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 /// Add code that checks at runtime if the accessed arrays overlap.
179 /// Returns the comparator value or NULL if no check is needed.
180 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
182 /// Create an empty loop, based on the loop ranges of the old loop.
183 void createEmptyLoop(LoopVectorizationLegality *Legal);
184 /// Copy and widen the instructions from the old loop.
185 void vectorizeLoop(LoopVectorizationLegality *Legal);
187 /// A helper function that computes the predicate of the block BB, assuming
188 /// that the header block of the loop is set to True. It returns the *entry*
189 /// mask for the block BB.
190 VectorParts createBlockInMask(BasicBlock *BB);
191 /// A helper function that computes the predicate of the edge between SRC
193 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
195 /// A helper function to vectorize a single BB within the innermost loop.
196 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
199 /// Insert the new loop to the loop hierarchy and pass manager
200 /// and update the analysis passes.
201 void updateAnalysis();
203 /// This instruction is un-vectorizable. Implement it as a sequence
205 void scalarizeInstruction(Instruction *Instr);
207 /// Vectorize Load and Store instructions,
208 void vectorizeMemoryInstruction(Instruction *Instr,
209 LoopVectorizationLegality *Legal);
211 /// Create a broadcast instruction. This method generates a broadcast
212 /// instruction (shuffle) for loop invariant values and for the induction
213 /// value. If this is the induction variable then we extend it to N, N+1, ...
214 /// this is needed because each iteration in the loop corresponds to a SIMD
216 Value *getBroadcastInstrs(Value *V);
218 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
219 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
220 /// The sequence starts at StartIndex.
221 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
223 /// When we go over instructions in the basic block we rely on previous
224 /// values within the current basic block or on loop invariant values.
225 /// When we widen (vectorize) values we place them in the map. If the values
226 /// are not within the map, they have to be loop invariant, so we simply
227 /// broadcast them into a vector.
228 VectorParts &getVectorValue(Value *V);
230 /// Generate a shuffle sequence that will reverse the vector Vec.
231 Value *reverseVector(Value *Vec);
233 /// This is a helper class that holds the vectorizer state. It maps scalar
234 /// instructions to vector instructions. When the code is 'unrolled' then
235 /// then a single scalar value is mapped to multiple vector parts. The parts
236 /// are stored in the VectorPart type.
238 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
240 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
242 /// \return True if 'Key' is saved in the Value Map.
243 bool has(Value *Key) const { return MapStorage.count(Key); }
245 /// Initializes a new entry in the map. Sets all of the vector parts to the
246 /// save value in 'Val'.
247 /// \return A reference to a vector with splat values.
248 VectorParts &splat(Value *Key, Value *Val) {
249 VectorParts &Entry = MapStorage[Key];
250 Entry.assign(UF, Val);
254 ///\return A reference to the value that is stored at 'Key'.
255 VectorParts &get(Value *Key) {
256 VectorParts &Entry = MapStorage[Key];
259 assert(Entry.size() == UF);
264 /// The unroll factor. Each entry in the map stores this number of vector
268 /// Map storage. We use std::map and not DenseMap because insertions to a
269 /// dense map invalidates its iterators.
270 std::map<Value *, VectorParts> MapStorage;
273 /// The original loop.
275 /// Scev analysis to use.
283 /// Target Library Info.
284 const TargetLibraryInfo *TLI;
286 /// The vectorization SIMD factor to use. Each vector will have this many
289 /// The vectorization unroll factor to use. Each scalar is vectorized to this
290 /// many different vector instructions.
293 /// The builder that we use
296 // --- Vectorization state ---
298 /// The vector-loop preheader.
299 BasicBlock *LoopVectorPreHeader;
300 /// The scalar-loop preheader.
301 BasicBlock *LoopScalarPreHeader;
302 /// Middle Block between the vector and the scalar.
303 BasicBlock *LoopMiddleBlock;
304 ///The ExitBlock of the scalar loop.
305 BasicBlock *LoopExitBlock;
306 ///The vector loop body.
307 BasicBlock *LoopVectorBody;
308 ///The scalar loop body.
309 BasicBlock *LoopScalarBody;
310 /// A list of all bypass blocks. The first block is the entry of the loop.
311 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
313 /// The new Induction variable which was added to the new block.
315 /// The induction variable of the old basic block.
316 PHINode *OldInduction;
317 /// Holds the extended (to the widest induction type) start index.
319 /// Maps scalars to widened vectors.
323 /// \brief Check if conditionally executed loads are hoistable.
325 /// This class has two functions: isHoistableLoad and canHoistAllLoads.
326 /// isHoistableLoad should be called on all load instructions that are executed
327 /// conditionally. After all conditional loads are processed, the client should
328 /// call canHoistAllLoads to determine if all of the conditional executed loads
329 /// have an unconditional memory access to the same memory address in the loop.
331 typedef SmallPtrSet<Value *, 8> MemorySet;
335 MemorySet CondLoadAddrSet;
338 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
340 /// \brief Check if the instruction is a load with a identifiable address.
341 bool isHoistableLoad(Instruction *L);
343 /// \brief Check if all of the conditional loads are hoistable because there
344 /// exists an unconditional memory access to the same address in the loop.
345 bool canHoistAllLoads();
348 bool LoadHoisting::isHoistableLoad(Instruction *L) {
349 LoadInst *LI = dyn_cast<LoadInst>(L);
353 CondLoadAddrSet.insert(LI->getPointerOperand());
357 static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
358 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
359 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
360 Set.insert(LI->getPointerOperand());
361 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
362 Set.insert(SI->getPointerOperand());
366 bool LoadHoisting::canHoistAllLoads() {
367 // No conditional loads.
368 if (CondLoadAddrSet.empty())
371 MemorySet UncondMemAccesses;
372 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
373 BasicBlock *LoopLatch = TheLoop->getLoopLatch();
375 // Iterate over the unconditional blocks and collect memory access addresses.
376 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
377 BasicBlock *BB = LoopBlocks[i];
379 // Ignore conditional blocks.
380 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
383 addMemAccesses(BB, UncondMemAccesses);
386 // And make sure there is a matching unconditional access for every
388 for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
389 ME = CondLoadAddrSet.end(); MI != ME; ++MI)
390 if (!UncondMemAccesses.count(*MI))
396 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
397 /// to what vectorization factor.
398 /// This class does not look at the profitability of vectorization, only the
399 /// legality. This class has two main kinds of checks:
400 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
401 /// will change the order of memory accesses in a way that will change the
402 /// correctness of the program.
403 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
404 /// checks for a number of different conditions, such as the availability of a
405 /// single induction variable, that all types are supported and vectorize-able,
406 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
407 /// This class is also used by InnerLoopVectorizer for identifying
408 /// induction variable and the different reduction variables.
409 class LoopVectorizationLegality {
411 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
412 DominatorTree *DT, TargetLibraryInfo *TLI)
413 : TheLoop(L), SE(SE), DL(DL), DT(DT), TLI(TLI),
414 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
415 MaxSafeDepDistBytes(-1U), LoadSpeculation(L, DT) {}
417 /// This enum represents the kinds of reductions that we support.
419 RK_NoReduction, ///< Not a reduction.
420 RK_IntegerAdd, ///< Sum of integers.
421 RK_IntegerMult, ///< Product of integers.
422 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
423 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
424 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
425 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
426 RK_FloatAdd, ///< Sum of floats.
427 RK_FloatMult, ///< Product of floats.
428 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
431 /// This enum represents the kinds of inductions that we support.
433 IK_NoInduction, ///< Not an induction variable.
434 IK_IntInduction, ///< Integer induction variable. Step = 1.
435 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
436 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
437 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
440 // This enum represents the kind of minmax reduction.
441 enum MinMaxReductionKind {
451 /// This POD struct holds information about reduction variables.
452 struct ReductionDescriptor {
453 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
454 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
456 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
457 MinMaxReductionKind MK)
458 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
460 // The starting value of the reduction.
461 // It does not have to be zero!
462 TrackingVH<Value> StartValue;
463 // The instruction who's value is used outside the loop.
464 Instruction *LoopExitInstr;
465 // The kind of the reduction.
467 // If this a min/max reduction the kind of reduction.
468 MinMaxReductionKind MinMaxKind;
471 /// This POD struct holds information about a potential reduction operation.
472 struct ReductionInstDesc {
473 ReductionInstDesc(bool IsRedux, Instruction *I) :
474 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
476 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
477 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
479 // Is this instruction a reduction candidate.
481 // The last instruction in a min/max pattern (select of the select(icmp())
482 // pattern), or the current reduction instruction otherwise.
483 Instruction *PatternLastInst;
484 // If this is a min/max pattern the comparison predicate.
485 MinMaxReductionKind MinMaxKind;
488 // This POD struct holds information about the memory runtime legality
489 // check that a group of pointers do not overlap.
490 struct RuntimePointerCheck {
491 RuntimePointerCheck() : Need(false) {}
493 /// Reset the state of the pointer runtime information.
501 /// Insert a pointer and calculate the start and end SCEVs.
502 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
505 /// This flag indicates if we need to add the runtime check.
507 /// Holds the pointers that we need to check.
508 SmallVector<TrackingVH<Value>, 2> Pointers;
509 /// Holds the pointer value at the beginning of the loop.
510 SmallVector<const SCEV*, 2> Starts;
511 /// Holds the pointer value at the end of the loop.
512 SmallVector<const SCEV*, 2> Ends;
513 /// Holds the information if this pointer is used for writing to memory.
514 SmallVector<bool, 2> IsWritePtr;
515 /// Holds the id of the set of pointers that could be dependent because of a
516 /// shared underlying object.
517 SmallVector<unsigned, 2> DependencySetId;
520 /// A POD for saving information about induction variables.
521 struct InductionInfo {
522 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
523 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
525 TrackingVH<Value> StartValue;
530 /// ReductionList contains the reduction descriptors for all
531 /// of the reductions that were found in the loop.
532 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
534 /// InductionList saves induction variables and maps them to the
535 /// induction descriptor.
536 typedef MapVector<PHINode*, InductionInfo> InductionList;
538 /// Returns true if it is legal to vectorize this loop.
539 /// This does not mean that it is profitable to vectorize this
540 /// loop, only that it is legal to do so.
543 /// Returns the Induction variable.
544 PHINode *getInduction() { return Induction; }
546 /// Returns the reduction variables found in the loop.
547 ReductionList *getReductionVars() { return &Reductions; }
549 /// Returns the induction variables found in the loop.
550 InductionList *getInductionVars() { return &Inductions; }
552 /// Returns the widest induction type.
553 Type *getWidestInductionType() { return WidestIndTy; }
555 /// Returns True if V is an induction variable in this loop.
556 bool isInductionVariable(const Value *V);
558 /// Return true if the block BB needs to be predicated in order for the loop
559 /// to be vectorized.
560 bool blockNeedsPredication(BasicBlock *BB);
562 /// Check if this pointer is consecutive when vectorizing. This happens
563 /// when the last index of the GEP is the induction variable, or that the
564 /// pointer itself is an induction variable.
565 /// This check allows us to vectorize A[idx] into a wide load/store.
567 /// 0 - Stride is unknown or non consecutive.
568 /// 1 - Address is consecutive.
569 /// -1 - Address is consecutive, and decreasing.
570 int isConsecutivePtr(Value *Ptr);
572 /// Returns true if the value V is uniform within the loop.
573 bool isUniform(Value *V);
575 /// Returns true if this instruction will remain scalar after vectorization.
576 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
578 /// Returns the information that we collected about runtime memory check.
579 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
581 /// This function returns the identity element (or neutral element) for
583 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
585 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
588 /// Check if a single basic block loop is vectorizable.
589 /// At this point we know that this is a loop with a constant trip count
590 /// and we only need to check individual instructions.
591 bool canVectorizeInstrs();
593 /// When we vectorize loops we may change the order in which
594 /// we read and write from memory. This method checks if it is
595 /// legal to vectorize the code, considering only memory constrains.
596 /// Returns true if the loop is vectorizable
597 bool canVectorizeMemory();
599 /// Return true if we can vectorize this loop using the IF-conversion
601 bool canVectorizeWithIfConvert();
603 /// Collect the variables that need to stay uniform after vectorization.
604 void collectLoopUniforms();
606 /// Return true if all of the instructions in the block can be speculatively
608 bool blockCanBePredicated(BasicBlock *BB);
610 /// Returns True, if 'Phi' is the kind of reduction variable for type
611 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
612 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
613 /// Returns a struct describing if the instruction 'I' can be a reduction
614 /// variable of type 'Kind'. If the reduction is a min/max pattern of
615 /// select(icmp()) this function advances the instruction pointer 'I' from the
616 /// compare instruction to the select instruction and stores this pointer in
617 /// 'PatternLastInst' member of the returned struct.
618 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
619 ReductionInstDesc &Desc);
620 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
621 /// pattern corresponding to a min(X, Y) or max(X, Y).
622 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
623 ReductionInstDesc &Prev);
624 /// Returns the induction kind of Phi. This function may return NoInduction
625 /// if the PHI is not an induction variable.
626 InductionKind isInductionVariable(PHINode *Phi);
628 /// The loop that we evaluate.
632 /// DataLayout analysis.
636 /// Target Library Info.
637 TargetLibraryInfo *TLI;
639 // --- vectorization state --- //
641 /// Holds the integer induction variable. This is the counter of the
644 /// Holds the reduction variables.
645 ReductionList Reductions;
646 /// Holds all of the induction variables that we found in the loop.
647 /// Notice that inductions don't need to start at zero and that induction
648 /// variables can be pointers.
649 InductionList Inductions;
650 /// Holds the widest induction type encountered.
653 /// Allowed outside users. This holds the reduction
654 /// vars which can be accessed from outside the loop.
655 SmallPtrSet<Value*, 4> AllowedExit;
656 /// This set holds the variables which are known to be uniform after
658 SmallPtrSet<Instruction*, 4> Uniforms;
659 /// We need to check that all of the pointers in this list are disjoint
661 RuntimePointerCheck PtrRtCheck;
662 /// Can we assume the absence of NaNs.
663 bool HasFunNoNaNAttr;
665 unsigned MaxSafeDepDistBytes;
667 /// Utility to determine whether loads can be speculated.
668 LoadHoisting LoadSpeculation;
671 /// LoopVectorizationCostModel - estimates the expected speedups due to
673 /// In many cases vectorization is not profitable. This can happen because of
674 /// a number of reasons. In this class we mainly attempt to predict the
675 /// expected speedup/slowdowns due to the supported instruction set. We use the
676 /// TargetTransformInfo to query the different backends for the cost of
677 /// different operations.
678 class LoopVectorizationCostModel {
680 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
681 LoopVectorizationLegality *Legal,
682 const TargetTransformInfo &TTI,
683 DataLayout *DL, const TargetLibraryInfo *TLI)
684 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
686 /// Information about vectorization costs
687 struct VectorizationFactor {
688 unsigned Width; // Vector width with best cost
689 unsigned Cost; // Cost of the loop with that width
691 /// \return The most profitable vectorization factor and the cost of that VF.
692 /// This method checks every power of two up to VF. If UserVF is not ZERO
693 /// then this vectorization factor will be selected if vectorization is
695 VectorizationFactor selectVectorizationFactor(bool OptForSize,
698 /// \return The size (in bits) of the widest type in the code that
699 /// needs to be vectorized. We ignore values that remain scalar such as
700 /// 64 bit loop indices.
701 unsigned getWidestType();
703 /// \return The most profitable unroll factor.
704 /// If UserUF is non-zero then this method finds the best unroll-factor
705 /// based on register pressure and other parameters.
706 /// VF and LoopCost are the selected vectorization factor and the cost of the
708 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
711 /// \brief A struct that represents some properties of the register usage
713 struct RegisterUsage {
714 /// Holds the number of loop invariant values that are used in the loop.
715 unsigned LoopInvariantRegs;
716 /// Holds the maximum number of concurrent live intervals in the loop.
717 unsigned MaxLocalUsers;
718 /// Holds the number of instructions in the loop.
719 unsigned NumInstructions;
722 /// \return information about the register usage of the loop.
723 RegisterUsage calculateRegisterUsage();
726 /// Returns the expected execution cost. The unit of the cost does
727 /// not matter because we use the 'cost' units to compare different
728 /// vector widths. The cost that is returned is *not* normalized by
729 /// the factor width.
730 unsigned expectedCost(unsigned VF);
732 /// Returns the execution time cost of an instruction for a given vector
733 /// width. Vector width of one means scalar.
734 unsigned getInstructionCost(Instruction *I, unsigned VF);
736 /// A helper function for converting Scalar types to vector types.
737 /// If the incoming type is void, we return void. If the VF is 1, we return
739 static Type* ToVectorTy(Type *Scalar, unsigned VF);
741 /// Returns whether the instruction is a load or store and will be a emitted
742 /// as a vector operation.
743 bool isConsecutiveLoadOrStore(Instruction *I);
745 /// The loop that we evaluate.
749 /// Loop Info analysis.
751 /// Vectorization legality.
752 LoopVectorizationLegality *Legal;
753 /// Vector target information.
754 const TargetTransformInfo &TTI;
755 /// Target data layout information.
757 /// Target Library Info.
758 const TargetLibraryInfo *TLI;
761 /// Utility class for getting and setting loop vectorizer hints in the form
762 /// of loop metadata.
763 struct LoopVectorizeHints {
764 /// Vectorization width.
766 /// Vectorization unroll factor.
769 LoopVectorizeHints(const Loop *L)
770 : Width(VectorizationFactor)
771 , Unroll(VectorizationUnroll)
772 , LoopID(L->getLoopID()) {
774 // The command line options override any loop metadata except for when
775 // width == 1 which is used to indicate the loop is already vectorized.
776 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
777 Width = VectorizationFactor;
778 if (VectorizationUnroll.getNumOccurrences() > 0)
779 Unroll = VectorizationUnroll;
782 /// Return the loop vectorizer metadata prefix.
783 static StringRef Prefix() { return "llvm.vectorizer."; }
785 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
786 SmallVector<Value*, 2> Vals;
787 Vals.push_back(MDString::get(Context, Name));
788 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
789 return MDNode::get(Context, Vals);
792 /// Mark the loop L as already vectorized by setting the width to 1.
793 void setAlreadyVectorized(Loop *L) {
794 LLVMContext &Context = L->getHeader()->getContext();
798 // Create a new loop id with one more operand for the already_vectorized
799 // hint. If the loop already has a loop id then copy the existing operands.
800 SmallVector<Value*, 4> Vals(1);
802 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
803 Vals.push_back(LoopID->getOperand(i));
805 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
807 MDNode *NewLoopID = MDNode::get(Context, Vals);
808 // Set operand 0 to refer to the loop id itself.
809 NewLoopID->replaceOperandWith(0, NewLoopID);
811 L->setLoopID(NewLoopID);
813 LoopID->replaceAllUsesWith(NewLoopID);
821 /// Find hints specified in the loop metadata.
822 void getHints(const Loop *L) {
826 // First operand should refer to the loop id itself.
827 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
828 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
830 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
831 const MDString *S = 0;
832 SmallVector<Value*, 4> Args;
834 // The expected hint is either a MDString or a MDNode with the first
835 // operand a MDString.
836 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
837 if (!MD || MD->getNumOperands() == 0)
839 S = dyn_cast<MDString>(MD->getOperand(0));
840 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
841 Args.push_back(MD->getOperand(i));
843 S = dyn_cast<MDString>(LoopID->getOperand(i));
844 assert(Args.size() == 0 && "too many arguments for MDString");
850 // Check if the hint starts with the vectorizer prefix.
851 StringRef Hint = S->getString();
852 if (!Hint.startswith(Prefix()))
854 // Remove the prefix.
855 Hint = Hint.substr(Prefix().size(), StringRef::npos);
857 if (Args.size() == 1)
858 getHint(Hint, Args[0]);
862 // Check string hint with one operand.
863 void getHint(StringRef Hint, Value *Arg) {
864 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
866 unsigned Val = C->getZExtValue();
868 if (Hint == "width") {
869 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
870 "Invalid width metadata");
872 } else if (Hint == "unroll") {
873 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
874 "Invalid unroll metadata");
877 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
881 /// The LoopVectorize Pass.
882 struct LoopVectorize : public LoopPass {
883 /// Pass identification, replacement for typeid
886 explicit LoopVectorize() : LoopPass(ID) {
887 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
893 TargetTransformInfo *TTI;
895 TargetLibraryInfo *TLI;
897 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
898 // We only vectorize innermost loops.
902 SE = &getAnalysis<ScalarEvolution>();
903 DL = getAnalysisIfAvailable<DataLayout>();
904 LI = &getAnalysis<LoopInfo>();
905 TTI = &getAnalysis<TargetTransformInfo>();
906 DT = &getAnalysis<DominatorTree>();
907 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
910 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
914 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
915 L->getHeader()->getParent()->getName() << "\"\n");
917 LoopVectorizeHints Hints(L);
919 if (Hints.Width == 1) {
920 DEBUG(dbgs() << "LV: Not vectorizing.\n");
924 // Check if it is legal to vectorize the loop.
925 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
926 if (!LVL.canVectorize()) {
927 DEBUG(dbgs() << "LV: Not vectorizing.\n");
931 // Use the cost model.
932 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
934 // Check the function attributes to find out if this function should be
935 // optimized for size.
936 Function *F = L->getHeader()->getParent();
937 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
938 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
939 unsigned FnIndex = AttributeSet::FunctionIndex;
940 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
941 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
944 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
945 "attribute is used.\n");
949 // Select the optimal vectorization factor.
950 LoopVectorizationCostModel::VectorizationFactor VF;
951 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
952 // Select the unroll factor.
953 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
957 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
961 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
962 F->getParent()->getModuleIdentifier()<<"\n");
963 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
965 // If we decided that it is *legal* to vectorize the loop then do it.
966 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
969 // Mark the loop as already vectorized to avoid vectorizing again.
970 Hints.setAlreadyVectorized(L);
972 DEBUG(verifyFunction(*L->getHeader()->getParent()));
976 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
977 LoopPass::getAnalysisUsage(AU);
978 AU.addRequiredID(LoopSimplifyID);
979 AU.addRequiredID(LCSSAID);
980 AU.addRequired<DominatorTree>();
981 AU.addRequired<LoopInfo>();
982 AU.addRequired<ScalarEvolution>();
983 AU.addRequired<TargetTransformInfo>();
984 AU.addPreserved<LoopInfo>();
985 AU.addPreserved<DominatorTree>();
990 } // end anonymous namespace
992 //===----------------------------------------------------------------------===//
993 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
994 // LoopVectorizationCostModel.
995 //===----------------------------------------------------------------------===//
998 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
999 Loop *Lp, Value *Ptr,
1001 unsigned DepSetId) {
1002 const SCEV *Sc = SE->getSCEV(Ptr);
1003 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1004 assert(AR && "Invalid addrec expression");
1005 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1006 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1007 Pointers.push_back(Ptr);
1008 Starts.push_back(AR->getStart());
1009 Ends.push_back(ScEnd);
1010 IsWritePtr.push_back(WritePtr);
1011 DependencySetId.push_back(DepSetId);
1014 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1015 // Save the current insertion location.
1016 Instruction *Loc = Builder.GetInsertPoint();
1018 // We need to place the broadcast of invariant variables outside the loop.
1019 Instruction *Instr = dyn_cast<Instruction>(V);
1020 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1021 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1023 // Place the code for broadcasting invariant variables in the new preheader.
1025 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1027 // Broadcast the scalar into all locations in the vector.
1028 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1030 // Restore the builder insertion point.
1032 Builder.SetInsertPoint(Loc);
1037 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1039 assert(Val->getType()->isVectorTy() && "Must be a vector");
1040 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1041 "Elem must be an integer");
1042 // Create the types.
1043 Type *ITy = Val->getType()->getScalarType();
1044 VectorType *Ty = cast<VectorType>(Val->getType());
1045 int VLen = Ty->getNumElements();
1046 SmallVector<Constant*, 8> Indices;
1048 // Create a vector of consecutive numbers from zero to VF.
1049 for (int i = 0; i < VLen; ++i) {
1050 int64_t Idx = Negate ? (-i) : i;
1051 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1054 // Add the consecutive indices to the vector value.
1055 Constant *Cv = ConstantVector::get(Indices);
1056 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1057 return Builder.CreateAdd(Val, Cv, "induction");
1060 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1061 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1062 // Make sure that the pointer does not point to structs.
1063 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1066 // If this value is a pointer induction variable we know it is consecutive.
1067 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1068 if (Phi && Inductions.count(Phi)) {
1069 InductionInfo II = Inductions[Phi];
1070 if (IK_PtrInduction == II.IK)
1072 else if (IK_ReversePtrInduction == II.IK)
1076 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1080 unsigned NumOperands = Gep->getNumOperands();
1081 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1083 Value *GpPtr = Gep->getPointerOperand();
1084 // If this GEP value is a consecutive pointer induction variable and all of
1085 // the indices are constant then we know it is consecutive. We can
1086 Phi = dyn_cast<PHINode>(GpPtr);
1087 if (Phi && Inductions.count(Phi)) {
1089 // Make sure that the pointer does not point to structs.
1090 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1091 if (GepPtrType->getElementType()->isAggregateType())
1094 // Make sure that all of the index operands are loop invariant.
1095 for (unsigned i = 1; i < NumOperands; ++i)
1096 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1099 InductionInfo II = Inductions[Phi];
1100 if (IK_PtrInduction == II.IK)
1102 else if (IK_ReversePtrInduction == II.IK)
1106 // Check that all of the gep indices are uniform except for the last.
1107 for (unsigned i = 0; i < NumOperands - 1; ++i)
1108 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1111 // We can emit wide load/stores only if the last index is the induction
1113 const SCEV *Last = SE->getSCEV(LastIndex);
1114 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1115 const SCEV *Step = AR->getStepRecurrence(*SE);
1117 // The memory is consecutive because the last index is consecutive
1118 // and all other indices are loop invariant.
1121 if (Step->isAllOnesValue())
1128 bool LoopVectorizationLegality::isUniform(Value *V) {
1129 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1132 InnerLoopVectorizer::VectorParts&
1133 InnerLoopVectorizer::getVectorValue(Value *V) {
1134 assert(V != Induction && "The new induction variable should not be used.");
1135 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1137 // If we have this scalar in the map, return it.
1138 if (WidenMap.has(V))
1139 return WidenMap.get(V);
1141 // If this scalar is unknown, assume that it is a constant or that it is
1142 // loop invariant. Broadcast V and save the value for future uses.
1143 Value *B = getBroadcastInstrs(V);
1144 return WidenMap.splat(V, B);
1147 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1148 assert(Vec->getType()->isVectorTy() && "Invalid type");
1149 SmallVector<Constant*, 8> ShuffleMask;
1150 for (unsigned i = 0; i < VF; ++i)
1151 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1153 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1154 ConstantVector::get(ShuffleMask),
1159 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1160 LoopVectorizationLegality *Legal) {
1161 // Attempt to issue a wide load.
1162 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1163 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1165 assert((LI || SI) && "Invalid Load/Store instruction");
1167 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1168 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1169 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1170 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1171 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1172 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1173 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1175 if (ScalarAllocatedSize != VectorElementSize)
1176 return scalarizeInstruction(Instr);
1178 // If the pointer is loop invariant or if it is non consecutive,
1179 // scalarize the load.
1180 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1181 bool Reverse = ConsecutiveStride < 0;
1182 bool UniformLoad = LI && Legal->isUniform(Ptr);
1183 if (!ConsecutiveStride || UniformLoad)
1184 return scalarizeInstruction(Instr);
1186 Constant *Zero = Builder.getInt32(0);
1187 VectorParts &Entry = WidenMap.get(Instr);
1189 // Handle consecutive loads/stores.
1190 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1191 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1192 Value *PtrOperand = Gep->getPointerOperand();
1193 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1194 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1196 // Create the new GEP with the new induction variable.
1197 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1198 Gep2->setOperand(0, FirstBasePtr);
1199 Gep2->setName("gep.indvar.base");
1200 Ptr = Builder.Insert(Gep2);
1202 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1203 OrigLoop) && "Base ptr must be invariant");
1205 // The last index does not have to be the induction. It can be
1206 // consecutive and be a function of the index. For example A[I+1];
1207 unsigned NumOperands = Gep->getNumOperands();
1208 unsigned LastOperand = NumOperands - 1;
1209 // Create the new GEP with the new induction variable.
1210 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1212 for (unsigned i = 0; i < NumOperands; ++i) {
1213 Value *GepOperand = Gep->getOperand(i);
1214 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1216 // Update last index or loop invariant instruction anchored in loop.
1217 if (i == LastOperand ||
1218 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1219 assert((i == LastOperand ||
1220 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1221 "Must be last index or loop invariant");
1223 VectorParts &GEPParts = getVectorValue(GepOperand);
1224 Value *Index = GEPParts[0];
1225 Index = Builder.CreateExtractElement(Index, Zero);
1226 Gep2->setOperand(i, Index);
1227 Gep2->setName("gep.indvar.idx");
1230 Ptr = Builder.Insert(Gep2);
1232 // Use the induction element ptr.
1233 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1234 VectorParts &PtrVal = getVectorValue(Ptr);
1235 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1240 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1241 "We do not allow storing to uniform addresses");
1242 // We don't want to update the value in the map as it might be used in
1243 // another expression. So don't use a reference type for "StoredVal".
1244 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1246 for (unsigned Part = 0; Part < UF; ++Part) {
1247 // Calculate the pointer for the specific unroll-part.
1248 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1251 // If we store to reverse consecutive memory locations then we need
1252 // to reverse the order of elements in the stored value.
1253 StoredVal[Part] = reverseVector(StoredVal[Part]);
1254 // If the address is consecutive but reversed, then the
1255 // wide store needs to start at the last vector element.
1256 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1257 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1260 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1261 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1266 for (unsigned Part = 0; Part < UF; ++Part) {
1267 // Calculate the pointer for the specific unroll-part.
1268 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1271 // If the address is consecutive but reversed, then the
1272 // wide store needs to start at the last vector element.
1273 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1274 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1277 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1278 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1279 cast<LoadInst>(LI)->setAlignment(Alignment);
1280 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1284 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1285 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1286 // Holds vector parameters or scalars, in case of uniform vals.
1287 SmallVector<VectorParts, 4> Params;
1289 // Find all of the vectorized parameters.
1290 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1291 Value *SrcOp = Instr->getOperand(op);
1293 // If we are accessing the old induction variable, use the new one.
1294 if (SrcOp == OldInduction) {
1295 Params.push_back(getVectorValue(SrcOp));
1299 // Try using previously calculated values.
1300 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1302 // If the src is an instruction that appeared earlier in the basic block
1303 // then it should already be vectorized.
1304 if (SrcInst && OrigLoop->contains(SrcInst)) {
1305 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1306 // The parameter is a vector value from earlier.
1307 Params.push_back(WidenMap.get(SrcInst));
1309 // The parameter is a scalar from outside the loop. Maybe even a constant.
1310 VectorParts Scalars;
1311 Scalars.append(UF, SrcOp);
1312 Params.push_back(Scalars);
1316 assert(Params.size() == Instr->getNumOperands() &&
1317 "Invalid number of operands");
1319 // Does this instruction return a value ?
1320 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1322 Value *UndefVec = IsVoidRetTy ? 0 :
1323 UndefValue::get(VectorType::get(Instr->getType(), VF));
1324 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1325 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1327 // For each vector unroll 'part':
1328 for (unsigned Part = 0; Part < UF; ++Part) {
1329 // For each scalar that we create:
1330 for (unsigned Width = 0; Width < VF; ++Width) {
1331 Instruction *Cloned = Instr->clone();
1333 Cloned->setName(Instr->getName() + ".cloned");
1334 // Replace the operands of the cloned instrucions with extracted scalars.
1335 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1336 Value *Op = Params[op][Part];
1337 // Param is a vector. Need to extract the right lane.
1338 if (Op->getType()->isVectorTy())
1339 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1340 Cloned->setOperand(op, Op);
1343 // Place the cloned scalar in the new loop.
1344 Builder.Insert(Cloned);
1346 // If the original scalar returns a value we need to place it in a vector
1347 // so that future users will be able to use it.
1349 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1350 Builder.getInt32(Width));
1356 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1358 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1359 Legal->getRuntimePointerCheck();
1361 if (!PtrRtCheck->Need)
1364 unsigned NumPointers = PtrRtCheck->Pointers.size();
1365 SmallVector<TrackingVH<Value> , 2> Starts;
1366 SmallVector<TrackingVH<Value> , 2> Ends;
1368 SCEVExpander Exp(*SE, "induction");
1370 // Use this type for pointer arithmetic.
1371 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1373 for (unsigned i = 0; i < NumPointers; ++i) {
1374 Value *Ptr = PtrRtCheck->Pointers[i];
1375 const SCEV *Sc = SE->getSCEV(Ptr);
1377 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1378 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1380 Starts.push_back(Ptr);
1381 Ends.push_back(Ptr);
1383 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1385 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1386 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1387 Starts.push_back(Start);
1388 Ends.push_back(End);
1392 IRBuilder<> ChkBuilder(Loc);
1393 // Our instructions might fold to a constant.
1394 Value *MemoryRuntimeCheck = 0;
1395 for (unsigned i = 0; i < NumPointers; ++i) {
1396 for (unsigned j = i+1; j < NumPointers; ++j) {
1397 // No need to check if two readonly pointers intersect.
1398 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1401 // Only need to check pointers between two different dependency sets.
1402 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1405 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1406 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1407 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1408 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1410 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1411 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1412 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1413 if (MemoryRuntimeCheck)
1414 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1416 MemoryRuntimeCheck = IsConflict;
1420 // We have to do this trickery because the IRBuilder might fold the check to a
1421 // constant expression in which case there is no Instruction anchored in a
1423 LLVMContext &Ctx = Loc->getContext();
1424 Instruction * Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1425 ConstantInt::getTrue(Ctx));
1426 ChkBuilder.Insert(Check, "memcheck.conflict");
1431 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1433 In this function we generate a new loop. The new loop will contain
1434 the vectorized instructions while the old loop will continue to run the
1437 [ ] <-- vector loop bypass (may consist of multiple blocks).
1440 | [ ] <-- vector pre header.
1444 | [ ]_| <-- vector loop.
1447 >[ ] <--- middle-block.
1450 | [ ] <--- new preheader.
1454 | [ ]_| <-- old scalar loop to handle remainder.
1457 >[ ] <-- exit block.
1461 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1462 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1463 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1464 assert(ExitBlock && "Must have an exit block");
1466 // Some loops have a single integer induction variable, while other loops
1467 // don't. One example is c++ iterators that often have multiple pointer
1468 // induction variables. In the code below we also support a case where we
1469 // don't have a single induction variable.
1470 OldInduction = Legal->getInduction();
1471 Type *IdxTy = Legal->getWidestInductionType();
1473 // Find the loop boundaries.
1474 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1475 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1477 // Get the total trip count from the count by adding 1.
1478 ExitCount = SE->getAddExpr(ExitCount,
1479 SE->getConstant(ExitCount->getType(), 1));
1481 // Expand the trip count and place the new instructions in the preheader.
1482 // Notice that the pre-header does not change, only the loop body.
1483 SCEVExpander Exp(*SE, "induction");
1485 // Count holds the overall loop count (N).
1486 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1487 BypassBlock->getTerminator());
1489 // The loop index does not have to start at Zero. Find the original start
1490 // value from the induction PHI node. If we don't have an induction variable
1491 // then we know that it starts at zero.
1492 Builder.SetInsertPoint(BypassBlock->getTerminator());
1493 Value *StartIdx = ExtendedIdx = OldInduction ?
1494 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1496 ConstantInt::get(IdxTy, 0);
1498 assert(BypassBlock && "Invalid loop structure");
1499 LoopBypassBlocks.push_back(BypassBlock);
1501 // Split the single block loop into the two loop structure described above.
1502 BasicBlock *VectorPH =
1503 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1504 BasicBlock *VecBody =
1505 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1506 BasicBlock *MiddleBlock =
1507 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1508 BasicBlock *ScalarPH =
1509 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1511 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1513 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1515 // Generate the induction variable.
1516 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1517 // The loop step is equal to the vectorization factor (num of SIMD elements)
1518 // times the unroll factor (num of SIMD instructions).
1519 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1521 // This is the IR builder that we use to add all of the logic for bypassing
1522 // the new vector loop.
1523 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1525 // We may need to extend the index in case there is a type mismatch.
1526 // We know that the count starts at zero and does not overflow.
1527 if (Count->getType() != IdxTy) {
1528 // The exit count can be of pointer type. Convert it to the correct
1530 if (ExitCount->getType()->isPointerTy())
1531 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1533 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1536 // Add the start index to the loop count to get the new end index.
1537 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1539 // Now we need to generate the expression for N - (N % VF), which is
1540 // the part that the vectorized body will execute.
1541 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1542 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1543 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1544 "end.idx.rnd.down");
1546 // Now, compare the new count to zero. If it is zero skip the vector loop and
1547 // jump to the scalar loop.
1548 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1551 BasicBlock *LastBypassBlock = BypassBlock;
1553 // Generate the code that checks in runtime if arrays overlap. We put the
1554 // checks into a separate block to make the more common case of few elements
1556 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1557 BypassBlock->getTerminator());
1558 if (MemRuntimeCheck) {
1559 // Create a new block containing the memory check.
1560 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1562 LoopBypassBlocks.push_back(CheckBlock);
1564 // Replace the branch into the memory check block with a conditional branch
1565 // for the "few elements case".
1566 Instruction *OldTerm = BypassBlock->getTerminator();
1567 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1568 OldTerm->eraseFromParent();
1570 Cmp = MemRuntimeCheck;
1571 LastBypassBlock = CheckBlock;
1574 LastBypassBlock->getTerminator()->eraseFromParent();
1575 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1578 // We are going to resume the execution of the scalar loop.
1579 // Go over all of the induction variables that we found and fix the
1580 // PHIs that are left in the scalar version of the loop.
1581 // The starting values of PHI nodes depend on the counter of the last
1582 // iteration in the vectorized loop.
1583 // If we come from a bypass edge then we need to start from the original
1586 // This variable saves the new starting index for the scalar loop.
1587 PHINode *ResumeIndex = 0;
1588 LoopVectorizationLegality::InductionList::iterator I, E;
1589 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1590 // Set builder to point to last bypass block.
1591 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1592 for (I = List->begin(), E = List->end(); I != E; ++I) {
1593 PHINode *OrigPhi = I->first;
1594 LoopVectorizationLegality::InductionInfo II = I->second;
1596 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1597 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1598 MiddleBlock->getTerminator());
1599 // We might have extended the type of the induction variable but we need a
1600 // truncated version for the scalar loop.
1601 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1602 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1603 MiddleBlock->getTerminator()) : 0;
1605 Value *EndValue = 0;
1607 case LoopVectorizationLegality::IK_NoInduction:
1608 llvm_unreachable("Unknown induction");
1609 case LoopVectorizationLegality::IK_IntInduction: {
1610 // Handle the integer induction counter.
1611 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1613 // We have the canonical induction variable.
1614 if (OrigPhi == OldInduction) {
1615 // Create a truncated version of the resume value for the scalar loop,
1616 // we might have promoted the type to a larger width.
1618 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1619 // The new PHI merges the original incoming value, in case of a bypass,
1620 // or the value at the end of the vectorized loop.
1621 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1622 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1623 TruncResumeVal->addIncoming(EndValue, VecBody);
1625 // We know what the end value is.
1626 EndValue = IdxEndRoundDown;
1627 // We also know which PHI node holds it.
1628 ResumeIndex = ResumeVal;
1632 // Not the canonical induction variable - add the vector loop count to the
1634 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1635 II.StartValue->getType(),
1637 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1640 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1641 // Convert the CountRoundDown variable to the PHI size.
1642 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1643 II.StartValue->getType(),
1645 // Handle reverse integer induction counter.
1646 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1649 case LoopVectorizationLegality::IK_PtrInduction: {
1650 // For pointer induction variables, calculate the offset using
1652 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1656 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1657 // The value at the end of the loop for the reverse pointer is calculated
1658 // by creating a GEP with a negative index starting from the start value.
1659 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1660 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1662 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1668 // The new PHI merges the original incoming value, in case of a bypass,
1669 // or the value at the end of the vectorized loop.
1670 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1671 if (OrigPhi == OldInduction)
1672 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1674 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1676 ResumeVal->addIncoming(EndValue, VecBody);
1678 // Fix the scalar body counter (PHI node).
1679 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1680 // The old inductions phi node in the scalar body needs the truncated value.
1681 if (OrigPhi == OldInduction)
1682 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1684 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1687 // If we are generating a new induction variable then we also need to
1688 // generate the code that calculates the exit value. This value is not
1689 // simply the end of the counter because we may skip the vectorized body
1690 // in case of a runtime check.
1692 assert(!ResumeIndex && "Unexpected resume value found");
1693 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1694 MiddleBlock->getTerminator());
1695 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1696 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1697 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1700 // Make sure that we found the index where scalar loop needs to continue.
1701 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1702 "Invalid resume Index");
1704 // Add a check in the middle block to see if we have completed
1705 // all of the iterations in the first vector loop.
1706 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1707 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1708 ResumeIndex, "cmp.n",
1709 MiddleBlock->getTerminator());
1711 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1712 // Remove the old terminator.
1713 MiddleBlock->getTerminator()->eraseFromParent();
1715 // Create i+1 and fill the PHINode.
1716 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1717 Induction->addIncoming(StartIdx, VectorPH);
1718 Induction->addIncoming(NextIdx, VecBody);
1719 // Create the compare.
1720 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1721 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1723 // Now we have two terminators. Remove the old one from the block.
1724 VecBody->getTerminator()->eraseFromParent();
1726 // Get ready to start creating new instructions into the vectorized body.
1727 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1729 // Create and register the new vector loop.
1730 Loop* Lp = new Loop();
1731 Loop *ParentLoop = OrigLoop->getParentLoop();
1733 // Insert the new loop into the loop nest and register the new basic blocks.
1735 ParentLoop->addChildLoop(Lp);
1736 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1737 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1738 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1739 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1740 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1742 LI->addTopLevelLoop(Lp);
1745 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1748 LoopVectorPreHeader = VectorPH;
1749 LoopScalarPreHeader = ScalarPH;
1750 LoopMiddleBlock = MiddleBlock;
1751 LoopExitBlock = ExitBlock;
1752 LoopVectorBody = VecBody;
1753 LoopScalarBody = OldBasicBlock;
1756 /// This function returns the identity element (or neutral element) for
1757 /// the operation K.
1759 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1764 // Adding, Xoring, Oring zero to a number does not change it.
1765 return ConstantInt::get(Tp, 0);
1766 case RK_IntegerMult:
1767 // Multiplying a number by 1 does not change it.
1768 return ConstantInt::get(Tp, 1);
1770 // AND-ing a number with an all-1 value does not change it.
1771 return ConstantInt::get(Tp, -1, true);
1773 // Multiplying a number by 1 does not change it.
1774 return ConstantFP::get(Tp, 1.0L);
1776 // Adding zero to a number does not change it.
1777 return ConstantFP::get(Tp, 0.0L);
1779 llvm_unreachable("Unknown reduction kind");
1783 static Intrinsic::ID
1784 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1785 // If we have an intrinsic call, check if it is trivially vectorizable.
1786 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1787 switch (II->getIntrinsicID()) {
1788 case Intrinsic::sqrt:
1789 case Intrinsic::sin:
1790 case Intrinsic::cos:
1791 case Intrinsic::exp:
1792 case Intrinsic::exp2:
1793 case Intrinsic::log:
1794 case Intrinsic::log10:
1795 case Intrinsic::log2:
1796 case Intrinsic::fabs:
1797 case Intrinsic::floor:
1798 case Intrinsic::ceil:
1799 case Intrinsic::trunc:
1800 case Intrinsic::rint:
1801 case Intrinsic::nearbyint:
1802 case Intrinsic::pow:
1803 case Intrinsic::fma:
1804 case Intrinsic::fmuladd:
1805 return II->getIntrinsicID();
1807 return Intrinsic::not_intrinsic;
1812 return Intrinsic::not_intrinsic;
1815 Function *F = CI->getCalledFunction();
1816 // We're going to make assumptions on the semantics of the functions, check
1817 // that the target knows that it's available in this environment.
1818 if (!F || !TLI->getLibFunc(F->getName(), Func))
1819 return Intrinsic::not_intrinsic;
1821 // Otherwise check if we have a call to a function that can be turned into a
1822 // vector intrinsic.
1829 return Intrinsic::sin;
1833 return Intrinsic::cos;
1837 return Intrinsic::exp;
1839 case LibFunc::exp2f:
1840 case LibFunc::exp2l:
1841 return Intrinsic::exp2;
1845 return Intrinsic::log;
1846 case LibFunc::log10:
1847 case LibFunc::log10f:
1848 case LibFunc::log10l:
1849 return Intrinsic::log10;
1851 case LibFunc::log2f:
1852 case LibFunc::log2l:
1853 return Intrinsic::log2;
1855 case LibFunc::fabsf:
1856 case LibFunc::fabsl:
1857 return Intrinsic::fabs;
1858 case LibFunc::floor:
1859 case LibFunc::floorf:
1860 case LibFunc::floorl:
1861 return Intrinsic::floor;
1863 case LibFunc::ceilf:
1864 case LibFunc::ceill:
1865 return Intrinsic::ceil;
1866 case LibFunc::trunc:
1867 case LibFunc::truncf:
1868 case LibFunc::truncl:
1869 return Intrinsic::trunc;
1871 case LibFunc::rintf:
1872 case LibFunc::rintl:
1873 return Intrinsic::rint;
1874 case LibFunc::nearbyint:
1875 case LibFunc::nearbyintf:
1876 case LibFunc::nearbyintl:
1877 return Intrinsic::nearbyint;
1881 return Intrinsic::pow;
1884 return Intrinsic::not_intrinsic;
1887 /// This function translates the reduction kind to an LLVM binary operator.
1889 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1891 case LoopVectorizationLegality::RK_IntegerAdd:
1892 return Instruction::Add;
1893 case LoopVectorizationLegality::RK_IntegerMult:
1894 return Instruction::Mul;
1895 case LoopVectorizationLegality::RK_IntegerOr:
1896 return Instruction::Or;
1897 case LoopVectorizationLegality::RK_IntegerAnd:
1898 return Instruction::And;
1899 case LoopVectorizationLegality::RK_IntegerXor:
1900 return Instruction::Xor;
1901 case LoopVectorizationLegality::RK_FloatMult:
1902 return Instruction::FMul;
1903 case LoopVectorizationLegality::RK_FloatAdd:
1904 return Instruction::FAdd;
1905 case LoopVectorizationLegality::RK_IntegerMinMax:
1906 return Instruction::ICmp;
1907 case LoopVectorizationLegality::RK_FloatMinMax:
1908 return Instruction::FCmp;
1910 llvm_unreachable("Unknown reduction operation");
1914 Value *createMinMaxOp(IRBuilder<> &Builder,
1915 LoopVectorizationLegality::MinMaxReductionKind RK,
1918 CmpInst::Predicate P = CmpInst::ICMP_NE;
1921 llvm_unreachable("Unknown min/max reduction kind");
1922 case LoopVectorizationLegality::MRK_UIntMin:
1923 P = CmpInst::ICMP_ULT;
1925 case LoopVectorizationLegality::MRK_UIntMax:
1926 P = CmpInst::ICMP_UGT;
1928 case LoopVectorizationLegality::MRK_SIntMin:
1929 P = CmpInst::ICMP_SLT;
1931 case LoopVectorizationLegality::MRK_SIntMax:
1932 P = CmpInst::ICMP_SGT;
1934 case LoopVectorizationLegality::MRK_FloatMin:
1935 P = CmpInst::FCMP_OLT;
1937 case LoopVectorizationLegality::MRK_FloatMax:
1938 P = CmpInst::FCMP_OGT;
1943 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1944 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1946 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1948 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1953 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1954 //===------------------------------------------------===//
1956 // Notice: any optimization or new instruction that go
1957 // into the code below should be also be implemented in
1960 //===------------------------------------------------===//
1961 Constant *Zero = Builder.getInt32(0);
1963 // In order to support reduction variables we need to be able to vectorize
1964 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1965 // stages. First, we create a new vector PHI node with no incoming edges.
1966 // We use this value when we vectorize all of the instructions that use the
1967 // PHI. Next, after all of the instructions in the block are complete we
1968 // add the new incoming edges to the PHI. At this point all of the
1969 // instructions in the basic block are vectorized, so we can use them to
1970 // construct the PHI.
1971 PhiVector RdxPHIsToFix;
1973 // Scan the loop in a topological order to ensure that defs are vectorized
1975 LoopBlocksDFS DFS(OrigLoop);
1978 // Vectorize all of the blocks in the original loop.
1979 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1980 be = DFS.endRPO(); bb != be; ++bb)
1981 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1983 // At this point every instruction in the original loop is widened to
1984 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1985 // that we vectorized. The PHI nodes are currently empty because we did
1986 // not want to introduce cycles. Notice that the remaining PHI nodes
1987 // that we need to fix are reduction variables.
1989 // Create the 'reduced' values for each of the induction vars.
1990 // The reduced values are the vector values that we scalarize and combine
1991 // after the loop is finished.
1992 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1994 PHINode *RdxPhi = *it;
1995 assert(RdxPhi && "Unable to recover vectorized PHI");
1997 // Find the reduction variable descriptor.
1998 assert(Legal->getReductionVars()->count(RdxPhi) &&
1999 "Unable to find the reduction variable");
2000 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2001 (*Legal->getReductionVars())[RdxPhi];
2003 // We need to generate a reduction vector from the incoming scalar.
2004 // To do so, we need to generate the 'identity' vector and overide
2005 // one of the elements with the incoming scalar reduction. We need
2006 // to do it in the vector-loop preheader.
2007 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2009 // This is the vector-clone of the value that leaves the loop.
2010 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2011 Type *VecTy = VectorExit[0]->getType();
2013 // Find the reduction identity variable. Zero for addition, or, xor,
2014 // one for multiplication, -1 for And.
2017 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2018 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2019 // MinMax reduction have the start value as their identify.
2020 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2024 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2025 VecTy->getScalarType());
2026 Identity = ConstantVector::getSplat(VF, Iden);
2028 // This vector is the Identity vector where the first element is the
2029 // incoming scalar reduction.
2030 VectorStart = Builder.CreateInsertElement(Identity,
2031 RdxDesc.StartValue, Zero);
2034 // Fix the vector-loop phi.
2035 // We created the induction variable so we know that the
2036 // preheader is the first entry.
2037 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2039 // Reductions do not have to start at zero. They can start with
2040 // any loop invariant values.
2041 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2042 BasicBlock *Latch = OrigLoop->getLoopLatch();
2043 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2044 VectorParts &Val = getVectorValue(LoopVal);
2045 for (unsigned part = 0; part < UF; ++part) {
2046 // Make sure to add the reduction stat value only to the
2047 // first unroll part.
2048 Value *StartVal = (part == 0) ? VectorStart : Identity;
2049 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2050 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2053 // Before each round, move the insertion point right between
2054 // the PHIs and the values we are going to write.
2055 // This allows us to write both PHINodes and the extractelement
2057 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2059 VectorParts RdxParts;
2060 for (unsigned part = 0; part < UF; ++part) {
2061 // This PHINode contains the vectorized reduction variable, or
2062 // the initial value vector, if we bypass the vector loop.
2063 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2064 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2065 Value *StartVal = (part == 0) ? VectorStart : Identity;
2066 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2067 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2068 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2069 RdxParts.push_back(NewPhi);
2072 // Reduce all of the unrolled parts into a single vector.
2073 Value *ReducedPartRdx = RdxParts[0];
2074 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2075 for (unsigned part = 1; part < UF; ++part) {
2076 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2077 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2078 RdxParts[part], ReducedPartRdx,
2081 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2082 ReducedPartRdx, RdxParts[part]);
2085 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2086 // and vector ops, reducing the set of values being computed by half each
2088 assert(isPowerOf2_32(VF) &&
2089 "Reduction emission only supported for pow2 vectors!");
2090 Value *TmpVec = ReducedPartRdx;
2091 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2092 for (unsigned i = VF; i != 1; i >>= 1) {
2093 // Move the upper half of the vector to the lower half.
2094 for (unsigned j = 0; j != i/2; ++j)
2095 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2097 // Fill the rest of the mask with undef.
2098 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2099 UndefValue::get(Builder.getInt32Ty()));
2102 Builder.CreateShuffleVector(TmpVec,
2103 UndefValue::get(TmpVec->getType()),
2104 ConstantVector::get(ShuffleMask),
2107 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2108 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2111 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2114 // The result is in the first element of the vector.
2115 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2117 // Now, we need to fix the users of the reduction variable
2118 // inside and outside of the scalar remainder loop.
2119 // We know that the loop is in LCSSA form. We need to update the
2120 // PHI nodes in the exit blocks.
2121 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2122 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2123 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2124 if (!LCSSAPhi) continue;
2126 // All PHINodes need to have a single entry edge, or two if
2127 // we already fixed them.
2128 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2130 // We found our reduction value exit-PHI. Update it with the
2131 // incoming bypass edge.
2132 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2133 // Add an edge coming from the bypass.
2134 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2137 }// end of the LCSSA phi scan.
2139 // Fix the scalar loop reduction variable with the incoming reduction sum
2140 // from the vector body and from the backedge value.
2141 int IncomingEdgeBlockIdx =
2142 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2143 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2144 // Pick the other block.
2145 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2146 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2147 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2148 }// end of for each redux variable.
2150 // The Loop exit block may have single value PHI nodes where the incoming
2151 // value is 'undef'. While vectorizing we only handled real values that
2152 // were defined inside the loop. Here we handle the 'undef case'.
2154 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2155 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2156 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2157 if (!LCSSAPhi) continue;
2158 if (LCSSAPhi->getNumIncomingValues() == 1)
2159 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2164 InnerLoopVectorizer::VectorParts
2165 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2166 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2169 VectorParts SrcMask = createBlockInMask(Src);
2171 // The terminator has to be a branch inst!
2172 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2173 assert(BI && "Unexpected terminator found");
2175 if (BI->isConditional()) {
2176 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2178 if (BI->getSuccessor(0) != Dst)
2179 for (unsigned part = 0; part < UF; ++part)
2180 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2182 for (unsigned part = 0; part < UF; ++part)
2183 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2190 InnerLoopVectorizer::VectorParts
2191 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2192 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2194 // Loop incoming mask is all-one.
2195 if (OrigLoop->getHeader() == BB) {
2196 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2197 return getVectorValue(C);
2200 // This is the block mask. We OR all incoming edges, and with zero.
2201 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2202 VectorParts BlockMask = getVectorValue(Zero);
2205 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2206 VectorParts EM = createEdgeMask(*it, BB);
2207 for (unsigned part = 0; part < UF; ++part)
2208 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2215 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2216 BasicBlock *BB, PhiVector *PV) {
2217 // For each instruction in the old loop.
2218 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2219 VectorParts &Entry = WidenMap.get(it);
2220 switch (it->getOpcode()) {
2221 case Instruction::Br:
2222 // Nothing to do for PHIs and BR, since we already took care of the
2223 // loop control flow instructions.
2225 case Instruction::PHI:{
2226 PHINode* P = cast<PHINode>(it);
2227 // Handle reduction variables:
2228 if (Legal->getReductionVars()->count(P)) {
2229 for (unsigned part = 0; part < UF; ++part) {
2230 // This is phase one of vectorizing PHIs.
2231 Type *VecTy = VectorType::get(it->getType(), VF);
2232 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2233 LoopVectorBody-> getFirstInsertionPt());
2239 // Check for PHI nodes that are lowered to vector selects.
2240 if (P->getParent() != OrigLoop->getHeader()) {
2241 // We know that all PHIs in non header blocks are converted into
2242 // selects, so we don't have to worry about the insertion order and we
2243 // can just use the builder.
2244 // At this point we generate the predication tree. There may be
2245 // duplications since this is a simple recursive scan, but future
2246 // optimizations will clean it up.
2248 unsigned NumIncoming = P->getNumIncomingValues();
2250 // Generate a sequence of selects of the form:
2251 // SELECT(Mask3, In3,
2252 // SELECT(Mask2, In2,
2254 for (unsigned In = 0; In < NumIncoming; In++) {
2255 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2257 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2259 for (unsigned part = 0; part < UF; ++part) {
2260 // We might have single edge PHIs (blocks) - use an identity
2261 // 'select' for the first PHI operand.
2263 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2266 // Select between the current value and the previous incoming edge
2267 // based on the incoming mask.
2268 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2269 Entry[part], "predphi");
2275 // This PHINode must be an induction variable.
2276 // Make sure that we know about it.
2277 assert(Legal->getInductionVars()->count(P) &&
2278 "Not an induction variable");
2280 LoopVectorizationLegality::InductionInfo II =
2281 Legal->getInductionVars()->lookup(P);
2284 case LoopVectorizationLegality::IK_NoInduction:
2285 llvm_unreachable("Unknown induction");
2286 case LoopVectorizationLegality::IK_IntInduction: {
2287 assert(P->getType() == II.StartValue->getType() && "Types must match");
2288 Type *PhiTy = P->getType();
2290 if (P == OldInduction) {
2291 // Handle the canonical induction variable. We might have had to
2293 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2295 // Handle other induction variables that are now based on the
2297 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2299 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2300 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2303 Broadcasted = getBroadcastInstrs(Broadcasted);
2304 // After broadcasting the induction variable we need to make the vector
2305 // consecutive by adding 0, 1, 2, etc.
2306 for (unsigned part = 0; part < UF; ++part)
2307 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2310 case LoopVectorizationLegality::IK_ReverseIntInduction:
2311 case LoopVectorizationLegality::IK_PtrInduction:
2312 case LoopVectorizationLegality::IK_ReversePtrInduction:
2313 // Handle reverse integer and pointer inductions.
2314 Value *StartIdx = ExtendedIdx;
2315 // This is the normalized GEP that starts counting at zero.
2316 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2319 // Handle the reverse integer induction variable case.
2320 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2321 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2322 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2324 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2327 // This is a new value so do not hoist it out.
2328 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2329 // After broadcasting the induction variable we need to make the
2330 // vector consecutive by adding ... -3, -2, -1, 0.
2331 for (unsigned part = 0; part < UF; ++part)
2332 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2337 // Handle the pointer induction variable case.
2338 assert(P->getType()->isPointerTy() && "Unexpected type.");
2340 // Is this a reverse induction ptr or a consecutive induction ptr.
2341 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2344 // This is the vector of results. Notice that we don't generate
2345 // vector geps because scalar geps result in better code.
2346 for (unsigned part = 0; part < UF; ++part) {
2347 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2348 for (unsigned int i = 0; i < VF; ++i) {
2349 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2350 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2353 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2355 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2357 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2359 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2360 Builder.getInt32(i),
2363 Entry[part] = VecVal;
2370 case Instruction::Add:
2371 case Instruction::FAdd:
2372 case Instruction::Sub:
2373 case Instruction::FSub:
2374 case Instruction::Mul:
2375 case Instruction::FMul:
2376 case Instruction::UDiv:
2377 case Instruction::SDiv:
2378 case Instruction::FDiv:
2379 case Instruction::URem:
2380 case Instruction::SRem:
2381 case Instruction::FRem:
2382 case Instruction::Shl:
2383 case Instruction::LShr:
2384 case Instruction::AShr:
2385 case Instruction::And:
2386 case Instruction::Or:
2387 case Instruction::Xor: {
2388 // Just widen binops.
2389 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2390 VectorParts &A = getVectorValue(it->getOperand(0));
2391 VectorParts &B = getVectorValue(it->getOperand(1));
2393 // Use this vector value for all users of the original instruction.
2394 for (unsigned Part = 0; Part < UF; ++Part) {
2395 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2397 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2398 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2399 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2400 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2401 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2403 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2404 VecOp->setIsExact(BinOp->isExact());
2410 case Instruction::Select: {
2412 // If the selector is loop invariant we can create a select
2413 // instruction with a scalar condition. Otherwise, use vector-select.
2414 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2417 // The condition can be loop invariant but still defined inside the
2418 // loop. This means that we can't just use the original 'cond' value.
2419 // We have to take the 'vectorized' value and pick the first lane.
2420 // Instcombine will make this a no-op.
2421 VectorParts &Cond = getVectorValue(it->getOperand(0));
2422 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2423 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2424 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2425 Builder.getInt32(0));
2426 for (unsigned Part = 0; Part < UF; ++Part) {
2427 Entry[Part] = Builder.CreateSelect(
2428 InvariantCond ? ScalarCond : Cond[Part],
2435 case Instruction::ICmp:
2436 case Instruction::FCmp: {
2437 // Widen compares. Generate vector compares.
2438 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2439 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2440 VectorParts &A = getVectorValue(it->getOperand(0));
2441 VectorParts &B = getVectorValue(it->getOperand(1));
2442 for (unsigned Part = 0; Part < UF; ++Part) {
2445 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2447 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2453 case Instruction::Store:
2454 case Instruction::Load:
2455 vectorizeMemoryInstruction(it, Legal);
2457 case Instruction::ZExt:
2458 case Instruction::SExt:
2459 case Instruction::FPToUI:
2460 case Instruction::FPToSI:
2461 case Instruction::FPExt:
2462 case Instruction::PtrToInt:
2463 case Instruction::IntToPtr:
2464 case Instruction::SIToFP:
2465 case Instruction::UIToFP:
2466 case Instruction::Trunc:
2467 case Instruction::FPTrunc:
2468 case Instruction::BitCast: {
2469 CastInst *CI = dyn_cast<CastInst>(it);
2470 /// Optimize the special case where the source is the induction
2471 /// variable. Notice that we can only optimize the 'trunc' case
2472 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2473 /// c. other casts depend on pointer size.
2474 if (CI->getOperand(0) == OldInduction &&
2475 it->getOpcode() == Instruction::Trunc) {
2476 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2478 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2479 for (unsigned Part = 0; Part < UF; ++Part)
2480 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2483 /// Vectorize casts.
2484 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2486 VectorParts &A = getVectorValue(it->getOperand(0));
2487 for (unsigned Part = 0; Part < UF; ++Part)
2488 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2492 case Instruction::Call: {
2493 // Ignore dbg intrinsics.
2494 if (isa<DbgInfoIntrinsic>(it))
2497 Module *M = BB->getParent()->getParent();
2498 CallInst *CI = cast<CallInst>(it);
2499 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2500 assert(ID && "Not an intrinsic call!");
2501 for (unsigned Part = 0; Part < UF; ++Part) {
2502 SmallVector<Value*, 4> Args;
2503 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2504 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2505 Args.push_back(Arg[Part]);
2507 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2508 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2509 Entry[Part] = Builder.CreateCall(F, Args);
2515 // All other instructions are unsupported. Scalarize them.
2516 scalarizeInstruction(it);
2519 }// end of for_each instr.
2522 void InnerLoopVectorizer::updateAnalysis() {
2523 // Forget the original basic block.
2524 SE->forgetLoop(OrigLoop);
2526 // Update the dominator tree information.
2527 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2528 "Entry does not dominate exit.");
2530 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2531 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2532 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2533 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2534 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2535 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2536 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2537 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2539 DEBUG(DT->verifyAnalysis());
2542 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2543 if (!EnableIfConversion)
2546 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2547 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2549 // Collect the blocks that need predication.
2550 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2551 BasicBlock *BB = LoopBlocks[i];
2553 // We don't support switch statements inside loops.
2554 if (!isa<BranchInst>(BB->getTerminator()))
2557 // We must be able to predicate all blocks that need to be predicated.
2558 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2562 // Check that we can actually speculate the hoistable loads.
2563 if (!LoadSpeculation.canHoistAllLoads())
2566 // We can if-convert this loop.
2570 bool LoopVectorizationLegality::canVectorize() {
2571 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2572 // be canonicalized.
2573 if (!TheLoop->getLoopPreheader())
2576 // We can only vectorize innermost loops.
2577 if (TheLoop->getSubLoopsVector().size())
2580 // We must have a single backedge.
2581 if (TheLoop->getNumBackEdges() != 1)
2584 // We must have a single exiting block.
2585 if (!TheLoop->getExitingBlock())
2588 unsigned NumBlocks = TheLoop->getNumBlocks();
2590 // Check if we can if-convert non single-bb loops.
2591 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2592 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2596 // We need to have a loop header.
2597 BasicBlock *Latch = TheLoop->getLoopLatch();
2598 DEBUG(dbgs() << "LV: Found a loop: " <<
2599 TheLoop->getHeader()->getName() << "\n");
2601 // ScalarEvolution needs to be able to find the exit count.
2602 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2603 if (ExitCount == SE->getCouldNotCompute()) {
2604 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2608 // Do not loop-vectorize loops with a tiny trip count.
2609 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2610 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2611 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2612 "This loop is not worth vectorizing.\n");
2616 // Check if we can vectorize the instructions and CFG in this loop.
2617 if (!canVectorizeInstrs()) {
2618 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2622 // Go over each instruction and look at memory deps.
2623 if (!canVectorizeMemory()) {
2624 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2628 // Collect all of the variables that remain uniform after vectorization.
2629 collectLoopUniforms();
2631 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2632 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2635 // Okay! We can vectorize. At this point we don't have any other mem analysis
2636 // which may limit our maximum vectorization factor, so just return true with
2641 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2642 if (Ty->isPointerTy())
2643 return DL.getIntPtrType(Ty->getContext());
2647 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2648 Ty0 = convertPointerToIntegerType(DL, Ty0);
2649 Ty1 = convertPointerToIntegerType(DL, Ty1);
2650 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2655 /// \brief Check that the instruction has outside loop users and is not an
2656 /// identified reduction variable.
2657 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2658 SmallPtrSet<Value *, 4> &Reductions) {
2659 // Reduction instructions are allowed to have exit users. All other
2660 // instructions must not have external users.
2661 if (!Reductions.count(Inst))
2662 //Check that all of the users of the loop are inside the BB.
2663 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2665 Instruction *U = cast<Instruction>(*I);
2666 // This user may be a reduction exit value.
2667 if (!TheLoop->contains(U)) {
2668 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2675 bool LoopVectorizationLegality::canVectorizeInstrs() {
2676 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2677 BasicBlock *Header = TheLoop->getHeader();
2679 // Look for the attribute signaling the absence of NaNs.
2680 Function &F = *Header->getParent();
2681 if (F.hasFnAttribute("no-nans-fp-math"))
2682 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2683 AttributeSet::FunctionIndex,
2684 "no-nans-fp-math").getValueAsString() == "true";
2686 // For each block in the loop.
2687 for (Loop::block_iterator bb = TheLoop->block_begin(),
2688 be = TheLoop->block_end(); bb != be; ++bb) {
2690 // Scan the instructions in the block and look for hazards.
2691 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2694 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2695 Type *PhiTy = Phi->getType();
2696 // Check that this PHI type is allowed.
2697 if (!PhiTy->isIntegerTy() &&
2698 !PhiTy->isFloatingPointTy() &&
2699 !PhiTy->isPointerTy()) {
2700 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2704 // If this PHINode is not in the header block, then we know that we
2705 // can convert it to select during if-conversion. No need to check if
2706 // the PHIs in this block are induction or reduction variables.
2707 if (*bb != Header) {
2708 // Check that this instruction has no outside users or is an
2709 // identified reduction value with an outside user.
2710 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2715 // We only allow if-converted PHIs with more than two incoming values.
2716 if (Phi->getNumIncomingValues() != 2) {
2717 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2721 // This is the value coming from the preheader.
2722 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2723 // Check if this is an induction variable.
2724 InductionKind IK = isInductionVariable(Phi);
2726 if (IK_NoInduction != IK) {
2727 // Get the widest type.
2729 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2731 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2733 // Int inductions are special because we only allow one IV.
2734 if (IK == IK_IntInduction) {
2735 // Use the phi node with the widest type as induction. Use the last
2736 // one if there are multiple (no good reason for doing this other
2737 // than it is expedient).
2738 if (!Induction || PhiTy == WidestIndTy)
2742 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2743 Inductions[Phi] = InductionInfo(StartValue, IK);
2747 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2748 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2751 if (AddReductionVar(Phi, RK_IntegerMult)) {
2752 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2755 if (AddReductionVar(Phi, RK_IntegerOr)) {
2756 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2759 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2760 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2763 if (AddReductionVar(Phi, RK_IntegerXor)) {
2764 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2767 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2768 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2771 if (AddReductionVar(Phi, RK_FloatMult)) {
2772 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2775 if (AddReductionVar(Phi, RK_FloatAdd)) {
2776 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2779 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2780 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2784 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2786 }// end of PHI handling
2788 // We still don't handle functions. However, we can ignore dbg intrinsic
2789 // calls and we do handle certain intrinsic and libm functions.
2790 CallInst *CI = dyn_cast<CallInst>(it);
2791 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2792 DEBUG(dbgs() << "LV: Found a call site.\n");
2796 // Check that the instruction return type is vectorizable.
2797 if (!VectorType::isValidElementType(it->getType()) &&
2798 !it->getType()->isVoidTy()) {
2799 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2803 // Check that the stored type is vectorizable.
2804 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2805 Type *T = ST->getValueOperand()->getType();
2806 if (!VectorType::isValidElementType(T))
2810 // Reduction instructions are allowed to have exit users.
2811 // All other instructions must not have external users.
2812 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2820 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2821 if (Inductions.empty())
2828 void LoopVectorizationLegality::collectLoopUniforms() {
2829 // We now know that the loop is vectorizable!
2830 // Collect variables that will remain uniform after vectorization.
2831 std::vector<Value*> Worklist;
2832 BasicBlock *Latch = TheLoop->getLoopLatch();
2834 // Start with the conditional branch and walk up the block.
2835 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2837 while (Worklist.size()) {
2838 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2839 Worklist.pop_back();
2841 // Look at instructions inside this loop.
2842 // Stop when reaching PHI nodes.
2843 // TODO: we need to follow values all over the loop, not only in this block.
2844 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2847 // This is a known uniform.
2850 // Insert all operands.
2851 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2855 /// \brief Analyses memory accesses in a loop.
2857 /// Checks whether run time pointer checks are needed and builds sets for data
2858 /// dependence checking.
2859 class AccessAnalysis {
2861 /// \brief Read or write access location.
2862 typedef std::pair<Value*, char> MemAccessInfo;
2864 /// \brief Set of potential dependent memory accesses.
2865 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2867 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2868 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2869 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2871 /// \brief Register a load and whether it is only read from.
2872 void addLoad(Value *Ptr, bool IsReadOnly) {
2873 Accesses.insert(std::make_pair(Ptr, false));
2875 ReadOnlyPtr.insert(Ptr);
2878 /// \brief Register a store.
2879 void addStore(Value *Ptr) {
2880 Accesses.insert(std::make_pair(Ptr, true));
2883 /// \brief Check whether we can check the pointers at runtime for
2884 /// non-intersection.
2885 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2886 unsigned &NumComparisons, ScalarEvolution *SE,
2889 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2890 /// and builds sets of dependent accesses.
2891 void buildDependenceSets() {
2892 // Process read-write pointers first.
2893 processMemAccesses(false);
2894 // Next, process read pointers.
2895 processMemAccesses(true);
2898 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
2900 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
2902 DenseSet<MemAccessInfo> &getDependenciesToCheck() { return CheckDeps; }
2905 typedef SetVector<MemAccessInfo> PtrAccessSet;
2906 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
2908 /// \brief Go over all memory access or only the deferred ones if
2909 /// \p UseDeferred is true and check whether runtime pointer checks are needed
2910 /// and build sets of dependency check candidates.
2911 void processMemAccesses(bool UseDeferred);
2913 /// Set of all accesses.
2914 PtrAccessSet Accesses;
2916 /// Set of access to check after all writes have been processed.
2917 PtrAccessSet DeferredAccesses;
2919 /// Map of pointers to last access encountered.
2920 UnderlyingObjToAccessMap ObjToLastAccess;
2922 /// Set of accesses that need a further dependence check.
2923 DenseSet<MemAccessInfo> CheckDeps;
2925 /// Set of pointers that are read only.
2926 SmallPtrSet<Value*, 16> ReadOnlyPtr;
2928 /// Set of underlying objects already written to.
2929 SmallPtrSet<Value*, 16> WriteObjects;
2933 /// Sets of potentially dependent accesses - members of one set share an
2934 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
2935 /// dependence check.
2936 DepCandidates &DepCands;
2938 bool AreAllWritesIdentified;
2939 bool AreAllReadsIdentified;
2940 bool IsRTCheckNeeded;
2943 /// \brief Check whether a pointer can participate in a runtime bounds check.
2944 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
2945 const SCEV *PtrScev = SE->getSCEV(Ptr);
2946 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
2950 return AR->isAffine();
2953 bool AccessAnalysis::canCheckPtrAtRT(
2954 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2955 unsigned &NumComparisons, ScalarEvolution *SE,
2957 // Find pointers with computable bounds. We are going to use this information
2958 // to place a runtime bound check.
2959 unsigned NumReadPtrChecks = 0;
2960 unsigned NumWritePtrChecks = 0;
2961 bool CanDoRT = true;
2963 bool IsDepCheckNeeded = isDependencyCheckNeeded();
2964 // We assign consecutive id to access from different dependence sets.
2965 // Accesses within the same set don't need a runtime check.
2966 unsigned RunningDepId = 1;
2967 DenseMap<Value *, unsigned> DepSetId;
2969 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
2971 const MemAccessInfo &Access = *AI;
2972 Value *Ptr = Access.first;
2973 bool IsWrite = Access.second;
2975 // Just add write checks if we have both.
2976 if (!IsWrite && Accesses.count(std::make_pair(Ptr, true)))
2980 ++NumWritePtrChecks;
2984 if (hasComputableBounds(SE, Ptr)) {
2985 // The id of the dependence set.
2988 if (IsDepCheckNeeded) {
2989 Value *Leader = DepCands.getLeaderValue(Access).first;
2990 unsigned &LeaderId = DepSetId[Leader];
2992 LeaderId = RunningDepId++;
2995 // Each access has its own dependence set.
2996 DepId = RunningDepId++;
2998 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
3000 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
3006 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3007 NumComparisons = 0; // Only one dependence set.
3009 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3010 NumWritePtrChecks - 1));
3014 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3015 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3018 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3019 // We process the set twice: first we process read-write pointers, last we
3020 // process read-only pointers. This allows us to skip dependence tests for
3021 // read-only pointers.
3023 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3024 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3025 const MemAccessInfo &Access = *AI;
3026 Value *Ptr = Access.first;
3027 bool IsWrite = Access.second;
3029 DepCands.insert(Access);
3031 // Memorize read-only pointers for later processing and skip them in the
3032 // first round (they need to be checked after we have seen all write
3033 // pointers). Note: we also mark pointer that are not consecutive as
3034 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3035 // second check for "!IsWrite".
3036 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3037 if (!UseDeferred && IsReadOnlyPtr) {
3038 DeferredAccesses.insert(Access);
3042 bool NeedDepCheck = false;
3043 // Check whether there is the possiblity of dependency because of underlying
3044 // objects being the same.
3045 typedef SmallVector<Value*, 16> ValueVector;
3046 ValueVector TempObjects;
3047 GetUnderlyingObjects(Ptr, TempObjects, DL);
3048 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3050 Value *UnderlyingObj = *UI;
3052 // If this is a write then it needs to be an identified object. If this a
3053 // read and all writes (so far) are identified function scope objects we
3054 // don't need an identified underlying object but only an Argument (the
3055 // next write is going to invalidate this assumption if it is
3057 // This is a micro-optimization for the case where all writes are
3058 // identified and we have one argument pointer.
3059 // Otherwise, we do need a runtime check.
3060 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3061 (!IsWrite && (!AreAllWritesIdentified ||
3062 !isa<Argument>(UnderlyingObj)) &&
3063 !isIdentifiedObject(UnderlyingObj))) {
3064 DEBUG(dbgs() << "LV: Found an unidentified " <<
3065 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3067 IsRTCheckNeeded = (IsRTCheckNeeded ||
3068 !isIdentifiedObject(UnderlyingObj) ||
3069 !AreAllReadsIdentified);
3072 AreAllWritesIdentified = false;
3074 AreAllReadsIdentified = false;
3077 // If this is a write - check other reads and writes for conflicts. If
3078 // this is a read only check other writes for conflicts (but only if there
3079 // is no other write to the ptr - this is an optimization to catch "a[i] =
3080 // a[i] + " without having to do a dependence check).
3081 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3082 NeedDepCheck = true;
3085 WriteObjects.insert(UnderlyingObj);
3087 // Create sets of pointers connected by shared underlying objects.
3088 UnderlyingObjToAccessMap::iterator Prev =
3089 ObjToLastAccess.find(UnderlyingObj);
3090 if (Prev != ObjToLastAccess.end())
3091 DepCands.unionSets(Access, Prev->second);
3093 ObjToLastAccess[UnderlyingObj] = Access;
3097 CheckDeps.insert(Access);
3101 /// \brief Checks memory dependences among accesses to the same underlying
3102 /// object to determine whether there vectorization is legal or not (and at
3103 /// which vectorization factor).
3105 /// This class works under the assumption that we already checked that memory
3106 /// locations with different underlying pointers are "must-not alias".
3107 /// We use the ScalarEvolution framework to symbolically evalutate access
3108 /// functions pairs. Since we currently don't restructure the loop we can rely
3109 /// on the program order of memory accesses to determine their safety.
3110 /// At the moment we will only deem accesses as safe for:
3111 /// * A negative constant distance assuming program order.
3113 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3114 /// a[i] = tmp; y = a[i];
3116 /// The latter case is safe because later checks guarantuee that there can't
3117 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3118 /// the same variable: a header phi can only be an induction or a reduction, a
3119 /// reduction can't have a memory sink, an induction can't have a memory
3120 /// source). This is important and must not be violated (or we have to
3121 /// resort to checking for cycles through memory).
3123 /// * A positive constant distance assuming program order that is bigger
3124 /// than the biggest memory access.
3126 /// tmp = a[i] OR b[i] = x
3127 /// a[i+2] = tmp y = b[i+2];
3129 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3131 /// * Zero distances and all accesses have the same size.
3133 class MemoryDepChecker {
3135 typedef std::pair<Value*, char> MemAccessInfo;
3137 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3138 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3140 /// \brief Register the location (instructions are given increasing numbers)
3141 /// of a write access.
3142 void addAccess(StoreInst *SI) {
3143 Value *Ptr = SI->getPointerOperand();
3144 Accesses[std::make_pair(Ptr, true)].push_back(AccessIdx);
3145 InstMap.push_back(SI);
3149 /// \brief Register the location (instructions are given increasing numbers)
3150 /// of a write access.
3151 void addAccess(LoadInst *LI) {
3152 Value *Ptr = LI->getPointerOperand();
3153 Accesses[std::make_pair(Ptr, false)].push_back(AccessIdx);
3154 InstMap.push_back(LI);
3158 /// \brief Check whether the dependencies between the accesses are safe.
3160 /// Only checks sets with elements in \p CheckDeps.
3161 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3162 DenseSet<MemAccessInfo> &CheckDeps);
3164 /// \brief The maximum number of bytes of a vector register we can vectorize
3165 /// the accesses safely with.
3166 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3169 ScalarEvolution *SE;
3171 const Loop *InnermostLoop;
3173 /// \brief Maps access locations (ptr, read/write) to program order.
3174 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3176 /// \brief Memory access instructions in program order.
3177 SmallVector<Instruction *, 16> InstMap;
3179 /// \brief The program order index to be used for the next instruction.
3182 // We can access this many bytes in parallel safely.
3183 unsigned MaxSafeDepDistBytes;
3185 /// \brief Check whether there is a plausible dependence between the two
3188 /// Access \p A must happen before \p B in program order. The two indices
3189 /// identify the index into the program order map.
3191 /// This function checks whether there is a plausible dependence (or the
3192 /// absence of such can't be proved) between the two accesses. If there is a
3193 /// plausible dependence but the dependence distance is bigger than one
3194 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3195 /// distance is smaller than any other distance encountered so far).
3196 /// Otherwise, this function returns true signaling a possible dependence.
3197 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3198 const MemAccessInfo &B, unsigned BIdx);
3200 /// \brief Check whether the data dependence could prevent store-load
3202 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3205 static bool isInBoundsGep(Value *Ptr) {
3206 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3207 return GEP->isInBounds();
3211 /// \brief Check whether the access through \p Ptr has a constant stride.
3212 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3214 const Type *PtrTy = Ptr->getType();
3215 assert(PtrTy->isPointerTy() && "Unexpected non ptr");
3217 // Make sure that the pointer does not point to aggregate types.
3218 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) {
3219 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr
3224 const SCEV *PtrScev = SE->getSCEV(Ptr);
3225 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3227 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3228 << *Ptr << " SCEV: " << *PtrScev << "\n");
3232 // The accesss function must stride over the innermost loop.
3233 if (Lp != AR->getLoop()) {
3234 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << *Ptr
3235 << " SCEV: " << *PtrScev << "\n");
3238 // The address calculation must not wrap. Otherwise, a dependence could be
3239 // inverted. An inbounds getelementptr that is a AddRec with a unit stride
3240 // cannot wrap per definition. The unit stride requirement is checked later.
3241 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3242 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3243 if (!IsNoWrapAddRec && !IsInBoundsGEP) {
3244 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3245 << *Ptr << " SCEV: " << *PtrScev << "\n");
3249 // Check the step is constant.
3250 const SCEV *Step = AR->getStepRecurrence(*SE);
3252 // Calculate the pointer stride and check if it is consecutive.
3253 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3255 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3256 " SCEV: " << *PtrScev << "\n");
3260 int64_t Size = DL->getTypeAllocSize(PtrTy->getPointerElementType());
3261 const APInt &APStepVal = C->getValue()->getValue();
3263 // Huge step value - give up.
3264 if (APStepVal.getBitWidth() > 64)
3267 int64_t StepVal = APStepVal.getSExtValue();
3270 int64_t Stride = StepVal / Size;
3271 int64_t Rem = StepVal % Size;
3275 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3276 // know we can't "wrap around the address space".
3277 if (!IsNoWrapAddRec && IsInBoundsGEP && Stride != 1 && Stride != -1)
3283 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3284 unsigned TypeByteSize) {
3285 // If loads occur at a distance that is not a multiple of a feasible vector
3286 // factor store-load forwarding does not take place.
3287 // Positive dependences might cause troubles because vectorizing them might
3288 // prevent store-load forwarding making vectorized code run a lot slower.
3289 // a[i] = a[i-3] ^ a[i-8];
3290 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3291 // hence on your typical architecture store-load forwarding does not take
3292 // place. Vectorizing in such cases does not make sense.
3293 // Store-load forwarding distance.
3294 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3295 // Maximum vector factor.
3296 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3297 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3298 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3300 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3302 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3303 MaxVFWithoutSLForwardIssues = (vf >>=1);
3308 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3309 DEBUG(dbgs() << "LV: Distance " << Distance <<
3310 " that could cause a store-load forwarding conflict\n");
3314 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3315 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3316 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3320 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3321 const MemAccessInfo &B, unsigned BIdx) {
3322 assert (AIdx < BIdx && "Must pass arguments in program order");
3324 Value *APtr = A.first;
3325 Value *BPtr = B.first;
3326 bool AIsWrite = A.second;
3327 bool BIsWrite = B.second;
3329 // Two reads are independent.
3330 if (!AIsWrite && !BIsWrite)
3333 const SCEV *AScev = SE->getSCEV(APtr);
3334 const SCEV *BScev = SE->getSCEV(BPtr);
3336 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3337 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3339 const SCEV *Src = AScev;
3340 const SCEV *Sink = BScev;
3342 // If the induction step is negative we have to invert source and sink of the
3344 if (StrideAPtr < 0) {
3347 std::swap(APtr, BPtr);
3348 std::swap(Src, Sink);
3349 std::swap(AIsWrite, BIsWrite);
3350 std::swap(AIdx, BIdx);
3351 std::swap(StrideAPtr, StrideBPtr);
3354 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3356 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3357 << "(Induction step: " << StrideAPtr << ")\n");
3358 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3359 << *InstMap[BIdx] << ": " << *Dist << "\n");
3361 // Need consecutive accesses. We don't want to vectorize
3362 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3363 // the address space.
3364 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3365 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3369 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3371 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3375 Type *ATy = APtr->getType()->getPointerElementType();
3376 Type *BTy = BPtr->getType()->getPointerElementType();
3377 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3379 // Negative distances are not plausible dependencies.
3380 const APInt &Val = C->getValue()->getValue();
3381 if (Val.isNegative()) {
3382 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3383 if (IsTrueDataDependence &&
3384 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3388 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3392 // Write to the same location with the same size.
3393 // Could be improved to assert type sizes are the same (i32 == float, etc).
3397 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3401 assert(Val.isStrictlyPositive() && "Expect a positive value");
3403 // Positive distance bigger than max vectorization factor.
3406 "LV: ReadWrite-Write positive dependency with different types");
3410 unsigned Distance = (unsigned) Val.getZExtValue();
3412 // Bail out early if passed-in parameters make vectorization not feasible.
3413 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3414 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3416 // The distance must be bigger than the size needed for a vectorized version
3417 // of the operation and the size of the vectorized operation must not be
3418 // bigger than the currrent maximum size.
3419 if (Distance < 2*TypeByteSize ||
3420 2*TypeByteSize > MaxSafeDepDistBytes ||
3421 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3422 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3423 << Val.getSExtValue() << "\n");
3427 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3428 Distance : MaxSafeDepDistBytes;
3430 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3431 if (IsTrueDataDependence &&
3432 couldPreventStoreLoadForward(Distance, TypeByteSize))
3435 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3436 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3442 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3443 DenseSet<MemAccessInfo> &CheckDeps) {
3445 MaxSafeDepDistBytes = -1U;
3446 while (!CheckDeps.empty()) {
3447 MemAccessInfo CurAccess = *CheckDeps.begin();
3449 // Get the relevant memory access set.
3450 EquivalenceClasses<MemAccessInfo>::iterator I =
3451 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3453 // Check accesses within this set.
3454 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3455 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3457 // Check every access pair.
3459 CheckDeps.erase(*AI);
3460 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3462 // Check every accessing instruction pair in program order.
3463 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3464 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3465 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3466 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3467 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3469 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3480 bool LoopVectorizationLegality::canVectorizeMemory() {
3482 typedef SmallVector<Value*, 16> ValueVector;
3483 typedef SmallPtrSet<Value*, 16> ValueSet;
3485 // Stores a pair of memory access location and whether the access is a store
3486 // (true) or a load (false).
3487 typedef std::pair<Value*, char> MemAccessInfo;
3488 typedef DenseSet<MemAccessInfo> PtrAccessSet;
3490 // Holds the Load and Store *instructions*.
3494 // Holds all the different accesses in the loop.
3495 unsigned NumReads = 0;
3496 unsigned NumReadWrites = 0;
3498 PtrRtCheck.Pointers.clear();
3499 PtrRtCheck.Need = false;
3501 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3502 MemoryDepChecker DepChecker(SE, DL, TheLoop);
3505 for (Loop::block_iterator bb = TheLoop->block_begin(),
3506 be = TheLoop->block_end(); bb != be; ++bb) {
3508 // Scan the BB and collect legal loads and stores.
3509 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3512 // If this is a load, save it. If this instruction can read from memory
3513 // but is not a load, then we quit. Notice that we don't handle function
3514 // calls that read or write.
3515 if (it->mayReadFromMemory()) {
3516 LoadInst *Ld = dyn_cast<LoadInst>(it);
3517 if (!Ld) return false;
3518 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3519 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3522 Loads.push_back(Ld);
3523 DepChecker.addAccess(Ld);
3527 // Save 'store' instructions. Abort if other instructions write to memory.
3528 if (it->mayWriteToMemory()) {
3529 StoreInst *St = dyn_cast<StoreInst>(it);
3530 if (!St) return false;
3531 if (!St->isSimple() && !IsAnnotatedParallel) {
3532 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3535 Stores.push_back(St);
3536 DepChecker.addAccess(St);
3541 // Now we have two lists that hold the loads and the stores.
3542 // Next, we find the pointers that they use.
3544 // Check if we see any stores. If there are no stores, then we don't
3545 // care if the pointers are *restrict*.
3546 if (!Stores.size()) {
3547 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3551 AccessAnalysis::DepCandidates DependentAccesses;
3552 AccessAnalysis Accesses(DL, DependentAccesses);
3554 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3555 // multiple times on the same object. If the ptr is accessed twice, once
3556 // for read and once for write, it will only appear once (on the write
3557 // list). This is okay, since we are going to check for conflicts between
3558 // writes and between reads and writes, but not between reads and reads.
3561 ValueVector::iterator I, IE;
3562 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3563 StoreInst *ST = cast<StoreInst>(*I);
3564 Value* Ptr = ST->getPointerOperand();
3566 if (isUniform(Ptr)) {
3567 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3571 // If we did *not* see this pointer before, insert it to the read-write
3572 // list. At this phase it is only a 'write' list.
3573 if (Seen.insert(Ptr)) {
3575 Accesses.addStore(Ptr);
3579 if (IsAnnotatedParallel) {
3581 << "LV: A loop annotated parallel, ignore memory dependency "
3586 SmallPtrSet<Value *, 16> ReadOnlyPtr;
3587 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3588 LoadInst *LD = cast<LoadInst>(*I);
3589 Value* Ptr = LD->getPointerOperand();
3590 // If we did *not* see this pointer before, insert it to the
3591 // read list. If we *did* see it before, then it is already in
3592 // the read-write list. This allows us to vectorize expressions
3593 // such as A[i] += x; Because the address of A[i] is a read-write
3594 // pointer. This only works if the index of A[i] is consecutive.
3595 // If the address of i is unknown (for example A[B[i]]) then we may
3596 // read a few words, modify, and write a few words, and some of the
3597 // words may be written to the same address.
3598 bool IsReadOnlyPtr = false;
3599 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop)) {
3601 IsReadOnlyPtr = true;
3603 Accesses.addLoad(Ptr, IsReadOnlyPtr);
3606 // If we write (or read-write) to a single destination and there are no
3607 // other reads in this loop then is it safe to vectorize.
3608 if (NumReadWrites == 1 && NumReads == 0) {
3609 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3613 // Build dependence sets and check whether we need a runtime pointer bounds
3615 Accesses.buildDependenceSets();
3616 bool NeedRTCheck = Accesses.isRTCheckNeeded();
3618 // Find pointers with computable bounds. We are going to use this information
3619 // to place a runtime bound check.
3620 unsigned NumComparisons = 0;
3621 bool CanDoRT = false;
3623 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop);
3626 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
3627 " pointer comparisons.\n");
3629 // If we only have one set of dependences to check pointers among we don't
3630 // need a runtime check.
3631 if (NumComparisons == 0 && NeedRTCheck)
3632 NeedRTCheck = false;
3634 // Check that we did not collect too many pointers or found a unsizeable
3636 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3642 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3645 if (NeedRTCheck && !CanDoRT) {
3646 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3647 "the array bounds.\n");
3652 PtrRtCheck.Need = NeedRTCheck;
3654 bool CanVecMem = true;
3655 if (Accesses.isDependencyCheckNeeded()) {
3656 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
3657 CanVecMem = DepChecker.areDepsSafe(DependentAccesses,
3658 Accesses.getDependenciesToCheck());
3659 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
3662 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3663 " need a runtime memory check.\n");
3668 static bool hasMultipleUsesOf(Instruction *I,
3669 SmallPtrSet<Instruction *, 8> &Insts) {
3670 unsigned NumUses = 0;
3671 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3672 if (Insts.count(dyn_cast<Instruction>(*Use)))
3681 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3682 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3683 if (!Set.count(dyn_cast<Instruction>(*Use)))
3688 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3689 ReductionKind Kind) {
3690 if (Phi->getNumIncomingValues() != 2)
3693 // Reduction variables are only found in the loop header block.
3694 if (Phi->getParent() != TheLoop->getHeader())
3697 // Obtain the reduction start value from the value that comes from the loop
3699 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3701 // ExitInstruction is the single value which is used outside the loop.
3702 // We only allow for a single reduction value to be used outside the loop.
3703 // This includes users of the reduction, variables (which form a cycle
3704 // which ends in the phi node).
3705 Instruction *ExitInstruction = 0;
3706 // Indicates that we found a reduction operation in our scan.
3707 bool FoundReduxOp = false;
3709 // We start with the PHI node and scan for all of the users of this
3710 // instruction. All users must be instructions that can be used as reduction
3711 // variables (such as ADD). We must have a single out-of-block user. The cycle
3712 // must include the original PHI.
3713 bool FoundStartPHI = false;
3715 // To recognize min/max patterns formed by a icmp select sequence, we store
3716 // the number of instruction we saw from the recognized min/max pattern,
3717 // to make sure we only see exactly the two instructions.
3718 unsigned NumCmpSelectPatternInst = 0;
3719 ReductionInstDesc ReduxDesc(false, 0);
3721 SmallPtrSet<Instruction *, 8> VisitedInsts;
3722 SmallVector<Instruction *, 8> Worklist;
3723 Worklist.push_back(Phi);
3724 VisitedInsts.insert(Phi);
3726 // A value in the reduction can be used:
3727 // - By the reduction:
3728 // - Reduction operation:
3729 // - One use of reduction value (safe).
3730 // - Multiple use of reduction value (not safe).
3732 // - All uses of the PHI must be the reduction (safe).
3733 // - Otherwise, not safe.
3734 // - By one instruction outside of the loop (safe).
3735 // - By further instructions outside of the loop (not safe).
3736 // - By an instruction that is not part of the reduction (not safe).
3738 // * An instruction type other than PHI or the reduction operation.
3739 // * A PHI in the header other than the initial PHI.
3740 while (!Worklist.empty()) {
3741 Instruction *Cur = Worklist.back();
3742 Worklist.pop_back();
3745 // If the instruction has no users then this is a broken chain and can't be
3746 // a reduction variable.
3747 if (Cur->use_empty())
3750 bool IsAPhi = isa<PHINode>(Cur);
3752 // A header PHI use other than the original PHI.
3753 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3756 // Reductions of instructions such as Div, and Sub is only possible if the
3757 // LHS is the reduction variable.
3758 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3759 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3760 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3763 // Any reduction instruction must be of one of the allowed kinds.
3764 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3765 if (!ReduxDesc.IsReduction)
3768 // A reduction operation must only have one use of the reduction value.
3769 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3770 hasMultipleUsesOf(Cur, VisitedInsts))
3773 // All inputs to a PHI node must be a reduction value.
3774 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3777 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3778 isa<SelectInst>(Cur)))
3779 ++NumCmpSelectPatternInst;
3780 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3781 isa<SelectInst>(Cur)))
3782 ++NumCmpSelectPatternInst;
3784 // Check whether we found a reduction operator.
3785 FoundReduxOp |= !IsAPhi;
3787 // Process users of current instruction. Push non PHI nodes after PHI nodes
3788 // onto the stack. This way we are going to have seen all inputs to PHI
3789 // nodes once we get to them.
3790 SmallVector<Instruction *, 8> NonPHIs;
3791 SmallVector<Instruction *, 8> PHIs;
3792 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3794 Instruction *Usr = cast<Instruction>(*UI);
3796 // Check if we found the exit user.
3797 BasicBlock *Parent = Usr->getParent();
3798 if (!TheLoop->contains(Parent)) {
3799 // Exit if you find multiple outside users.
3800 if (ExitInstruction != 0)
3802 ExitInstruction = Cur;
3806 // Process instructions only once (termination).
3807 if (VisitedInsts.insert(Usr)) {
3808 if (isa<PHINode>(Usr))
3809 PHIs.push_back(Usr);
3811 NonPHIs.push_back(Usr);
3813 // Remember that we completed the cycle.
3815 FoundStartPHI = true;
3817 Worklist.append(PHIs.begin(), PHIs.end());
3818 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3821 // This means we have seen one but not the other instruction of the
3822 // pattern or more than just a select and cmp.
3823 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3824 NumCmpSelectPatternInst != 2)
3827 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3830 // We found a reduction var if we have reached the original phi node and we
3831 // only have a single instruction with out-of-loop users.
3833 // This instruction is allowed to have out-of-loop users.
3834 AllowedExit.insert(ExitInstruction);
3836 // Save the description of this reduction variable.
3837 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3838 ReduxDesc.MinMaxKind);
3839 Reductions[Phi] = RD;
3840 // We've ended the cycle. This is a reduction variable if we have an
3841 // outside user and it has a binary op.
3846 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3847 /// pattern corresponding to a min(X, Y) or max(X, Y).
3848 LoopVectorizationLegality::ReductionInstDesc
3849 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3850 ReductionInstDesc &Prev) {
3852 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3853 "Expect a select instruction");
3854 Instruction *Cmp = 0;
3855 SelectInst *Select = 0;
3857 // We must handle the select(cmp()) as a single instruction. Advance to the
3859 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3860 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3861 return ReductionInstDesc(false, I);
3862 return ReductionInstDesc(Select, Prev.MinMaxKind);
3865 // Only handle single use cases for now.
3866 if (!(Select = dyn_cast<SelectInst>(I)))
3867 return ReductionInstDesc(false, I);
3868 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3869 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3870 return ReductionInstDesc(false, I);
3871 if (!Cmp->hasOneUse())
3872 return ReductionInstDesc(false, I);
3877 // Look for a min/max pattern.
3878 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3879 return ReductionInstDesc(Select, MRK_UIntMin);
3880 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3881 return ReductionInstDesc(Select, MRK_UIntMax);
3882 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3883 return ReductionInstDesc(Select, MRK_SIntMax);
3884 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3885 return ReductionInstDesc(Select, MRK_SIntMin);
3886 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3887 return ReductionInstDesc(Select, MRK_FloatMin);
3888 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3889 return ReductionInstDesc(Select, MRK_FloatMax);
3890 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3891 return ReductionInstDesc(Select, MRK_FloatMin);
3892 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3893 return ReductionInstDesc(Select, MRK_FloatMax);
3895 return ReductionInstDesc(false, I);
3898 LoopVectorizationLegality::ReductionInstDesc
3899 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3901 ReductionInstDesc &Prev) {
3902 bool FP = I->getType()->isFloatingPointTy();
3903 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3904 switch (I->getOpcode()) {
3906 return ReductionInstDesc(false, I);
3907 case Instruction::PHI:
3908 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3909 Kind != RK_FloatMinMax))
3910 return ReductionInstDesc(false, I);
3911 return ReductionInstDesc(I, Prev.MinMaxKind);
3912 case Instruction::Sub:
3913 case Instruction::Add:
3914 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3915 case Instruction::Mul:
3916 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3917 case Instruction::And:
3918 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3919 case Instruction::Or:
3920 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3921 case Instruction::Xor:
3922 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3923 case Instruction::FMul:
3924 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3925 case Instruction::FAdd:
3926 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3927 case Instruction::FCmp:
3928 case Instruction::ICmp:
3929 case Instruction::Select:
3930 if (Kind != RK_IntegerMinMax &&
3931 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3932 return ReductionInstDesc(false, I);
3933 return isMinMaxSelectCmpPattern(I, Prev);
3937 LoopVectorizationLegality::InductionKind
3938 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3939 Type *PhiTy = Phi->getType();
3940 // We only handle integer and pointer inductions variables.
3941 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3942 return IK_NoInduction;
3944 // Check that the PHI is consecutive.
3945 const SCEV *PhiScev = SE->getSCEV(Phi);
3946 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3948 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3949 return IK_NoInduction;
3951 const SCEV *Step = AR->getStepRecurrence(*SE);
3953 // Integer inductions need to have a stride of one.
3954 if (PhiTy->isIntegerTy()) {
3956 return IK_IntInduction;
3957 if (Step->isAllOnesValue())
3958 return IK_ReverseIntInduction;
3959 return IK_NoInduction;
3962 // Calculate the pointer stride and check if it is consecutive.
3963 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3965 return IK_NoInduction;
3967 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3968 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3969 if (C->getValue()->equalsInt(Size))
3970 return IK_PtrInduction;
3971 else if (C->getValue()->equalsInt(0 - Size))
3972 return IK_ReversePtrInduction;
3974 return IK_NoInduction;
3977 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3978 Value *In0 = const_cast<Value*>(V);
3979 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3983 return Inductions.count(PN);
3986 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3987 assert(TheLoop->contains(BB) && "Unknown block used");
3989 // Blocks that do not dominate the latch need predication.
3990 BasicBlock* Latch = TheLoop->getLoopLatch();
3991 return !DT->dominates(BB, Latch);
3994 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3995 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3996 // We might be able to hoist the load.
3997 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
4000 // We don't predicate stores at the moment.
4001 if (it->mayWriteToMemory() || it->mayThrow())
4004 // The instructions below can trap.
4005 switch (it->getOpcode()) {
4007 case Instruction::UDiv:
4008 case Instruction::SDiv:
4009 case Instruction::URem:
4010 case Instruction::SRem:
4018 LoopVectorizationCostModel::VectorizationFactor
4019 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4021 // Width 1 means no vectorize
4022 VectorizationFactor Factor = { 1U, 0U };
4023 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4024 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4028 // Find the trip count.
4029 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4030 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4032 unsigned WidestType = getWidestType();
4033 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4034 unsigned MaxSafeDepDist = -1U;
4035 if (Legal->getMaxSafeDepDistBytes() != -1U)
4036 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4037 WidestRegister = WidestRegister < MaxSafeDepDist ? WidestRegister : MaxSafeDepDist;
4038 unsigned MaxVectorSize = WidestRegister / WidestType;
4039 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4040 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4042 if (MaxVectorSize == 0) {
4043 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4047 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4048 " into one vector!");
4050 unsigned VF = MaxVectorSize;
4052 // If we optimize the program for size, avoid creating the tail loop.
4054 // If we are unable to calculate the trip count then don't try to vectorize.
4056 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4060 // Find the maximum SIMD width that can fit within the trip count.
4061 VF = TC % MaxVectorSize;
4066 // If the trip count that we found modulo the vectorization factor is not
4067 // zero then we require a tail.
4069 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4075 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4076 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4078 Factor.Width = UserVF;
4082 float Cost = expectedCost(1);
4084 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4085 for (unsigned i=2; i <= VF; i*=2) {
4086 // Notice that the vector loop needs to be executed less times, so
4087 // we need to divide the cost of the vector loops by the width of
4088 // the vector elements.
4089 float VectorCost = expectedCost(i) / (float)i;
4090 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4091 (int)VectorCost << ".\n");
4092 if (VectorCost < Cost) {
4098 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4099 Factor.Width = Width;
4100 Factor.Cost = Width * Cost;
4104 unsigned LoopVectorizationCostModel::getWidestType() {
4105 unsigned MaxWidth = 8;
4108 for (Loop::block_iterator bb = TheLoop->block_begin(),
4109 be = TheLoop->block_end(); bb != be; ++bb) {
4110 BasicBlock *BB = *bb;
4112 // For each instruction in the loop.
4113 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4114 Type *T = it->getType();
4116 // Only examine Loads, Stores and PHINodes.
4117 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4120 // Examine PHI nodes that are reduction variables.
4121 if (PHINode *PN = dyn_cast<PHINode>(it))
4122 if (!Legal->getReductionVars()->count(PN))
4125 // Examine the stored values.
4126 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4127 T = ST->getValueOperand()->getType();
4129 // Ignore loaded pointer types and stored pointer types that are not
4130 // consecutive. However, we do want to take consecutive stores/loads of
4131 // pointer vectors into account.
4132 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4135 MaxWidth = std::max(MaxWidth,
4136 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4144 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4147 unsigned LoopCost) {
4149 // -- The unroll heuristics --
4150 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4151 // There are many micro-architectural considerations that we can't predict
4152 // at this level. For example frontend pressure (on decode or fetch) due to
4153 // code size, or the number and capabilities of the execution ports.
4155 // We use the following heuristics to select the unroll factor:
4156 // 1. If the code has reductions the we unroll in order to break the cross
4157 // iteration dependency.
4158 // 2. If the loop is really small then we unroll in order to reduce the loop
4160 // 3. We don't unroll if we think that we will spill registers to memory due
4161 // to the increased register pressure.
4163 // Use the user preference, unless 'auto' is selected.
4167 // When we optimize for size we don't unroll.
4171 // We used the distance for the unroll factor.
4172 if (Legal->getMaxSafeDepDistBytes() != -1U)
4175 // Do not unroll loops with a relatively small trip count.
4176 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4177 TheLoop->getLoopLatch());
4178 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4181 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4182 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4183 " vector registers\n");
4185 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4186 // We divide by these constants so assume that we have at least one
4187 // instruction that uses at least one register.
4188 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4189 R.NumInstructions = std::max(R.NumInstructions, 1U);
4191 // We calculate the unroll factor using the following formula.
4192 // Subtract the number of loop invariants from the number of available
4193 // registers. These registers are used by all of the unrolled instances.
4194 // Next, divide the remaining registers by the number of registers that is
4195 // required by the loop, in order to estimate how many parallel instances
4196 // fit without causing spills.
4197 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4199 // Clamp the unroll factor ranges to reasonable factors.
4200 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4202 // If we did not calculate the cost for VF (because the user selected the VF)
4203 // then we calculate the cost of VF here.
4205 LoopCost = expectedCost(VF);
4207 // Clamp the calculated UF to be between the 1 and the max unroll factor
4208 // that the target allows.
4209 if (UF > MaxUnrollSize)
4214 if (Legal->getReductionVars()->size()) {
4215 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4219 // We want to unroll tiny loops in order to reduce the loop overhead.
4220 // We assume that the cost overhead is 1 and we use the cost model
4221 // to estimate the cost of the loop and unroll until the cost of the
4222 // loop overhead is about 5% of the cost of the loop.
4223 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4224 if (LoopCost < 20) {
4225 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4226 unsigned NewUF = 20/LoopCost + 1;
4227 return std::min(NewUF, UF);
4230 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4234 LoopVectorizationCostModel::RegisterUsage
4235 LoopVectorizationCostModel::calculateRegisterUsage() {
4236 // This function calculates the register usage by measuring the highest number
4237 // of values that are alive at a single location. Obviously, this is a very
4238 // rough estimation. We scan the loop in a topological order in order and
4239 // assign a number to each instruction. We use RPO to ensure that defs are
4240 // met before their users. We assume that each instruction that has in-loop
4241 // users starts an interval. We record every time that an in-loop value is
4242 // used, so we have a list of the first and last occurrences of each
4243 // instruction. Next, we transpose this data structure into a multi map that
4244 // holds the list of intervals that *end* at a specific location. This multi
4245 // map allows us to perform a linear search. We scan the instructions linearly
4246 // and record each time that a new interval starts, by placing it in a set.
4247 // If we find this value in the multi-map then we remove it from the set.
4248 // The max register usage is the maximum size of the set.
4249 // We also search for instructions that are defined outside the loop, but are
4250 // used inside the loop. We need this number separately from the max-interval
4251 // usage number because when we unroll, loop-invariant values do not take
4253 LoopBlocksDFS DFS(TheLoop);
4257 R.NumInstructions = 0;
4259 // Each 'key' in the map opens a new interval. The values
4260 // of the map are the index of the 'last seen' usage of the
4261 // instruction that is the key.
4262 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4263 // Maps instruction to its index.
4264 DenseMap<unsigned, Instruction*> IdxToInstr;
4265 // Marks the end of each interval.
4266 IntervalMap EndPoint;
4267 // Saves the list of instruction indices that are used in the loop.
4268 SmallSet<Instruction*, 8> Ends;
4269 // Saves the list of values that are used in the loop but are
4270 // defined outside the loop, such as arguments and constants.
4271 SmallPtrSet<Value*, 8> LoopInvariants;
4274 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4275 be = DFS.endRPO(); bb != be; ++bb) {
4276 R.NumInstructions += (*bb)->size();
4277 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4279 Instruction *I = it;
4280 IdxToInstr[Index++] = I;
4282 // Save the end location of each USE.
4283 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4284 Value *U = I->getOperand(i);
4285 Instruction *Instr = dyn_cast<Instruction>(U);
4287 // Ignore non-instruction values such as arguments, constants, etc.
4288 if (!Instr) continue;
4290 // If this instruction is outside the loop then record it and continue.
4291 if (!TheLoop->contains(Instr)) {
4292 LoopInvariants.insert(Instr);
4296 // Overwrite previous end points.
4297 EndPoint[Instr] = Index;
4303 // Saves the list of intervals that end with the index in 'key'.
4304 typedef SmallVector<Instruction*, 2> InstrList;
4305 DenseMap<unsigned, InstrList> TransposeEnds;
4307 // Transpose the EndPoints to a list of values that end at each index.
4308 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4310 TransposeEnds[it->second].push_back(it->first);
4312 SmallSet<Instruction*, 8> OpenIntervals;
4313 unsigned MaxUsage = 0;
4316 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4317 for (unsigned int i = 0; i < Index; ++i) {
4318 Instruction *I = IdxToInstr[i];
4319 // Ignore instructions that are never used within the loop.
4320 if (!Ends.count(I)) continue;
4322 // Remove all of the instructions that end at this location.
4323 InstrList &List = TransposeEnds[i];
4324 for (unsigned int j=0, e = List.size(); j < e; ++j)
4325 OpenIntervals.erase(List[j]);
4327 // Count the number of live interals.
4328 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4330 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4331 OpenIntervals.size() <<"\n");
4333 // Add the current instruction to the list of open intervals.
4334 OpenIntervals.insert(I);
4337 unsigned Invariant = LoopInvariants.size();
4338 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4339 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4340 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4342 R.LoopInvariantRegs = Invariant;
4343 R.MaxLocalUsers = MaxUsage;
4347 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4351 for (Loop::block_iterator bb = TheLoop->block_begin(),
4352 be = TheLoop->block_end(); bb != be; ++bb) {
4353 unsigned BlockCost = 0;
4354 BasicBlock *BB = *bb;
4356 // For each instruction in the old loop.
4357 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4358 // Skip dbg intrinsics.
4359 if (isa<DbgInfoIntrinsic>(it))
4362 unsigned C = getInstructionCost(it, VF);
4364 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4365 VF << " For instruction: "<< *it << "\n");
4368 // We assume that if-converted blocks have a 50% chance of being executed.
4369 // When the code is scalar then some of the blocks are avoided due to CF.
4370 // When the code is vectorized we execute all code paths.
4371 if (Legal->blockNeedsPredication(*bb) && VF == 1)
4381 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4382 // If we know that this instruction will remain uniform, check the cost of
4383 // the scalar version.
4384 if (Legal->isUniformAfterVectorization(I))
4387 Type *RetTy = I->getType();
4388 Type *VectorTy = ToVectorTy(RetTy, VF);
4390 // TODO: We need to estimate the cost of intrinsic calls.
4391 switch (I->getOpcode()) {
4392 case Instruction::GetElementPtr:
4393 // We mark this instruction as zero-cost because the cost of GEPs in
4394 // vectorized code depends on whether the corresponding memory instruction
4395 // is scalarized or not. Therefore, we handle GEPs with the memory
4396 // instruction cost.
4398 case Instruction::Br: {
4399 return TTI.getCFInstrCost(I->getOpcode());
4401 case Instruction::PHI:
4402 //TODO: IF-converted IFs become selects.
4404 case Instruction::Add:
4405 case Instruction::FAdd:
4406 case Instruction::Sub:
4407 case Instruction::FSub:
4408 case Instruction::Mul:
4409 case Instruction::FMul:
4410 case Instruction::UDiv:
4411 case Instruction::SDiv:
4412 case Instruction::FDiv:
4413 case Instruction::URem:
4414 case Instruction::SRem:
4415 case Instruction::FRem:
4416 case Instruction::Shl:
4417 case Instruction::LShr:
4418 case Instruction::AShr:
4419 case Instruction::And:
4420 case Instruction::Or:
4421 case Instruction::Xor: {
4422 // Certain instructions can be cheaper to vectorize if they have a constant
4423 // second vector operand. One example of this are shifts on x86.
4424 TargetTransformInfo::OperandValueKind Op1VK =
4425 TargetTransformInfo::OK_AnyValue;
4426 TargetTransformInfo::OperandValueKind Op2VK =
4427 TargetTransformInfo::OK_AnyValue;
4429 if (isa<ConstantInt>(I->getOperand(1)))
4430 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4432 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4434 case Instruction::Select: {
4435 SelectInst *SI = cast<SelectInst>(I);
4436 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4437 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4438 Type *CondTy = SI->getCondition()->getType();
4440 CondTy = VectorType::get(CondTy, VF);
4442 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4444 case Instruction::ICmp:
4445 case Instruction::FCmp: {
4446 Type *ValTy = I->getOperand(0)->getType();
4447 VectorTy = ToVectorTy(ValTy, VF);
4448 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4450 case Instruction::Store:
4451 case Instruction::Load: {
4452 StoreInst *SI = dyn_cast<StoreInst>(I);
4453 LoadInst *LI = dyn_cast<LoadInst>(I);
4454 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4456 VectorTy = ToVectorTy(ValTy, VF);
4458 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4459 unsigned AS = SI ? SI->getPointerAddressSpace() :
4460 LI->getPointerAddressSpace();
4461 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4462 // We add the cost of address computation here instead of with the gep
4463 // instruction because only here we know whether the operation is
4466 return TTI.getAddressComputationCost(VectorTy) +
4467 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4469 // Scalarized loads/stores.
4470 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4471 bool Reverse = ConsecutiveStride < 0;
4472 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4473 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4474 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4476 // The cost of extracting from the value vector and pointer vector.
4477 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4478 for (unsigned i = 0; i < VF; ++i) {
4479 // The cost of extracting the pointer operand.
4480 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4481 // In case of STORE, the cost of ExtractElement from the vector.
4482 // In case of LOAD, the cost of InsertElement into the returned
4484 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4485 Instruction::InsertElement,
4489 // The cost of the scalar loads/stores.
4490 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
4491 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4496 // Wide load/stores.
4497 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4498 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4501 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4505 case Instruction::ZExt:
4506 case Instruction::SExt:
4507 case Instruction::FPToUI:
4508 case Instruction::FPToSI:
4509 case Instruction::FPExt:
4510 case Instruction::PtrToInt:
4511 case Instruction::IntToPtr:
4512 case Instruction::SIToFP:
4513 case Instruction::UIToFP:
4514 case Instruction::Trunc:
4515 case Instruction::FPTrunc:
4516 case Instruction::BitCast: {
4517 // We optimize the truncation of induction variable.
4518 // The cost of these is the same as the scalar operation.
4519 if (I->getOpcode() == Instruction::Trunc &&
4520 Legal->isInductionVariable(I->getOperand(0)))
4521 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4522 I->getOperand(0)->getType());
4524 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4525 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4527 case Instruction::Call: {
4528 CallInst *CI = cast<CallInst>(I);
4529 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4530 assert(ID && "Not an intrinsic call!");
4531 Type *RetTy = ToVectorTy(CI->getType(), VF);
4532 SmallVector<Type*, 4> Tys;
4533 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4534 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4535 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4538 // We are scalarizing the instruction. Return the cost of the scalar
4539 // instruction, plus the cost of insert and extract into vector
4540 // elements, times the vector width.
4543 if (!RetTy->isVoidTy() && VF != 1) {
4544 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4546 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4549 // The cost of inserting the results plus extracting each one of the
4551 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4554 // The cost of executing VF copies of the scalar instruction. This opcode
4555 // is unknown. Assume that it is the same as 'mul'.
4556 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4562 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4563 if (Scalar->isVoidTy() || VF == 1)
4565 return VectorType::get(Scalar, VF);
4568 char LoopVectorize::ID = 0;
4569 static const char lv_name[] = "Loop Vectorization";
4570 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4571 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4572 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4573 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4574 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4577 Pass *createLoopVectorizePass() {
4578 return new LoopVectorize();
4582 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4583 // Check for a store.
4584 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4585 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4587 // Check for a load.
4588 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4589 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;