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, DataTy->getPointerTo(AddressSpace));
1224 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1230 assert(LI && "Must have a load instruction");
1231 setDebugLocFromInst(Builder, LI);
1232 for (unsigned Part = 0; Part < UF; ++Part) {
1233 // Calculate the pointer for the specific unroll-part.
1234 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1237 // If the address is consecutive but reversed, then the
1238 // wide store needs to start at the last vector element.
1239 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1240 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1243 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1244 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1245 cast<LoadInst>(LI)->setAlignment(Alignment);
1246 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1250 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1251 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1252 // Holds vector parameters or scalars, in case of uniform vals.
1253 SmallVector<VectorParts, 4> Params;
1255 setDebugLocFromInst(Builder, Instr);
1257 // Find all of the vectorized parameters.
1258 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1259 Value *SrcOp = Instr->getOperand(op);
1261 // If we are accessing the old induction variable, use the new one.
1262 if (SrcOp == OldInduction) {
1263 Params.push_back(getVectorValue(SrcOp));
1267 // Try using previously calculated values.
1268 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1270 // If the src is an instruction that appeared earlier in the basic block
1271 // then it should already be vectorized.
1272 if (SrcInst && OrigLoop->contains(SrcInst)) {
1273 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1274 // The parameter is a vector value from earlier.
1275 Params.push_back(WidenMap.get(SrcInst));
1277 // The parameter is a scalar from outside the loop. Maybe even a constant.
1278 VectorParts Scalars;
1279 Scalars.append(UF, SrcOp);
1280 Params.push_back(Scalars);
1284 assert(Params.size() == Instr->getNumOperands() &&
1285 "Invalid number of operands");
1287 // Does this instruction return a value ?
1288 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1290 Value *UndefVec = IsVoidRetTy ? 0 :
1291 UndefValue::get(VectorType::get(Instr->getType(), VF));
1292 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1293 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1295 // For each vector unroll 'part':
1296 for (unsigned Part = 0; Part < UF; ++Part) {
1297 // For each scalar that we create:
1298 for (unsigned Width = 0; Width < VF; ++Width) {
1299 Instruction *Cloned = Instr->clone();
1301 Cloned->setName(Instr->getName() + ".cloned");
1302 // Replace the operands of the cloned instrucions with extracted scalars.
1303 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1304 Value *Op = Params[op][Part];
1305 // Param is a vector. Need to extract the right lane.
1306 if (Op->getType()->isVectorTy())
1307 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1308 Cloned->setOperand(op, Op);
1311 // Place the cloned scalar in the new loop.
1312 Builder.Insert(Cloned);
1314 // If the original scalar returns a value we need to place it in a vector
1315 // so that future users will be able to use it.
1317 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1318 Builder.getInt32(Width));
1324 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1326 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1327 Legal->getRuntimePointerCheck();
1329 if (!PtrRtCheck->Need)
1332 unsigned NumPointers = PtrRtCheck->Pointers.size();
1333 SmallVector<TrackingVH<Value> , 2> Starts;
1334 SmallVector<TrackingVH<Value> , 2> Ends;
1336 SCEVExpander Exp(*SE, "induction");
1338 // Use this type for pointer arithmetic.
1339 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1341 for (unsigned i = 0; i < NumPointers; ++i) {
1342 Value *Ptr = PtrRtCheck->Pointers[i];
1343 const SCEV *Sc = SE->getSCEV(Ptr);
1345 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1346 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1348 Starts.push_back(Ptr);
1349 Ends.push_back(Ptr);
1351 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1353 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1354 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1355 Starts.push_back(Start);
1356 Ends.push_back(End);
1360 IRBuilder<> ChkBuilder(Loc);
1361 // Our instructions might fold to a constant.
1362 Value *MemoryRuntimeCheck = 0;
1363 for (unsigned i = 0; i < NumPointers; ++i) {
1364 for (unsigned j = i+1; j < NumPointers; ++j) {
1365 // No need to check if two readonly pointers intersect.
1366 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1369 // Only need to check pointers between two different dependency sets.
1370 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1373 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1374 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1375 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1376 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1378 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1379 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1380 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1381 if (MemoryRuntimeCheck)
1382 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1384 MemoryRuntimeCheck = IsConflict;
1388 // We have to do this trickery because the IRBuilder might fold the check to a
1389 // constant expression in which case there is no Instruction anchored in a
1391 LLVMContext &Ctx = Loc->getContext();
1392 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1393 ConstantInt::getTrue(Ctx));
1394 ChkBuilder.Insert(Check, "memcheck.conflict");
1399 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1401 In this function we generate a new loop. The new loop will contain
1402 the vectorized instructions while the old loop will continue to run the
1405 [ ] <-- vector loop bypass (may consist of multiple blocks).
1408 | [ ] <-- vector pre header.
1412 | [ ]_| <-- vector loop.
1415 >[ ] <--- middle-block.
1418 | [ ] <--- new preheader.
1422 | [ ]_| <-- old scalar loop to handle remainder.
1425 >[ ] <-- exit block.
1429 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1430 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1431 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1432 assert(ExitBlock && "Must have an exit block");
1434 // Some loops have a single integer induction variable, while other loops
1435 // don't. One example is c++ iterators that often have multiple pointer
1436 // induction variables. In the code below we also support a case where we
1437 // don't have a single induction variable.
1438 OldInduction = Legal->getInduction();
1439 Type *IdxTy = Legal->getWidestInductionType();
1441 // Find the loop boundaries.
1442 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1443 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1445 // Get the total trip count from the count by adding 1.
1446 ExitCount = SE->getAddExpr(ExitCount,
1447 SE->getConstant(ExitCount->getType(), 1));
1449 // Expand the trip count and place the new instructions in the preheader.
1450 // Notice that the pre-header does not change, only the loop body.
1451 SCEVExpander Exp(*SE, "induction");
1453 // Count holds the overall loop count (N).
1454 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1455 BypassBlock->getTerminator());
1457 // The loop index does not have to start at Zero. Find the original start
1458 // value from the induction PHI node. If we don't have an induction variable
1459 // then we know that it starts at zero.
1460 Builder.SetInsertPoint(BypassBlock->getTerminator());
1461 Value *StartIdx = ExtendedIdx = OldInduction ?
1462 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1464 ConstantInt::get(IdxTy, 0);
1466 assert(BypassBlock && "Invalid loop structure");
1467 LoopBypassBlocks.push_back(BypassBlock);
1469 // Split the single block loop into the two loop structure described above.
1470 BasicBlock *VectorPH =
1471 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1472 BasicBlock *VecBody =
1473 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1474 BasicBlock *MiddleBlock =
1475 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1476 BasicBlock *ScalarPH =
1477 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1479 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1481 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1483 // Generate the induction variable.
1484 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1485 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1486 // The loop step is equal to the vectorization factor (num of SIMD elements)
1487 // times the unroll factor (num of SIMD instructions).
1488 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1490 // This is the IR builder that we use to add all of the logic for bypassing
1491 // the new vector loop.
1492 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1493 setDebugLocFromInst(BypassBuilder,
1494 getDebugLocFromInstOrOperands(OldInduction));
1496 // We may need to extend the index in case there is a type mismatch.
1497 // We know that the count starts at zero and does not overflow.
1498 if (Count->getType() != IdxTy) {
1499 // The exit count can be of pointer type. Convert it to the correct
1501 if (ExitCount->getType()->isPointerTy())
1502 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1504 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1507 // Add the start index to the loop count to get the new end index.
1508 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1510 // Now we need to generate the expression for N - (N % VF), which is
1511 // the part that the vectorized body will execute.
1512 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1513 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1514 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1515 "end.idx.rnd.down");
1517 // Now, compare the new count to zero. If it is zero skip the vector loop and
1518 // jump to the scalar loop.
1519 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1522 BasicBlock *LastBypassBlock = BypassBlock;
1524 // Generate the code that checks in runtime if arrays overlap. We put the
1525 // checks into a separate block to make the more common case of few elements
1527 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1528 BypassBlock->getTerminator());
1529 if (MemRuntimeCheck) {
1530 // Create a new block containing the memory check.
1531 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1533 LoopBypassBlocks.push_back(CheckBlock);
1535 // Replace the branch into the memory check block with a conditional branch
1536 // for the "few elements case".
1537 Instruction *OldTerm = BypassBlock->getTerminator();
1538 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1539 OldTerm->eraseFromParent();
1541 Cmp = MemRuntimeCheck;
1542 LastBypassBlock = CheckBlock;
1545 LastBypassBlock->getTerminator()->eraseFromParent();
1546 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1549 // We are going to resume the execution of the scalar loop.
1550 // Go over all of the induction variables that we found and fix the
1551 // PHIs that are left in the scalar version of the loop.
1552 // The starting values of PHI nodes depend on the counter of the last
1553 // iteration in the vectorized loop.
1554 // If we come from a bypass edge then we need to start from the original
1557 // This variable saves the new starting index for the scalar loop.
1558 PHINode *ResumeIndex = 0;
1559 LoopVectorizationLegality::InductionList::iterator I, E;
1560 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1561 // Set builder to point to last bypass block.
1562 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1563 for (I = List->begin(), E = List->end(); I != E; ++I) {
1564 PHINode *OrigPhi = I->first;
1565 LoopVectorizationLegality::InductionInfo II = I->second;
1567 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1568 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1569 MiddleBlock->getTerminator());
1570 // We might have extended the type of the induction variable but we need a
1571 // truncated version for the scalar loop.
1572 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1573 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1574 MiddleBlock->getTerminator()) : 0;
1576 Value *EndValue = 0;
1578 case LoopVectorizationLegality::IK_NoInduction:
1579 llvm_unreachable("Unknown induction");
1580 case LoopVectorizationLegality::IK_IntInduction: {
1581 // Handle the integer induction counter.
1582 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1584 // We have the canonical induction variable.
1585 if (OrigPhi == OldInduction) {
1586 // Create a truncated version of the resume value for the scalar loop,
1587 // we might have promoted the type to a larger width.
1589 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1590 // The new PHI merges the original incoming value, in case of a bypass,
1591 // or the value at the end of the vectorized loop.
1592 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1593 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1594 TruncResumeVal->addIncoming(EndValue, VecBody);
1596 // We know what the end value is.
1597 EndValue = IdxEndRoundDown;
1598 // We also know which PHI node holds it.
1599 ResumeIndex = ResumeVal;
1603 // Not the canonical induction variable - add the vector loop count to the
1605 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1606 II.StartValue->getType(),
1608 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1611 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1612 // Convert the CountRoundDown variable to the PHI size.
1613 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1614 II.StartValue->getType(),
1616 // Handle reverse integer induction counter.
1617 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1620 case LoopVectorizationLegality::IK_PtrInduction: {
1621 // For pointer induction variables, calculate the offset using
1623 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1627 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1628 // The value at the end of the loop for the reverse pointer is calculated
1629 // by creating a GEP with a negative index starting from the start value.
1630 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1631 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1633 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1639 // The new PHI merges the original incoming value, in case of a bypass,
1640 // or the value at the end of the vectorized loop.
1641 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1642 if (OrigPhi == OldInduction)
1643 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1645 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1647 ResumeVal->addIncoming(EndValue, VecBody);
1649 // Fix the scalar body counter (PHI node).
1650 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1651 // The old inductions phi node in the scalar body needs the truncated value.
1652 if (OrigPhi == OldInduction)
1653 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1655 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1658 // If we are generating a new induction variable then we also need to
1659 // generate the code that calculates the exit value. This value is not
1660 // simply the end of the counter because we may skip the vectorized body
1661 // in case of a runtime check.
1663 assert(!ResumeIndex && "Unexpected resume value found");
1664 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1665 MiddleBlock->getTerminator());
1666 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1667 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1668 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1671 // Make sure that we found the index where scalar loop needs to continue.
1672 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1673 "Invalid resume Index");
1675 // Add a check in the middle block to see if we have completed
1676 // all of the iterations in the first vector loop.
1677 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1678 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1679 ResumeIndex, "cmp.n",
1680 MiddleBlock->getTerminator());
1682 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1683 // Remove the old terminator.
1684 MiddleBlock->getTerminator()->eraseFromParent();
1686 // Create i+1 and fill the PHINode.
1687 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1688 Induction->addIncoming(StartIdx, VectorPH);
1689 Induction->addIncoming(NextIdx, VecBody);
1690 // Create the compare.
1691 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1692 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1694 // Now we have two terminators. Remove the old one from the block.
1695 VecBody->getTerminator()->eraseFromParent();
1697 // Get ready to start creating new instructions into the vectorized body.
1698 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1700 // Create and register the new vector loop.
1701 Loop* Lp = new Loop();
1702 Loop *ParentLoop = OrigLoop->getParentLoop();
1704 // Insert the new loop into the loop nest and register the new basic blocks.
1706 ParentLoop->addChildLoop(Lp);
1707 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1708 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1709 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1710 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1711 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1713 LI->addTopLevelLoop(Lp);
1716 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1719 LoopVectorPreHeader = VectorPH;
1720 LoopScalarPreHeader = ScalarPH;
1721 LoopMiddleBlock = MiddleBlock;
1722 LoopExitBlock = ExitBlock;
1723 LoopVectorBody = VecBody;
1724 LoopScalarBody = OldBasicBlock;
1727 /// This function returns the identity element (or neutral element) for
1728 /// the operation K.
1730 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1735 // Adding, Xoring, Oring zero to a number does not change it.
1736 return ConstantInt::get(Tp, 0);
1737 case RK_IntegerMult:
1738 // Multiplying a number by 1 does not change it.
1739 return ConstantInt::get(Tp, 1);
1741 // AND-ing a number with an all-1 value does not change it.
1742 return ConstantInt::get(Tp, -1, true);
1744 // Multiplying a number by 1 does not change it.
1745 return ConstantFP::get(Tp, 1.0L);
1747 // Adding zero to a number does not change it.
1748 return ConstantFP::get(Tp, 0.0L);
1750 llvm_unreachable("Unknown reduction kind");
1754 static Intrinsic::ID
1755 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1756 // If we have an intrinsic call, check if it is trivially vectorizable.
1757 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1758 switch (II->getIntrinsicID()) {
1759 case Intrinsic::sqrt:
1760 case Intrinsic::sin:
1761 case Intrinsic::cos:
1762 case Intrinsic::exp:
1763 case Intrinsic::exp2:
1764 case Intrinsic::log:
1765 case Intrinsic::log10:
1766 case Intrinsic::log2:
1767 case Intrinsic::fabs:
1768 case Intrinsic::floor:
1769 case Intrinsic::ceil:
1770 case Intrinsic::trunc:
1771 case Intrinsic::rint:
1772 case Intrinsic::nearbyint:
1773 case Intrinsic::pow:
1774 case Intrinsic::fma:
1775 case Intrinsic::fmuladd:
1776 return II->getIntrinsicID();
1778 return Intrinsic::not_intrinsic;
1783 return Intrinsic::not_intrinsic;
1786 Function *F = CI->getCalledFunction();
1787 // We're going to make assumptions on the semantics of the functions, check
1788 // that the target knows that it's available in this environment.
1789 if (!F || !TLI->getLibFunc(F->getName(), Func))
1790 return Intrinsic::not_intrinsic;
1792 // Otherwise check if we have a call to a function that can be turned into a
1793 // vector intrinsic.
1800 return Intrinsic::sin;
1804 return Intrinsic::cos;
1808 return Intrinsic::exp;
1810 case LibFunc::exp2f:
1811 case LibFunc::exp2l:
1812 return Intrinsic::exp2;
1816 return Intrinsic::log;
1817 case LibFunc::log10:
1818 case LibFunc::log10f:
1819 case LibFunc::log10l:
1820 return Intrinsic::log10;
1822 case LibFunc::log2f:
1823 case LibFunc::log2l:
1824 return Intrinsic::log2;
1826 case LibFunc::fabsf:
1827 case LibFunc::fabsl:
1828 return Intrinsic::fabs;
1829 case LibFunc::floor:
1830 case LibFunc::floorf:
1831 case LibFunc::floorl:
1832 return Intrinsic::floor;
1834 case LibFunc::ceilf:
1835 case LibFunc::ceill:
1836 return Intrinsic::ceil;
1837 case LibFunc::trunc:
1838 case LibFunc::truncf:
1839 case LibFunc::truncl:
1840 return Intrinsic::trunc;
1842 case LibFunc::rintf:
1843 case LibFunc::rintl:
1844 return Intrinsic::rint;
1845 case LibFunc::nearbyint:
1846 case LibFunc::nearbyintf:
1847 case LibFunc::nearbyintl:
1848 return Intrinsic::nearbyint;
1852 return Intrinsic::pow;
1855 return Intrinsic::not_intrinsic;
1858 /// This function translates the reduction kind to an LLVM binary operator.
1860 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1862 case LoopVectorizationLegality::RK_IntegerAdd:
1863 return Instruction::Add;
1864 case LoopVectorizationLegality::RK_IntegerMult:
1865 return Instruction::Mul;
1866 case LoopVectorizationLegality::RK_IntegerOr:
1867 return Instruction::Or;
1868 case LoopVectorizationLegality::RK_IntegerAnd:
1869 return Instruction::And;
1870 case LoopVectorizationLegality::RK_IntegerXor:
1871 return Instruction::Xor;
1872 case LoopVectorizationLegality::RK_FloatMult:
1873 return Instruction::FMul;
1874 case LoopVectorizationLegality::RK_FloatAdd:
1875 return Instruction::FAdd;
1876 case LoopVectorizationLegality::RK_IntegerMinMax:
1877 return Instruction::ICmp;
1878 case LoopVectorizationLegality::RK_FloatMinMax:
1879 return Instruction::FCmp;
1881 llvm_unreachable("Unknown reduction operation");
1885 Value *createMinMaxOp(IRBuilder<> &Builder,
1886 LoopVectorizationLegality::MinMaxReductionKind RK,
1889 CmpInst::Predicate P = CmpInst::ICMP_NE;
1892 llvm_unreachable("Unknown min/max reduction kind");
1893 case LoopVectorizationLegality::MRK_UIntMin:
1894 P = CmpInst::ICMP_ULT;
1896 case LoopVectorizationLegality::MRK_UIntMax:
1897 P = CmpInst::ICMP_UGT;
1899 case LoopVectorizationLegality::MRK_SIntMin:
1900 P = CmpInst::ICMP_SLT;
1902 case LoopVectorizationLegality::MRK_SIntMax:
1903 P = CmpInst::ICMP_SGT;
1905 case LoopVectorizationLegality::MRK_FloatMin:
1906 P = CmpInst::FCMP_OLT;
1908 case LoopVectorizationLegality::MRK_FloatMax:
1909 P = CmpInst::FCMP_OGT;
1914 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1915 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1917 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1919 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1924 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1925 //===------------------------------------------------===//
1927 // Notice: any optimization or new instruction that go
1928 // into the code below should be also be implemented in
1931 //===------------------------------------------------===//
1932 Constant *Zero = Builder.getInt32(0);
1934 // In order to support reduction variables we need to be able to vectorize
1935 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1936 // stages. First, we create a new vector PHI node with no incoming edges.
1937 // We use this value when we vectorize all of the instructions that use the
1938 // PHI. Next, after all of the instructions in the block are complete we
1939 // add the new incoming edges to the PHI. At this point all of the
1940 // instructions in the basic block are vectorized, so we can use them to
1941 // construct the PHI.
1942 PhiVector RdxPHIsToFix;
1944 // Scan the loop in a topological order to ensure that defs are vectorized
1946 LoopBlocksDFS DFS(OrigLoop);
1949 // Vectorize all of the blocks in the original loop.
1950 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1951 be = DFS.endRPO(); bb != be; ++bb)
1952 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1954 // At this point every instruction in the original loop is widened to
1955 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1956 // that we vectorized. The PHI nodes are currently empty because we did
1957 // not want to introduce cycles. Notice that the remaining PHI nodes
1958 // that we need to fix are reduction variables.
1960 // Create the 'reduced' values for each of the induction vars.
1961 // The reduced values are the vector values that we scalarize and combine
1962 // after the loop is finished.
1963 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1965 PHINode *RdxPhi = *it;
1966 assert(RdxPhi && "Unable to recover vectorized PHI");
1968 // Find the reduction variable descriptor.
1969 assert(Legal->getReductionVars()->count(RdxPhi) &&
1970 "Unable to find the reduction variable");
1971 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1972 (*Legal->getReductionVars())[RdxPhi];
1974 setDebugLocFromInst(Builder, RdxDesc.StartValue);
1976 // We need to generate a reduction vector from the incoming scalar.
1977 // To do so, we need to generate the 'identity' vector and overide
1978 // one of the elements with the incoming scalar reduction. We need
1979 // to do it in the vector-loop preheader.
1980 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1982 // This is the vector-clone of the value that leaves the loop.
1983 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1984 Type *VecTy = VectorExit[0]->getType();
1986 // Find the reduction identity variable. Zero for addition, or, xor,
1987 // one for multiplication, -1 for And.
1990 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
1991 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
1992 // MinMax reduction have the start value as their identify.
1993 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
1997 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1998 VecTy->getScalarType());
1999 Identity = ConstantVector::getSplat(VF, Iden);
2001 // This vector is the Identity vector where the first element is the
2002 // incoming scalar reduction.
2003 VectorStart = Builder.CreateInsertElement(Identity,
2004 RdxDesc.StartValue, Zero);
2007 // Fix the vector-loop phi.
2008 // We created the induction variable so we know that the
2009 // preheader is the first entry.
2010 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2012 // Reductions do not have to start at zero. They can start with
2013 // any loop invariant values.
2014 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2015 BasicBlock *Latch = OrigLoop->getLoopLatch();
2016 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2017 VectorParts &Val = getVectorValue(LoopVal);
2018 for (unsigned part = 0; part < UF; ++part) {
2019 // Make sure to add the reduction stat value only to the
2020 // first unroll part.
2021 Value *StartVal = (part == 0) ? VectorStart : Identity;
2022 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2023 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2026 // Before each round, move the insertion point right between
2027 // the PHIs and the values we are going to write.
2028 // This allows us to write both PHINodes and the extractelement
2030 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2032 VectorParts RdxParts;
2033 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2034 for (unsigned part = 0; part < UF; ++part) {
2035 // This PHINode contains the vectorized reduction variable, or
2036 // the initial value vector, if we bypass the vector loop.
2037 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2038 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2039 Value *StartVal = (part == 0) ? VectorStart : Identity;
2040 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2041 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2042 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2043 RdxParts.push_back(NewPhi);
2046 // Reduce all of the unrolled parts into a single vector.
2047 Value *ReducedPartRdx = RdxParts[0];
2048 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2049 setDebugLocFromInst(Builder, ReducedPartRdx);
2050 for (unsigned part = 1; part < UF; ++part) {
2051 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2052 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2053 RdxParts[part], ReducedPartRdx,
2056 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2057 ReducedPartRdx, RdxParts[part]);
2060 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2061 // and vector ops, reducing the set of values being computed by half each
2063 assert(isPowerOf2_32(VF) &&
2064 "Reduction emission only supported for pow2 vectors!");
2065 Value *TmpVec = ReducedPartRdx;
2066 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2067 for (unsigned i = VF; i != 1; i >>= 1) {
2068 // Move the upper half of the vector to the lower half.
2069 for (unsigned j = 0; j != i/2; ++j)
2070 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2072 // Fill the rest of the mask with undef.
2073 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2074 UndefValue::get(Builder.getInt32Ty()));
2077 Builder.CreateShuffleVector(TmpVec,
2078 UndefValue::get(TmpVec->getType()),
2079 ConstantVector::get(ShuffleMask),
2082 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2083 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2086 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2089 // The result is in the first element of the vector.
2090 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2092 // Now, we need to fix the users of the reduction variable
2093 // inside and outside of the scalar remainder loop.
2094 // We know that the loop is in LCSSA form. We need to update the
2095 // PHI nodes in the exit blocks.
2096 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2097 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2098 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2099 if (!LCSSAPhi) continue;
2101 // All PHINodes need to have a single entry edge, or two if
2102 // we already fixed them.
2103 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2105 // We found our reduction value exit-PHI. Update it with the
2106 // incoming bypass edge.
2107 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2108 // Add an edge coming from the bypass.
2109 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2112 }// end of the LCSSA phi scan.
2114 // Fix the scalar loop reduction variable with the incoming reduction sum
2115 // from the vector body and from the backedge value.
2116 int IncomingEdgeBlockIdx =
2117 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2118 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2119 // Pick the other block.
2120 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2121 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2122 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2123 }// end of for each redux variable.
2125 // The Loop exit block may have single value PHI nodes where the incoming
2126 // value is 'undef'. While vectorizing we only handled real values that
2127 // were defined inside the loop. Here we handle the 'undef case'.
2129 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2130 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2131 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2132 if (!LCSSAPhi) continue;
2133 if (LCSSAPhi->getNumIncomingValues() == 1)
2134 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2139 InnerLoopVectorizer::VectorParts
2140 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2141 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2144 // Look for cached value.
2145 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2146 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2147 if (ECEntryIt != MaskCache.end())
2148 return ECEntryIt->second;
2150 VectorParts SrcMask = createBlockInMask(Src);
2152 // The terminator has to be a branch inst!
2153 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2154 assert(BI && "Unexpected terminator found");
2156 if (BI->isConditional()) {
2157 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2159 if (BI->getSuccessor(0) != Dst)
2160 for (unsigned part = 0; part < UF; ++part)
2161 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2163 for (unsigned part = 0; part < UF; ++part)
2164 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2166 MaskCache[Edge] = EdgeMask;
2170 MaskCache[Edge] = SrcMask;
2174 InnerLoopVectorizer::VectorParts
2175 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2176 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2178 // Loop incoming mask is all-one.
2179 if (OrigLoop->getHeader() == BB) {
2180 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2181 return getVectorValue(C);
2184 // This is the block mask. We OR all incoming edges, and with zero.
2185 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2186 VectorParts BlockMask = getVectorValue(Zero);
2189 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2190 VectorParts EM = createEdgeMask(*it, BB);
2191 for (unsigned part = 0; part < UF; ++part)
2192 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2199 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2200 BasicBlock *BB, PhiVector *PV) {
2201 // For each instruction in the old loop.
2202 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2203 VectorParts &Entry = WidenMap.get(it);
2204 switch (it->getOpcode()) {
2205 case Instruction::Br:
2206 // Nothing to do for PHIs and BR, since we already took care of the
2207 // loop control flow instructions.
2209 case Instruction::PHI:{
2210 PHINode* P = cast<PHINode>(it);
2211 // Handle reduction variables:
2212 if (Legal->getReductionVars()->count(P)) {
2213 for (unsigned part = 0; part < UF; ++part) {
2214 // This is phase one of vectorizing PHIs.
2215 Type *VecTy = VectorType::get(it->getType(), VF);
2216 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2217 LoopVectorBody-> getFirstInsertionPt());
2223 setDebugLocFromInst(Builder, P);
2224 // Check for PHI nodes that are lowered to vector selects.
2225 if (P->getParent() != OrigLoop->getHeader()) {
2226 // We know that all PHIs in non header blocks are converted into
2227 // selects, so we don't have to worry about the insertion order and we
2228 // can just use the builder.
2229 // At this point we generate the predication tree. There may be
2230 // duplications since this is a simple recursive scan, but future
2231 // optimizations will clean it up.
2233 unsigned NumIncoming = P->getNumIncomingValues();
2235 // Generate a sequence of selects of the form:
2236 // SELECT(Mask3, In3,
2237 // SELECT(Mask2, In2,
2239 for (unsigned In = 0; In < NumIncoming; In++) {
2240 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2242 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2244 for (unsigned part = 0; part < UF; ++part) {
2245 // We might have single edge PHIs (blocks) - use an identity
2246 // 'select' for the first PHI operand.
2248 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2251 // Select between the current value and the previous incoming edge
2252 // based on the incoming mask.
2253 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2254 Entry[part], "predphi");
2260 // This PHINode must be an induction variable.
2261 // Make sure that we know about it.
2262 assert(Legal->getInductionVars()->count(P) &&
2263 "Not an induction variable");
2265 LoopVectorizationLegality::InductionInfo II =
2266 Legal->getInductionVars()->lookup(P);
2269 case LoopVectorizationLegality::IK_NoInduction:
2270 llvm_unreachable("Unknown induction");
2271 case LoopVectorizationLegality::IK_IntInduction: {
2272 assert(P->getType() == II.StartValue->getType() && "Types must match");
2273 Type *PhiTy = P->getType();
2275 if (P == OldInduction) {
2276 // Handle the canonical induction variable. We might have had to
2278 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2280 // Handle other induction variables that are now based on the
2282 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2284 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2285 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2288 Broadcasted = getBroadcastInstrs(Broadcasted);
2289 // After broadcasting the induction variable we need to make the vector
2290 // consecutive by adding 0, 1, 2, etc.
2291 for (unsigned part = 0; part < UF; ++part)
2292 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2295 case LoopVectorizationLegality::IK_ReverseIntInduction:
2296 case LoopVectorizationLegality::IK_PtrInduction:
2297 case LoopVectorizationLegality::IK_ReversePtrInduction:
2298 // Handle reverse integer and pointer inductions.
2299 Value *StartIdx = ExtendedIdx;
2300 // This is the normalized GEP that starts counting at zero.
2301 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2304 // Handle the reverse integer induction variable case.
2305 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2306 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2307 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2309 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2312 // This is a new value so do not hoist it out.
2313 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2314 // After broadcasting the induction variable we need to make the
2315 // vector consecutive by adding ... -3, -2, -1, 0.
2316 for (unsigned part = 0; part < UF; ++part)
2317 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2322 // Handle the pointer induction variable case.
2323 assert(P->getType()->isPointerTy() && "Unexpected type.");
2325 // Is this a reverse induction ptr or a consecutive induction ptr.
2326 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2329 // This is the vector of results. Notice that we don't generate
2330 // vector geps because scalar geps result in better code.
2331 for (unsigned part = 0; part < UF; ++part) {
2332 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2333 for (unsigned int i = 0; i < VF; ++i) {
2334 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2335 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2338 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2340 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2342 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2344 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2345 Builder.getInt32(i),
2348 Entry[part] = VecVal;
2355 case Instruction::Add:
2356 case Instruction::FAdd:
2357 case Instruction::Sub:
2358 case Instruction::FSub:
2359 case Instruction::Mul:
2360 case Instruction::FMul:
2361 case Instruction::UDiv:
2362 case Instruction::SDiv:
2363 case Instruction::FDiv:
2364 case Instruction::URem:
2365 case Instruction::SRem:
2366 case Instruction::FRem:
2367 case Instruction::Shl:
2368 case Instruction::LShr:
2369 case Instruction::AShr:
2370 case Instruction::And:
2371 case Instruction::Or:
2372 case Instruction::Xor: {
2373 // Just widen binops.
2374 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2375 setDebugLocFromInst(Builder, BinOp);
2376 VectorParts &A = getVectorValue(it->getOperand(0));
2377 VectorParts &B = getVectorValue(it->getOperand(1));
2379 // Use this vector value for all users of the original instruction.
2380 for (unsigned Part = 0; Part < UF; ++Part) {
2381 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2383 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2384 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2385 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2386 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2387 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2389 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2390 VecOp->setIsExact(BinOp->isExact());
2396 case Instruction::Select: {
2398 // If the selector is loop invariant we can create a select
2399 // instruction with a scalar condition. Otherwise, use vector-select.
2400 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2402 setDebugLocFromInst(Builder, it);
2404 // The condition can be loop invariant but still defined inside the
2405 // loop. This means that we can't just use the original 'cond' value.
2406 // We have to take the 'vectorized' value and pick the first lane.
2407 // Instcombine will make this a no-op.
2408 VectorParts &Cond = getVectorValue(it->getOperand(0));
2409 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2410 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2411 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2412 Builder.getInt32(0));
2413 for (unsigned Part = 0; Part < UF; ++Part) {
2414 Entry[Part] = Builder.CreateSelect(
2415 InvariantCond ? ScalarCond : Cond[Part],
2422 case Instruction::ICmp:
2423 case Instruction::FCmp: {
2424 // Widen compares. Generate vector compares.
2425 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2426 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2427 setDebugLocFromInst(Builder, it);
2428 VectorParts &A = getVectorValue(it->getOperand(0));
2429 VectorParts &B = getVectorValue(it->getOperand(1));
2430 for (unsigned Part = 0; Part < UF; ++Part) {
2433 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2435 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2441 case Instruction::Store:
2442 case Instruction::Load:
2443 vectorizeMemoryInstruction(it, Legal);
2445 case Instruction::ZExt:
2446 case Instruction::SExt:
2447 case Instruction::FPToUI:
2448 case Instruction::FPToSI:
2449 case Instruction::FPExt:
2450 case Instruction::PtrToInt:
2451 case Instruction::IntToPtr:
2452 case Instruction::SIToFP:
2453 case Instruction::UIToFP:
2454 case Instruction::Trunc:
2455 case Instruction::FPTrunc:
2456 case Instruction::BitCast: {
2457 CastInst *CI = dyn_cast<CastInst>(it);
2458 setDebugLocFromInst(Builder, it);
2459 /// Optimize the special case where the source is the induction
2460 /// variable. Notice that we can only optimize the 'trunc' case
2461 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2462 /// c. other casts depend on pointer size.
2463 if (CI->getOperand(0) == OldInduction &&
2464 it->getOpcode() == Instruction::Trunc) {
2465 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2467 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2468 for (unsigned Part = 0; Part < UF; ++Part)
2469 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2472 /// Vectorize casts.
2473 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2475 VectorParts &A = getVectorValue(it->getOperand(0));
2476 for (unsigned Part = 0; Part < UF; ++Part)
2477 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2481 case Instruction::Call: {
2482 // Ignore dbg intrinsics.
2483 if (isa<DbgInfoIntrinsic>(it))
2485 setDebugLocFromInst(Builder, it);
2487 Module *M = BB->getParent()->getParent();
2488 CallInst *CI = cast<CallInst>(it);
2489 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2490 assert(ID && "Not an intrinsic call!");
2491 for (unsigned Part = 0; Part < UF; ++Part) {
2492 SmallVector<Value*, 4> Args;
2493 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2494 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2495 Args.push_back(Arg[Part]);
2497 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2498 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2499 Entry[Part] = Builder.CreateCall(F, Args);
2505 // All other instructions are unsupported. Scalarize them.
2506 scalarizeInstruction(it);
2509 }// end of for_each instr.
2512 void InnerLoopVectorizer::updateAnalysis() {
2513 // Forget the original basic block.
2514 SE->forgetLoop(OrigLoop);
2516 // Update the dominator tree information.
2517 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2518 "Entry does not dominate exit.");
2520 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2521 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2522 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2523 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2524 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2525 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2526 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2527 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2529 DEBUG(DT->verifyAnalysis());
2532 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2533 if (!EnableIfConversion)
2536 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2537 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2539 // A list of pointers that we can safely read and write to.
2540 SmallPtrSet<Value *, 8> SafePointes;
2542 // Collect safe addresses.
2543 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2544 BasicBlock *BB = LoopBlocks[i];
2546 if (blockNeedsPredication(BB))
2549 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
2550 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2551 SafePointes.insert(LI->getPointerOperand());
2552 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
2553 SafePointes.insert(SI->getPointerOperand());
2557 // Collect the blocks that need predication.
2558 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2559 BasicBlock *BB = LoopBlocks[i];
2561 // We don't support switch statements inside loops.
2562 if (!isa<BranchInst>(BB->getTerminator()))
2565 // We must be able to predicate all blocks that need to be predicated.
2566 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB, SafePointes))
2570 // We can if-convert this loop.
2574 bool LoopVectorizationLegality::canVectorize() {
2575 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2576 // be canonicalized.
2577 if (!TheLoop->getLoopPreheader())
2580 // We can only vectorize innermost loops.
2581 if (TheLoop->getSubLoopsVector().size())
2584 // We must have a single backedge.
2585 if (TheLoop->getNumBackEdges() != 1)
2588 // We must have a single exiting block.
2589 if (!TheLoop->getExitingBlock())
2592 unsigned NumBlocks = TheLoop->getNumBlocks();
2594 // Check if we can if-convert non single-bb loops.
2595 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2596 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2600 // We need to have a loop header.
2601 BasicBlock *Latch = TheLoop->getLoopLatch();
2602 DEBUG(dbgs() << "LV: Found a loop: " <<
2603 TheLoop->getHeader()->getName() << "\n");
2605 // ScalarEvolution needs to be able to find the exit count.
2606 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2607 if (ExitCount == SE->getCouldNotCompute()) {
2608 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2612 // Do not loop-vectorize loops with a tiny trip count.
2613 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2614 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2615 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2616 "This loop is not worth vectorizing.\n");
2620 // Check if we can vectorize the instructions and CFG in this loop.
2621 if (!canVectorizeInstrs()) {
2622 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2626 // Go over each instruction and look at memory deps.
2627 if (!canVectorizeMemory()) {
2628 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2632 // Collect all of the variables that remain uniform after vectorization.
2633 collectLoopUniforms();
2635 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2636 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2639 // Okay! We can vectorize. At this point we don't have any other mem analysis
2640 // which may limit our maximum vectorization factor, so just return true with
2645 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2646 if (Ty->isPointerTy())
2647 return DL.getIntPtrType(Ty->getContext());
2651 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2652 Ty0 = convertPointerToIntegerType(DL, Ty0);
2653 Ty1 = convertPointerToIntegerType(DL, Ty1);
2654 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2659 /// \brief Check that the instruction has outside loop users and is not an
2660 /// identified reduction variable.
2661 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2662 SmallPtrSet<Value *, 4> &Reductions) {
2663 // Reduction instructions are allowed to have exit users. All other
2664 // instructions must not have external users.
2665 if (!Reductions.count(Inst))
2666 //Check that all of the users of the loop are inside the BB.
2667 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2669 Instruction *U = cast<Instruction>(*I);
2670 // This user may be a reduction exit value.
2671 if (!TheLoop->contains(U)) {
2672 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2679 bool LoopVectorizationLegality::canVectorizeInstrs() {
2680 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2681 BasicBlock *Header = TheLoop->getHeader();
2683 // Look for the attribute signaling the absence of NaNs.
2684 Function &F = *Header->getParent();
2685 if (F.hasFnAttribute("no-nans-fp-math"))
2686 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2687 AttributeSet::FunctionIndex,
2688 "no-nans-fp-math").getValueAsString() == "true";
2690 // For each block in the loop.
2691 for (Loop::block_iterator bb = TheLoop->block_begin(),
2692 be = TheLoop->block_end(); bb != be; ++bb) {
2694 // Scan the instructions in the block and look for hazards.
2695 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2698 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2699 Type *PhiTy = Phi->getType();
2700 // Check that this PHI type is allowed.
2701 if (!PhiTy->isIntegerTy() &&
2702 !PhiTy->isFloatingPointTy() &&
2703 !PhiTy->isPointerTy()) {
2704 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2708 // If this PHINode is not in the header block, then we know that we
2709 // can convert it to select during if-conversion. No need to check if
2710 // the PHIs in this block are induction or reduction variables.
2711 if (*bb != Header) {
2712 // Check that this instruction has no outside users or is an
2713 // identified reduction value with an outside user.
2714 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2719 // We only allow if-converted PHIs with more than two incoming values.
2720 if (Phi->getNumIncomingValues() != 2) {
2721 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2725 // This is the value coming from the preheader.
2726 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2727 // Check if this is an induction variable.
2728 InductionKind IK = isInductionVariable(Phi);
2730 if (IK_NoInduction != IK) {
2731 // Get the widest type.
2733 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2735 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2737 // Int inductions are special because we only allow one IV.
2738 if (IK == IK_IntInduction) {
2739 // Use the phi node with the widest type as induction. Use the last
2740 // one if there are multiple (no good reason for doing this other
2741 // than it is expedient).
2742 if (!Induction || PhiTy == WidestIndTy)
2746 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2747 Inductions[Phi] = InductionInfo(StartValue, IK);
2751 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2752 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2755 if (AddReductionVar(Phi, RK_IntegerMult)) {
2756 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2759 if (AddReductionVar(Phi, RK_IntegerOr)) {
2760 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2763 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2764 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2767 if (AddReductionVar(Phi, RK_IntegerXor)) {
2768 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2771 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2772 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2775 if (AddReductionVar(Phi, RK_FloatMult)) {
2776 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2779 if (AddReductionVar(Phi, RK_FloatAdd)) {
2780 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2783 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2784 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2788 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2790 }// end of PHI handling
2792 // We still don't handle functions. However, we can ignore dbg intrinsic
2793 // calls and we do handle certain intrinsic and libm functions.
2794 CallInst *CI = dyn_cast<CallInst>(it);
2795 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2796 DEBUG(dbgs() << "LV: Found a call site.\n");
2800 // Check that the instruction return type is vectorizable.
2801 if (!VectorType::isValidElementType(it->getType()) &&
2802 !it->getType()->isVoidTy()) {
2803 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2807 // Check that the stored type is vectorizable.
2808 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2809 Type *T = ST->getValueOperand()->getType();
2810 if (!VectorType::isValidElementType(T))
2814 // Reduction instructions are allowed to have exit users.
2815 // All other instructions must not have external users.
2816 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2824 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2825 if (Inductions.empty())
2832 void LoopVectorizationLegality::collectLoopUniforms() {
2833 // We now know that the loop is vectorizable!
2834 // Collect variables that will remain uniform after vectorization.
2835 std::vector<Value*> Worklist;
2836 BasicBlock *Latch = TheLoop->getLoopLatch();
2838 // Start with the conditional branch and walk up the block.
2839 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2841 while (Worklist.size()) {
2842 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2843 Worklist.pop_back();
2845 // Look at instructions inside this loop.
2846 // Stop when reaching PHI nodes.
2847 // TODO: we need to follow values all over the loop, not only in this block.
2848 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2851 // This is a known uniform.
2854 // Insert all operands.
2855 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2860 /// \brief Analyses memory accesses in a loop.
2862 /// Checks whether run time pointer checks are needed and builds sets for data
2863 /// dependence checking.
2864 class AccessAnalysis {
2866 /// \brief Read or write access location.
2867 typedef std::pair<Value*, char> MemAccessInfo;
2869 /// \brief Set of potential dependent memory accesses.
2870 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2872 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2873 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2874 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2876 /// \brief Register a load and whether it is only read from.
2877 void addLoad(Value *Ptr, bool IsReadOnly) {
2878 Accesses.insert(std::make_pair(Ptr, false));
2880 ReadOnlyPtr.insert(Ptr);
2883 /// \brief Register a store.
2884 void addStore(Value *Ptr) {
2885 Accesses.insert(std::make_pair(Ptr, true));
2888 /// \brief Check whether we can check the pointers at runtime for
2889 /// non-intersection.
2890 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2891 unsigned &NumComparisons, ScalarEvolution *SE,
2894 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2895 /// and builds sets of dependent accesses.
2896 void buildDependenceSets() {
2897 // Process read-write pointers first.
2898 processMemAccesses(false);
2899 // Next, process read pointers.
2900 processMemAccesses(true);
2903 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
2905 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
2907 DenseSet<MemAccessInfo> &getDependenciesToCheck() { return CheckDeps; }
2910 typedef SetVector<MemAccessInfo> PtrAccessSet;
2911 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
2913 /// \brief Go over all memory access or only the deferred ones if
2914 /// \p UseDeferred is true and check whether runtime pointer checks are needed
2915 /// and build sets of dependency check candidates.
2916 void processMemAccesses(bool UseDeferred);
2918 /// Set of all accesses.
2919 PtrAccessSet Accesses;
2921 /// Set of access to check after all writes have been processed.
2922 PtrAccessSet DeferredAccesses;
2924 /// Map of pointers to last access encountered.
2925 UnderlyingObjToAccessMap ObjToLastAccess;
2927 /// Set of accesses that need a further dependence check.
2928 DenseSet<MemAccessInfo> CheckDeps;
2930 /// Set of pointers that are read only.
2931 SmallPtrSet<Value*, 16> ReadOnlyPtr;
2933 /// Set of underlying objects already written to.
2934 SmallPtrSet<Value*, 16> WriteObjects;
2938 /// Sets of potentially dependent accesses - members of one set share an
2939 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
2940 /// dependence check.
2941 DepCandidates &DepCands;
2943 bool AreAllWritesIdentified;
2944 bool AreAllReadsIdentified;
2945 bool IsRTCheckNeeded;
2948 } // end anonymous namespace
2950 /// \brief Check whether a pointer can participate in a runtime bounds check.
2951 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
2952 const SCEV *PtrScev = SE->getSCEV(Ptr);
2953 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
2957 return AR->isAffine();
2960 bool AccessAnalysis::canCheckPtrAtRT(
2961 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2962 unsigned &NumComparisons, ScalarEvolution *SE,
2964 // Find pointers with computable bounds. We are going to use this information
2965 // to place a runtime bound check.
2966 unsigned NumReadPtrChecks = 0;
2967 unsigned NumWritePtrChecks = 0;
2968 bool CanDoRT = true;
2970 bool IsDepCheckNeeded = isDependencyCheckNeeded();
2971 // We assign consecutive id to access from different dependence sets.
2972 // Accesses within the same set don't need a runtime check.
2973 unsigned RunningDepId = 1;
2974 DenseMap<Value *, unsigned> DepSetId;
2976 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
2978 const MemAccessInfo &Access = *AI;
2979 Value *Ptr = Access.first;
2980 bool IsWrite = Access.second;
2982 // Just add write checks if we have both.
2983 if (!IsWrite && Accesses.count(std::make_pair(Ptr, true)))
2987 ++NumWritePtrChecks;
2991 if (hasComputableBounds(SE, Ptr)) {
2992 // The id of the dependence set.
2995 if (IsDepCheckNeeded) {
2996 Value *Leader = DepCands.getLeaderValue(Access).first;
2997 unsigned &LeaderId = DepSetId[Leader];
2999 LeaderId = RunningDepId++;
3002 // Each access has its own dependence set.
3003 DepId = RunningDepId++;
3005 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3007 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3013 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3014 NumComparisons = 0; // Only one dependence set.
3016 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3017 NumWritePtrChecks - 1));
3021 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3022 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3025 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3026 // We process the set twice: first we process read-write pointers, last we
3027 // process read-only pointers. This allows us to skip dependence tests for
3028 // read-only pointers.
3030 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3031 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3032 const MemAccessInfo &Access = *AI;
3033 Value *Ptr = Access.first;
3034 bool IsWrite = Access.second;
3036 DepCands.insert(Access);
3038 // Memorize read-only pointers for later processing and skip them in the
3039 // first round (they need to be checked after we have seen all write
3040 // pointers). Note: we also mark pointer that are not consecutive as
3041 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3042 // second check for "!IsWrite".
3043 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3044 if (!UseDeferred && IsReadOnlyPtr) {
3045 DeferredAccesses.insert(Access);
3049 bool NeedDepCheck = false;
3050 // Check whether there is the possiblity of dependency because of underlying
3051 // objects being the same.
3052 typedef SmallVector<Value*, 16> ValueVector;
3053 ValueVector TempObjects;
3054 GetUnderlyingObjects(Ptr, TempObjects, DL);
3055 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3057 Value *UnderlyingObj = *UI;
3059 // If this is a write then it needs to be an identified object. If this a
3060 // read and all writes (so far) are identified function scope objects we
3061 // don't need an identified underlying object but only an Argument (the
3062 // next write is going to invalidate this assumption if it is
3064 // This is a micro-optimization for the case where all writes are
3065 // identified and we have one argument pointer.
3066 // Otherwise, we do need a runtime check.
3067 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3068 (!IsWrite && (!AreAllWritesIdentified ||
3069 !isa<Argument>(UnderlyingObj)) &&
3070 !isIdentifiedObject(UnderlyingObj))) {
3071 DEBUG(dbgs() << "LV: Found an unidentified " <<
3072 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3074 IsRTCheckNeeded = (IsRTCheckNeeded ||
3075 !isIdentifiedObject(UnderlyingObj) ||
3076 !AreAllReadsIdentified);
3079 AreAllWritesIdentified = false;
3081 AreAllReadsIdentified = false;
3084 // If this is a write - check other reads and writes for conflicts. If
3085 // this is a read only check other writes for conflicts (but only if there
3086 // is no other write to the ptr - this is an optimization to catch "a[i] =
3087 // a[i] + " without having to do a dependence check).
3088 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3089 NeedDepCheck = true;
3092 WriteObjects.insert(UnderlyingObj);
3094 // Create sets of pointers connected by shared underlying objects.
3095 UnderlyingObjToAccessMap::iterator Prev =
3096 ObjToLastAccess.find(UnderlyingObj);
3097 if (Prev != ObjToLastAccess.end())
3098 DepCands.unionSets(Access, Prev->second);
3100 ObjToLastAccess[UnderlyingObj] = Access;
3104 CheckDeps.insert(Access);
3109 /// \brief Checks memory dependences among accesses to the same underlying
3110 /// object to determine whether there vectorization is legal or not (and at
3111 /// which vectorization factor).
3113 /// This class works under the assumption that we already checked that memory
3114 /// locations with different underlying pointers are "must-not alias".
3115 /// We use the ScalarEvolution framework to symbolically evalutate access
3116 /// functions pairs. Since we currently don't restructure the loop we can rely
3117 /// on the program order of memory accesses to determine their safety.
3118 /// At the moment we will only deem accesses as safe for:
3119 /// * A negative constant distance assuming program order.
3121 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3122 /// a[i] = tmp; y = a[i];
3124 /// The latter case is safe because later checks guarantuee that there can't
3125 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3126 /// the same variable: a header phi can only be an induction or a reduction, a
3127 /// reduction can't have a memory sink, an induction can't have a memory
3128 /// source). This is important and must not be violated (or we have to
3129 /// resort to checking for cycles through memory).
3131 /// * A positive constant distance assuming program order that is bigger
3132 /// than the biggest memory access.
3134 /// tmp = a[i] OR b[i] = x
3135 /// a[i+2] = tmp y = b[i+2];
3137 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3139 /// * Zero distances and all accesses have the same size.
3141 class MemoryDepChecker {
3143 typedef std::pair<Value*, char> MemAccessInfo;
3145 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3146 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3148 /// \brief Register the location (instructions are given increasing numbers)
3149 /// of a write access.
3150 void addAccess(StoreInst *SI) {
3151 Value *Ptr = SI->getPointerOperand();
3152 Accesses[std::make_pair(Ptr, true)].push_back(AccessIdx);
3153 InstMap.push_back(SI);
3157 /// \brief Register the location (instructions are given increasing numbers)
3158 /// of a write access.
3159 void addAccess(LoadInst *LI) {
3160 Value *Ptr = LI->getPointerOperand();
3161 Accesses[std::make_pair(Ptr, false)].push_back(AccessIdx);
3162 InstMap.push_back(LI);
3166 /// \brief Check whether the dependencies between the accesses are safe.
3168 /// Only checks sets with elements in \p CheckDeps.
3169 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3170 DenseSet<MemAccessInfo> &CheckDeps);
3172 /// \brief The maximum number of bytes of a vector register we can vectorize
3173 /// the accesses safely with.
3174 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3177 ScalarEvolution *SE;
3179 const Loop *InnermostLoop;
3181 /// \brief Maps access locations (ptr, read/write) to program order.
3182 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3184 /// \brief Memory access instructions in program order.
3185 SmallVector<Instruction *, 16> InstMap;
3187 /// \brief The program order index to be used for the next instruction.
3190 // We can access this many bytes in parallel safely.
3191 unsigned MaxSafeDepDistBytes;
3193 /// \brief Check whether there is a plausible dependence between the two
3196 /// Access \p A must happen before \p B in program order. The two indices
3197 /// identify the index into the program order map.
3199 /// This function checks whether there is a plausible dependence (or the
3200 /// absence of such can't be proved) between the two accesses. If there is a
3201 /// plausible dependence but the dependence distance is bigger than one
3202 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3203 /// distance is smaller than any other distance encountered so far).
3204 /// Otherwise, this function returns true signaling a possible dependence.
3205 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3206 const MemAccessInfo &B, unsigned BIdx);
3208 /// \brief Check whether the data dependence could prevent store-load
3210 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3213 } // end anonymous namespace
3215 static bool isInBoundsGep(Value *Ptr) {
3216 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3217 return GEP->isInBounds();
3221 /// \brief Check whether the access through \p Ptr has a constant stride.
3222 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3224 const Type *PtrTy = Ptr->getType();
3225 assert(PtrTy->isPointerTy() && "Unexpected non ptr");
3227 // Make sure that the pointer does not point to aggregate types.
3228 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) {
3229 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr
3234 const SCEV *PtrScev = SE->getSCEV(Ptr);
3235 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3237 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3238 << *Ptr << " SCEV: " << *PtrScev << "\n");
3242 // The accesss function must stride over the innermost loop.
3243 if (Lp != AR->getLoop()) {
3244 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << *Ptr
3245 << " SCEV: " << *PtrScev << "\n");
3248 // The address calculation must not wrap. Otherwise, a dependence could be
3249 // inverted. An inbounds getelementptr that is a AddRec with a unit stride
3250 // cannot wrap per definition. The unit stride requirement is checked later.
3251 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3252 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3253 if (!IsNoWrapAddRec && !IsInBoundsGEP) {
3254 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3255 << *Ptr << " SCEV: " << *PtrScev << "\n");
3259 // Check the step is constant.
3260 const SCEV *Step = AR->getStepRecurrence(*SE);
3262 // Calculate the pointer stride and check if it is consecutive.
3263 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3265 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3266 " SCEV: " << *PtrScev << "\n");
3270 int64_t Size = DL->getTypeAllocSize(PtrTy->getPointerElementType());
3271 const APInt &APStepVal = C->getValue()->getValue();
3273 // Huge step value - give up.
3274 if (APStepVal.getBitWidth() > 64)
3277 int64_t StepVal = APStepVal.getSExtValue();
3280 int64_t Stride = StepVal / Size;
3281 int64_t Rem = StepVal % Size;
3285 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3286 // know we can't "wrap around the address space".
3287 if (!IsNoWrapAddRec && IsInBoundsGEP && Stride != 1 && Stride != -1)
3293 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3294 unsigned TypeByteSize) {
3295 // If loads occur at a distance that is not a multiple of a feasible vector
3296 // factor store-load forwarding does not take place.
3297 // Positive dependences might cause troubles because vectorizing them might
3298 // prevent store-load forwarding making vectorized code run a lot slower.
3299 // a[i] = a[i-3] ^ a[i-8];
3300 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3301 // hence on your typical architecture store-load forwarding does not take
3302 // place. Vectorizing in such cases does not make sense.
3303 // Store-load forwarding distance.
3304 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3305 // Maximum vector factor.
3306 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3307 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3308 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3310 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3312 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3313 MaxVFWithoutSLForwardIssues = (vf >>=1);
3318 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3319 DEBUG(dbgs() << "LV: Distance " << Distance <<
3320 " that could cause a store-load forwarding conflict\n");
3324 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3325 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3326 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3330 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3331 const MemAccessInfo &B, unsigned BIdx) {
3332 assert (AIdx < BIdx && "Must pass arguments in program order");
3334 Value *APtr = A.first;
3335 Value *BPtr = B.first;
3336 bool AIsWrite = A.second;
3337 bool BIsWrite = B.second;
3339 // Two reads are independent.
3340 if (!AIsWrite && !BIsWrite)
3343 const SCEV *AScev = SE->getSCEV(APtr);
3344 const SCEV *BScev = SE->getSCEV(BPtr);
3346 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3347 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3349 const SCEV *Src = AScev;
3350 const SCEV *Sink = BScev;
3352 // If the induction step is negative we have to invert source and sink of the
3354 if (StrideAPtr < 0) {
3357 std::swap(APtr, BPtr);
3358 std::swap(Src, Sink);
3359 std::swap(AIsWrite, BIsWrite);
3360 std::swap(AIdx, BIdx);
3361 std::swap(StrideAPtr, StrideBPtr);
3364 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3366 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3367 << "(Induction step: " << StrideAPtr << ")\n");
3368 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3369 << *InstMap[BIdx] << ": " << *Dist << "\n");
3371 // Need consecutive accesses. We don't want to vectorize
3372 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3373 // the address space.
3374 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3375 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3379 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3381 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3385 Type *ATy = APtr->getType()->getPointerElementType();
3386 Type *BTy = BPtr->getType()->getPointerElementType();
3387 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3389 // Negative distances are not plausible dependencies.
3390 const APInt &Val = C->getValue()->getValue();
3391 if (Val.isNegative()) {
3392 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3393 if (IsTrueDataDependence &&
3394 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3398 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3402 // Write to the same location with the same size.
3403 // Could be improved to assert type sizes are the same (i32 == float, etc).
3407 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3411 assert(Val.isStrictlyPositive() && "Expect a positive value");
3413 // Positive distance bigger than max vectorization factor.
3416 "LV: ReadWrite-Write positive dependency with different types");
3420 unsigned Distance = (unsigned) Val.getZExtValue();
3422 // Bail out early if passed-in parameters make vectorization not feasible.
3423 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3424 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3426 // The distance must be bigger than the size needed for a vectorized version
3427 // of the operation and the size of the vectorized operation must not be
3428 // bigger than the currrent maximum size.
3429 if (Distance < 2*TypeByteSize ||
3430 2*TypeByteSize > MaxSafeDepDistBytes ||
3431 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3432 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3433 << Val.getSExtValue() << "\n");
3437 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3438 Distance : MaxSafeDepDistBytes;
3440 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3441 if (IsTrueDataDependence &&
3442 couldPreventStoreLoadForward(Distance, TypeByteSize))
3445 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3446 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3452 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3453 DenseSet<MemAccessInfo> &CheckDeps) {
3455 MaxSafeDepDistBytes = -1U;
3456 while (!CheckDeps.empty()) {
3457 MemAccessInfo CurAccess = *CheckDeps.begin();
3459 // Get the relevant memory access set.
3460 EquivalenceClasses<MemAccessInfo>::iterator I =
3461 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3463 // Check accesses within this set.
3464 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3465 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3467 // Check every access pair.
3469 CheckDeps.erase(*AI);
3470 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3472 // Check every accessing instruction pair in program order.
3473 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3474 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3475 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3476 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3477 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3479 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3490 bool LoopVectorizationLegality::canVectorizeMemory() {
3492 typedef SmallVector<Value*, 16> ValueVector;
3493 typedef SmallPtrSet<Value*, 16> ValueSet;
3495 // Stores a pair of memory access location and whether the access is a store
3496 // (true) or a load (false).
3497 typedef std::pair<Value*, char> MemAccessInfo;
3498 typedef DenseSet<MemAccessInfo> PtrAccessSet;
3500 // Holds the Load and Store *instructions*.
3504 // Holds all the different accesses in the loop.
3505 unsigned NumReads = 0;
3506 unsigned NumReadWrites = 0;
3508 PtrRtCheck.Pointers.clear();
3509 PtrRtCheck.Need = false;
3511 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3512 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3515 for (Loop::block_iterator bb = TheLoop->block_begin(),
3516 be = TheLoop->block_end(); bb != be; ++bb) {
3518 // Scan the BB and collect legal loads and stores.
3519 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3522 // If this is a load, save it. If this instruction can read from memory
3523 // but is not a load, then we quit. Notice that we don't handle function
3524 // calls that read or write.
3525 if (it->mayReadFromMemory()) {
3526 LoadInst *Ld = dyn_cast<LoadInst>(it);
3527 if (!Ld) return false;
3528 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3529 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3532 Loads.push_back(Ld);
3533 DepChecker.addAccess(Ld);
3537 // Save 'store' instructions. Abort if other instructions write to memory.
3538 if (it->mayWriteToMemory()) {
3539 StoreInst *St = dyn_cast<StoreInst>(it);
3540 if (!St) return false;
3541 if (!St->isSimple() && !IsAnnotatedParallel) {
3542 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3545 Stores.push_back(St);
3546 DepChecker.addAccess(St);
3551 // Now we have two lists that hold the loads and the stores.
3552 // Next, we find the pointers that they use.
3554 // Check if we see any stores. If there are no stores, then we don't
3555 // care if the pointers are *restrict*.
3556 if (!Stores.size()) {
3557 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3561 AccessAnalysis::DepCandidates DependentAccesses;
3562 AccessAnalysis Accesses(DL, DependentAccesses);
3564 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3565 // multiple times on the same object. If the ptr is accessed twice, once
3566 // for read and once for write, it will only appear once (on the write
3567 // list). This is okay, since we are going to check for conflicts between
3568 // writes and between reads and writes, but not between reads and reads.
3571 ValueVector::iterator I, IE;
3572 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3573 StoreInst *ST = cast<StoreInst>(*I);
3574 Value* Ptr = ST->getPointerOperand();
3576 if (isUniform(Ptr)) {
3577 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3581 // If we did *not* see this pointer before, insert it to the read-write
3582 // list. At this phase it is only a 'write' list.
3583 if (Seen.insert(Ptr)) {
3585 Accesses.addStore(Ptr);
3589 if (IsAnnotatedParallel) {
3591 << "LV: A loop annotated parallel, ignore memory dependency "
3596 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3597 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3598 LoadInst *LD = cast<LoadInst>(*I);
3599 Value* Ptr = LD->getPointerOperand();
3600 // If we did *not* see this pointer before, insert it to the
3601 // read list. If we *did* see it before, then it is already in
3602 // the read-write list. This allows us to vectorize expressions
3603 // such as A[i] += x; Because the address of A[i] is a read-write
3604 // pointer. This only works if the index of A[i] is consecutive.
3605 // If the address of i is unknown (for example A[B[i]]) then we may
3606 // read a few words, modify, and write a few words, and some of the
3607 // words may be written to the same address.
3608 bool IsReadOnlyPtr = false;
3609 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3611 IsReadOnlyPtr = true;
3613 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3616 // If we write (or read-write) to a single destination and there are no
3617 // other reads in this loop then is it safe to vectorize.
3618 if (NumReadWrites == 1 && NumReads == 0) {
3619 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3623 // Build dependence sets and check whether we need a runtime pointer bounds
3625 Accesses.buildDependenceSets();
3626 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3628 // Find pointers with computable bounds. We are going to use this information
3629 // to place a runtime bound check.
3630 unsigned NumComparisons = 0;
3631 bool CanDoRT = false;
3633 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3636 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3637 " pointer comparisons.\n");
3639 // If we only have one set of dependences to check pointers among we don't
3640 // need a runtime check.
3641 if (NumComparisons == 0 && NeedRTCheck)
3642 NeedRTCheck = false;
3644 // Check that we did not collect too many pointers or found a unsizeable
3646 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3652 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3655 if (NeedRTCheck && !CanDoRT) {
3656 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3657 "the array bounds.\n");
3662 PtrRtCheck.Need = NeedRTCheck;
3664 bool CanVecMem = true;
3665 if (Accesses.isDependencyCheckNeeded()) {
3666 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3667 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3668 Accesses.getDependenciesToCheck());
3669 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3672 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3673 " need a runtime memory check.\n");
3678 static bool hasMultipleUsesOf(Instruction *I,
3679 SmallPtrSet<Instruction *, 8> &Insts) {
3680 unsigned NumUses = 0;
3681 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3682 if (Insts.count(dyn_cast<Instruction>(*Use)))
3691 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3692 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3693 if (!Set.count(dyn_cast<Instruction>(*Use)))
3698 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3699 ReductionKind Kind) {
3700 if (Phi->getNumIncomingValues() != 2)
3703 // Reduction variables are only found in the loop header block.
3704 if (Phi->getParent() != TheLoop->getHeader())
3707 // Obtain the reduction start value from the value that comes from the loop
3709 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3711 // ExitInstruction is the single value which is used outside the loop.
3712 // We only allow for a single reduction value to be used outside the loop.
3713 // This includes users of the reduction, variables (which form a cycle
3714 // which ends in the phi node).
3715 Instruction *ExitInstruction = 0;
3716 // Indicates that we found a reduction operation in our scan.
3717 bool FoundReduxOp = false;
3719 // We start with the PHI node and scan for all of the users of this
3720 // instruction. All users must be instructions that can be used as reduction
3721 // variables (such as ADD). We must have a single out-of-block user. The cycle
3722 // must include the original PHI.
3723 bool FoundStartPHI = false;
3725 // To recognize min/max patterns formed by a icmp select sequence, we store
3726 // the number of instruction we saw from the recognized min/max pattern,
3727 // to make sure we only see exactly the two instructions.
3728 unsigned NumCmpSelectPatternInst = 0;
3729 ReductionInstDesc ReduxDesc(false, 0);
3731 SmallPtrSet<Instruction *, 8> VisitedInsts;
3732 SmallVector<Instruction *, 8> Worklist;
3733 Worklist.push_back(Phi);
3734 VisitedInsts.insert(Phi);
3736 // A value in the reduction can be used:
3737 // - By the reduction:
3738 // - Reduction operation:
3739 // - One use of reduction value (safe).
3740 // - Multiple use of reduction value (not safe).
3742 // - All uses of the PHI must be the reduction (safe).
3743 // - Otherwise, not safe.
3744 // - By one instruction outside of the loop (safe).
3745 // - By further instructions outside of the loop (not safe).
3746 // - By an instruction that is not part of the reduction (not safe).
3748 // * An instruction type other than PHI or the reduction operation.
3749 // * A PHI in the header other than the initial PHI.
3750 while (!Worklist.empty()) {
3751 Instruction *Cur = Worklist.back();
3752 Worklist.pop_back();
3755 // If the instruction has no users then this is a broken chain and can't be
3756 // a reduction variable.
3757 if (Cur->use_empty())
3760 bool IsAPhi = isa<PHINode>(Cur);
3762 // A header PHI use other than the original PHI.
3763 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3766 // Reductions of instructions such as Div, and Sub is only possible if the
3767 // LHS is the reduction variable.
3768 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3769 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3770 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3773 // Any reduction instruction must be of one of the allowed kinds.
3774 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3775 if (!ReduxDesc.IsReduction)
3778 // A reduction operation must only have one use of the reduction value.
3779 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3780 hasMultipleUsesOf(Cur, VisitedInsts))
3783 // All inputs to a PHI node must be a reduction value.
3784 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3787 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3788 isa<SelectInst>(Cur)))
3789 ++NumCmpSelectPatternInst;
3790 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3791 isa<SelectInst>(Cur)))
3792 ++NumCmpSelectPatternInst;
3794 // Check whether we found a reduction operator.
3795 FoundReduxOp |= !IsAPhi;
3797 // Process users of current instruction. Push non PHI nodes after PHI nodes
3798 // onto the stack. This way we are going to have seen all inputs to PHI
3799 // nodes once we get to them.
3800 SmallVector<Instruction *, 8> NonPHIs;
3801 SmallVector<Instruction *, 8> PHIs;
3802 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3804 Instruction *Usr = cast<Instruction>(*UI);
3806 // Check if we found the exit user.
3807 BasicBlock *Parent = Usr->getParent();
3808 if (!TheLoop->contains(Parent)) {
3809 // Exit if you find multiple outside users.
3810 if (ExitInstruction != 0)
3812 ExitInstruction = Cur;
3816 // Process instructions only once (termination).
3817 if (VisitedInsts.insert(Usr)) {
3818 if (isa<PHINode>(Usr))
3819 PHIs.push_back(Usr);
3821 NonPHIs.push_back(Usr);
3823 // Remember that we completed the cycle.
3825 FoundStartPHI = true;
3827 Worklist.append(PHIs.begin(), PHIs.end());
3828 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3831 // This means we have seen one but not the other instruction of the
3832 // pattern or more than just a select and cmp.
3833 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3834 NumCmpSelectPatternInst != 2)
3837 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3840 // We found a reduction var if we have reached the original phi node and we
3841 // only have a single instruction with out-of-loop users.
3843 // This instruction is allowed to have out-of-loop users.
3844 AllowedExit.insert(ExitInstruction);
3846 // Save the description of this reduction variable.
3847 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3848 ReduxDesc.MinMaxKind);
3849 Reductions[Phi] = RD;
3850 // We've ended the cycle. This is a reduction variable if we have an
3851 // outside user and it has a binary op.
3856 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3857 /// pattern corresponding to a min(X, Y) or max(X, Y).
3858 LoopVectorizationLegality::ReductionInstDesc
3859 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3860 ReductionInstDesc &Prev) {
3862 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3863 "Expect a select instruction");
3864 Instruction *Cmp = 0;
3865 SelectInst *Select = 0;
3867 // We must handle the select(cmp()) as a single instruction. Advance to the
3869 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3870 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3871 return ReductionInstDesc(false, I);
3872 return ReductionInstDesc(Select, Prev.MinMaxKind);
3875 // Only handle single use cases for now.
3876 if (!(Select = dyn_cast<SelectInst>(I)))
3877 return ReductionInstDesc(false, I);
3878 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3879 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3880 return ReductionInstDesc(false, I);
3881 if (!Cmp->hasOneUse())
3882 return ReductionInstDesc(false, I);
3887 // Look for a min/max pattern.
3888 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3889 return ReductionInstDesc(Select, MRK_UIntMin);
3890 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3891 return ReductionInstDesc(Select, MRK_UIntMax);
3892 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3893 return ReductionInstDesc(Select, MRK_SIntMax);
3894 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3895 return ReductionInstDesc(Select, MRK_SIntMin);
3896 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3897 return ReductionInstDesc(Select, MRK_FloatMin);
3898 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3899 return ReductionInstDesc(Select, MRK_FloatMax);
3900 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3901 return ReductionInstDesc(Select, MRK_FloatMin);
3902 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3903 return ReductionInstDesc(Select, MRK_FloatMax);
3905 return ReductionInstDesc(false, I);
3908 LoopVectorizationLegality::ReductionInstDesc
3909 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3911 ReductionInstDesc &Prev) {
3912 bool FP = I->getType()->isFloatingPointTy();
3913 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3914 switch (I->getOpcode()) {
3916 return ReductionInstDesc(false, I);
3917 case Instruction::PHI:
3918 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3919 Kind != RK_FloatMinMax))
3920 return ReductionInstDesc(false, I);
3921 return ReductionInstDesc(I, Prev.MinMaxKind);
3922 case Instruction::Sub:
3923 case Instruction::Add:
3924 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3925 case Instruction::Mul:
3926 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3927 case Instruction::And:
3928 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3929 case Instruction::Or:
3930 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3931 case Instruction::Xor:
3932 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3933 case Instruction::FMul:
3934 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3935 case Instruction::FAdd:
3936 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3937 case Instruction::FCmp:
3938 case Instruction::ICmp:
3939 case Instruction::Select:
3940 if (Kind != RK_IntegerMinMax &&
3941 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3942 return ReductionInstDesc(false, I);
3943 return isMinMaxSelectCmpPattern(I, Prev);
3947 LoopVectorizationLegality::InductionKind
3948 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3949 Type *PhiTy = Phi->getType();
3950 // We only handle integer and pointer inductions variables.
3951 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3952 return IK_NoInduction;
3954 // Check that the PHI is consecutive.
3955 const SCEV *PhiScev = SE->getSCEV(Phi);
3956 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3958 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3959 return IK_NoInduction;
3961 const SCEV *Step = AR->getStepRecurrence(*SE);
3963 // Integer inductions need to have a stride of one.
3964 if (PhiTy->isIntegerTy()) {
3966 return IK_IntInduction;
3967 if (Step->isAllOnesValue())
3968 return IK_ReverseIntInduction;
3969 return IK_NoInduction;
3972 // Calculate the pointer stride and check if it is consecutive.
3973 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3975 return IK_NoInduction;
3977 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3978 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3979 if (C->getValue()->equalsInt(Size))
3980 return IK_PtrInduction;
3981 else if (C->getValue()->equalsInt(0 - Size))
3982 return IK_ReversePtrInduction;
3984 return IK_NoInduction;
3987 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3988 Value *In0 = const_cast<Value*>(V);
3989 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3993 return Inductions.count(PN);
3996 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3997 assert(TheLoop->contains(BB) && "Unknown block used");
3999 // Blocks that do not dominate the latch need predication.
4000 BasicBlock* Latch = TheLoop->getLoopLatch();
4001 return !DT->dominates(BB, Latch);
4004 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4005 SmallPtrSet<Value *, 8>& SafePtrs) {
4006 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4007 // We might be able to hoist the load.
4008 if (it->mayReadFromMemory()) {
4009 LoadInst *LI = dyn_cast<LoadInst>(it);
4010 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4014 // We don't predicate stores at the moment.
4015 if (it->mayWriteToMemory() || it->mayThrow())
4018 // The instructions below can trap.
4019 switch (it->getOpcode()) {
4021 case Instruction::UDiv:
4022 case Instruction::SDiv:
4023 case Instruction::URem:
4024 case Instruction::SRem:
4032 LoopVectorizationCostModel::VectorizationFactor
4033 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4035 // Width 1 means no vectorize
4036 VectorizationFactor Factor = { 1U, 0U };
4037 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4038 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4042 // Find the trip count.
4043 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4044 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4046 unsigned WidestType = getWidestType();
4047 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4048 unsigned MaxSafeDepDist = -1U;
4049 if (Legal->getMaxSafeDepDistBytes() != -1U)
4050 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4051 WidestRegister = WidestRegister < MaxSafeDepDist ? WidestRegister : MaxSafeDepDist;
4052 unsigned MaxVectorSize = WidestRegister / WidestType;
4053 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4054 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4056 if (MaxVectorSize == 0) {
4057 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4061 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4062 " into one vector!");
4064 unsigned VF = MaxVectorSize;
4066 // If we optimize the program for size, avoid creating the tail loop.
4068 // If we are unable to calculate the trip count then don't try to vectorize.
4070 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4074 // Find the maximum SIMD width that can fit within the trip count.
4075 VF = TC % MaxVectorSize;
4080 // If the trip count that we found modulo the vectorization factor is not
4081 // zero then we require a tail.
4083 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4089 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4090 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4092 Factor.Width = UserVF;
4096 float Cost = expectedCost(1);
4098 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4099 for (unsigned i=2; i <= VF; i*=2) {
4100 // Notice that the vector loop needs to be executed less times, so
4101 // we need to divide the cost of the vector loops by the width of
4102 // the vector elements.
4103 float VectorCost = expectedCost(i) / (float)i;
4104 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4105 (int)VectorCost << ".\n");
4106 if (VectorCost < Cost) {
4112 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4113 Factor.Width = Width;
4114 Factor.Cost = Width * Cost;
4118 unsigned LoopVectorizationCostModel::getWidestType() {
4119 unsigned MaxWidth = 8;
4122 for (Loop::block_iterator bb = TheLoop->block_begin(),
4123 be = TheLoop->block_end(); bb != be; ++bb) {
4124 BasicBlock *BB = *bb;
4126 // For each instruction in the loop.
4127 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4128 Type *T = it->getType();
4130 // Only examine Loads, Stores and PHINodes.
4131 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4134 // Examine PHI nodes that are reduction variables.
4135 if (PHINode *PN = dyn_cast<PHINode>(it))
4136 if (!Legal->getReductionVars()->count(PN))
4139 // Examine the stored values.
4140 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4141 T = ST->getValueOperand()->getType();
4143 // Ignore loaded pointer types and stored pointer types that are not
4144 // consecutive. However, we do want to take consecutive stores/loads of
4145 // pointer vectors into account.
4146 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4149 MaxWidth = std::max(MaxWidth,
4150 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4158 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4161 unsigned LoopCost) {
4163 // -- The unroll heuristics --
4164 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4165 // There are many micro-architectural considerations that we can't predict
4166 // at this level. For example frontend pressure (on decode or fetch) due to
4167 // code size, or the number and capabilities of the execution ports.
4169 // We use the following heuristics to select the unroll factor:
4170 // 1. If the code has reductions the we unroll in order to break the cross
4171 // iteration dependency.
4172 // 2. If the loop is really small then we unroll in order to reduce the loop
4174 // 3. We don't unroll if we think that we will spill registers to memory due
4175 // to the increased register pressure.
4177 // Use the user preference, unless 'auto' is selected.
4181 // When we optimize for size we don't unroll.
4185 // We used the distance for the unroll factor.
4186 if (Legal->getMaxSafeDepDistBytes() != -1U)
4189 // Do not unroll loops with a relatively small trip count.
4190 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4191 TheLoop->getLoopLatch());
4192 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4195 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4196 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4197 " vector registers\n");
4199 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4200 // We divide by these constants so assume that we have at least one
4201 // instruction that uses at least one register.
4202 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4203 R.NumInstructions = std::max(R.NumInstructions, 1U);
4205 // We calculate the unroll factor using the following formula.
4206 // Subtract the number of loop invariants from the number of available
4207 // registers. These registers are used by all of the unrolled instances.
4208 // Next, divide the remaining registers by the number of registers that is
4209 // required by the loop, in order to estimate how many parallel instances
4210 // fit without causing spills.
4211 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4213 // Clamp the unroll factor ranges to reasonable factors.
4214 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4216 // If we did not calculate the cost for VF (because the user selected the VF)
4217 // then we calculate the cost of VF here.
4219 LoopCost = expectedCost(VF);
4221 // Clamp the calculated UF to be between the 1 and the max unroll factor
4222 // that the target allows.
4223 if (UF > MaxUnrollSize)
4228 if (Legal->getReductionVars()->size()) {
4229 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4233 // We want to unroll tiny loops in order to reduce the loop overhead.
4234 // We assume that the cost overhead is 1 and we use the cost model
4235 // to estimate the cost of the loop and unroll until the cost of the
4236 // loop overhead is about 5% of the cost of the loop.
4237 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4238 if (LoopCost < 20) {
4239 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4240 unsigned NewUF = 20/LoopCost + 1;
4241 return std::min(NewUF, UF);
4244 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4248 LoopVectorizationCostModel::RegisterUsage
4249 LoopVectorizationCostModel::calculateRegisterUsage() {
4250 // This function calculates the register usage by measuring the highest number
4251 // of values that are alive at a single location. Obviously, this is a very
4252 // rough estimation. We scan the loop in a topological order in order and
4253 // assign a number to each instruction. We use RPO to ensure that defs are
4254 // met before their users. We assume that each instruction that has in-loop
4255 // users starts an interval. We record every time that an in-loop value is
4256 // used, so we have a list of the first and last occurrences of each
4257 // instruction. Next, we transpose this data structure into a multi map that
4258 // holds the list of intervals that *end* at a specific location. This multi
4259 // map allows us to perform a linear search. We scan the instructions linearly
4260 // and record each time that a new interval starts, by placing it in a set.
4261 // If we find this value in the multi-map then we remove it from the set.
4262 // The max register usage is the maximum size of the set.
4263 // We also search for instructions that are defined outside the loop, but are
4264 // used inside the loop. We need this number separately from the max-interval
4265 // usage number because when we unroll, loop-invariant values do not take
4267 LoopBlocksDFS DFS(TheLoop);
4271 R.NumInstructions = 0;
4273 // Each 'key' in the map opens a new interval. The values
4274 // of the map are the index of the 'last seen' usage of the
4275 // instruction that is the key.
4276 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4277 // Maps instruction to its index.
4278 DenseMap<unsigned, Instruction*> IdxToInstr;
4279 // Marks the end of each interval.
4280 IntervalMap EndPoint;
4281 // Saves the list of instruction indices that are used in the loop.
4282 SmallSet<Instruction*, 8> Ends;
4283 // Saves the list of values that are used in the loop but are
4284 // defined outside the loop, such as arguments and constants.
4285 SmallPtrSet<Value*, 8> LoopInvariants;
4288 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4289 be = DFS.endRPO(); bb != be; ++bb) {
4290 R.NumInstructions += (*bb)->size();
4291 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4293 Instruction *I = it;
4294 IdxToInstr[Index++] = I;
4296 // Save the end location of each USE.
4297 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4298 Value *U = I->getOperand(i);
4299 Instruction *Instr = dyn_cast<Instruction>(U);
4301 // Ignore non-instruction values such as arguments, constants, etc.
4302 if (!Instr) continue;
4304 // If this instruction is outside the loop then record it and continue.
4305 if (!TheLoop->contains(Instr)) {
4306 LoopInvariants.insert(Instr);
4310 // Overwrite previous end points.
4311 EndPoint[Instr] = Index;
4317 // Saves the list of intervals that end with the index in 'key'.
4318 typedef SmallVector<Instruction*, 2> InstrList;
4319 DenseMap<unsigned, InstrList> TransposeEnds;
4321 // Transpose the EndPoints to a list of values that end at each index.
4322 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4324 TransposeEnds[it->second].push_back(it->first);
4326 SmallSet<Instruction*, 8> OpenIntervals;
4327 unsigned MaxUsage = 0;
4330 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4331 for (unsigned int i = 0; i < Index; ++i) {
4332 Instruction *I = IdxToInstr[i];
4333 // Ignore instructions that are never used within the loop.
4334 if (!Ends.count(I)) continue;
4336 // Remove all of the instructions that end at this location.
4337 InstrList &List = TransposeEnds[i];
4338 for (unsigned int j=0, e = List.size(); j < e; ++j)
4339 OpenIntervals.erase(List[j]);
4341 // Count the number of live interals.
4342 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4344 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4345 OpenIntervals.size() <<"\n");
4347 // Add the current instruction to the list of open intervals.
4348 OpenIntervals.insert(I);
4351 unsigned Invariant = LoopInvariants.size();
4352 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4353 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4354 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4356 R.LoopInvariantRegs = Invariant;
4357 R.MaxLocalUsers = MaxUsage;
4361 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4365 for (Loop::block_iterator bb = TheLoop->block_begin(),
4366 be = TheLoop->block_end(); bb != be; ++bb) {
4367 unsigned BlockCost = 0;
4368 BasicBlock *BB = *bb;
4370 // For each instruction in the old loop.
4371 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4372 // Skip dbg intrinsics.
4373 if (isa<DbgInfoIntrinsic>(it))
4376 unsigned C = getInstructionCost(it, VF);
4378 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4379 VF << " For instruction: "<< *it << "\n");
4382 // We assume that if-converted blocks have a 50% chance of being executed.
4383 // When the code is scalar then some of the blocks are avoided due to CF.
4384 // When the code is vectorized we execute all code paths.
4385 if (Legal->blockNeedsPredication(*bb) && VF == 1)
4395 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4396 // If we know that this instruction will remain uniform, check the cost of
4397 // the scalar version.
4398 if (Legal->isUniformAfterVectorization(I))
4401 Type *RetTy = I->getType();
4402 Type *VectorTy = ToVectorTy(RetTy, VF);
4404 // TODO: We need to estimate the cost of intrinsic calls.
4405 switch (I->getOpcode()) {
4406 case Instruction::GetElementPtr:
4407 // We mark this instruction as zero-cost because the cost of GEPs in
4408 // vectorized code depends on whether the corresponding memory instruction
4409 // is scalarized or not. Therefore, we handle GEPs with the memory
4410 // instruction cost.
4412 case Instruction::Br: {
4413 return TTI.getCFInstrCost(I->getOpcode());
4415 case Instruction::PHI:
4416 //TODO: IF-converted IFs become selects.
4418 case Instruction::Add:
4419 case Instruction::FAdd:
4420 case Instruction::Sub:
4421 case Instruction::FSub:
4422 case Instruction::Mul:
4423 case Instruction::FMul:
4424 case Instruction::UDiv:
4425 case Instruction::SDiv:
4426 case Instruction::FDiv:
4427 case Instruction::URem:
4428 case Instruction::SRem:
4429 case Instruction::FRem:
4430 case Instruction::Shl:
4431 case Instruction::LShr:
4432 case Instruction::AShr:
4433 case Instruction::And:
4434 case Instruction::Or:
4435 case Instruction::Xor: {
4436 // Certain instructions can be cheaper to vectorize if they have a constant
4437 // second vector operand. One example of this are shifts on x86.
4438 TargetTransformInfo::OperandValueKind Op1VK =
4439 TargetTransformInfo::OK_AnyValue;
4440 TargetTransformInfo::OperandValueKind Op2VK =
4441 TargetTransformInfo::OK_AnyValue;
4443 if (isa<ConstantInt>(I->getOperand(1)))
4444 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4446 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4448 case Instruction::Select: {
4449 SelectInst *SI = cast<SelectInst>(I);
4450 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4451 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4452 Type *CondTy = SI->getCondition()->getType();
4454 CondTy = VectorType::get(CondTy, VF);
4456 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4458 case Instruction::ICmp:
4459 case Instruction::FCmp: {
4460 Type *ValTy = I->getOperand(0)->getType();
4461 VectorTy = ToVectorTy(ValTy, VF);
4462 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4464 case Instruction::Store:
4465 case Instruction::Load: {
4466 StoreInst *SI = dyn_cast<StoreInst>(I);
4467 LoadInst *LI = dyn_cast<LoadInst>(I);
4468 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4470 VectorTy = ToVectorTy(ValTy, VF);
4472 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4473 unsigned AS = SI ? SI->getPointerAddressSpace() :
4474 LI->getPointerAddressSpace();
4475 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4476 // We add the cost of address computation here instead of with the gep
4477 // instruction because only here we know whether the operation is
4480 return TTI.getAddressComputationCost(VectorTy) +
4481 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4483 // Scalarized loads/stores.
4484 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4485 bool Reverse = ConsecutiveStride < 0;
4486 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4487 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4488 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4490 // The cost of extracting from the value vector and pointer vector.
4491 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4492 for (unsigned i = 0; i < VF; ++i) {
4493 // The cost of extracting the pointer operand.
4494 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4495 // In case of STORE, the cost of ExtractElement from the vector.
4496 // In case of LOAD, the cost of InsertElement into the returned
4498 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4499 Instruction::InsertElement,
4503 // The cost of the scalar loads/stores.
4504 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
4505 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4510 // Wide load/stores.
4511 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4512 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4515 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4519 case Instruction::ZExt:
4520 case Instruction::SExt:
4521 case Instruction::FPToUI:
4522 case Instruction::FPToSI:
4523 case Instruction::FPExt:
4524 case Instruction::PtrToInt:
4525 case Instruction::IntToPtr:
4526 case Instruction::SIToFP:
4527 case Instruction::UIToFP:
4528 case Instruction::Trunc:
4529 case Instruction::FPTrunc:
4530 case Instruction::BitCast: {
4531 // We optimize the truncation of induction variable.
4532 // The cost of these is the same as the scalar operation.
4533 if (I->getOpcode() == Instruction::Trunc &&
4534 Legal->isInductionVariable(I->getOperand(0)))
4535 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4536 I->getOperand(0)->getType());
4538 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4539 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4541 case Instruction::Call: {
4542 CallInst *CI = cast<CallInst>(I);
4543 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4544 assert(ID && "Not an intrinsic call!");
4545 Type *RetTy = ToVectorTy(CI->getType(), VF);
4546 SmallVector<Type*, 4> Tys;
4547 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4548 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4549 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4552 // We are scalarizing the instruction. Return the cost of the scalar
4553 // instruction, plus the cost of insert and extract into vector
4554 // elements, times the vector width.
4557 if (!RetTy->isVoidTy() && VF != 1) {
4558 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4560 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4563 // The cost of inserting the results plus extracting each one of the
4565 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4568 // The cost of executing VF copies of the scalar instruction. This opcode
4569 // is unknown. Assume that it is the same as 'mul'.
4570 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4576 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4577 if (Scalar->isVoidTy() || VF == 1)
4579 return VectorType::get(Scalar, VF);
4582 char LoopVectorize::ID = 0;
4583 static const char lv_name[] = "Loop Vectorization";
4584 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4585 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4586 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4587 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4588 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4591 Pass *createLoopVectorizePass() {
4592 return new LoopVectorize();
4596 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4597 // Check for a store.
4598 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4599 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4601 // Check for a load.
4602 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4603 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;