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.
767 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
768 : Width(VectorizationFactor)
769 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
770 , LoopID(L->getLoopID()) {
772 // The command line options override any loop metadata except for when
773 // width == 1 which is used to indicate the loop is already vectorized.
774 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
775 Width = VectorizationFactor;
776 if (VectorizationUnroll.getNumOccurrences() > 0)
777 Unroll = VectorizationUnroll;
779 DEBUG(if (DisableUnrolling && Unroll == 1)
780 dbgs() << "LV: Unrolling disabled by the pass manager\n");
783 /// Return the loop vectorizer metadata prefix.
784 static StringRef Prefix() { return "llvm.vectorizer."; }
786 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
787 SmallVector<Value*, 2> Vals;
788 Vals.push_back(MDString::get(Context, Name));
789 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
790 return MDNode::get(Context, Vals);
793 /// Mark the loop L as already vectorized by setting the width to 1.
794 void setAlreadyVectorized(Loop *L) {
795 LLVMContext &Context = L->getHeader()->getContext();
799 // Create a new loop id with one more operand for the already_vectorized
800 // hint. If the loop already has a loop id then copy the existing operands.
801 SmallVector<Value*, 4> Vals(1);
803 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
804 Vals.push_back(LoopID->getOperand(i));
806 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
807 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
809 MDNode *NewLoopID = MDNode::get(Context, Vals);
810 // Set operand 0 to refer to the loop id itself.
811 NewLoopID->replaceOperandWith(0, NewLoopID);
813 L->setLoopID(NewLoopID);
815 LoopID->replaceAllUsesWith(NewLoopID);
823 /// Find hints specified in the loop metadata.
824 void getHints(const Loop *L) {
828 // First operand should refer to the loop id itself.
829 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
830 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
832 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
833 const MDString *S = 0;
834 SmallVector<Value*, 4> Args;
836 // The expected hint is either a MDString or a MDNode with the first
837 // operand a MDString.
838 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
839 if (!MD || MD->getNumOperands() == 0)
841 S = dyn_cast<MDString>(MD->getOperand(0));
842 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
843 Args.push_back(MD->getOperand(i));
845 S = dyn_cast<MDString>(LoopID->getOperand(i));
846 assert(Args.size() == 0 && "too many arguments for MDString");
852 // Check if the hint starts with the vectorizer prefix.
853 StringRef Hint = S->getString();
854 if (!Hint.startswith(Prefix()))
856 // Remove the prefix.
857 Hint = Hint.substr(Prefix().size(), StringRef::npos);
859 if (Args.size() == 1)
860 getHint(Hint, Args[0]);
864 // Check string hint with one operand.
865 void getHint(StringRef Hint, Value *Arg) {
866 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
868 unsigned Val = C->getZExtValue();
870 if (Hint == "width") {
871 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
874 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
875 } else if (Hint == "unroll") {
876 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
879 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
881 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
886 /// The LoopVectorize Pass.
887 struct LoopVectorize : public LoopPass {
888 /// Pass identification, replacement for typeid
891 explicit LoopVectorize(bool NoUnrolling = false)
892 : LoopPass(ID), DisableUnrolling(NoUnrolling) {
893 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
899 TargetTransformInfo *TTI;
901 TargetLibraryInfo *TLI;
902 bool DisableUnrolling;
904 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
905 // We only vectorize innermost loops.
909 SE = &getAnalysis<ScalarEvolution>();
910 DL = getAnalysisIfAvailable<DataLayout>();
911 LI = &getAnalysis<LoopInfo>();
912 TTI = &getAnalysis<TargetTransformInfo>();
913 DT = &getAnalysis<DominatorTree>();
914 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
916 // If the target claims to have no vector registers don't attempt
918 if (!TTI->getNumberOfRegisters(true))
922 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout\n");
926 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
927 L->getHeader()->getParent()->getName() << "\"\n");
929 LoopVectorizeHints Hints(L, DisableUnrolling);
931 if (Hints.Width == 1 && Hints.Unroll == 1) {
932 DEBUG(dbgs() << "LV: Not vectorizing.\n");
936 // Check if it is legal to vectorize the loop.
937 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
938 if (!LVL.canVectorize()) {
939 DEBUG(dbgs() << "LV: Not vectorizing.\n");
943 // Use the cost model.
944 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
946 // Check the function attributes to find out if this function should be
947 // optimized for size.
948 Function *F = L->getHeader()->getParent();
949 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
950 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
951 unsigned FnIndex = AttributeSet::FunctionIndex;
952 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
953 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
956 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
957 "attribute is used.\n");
961 // Select the optimal vectorization factor.
962 LoopVectorizationCostModel::VectorizationFactor VF;
963 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
964 // Select the unroll factor.
965 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
968 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
969 F->getParent()->getModuleIdentifier() << '\n');
970 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
973 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
976 // We decided not to vectorize, but we may want to unroll.
977 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
978 Unroller.vectorize(&LVL);
980 // If we decided that it is *legal* to vectorize the loop then do it.
981 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
985 // Mark the loop as already vectorized to avoid vectorizing again.
986 Hints.setAlreadyVectorized(L);
988 DEBUG(verifyFunction(*L->getHeader()->getParent()));
992 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
993 LoopPass::getAnalysisUsage(AU);
994 AU.addRequiredID(LoopSimplifyID);
995 AU.addRequiredID(LCSSAID);
996 AU.addRequired<DominatorTree>();
997 AU.addRequired<LoopInfo>();
998 AU.addRequired<ScalarEvolution>();
999 AU.addRequired<TargetTransformInfo>();
1000 AU.addPreserved<LoopInfo>();
1001 AU.addPreserved<DominatorTree>();
1006 } // end anonymous namespace
1008 //===----------------------------------------------------------------------===//
1009 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1010 // LoopVectorizationCostModel.
1011 //===----------------------------------------------------------------------===//
1014 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1015 Loop *Lp, Value *Ptr,
1017 unsigned DepSetId) {
1018 const SCEV *Sc = SE->getSCEV(Ptr);
1019 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1020 assert(AR && "Invalid addrec expression");
1021 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1022 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1023 Pointers.push_back(Ptr);
1024 Starts.push_back(AR->getStart());
1025 Ends.push_back(ScEnd);
1026 IsWritePtr.push_back(WritePtr);
1027 DependencySetId.push_back(DepSetId);
1030 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1031 // We need to place the broadcast of invariant variables outside the loop.
1032 Instruction *Instr = dyn_cast<Instruction>(V);
1033 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1034 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1036 // Place the code for broadcasting invariant variables in the new preheader.
1037 IRBuilder<>::InsertPointGuard Guard(Builder);
1039 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1041 // Broadcast the scalar into all locations in the vector.
1042 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1047 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1049 assert(Val->getType()->isVectorTy() && "Must be a vector");
1050 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1051 "Elem must be an integer");
1052 // Create the types.
1053 Type *ITy = Val->getType()->getScalarType();
1054 VectorType *Ty = cast<VectorType>(Val->getType());
1055 int VLen = Ty->getNumElements();
1056 SmallVector<Constant*, 8> Indices;
1058 // Create a vector of consecutive numbers from zero to VF.
1059 for (int i = 0; i < VLen; ++i) {
1060 int64_t Idx = Negate ? (-i) : i;
1061 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1064 // Add the consecutive indices to the vector value.
1065 Constant *Cv = ConstantVector::get(Indices);
1066 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1067 return Builder.CreateAdd(Val, Cv, "induction");
1070 /// \brief Find the operand of the GEP that should be checked for consecutive
1071 /// stores. This ignores trailing indices that have no effect on the final
1073 static unsigned getGEPInductionOperand(DataLayout *DL,
1074 const GetElementPtrInst *Gep) {
1075 unsigned LastOperand = Gep->getNumOperands() - 1;
1076 unsigned GEPAllocSize = DL->getTypeAllocSize(
1077 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1079 // Walk backwards and try to peel off zeros.
1080 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1081 // Find the type we're currently indexing into.
1082 gep_type_iterator GEPTI = gep_type_begin(Gep);
1083 std::advance(GEPTI, LastOperand - 1);
1085 // If it's a type with the same allocation size as the result of the GEP we
1086 // can peel off the zero index.
1087 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1095 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1096 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1097 // Make sure that the pointer does not point to structs.
1098 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1101 // If this value is a pointer induction variable we know it is consecutive.
1102 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1103 if (Phi && Inductions.count(Phi)) {
1104 InductionInfo II = Inductions[Phi];
1105 if (IK_PtrInduction == II.IK)
1107 else if (IK_ReversePtrInduction == II.IK)
1111 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1115 unsigned NumOperands = Gep->getNumOperands();
1116 Value *GpPtr = Gep->getPointerOperand();
1117 // If this GEP value is a consecutive pointer induction variable and all of
1118 // the indices are constant then we know it is consecutive. We can
1119 Phi = dyn_cast<PHINode>(GpPtr);
1120 if (Phi && Inductions.count(Phi)) {
1122 // Make sure that the pointer does not point to structs.
1123 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1124 if (GepPtrType->getElementType()->isAggregateType())
1127 // Make sure that all of the index operands are loop invariant.
1128 for (unsigned i = 1; i < NumOperands; ++i)
1129 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1132 InductionInfo II = Inductions[Phi];
1133 if (IK_PtrInduction == II.IK)
1135 else if (IK_ReversePtrInduction == II.IK)
1139 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1141 // Check that all of the gep indices are uniform except for our induction
1143 for (unsigned i = 0; i != NumOperands; ++i)
1144 if (i != InductionOperand &&
1145 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1148 // We can emit wide load/stores only if the last non-zero index is the
1149 // induction variable.
1150 const SCEV *Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1151 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1152 const SCEV *Step = AR->getStepRecurrence(*SE);
1154 // The memory is consecutive because the last index is consecutive
1155 // and all other indices are loop invariant.
1158 if (Step->isAllOnesValue())
1165 bool LoopVectorizationLegality::isUniform(Value *V) {
1166 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1169 InnerLoopVectorizer::VectorParts&
1170 InnerLoopVectorizer::getVectorValue(Value *V) {
1171 assert(V != Induction && "The new induction variable should not be used.");
1172 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1174 // If we have this scalar in the map, return it.
1175 if (WidenMap.has(V))
1176 return WidenMap.get(V);
1178 // If this scalar is unknown, assume that it is a constant or that it is
1179 // loop invariant. Broadcast V and save the value for future uses.
1180 Value *B = getBroadcastInstrs(V);
1181 return WidenMap.splat(V, B);
1184 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1185 assert(Vec->getType()->isVectorTy() && "Invalid type");
1186 SmallVector<Constant*, 8> ShuffleMask;
1187 for (unsigned i = 0; i < VF; ++i)
1188 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1190 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1191 ConstantVector::get(ShuffleMask),
1196 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1197 LoopVectorizationLegality *Legal) {
1198 // Attempt to issue a wide load.
1199 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1200 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1202 assert((LI || SI) && "Invalid Load/Store instruction");
1204 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1205 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1206 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1207 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1208 // An alignment of 0 means target abi alignment. We need to use the scalar's
1209 // target abi alignment in such a case.
1211 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1212 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1213 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1214 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1216 if (ScalarAllocatedSize != VectorElementSize)
1217 return scalarizeInstruction(Instr);
1219 // If the pointer is loop invariant or if it is non-consecutive,
1220 // scalarize the load.
1221 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1222 bool Reverse = ConsecutiveStride < 0;
1223 bool UniformLoad = LI && Legal->isUniform(Ptr);
1224 if (!ConsecutiveStride || UniformLoad)
1225 return scalarizeInstruction(Instr);
1227 Constant *Zero = Builder.getInt32(0);
1228 VectorParts &Entry = WidenMap.get(Instr);
1230 // Handle consecutive loads/stores.
1231 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1232 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1233 setDebugLocFromInst(Builder, Gep);
1234 Value *PtrOperand = Gep->getPointerOperand();
1235 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1236 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1238 // Create the new GEP with the new induction variable.
1239 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1240 Gep2->setOperand(0, FirstBasePtr);
1241 Gep2->setName("gep.indvar.base");
1242 Ptr = Builder.Insert(Gep2);
1244 setDebugLocFromInst(Builder, Gep);
1245 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1246 OrigLoop) && "Base ptr must be invariant");
1248 // The last index does not have to be the induction. It can be
1249 // consecutive and be a function of the index. For example A[I+1];
1250 unsigned NumOperands = Gep->getNumOperands();
1251 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1252 // Create the new GEP with the new induction variable.
1253 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1255 for (unsigned i = 0; i < NumOperands; ++i) {
1256 Value *GepOperand = Gep->getOperand(i);
1257 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1259 // Update last index or loop invariant instruction anchored in loop.
1260 if (i == InductionOperand ||
1261 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1262 assert((i == InductionOperand ||
1263 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1264 "Must be last index or loop invariant");
1266 VectorParts &GEPParts = getVectorValue(GepOperand);
1267 Value *Index = GEPParts[0];
1268 Index = Builder.CreateExtractElement(Index, Zero);
1269 Gep2->setOperand(i, Index);
1270 Gep2->setName("gep.indvar.idx");
1273 Ptr = Builder.Insert(Gep2);
1275 // Use the induction element ptr.
1276 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1277 setDebugLocFromInst(Builder, Ptr);
1278 VectorParts &PtrVal = getVectorValue(Ptr);
1279 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1284 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1285 "We do not allow storing to uniform addresses");
1286 setDebugLocFromInst(Builder, SI);
1287 // We don't want to update the value in the map as it might be used in
1288 // another expression. So don't use a reference type for "StoredVal".
1289 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1291 for (unsigned Part = 0; Part < UF; ++Part) {
1292 // Calculate the pointer for the specific unroll-part.
1293 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1296 // If we store to reverse consecutive memory locations then we need
1297 // to reverse the order of elements in the stored value.
1298 StoredVal[Part] = reverseVector(StoredVal[Part]);
1299 // If the address is consecutive but reversed, then the
1300 // wide store needs to start at the last vector element.
1301 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1302 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1305 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1306 DataTy->getPointerTo(AddressSpace));
1307 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1313 assert(LI && "Must have a load instruction");
1314 setDebugLocFromInst(Builder, LI);
1315 for (unsigned Part = 0; Part < UF; ++Part) {
1316 // Calculate the pointer for the specific unroll-part.
1317 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1320 // If the address is consecutive but reversed, then the
1321 // wide store needs to start at the last vector element.
1322 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1323 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1326 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1327 DataTy->getPointerTo(AddressSpace));
1328 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1329 cast<LoadInst>(LI)->setAlignment(Alignment);
1330 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1334 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1335 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1336 // Holds vector parameters or scalars, in case of uniform vals.
1337 SmallVector<VectorParts, 4> Params;
1339 setDebugLocFromInst(Builder, Instr);
1341 // Find all of the vectorized parameters.
1342 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1343 Value *SrcOp = Instr->getOperand(op);
1345 // If we are accessing the old induction variable, use the new one.
1346 if (SrcOp == OldInduction) {
1347 Params.push_back(getVectorValue(SrcOp));
1351 // Try using previously calculated values.
1352 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1354 // If the src is an instruction that appeared earlier in the basic block
1355 // then it should already be vectorized.
1356 if (SrcInst && OrigLoop->contains(SrcInst)) {
1357 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1358 // The parameter is a vector value from earlier.
1359 Params.push_back(WidenMap.get(SrcInst));
1361 // The parameter is a scalar from outside the loop. Maybe even a constant.
1362 VectorParts Scalars;
1363 Scalars.append(UF, SrcOp);
1364 Params.push_back(Scalars);
1368 assert(Params.size() == Instr->getNumOperands() &&
1369 "Invalid number of operands");
1371 // Does this instruction return a value ?
1372 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1374 Value *UndefVec = IsVoidRetTy ? 0 :
1375 UndefValue::get(VectorType::get(Instr->getType(), VF));
1376 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1377 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1379 // For each vector unroll 'part':
1380 for (unsigned Part = 0; Part < UF; ++Part) {
1381 // For each scalar that we create:
1382 for (unsigned Width = 0; Width < VF; ++Width) {
1383 Instruction *Cloned = Instr->clone();
1385 Cloned->setName(Instr->getName() + ".cloned");
1386 // Replace the operands of the cloned instructions with extracted scalars.
1387 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1388 Value *Op = Params[op][Part];
1389 // Param is a vector. Need to extract the right lane.
1390 if (Op->getType()->isVectorTy())
1391 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1392 Cloned->setOperand(op, Op);
1395 // Place the cloned scalar in the new loop.
1396 Builder.Insert(Cloned);
1398 // If the original scalar returns a value we need to place it in a vector
1399 // so that future users will be able to use it.
1401 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1402 Builder.getInt32(Width));
1408 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1410 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1411 Legal->getRuntimePointerCheck();
1413 if (!PtrRtCheck->Need)
1416 unsigned NumPointers = PtrRtCheck->Pointers.size();
1417 SmallVector<TrackingVH<Value> , 2> Starts;
1418 SmallVector<TrackingVH<Value> , 2> Ends;
1420 LLVMContext &Ctx = Loc->getContext();
1421 SCEVExpander Exp(*SE, "induction");
1423 for (unsigned i = 0; i < NumPointers; ++i) {
1424 Value *Ptr = PtrRtCheck->Pointers[i];
1425 const SCEV *Sc = SE->getSCEV(Ptr);
1427 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1428 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1430 Starts.push_back(Ptr);
1431 Ends.push_back(Ptr);
1433 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1434 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1436 // Use this type for pointer arithmetic.
1437 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1439 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1440 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1441 Starts.push_back(Start);
1442 Ends.push_back(End);
1446 IRBuilder<> ChkBuilder(Loc);
1447 // Our instructions might fold to a constant.
1448 Value *MemoryRuntimeCheck = 0;
1449 for (unsigned i = 0; i < NumPointers; ++i) {
1450 for (unsigned j = i+1; j < NumPointers; ++j) {
1451 // No need to check if two readonly pointers intersect.
1452 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1455 // Only need to check pointers between two different dependency sets.
1456 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1459 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1460 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1462 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1463 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1464 "Trying to bounds check pointers with different address spaces");
1466 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1467 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1469 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1470 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1471 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1472 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1474 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1475 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1476 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1477 if (MemoryRuntimeCheck)
1478 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1480 MemoryRuntimeCheck = IsConflict;
1484 // We have to do this trickery because the IRBuilder might fold the check to a
1485 // constant expression in which case there is no Instruction anchored in a
1487 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1488 ConstantInt::getTrue(Ctx));
1489 ChkBuilder.Insert(Check, "memcheck.conflict");
1494 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1496 In this function we generate a new loop. The new loop will contain
1497 the vectorized instructions while the old loop will continue to run the
1500 [ ] <-- vector loop bypass (may consist of multiple blocks).
1503 | [ ] <-- vector pre header.
1507 | [ ]_| <-- vector loop.
1510 >[ ] <--- middle-block.
1513 | [ ] <--- new preheader.
1517 | [ ]_| <-- old scalar loop to handle remainder.
1520 >[ ] <-- exit block.
1524 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1525 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1526 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1527 assert(ExitBlock && "Must have an exit block");
1529 // Some loops have a single integer induction variable, while other loops
1530 // don't. One example is c++ iterators that often have multiple pointer
1531 // induction variables. In the code below we also support a case where we
1532 // don't have a single induction variable.
1533 OldInduction = Legal->getInduction();
1534 Type *IdxTy = Legal->getWidestInductionType();
1536 // Find the loop boundaries.
1537 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1538 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1540 // The exit count might have the type of i64 while the phi is i32. This can
1541 // happen if we have an induction variable that is sign extended before the
1542 // compare. The only way that we get a backedge taken count is that the
1543 // induction variable was signed and as such will not overflow. In such a case
1544 // truncation is legal.
1545 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1546 IdxTy->getPrimitiveSizeInBits())
1547 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1549 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1550 // Get the total trip count from the count by adding 1.
1551 ExitCount = SE->getAddExpr(ExitCount,
1552 SE->getConstant(ExitCount->getType(), 1));
1554 // Expand the trip count and place the new instructions in the preheader.
1555 // Notice that the pre-header does not change, only the loop body.
1556 SCEVExpander Exp(*SE, "induction");
1558 // Count holds the overall loop count (N).
1559 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1560 BypassBlock->getTerminator());
1562 // The loop index does not have to start at Zero. Find the original start
1563 // value from the induction PHI node. If we don't have an induction variable
1564 // then we know that it starts at zero.
1565 Builder.SetInsertPoint(BypassBlock->getTerminator());
1566 Value *StartIdx = ExtendedIdx = OldInduction ?
1567 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1569 ConstantInt::get(IdxTy, 0);
1571 assert(BypassBlock && "Invalid loop structure");
1572 LoopBypassBlocks.push_back(BypassBlock);
1574 // Split the single block loop into the two loop structure described above.
1575 BasicBlock *VectorPH =
1576 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1577 BasicBlock *VecBody =
1578 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1579 BasicBlock *MiddleBlock =
1580 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1581 BasicBlock *ScalarPH =
1582 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1584 // Create and register the new vector loop.
1585 Loop* Lp = new Loop();
1586 Loop *ParentLoop = OrigLoop->getParentLoop();
1588 // Insert the new loop into the loop nest and register the new basic blocks
1589 // before calling any utilities such as SCEV that require valid LoopInfo.
1591 ParentLoop->addChildLoop(Lp);
1592 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1593 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1594 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1596 LI->addTopLevelLoop(Lp);
1598 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1600 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1602 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1604 // Generate the induction variable.
1605 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1606 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1607 // The loop step is equal to the vectorization factor (num of SIMD elements)
1608 // times the unroll factor (num of SIMD instructions).
1609 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1611 // This is the IR builder that we use to add all of the logic for bypassing
1612 // the new vector loop.
1613 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1614 setDebugLocFromInst(BypassBuilder,
1615 getDebugLocFromInstOrOperands(OldInduction));
1617 // We may need to extend the index in case there is a type mismatch.
1618 // We know that the count starts at zero and does not overflow.
1619 if (Count->getType() != IdxTy) {
1620 // The exit count can be of pointer type. Convert it to the correct
1622 if (ExitCount->getType()->isPointerTy())
1623 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1625 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1628 // Add the start index to the loop count to get the new end index.
1629 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1631 // Now we need to generate the expression for N - (N % VF), which is
1632 // the part that the vectorized body will execute.
1633 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1634 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1635 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1636 "end.idx.rnd.down");
1638 // Now, compare the new count to zero. If it is zero skip the vector loop and
1639 // jump to the scalar loop.
1640 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1643 BasicBlock *LastBypassBlock = BypassBlock;
1645 // Generate the code that checks in runtime if arrays overlap. We put the
1646 // checks into a separate block to make the more common case of few elements
1648 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1649 BypassBlock->getTerminator());
1650 if (MemRuntimeCheck) {
1651 // Create a new block containing the memory check.
1652 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1655 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1656 LoopBypassBlocks.push_back(CheckBlock);
1658 // Replace the branch into the memory check block with a conditional branch
1659 // for the "few elements case".
1660 Instruction *OldTerm = BypassBlock->getTerminator();
1661 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1662 OldTerm->eraseFromParent();
1664 Cmp = MemRuntimeCheck;
1665 LastBypassBlock = CheckBlock;
1668 LastBypassBlock->getTerminator()->eraseFromParent();
1669 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1672 // We are going to resume the execution of the scalar loop.
1673 // Go over all of the induction variables that we found and fix the
1674 // PHIs that are left in the scalar version of the loop.
1675 // The starting values of PHI nodes depend on the counter of the last
1676 // iteration in the vectorized loop.
1677 // If we come from a bypass edge then we need to start from the original
1680 // This variable saves the new starting index for the scalar loop.
1681 PHINode *ResumeIndex = 0;
1682 LoopVectorizationLegality::InductionList::iterator I, E;
1683 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1684 // Set builder to point to last bypass block.
1685 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1686 for (I = List->begin(), E = List->end(); I != E; ++I) {
1687 PHINode *OrigPhi = I->first;
1688 LoopVectorizationLegality::InductionInfo II = I->second;
1690 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1691 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1692 MiddleBlock->getTerminator());
1693 // We might have extended the type of the induction variable but we need a
1694 // truncated version for the scalar loop.
1695 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1696 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1697 MiddleBlock->getTerminator()) : 0;
1699 Value *EndValue = 0;
1701 case LoopVectorizationLegality::IK_NoInduction:
1702 llvm_unreachable("Unknown induction");
1703 case LoopVectorizationLegality::IK_IntInduction: {
1704 // Handle the integer induction counter.
1705 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1707 // We have the canonical induction variable.
1708 if (OrigPhi == OldInduction) {
1709 // Create a truncated version of the resume value for the scalar loop,
1710 // we might have promoted the type to a larger width.
1712 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1713 // The new PHI merges the original incoming value, in case of a bypass,
1714 // or the value at the end of the vectorized loop.
1715 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1716 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1717 TruncResumeVal->addIncoming(EndValue, VecBody);
1719 // We know what the end value is.
1720 EndValue = IdxEndRoundDown;
1721 // We also know which PHI node holds it.
1722 ResumeIndex = ResumeVal;
1726 // Not the canonical induction variable - add the vector loop count to the
1728 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1729 II.StartValue->getType(),
1731 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1734 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1735 // Convert the CountRoundDown variable to the PHI size.
1736 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1737 II.StartValue->getType(),
1739 // Handle reverse integer induction counter.
1740 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1743 case LoopVectorizationLegality::IK_PtrInduction: {
1744 // For pointer induction variables, calculate the offset using
1746 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1750 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1751 // The value at the end of the loop for the reverse pointer is calculated
1752 // by creating a GEP with a negative index starting from the start value.
1753 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1754 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1756 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1762 // The new PHI merges the original incoming value, in case of a bypass,
1763 // or the value at the end of the vectorized loop.
1764 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1765 if (OrigPhi == OldInduction)
1766 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1768 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1770 ResumeVal->addIncoming(EndValue, VecBody);
1772 // Fix the scalar body counter (PHI node).
1773 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1774 // The old inductions phi node in the scalar body needs the truncated value.
1775 if (OrigPhi == OldInduction)
1776 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1778 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1781 // If we are generating a new induction variable then we also need to
1782 // generate the code that calculates the exit value. This value is not
1783 // simply the end of the counter because we may skip the vectorized body
1784 // in case of a runtime check.
1786 assert(!ResumeIndex && "Unexpected resume value found");
1787 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1788 MiddleBlock->getTerminator());
1789 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1790 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1791 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1794 // Make sure that we found the index where scalar loop needs to continue.
1795 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1796 "Invalid resume Index");
1798 // Add a check in the middle block to see if we have completed
1799 // all of the iterations in the first vector loop.
1800 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1801 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1802 ResumeIndex, "cmp.n",
1803 MiddleBlock->getTerminator());
1805 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1806 // Remove the old terminator.
1807 MiddleBlock->getTerminator()->eraseFromParent();
1809 // Create i+1 and fill the PHINode.
1810 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1811 Induction->addIncoming(StartIdx, VectorPH);
1812 Induction->addIncoming(NextIdx, VecBody);
1813 // Create the compare.
1814 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1815 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1817 // Now we have two terminators. Remove the old one from the block.
1818 VecBody->getTerminator()->eraseFromParent();
1820 // Get ready to start creating new instructions into the vectorized body.
1821 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1824 LoopVectorPreHeader = VectorPH;
1825 LoopScalarPreHeader = ScalarPH;
1826 LoopMiddleBlock = MiddleBlock;
1827 LoopExitBlock = ExitBlock;
1828 LoopVectorBody = VecBody;
1829 LoopScalarBody = OldBasicBlock;
1831 LoopVectorizeHints Hints(Lp, true);
1832 Hints.setAlreadyVectorized(Lp);
1835 /// This function returns the identity element (or neutral element) for
1836 /// the operation K.
1838 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1843 // Adding, Xoring, Oring zero to a number does not change it.
1844 return ConstantInt::get(Tp, 0);
1845 case RK_IntegerMult:
1846 // Multiplying a number by 1 does not change it.
1847 return ConstantInt::get(Tp, 1);
1849 // AND-ing a number with an all-1 value does not change it.
1850 return ConstantInt::get(Tp, -1, true);
1852 // Multiplying a number by 1 does not change it.
1853 return ConstantFP::get(Tp, 1.0L);
1855 // Adding zero to a number does not change it.
1856 return ConstantFP::get(Tp, 0.0L);
1858 llvm_unreachable("Unknown reduction kind");
1862 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
1863 Intrinsic::ID ValidIntrinsicID) {
1864 if (I.getNumArgOperands() != 1 ||
1865 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1866 I.getType() != I.getArgOperand(0)->getType() ||
1867 !I.onlyReadsMemory())
1868 return Intrinsic::not_intrinsic;
1870 return ValidIntrinsicID;
1873 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
1874 Intrinsic::ID ValidIntrinsicID) {
1875 if (I.getNumArgOperands() != 2 ||
1876 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
1877 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
1878 I.getType() != I.getArgOperand(0)->getType() ||
1879 I.getType() != I.getArgOperand(1)->getType() ||
1880 !I.onlyReadsMemory())
1881 return Intrinsic::not_intrinsic;
1883 return ValidIntrinsicID;
1887 static Intrinsic::ID
1888 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1889 // If we have an intrinsic call, check if it is trivially vectorizable.
1890 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1891 switch (II->getIntrinsicID()) {
1892 case Intrinsic::sqrt:
1893 case Intrinsic::sin:
1894 case Intrinsic::cos:
1895 case Intrinsic::exp:
1896 case Intrinsic::exp2:
1897 case Intrinsic::log:
1898 case Intrinsic::log10:
1899 case Intrinsic::log2:
1900 case Intrinsic::fabs:
1901 case Intrinsic::copysign:
1902 case Intrinsic::floor:
1903 case Intrinsic::ceil:
1904 case Intrinsic::trunc:
1905 case Intrinsic::rint:
1906 case Intrinsic::nearbyint:
1907 case Intrinsic::round:
1908 case Intrinsic::pow:
1909 case Intrinsic::fma:
1910 case Intrinsic::fmuladd:
1911 case Intrinsic::lifetime_start:
1912 case Intrinsic::lifetime_end:
1913 return II->getIntrinsicID();
1915 return Intrinsic::not_intrinsic;
1920 return Intrinsic::not_intrinsic;
1923 Function *F = CI->getCalledFunction();
1924 // We're going to make assumptions on the semantics of the functions, check
1925 // that the target knows that it's available in this environment and it does
1926 // not have local linkage.
1927 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
1928 return Intrinsic::not_intrinsic;
1930 // Otherwise check if we have a call to a function that can be turned into a
1931 // vector intrinsic.
1938 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
1942 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
1946 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
1948 case LibFunc::exp2f:
1949 case LibFunc::exp2l:
1950 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
1954 return checkUnaryFloatSignature(*CI, Intrinsic::log);
1955 case LibFunc::log10:
1956 case LibFunc::log10f:
1957 case LibFunc::log10l:
1958 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
1960 case LibFunc::log2f:
1961 case LibFunc::log2l:
1962 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
1964 case LibFunc::fabsf:
1965 case LibFunc::fabsl:
1966 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
1967 case LibFunc::copysign:
1968 case LibFunc::copysignf:
1969 case LibFunc::copysignl:
1970 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
1971 case LibFunc::floor:
1972 case LibFunc::floorf:
1973 case LibFunc::floorl:
1974 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
1976 case LibFunc::ceilf:
1977 case LibFunc::ceill:
1978 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
1979 case LibFunc::trunc:
1980 case LibFunc::truncf:
1981 case LibFunc::truncl:
1982 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
1984 case LibFunc::rintf:
1985 case LibFunc::rintl:
1986 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
1987 case LibFunc::nearbyint:
1988 case LibFunc::nearbyintf:
1989 case LibFunc::nearbyintl:
1990 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
1991 case LibFunc::round:
1992 case LibFunc::roundf:
1993 case LibFunc::roundl:
1994 return checkUnaryFloatSignature(*CI, Intrinsic::round);
1998 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2001 return Intrinsic::not_intrinsic;
2004 /// This function translates the reduction kind to an LLVM binary operator.
2006 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2008 case LoopVectorizationLegality::RK_IntegerAdd:
2009 return Instruction::Add;
2010 case LoopVectorizationLegality::RK_IntegerMult:
2011 return Instruction::Mul;
2012 case LoopVectorizationLegality::RK_IntegerOr:
2013 return Instruction::Or;
2014 case LoopVectorizationLegality::RK_IntegerAnd:
2015 return Instruction::And;
2016 case LoopVectorizationLegality::RK_IntegerXor:
2017 return Instruction::Xor;
2018 case LoopVectorizationLegality::RK_FloatMult:
2019 return Instruction::FMul;
2020 case LoopVectorizationLegality::RK_FloatAdd:
2021 return Instruction::FAdd;
2022 case LoopVectorizationLegality::RK_IntegerMinMax:
2023 return Instruction::ICmp;
2024 case LoopVectorizationLegality::RK_FloatMinMax:
2025 return Instruction::FCmp;
2027 llvm_unreachable("Unknown reduction operation");
2031 Value *createMinMaxOp(IRBuilder<> &Builder,
2032 LoopVectorizationLegality::MinMaxReductionKind RK,
2035 CmpInst::Predicate P = CmpInst::ICMP_NE;
2038 llvm_unreachable("Unknown min/max reduction kind");
2039 case LoopVectorizationLegality::MRK_UIntMin:
2040 P = CmpInst::ICMP_ULT;
2042 case LoopVectorizationLegality::MRK_UIntMax:
2043 P = CmpInst::ICMP_UGT;
2045 case LoopVectorizationLegality::MRK_SIntMin:
2046 P = CmpInst::ICMP_SLT;
2048 case LoopVectorizationLegality::MRK_SIntMax:
2049 P = CmpInst::ICMP_SGT;
2051 case LoopVectorizationLegality::MRK_FloatMin:
2052 P = CmpInst::FCMP_OLT;
2054 case LoopVectorizationLegality::MRK_FloatMax:
2055 P = CmpInst::FCMP_OGT;
2060 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2061 RK == LoopVectorizationLegality::MRK_FloatMax)
2062 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2064 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2066 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2071 struct CSEDenseMapInfo {
2072 static bool canHandle(Instruction *I) {
2073 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2074 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2076 static inline Instruction *getEmptyKey() {
2077 return DenseMapInfo<Instruction *>::getEmptyKey();
2079 static inline Instruction *getTombstoneKey() {
2080 return DenseMapInfo<Instruction *>::getTombstoneKey();
2082 static unsigned getHashValue(Instruction *I) {
2083 assert(canHandle(I) && "Unknown instruction!");
2084 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2085 I->value_op_end()));
2087 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2088 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2089 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2091 return LHS->isIdenticalTo(RHS);
2096 ///\brief Perform cse of induction variable instructions.
2097 static void cse(BasicBlock *BB) {
2098 // Perform simple cse.
2099 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2100 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2101 Instruction *In = I++;
2103 if (!CSEDenseMapInfo::canHandle(In))
2106 // Check if we can replace this instruction with any of the
2107 // visited instructions.
2108 if (Instruction *V = CSEMap.lookup(In)) {
2109 In->replaceAllUsesWith(V);
2110 In->eraseFromParent();
2119 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
2120 //===------------------------------------------------===//
2122 // Notice: any optimization or new instruction that go
2123 // into the code below should be also be implemented in
2126 //===------------------------------------------------===//
2127 Constant *Zero = Builder.getInt32(0);
2129 // In order to support reduction variables we need to be able to vectorize
2130 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2131 // stages. First, we create a new vector PHI node with no incoming edges.
2132 // We use this value when we vectorize all of the instructions that use the
2133 // PHI. Next, after all of the instructions in the block are complete we
2134 // add the new incoming edges to the PHI. At this point all of the
2135 // instructions in the basic block are vectorized, so we can use them to
2136 // construct the PHI.
2137 PhiVector RdxPHIsToFix;
2139 // Scan the loop in a topological order to ensure that defs are vectorized
2141 LoopBlocksDFS DFS(OrigLoop);
2144 // Vectorize all of the blocks in the original loop.
2145 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2146 be = DFS.endRPO(); bb != be; ++bb)
2147 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
2149 // At this point every instruction in the original loop is widened to
2150 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2151 // that we vectorized. The PHI nodes are currently empty because we did
2152 // not want to introduce cycles. Notice that the remaining PHI nodes
2153 // that we need to fix are reduction variables.
2155 // Create the 'reduced' values for each of the induction vars.
2156 // The reduced values are the vector values that we scalarize and combine
2157 // after the loop is finished.
2158 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2160 PHINode *RdxPhi = *it;
2161 assert(RdxPhi && "Unable to recover vectorized PHI");
2163 // Find the reduction variable descriptor.
2164 assert(Legal->getReductionVars()->count(RdxPhi) &&
2165 "Unable to find the reduction variable");
2166 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2167 (*Legal->getReductionVars())[RdxPhi];
2169 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2171 // We need to generate a reduction vector from the incoming scalar.
2172 // To do so, we need to generate the 'identity' vector and overide
2173 // one of the elements with the incoming scalar reduction. We need
2174 // to do it in the vector-loop preheader.
2175 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2177 // This is the vector-clone of the value that leaves the loop.
2178 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2179 Type *VecTy = VectorExit[0]->getType();
2181 // Find the reduction identity variable. Zero for addition, or, xor,
2182 // one for multiplication, -1 for And.
2185 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2186 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2187 // MinMax reduction have the start value as their identify.
2189 VectorStart = Identity = RdxDesc.StartValue;
2191 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2196 // Handle other reduction kinds:
2198 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2199 VecTy->getScalarType());
2202 // This vector is the Identity vector where the first element is the
2203 // incoming scalar reduction.
2204 VectorStart = RdxDesc.StartValue;
2206 Identity = ConstantVector::getSplat(VF, Iden);
2208 // This vector is the Identity vector where the first element is the
2209 // incoming scalar reduction.
2210 VectorStart = Builder.CreateInsertElement(Identity,
2211 RdxDesc.StartValue, Zero);
2215 // Fix the vector-loop phi.
2216 // We created the induction variable so we know that the
2217 // preheader is the first entry.
2218 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2220 // Reductions do not have to start at zero. They can start with
2221 // any loop invariant values.
2222 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2223 BasicBlock *Latch = OrigLoop->getLoopLatch();
2224 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2225 VectorParts &Val = getVectorValue(LoopVal);
2226 for (unsigned part = 0; part < UF; ++part) {
2227 // Make sure to add the reduction stat value only to the
2228 // first unroll part.
2229 Value *StartVal = (part == 0) ? VectorStart : Identity;
2230 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2231 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2234 // Before each round, move the insertion point right between
2235 // the PHIs and the values we are going to write.
2236 // This allows us to write both PHINodes and the extractelement
2238 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2240 VectorParts RdxParts;
2241 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2242 for (unsigned part = 0; part < UF; ++part) {
2243 // This PHINode contains the vectorized reduction variable, or
2244 // the initial value vector, if we bypass the vector loop.
2245 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2246 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2247 Value *StartVal = (part == 0) ? VectorStart : Identity;
2248 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2249 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2250 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2251 RdxParts.push_back(NewPhi);
2254 // Reduce all of the unrolled parts into a single vector.
2255 Value *ReducedPartRdx = RdxParts[0];
2256 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2257 setDebugLocFromInst(Builder, ReducedPartRdx);
2258 for (unsigned part = 1; part < UF; ++part) {
2259 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2260 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2261 RdxParts[part], ReducedPartRdx,
2264 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2265 ReducedPartRdx, RdxParts[part]);
2269 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2270 // and vector ops, reducing the set of values being computed by half each
2272 assert(isPowerOf2_32(VF) &&
2273 "Reduction emission only supported for pow2 vectors!");
2274 Value *TmpVec = ReducedPartRdx;
2275 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2276 for (unsigned i = VF; i != 1; i >>= 1) {
2277 // Move the upper half of the vector to the lower half.
2278 for (unsigned j = 0; j != i/2; ++j)
2279 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2281 // Fill the rest of the mask with undef.
2282 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2283 UndefValue::get(Builder.getInt32Ty()));
2286 Builder.CreateShuffleVector(TmpVec,
2287 UndefValue::get(TmpVec->getType()),
2288 ConstantVector::get(ShuffleMask),
2291 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2292 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2295 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2298 // The result is in the first element of the vector.
2299 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2300 Builder.getInt32(0));
2303 // Now, we need to fix the users of the reduction variable
2304 // inside and outside of the scalar remainder loop.
2305 // We know that the loop is in LCSSA form. We need to update the
2306 // PHI nodes in the exit blocks.
2307 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2308 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2309 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2310 if (!LCSSAPhi) break;
2312 // All PHINodes need to have a single entry edge, or two if
2313 // we already fixed them.
2314 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2316 // We found our reduction value exit-PHI. Update it with the
2317 // incoming bypass edge.
2318 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2319 // Add an edge coming from the bypass.
2320 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2323 }// end of the LCSSA phi scan.
2325 // Fix the scalar loop reduction variable with the incoming reduction sum
2326 // from the vector body and from the backedge value.
2327 int IncomingEdgeBlockIdx =
2328 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2329 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2330 // Pick the other block.
2331 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2332 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2333 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2334 }// end of for each redux variable.
2338 // Remove redundant induction instructions.
2339 cse(LoopVectorBody);
2342 void InnerLoopVectorizer::fixLCSSAPHIs() {
2343 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2344 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2345 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2346 if (!LCSSAPhi) break;
2347 if (LCSSAPhi->getNumIncomingValues() == 1)
2348 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2353 InnerLoopVectorizer::VectorParts
2354 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2355 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2358 // Look for cached value.
2359 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2360 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2361 if (ECEntryIt != MaskCache.end())
2362 return ECEntryIt->second;
2364 VectorParts SrcMask = createBlockInMask(Src);
2366 // The terminator has to be a branch inst!
2367 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2368 assert(BI && "Unexpected terminator found");
2370 if (BI->isConditional()) {
2371 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2373 if (BI->getSuccessor(0) != Dst)
2374 for (unsigned part = 0; part < UF; ++part)
2375 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2377 for (unsigned part = 0; part < UF; ++part)
2378 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2380 MaskCache[Edge] = EdgeMask;
2384 MaskCache[Edge] = SrcMask;
2388 InnerLoopVectorizer::VectorParts
2389 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2390 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2392 // Loop incoming mask is all-one.
2393 if (OrigLoop->getHeader() == BB) {
2394 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2395 return getVectorValue(C);
2398 // This is the block mask. We OR all incoming edges, and with zero.
2399 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2400 VectorParts BlockMask = getVectorValue(Zero);
2403 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2404 VectorParts EM = createEdgeMask(*it, BB);
2405 for (unsigned part = 0; part < UF; ++part)
2406 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2412 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2413 InnerLoopVectorizer::VectorParts &Entry,
2414 LoopVectorizationLegality *Legal,
2415 unsigned UF, unsigned VF, PhiVector *PV) {
2416 PHINode* P = cast<PHINode>(PN);
2417 // Handle reduction variables:
2418 if (Legal->getReductionVars()->count(P)) {
2419 for (unsigned part = 0; part < UF; ++part) {
2420 // This is phase one of vectorizing PHIs.
2421 Type *VecTy = (VF == 1) ? PN->getType() :
2422 VectorType::get(PN->getType(), VF);
2423 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2424 LoopVectorBody-> getFirstInsertionPt());
2430 setDebugLocFromInst(Builder, P);
2431 // Check for PHI nodes that are lowered to vector selects.
2432 if (P->getParent() != OrigLoop->getHeader()) {
2433 // We know that all PHIs in non-header blocks are converted into
2434 // selects, so we don't have to worry about the insertion order and we
2435 // can just use the builder.
2436 // At this point we generate the predication tree. There may be
2437 // duplications since this is a simple recursive scan, but future
2438 // optimizations will clean it up.
2440 unsigned NumIncoming = P->getNumIncomingValues();
2442 // Generate a sequence of selects of the form:
2443 // SELECT(Mask3, In3,
2444 // SELECT(Mask2, In2,
2446 for (unsigned In = 0; In < NumIncoming; In++) {
2447 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2449 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2451 for (unsigned part = 0; part < UF; ++part) {
2452 // We might have single edge PHIs (blocks) - use an identity
2453 // 'select' for the first PHI operand.
2455 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2458 // Select between the current value and the previous incoming edge
2459 // based on the incoming mask.
2460 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2461 Entry[part], "predphi");
2467 // This PHINode must be an induction variable.
2468 // Make sure that we know about it.
2469 assert(Legal->getInductionVars()->count(P) &&
2470 "Not an induction variable");
2472 LoopVectorizationLegality::InductionInfo II =
2473 Legal->getInductionVars()->lookup(P);
2476 case LoopVectorizationLegality::IK_NoInduction:
2477 llvm_unreachable("Unknown induction");
2478 case LoopVectorizationLegality::IK_IntInduction: {
2479 assert(P->getType() == II.StartValue->getType() && "Types must match");
2480 Type *PhiTy = P->getType();
2482 if (P == OldInduction) {
2483 // Handle the canonical induction variable. We might have had to
2485 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2487 // Handle other induction variables that are now based on the
2489 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2491 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2492 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2495 Broadcasted = getBroadcastInstrs(Broadcasted);
2496 // After broadcasting the induction variable we need to make the vector
2497 // consecutive by adding 0, 1, 2, etc.
2498 for (unsigned part = 0; part < UF; ++part)
2499 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2502 case LoopVectorizationLegality::IK_ReverseIntInduction:
2503 case LoopVectorizationLegality::IK_PtrInduction:
2504 case LoopVectorizationLegality::IK_ReversePtrInduction:
2505 // Handle reverse integer and pointer inductions.
2506 Value *StartIdx = ExtendedIdx;
2507 // This is the normalized GEP that starts counting at zero.
2508 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2511 // Handle the reverse integer induction variable case.
2512 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2513 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2514 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2516 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2519 // This is a new value so do not hoist it out.
2520 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2521 // After broadcasting the induction variable we need to make the
2522 // vector consecutive by adding ... -3, -2, -1, 0.
2523 for (unsigned part = 0; part < UF; ++part)
2524 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2529 // Handle the pointer induction variable case.
2530 assert(P->getType()->isPointerTy() && "Unexpected type.");
2532 // Is this a reverse induction ptr or a consecutive induction ptr.
2533 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2536 // This is the vector of results. Notice that we don't generate
2537 // vector geps because scalar geps result in better code.
2538 for (unsigned part = 0; part < UF; ++part) {
2540 int EltIndex = (part) * (Reverse ? -1 : 1);
2541 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2544 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2546 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2548 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2550 Entry[part] = SclrGep;
2554 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2555 for (unsigned int i = 0; i < VF; ++i) {
2556 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2557 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2560 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2562 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2564 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2566 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2567 Builder.getInt32(i),
2570 Entry[part] = VecVal;
2577 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2578 BasicBlock *BB, PhiVector *PV) {
2579 // For each instruction in the old loop.
2580 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2581 VectorParts &Entry = WidenMap.get(it);
2582 switch (it->getOpcode()) {
2583 case Instruction::Br:
2584 // Nothing to do for PHIs and BR, since we already took care of the
2585 // loop control flow instructions.
2587 case Instruction::PHI:{
2588 // Vectorize PHINodes.
2589 widenPHIInstruction(it, Entry, Legal, UF, VF, PV);
2593 case Instruction::Add:
2594 case Instruction::FAdd:
2595 case Instruction::Sub:
2596 case Instruction::FSub:
2597 case Instruction::Mul:
2598 case Instruction::FMul:
2599 case Instruction::UDiv:
2600 case Instruction::SDiv:
2601 case Instruction::FDiv:
2602 case Instruction::URem:
2603 case Instruction::SRem:
2604 case Instruction::FRem:
2605 case Instruction::Shl:
2606 case Instruction::LShr:
2607 case Instruction::AShr:
2608 case Instruction::And:
2609 case Instruction::Or:
2610 case Instruction::Xor: {
2611 // Just widen binops.
2612 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2613 setDebugLocFromInst(Builder, BinOp);
2614 VectorParts &A = getVectorValue(it->getOperand(0));
2615 VectorParts &B = getVectorValue(it->getOperand(1));
2617 // Use this vector value for all users of the original instruction.
2618 for (unsigned Part = 0; Part < UF; ++Part) {
2619 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2621 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2622 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2623 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2624 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2625 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2627 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2628 VecOp->setIsExact(BinOp->isExact());
2634 case Instruction::Select: {
2636 // If the selector is loop invariant we can create a select
2637 // instruction with a scalar condition. Otherwise, use vector-select.
2638 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2640 setDebugLocFromInst(Builder, it);
2642 // The condition can be loop invariant but still defined inside the
2643 // loop. This means that we can't just use the original 'cond' value.
2644 // We have to take the 'vectorized' value and pick the first lane.
2645 // Instcombine will make this a no-op.
2646 VectorParts &Cond = getVectorValue(it->getOperand(0));
2647 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2648 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2650 Value *ScalarCond = (VF == 1) ? Cond[0] :
2651 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2653 for (unsigned Part = 0; Part < UF; ++Part) {
2654 Entry[Part] = Builder.CreateSelect(
2655 InvariantCond ? ScalarCond : Cond[Part],
2662 case Instruction::ICmp:
2663 case Instruction::FCmp: {
2664 // Widen compares. Generate vector compares.
2665 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2666 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2667 setDebugLocFromInst(Builder, it);
2668 VectorParts &A = getVectorValue(it->getOperand(0));
2669 VectorParts &B = getVectorValue(it->getOperand(1));
2670 for (unsigned Part = 0; Part < UF; ++Part) {
2673 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2675 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2681 case Instruction::Store:
2682 case Instruction::Load:
2683 vectorizeMemoryInstruction(it, Legal);
2685 case Instruction::ZExt:
2686 case Instruction::SExt:
2687 case Instruction::FPToUI:
2688 case Instruction::FPToSI:
2689 case Instruction::FPExt:
2690 case Instruction::PtrToInt:
2691 case Instruction::IntToPtr:
2692 case Instruction::SIToFP:
2693 case Instruction::UIToFP:
2694 case Instruction::Trunc:
2695 case Instruction::FPTrunc:
2696 case Instruction::BitCast: {
2697 CastInst *CI = dyn_cast<CastInst>(it);
2698 setDebugLocFromInst(Builder, it);
2699 /// Optimize the special case where the source is the induction
2700 /// variable. Notice that we can only optimize the 'trunc' case
2701 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2702 /// c. other casts depend on pointer size.
2703 if (CI->getOperand(0) == OldInduction &&
2704 it->getOpcode() == Instruction::Trunc) {
2705 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2707 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2708 for (unsigned Part = 0; Part < UF; ++Part)
2709 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2712 /// Vectorize casts.
2713 Type *DestTy = (VF == 1) ? CI->getType() :
2714 VectorType::get(CI->getType(), VF);
2716 VectorParts &A = getVectorValue(it->getOperand(0));
2717 for (unsigned Part = 0; Part < UF; ++Part)
2718 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2722 case Instruction::Call: {
2723 // Ignore dbg intrinsics.
2724 if (isa<DbgInfoIntrinsic>(it))
2726 setDebugLocFromInst(Builder, it);
2728 Module *M = BB->getParent()->getParent();
2729 CallInst *CI = cast<CallInst>(it);
2730 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2731 assert(ID && "Not an intrinsic call!");
2733 case Intrinsic::lifetime_end:
2734 case Intrinsic::lifetime_start:
2735 scalarizeInstruction(it);
2738 for (unsigned Part = 0; Part < UF; ++Part) {
2739 SmallVector<Value *, 4> Args;
2740 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2741 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2742 Args.push_back(Arg[Part]);
2744 Type *Tys[] = {CI->getType()};
2746 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
2748 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2749 Entry[Part] = Builder.CreateCall(F, Args);
2757 // All other instructions are unsupported. Scalarize them.
2758 scalarizeInstruction(it);
2761 }// end of for_each instr.
2764 void InnerLoopVectorizer::updateAnalysis() {
2765 // Forget the original basic block.
2766 SE->forgetLoop(OrigLoop);
2768 // Update the dominator tree information.
2769 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2770 "Entry does not dominate exit.");
2772 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2773 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2774 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2775 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2776 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2777 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2778 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2779 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2781 DEBUG(DT->verifyAnalysis());
2784 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2785 if (!EnableIfConversion)
2788 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2790 // A list of pointers that we can safely read and write to.
2791 SmallPtrSet<Value *, 8> SafePointes;
2793 // Collect safe addresses.
2794 for (Loop::block_iterator BI = TheLoop->block_begin(),
2795 BE = TheLoop->block_end(); BI != BE; ++BI) {
2796 BasicBlock *BB = *BI;
2798 if (blockNeedsPredication(BB))
2801 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2802 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2803 SafePointes.insert(LI->getPointerOperand());
2804 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2805 SafePointes.insert(SI->getPointerOperand());
2809 // Collect the blocks that need predication.
2810 for (Loop::block_iterator BI = TheLoop->block_begin(),
2811 BE = TheLoop->block_end(); BI != BE; ++BI) {
2812 BasicBlock *BB = *BI;
2814 // We don't support switch statements inside loops.
2815 if (!isa<BranchInst>(BB->getTerminator()))
2818 // We must be able to predicate all blocks that need to be predicated.
2819 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2823 // We can if-convert this loop.
2827 bool LoopVectorizationLegality::canVectorize() {
2828 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2829 // be canonicalized.
2830 if (!TheLoop->getLoopPreheader())
2833 // We can only vectorize innermost loops.
2834 if (TheLoop->getSubLoopsVector().size())
2837 // We must have a single backedge.
2838 if (TheLoop->getNumBackEdges() != 1)
2841 // We must have a single exiting block.
2842 if (!TheLoop->getExitingBlock())
2845 // We need to have a loop header.
2846 DEBUG(dbgs() << "LV: Found a loop: " <<
2847 TheLoop->getHeader()->getName() << '\n');
2849 // Check if we can if-convert non-single-bb loops.
2850 unsigned NumBlocks = TheLoop->getNumBlocks();
2851 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2852 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2856 // ScalarEvolution needs to be able to find the exit count.
2857 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2858 if (ExitCount == SE->getCouldNotCompute()) {
2859 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2863 // Do not loop-vectorize loops with a tiny trip count.
2864 BasicBlock *Latch = TheLoop->getLoopLatch();
2865 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2866 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2867 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2868 "This loop is not worth vectorizing.\n");
2872 // Check if we can vectorize the instructions and CFG in this loop.
2873 if (!canVectorizeInstrs()) {
2874 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2878 // Go over each instruction and look at memory deps.
2879 if (!canVectorizeMemory()) {
2880 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2884 // Collect all of the variables that remain uniform after vectorization.
2885 collectLoopUniforms();
2887 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2888 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2891 // Okay! We can vectorize. At this point we don't have any other mem analysis
2892 // which may limit our maximum vectorization factor, so just return true with
2897 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2898 if (Ty->isPointerTy())
2899 return DL.getIntPtrType(Ty);
2901 // It is possible that char's or short's overflow when we ask for the loop's
2902 // trip count, work around this by changing the type size.
2903 if (Ty->getScalarSizeInBits() < 32)
2904 return Type::getInt32Ty(Ty->getContext());
2909 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2910 Ty0 = convertPointerToIntegerType(DL, Ty0);
2911 Ty1 = convertPointerToIntegerType(DL, Ty1);
2912 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2917 /// \brief Check that the instruction has outside loop users and is not an
2918 /// identified reduction variable.
2919 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2920 SmallPtrSet<Value *, 4> &Reductions) {
2921 // Reduction instructions are allowed to have exit users. All other
2922 // instructions must not have external users.
2923 if (!Reductions.count(Inst))
2924 //Check that all of the users of the loop are inside the BB.
2925 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2927 Instruction *U = cast<Instruction>(*I);
2928 // This user may be a reduction exit value.
2929 if (!TheLoop->contains(U)) {
2930 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
2937 bool LoopVectorizationLegality::canVectorizeInstrs() {
2938 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2939 BasicBlock *Header = TheLoop->getHeader();
2941 // Look for the attribute signaling the absence of NaNs.
2942 Function &F = *Header->getParent();
2943 if (F.hasFnAttribute("no-nans-fp-math"))
2944 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2945 AttributeSet::FunctionIndex,
2946 "no-nans-fp-math").getValueAsString() == "true";
2948 // For each block in the loop.
2949 for (Loop::block_iterator bb = TheLoop->block_begin(),
2950 be = TheLoop->block_end(); bb != be; ++bb) {
2952 // Scan the instructions in the block and look for hazards.
2953 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2956 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2957 Type *PhiTy = Phi->getType();
2958 // Check that this PHI type is allowed.
2959 if (!PhiTy->isIntegerTy() &&
2960 !PhiTy->isFloatingPointTy() &&
2961 !PhiTy->isPointerTy()) {
2962 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2966 // If this PHINode is not in the header block, then we know that we
2967 // can convert it to select during if-conversion. No need to check if
2968 // the PHIs in this block are induction or reduction variables.
2969 if (*bb != Header) {
2970 // Check that this instruction has no outside users or is an
2971 // identified reduction value with an outside user.
2972 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2977 // We only allow if-converted PHIs with more than two incoming values.
2978 if (Phi->getNumIncomingValues() != 2) {
2979 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2983 // This is the value coming from the preheader.
2984 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2985 // Check if this is an induction variable.
2986 InductionKind IK = isInductionVariable(Phi);
2988 if (IK_NoInduction != IK) {
2989 // Get the widest type.
2991 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2993 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2995 // Int inductions are special because we only allow one IV.
2996 if (IK == IK_IntInduction) {
2997 // Use the phi node with the widest type as induction. Use the last
2998 // one if there are multiple (no good reason for doing this other
2999 // than it is expedient).
3000 if (!Induction || PhiTy == WidestIndTy)
3004 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3005 Inductions[Phi] = InductionInfo(StartValue, IK);
3007 // Until we explicitly handle the case of an induction variable with
3008 // an outside loop user we have to give up vectorizing this loop.
3009 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3015 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3016 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3019 if (AddReductionVar(Phi, RK_IntegerMult)) {
3020 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3023 if (AddReductionVar(Phi, RK_IntegerOr)) {
3024 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3027 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3028 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3031 if (AddReductionVar(Phi, RK_IntegerXor)) {
3032 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3035 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3036 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3039 if (AddReductionVar(Phi, RK_FloatMult)) {
3040 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3043 if (AddReductionVar(Phi, RK_FloatAdd)) {
3044 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3047 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3048 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3053 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3055 }// end of PHI handling
3057 // We still don't handle functions. However, we can ignore dbg intrinsic
3058 // calls and we do handle certain intrinsic and libm functions.
3059 CallInst *CI = dyn_cast<CallInst>(it);
3060 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3061 DEBUG(dbgs() << "LV: Found a call site.\n");
3065 // Check that the instruction return type is vectorizable.
3066 // Also, we can't vectorize extractelement instructions.
3067 if ((!VectorType::isValidElementType(it->getType()) &&
3068 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3069 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3073 // Check that the stored type is vectorizable.
3074 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3075 Type *T = ST->getValueOperand()->getType();
3076 if (!VectorType::isValidElementType(T))
3080 // Reduction instructions are allowed to have exit users.
3081 // All other instructions must not have external users.
3082 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3090 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3091 if (Inductions.empty())
3098 void LoopVectorizationLegality::collectLoopUniforms() {
3099 // We now know that the loop is vectorizable!
3100 // Collect variables that will remain uniform after vectorization.
3101 std::vector<Value*> Worklist;
3102 BasicBlock *Latch = TheLoop->getLoopLatch();
3104 // Start with the conditional branch and walk up the block.
3105 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3107 while (Worklist.size()) {
3108 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3109 Worklist.pop_back();
3111 // Look at instructions inside this loop.
3112 // Stop when reaching PHI nodes.
3113 // TODO: we need to follow values all over the loop, not only in this block.
3114 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3117 // This is a known uniform.
3120 // Insert all operands.
3121 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3126 /// \brief Analyses memory accesses in a loop.
3128 /// Checks whether run time pointer checks are needed and builds sets for data
3129 /// dependence checking.
3130 class AccessAnalysis {
3132 /// \brief Read or write access location.
3133 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3134 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3136 /// \brief Set of potential dependent memory accesses.
3137 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3139 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3140 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3141 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3143 /// \brief Register a load and whether it is only read from.
3144 void addLoad(Value *Ptr, bool IsReadOnly) {
3145 Accesses.insert(MemAccessInfo(Ptr, false));
3147 ReadOnlyPtr.insert(Ptr);
3150 /// \brief Register a store.
3151 void addStore(Value *Ptr) {
3152 Accesses.insert(MemAccessInfo(Ptr, true));
3155 /// \brief Check whether we can check the pointers at runtime for
3156 /// non-intersection.
3157 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3158 unsigned &NumComparisons, ScalarEvolution *SE,
3159 Loop *TheLoop, bool ShouldCheckStride = false);
3161 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3162 /// and builds sets of dependent accesses.
3163 void buildDependenceSets() {
3164 // Process read-write pointers first.
3165 processMemAccesses(false);
3166 // Next, process read pointers.
3167 processMemAccesses(true);
3170 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3172 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3173 void resetDepChecks() { CheckDeps.clear(); }
3175 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3178 typedef SetVector<MemAccessInfo> PtrAccessSet;
3179 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3181 /// \brief Go over all memory access or only the deferred ones if
3182 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3183 /// and build sets of dependency check candidates.
3184 void processMemAccesses(bool UseDeferred);
3186 /// Set of all accesses.
3187 PtrAccessSet Accesses;
3189 /// Set of access to check after all writes have been processed.
3190 PtrAccessSet DeferredAccesses;
3192 /// Map of pointers to last access encountered.
3193 UnderlyingObjToAccessMap ObjToLastAccess;
3195 /// Set of accesses that need a further dependence check.
3196 MemAccessInfoSet CheckDeps;
3198 /// Set of pointers that are read only.
3199 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3201 /// Set of underlying objects already written to.
3202 SmallPtrSet<Value*, 16> WriteObjects;
3206 /// Sets of potentially dependent accesses - members of one set share an
3207 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3208 /// dependence check.
3209 DepCandidates &DepCands;
3211 bool AreAllWritesIdentified;
3212 bool AreAllReadsIdentified;
3213 bool IsRTCheckNeeded;
3216 } // end anonymous namespace
3218 /// \brief Check whether a pointer can participate in a runtime bounds check.
3219 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
3220 const SCEV *PtrScev = SE->getSCEV(Ptr);
3221 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3225 return AR->isAffine();
3228 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3229 /// the address space.
3230 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3233 bool AccessAnalysis::canCheckPtrAtRT(
3234 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3235 unsigned &NumComparisons, ScalarEvolution *SE,
3236 Loop *TheLoop, bool ShouldCheckStride) {
3237 // Find pointers with computable bounds. We are going to use this information
3238 // to place a runtime bound check.
3239 unsigned NumReadPtrChecks = 0;
3240 unsigned NumWritePtrChecks = 0;
3241 bool CanDoRT = true;
3243 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3244 // We assign consecutive id to access from different dependence sets.
3245 // Accesses within the same set don't need a runtime check.
3246 unsigned RunningDepId = 1;
3247 DenseMap<Value *, unsigned> DepSetId;
3249 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3251 const MemAccessInfo &Access = *AI;
3252 Value *Ptr = Access.getPointer();
3253 bool IsWrite = Access.getInt();
3255 // Just add write checks if we have both.
3256 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3260 ++NumWritePtrChecks;
3264 if (hasComputableBounds(SE, Ptr) &&
3265 // When we run after a failing dependency check we have to make sure we
3266 // don't have wrapping pointers.
3267 (!ShouldCheckStride || isStridedPtr(SE, DL, Ptr, TheLoop) == 1)) {
3268 // The id of the dependence set.
3271 if (IsDepCheckNeeded) {
3272 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3273 unsigned &LeaderId = DepSetId[Leader];
3275 LeaderId = RunningDepId++;
3278 // Each access has its own dependence set.
3279 DepId = RunningDepId++;
3281 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3283 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3289 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3290 NumComparisons = 0; // Only one dependence set.
3292 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3293 NumWritePtrChecks - 1));
3296 // If the pointers that we would use for the bounds comparison have different
3297 // address spaces, assume the values aren't directly comparable, so we can't
3298 // use them for the runtime check. We also have to assume they could
3299 // overlap. In the future there should be metadata for whether address spaces
3301 unsigned NumPointers = RtCheck.Pointers.size();
3302 for (unsigned i = 0; i < NumPointers; ++i) {
3303 for (unsigned j = i + 1; j < NumPointers; ++j) {
3304 // Only need to check pointers between two different dependency sets.
3305 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3308 Value *PtrI = RtCheck.Pointers[i];
3309 Value *PtrJ = RtCheck.Pointers[j];
3311 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3312 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3314 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3315 " different address spaces\n");
3324 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3325 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3328 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3329 // We process the set twice: first we process read-write pointers, last we
3330 // process read-only pointers. This allows us to skip dependence tests for
3331 // read-only pointers.
3333 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3334 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3335 const MemAccessInfo &Access = *AI;
3336 Value *Ptr = Access.getPointer();
3337 bool IsWrite = Access.getInt();
3339 DepCands.insert(Access);
3341 // Memorize read-only pointers for later processing and skip them in the
3342 // first round (they need to be checked after we have seen all write
3343 // pointers). Note: we also mark pointer that are not consecutive as
3344 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3345 // second check for "!IsWrite".
3346 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3347 if (!UseDeferred && IsReadOnlyPtr) {
3348 DeferredAccesses.insert(Access);
3352 bool NeedDepCheck = false;
3353 // Check whether there is the possiblity of dependency because of underlying
3354 // objects being the same.
3355 typedef SmallVector<Value*, 16> ValueVector;
3356 ValueVector TempObjects;
3357 GetUnderlyingObjects(Ptr, TempObjects, DL);
3358 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3360 Value *UnderlyingObj = *UI;
3362 // If this is a write then it needs to be an identified object. If this a
3363 // read and all writes (so far) are identified function scope objects we
3364 // don't need an identified underlying object but only an Argument (the
3365 // next write is going to invalidate this assumption if it is
3367 // This is a micro-optimization for the case where all writes are
3368 // identified and we have one argument pointer.
3369 // Otherwise, we do need a runtime check.
3370 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3371 (!IsWrite && (!AreAllWritesIdentified ||
3372 !isa<Argument>(UnderlyingObj)) &&
3373 !isIdentifiedObject(UnderlyingObj))) {
3374 DEBUG(dbgs() << "LV: Found an unidentified " <<
3375 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3377 IsRTCheckNeeded = (IsRTCheckNeeded ||
3378 !isIdentifiedObject(UnderlyingObj) ||
3379 !AreAllReadsIdentified);
3382 AreAllWritesIdentified = false;
3384 AreAllReadsIdentified = false;
3387 // If this is a write - check other reads and writes for conflicts. If
3388 // this is a read only check other writes for conflicts (but only if there
3389 // is no other write to the ptr - this is an optimization to catch "a[i] =
3390 // a[i] + " without having to do a dependence check).
3391 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3392 NeedDepCheck = true;
3395 WriteObjects.insert(UnderlyingObj);
3397 // Create sets of pointers connected by shared underlying objects.
3398 UnderlyingObjToAccessMap::iterator Prev =
3399 ObjToLastAccess.find(UnderlyingObj);
3400 if (Prev != ObjToLastAccess.end())
3401 DepCands.unionSets(Access, Prev->second);
3403 ObjToLastAccess[UnderlyingObj] = Access;
3407 CheckDeps.insert(Access);
3412 /// \brief Checks memory dependences among accesses to the same underlying
3413 /// object to determine whether there vectorization is legal or not (and at
3414 /// which vectorization factor).
3416 /// This class works under the assumption that we already checked that memory
3417 /// locations with different underlying pointers are "must-not alias".
3418 /// We use the ScalarEvolution framework to symbolically evalutate access
3419 /// functions pairs. Since we currently don't restructure the loop we can rely
3420 /// on the program order of memory accesses to determine their safety.
3421 /// At the moment we will only deem accesses as safe for:
3422 /// * A negative constant distance assuming program order.
3424 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3425 /// a[i] = tmp; y = a[i];
3427 /// The latter case is safe because later checks guarantuee that there can't
3428 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3429 /// the same variable: a header phi can only be an induction or a reduction, a
3430 /// reduction can't have a memory sink, an induction can't have a memory
3431 /// source). This is important and must not be violated (or we have to
3432 /// resort to checking for cycles through memory).
3434 /// * A positive constant distance assuming program order that is bigger
3435 /// than the biggest memory access.
3437 /// tmp = a[i] OR b[i] = x
3438 /// a[i+2] = tmp y = b[i+2];
3440 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3442 /// * Zero distances and all accesses have the same size.
3444 class MemoryDepChecker {
3446 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3447 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3449 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3450 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3451 ShouldRetryWithRuntimeCheck(false) {}
3453 /// \brief Register the location (instructions are given increasing numbers)
3454 /// of a write access.
3455 void addAccess(StoreInst *SI) {
3456 Value *Ptr = SI->getPointerOperand();
3457 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3458 InstMap.push_back(SI);
3462 /// \brief Register the location (instructions are given increasing numbers)
3463 /// of a write access.
3464 void addAccess(LoadInst *LI) {
3465 Value *Ptr = LI->getPointerOperand();
3466 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3467 InstMap.push_back(LI);
3471 /// \brief Check whether the dependencies between the accesses are safe.
3473 /// Only checks sets with elements in \p CheckDeps.
3474 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3475 MemAccessInfoSet &CheckDeps);
3477 /// \brief The maximum number of bytes of a vector register we can vectorize
3478 /// the accesses safely with.
3479 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3481 /// \brief In same cases when the dependency check fails we can still
3482 /// vectorize the loop with a dynamic array access check.
3483 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3486 ScalarEvolution *SE;
3488 const Loop *InnermostLoop;
3490 /// \brief Maps access locations (ptr, read/write) to program order.
3491 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3493 /// \brief Memory access instructions in program order.
3494 SmallVector<Instruction *, 16> InstMap;
3496 /// \brief The program order index to be used for the next instruction.
3499 // We can access this many bytes in parallel safely.
3500 unsigned MaxSafeDepDistBytes;
3502 /// \brief If we see a non-constant dependence distance we can still try to
3503 /// vectorize this loop with runtime checks.
3504 bool ShouldRetryWithRuntimeCheck;
3506 /// \brief Check whether there is a plausible dependence between the two
3509 /// Access \p A must happen before \p B in program order. The two indices
3510 /// identify the index into the program order map.
3512 /// This function checks whether there is a plausible dependence (or the
3513 /// absence of such can't be proved) between the two accesses. If there is a
3514 /// plausible dependence but the dependence distance is bigger than one
3515 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3516 /// distance is smaller than any other distance encountered so far).
3517 /// Otherwise, this function returns true signaling a possible dependence.
3518 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3519 const MemAccessInfo &B, unsigned BIdx);
3521 /// \brief Check whether the data dependence could prevent store-load
3523 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3526 } // end anonymous namespace
3528 static bool isInBoundsGep(Value *Ptr) {
3529 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3530 return GEP->isInBounds();
3534 /// \brief Check whether the access through \p Ptr has a constant stride.
3535 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3537 const Type *Ty = Ptr->getType();
3538 assert(Ty->isPointerTy() && "Unexpected non-ptr");
3540 // Make sure that the pointer does not point to aggregate types.
3541 const PointerType *PtrTy = cast<PointerType>(Ty);
3542 if (PtrTy->getElementType()->isAggregateType()) {
3543 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3548 const SCEV *PtrScev = SE->getSCEV(Ptr);
3549 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3551 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3552 << *Ptr << " SCEV: " << *PtrScev << "\n");
3556 // The accesss function must stride over the innermost loop.
3557 if (Lp != AR->getLoop()) {
3558 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3559 *Ptr << " SCEV: " << *PtrScev << "\n");
3562 // The address calculation must not wrap. Otherwise, a dependence could be
3564 // An inbounds getelementptr that is a AddRec with a unit stride
3565 // cannot wrap per definition. The unit stride requirement is checked later.
3566 // An getelementptr without an inbounds attribute and unit stride would have
3567 // to access the pointer value "0" which is undefined behavior in address
3568 // space 0, therefore we can also vectorize this case.
3569 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3570 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3571 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3572 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3573 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3574 << *Ptr << " SCEV: " << *PtrScev << "\n");
3578 // Check the step is constant.
3579 const SCEV *Step = AR->getStepRecurrence(*SE);
3581 // Calculate the pointer stride and check if it is consecutive.
3582 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3584 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3585 " SCEV: " << *PtrScev << "\n");
3589 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3590 const APInt &APStepVal = C->getValue()->getValue();
3592 // Huge step value - give up.
3593 if (APStepVal.getBitWidth() > 64)
3596 int64_t StepVal = APStepVal.getSExtValue();
3599 int64_t Stride = StepVal / Size;
3600 int64_t Rem = StepVal % Size;
3604 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3605 // know we can't "wrap around the address space". In case of address space
3606 // zero we know that this won't happen without triggering undefined behavior.
3607 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3608 Stride != 1 && Stride != -1)
3614 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3615 unsigned TypeByteSize) {
3616 // If loads occur at a distance that is not a multiple of a feasible vector
3617 // factor store-load forwarding does not take place.
3618 // Positive dependences might cause troubles because vectorizing them might
3619 // prevent store-load forwarding making vectorized code run a lot slower.
3620 // a[i] = a[i-3] ^ a[i-8];
3621 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3622 // hence on your typical architecture store-load forwarding does not take
3623 // place. Vectorizing in such cases does not make sense.
3624 // Store-load forwarding distance.
3625 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3626 // Maximum vector factor.
3627 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3628 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3629 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3631 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3633 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3634 MaxVFWithoutSLForwardIssues = (vf >>=1);
3639 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3640 DEBUG(dbgs() << "LV: Distance " << Distance <<
3641 " that could cause a store-load forwarding conflict\n");
3645 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3646 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3647 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3651 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3652 const MemAccessInfo &B, unsigned BIdx) {
3653 assert (AIdx < BIdx && "Must pass arguments in program order");
3655 Value *APtr = A.getPointer();
3656 Value *BPtr = B.getPointer();
3657 bool AIsWrite = A.getInt();
3658 bool BIsWrite = B.getInt();
3660 // Two reads are independent.
3661 if (!AIsWrite && !BIsWrite)
3664 const SCEV *AScev = SE->getSCEV(APtr);
3665 const SCEV *BScev = SE->getSCEV(BPtr);
3667 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3668 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3670 const SCEV *Src = AScev;
3671 const SCEV *Sink = BScev;
3673 // If the induction step is negative we have to invert source and sink of the
3675 if (StrideAPtr < 0) {
3678 std::swap(APtr, BPtr);
3679 std::swap(Src, Sink);
3680 std::swap(AIsWrite, BIsWrite);
3681 std::swap(AIdx, BIdx);
3682 std::swap(StrideAPtr, StrideBPtr);
3685 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3687 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3688 << "(Induction step: " << StrideAPtr << ")\n");
3689 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3690 << *InstMap[BIdx] << ": " << *Dist << "\n");
3692 // Need consecutive accesses. We don't want to vectorize
3693 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3694 // the address space.
3695 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3696 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3700 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3702 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
3703 ShouldRetryWithRuntimeCheck = true;
3707 Type *ATy = APtr->getType()->getPointerElementType();
3708 Type *BTy = BPtr->getType()->getPointerElementType();
3709 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3711 // Negative distances are not plausible dependencies.
3712 const APInt &Val = C->getValue()->getValue();
3713 if (Val.isNegative()) {
3714 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3715 if (IsTrueDataDependence &&
3716 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3720 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3724 // Write to the same location with the same size.
3725 // Could be improved to assert type sizes are the same (i32 == float, etc).
3729 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
3733 assert(Val.isStrictlyPositive() && "Expect a positive value");
3735 // Positive distance bigger than max vectorization factor.
3738 "LV: ReadWrite-Write positive dependency with different types\n");
3742 unsigned Distance = (unsigned) Val.getZExtValue();
3744 // Bail out early if passed-in parameters make vectorization not feasible.
3745 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3746 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3748 // The distance must be bigger than the size needed for a vectorized version
3749 // of the operation and the size of the vectorized operation must not be
3750 // bigger than the currrent maximum size.
3751 if (Distance < 2*TypeByteSize ||
3752 2*TypeByteSize > MaxSafeDepDistBytes ||
3753 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3754 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3755 << Val.getSExtValue() << '\n');
3759 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3760 Distance : MaxSafeDepDistBytes;
3762 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3763 if (IsTrueDataDependence &&
3764 couldPreventStoreLoadForward(Distance, TypeByteSize))
3767 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3768 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
3774 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3775 MemAccessInfoSet &CheckDeps) {
3777 MaxSafeDepDistBytes = -1U;
3778 while (!CheckDeps.empty()) {
3779 MemAccessInfo CurAccess = *CheckDeps.begin();
3781 // Get the relevant memory access set.
3782 EquivalenceClasses<MemAccessInfo>::iterator I =
3783 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3785 // Check accesses within this set.
3786 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3787 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3789 // Check every access pair.
3791 CheckDeps.erase(*AI);
3792 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3794 // Check every accessing instruction pair in program order.
3795 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3796 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3797 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3798 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3799 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3801 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3812 bool LoopVectorizationLegality::canVectorizeMemory() {
3814 typedef SmallVector<Value*, 16> ValueVector;
3815 typedef SmallPtrSet<Value*, 16> ValueSet;
3817 // Holds the Load and Store *instructions*.
3821 // Holds all the different accesses in the loop.
3822 unsigned NumReads = 0;
3823 unsigned NumReadWrites = 0;
3825 PtrRtCheck.Pointers.clear();
3826 PtrRtCheck.Need = false;
3828 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3829 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3832 for (Loop::block_iterator bb = TheLoop->block_begin(),
3833 be = TheLoop->block_end(); bb != be; ++bb) {
3835 // Scan the BB and collect legal loads and stores.
3836 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3839 // If this is a load, save it. If this instruction can read from memory
3840 // but is not a load, then we quit. Notice that we don't handle function
3841 // calls that read or write.
3842 if (it->mayReadFromMemory()) {
3843 // Many math library functions read the rounding mode. We will only
3844 // vectorize a loop if it contains known function calls that don't set
3845 // the flag. Therefore, it is safe to ignore this read from memory.
3846 CallInst *Call = dyn_cast<CallInst>(it);
3847 if (Call && getIntrinsicIDForCall(Call, TLI))
3850 LoadInst *Ld = dyn_cast<LoadInst>(it);
3851 if (!Ld) return false;
3852 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3853 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3856 Loads.push_back(Ld);
3857 DepChecker.addAccess(Ld);
3861 // Save 'store' instructions. Abort if other instructions write to memory.
3862 if (it->mayWriteToMemory()) {
3863 StoreInst *St = dyn_cast<StoreInst>(it);
3864 if (!St) return false;
3865 if (!St->isSimple() && !IsAnnotatedParallel) {
3866 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3869 Stores.push_back(St);
3870 DepChecker.addAccess(St);
3875 // Now we have two lists that hold the loads and the stores.
3876 // Next, we find the pointers that they use.
3878 // Check if we see any stores. If there are no stores, then we don't
3879 // care if the pointers are *restrict*.
3880 if (!Stores.size()) {
3881 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3885 AccessAnalysis::DepCandidates DependentAccesses;
3886 AccessAnalysis Accesses(DL, DependentAccesses);
3888 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3889 // multiple times on the same object. If the ptr is accessed twice, once
3890 // for read and once for write, it will only appear once (on the write
3891 // list). This is okay, since we are going to check for conflicts between
3892 // writes and between reads and writes, but not between reads and reads.
3895 ValueVector::iterator I, IE;
3896 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3897 StoreInst *ST = cast<StoreInst>(*I);
3898 Value* Ptr = ST->getPointerOperand();
3900 if (isUniform(Ptr)) {
3901 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3905 // If we did *not* see this pointer before, insert it to the read-write
3906 // list. At this phase it is only a 'write' list.
3907 if (Seen.insert(Ptr)) {
3909 Accesses.addStore(Ptr);
3913 if (IsAnnotatedParallel) {
3915 << "LV: A loop annotated parallel, ignore memory dependency "
3920 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3921 LoadInst *LD = cast<LoadInst>(*I);
3922 Value* Ptr = LD->getPointerOperand();
3923 // If we did *not* see this pointer before, insert it to the
3924 // read list. If we *did* see it before, then it is already in
3925 // the read-write list. This allows us to vectorize expressions
3926 // such as A[i] += x; Because the address of A[i] is a read-write
3927 // pointer. This only works if the index of A[i] is consecutive.
3928 // If the address of i is unknown (for example A[B[i]]) then we may
3929 // read a few words, modify, and write a few words, and some of the
3930 // words may be written to the same address.
3931 bool IsReadOnlyPtr = false;
3932 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3934 IsReadOnlyPtr = true;
3936 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3939 // If we write (or read-write) to a single destination and there are no
3940 // other reads in this loop then is it safe to vectorize.
3941 if (NumReadWrites == 1 && NumReads == 0) {
3942 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3946 // Build dependence sets and check whether we need a runtime pointer bounds
3948 Accesses.buildDependenceSets();
3949 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3951 // Find pointers with computable bounds. We are going to use this information
3952 // to place a runtime bound check.
3953 unsigned NumComparisons = 0;
3954 bool CanDoRT = false;
3956 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3959 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3960 " pointer comparisons.\n");
3962 // If we only have one set of dependences to check pointers among we don't
3963 // need a runtime check.
3964 if (NumComparisons == 0 && NeedRTCheck)
3965 NeedRTCheck = false;
3967 // Check that we did not collect too many pointers or found an unsizeable
3969 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3975 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3978 if (NeedRTCheck && !CanDoRT) {
3979 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3980 "the array bounds.\n");
3985 PtrRtCheck.Need = NeedRTCheck;
3987 bool CanVecMem = true;
3988 if (Accesses.isDependencyCheckNeeded()) {
3989 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3990 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3991 Accesses.getDependenciesToCheck());
3992 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3994 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
3995 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
3998 // Clear the dependency checks. We assume they are not needed.
3999 Accesses.resetDepChecks();
4002 PtrRtCheck.Need = true;
4004 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4006 // Check that we did not collect too many pointers or found an unsizeable
4008 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4009 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4018 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4019 " need a runtime memory check.\n");
4024 static bool hasMultipleUsesOf(Instruction *I,
4025 SmallPtrSet<Instruction *, 8> &Insts) {
4026 unsigned NumUses = 0;
4027 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4028 if (Insts.count(dyn_cast<Instruction>(*Use)))
4037 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4038 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4039 if (!Set.count(dyn_cast<Instruction>(*Use)))
4044 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4045 ReductionKind Kind) {
4046 if (Phi->getNumIncomingValues() != 2)
4049 // Reduction variables are only found in the loop header block.
4050 if (Phi->getParent() != TheLoop->getHeader())
4053 // Obtain the reduction start value from the value that comes from the loop
4055 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4057 // ExitInstruction is the single value which is used outside the loop.
4058 // We only allow for a single reduction value to be used outside the loop.
4059 // This includes users of the reduction, variables (which form a cycle
4060 // which ends in the phi node).
4061 Instruction *ExitInstruction = 0;
4062 // Indicates that we found a reduction operation in our scan.
4063 bool FoundReduxOp = false;
4065 // We start with the PHI node and scan for all of the users of this
4066 // instruction. All users must be instructions that can be used as reduction
4067 // variables (such as ADD). We must have a single out-of-block user. The cycle
4068 // must include the original PHI.
4069 bool FoundStartPHI = false;
4071 // To recognize min/max patterns formed by a icmp select sequence, we store
4072 // the number of instruction we saw from the recognized min/max pattern,
4073 // to make sure we only see exactly the two instructions.
4074 unsigned NumCmpSelectPatternInst = 0;
4075 ReductionInstDesc ReduxDesc(false, 0);
4077 SmallPtrSet<Instruction *, 8> VisitedInsts;
4078 SmallVector<Instruction *, 8> Worklist;
4079 Worklist.push_back(Phi);
4080 VisitedInsts.insert(Phi);
4082 // A value in the reduction can be used:
4083 // - By the reduction:
4084 // - Reduction operation:
4085 // - One use of reduction value (safe).
4086 // - Multiple use of reduction value (not safe).
4088 // - All uses of the PHI must be the reduction (safe).
4089 // - Otherwise, not safe.
4090 // - By one instruction outside of the loop (safe).
4091 // - By further instructions outside of the loop (not safe).
4092 // - By an instruction that is not part of the reduction (not safe).
4094 // * An instruction type other than PHI or the reduction operation.
4095 // * A PHI in the header other than the initial PHI.
4096 while (!Worklist.empty()) {
4097 Instruction *Cur = Worklist.back();
4098 Worklist.pop_back();
4101 // If the instruction has no users then this is a broken chain and can't be
4102 // a reduction variable.
4103 if (Cur->use_empty())
4106 bool IsAPhi = isa<PHINode>(Cur);
4108 // A header PHI use other than the original PHI.
4109 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4112 // Reductions of instructions such as Div, and Sub is only possible if the
4113 // LHS is the reduction variable.
4114 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4115 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4116 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4119 // Any reduction instruction must be of one of the allowed kinds.
4120 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4121 if (!ReduxDesc.IsReduction)
4124 // A reduction operation must only have one use of the reduction value.
4125 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4126 hasMultipleUsesOf(Cur, VisitedInsts))
4129 // All inputs to a PHI node must be a reduction value.
4130 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4133 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4134 isa<SelectInst>(Cur)))
4135 ++NumCmpSelectPatternInst;
4136 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4137 isa<SelectInst>(Cur)))
4138 ++NumCmpSelectPatternInst;
4140 // Check whether we found a reduction operator.
4141 FoundReduxOp |= !IsAPhi;
4143 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4144 // onto the stack. This way we are going to have seen all inputs to PHI
4145 // nodes once we get to them.
4146 SmallVector<Instruction *, 8> NonPHIs;
4147 SmallVector<Instruction *, 8> PHIs;
4148 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4150 Instruction *Usr = cast<Instruction>(*UI);
4152 // Check if we found the exit user.
4153 BasicBlock *Parent = Usr->getParent();
4154 if (!TheLoop->contains(Parent)) {
4155 // Exit if you find multiple outside users or if the header phi node is
4156 // being used. In this case the user uses the value of the previous
4157 // iteration, in which case we would loose "VF-1" iterations of the
4158 // reduction operation if we vectorize.
4159 if (ExitInstruction != 0 || Cur == Phi)
4162 // The instruction used by an outside user must be the last instruction
4163 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4164 // operations on the value.
4165 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4168 ExitInstruction = Cur;
4172 // Process instructions only once (termination).
4173 if (VisitedInsts.insert(Usr)) {
4174 if (isa<PHINode>(Usr))
4175 PHIs.push_back(Usr);
4177 NonPHIs.push_back(Usr);
4179 // Remember that we completed the cycle.
4181 FoundStartPHI = true;
4183 Worklist.append(PHIs.begin(), PHIs.end());
4184 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4187 // This means we have seen one but not the other instruction of the
4188 // pattern or more than just a select and cmp.
4189 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4190 NumCmpSelectPatternInst != 2)
4193 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4196 // We found a reduction var if we have reached the original phi node and we
4197 // only have a single instruction with out-of-loop users.
4199 // This instruction is allowed to have out-of-loop users.
4200 AllowedExit.insert(ExitInstruction);
4202 // Save the description of this reduction variable.
4203 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4204 ReduxDesc.MinMaxKind);
4205 Reductions[Phi] = RD;
4206 // We've ended the cycle. This is a reduction variable if we have an
4207 // outside user and it has a binary op.
4212 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4213 /// pattern corresponding to a min(X, Y) or max(X, Y).
4214 LoopVectorizationLegality::ReductionInstDesc
4215 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4216 ReductionInstDesc &Prev) {
4218 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4219 "Expect a select instruction");
4220 Instruction *Cmp = 0;
4221 SelectInst *Select = 0;
4223 // We must handle the select(cmp()) as a single instruction. Advance to the
4225 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4226 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4227 return ReductionInstDesc(false, I);
4228 return ReductionInstDesc(Select, Prev.MinMaxKind);
4231 // Only handle single use cases for now.
4232 if (!(Select = dyn_cast<SelectInst>(I)))
4233 return ReductionInstDesc(false, I);
4234 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4235 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4236 return ReductionInstDesc(false, I);
4237 if (!Cmp->hasOneUse())
4238 return ReductionInstDesc(false, I);
4243 // Look for a min/max pattern.
4244 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4245 return ReductionInstDesc(Select, MRK_UIntMin);
4246 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4247 return ReductionInstDesc(Select, MRK_UIntMax);
4248 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4249 return ReductionInstDesc(Select, MRK_SIntMax);
4250 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4251 return ReductionInstDesc(Select, MRK_SIntMin);
4252 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4253 return ReductionInstDesc(Select, MRK_FloatMin);
4254 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4255 return ReductionInstDesc(Select, MRK_FloatMax);
4256 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4257 return ReductionInstDesc(Select, MRK_FloatMin);
4258 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4259 return ReductionInstDesc(Select, MRK_FloatMax);
4261 return ReductionInstDesc(false, I);
4264 LoopVectorizationLegality::ReductionInstDesc
4265 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4267 ReductionInstDesc &Prev) {
4268 bool FP = I->getType()->isFloatingPointTy();
4269 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4270 switch (I->getOpcode()) {
4272 return ReductionInstDesc(false, I);
4273 case Instruction::PHI:
4274 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4275 Kind != RK_FloatMinMax))
4276 return ReductionInstDesc(false, I);
4277 return ReductionInstDesc(I, Prev.MinMaxKind);
4278 case Instruction::Sub:
4279 case Instruction::Add:
4280 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4281 case Instruction::Mul:
4282 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4283 case Instruction::And:
4284 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4285 case Instruction::Or:
4286 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4287 case Instruction::Xor:
4288 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4289 case Instruction::FMul:
4290 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4291 case Instruction::FAdd:
4292 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4293 case Instruction::FCmp:
4294 case Instruction::ICmp:
4295 case Instruction::Select:
4296 if (Kind != RK_IntegerMinMax &&
4297 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4298 return ReductionInstDesc(false, I);
4299 return isMinMaxSelectCmpPattern(I, Prev);
4303 LoopVectorizationLegality::InductionKind
4304 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4305 Type *PhiTy = Phi->getType();
4306 // We only handle integer and pointer inductions variables.
4307 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4308 return IK_NoInduction;
4310 // Check that the PHI is consecutive.
4311 const SCEV *PhiScev = SE->getSCEV(Phi);
4312 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4314 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4315 return IK_NoInduction;
4317 const SCEV *Step = AR->getStepRecurrence(*SE);
4319 // Integer inductions need to have a stride of one.
4320 if (PhiTy->isIntegerTy()) {
4322 return IK_IntInduction;
4323 if (Step->isAllOnesValue())
4324 return IK_ReverseIntInduction;
4325 return IK_NoInduction;
4328 // Calculate the pointer stride and check if it is consecutive.
4329 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4331 return IK_NoInduction;
4333 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4334 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4335 if (C->getValue()->equalsInt(Size))
4336 return IK_PtrInduction;
4337 else if (C->getValue()->equalsInt(0 - Size))
4338 return IK_ReversePtrInduction;
4340 return IK_NoInduction;
4343 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4344 Value *In0 = const_cast<Value*>(V);
4345 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4349 return Inductions.count(PN);
4352 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4353 assert(TheLoop->contains(BB) && "Unknown block used");
4355 // Blocks that do not dominate the latch need predication.
4356 BasicBlock* Latch = TheLoop->getLoopLatch();
4357 return !DT->dominates(BB, Latch);
4360 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4361 SmallPtrSet<Value *, 8>& SafePtrs) {
4362 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4363 // We might be able to hoist the load.
4364 if (it->mayReadFromMemory()) {
4365 LoadInst *LI = dyn_cast<LoadInst>(it);
4366 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4370 // We don't predicate stores at the moment.
4371 if (it->mayWriteToMemory() || it->mayThrow())
4374 // The instructions below can trap.
4375 switch (it->getOpcode()) {
4377 case Instruction::UDiv:
4378 case Instruction::SDiv:
4379 case Instruction::URem:
4380 case Instruction::SRem:
4388 LoopVectorizationCostModel::VectorizationFactor
4389 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4391 // Width 1 means no vectorize
4392 VectorizationFactor Factor = { 1U, 0U };
4393 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4394 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4398 // Find the trip count.
4399 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4400 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4402 unsigned WidestType = getWidestType();
4403 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4404 unsigned MaxSafeDepDist = -1U;
4405 if (Legal->getMaxSafeDepDistBytes() != -1U)
4406 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4407 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4408 WidestRegister : MaxSafeDepDist);
4409 unsigned MaxVectorSize = WidestRegister / WidestType;
4410 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4411 DEBUG(dbgs() << "LV: The Widest register is: "
4412 << WidestRegister << " bits.\n");
4414 if (MaxVectorSize == 0) {
4415 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4419 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4420 " into one vector!");
4422 unsigned VF = MaxVectorSize;
4424 // If we optimize the program for size, avoid creating the tail loop.
4426 // If we are unable to calculate the trip count then don't try to vectorize.
4428 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4432 // Find the maximum SIMD width that can fit within the trip count.
4433 VF = TC % MaxVectorSize;
4438 // If the trip count that we found modulo the vectorization factor is not
4439 // zero then we require a tail.
4441 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4447 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4448 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4450 Factor.Width = UserVF;
4454 float Cost = expectedCost(1);
4456 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
4457 for (unsigned i=2; i <= VF; i*=2) {
4458 // Notice that the vector loop needs to be executed less times, so
4459 // we need to divide the cost of the vector loops by the width of
4460 // the vector elements.
4461 float VectorCost = expectedCost(i) / (float)i;
4462 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4463 (int)VectorCost << ".\n");
4464 if (VectorCost < Cost) {
4470 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4471 Factor.Width = Width;
4472 Factor.Cost = Width * Cost;
4476 unsigned LoopVectorizationCostModel::getWidestType() {
4477 unsigned MaxWidth = 8;
4480 for (Loop::block_iterator bb = TheLoop->block_begin(),
4481 be = TheLoop->block_end(); bb != be; ++bb) {
4482 BasicBlock *BB = *bb;
4484 // For each instruction in the loop.
4485 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4486 Type *T = it->getType();
4488 // Only examine Loads, Stores and PHINodes.
4489 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4492 // Examine PHI nodes that are reduction variables.
4493 if (PHINode *PN = dyn_cast<PHINode>(it))
4494 if (!Legal->getReductionVars()->count(PN))
4497 // Examine the stored values.
4498 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4499 T = ST->getValueOperand()->getType();
4501 // Ignore loaded pointer types and stored pointer types that are not
4502 // consecutive. However, we do want to take consecutive stores/loads of
4503 // pointer vectors into account.
4504 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4507 MaxWidth = std::max(MaxWidth,
4508 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4516 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4519 unsigned LoopCost) {
4521 // -- The unroll heuristics --
4522 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4523 // There are many micro-architectural considerations that we can't predict
4524 // at this level. For example frontend pressure (on decode or fetch) due to
4525 // code size, or the number and capabilities of the execution ports.
4527 // We use the following heuristics to select the unroll factor:
4528 // 1. If the code has reductions the we unroll in order to break the cross
4529 // iteration dependency.
4530 // 2. If the loop is really small then we unroll in order to reduce the loop
4532 // 3. We don't unroll if we think that we will spill registers to memory due
4533 // to the increased register pressure.
4535 // Use the user preference, unless 'auto' is selected.
4539 // When we optimize for size we don't unroll.
4543 // We used the distance for the unroll factor.
4544 if (Legal->getMaxSafeDepDistBytes() != -1U)
4547 // Do not unroll loops with a relatively small trip count.
4548 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4549 TheLoop->getLoopLatch());
4550 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4553 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4554 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4555 " vector registers\n");
4557 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4558 // We divide by these constants so assume that we have at least one
4559 // instruction that uses at least one register.
4560 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4561 R.NumInstructions = std::max(R.NumInstructions, 1U);
4563 // We calculate the unroll factor using the following formula.
4564 // Subtract the number of loop invariants from the number of available
4565 // registers. These registers are used by all of the unrolled instances.
4566 // Next, divide the remaining registers by the number of registers that is
4567 // required by the loop, in order to estimate how many parallel instances
4568 // fit without causing spills.
4569 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4571 // Clamp the unroll factor ranges to reasonable factors.
4572 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4574 // If we did not calculate the cost for VF (because the user selected the VF)
4575 // then we calculate the cost of VF here.
4577 LoopCost = expectedCost(VF);
4579 // Clamp the calculated UF to be between the 1 and the max unroll factor
4580 // that the target allows.
4581 if (UF > MaxUnrollSize)
4586 bool HasReductions = Legal->getReductionVars()->size();
4588 // Decide if we want to unroll if we decided that it is legal to vectorize
4589 // but not profitable.
4591 if (TheLoop->getNumBlocks() > 1 || !HasReductions ||
4592 LoopCost > SmallLoopCost)
4598 if (HasReductions) {
4599 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4603 // We want to unroll tiny loops in order to reduce the loop overhead.
4604 // We assume that the cost overhead is 1 and we use the cost model
4605 // to estimate the cost of the loop and unroll until the cost of the
4606 // loop overhead is about 5% of the cost of the loop.
4607 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4608 if (LoopCost < SmallLoopCost) {
4609 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4610 unsigned NewUF = SmallLoopCost / (LoopCost + 1);
4611 return std::min(NewUF, UF);
4614 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4618 LoopVectorizationCostModel::RegisterUsage
4619 LoopVectorizationCostModel::calculateRegisterUsage() {
4620 // This function calculates the register usage by measuring the highest number
4621 // of values that are alive at a single location. Obviously, this is a very
4622 // rough estimation. We scan the loop in a topological order in order and
4623 // assign a number to each instruction. We use RPO to ensure that defs are
4624 // met before their users. We assume that each instruction that has in-loop
4625 // users starts an interval. We record every time that an in-loop value is
4626 // used, so we have a list of the first and last occurrences of each
4627 // instruction. Next, we transpose this data structure into a multi map that
4628 // holds the list of intervals that *end* at a specific location. This multi
4629 // map allows us to perform a linear search. We scan the instructions linearly
4630 // and record each time that a new interval starts, by placing it in a set.
4631 // If we find this value in the multi-map then we remove it from the set.
4632 // The max register usage is the maximum size of the set.
4633 // We also search for instructions that are defined outside the loop, but are
4634 // used inside the loop. We need this number separately from the max-interval
4635 // usage number because when we unroll, loop-invariant values do not take
4637 LoopBlocksDFS DFS(TheLoop);
4641 R.NumInstructions = 0;
4643 // Each 'key' in the map opens a new interval. The values
4644 // of the map are the index of the 'last seen' usage of the
4645 // instruction that is the key.
4646 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4647 // Maps instruction to its index.
4648 DenseMap<unsigned, Instruction*> IdxToInstr;
4649 // Marks the end of each interval.
4650 IntervalMap EndPoint;
4651 // Saves the list of instruction indices that are used in the loop.
4652 SmallSet<Instruction*, 8> Ends;
4653 // Saves the list of values that are used in the loop but are
4654 // defined outside the loop, such as arguments and constants.
4655 SmallPtrSet<Value*, 8> LoopInvariants;
4658 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4659 be = DFS.endRPO(); bb != be; ++bb) {
4660 R.NumInstructions += (*bb)->size();
4661 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4663 Instruction *I = it;
4664 IdxToInstr[Index++] = I;
4666 // Save the end location of each USE.
4667 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4668 Value *U = I->getOperand(i);
4669 Instruction *Instr = dyn_cast<Instruction>(U);
4671 // Ignore non-instruction values such as arguments, constants, etc.
4672 if (!Instr) continue;
4674 // If this instruction is outside the loop then record it and continue.
4675 if (!TheLoop->contains(Instr)) {
4676 LoopInvariants.insert(Instr);
4680 // Overwrite previous end points.
4681 EndPoint[Instr] = Index;
4687 // Saves the list of intervals that end with the index in 'key'.
4688 typedef SmallVector<Instruction*, 2> InstrList;
4689 DenseMap<unsigned, InstrList> TransposeEnds;
4691 // Transpose the EndPoints to a list of values that end at each index.
4692 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4694 TransposeEnds[it->second].push_back(it->first);
4696 SmallSet<Instruction*, 8> OpenIntervals;
4697 unsigned MaxUsage = 0;
4700 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4701 for (unsigned int i = 0; i < Index; ++i) {
4702 Instruction *I = IdxToInstr[i];
4703 // Ignore instructions that are never used within the loop.
4704 if (!Ends.count(I)) continue;
4706 // Remove all of the instructions that end at this location.
4707 InstrList &List = TransposeEnds[i];
4708 for (unsigned int j=0, e = List.size(); j < e; ++j)
4709 OpenIntervals.erase(List[j]);
4711 // Count the number of live interals.
4712 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4714 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4715 OpenIntervals.size() << '\n');
4717 // Add the current instruction to the list of open intervals.
4718 OpenIntervals.insert(I);
4721 unsigned Invariant = LoopInvariants.size();
4722 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4723 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4724 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4726 R.LoopInvariantRegs = Invariant;
4727 R.MaxLocalUsers = MaxUsage;
4731 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4735 for (Loop::block_iterator bb = TheLoop->block_begin(),
4736 be = TheLoop->block_end(); bb != be; ++bb) {
4737 unsigned BlockCost = 0;
4738 BasicBlock *BB = *bb;
4740 // For each instruction in the old loop.
4741 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4742 // Skip dbg intrinsics.
4743 if (isa<DbgInfoIntrinsic>(it))
4746 unsigned C = getInstructionCost(it, VF);
4748 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4749 VF << " For instruction: " << *it << '\n');
4752 // We assume that if-converted blocks have a 50% chance of being executed.
4753 // When the code is scalar then some of the blocks are avoided due to CF.
4754 // When the code is vectorized we execute all code paths.
4755 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4764 /// \brief Check whether the address computation for a non-consecutive memory
4765 /// access looks like an unlikely candidate for being merged into the indexing
4768 /// We look for a GEP which has one index that is an induction variable and all
4769 /// other indices are loop invariant. If the stride of this access is also
4770 /// within a small bound we decide that this address computation can likely be
4771 /// merged into the addressing mode.
4772 /// In all other cases, we identify the address computation as complex.
4773 static bool isLikelyComplexAddressComputation(Value *Ptr,
4774 LoopVectorizationLegality *Legal,
4775 ScalarEvolution *SE,
4776 const Loop *TheLoop) {
4777 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4781 // We are looking for a gep with all loop invariant indices except for one
4782 // which should be an induction variable.
4783 unsigned NumOperands = Gep->getNumOperands();
4784 for (unsigned i = 1; i < NumOperands; ++i) {
4785 Value *Opd = Gep->getOperand(i);
4786 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4787 !Legal->isInductionVariable(Opd))
4791 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4792 // can likely be merged into the address computation.
4793 unsigned MaxMergeDistance = 64;
4795 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4799 // Check the step is constant.
4800 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4801 // Calculate the pointer stride and check if it is consecutive.
4802 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4806 const APInt &APStepVal = C->getValue()->getValue();
4808 // Huge step value - give up.
4809 if (APStepVal.getBitWidth() > 64)
4812 int64_t StepVal = APStepVal.getSExtValue();
4814 return StepVal > MaxMergeDistance;
4818 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4819 // If we know that this instruction will remain uniform, check the cost of
4820 // the scalar version.
4821 if (Legal->isUniformAfterVectorization(I))
4824 Type *RetTy = I->getType();
4825 Type *VectorTy = ToVectorTy(RetTy, VF);
4827 // TODO: We need to estimate the cost of intrinsic calls.
4828 switch (I->getOpcode()) {
4829 case Instruction::GetElementPtr:
4830 // We mark this instruction as zero-cost because the cost of GEPs in
4831 // vectorized code depends on whether the corresponding memory instruction
4832 // is scalarized or not. Therefore, we handle GEPs with the memory
4833 // instruction cost.
4835 case Instruction::Br: {
4836 return TTI.getCFInstrCost(I->getOpcode());
4838 case Instruction::PHI:
4839 //TODO: IF-converted IFs become selects.
4841 case Instruction::Add:
4842 case Instruction::FAdd:
4843 case Instruction::Sub:
4844 case Instruction::FSub:
4845 case Instruction::Mul:
4846 case Instruction::FMul:
4847 case Instruction::UDiv:
4848 case Instruction::SDiv:
4849 case Instruction::FDiv:
4850 case Instruction::URem:
4851 case Instruction::SRem:
4852 case Instruction::FRem:
4853 case Instruction::Shl:
4854 case Instruction::LShr:
4855 case Instruction::AShr:
4856 case Instruction::And:
4857 case Instruction::Or:
4858 case Instruction::Xor: {
4859 // Certain instructions can be cheaper to vectorize if they have a constant
4860 // second vector operand. One example of this are shifts on x86.
4861 TargetTransformInfo::OperandValueKind Op1VK =
4862 TargetTransformInfo::OK_AnyValue;
4863 TargetTransformInfo::OperandValueKind Op2VK =
4864 TargetTransformInfo::OK_AnyValue;
4866 if (isa<ConstantInt>(I->getOperand(1)))
4867 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4869 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4871 case Instruction::Select: {
4872 SelectInst *SI = cast<SelectInst>(I);
4873 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4874 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4875 Type *CondTy = SI->getCondition()->getType();
4877 CondTy = VectorType::get(CondTy, VF);
4879 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4881 case Instruction::ICmp:
4882 case Instruction::FCmp: {
4883 Type *ValTy = I->getOperand(0)->getType();
4884 VectorTy = ToVectorTy(ValTy, VF);
4885 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4887 case Instruction::Store:
4888 case Instruction::Load: {
4889 StoreInst *SI = dyn_cast<StoreInst>(I);
4890 LoadInst *LI = dyn_cast<LoadInst>(I);
4891 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4893 VectorTy = ToVectorTy(ValTy, VF);
4895 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4896 unsigned AS = SI ? SI->getPointerAddressSpace() :
4897 LI->getPointerAddressSpace();
4898 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4899 // We add the cost of address computation here instead of with the gep
4900 // instruction because only here we know whether the operation is
4903 return TTI.getAddressComputationCost(VectorTy) +
4904 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4906 // Scalarized loads/stores.
4907 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4908 bool Reverse = ConsecutiveStride < 0;
4909 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4910 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4911 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4912 bool IsComplexComputation =
4913 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4915 // The cost of extracting from the value vector and pointer vector.
4916 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4917 for (unsigned i = 0; i < VF; ++i) {
4918 // The cost of extracting the pointer operand.
4919 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4920 // In case of STORE, the cost of ExtractElement from the vector.
4921 // In case of LOAD, the cost of InsertElement into the returned
4923 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4924 Instruction::InsertElement,
4928 // The cost of the scalar loads/stores.
4929 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4930 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4935 // Wide load/stores.
4936 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4937 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4940 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4944 case Instruction::ZExt:
4945 case Instruction::SExt:
4946 case Instruction::FPToUI:
4947 case Instruction::FPToSI:
4948 case Instruction::FPExt:
4949 case Instruction::PtrToInt:
4950 case Instruction::IntToPtr:
4951 case Instruction::SIToFP:
4952 case Instruction::UIToFP:
4953 case Instruction::Trunc:
4954 case Instruction::FPTrunc:
4955 case Instruction::BitCast: {
4956 // We optimize the truncation of induction variable.
4957 // The cost of these is the same as the scalar operation.
4958 if (I->getOpcode() == Instruction::Trunc &&
4959 Legal->isInductionVariable(I->getOperand(0)))
4960 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4961 I->getOperand(0)->getType());
4963 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4964 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4966 case Instruction::Call: {
4967 CallInst *CI = cast<CallInst>(I);
4968 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4969 assert(ID && "Not an intrinsic call!");
4970 Type *RetTy = ToVectorTy(CI->getType(), VF);
4971 SmallVector<Type*, 4> Tys;
4972 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4973 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4974 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4977 // We are scalarizing the instruction. Return the cost of the scalar
4978 // instruction, plus the cost of insert and extract into vector
4979 // elements, times the vector width.
4982 if (!RetTy->isVoidTy() && VF != 1) {
4983 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4985 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4988 // The cost of inserting the results plus extracting each one of the
4990 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4993 // The cost of executing VF copies of the scalar instruction. This opcode
4994 // is unknown. Assume that it is the same as 'mul'.
4995 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5001 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5002 if (Scalar->isVoidTy() || VF == 1)
5004 return VectorType::get(Scalar, VF);
5007 char LoopVectorize::ID = 0;
5008 static const char lv_name[] = "Loop Vectorization";
5009 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5010 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5011 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
5012 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5013 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5014 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5015 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5016 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5019 Pass *createLoopVectorizePass(bool NoUnrolling) {
5020 return new LoopVectorize(NoUnrolling);
5024 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5025 // Check for a store.
5026 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5027 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5029 // Check for a load.
5030 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5031 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5037 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr) {
5038 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5039 // Holds vector parameters or scalars, in case of uniform vals.
5040 SmallVector<VectorParts, 4> Params;
5042 setDebugLocFromInst(Builder, Instr);
5044 // Find all of the vectorized parameters.
5045 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5046 Value *SrcOp = Instr->getOperand(op);
5048 // If we are accessing the old induction variable, use the new one.
5049 if (SrcOp == OldInduction) {
5050 Params.push_back(getVectorValue(SrcOp));
5054 // Try using previously calculated values.
5055 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5057 // If the src is an instruction that appeared earlier in the basic block
5058 // then it should already be vectorized.
5059 if (SrcInst && OrigLoop->contains(SrcInst)) {
5060 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5061 // The parameter is a vector value from earlier.
5062 Params.push_back(WidenMap.get(SrcInst));
5064 // The parameter is a scalar from outside the loop. Maybe even a constant.
5065 VectorParts Scalars;
5066 Scalars.append(UF, SrcOp);
5067 Params.push_back(Scalars);
5071 assert(Params.size() == Instr->getNumOperands() &&
5072 "Invalid number of operands");
5074 // Does this instruction return a value ?
5075 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5077 Value *UndefVec = IsVoidRetTy ? 0 :
5078 UndefValue::get(Instr->getType());
5079 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5080 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5082 // For each vector unroll 'part':
5083 for (unsigned Part = 0; Part < UF; ++Part) {
5084 // For each scalar that we create:
5086 Instruction *Cloned = Instr->clone();
5088 Cloned->setName(Instr->getName() + ".cloned");
5089 // Replace the operands of the cloned instructions with extracted scalars.
5090 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5091 Value *Op = Params[op][Part];
5092 Cloned->setOperand(op, Op);
5095 // Place the cloned scalar in the new loop.
5096 Builder.Insert(Cloned);
5098 // If the original scalar returns a value we need to place it in a vector
5099 // so that future users will be able to use it.
5101 VecResults[Part] = Cloned;
5106 InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr,
5107 LoopVectorizationLegality*) {
5108 return scalarizeInstruction(Instr);
5111 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5115 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5119 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5121 // When unrolling and the VF is 1, we only need to add a simple scalar.
5122 Type *ITy = Val->getType();
5123 assert(!ITy->isVectorTy() && "Val must be a scalar");
5124 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5125 return Builder.CreateAdd(Val, C, "induction");