends on whether the corresponding memory instruction
4605 // is scalarized or not. Therefore, we handle GEPs with the memory
4606 // instruction cost.
4608 case Instruction::Br: {
4609 return TTI.getCFInstrCost(I->getOpcode());
4611 case Instruction::PHI:
4612 //TODO: IF-converted IFs become selects.
4614 case Instruction::Add:
4615 case Instruction::FAdd:
4616 case Instruction::Sub:
4617 case Instruction::FSub:
4618 case Instruction::Mul:
4619 case Instruction::FMul:
4620 case Instruction::UDiv:
4621 case Instruction::SDiv:
4622 case Instruction::FDiv:
4623 case Instruction::URem:
4624 case Instruction::SRem:
4625 case Instruction::FRem:
4626 case Instruction::Shl:
4627 case Instruction::LShr:
4628 case Instruction::AShr:
4629 case Instruction::And:
4630 case Instruction::Or:
4631 case Instruction::Xor: {
4632 // Certain instructions can be cheaper to vectorize if they have a constant
4633 // second vector operand. One example of this are shifts on x86.
4634 TargetTransformInfo::OperandValueKind Op1VK =
4635 TargetTransformInfo::OK_AnyValue;
4636 TargetTransformInfo::OperandValueKind Op2VK =
4637 TargetTransformInfo::OK_AnyValue;
4639 if (isa<ConstantInt>(I->getOperand(1)))
4640 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4642 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4644 case Instruction::Select: {
4645 SelectInst *SI = cast<SelectInst>(I);
4646 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4647 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4648 Type *CondTy = SI->getCondition()->getType();
4650 CondTy = VectorType::get(CondTy, VF);
4652 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4654 case Instruction::ICmp:
4655 case Instruction::FCmp: {
4656 Type *ValTy = I->getOperand(0)->getType();
4657 VectorTy = ToVectorTy(ValTy, VF);
4658 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4660 case Instruction::Store:
4661 case Instruction::Load: {
4662 StoreInst *SI = dyn_cast<StoreInst>(I);
4663 LoadInst *LI = dyn_cast<LoadInst>(I);
4664 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4666 VectorTy = ToVectorTy(ValTy, VF);
4668 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4669 unsigned AS = SI ? SI->getPointerAddressSpace() :
4670 LI->getPointerAddressSpace();
4671 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4672 // We add the cost of address computation here instead of with the gep
4673 // instruction because only here we know whether the operation is
4676 return TTI.getAddressComputationCost(VectorTy) +
4677 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4679 // Scalarized loads/stores.
4680 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4681 bool Reverse = ConsecutiveStride < 0;
4682 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4683 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4684 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4685 bool IsComplexComputation =
4686 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4688 // The cost of extracting from the value vector and pointer vector.
4689 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4690 for (unsigned i = 0; i < VF; ++i) {
4691 // The cost of extracting the pointer operand.
4692 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4693 // In case of STORE, the cost of ExtractElement from the vector.
4694 // In case of LOAD, the cost of InsertElement into the returned
4696 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4697 Instruction::InsertElement,
4701 // The cost of the scalar loads/stores.
4702 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4703 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4708 // Wide load/stores.
4709 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4710 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4713 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4717 case Instruction::ZExt:
4718 case Instruction::SExt:
4719 case Instruction::FPToUI:
4720 case Instruction::FPToSI:
4721 case Instruction::FPExt:
4722 case Instruction::PtrToInt:
4723 case Instruction::IntToPtr:
4724 case Instruction::SIToFP:
4725 case Instruction::UIToFP:
4726 case Instruction::Trunc:
4727 case Instruction::FPTrunc:
4728 case Instruction::BitCast: {
4729 // We optimize the truncation of induction variable.
4730 // The cost of these is the same as the scalar operation.
4731 if (I->getOpcode() == Instruction::Trunc &&
4732 Legal->isInductionVariable(I->getOperand(0)))
4733 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4734 I->getOperand(0)->getType());
4736 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4737 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4739 case Instruction::Call: {
4740 CallInst *CI = cast<CallInst>(I);
4741 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4742 assert(ID && "Not an intrinsic call!");
4743 Type *RetTy = ToVectorTy(CI->getType(), VF);
4744 SmallVector<Type*, 4> Tys;
4745 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4746 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4747 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4750 // We are scalarizing the instruction. Return the cost of the scalar
4751 // instruction, plus the cost of insert and extract into vector
4752 // elements, times the vector width.
4755 if (!RetTy->isVoidTy() && VF != 1) {
4756 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4758 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4761 // The cost of inserting the results plus extracting each one of the
4763 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4766 // The cost of executing VF copies of the scalar instruction. This opcode
4767 // is unknown. Assume that it is the same as 'mul'.
4768 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4774 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4775 if (Scalar->isVoidTy() || VF == 1)
4777 return VectorType::get(Scalar, VF);
4780 char LoopVectorize::ID = 0;
4781 static const char lv_name[] = "Loop Vectorization";
4782 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4783 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4784 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4785 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4786 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4789 Pass *createLoopVectorizePass() {
4790 return new LoopVectorize();
4794 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4795 // Check for a store.
4796 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4797 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4799 // Check for a load.
4800 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4801 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
4807 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
4808 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
4809 // Holds vector parameters or scalars, in case of uniform vals.
4810 SmallVector<VectorParts, 4> Params;
4812 setDebugLocFromInst(Builder, Instr);
4814 // Find all of the vectorized parameters.
4815 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
4816 Value *SrcOp = Instr->getOperand(op);
4818 // If we are accessing the old induction variable, use the new one.
4819 if (SrcOp == OldInduction) {
4820 Params.push_back(getVectorValue(SrcOp));
4824 // Try using previously calculated values.
4825 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
4827 // If the src is an instruction that appeared earlier in the basic block
4828 // then it should already be vectorized.
4829 if (SrcInst && OrigLoop->contains(SrcInst)) {
4830 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
4831 // The parameter is a vector value from earlier.
4832 Params.push_back(WidenMap.get(SrcInst));
4834 // The parameter is a scalar from outside the loop. Maybe even a constant.
4835 VectorParts Scalars;
4836 Scalars.append(UF, SrcOp);
4837 Params.push_back(Scalars);
4841 assert(Params.size() == Instr->getNumOperands() &&
4842 "Invalid number of operands");
4844 // Does this instruction return a value ?
4845 bool IsVoidRetTy = Instr->getType()->isVoidTy();
4847 Value *UndefVec = IsVoidRetTy ? 0 :
4848 UndefValue::get(Instr->getType());
4849 // Create a new entry in the WidenMap and initialize it to Undef or Null.
4850 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
4852 // For each vector unroll 'part':
4853 for (unsigned Part = 0; Part < UF; ++Part) {
4854 // For each scalar that we create:
4856 Instruction *Cloned = Instr->clone();
4858 Cloned->setName(Instr->getName() + ".cloned");
4859 // Replace the operands of the cloned instrucions with extracted scalars.
4860 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
4861 Value *Op = Params[op][Part];
4862 Cloned->setOperand(op, Op);
4865 // Place the cloned scalar in the new loop.
4866 Builder.Insert(Cloned);
4868 // If the original scalar returns a value we need to place it in a vector
4869 // so that future users will be able to use it.
4871 VecResults[Part] = Cloned;
4876 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
4877 LoopVectorizationLegality*) {
4878 return scalarizeInstruction(Instr);
4881 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
4885 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
4889 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
4891 // When unrolling and the VF is 1, we only need to add a simple scalar.
4892 Type *ITy = Val->getType();
4893 assert(!ITy->isVectorTy() && "Val must be a scalar");
4894 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
4895 return Builder.CreateAdd(Val, C, "induction");