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, bool DisableUnrolling)
765 : Width(VectorizationFactor)
766 , Unroll(DisableUnrolling ? 1 : 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;
776 DEBUG(if (DisableUnrolling && Unroll == 1)
777 dbgs() << "LV: Unrolling disabled by the pass manager\n");
780 /// Return the loop vectorizer metadata prefix.
781 static StringRef Prefix() { return "llvm.vectorizer."; }
783 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
784 SmallVector<Value*, 2> Vals;
785 Vals.push_back(MDString::get(Context, Name));
786 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
787 return MDNode::get(Context, Vals);
790 /// Mark the loop L as already vectorized by setting the width to 1.
791 void setAlreadyVectorized(Loop *L) {
792 LLVMContext &Context = L->getHeader()->getContext();
796 // Create a new loop id with one more operand for the already_vectorized
797 // hint. If the loop already has a loop id then copy the existing operands.
798 SmallVector<Value*, 4> Vals(1);
800 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
801 Vals.push_back(LoopID->getOperand(i));
803 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
805 MDNode *NewLoopID = MDNode::get(Context, Vals);
806 // Set operand 0 to refer to the loop id itself.
807 NewLoopID->replaceOperandWith(0, NewLoopID);
809 L->setLoopID(NewLoopID);
811 LoopID->replaceAllUsesWith(NewLoopID);
819 /// Find hints specified in the loop metadata.
820 void getHints(const Loop *L) {
824 // First operand should refer to the loop id itself.
825 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
826 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
828 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
829 const MDString *S = 0;
830 SmallVector<Value*, 4> Args;
832 // The expected hint is either a MDString or a MDNode with the first
833 // operand a MDString.
834 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
835 if (!MD || MD->getNumOperands() == 0)
837 S = dyn_cast<MDString>(MD->getOperand(0));
838 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
839 Args.push_back(MD->getOperand(i));
841 S = dyn_cast<MDString>(LoopID->getOperand(i));
842 assert(Args.size() == 0 && "too many arguments for MDString");
848 // Check if the hint starts with the vectorizer prefix.
849 StringRef Hint = S->getString();
850 if (!Hint.startswith(Prefix()))
852 // Remove the prefix.
853 Hint = Hint.substr(Prefix().size(), StringRef::npos);
855 if (Args.size() == 1)
856 getHint(Hint, Args[0]);
860 // Check string hint with one operand.
861 void getHint(StringRef Hint, Value *Arg) {
862 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
864 unsigned Val = C->getZExtValue();
866 if (Hint == "width") {
867 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
870 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata");
871 } else if (Hint == "unroll") {
872 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
875 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata");
877 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
882 /// The LoopVectorize Pass.
883 struct LoopVectorize : public LoopPass {
884 /// Pass identification, replacement for typeid
887 explicit LoopVectorize(bool NoUnrolling = false)
888 : LoopPass(ID), DisableUnrolling(NoUnrolling) {
889 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
895 TargetTransformInfo *TTI;
897 TargetLibraryInfo *TLI;
898 bool DisableUnrolling;
900 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
901 // We only vectorize innermost loops.
905 SE = &getAnalysis<ScalarEvolution>();
906 DL = getAnalysisIfAvailable<DataLayout>();
907 LI = &getAnalysis<LoopInfo>();
908 TTI = &getAnalysis<TargetTransformInfo>();
909 DT = &getAnalysis<DominatorTree>();
910 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
912 // If the target claims to have no vector registers don't attempt
914 if (!TTI->getNumberOfRegisters(true))
918 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
922 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
923 L->getHeader()->getParent()->getName() << "\"\n");
925 LoopVectorizeHints Hints(L, DisableUnrolling);
927 if (Hints.Width == 1 && Hints.Unroll == 1) {
928 DEBUG(dbgs() << "LV: Not vectorizing.\n");
932 // Check if it is legal to vectorize the loop.
933 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
934 if (!LVL.canVectorize()) {
935 DEBUG(dbgs() << "LV: Not vectorizing.\n");
939 // Use the cost model.
940 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
942 // Check the function attributes to find out if this function should be
943 // optimized for size.
944 Function *F = L->getHeader()->getParent();
945 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
946 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
947 unsigned FnIndex = AttributeSet::FunctionIndex;
948 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
949 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
952 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
953 "attribute is used.\n");
957 // Select the optimal vectorization factor.
958 LoopVectorizationCostModel::VectorizationFactor VF;
959 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
960 // Select the unroll factor.
961 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
965 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
968 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
969 F->getParent()->getModuleIdentifier()<<"\n");
970 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
975 // We decided not to vectorize, but we may want to unroll.
976 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
977 Unroller.vectorize(&LVL);
979 // If we decided that it is *legal* to vectorize the loop then do it.
980 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
984 // Mark the loop as already vectorized to avoid vectorizing again.
985 Hints.setAlreadyVectorized(L);
987 DEBUG(verifyFunction(*L->getHeader()->getParent()));
991 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
992 LoopPass::getAnalysisUsage(AU);
993 AU.addRequiredID(LoopSimplifyID);
994 AU.addRequiredID(LCSSAID);
995 AU.addRequired<DominatorTree>();
996 AU.addRequired<LoopInfo>();
997 AU.addRequired<ScalarEvolution>();
998 AU.addRequired<TargetTransformInfo>();
999 AU.addPreserved<LoopInfo>();
1000 AU.addPreserved<DominatorTree>();
1005 } // end anonymous namespace
1007 //===----------------------------------------------------------------------===//
1008 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1009 // LoopVectorizationCostModel.
1010 //===----------------------------------------------------------------------===//
1013 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1014 Loop *Lp, Value *Ptr,
1016 unsigned DepSetId) {
1017 const SCEV *Sc = SE->getSCEV(Ptr);
1018 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1019 assert(AR && "Invalid addrec expression");
1020 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1021 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1022 Pointers.push_back(Ptr);
1023 Starts.push_back(AR->getStart());
1024 Ends.push_back(ScEnd);
1025 IsWritePtr.push_back(WritePtr);
1026 DependencySetId.push_back(DepSetId);
1029 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1030 // We need to place the broadcast of invariant variables outside the loop.
1031 Instruction *Instr = dyn_cast<Instruction>(V);
1032 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1033 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1035 // Place the code for broadcasting invariant variables in the new preheader.
1036 IRBuilder<>::InsertPointGuard Guard(Builder);
1038 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1040 // Broadcast the scalar into all locations in the vector.
1041 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1046 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1048 assert(Val->getType()->isVectorTy() && "Must be a vector");
1049 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1050 "Elem must be an integer");
1051 // Create the types.
1052 Type *ITy = Val->getType()->getScalarType();
1053 VectorType *Ty = cast<VectorType>(Val->getType());
1054 int VLen = Ty->getNumElements();
1055 SmallVector<Constant*, 8> Indices;
1057 // Create a vector of consecutive numbers from zero to VF.
1058 for (int i = 0; i < VLen; ++i) {
1059 int64_t Idx = Negate ? (-i) : i;
1060 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1063 // Add the consecutive indices to the vector value.
1064 Constant *Cv = ConstantVector::get(Indices);
1065 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1066 return Builder.CreateAdd(Val, Cv, "induction");
1069 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1070 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1071 // Make sure that the pointer does not point to structs.
1072 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1075 // If this value is a pointer induction variable we know it is consecutive.
1076 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1077 if (Phi && Inductions.count(Phi)) {
1078 InductionInfo II = Inductions[Phi];
1079 if (IK_PtrInduction == II.IK)
1081 else if (IK_ReversePtrInduction == II.IK)
1085 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1089 unsigned NumOperands = Gep->getNumOperands();
1090 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1092 Value *GpPtr = Gep->getPointerOperand();
1093 // If this GEP value is a consecutive pointer induction variable and all of
1094 // the indices are constant then we know it is consecutive. We can
1095 Phi = dyn_cast<PHINode>(GpPtr);
1096 if (Phi && Inductions.count(Phi)) {
1098 // Make sure that the pointer does not point to structs.
1099 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1100 if (GepPtrType->getElementType()->isAggregateType())
1103 // Make sure that all of the index operands are loop invariant.
1104 for (unsigned i = 1; i < NumOperands; ++i)
1105 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1108 InductionInfo II = Inductions[Phi];
1109 if (IK_PtrInduction == II.IK)
1111 else if (IK_ReversePtrInduction == II.IK)
1115 // Check that all of the gep indices are uniform except for the last.
1116 for (unsigned i = 0; i < NumOperands - 1; ++i)
1117 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1120 // We can emit wide load/stores only if the last index is the induction
1122 const SCEV *Last = SE->getSCEV(LastIndex);
1123 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1124 const SCEV *Step = AR->getStepRecurrence(*SE);
1126 // The memory is consecutive because the last index is consecutive
1127 // and all other indices are loop invariant.
1130 if (Step->isAllOnesValue())
1137 bool LoopVectorizationLegality::isUniform(Value *V) {
1138 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1141 InnerLoopVectorizer::VectorParts&
1142 InnerLoopVectorizer::getVectorValue(Value *V) {
1143 assert(V != Induction && "The new induction variable should not be used.");
1144 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1146 // If we have this scalar in the map, return it.
1147 if (WidenMap.has(V))
1148 return WidenMap.get(V);
1150 // If this scalar is unknown, assume that it is a constant or that it is
1151 // loop invariant. Broadcast V and save the value for future uses.
1152 Value *B = getBroadcastInstrs(V);
1153 return WidenMap.splat(V, B);
1156 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1157 assert(Vec->getType()->isVectorTy() && "Invalid type");
1158 SmallVector<Constant*, 8> ShuffleMask;
1159 for (unsigned i = 0; i < VF; ++i)
1160 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1162 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1163 ConstantVector::get(ShuffleMask),
1168 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1169 LoopVectorizationLegality *Legal) {
1170 // Attempt to issue a wide load.
1171 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1172 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1174 assert((LI || SI) && "Invalid Load/Store instruction");
1176 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1177 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1178 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1179 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1180 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1181 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1182 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1184 if (ScalarAllocatedSize != VectorElementSize)
1185 return scalarizeInstruction(Instr);
1187 // If the pointer is loop invariant or if it is non consecutive,
1188 // scalarize the load.
1189 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1190 bool Reverse = ConsecutiveStride < 0;
1191 bool UniformLoad = LI && Legal->isUniform(Ptr);
1192 if (!ConsecutiveStride || UniformLoad)
1193 return scalarizeInstruction(Instr);
1195 Constant *Zero = Builder.getInt32(0);
1196 VectorParts &Entry = WidenMap.get(Instr);
1198 // Handle consecutive loads/stores.
1199 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1200 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1201 setDebugLocFromInst(Builder, Gep);
1202 Value *PtrOperand = Gep->getPointerOperand();
1203 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1204 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1206 // Create the new GEP with the new induction variable.
1207 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1208 Gep2->setOperand(0, FirstBasePtr);
1209 Gep2->setName("gep.indvar.base");
1210 Ptr = Builder.Insert(Gep2);
1212 setDebugLocFromInst(Builder, Gep);
1213 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1214 OrigLoop) && "Base ptr must be invariant");
1216 // The last index does not have to be the induction. It can be
1217 // consecutive and be a function of the index. For example A[I+1];
1218 unsigned NumOperands = Gep->getNumOperands();
1219 unsigned LastOperand = NumOperands - 1;
1220 // Create the new GEP with the new induction variable.
1221 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1223 for (unsigned i = 0; i < NumOperands; ++i) {
1224 Value *GepOperand = Gep->getOperand(i);
1225 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1227 // Update last index or loop invariant instruction anchored in loop.
1228 if (i == LastOperand ||
1229 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1230 assert((i == LastOperand ||
1231 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1232 "Must be last index or loop invariant");
1234 VectorParts &GEPParts = getVectorValue(GepOperand);
1235 Value *Index = GEPParts[0];
1236 Index = Builder.CreateExtractElement(Index, Zero);
1237 Gep2->setOperand(i, Index);
1238 Gep2->setName("gep.indvar.idx");
1241 Ptr = Builder.Insert(Gep2);
1243 // Use the induction element ptr.
1244 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1245 setDebugLocFromInst(Builder, Ptr);
1246 VectorParts &PtrVal = getVectorValue(Ptr);
1247 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1252 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1253 "We do not allow storing to uniform addresses");
1254 setDebugLocFromInst(Builder, SI);
1255 // We don't want to update the value in the map as it might be used in
1256 // another expression. So don't use a reference type for "StoredVal".
1257 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1259 for (unsigned Part = 0; Part < UF; ++Part) {
1260 // Calculate the pointer for the specific unroll-part.
1261 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1264 // If we store to reverse consecutive memory locations then we need
1265 // to reverse the order of elements in the stored value.
1266 StoredVal[Part] = reverseVector(StoredVal[Part]);
1267 // If the address is consecutive but reversed, then the
1268 // wide store needs to start at the last vector element.
1269 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1270 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1273 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1274 DataTy->getPointerTo(AddressSpace));
1275 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1281 assert(LI && "Must have a load instruction");
1282 setDebugLocFromInst(Builder, LI);
1283 for (unsigned Part = 0; Part < UF; ++Part) {
1284 // Calculate the pointer for the specific unroll-part.
1285 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1288 // If the address is consecutive but reversed, then the
1289 // wide store needs to start at the last vector element.
1290 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1291 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1294 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1295 DataTy->getPointerTo(AddressSpace));
1296 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1297 cast<LoadInst>(LI)->setAlignment(Alignment);
1298 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1302 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1303 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1304 // Holds vector parameters or scalars, in case of uniform vals.
1305 SmallVector<VectorParts, 4> Params;
1307 setDebugLocFromInst(Builder, Instr);
1309 // Find all of the vectorized parameters.
1310 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1311 Value *SrcOp = Instr->getOperand(op);
1313 // If we are accessing the old induction variable, use the new one.
1314 if (SrcOp == OldInduction) {
1315 Params.push_back(getVectorValue(SrcOp));
1319 // Try using previously calculated values.
1320 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1322 // If the src is an instruction that appeared earlier in the basic block
1323 // then it should already be vectorized.
1324 if (SrcInst && OrigLoop->contains(SrcInst)) {
1325 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1326 // The parameter is a vector value from earlier.
1327 Params.push_back(WidenMap.get(SrcInst));
1329 // The parameter is a scalar from outside the loop. Maybe even a constant.
1330 VectorParts Scalars;
1331 Scalars.append(UF, SrcOp);
1332 Params.push_back(Scalars);
1336 assert(Params.size() == Instr->getNumOperands() &&
1337 "Invalid number of operands");
1339 // Does this instruction return a value ?
1340 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1342 Value *UndefVec = IsVoidRetTy ? 0 :
1343 UndefValue::get(VectorType::get(Instr->getType(), VF));
1344 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1345 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1347 // For each vector unroll 'part':
1348 for (unsigned Part = 0; Part < UF; ++Part) {
1349 // For each scalar that we create:
1350 for (unsigned Width = 0; Width < VF; ++Width) {
1351 Instruction *Cloned = Instr->clone();
1353 Cloned->setName(Instr->getName() + ".cloned");
1354 // Replace the operands of the cloned instructions with extracted scalars.
1355 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1356 Value *Op = Params[op][Part];
1357 // Param is a vector. Need to extract the right lane.
1358 if (Op->getType()->isVectorTy())
1359 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1360 Cloned->setOperand(op, Op);
1363 // Place the cloned scalar in the new loop.
1364 Builder.Insert(Cloned);
1366 // If the original scalar returns a value we need to place it in a vector
1367 // so that future users will be able to use it.
1369 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1370 Builder.getInt32(Width));
1376 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1378 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1379 Legal->getRuntimePointerCheck();
1381 if (!PtrRtCheck->Need)
1384 unsigned NumPointers = PtrRtCheck->Pointers.size();
1385 SmallVector<TrackingVH<Value> , 2> Starts;
1386 SmallVector<TrackingVH<Value> , 2> Ends;
1388 SCEVExpander Exp(*SE, "induction");
1390 // Use this type for pointer arithmetic.
1391 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1393 for (unsigned i = 0; i < NumPointers; ++i) {
1394 Value *Ptr = PtrRtCheck->Pointers[i];
1395 const SCEV *Sc = SE->getSCEV(Ptr);
1397 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1398 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1400 Starts.push_back(Ptr);
1401 Ends.push_back(Ptr);
1403 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1405 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1406 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1407 Starts.push_back(Start);
1408 Ends.push_back(End);
1412 IRBuilder<> ChkBuilder(Loc);
1413 // Our instructions might fold to a constant.
1414 Value *MemoryRuntimeCheck = 0;
1415 for (unsigned i = 0; i < NumPointers; ++i) {
1416 for (unsigned j = i+1; j < NumPointers; ++j) {
1417 // No need to check if two readonly pointers intersect.
1418 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1421 // Only need to check pointers between two different dependency sets.
1422 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1425 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1426 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1427 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1428 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1430 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1431 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1432 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1433 if (MemoryRuntimeCheck)
1434 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1436 MemoryRuntimeCheck = IsConflict;
1440 // We have to do this trickery because the IRBuilder might fold the check to a
1441 // constant expression in which case there is no Instruction anchored in a
1443 LLVMContext &Ctx = Loc->getContext();
1444 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1445 ConstantInt::getTrue(Ctx));
1446 ChkBuilder.Insert(Check, "memcheck.conflict");
1451 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1453 In this function we generate a new loop. The new loop will contain
1454 the vectorized instructions while the old loop will continue to run the
1457 [ ] <-- vector loop bypass (may consist of multiple blocks).
1460 | [ ] <-- vector pre header.
1464 | [ ]_| <-- vector loop.
1467 >[ ] <--- middle-block.
1470 | [ ] <--- new preheader.
1474 | [ ]_| <-- old scalar loop to handle remainder.
1477 >[ ] <-- exit block.
1481 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1482 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1483 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1484 assert(ExitBlock && "Must have an exit block");
1486 // Some loops have a single integer induction variable, while other loops
1487 // don't. One example is c++ iterators that often have multiple pointer
1488 // induction variables. In the code below we also support a case where we
1489 // don't have a single induction variable.
1490 OldInduction = Legal->getInduction();
1491 Type *IdxTy = Legal->getWidestInductionType();
1493 // Find the loop boundaries.
1494 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1495 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1497 // Get the total trip count from the count by adding 1.
1498 ExitCount = SE->getAddExpr(ExitCount,
1499 SE->getConstant(ExitCount->getType(), 1));
1501 // Expand the trip count and place the new instructions in the preheader.
1502 // Notice that the pre-header does not change, only the loop body.
1503 SCEVExpander Exp(*SE, "induction");
1505 // Count holds the overall loop count (N).
1506 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1507 BypassBlock->getTerminator());
1509 // The loop index does not have to start at Zero. Find the original start
1510 // value from the induction PHI node. If we don't have an induction variable
1511 // then we know that it starts at zero.
1512 Builder.SetInsertPoint(BypassBlock->getTerminator());
1513 Value *StartIdx = ExtendedIdx = OldInduction ?
1514 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1516 ConstantInt::get(IdxTy, 0);
1518 assert(BypassBlock && "Invalid loop structure");
1519 LoopBypassBlocks.push_back(BypassBlock);
1521 // Split the single block loop into the two loop structure described above.
1522 BasicBlock *VectorPH =
1523 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1524 BasicBlock *VecBody =
1525 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1526 BasicBlock *MiddleBlock =
1527 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1528 BasicBlock *ScalarPH =
1529 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1531 // Create and register the new vector loop.
1532 Loop* Lp = new Loop();
1533 Loop *ParentLoop = OrigLoop->getParentLoop();
1535 // Insert the new loop into the loop nest and register the new basic blocks
1536 // before calling any utilities such as SCEV that require valid LoopInfo.
1538 ParentLoop->addChildLoop(Lp);
1539 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1540 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1541 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1543 LI->addTopLevelLoop(Lp);
1545 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1547 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1549 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1551 // Generate the induction variable.
1552 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1553 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1554 // The loop step is equal to the vectorization factor (num of SIMD elements)
1555 // times the unroll factor (num of SIMD instructions).
1556 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1558 // This is the IR builder that we use to add all of the logic for bypassing
1559 // the new vector loop.
1560 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1561 setDebugLocFromInst(BypassBuilder,
1562 getDebugLocFromInstOrOperands(OldInduction));
1564 // We may need to extend the index in case there is a type mismatch.
1565 // We know that the count starts at zero and does not overflow.
1566 if (Count->getType() != IdxTy) {
1567 // The exit count can be of pointer type. Convert it to the correct
1569 if (ExitCount->getType()->isPointerTy())
1570 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1572 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1575 // Add the start index to the loop count to get the new end index.
1576 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1578 // Now we need to generate the expression for N - (N % VF), which is
1579 // the part that the vectorized body will execute.
1580 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1581 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1582 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1583 "end.idx.rnd.down");
1585 // Now, compare the new count to zero. If it is zero skip the vector loop and
1586 // jump to the scalar loop.
1587 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1590 BasicBlock *LastBypassBlock = BypassBlock;
1592 // Generate the code that checks in runtime if arrays overlap. We put the
1593 // checks into a separate block to make the more common case of few elements
1595 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1596 BypassBlock->getTerminator());
1597 if (MemRuntimeCheck) {
1598 // Create a new block containing the memory check.
1599 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1602 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1603 LoopBypassBlocks.push_back(CheckBlock);
1605 // Replace the branch into the memory check block with a conditional branch
1606 // for the "few elements case".
1607 Instruction *OldTerm = BypassBlock->getTerminator();
1608 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1609 OldTerm->eraseFromParent();
1611 Cmp = MemRuntimeCheck;
1612 LastBypassBlock = CheckBlock;
1615 LastBypassBlock->getTerminator()->eraseFromParent();
1616 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1619 // We are going to resume the execution of the scalar loop.
1620 // Go over all of the induction variables that we found and fix the
1621 // PHIs that are left in the scalar version of the loop.
1622 // The starting values of PHI nodes depend on the counter of the last
1623 // iteration in the vectorized loop.
1624 // If we come from a bypass edge then we need to start from the original
1627 // This variable saves the new starting index for the scalar loop.
1628 PHINode *ResumeIndex = 0;
1629 LoopVectorizationLegality::InductionList::iterator I, E;
1630 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1631 // Set builder to point to last bypass block.
1632 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1633 for (I = List->begin(), E = List->end(); I != E; ++I) {
1634 PHINode *OrigPhi = I->first;
1635 LoopVectorizationLegality::InductionInfo II = I->second;
1637 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1638 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1639 MiddleBlock->getTerminator());
1640 // We might have extended the type of the induction variable but we need a
1641 // truncated version for the scalar loop.
1642 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1643 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1644 MiddleBlock->getTerminator()) : 0;
1646 Value *EndValue = 0;
1648 case LoopVectorizationLegality::IK_NoInduction:
1649 llvm_unreachable("Unknown induction");
1650 case LoopVectorizationLegality::IK_IntInduction: {
1651 // Handle the integer induction counter.
1652 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1654 // We have the canonical induction variable.
1655 if (OrigPhi == OldInduction) {
1656 // Create a truncated version of the resume value for the scalar loop,
1657 // we might have promoted the type to a larger width.
1659 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1660 // The new PHI merges the original incoming value, in case of a bypass,
1661 // or the value at the end of the vectorized loop.
1662 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1663 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1664 TruncResumeVal->addIncoming(EndValue, VecBody);
1666 // We know what the end value is.
1667 EndValue = IdxEndRoundDown;
1668 // We also know which PHI node holds it.
1669 ResumeIndex = ResumeVal;
1673 // Not the canonical induction variable - add the vector loop count to the
1675 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1676 II.StartValue->getType(),
1678 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1681 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1682 // Convert the CountRoundDown variable to the PHI size.
1683 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1684 II.StartValue->getType(),
1686 // Handle reverse integer induction counter.
1687 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1690 case LoopVectorizationLegality::IK_PtrInduction: {
1691 // For pointer induction variables, calculate the offset using
1693 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1697 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1698 // The value at the end of the loop for the reverse pointer is calculated
1699 // by creating a GEP with a negative index starting from the start value.
1700 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1701 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1703 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1709 // The new PHI merges the original incoming value, in case of a bypass,
1710 // or the value at the end of the vectorized loop.
1711 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1712 if (OrigPhi == OldInduction)
1713 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1715 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1717 ResumeVal->addIncoming(EndValue, VecBody);
1719 // Fix the scalar body counter (PHI node).
1720 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1721 // The old inductions phi node in the scalar body needs the truncated value.
1722 if (OrigPhi == OldInduction)
1723 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1725 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1728 // If we are generating a new induction variable then we also need to
1729 // generate the code that calculates the exit value. This value is not
1730 // simply the end of the counter because we may skip the vectorized body
1731 // in case of a runtime check.
1733 assert(!ResumeIndex && "Unexpected resume value found");
1734 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1735 MiddleBlock->getTerminator());
1736 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1737 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1738 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1741 // Make sure that we found the index where scalar loop needs to continue.
1742 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1743 "Invalid resume Index");
1745 // Add a check in the middle block to see if we have completed
1746 // all of the iterations in the first vector loop.
1747 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1748 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1749 ResumeIndex, "cmp.n",
1750 MiddleBlock->getTerminator());
1752 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1753 // Remove the old terminator.
1754 MiddleBlock->getTerminator()->eraseFromParent();
1756 // Create i+1 and fill the PHINode.
1757 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1758 Induction->addIncoming(StartIdx, VectorPH);
1759 Induction->addIncoming(NextIdx, VecBody);
1760 // Create the compare.
1761 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1762 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1764 // Now we have two terminators. Remove the old one from the block.
1765 VecBody->getTerminator()->eraseFromParent();
1767 // Get ready to start creating new instructions into the vectorized body.
1768 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1771 LoopVectorPreHeader = VectorPH;
1772 LoopScalarPreHeader = ScalarPH;
1773 LoopMiddleBlock = MiddleBlock;
1774 LoopExitBlock = ExitBlock;
1775 LoopVectorBody = VecBody;
1776 LoopScalarBody = OldBasicBlock;
1779 /// This function returns the identity element (or neutral element) for
1780 /// the operation K.
1782 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1787 // Adding, Xoring, Oring zero to a number does not change it.
1788 return ConstantInt::get(Tp, 0);
1789 case RK_IntegerMult:
1790 // Multiplying a number by 1 does not change it.
1791 return ConstantInt::get(Tp, 1);
1793 // AND-ing a number with an all-1 value does not change it.
1794 return ConstantInt::get(Tp, -1, true);
1796 // Multiplying a number by 1 does not change it.
1797 return ConstantFP::get(Tp, 1.0L);
1799 // Adding zero to a number does not change it.
1800 return ConstantFP::get(Tp, 0.0L);
1802 llvm_unreachable("Unknown reduction kind");
1806 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
1807 Intrinsic::ID ValidIntrinsicID) {
1808 if (I.getNumArgOperands() != 1 ||
1809 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1810 I.getType() != I.getArgOperand(0)->getType() ||
1811 !I.onlyReadsMemory())
1812 return Intrinsic::not_intrinsic;
1814 return ValidIntrinsicID;
1817 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
1818 Intrinsic::ID ValidIntrinsicID) {
1819 if (I.getNumArgOperands() != 2 ||
1820 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1821 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
1822 I.getType() != I.getArgOperand(0)->getType() ||
1823 I.getType() != I.getArgOperand(1)->getType() ||
1824 !I.onlyReadsMemory())
1825 return Intrinsic::not_intrinsic;
1827 return ValidIntrinsicID;
1831 static Intrinsic::ID
1832 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1833 // If we have an intrinsic call, check if it is trivially vectorizable.
1834 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1835 switch (II->getIntrinsicID()) {
1836 case Intrinsic::sqrt:
1837 case Intrinsic::sin:
1838 case Intrinsic::cos:
1839 case Intrinsic::exp:
1840 case Intrinsic::exp2:
1841 case Intrinsic::log:
1842 case Intrinsic::log10:
1843 case Intrinsic::log2:
1844 case Intrinsic::fabs:
1845 case Intrinsic::copysign:
1846 case Intrinsic::floor:
1847 case Intrinsic::ceil:
1848 case Intrinsic::trunc:
1849 case Intrinsic::rint:
1850 case Intrinsic::nearbyint:
1851 case Intrinsic::round:
1852 case Intrinsic::pow:
1853 case Intrinsic::fma:
1854 case Intrinsic::fmuladd:
1855 case Intrinsic::lifetime_start:
1856 case Intrinsic::lifetime_end:
1857 return II->getIntrinsicID();
1859 return Intrinsic::not_intrinsic;
1864 return Intrinsic::not_intrinsic;
1867 Function *F = CI->getCalledFunction();
1868 // We're going to make assumptions on the semantics of the functions, check
1869 // that the target knows that it's available in this environment and it does
1870 // not have local linkage.
1871 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
1872 return Intrinsic::not_intrinsic;
1874 // Otherwise check if we have a call to a function that can be turned into a
1875 // vector intrinsic.
1882 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
1886 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
1890 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
1892 case LibFunc::exp2f:
1893 case LibFunc::exp2l:
1894 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
1898 return checkUnaryFloatSignature(*CI, Intrinsic::log);
1899 case LibFunc::log10:
1900 case LibFunc::log10f:
1901 case LibFunc::log10l:
1902 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
1904 case LibFunc::log2f:
1905 case LibFunc::log2l:
1906 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
1908 case LibFunc::fabsf:
1909 case LibFunc::fabsl:
1910 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
1911 case LibFunc::copysign:
1912 case LibFunc::copysignf:
1913 case LibFunc::copysignl:
1914 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
1915 case LibFunc::floor:
1916 case LibFunc::floorf:
1917 case LibFunc::floorl:
1918 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
1920 case LibFunc::ceilf:
1921 case LibFunc::ceill:
1922 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
1923 case LibFunc::trunc:
1924 case LibFunc::truncf:
1925 case LibFunc::truncl:
1926 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
1928 case LibFunc::rintf:
1929 case LibFunc::rintl:
1930 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
1931 case LibFunc::nearbyint:
1932 case LibFunc::nearbyintf:
1933 case LibFunc::nearbyintl:
1934 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
1935 case LibFunc::round:
1936 case LibFunc::roundf:
1937 case LibFunc::roundl:
1938 return checkUnaryFloatSignature(*CI, Intrinsic::round);
1942 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
1945 return Intrinsic::not_intrinsic;
1948 /// This function translates the reduction kind to an LLVM binary operator.
1950 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1952 case LoopVectorizationLegality::RK_IntegerAdd:
1953 return Instruction::Add;
1954 case LoopVectorizationLegality::RK_IntegerMult:
1955 return Instruction::Mul;
1956 case LoopVectorizationLegality::RK_IntegerOr:
1957 return Instruction::Or;
1958 case LoopVectorizationLegality::RK_IntegerAnd:
1959 return Instruction::And;
1960 case LoopVectorizationLegality::RK_IntegerXor:
1961 return Instruction::Xor;
1962 case LoopVectorizationLegality::RK_FloatMult:
1963 return Instruction::FMul;
1964 case LoopVectorizationLegality::RK_FloatAdd:
1965 return Instruction::FAdd;
1966 case LoopVectorizationLegality::RK_IntegerMinMax:
1967 return Instruction::ICmp;
1968 case LoopVectorizationLegality::RK_FloatMinMax:
1969 return Instruction::FCmp;
1971 llvm_unreachable("Unknown reduction operation");
1975 Value *createMinMaxOp(IRBuilder<> &Builder,
1976 LoopVectorizationLegality::MinMaxReductionKind RK,
1979 CmpInst::Predicate P = CmpInst::ICMP_NE;
1982 llvm_unreachable("Unknown min/max reduction kind");
1983 case LoopVectorizationLegality::MRK_UIntMin:
1984 P = CmpInst::ICMP_ULT;
1986 case LoopVectorizationLegality::MRK_UIntMax:
1987 P = CmpInst::ICMP_UGT;
1989 case LoopVectorizationLegality::MRK_SIntMin:
1990 P = CmpInst::ICMP_SLT;
1992 case LoopVectorizationLegality::MRK_SIntMax:
1993 P = CmpInst::ICMP_SGT;
1995 case LoopVectorizationLegality::MRK_FloatMin:
1996 P = CmpInst::FCMP_OLT;
1998 case LoopVectorizationLegality::MRK_FloatMax:
1999 P = CmpInst::FCMP_OGT;
2004 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2005 RK == LoopVectorizationLegality::MRK_FloatMax)
2006 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2008 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2010 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2015 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2016 //===------------------------------------------------===//
2018 // Notice: any optimization or new instruction that go
2019 // into the code below should be also be implemented in
2022 //===------------------------------------------------===//
2023 Constant *Zero = Builder.getInt32(0);
2025 // In order to support reduction variables we need to be able to vectorize
2026 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2027 // stages. First, we create a new vector PHI node with no incoming edges.
2028 // We use this value when we vectorize all of the instructions that use the
2029 // PHI. Next, after all of the instructions in the block are complete we
2030 // add the new incoming edges to the PHI. At this point all of the
2031 // instructions in the basic block are vectorized, so we can use them to
2032 // construct the PHI.
2033 PhiVector RdxPHIsToFix;
2035 // Scan the loop in a topological order to ensure that defs are vectorized
2037 LoopBlocksDFS DFS(OrigLoop);
2040 // Vectorize all of the blocks in the original loop.
2041 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2042 be = DFS.endRPO(); bb != be; ++bb)
2043 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2045 // At this point every instruction in the original loop is widened to
2046 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2047 // that we vectorized. The PHI nodes are currently empty because we did
2048 // not want to introduce cycles. Notice that the remaining PHI nodes
2049 // that we need to fix are reduction variables.
2051 // Create the 'reduced' values for each of the induction vars.
2052 // The reduced values are the vector values that we scalarize and combine
2053 // after the loop is finished.
2054 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2056 PHINode *RdxPhi = *it;
2057 assert(RdxPhi && "Unable to recover vectorized PHI");
2059 // Find the reduction variable descriptor.
2060 assert(Legal->getReductionVars()->count(RdxPhi) &&
2061 "Unable to find the reduction variable");
2062 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2063 (*Legal->getReductionVars())[RdxPhi];
2065 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2067 // We need to generate a reduction vector from the incoming scalar.
2068 // To do so, we need to generate the 'identity' vector and overide
2069 // one of the elements with the incoming scalar reduction. We need
2070 // to do it in the vector-loop preheader.
2071 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2073 // This is the vector-clone of the value that leaves the loop.
2074 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2075 Type *VecTy = VectorExit[0]->getType();
2077 // Find the reduction identity variable. Zero for addition, or, xor,
2078 // one for multiplication, -1 for And.
2081 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2082 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2083 // MinMax reduction have the start value as their identify.
2085 VectorStart = Identity = RdxDesc.StartValue;
2087 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2092 // Handle other reduction kinds:
2094 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2095 VecTy->getScalarType());
2098 // This vector is the Identity vector where the first element is the
2099 // incoming scalar reduction.
2100 VectorStart = RdxDesc.StartValue;
2102 Identity = ConstantVector::getSplat(VF, Iden);
2104 // This vector is the Identity vector where the first element is the
2105 // incoming scalar reduction.
2106 VectorStart = Builder.CreateInsertElement(Identity,
2107 RdxDesc.StartValue, Zero);
2111 // Fix the vector-loop phi.
2112 // We created the induction variable so we know that the
2113 // preheader is the first entry.
2114 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2116 // Reductions do not have to start at zero. They can start with
2117 // any loop invariant values.
2118 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2119 BasicBlock *Latch = OrigLoop->getLoopLatch();
2120 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2121 VectorParts &Val = getVectorValue(LoopVal);
2122 for (unsigned part = 0; part < UF; ++part) {
2123 // Make sure to add the reduction stat value only to the
2124 // first unroll part.
2125 Value *StartVal = (part == 0) ? VectorStart : Identity;
2126 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2127 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2130 // Before each round, move the insertion point right between
2131 // the PHIs and the values we are going to write.
2132 // This allows us to write both PHINodes and the extractelement
2134 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2136 VectorParts RdxParts;
2137 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2138 for (unsigned part = 0; part < UF; ++part) {
2139 // This PHINode contains the vectorized reduction variable, or
2140 // the initial value vector, if we bypass the vector loop.
2141 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2142 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2143 Value *StartVal = (part == 0) ? VectorStart : Identity;
2144 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2145 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2146 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2147 RdxParts.push_back(NewPhi);
2150 // Reduce all of the unrolled parts into a single vector.
2151 Value *ReducedPartRdx = RdxParts[0];
2152 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2153 setDebugLocFromInst(Builder, ReducedPartRdx);
2154 for (unsigned part = 1; part < UF; ++part) {
2155 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2156 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2157 RdxParts[part], ReducedPartRdx,
2160 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2161 ReducedPartRdx, RdxParts[part]);
2165 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2166 // and vector ops, reducing the set of values being computed by half each
2168 assert(isPowerOf2_32(VF) &&
2169 "Reduction emission only supported for pow2 vectors!");
2170 Value *TmpVec = ReducedPartRdx;
2171 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2172 for (unsigned i = VF; i != 1; i >>= 1) {
2173 // Move the upper half of the vector to the lower half.
2174 for (unsigned j = 0; j != i/2; ++j)
2175 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2177 // Fill the rest of the mask with undef.
2178 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2179 UndefValue::get(Builder.getInt32Ty()));
2182 Builder.CreateShuffleVector(TmpVec,
2183 UndefValue::get(TmpVec->getType()),
2184 ConstantVector::get(ShuffleMask),
2187 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2188 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2191 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2194 // The result is in the first element of the vector.
2195 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2196 Builder.getInt32(0));
2199 // Now, we need to fix the users of the reduction variable
2200 // inside and outside of the scalar remainder loop.
2201 // We know that the loop is in LCSSA form. We need to update the
2202 // PHI nodes in the exit blocks.
2203 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2204 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2205 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2206 if (!LCSSAPhi) break;
2208 // All PHINodes need to have a single entry edge, or two if
2209 // we already fixed them.
2210 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2212 // We found our reduction value exit-PHI. Update it with the
2213 // incoming bypass edge.
2214 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2215 // Add an edge coming from the bypass.
2216 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2219 }// end of the LCSSA phi scan.
2221 // Fix the scalar loop reduction variable with the incoming reduction sum
2222 // from the vector body and from the backedge value.
2223 int IncomingEdgeBlockIdx =
2224 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2225 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2226 // Pick the other block.
2227 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2228 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2229 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2230 }// end of for each redux variable.
2235 void InnerLoopVectorizer::fixLCSSAPHIs() {
2236 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2237 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2238 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2239 if (!LCSSAPhi) break;
2240 if (LCSSAPhi->getNumIncomingValues() == 1)
2241 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2246 InnerLoopVectorizer::VectorParts
2247 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2248 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2251 // Look for cached value.
2252 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2253 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2254 if (ECEntryIt != MaskCache.end())
2255 return ECEntryIt->second;
2257 VectorParts SrcMask = createBlockInMask(Src);
2259 // The terminator has to be a branch inst!
2260 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2261 assert(BI && "Unexpected terminator found");
2263 if (BI->isConditional()) {
2264 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2266 if (BI->getSuccessor(0) != Dst)
2267 for (unsigned part = 0; part < UF; ++part)
2268 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2270 for (unsigned part = 0; part < UF; ++part)
2271 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2273 MaskCache[Edge] = EdgeMask;
2277 MaskCache[Edge] = SrcMask;
2281 InnerLoopVectorizer::VectorParts
2282 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2283 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2285 // Loop incoming mask is all-one.
2286 if (OrigLoop->getHeader() == BB) {
2287 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2288 return getVectorValue(C);
2291 // This is the block mask. We OR all incoming edges, and with zero.
2292 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2293 VectorParts BlockMask = getVectorValue(Zero);
2296 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2297 VectorParts EM = createEdgeMask(*it, BB);
2298 for (unsigned part = 0; part < UF; ++part)
2299 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2305 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2306 InnerLoopVectorizer::VectorParts &Entry,
2307 LoopVectorizationLegality *Legal,
2308 unsigned UF, unsigned VF, PhiVector *PV) {
2309 PHINode* P = cast<PHINode>(PN);
2310 // Handle reduction variables:
2311 if (Legal->getReductionVars()->count(P)) {
2312 for (unsigned part = 0; part < UF; ++part) {
2313 // This is phase one of vectorizing PHIs.
2314 Type *VecTy = (VF == 1) ? PN->getType() :
2315 VectorType::get(PN->getType(), VF);
2316 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2317 LoopVectorBody-> getFirstInsertionPt());
2323 setDebugLocFromInst(Builder, P);
2324 // Check for PHI nodes that are lowered to vector selects.
2325 if (P->getParent() != OrigLoop->getHeader()) {
2326 // We know that all PHIs in non header blocks are converted into
2327 // selects, so we don't have to worry about the insertion order and we
2328 // can just use the builder.
2329 // At this point we generate the predication tree. There may be
2330 // duplications since this is a simple recursive scan, but future
2331 // optimizations will clean it up.
2333 unsigned NumIncoming = P->getNumIncomingValues();
2335 // Generate a sequence of selects of the form:
2336 // SELECT(Mask3, In3,
2337 // SELECT(Mask2, In2,
2339 for (unsigned In = 0; In < NumIncoming; In++) {
2340 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2342 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2344 for (unsigned part = 0; part < UF; ++part) {
2345 // We might have single edge PHIs (blocks) - use an identity
2346 // 'select' for the first PHI operand.
2348 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2351 // Select between the current value and the previous incoming edge
2352 // based on the incoming mask.
2353 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2354 Entry[part], "predphi");
2360 // This PHINode must be an induction variable.
2361 // Make sure that we know about it.
2362 assert(Legal->getInductionVars()->count(P) &&
2363 "Not an induction variable");
2365 LoopVectorizationLegality::InductionInfo II =
2366 Legal->getInductionVars()->lookup(P);
2369 case LoopVectorizationLegality::IK_NoInduction:
2370 llvm_unreachable("Unknown induction");
2371 case LoopVectorizationLegality::IK_IntInduction: {
2372 assert(P->getType() == II.StartValue->getType() && "Types must match");
2373 Type *PhiTy = P->getType();
2375 if (P == OldInduction) {
2376 // Handle the canonical induction variable. We might have had to
2378 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2380 // Handle other induction variables that are now based on the
2382 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2384 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2385 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2388 Broadcasted = getBroadcastInstrs(Broadcasted);
2389 // After broadcasting the induction variable we need to make the vector
2390 // consecutive by adding 0, 1, 2, etc.
2391 for (unsigned part = 0; part < UF; ++part)
2392 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2395 case LoopVectorizationLegality::IK_ReverseIntInduction:
2396 case LoopVectorizationLegality::IK_PtrInduction:
2397 case LoopVectorizationLegality::IK_ReversePtrInduction:
2398 // Handle reverse integer and pointer inductions.
2399 Value *StartIdx = ExtendedIdx;
2400 // This is the normalized GEP that starts counting at zero.
2401 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2404 // Handle the reverse integer induction variable case.
2405 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2406 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2407 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2409 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2412 // This is a new value so do not hoist it out.
2413 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2414 // After broadcasting the induction variable we need to make the
2415 // vector consecutive by adding ... -3, -2, -1, 0.
2416 for (unsigned part = 0; part < UF; ++part)
2417 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2422 // Handle the pointer induction variable case.
2423 assert(P->getType()->isPointerTy() && "Unexpected type.");
2425 // Is this a reverse induction ptr or a consecutive induction ptr.
2426 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2429 // This is the vector of results. Notice that we don't generate
2430 // vector geps because scalar geps result in better code.
2431 for (unsigned part = 0; part < UF; ++part) {
2433 int EltIndex = (part) * (Reverse ? -1 : 1);
2434 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2437 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2439 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2441 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2443 Entry[part] = SclrGep;
2447 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2448 for (unsigned int i = 0; i < VF; ++i) {
2449 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2450 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2453 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2455 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2457 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2459 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2460 Builder.getInt32(i),
2463 Entry[part] = VecVal;
2470 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2471 BasicBlock *BB, PhiVector *PV) {
2472 // For each instruction in the old loop.
2473 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2474 VectorParts &Entry = WidenMap.get(it);
2475 switch (it->getOpcode()) {
2476 case Instruction::Br:
2477 // Nothing to do for PHIs and BR, since we already took care of the
2478 // loop control flow instructions.
2480 case Instruction::PHI:{
2481 // Vectorize PHINodes.
2482 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2486 case Instruction::Add:
2487 case Instruction::FAdd:
2488 case Instruction::Sub:
2489 case Instruction::FSub:
2490 case Instruction::Mul:
2491 case Instruction::FMul:
2492 case Instruction::UDiv:
2493 case Instruction::SDiv:
2494 case Instruction::FDiv:
2495 case Instruction::URem:
2496 case Instruction::SRem:
2497 case Instruction::FRem:
2498 case Instruction::Shl:
2499 case Instruction::LShr:
2500 case Instruction::AShr:
2501 case Instruction::And:
2502 case Instruction::Or:
2503 case Instruction::Xor: {
2504 // Just widen binops.
2505 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2506 setDebugLocFromInst(Builder, BinOp);
2507 VectorParts &A = getVectorValue(it->getOperand(0));
2508 VectorParts &B = getVectorValue(it->getOperand(1));
2510 // Use this vector value for all users of the original instruction.
2511 for (unsigned Part = 0; Part < UF; ++Part) {
2512 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2514 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2515 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2516 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2517 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2518 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2520 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2521 VecOp->setIsExact(BinOp->isExact());
2527 case Instruction::Select: {
2529 // If the selector is loop invariant we can create a select
2530 // instruction with a scalar condition. Otherwise, use vector-select.
2531 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2533 setDebugLocFromInst(Builder, it);
2535 // The condition can be loop invariant but still defined inside the
2536 // loop. This means that we can't just use the original 'cond' value.
2537 // We have to take the 'vectorized' value and pick the first lane.
2538 // Instcombine will make this a no-op.
2539 VectorParts &Cond = getVectorValue(it->getOperand(0));
2540 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2541 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2543 Value *ScalarCond = (VF == 1) ? Cond[0] :
2544 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2546 for (unsigned Part = 0; Part < UF; ++Part) {
2547 Entry[Part] = Builder.CreateSelect(
2548 InvariantCond ? ScalarCond : Cond[Part],
2555 case Instruction::ICmp:
2556 case Instruction::FCmp: {
2557 // Widen compares. Generate vector compares.
2558 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2559 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2560 setDebugLocFromInst(Builder, it);
2561 VectorParts &A = getVectorValue(it->getOperand(0));
2562 VectorParts &B = getVectorValue(it->getOperand(1));
2563 for (unsigned Part = 0; Part < UF; ++Part) {
2566 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2568 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2574 case Instruction::Store:
2575 case Instruction::Load:
2576 vectorizeMemoryInstruction(it, Legal);
2578 case Instruction::ZExt:
2579 case Instruction::SExt:
2580 case Instruction::FPToUI:
2581 case Instruction::FPToSI:
2582 case Instruction::FPExt:
2583 case Instruction::PtrToInt:
2584 case Instruction::IntToPtr:
2585 case Instruction::SIToFP:
2586 case Instruction::UIToFP:
2587 case Instruction::Trunc:
2588 case Instruction::FPTrunc:
2589 case Instruction::BitCast: {
2590 CastInst *CI = dyn_cast<CastInst>(it);
2591 setDebugLocFromInst(Builder, it);
2592 /// Optimize the special case where the source is the induction
2593 /// variable. Notice that we can only optimize the 'trunc' case
2594 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2595 /// c. other casts depend on pointer size.
2596 if (CI->getOperand(0) == OldInduction &&
2597 it->getOpcode() == Instruction::Trunc) {
2598 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2600 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2601 for (unsigned Part = 0; Part < UF; ++Part)
2602 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2605 /// Vectorize casts.
2606 Type *DestTy = (VF == 1) ? CI->getType() :
2607 VectorType::get(CI->getType(), VF);
2609 VectorParts &A = getVectorValue(it->getOperand(0));
2610 for (unsigned Part = 0; Part < UF; ++Part)
2611 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2615 case Instruction::Call: {
2616 // Ignore dbg intrinsics.
2617 if (isa<DbgInfoIntrinsic>(it))
2619 setDebugLocFromInst(Builder, it);
2621 Module *M = BB->getParent()->getParent();
2622 CallInst *CI = cast<CallInst>(it);
2623 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2624 assert(ID && "Not an intrinsic call!");
2626 case Intrinsic::lifetime_end:
2627 case Intrinsic::lifetime_start:
2628 scalarizeInstruction(it);
2631 for (unsigned Part = 0; Part < UF; ++Part) {
2632 SmallVector<Value *, 4> Args;
2633 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2634 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2635 Args.push_back(Arg[Part]);
2637 Type *Tys[] = {CI->getType()};
2639 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2641 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2642 Entry[Part] = Builder.CreateCall(F, Args);
2650 // All other instructions are unsupported. Scalarize them.
2651 scalarizeInstruction(it);
2654 }// end of for_each instr.
2657 void InnerLoopVectorizer::updateAnalysis() {
2658 // Forget the original basic block.
2659 SE->forgetLoop(OrigLoop);
2661 // Update the dominator tree information.
2662 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2663 "Entry does not dominate exit.");
2665 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2666 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2667 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2668 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2669 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2670 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2671 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2672 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2674 DEBUG(DT->verifyAnalysis());
2677 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2678 if (!EnableIfConversion)
2681 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2682 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2684 // A list of pointers that we can safely read and write to.
2685 SmallPtrSet<Value *, 8> SafePointes;
2687 // Collect safe addresses.
2688 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2689 BasicBlock *BB = LoopBlocks[i];
2691 if (blockNeedsPredication(BB))
2694 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2695 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2696 SafePointes.insert(LI->getPointerOperand());
2697 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2698 SafePointes.insert(SI->getPointerOperand());
2702 // Collect the blocks that need predication.
2703 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2704 BasicBlock *BB = LoopBlocks[i];
2706 // We don't support switch statements inside loops.
2707 if (!isa<BranchInst>(BB->getTerminator()))
2710 // We must be able to predicate all blocks that need to be predicated.
2711 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2715 // We can if-convert this loop.
2719 bool LoopVectorizationLegality::canVectorize() {
2720 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2721 // be canonicalized.
2722 if (!TheLoop->getLoopPreheader())
2725 // We can only vectorize innermost loops.
2726 if (TheLoop->getSubLoopsVector().size())
2729 // We must have a single backedge.
2730 if (TheLoop->getNumBackEdges() != 1)
2733 // We must have a single exiting block.
2734 if (!TheLoop->getExitingBlock())
2737 unsigned NumBlocks = TheLoop->getNumBlocks();
2739 // Check if we can if-convert non single-bb loops.
2740 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2741 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2745 // We need to have a loop header.
2746 BasicBlock *Latch = TheLoop->getLoopLatch();
2747 DEBUG(dbgs() << "LV: Found a loop: " <<
2748 TheLoop->getHeader()->getName() << "\n");
2750 // ScalarEvolution needs to be able to find the exit count.
2751 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2752 if (ExitCount == SE->getCouldNotCompute()) {
2753 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2757 // Do not loop-vectorize loops with a tiny trip count.
2758 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2759 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2760 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2761 "This loop is not worth vectorizing.\n");
2765 // Check if we can vectorize the instructions and CFG in this loop.
2766 if (!canVectorizeInstrs()) {
2767 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2771 // Go over each instruction and look at memory deps.
2772 if (!canVectorizeMemory()) {
2773 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2777 // Collect all of the variables that remain uniform after vectorization.
2778 collectLoopUniforms();
2780 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2781 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2784 // Okay! We can vectorize. At this point we don't have any other mem analysis
2785 // which may limit our maximum vectorization factor, so just return true with
2790 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2791 if (Ty->isPointerTy())
2792 return DL.getIntPtrType(Ty);
2797 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2798 Ty0 = convertPointerToIntegerType(DL, Ty0);
2799 Ty1 = convertPointerToIntegerType(DL, Ty1);
2800 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2805 /// \brief Check that the instruction has outside loop users and is not an
2806 /// identified reduction variable.
2807 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2808 SmallPtrSet<Value *, 4> &Reductions) {
2809 // Reduction instructions are allowed to have exit users. All other
2810 // instructions must not have external users.
2811 if (!Reductions.count(Inst))
2812 //Check that all of the users of the loop are inside the BB.
2813 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2815 Instruction *U = cast<Instruction>(*I);
2816 // This user may be a reduction exit value.
2817 if (!TheLoop->contains(U)) {
2818 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2825 bool LoopVectorizationLegality::canVectorizeInstrs() {
2826 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2827 BasicBlock *Header = TheLoop->getHeader();
2829 // Look for the attribute signaling the absence of NaNs.
2830 Function &F = *Header->getParent();
2831 if (F.hasFnAttribute("no-nans-fp-math"))
2832 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2833 AttributeSet::FunctionIndex,
2834 "no-nans-fp-math").getValueAsString() == "true";
2836 // For each block in the loop.
2837 for (Loop::block_iterator bb = TheLoop->block_begin(),
2838 be = TheLoop->block_end(); bb != be; ++bb) {
2840 // Scan the instructions in the block and look for hazards.
2841 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2844 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2845 Type *PhiTy = Phi->getType();
2846 // Check that this PHI type is allowed.
2847 if (!PhiTy->isIntegerTy() &&
2848 !PhiTy->isFloatingPointTy() &&
2849 !PhiTy->isPointerTy()) {
2850 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2854 // If this PHINode is not in the header block, then we know that we
2855 // can convert it to select during if-conversion. No need to check if
2856 // the PHIs in this block are induction or reduction variables.
2857 if (*bb != Header) {
2858 // Check that this instruction has no outside users or is an
2859 // identified reduction value with an outside user.
2860 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2865 // We only allow if-converted PHIs with more than two incoming values.
2866 if (Phi->getNumIncomingValues() != 2) {
2867 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2871 // This is the value coming from the preheader.
2872 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2873 // Check if this is an induction variable.
2874 InductionKind IK = isInductionVariable(Phi);
2876 if (IK_NoInduction != IK) {
2877 // Get the widest type.
2879 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2881 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2883 // Int inductions are special because we only allow one IV.
2884 if (IK == IK_IntInduction) {
2885 // Use the phi node with the widest type as induction. Use the last
2886 // one if there are multiple (no good reason for doing this other
2887 // than it is expedient).
2888 if (!Induction || PhiTy == WidestIndTy)
2892 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2893 Inductions[Phi] = InductionInfo(StartValue, IK);
2895 // Until we explicitly handle the case of an induction variable with
2896 // an outside loop user we have to give up vectorizing this loop.
2897 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2903 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2904 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2907 if (AddReductionVar(Phi, RK_IntegerMult)) {
2908 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2911 if (AddReductionVar(Phi, RK_IntegerOr)) {
2912 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2915 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2916 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2919 if (AddReductionVar(Phi, RK_IntegerXor)) {
2920 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2923 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2924 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2927 if (AddReductionVar(Phi, RK_FloatMult)) {
2928 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2931 if (AddReductionVar(Phi, RK_FloatAdd)) {
2932 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2935 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2936 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
2941 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2943 }// end of PHI handling
2945 // We still don't handle functions. However, we can ignore dbg intrinsic
2946 // calls and we do handle certain intrinsic and libm functions.
2947 CallInst *CI = dyn_cast<CallInst>(it);
2948 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2949 DEBUG(dbgs() << "LV: Found a call site.\n");
2953 // Check that the instruction return type is vectorizable.
2954 if (!VectorType::isValidElementType(it->getType()) &&
2955 !it->getType()->isVoidTy()) {
2956 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2960 // Check that the stored type is vectorizable.
2961 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2962 Type *T = ST->getValueOperand()->getType();
2963 if (!VectorType::isValidElementType(T))
2967 // Reduction instructions are allowed to have exit users.
2968 // All other instructions must not have external users.
2969 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2977 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2978 if (Inductions.empty())
2985 void LoopVectorizationLegality::collectLoopUniforms() {
2986 // We now know that the loop is vectorizable!
2987 // Collect variables that will remain uniform after vectorization.
2988 std::vector<Value*> Worklist;
2989 BasicBlock *Latch = TheLoop->getLoopLatch();
2991 // Start with the conditional branch and walk up the block.
2992 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2994 while (Worklist.size()) {
2995 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2996 Worklist.pop_back();
2998 // Look at instructions inside this loop.
2999 // Stop when reaching PHI nodes.
3000 // TODO: we need to follow values all over the loop, not only in this block.
3001 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3004 // This is a known uniform.
3007 // Insert all operands.
3008 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3013 /// \brief Analyses memory accesses in a loop.
3015 /// Checks whether run time pointer checks are needed and builds sets for data
3016 /// dependence checking.
3017 class AccessAnalysis {
3019 /// \brief Read or write access location.
3020 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3021 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3023 /// \brief Set of potential dependent memory accesses.
3024 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3026 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3027 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3028 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3030 /// \brief Register a load and whether it is only read from.
3031 void addLoad(Value *Ptr, bool IsReadOnly) {
3032 Accesses.insert(MemAccessInfo(Ptr, false));
3034 ReadOnlyPtr.insert(Ptr);
3037 /// \brief Register a store.
3038 void addStore(Value *Ptr) {
3039 Accesses.insert(MemAccessInfo(Ptr, true));
3042 /// \brief Check whether we can check the pointers at runtime for
3043 /// non-intersection.
3044 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3045 unsigned &NumComparisons, ScalarEvolution *SE,
3048 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3049 /// and builds sets of dependent accesses.
3050 void buildDependenceSets() {
3051 // Process read-write pointers first.
3052 processMemAccesses(false);
3053 // Next, process read pointers.
3054 processMemAccesses(true);
3057 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3059 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3061 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3064 typedef SetVector<MemAccessInfo> PtrAccessSet;
3065 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3067 /// \brief Go over all memory access or only the deferred ones if
3068 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3069 /// and build sets of dependency check candidates.
3070 void processMemAccesses(bool UseDeferred);
3072 /// Set of all accesses.
3073 PtrAccessSet Accesses;
3075 /// Set of access to check after all writes have been processed.
3076 PtrAccessSet DeferredAccesses;
3078 /// Map of pointers to last access encountered.
3079 UnderlyingObjToAccessMap ObjToLastAccess;
3081 /// Set of accesses that need a further dependence check.
3082 MemAccessInfoSet CheckDeps;
3084 /// Set of pointers that are read only.
3085 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3087 /// Set of underlying objects already written to.
3088 SmallPtrSet<Value*, 16> WriteObjects;
3092 /// Sets of potentially dependent accesses - members of one set share an
3093 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3094 /// dependence check.
3095 DepCandidates &DepCands;
3097 bool AreAllWritesIdentified;
3098 bool AreAllReadsIdentified;
3099 bool IsRTCheckNeeded;
3102 } // end anonymous namespace
3104 /// \brief Check whether a pointer can participate in a runtime bounds check.
3105 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3106 const SCEV *PtrScev = SE->getSCEV(Ptr);
3107 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3111 return AR->isAffine();
3114 bool AccessAnalysis::canCheckPtrAtRT(
3115 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3116 unsigned &NumComparisons, ScalarEvolution *SE,
3118 // Find pointers with computable bounds. We are going to use this information
3119 // to place a runtime bound check.
3120 unsigned NumReadPtrChecks = 0;
3121 unsigned NumWritePtrChecks = 0;
3122 bool CanDoRT = true;
3124 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3125 // We assign consecutive id to access from different dependence sets.
3126 // Accesses within the same set don't need a runtime check.
3127 unsigned RunningDepId = 1;
3128 DenseMap<Value *, unsigned> DepSetId;
3130 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3132 const MemAccessInfo &Access = *AI;
3133 Value *Ptr = Access.getPointer();
3134 bool IsWrite = Access.getInt();
3136 // Just add write checks if we have both.
3137 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3141 ++NumWritePtrChecks;
3145 if (hasComputableBounds(SE, Ptr)) {
3146 // The id of the dependence set.
3149 if (IsDepCheckNeeded) {
3150 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3151 unsigned &LeaderId = DepSetId[Leader];
3153 LeaderId = RunningDepId++;
3156 // Each access has its own dependence set.
3157 DepId = RunningDepId++;
3159 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3161 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3167 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3168 NumComparisons = 0; // Only one dependence set.
3170 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3171 NumWritePtrChecks - 1));
3175 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3176 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3179 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3180 // We process the set twice: first we process read-write pointers, last we
3181 // process read-only pointers. This allows us to skip dependence tests for
3182 // read-only pointers.
3184 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3185 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3186 const MemAccessInfo &Access = *AI;
3187 Value *Ptr = Access.getPointer();
3188 bool IsWrite = Access.getInt();
3190 DepCands.insert(Access);
3192 // Memorize read-only pointers for later processing and skip them in the
3193 // first round (they need to be checked after we have seen all write
3194 // pointers). Note: we also mark pointer that are not consecutive as
3195 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3196 // second check for "!IsWrite".
3197 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3198 if (!UseDeferred && IsReadOnlyPtr) {
3199 DeferredAccesses.insert(Access);
3203 bool NeedDepCheck = false;
3204 // Check whether there is the possiblity of dependency because of underlying
3205 // objects being the same.
3206 typedef SmallVector<Value*, 16> ValueVector;
3207 ValueVector TempObjects;
3208 GetUnderlyingObjects(Ptr, TempObjects, DL);
3209 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3211 Value *UnderlyingObj = *UI;
3213 // If this is a write then it needs to be an identified object. If this a
3214 // read and all writes (so far) are identified function scope objects we
3215 // don't need an identified underlying object but only an Argument (the
3216 // next write is going to invalidate this assumption if it is
3218 // This is a micro-optimization for the case where all writes are
3219 // identified and we have one argument pointer.
3220 // Otherwise, we do need a runtime check.
3221 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3222 (!IsWrite && (!AreAllWritesIdentified ||
3223 !isa<Argument>(UnderlyingObj)) &&
3224 !isIdentifiedObject(UnderlyingObj))) {
3225 DEBUG(dbgs() << "LV: Found an unidentified " <<
3226 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3228 IsRTCheckNeeded = (IsRTCheckNeeded ||
3229 !isIdentifiedObject(UnderlyingObj) ||
3230 !AreAllReadsIdentified);
3233 AreAllWritesIdentified = false;
3235 AreAllReadsIdentified = false;
3238 // If this is a write - check other reads and writes for conflicts. If
3239 // this is a read only check other writes for conflicts (but only if there
3240 // is no other write to the ptr - this is an optimization to catch "a[i] =
3241 // a[i] + " without having to do a dependence check).
3242 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3243 NeedDepCheck = true;
3246 WriteObjects.insert(UnderlyingObj);
3248 // Create sets of pointers connected by shared underlying objects.
3249 UnderlyingObjToAccessMap::iterator Prev =
3250 ObjToLastAccess.find(UnderlyingObj);
3251 if (Prev != ObjToLastAccess.end())
3252 DepCands.unionSets(Access, Prev->second);
3254 ObjToLastAccess[UnderlyingObj] = Access;
3258 CheckDeps.insert(Access);
3263 /// \brief Checks memory dependences among accesses to the same underlying
3264 /// object to determine whether there vectorization is legal or not (and at
3265 /// which vectorization factor).
3267 /// This class works under the assumption that we already checked that memory
3268 /// locations with different underlying pointers are "must-not alias".
3269 /// We use the ScalarEvolution framework to symbolically evalutate access
3270 /// functions pairs. Since we currently don't restructure the loop we can rely
3271 /// on the program order of memory accesses to determine their safety.
3272 /// At the moment we will only deem accesses as safe for:
3273 /// * A negative constant distance assuming program order.
3275 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3276 /// a[i] = tmp; y = a[i];
3278 /// The latter case is safe because later checks guarantuee that there can't
3279 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3280 /// the same variable: a header phi can only be an induction or a reduction, a
3281 /// reduction can't have a memory sink, an induction can't have a memory
3282 /// source). This is important and must not be violated (or we have to
3283 /// resort to checking for cycles through memory).
3285 /// * A positive constant distance assuming program order that is bigger
3286 /// than the biggest memory access.
3288 /// tmp = a[i] OR b[i] = x
3289 /// a[i+2] = tmp y = b[i+2];
3291 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3293 /// * Zero distances and all accesses have the same size.
3295 class MemoryDepChecker {
3297 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3298 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3300 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3301 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3303 /// \brief Register the location (instructions are given increasing numbers)
3304 /// of a write access.
3305 void addAccess(StoreInst *SI) {
3306 Value *Ptr = SI->getPointerOperand();
3307 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3308 InstMap.push_back(SI);
3312 /// \brief Register the location (instructions are given increasing numbers)
3313 /// of a write access.
3314 void addAccess(LoadInst *LI) {
3315 Value *Ptr = LI->getPointerOperand();
3316 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3317 InstMap.push_back(LI);
3321 /// \brief Check whether the dependencies between the accesses are safe.
3323 /// Only checks sets with elements in \p CheckDeps.
3324 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3325 MemAccessInfoSet &CheckDeps);
3327 /// \brief The maximum number of bytes of a vector register we can vectorize
3328 /// the accesses safely with.
3329 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3332 ScalarEvolution *SE;
3334 const Loop *InnermostLoop;
3336 /// \brief Maps access locations (ptr, read/write) to program order.
3337 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3339 /// \brief Memory access instructions in program order.
3340 SmallVector<Instruction *, 16> InstMap;
3342 /// \brief The program order index to be used for the next instruction.
3345 // We can access this many bytes in parallel safely.
3346 unsigned MaxSafeDepDistBytes;
3348 /// \brief Check whether there is a plausible dependence between the two
3351 /// Access \p A must happen before \p B in program order. The two indices
3352 /// identify the index into the program order map.
3354 /// This function checks whether there is a plausible dependence (or the
3355 /// absence of such can't be proved) between the two accesses. If there is a
3356 /// plausible dependence but the dependence distance is bigger than one
3357 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3358 /// distance is smaller than any other distance encountered so far).
3359 /// Otherwise, this function returns true signaling a possible dependence.
3360 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3361 const MemAccessInfo &B, unsigned BIdx);
3363 /// \brief Check whether the data dependence could prevent store-load
3365 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3368 } // end anonymous namespace
3370 static bool isInBoundsGep(Value *Ptr) {
3371 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3372 return GEP->isInBounds();
3376 /// \brief Check whether the access through \p Ptr has a constant stride.
3377 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3379 const Type *Ty = Ptr->getType();
3380 assert(Ty->isPointerTy() && "Unexpected non ptr");
3382 // Make sure that the pointer does not point to aggregate types.
3383 const PointerType *PtrTy = cast<PointerType>(Ty);
3384 if (PtrTy->getElementType()->isAggregateType()) {
3385 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3390 const SCEV *PtrScev = SE->getSCEV(Ptr);
3391 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3393 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3394 << *Ptr << " SCEV: " << *PtrScev << "\n");
3398 // The accesss function must stride over the innermost loop.
3399 if (Lp != AR->getLoop()) {
3400 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3401 *Ptr << " SCEV: " << *PtrScev << "\n");
3404 // The address calculation must not wrap. Otherwise, a dependence could be
3406 // An inbounds getelementptr that is a AddRec with a unit stride
3407 // cannot wrap per definition. The unit stride requirement is checked later.
3408 // An getelementptr without an inbounds attribute and unit stride would have
3409 // to access the pointer value "0" which is undefined behavior in address
3410 // space 0, therefore we can also vectorize this case.
3411 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3412 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3413 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3414 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3415 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3416 << *Ptr << " SCEV: " << *PtrScev << "\n");
3420 // Check the step is constant.
3421 const SCEV *Step = AR->getStepRecurrence(*SE);
3423 // Calculate the pointer stride and check if it is consecutive.
3424 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3426 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3427 " SCEV: " << *PtrScev << "\n");
3431 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3432 const APInt &APStepVal = C->getValue()->getValue();
3434 // Huge step value - give up.
3435 if (APStepVal.getBitWidth() > 64)
3438 int64_t StepVal = APStepVal.getSExtValue();
3441 int64_t Stride = StepVal / Size;
3442 int64_t Rem = StepVal % Size;
3446 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3447 // know we can't "wrap around the address space". In case of address space
3448 // zero we know that this won't happen without triggering undefined behavior.
3449 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3450 Stride != 1 && Stride != -1)
3456 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3457 unsigned TypeByteSize) {
3458 // If loads occur at a distance that is not a multiple of a feasible vector
3459 // factor store-load forwarding does not take place.
3460 // Positive dependences might cause troubles because vectorizing them might
3461 // prevent store-load forwarding making vectorized code run a lot slower.
3462 // a[i] = a[i-3] ^ a[i-8];
3463 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3464 // hence on your typical architecture store-load forwarding does not take
3465 // place. Vectorizing in such cases does not make sense.
3466 // Store-load forwarding distance.
3467 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3468 // Maximum vector factor.
3469 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3470 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3471 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3473 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3475 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3476 MaxVFWithoutSLForwardIssues = (vf >>=1);
3481 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3482 DEBUG(dbgs() << "LV: Distance " << Distance <<
3483 " that could cause a store-load forwarding conflict\n");
3487 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3488 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3489 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3493 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3494 const MemAccessInfo &B, unsigned BIdx) {
3495 assert (AIdx < BIdx && "Must pass arguments in program order");
3497 Value *APtr = A.getPointer();
3498 Value *BPtr = B.getPointer();
3499 bool AIsWrite = A.getInt();
3500 bool BIsWrite = B.getInt();
3502 // Two reads are independent.
3503 if (!AIsWrite && !BIsWrite)
3506 const SCEV *AScev = SE->getSCEV(APtr);
3507 const SCEV *BScev = SE->getSCEV(BPtr);
3509 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3510 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3512 const SCEV *Src = AScev;
3513 const SCEV *Sink = BScev;
3515 // If the induction step is negative we have to invert source and sink of the
3517 if (StrideAPtr < 0) {
3520 std::swap(APtr, BPtr);
3521 std::swap(Src, Sink);
3522 std::swap(AIsWrite, BIsWrite);
3523 std::swap(AIdx, BIdx);
3524 std::swap(StrideAPtr, StrideBPtr);
3527 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3529 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3530 << "(Induction step: " << StrideAPtr << ")\n");
3531 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3532 << *InstMap[BIdx] << ": " << *Dist << "\n");
3534 // Need consecutive accesses. We don't want to vectorize
3535 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3536 // the address space.
3537 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3538 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3542 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3544 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3548 Type *ATy = APtr->getType()->getPointerElementType();
3549 Type *BTy = BPtr->getType()->getPointerElementType();
3550 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3552 // Negative distances are not plausible dependencies.
3553 const APInt &Val = C->getValue()->getValue();
3554 if (Val.isNegative()) {
3555 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3556 if (IsTrueDataDependence &&
3557 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3561 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3565 // Write to the same location with the same size.
3566 // Could be improved to assert type sizes are the same (i32 == float, etc).
3570 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3574 assert(Val.isStrictlyPositive() && "Expect a positive value");
3576 // Positive distance bigger than max vectorization factor.
3579 "LV: ReadWrite-Write positive dependency with different types");
3583 unsigned Distance = (unsigned) Val.getZExtValue();
3585 // Bail out early if passed-in parameters make vectorization not feasible.
3586 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3587 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3589 // The distance must be bigger than the size needed for a vectorized version
3590 // of the operation and the size of the vectorized operation must not be
3591 // bigger than the currrent maximum size.
3592 if (Distance < 2*TypeByteSize ||
3593 2*TypeByteSize > MaxSafeDepDistBytes ||
3594 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3595 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3596 << Val.getSExtValue() << "\n");
3600 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3601 Distance : MaxSafeDepDistBytes;
3603 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3604 if (IsTrueDataDependence &&
3605 couldPreventStoreLoadForward(Distance, TypeByteSize))
3608 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3609 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3615 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3616 MemAccessInfoSet &CheckDeps) {
3618 MaxSafeDepDistBytes = -1U;
3619 while (!CheckDeps.empty()) {
3620 MemAccessInfo CurAccess = *CheckDeps.begin();
3622 // Get the relevant memory access set.
3623 EquivalenceClasses<MemAccessInfo>::iterator I =
3624 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3626 // Check accesses within this set.
3627 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3628 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3630 // Check every access pair.
3632 CheckDeps.erase(*AI);
3633 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3635 // Check every accessing instruction pair in program order.
3636 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3637 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3638 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3639 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3640 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3642 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3653 bool LoopVectorizationLegality::canVectorizeMemory() {
3655 typedef SmallVector<Value*, 16> ValueVector;
3656 typedef SmallPtrSet<Value*, 16> ValueSet;
3658 // Holds the Load and Store *instructions*.
3662 // Holds all the different accesses in the loop.
3663 unsigned NumReads = 0;
3664 unsigned NumReadWrites = 0;
3666 PtrRtCheck.Pointers.clear();
3667 PtrRtCheck.Need = false;
3669 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3670 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3673 for (Loop::block_iterator bb = TheLoop->block_begin(),
3674 be = TheLoop->block_end(); bb != be; ++bb) {
3676 // Scan the BB and collect legal loads and stores.
3677 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3680 // If this is a load, save it. If this instruction can read from memory
3681 // but is not a load, then we quit. Notice that we don't handle function
3682 // calls that read or write.
3683 if (it->mayReadFromMemory()) {
3684 // Many math library functions read the rounding mode. We will only
3685 // vectorize a loop if it contains known function calls that don't set
3686 // the flag. Therefore, it is safe to ignore this read from memory.
3687 CallInst *Call = dyn_cast<CallInst>(it);
3688 if (Call && getIntrinsicIDForCall(Call, TLI))
3691 LoadInst *Ld = dyn_cast<LoadInst>(it);
3692 if (!Ld) return false;
3693 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3694 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3697 Loads.push_back(Ld);
3698 DepChecker.addAccess(Ld);
3702 // Save 'store' instructions. Abort if other instructions write to memory.
3703 if (it->mayWriteToMemory()) {
3704 StoreInst *St = dyn_cast<StoreInst>(it);
3705 if (!St) return false;
3706 if (!St->isSimple() && !IsAnnotatedParallel) {
3707 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3710 Stores.push_back(St);
3711 DepChecker.addAccess(St);
3716 // Now we have two lists that hold the loads and the stores.
3717 // Next, we find the pointers that they use.
3719 // Check if we see any stores. If there are no stores, then we don't
3720 // care if the pointers are *restrict*.
3721 if (!Stores.size()) {
3722 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3726 AccessAnalysis::DepCandidates DependentAccesses;
3727 AccessAnalysis Accesses(DL, DependentAccesses);
3729 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3730 // multiple times on the same object. If the ptr is accessed twice, once
3731 // for read and once for write, it will only appear once (on the write
3732 // list). This is okay, since we are going to check for conflicts between
3733 // writes and between reads and writes, but not between reads and reads.
3736 ValueVector::iterator I, IE;
3737 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3738 StoreInst *ST = cast<StoreInst>(*I);
3739 Value* Ptr = ST->getPointerOperand();
3741 if (isUniform(Ptr)) {
3742 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3746 // If we did *not* see this pointer before, insert it to the read-write
3747 // list. At this phase it is only a 'write' list.
3748 if (Seen.insert(Ptr)) {
3750 Accesses.addStore(Ptr);
3754 if (IsAnnotatedParallel) {
3756 << "LV: A loop annotated parallel, ignore memory dependency "
3761 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3762 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3763 LoadInst *LD = cast<LoadInst>(*I);
3764 Value* Ptr = LD->getPointerOperand();
3765 // If we did *not* see this pointer before, insert it to the
3766 // read list. If we *did* see it before, then it is already in
3767 // the read-write list. This allows us to vectorize expressions
3768 // such as A[i] += x; Because the address of A[i] is a read-write
3769 // pointer. This only works if the index of A[i] is consecutive.
3770 // If the address of i is unknown (for example A[B[i]]) then we may
3771 // read a few words, modify, and write a few words, and some of the
3772 // words may be written to the same address.
3773 bool IsReadOnlyPtr = false;
3774 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3776 IsReadOnlyPtr = true;
3778 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3781 // If we write (or read-write) to a single destination and there are no
3782 // other reads in this loop then is it safe to vectorize.
3783 if (NumReadWrites == 1 && NumReads == 0) {
3784 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3788 // Build dependence sets and check whether we need a runtime pointer bounds
3790 Accesses.buildDependenceSets();
3791 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3793 // Find pointers with computable bounds. We are going to use this information
3794 // to place a runtime bound check.
3795 unsigned NumComparisons = 0;
3796 bool CanDoRT = false;
3798 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3801 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3802 " pointer comparisons.\n");
3804 // If we only have one set of dependences to check pointers among we don't
3805 // need a runtime check.
3806 if (NumComparisons == 0 && NeedRTCheck)
3807 NeedRTCheck = false;
3809 // Check that we did not collect too many pointers or found a unsizeable
3811 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3817 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3820 if (NeedRTCheck && !CanDoRT) {
3821 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3822 "the array bounds.\n");
3827 PtrRtCheck.Need = NeedRTCheck;
3829 bool CanVecMem = true;
3830 if (Accesses.isDependencyCheckNeeded()) {
3831 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3832 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3833 Accesses.getDependenciesToCheck());
3834 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3837 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3838 " need a runtime memory check.\n");
3843 static bool hasMultipleUsesOf(Instruction *I,
3844 SmallPtrSet<Instruction *, 8> &Insts) {
3845 unsigned NumUses = 0;
3846 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3847 if (Insts.count(dyn_cast<Instruction>(*Use)))
3856 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3857 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3858 if (!Set.count(dyn_cast<Instruction>(*Use)))
3863 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3864 ReductionKind Kind) {
3865 if (Phi->getNumIncomingValues() != 2)
3868 // Reduction variables are only found in the loop header block.
3869 if (Phi->getParent() != TheLoop->getHeader())
3872 // Obtain the reduction start value from the value that comes from the loop
3874 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3876 // ExitInstruction is the single value which is used outside the loop.
3877 // We only allow for a single reduction value to be used outside the loop.
3878 // This includes users of the reduction, variables (which form a cycle
3879 // which ends in the phi node).
3880 Instruction *ExitInstruction = 0;
3881 // Indicates that we found a reduction operation in our scan.
3882 bool FoundReduxOp = false;
3884 // We start with the PHI node and scan for all of the users of this
3885 // instruction. All users must be instructions that can be used as reduction
3886 // variables (such as ADD). We must have a single out-of-block user. The cycle
3887 // must include the original PHI.
3888 bool FoundStartPHI = false;
3890 // To recognize min/max patterns formed by a icmp select sequence, we store
3891 // the number of instruction we saw from the recognized min/max pattern,
3892 // to make sure we only see exactly the two instructions.
3893 unsigned NumCmpSelectPatternInst = 0;
3894 ReductionInstDesc ReduxDesc(false, 0);
3896 SmallPtrSet<Instruction *, 8> VisitedInsts;
3897 SmallVector<Instruction *, 8> Worklist;
3898 Worklist.push_back(Phi);
3899 VisitedInsts.insert(Phi);
3901 // A value in the reduction can be used:
3902 // - By the reduction:
3903 // - Reduction operation:
3904 // - One use of reduction value (safe).
3905 // - Multiple use of reduction value (not safe).
3907 // - All uses of the PHI must be the reduction (safe).
3908 // - Otherwise, not safe.
3909 // - By one instruction outside of the loop (safe).
3910 // - By further instructions outside of the loop (not safe).
3911 // - By an instruction that is not part of the reduction (not safe).
3913 // * An instruction type other than PHI or the reduction operation.
3914 // * A PHI in the header other than the initial PHI.
3915 while (!Worklist.empty()) {
3916 Instruction *Cur = Worklist.back();
3917 Worklist.pop_back();
3920 // If the instruction has no users then this is a broken chain and can't be
3921 // a reduction variable.
3922 if (Cur->use_empty())
3925 bool IsAPhi = isa<PHINode>(Cur);
3927 // A header PHI use other than the original PHI.
3928 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3931 // Reductions of instructions such as Div, and Sub is only possible if the
3932 // LHS is the reduction variable.
3933 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3934 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3935 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3938 // Any reduction instruction must be of one of the allowed kinds.
3939 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3940 if (!ReduxDesc.IsReduction)
3943 // A reduction operation must only have one use of the reduction value.
3944 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3945 hasMultipleUsesOf(Cur, VisitedInsts))
3948 // All inputs to a PHI node must be a reduction value.
3949 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3952 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3953 isa<SelectInst>(Cur)))
3954 ++NumCmpSelectPatternInst;
3955 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3956 isa<SelectInst>(Cur)))
3957 ++NumCmpSelectPatternInst;
3959 // Check whether we found a reduction operator.
3960 FoundReduxOp |= !IsAPhi;
3962 // Process users of current instruction. Push non PHI nodes after PHI nodes
3963 // onto the stack. This way we are going to have seen all inputs to PHI
3964 // nodes once we get to them.
3965 SmallVector<Instruction *, 8> NonPHIs;
3966 SmallVector<Instruction *, 8> PHIs;
3967 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3969 Instruction *Usr = cast<Instruction>(*UI);
3971 // Check if we found the exit user.
3972 BasicBlock *Parent = Usr->getParent();
3973 if (!TheLoop->contains(Parent)) {
3974 // Exit if you find multiple outside users or if the header phi node is
3975 // being used. In this case the user uses the value of the previous
3976 // iteration, in which case we would loose "VF-1" iterations of the
3977 // reduction operation if we vectorize.
3978 if (ExitInstruction != 0 || Cur == Phi)
3981 ExitInstruction = Cur;
3985 // Process instructions only once (termination).
3986 if (VisitedInsts.insert(Usr)) {
3987 if (isa<PHINode>(Usr))
3988 PHIs.push_back(Usr);
3990 NonPHIs.push_back(Usr);
3992 // Remember that we completed the cycle.
3994 FoundStartPHI = true;
3996 Worklist.append(PHIs.begin(), PHIs.end());
3997 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4000 // This means we have seen one but not the other instruction of the
4001 // pattern or more than just a select and cmp.
4002 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4003 NumCmpSelectPatternInst != 2)
4006 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4009 // We found a reduction var if we have reached the original phi node and we
4010 // only have a single instruction with out-of-loop users.
4012 // This instruction is allowed to have out-of-loop users.
4013 AllowedExit.insert(ExitInstruction);
4015 // Save the description of this reduction variable.
4016 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4017 ReduxDesc.MinMaxKind);
4018 Reductions[Phi] = RD;
4019 // We've ended the cycle. This is a reduction variable if we have an
4020 // outside user and it has a binary op.
4025 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4026 /// pattern corresponding to a min(X, Y) or max(X, Y).
4027 LoopVectorizationLegality::ReductionInstDesc
4028 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4029 ReductionInstDesc &Prev) {
4031 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4032 "Expect a select instruction");
4033 Instruction *Cmp = 0;
4034 SelectInst *Select = 0;
4036 // We must handle the select(cmp()) as a single instruction. Advance to the
4038 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4039 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4040 return ReductionInstDesc(false, I);
4041 return ReductionInstDesc(Select, Prev.MinMaxKind);
4044 // Only handle single use cases for now.
4045 if (!(Select = dyn_cast<SelectInst>(I)))
4046 return ReductionInstDesc(false, I);
4047 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4048 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4049 return ReductionInstDesc(false, I);
4050 if (!Cmp->hasOneUse())
4051 return ReductionInstDesc(false, I);
4056 // Look for a min/max pattern.
4057 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4058 return ReductionInstDesc(Select, MRK_UIntMin);
4059 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4060 return ReductionInstDesc(Select, MRK_UIntMax);
4061 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4062 return ReductionInstDesc(Select, MRK_SIntMax);
4063 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4064 return ReductionInstDesc(Select, MRK_SIntMin);
4065 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4066 return ReductionInstDesc(Select, MRK_FloatMin);
4067 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4068 return ReductionInstDesc(Select, MRK_FloatMax);
4069 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4070 return ReductionInstDesc(Select, MRK_FloatMin);
4071 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4072 return ReductionInstDesc(Select, MRK_FloatMax);
4074 return ReductionInstDesc(false, I);
4077 LoopVectorizationLegality::ReductionInstDesc
4078 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4080 ReductionInstDesc &Prev) {
4081 bool FP = I->getType()->isFloatingPointTy();
4082 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4083 switch (I->getOpcode()) {
4085 return ReductionInstDesc(false, I);
4086 case Instruction::PHI:
4087 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4088 Kind != RK_FloatMinMax))
4089 return ReductionInstDesc(false, I);
4090 return ReductionInstDesc(I, Prev.MinMaxKind);
4091 case Instruction::Sub:
4092 case Instruction::Add:
4093 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4094 case Instruction::Mul:
4095 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4096 case Instruction::And:
4097 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4098 case Instruction::Or:
4099 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4100 case Instruction::Xor:
4101 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4102 case Instruction::FMul:
4103 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4104 case Instruction::FAdd:
4105 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4106 case Instruction::FCmp:
4107 case Instruction::ICmp:
4108 case Instruction::Select:
4109 if (Kind != RK_IntegerMinMax &&
4110 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4111 return ReductionInstDesc(false, I);
4112 return isMinMaxSelectCmpPattern(I, Prev);
4116 LoopVectorizationLegality::InductionKind
4117 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4118 Type *PhiTy = Phi->getType();
4119 // We only handle integer and pointer inductions variables.
4120 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4121 return IK_NoInduction;
4123 // Check that the PHI is consecutive.
4124 const SCEV *PhiScev = SE->getSCEV(Phi);
4125 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4127 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4128 return IK_NoInduction;
4130 const SCEV *Step = AR->getStepRecurrence(*SE);
4132 // Integer inductions need to have a stride of one.
4133 if (PhiTy->isIntegerTy()) {
4135 return IK_IntInduction;
4136 if (Step->isAllOnesValue())
4137 return IK_ReverseIntInduction;
4138 return IK_NoInduction;
4141 // Calculate the pointer stride and check if it is consecutive.
4142 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4144 return IK_NoInduction;
4146 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4147 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4148 if (C->getValue()->equalsInt(Size))
4149 return IK_PtrInduction;
4150 else if (C->getValue()->equalsInt(0 - Size))
4151 return IK_ReversePtrInduction;
4153 return IK_NoInduction;
4156 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4157 Value *In0 = const_cast<Value*>(V);
4158 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4162 return Inductions.count(PN);
4165 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4166 assert(TheLoop->contains(BB) && "Unknown block used");
4168 // Blocks that do not dominate the latch need predication.
4169 BasicBlock* Latch = TheLoop->getLoopLatch();
4170 return !DT->dominates(BB, Latch);
4173 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4174 SmallPtrSet<Value *, 8>& SafePtrs) {
4175 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4176 // We might be able to hoist the load.
4177 if (it->mayReadFromMemory()) {
4178 LoadInst *LI = dyn_cast<LoadInst>(it);
4179 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4183 // We don't predicate stores at the moment.
4184 if (it->mayWriteToMemory() || it->mayThrow())
4187 // The instructions below can trap.
4188 switch (it->getOpcode()) {
4190 case Instruction::UDiv:
4191 case Instruction::SDiv:
4192 case Instruction::URem:
4193 case Instruction::SRem:
4201 LoopVectorizationCostModel::VectorizationFactor
4202 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4204 // Width 1 means no vectorize
4205 VectorizationFactor Factor = { 1U, 0U };
4206 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4207 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4211 // Find the trip count.
4212 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4213 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4215 unsigned WidestType = getWidestType();
4216 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4217 unsigned MaxSafeDepDist = -1U;
4218 if (Legal->getMaxSafeDepDistBytes() != -1U)
4219 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4220 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4221 WidestRegister : MaxSafeDepDist);
4222 unsigned MaxVectorSize = WidestRegister / WidestType;
4223 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4224 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4226 if (MaxVectorSize == 0) {
4227 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4231 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4232 " into one vector!");
4234 unsigned VF = MaxVectorSize;
4236 // If we optimize the program for size, avoid creating the tail loop.
4238 // If we are unable to calculate the trip count then don't try to vectorize.
4240 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4244 // Find the maximum SIMD width that can fit within the trip count.
4245 VF = TC % MaxVectorSize;
4250 // If the trip count that we found modulo the vectorization factor is not
4251 // zero then we require a tail.
4253 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4259 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4260 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4262 Factor.Width = UserVF;
4266 float Cost = expectedCost(1);
4268 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4269 for (unsigned i=2; i <= VF; i*=2) {
4270 // Notice that the vector loop needs to be executed less times, so
4271 // we need to divide the cost of the vector loops by the width of
4272 // the vector elements.
4273 float VectorCost = expectedCost(i) / (float)i;
4274 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4275 (int)VectorCost << ".\n");
4276 if (VectorCost < Cost) {
4282 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4283 Factor.Width = Width;
4284 Factor.Cost = Width * Cost;
4288 unsigned LoopVectorizationCostModel::getWidestType() {
4289 unsigned MaxWidth = 8;
4292 for (Loop::block_iterator bb = TheLoop->block_begin(),
4293 be = TheLoop->block_end(); bb != be; ++bb) {
4294 BasicBlock *BB = *bb;
4296 // For each instruction in the loop.
4297 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4298 Type *T = it->getType();
4300 // Only examine Loads, Stores and PHINodes.
4301 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4304 // Examine PHI nodes that are reduction variables.
4305 if (PHINode *PN = dyn_cast<PHINode>(it))
4306 if (!Legal->getReductionVars()->count(PN))
4309 // Examine the stored values.
4310 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4311 T = ST->getValueOperand()->getType();
4313 // Ignore loaded pointer types and stored pointer types that are not
4314 // consecutive. However, we do want to take consecutive stores/loads of
4315 // pointer vectors into account.
4316 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4319 MaxWidth = std::max(MaxWidth,
4320 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4328 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4331 unsigned LoopCost) {
4333 // -- The unroll heuristics --
4334 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4335 // There are many micro-architectural considerations that we can't predict
4336 // at this level. For example frontend pressure (on decode or fetch) due to
4337 // code size, or the number and capabilities of the execution ports.
4339 // We use the following heuristics to select the unroll factor:
4340 // 1. If the code has reductions the we unroll in order to break the cross
4341 // iteration dependency.
4342 // 2. If the loop is really small then we unroll in order to reduce the loop
4344 // 3. We don't unroll if we think that we will spill registers to memory due
4345 // to the increased register pressure.
4347 // Use the user preference, unless 'auto' is selected.
4351 // When we optimize for size we don't unroll.
4355 // We used the distance for the unroll factor.
4356 if (Legal->getMaxSafeDepDistBytes() != -1U)
4359 // Do not unroll loops with a relatively small trip count.
4360 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4361 TheLoop->getLoopLatch());
4362 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4365 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4366 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4367 " vector registers\n");
4369 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4370 // We divide by these constants so assume that we have at least one
4371 // instruction that uses at least one register.
4372 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4373 R.NumInstructions = std::max(R.NumInstructions, 1U);
4375 // We calculate the unroll factor using the following formula.
4376 // Subtract the number of loop invariants from the number of available
4377 // registers. These registers are used by all of the unrolled instances.
4378 // Next, divide the remaining registers by the number of registers that is
4379 // required by the loop, in order to estimate how many parallel instances
4380 // fit without causing spills.
4381 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4383 // Clamp the unroll factor ranges to reasonable factors.
4384 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4386 // If we did not calculate the cost for VF (because the user selected the VF)
4387 // then we calculate the cost of VF here.
4389 LoopCost = expectedCost(VF);
4391 // Clamp the calculated UF to be between the 1 and the max unroll factor
4392 // that the target allows.
4393 if (UF > MaxUnrollSize)
4398 bool HasReductions = Legal->getReductionVars()->size();
4400 // Decide if we want to unroll if we decided that it is legal to vectorize
4401 // but not profitable.
4403 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4404 LoopCost > SmallLoopCost)
4410 if (HasReductions) {
4411 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4415 // We want to unroll tiny loops in order to reduce the loop overhead.
4416 // We assume that the cost overhead is 1 and we use the cost model
4417 // to estimate the cost of the loop and unroll until the cost of the
4418 // loop overhead is about 5% of the cost of the loop.
4419 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4420 if (LoopCost < SmallLoopCost) {
4421 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4422 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4423 return std::min(NewUF, UF);
4426 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4430 LoopVectorizationCostModel::RegisterUsage
4431 LoopVectorizationCostModel::calculateRegisterUsage() {
4432 // This function calculates the register usage by measuring the highest number
4433 // of values that are alive at a single location. Obviously, this is a very
4434 // rough estimation. We scan the loop in a topological order in order and
4435 // assign a number to each instruction. We use RPO to ensure that defs are
4436 // met before their users. We assume that each instruction that has in-loop
4437 // users starts an interval. We record every time that an in-loop value is
4438 // used, so we have a list of the first and last occurrences of each
4439 // instruction. Next, we transpose this data structure into a multi map that
4440 // holds the list of intervals that *end* at a specific location. This multi
4441 // map allows us to perform a linear search. We scan the instructions linearly
4442 // and record each time that a new interval starts, by placing it in a set.
4443 // If we find this value in the multi-map then we remove it from the set.
4444 // The max register usage is the maximum size of the set.
4445 // We also search for instructions that are defined outside the loop, but are
4446 // used inside the loop. We need this number separately from the max-interval
4447 // usage number because when we unroll, loop-invariant values do not take
4449 LoopBlocksDFS DFS(TheLoop);
4453 R.NumInstructions = 0;
4455 // Each 'key' in the map opens a new interval. The values
4456 // of the map are the index of the 'last seen' usage of the
4457 // instruction that is the key.
4458 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4459 // Maps instruction to its index.
4460 DenseMap<unsigned, Instruction*> IdxToInstr;
4461 // Marks the end of each interval.
4462 IntervalMap EndPoint;
4463 // Saves the list of instruction indices that are used in the loop.
4464 SmallSet<Instruction*, 8> Ends;
4465 // Saves the list of values that are used in the loop but are
4466 // defined outside the loop, such as arguments and constants.
4467 SmallPtrSet<Value*, 8> LoopInvariants;
4470 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4471 be = DFS.endRPO(); bb != be; ++bb) {
4472 R.NumInstructions += (*bb)->size();
4473 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4475 Instruction *I = it;
4476 IdxToInstr[Index++] = I;
4478 // Save the end location of each USE.
4479 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4480 Value *U = I->getOperand(i);
4481 Instruction *Instr = dyn_cast<Instruction>(U);
4483 // Ignore non-instruction values such as arguments, constants, etc.
4484 if (!Instr) continue;
4486 // If this instruction is outside the loop then record it and continue.
4487 if (!TheLoop->contains(Instr)) {
4488 LoopInvariants.insert(Instr);
4492 // Overwrite previous end points.
4493 EndPoint[Instr] = Index;
4499 // Saves the list of intervals that end with the index in 'key'.
4500 typedef SmallVector<Instruction*, 2> InstrList;
4501 DenseMap<unsigned, InstrList> TransposeEnds;
4503 // Transpose the EndPoints to a list of values that end at each index.
4504 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4506 TransposeEnds[it->second].push_back(it->first);
4508 SmallSet<Instruction*, 8> OpenIntervals;
4509 unsigned MaxUsage = 0;
4512 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4513 for (unsigned int i = 0; i < Index; ++i) {
4514 Instruction *I = IdxToInstr[i];
4515 // Ignore instructions that are never used within the loop.
4516 if (!Ends.count(I)) continue;
4518 // Remove all of the instructions that end at this location.
4519 InstrList &List = TransposeEnds[i];
4520 for (unsigned int j=0, e = List.size(); j < e; ++j)
4521 OpenIntervals.erase(List[j]);
4523 // Count the number of live interals.
4524 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4526 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4527 OpenIntervals.size() <<"\n");
4529 // Add the current instruction to the list of open intervals.
4530 OpenIntervals.insert(I);
4533 unsigned Invariant = LoopInvariants.size();
4534 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4535 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4536 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4538 R.LoopInvariantRegs = Invariant;
4539 R.MaxLocalUsers = MaxUsage;
4543 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4547 for (Loop::block_iterator bb = TheLoop->block_begin(),
4548 be = TheLoop->block_end(); bb != be; ++bb) {
4549 unsigned BlockCost = 0;
4550 BasicBlock *BB = *bb;
4552 // For each instruction in the old loop.
4553 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4554 // Skip dbg intrinsics.
4555 if (isa<DbgInfoIntrinsic>(it))
4558 unsigned C = getInstructionCost(it, VF);
4560 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4561 VF << " For instruction: "<< *it << "\n");
4564 // We assume that if-converted blocks have a 50% chance of being executed.
4565 // When the code is scalar then some of the blocks are avoided due to CF.
4566 // When the code is vectorized we execute all code paths.
4567 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4576 /// \brief Check whether the address computation for a non-consecutive memory
4577 /// access looks like an unlikely candidate for being merged into the indexing
4580 /// We look for a GEP which has one index that is an induction variable and all
4581 /// other indices are loop invariant. If the stride of this access is also
4582 /// within a small bound we decide that this address computation can likely be
4583 /// merged into the addressing mode.
4584 /// In all other cases, we identify the address computation as complex.
4585 static bool isLikelyComplexAddressComputation(Value *Ptr,
4586 LoopVectorizationLegality *Legal,
4587 ScalarEvolution *SE,
4588 const Loop *TheLoop) {
4589 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4593 // We are looking for a gep with all loop invariant indices except for one
4594 // which should be an induction variable.
4595 unsigned NumOperands = Gep->getNumOperands();
4596 for (unsigned i = 1; i < NumOperands; ++i) {
4597 Value *Opd = Gep->getOperand(i);
4598 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4599 !Legal->isInductionVariable(Opd))
4603 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4604 // can likely be merged into the address computation.
4605 unsigned MaxMergeDistance = 64;
4607 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4611 // Check the step is constant.
4612 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4613 // Calculate the pointer stride and check if it is consecutive.
4614 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4618 const APInt &APStepVal = C->getValue()->getValue();
4620 // Huge step value - give up.
4621 if (APStepVal.getBitWidth() > 64)
4624 int64_t StepVal = APStepVal.getSExtValue();
4626 return StepVal > MaxMergeDistance;
4630 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4631 // If we know that this instruction will remain uniform, check the cost of
4632 // the scalar version.
4633 if (Legal->isUniformAfterVectorization(I))
4636 Type *RetTy = I->getType();
4637 Type *VectorTy = ToVectorTy(RetTy, VF);
4639 // TODO: We need to estimate the cost of intrinsic calls.
4640 switch (I->getOpcode()) {
4641 case Instruction::GetElementPtr:
4642 // We mark this instruction as zero-cost because the cost of GEPs in
4643 // vectorized code depends on whether the corresponding memory instruction
4644 // is scalarized or not. Therefore, we handle GEPs with the memory
4645 // instruction cost.
4647 case Instruction::Br: {
4648 return TTI.getCFInstrCost(I->getOpcode());
4650 case Instruction::PHI:
4651 //TODO: IF-converted IFs become selects.
4653 case Instruction::Add:
4654 case Instruction::FAdd:
4655 case Instruction::Sub:
4656 case Instruction::FSub:
4657 case Instruction::Mul:
4658 case Instruction::FMul:
4659 case Instruction::UDiv:
4660 case Instruction::SDiv:
4661 case Instruction::FDiv:
4662 case Instruction::URem:
4663 case Instruction::SRem:
4664 case Instruction::FRem:
4665 case Instruction::Shl:
4666 case Instruction::LShr:
4667 case Instruction::AShr:
4668 case Instruction::And:
4669 case Instruction::Or:
4670 case Instruction::Xor: {
4671 // Certain instructions can be cheaper to vectorize if they have a constant
4672 // second vector operand. One example of this are shifts on x86.
4673 TargetTransformInfo::OperandValueKind Op1VK =
4674 TargetTransformInfo::OK_AnyValue;
4675 TargetTransformInfo::OperandValueKind Op2VK =
4676 TargetTransformInfo::OK_AnyValue;
4678 if (isa<ConstantInt>(I->getOperand(1)))
4679 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4681 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4683 case Instruction::Select: {
4684 SelectInst *SI = cast<SelectInst>(I);
4685 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4686 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4687 Type *CondTy = SI->getCondition()->getType();
4689 CondTy = VectorType::get(CondTy, VF);
4691 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4693 case Instruction::ICmp:
4694 case Instruction::FCmp: {
4695 Type *ValTy = I->getOperand(0)->getType();
4696 VectorTy = ToVectorTy(ValTy, VF);
4697 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4699 case Instruction::Store:
4700 case Instruction::Load: {
4701 StoreInst *SI = dyn_cast<StoreInst>(I);
4702 LoadInst *LI = dyn_cast<LoadInst>(I);
4703 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4705 VectorTy = ToVectorTy(ValTy, VF);
4707 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4708 unsigned AS = SI ? SI->getPointerAddressSpace() :
4709 LI->getPointerAddressSpace();
4710 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4711 // We add the cost of address computation here instead of with the gep
4712 // instruction because only here we know whether the operation is
4715 return TTI.getAddressComputationCost(VectorTy) +
4716 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4718 // Scalarized loads/stores.
4719 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4720 bool Reverse = ConsecutiveStride < 0;
4721 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4722 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4723 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4724 bool IsComplexComputation =
4725 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4727 // The cost of extracting from the value vector and pointer vector.
4728 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4729 for (unsigned i = 0; i < VF; ++i) {
4730 // The cost of extracting the pointer operand.
4731 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4732 // In case of STORE, the cost of ExtractElement from the vector.
4733 // In case of LOAD, the cost of InsertElement into the returned
4735 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4736 Instruction::InsertElement,
4740 // The cost of the scalar loads/stores.
4741 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4742 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4747 // Wide load/stores.
4748 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4749 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4752 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4756 case Instruction::ZExt:
4757 case Instruction::SExt:
4758 case Instruction::FPToUI:
4759 case Instruction::FPToSI:
4760 case Instruction::FPExt:
4761 case Instruction::PtrToInt:
4762 case Instruction::IntToPtr:
4763 case Instruction::SIToFP:
4764 case Instruction::UIToFP:
4765 case Instruction::Trunc:
4766 case Instruction::FPTrunc:
4767 case Instruction::BitCast: {
4768 // We optimize the truncation of induction variable.
4769 // The cost of these is the same as the scalar operation.
4770 if (I->getOpcode() == Instruction::Trunc &&
4771 Legal->isInductionVariable(I->getOperand(0)))
4772 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4773 I->getOperand(0)->getType());
4775 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4776 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4778 case Instruction::Call: {
4779 CallInst *CI = cast<CallInst>(I);
4780 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4781 assert(ID && "Not an intrinsic call!");
4782 Type *RetTy = ToVectorTy(CI->getType(), VF);
4783 SmallVector<Type*, 4> Tys;
4784 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4785 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4786 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4789 // We are scalarizing the instruction. Return the cost of the scalar
4790 // instruction, plus the cost of insert and extract into vector
4791 // elements, times the vector width.
4794 if (!RetTy->isVoidTy() && VF != 1) {
4795 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4797 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4800 // The cost of inserting the results plus extracting each one of the
4802 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4805 // The cost of executing VF copies of the scalar instruction. This opcode
4806 // is unknown. Assume that it is the same as 'mul'.
4807 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4813 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4814 if (Scalar->isVoidTy() || VF == 1)
4816 return VectorType::get(Scalar, VF);
4819 char LoopVectorize::ID = 0;
4820 static const char lv_name[] = "Loop Vectorization";
4821 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4822 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4823 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4824 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4825 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4828 Pass *createLoopVectorizePass(bool NoUnrolling) {
4829 return new LoopVectorize(NoUnrolling);
4833 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4834 // Check for a store.
4835 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4836 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4838 // Check for a load.
4839 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4840 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
4846 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
4847 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
4848 // Holds vector parameters or scalars, in case of uniform vals.
4849 SmallVector<VectorParts, 4> Params;
4851 setDebugLocFromInst(Builder, Instr);
4853 // Find all of the vectorized parameters.
4854 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
4855 Value *SrcOp = Instr->getOperand(op);
4857 // If we are accessing the old induction variable, use the new one.
4858 if (SrcOp == OldInduction) {
4859 Params.push_back(getVectorValue(SrcOp));
4863 // Try using previously calculated values.
4864 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
4866 // If the src is an instruction that appeared earlier in the basic block
4867 // then it should already be vectorized.
4868 if (SrcInst && OrigLoop->contains(SrcInst)) {
4869 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
4870 // The parameter is a vector value from earlier.
4871 Params.push_back(WidenMap.get(SrcInst));
4873 // The parameter is a scalar from outside the loop. Maybe even a constant.
4874 VectorParts Scalars;
4875 Scalars.append(UF, SrcOp);
4876 Params.push_back(Scalars);
4880 assert(Params.size() == Instr->getNumOperands() &&
4881 "Invalid number of operands");
4883 // Does this instruction return a value ?
4884 bool IsVoidRetTy = Instr->getType()->isVoidTy();
4886 Value *UndefVec = IsVoidRetTy ? 0 :
4887 UndefValue::get(Instr->getType());
4888 // Create a new entry in the WidenMap and initialize it to Undef or Null.
4889 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
4891 // For each vector unroll 'part':
4892 for (unsigned Part = 0; Part < UF; ++Part) {
4893 // For each scalar that we create:
4895 Instruction *Cloned = Instr->clone();
4897 Cloned->setName(Instr->getName() + ".cloned");
4898 // Replace the operands of the cloned instructions with extracted scalars.
4899 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
4900 Value *Op = Params[op][Part];
4901 Cloned->setOperand(op, Op);
4904 // Place the cloned scalar in the new loop.
4905 Builder.Insert(Cloned);
4907 // If the original scalar returns a value we need to place it in a vector
4908 // so that future users will be able to use it.
4910 VecResults[Part] = Cloned;
4915 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
4916 LoopVectorizationLegality*) {
4917 return scalarizeInstruction(Instr);
4920 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
4924 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
4928 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
4930 // When unrolling and the VF is 1, we only need to add a simple scalar.
4931 Type *ITy = Val->getType();
4932 assert(!ITy->isVectorTy() && "Val must be a scalar");
4933 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
4934 return Builder.CreateAdd(Val, C, "induction");