1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/EquivalenceClasses.h"
51 #include "llvm/ADT/MapVector.h"
52 #include "llvm/ADT/SetVector.h"
53 #include "llvm/ADT/SmallPtrSet.h"
54 #include "llvm/ADT/SmallSet.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/StringExtras.h"
57 #include "llvm/Analysis/AliasAnalysis.h"
58 #include "llvm/Analysis/Dominators.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/Analysis/Verifier.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/LLVMContext.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/IR/Type.h"
78 #include "llvm/IR/Value.h"
79 #include "llvm/Pass.h"
80 #include "llvm/Support/CommandLine.h"
81 #include "llvm/Support/Debug.h"
82 #include "llvm/Support/PatternMatch.h"
83 #include "llvm/Support/raw_ostream.h"
84 #include "llvm/Support/ValueHandle.h"
85 #include "llvm/Target/TargetLibraryInfo.h"
86 #include "llvm/Transforms/Scalar.h"
87 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
88 #include "llvm/Transforms/Utils/Local.h"
93 using namespace llvm::PatternMatch;
95 static cl::opt<unsigned>
96 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
97 cl::desc("Sets the SIMD width. Zero is autoselect."));
99 static cl::opt<unsigned>
100 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
101 cl::desc("Sets the vectorization unroll count. "
102 "Zero is autoselect."));
105 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
106 cl::desc("Enable if-conversion during vectorization."));
108 /// We don't vectorize loops with a known constant trip count below this number.
109 static cl::opt<unsigned>
110 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
112 cl::desc("Don't vectorize loops with a constant "
113 "trip count that is smaller than this "
116 /// We don't unroll loops with a known constant trip count below this number.
117 static const unsigned TinyTripCountUnrollThreshold = 128;
119 /// When performing memory disambiguation checks at runtime do not make more
120 /// than this number of comparisons.
121 static const unsigned RuntimeMemoryCheckThreshold = 8;
123 /// Maximum simd width.
124 static const unsigned MaxVectorWidth = 64;
126 /// Maximum vectorization unroll count.
127 static const unsigned MaxUnrollFactor = 16;
131 // Forward declarations.
132 class LoopVectorizationLegality;
133 class LoopVectorizationCostModel;
135 /// InnerLoopVectorizer vectorizes loops which contain only one basic
136 /// block to a specified vectorization factor (VF).
137 /// This class performs the widening of scalars into vectors, or multiple
138 /// scalars. This class also implements the following features:
139 /// * It inserts an epilogue loop for handling loops that don't have iteration
140 /// counts that are known to be a multiple of the vectorization factor.
141 /// * It handles the code generation for reduction variables.
142 /// * Scalarization (implementation using scalars) of un-vectorizable
144 /// InnerLoopVectorizer does not perform any vectorization-legality
145 /// checks, and relies on the caller to check for the different legality
146 /// aspects. The InnerLoopVectorizer relies on the
147 /// LoopVectorizationLegality class to provide information about the induction
148 /// and reduction variables that were found to a given vectorization factor.
149 class InnerLoopVectorizer {
151 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
152 DominatorTree *DT, DataLayout *DL,
153 const TargetLibraryInfo *TLI, unsigned VecWidth,
154 unsigned UnrollFactor)
155 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
156 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
157 OldInduction(0), WidenMap(UnrollFactor) {}
159 // Perform the actual loop widening (vectorization).
160 void vectorize(LoopVectorizationLegality *Legal) {
161 // Create a new empty loop. Unlink the old loop and connect the new one.
162 createEmptyLoop(Legal);
163 // Widen each instruction in the old loop to a new one in the new loop.
164 // Use the Legality module to find the induction and reduction variables.
165 vectorizeLoop(Legal);
166 // Register the new loop and update the analysis passes.
171 /// A small list of PHINodes.
172 typedef SmallVector<PHINode*, 4> PhiVector;
173 /// When we unroll loops we have multiple vector values for each scalar.
174 /// This data structure holds the unrolled and vectorized values that
175 /// originated from one scalar instruction.
176 typedef SmallVector<Value*, 2> VectorParts;
178 // When we if-convert we need create edge masks. We have to cache values so
179 // that we don't end up with exponential recursion/IR.
180 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
181 VectorParts> EdgeMaskCache;
183 /// Add code that checks at runtime if the accessed arrays overlap.
184 /// Returns the comparator value or NULL if no check is needed.
185 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
187 /// Create an empty loop, based on the loop ranges of the old loop.
188 void createEmptyLoop(LoopVectorizationLegality *Legal);
189 /// Copy and widen the instructions from the old loop.
190 void vectorizeLoop(LoopVectorizationLegality *Legal);
192 /// A helper function that computes the predicate of the block BB, assuming
193 /// that the header block of the loop is set to True. It returns the *entry*
194 /// mask for the block BB.
195 VectorParts createBlockInMask(BasicBlock *BB);
196 /// A helper function that computes the predicate of the edge between SRC
198 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
200 /// A helper function to vectorize a single BB within the innermost loop.
201 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
204 /// Insert the new loop to the loop hierarchy and pass manager
205 /// and update the analysis passes.
206 void updateAnalysis();
208 /// This instruction is un-vectorizable. Implement it as a sequence
210 void scalarizeInstruction(Instruction *Instr);
212 /// Vectorize Load and Store instructions,
213 void vectorizeMemoryInstruction(Instruction *Instr,
214 LoopVectorizationLegality *Legal);
216 /// Create a broadcast instruction. This method generates a broadcast
217 /// instruction (shuffle) for loop invariant values and for the induction
218 /// value. If this is the induction variable then we extend it to N, N+1, ...
219 /// this is needed because each iteration in the loop corresponds to a SIMD
221 Value *getBroadcastInstrs(Value *V);
223 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
224 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
225 /// The sequence starts at StartIndex.
226 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
228 /// When we go over instructions in the basic block we rely on previous
229 /// values within the current basic block or on loop invariant values.
230 /// When we widen (vectorize) values we place them in the map. If the values
231 /// are not within the map, they have to be loop invariant, so we simply
232 /// broadcast them into a vector.
233 VectorParts &getVectorValue(Value *V);
235 /// Generate a shuffle sequence that will reverse the vector Vec.
236 Value *reverseVector(Value *Vec);
238 /// This is a helper class that holds the vectorizer state. It maps scalar
239 /// instructions to vector instructions. When the code is 'unrolled' then
240 /// then a single scalar value is mapped to multiple vector parts. The parts
241 /// are stored in the VectorPart type.
243 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
245 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
247 /// \return True if 'Key' is saved in the Value Map.
248 bool has(Value *Key) const { return MapStorage.count(Key); }
250 /// Initializes a new entry in the map. Sets all of the vector parts to the
251 /// save value in 'Val'.
252 /// \return A reference to a vector with splat values.
253 VectorParts &splat(Value *Key, Value *Val) {
254 VectorParts &Entry = MapStorage[Key];
255 Entry.assign(UF, Val);
259 ///\return A reference to the value that is stored at 'Key'.
260 VectorParts &get(Value *Key) {
261 VectorParts &Entry = MapStorage[Key];
264 assert(Entry.size() == UF);
269 /// The unroll factor. Each entry in the map stores this number of vector
273 /// Map storage. We use std::map and not DenseMap because insertions to a
274 /// dense map invalidates its iterators.
275 std::map<Value *, VectorParts> MapStorage;
278 /// The original loop.
280 /// Scev analysis to use.
288 /// Target Library Info.
289 const TargetLibraryInfo *TLI;
291 /// The vectorization SIMD factor to use. Each vector will have this many
294 /// The vectorization unroll factor to use. Each scalar is vectorized to this
295 /// many different vector instructions.
298 /// The builder that we use
301 // --- Vectorization state ---
303 /// The vector-loop preheader.
304 BasicBlock *LoopVectorPreHeader;
305 /// The scalar-loop preheader.
306 BasicBlock *LoopScalarPreHeader;
307 /// Middle Block between the vector and the scalar.
308 BasicBlock *LoopMiddleBlock;
309 ///The ExitBlock of the scalar loop.
310 BasicBlock *LoopExitBlock;
311 ///The vector loop body.
312 BasicBlock *LoopVectorBody;
313 ///The scalar loop body.
314 BasicBlock *LoopScalarBody;
315 /// A list of all bypass blocks. The first block is the entry of the loop.
316 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
318 /// The new Induction variable which was added to the new block.
320 /// The induction variable of the old basic block.
321 PHINode *OldInduction;
322 /// Holds the extended (to the widest induction type) start index.
324 /// Maps scalars to widened vectors.
326 EdgeMaskCache MaskCache;
329 /// \brief Look for a meaningful debug location on the instruction or it's
331 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
336 if (I->getDebugLoc() != Empty)
339 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
340 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
341 if (OpInst->getDebugLoc() != Empty)
348 /// \brief Set the debug location in the builder using the debug location in the
350 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
351 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
352 B.SetCurrentDebugLocation(Inst->getDebugLoc());
354 B.SetCurrentDebugLocation(DebugLoc());
357 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
358 /// to what vectorization factor.
359 /// This class does not look at the profitability of vectorization, only the
360 /// legality. This class has two main kinds of checks:
361 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
362 /// will change the order of memory accesses in a way that will change the
363 /// correctness of the program.
364 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
365 /// checks for a number of different conditions, such as the availability of a
366 /// single induction variable, that all types are supported and vectorize-able,
367 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
368 /// This class is also used by InnerLoopVectorizer for identifying
369 /// induction variable and the different reduction variables.
370 class LoopVectorizationLegality {
372 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
373 DominatorTree *DT, TargetLibraryInfo *TLI)
374 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
375 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
376 MaxSafeDepDistBytes(-1U) {}
378 /// This enum represents the kinds of reductions that we support.
380 RK_NoReduction, ///< Not a reduction.
381 RK_IntegerAdd, ///< Sum of integers.
382 RK_IntegerMult, ///< Product of integers.
383 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
384 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
385 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
386 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
387 RK_FloatAdd, ///< Sum of floats.
388 RK_FloatMult, ///< Product of floats.
389 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
392 /// This enum represents the kinds of inductions that we support.
394 IK_NoInduction, ///< Not an induction variable.
395 IK_IntInduction, ///< Integer induction variable. Step = 1.
396 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
397 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
398 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
401 // This enum represents the kind of minmax reduction.
402 enum MinMaxReductionKind {
412 /// This POD struct holds information about reduction variables.
413 struct ReductionDescriptor {
414 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
415 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
417 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
418 MinMaxReductionKind MK)
419 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
421 // The starting value of the reduction.
422 // It does not have to be zero!
423 TrackingVH<Value> StartValue;
424 // The instruction who's value is used outside the loop.
425 Instruction *LoopExitInstr;
426 // The kind of the reduction.
428 // If this a min/max reduction the kind of reduction.
429 MinMaxReductionKind MinMaxKind;
432 /// This POD struct holds information about a potential reduction operation.
433 struct ReductionInstDesc {
434 ReductionInstDesc(bool IsRedux, Instruction *I) :
435 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
437 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
438 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
440 // Is this instruction a reduction candidate.
442 // The last instruction in a min/max pattern (select of the select(icmp())
443 // pattern), or the current reduction instruction otherwise.
444 Instruction *PatternLastInst;
445 // If this is a min/max pattern the comparison predicate.
446 MinMaxReductionKind MinMaxKind;
449 // This POD struct holds information about the memory runtime legality
450 // check that a group of pointers do not overlap.
451 struct RuntimePointerCheck {
452 RuntimePointerCheck() : Need(false) {}
454 /// Reset the state of the pointer runtime information.
462 /// Insert a pointer and calculate the start and end SCEVs.
463 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
466 /// This flag indicates if we need to add the runtime check.
468 /// Holds the pointers that we need to check.
469 SmallVector<TrackingVH<Value>, 2> Pointers;
470 /// Holds the pointer value at the beginning of the loop.
471 SmallVector<const SCEV*, 2> Starts;
472 /// Holds the pointer value at the end of the loop.
473 SmallVector<const SCEV*, 2> Ends;
474 /// Holds the information if this pointer is used for writing to memory.
475 SmallVector<bool, 2> IsWritePtr;
476 /// Holds the id of the set of pointers that could be dependent because of a
477 /// shared underlying object.
478 SmallVector<unsigned, 2> DependencySetId;
481 /// A POD for saving information about induction variables.
482 struct InductionInfo {
483 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
484 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
486 TrackingVH<Value> StartValue;
491 /// ReductionList contains the reduction descriptors for all
492 /// of the reductions that were found in the loop.
493 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
495 /// InductionList saves induction variables and maps them to the
496 /// induction descriptor.
497 typedef MapVector<PHINode*, InductionInfo> InductionList;
499 /// Returns true if it is legal to vectorize this loop.
500 /// This does not mean that it is profitable to vectorize this
501 /// loop, only that it is legal to do so.
504 /// Returns the Induction variable.
505 PHINode *getInduction() { return Induction; }
507 /// Returns the reduction variables found in the loop.
508 ReductionList *getReductionVars() { return &Reductions; }
510 /// Returns the induction variables found in the loop.
511 InductionList *getInductionVars() { return &Inductions; }
513 /// Returns the widest induction type.
514 Type *getWidestInductionType() { return WidestIndTy; }
516 /// Returns True if V is an induction variable in this loop.
517 bool isInductionVariable(const Value *V);
519 /// Return true if the block BB needs to be predicated in order for the loop
520 /// to be vectorized.
521 bool blockNeedsPredication(BasicBlock *BB);
523 /// Check if this pointer is consecutive when vectorizing. This happens
524 /// when the last index of the GEP is the induction variable, or that the
525 /// pointer itself is an induction variable.
526 /// This check allows us to vectorize A[idx] into a wide load/store.
528 /// 0 - Stride is unknown or non consecutive.
529 /// 1 - Address is consecutive.
530 /// -1 - Address is consecutive, and decreasing.
531 int isConsecutivePtr(Value *Ptr);
533 /// Returns true if the value V is uniform within the loop.
534 bool isUniform(Value *V);
536 /// Returns true if this instruction will remain scalar after vectorization.
537 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
539 /// Returns the information that we collected about runtime memory check.
540 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
542 /// This function returns the identity element (or neutral element) for
544 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
546 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
549 /// Check if a single basic block loop is vectorizable.
550 /// At this point we know that this is a loop with a constant trip count
551 /// and we only need to check individual instructions.
552 bool canVectorizeInstrs();
554 /// When we vectorize loops we may change the order in which
555 /// we read and write from memory. This method checks if it is
556 /// legal to vectorize the code, considering only memory constrains.
557 /// Returns true if the loop is vectorizable
558 bool canVectorizeMemory();
560 /// Return true if we can vectorize this loop using the IF-conversion
562 bool canVectorizeWithIfConvert();
564 /// Collect the variables that need to stay uniform after vectorization.
565 void collectLoopUniforms();
567 /// Return true if all of the instructions in the block can be speculatively
568 /// executed. \p SafePtrs is a list of addresses that are known to be legal
569 /// and we know that we can read from them without segfault.
570 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
572 /// Returns True, if 'Phi' is the kind of reduction variable for type
573 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
574 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
575 /// Returns a struct describing if the instruction 'I' can be a reduction
576 /// variable of type 'Kind'. If the reduction is a min/max pattern of
577 /// select(icmp()) this function advances the instruction pointer 'I' from the
578 /// compare instruction to the select instruction and stores this pointer in
579 /// 'PatternLastInst' member of the returned struct.
580 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
581 ReductionInstDesc &Desc);
582 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
583 /// pattern corresponding to a min(X, Y) or max(X, Y).
584 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
585 ReductionInstDesc &Prev);
586 /// Returns the induction kind of Phi. This function may return NoInduction
587 /// if the PHI is not an induction variable.
588 InductionKind isInductionVariable(PHINode *Phi);
590 /// The loop that we evaluate.
594 /// DataLayout analysis.
598 /// Target Library Info.
599 TargetLibraryInfo *TLI;
601 // --- vectorization state --- //
603 /// Holds the integer induction variable. This is the counter of the
606 /// Holds the reduction variables.
607 ReductionList Reductions;
608 /// Holds all of the induction variables that we found in the loop.
609 /// Notice that inductions don't need to start at zero and that induction
610 /// variables can be pointers.
611 InductionList Inductions;
612 /// Holds the widest induction type encountered.
615 /// Allowed outside users. This holds the reduction
616 /// vars which can be accessed from outside the loop.
617 SmallPtrSet<Value*, 4> AllowedExit;
618 /// This set holds the variables which are known to be uniform after
620 SmallPtrSet<Instruction*, 4> Uniforms;
621 /// We need to check that all of the pointers in this list are disjoint
623 RuntimePointerCheck PtrRtCheck;
624 /// Can we assume the absence of NaNs.
625 bool HasFunNoNaNAttr;
627 unsigned MaxSafeDepDistBytes;
630 /// LoopVectorizationCostModel - estimates the expected speedups due to
632 /// In many cases vectorization is not profitable. This can happen because of
633 /// a number of reasons. In this class we mainly attempt to predict the
634 /// expected speedup/slowdowns due to the supported instruction set. We use the
635 /// TargetTransformInfo to query the different backends for the cost of
636 /// different operations.
637 class LoopVectorizationCostModel {
639 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
640 LoopVectorizationLegality *Legal,
641 const TargetTransformInfo &TTI,
642 DataLayout *DL, const TargetLibraryInfo *TLI)
643 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
645 /// Information about vectorization costs
646 struct VectorizationFactor {
647 unsigned Width; // Vector width with best cost
648 unsigned Cost; // Cost of the loop with that width
650 /// \return The most profitable vectorization factor and the cost of that VF.
651 /// This method checks every power of two up to VF. If UserVF is not ZERO
652 /// then this vectorization factor will be selected if vectorization is
654 VectorizationFactor selectVectorizationFactor(bool OptForSize,
657 /// \return The size (in bits) of the widest type in the code that
658 /// needs to be vectorized. We ignore values that remain scalar such as
659 /// 64 bit loop indices.
660 unsigned getWidestType();
662 /// \return The most profitable unroll factor.
663 /// If UserUF is non-zero then this method finds the best unroll-factor
664 /// based on register pressure and other parameters.
665 /// VF and LoopCost are the selected vectorization factor and the cost of the
667 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
670 /// \brief A struct that represents some properties of the register usage
672 struct RegisterUsage {
673 /// Holds the number of loop invariant values that are used in the loop.
674 unsigned LoopInvariantRegs;
675 /// Holds the maximum number of concurrent live intervals in the loop.
676 unsigned MaxLocalUsers;
677 /// Holds the number of instructions in the loop.
678 unsigned NumInstructions;
681 /// \return information about the register usage of the loop.
682 RegisterUsage calculateRegisterUsage();
685 /// Returns the expected execution cost. The unit of the cost does
686 /// not matter because we use the 'cost' units to compare different
687 /// vector widths. The cost that is returned is *not* normalized by
688 /// the factor width.
689 unsigned expectedCost(unsigned VF);
691 /// Returns the execution time cost of an instruction for a given vector
692 /// width. Vector width of one means scalar.
693 unsigned getInstructionCost(Instruction *I, unsigned VF);
695 /// A helper function for converting Scalar types to vector types.
696 /// If the incoming type is void, we return void. If the VF is 1, we return
698 static Type* ToVectorTy(Type *Scalar, unsigned VF);
700 /// Returns whether the instruction is a load or store and will be a emitted
701 /// as a vector operation.
702 bool isConsecutiveLoadOrStore(Instruction *I);
704 /// The loop that we evaluate.
708 /// Loop Info analysis.
710 /// Vectorization legality.
711 LoopVectorizationLegality *Legal;
712 /// Vector target information.
713 const TargetTransformInfo &TTI;
714 /// Target data layout information.
716 /// Target Library Info.
717 const TargetLibraryInfo *TLI;
720 /// Utility class for getting and setting loop vectorizer hints in the form
721 /// of loop metadata.
722 struct LoopVectorizeHints {
723 /// Vectorization width.
725 /// Vectorization unroll factor.
728 LoopVectorizeHints(const Loop *L)
729 : Width(VectorizationFactor)
730 , Unroll(VectorizationUnroll)
731 , LoopID(L->getLoopID()) {
733 // The command line options override any loop metadata except for when
734 // width == 1 which is used to indicate the loop is already vectorized.
735 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
736 Width = VectorizationFactor;
737 if (VectorizationUnroll.getNumOccurrences() > 0)
738 Unroll = VectorizationUnroll;
741 /// Return the loop vectorizer metadata prefix.
742 static StringRef Prefix() { return "llvm.vectorizer."; }
744 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
745 SmallVector<Value*, 2> Vals;
746 Vals.push_back(MDString::get(Context, Name));
747 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
748 return MDNode::get(Context, Vals);
751 /// Mark the loop L as already vectorized by setting the width to 1.
752 void setAlreadyVectorized(Loop *L) {
753 LLVMContext &Context = L->getHeader()->getContext();
757 // Create a new loop id with one more operand for the already_vectorized
758 // hint. If the loop already has a loop id then copy the existing operands.
759 SmallVector<Value*, 4> Vals(1);
761 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
762 Vals.push_back(LoopID->getOperand(i));
764 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
766 MDNode *NewLoopID = MDNode::get(Context, Vals);
767 // Set operand 0 to refer to the loop id itself.
768 NewLoopID->replaceOperandWith(0, NewLoopID);
770 L->setLoopID(NewLoopID);
772 LoopID->replaceAllUsesWith(NewLoopID);
780 /// Find hints specified in the loop metadata.
781 void getHints(const Loop *L) {
785 // First operand should refer to the loop id itself.
786 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
787 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
789 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
790 const MDString *S = 0;
791 SmallVector<Value*, 4> Args;
793 // The expected hint is either a MDString or a MDNode with the first
794 // operand a MDString.
795 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
796 if (!MD || MD->getNumOperands() == 0)
798 S = dyn_cast<MDString>(MD->getOperand(0));
799 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
800 Args.push_back(MD->getOperand(i));
802 S = dyn_cast<MDString>(LoopID->getOperand(i));
803 assert(Args.size() == 0 && "too many arguments for MDString");
809 // Check if the hint starts with the vectorizer prefix.
810 StringRef Hint = S->getString();
811 if (!Hint.startswith(Prefix()))
813 // Remove the prefix.
814 Hint = Hint.substr(Prefix().size(), StringRef::npos);
816 if (Args.size() == 1)
817 getHint(Hint, Args[0]);
821 // Check string hint with one operand.
822 void getHint(StringRef Hint, Value *Arg) {
823 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
825 unsigned Val = C->getZExtValue();
827 if (Hint == "width") {
828 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
829 "Invalid width metadata");
831 } else if (Hint == "unroll") {
832 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
833 "Invalid unroll metadata");
836 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
840 /// The LoopVectorize Pass.
841 struct LoopVectorize : public LoopPass {
842 /// Pass identification, replacement for typeid
845 explicit LoopVectorize() : LoopPass(ID) {
846 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
852 TargetTransformInfo *TTI;
854 TargetLibraryInfo *TLI;
856 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
857 // We only vectorize innermost loops.
861 SE = &getAnalysis<ScalarEvolution>();
862 DL = getAnalysisIfAvailable<DataLayout>();
863 LI = &getAnalysis<LoopInfo>();
864 TTI = &getAnalysis<TargetTransformInfo>();
865 DT = &getAnalysis<DominatorTree>();
866 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
869 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
873 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
874 L->getHeader()->getParent()->getName() << "\"\n");
876 LoopVectorizeHints Hints(L);
878 if (Hints.Width == 1) {
879 DEBUG(dbgs() << "LV: Not vectorizing.\n");
883 // Check if it is legal to vectorize the loop.
884 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
885 if (!LVL.canVectorize()) {
886 DEBUG(dbgs() << "LV: Not vectorizing.\n");
890 // Use the cost model.
891 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
893 // Check the function attributes to find out if this function should be
894 // optimized for size.
895 Function *F = L->getHeader()->getParent();
896 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
897 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
898 unsigned FnIndex = AttributeSet::FunctionIndex;
899 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
900 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
903 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
904 "attribute is used.\n");
908 // Select the optimal vectorization factor.
909 LoopVectorizationCostModel::VectorizationFactor VF;
910 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
911 // Select the unroll factor.
912 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
916 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
920 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
921 F->getParent()->getModuleIdentifier()<<"\n");
922 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
924 // If we decided that it is *legal* to vectorize the loop then do it.
925 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
928 // Mark the loop as already vectorized to avoid vectorizing again.
929 Hints.setAlreadyVectorized(L);
931 DEBUG(verifyFunction(*L->getHeader()->getParent()));
935 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
936 LoopPass::getAnalysisUsage(AU);
937 AU.addRequiredID(LoopSimplifyID);
938 AU.addRequiredID(LCSSAID);
939 AU.addRequired<DominatorTree>();
940 AU.addRequired<LoopInfo>();
941 AU.addRequired<ScalarEvolution>();
942 AU.addRequired<TargetTransformInfo>();
943 AU.addPreserved<LoopInfo>();
944 AU.addPreserved<DominatorTree>();
949 } // end anonymous namespace
951 //===----------------------------------------------------------------------===//
952 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
953 // LoopVectorizationCostModel.
954 //===----------------------------------------------------------------------===//
957 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
958 Loop *Lp, Value *Ptr,
961 const SCEV *Sc = SE->getSCEV(Ptr);
962 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
963 assert(AR && "Invalid addrec expression");
964 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
965 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
966 Pointers.push_back(Ptr);
967 Starts.push_back(AR->getStart());
968 Ends.push_back(ScEnd);
969 IsWritePtr.push_back(WritePtr);
970 DependencySetId.push_back(DepSetId);
973 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
974 // Save the current insertion location.
975 Instruction *Loc = Builder.GetInsertPoint();
977 // We need to place the broadcast of invariant variables outside the loop.
978 Instruction *Instr = dyn_cast<Instruction>(V);
979 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
980 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
982 // Place the code for broadcasting invariant variables in the new preheader.
984 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
986 // Broadcast the scalar into all locations in the vector.
987 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
989 // Restore the builder insertion point.
991 Builder.SetInsertPoint(Loc);
996 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
998 assert(Val->getType()->isVectorTy() && "Must be a vector");
999 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1000 "Elem must be an integer");
1001 // Create the types.
1002 Type *ITy = Val->getType()->getScalarType();
1003 VectorType *Ty = cast<VectorType>(Val->getType());
1004 int VLen = Ty->getNumElements();
1005 SmallVector<Constant*, 8> Indices;
1007 // Create a vector of consecutive numbers from zero to VF.
1008 for (int i = 0; i < VLen; ++i) {
1009 int64_t Idx = Negate ? (-i) : i;
1010 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1013 // Add the consecutive indices to the vector value.
1014 Constant *Cv = ConstantVector::get(Indices);
1015 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1016 return Builder.CreateAdd(Val, Cv, "induction");
1019 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1020 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1021 // Make sure that the pointer does not point to structs.
1022 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1025 // If this value is a pointer induction variable we know it is consecutive.
1026 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1027 if (Phi && Inductions.count(Phi)) {
1028 InductionInfo II = Inductions[Phi];
1029 if (IK_PtrInduction == II.IK)
1031 else if (IK_ReversePtrInduction == II.IK)
1035 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1039 unsigned NumOperands = Gep->getNumOperands();
1040 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1042 Value *GpPtr = Gep->getPointerOperand();
1043 // If this GEP value is a consecutive pointer induction variable and all of
1044 // the indices are constant then we know it is consecutive. We can
1045 Phi = dyn_cast<PHINode>(GpPtr);
1046 if (Phi && Inductions.count(Phi)) {
1048 // Make sure that the pointer does not point to structs.
1049 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1050 if (GepPtrType->getElementType()->isAggregateType())
1053 // Make sure that all of the index operands are loop invariant.
1054 for (unsigned i = 1; i < NumOperands; ++i)
1055 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1058 InductionInfo II = Inductions[Phi];
1059 if (IK_PtrInduction == II.IK)
1061 else if (IK_ReversePtrInduction == II.IK)
1065 // Check that all of the gep indices are uniform except for the last.
1066 for (unsigned i = 0; i < NumOperands - 1; ++i)
1067 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1070 // We can emit wide load/stores only if the last index is the induction
1072 const SCEV *Last = SE->getSCEV(LastIndex);
1073 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1074 const SCEV *Step = AR->getStepRecurrence(*SE);
1076 // The memory is consecutive because the last index is consecutive
1077 // and all other indices are loop invariant.
1080 if (Step->isAllOnesValue())
1087 bool LoopVectorizationLegality::isUniform(Value *V) {
1088 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1091 InnerLoopVectorizer::VectorParts&
1092 InnerLoopVectorizer::getVectorValue(Value *V) {
1093 assert(V != Induction && "The new induction variable should not be used.");
1094 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1096 // If we have this scalar in the map, return it.
1097 if (WidenMap.has(V))
1098 return WidenMap.get(V);
1100 // If this scalar is unknown, assume that it is a constant or that it is
1101 // loop invariant. Broadcast V and save the value for future uses.
1102 Value *B = getBroadcastInstrs(V);
1103 return WidenMap.splat(V, B);
1106 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1107 assert(Vec->getType()->isVectorTy() && "Invalid type");
1108 SmallVector<Constant*, 8> ShuffleMask;
1109 for (unsigned i = 0; i < VF; ++i)
1110 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1112 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1113 ConstantVector::get(ShuffleMask),
1118 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1119 LoopVectorizationLegality *Legal) {
1120 // Attempt to issue a wide load.
1121 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1122 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1124 assert((LI || SI) && "Invalid Load/Store instruction");
1126 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1127 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1128 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1129 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1130 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1131 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1132 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1134 if (ScalarAllocatedSize != VectorElementSize)
1135 return scalarizeInstruction(Instr);
1137 // If the pointer is loop invariant or if it is non consecutive,
1138 // scalarize the load.
1139 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1140 bool Reverse = ConsecutiveStride < 0;
1141 bool UniformLoad = LI && Legal->isUniform(Ptr);
1142 if (!ConsecutiveStride || UniformLoad)
1143 return scalarizeInstruction(Instr);
1145 Constant *Zero = Builder.getInt32(0);
1146 VectorParts &Entry = WidenMap.get(Instr);
1148 // Handle consecutive loads/stores.
1149 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1150 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1151 setDebugLocFromInst(Builder, Gep);
1152 Value *PtrOperand = Gep->getPointerOperand();
1153 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1154 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1156 // Create the new GEP with the new induction variable.
1157 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1158 Gep2->setOperand(0, FirstBasePtr);
1159 Gep2->setName("gep.indvar.base");
1160 Ptr = Builder.Insert(Gep2);
1162 setDebugLocFromInst(Builder, Gep);
1163 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1164 OrigLoop) && "Base ptr must be invariant");
1166 // The last index does not have to be the induction. It can be
1167 // consecutive and be a function of the index. For example A[I+1];
1168 unsigned NumOperands = Gep->getNumOperands();
1169 unsigned LastOperand = NumOperands - 1;
1170 // Create the new GEP with the new induction variable.
1171 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1173 for (unsigned i = 0; i < NumOperands; ++i) {
1174 Value *GepOperand = Gep->getOperand(i);
1175 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1177 // Update last index or loop invariant instruction anchored in loop.
1178 if (i == LastOperand ||
1179 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1180 assert((i == LastOperand ||
1181 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1182 "Must be last index or loop invariant");
1184 VectorParts &GEPParts = getVectorValue(GepOperand);
1185 Value *Index = GEPParts[0];
1186 Index = Builder.CreateExtractElement(Index, Zero);
1187 Gep2->setOperand(i, Index);
1188 Gep2->setName("gep.indvar.idx");
1191 Ptr = Builder.Insert(Gep2);
1193 // Use the induction element ptr.
1194 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1195 setDebugLocFromInst(Builder, Ptr);
1196 VectorParts &PtrVal = getVectorValue(Ptr);
1197 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1202 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1203 "We do not allow storing to uniform addresses");
1204 setDebugLocFromInst(Builder, SI);
1205 // We don't want to update the value in the map as it might be used in
1206 // another expression. So don't use a reference type for "StoredVal".
1207 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1209 for (unsigned Part = 0; Part < UF; ++Part) {
1210 // Calculate the pointer for the specific unroll-part.
1211 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1214 // If we store to reverse consecutive memory locations then we need
1215 // to reverse the order of elements in the stored value.
1216 StoredVal[Part] = reverseVector(StoredVal[Part]);
1217 // If the address is consecutive but reversed, then the
1218 // wide store needs to start at the last vector element.
1219 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1220 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1223 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1224 DataTy->getPointerTo(AddressSpace));
1225 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1231 assert(LI && "Must have a load instruction");
1232 setDebugLocFromInst(Builder, LI);
1233 for (unsigned Part = 0; Part < UF; ++Part) {
1234 // Calculate the pointer for the specific unroll-part.
1235 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1238 // If the address is consecutive but reversed, then the
1239 // wide store needs to start at the last vector element.
1240 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1241 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1244 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1245 DataTy->getPointerTo(AddressSpace));
1246 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1247 cast<LoadInst>(LI)->setAlignment(Alignment);
1248 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1252 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1253 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1254 // Holds vector parameters or scalars, in case of uniform vals.
1255 SmallVector<VectorParts, 4> Params;
1257 setDebugLocFromInst(Builder, Instr);
1259 // Find all of the vectorized parameters.
1260 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1261 Value *SrcOp = Instr->getOperand(op);
1263 // If we are accessing the old induction variable, use the new one.
1264 if (SrcOp == OldInduction) {
1265 Params.push_back(getVectorValue(SrcOp));
1269 // Try using previously calculated values.
1270 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1272 // If the src is an instruction that appeared earlier in the basic block
1273 // then it should already be vectorized.
1274 if (SrcInst && OrigLoop->contains(SrcInst)) {
1275 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1276 // The parameter is a vector value from earlier.
1277 Params.push_back(WidenMap.get(SrcInst));
1279 // The parameter is a scalar from outside the loop. Maybe even a constant.
1280 VectorParts Scalars;
1281 Scalars.append(UF, SrcOp);
1282 Params.push_back(Scalars);
1286 assert(Params.size() == Instr->getNumOperands() &&
1287 "Invalid number of operands");
1289 // Does this instruction return a value ?
1290 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1292 Value *UndefVec = IsVoidRetTy ? 0 :
1293 UndefValue::get(VectorType::get(Instr->getType(), VF));
1294 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1295 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1297 // For each vector unroll 'part':
1298 for (unsigned Part = 0; Part < UF; ++Part) {
1299 // For each scalar that we create:
1300 for (unsigned Width = 0; Width < VF; ++Width) {
1301 Instruction *Cloned = Instr->clone();
1303 Cloned->setName(Instr->getName() + ".cloned");
1304 // Replace the operands of the cloned instrucions with extracted scalars.
1305 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1306 Value *Op = Params[op][Part];
1307 // Param is a vector. Need to extract the right lane.
1308 if (Op->getType()->isVectorTy())
1309 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1310 Cloned->setOperand(op, Op);
1313 // Place the cloned scalar in the new loop.
1314 Builder.Insert(Cloned);
1316 // If the original scalar returns a value we need to place it in a vector
1317 // so that future users will be able to use it.
1319 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1320 Builder.getInt32(Width));
1326 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1328 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1329 Legal->getRuntimePointerCheck();
1331 if (!PtrRtCheck->Need)
1334 unsigned NumPointers = PtrRtCheck->Pointers.size();
1335 SmallVector<TrackingVH<Value> , 2> Starts;
1336 SmallVector<TrackingVH<Value> , 2> Ends;
1338 SCEVExpander Exp(*SE, "induction");
1340 // Use this type for pointer arithmetic.
1341 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1343 for (unsigned i = 0; i < NumPointers; ++i) {
1344 Value *Ptr = PtrRtCheck->Pointers[i];
1345 const SCEV *Sc = SE->getSCEV(Ptr);
1347 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1348 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1350 Starts.push_back(Ptr);
1351 Ends.push_back(Ptr);
1353 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1355 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1356 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1357 Starts.push_back(Start);
1358 Ends.push_back(End);
1362 IRBuilder<> ChkBuilder(Loc);
1363 // Our instructions might fold to a constant.
1364 Value *MemoryRuntimeCheck = 0;
1365 for (unsigned i = 0; i < NumPointers; ++i) {
1366 for (unsigned j = i+1; j < NumPointers; ++j) {
1367 // No need to check if two readonly pointers intersect.
1368 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1371 // Only need to check pointers between two different dependency sets.
1372 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1375 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1376 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1377 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1378 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1380 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1381 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1382 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1383 if (MemoryRuntimeCheck)
1384 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1386 MemoryRuntimeCheck = IsConflict;
1390 // We have to do this trickery because the IRBuilder might fold the check to a
1391 // constant expression in which case there is no Instruction anchored in a
1393 LLVMContext &Ctx = Loc->getContext();
1394 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1395 ConstantInt::getTrue(Ctx));
1396 ChkBuilder.Insert(Check, "memcheck.conflict");
1401 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1403 In this function we generate a new loop. The new loop will contain
1404 the vectorized instructions while the old loop will continue to run the
1407 [ ] <-- vector loop bypass (may consist of multiple blocks).
1410 | [ ] <-- vector pre header.
1414 | [ ]_| <-- vector loop.
1417 >[ ] <--- middle-block.
1420 | [ ] <--- new preheader.
1424 | [ ]_| <-- old scalar loop to handle remainder.
1427 >[ ] <-- exit block.
1431 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1432 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1433 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1434 assert(ExitBlock && "Must have an exit block");
1436 // Some loops have a single integer induction variable, while other loops
1437 // don't. One example is c++ iterators that often have multiple pointer
1438 // induction variables. In the code below we also support a case where we
1439 // don't have a single induction variable.
1440 OldInduction = Legal->getInduction();
1441 Type *IdxTy = Legal->getWidestInductionType();
1443 // Find the loop boundaries.
1444 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1445 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1447 // Get the total trip count from the count by adding 1.
1448 ExitCount = SE->getAddExpr(ExitCount,
1449 SE->getConstant(ExitCount->getType(), 1));
1451 // Expand the trip count and place the new instructions in the preheader.
1452 // Notice that the pre-header does not change, only the loop body.
1453 SCEVExpander Exp(*SE, "induction");
1455 // Count holds the overall loop count (N).
1456 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1457 BypassBlock->getTerminator());
1459 // The loop index does not have to start at Zero. Find the original start
1460 // value from the induction PHI node. If we don't have an induction variable
1461 // then we know that it starts at zero.
1462 Builder.SetInsertPoint(BypassBlock->getTerminator());
1463 Value *StartIdx = ExtendedIdx = OldInduction ?
1464 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1466 ConstantInt::get(IdxTy, 0);
1468 assert(BypassBlock && "Invalid loop structure");
1469 LoopBypassBlocks.push_back(BypassBlock);
1471 // Split the single block loop into the two loop structure described above.
1472 BasicBlock *VectorPH =
1473 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1474 BasicBlock *VecBody =
1475 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1476 BasicBlock *MiddleBlock =
1477 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1478 BasicBlock *ScalarPH =
1479 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1481 // Create and register the new vector loop.
1482 Loop* Lp = new Loop();
1483 Loop *ParentLoop = OrigLoop->getParentLoop();
1485 // Insert the new loop into the loop nest and register the new basic blocks
1486 // before calling any utilities such as SCEV that require valid LoopInfo.
1488 ParentLoop->addChildLoop(Lp);
1489 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1490 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1491 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1493 LI->addTopLevelLoop(Lp);
1495 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1497 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1499 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1501 // Generate the induction variable.
1502 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1503 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1504 // The loop step is equal to the vectorization factor (num of SIMD elements)
1505 // times the unroll factor (num of SIMD instructions).
1506 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1508 // This is the IR builder that we use to add all of the logic for bypassing
1509 // the new vector loop.
1510 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1511 setDebugLocFromInst(BypassBuilder,
1512 getDebugLocFromInstOrOperands(OldInduction));
1514 // We may need to extend the index in case there is a type mismatch.
1515 // We know that the count starts at zero and does not overflow.
1516 if (Count->getType() != IdxTy) {
1517 // The exit count can be of pointer type. Convert it to the correct
1519 if (ExitCount->getType()->isPointerTy())
1520 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1522 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1525 // Add the start index to the loop count to get the new end index.
1526 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1528 // Now we need to generate the expression for N - (N % VF), which is
1529 // the part that the vectorized body will execute.
1530 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1531 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1532 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1533 "end.idx.rnd.down");
1535 // Now, compare the new count to zero. If it is zero skip the vector loop and
1536 // jump to the scalar loop.
1537 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1540 BasicBlock *LastBypassBlock = BypassBlock;
1542 // Generate the code that checks in runtime if arrays overlap. We put the
1543 // checks into a separate block to make the more common case of few elements
1545 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1546 BypassBlock->getTerminator());
1547 if (MemRuntimeCheck) {
1548 // Create a new block containing the memory check.
1549 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1552 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1553 LoopBypassBlocks.push_back(CheckBlock);
1555 // Replace the branch into the memory check block with a conditional branch
1556 // for the "few elements case".
1557 Instruction *OldTerm = BypassBlock->getTerminator();
1558 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1559 OldTerm->eraseFromParent();
1561 Cmp = MemRuntimeCheck;
1562 LastBypassBlock = CheckBlock;
1565 LastBypassBlock->getTerminator()->eraseFromParent();
1566 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1569 // We are going to resume the execution of the scalar loop.
1570 // Go over all of the induction variables that we found and fix the
1571 // PHIs that are left in the scalar version of the loop.
1572 // The starting values of PHI nodes depend on the counter of the last
1573 // iteration in the vectorized loop.
1574 // If we come from a bypass edge then we need to start from the original
1577 // This variable saves the new starting index for the scalar loop.
1578 PHINode *ResumeIndex = 0;
1579 LoopVectorizationLegality::InductionList::iterator I, E;
1580 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1581 // Set builder to point to last bypass block.
1582 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1583 for (I = List->begin(), E = List->end(); I != E; ++I) {
1584 PHINode *OrigPhi = I->first;
1585 LoopVectorizationLegality::InductionInfo II = I->second;
1587 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1588 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1589 MiddleBlock->getTerminator());
1590 // We might have extended the type of the induction variable but we need a
1591 // truncated version for the scalar loop.
1592 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1593 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1594 MiddleBlock->getTerminator()) : 0;
1596 Value *EndValue = 0;
1598 case LoopVectorizationLegality::IK_NoInduction:
1599 llvm_unreachable("Unknown induction");
1600 case LoopVectorizationLegality::IK_IntInduction: {
1601 // Handle the integer induction counter.
1602 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1604 // We have the canonical induction variable.
1605 if (OrigPhi == OldInduction) {
1606 // Create a truncated version of the resume value for the scalar loop,
1607 // we might have promoted the type to a larger width.
1609 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1610 // The new PHI merges the original incoming value, in case of a bypass,
1611 // or the value at the end of the vectorized loop.
1612 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1613 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1614 TruncResumeVal->addIncoming(EndValue, VecBody);
1616 // We know what the end value is.
1617 EndValue = IdxEndRoundDown;
1618 // We also know which PHI node holds it.
1619 ResumeIndex = ResumeVal;
1623 // Not the canonical induction variable - add the vector loop count to the
1625 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1626 II.StartValue->getType(),
1628 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1631 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1632 // Convert the CountRoundDown variable to the PHI size.
1633 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1634 II.StartValue->getType(),
1636 // Handle reverse integer induction counter.
1637 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1640 case LoopVectorizationLegality::IK_PtrInduction: {
1641 // For pointer induction variables, calculate the offset using
1643 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1647 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1648 // The value at the end of the loop for the reverse pointer is calculated
1649 // by creating a GEP with a negative index starting from the start value.
1650 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1651 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1653 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1659 // The new PHI merges the original incoming value, in case of a bypass,
1660 // or the value at the end of the vectorized loop.
1661 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1662 if (OrigPhi == OldInduction)
1663 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1665 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1667 ResumeVal->addIncoming(EndValue, VecBody);
1669 // Fix the scalar body counter (PHI node).
1670 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1671 // The old inductions phi node in the scalar body needs the truncated value.
1672 if (OrigPhi == OldInduction)
1673 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1675 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1678 // If we are generating a new induction variable then we also need to
1679 // generate the code that calculates the exit value. This value is not
1680 // simply the end of the counter because we may skip the vectorized body
1681 // in case of a runtime check.
1683 assert(!ResumeIndex && "Unexpected resume value found");
1684 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1685 MiddleBlock->getTerminator());
1686 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1687 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1688 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1691 // Make sure that we found the index where scalar loop needs to continue.
1692 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1693 "Invalid resume Index");
1695 // Add a check in the middle block to see if we have completed
1696 // all of the iterations in the first vector loop.
1697 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1698 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1699 ResumeIndex, "cmp.n",
1700 MiddleBlock->getTerminator());
1702 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1703 // Remove the old terminator.
1704 MiddleBlock->getTerminator()->eraseFromParent();
1706 // Create i+1 and fill the PHINode.
1707 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1708 Induction->addIncoming(StartIdx, VectorPH);
1709 Induction->addIncoming(NextIdx, VecBody);
1710 // Create the compare.
1711 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1712 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1714 // Now we have two terminators. Remove the old one from the block.
1715 VecBody->getTerminator()->eraseFromParent();
1717 // Get ready to start creating new instructions into the vectorized body.
1718 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1721 LoopVectorPreHeader = VectorPH;
1722 LoopScalarPreHeader = ScalarPH;
1723 LoopMiddleBlock = MiddleBlock;
1724 LoopExitBlock = ExitBlock;
1725 LoopVectorBody = VecBody;
1726 LoopScalarBody = OldBasicBlock;
1729 /// This function returns the identity element (or neutral element) for
1730 /// the operation K.
1732 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1737 // Adding, Xoring, Oring zero to a number does not change it.
1738 return ConstantInt::get(Tp, 0);
1739 case RK_IntegerMult:
1740 // Multiplying a number by 1 does not change it.
1741 return ConstantInt::get(Tp, 1);
1743 // AND-ing a number with an all-1 value does not change it.
1744 return ConstantInt::get(Tp, -1, true);
1746 // Multiplying a number by 1 does not change it.
1747 return ConstantFP::get(Tp, 1.0L);
1749 // Adding zero to a number does not change it.
1750 return ConstantFP::get(Tp, 0.0L);
1752 llvm_unreachable("Unknown reduction kind");
1756 static Intrinsic::ID
1757 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1758 // If we have an intrinsic call, check if it is trivially vectorizable.
1759 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1760 switch (II->getIntrinsicID()) {
1761 case Intrinsic::sqrt:
1762 case Intrinsic::sin:
1763 case Intrinsic::cos:
1764 case Intrinsic::exp:
1765 case Intrinsic::exp2:
1766 case Intrinsic::log:
1767 case Intrinsic::log10:
1768 case Intrinsic::log2:
1769 case Intrinsic::fabs:
1770 case Intrinsic::copysign:
1771 case Intrinsic::floor:
1772 case Intrinsic::ceil:
1773 case Intrinsic::trunc:
1774 case Intrinsic::rint:
1775 case Intrinsic::nearbyint:
1776 case Intrinsic::round:
1777 case Intrinsic::pow:
1778 case Intrinsic::fma:
1779 case Intrinsic::fmuladd:
1780 case Intrinsic::lifetime_start:
1781 case Intrinsic::lifetime_end:
1782 return II->getIntrinsicID();
1784 return Intrinsic::not_intrinsic;
1789 return Intrinsic::not_intrinsic;
1792 Function *F = CI->getCalledFunction();
1793 // We're going to make assumptions on the semantics of the functions, check
1794 // that the target knows that it's available in this environment.
1795 if (!F || !TLI->getLibFunc(F->getName(), Func))
1796 return Intrinsic::not_intrinsic;
1798 // Otherwise check if we have a call to a function that can be turned into a
1799 // vector intrinsic.
1806 return Intrinsic::sin;
1810 return Intrinsic::cos;
1814 return Intrinsic::exp;
1816 case LibFunc::exp2f:
1817 case LibFunc::exp2l:
1818 return Intrinsic::exp2;
1822 return Intrinsic::log;
1823 case LibFunc::log10:
1824 case LibFunc::log10f:
1825 case LibFunc::log10l:
1826 return Intrinsic::log10;
1828 case LibFunc::log2f:
1829 case LibFunc::log2l:
1830 return Intrinsic::log2;
1832 case LibFunc::fabsf:
1833 case LibFunc::fabsl:
1834 return Intrinsic::fabs;
1835 case LibFunc::copysign:
1836 case LibFunc::copysignf:
1837 case LibFunc::copysignl:
1838 return Intrinsic::copysign;
1839 case LibFunc::floor:
1840 case LibFunc::floorf:
1841 case LibFunc::floorl:
1842 return Intrinsic::floor;
1844 case LibFunc::ceilf:
1845 case LibFunc::ceill:
1846 return Intrinsic::ceil;
1847 case LibFunc::trunc:
1848 case LibFunc::truncf:
1849 case LibFunc::truncl:
1850 return Intrinsic::trunc;
1852 case LibFunc::rintf:
1853 case LibFunc::rintl:
1854 return Intrinsic::rint;
1855 case LibFunc::nearbyint:
1856 case LibFunc::nearbyintf:
1857 case LibFunc::nearbyintl:
1858 return Intrinsic::nearbyint;
1859 case LibFunc::round:
1860 case LibFunc::roundf:
1861 case LibFunc::roundl:
1862 return Intrinsic::round;
1866 return Intrinsic::pow;
1869 return Intrinsic::not_intrinsic;
1872 /// This function translates the reduction kind to an LLVM binary operator.
1874 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1876 case LoopVectorizationLegality::RK_IntegerAdd:
1877 return Instruction::Add;
1878 case LoopVectorizationLegality::RK_IntegerMult:
1879 return Instruction::Mul;
1880 case LoopVectorizationLegality::RK_IntegerOr:
1881 return Instruction::Or;
1882 case LoopVectorizationLegality::RK_IntegerAnd:
1883 return Instruction::And;
1884 case LoopVectorizationLegality::RK_IntegerXor:
1885 return Instruction::Xor;
1886 case LoopVectorizationLegality::RK_FloatMult:
1887 return Instruction::FMul;
1888 case LoopVectorizationLegality::RK_FloatAdd:
1889 return Instruction::FAdd;
1890 case LoopVectorizationLegality::RK_IntegerMinMax:
1891 return Instruction::ICmp;
1892 case LoopVectorizationLegality::RK_FloatMinMax:
1893 return Instruction::FCmp;
1895 llvm_unreachable("Unknown reduction operation");
1899 Value *createMinMaxOp(IRBuilder<> &Builder,
1900 LoopVectorizationLegality::MinMaxReductionKind RK,
1903 CmpInst::Predicate P = CmpInst::ICMP_NE;
1906 llvm_unreachable("Unknown min/max reduction kind");
1907 case LoopVectorizationLegality::MRK_UIntMin:
1908 P = CmpInst::ICMP_ULT;
1910 case LoopVectorizationLegality::MRK_UIntMax:
1911 P = CmpInst::ICMP_UGT;
1913 case LoopVectorizationLegality::MRK_SIntMin:
1914 P = CmpInst::ICMP_SLT;
1916 case LoopVectorizationLegality::MRK_SIntMax:
1917 P = CmpInst::ICMP_SGT;
1919 case LoopVectorizationLegality::MRK_FloatMin:
1920 P = CmpInst::FCMP_OLT;
1922 case LoopVectorizationLegality::MRK_FloatMax:
1923 P = CmpInst::FCMP_OGT;
1928 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
1929 RK == LoopVectorizationLegality::MRK_FloatMax)
1930 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1932 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1934 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1939 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1940 //===------------------------------------------------===//
1942 // Notice: any optimization or new instruction that go
1943 // into the code below should be also be implemented in
1946 //===------------------------------------------------===//
1947 Constant *Zero = Builder.getInt32(0);
1949 // In order to support reduction variables we need to be able to vectorize
1950 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1951 // stages. First, we create a new vector PHI node with no incoming edges.
1952 // We use this value when we vectorize all of the instructions that use the
1953 // PHI. Next, after all of the instructions in the block are complete we
1954 // add the new incoming edges to the PHI. At this point all of the
1955 // instructions in the basic block are vectorized, so we can use them to
1956 // construct the PHI.
1957 PhiVector RdxPHIsToFix;
1959 // Scan the loop in a topological order to ensure that defs are vectorized
1961 LoopBlocksDFS DFS(OrigLoop);
1964 // Vectorize all of the blocks in the original loop.
1965 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1966 be = DFS.endRPO(); bb != be; ++bb)
1967 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1969 // At this point every instruction in the original loop is widened to
1970 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1971 // that we vectorized. The PHI nodes are currently empty because we did
1972 // not want to introduce cycles. Notice that the remaining PHI nodes
1973 // that we need to fix are reduction variables.
1975 // Create the 'reduced' values for each of the induction vars.
1976 // The reduced values are the vector values that we scalarize and combine
1977 // after the loop is finished.
1978 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1980 PHINode *RdxPhi = *it;
1981 assert(RdxPhi && "Unable to recover vectorized PHI");
1983 // Find the reduction variable descriptor.
1984 assert(Legal->getReductionVars()->count(RdxPhi) &&
1985 "Unable to find the reduction variable");
1986 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1987 (*Legal->getReductionVars())[RdxPhi];
1989 setDebugLocFromInst(Builder, RdxDesc.StartValue);
1991 // We need to generate a reduction vector from the incoming scalar.
1992 // To do so, we need to generate the 'identity' vector and overide
1993 // one of the elements with the incoming scalar reduction. We need
1994 // to do it in the vector-loop preheader.
1995 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1997 // This is the vector-clone of the value that leaves the loop.
1998 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1999 Type *VecTy = VectorExit[0]->getType();
2001 // Find the reduction identity variable. Zero for addition, or, xor,
2002 // one for multiplication, -1 for And.
2005 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2006 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2007 // MinMax reduction have the start value as their identify.
2008 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2012 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2013 VecTy->getScalarType());
2014 Identity = ConstantVector::getSplat(VF, Iden);
2016 // This vector is the Identity vector where the first element is the
2017 // incoming scalar reduction.
2018 VectorStart = Builder.CreateInsertElement(Identity,
2019 RdxDesc.StartValue, Zero);
2022 // Fix the vector-loop phi.
2023 // We created the induction variable so we know that the
2024 // preheader is the first entry.
2025 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2027 // Reductions do not have to start at zero. They can start with
2028 // any loop invariant values.
2029 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2030 BasicBlock *Latch = OrigLoop->getLoopLatch();
2031 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2032 VectorParts &Val = getVectorValue(LoopVal);
2033 for (unsigned part = 0; part < UF; ++part) {
2034 // Make sure to add the reduction stat value only to the
2035 // first unroll part.
2036 Value *StartVal = (part == 0) ? VectorStart : Identity;
2037 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2038 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2041 // Before each round, move the insertion point right between
2042 // the PHIs and the values we are going to write.
2043 // This allows us to write both PHINodes and the extractelement
2045 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2047 VectorParts RdxParts;
2048 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2049 for (unsigned part = 0; part < UF; ++part) {
2050 // This PHINode contains the vectorized reduction variable, or
2051 // the initial value vector, if we bypass the vector loop.
2052 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2053 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2054 Value *StartVal = (part == 0) ? VectorStart : Identity;
2055 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2056 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2057 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2058 RdxParts.push_back(NewPhi);
2061 // Reduce all of the unrolled parts into a single vector.
2062 Value *ReducedPartRdx = RdxParts[0];
2063 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2064 setDebugLocFromInst(Builder, ReducedPartRdx);
2065 for (unsigned part = 1; part < UF; ++part) {
2066 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2067 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2068 RdxParts[part], ReducedPartRdx,
2071 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2072 ReducedPartRdx, RdxParts[part]);
2075 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2076 // and vector ops, reducing the set of values being computed by half each
2078 assert(isPowerOf2_32(VF) &&
2079 "Reduction emission only supported for pow2 vectors!");
2080 Value *TmpVec = ReducedPartRdx;
2081 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2082 for (unsigned i = VF; i != 1; i >>= 1) {
2083 // Move the upper half of the vector to the lower half.
2084 for (unsigned j = 0; j != i/2; ++j)
2085 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2087 // Fill the rest of the mask with undef.
2088 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2089 UndefValue::get(Builder.getInt32Ty()));
2092 Builder.CreateShuffleVector(TmpVec,
2093 UndefValue::get(TmpVec->getType()),
2094 ConstantVector::get(ShuffleMask),
2097 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2098 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2101 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2104 // The result is in the first element of the vector.
2105 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2107 // Now, we need to fix the users of the reduction variable
2108 // inside and outside of the scalar remainder loop.
2109 // We know that the loop is in LCSSA form. We need to update the
2110 // PHI nodes in the exit blocks.
2111 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2112 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2113 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2114 if (!LCSSAPhi) continue;
2116 // All PHINodes need to have a single entry edge, or two if
2117 // we already fixed them.
2118 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2120 // We found our reduction value exit-PHI. Update it with the
2121 // incoming bypass edge.
2122 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2123 // Add an edge coming from the bypass.
2124 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2127 }// end of the LCSSA phi scan.
2129 // Fix the scalar loop reduction variable with the incoming reduction sum
2130 // from the vector body and from the backedge value.
2131 int IncomingEdgeBlockIdx =
2132 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2133 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2134 // Pick the other block.
2135 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2136 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2137 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2138 }// end of for each redux variable.
2140 // The Loop exit block may have single value PHI nodes where the incoming
2141 // value is 'undef'. While vectorizing we only handled real values that
2142 // were defined inside the loop. Here we handle the 'undef case'.
2144 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2145 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2146 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2147 if (!LCSSAPhi) continue;
2148 if (LCSSAPhi->getNumIncomingValues() == 1)
2149 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2154 InnerLoopVectorizer::VectorParts
2155 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2156 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2159 // Look for cached value.
2160 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2161 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2162 if (ECEntryIt != MaskCache.end())
2163 return ECEntryIt->second;
2165 VectorParts SrcMask = createBlockInMask(Src);
2167 // The terminator has to be a branch inst!
2168 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2169 assert(BI && "Unexpected terminator found");
2171 if (BI->isConditional()) {
2172 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2174 if (BI->getSuccessor(0) != Dst)
2175 for (unsigned part = 0; part < UF; ++part)
2176 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2178 for (unsigned part = 0; part < UF; ++part)
2179 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2181 MaskCache[Edge] = EdgeMask;
2185 MaskCache[Edge] = SrcMask;
2189 InnerLoopVectorizer::VectorParts
2190 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2191 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2193 // Loop incoming mask is all-one.
2194 if (OrigLoop->getHeader() == BB) {
2195 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2196 return getVectorValue(C);
2199 // This is the block mask. We OR all incoming edges, and with zero.
2200 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2201 VectorParts BlockMask = getVectorValue(Zero);
2204 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2205 VectorParts EM = createEdgeMask(*it, BB);
2206 for (unsigned part = 0; part < UF; ++part)
2207 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2214 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2215 BasicBlock *BB, PhiVector *PV) {
2216 // For each instruction in the old loop.
2217 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2218 VectorParts &Entry = WidenMap.get(it);
2219 switch (it->getOpcode()) {
2220 case Instruction::Br:
2221 // Nothing to do for PHIs and BR, since we already took care of the
2222 // loop control flow instructions.
2224 case Instruction::PHI:{
2225 PHINode* P = cast<PHINode>(it);
2226 // Handle reduction variables:
2227 if (Legal->getReductionVars()->count(P)) {
2228 for (unsigned part = 0; part < UF; ++part) {
2229 // This is phase one of vectorizing PHIs.
2230 Type *VecTy = VectorType::get(it->getType(), VF);
2231 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2232 LoopVectorBody-> getFirstInsertionPt());
2238 setDebugLocFromInst(Builder, P);
2239 // Check for PHI nodes that are lowered to vector selects.
2240 if (P->getParent() != OrigLoop->getHeader()) {
2241 // We know that all PHIs in non header blocks are converted into
2242 // selects, so we don't have to worry about the insertion order and we
2243 // can just use the builder.
2244 // At this point we generate the predication tree. There may be
2245 // duplications since this is a simple recursive scan, but future
2246 // optimizations will clean it up.
2248 unsigned NumIncoming = P->getNumIncomingValues();
2250 // Generate a sequence of selects of the form:
2251 // SELECT(Mask3, In3,
2252 // SELECT(Mask2, In2,
2254 for (unsigned In = 0; In < NumIncoming; In++) {
2255 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2257 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2259 for (unsigned part = 0; part < UF; ++part) {
2260 // We might have single edge PHIs (blocks) - use an identity
2261 // 'select' for the first PHI operand.
2263 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2266 // Select between the current value and the previous incoming edge
2267 // based on the incoming mask.
2268 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2269 Entry[part], "predphi");
2275 // This PHINode must be an induction variable.
2276 // Make sure that we know about it.
2277 assert(Legal->getInductionVars()->count(P) &&
2278 "Not an induction variable");
2280 LoopVectorizationLegality::InductionInfo II =
2281 Legal->getInductionVars()->lookup(P);
2284 case LoopVectorizationLegality::IK_NoInduction:
2285 llvm_unreachable("Unknown induction");
2286 case LoopVectorizationLegality::IK_IntInduction: {
2287 assert(P->getType() == II.StartValue->getType() && "Types must match");
2288 Type *PhiTy = P->getType();
2290 if (P == OldInduction) {
2291 // Handle the canonical induction variable. We might have had to
2293 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2295 // Handle other induction variables that are now based on the
2297 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2299 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2300 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2303 Broadcasted = getBroadcastInstrs(Broadcasted);
2304 // After broadcasting the induction variable we need to make the vector
2305 // consecutive by adding 0, 1, 2, etc.
2306 for (unsigned part = 0; part < UF; ++part)
2307 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2310 case LoopVectorizationLegality::IK_ReverseIntInduction:
2311 case LoopVectorizationLegality::IK_PtrInduction:
2312 case LoopVectorizationLegality::IK_ReversePtrInduction:
2313 // Handle reverse integer and pointer inductions.
2314 Value *StartIdx = ExtendedIdx;
2315 // This is the normalized GEP that starts counting at zero.
2316 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2319 // Handle the reverse integer induction variable case.
2320 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2321 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2322 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2324 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2327 // This is a new value so do not hoist it out.
2328 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2329 // After broadcasting the induction variable we need to make the
2330 // vector consecutive by adding ... -3, -2, -1, 0.
2331 for (unsigned part = 0; part < UF; ++part)
2332 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2337 // Handle the pointer induction variable case.
2338 assert(P->getType()->isPointerTy() && "Unexpected type.");
2340 // Is this a reverse induction ptr or a consecutive induction ptr.
2341 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2344 // This is the vector of results. Notice that we don't generate
2345 // vector geps because scalar geps result in better code.
2346 for (unsigned part = 0; part < UF; ++part) {
2347 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2348 for (unsigned int i = 0; i < VF; ++i) {
2349 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2350 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2353 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2355 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2357 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2359 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2360 Builder.getInt32(i),
2363 Entry[part] = VecVal;
2370 case Instruction::Add:
2371 case Instruction::FAdd:
2372 case Instruction::Sub:
2373 case Instruction::FSub:
2374 case Instruction::Mul:
2375 case Instruction::FMul:
2376 case Instruction::UDiv:
2377 case Instruction::SDiv:
2378 case Instruction::FDiv:
2379 case Instruction::URem:
2380 case Instruction::SRem:
2381 case Instruction::FRem:
2382 case Instruction::Shl:
2383 case Instruction::LShr:
2384 case Instruction::AShr:
2385 case Instruction::And:
2386 case Instruction::Or:
2387 case Instruction::Xor: {
2388 // Just widen binops.
2389 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2390 setDebugLocFromInst(Builder, BinOp);
2391 VectorParts &A = getVectorValue(it->getOperand(0));
2392 VectorParts &B = getVectorValue(it->getOperand(1));
2394 // Use this vector value for all users of the original instruction.
2395 for (unsigned Part = 0; Part < UF; ++Part) {
2396 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2398 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2399 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2400 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2401 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2402 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2404 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2405 VecOp->setIsExact(BinOp->isExact());
2411 case Instruction::Select: {
2413 // If the selector is loop invariant we can create a select
2414 // instruction with a scalar condition. Otherwise, use vector-select.
2415 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2417 setDebugLocFromInst(Builder, it);
2419 // The condition can be loop invariant but still defined inside the
2420 // loop. This means that we can't just use the original 'cond' value.
2421 // We have to take the 'vectorized' value and pick the first lane.
2422 // Instcombine will make this a no-op.
2423 VectorParts &Cond = getVectorValue(it->getOperand(0));
2424 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2425 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2426 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2427 Builder.getInt32(0));
2428 for (unsigned Part = 0; Part < UF; ++Part) {
2429 Entry[Part] = Builder.CreateSelect(
2430 InvariantCond ? ScalarCond : Cond[Part],
2437 case Instruction::ICmp:
2438 case Instruction::FCmp: {
2439 // Widen compares. Generate vector compares.
2440 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2441 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2442 setDebugLocFromInst(Builder, it);
2443 VectorParts &A = getVectorValue(it->getOperand(0));
2444 VectorParts &B = getVectorValue(it->getOperand(1));
2445 for (unsigned Part = 0; Part < UF; ++Part) {
2448 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2450 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2456 case Instruction::Store:
2457 case Instruction::Load:
2458 vectorizeMemoryInstruction(it, Legal);
2460 case Instruction::ZExt:
2461 case Instruction::SExt:
2462 case Instruction::FPToUI:
2463 case Instruction::FPToSI:
2464 case Instruction::FPExt:
2465 case Instruction::PtrToInt:
2466 case Instruction::IntToPtr:
2467 case Instruction::SIToFP:
2468 case Instruction::UIToFP:
2469 case Instruction::Trunc:
2470 case Instruction::FPTrunc:
2471 case Instruction::BitCast: {
2472 CastInst *CI = dyn_cast<CastInst>(it);
2473 setDebugLocFromInst(Builder, it);
2474 /// Optimize the special case where the source is the induction
2475 /// variable. Notice that we can only optimize the 'trunc' case
2476 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2477 /// c. other casts depend on pointer size.
2478 if (CI->getOperand(0) == OldInduction &&
2479 it->getOpcode() == Instruction::Trunc) {
2480 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2482 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2483 for (unsigned Part = 0; Part < UF; ++Part)
2484 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2487 /// Vectorize casts.
2488 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2490 VectorParts &A = getVectorValue(it->getOperand(0));
2491 for (unsigned Part = 0; Part < UF; ++Part)
2492 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2496 case Instruction::Call: {
2497 // Ignore dbg intrinsics.
2498 if (isa<DbgInfoIntrinsic>(it))
2500 setDebugLocFromInst(Builder, it);
2502 Module *M = BB->getParent()->getParent();
2503 CallInst *CI = cast<CallInst>(it);
2504 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2505 assert(ID && "Not an intrinsic call!");
2507 case Intrinsic::lifetime_end:
2508 case Intrinsic::lifetime_start:
2509 scalarizeInstruction(it);
2512 for (unsigned Part = 0; Part < UF; ++Part) {
2513 SmallVector<Value *, 4> Args;
2514 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2515 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2516 Args.push_back(Arg[Part]);
2518 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2519 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2520 Entry[Part] = Builder.CreateCall(F, Args);
2528 // All other instructions are unsupported. Scalarize them.
2529 scalarizeInstruction(it);
2532 }// end of for_each instr.
2535 void InnerLoopVectorizer::updateAnalysis() {
2536 // Forget the original basic block.
2537 SE->forgetLoop(OrigLoop);
2539 // Update the dominator tree information.
2540 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2541 "Entry does not dominate exit.");
2543 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2544 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2545 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2546 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2547 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2548 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2549 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2550 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2552 DEBUG(DT->verifyAnalysis());
2555 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2556 if (!EnableIfConversion)
2559 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2560 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2562 // A list of pointers that we can safely read and write to.
2563 SmallPtrSet<Value *, 8> SafePointes;
2565 // Collect safe addresses.
2566 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2567 BasicBlock *BB = LoopBlocks[i];
2569 if (blockNeedsPredication(BB))
2572 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2573 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2574 SafePointes.insert(LI->getPointerOperand());
2575 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2576 SafePointes.insert(SI->getPointerOperand());
2580 // Collect the blocks that need predication.
2581 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2582 BasicBlock *BB = LoopBlocks[i];
2584 // We don't support switch statements inside loops.
2585 if (!isa<BranchInst>(BB->getTerminator()))
2588 // We must be able to predicate all blocks that need to be predicated.
2589 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2593 // We can if-convert this loop.
2597 bool LoopVectorizationLegality::canVectorize() {
2598 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2599 // be canonicalized.
2600 if (!TheLoop->getLoopPreheader())
2603 // We can only vectorize innermost loops.
2604 if (TheLoop->getSubLoopsVector().size())
2607 // We must have a single backedge.
2608 if (TheLoop->getNumBackEdges() != 1)
2611 // We must have a single exiting block.
2612 if (!TheLoop->getExitingBlock())
2615 unsigned NumBlocks = TheLoop->getNumBlocks();
2617 // Check if we can if-convert non single-bb loops.
2618 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2619 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2623 // We need to have a loop header.
2624 BasicBlock *Latch = TheLoop->getLoopLatch();
2625 DEBUG(dbgs() << "LV: Found a loop: " <<
2626 TheLoop->getHeader()->getName() << "\n");
2628 // ScalarEvolution needs to be able to find the exit count.
2629 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2630 if (ExitCount == SE->getCouldNotCompute()) {
2631 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2635 // Do not loop-vectorize loops with a tiny trip count.
2636 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2637 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2638 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2639 "This loop is not worth vectorizing.\n");
2643 // Check if we can vectorize the instructions and CFG in this loop.
2644 if (!canVectorizeInstrs()) {
2645 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2649 // Go over each instruction and look at memory deps.
2650 if (!canVectorizeMemory()) {
2651 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2655 // Collect all of the variables that remain uniform after vectorization.
2656 collectLoopUniforms();
2658 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2659 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2662 // Okay! We can vectorize. At this point we don't have any other mem analysis
2663 // which may limit our maximum vectorization factor, so just return true with
2668 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2669 if (Ty->isPointerTy())
2670 return DL.getIntPtrType(Ty->getContext());
2674 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2675 Ty0 = convertPointerToIntegerType(DL, Ty0);
2676 Ty1 = convertPointerToIntegerType(DL, Ty1);
2677 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2682 /// \brief Check that the instruction has outside loop users and is not an
2683 /// identified reduction variable.
2684 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2685 SmallPtrSet<Value *, 4> &Reductions) {
2686 // Reduction instructions are allowed to have exit users. All other
2687 // instructions must not have external users.
2688 if (!Reductions.count(Inst))
2689 //Check that all of the users of the loop are inside the BB.
2690 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2692 Instruction *U = cast<Instruction>(*I);
2693 // This user may be a reduction exit value.
2694 if (!TheLoop->contains(U)) {
2695 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2702 bool LoopVectorizationLegality::canVectorizeInstrs() {
2703 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2704 BasicBlock *Header = TheLoop->getHeader();
2706 // Look for the attribute signaling the absence of NaNs.
2707 Function &F = *Header->getParent();
2708 if (F.hasFnAttribute("no-nans-fp-math"))
2709 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2710 AttributeSet::FunctionIndex,
2711 "no-nans-fp-math").getValueAsString() == "true";
2713 // For each block in the loop.
2714 for (Loop::block_iterator bb = TheLoop->block_begin(),
2715 be = TheLoop->block_end(); bb != be; ++bb) {
2717 // Scan the instructions in the block and look for hazards.
2718 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2721 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2722 Type *PhiTy = Phi->getType();
2723 // Check that this PHI type is allowed.
2724 if (!PhiTy->isIntegerTy() &&
2725 !PhiTy->isFloatingPointTy() &&
2726 !PhiTy->isPointerTy()) {
2727 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2731 // If this PHINode is not in the header block, then we know that we
2732 // can convert it to select during if-conversion. No need to check if
2733 // the PHIs in this block are induction or reduction variables.
2734 if (*bb != Header) {
2735 // Check that this instruction has no outside users or is an
2736 // identified reduction value with an outside user.
2737 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2742 // We only allow if-converted PHIs with more than two incoming values.
2743 if (Phi->getNumIncomingValues() != 2) {
2744 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2748 // This is the value coming from the preheader.
2749 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2750 // Check if this is an induction variable.
2751 InductionKind IK = isInductionVariable(Phi);
2753 if (IK_NoInduction != IK) {
2754 // Get the widest type.
2756 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2758 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2760 // Int inductions are special because we only allow one IV.
2761 if (IK == IK_IntInduction) {
2762 // Use the phi node with the widest type as induction. Use the last
2763 // one if there are multiple (no good reason for doing this other
2764 // than it is expedient).
2765 if (!Induction || PhiTy == WidestIndTy)
2769 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2770 Inductions[Phi] = InductionInfo(StartValue, IK);
2774 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2775 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2778 if (AddReductionVar(Phi, RK_IntegerMult)) {
2779 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2782 if (AddReductionVar(Phi, RK_IntegerOr)) {
2783 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2786 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2787 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2790 if (AddReductionVar(Phi, RK_IntegerXor)) {
2791 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2794 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2795 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2798 if (AddReductionVar(Phi, RK_FloatMult)) {
2799 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2802 if (AddReductionVar(Phi, RK_FloatAdd)) {
2803 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2806 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2807 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
2812 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2814 }// end of PHI handling
2816 // We still don't handle functions. However, we can ignore dbg intrinsic
2817 // calls and we do handle certain intrinsic and libm functions.
2818 CallInst *CI = dyn_cast<CallInst>(it);
2819 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2820 DEBUG(dbgs() << "LV: Found a call site.\n");
2824 // Check that the instruction return type is vectorizable.
2825 if (!VectorType::isValidElementType(it->getType()) &&
2826 !it->getType()->isVoidTy()) {
2827 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2831 // Check that the stored type is vectorizable.
2832 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2833 Type *T = ST->getValueOperand()->getType();
2834 if (!VectorType::isValidElementType(T))
2838 // Reduction instructions are allowed to have exit users.
2839 // All other instructions must not have external users.
2840 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2848 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2849 if (Inductions.empty())
2856 void LoopVectorizationLegality::collectLoopUniforms() {
2857 // We now know that the loop is vectorizable!
2858 // Collect variables that will remain uniform after vectorization.
2859 std::vector<Value*> Worklist;
2860 BasicBlock *Latch = TheLoop->getLoopLatch();
2862 // Start with the conditional branch and walk up the block.
2863 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2865 while (Worklist.size()) {
2866 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2867 Worklist.pop_back();
2869 // Look at instructions inside this loop.
2870 // Stop when reaching PHI nodes.
2871 // TODO: we need to follow values all over the loop, not only in this block.
2872 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2875 // This is a known uniform.
2878 // Insert all operands.
2879 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2884 /// \brief Analyses memory accesses in a loop.
2886 /// Checks whether run time pointer checks are needed and builds sets for data
2887 /// dependence checking.
2888 class AccessAnalysis {
2890 /// \brief Read or write access location.
2891 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
2892 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
2894 /// \brief Set of potential dependent memory accesses.
2895 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2897 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2898 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2899 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2901 /// \brief Register a load and whether it is only read from.
2902 void addLoad(Value *Ptr, bool IsReadOnly) {
2903 Accesses.insert(MemAccessInfo(Ptr, false));
2905 ReadOnlyPtr.insert(Ptr);
2908 /// \brief Register a store.
2909 void addStore(Value *Ptr) {
2910 Accesses.insert(MemAccessInfo(Ptr, true));
2913 /// \brief Check whether we can check the pointers at runtime for
2914 /// non-intersection.
2915 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2916 unsigned &NumComparisons, ScalarEvolution *SE,
2919 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2920 /// and builds sets of dependent accesses.
2921 void buildDependenceSets() {
2922 // Process read-write pointers first.
2923 processMemAccesses(false);
2924 // Next, process read pointers.
2925 processMemAccesses(true);
2928 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
2930 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
2932 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
2935 typedef SetVector<MemAccessInfo> PtrAccessSet;
2936 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
2938 /// \brief Go over all memory access or only the deferred ones if
2939 /// \p UseDeferred is true and check whether runtime pointer checks are needed
2940 /// and build sets of dependency check candidates.
2941 void processMemAccesses(bool UseDeferred);
2943 /// Set of all accesses.
2944 PtrAccessSet Accesses;
2946 /// Set of access to check after all writes have been processed.
2947 PtrAccessSet DeferredAccesses;
2949 /// Map of pointers to last access encountered.
2950 UnderlyingObjToAccessMap ObjToLastAccess;
2952 /// Set of accesses that need a further dependence check.
2953 MemAccessInfoSet CheckDeps;
2955 /// Set of pointers that are read only.
2956 SmallPtrSet<Value*, 16> ReadOnlyPtr;
2958 /// Set of underlying objects already written to.
2959 SmallPtrSet<Value*, 16> WriteObjects;
2963 /// Sets of potentially dependent accesses - members of one set share an
2964 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
2965 /// dependence check.
2966 DepCandidates &DepCands;
2968 bool AreAllWritesIdentified;
2969 bool AreAllReadsIdentified;
2970 bool IsRTCheckNeeded;
2973 } // end anonymous namespace
2975 /// \brief Check whether a pointer can participate in a runtime bounds check.
2976 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
2977 const SCEV *PtrScev = SE->getSCEV(Ptr);
2978 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
2982 return AR->isAffine();
2985 bool AccessAnalysis::canCheckPtrAtRT(
2986 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2987 unsigned &NumComparisons, ScalarEvolution *SE,
2989 // Find pointers with computable bounds. We are going to use this information
2990 // to place a runtime bound check.
2991 unsigned NumReadPtrChecks = 0;
2992 unsigned NumWritePtrChecks = 0;
2993 bool CanDoRT = true;
2995 bool IsDepCheckNeeded = isDependencyCheckNeeded();
2996 // We assign consecutive id to access from different dependence sets.
2997 // Accesses within the same set don't need a runtime check.
2998 unsigned RunningDepId = 1;
2999 DenseMap<Value *, unsigned> DepSetId;
3001 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3003 const MemAccessInfo &Access = *AI;
3004 Value *Ptr = Access.getPointer();
3005 bool IsWrite = Access.getInt();
3007 // Just add write checks if we have both.
3008 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3012 ++NumWritePtrChecks;
3016 if (hasComputableBounds(SE, Ptr)) {
3017 // The id of the dependence set.
3020 if (IsDepCheckNeeded) {
3021 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3022 unsigned &LeaderId = DepSetId[Leader];
3024 LeaderId = RunningDepId++;
3027 // Each access has its own dependence set.
3028 DepId = RunningDepId++;
3030 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3032 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3038 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3039 NumComparisons = 0; // Only one dependence set.
3041 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3042 NumWritePtrChecks - 1));
3046 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3047 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3050 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3051 // We process the set twice: first we process read-write pointers, last we
3052 // process read-only pointers. This allows us to skip dependence tests for
3053 // read-only pointers.
3055 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3056 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3057 const MemAccessInfo &Access = *AI;
3058 Value *Ptr = Access.getPointer();
3059 bool IsWrite = Access.getInt();
3061 DepCands.insert(Access);
3063 // Memorize read-only pointers for later processing and skip them in the
3064 // first round (they need to be checked after we have seen all write
3065 // pointers). Note: we also mark pointer that are not consecutive as
3066 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3067 // second check for "!IsWrite".
3068 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3069 if (!UseDeferred && IsReadOnlyPtr) {
3070 DeferredAccesses.insert(Access);
3074 bool NeedDepCheck = false;
3075 // Check whether there is the possiblity of dependency because of underlying
3076 // objects being the same.
3077 typedef SmallVector<Value*, 16> ValueVector;
3078 ValueVector TempObjects;
3079 GetUnderlyingObjects(Ptr, TempObjects, DL);
3080 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3082 Value *UnderlyingObj = *UI;
3084 // If this is a write then it needs to be an identified object. If this a
3085 // read and all writes (so far) are identified function scope objects we
3086 // don't need an identified underlying object but only an Argument (the
3087 // next write is going to invalidate this assumption if it is
3089 // This is a micro-optimization for the case where all writes are
3090 // identified and we have one argument pointer.
3091 // Otherwise, we do need a runtime check.
3092 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3093 (!IsWrite && (!AreAllWritesIdentified ||
3094 !isa<Argument>(UnderlyingObj)) &&
3095 !isIdentifiedObject(UnderlyingObj))) {
3096 DEBUG(dbgs() << "LV: Found an unidentified " <<
3097 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3099 IsRTCheckNeeded = (IsRTCheckNeeded ||
3100 !isIdentifiedObject(UnderlyingObj) ||
3101 !AreAllReadsIdentified);
3104 AreAllWritesIdentified = false;
3106 AreAllReadsIdentified = false;
3109 // If this is a write - check other reads and writes for conflicts. If
3110 // this is a read only check other writes for conflicts (but only if there
3111 // is no other write to the ptr - this is an optimization to catch "a[i] =
3112 // a[i] + " without having to do a dependence check).
3113 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3114 NeedDepCheck = true;
3117 WriteObjects.insert(UnderlyingObj);
3119 // Create sets of pointers connected by shared underlying objects.
3120 UnderlyingObjToAccessMap::iterator Prev =
3121 ObjToLastAccess.find(UnderlyingObj);
3122 if (Prev != ObjToLastAccess.end())
3123 DepCands.unionSets(Access, Prev->second);
3125 ObjToLastAccess[UnderlyingObj] = Access;
3129 CheckDeps.insert(Access);
3134 /// \brief Checks memory dependences among accesses to the same underlying
3135 /// object to determine whether there vectorization is legal or not (and at
3136 /// which vectorization factor).
3138 /// This class works under the assumption that we already checked that memory
3139 /// locations with different underlying pointers are "must-not alias".
3140 /// We use the ScalarEvolution framework to symbolically evalutate access
3141 /// functions pairs. Since we currently don't restructure the loop we can rely
3142 /// on the program order of memory accesses to determine their safety.
3143 /// At the moment we will only deem accesses as safe for:
3144 /// * A negative constant distance assuming program order.
3146 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3147 /// a[i] = tmp; y = a[i];
3149 /// The latter case is safe because later checks guarantuee that there can't
3150 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3151 /// the same variable: a header phi can only be an induction or a reduction, a
3152 /// reduction can't have a memory sink, an induction can't have a memory
3153 /// source). This is important and must not be violated (or we have to
3154 /// resort to checking for cycles through memory).
3156 /// * A positive constant distance assuming program order that is bigger
3157 /// than the biggest memory access.
3159 /// tmp = a[i] OR b[i] = x
3160 /// a[i+2] = tmp y = b[i+2];
3162 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3164 /// * Zero distances and all accesses have the same size.
3166 class MemoryDepChecker {
3168 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3169 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3171 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3172 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3174 /// \brief Register the location (instructions are given increasing numbers)
3175 /// of a write access.
3176 void addAccess(StoreInst *SI) {
3177 Value *Ptr = SI->getPointerOperand();
3178 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3179 InstMap.push_back(SI);
3183 /// \brief Register the location (instructions are given increasing numbers)
3184 /// of a write access.
3185 void addAccess(LoadInst *LI) {
3186 Value *Ptr = LI->getPointerOperand();
3187 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3188 InstMap.push_back(LI);
3192 /// \brief Check whether the dependencies between the accesses are safe.
3194 /// Only checks sets with elements in \p CheckDeps.
3195 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3196 MemAccessInfoSet &CheckDeps);
3198 /// \brief The maximum number of bytes of a vector register we can vectorize
3199 /// the accesses safely with.
3200 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3203 ScalarEvolution *SE;
3205 const Loop *InnermostLoop;
3207 /// \brief Maps access locations (ptr, read/write) to program order.
3208 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3210 /// \brief Memory access instructions in program order.
3211 SmallVector<Instruction *, 16> InstMap;
3213 /// \brief The program order index to be used for the next instruction.
3216 // We can access this many bytes in parallel safely.
3217 unsigned MaxSafeDepDistBytes;
3219 /// \brief Check whether there is a plausible dependence between the two
3222 /// Access \p A must happen before \p B in program order. The two indices
3223 /// identify the index into the program order map.
3225 /// This function checks whether there is a plausible dependence (or the
3226 /// absence of such can't be proved) between the two accesses. If there is a
3227 /// plausible dependence but the dependence distance is bigger than one
3228 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3229 /// distance is smaller than any other distance encountered so far).
3230 /// Otherwise, this function returns true signaling a possible dependence.
3231 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3232 const MemAccessInfo &B, unsigned BIdx);
3234 /// \brief Check whether the data dependence could prevent store-load
3236 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3239 } // end anonymous namespace
3241 static bool isInBoundsGep(Value *Ptr) {
3242 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3243 return GEP->isInBounds();
3247 /// \brief Check whether the access through \p Ptr has a constant stride.
3248 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3250 const Type *Ty = Ptr->getType();
3251 assert(Ty->isPointerTy() && "Unexpected non ptr");
3253 // Make sure that the pointer does not point to aggregate types.
3254 const PointerType *PtrTy = cast<PointerType>(Ty);
3255 if (PtrTy->getElementType()->isAggregateType()) {
3256 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
3261 const SCEV *PtrScev = SE->getSCEV(Ptr);
3262 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3264 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3265 << *Ptr << " SCEV: " << *PtrScev << "\n");
3269 // The accesss function must stride over the innermost loop.
3270 if (Lp != AR->getLoop()) {
3271 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
3272 *Ptr << " SCEV: " << *PtrScev << "\n");
3275 // The address calculation must not wrap. Otherwise, a dependence could be
3277 // An inbounds getelementptr that is a AddRec with a unit stride
3278 // cannot wrap per definition. The unit stride requirement is checked later.
3279 // An getelementptr without an inbounds attribute and unit stride would have
3280 // to access the pointer value "0" which is undefined behavior in address
3281 // space 0, therefore we can also vectorize this case.
3282 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3283 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3284 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
3285 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
3286 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3287 << *Ptr << " SCEV: " << *PtrScev << "\n");
3291 // Check the step is constant.
3292 const SCEV *Step = AR->getStepRecurrence(*SE);
3294 // Calculate the pointer stride and check if it is consecutive.
3295 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3297 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3298 " SCEV: " << *PtrScev << "\n");
3302 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
3303 const APInt &APStepVal = C->getValue()->getValue();
3305 // Huge step value - give up.
3306 if (APStepVal.getBitWidth() > 64)
3309 int64_t StepVal = APStepVal.getSExtValue();
3312 int64_t Stride = StepVal / Size;
3313 int64_t Rem = StepVal % Size;
3317 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3318 // know we can't "wrap around the address space". In case of address space
3319 // zero we know that this won't happen without triggering undefined behavior.
3320 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
3321 Stride != 1 && Stride != -1)
3327 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3328 unsigned TypeByteSize) {
3329 // If loads occur at a distance that is not a multiple of a feasible vector
3330 // factor store-load forwarding does not take place.
3331 // Positive dependences might cause troubles because vectorizing them might
3332 // prevent store-load forwarding making vectorized code run a lot slower.
3333 // a[i] = a[i-3] ^ a[i-8];
3334 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3335 // hence on your typical architecture store-load forwarding does not take
3336 // place. Vectorizing in such cases does not make sense.
3337 // Store-load forwarding distance.
3338 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3339 // Maximum vector factor.
3340 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3341 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3342 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3344 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3346 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3347 MaxVFWithoutSLForwardIssues = (vf >>=1);
3352 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3353 DEBUG(dbgs() << "LV: Distance " << Distance <<
3354 " that could cause a store-load forwarding conflict\n");
3358 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3359 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3360 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3364 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3365 const MemAccessInfo &B, unsigned BIdx) {
3366 assert (AIdx < BIdx && "Must pass arguments in program order");
3368 Value *APtr = A.getPointer();
3369 Value *BPtr = B.getPointer();
3370 bool AIsWrite = A.getInt();
3371 bool BIsWrite = B.getInt();
3373 // Two reads are independent.
3374 if (!AIsWrite && !BIsWrite)
3377 const SCEV *AScev = SE->getSCEV(APtr);
3378 const SCEV *BScev = SE->getSCEV(BPtr);
3380 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3381 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3383 const SCEV *Src = AScev;
3384 const SCEV *Sink = BScev;
3386 // If the induction step is negative we have to invert source and sink of the
3388 if (StrideAPtr < 0) {
3391 std::swap(APtr, BPtr);
3392 std::swap(Src, Sink);
3393 std::swap(AIsWrite, BIsWrite);
3394 std::swap(AIdx, BIdx);
3395 std::swap(StrideAPtr, StrideBPtr);
3398 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3400 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3401 << "(Induction step: " << StrideAPtr << ")\n");
3402 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3403 << *InstMap[BIdx] << ": " << *Dist << "\n");
3405 // Need consecutive accesses. We don't want to vectorize
3406 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3407 // the address space.
3408 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3409 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3413 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3415 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3419 Type *ATy = APtr->getType()->getPointerElementType();
3420 Type *BTy = BPtr->getType()->getPointerElementType();
3421 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3423 // Negative distances are not plausible dependencies.
3424 const APInt &Val = C->getValue()->getValue();
3425 if (Val.isNegative()) {
3426 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3427 if (IsTrueDataDependence &&
3428 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3432 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3436 // Write to the same location with the same size.
3437 // Could be improved to assert type sizes are the same (i32 == float, etc).
3441 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3445 assert(Val.isStrictlyPositive() && "Expect a positive value");
3447 // Positive distance bigger than max vectorization factor.
3450 "LV: ReadWrite-Write positive dependency with different types");
3454 unsigned Distance = (unsigned) Val.getZExtValue();
3456 // Bail out early if passed-in parameters make vectorization not feasible.
3457 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3458 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3460 // The distance must be bigger than the size needed for a vectorized version
3461 // of the operation and the size of the vectorized operation must not be
3462 // bigger than the currrent maximum size.
3463 if (Distance < 2*TypeByteSize ||
3464 2*TypeByteSize > MaxSafeDepDistBytes ||
3465 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3466 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3467 << Val.getSExtValue() << "\n");
3471 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3472 Distance : MaxSafeDepDistBytes;
3474 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3475 if (IsTrueDataDependence &&
3476 couldPreventStoreLoadForward(Distance, TypeByteSize))
3479 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3480 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3486 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3487 MemAccessInfoSet &CheckDeps) {
3489 MaxSafeDepDistBytes = -1U;
3490 while (!CheckDeps.empty()) {
3491 MemAccessInfo CurAccess = *CheckDeps.begin();
3493 // Get the relevant memory access set.
3494 EquivalenceClasses<MemAccessInfo>::iterator I =
3495 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3497 // Check accesses within this set.
3498 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3499 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3501 // Check every access pair.
3503 CheckDeps.erase(*AI);
3504 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3506 // Check every accessing instruction pair in program order.
3507 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3508 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3509 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3510 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3511 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3513 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3524 bool LoopVectorizationLegality::canVectorizeMemory() {
3526 typedef SmallVector<Value*, 16> ValueVector;
3527 typedef SmallPtrSet<Value*, 16> ValueSet;
3529 // Holds the Load and Store *instructions*.
3533 // Holds all the different accesses in the loop.
3534 unsigned NumReads = 0;
3535 unsigned NumReadWrites = 0;
3537 PtrRtCheck.Pointers.clear();
3538 PtrRtCheck.Need = false;
3540 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3541 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3544 for (Loop::block_iterator bb = TheLoop->block_begin(),
3545 be = TheLoop->block_end(); bb != be; ++bb) {
3547 // Scan the BB and collect legal loads and stores.
3548 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3551 // If this is a load, save it. If this instruction can read from memory
3552 // but is not a load, then we quit. Notice that we don't handle function
3553 // calls that read or write.
3554 if (it->mayReadFromMemory()) {
3555 // Many math library functions read the rounding mode. We will only
3556 // vectorize a loop if it contains known function calls that don't set
3557 // the flag. Therefore, it is safe to ignore this read from memory.
3558 CallInst *Call = dyn_cast<CallInst>(it);
3559 if (Call && getIntrinsicIDForCall(Call, TLI))
3562 LoadInst *Ld = dyn_cast<LoadInst>(it);
3563 if (!Ld) return false;
3564 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3565 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3568 Loads.push_back(Ld);
3569 DepChecker.addAccess(Ld);
3573 // Save 'store' instructions. Abort if other instructions write to memory.
3574 if (it->mayWriteToMemory()) {
3575 StoreInst *St = dyn_cast<StoreInst>(it);
3576 if (!St) return false;
3577 if (!St->isSimple() && !IsAnnotatedParallel) {
3578 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3581 Stores.push_back(St);
3582 DepChecker.addAccess(St);
3587 // Now we have two lists that hold the loads and the stores.
3588 // Next, we find the pointers that they use.
3590 // Check if we see any stores. If there are no stores, then we don't
3591 // care if the pointers are *restrict*.
3592 if (!Stores.size()) {
3593 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3597 AccessAnalysis::DepCandidates DependentAccesses;
3598 AccessAnalysis Accesses(DL, DependentAccesses);
3600 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3601 // multiple times on the same object. If the ptr is accessed twice, once
3602 // for read and once for write, it will only appear once (on the write
3603 // list). This is okay, since we are going to check for conflicts between
3604 // writes and between reads and writes, but not between reads and reads.
3607 ValueVector::iterator I, IE;
3608 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3609 StoreInst *ST = cast<StoreInst>(*I);
3610 Value* Ptr = ST->getPointerOperand();
3612 if (isUniform(Ptr)) {
3613 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3617 // If we did *not* see this pointer before, insert it to the read-write
3618 // list. At this phase it is only a 'write' list.
3619 if (Seen.insert(Ptr)) {
3621 Accesses.addStore(Ptr);
3625 if (IsAnnotatedParallel) {
3627 << "LV: A loop annotated parallel, ignore memory dependency "
3632 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3633 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3634 LoadInst *LD = cast<LoadInst>(*I);
3635 Value* Ptr = LD->getPointerOperand();
3636 // If we did *not* see this pointer before, insert it to the
3637 // read list. If we *did* see it before, then it is already in
3638 // the read-write list. This allows us to vectorize expressions
3639 // such as A[i] += x; Because the address of A[i] is a read-write
3640 // pointer. This only works if the index of A[i] is consecutive.
3641 // If the address of i is unknown (for example A[B[i]]) then we may
3642 // read a few words, modify, and write a few words, and some of the
3643 // words may be written to the same address.
3644 bool IsReadOnlyPtr = false;
3645 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3647 IsReadOnlyPtr = true;
3649 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3652 // If we write (or read-write) to a single destination and there are no
3653 // other reads in this loop then is it safe to vectorize.
3654 if (NumReadWrites == 1 && NumReads == 0) {
3655 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3659 // Build dependence sets and check whether we need a runtime pointer bounds
3661 Accesses.buildDependenceSets();
3662 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3664 // Find pointers with computable bounds. We are going to use this information
3665 // to place a runtime bound check.
3666 unsigned NumComparisons = 0;
3667 bool CanDoRT = false;
3669 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3672 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3673 " pointer comparisons.\n");
3675 // If we only have one set of dependences to check pointers among we don't
3676 // need a runtime check.
3677 if (NumComparisons == 0 && NeedRTCheck)
3678 NeedRTCheck = false;
3680 // Check that we did not collect too many pointers or found a unsizeable
3682 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3688 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3691 if (NeedRTCheck && !CanDoRT) {
3692 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3693 "the array bounds.\n");
3698 PtrRtCheck.Need = NeedRTCheck;
3700 bool CanVecMem = true;
3701 if (Accesses.isDependencyCheckNeeded()) {
3702 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3703 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3704 Accesses.getDependenciesToCheck());
3705 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3708 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3709 " need a runtime memory check.\n");
3714 static bool hasMultipleUsesOf(Instruction *I,
3715 SmallPtrSet<Instruction *, 8> &Insts) {
3716 unsigned NumUses = 0;
3717 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3718 if (Insts.count(dyn_cast<Instruction>(*Use)))
3727 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3728 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3729 if (!Set.count(dyn_cast<Instruction>(*Use)))
3734 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3735 ReductionKind Kind) {
3736 if (Phi->getNumIncomingValues() != 2)
3739 // Reduction variables are only found in the loop header block.
3740 if (Phi->getParent() != TheLoop->getHeader())
3743 // Obtain the reduction start value from the value that comes from the loop
3745 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3747 // ExitInstruction is the single value which is used outside the loop.
3748 // We only allow for a single reduction value to be used outside the loop.
3749 // This includes users of the reduction, variables (which form a cycle
3750 // which ends in the phi node).
3751 Instruction *ExitInstruction = 0;
3752 // Indicates that we found a reduction operation in our scan.
3753 bool FoundReduxOp = false;
3755 // We start with the PHI node and scan for all of the users of this
3756 // instruction. All users must be instructions that can be used as reduction
3757 // variables (such as ADD). We must have a single out-of-block user. The cycle
3758 // must include the original PHI.
3759 bool FoundStartPHI = false;
3761 // To recognize min/max patterns formed by a icmp select sequence, we store
3762 // the number of instruction we saw from the recognized min/max pattern,
3763 // to make sure we only see exactly the two instructions.
3764 unsigned NumCmpSelectPatternInst = 0;
3765 ReductionInstDesc ReduxDesc(false, 0);
3767 SmallPtrSet<Instruction *, 8> VisitedInsts;
3768 SmallVector<Instruction *, 8> Worklist;
3769 Worklist.push_back(Phi);
3770 VisitedInsts.insert(Phi);
3772 // A value in the reduction can be used:
3773 // - By the reduction:
3774 // - Reduction operation:
3775 // - One use of reduction value (safe).
3776 // - Multiple use of reduction value (not safe).
3778 // - All uses of the PHI must be the reduction (safe).
3779 // - Otherwise, not safe.
3780 // - By one instruction outside of the loop (safe).
3781 // - By further instructions outside of the loop (not safe).
3782 // - By an instruction that is not part of the reduction (not safe).
3784 // * An instruction type other than PHI or the reduction operation.
3785 // * A PHI in the header other than the initial PHI.
3786 while (!Worklist.empty()) {
3787 Instruction *Cur = Worklist.back();
3788 Worklist.pop_back();
3791 // If the instruction has no users then this is a broken chain and can't be
3792 // a reduction variable.
3793 if (Cur->use_empty())
3796 bool IsAPhi = isa<PHINode>(Cur);
3798 // A header PHI use other than the original PHI.
3799 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3802 // Reductions of instructions such as Div, and Sub is only possible if the
3803 // LHS is the reduction variable.
3804 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3805 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3806 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3809 // Any reduction instruction must be of one of the allowed kinds.
3810 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3811 if (!ReduxDesc.IsReduction)
3814 // A reduction operation must only have one use of the reduction value.
3815 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3816 hasMultipleUsesOf(Cur, VisitedInsts))
3819 // All inputs to a PHI node must be a reduction value.
3820 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3823 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3824 isa<SelectInst>(Cur)))
3825 ++NumCmpSelectPatternInst;
3826 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3827 isa<SelectInst>(Cur)))
3828 ++NumCmpSelectPatternInst;
3830 // Check whether we found a reduction operator.
3831 FoundReduxOp |= !IsAPhi;
3833 // Process users of current instruction. Push non PHI nodes after PHI nodes
3834 // onto the stack. This way we are going to have seen all inputs to PHI
3835 // nodes once we get to them.
3836 SmallVector<Instruction *, 8> NonPHIs;
3837 SmallVector<Instruction *, 8> PHIs;
3838 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3840 Instruction *Usr = cast<Instruction>(*UI);
3842 // Check if we found the exit user.
3843 BasicBlock *Parent = Usr->getParent();
3844 if (!TheLoop->contains(Parent)) {
3845 // Exit if you find multiple outside users or if the header phi node is
3846 // being used. In this case the user uses the value of the previous
3847 // iteration, in which case we would loose "VF-1" iterations of the
3848 // reduction operation if we vectorize.
3849 if (ExitInstruction != 0 || Cur == Phi)
3852 ExitInstruction = Cur;
3856 // Process instructions only once (termination).
3857 if (VisitedInsts.insert(Usr)) {
3858 if (isa<PHINode>(Usr))
3859 PHIs.push_back(Usr);
3861 NonPHIs.push_back(Usr);
3863 // Remember that we completed the cycle.
3865 FoundStartPHI = true;
3867 Worklist.append(PHIs.begin(), PHIs.end());
3868 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3871 // This means we have seen one but not the other instruction of the
3872 // pattern or more than just a select and cmp.
3873 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3874 NumCmpSelectPatternInst != 2)
3877 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3880 // We found a reduction var if we have reached the original phi node and we
3881 // only have a single instruction with out-of-loop users.
3883 // This instruction is allowed to have out-of-loop users.
3884 AllowedExit.insert(ExitInstruction);
3886 // Save the description of this reduction variable.
3887 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3888 ReduxDesc.MinMaxKind);
3889 Reductions[Phi] = RD;
3890 // We've ended the cycle. This is a reduction variable if we have an
3891 // outside user and it has a binary op.
3896 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3897 /// pattern corresponding to a min(X, Y) or max(X, Y).
3898 LoopVectorizationLegality::ReductionInstDesc
3899 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3900 ReductionInstDesc &Prev) {
3902 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3903 "Expect a select instruction");
3904 Instruction *Cmp = 0;
3905 SelectInst *Select = 0;
3907 // We must handle the select(cmp()) as a single instruction. Advance to the
3909 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3910 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3911 return ReductionInstDesc(false, I);
3912 return ReductionInstDesc(Select, Prev.MinMaxKind);
3915 // Only handle single use cases for now.
3916 if (!(Select = dyn_cast<SelectInst>(I)))
3917 return ReductionInstDesc(false, I);
3918 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3919 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3920 return ReductionInstDesc(false, I);
3921 if (!Cmp->hasOneUse())
3922 return ReductionInstDesc(false, I);
3927 // Look for a min/max pattern.
3928 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3929 return ReductionInstDesc(Select, MRK_UIntMin);
3930 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3931 return ReductionInstDesc(Select, MRK_UIntMax);
3932 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3933 return ReductionInstDesc(Select, MRK_SIntMax);
3934 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3935 return ReductionInstDesc(Select, MRK_SIntMin);
3936 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3937 return ReductionInstDesc(Select, MRK_FloatMin);
3938 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3939 return ReductionInstDesc(Select, MRK_FloatMax);
3940 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3941 return ReductionInstDesc(Select, MRK_FloatMin);
3942 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3943 return ReductionInstDesc(Select, MRK_FloatMax);
3945 return ReductionInstDesc(false, I);
3948 LoopVectorizationLegality::ReductionInstDesc
3949 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3951 ReductionInstDesc &Prev) {
3952 bool FP = I->getType()->isFloatingPointTy();
3953 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3954 switch (I->getOpcode()) {
3956 return ReductionInstDesc(false, I);
3957 case Instruction::PHI:
3958 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3959 Kind != RK_FloatMinMax))
3960 return ReductionInstDesc(false, I);
3961 return ReductionInstDesc(I, Prev.MinMaxKind);
3962 case Instruction::Sub:
3963 case Instruction::Add:
3964 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3965 case Instruction::Mul:
3966 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3967 case Instruction::And:
3968 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3969 case Instruction::Or:
3970 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3971 case Instruction::Xor:
3972 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3973 case Instruction::FMul:
3974 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3975 case Instruction::FAdd:
3976 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3977 case Instruction::FCmp:
3978 case Instruction::ICmp:
3979 case Instruction::Select:
3980 if (Kind != RK_IntegerMinMax &&
3981 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3982 return ReductionInstDesc(false, I);
3983 return isMinMaxSelectCmpPattern(I, Prev);
3987 LoopVectorizationLegality::InductionKind
3988 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3989 Type *PhiTy = Phi->getType();
3990 // We only handle integer and pointer inductions variables.
3991 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3992 return IK_NoInduction;
3994 // Check that the PHI is consecutive.
3995 const SCEV *PhiScev = SE->getSCEV(Phi);
3996 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3998 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3999 return IK_NoInduction;
4001 const SCEV *Step = AR->getStepRecurrence(*SE);
4003 // Integer inductions need to have a stride of one.
4004 if (PhiTy->isIntegerTy()) {
4006 return IK_IntInduction;
4007 if (Step->isAllOnesValue())
4008 return IK_ReverseIntInduction;
4009 return IK_NoInduction;
4012 // Calculate the pointer stride and check if it is consecutive.
4013 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4015 return IK_NoInduction;
4017 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4018 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4019 if (C->getValue()->equalsInt(Size))
4020 return IK_PtrInduction;
4021 else if (C->getValue()->equalsInt(0 - Size))
4022 return IK_ReversePtrInduction;
4024 return IK_NoInduction;
4027 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4028 Value *In0 = const_cast<Value*>(V);
4029 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4033 return Inductions.count(PN);
4036 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4037 assert(TheLoop->contains(BB) && "Unknown block used");
4039 // Blocks that do not dominate the latch need predication.
4040 BasicBlock* Latch = TheLoop->getLoopLatch();
4041 return !DT->dominates(BB, Latch);
4044 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4045 SmallPtrSet<Value *, 8>& SafePtrs) {
4046 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4047 // We might be able to hoist the load.
4048 if (it->mayReadFromMemory()) {
4049 LoadInst *LI = dyn_cast<LoadInst>(it);
4050 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4054 // We don't predicate stores at the moment.
4055 if (it->mayWriteToMemory() || it->mayThrow())
4058 // The instructions below can trap.
4059 switch (it->getOpcode()) {
4061 case Instruction::UDiv:
4062 case Instruction::SDiv:
4063 case Instruction::URem:
4064 case Instruction::SRem:
4072 LoopVectorizationCostModel::VectorizationFactor
4073 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4075 // Width 1 means no vectorize
4076 VectorizationFactor Factor = { 1U, 0U };
4077 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4078 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4082 // Find the trip count.
4083 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4084 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4086 unsigned WidestType = getWidestType();
4087 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4088 unsigned MaxSafeDepDist = -1U;
4089 if (Legal->getMaxSafeDepDistBytes() != -1U)
4090 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4091 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4092 WidestRegister : MaxSafeDepDist);
4093 unsigned MaxVectorSize = WidestRegister / WidestType;
4094 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4095 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4097 if (MaxVectorSize == 0) {
4098 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4102 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4103 " into one vector!");
4105 unsigned VF = MaxVectorSize;
4107 // If we optimize the program for size, avoid creating the tail loop.
4109 // If we are unable to calculate the trip count then don't try to vectorize.
4111 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4115 // Find the maximum SIMD width that can fit within the trip count.
4116 VF = TC % MaxVectorSize;
4121 // If the trip count that we found modulo the vectorization factor is not
4122 // zero then we require a tail.
4124 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4130 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4131 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4133 Factor.Width = UserVF;
4137 float Cost = expectedCost(1);
4139 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4140 for (unsigned i=2; i <= VF; i*=2) {
4141 // Notice that the vector loop needs to be executed less times, so
4142 // we need to divide the cost of the vector loops by the width of
4143 // the vector elements.
4144 float VectorCost = expectedCost(i) / (float)i;
4145 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4146 (int)VectorCost << ".\n");
4147 if (VectorCost < Cost) {
4153 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4154 Factor.Width = Width;
4155 Factor.Cost = Width * Cost;
4159 unsigned LoopVectorizationCostModel::getWidestType() {
4160 unsigned MaxWidth = 8;
4163 for (Loop::block_iterator bb = TheLoop->block_begin(),
4164 be = TheLoop->block_end(); bb != be; ++bb) {
4165 BasicBlock *BB = *bb;
4167 // For each instruction in the loop.
4168 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4169 Type *T = it->getType();
4171 // Only examine Loads, Stores and PHINodes.
4172 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4175 // Examine PHI nodes that are reduction variables.
4176 if (PHINode *PN = dyn_cast<PHINode>(it))
4177 if (!Legal->getReductionVars()->count(PN))
4180 // Examine the stored values.
4181 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4182 T = ST->getValueOperand()->getType();
4184 // Ignore loaded pointer types and stored pointer types that are not
4185 // consecutive. However, we do want to take consecutive stores/loads of
4186 // pointer vectors into account.
4187 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4190 MaxWidth = std::max(MaxWidth,
4191 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4199 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4202 unsigned LoopCost) {
4204 // -- The unroll heuristics --
4205 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4206 // There are many micro-architectural considerations that we can't predict
4207 // at this level. For example frontend pressure (on decode or fetch) due to
4208 // code size, or the number and capabilities of the execution ports.
4210 // We use the following heuristics to select the unroll factor:
4211 // 1. If the code has reductions the we unroll in order to break the cross
4212 // iteration dependency.
4213 // 2. If the loop is really small then we unroll in order to reduce the loop
4215 // 3. We don't unroll if we think that we will spill registers to memory due
4216 // to the increased register pressure.
4218 // Use the user preference, unless 'auto' is selected.
4222 // When we optimize for size we don't unroll.
4226 // We used the distance for the unroll factor.
4227 if (Legal->getMaxSafeDepDistBytes() != -1U)
4230 // Do not unroll loops with a relatively small trip count.
4231 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4232 TheLoop->getLoopLatch());
4233 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4236 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4237 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4238 " vector registers\n");
4240 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4241 // We divide by these constants so assume that we have at least one
4242 // instruction that uses at least one register.
4243 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4244 R.NumInstructions = std::max(R.NumInstructions, 1U);
4246 // We calculate the unroll factor using the following formula.
4247 // Subtract the number of loop invariants from the number of available
4248 // registers. These registers are used by all of the unrolled instances.
4249 // Next, divide the remaining registers by the number of registers that is
4250 // required by the loop, in order to estimate how many parallel instances
4251 // fit without causing spills.
4252 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4254 // Clamp the unroll factor ranges to reasonable factors.
4255 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4257 // If we did not calculate the cost for VF (because the user selected the VF)
4258 // then we calculate the cost of VF here.
4260 LoopCost = expectedCost(VF);
4262 // Clamp the calculated UF to be between the 1 and the max unroll factor
4263 // that the target allows.
4264 if (UF > MaxUnrollSize)
4269 if (Legal->getReductionVars()->size()) {
4270 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4274 // We want to unroll tiny loops in order to reduce the loop overhead.
4275 // We assume that the cost overhead is 1 and we use the cost model
4276 // to estimate the cost of the loop and unroll until the cost of the
4277 // loop overhead is about 5% of the cost of the loop.
4278 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4279 if (LoopCost < 20) {
4280 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4281 unsigned NewUF = 20/LoopCost + 1;
4282 return std::min(NewUF, UF);
4285 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4289 LoopVectorizationCostModel::RegisterUsage
4290 LoopVectorizationCostModel::calculateRegisterUsage() {
4291 // This function calculates the register usage by measuring the highest number
4292 // of values that are alive at a single location. Obviously, this is a very
4293 // rough estimation. We scan the loop in a topological order in order and
4294 // assign a number to each instruction. We use RPO to ensure that defs are
4295 // met before their users. We assume that each instruction that has in-loop
4296 // users starts an interval. We record every time that an in-loop value is
4297 // used, so we have a list of the first and last occurrences of each
4298 // instruction. Next, we transpose this data structure into a multi map that
4299 // holds the list of intervals that *end* at a specific location. This multi
4300 // map allows us to perform a linear search. We scan the instructions linearly
4301 // and record each time that a new interval starts, by placing it in a set.
4302 // If we find this value in the multi-map then we remove it from the set.
4303 // The max register usage is the maximum size of the set.
4304 // We also search for instructions that are defined outside the loop, but are
4305 // used inside the loop. We need this number separately from the max-interval
4306 // usage number because when we unroll, loop-invariant values do not take
4308 LoopBlocksDFS DFS(TheLoop);
4312 R.NumInstructions = 0;
4314 // Each 'key' in the map opens a new interval. The values
4315 // of the map are the index of the 'last seen' usage of the
4316 // instruction that is the key.
4317 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4318 // Maps instruction to its index.
4319 DenseMap<unsigned, Instruction*> IdxToInstr;
4320 // Marks the end of each interval.
4321 IntervalMap EndPoint;
4322 // Saves the list of instruction indices that are used in the loop.
4323 SmallSet<Instruction*, 8> Ends;
4324 // Saves the list of values that are used in the loop but are
4325 // defined outside the loop, such as arguments and constants.
4326 SmallPtrSet<Value*, 8> LoopInvariants;
4329 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4330 be = DFS.endRPO(); bb != be; ++bb) {
4331 R.NumInstructions += (*bb)->size();
4332 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4334 Instruction *I = it;
4335 IdxToInstr[Index++] = I;
4337 // Save the end location of each USE.
4338 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4339 Value *U = I->getOperand(i);
4340 Instruction *Instr = dyn_cast<Instruction>(U);
4342 // Ignore non-instruction values such as arguments, constants, etc.
4343 if (!Instr) continue;
4345 // If this instruction is outside the loop then record it and continue.
4346 if (!TheLoop->contains(Instr)) {
4347 LoopInvariants.insert(Instr);
4351 // Overwrite previous end points.
4352 EndPoint[Instr] = Index;
4358 // Saves the list of intervals that end with the index in 'key'.
4359 typedef SmallVector<Instruction*, 2> InstrList;
4360 DenseMap<unsigned, InstrList> TransposeEnds;
4362 // Transpose the EndPoints to a list of values that end at each index.
4363 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4365 TransposeEnds[it->second].push_back(it->first);
4367 SmallSet<Instruction*, 8> OpenIntervals;
4368 unsigned MaxUsage = 0;
4371 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4372 for (unsigned int i = 0; i < Index; ++i) {
4373 Instruction *I = IdxToInstr[i];
4374 // Ignore instructions that are never used within the loop.
4375 if (!Ends.count(I)) continue;
4377 // Remove all of the instructions that end at this location.
4378 InstrList &List = TransposeEnds[i];
4379 for (unsigned int j=0, e = List.size(); j < e; ++j)
4380 OpenIntervals.erase(List[j]);
4382 // Count the number of live interals.
4383 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4385 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4386 OpenIntervals.size() <<"\n");
4388 // Add the current instruction to the list of open intervals.
4389 OpenIntervals.insert(I);
4392 unsigned Invariant = LoopInvariants.size();
4393 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4394 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4395 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4397 R.LoopInvariantRegs = Invariant;
4398 R.MaxLocalUsers = MaxUsage;
4402 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4406 for (Loop::block_iterator bb = TheLoop->block_begin(),
4407 be = TheLoop->block_end(); bb != be; ++bb) {
4408 unsigned BlockCost = 0;
4409 BasicBlock *BB = *bb;
4411 // For each instruction in the old loop.
4412 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4413 // Skip dbg intrinsics.
4414 if (isa<DbgInfoIntrinsic>(it))
4417 unsigned C = getInstructionCost(it, VF);
4419 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4420 VF << " For instruction: "<< *it << "\n");
4423 // We assume that if-converted blocks have a 50% chance of being executed.
4424 // When the code is scalar then some of the blocks are avoided due to CF.
4425 // When the code is vectorized we execute all code paths.
4426 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4435 /// \brief Check whether the address computation for a non-consecutive memory
4436 /// access looks like an unlikely candidate for being merged into the indexing
4439 /// We look for a GEP which has one index that is an induction variable and all
4440 /// other indices are loop invariant. If the stride of this access is also
4441 /// within a small bound we decide that this address computation can likely be
4442 /// merged into the addressing mode.
4443 /// In all other cases, we identify the address computation as complex.
4444 static bool isLikelyComplexAddressComputation(Value *Ptr,
4445 LoopVectorizationLegality *Legal,
4446 ScalarEvolution *SE,
4447 const Loop *TheLoop) {
4448 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4452 // We are looking for a gep with all loop invariant indices except for one
4453 // which should be an induction variable.
4454 unsigned NumOperands = Gep->getNumOperands();
4455 for (unsigned i = 1; i < NumOperands; ++i) {
4456 Value *Opd = Gep->getOperand(i);
4457 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4458 !Legal->isInductionVariable(Opd))
4462 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4463 // can likely be merged into the address computation.
4464 unsigned MaxMergeDistance = 64;
4466 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4470 // Check the step is constant.
4471 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4472 // Calculate the pointer stride and check if it is consecutive.
4473 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4477 const APInt &APStepVal = C->getValue()->getValue();
4479 // Huge step value - give up.
4480 if (APStepVal.getBitWidth() > 64)
4483 int64_t StepVal = APStepVal.getSExtValue();
4485 return StepVal > MaxMergeDistance;
4489 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4490 // If we know that this instruction will remain uniform, check the cost of
4491 // the scalar version.
4492 if (Legal->isUniformAfterVectorization(I))
4495 Type *RetTy = I->getType();
4496 Type *VectorTy = ToVectorTy(RetTy, VF);
4498 // TODO: We need to estimate the cost of intrinsic calls.
4499 switch (I->getOpcode()) {
4500 case Instruction::GetElementPtr:
4501 // We mark this instruction as zero-cost because the cost of GEPs in
4502 // vectorized code depends on whether the corresponding memory instruction
4503 // is scalarized or not. Therefore, we handle GEPs with the memory
4504 // instruction cost.
4506 case Instruction::Br: {
4507 return TTI.getCFInstrCost(I->getOpcode());
4509 case Instruction::PHI:
4510 //TODO: IF-converted IFs become selects.
4512 case Instruction::Add:
4513 case Instruction::FAdd:
4514 case Instruction::Sub:
4515 case Instruction::FSub:
4516 case Instruction::Mul:
4517 case Instruction::FMul:
4518 case Instruction::UDiv:
4519 case Instruction::SDiv:
4520 case Instruction::FDiv:
4521 case Instruction::URem:
4522 case Instruction::SRem:
4523 case Instruction::FRem:
4524 case Instruction::Shl:
4525 case Instruction::LShr:
4526 case Instruction::AShr:
4527 case Instruction::And:
4528 case Instruction::Or:
4529 case Instruction::Xor: {
4530 // Certain instructions can be cheaper to vectorize if they have a constant
4531 // second vector operand. One example of this are shifts on x86.
4532 TargetTransformInfo::OperandValueKind Op1VK =
4533 TargetTransformInfo::OK_AnyValue;
4534 TargetTransformInfo::OperandValueKind Op2VK =
4535 TargetTransformInfo::OK_AnyValue;
4537 if (isa<ConstantInt>(I->getOperand(1)))
4538 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4540 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4542 case Instruction::Select: {
4543 SelectInst *SI = cast<SelectInst>(I);
4544 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4545 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4546 Type *CondTy = SI->getCondition()->getType();
4548 CondTy = VectorType::get(CondTy, VF);
4550 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4552 case Instruction::ICmp:
4553 case Instruction::FCmp: {
4554 Type *ValTy = I->getOperand(0)->getType();
4555 VectorTy = ToVectorTy(ValTy, VF);
4556 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4558 case Instruction::Store:
4559 case Instruction::Load: {
4560 StoreInst *SI = dyn_cast<StoreInst>(I);
4561 LoadInst *LI = dyn_cast<LoadInst>(I);
4562 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4564 VectorTy = ToVectorTy(ValTy, VF);
4566 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4567 unsigned AS = SI ? SI->getPointerAddressSpace() :
4568 LI->getPointerAddressSpace();
4569 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4570 // We add the cost of address computation here instead of with the gep
4571 // instruction because only here we know whether the operation is
4574 return TTI.getAddressComputationCost(VectorTy) +
4575 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4577 // Scalarized loads/stores.
4578 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4579 bool Reverse = ConsecutiveStride < 0;
4580 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4581 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4582 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4583 bool IsComplexComputation =
4584 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4586 // The cost of extracting from the value vector and pointer vector.
4587 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4588 for (unsigned i = 0; i < VF; ++i) {
4589 // The cost of extracting the pointer operand.
4590 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4591 // In case of STORE, the cost of ExtractElement from the vector.
4592 // In case of LOAD, the cost of InsertElement into the returned
4594 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4595 Instruction::InsertElement,
4599 // The cost of the scalar loads/stores.
4600 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4601 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4606 // Wide load/stores.
4607 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4608 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4611 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4615 case Instruction::ZExt:
4616 case Instruction::SExt:
4617 case Instruction::FPToUI:
4618 case Instruction::FPToSI:
4619 case Instruction::FPExt:
4620 case Instruction::PtrToInt:
4621 case Instruction::IntToPtr:
4622 case Instruction::SIToFP:
4623 case Instruction::UIToFP:
4624 case Instruction::Trunc:
4625 case Instruction::FPTrunc:
4626 case Instruction::BitCast: {
4627 // We optimize the truncation of induction variable.
4628 // The cost of these is the same as the scalar operation.
4629 if (I->getOpcode() == Instruction::Trunc &&
4630 Legal->isInductionVariable(I->getOperand(0)))
4631 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4632 I->getOperand(0)->getType());
4634 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4635 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4637 case Instruction::Call: {
4638 CallInst *CI = cast<CallInst>(I);
4639 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4640 assert(ID && "Not an intrinsic call!");
4641 Type *RetTy = ToVectorTy(CI->getType(), VF);
4642 SmallVector<Type*, 4> Tys;
4643 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4644 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4645 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4648 // We are scalarizing the instruction. Return the cost of the scalar
4649 // instruction, plus the cost of insert and extract into vector
4650 // elements, times the vector width.
4653 if (!RetTy->isVoidTy() && VF != 1) {
4654 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4656 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4659 // The cost of inserting the results plus extracting each one of the
4661 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4664 // The cost of executing VF copies of the scalar instruction. This opcode
4665 // is unknown. Assume that it is the same as 'mul'.
4666 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4672 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4673 if (Scalar->isVoidTy() || VF == 1)
4675 return VectorType::get(Scalar, VF);
4678 char LoopVectorize::ID = 0;
4679 static const char lv_name[] = "Loop Vectorization";
4680 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4681 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4682 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4683 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4684 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4687 Pass *createLoopVectorizePass() {
4688 return new LoopVectorize();
4692 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4693 // Check for a store.
4694 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4695 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4697 // Check for a load.
4698 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4699 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;