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/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/StringExtras.h"
58 #include "llvm/Analysis/AliasAnalysis.h"
59 #include "llvm/Analysis/Dominators.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/Analysis/Verifier.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DerivedTypes.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.h"
80 #include "llvm/Pass.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/Debug.h"
83 #include "llvm/Support/PatternMatch.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/Support/ValueHandle.h"
86 #include "llvm/Target/TargetLibraryInfo.h"
87 #include "llvm/Transforms/Scalar.h"
88 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
89 #include "llvm/Transforms/Utils/Local.h"
94 using namespace llvm::PatternMatch;
96 static cl::opt<unsigned>
97 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
98 cl::desc("Sets the SIMD width. Zero is autoselect."));
100 static cl::opt<unsigned>
101 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
102 cl::desc("Sets the vectorization unroll count. "
103 "Zero is autoselect."));
106 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
107 cl::desc("Enable if-conversion during vectorization."));
109 /// We don't vectorize loops with a known constant trip count below this number.
110 static cl::opt<unsigned>
111 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
113 cl::desc("Don't vectorize loops with a constant "
114 "trip count that is smaller than this "
117 /// We don't unroll loops with a known constant trip count below this number.
118 static const unsigned TinyTripCountUnrollThreshold = 128;
120 /// When performing memory disambiguation checks at runtime do not make more
121 /// than this number of comparisons.
122 static const unsigned RuntimeMemoryCheckThreshold = 8;
124 /// Maximum simd width.
125 static const unsigned MaxVectorWidth = 64;
127 /// Maximum vectorization unroll count.
128 static const unsigned MaxUnrollFactor = 16;
130 /// The cost of a loop that is considered 'small' by the unroller.
131 static const unsigned SmallLoopCost = 20;
135 // Forward declarations.
136 class LoopVectorizationLegality;
137 class LoopVectorizationCostModel;
139 /// InnerLoopVectorizer vectorizes loops which contain only one basic
140 /// block to a specified vectorization factor (VF).
141 /// This class performs the widening of scalars into vectors, or multiple
142 /// scalars. This class also implements the following features:
143 /// * It inserts an epilogue loop for handling loops that don't have iteration
144 /// counts that are known to be a multiple of the vectorization factor.
145 /// * It handles the code generation for reduction variables.
146 /// * Scalarization (implementation using scalars) of un-vectorizable
148 /// InnerLoopVectorizer does not perform any vectorization-legality
149 /// checks, and relies on the caller to check for the different legality
150 /// aspects. The InnerLoopVectorizer relies on the
151 /// LoopVectorizationLegality class to provide information about the induction
152 /// and reduction variables that were found to a given vectorization factor.
153 class InnerLoopVectorizer {
155 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
156 DominatorTree *DT, DataLayout *DL,
157 const TargetLibraryInfo *TLI, unsigned VecWidth,
158 unsigned UnrollFactor)
159 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
160 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
161 OldInduction(0), WidenMap(UnrollFactor) {}
163 // Perform the actual loop widening (vectorization).
164 void vectorize(LoopVectorizationLegality *Legal) {
165 // Create a new empty loop. Unlink the old loop and connect the new one.
166 createEmptyLoop(Legal);
167 // Widen each instruction in the old loop to a new one in the new loop.
168 // Use the Legality module to find the induction and reduction variables.
169 vectorizeLoop(Legal);
170 // Register the new loop and update the analysis passes.
174 virtual ~InnerLoopVectorizer() {}
177 /// A small list of PHINodes.
178 typedef SmallVector<PHINode*, 4> PhiVector;
179 /// When we unroll loops we have multiple vector values for each scalar.
180 /// This data structure holds the unrolled and vectorized values that
181 /// originated from one scalar instruction.
182 typedef SmallVector<Value*, 2> VectorParts;
184 // When we if-convert we need create edge masks. We have to cache values so
185 // that we don't end up with exponential recursion/IR.
186 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
187 VectorParts> EdgeMaskCache;
189 /// Add code that checks at runtime if the accessed arrays overlap.
190 /// Returns the comparator value or NULL if no check is needed.
191 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
193 /// Create an empty loop, based on the loop ranges of the old loop.
194 void createEmptyLoop(LoopVectorizationLegality *Legal);
195 /// Copy and widen the instructions from the old loop.
196 virtual void vectorizeLoop(LoopVectorizationLegality *Legal);
198 /// \brief The Loop exit block may have single value PHI nodes where the
199 /// incoming value is 'Undef'. While vectorizing we only handled real values
200 /// that were defined inside the loop. Here we fix the 'undef case'.
204 /// A helper function that computes the predicate of the block BB, assuming
205 /// that the header block of the loop is set to True. It returns the *entry*
206 /// mask for the block BB.
207 VectorParts createBlockInMask(BasicBlock *BB);
208 /// A helper function that computes the predicate of the edge between SRC
210 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
212 /// A helper function to vectorize a single BB within the innermost loop.
213 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
216 /// Vectorize a single PHINode in a block. This method handles the induction
217 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
218 /// arbitrary length vectors.
219 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
220 LoopVectorizationLegality *Legal,
221 unsigned UF, unsigned VF, PhiVector *PV);
223 /// Insert the new loop to the loop hierarchy and pass manager
224 /// and update the analysis passes.
225 void updateAnalysis();
227 /// This instruction is un-vectorizable. Implement it as a sequence
229 virtual void scalarizeInstruction(Instruction *Instr);
231 /// Vectorize Load and Store instructions,
232 virtual void vectorizeMemoryInstruction(Instruction *Instr,
233 LoopVectorizationLegality *Legal);
235 /// Create a broadcast instruction. This method generates a broadcast
236 /// instruction (shuffle) for loop invariant values and for the induction
237 /// value. If this is the induction variable then we extend it to N, N+1, ...
238 /// this is needed because each iteration in the loop corresponds to a SIMD
240 virtual Value *getBroadcastInstrs(Value *V);
242 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
243 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
244 /// The sequence starts at StartIndex.
245 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
247 /// When we go over instructions in the basic block we rely on previous
248 /// values within the current basic block or on loop invariant values.
249 /// When we widen (vectorize) values we place them in the map. If the values
250 /// are not within the map, they have to be loop invariant, so we simply
251 /// broadcast them into a vector.
252 VectorParts &getVectorValue(Value *V);
254 /// Generate a shuffle sequence that will reverse the vector Vec.
255 virtual Value *reverseVector(Value *Vec);
257 /// This is a helper class that holds the vectorizer state. It maps scalar
258 /// instructions to vector instructions. When the code is 'unrolled' then
259 /// then a single scalar value is mapped to multiple vector parts. The parts
260 /// are stored in the VectorPart type.
262 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
264 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
266 /// \return True if 'Key' is saved in the Value Map.
267 bool has(Value *Key) const { return MapStorage.count(Key); }
269 /// Initializes a new entry in the map. Sets all of the vector parts to the
270 /// save value in 'Val'.
271 /// \return A reference to a vector with splat values.
272 VectorParts &splat(Value *Key, Value *Val) {
273 VectorParts &Entry = MapStorage[Key];
274 Entry.assign(UF, Val);
278 ///\return A reference to the value that is stored at 'Key'.
279 VectorParts &get(Value *Key) {
280 VectorParts &Entry = MapStorage[Key];
283 assert(Entry.size() == UF);
288 /// The unroll factor. Each entry in the map stores this number of vector
292 /// Map storage. We use std::map and not DenseMap because insertions to a
293 /// dense map invalidates its iterators.
294 std::map<Value *, VectorParts> MapStorage;
297 /// The original loop.
299 /// Scev analysis to use.
307 /// Target Library Info.
308 const TargetLibraryInfo *TLI;
310 /// The vectorization SIMD factor to use. Each vector will have this many
315 /// The vectorization unroll factor to use. Each scalar is vectorized to this
316 /// many different vector instructions.
319 /// The builder that we use
322 // --- Vectorization state ---
324 /// The vector-loop preheader.
325 BasicBlock *LoopVectorPreHeader;
326 /// The scalar-loop preheader.
327 BasicBlock *LoopScalarPreHeader;
328 /// Middle Block between the vector and the scalar.
329 BasicBlock *LoopMiddleBlock;
330 ///The ExitBlock of the scalar loop.
331 BasicBlock *LoopExitBlock;
332 ///The vector loop body.
333 BasicBlock *LoopVectorBody;
334 ///The scalar loop body.
335 BasicBlock *LoopScalarBody;
336 /// A list of all bypass blocks. The first block is the entry of the loop.
337 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
339 /// The new Induction variable which was added to the new block.
341 /// The induction variable of the old basic block.
342 PHINode *OldInduction;
343 /// Holds the extended (to the widest induction type) start index.
345 /// Maps scalars to widened vectors.
347 EdgeMaskCache MaskCache;
350 class InnerLoopUnroller : public InnerLoopVectorizer {
352 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
353 DominatorTree *DT, DataLayout *DL,
354 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
355 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
358 virtual void scalarizeInstruction(Instruction *Instr);
359 virtual void vectorizeMemoryInstruction(Instruction *Instr,
360 LoopVectorizationLegality *Legal);
361 virtual Value *getBroadcastInstrs(Value *V);
362 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
363 virtual Value *reverseVector(Value *Vec);
366 /// \brief Look for a meaningful debug location on the instruction or it's
368 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
373 if (I->getDebugLoc() != Empty)
376 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
377 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
378 if (OpInst->getDebugLoc() != Empty)
385 /// \brief Set the debug location in the builder using the debug location in the
387 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
388 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
389 B.SetCurrentDebugLocation(Inst->getDebugLoc());
391 B.SetCurrentDebugLocation(DebugLoc());
394 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
395 /// to what vectorization factor.
396 /// This class does not look at the profitability of vectorization, only the
397 /// legality. This class has two main kinds of checks:
398 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
399 /// will change the order of memory accesses in a way that will change the
400 /// correctness of the program.
401 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
402 /// checks for a number of different conditions, such as the availability of a
403 /// single induction variable, that all types are supported and vectorize-able,
404 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
405 /// This class is also used by InnerLoopVectorizer for identifying
406 /// induction variable and the different reduction variables.
407 class LoopVectorizationLegality {
409 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
410 DominatorTree *DT, TargetLibraryInfo *TLI)
411 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
412 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
413 MaxSafeDepDistBytes(-1U) {}
415 /// This enum represents the kinds of reductions that we support.
417 RK_NoReduction, ///< Not a reduction.
418 RK_IntegerAdd, ///< Sum of integers.
419 RK_IntegerMult, ///< Product of integers.
420 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
421 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
422 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
423 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
424 RK_FloatAdd, ///< Sum of floats.
425 RK_FloatMult, ///< Product of floats.
426 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
429 /// This enum represents the kinds of inductions that we support.
431 IK_NoInduction, ///< Not an induction variable.
432 IK_IntInduction, ///< Integer induction variable. Step = 1.
433 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
434 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
435 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
438 // This enum represents the kind of minmax reduction.
439 enum MinMaxReductionKind {
449 /// This struct holds information about reduction variables.
450 struct ReductionDescriptor {
451 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
452 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
454 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
455 MinMaxReductionKind MK)
456 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
458 // The starting value of the reduction.
459 // It does not have to be zero!
460 TrackingVH<Value> StartValue;
461 // The instruction who's value is used outside the loop.
462 Instruction *LoopExitInstr;
463 // The kind of the reduction.
465 // If this a min/max reduction the kind of reduction.
466 MinMaxReductionKind MinMaxKind;
469 /// This POD struct holds information about a potential reduction operation.
470 struct ReductionInstDesc {
471 ReductionInstDesc(bool IsRedux, Instruction *I) :
472 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
474 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
475 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
477 // Is this instruction a reduction candidate.
479 // The last instruction in a min/max pattern (select of the select(icmp())
480 // pattern), or the current reduction instruction otherwise.
481 Instruction *PatternLastInst;
482 // If this is a min/max pattern the comparison predicate.
483 MinMaxReductionKind MinMaxKind;
486 /// This struct holds information about the memory runtime legality
487 /// check that a group of pointers do not overlap.
488 struct RuntimePointerCheck {
489 RuntimePointerCheck() : Need(false) {}
491 /// Reset the state of the pointer runtime information.
498 DependencySetId.clear();
501 /// Insert a pointer and calculate the start and end SCEVs.
502 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
505 /// This flag indicates if we need to add the runtime check.
507 /// Holds the pointers that we need to check.
508 SmallVector<TrackingVH<Value>, 2> Pointers;
509 /// Holds the pointer value at the beginning of the loop.
510 SmallVector<const SCEV*, 2> Starts;
511 /// Holds the pointer value at the end of the loop.
512 SmallVector<const SCEV*, 2> Ends;
513 /// Holds the information if this pointer is used for writing to memory.
514 SmallVector<bool, 2> IsWritePtr;
515 /// Holds the id of the set of pointers that could be dependent because of a
516 /// shared underlying object.
517 SmallVector<unsigned, 2> DependencySetId;
520 /// A struct for saving information about induction variables.
521 struct InductionInfo {
522 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
523 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
525 TrackingVH<Value> StartValue;
530 /// ReductionList contains the reduction descriptors for all
531 /// of the reductions that were found in the loop.
532 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
534 /// InductionList saves induction variables and maps them to the
535 /// induction descriptor.
536 typedef MapVector<PHINode*, InductionInfo> InductionList;
538 /// Returns true if it is legal to vectorize this loop.
539 /// This does not mean that it is profitable to vectorize this
540 /// loop, only that it is legal to do so.
543 /// Returns the Induction variable.
544 PHINode *getInduction() { return Induction; }
546 /// Returns the reduction variables found in the loop.
547 ReductionList *getReductionVars() { return &Reductions; }
549 /// Returns the induction variables found in the loop.
550 InductionList *getInductionVars() { return &Inductions; }
552 /// Returns the widest induction type.
553 Type *getWidestInductionType() { return WidestIndTy; }
555 /// Returns True if V is an induction variable in this loop.
556 bool isInductionVariable(const Value *V);
558 /// Return true if the block BB needs to be predicated in order for the loop
559 /// to be vectorized.
560 bool blockNeedsPredication(BasicBlock *BB);
562 /// Check if this pointer is consecutive when vectorizing. This happens
563 /// when the last index of the GEP is the induction variable, or that the
564 /// pointer itself is an induction variable.
565 /// This check allows us to vectorize A[idx] into a wide load/store.
567 /// 0 - Stride is unknown or non-consecutive.
568 /// 1 - Address is consecutive.
569 /// -1 - Address is consecutive, and decreasing.
570 int isConsecutivePtr(Value *Ptr);
572 /// Returns true if the value V is uniform within the loop.
573 bool isUniform(Value *V);
575 /// Returns true if this instruction will remain scalar after vectorization.
576 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
578 /// Returns the information that we collected about runtime memory check.
579 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
581 /// This function returns the identity element (or neutral element) for
583 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
585 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
588 /// Check if a single basic block loop is vectorizable.
589 /// At this point we know that this is a loop with a constant trip count
590 /// and we only need to check individual instructions.
591 bool canVectorizeInstrs();
593 /// When we vectorize loops we may change the order in which
594 /// we read and write from memory. This method checks if it is
595 /// legal to vectorize the code, considering only memory constrains.
596 /// Returns true if the loop is vectorizable
597 bool canVectorizeMemory();
599 /// Return true if we can vectorize this loop using the IF-conversion
601 bool canVectorizeWithIfConvert();
603 /// Collect the variables that need to stay uniform after vectorization.
604 void collectLoopUniforms();
606 /// Return true if all of the instructions in the block can be speculatively
607 /// executed. \p SafePtrs is a list of addresses that are known to be legal
608 /// and we know that we can read from them without segfault.
609 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
611 /// Returns True, if 'Phi' is the kind of reduction variable for type
612 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
613 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
614 /// Returns a struct describing if the instruction 'I' can be a reduction
615 /// variable of type 'Kind'. If the reduction is a min/max pattern of
616 /// select(icmp()) this function advances the instruction pointer 'I' from the
617 /// compare instruction to the select instruction and stores this pointer in
618 /// 'PatternLastInst' member of the returned struct.
619 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
620 ReductionInstDesc &Desc);
621 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
622 /// pattern corresponding to a min(X, Y) or max(X, Y).
623 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
624 ReductionInstDesc &Prev);
625 /// Returns the induction kind of Phi. This function may return NoInduction
626 /// if the PHI is not an induction variable.
627 InductionKind isInductionVariable(PHINode *Phi);
629 /// The loop that we evaluate.
633 /// DataLayout analysis.
637 /// Target Library Info.
638 TargetLibraryInfo *TLI;
640 // --- vectorization state --- //
642 /// Holds the integer induction variable. This is the counter of the
645 /// Holds the reduction variables.
646 ReductionList Reductions;
647 /// Holds all of the induction variables that we found in the loop.
648 /// Notice that inductions don't need to start at zero and that induction
649 /// variables can be pointers.
650 InductionList Inductions;
651 /// Holds the widest induction type encountered.
654 /// Allowed outside users. This holds the reduction
655 /// vars which can be accessed from outside the loop.
656 SmallPtrSet<Value*, 4> AllowedExit;
657 /// This set holds the variables which are known to be uniform after
659 SmallPtrSet<Instruction*, 4> Uniforms;
660 /// We need to check that all of the pointers in this list are disjoint
662 RuntimePointerCheck PtrRtCheck;
663 /// Can we assume the absence of NaNs.
664 bool HasFunNoNaNAttr;
666 unsigned MaxSafeDepDistBytes;
669 /// LoopVectorizationCostModel - estimates the expected speedups due to
671 /// In many cases vectorization is not profitable. This can happen because of
672 /// a number of reasons. In this class we mainly attempt to predict the
673 /// expected speedup/slowdowns due to the supported instruction set. We use the
674 /// TargetTransformInfo to query the different backends for the cost of
675 /// different operations.
676 class LoopVectorizationCostModel {
678 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
679 LoopVectorizationLegality *Legal,
680 const TargetTransformInfo &TTI,
681 DataLayout *DL, const TargetLibraryInfo *TLI)
682 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
684 /// Information about vectorization costs
685 struct VectorizationFactor {
686 unsigned Width; // Vector width with best cost
687 unsigned Cost; // Cost of the loop with that width
689 /// \return The most profitable vectorization factor and the cost of that VF.
690 /// This method checks every power of two up to VF. If UserVF is not ZERO
691 /// then this vectorization factor will be selected if vectorization is
693 VectorizationFactor selectVectorizationFactor(bool OptForSize,
696 /// \return The size (in bits) of the widest type in the code that
697 /// needs to be vectorized. We ignore values that remain scalar such as
698 /// 64 bit loop indices.
699 unsigned getWidestType();
701 /// \return The most profitable unroll factor.
702 /// If UserUF is non-zero then this method finds the best unroll-factor
703 /// based on register pressure and other parameters.
704 /// VF and LoopCost are the selected vectorization factor and the cost of the
706 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
709 /// \brief A struct that represents some properties of the register usage
711 struct RegisterUsage {
712 /// Holds the number of loop invariant values that are used in the loop.
713 unsigned LoopInvariantRegs;
714 /// Holds the maximum number of concurrent live intervals in the loop.
715 unsigned MaxLocalUsers;
716 /// Holds the number of instructions in the loop.
717 unsigned NumInstructions;
720 /// \return information about the register usage of the loop.
721 RegisterUsage calculateRegisterUsage();
724 /// Returns the expected execution cost. The unit of the cost does
725 /// not matter because we use the 'cost' units to compare different
726 /// vector widths. The cost that is returned is *not* normalized by
727 /// the factor width.
728 unsigned expectedCost(unsigned VF);
730 /// Returns the execution time cost of an instruction for a given vector
731 /// width. Vector width of one means scalar.
732 unsigned getInstructionCost(Instruction *I, unsigned VF);
734 /// A helper function for converting Scalar types to vector types.
735 /// If the incoming type is void, we return void. If the VF is 1, we return
737 static Type* ToVectorTy(Type *Scalar, unsigned VF);
739 /// Returns whether the instruction is a load or store and will be a emitted
740 /// as a vector operation.
741 bool isConsecutiveLoadOrStore(Instruction *I);
743 /// The loop that we evaluate.
747 /// Loop Info analysis.
749 /// Vectorization legality.
750 LoopVectorizationLegality *Legal;
751 /// Vector target information.
752 const TargetTransformInfo &TTI;
753 /// Target data layout information.
755 /// Target Library Info.
756 const TargetLibraryInfo *TLI;
759 /// Utility class for getting and setting loop vectorizer hints in the form
760 /// of loop metadata.
761 struct LoopVectorizeHints {
762 /// Vectorization width.
764 /// Vectorization unroll factor.
766 /// Vectorization forced (-1 not selected, 0 force disabled, 1 force enabled)
769 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
770 : Width(VectorizationFactor)
771 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
773 , LoopID(L->getLoopID()) {
775 // The command line options override any loop metadata except for when
776 // width == 1 which is used to indicate the loop is already vectorized.
777 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
778 Width = VectorizationFactor;
779 if (VectorizationUnroll.getNumOccurrences() > 0)
780 Unroll = VectorizationUnroll;
782 DEBUG(if (DisableUnrolling && Unroll == 1)
783 dbgs() << "LV: Unrolling disabled by the pass manager\n");
786 /// Return the loop vectorizer metadata prefix.
787 static StringRef Prefix() { return "llvm.vectorizer."; }
789 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
790 SmallVector<Value*, 2> Vals;
791 Vals.push_back(MDString::get(Context, Name));
792 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
793 return MDNode::get(Context, Vals);
796 /// Mark the loop L as already vectorized by setting the width to 1.
797 void setAlreadyVectorized(Loop *L) {
798 LLVMContext &Context = L->getHeader()->getContext();
802 // Create a new loop id with one more operand for the already_vectorized
803 // hint. If the loop already has a loop id then copy the existing operands.
804 SmallVector<Value*, 4> Vals(1);
806 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
807 Vals.push_back(LoopID->getOperand(i));
809 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
810 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
812 MDNode *NewLoopID = MDNode::get(Context, Vals);
813 // Set operand 0 to refer to the loop id itself.
814 NewLoopID->replaceOperandWith(0, NewLoopID);
816 L->setLoopID(NewLoopID);
818 LoopID->replaceAllUsesWith(NewLoopID);
826 /// Find hints specified in the loop metadata.
827 void getHints(const Loop *L) {
831 // First operand should refer to the loop id itself.
832 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
833 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
835 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
836 const MDString *S = 0;
837 SmallVector<Value*, 4> Args;
839 // The expected hint is either a MDString or a MDNode with the first
840 // operand a MDString.
841 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
842 if (!MD || MD->getNumOperands() == 0)
844 S = dyn_cast<MDString>(MD->getOperand(0));
845 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
846 Args.push_back(MD->getOperand(i));
848 S = dyn_cast<MDString>(LoopID->getOperand(i));
849 assert(Args.size() == 0 && "too many arguments for MDString");
855 // Check if the hint starts with the vectorizer prefix.
856 StringRef Hint = S->getString();
857 if (!Hint.startswith(Prefix()))
859 // Remove the prefix.
860 Hint = Hint.substr(Prefix().size(), StringRef::npos);
862 if (Args.size() == 1)
863 getHint(Hint, Args[0]);
867 // Check string hint with one operand.
868 void getHint(StringRef Hint, Value *Arg) {
869 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
871 unsigned Val = C->getZExtValue();
873 if (Hint == "width") {
874 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
877 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
878 } else if (Hint == "unroll") {
879 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
882 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
883 } else if (Hint == "enable") {
884 if (C->getBitWidth() == 1)
887 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
889 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
894 /// The LoopVectorize Pass.
895 struct LoopVectorize : public LoopPass {
896 /// Pass identification, replacement for typeid
899 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
901 DisableUnrolling(NoUnrolling),
902 AlwaysVectorize(AlwaysVectorize) {
903 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
909 TargetTransformInfo *TTI;
911 TargetLibraryInfo *TLI;
912 bool DisableUnrolling;
913 bool AlwaysVectorize;
915 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
916 // We only vectorize innermost loops.
920 SE = &getAnalysis<ScalarEvolution>();
921 DL = getAnalysisIfAvailable<DataLayout>();
922 LI = &getAnalysis<LoopInfo>();
923 TTI = &getAnalysis<TargetTransformInfo>();
924 DT = &getAnalysis<DominatorTree>();
925 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
927 // If the target claims to have no vector registers don't attempt
929 if (!TTI->getNumberOfRegisters(true))
933 DEBUG(dbgs() << "LV: Not vectorizing: Missing data layout\n");
937 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
938 L->getHeader()->getParent()->getName() << "\"\n");
940 LoopVectorizeHints Hints(L, DisableUnrolling);
942 if (Hints.Force == 0) {
943 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
947 if (!AlwaysVectorize && Hints.Force != 1) {
948 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
952 if (Hints.Width == 1 && Hints.Unroll == 1) {
953 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
957 // Check if it is legal to vectorize the loop.
958 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
959 if (!LVL.canVectorize()) {
960 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
964 // Use the cost model.
965 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
967 // Check the function attributes to find out if this function should be
968 // optimized for size.
969 Function *F = L->getHeader()->getParent();
970 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
971 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
972 unsigned FnIndex = AttributeSet::FunctionIndex;
973 bool OptForSize = Hints.Force != 1 &&
974 F->getAttributes().hasAttribute(FnIndex, SzAttr);
975 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
978 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
979 "attribute is used.\n");
983 // Select the optimal vectorization factor.
984 LoopVectorizationCostModel::VectorizationFactor VF;
985 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
986 // Select the unroll factor.
987 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
990 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
991 F->getParent()->getModuleIdentifier() << '\n');
992 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
995 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
998 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
999 // We decided not to vectorize, but we may want to unroll.
1000 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1001 Unroller.vectorize(&LVL);
1003 // If we decided that it is *legal* to vectorize the loop then do it.
1004 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1008 // Mark the loop as already vectorized to avoid vectorizing again.
1009 Hints.setAlreadyVectorized(L);
1011 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1015 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1016 LoopPass::getAnalysisUsage(AU);
1017 AU.addRequiredID(LoopSimplifyID);
1018 AU.addRequiredID(LCSSAID);
1019 AU.addRequired<DominatorTree>();
1020 AU.addRequired<LoopInfo>();
1021 AU.addRequired<ScalarEvolution>();
1022 AU.addRequired<TargetTransformInfo>();
1023 AU.addPreserved<LoopInfo>();
1024 AU.addPreserved<DominatorTree>();
1029 } // end anonymous namespace
1031 //===----------------------------------------------------------------------===//
1032 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1033 // LoopVectorizationCostModel.
1034 //===----------------------------------------------------------------------===//
1037 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1038 Loop *Lp, Value *Ptr,
1040 unsigned DepSetId) {
1041 const SCEV *Sc = SE->getSCEV(Ptr);
1042 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1043 assert(AR && "Invalid addrec expression");
1044 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1045 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1046 Pointers.push_back(Ptr);
1047 Starts.push_back(AR->getStart());
1048 Ends.push_back(ScEnd);
1049 IsWritePtr.push_back(WritePtr);
1050 DependencySetId.push_back(DepSetId);
1053 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1054 // We need to place the broadcast of invariant variables outside the loop.
1055 Instruction *Instr = dyn_cast<Instruction>(V);
1056 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1057 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1059 // Place the code for broadcasting invariant variables in the new preheader.
1060 IRBuilder<>::InsertPointGuard Guard(Builder);
1062 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1064 // Broadcast the scalar into all locations in the vector.
1065 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1070 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1072 assert(Val->getType()->isVectorTy() && "Must be a vector");
1073 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1074 "Elem must be an integer");
1075 // Create the types.
1076 Type *ITy = Val->getType()->getScalarType();
1077 VectorType *Ty = cast<VectorType>(Val->getType());
1078 int VLen = Ty->getNumElements();
1079 SmallVector<Constant*, 8> Indices;
1081 // Create a vector of consecutive numbers from zero to VF.
1082 for (int i = 0; i < VLen; ++i) {
1083 int64_t Idx = Negate ? (-i) : i;
1084 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1087 // Add the consecutive indices to the vector value.
1088 Constant *Cv = ConstantVector::get(Indices);
1089 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1090 return Builder.CreateAdd(Val, Cv, "induction");
1093 /// \brief Find the operand of the GEP that should be checked for consecutive
1094 /// stores. This ignores trailing indices that have no effect on the final
1096 static unsigned getGEPInductionOperand(DataLayout *DL,
1097 const GetElementPtrInst *Gep) {
1098 unsigned LastOperand = Gep->getNumOperands() - 1;
1099 unsigned GEPAllocSize = DL->getTypeAllocSize(
1100 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1102 // Walk backwards and try to peel off zeros.
1103 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1104 // Find the type we're currently indexing into.
1105 gep_type_iterator GEPTI = gep_type_begin(Gep);
1106 std::advance(GEPTI, LastOperand - 1);
1108 // If it's a type with the same allocation size as the result of the GEP we
1109 // can peel off the zero index.
1110 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1118 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1119 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1120 // Make sure that the pointer does not point to structs.
1121 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1124 // If this value is a pointer induction variable we know it is consecutive.
1125 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1126 if (Phi && Inductions.count(Phi)) {
1127 InductionInfo II = Inductions[Phi];
1128 if (IK_PtrInduction == II.IK)
1130 else if (IK_ReversePtrInduction == II.IK)
1134 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1138 unsigned NumOperands = Gep->getNumOperands();
1139 Value *GpPtr = Gep->getPointerOperand();
1140 // If this GEP value is a consecutive pointer induction variable and all of
1141 // the indices are constant then we know it is consecutive. We can
1142 Phi = dyn_cast<PHINode>(GpPtr);
1143 if (Phi && Inductions.count(Phi)) {
1145 // Make sure that the pointer does not point to structs.
1146 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1147 if (GepPtrType->getElementType()->isAggregateType())
1150 // Make sure that all of the index operands are loop invariant.
1151 for (unsigned i = 1; i < NumOperands; ++i)
1152 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1155 InductionInfo II = Inductions[Phi];
1156 if (IK_PtrInduction == II.IK)
1158 else if (IK_ReversePtrInduction == II.IK)
1162 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1164 // Check that all of the gep indices are uniform except for our induction
1166 for (unsigned i = 0; i != NumOperands; ++i)
1167 if (i != InductionOperand &&
1168 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1171 // We can emit wide load/stores only if the last non-zero index is the
1172 // induction variable.
1173 const SCEV *Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1174 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1175 const SCEV *Step = AR->getStepRecurrence(*SE);
1177 // The memory is consecutive because the last index is consecutive
1178 // and all other indices are loop invariant.
1181 if (Step->isAllOnesValue())
1188 bool LoopVectorizationLegality::isUniform(Value *V) {
1189 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1192 InnerLoopVectorizer::VectorParts&
1193 InnerLoopVectorizer::getVectorValue(Value *V) {
1194 assert(V != Induction && "The new induction variable should not be used.");
1195 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1197 // If we have this scalar in the map, return it.
1198 if (WidenMap.has(V))
1199 return WidenMap.get(V);
1201 // If this scalar is unknown, assume that it is a constant or that it is
1202 // loop invariant. Broadcast V and save the value for future uses.
1203 Value *B = getBroadcastInstrs(V);
1204 return WidenMap.splat(V, B);
1207 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1208 assert(Vec->getType()->isVectorTy() && "Invalid type");
1209 SmallVector<Constant*, 8> ShuffleMask;
1210 for (unsigned i = 0; i < VF; ++i)
1211 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1213 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1214 ConstantVector::get(ShuffleMask),
1219 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1220 LoopVectorizationLegality *Legal) {
1221 // Attempt to issue a wide load.
1222 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1223 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1225 assert((LI || SI) && "Invalid Load/Store instruction");
1227 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1228 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1229 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1230 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1231 // An alignment of 0 means target abi alignment. We need to use the scalar's
1232 // target abi alignment in such a case.
1234 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1235 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1236 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1237 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1239 if (ScalarAllocatedSize != VectorElementSize)
1240 return scalarizeInstruction(Instr);
1242 // If the pointer is loop invariant or if it is non-consecutive,
1243 // scalarize the load.
1244 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1245 bool Reverse = ConsecutiveStride < 0;
1246 bool UniformLoad = LI && Legal->isUniform(Ptr);
1247 if (!ConsecutiveStride || UniformLoad)
1248 return scalarizeInstruction(Instr);
1250 Constant *Zero = Builder.getInt32(0);
1251 VectorParts &Entry = WidenMap.get(Instr);
1253 // Handle consecutive loads/stores.
1254 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1255 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1256 setDebugLocFromInst(Builder, Gep);
1257 Value *PtrOperand = Gep->getPointerOperand();
1258 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1259 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1261 // Create the new GEP with the new induction variable.
1262 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1263 Gep2->setOperand(0, FirstBasePtr);
1264 Gep2->setName("gep.indvar.base");
1265 Ptr = Builder.Insert(Gep2);
1267 setDebugLocFromInst(Builder, Gep);
1268 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1269 OrigLoop) && "Base ptr must be invariant");
1271 // The last index does not have to be the induction. It can be
1272 // consecutive and be a function of the index. For example A[I+1];
1273 unsigned NumOperands = Gep->getNumOperands();
1274 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1275 // Create the new GEP with the new induction variable.
1276 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1278 for (unsigned i = 0; i < NumOperands; ++i) {
1279 Value *GepOperand = Gep->getOperand(i);
1280 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1282 // Update last index or loop invariant instruction anchored in loop.
1283 if (i == InductionOperand ||
1284 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1285 assert((i == InductionOperand ||
1286 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1287 "Must be last index or loop invariant");
1289 VectorParts &GEPParts = getVectorValue(GepOperand);
1290 Value *Index = GEPParts[0];
1291 Index = Builder.CreateExtractElement(Index, Zero);
1292 Gep2->setOperand(i, Index);
1293 Gep2->setName("gep.indvar.idx");
1296 Ptr = Builder.Insert(Gep2);
1298 // Use the induction element ptr.
1299 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1300 setDebugLocFromInst(Builder, Ptr);
1301 VectorParts &PtrVal = getVectorValue(Ptr);
1302 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1307 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1308 "We do not allow storing to uniform addresses");
1309 setDebugLocFromInst(Builder, SI);
1310 // We don't want to update the value in the map as it might be used in
1311 // another expression. So don't use a reference type for "StoredVal".
1312 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1314 for (unsigned Part = 0; Part < UF; ++Part) {
1315 // Calculate the pointer for the specific unroll-part.
1316 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1319 // If we store to reverse consecutive memory locations then we need
1320 // to reverse the order of elements in the stored value.
1321 StoredVal[Part] = reverseVector(StoredVal[Part]);
1322 // If the address is consecutive but reversed, then the
1323 // wide store needs to start at the last vector element.
1324 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1325 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1328 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1329 DataTy->getPointerTo(AddressSpace));
1330 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1336 assert(LI && "Must have a load instruction");
1337 setDebugLocFromInst(Builder, LI);
1338 for (unsigned Part = 0; Part < UF; ++Part) {
1339 // Calculate the pointer for the specific unroll-part.
1340 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1343 // If the address is consecutive but reversed, then the
1344 // wide store needs to start at the last vector element.
1345 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1346 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1349 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1350 DataTy->getPointerTo(AddressSpace));
1351 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1352 cast<LoadInst>(LI)->setAlignment(Alignment);
1353 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1357 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1358 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1359 // Holds vector parameters or scalars, in case of uniform vals.
1360 SmallVector<VectorParts, 4> Params;
1362 setDebugLocFromInst(Builder, Instr);
1364 // Find all of the vectorized parameters.
1365 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1366 Value *SrcOp = Instr->getOperand(op);
1368 // If we are accessing the old induction variable, use the new one.
1369 if (SrcOp == OldInduction) {
1370 Params.push_back(getVectorValue(SrcOp));
1374 // Try using previously calculated values.
1375 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1377 // If the src is an instruction that appeared earlier in the basic block
1378 // then it should already be vectorized.
1379 if (SrcInst && OrigLoop->contains(SrcInst)) {
1380 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1381 // The parameter is a vector value from earlier.
1382 Params.push_back(WidenMap.get(SrcInst));
1384 // The parameter is a scalar from outside the loop. Maybe even a constant.
1385 VectorParts Scalars;
1386 Scalars.append(UF, SrcOp);
1387 Params.push_back(Scalars);
1391 assert(Params.size() == Instr->getNumOperands() &&
1392 "Invalid number of operands");
1394 // Does this instruction return a value ?
1395 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1397 Value *UndefVec = IsVoidRetTy ? 0 :
1398 UndefValue::get(VectorType::get(Instr->getType(), VF));
1399 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1400 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1402 // For each vector unroll 'part':
1403 for (unsigned Part = 0; Part < UF; ++Part) {
1404 // For each scalar that we create:
1405 for (unsigned Width = 0; Width < VF; ++Width) {
1406 Instruction *Cloned = Instr->clone();
1408 Cloned->setName(Instr->getName() + ".cloned");
1409 // Replace the operands of the cloned instructions with extracted scalars.
1410 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1411 Value *Op = Params[op][Part];
1412 // Param is a vector. Need to extract the right lane.
1413 if (Op->getType()->isVectorTy())
1414 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1415 Cloned->setOperand(op, Op);
1418 // Place the cloned scalar in the new loop.
1419 Builder.Insert(Cloned);
1421 // If the original scalar returns a value we need to place it in a vector
1422 // so that future users will be able to use it.
1424 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1425 Builder.getInt32(Width));
1431 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1433 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1434 Legal->getRuntimePointerCheck();
1436 if (!PtrRtCheck->Need)
1439 unsigned NumPointers = PtrRtCheck->Pointers.size();
1440 SmallVector<TrackingVH<Value> , 2> Starts;
1441 SmallVector<TrackingVH<Value> , 2> Ends;
1443 LLVMContext &Ctx = Loc->getContext();
1444 SCEVExpander Exp(*SE, "induction");
1446 for (unsigned i = 0; i < NumPointers; ++i) {
1447 Value *Ptr = PtrRtCheck->Pointers[i];
1448 const SCEV *Sc = SE->getSCEV(Ptr);
1450 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1451 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1453 Starts.push_back(Ptr);
1454 Ends.push_back(Ptr);
1456 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1457 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1459 // Use this type for pointer arithmetic.
1460 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1462 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1463 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1464 Starts.push_back(Start);
1465 Ends.push_back(End);
1469 IRBuilder<> ChkBuilder(Loc);
1470 // Our instructions might fold to a constant.
1471 Value *MemoryRuntimeCheck = 0;
1472 for (unsigned i = 0; i < NumPointers; ++i) {
1473 for (unsigned j = i+1; j < NumPointers; ++j) {
1474 // No need to check if two readonly pointers intersect.
1475 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1478 // Only need to check pointers between two different dependency sets.
1479 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1482 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1483 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1485 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1486 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1487 "Trying to bounds check pointers with different address spaces");
1489 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1490 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1492 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1493 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1494 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1495 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1497 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1498 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1499 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1500 if (MemoryRuntimeCheck)
1501 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1503 MemoryRuntimeCheck = IsConflict;
1507 // We have to do this trickery because the IRBuilder might fold the check to a
1508 // constant expression in which case there is no Instruction anchored in a
1510 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1511 ConstantInt::getTrue(Ctx));
1512 ChkBuilder.Insert(Check, "memcheck.conflict");
1517 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1519 In this function we generate a new loop. The new loop will contain
1520 the vectorized instructions while the old loop will continue to run the
1523 [ ] <-- vector loop bypass (may consist of multiple blocks).
1526 | [ ] <-- vector pre header.
1530 | [ ]_| <-- vector loop.
1533 >[ ] <--- middle-block.
1536 | [ ] <--- new preheader.
1540 | [ ]_| <-- old scalar loop to handle remainder.
1543 >[ ] <-- exit block.
1547 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1548 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1549 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1550 assert(ExitBlock && "Must have an exit block");
1552 // Some loops have a single integer induction variable, while other loops
1553 // don't. One example is c++ iterators that often have multiple pointer
1554 // induction variables. In the code below we also support a case where we
1555 // don't have a single induction variable.
1556 OldInduction = Legal->getInduction();
1557 Type *IdxTy = Legal->getWidestInductionType();
1559 // Find the loop boundaries.
1560 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1561 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1563 // The exit count might have the type of i64 while the phi is i32. This can
1564 // happen if we have an induction variable that is sign extended before the
1565 // compare. The only way that we get a backedge taken count is that the
1566 // induction variable was signed and as such will not overflow. In such a case
1567 // truncation is legal.
1568 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1569 IdxTy->getPrimitiveSizeInBits())
1570 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1572 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1573 // Get the total trip count from the count by adding 1.
1574 ExitCount = SE->getAddExpr(ExitCount,
1575 SE->getConstant(ExitCount->getType(), 1));
1577 // Expand the trip count and place the new instructions in the preheader.
1578 // Notice that the pre-header does not change, only the loop body.
1579 SCEVExpander Exp(*SE, "induction");
1581 // Count holds the overall loop count (N).
1582 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1583 BypassBlock->getTerminator());
1585 // The loop index does not have to start at Zero. Find the original start
1586 // value from the induction PHI node. If we don't have an induction variable
1587 // then we know that it starts at zero.
1588 Builder.SetInsertPoint(BypassBlock->getTerminator());
1589 Value *StartIdx = ExtendedIdx = OldInduction ?
1590 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1592 ConstantInt::get(IdxTy, 0);
1594 assert(BypassBlock && "Invalid loop structure");
1595 LoopBypassBlocks.push_back(BypassBlock);
1597 // Split the single block loop into the two loop structure described above.
1598 BasicBlock *VectorPH =
1599 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1600 BasicBlock *VecBody =
1601 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1602 BasicBlock *MiddleBlock =
1603 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1604 BasicBlock *ScalarPH =
1605 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1607 // Create and register the new vector loop.
1608 Loop* Lp = new Loop();
1609 Loop *ParentLoop = OrigLoop->getParentLoop();
1611 // Insert the new loop into the loop nest and register the new basic blocks
1612 // before calling any utilities such as SCEV that require valid LoopInfo.
1614 ParentLoop->addChildLoop(Lp);
1615 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1616 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1617 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1619 LI->addTopLevelLoop(Lp);
1621 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1623 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1625 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1627 // Generate the induction variable.
1628 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1629 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1630 // The loop step is equal to the vectorization factor (num of SIMD elements)
1631 // times the unroll factor (num of SIMD instructions).
1632 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1634 // This is the IR builder that we use to add all of the logic for bypassing
1635 // the new vector loop.
1636 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1637 setDebugLocFromInst(BypassBuilder,
1638 getDebugLocFromInstOrOperands(OldInduction));
1640 // We may need to extend the index in case there is a type mismatch.
1641 // We know that the count starts at zero and does not overflow.
1642 if (Count->getType() != IdxTy) {
1643 // The exit count can be of pointer type. Convert it to the correct
1645 if (ExitCount->getType()->isPointerTy())
1646 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1648 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1651 // Add the start index to the loop count to get the new end index.
1652 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1654 // Now we need to generate the expression for N - (N % VF), which is
1655 // the part that the vectorized body will execute.
1656 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1657 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1658 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1659 "end.idx.rnd.down");
1661 // Now, compare the new count to zero. If it is zero skip the vector loop and
1662 // jump to the scalar loop.
1663 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1666 BasicBlock *LastBypassBlock = BypassBlock;
1668 // Generate the code that checks in runtime if arrays overlap. We put the
1669 // checks into a separate block to make the more common case of few elements
1671 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1672 BypassBlock->getTerminator());
1673 if (MemRuntimeCheck) {
1674 // Create a new block containing the memory check.
1675 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1678 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1679 LoopBypassBlocks.push_back(CheckBlock);
1681 // Replace the branch into the memory check block with a conditional branch
1682 // for the "few elements case".
1683 Instruction *OldTerm = BypassBlock->getTerminator();
1684 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1685 OldTerm->eraseFromParent();
1687 Cmp = MemRuntimeCheck;
1688 LastBypassBlock = CheckBlock;
1691 LastBypassBlock->getTerminator()->eraseFromParent();
1692 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1695 // We are going to resume the execution of the scalar loop.
1696 // Go over all of the induction variables that we found and fix the
1697 // PHIs that are left in the scalar version of the loop.
1698 // The starting values of PHI nodes depend on the counter of the last
1699 // iteration in the vectorized loop.
1700 // If we come from a bypass edge then we need to start from the original
1703 // This variable saves the new starting index for the scalar loop.
1704 PHINode *ResumeIndex = 0;
1705 LoopVectorizationLegality::InductionList::iterator I, E;
1706 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1707 // Set builder to point to last bypass block.
1708 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1709 for (I = List->begin(), E = List->end(); I != E; ++I) {
1710 PHINode *OrigPhi = I->first;
1711 LoopVectorizationLegality::InductionInfo II = I->second;
1713 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1714 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1715 MiddleBlock->getTerminator());
1716 // We might have extended the type of the induction variable but we need a
1717 // truncated version for the scalar loop.
1718 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1719 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1720 MiddleBlock->getTerminator()) : 0;
1722 Value *EndValue = 0;
1724 case LoopVectorizationLegality::IK_NoInduction:
1725 llvm_unreachable("Unknown induction");
1726 case LoopVectorizationLegality::IK_IntInduction: {
1727 // Handle the integer induction counter.
1728 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1730 // We have the canonical induction variable.
1731 if (OrigPhi == OldInduction) {
1732 // Create a truncated version of the resume value for the scalar loop,
1733 // we might have promoted the type to a larger width.
1735 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1736 // The new PHI merges the original incoming value, in case of a bypass,
1737 // or the value at the end of the vectorized loop.
1738 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1739 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1740 TruncResumeVal->addIncoming(EndValue, VecBody);
1742 // We know what the end value is.
1743 EndValue = IdxEndRoundDown;
1744 // We also know which PHI node holds it.
1745 ResumeIndex = ResumeVal;
1749 // Not the canonical induction variable - add the vector loop count to the
1751 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1752 II.StartValue->getType(),
1754 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1757 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1758 // Convert the CountRoundDown variable to the PHI size.
1759 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1760 II.StartValue->getType(),
1762 // Handle reverse integer induction counter.
1763 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1766 case LoopVectorizationLegality::IK_PtrInduction: {
1767 // For pointer induction variables, calculate the offset using
1769 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1773 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1774 // The value at the end of the loop for the reverse pointer is calculated
1775 // by creating a GEP with a negative index starting from the start value.
1776 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1777 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1779 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1785 // The new PHI merges the original incoming value, in case of a bypass,
1786 // or the value at the end of the vectorized loop.
1787 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1788 if (OrigPhi == OldInduction)
1789 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1791 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1793 ResumeVal->addIncoming(EndValue, VecBody);
1795 // Fix the scalar body counter (PHI node).
1796 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1797 // The old inductions phi node in the scalar body needs the truncated value.
1798 if (OrigPhi == OldInduction)
1799 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1801 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1804 // If we are generating a new induction variable then we also need to
1805 // generate the code that calculates the exit value. This value is not
1806 // simply the end of the counter because we may skip the vectorized body
1807 // in case of a runtime check.
1809 assert(!ResumeIndex && "Unexpected resume value found");
1810 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1811 MiddleBlock->getTerminator());
1812 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1813 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1814 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1817 // Make sure that we found the index where scalar loop needs to continue.
1818 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1819 "Invalid resume Index");
1821 // Add a check in the middle block to see if we have completed
1822 // all of the iterations in the first vector loop.
1823 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1824 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1825 ResumeIndex, "cmp.n",
1826 MiddleBlock->getTerminator());
1828 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1829 // Remove the old terminator.
1830 MiddleBlock->getTerminator()->eraseFromParent();
1832 // Create i+1 and fill the PHINode.
1833 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1834 Induction->addIncoming(StartIdx, VectorPH);
1835 Induction->addIncoming(NextIdx, VecBody);
1836 // Create the compare.
1837 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1838 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1840 // Now we have two terminators. Remove the old one from the block.
1841 VecBody->getTerminator()->eraseFromParent();
1843 // Get ready to start creating new instructions into the vectorized body.
1844 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1847 LoopVectorPreHeader = VectorPH;
1848 LoopScalarPreHeader = ScalarPH;
1849 LoopMiddleBlock = MiddleBlock;
1850 LoopExitBlock = ExitBlock;
1851 LoopVectorBody = VecBody;
1852 LoopScalarBody = OldBasicBlock;
1854 LoopVectorizeHints Hints(Lp, true);
1855 Hints.setAlreadyVectorized(Lp);
1858 /// This function returns the identity element (or neutral element) for
1859 /// the operation K.
1861 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1866 // Adding, Xoring, Oring zero to a number does not change it.
1867 return ConstantInt::get(Tp, 0);
1868 case RK_IntegerMult:
1869 // Multiplying a number by 1 does not change it.
1870 return ConstantInt::get(Tp, 1);
1872 // AND-ing a number with an all-1 value does not change it.
1873 return ConstantInt::get(Tp, -1, true);
1875 // Multiplying a number by 1 does not change it.
1876 return ConstantFP::get(Tp, 1.0L);
1878 // Adding zero to a number does not change it.
1879 return ConstantFP::get(Tp, 0.0L);
1881 llvm_unreachable("Unknown reduction kind");
1885 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
1886 Intrinsic::ID ValidIntrinsicID) {
1887 if (I.getNumArgOperands() != 1 ||
1888 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1889 I.getType() != I.getArgOperand(0)->getType() ||
1890 !I.onlyReadsMemory())
1891 return Intrinsic::not_intrinsic;
1893 return ValidIntrinsicID;
1896 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
1897 Intrinsic::ID ValidIntrinsicID) {
1898 if (I.getNumArgOperands() != 2 ||
1899 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1900 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
1901 I.getType() != I.getArgOperand(0)->getType() ||
1902 I.getType() != I.getArgOperand(1)->getType() ||
1903 !I.onlyReadsMemory())
1904 return Intrinsic::not_intrinsic;
1906 return ValidIntrinsicID;
1910 static Intrinsic::ID
1911 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1912 // If we have an intrinsic call, check if it is trivially vectorizable.
1913 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1914 switch (II->getIntrinsicID()) {
1915 case Intrinsic::sqrt:
1916 case Intrinsic::sin:
1917 case Intrinsic::cos:
1918 case Intrinsic::exp:
1919 case Intrinsic::exp2:
1920 case Intrinsic::log:
1921 case Intrinsic::log10:
1922 case Intrinsic::log2:
1923 case Intrinsic::fabs:
1924 case Intrinsic::copysign:
1925 case Intrinsic::floor:
1926 case Intrinsic::ceil:
1927 case Intrinsic::trunc:
1928 case Intrinsic::rint:
1929 case Intrinsic::nearbyint:
1930 case Intrinsic::round:
1931 case Intrinsic::pow:
1932 case Intrinsic::fma:
1933 case Intrinsic::fmuladd:
1934 case Intrinsic::lifetime_start:
1935 case Intrinsic::lifetime_end:
1936 return II->getIntrinsicID();
1938 return Intrinsic::not_intrinsic;
1943 return Intrinsic::not_intrinsic;
1946 Function *F = CI->getCalledFunction();
1947 // We're going to make assumptions on the semantics of the functions, check
1948 // that the target knows that it's available in this environment and it does
1949 // not have local linkage.
1950 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
1951 return Intrinsic::not_intrinsic;
1953 // Otherwise check if we have a call to a function that can be turned into a
1954 // vector intrinsic.
1961 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
1965 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
1969 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
1971 case LibFunc::exp2f:
1972 case LibFunc::exp2l:
1973 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
1977 return checkUnaryFloatSignature(*CI, Intrinsic::log);
1978 case LibFunc::log10:
1979 case LibFunc::log10f:
1980 case LibFunc::log10l:
1981 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
1983 case LibFunc::log2f:
1984 case LibFunc::log2l:
1985 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
1987 case LibFunc::fabsf:
1988 case LibFunc::fabsl:
1989 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
1990 case LibFunc::copysign:
1991 case LibFunc::copysignf:
1992 case LibFunc::copysignl:
1993 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
1994 case LibFunc::floor:
1995 case LibFunc::floorf:
1996 case LibFunc::floorl:
1997 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
1999 case LibFunc::ceilf:
2000 case LibFunc::ceill:
2001 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2002 case LibFunc::trunc:
2003 case LibFunc::truncf:
2004 case LibFunc::truncl:
2005 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2007 case LibFunc::rintf:
2008 case LibFunc::rintl:
2009 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2010 case LibFunc::nearbyint:
2011 case LibFunc::nearbyintf:
2012 case LibFunc::nearbyintl:
2013 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2014 case LibFunc::round:
2015 case LibFunc::roundf:
2016 case LibFunc::roundl:
2017 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2021 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2024 return Intrinsic::not_intrinsic;
2027 /// This function translates the reduction kind to an LLVM binary operator.
2029 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2031 case LoopVectorizationLegality::RK_IntegerAdd:
2032 return Instruction::Add;
2033 case LoopVectorizationLegality::RK_IntegerMult:
2034 return Instruction::Mul;
2035 case LoopVectorizationLegality::RK_IntegerOr:
2036 return Instruction::Or;
2037 case LoopVectorizationLegality::RK_IntegerAnd:
2038 return Instruction::And;
2039 case LoopVectorizationLegality::RK_IntegerXor:
2040 return Instruction::Xor;
2041 case LoopVectorizationLegality::RK_FloatMult:
2042 return Instruction::FMul;
2043 case LoopVectorizationLegality::RK_FloatAdd:
2044 return Instruction::FAdd;
2045 case LoopVectorizationLegality::RK_IntegerMinMax:
2046 return Instruction::ICmp;
2047 case LoopVectorizationLegality::RK_FloatMinMax:
2048 return Instruction::FCmp;
2050 llvm_unreachable("Unknown reduction operation");
2054 Value *createMinMaxOp(IRBuilder<> &Builder,
2055 LoopVectorizationLegality::MinMaxReductionKind RK,
2058 CmpInst::Predicate P = CmpInst::ICMP_NE;
2061 llvm_unreachable("Unknown min/max reduction kind");
2062 case LoopVectorizationLegality::MRK_UIntMin:
2063 P = CmpInst::ICMP_ULT;
2065 case LoopVectorizationLegality::MRK_UIntMax:
2066 P = CmpInst::ICMP_UGT;
2068 case LoopVectorizationLegality::MRK_SIntMin:
2069 P = CmpInst::ICMP_SLT;
2071 case LoopVectorizationLegality::MRK_SIntMax:
2072 P = CmpInst::ICMP_SGT;
2074 case LoopVectorizationLegality::MRK_FloatMin:
2075 P = CmpInst::FCMP_OLT;
2077 case LoopVectorizationLegality::MRK_FloatMax:
2078 P = CmpInst::FCMP_OGT;
2083 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2084 RK == LoopVectorizationLegality::MRK_FloatMax)
2085 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2087 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2089 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2094 struct CSEDenseMapInfo {
2095 static bool canHandle(Instruction *I) {
2096 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2097 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2099 static inline Instruction *getEmptyKey() {
2100 return DenseMapInfo<Instruction *>::getEmptyKey();
2102 static inline Instruction *getTombstoneKey() {
2103 return DenseMapInfo<Instruction *>::getTombstoneKey();
2105 static unsigned getHashValue(Instruction *I) {
2106 assert(canHandle(I) && "Unknown instruction!");
2107 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2108 I->value_op_end()));
2110 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2111 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2112 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2114 return LHS->isIdenticalTo(RHS);
2119 ///\brief Perform cse of induction variable instructions.
2120 static void cse(BasicBlock *BB) {
2121 // Perform simple cse.
2122 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2123 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2124 Instruction *In = I++;
2126 if (!CSEDenseMapInfo::canHandle(In))
2129 // Check if we can replace this instruction with any of the
2130 // visited instructions.
2131 if (Instruction *V = CSEMap.lookup(In)) {
2132 In->replaceAllUsesWith(V);
2133 In->eraseFromParent();
2142 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2143 //===------------------------------------------------===//
2145 // Notice: any optimization or new instruction that go
2146 // into the code below should be also be implemented in
2149 //===------------------------------------------------===//
2150 Constant *Zero = Builder.getInt32(0);
2152 // In order to support reduction variables we need to be able to vectorize
2153 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2154 // stages. First, we create a new vector PHI node with no incoming edges.
2155 // We use this value when we vectorize all of the instructions that use the
2156 // PHI. Next, after all of the instructions in the block are complete we
2157 // add the new incoming edges to the PHI. At this point all of the
2158 // instructions in the basic block are vectorized, so we can use them to
2159 // construct the PHI.
2160 PhiVector RdxPHIsToFix;
2162 // Scan the loop in a topological order to ensure that defs are vectorized
2164 LoopBlocksDFS DFS(OrigLoop);
2167 // Vectorize all of the blocks in the original loop.
2168 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2169 be = DFS.endRPO(); bb != be; ++bb)
2170 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2172 // At this point every instruction in the original loop is widened to
2173 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2174 // that we vectorized. The PHI nodes are currently empty because we did
2175 // not want to introduce cycles. Notice that the remaining PHI nodes
2176 // that we need to fix are reduction variables.
2178 // Create the 'reduced' values for each of the induction vars.
2179 // The reduced values are the vector values that we scalarize and combine
2180 // after the loop is finished.
2181 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2183 PHINode *RdxPhi = *it;
2184 assert(RdxPhi && "Unable to recover vectorized PHI");
2186 // Find the reduction variable descriptor.
2187 assert(Legal->getReductionVars()->count(RdxPhi) &&
2188 "Unable to find the reduction variable");
2189 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2190 (*Legal->getReductionVars())[RdxPhi];
2192 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2194 // We need to generate a reduction vector from the incoming scalar.
2195 // To do so, we need to generate the 'identity' vector and overide
2196 // one of the elements with the incoming scalar reduction. We need
2197 // to do it in the vector-loop preheader.
2198 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2200 // This is the vector-clone of the value that leaves the loop.
2201 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2202 Type *VecTy = VectorExit[0]->getType();
2204 // Find the reduction identity variable. Zero for addition, or, xor,
2205 // one for multiplication, -1 for And.
2208 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2209 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2210 // MinMax reduction have the start value as their identify.
2212 VectorStart = Identity = RdxDesc.StartValue;
2214 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2219 // Handle other reduction kinds:
2221 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2222 VecTy->getScalarType());
2225 // This vector is the Identity vector where the first element is the
2226 // incoming scalar reduction.
2227 VectorStart = RdxDesc.StartValue;
2229 Identity = ConstantVector::getSplat(VF, Iden);
2231 // This vector is the Identity vector where the first element is the
2232 // incoming scalar reduction.
2233 VectorStart = Builder.CreateInsertElement(Identity,
2234 RdxDesc.StartValue, Zero);
2238 // Fix the vector-loop phi.
2239 // We created the induction variable so we know that the
2240 // preheader is the first entry.
2241 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2243 // Reductions do not have to start at zero. They can start with
2244 // any loop invariant values.
2245 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2246 BasicBlock *Latch = OrigLoop->getLoopLatch();
2247 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2248 VectorParts &Val = getVectorValue(LoopVal);
2249 for (unsigned part = 0; part < UF; ++part) {
2250 // Make sure to add the reduction stat value only to the
2251 // first unroll part.
2252 Value *StartVal = (part == 0) ? VectorStart : Identity;
2253 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2254 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2257 // Before each round, move the insertion point right between
2258 // the PHIs and the values we are going to write.
2259 // This allows us to write both PHINodes and the extractelement
2261 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2263 VectorParts RdxParts;
2264 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2265 for (unsigned part = 0; part < UF; ++part) {
2266 // This PHINode contains the vectorized reduction variable, or
2267 // the initial value vector, if we bypass the vector loop.
2268 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2269 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2270 Value *StartVal = (part == 0) ? VectorStart : Identity;
2271 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2272 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2273 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2274 RdxParts.push_back(NewPhi);
2277 // Reduce all of the unrolled parts into a single vector.
2278 Value *ReducedPartRdx = RdxParts[0];
2279 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2280 setDebugLocFromInst(Builder, ReducedPartRdx);
2281 for (unsigned part = 1; part < UF; ++part) {
2282 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2283 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2284 RdxParts[part], ReducedPartRdx,
2287 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2288 ReducedPartRdx, RdxParts[part]);
2292 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2293 // and vector ops, reducing the set of values being computed by half each
2295 assert(isPowerOf2_32(VF) &&
2296 "Reduction emission only supported for pow2 vectors!");
2297 Value *TmpVec = ReducedPartRdx;
2298 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2299 for (unsigned i = VF; i != 1; i >>= 1) {
2300 // Move the upper half of the vector to the lower half.
2301 for (unsigned j = 0; j != i/2; ++j)
2302 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2304 // Fill the rest of the mask with undef.
2305 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2306 UndefValue::get(Builder.getInt32Ty()));
2309 Builder.CreateShuffleVector(TmpVec,
2310 UndefValue::get(TmpVec->getType()),
2311 ConstantVector::get(ShuffleMask),
2314 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2315 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2318 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2321 // The result is in the first element of the vector.
2322 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2323 Builder.getInt32(0));
2326 // Now, we need to fix the users of the reduction variable
2327 // inside and outside of the scalar remainder loop.
2328 // We know that the loop is in LCSSA form. We need to update the
2329 // PHI nodes in the exit blocks.
2330 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2331 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2332 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2333 if (!LCSSAPhi) break;
2335 // All PHINodes need to have a single entry edge, or two if
2336 // we already fixed them.
2337 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2339 // We found our reduction value exit-PHI. Update it with the
2340 // incoming bypass edge.
2341 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2342 // Add an edge coming from the bypass.
2343 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2346 }// end of the LCSSA phi scan.
2348 // Fix the scalar loop reduction variable with the incoming reduction sum
2349 // from the vector body and from the backedge value.
2350 int IncomingEdgeBlockIdx =
2351 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2352 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2353 // Pick the other block.
2354 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2355 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2356 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2357 }// end of for each redux variable.
2361 // Remove redundant induction instructions.
2362 cse(LoopVectorBody);
2365 void InnerLoopVectorizer::fixLCSSAPHIs() {
2366 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2367 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2368 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2369 if (!LCSSAPhi) break;
2370 if (LCSSAPhi->getNumIncomingValues() == 1)
2371 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2376 InnerLoopVectorizer::VectorParts
2377 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2378 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2381 // Look for cached value.
2382 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2383 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2384 if (ECEntryIt != MaskCache.end())
2385 return ECEntryIt->second;
2387 VectorParts SrcMask = createBlockInMask(Src);
2389 // The terminator has to be a branch inst!
2390 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2391 assert(BI && "Unexpected terminator found");
2393 if (BI->isConditional()) {
2394 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2396 if (BI->getSuccessor(0) != Dst)
2397 for (unsigned part = 0; part < UF; ++part)
2398 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2400 for (unsigned part = 0; part < UF; ++part)
2401 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2403 MaskCache[Edge] = EdgeMask;
2407 MaskCache[Edge] = SrcMask;
2411 InnerLoopVectorizer::VectorParts
2412 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2413 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2415 // Loop incoming mask is all-one.
2416 if (OrigLoop->getHeader() == BB) {
2417 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2418 return getVectorValue(C);
2421 // This is the block mask. We OR all incoming edges, and with zero.
2422 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2423 VectorParts BlockMask = getVectorValue(Zero);
2426 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2427 VectorParts EM = createEdgeMask(*it, BB);
2428 for (unsigned part = 0; part < UF; ++part)
2429 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2435 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2436 InnerLoopVectorizer::VectorParts &Entry,
2437 LoopVectorizationLegality *Legal,
2438 unsigned UF, unsigned VF, PhiVector *PV) {
2439 PHINode* P = cast<PHINode>(PN);
2440 // Handle reduction variables:
2441 if (Legal->getReductionVars()->count(P)) {
2442 for (unsigned part = 0; part < UF; ++part) {
2443 // This is phase one of vectorizing PHIs.
2444 Type *VecTy = (VF == 1) ? PN->getType() :
2445 VectorType::get(PN->getType(), VF);
2446 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2447 LoopVectorBody-> getFirstInsertionPt());
2453 setDebugLocFromInst(Builder, P);
2454 // Check for PHI nodes that are lowered to vector selects.
2455 if (P->getParent() != OrigLoop->getHeader()) {
2456 // We know that all PHIs in non-header blocks are converted into
2457 // selects, so we don't have to worry about the insertion order and we
2458 // can just use the builder.
2459 // At this point we generate the predication tree. There may be
2460 // duplications since this is a simple recursive scan, but future
2461 // optimizations will clean it up.
2463 unsigned NumIncoming = P->getNumIncomingValues();
2465 // Generate a sequence of selects of the form:
2466 // SELECT(Mask3, In3,
2467 // SELECT(Mask2, In2,
2469 for (unsigned In = 0; In < NumIncoming; In++) {
2470 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2472 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2474 for (unsigned part = 0; part < UF; ++part) {
2475 // We might have single edge PHIs (blocks) - use an identity
2476 // 'select' for the first PHI operand.
2478 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2481 // Select between the current value and the previous incoming edge
2482 // based on the incoming mask.
2483 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2484 Entry[part], "predphi");
2490 // This PHINode must be an induction variable.
2491 // Make sure that we know about it.
2492 assert(Legal->getInductionVars()->count(P) &&
2493 "Not an induction variable");
2495 LoopVectorizationLegality::InductionInfo II =
2496 Legal->getInductionVars()->lookup(P);
2499 case LoopVectorizationLegality::IK_NoInduction:
2500 llvm_unreachable("Unknown induction");
2501 case LoopVectorizationLegality::IK_IntInduction: {
2502 assert(P->getType() == II.StartValue->getType() && "Types must match");
2503 Type *PhiTy = P->getType();
2505 if (P == OldInduction) {
2506 // Handle the canonical induction variable. We might have had to
2508 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2510 // Handle other induction variables that are now based on the
2512 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2514 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2515 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2518 Broadcasted = getBroadcastInstrs(Broadcasted);
2519 // After broadcasting the induction variable we need to make the vector
2520 // consecutive by adding 0, 1, 2, etc.
2521 for (unsigned part = 0; part < UF; ++part)
2522 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2525 case LoopVectorizationLegality::IK_ReverseIntInduction:
2526 case LoopVectorizationLegality::IK_PtrInduction:
2527 case LoopVectorizationLegality::IK_ReversePtrInduction:
2528 // Handle reverse integer and pointer inductions.
2529 Value *StartIdx = ExtendedIdx;
2530 // This is the normalized GEP that starts counting at zero.
2531 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2534 // Handle the reverse integer induction variable case.
2535 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2536 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2537 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2539 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2542 // This is a new value so do not hoist it out.
2543 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2544 // After broadcasting the induction variable we need to make the
2545 // vector consecutive by adding ... -3, -2, -1, 0.
2546 for (unsigned part = 0; part < UF; ++part)
2547 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2552 // Handle the pointer induction variable case.
2553 assert(P->getType()->isPointerTy() && "Unexpected type.");
2555 // Is this a reverse induction ptr or a consecutive induction ptr.
2556 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2559 // This is the vector of results. Notice that we don't generate
2560 // vector geps because scalar geps result in better code.
2561 for (unsigned part = 0; part < UF; ++part) {
2563 int EltIndex = (part) * (Reverse ? -1 : 1);
2564 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2567 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2569 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2571 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2573 Entry[part] = SclrGep;
2577 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2578 for (unsigned int i = 0; i < VF; ++i) {
2579 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2580 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2583 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2585 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2587 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2589 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2590 Builder.getInt32(i),
2593 Entry[part] = VecVal;
2600 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2601 BasicBlock *BB, PhiVector *PV) {
2602 // For each instruction in the old loop.
2603 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2604 VectorParts &Entry = WidenMap.get(it);
2605 switch (it->getOpcode()) {
2606 case Instruction::Br:
2607 // Nothing to do for PHIs and BR, since we already took care of the
2608 // loop control flow instructions.
2610 case Instruction::PHI:{
2611 // Vectorize PHINodes.
2612 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2616 case Instruction::Add:
2617 case Instruction::FAdd:
2618 case Instruction::Sub:
2619 case Instruction::FSub:
2620 case Instruction::Mul:
2621 case Instruction::FMul:
2622 case Instruction::UDiv:
2623 case Instruction::SDiv:
2624 case Instruction::FDiv:
2625 case Instruction::URem:
2626 case Instruction::SRem:
2627 case Instruction::FRem:
2628 case Instruction::Shl:
2629 case Instruction::LShr:
2630 case Instruction::AShr:
2631 case Instruction::And:
2632 case Instruction::Or:
2633 case Instruction::Xor: {
2634 // Just widen binops.
2635 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2636 setDebugLocFromInst(Builder, BinOp);
2637 VectorParts &A = getVectorValue(it->getOperand(0));
2638 VectorParts &B = getVectorValue(it->getOperand(1));
2640 // Use this vector value for all users of the original instruction.
2641 for (unsigned Part = 0; Part < UF; ++Part) {
2642 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2644 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2645 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2646 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2647 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2648 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2650 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2651 VecOp->setIsExact(BinOp->isExact());
2657 case Instruction::Select: {
2659 // If the selector is loop invariant we can create a select
2660 // instruction with a scalar condition. Otherwise, use vector-select.
2661 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2663 setDebugLocFromInst(Builder, it);
2665 // The condition can be loop invariant but still defined inside the
2666 // loop. This means that we can't just use the original 'cond' value.
2667 // We have to take the 'vectorized' value and pick the first lane.
2668 // Instcombine will make this a no-op.
2669 VectorParts &Cond = getVectorValue(it->getOperand(0));
2670 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2671 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2673 Value *ScalarCond = (VF == 1) ? Cond[0] :
2674 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2676 for (unsigned Part = 0; Part < UF; ++Part) {
2677 Entry[Part] = Builder.CreateSelect(
2678 InvariantCond ? ScalarCond : Cond[Part],
2685 case Instruction::ICmp:
2686 case Instruction::FCmp: {
2687 // Widen compares. Generate vector compares.
2688 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2689 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2690 setDebugLocFromInst(Builder, it);
2691 VectorParts &A = getVectorValue(it->getOperand(0));
2692 VectorParts &B = getVectorValue(it->getOperand(1));
2693 for (unsigned Part = 0; Part < UF; ++Part) {
2696 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2698 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2704 case Instruction::Store:
2705 case Instruction::Load:
2706 vectorizeMemoryInstruction(it, Legal);
2708 case Instruction::ZExt:
2709 case Instruction::SExt:
2710 case Instruction::FPToUI:
2711 case Instruction::FPToSI:
2712 case Instruction::FPExt:
2713 case Instruction::PtrToInt:
2714 case Instruction::IntToPtr:
2715 case Instruction::SIToFP:
2716 case Instruction::UIToFP:
2717 case Instruction::Trunc:
2718 case Instruction::FPTrunc:
2719 case Instruction::BitCast: {
2720 CastInst *CI = dyn_cast<CastInst>(it);
2721 setDebugLocFromInst(Builder, it);
2722 /// Optimize the special case where the source is the induction
2723 /// variable. Notice that we can only optimize the 'trunc' case
2724 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2725 /// c. other casts depend on pointer size.
2726 if (CI->getOperand(0) == OldInduction &&
2727 it->getOpcode() == Instruction::Trunc) {
2728 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2730 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2731 for (unsigned Part = 0; Part < UF; ++Part)
2732 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2735 /// Vectorize casts.
2736 Type *DestTy = (VF == 1) ? CI->getType() :
2737 VectorType::get(CI->getType(), VF);
2739 VectorParts &A = getVectorValue(it->getOperand(0));
2740 for (unsigned Part = 0; Part < UF; ++Part)
2741 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2745 case Instruction::Call: {
2746 // Ignore dbg intrinsics.
2747 if (isa<DbgInfoIntrinsic>(it))
2749 setDebugLocFromInst(Builder, it);
2751 Module *M = BB->getParent()->getParent();
2752 CallInst *CI = cast<CallInst>(it);
2753 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2754 assert(ID && "Not an intrinsic call!");
2756 case Intrinsic::lifetime_end:
2757 case Intrinsic::lifetime_start:
2758 scalarizeInstruction(it);
2761 for (unsigned Part = 0; Part < UF; ++Part) {
2762 SmallVector<Value *, 4> Args;
2763 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2764 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2765 Args.push_back(Arg[Part]);
2767 Type *Tys[] = {CI->getType()};
2769 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2771 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2772 Entry[Part] = Builder.CreateCall(F, Args);
2780 // All other instructions are unsupported. Scalarize them.
2781 scalarizeInstruction(it);
2784 }// end of for_each instr.
2787 void InnerLoopVectorizer::updateAnalysis() {
2788 // Forget the original basic block.
2789 SE->forgetLoop(OrigLoop);
2791 // Update the dominator tree information.
2792 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2793 "Entry does not dominate exit.");
2795 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2796 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2797 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2798 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2799 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2800 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2801 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2802 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2804 DEBUG(DT->verifyAnalysis());
2807 /// \brief Check whether it is safe to if-convert this phi node.
2809 /// Phi nodes with constant expressions that can trap are not safe to if
2811 static bool canIfConvertPHINodes(BasicBlock *BB) {
2812 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2813 PHINode *Phi = dyn_cast<PHINode>(I);
2816 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
2817 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
2824 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2825 if (!EnableIfConversion)
2828 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2830 // A list of pointers that we can safely read and write to.
2831 SmallPtrSet<Value *, 8> SafePointes;
2833 // Collect safe addresses.
2834 for (Loop::block_iterator BI = TheLoop->block_begin(),
2835 BE = TheLoop->block_end(); BI != BE; ++BI) {
2836 BasicBlock *BB = *BI;
2838 if (blockNeedsPredication(BB))
2841 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2842 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2843 SafePointes.insert(LI->getPointerOperand());
2844 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2845 SafePointes.insert(SI->getPointerOperand());
2849 // Collect the blocks that need predication.
2850 BasicBlock *Header = TheLoop->getHeader();
2851 for (Loop::block_iterator BI = TheLoop->block_begin(),
2852 BE = TheLoop->block_end(); BI != BE; ++BI) {
2853 BasicBlock *BB = *BI;
2855 // We don't support switch statements inside loops.
2856 if (!isa<BranchInst>(BB->getTerminator()))
2859 // We must be able to predicate all blocks that need to be predicated.
2860 if (blockNeedsPredication(BB)) {
2861 if (!blockCanBePredicated(BB, SafePointes))
2863 } else if (BB != Header && !canIfConvertPHINodes(BB))
2868 // We can if-convert this loop.
2872 bool LoopVectorizationLegality::canVectorize() {
2873 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2874 // be canonicalized.
2875 if (!TheLoop->getLoopPreheader())
2878 // We can only vectorize innermost loops.
2879 if (TheLoop->getSubLoopsVector().size())
2882 // We must have a single backedge.
2883 if (TheLoop->getNumBackEdges() != 1)
2886 // We must have a single exiting block.
2887 if (!TheLoop->getExitingBlock())
2890 // We need to have a loop header.
2891 DEBUG(dbgs() << "LV: Found a loop: " <<
2892 TheLoop->getHeader()->getName() << '\n');
2894 // Check if we can if-convert non-single-bb loops.
2895 unsigned NumBlocks = TheLoop->getNumBlocks();
2896 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2897 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2901 // ScalarEvolution needs to be able to find the exit count.
2902 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2903 if (ExitCount == SE->getCouldNotCompute()) {
2904 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2908 // Do not loop-vectorize loops with a tiny trip count.
2909 BasicBlock *Latch = TheLoop->getLoopLatch();
2910 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2911 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2912 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2913 "This loop is not worth vectorizing.\n");
2917 // Check if we can vectorize the instructions and CFG in this loop.
2918 if (!canVectorizeInstrs()) {
2919 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2923 // Go over each instruction and look at memory deps.
2924 if (!canVectorizeMemory()) {
2925 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2929 // Collect all of the variables that remain uniform after vectorization.
2930 collectLoopUniforms();
2932 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2933 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2936 // Okay! We can vectorize. At this point we don't have any other mem analysis
2937 // which may limit our maximum vectorization factor, so just return true with
2942 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2943 if (Ty->isPointerTy())
2944 return DL.getIntPtrType(Ty);
2946 // It is possible that char's or short's overflow when we ask for the loop's
2947 // trip count, work around this by changing the type size.
2948 if (Ty->getScalarSizeInBits() < 32)
2949 return Type::getInt32Ty(Ty->getContext());
2954 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2955 Ty0 = convertPointerToIntegerType(DL, Ty0);
2956 Ty1 = convertPointerToIntegerType(DL, Ty1);
2957 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2962 /// \brief Check that the instruction has outside loop users and is not an
2963 /// identified reduction variable.
2964 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2965 SmallPtrSet<Value *, 4> &Reductions) {
2966 // Reduction instructions are allowed to have exit users. All other
2967 // instructions must not have external users.
2968 if (!Reductions.count(Inst))
2969 //Check that all of the users of the loop are inside the BB.
2970 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2972 Instruction *U = cast<Instruction>(*I);
2973 // This user may be a reduction exit value.
2974 if (!TheLoop->contains(U)) {
2975 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
2982 bool LoopVectorizationLegality::canVectorizeInstrs() {
2983 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2984 BasicBlock *Header = TheLoop->getHeader();
2986 // Look for the attribute signaling the absence of NaNs.
2987 Function &F = *Header->getParent();
2988 if (F.hasFnAttribute("no-nans-fp-math"))
2989 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2990 AttributeSet::FunctionIndex,
2991 "no-nans-fp-math").getValueAsString() == "true";
2993 // For each block in the loop.
2994 for (Loop::block_iterator bb = TheLoop->block_begin(),
2995 be = TheLoop->block_end(); bb != be; ++bb) {
2997 // Scan the instructions in the block and look for hazards.
2998 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3001 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3002 Type *PhiTy = Phi->getType();
3003 // Check that this PHI type is allowed.
3004 if (!PhiTy->isIntegerTy() &&
3005 !PhiTy->isFloatingPointTy() &&
3006 !PhiTy->isPointerTy()) {
3007 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3011 // If this PHINode is not in the header block, then we know that we
3012 // can convert it to select during if-conversion. No need to check if
3013 // the PHIs in this block are induction or reduction variables.
3014 if (*bb != Header) {
3015 // Check that this instruction has no outside users or is an
3016 // identified reduction value with an outside user.
3017 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3022 // We only allow if-converted PHIs with more than two incoming values.
3023 if (Phi->getNumIncomingValues() != 2) {
3024 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3028 // This is the value coming from the preheader.
3029 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3030 // Check if this is an induction variable.
3031 InductionKind IK = isInductionVariable(Phi);
3033 if (IK_NoInduction != IK) {
3034 // Get the widest type.
3036 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3038 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3040 // Int inductions are special because we only allow one IV.
3041 if (IK == IK_IntInduction) {
3042 // Use the phi node with the widest type as induction. Use the last
3043 // one if there are multiple (no good reason for doing this other
3044 // than it is expedient).
3045 if (!Induction || PhiTy == WidestIndTy)
3049 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3050 Inductions[Phi] = InductionInfo(StartValue, IK);
3052 // Until we explicitly handle the case of an induction variable with
3053 // an outside loop user we have to give up vectorizing this loop.
3054 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3060 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3061 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3064 if (AddReductionVar(Phi, RK_IntegerMult)) {
3065 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3068 if (AddReductionVar(Phi, RK_IntegerOr)) {
3069 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3072 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3073 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3076 if (AddReductionVar(Phi, RK_IntegerXor)) {
3077 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3080 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3081 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3084 if (AddReductionVar(Phi, RK_FloatMult)) {
3085 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3088 if (AddReductionVar(Phi, RK_FloatAdd)) {
3089 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3092 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3093 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3098 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3100 }// end of PHI handling
3102 // We still don't handle functions. However, we can ignore dbg intrinsic
3103 // calls and we do handle certain intrinsic and libm functions.
3104 CallInst *CI = dyn_cast<CallInst>(it);
3105 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3106 DEBUG(dbgs() << "LV: Found a call site.\n");
3110 // Check that the instruction return type is vectorizable.
3111 // Also, we can't vectorize extractelement instructions.
3112 if ((!VectorType::isValidElementType(it->getType()) &&
3113 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3114 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3118 // Check that the stored type is vectorizable.
3119 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3120 Type *T = ST->getValueOperand()->getType();
3121 if (!VectorType::isValidElementType(T))
3125 // Reduction instructions are allowed to have exit users.
3126 // All other instructions must not have external users.
3127 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3135 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3136 if (Inductions.empty())
3143 void LoopVectorizationLegality::collectLoopUniforms() {
3144 // We now know that the loop is vectorizable!
3145 // Collect variables that will remain uniform after vectorization.
3146 std::vector<Value*> Worklist;
3147 BasicBlock *Latch = TheLoop->getLoopLatch();
3149 // Start with the conditional branch and walk up the block.
3150 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3152 while (Worklist.size()) {
3153 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3154 Worklist.pop_back();
3156 // Look at instructions inside this loop.
3157 // Stop when reaching PHI nodes.
3158 // TODO: we need to follow values all over the loop, not only in this block.
3159 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3162 // This is a known uniform.
3165 // Insert all operands.
3166 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3171 /// \brief Analyses memory accesses in a loop.
3173 /// Checks whether run time pointer checks are needed and builds sets for data
3174 /// dependence checking.
3175 class AccessAnalysis {
3177 /// \brief Read or write access location.
3178 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3179 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3181 /// \brief Set of potential dependent memory accesses.
3182 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3184 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3185 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3186 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3188 /// \brief Register a load and whether it is only read from.
3189 void addLoad(Value *Ptr, bool IsReadOnly) {
3190 Accesses.insert(MemAccessInfo(Ptr, false));
3192 ReadOnlyPtr.insert(Ptr);
3195 /// \brief Register a store.
3196 void addStore(Value *Ptr) {
3197 Accesses.insert(MemAccessInfo(Ptr, true));
3200 /// \brief Check whether we can check the pointers at runtime for
3201 /// non-intersection.
3202 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3203 unsigned &NumComparisons, ScalarEvolution *SE,
3204 Loop *TheLoop, bool ShouldCheckStride = false);
3206 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3207 /// and builds sets of dependent accesses.
3208 void buildDependenceSets() {
3209 // Process read-write pointers first.
3210 processMemAccesses(false);
3211 // Next, process read pointers.
3212 processMemAccesses(true);
3215 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3217 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3218 void resetDepChecks() { CheckDeps.clear(); }
3220 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3223 typedef SetVector<MemAccessInfo> PtrAccessSet;
3224 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3226 /// \brief Go over all memory access or only the deferred ones if
3227 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3228 /// and build sets of dependency check candidates.
3229 void processMemAccesses(bool UseDeferred);
3231 /// Set of all accesses.
3232 PtrAccessSet Accesses;
3234 /// Set of access to check after all writes have been processed.
3235 PtrAccessSet DeferredAccesses;
3237 /// Map of pointers to last access encountered.
3238 UnderlyingObjToAccessMap ObjToLastAccess;
3240 /// Set of accesses that need a further dependence check.
3241 MemAccessInfoSet CheckDeps;
3243 /// Set of pointers that are read only.
3244 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3246 /// Set of underlying objects already written to.
3247 SmallPtrSet<Value*, 16> WriteObjects;
3251 /// Sets of potentially dependent accesses - members of one set share an
3252 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3253 /// dependence check.
3254 DepCandidates &DepCands;
3256 bool AreAllWritesIdentified;
3257 bool AreAllReadsIdentified;
3258 bool IsRTCheckNeeded;
3261 } // end anonymous namespace
3263 /// \brief Check whether a pointer can participate in a runtime bounds check.
3264 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3265 const SCEV *PtrScev = SE->getSCEV(Ptr);
3266 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3270 return AR->isAffine();
3273 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3274 /// the address space.
3275 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3278 bool AccessAnalysis::canCheckPtrAtRT(
3279 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3280 unsigned &NumComparisons, ScalarEvolution *SE,
3281 Loop *TheLoop, bool ShouldCheckStride) {
3282 // Find pointers with computable bounds. We are going to use this information
3283 // to place a runtime bound check.
3284 unsigned NumReadPtrChecks = 0;
3285 unsigned NumWritePtrChecks = 0;
3286 bool CanDoRT = true;
3288 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3289 // We assign consecutive id to access from different dependence sets.
3290 // Accesses within the same set don't need a runtime check.
3291 unsigned RunningDepId = 1;
3292 DenseMap<Value *, unsigned> DepSetId;
3294 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3296 const MemAccessInfo &Access = *AI;
3297 Value *Ptr = Access.getPointer();
3298 bool IsWrite = Access.getInt();
3300 // Just add write checks if we have both.
3301 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3305 ++NumWritePtrChecks;
3309 if (hasComputableBounds(SE, Ptr) &&
3310 // When we run after a failing dependency check we have to make sure we
3311 // don't have wrapping pointers.
3312 (!ShouldCheckStride || isStridedPtr(SE, DL, Ptr, TheLoop) == 1)) {
3313 // The id of the dependence set.
3316 if (IsDepCheckNeeded) {
3317 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3318 unsigned &LeaderId = DepSetId[Leader];
3320 LeaderId = RunningDepId++;
3323 // Each access has its own dependence set.
3324 DepId = RunningDepId++;
3326 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3328 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3334 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3335 NumComparisons = 0; // Only one dependence set.
3337 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3338 NumWritePtrChecks - 1));
3341 // If the pointers that we would use for the bounds comparison have different
3342 // address spaces, assume the values aren't directly comparable, so we can't
3343 // use them for the runtime check. We also have to assume they could
3344 // overlap. In the future there should be metadata for whether address spaces
3346 unsigned NumPointers = RtCheck.Pointers.size();
3347 for (unsigned i = 0; i < NumPointers; ++i) {
3348 for (unsigned j = i + 1; j < NumPointers; ++j) {
3349 // Only need to check pointers between two different dependency sets.
3350 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3353 Value *PtrI = RtCheck.Pointers[i];
3354 Value *PtrJ = RtCheck.Pointers[j];
3356 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3357 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3359 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3360 " different address spaces\n");
3369 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3370 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3373 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3374 // We process the set twice: first we process read-write pointers, last we
3375 // process read-only pointers. This allows us to skip dependence tests for
3376 // read-only pointers.
3378 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3379 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3380 const MemAccessInfo &Access = *AI;
3381 Value *Ptr = Access.getPointer();
3382 bool IsWrite = Access.getInt();
3384 DepCands.insert(Access);
3386 // Memorize read-only pointers for later processing and skip them in the
3387 // first round (they need to be checked after we have seen all write
3388 // pointers). Note: we also mark pointer that are not consecutive as
3389 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3390 // second check for "!IsWrite".
3391 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3392 if (!UseDeferred && IsReadOnlyPtr) {
3393 DeferredAccesses.insert(Access);
3397 bool NeedDepCheck = false;
3398 // Check whether there is the possiblity of dependency because of underlying
3399 // objects being the same.
3400 typedef SmallVector<Value*, 16> ValueVector;
3401 ValueVector TempObjects;
3402 GetUnderlyingObjects(Ptr, TempObjects, DL);
3403 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3405 Value *UnderlyingObj = *UI;
3407 // If this is a write then it needs to be an identified object. If this a
3408 // read and all writes (so far) are identified function scope objects we
3409 // don't need an identified underlying object but only an Argument (the
3410 // next write is going to invalidate this assumption if it is
3412 // This is a micro-optimization for the case where all writes are
3413 // identified and we have one argument pointer.
3414 // Otherwise, we do need a runtime check.
3415 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3416 (!IsWrite && (!AreAllWritesIdentified ||
3417 !isa<Argument>(UnderlyingObj)) &&
3418 !isIdentifiedObject(UnderlyingObj))) {
3419 DEBUG(dbgs() << "LV: Found an unidentified " <<
3420 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3422 IsRTCheckNeeded = (IsRTCheckNeeded ||
3423 !isIdentifiedObject(UnderlyingObj) ||
3424 !AreAllReadsIdentified);
3427 AreAllWritesIdentified = false;
3429 AreAllReadsIdentified = false;
3432 // If this is a write - check other reads and writes for conflicts. If
3433 // this is a read only check other writes for conflicts (but only if there
3434 // is no other write to the ptr - this is an optimization to catch "a[i] =
3435 // a[i] + " without having to do a dependence check).
3436 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3437 NeedDepCheck = true;
3440 WriteObjects.insert(UnderlyingObj);
3442 // Create sets of pointers connected by shared underlying objects.
3443 UnderlyingObjToAccessMap::iterator Prev =
3444 ObjToLastAccess.find(UnderlyingObj);
3445 if (Prev != ObjToLastAccess.end())
3446 DepCands.unionSets(Access, Prev->second);
3448 ObjToLastAccess[UnderlyingObj] = Access;
3452 CheckDeps.insert(Access);
3457 /// \brief Checks memory dependences among accesses to the same underlying
3458 /// object to determine whether there vectorization is legal or not (and at
3459 /// which vectorization factor).
3461 /// This class works under the assumption that we already checked that memory
3462 /// locations with different underlying pointers are "must-not alias".
3463 /// We use the ScalarEvolution framework to symbolically evalutate access
3464 /// functions pairs. Since we currently don't restructure the loop we can rely
3465 /// on the program order of memory accesses to determine their safety.
3466 /// At the moment we will only deem accesses as safe for:
3467 /// * A negative constant distance assuming program order.
3469 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3470 /// a[i] = tmp; y = a[i];
3472 /// The latter case is safe because later checks guarantuee that there can't
3473 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3474 /// the same variable: a header phi can only be an induction or a reduction, a
3475 /// reduction can't have a memory sink, an induction can't have a memory
3476 /// source). This is important and must not be violated (or we have to
3477 /// resort to checking for cycles through memory).
3479 /// * A positive constant distance assuming program order that is bigger
3480 /// than the biggest memory access.
3482 /// tmp = a[i] OR b[i] = x
3483 /// a[i+2] = tmp y = b[i+2];
3485 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3487 /// * Zero distances and all accesses have the same size.
3489 class MemoryDepChecker {
3491 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3492 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3494 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3495 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3496 ShouldRetryWithRuntimeCheck(false) {}
3498 /// \brief Register the location (instructions are given increasing numbers)
3499 /// of a write access.
3500 void addAccess(StoreInst *SI) {
3501 Value *Ptr = SI->getPointerOperand();
3502 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3503 InstMap.push_back(SI);
3507 /// \brief Register the location (instructions are given increasing numbers)
3508 /// of a write access.
3509 void addAccess(LoadInst *LI) {
3510 Value *Ptr = LI->getPointerOperand();
3511 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3512 InstMap.push_back(LI);
3516 /// \brief Check whether the dependencies between the accesses are safe.
3518 /// Only checks sets with elements in \p CheckDeps.
3519 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3520 MemAccessInfoSet &CheckDeps);
3522 /// \brief The maximum number of bytes of a vector register we can vectorize
3523 /// the accesses safely with.
3524 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3526 /// \brief In same cases when the dependency check fails we can still
3527 /// vectorize the loop with a dynamic array access check.
3528 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3531 ScalarEvolution *SE;
3533 const Loop *InnermostLoop;
3535 /// \brief Maps access locations (ptr, read/write) to program order.
3536 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3538 /// \brief Memory access instructions in program order.
3539 SmallVector<Instruction *, 16> InstMap;
3541 /// \brief The program order index to be used for the next instruction.
3544 // We can access this many bytes in parallel safely.
3545 unsigned MaxSafeDepDistBytes;
3547 /// \brief If we see a non-constant dependence distance we can still try to
3548 /// vectorize this loop with runtime checks.
3549 bool ShouldRetryWithRuntimeCheck;
3551 /// \brief Check whether there is a plausible dependence between the two
3554 /// Access \p A must happen before \p B in program order. The two indices
3555 /// identify the index into the program order map.
3557 /// This function checks whether there is a plausible dependence (or the
3558 /// absence of such can't be proved) between the two accesses. If there is a
3559 /// plausible dependence but the dependence distance is bigger than one
3560 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3561 /// distance is smaller than any other distance encountered so far).
3562 /// Otherwise, this function returns true signaling a possible dependence.
3563 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3564 const MemAccessInfo &B, unsigned BIdx);
3566 /// \brief Check whether the data dependence could prevent store-load
3568 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3571 } // end anonymous namespace
3573 static bool isInBoundsGep(Value *Ptr) {
3574 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3575 return GEP->isInBounds();
3579 /// \brief Check whether the access through \p Ptr has a constant stride.
3580 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3582 const Type *Ty = Ptr->getType();
3583 assert(Ty->isPointerTy() && "Unexpected non-ptr");
3585 // Make sure that the pointer does not point to aggregate types.
3586 const PointerType *PtrTy = cast<PointerType>(Ty);
3587 if (PtrTy->getElementType()->isAggregateType()) {
3588 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3593 const SCEV *PtrScev = SE->getSCEV(Ptr);
3594 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3596 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3597 << *Ptr << " SCEV: " << *PtrScev << "\n");
3601 // The accesss function must stride over the innermost loop.
3602 if (Lp != AR->getLoop()) {
3603 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3604 *Ptr << " SCEV: " << *PtrScev << "\n");
3607 // The address calculation must not wrap. Otherwise, a dependence could be
3609 // An inbounds getelementptr that is a AddRec with a unit stride
3610 // cannot wrap per definition. The unit stride requirement is checked later.
3611 // An getelementptr without an inbounds attribute and unit stride would have
3612 // to access the pointer value "0" which is undefined behavior in address
3613 // space 0, therefore we can also vectorize this case.
3614 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3615 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3616 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3617 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3618 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3619 << *Ptr << " SCEV: " << *PtrScev << "\n");
3623 // Check the step is constant.
3624 const SCEV *Step = AR->getStepRecurrence(*SE);
3626 // Calculate the pointer stride and check if it is consecutive.
3627 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3629 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3630 " SCEV: " << *PtrScev << "\n");
3634 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3635 const APInt &APStepVal = C->getValue()->getValue();
3637 // Huge step value - give up.
3638 if (APStepVal.getBitWidth() > 64)
3641 int64_t StepVal = APStepVal.getSExtValue();
3644 int64_t Stride = StepVal / Size;
3645 int64_t Rem = StepVal % Size;
3649 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3650 // know we can't "wrap around the address space". In case of address space
3651 // zero we know that this won't happen without triggering undefined behavior.
3652 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3653 Stride != 1 && Stride != -1)
3659 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3660 unsigned TypeByteSize) {
3661 // If loads occur at a distance that is not a multiple of a feasible vector
3662 // factor store-load forwarding does not take place.
3663 // Positive dependences might cause troubles because vectorizing them might
3664 // prevent store-load forwarding making vectorized code run a lot slower.
3665 // a[i] = a[i-3] ^ a[i-8];
3666 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3667 // hence on your typical architecture store-load forwarding does not take
3668 // place. Vectorizing in such cases does not make sense.
3669 // Store-load forwarding distance.
3670 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3671 // Maximum vector factor.
3672 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3673 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3674 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3676 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3678 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3679 MaxVFWithoutSLForwardIssues = (vf >>=1);
3684 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3685 DEBUG(dbgs() << "LV: Distance " << Distance <<
3686 " that could cause a store-load forwarding conflict\n");
3690 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3691 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3692 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3696 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3697 const MemAccessInfo &B, unsigned BIdx) {
3698 assert (AIdx < BIdx && "Must pass arguments in program order");
3700 Value *APtr = A.getPointer();
3701 Value *BPtr = B.getPointer();
3702 bool AIsWrite = A.getInt();
3703 bool BIsWrite = B.getInt();
3705 // Two reads are independent.
3706 if (!AIsWrite && !BIsWrite)
3709 const SCEV *AScev = SE->getSCEV(APtr);
3710 const SCEV *BScev = SE->getSCEV(BPtr);
3712 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3713 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3715 const SCEV *Src = AScev;
3716 const SCEV *Sink = BScev;
3718 // If the induction step is negative we have to invert source and sink of the
3720 if (StrideAPtr < 0) {
3723 std::swap(APtr, BPtr);
3724 std::swap(Src, Sink);
3725 std::swap(AIsWrite, BIsWrite);
3726 std::swap(AIdx, BIdx);
3727 std::swap(StrideAPtr, StrideBPtr);
3730 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3732 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3733 << "(Induction step: " << StrideAPtr << ")\n");
3734 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3735 << *InstMap[BIdx] << ": " << *Dist << "\n");
3737 // Need consecutive accesses. We don't want to vectorize
3738 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3739 // the address space.
3740 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3741 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3745 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3747 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
3748 ShouldRetryWithRuntimeCheck = true;
3752 Type *ATy = APtr->getType()->getPointerElementType();
3753 Type *BTy = BPtr->getType()->getPointerElementType();
3754 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3756 // Negative distances are not plausible dependencies.
3757 const APInt &Val = C->getValue()->getValue();
3758 if (Val.isNegative()) {
3759 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3760 if (IsTrueDataDependence &&
3761 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3765 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3769 // Write to the same location with the same size.
3770 // Could be improved to assert type sizes are the same (i32 == float, etc).
3774 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
3778 assert(Val.isStrictlyPositive() && "Expect a positive value");
3780 // Positive distance bigger than max vectorization factor.
3783 "LV: ReadWrite-Write positive dependency with different types\n");
3787 unsigned Distance = (unsigned) Val.getZExtValue();
3789 // Bail out early if passed-in parameters make vectorization not feasible.
3790 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3791 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3793 // The distance must be bigger than the size needed for a vectorized version
3794 // of the operation and the size of the vectorized operation must not be
3795 // bigger than the currrent maximum size.
3796 if (Distance < 2*TypeByteSize ||
3797 2*TypeByteSize > MaxSafeDepDistBytes ||
3798 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3799 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3800 << Val.getSExtValue() << '\n');
3804 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3805 Distance : MaxSafeDepDistBytes;
3807 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3808 if (IsTrueDataDependence &&
3809 couldPreventStoreLoadForward(Distance, TypeByteSize))
3812 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3813 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
3819 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3820 MemAccessInfoSet &CheckDeps) {
3822 MaxSafeDepDistBytes = -1U;
3823 while (!CheckDeps.empty()) {
3824 MemAccessInfo CurAccess = *CheckDeps.begin();
3826 // Get the relevant memory access set.
3827 EquivalenceClasses<MemAccessInfo>::iterator I =
3828 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3830 // Check accesses within this set.
3831 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3832 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3834 // Check every access pair.
3836 CheckDeps.erase(*AI);
3837 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3839 // Check every accessing instruction pair in program order.
3840 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3841 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3842 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3843 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3844 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3846 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3857 bool LoopVectorizationLegality::canVectorizeMemory() {
3859 typedef SmallVector<Value*, 16> ValueVector;
3860 typedef SmallPtrSet<Value*, 16> ValueSet;
3862 // Holds the Load and Store *instructions*.
3866 // Holds all the different accesses in the loop.
3867 unsigned NumReads = 0;
3868 unsigned NumReadWrites = 0;
3870 PtrRtCheck.Pointers.clear();
3871 PtrRtCheck.Need = false;
3873 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3874 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3877 for (Loop::block_iterator bb = TheLoop->block_begin(),
3878 be = TheLoop->block_end(); bb != be; ++bb) {
3880 // Scan the BB and collect legal loads and stores.
3881 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3884 // If this is a load, save it. If this instruction can read from memory
3885 // but is not a load, then we quit. Notice that we don't handle function
3886 // calls that read or write.
3887 if (it->mayReadFromMemory()) {
3888 // Many math library functions read the rounding mode. We will only
3889 // vectorize a loop if it contains known function calls that don't set
3890 // the flag. Therefore, it is safe to ignore this read from memory.
3891 CallInst *Call = dyn_cast<CallInst>(it);
3892 if (Call && getIntrinsicIDForCall(Call, TLI))
3895 LoadInst *Ld = dyn_cast<LoadInst>(it);
3896 if (!Ld) return false;
3897 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3898 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3901 Loads.push_back(Ld);
3902 DepChecker.addAccess(Ld);
3906 // Save 'store' instructions. Abort if other instructions write to memory.
3907 if (it->mayWriteToMemory()) {
3908 StoreInst *St = dyn_cast<StoreInst>(it);
3909 if (!St) return false;
3910 if (!St->isSimple() && !IsAnnotatedParallel) {
3911 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3914 Stores.push_back(St);
3915 DepChecker.addAccess(St);
3920 // Now we have two lists that hold the loads and the stores.
3921 // Next, we find the pointers that they use.
3923 // Check if we see any stores. If there are no stores, then we don't
3924 // care if the pointers are *restrict*.
3925 if (!Stores.size()) {
3926 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3930 AccessAnalysis::DepCandidates DependentAccesses;
3931 AccessAnalysis Accesses(DL, DependentAccesses);
3933 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3934 // multiple times on the same object. If the ptr is accessed twice, once
3935 // for read and once for write, it will only appear once (on the write
3936 // list). This is okay, since we are going to check for conflicts between
3937 // writes and between reads and writes, but not between reads and reads.
3940 ValueVector::iterator I, IE;
3941 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3942 StoreInst *ST = cast<StoreInst>(*I);
3943 Value* Ptr = ST->getPointerOperand();
3945 if (isUniform(Ptr)) {
3946 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3950 // If we did *not* see this pointer before, insert it to the read-write
3951 // list. At this phase it is only a 'write' list.
3952 if (Seen.insert(Ptr)) {
3954 Accesses.addStore(Ptr);
3958 if (IsAnnotatedParallel) {
3960 << "LV: A loop annotated parallel, ignore memory dependency "
3965 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3966 LoadInst *LD = cast<LoadInst>(*I);
3967 Value* Ptr = LD->getPointerOperand();
3968 // If we did *not* see this pointer before, insert it to the
3969 // read list. If we *did* see it before, then it is already in
3970 // the read-write list. This allows us to vectorize expressions
3971 // such as A[i] += x; Because the address of A[i] is a read-write
3972 // pointer. This only works if the index of A[i] is consecutive.
3973 // If the address of i is unknown (for example A[B[i]]) then we may
3974 // read a few words, modify, and write a few words, and some of the
3975 // words may be written to the same address.
3976 bool IsReadOnlyPtr = false;
3977 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3979 IsReadOnlyPtr = true;
3981 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3984 // If we write (or read-write) to a single destination and there are no
3985 // other reads in this loop then is it safe to vectorize.
3986 if (NumReadWrites == 1 && NumReads == 0) {
3987 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3991 // Build dependence sets and check whether we need a runtime pointer bounds
3993 Accesses.buildDependenceSets();
3994 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3996 // Find pointers with computable bounds. We are going to use this information
3997 // to place a runtime bound check.
3998 unsigned NumComparisons = 0;
3999 bool CanDoRT = false;
4001 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
4004 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4005 " pointer comparisons.\n");
4007 // If we only have one set of dependences to check pointers among we don't
4008 // need a runtime check.
4009 if (NumComparisons == 0 && NeedRTCheck)
4010 NeedRTCheck = false;
4012 // Check that we did not collect too many pointers or found an unsizeable
4014 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4020 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4023 if (NeedRTCheck && !CanDoRT) {
4024 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4025 "the array bounds.\n");
4030 PtrRtCheck.Need = NeedRTCheck;
4032 bool CanVecMem = true;
4033 if (Accesses.isDependencyCheckNeeded()) {
4034 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4035 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
4036 Accesses.getDependenciesToCheck());
4037 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4039 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4040 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4043 // Clear the dependency checks. We assume they are not needed.
4044 Accesses.resetDepChecks();
4047 PtrRtCheck.Need = true;
4049 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4051 // Check that we did not collect too many pointers or found an unsizeable
4053 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4054 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4063 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4064 " need a runtime memory check.\n");
4069 static bool hasMultipleUsesOf(Instruction *I,
4070 SmallPtrSet<Instruction *, 8> &Insts) {
4071 unsigned NumUses = 0;
4072 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4073 if (Insts.count(dyn_cast<Instruction>(*Use)))
4082 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4083 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4084 if (!Set.count(dyn_cast<Instruction>(*Use)))
4089 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4090 ReductionKind Kind) {
4091 if (Phi->getNumIncomingValues() != 2)
4094 // Reduction variables are only found in the loop header block.
4095 if (Phi->getParent() != TheLoop->getHeader())
4098 // Obtain the reduction start value from the value that comes from the loop
4100 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4102 // ExitInstruction is the single value which is used outside the loop.
4103 // We only allow for a single reduction value to be used outside the loop.
4104 // This includes users of the reduction, variables (which form a cycle
4105 // which ends in the phi node).
4106 Instruction *ExitInstruction = 0;
4107 // Indicates that we found a reduction operation in our scan.
4108 bool FoundReduxOp = false;
4110 // We start with the PHI node and scan for all of the users of this
4111 // instruction. All users must be instructions that can be used as reduction
4112 // variables (such as ADD). We must have a single out-of-block user. The cycle
4113 // must include the original PHI.
4114 bool FoundStartPHI = false;
4116 // To recognize min/max patterns formed by a icmp select sequence, we store
4117 // the number of instruction we saw from the recognized min/max pattern,
4118 // to make sure we only see exactly the two instructions.
4119 unsigned NumCmpSelectPatternInst = 0;
4120 ReductionInstDesc ReduxDesc(false, 0);
4122 SmallPtrSet<Instruction *, 8> VisitedInsts;
4123 SmallVector<Instruction *, 8> Worklist;
4124 Worklist.push_back(Phi);
4125 VisitedInsts.insert(Phi);
4127 // A value in the reduction can be used:
4128 // - By the reduction:
4129 // - Reduction operation:
4130 // - One use of reduction value (safe).
4131 // - Multiple use of reduction value (not safe).
4133 // - All uses of the PHI must be the reduction (safe).
4134 // - Otherwise, not safe.
4135 // - By one instruction outside of the loop (safe).
4136 // - By further instructions outside of the loop (not safe).
4137 // - By an instruction that is not part of the reduction (not safe).
4139 // * An instruction type other than PHI or the reduction operation.
4140 // * A PHI in the header other than the initial PHI.
4141 while (!Worklist.empty()) {
4142 Instruction *Cur = Worklist.back();
4143 Worklist.pop_back();
4146 // If the instruction has no users then this is a broken chain and can't be
4147 // a reduction variable.
4148 if (Cur->use_empty())
4151 bool IsAPhi = isa<PHINode>(Cur);
4153 // A header PHI use other than the original PHI.
4154 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4157 // Reductions of instructions such as Div, and Sub is only possible if the
4158 // LHS is the reduction variable.
4159 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4160 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4161 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4164 // Any reduction instruction must be of one of the allowed kinds.
4165 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4166 if (!ReduxDesc.IsReduction)
4169 // A reduction operation must only have one use of the reduction value.
4170 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4171 hasMultipleUsesOf(Cur, VisitedInsts))
4174 // All inputs to a PHI node must be a reduction value.
4175 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4178 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4179 isa<SelectInst>(Cur)))
4180 ++NumCmpSelectPatternInst;
4181 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4182 isa<SelectInst>(Cur)))
4183 ++NumCmpSelectPatternInst;
4185 // Check whether we found a reduction operator.
4186 FoundReduxOp |= !IsAPhi;
4188 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4189 // onto the stack. This way we are going to have seen all inputs to PHI
4190 // nodes once we get to them.
4191 SmallVector<Instruction *, 8> NonPHIs;
4192 SmallVector<Instruction *, 8> PHIs;
4193 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4195 Instruction *Usr = cast<Instruction>(*UI);
4197 // Check if we found the exit user.
4198 BasicBlock *Parent = Usr->getParent();
4199 if (!TheLoop->contains(Parent)) {
4200 // Exit if you find multiple outside users or if the header phi node is
4201 // being used. In this case the user uses the value of the previous
4202 // iteration, in which case we would loose "VF-1" iterations of the
4203 // reduction operation if we vectorize.
4204 if (ExitInstruction != 0 || Cur == Phi)
4207 // The instruction used by an outside user must be the last instruction
4208 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4209 // operations on the value.
4210 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4213 ExitInstruction = Cur;
4217 // Process instructions only once (termination).
4218 if (VisitedInsts.insert(Usr)) {
4219 if (isa<PHINode>(Usr))
4220 PHIs.push_back(Usr);
4222 NonPHIs.push_back(Usr);
4224 // Remember that we completed the cycle.
4226 FoundStartPHI = true;
4228 Worklist.append(PHIs.begin(), PHIs.end());
4229 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4232 // This means we have seen one but not the other instruction of the
4233 // pattern or more than just a select and cmp.
4234 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4235 NumCmpSelectPatternInst != 2)
4238 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4241 // We found a reduction var if we have reached the original phi node and we
4242 // only have a single instruction with out-of-loop users.
4244 // This instruction is allowed to have out-of-loop users.
4245 AllowedExit.insert(ExitInstruction);
4247 // Save the description of this reduction variable.
4248 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4249 ReduxDesc.MinMaxKind);
4250 Reductions[Phi] = RD;
4251 // We've ended the cycle. This is a reduction variable if we have an
4252 // outside user and it has a binary op.
4257 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4258 /// pattern corresponding to a min(X, Y) or max(X, Y).
4259 LoopVectorizationLegality::ReductionInstDesc
4260 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4261 ReductionInstDesc &Prev) {
4263 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4264 "Expect a select instruction");
4265 Instruction *Cmp = 0;
4266 SelectInst *Select = 0;
4268 // We must handle the select(cmp()) as a single instruction. Advance to the
4270 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4271 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4272 return ReductionInstDesc(false, I);
4273 return ReductionInstDesc(Select, Prev.MinMaxKind);
4276 // Only handle single use cases for now.
4277 if (!(Select = dyn_cast<SelectInst>(I)))
4278 return ReductionInstDesc(false, I);
4279 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4280 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4281 return ReductionInstDesc(false, I);
4282 if (!Cmp->hasOneUse())
4283 return ReductionInstDesc(false, I);
4288 // Look for a min/max pattern.
4289 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4290 return ReductionInstDesc(Select, MRK_UIntMin);
4291 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4292 return ReductionInstDesc(Select, MRK_UIntMax);
4293 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4294 return ReductionInstDesc(Select, MRK_SIntMax);
4295 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4296 return ReductionInstDesc(Select, MRK_SIntMin);
4297 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4298 return ReductionInstDesc(Select, MRK_FloatMin);
4299 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4300 return ReductionInstDesc(Select, MRK_FloatMax);
4301 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4302 return ReductionInstDesc(Select, MRK_FloatMin);
4303 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4304 return ReductionInstDesc(Select, MRK_FloatMax);
4306 return ReductionInstDesc(false, I);
4309 LoopVectorizationLegality::ReductionInstDesc
4310 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4312 ReductionInstDesc &Prev) {
4313 bool FP = I->getType()->isFloatingPointTy();
4314 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4315 switch (I->getOpcode()) {
4317 return ReductionInstDesc(false, I);
4318 case Instruction::PHI:
4319 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4320 Kind != RK_FloatMinMax))
4321 return ReductionInstDesc(false, I);
4322 return ReductionInstDesc(I, Prev.MinMaxKind);
4323 case Instruction::Sub:
4324 case Instruction::Add:
4325 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4326 case Instruction::Mul:
4327 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4328 case Instruction::And:
4329 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4330 case Instruction::Or:
4331 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4332 case Instruction::Xor:
4333 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4334 case Instruction::FMul:
4335 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4336 case Instruction::FAdd:
4337 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4338 case Instruction::FCmp:
4339 case Instruction::ICmp:
4340 case Instruction::Select:
4341 if (Kind != RK_IntegerMinMax &&
4342 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4343 return ReductionInstDesc(false, I);
4344 return isMinMaxSelectCmpPattern(I, Prev);
4348 LoopVectorizationLegality::InductionKind
4349 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4350 Type *PhiTy = Phi->getType();
4351 // We only handle integer and pointer inductions variables.
4352 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4353 return IK_NoInduction;
4355 // Check that the PHI is consecutive.
4356 const SCEV *PhiScev = SE->getSCEV(Phi);
4357 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4359 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4360 return IK_NoInduction;
4362 const SCEV *Step = AR->getStepRecurrence(*SE);
4364 // Integer inductions need to have a stride of one.
4365 if (PhiTy->isIntegerTy()) {
4367 return IK_IntInduction;
4368 if (Step->isAllOnesValue())
4369 return IK_ReverseIntInduction;
4370 return IK_NoInduction;
4373 // Calculate the pointer stride and check if it is consecutive.
4374 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4376 return IK_NoInduction;
4378 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4379 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4380 if (C->getValue()->equalsInt(Size))
4381 return IK_PtrInduction;
4382 else if (C->getValue()->equalsInt(0 - Size))
4383 return IK_ReversePtrInduction;
4385 return IK_NoInduction;
4388 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4389 Value *In0 = const_cast<Value*>(V);
4390 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4394 return Inductions.count(PN);
4397 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4398 assert(TheLoop->contains(BB) && "Unknown block used");
4400 // Blocks that do not dominate the latch need predication.
4401 BasicBlock* Latch = TheLoop->getLoopLatch();
4402 return !DT->dominates(BB, Latch);
4405 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4406 SmallPtrSet<Value *, 8>& SafePtrs) {
4407 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4408 // We might be able to hoist the load.
4409 if (it->mayReadFromMemory()) {
4410 LoadInst *LI = dyn_cast<LoadInst>(it);
4411 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4415 // We don't predicate stores at the moment.
4416 if (it->mayWriteToMemory() || it->mayThrow())
4419 // Check that we don't have a constant expression that can trap as operand.
4420 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4422 if (Constant *C = dyn_cast<Constant>(*OI))
4427 // The instructions below can trap.
4428 switch (it->getOpcode()) {
4430 case Instruction::UDiv:
4431 case Instruction::SDiv:
4432 case Instruction::URem:
4433 case Instruction::SRem:
4441 LoopVectorizationCostModel::VectorizationFactor
4442 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4444 // Width 1 means no vectorize
4445 VectorizationFactor Factor = { 1U, 0U };
4446 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4447 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4451 // Find the trip count.
4452 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4453 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4455 unsigned WidestType = getWidestType();
4456 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4457 unsigned MaxSafeDepDist = -1U;
4458 if (Legal->getMaxSafeDepDistBytes() != -1U)
4459 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4460 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4461 WidestRegister : MaxSafeDepDist);
4462 unsigned MaxVectorSize = WidestRegister / WidestType;
4463 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4464 DEBUG(dbgs() << "LV: The Widest register is: "
4465 << WidestRegister << " bits.\n");
4467 if (MaxVectorSize == 0) {
4468 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4472 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4473 " into one vector!");
4475 unsigned VF = MaxVectorSize;
4477 // If we optimize the program for size, avoid creating the tail loop.
4479 // If we are unable to calculate the trip count then don't try to vectorize.
4481 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4485 // Find the maximum SIMD width that can fit within the trip count.
4486 VF = TC % MaxVectorSize;
4491 // If the trip count that we found modulo the vectorization factor is not
4492 // zero then we require a tail.
4494 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4500 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4501 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4503 Factor.Width = UserVF;
4507 float Cost = expectedCost(1);
4509 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
4510 for (unsigned i=2; i <= VF; i*=2) {
4511 // Notice that the vector loop needs to be executed less times, so
4512 // we need to divide the cost of the vector loops by the width of
4513 // the vector elements.
4514 float VectorCost = expectedCost(i) / (float)i;
4515 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4516 (int)VectorCost << ".\n");
4517 if (VectorCost < Cost) {
4523 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4524 Factor.Width = Width;
4525 Factor.Cost = Width * Cost;
4529 unsigned LoopVectorizationCostModel::getWidestType() {
4530 unsigned MaxWidth = 8;
4533 for (Loop::block_iterator bb = TheLoop->block_begin(),
4534 be = TheLoop->block_end(); bb != be; ++bb) {
4535 BasicBlock *BB = *bb;
4537 // For each instruction in the loop.
4538 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4539 Type *T = it->getType();
4541 // Only examine Loads, Stores and PHINodes.
4542 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4545 // Examine PHI nodes that are reduction variables.
4546 if (PHINode *PN = dyn_cast<PHINode>(it))
4547 if (!Legal->getReductionVars()->count(PN))
4550 // Examine the stored values.
4551 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4552 T = ST->getValueOperand()->getType();
4554 // Ignore loaded pointer types and stored pointer types that are not
4555 // consecutive. However, we do want to take consecutive stores/loads of
4556 // pointer vectors into account.
4557 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4560 MaxWidth = std::max(MaxWidth,
4561 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4569 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4572 unsigned LoopCost) {
4574 // -- The unroll heuristics --
4575 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4576 // There are many micro-architectural considerations that we can't predict
4577 // at this level. For example frontend pressure (on decode or fetch) due to
4578 // code size, or the number and capabilities of the execution ports.
4580 // We use the following heuristics to select the unroll factor:
4581 // 1. If the code has reductions the we unroll in order to break the cross
4582 // iteration dependency.
4583 // 2. If the loop is really small then we unroll in order to reduce the loop
4585 // 3. We don't unroll if we think that we will spill registers to memory due
4586 // to the increased register pressure.
4588 // Use the user preference, unless 'auto' is selected.
4592 // When we optimize for size we don't unroll.
4596 // We used the distance for the unroll factor.
4597 if (Legal->getMaxSafeDepDistBytes() != -1U)
4600 // Do not unroll loops with a relatively small trip count.
4601 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4602 TheLoop->getLoopLatch());
4603 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4606 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4607 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4608 " vector registers\n");
4610 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4611 // We divide by these constants so assume that we have at least one
4612 // instruction that uses at least one register.
4613 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4614 R.NumInstructions = std::max(R.NumInstructions, 1U);
4616 // We calculate the unroll factor using the following formula.
4617 // Subtract the number of loop invariants from the number of available
4618 // registers. These registers are used by all of the unrolled instances.
4619 // Next, divide the remaining registers by the number of registers that is
4620 // required by the loop, in order to estimate how many parallel instances
4621 // fit without causing spills.
4622 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4624 // Clamp the unroll factor ranges to reasonable factors.
4625 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4627 // If we did not calculate the cost for VF (because the user selected the VF)
4628 // then we calculate the cost of VF here.
4630 LoopCost = expectedCost(VF);
4632 // Clamp the calculated UF to be between the 1 and the max unroll factor
4633 // that the target allows.
4634 if (UF > MaxUnrollSize)
4639 bool HasReductions = Legal->getReductionVars()->size();
4641 // Decide if we want to unroll if we decided that it is legal to vectorize
4642 // but not profitable.
4644 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4645 LoopCost > SmallLoopCost)
4651 if (HasReductions) {
4652 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4656 // We want to unroll tiny loops in order to reduce the loop overhead.
4657 // We assume that the cost overhead is 1 and we use the cost model
4658 // to estimate the cost of the loop and unroll until the cost of the
4659 // loop overhead is about 5% of the cost of the loop.
4660 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4661 if (LoopCost < SmallLoopCost) {
4662 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4663 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4664 return std::min(NewUF, UF);
4667 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4671 LoopVectorizationCostModel::RegisterUsage
4672 LoopVectorizationCostModel::calculateRegisterUsage() {
4673 // This function calculates the register usage by measuring the highest number
4674 // of values that are alive at a single location. Obviously, this is a very
4675 // rough estimation. We scan the loop in a topological order in order and
4676 // assign a number to each instruction. We use RPO to ensure that defs are
4677 // met before their users. We assume that each instruction that has in-loop
4678 // users starts an interval. We record every time that an in-loop value is
4679 // used, so we have a list of the first and last occurrences of each
4680 // instruction. Next, we transpose this data structure into a multi map that
4681 // holds the list of intervals that *end* at a specific location. This multi
4682 // map allows us to perform a linear search. We scan the instructions linearly
4683 // and record each time that a new interval starts, by placing it in a set.
4684 // If we find this value in the multi-map then we remove it from the set.
4685 // The max register usage is the maximum size of the set.
4686 // We also search for instructions that are defined outside the loop, but are
4687 // used inside the loop. We need this number separately from the max-interval
4688 // usage number because when we unroll, loop-invariant values do not take
4690 LoopBlocksDFS DFS(TheLoop);
4694 R.NumInstructions = 0;
4696 // Each 'key' in the map opens a new interval. The values
4697 // of the map are the index of the 'last seen' usage of the
4698 // instruction that is the key.
4699 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4700 // Maps instruction to its index.
4701 DenseMap<unsigned, Instruction*> IdxToInstr;
4702 // Marks the end of each interval.
4703 IntervalMap EndPoint;
4704 // Saves the list of instruction indices that are used in the loop.
4705 SmallSet<Instruction*, 8> Ends;
4706 // Saves the list of values that are used in the loop but are
4707 // defined outside the loop, such as arguments and constants.
4708 SmallPtrSet<Value*, 8> LoopInvariants;
4711 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4712 be = DFS.endRPO(); bb != be; ++bb) {
4713 R.NumInstructions += (*bb)->size();
4714 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4716 Instruction *I = it;
4717 IdxToInstr[Index++] = I;
4719 // Save the end location of each USE.
4720 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4721 Value *U = I->getOperand(i);
4722 Instruction *Instr = dyn_cast<Instruction>(U);
4724 // Ignore non-instruction values such as arguments, constants, etc.
4725 if (!Instr) continue;
4727 // If this instruction is outside the loop then record it and continue.
4728 if (!TheLoop->contains(Instr)) {
4729 LoopInvariants.insert(Instr);
4733 // Overwrite previous end points.
4734 EndPoint[Instr] = Index;
4740 // Saves the list of intervals that end with the index in 'key'.
4741 typedef SmallVector<Instruction*, 2> InstrList;
4742 DenseMap<unsigned, InstrList> TransposeEnds;
4744 // Transpose the EndPoints to a list of values that end at each index.
4745 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4747 TransposeEnds[it->second].push_back(it->first);
4749 SmallSet<Instruction*, 8> OpenIntervals;
4750 unsigned MaxUsage = 0;
4753 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4754 for (unsigned int i = 0; i < Index; ++i) {
4755 Instruction *I = IdxToInstr[i];
4756 // Ignore instructions that are never used within the loop.
4757 if (!Ends.count(I)) continue;
4759 // Remove all of the instructions that end at this location.
4760 InstrList &List = TransposeEnds[i];
4761 for (unsigned int j=0, e = List.size(); j < e; ++j)
4762 OpenIntervals.erase(List[j]);
4764 // Count the number of live interals.
4765 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4767 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4768 OpenIntervals.size() << '\n');
4770 // Add the current instruction to the list of open intervals.
4771 OpenIntervals.insert(I);
4774 unsigned Invariant = LoopInvariants.size();
4775 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4776 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4777 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4779 R.LoopInvariantRegs = Invariant;
4780 R.MaxLocalUsers = MaxUsage;
4784 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4788 for (Loop::block_iterator bb = TheLoop->block_begin(),
4789 be = TheLoop->block_end(); bb != be; ++bb) {
4790 unsigned BlockCost = 0;
4791 BasicBlock *BB = *bb;
4793 // For each instruction in the old loop.
4794 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4795 // Skip dbg intrinsics.
4796 if (isa<DbgInfoIntrinsic>(it))
4799 unsigned C = getInstructionCost(it, VF);
4801 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4802 VF << " For instruction: " << *it << '\n');
4805 // We assume that if-converted blocks have a 50% chance of being executed.
4806 // When the code is scalar then some of the blocks are avoided due to CF.
4807 // When the code is vectorized we execute all code paths.
4808 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4817 /// \brief Check whether the address computation for a non-consecutive memory
4818 /// access looks like an unlikely candidate for being merged into the indexing
4821 /// We look for a GEP which has one index that is an induction variable and all
4822 /// other indices are loop invariant. If the stride of this access is also
4823 /// within a small bound we decide that this address computation can likely be
4824 /// merged into the addressing mode.
4825 /// In all other cases, we identify the address computation as complex.
4826 static bool isLikelyComplexAddressComputation(Value *Ptr,
4827 LoopVectorizationLegality *Legal,
4828 ScalarEvolution *SE,
4829 const Loop *TheLoop) {
4830 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4834 // We are looking for a gep with all loop invariant indices except for one
4835 // which should be an induction variable.
4836 unsigned NumOperands = Gep->getNumOperands();
4837 for (unsigned i = 1; i < NumOperands; ++i) {
4838 Value *Opd = Gep->getOperand(i);
4839 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4840 !Legal->isInductionVariable(Opd))
4844 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4845 // can likely be merged into the address computation.
4846 unsigned MaxMergeDistance = 64;
4848 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4852 // Check the step is constant.
4853 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4854 // Calculate the pointer stride and check if it is consecutive.
4855 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4859 const APInt &APStepVal = C->getValue()->getValue();
4861 // Huge step value - give up.
4862 if (APStepVal.getBitWidth() > 64)
4865 int64_t StepVal = APStepVal.getSExtValue();
4867 return StepVal > MaxMergeDistance;
4871 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4872 // If we know that this instruction will remain uniform, check the cost of
4873 // the scalar version.
4874 if (Legal->isUniformAfterVectorization(I))
4877 Type *RetTy = I->getType();
4878 Type *VectorTy = ToVectorTy(RetTy, VF);
4880 // TODO: We need to estimate the cost of intrinsic calls.
4881 switch (I->getOpcode()) {
4882 case Instruction::GetElementPtr:
4883 // We mark this instruction as zero-cost because the cost of GEPs in
4884 // vectorized code depends on whether the corresponding memory instruction
4885 // is scalarized or not. Therefore, we handle GEPs with the memory
4886 // instruction cost.
4888 case Instruction::Br: {
4889 return TTI.getCFInstrCost(I->getOpcode());
4891 case Instruction::PHI:
4892 //TODO: IF-converted IFs become selects.
4894 case Instruction::Add:
4895 case Instruction::FAdd:
4896 case Instruction::Sub:
4897 case Instruction::FSub:
4898 case Instruction::Mul:
4899 case Instruction::FMul:
4900 case Instruction::UDiv:
4901 case Instruction::SDiv:
4902 case Instruction::FDiv:
4903 case Instruction::URem:
4904 case Instruction::SRem:
4905 case Instruction::FRem:
4906 case Instruction::Shl:
4907 case Instruction::LShr:
4908 case Instruction::AShr:
4909 case Instruction::And:
4910 case Instruction::Or:
4911 case Instruction::Xor: {
4912 // Certain instructions can be cheaper to vectorize if they have a constant
4913 // second vector operand. One example of this are shifts on x86.
4914 TargetTransformInfo::OperandValueKind Op1VK =
4915 TargetTransformInfo::OK_AnyValue;
4916 TargetTransformInfo::OperandValueKind Op2VK =
4917 TargetTransformInfo::OK_AnyValue;
4919 if (isa<ConstantInt>(I->getOperand(1)))
4920 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4922 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4924 case Instruction::Select: {
4925 SelectInst *SI = cast<SelectInst>(I);
4926 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4927 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4928 Type *CondTy = SI->getCondition()->getType();
4930 CondTy = VectorType::get(CondTy, VF);
4932 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4934 case Instruction::ICmp:
4935 case Instruction::FCmp: {
4936 Type *ValTy = I->getOperand(0)->getType();
4937 VectorTy = ToVectorTy(ValTy, VF);
4938 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4940 case Instruction::Store:
4941 case Instruction::Load: {
4942 StoreInst *SI = dyn_cast<StoreInst>(I);
4943 LoadInst *LI = dyn_cast<LoadInst>(I);
4944 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4946 VectorTy = ToVectorTy(ValTy, VF);
4948 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4949 unsigned AS = SI ? SI->getPointerAddressSpace() :
4950 LI->getPointerAddressSpace();
4951 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4952 // We add the cost of address computation here instead of with the gep
4953 // instruction because only here we know whether the operation is
4956 return TTI.getAddressComputationCost(VectorTy) +
4957 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4959 // Scalarized loads/stores.
4960 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4961 bool Reverse = ConsecutiveStride < 0;
4962 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4963 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4964 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4965 bool IsComplexComputation =
4966 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4968 // The cost of extracting from the value vector and pointer vector.
4969 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4970 for (unsigned i = 0; i < VF; ++i) {
4971 // The cost of extracting the pointer operand.
4972 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4973 // In case of STORE, the cost of ExtractElement from the vector.
4974 // In case of LOAD, the cost of InsertElement into the returned
4976 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4977 Instruction::InsertElement,
4981 // The cost of the scalar loads/stores.
4982 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4983 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4988 // Wide load/stores.
4989 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4990 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4993 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4997 case Instruction::ZExt:
4998 case Instruction::SExt:
4999 case Instruction::FPToUI:
5000 case Instruction::FPToSI:
5001 case Instruction::FPExt:
5002 case Instruction::PtrToInt:
5003 case Instruction::IntToPtr:
5004 case Instruction::SIToFP:
5005 case Instruction::UIToFP:
5006 case Instruction::Trunc:
5007 case Instruction::FPTrunc:
5008 case Instruction::BitCast: {
5009 // We optimize the truncation of induction variable.
5010 // The cost of these is the same as the scalar operation.
5011 if (I->getOpcode() == Instruction::Trunc &&
5012 Legal->isInductionVariable(I->getOperand(0)))
5013 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5014 I->getOperand(0)->getType());
5016 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5017 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5019 case Instruction::Call: {
5020 CallInst *CI = cast<CallInst>(I);
5021 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5022 assert(ID && "Not an intrinsic call!");
5023 Type *RetTy = ToVectorTy(CI->getType(), VF);
5024 SmallVector<Type*, 4> Tys;
5025 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5026 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5027 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5030 // We are scalarizing the instruction. Return the cost of the scalar
5031 // instruction, plus the cost of insert and extract into vector
5032 // elements, times the vector width.
5035 if (!RetTy->isVoidTy() && VF != 1) {
5036 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5038 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5041 // The cost of inserting the results plus extracting each one of the
5043 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5046 // The cost of executing VF copies of the scalar instruction. This opcode
5047 // is unknown. Assume that it is the same as 'mul'.
5048 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5054 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5055 if (Scalar->isVoidTy() || VF == 1)
5057 return VectorType::get(Scalar, VF);
5060 char LoopVectorize::ID = 0;
5061 static const char lv_name[] = "Loop Vectorization";
5062 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5063 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5064 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
5065 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5066 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5067 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5068 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5069 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5072 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5073 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5077 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5078 // Check for a store.
5079 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5080 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5082 // Check for a load.
5083 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5084 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5090 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
5091 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5092 // Holds vector parameters or scalars, in case of uniform vals.
5093 SmallVector<VectorParts, 4> Params;
5095 setDebugLocFromInst(Builder, Instr);
5097 // Find all of the vectorized parameters.
5098 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5099 Value *SrcOp = Instr->getOperand(op);
5101 // If we are accessing the old induction variable, use the new one.
5102 if (SrcOp == OldInduction) {
5103 Params.push_back(getVectorValue(SrcOp));
5107 // Try using previously calculated values.
5108 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5110 // If the src is an instruction that appeared earlier in the basic block
5111 // then it should already be vectorized.
5112 if (SrcInst && OrigLoop->contains(SrcInst)) {
5113 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5114 // The parameter is a vector value from earlier.
5115 Params.push_back(WidenMap.get(SrcInst));
5117 // The parameter is a scalar from outside the loop. Maybe even a constant.
5118 VectorParts Scalars;
5119 Scalars.append(UF, SrcOp);
5120 Params.push_back(Scalars);
5124 assert(Params.size() == Instr->getNumOperands() &&
5125 "Invalid number of operands");
5127 // Does this instruction return a value ?
5128 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5130 Value *UndefVec = IsVoidRetTy ? 0 :
5131 UndefValue::get(Instr->getType());
5132 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5133 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5135 // For each vector unroll 'part':
5136 for (unsigned Part = 0; Part < UF; ++Part) {
5137 // For each scalar that we create:
5139 Instruction *Cloned = Instr->clone();
5141 Cloned->setName(Instr->getName() + ".cloned");
5142 // Replace the operands of the cloned instructions with extracted scalars.
5143 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5144 Value *Op = Params[op][Part];
5145 Cloned->setOperand(op, Op);
5148 // Place the cloned scalar in the new loop.
5149 Builder.Insert(Cloned);
5151 // If the original scalar returns a value we need to place it in a vector
5152 // so that future users will be able to use it.
5154 VecResults[Part] = Cloned;
5159 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
5160 LoopVectorizationLegality*) {
5161 return scalarizeInstruction(Instr);
5164 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5168 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5172 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5174 // When unrolling and the VF is 1, we only need to add a simple scalar.
5175 Type *ITy = Val->getType();
5176 assert(!ITy->isVectorTy() && "Val must be a scalar");
5177 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5178 return Builder.CreateAdd(Val, C, "induction");