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/MapVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/StringExtras.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/AliasSetTracker.h"
57 #include "llvm/Analysis/Dominators.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopIterator.h"
60 #include "llvm/Analysis/LoopPass.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpander.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/TargetTransformInfo.h"
65 #include "llvm/Analysis/ValueTracking.h"
66 #include "llvm/Analysis/Verifier.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Function.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/LLVMContext.h"
75 #include "llvm/IR/Module.h"
76 #include "llvm/IR/Type.h"
77 #include "llvm/IR/Value.h"
78 #include "llvm/Pass.h"
79 #include "llvm/Support/CommandLine.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/PatternMatch.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/Support/ValueHandle.h"
84 #include "llvm/Target/TargetLibraryInfo.h"
85 #include "llvm/Transforms/Scalar.h"
86 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
87 #include "llvm/Transforms/Utils/Local.h"
92 using namespace llvm::PatternMatch;
94 static cl::opt<unsigned>
95 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
96 cl::desc("Sets the SIMD width. Zero is autoselect."));
98 static cl::opt<unsigned>
99 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
100 cl::desc("Sets the vectorization unroll count. "
101 "Zero is autoselect."));
104 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
105 cl::desc("Enable if-conversion during vectorization."));
107 /// We don't vectorize loops with a known constant trip count below this number.
108 static cl::opt<unsigned>
109 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
111 cl::desc("Don't vectorize loops with a constant "
112 "trip count that is smaller than this "
115 /// We don't unroll loops with a known constant trip count below this number.
116 static const unsigned TinyTripCountUnrollThreshold = 128;
118 /// When performing memory disambiguation checks at runtime do not make more
119 /// than this number of comparisons.
120 static const unsigned RuntimeMemoryCheckThreshold = 8;
122 /// We use a metadata with this name to indicate that a scalar loop was
123 /// vectorized and that we don't need to re-vectorize it if we run into it
126 AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized";
130 // Forward declarations.
131 class LoopVectorizationLegality;
132 class LoopVectorizationCostModel;
134 /// InnerLoopVectorizer vectorizes loops which contain only one basic
135 /// block to a specified vectorization factor (VF).
136 /// This class performs the widening of scalars into vectors, or multiple
137 /// scalars. This class also implements the following features:
138 /// * It inserts an epilogue loop for handling loops that don't have iteration
139 /// counts that are known to be a multiple of the vectorization factor.
140 /// * It handles the code generation for reduction variables.
141 /// * Scalarization (implementation using scalars) of un-vectorizable
143 /// InnerLoopVectorizer does not perform any vectorization-legality
144 /// checks, and relies on the caller to check for the different legality
145 /// aspects. The InnerLoopVectorizer relies on the
146 /// LoopVectorizationLegality class to provide information about the induction
147 /// and reduction variables that were found to a given vectorization factor.
148 class InnerLoopVectorizer {
150 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
151 DominatorTree *DT, DataLayout *DL,
152 const TargetLibraryInfo *TLI, unsigned VecWidth,
153 unsigned UnrollFactor)
154 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
155 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
156 OldInduction(0), WidenMap(UnrollFactor) {}
158 // Perform the actual loop widening (vectorization).
159 void vectorize(LoopVectorizationLegality *Legal) {
160 // Create a new empty loop. Unlink the old loop and connect the new one.
161 createEmptyLoop(Legal);
162 // Widen each instruction in the old loop to a new one in the new loop.
163 // Use the Legality module to find the induction and reduction variables.
164 vectorizeLoop(Legal);
165 // Register the new loop and update the analysis passes.
170 /// A small list of PHINodes.
171 typedef SmallVector<PHINode*, 4> PhiVector;
172 /// When we unroll loops we have multiple vector values for each scalar.
173 /// This data structure holds the unrolled and vectorized values that
174 /// originated from one scalar instruction.
175 typedef SmallVector<Value*, 2> VectorParts;
177 /// Add code that checks at runtime if the accessed arrays overlap.
178 /// Returns the comparator value or NULL if no check is needed.
179 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
181 /// Create an empty loop, based on the loop ranges of the old loop.
182 void createEmptyLoop(LoopVectorizationLegality *Legal);
183 /// Copy and widen the instructions from the old loop.
184 void vectorizeLoop(LoopVectorizationLegality *Legal);
186 /// A helper function that computes the predicate of the block BB, assuming
187 /// that the header block of the loop is set to True. It returns the *entry*
188 /// mask for the block BB.
189 VectorParts createBlockInMask(BasicBlock *BB);
190 /// A helper function that computes the predicate of the edge between SRC
192 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
194 /// A helper function to vectorize a single BB within the innermost loop.
195 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
198 /// Insert the new loop to the loop hierarchy and pass manager
199 /// and update the analysis passes.
200 void updateAnalysis();
202 /// This instruction is un-vectorizable. Implement it as a sequence
204 void scalarizeInstruction(Instruction *Instr);
206 /// Vectorize Load and Store instructions,
207 void vectorizeMemoryInstruction(Instruction *Instr,
208 LoopVectorizationLegality *Legal);
210 /// Create a broadcast instruction. This method generates a broadcast
211 /// instruction (shuffle) for loop invariant values and for the induction
212 /// value. If this is the induction variable then we extend it to N, N+1, ...
213 /// this is needed because each iteration in the loop corresponds to a SIMD
215 Value *getBroadcastInstrs(Value *V);
217 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
218 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
219 /// The sequence starts at StartIndex.
220 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
222 /// When we go over instructions in the basic block we rely on previous
223 /// values within the current basic block or on loop invariant values.
224 /// When we widen (vectorize) values we place them in the map. If the values
225 /// are not within the map, they have to be loop invariant, so we simply
226 /// broadcast them into a vector.
227 VectorParts &getVectorValue(Value *V);
229 /// Generate a shuffle sequence that will reverse the vector Vec.
230 Value *reverseVector(Value *Vec);
232 /// This is a helper class that holds the vectorizer state. It maps scalar
233 /// instructions to vector instructions. When the code is 'unrolled' then
234 /// then a single scalar value is mapped to multiple vector parts. The parts
235 /// are stored in the VectorPart type.
237 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
239 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
241 /// \return True if 'Key' is saved in the Value Map.
242 bool has(Value *Key) const { return MapStorage.count(Key); }
244 /// Initializes a new entry in the map. Sets all of the vector parts to the
245 /// save value in 'Val'.
246 /// \return A reference to a vector with splat values.
247 VectorParts &splat(Value *Key, Value *Val) {
248 VectorParts &Entry = MapStorage[Key];
249 Entry.assign(UF, Val);
253 ///\return A reference to the value that is stored at 'Key'.
254 VectorParts &get(Value *Key) {
255 VectorParts &Entry = MapStorage[Key];
258 assert(Entry.size() == UF);
263 /// The unroll factor. Each entry in the map stores this number of vector
267 /// Map storage. We use std::map and not DenseMap because insertions to a
268 /// dense map invalidates its iterators.
269 std::map<Value *, VectorParts> MapStorage;
272 /// The original loop.
274 /// Scev analysis to use.
282 /// Target Library Info.
283 const TargetLibraryInfo *TLI;
285 /// The vectorization SIMD factor to use. Each vector will have this many
288 /// The vectorization unroll factor to use. Each scalar is vectorized to this
289 /// many different vector instructions.
292 /// The builder that we use
295 // --- Vectorization state ---
297 /// The vector-loop preheader.
298 BasicBlock *LoopVectorPreHeader;
299 /// The scalar-loop preheader.
300 BasicBlock *LoopScalarPreHeader;
301 /// Middle Block between the vector and the scalar.
302 BasicBlock *LoopMiddleBlock;
303 ///The ExitBlock of the scalar loop.
304 BasicBlock *LoopExitBlock;
305 ///The vector loop body.
306 BasicBlock *LoopVectorBody;
307 ///The scalar loop body.
308 BasicBlock *LoopScalarBody;
309 /// A list of all bypass blocks. The first block is the entry of the loop.
310 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
312 /// The new Induction variable which was added to the new block.
314 /// The induction variable of the old basic block.
315 PHINode *OldInduction;
316 /// Holds the extended (to the widest induction type) start index.
318 /// Maps scalars to widened vectors.
322 /// \brief Check if conditionally executed loads are hoistable.
324 /// This class has two functions: isHoistableLoad and canHoistAllLoads.
325 /// isHoistableLoad should be called on all load instructions that are executed
326 /// conditionally. After all conditional loads are processed, the client should
327 /// call canHoistAllLoads to determine if all of the conditional executed loads
328 /// have an unconditional memory access to the same memory address in the loop.
330 typedef SmallPtrSet<Value *, 8> MemorySet;
334 MemorySet CondLoadAddrSet;
337 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
339 /// \brief Check if the instruction is a load with a identifiable address.
340 bool isHoistableLoad(Instruction *L);
342 /// \brief Check if all of the conditional loads are hoistable because there
343 /// exists an unconditional memory access to the same address in the loop.
344 bool canHoistAllLoads();
347 bool LoadHoisting::isHoistableLoad(Instruction *L) {
348 LoadInst *LI = dyn_cast<LoadInst>(L);
352 CondLoadAddrSet.insert(LI->getPointerOperand());
356 static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
357 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
358 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
359 Set.insert(LI->getPointerOperand());
360 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
361 Set.insert(SI->getPointerOperand());
365 bool LoadHoisting::canHoistAllLoads() {
366 // No conditional loads.
367 if (CondLoadAddrSet.empty())
370 MemorySet UncondMemAccesses;
371 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
372 BasicBlock *LoopLatch = TheLoop->getLoopLatch();
374 // Iterate over the unconditional blocks and collect memory access addresses.
375 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
376 BasicBlock *BB = LoopBlocks[i];
378 // Ignore conditional blocks.
379 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
382 addMemAccesses(BB, UncondMemAccesses);
385 // And make sure there is a matching unconditional access for every
387 for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
388 ME = CondLoadAddrSet.end(); MI != ME; ++MI)
389 if (!UncondMemAccesses.count(*MI))
395 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
396 /// to what vectorization factor.
397 /// This class does not look at the profitability of vectorization, only the
398 /// legality. This class has two main kinds of checks:
399 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
400 /// will change the order of memory accesses in a way that will change the
401 /// correctness of the program.
402 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
403 /// checks for a number of different conditions, such as the availability of a
404 /// single induction variable, that all types are supported and vectorize-able,
405 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
406 /// This class is also used by InnerLoopVectorizer for identifying
407 /// induction variable and the different reduction variables.
408 class LoopVectorizationLegality {
410 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
411 DominatorTree *DT, TargetTransformInfo* TTI,
412 AliasAnalysis *AA, TargetLibraryInfo *TLI)
413 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
414 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
415 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);
504 /// This flag indicates if we need to add the runtime check.
506 /// Holds the pointers that we need to check.
507 SmallVector<TrackingVH<Value>, 2> Pointers;
508 /// Holds the pointer value at the beginning of the loop.
509 SmallVector<const SCEV*, 2> Starts;
510 /// Holds the pointer value at the end of the loop.
511 SmallVector<const SCEV*, 2> Ends;
512 /// Holds the information if this pointer is used for writing to memory.
513 SmallVector<bool, 2> IsWritePtr;
516 /// A POD for saving information about induction variables.
517 struct InductionInfo {
518 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
519 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
521 TrackingVH<Value> StartValue;
526 /// ReductionList contains the reduction descriptors for all
527 /// of the reductions that were found in the loop.
528 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
530 /// InductionList saves induction variables and maps them to the
531 /// induction descriptor.
532 typedef MapVector<PHINode*, InductionInfo> InductionList;
534 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
535 /// respective Store/Load instruction(s) to calculate aliasing.
536 typedef MapVector<Value*, Instruction* > AliasMap;
537 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
539 /// Returns true if it is legal to vectorize this loop.
540 /// This does not mean that it is profitable to vectorize this
541 /// loop, only that it is legal to do so.
544 /// Returns the Induction variable.
545 PHINode *getInduction() { return Induction; }
547 /// Returns the reduction variables found in the loop.
548 ReductionList *getReductionVars() { return &Reductions; }
550 /// Returns the induction variables found in the loop.
551 InductionList *getInductionVars() { return &Inductions; }
553 /// Returns the widest induction type.
554 Type *getWidestInductionType() { return WidestIndTy; }
556 /// Returns True if V is an induction variable in this loop.
557 bool isInductionVariable(const Value *V);
559 /// Return true if the block BB needs to be predicated in order for the loop
560 /// to be vectorized.
561 bool blockNeedsPredication(BasicBlock *BB);
563 /// Check if this pointer is consecutive when vectorizing. This happens
564 /// when the last index of the GEP is the induction variable, or that the
565 /// pointer itself is an induction variable.
566 /// This check allows us to vectorize A[idx] into a wide load/store.
568 /// 0 - Stride is unknown or non consecutive.
569 /// 1 - Address is consecutive.
570 /// -1 - Address is consecutive, and decreasing.
571 int isConsecutivePtr(Value *Ptr);
573 /// Returns true if the value V is uniform within the loop.
574 bool isUniform(Value *V);
576 /// Returns true if this instruction will remain scalar after vectorization.
577 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
579 /// Returns the information that we collected about runtime memory check.
580 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
582 /// This function returns the identity element (or neutral element) for
584 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
586 /// Check if a single basic block loop is vectorizable.
587 /// At this point we know that this is a loop with a constant trip count
588 /// and we only need to check individual instructions.
589 bool canVectorizeInstrs();
591 /// When we vectorize loops we may change the order in which
592 /// we read and write from memory. This method checks if it is
593 /// legal to vectorize the code, considering only memory constrains.
594 /// Returns true if the loop is vectorizable
595 bool canVectorizeMemory();
597 /// Return true if we can vectorize this loop using the IF-conversion
599 bool canVectorizeWithIfConvert();
601 /// Collect the variables that need to stay uniform after vectorization.
602 void collectLoopUniforms();
604 /// Return true if all of the instructions in the block can be speculatively
606 bool blockCanBePredicated(BasicBlock *BB);
608 /// Returns True, if 'Phi' is the kind of reduction variable for type
609 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
610 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
611 /// Returns a struct describing if the instruction 'I' can be a reduction
612 /// variable of type 'Kind'. If the reduction is a min/max pattern of
613 /// select(icmp()) this function advances the instruction pointer 'I' from the
614 /// compare instruction to the select instruction and stores this pointer in
615 /// 'PatternLastInst' member of the returned struct.
616 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
617 ReductionInstDesc &Desc);
618 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
619 /// pattern corresponding to a min(X, Y) or max(X, Y).
620 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
621 ReductionInstDesc &Prev);
622 /// Returns the induction kind of Phi. This function may return NoInduction
623 /// if the PHI is not an induction variable.
624 InductionKind isInductionVariable(PHINode *Phi);
625 /// Return true if can compute the address bounds of Ptr within the loop.
626 bool hasComputableBounds(Value *Ptr);
627 /// Return true if there is the chance of write reorder.
628 bool hasPossibleGlobalWriteReorder(Value *Object,
630 AliasMultiMap &WriteObjects,
631 unsigned MaxByteWidth);
632 /// Return the AA location for a load or a store.
633 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
636 /// The loop that we evaluate.
640 /// DataLayout analysis.
645 TargetTransformInfo *TTI;
648 /// Target Library Info.
649 TargetLibraryInfo *TLI;
651 // --- vectorization state --- //
653 /// Holds the integer induction variable. This is the counter of the
656 /// Holds the reduction variables.
657 ReductionList Reductions;
658 /// Holds all of the induction variables that we found in the loop.
659 /// Notice that inductions don't need to start at zero and that induction
660 /// variables can be pointers.
661 InductionList Inductions;
662 /// Holds the widest induction type encountered.
665 /// Allowed outside users. This holds the reduction
666 /// vars which can be accessed from outside the loop.
667 SmallPtrSet<Value*, 4> AllowedExit;
668 /// This set holds the variables which are known to be uniform after
670 SmallPtrSet<Instruction*, 4> Uniforms;
671 /// We need to check that all of the pointers in this list are disjoint
673 RuntimePointerCheck PtrRtCheck;
674 /// Can we assume the absence of NaNs.
675 bool HasFunNoNaNAttr;
677 /// Utility to determine whether loads can be speculated.
678 LoadHoisting LoadSpeculation;
681 /// LoopVectorizationCostModel - estimates the expected speedups due to
683 /// In many cases vectorization is not profitable. This can happen because of
684 /// a number of reasons. In this class we mainly attempt to predict the
685 /// expected speedup/slowdowns due to the supported instruction set. We use the
686 /// TargetTransformInfo to query the different backends for the cost of
687 /// different operations.
688 class LoopVectorizationCostModel {
690 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
691 LoopVectorizationLegality *Legal,
692 const TargetTransformInfo &TTI,
693 DataLayout *DL, const TargetLibraryInfo *TLI)
694 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
696 /// Information about vectorization costs
697 struct VectorizationFactor {
698 unsigned Width; // Vector width with best cost
699 unsigned Cost; // Cost of the loop with that width
701 /// \return The most profitable vectorization factor and the cost of that VF.
702 /// This method checks every power of two up to VF. If UserVF is not ZERO
703 /// then this vectorization factor will be selected if vectorization is
705 VectorizationFactor selectVectorizationFactor(bool OptForSize,
708 /// \return The size (in bits) of the widest type in the code that
709 /// needs to be vectorized. We ignore values that remain scalar such as
710 /// 64 bit loop indices.
711 unsigned getWidestType();
713 /// \return The most profitable unroll factor.
714 /// If UserUF is non-zero then this method finds the best unroll-factor
715 /// based on register pressure and other parameters.
716 /// VF and LoopCost are the selected vectorization factor and the cost of the
718 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
721 /// \brief A struct that represents some properties of the register usage
723 struct RegisterUsage {
724 /// Holds the number of loop invariant values that are used in the loop.
725 unsigned LoopInvariantRegs;
726 /// Holds the maximum number of concurrent live intervals in the loop.
727 unsigned MaxLocalUsers;
728 /// Holds the number of instructions in the loop.
729 unsigned NumInstructions;
732 /// \return information about the register usage of the loop.
733 RegisterUsage calculateRegisterUsage();
736 /// Returns the expected execution cost. The unit of the cost does
737 /// not matter because we use the 'cost' units to compare different
738 /// vector widths. The cost that is returned is *not* normalized by
739 /// the factor width.
740 unsigned expectedCost(unsigned VF);
742 /// Returns the execution time cost of an instruction for a given vector
743 /// width. Vector width of one means scalar.
744 unsigned getInstructionCost(Instruction *I, unsigned VF);
746 /// A helper function for converting Scalar types to vector types.
747 /// If the incoming type is void, we return void. If the VF is 1, we return
749 static Type* ToVectorTy(Type *Scalar, unsigned VF);
751 /// Returns whether the instruction is a load or store and will be a emitted
752 /// as a vector operation.
753 bool isConsecutiveLoadOrStore(Instruction *I);
755 /// The loop that we evaluate.
759 /// Loop Info analysis.
761 /// Vectorization legality.
762 LoopVectorizationLegality *Legal;
763 /// Vector target information.
764 const TargetTransformInfo &TTI;
765 /// Target data layout information.
767 /// Target Library Info.
768 const TargetLibraryInfo *TLI;
771 /// The LoopVectorize Pass.
772 struct LoopVectorize : public LoopPass {
773 /// Pass identification, replacement for typeid
776 explicit LoopVectorize() : LoopPass(ID) {
777 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
783 TargetTransformInfo *TTI;
786 TargetLibraryInfo *TLI;
788 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
789 // We only vectorize innermost loops.
793 SE = &getAnalysis<ScalarEvolution>();
794 DL = getAnalysisIfAvailable<DataLayout>();
795 LI = &getAnalysis<LoopInfo>();
796 TTI = &getAnalysis<TargetTransformInfo>();
797 DT = &getAnalysis<DominatorTree>();
798 AA = getAnalysisIfAvailable<AliasAnalysis>();
799 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
802 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
806 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
807 L->getHeader()->getParent()->getName() << "\"\n");
809 // Check if it is legal to vectorize the loop.
810 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
811 if (!LVL.canVectorize()) {
812 DEBUG(dbgs() << "LV: Not vectorizing.\n");
816 // Use the cost model.
817 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
819 // Check the function attributes to find out if this function should be
820 // optimized for size.
821 Function *F = L->getHeader()->getParent();
822 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
823 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
824 unsigned FnIndex = AttributeSet::FunctionIndex;
825 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
826 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
829 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
830 "attribute is used.\n");
834 // Select the optimal vectorization factor.
835 LoopVectorizationCostModel::VectorizationFactor VF;
836 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
837 // Select the unroll factor.
838 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
842 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
846 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
847 F->getParent()->getModuleIdentifier()<<"\n");
848 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
850 // If we decided that it is *legal* to vectorize the loop then do it.
851 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
854 DEBUG(verifyFunction(*L->getHeader()->getParent()));
858 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
859 LoopPass::getAnalysisUsage(AU);
860 AU.addRequiredID(LoopSimplifyID);
861 AU.addRequiredID(LCSSAID);
862 AU.addRequired<DominatorTree>();
863 AU.addRequired<LoopInfo>();
864 AU.addRequired<ScalarEvolution>();
865 AU.addRequired<TargetTransformInfo>();
866 AU.addPreserved<LoopInfo>();
867 AU.addPreserved<DominatorTree>();
872 } // end anonymous namespace
874 //===----------------------------------------------------------------------===//
875 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
876 // LoopVectorizationCostModel.
877 //===----------------------------------------------------------------------===//
880 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
881 Loop *Lp, Value *Ptr,
883 const SCEV *Sc = SE->getSCEV(Ptr);
884 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
885 assert(AR && "Invalid addrec expression");
886 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
887 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
888 Pointers.push_back(Ptr);
889 Starts.push_back(AR->getStart());
890 Ends.push_back(ScEnd);
891 IsWritePtr.push_back(WritePtr);
894 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
895 // Save the current insertion location.
896 Instruction *Loc = Builder.GetInsertPoint();
898 // We need to place the broadcast of invariant variables outside the loop.
899 Instruction *Instr = dyn_cast<Instruction>(V);
900 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
901 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
903 // Place the code for broadcasting invariant variables in the new preheader.
905 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
907 // Broadcast the scalar into all locations in the vector.
908 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
910 // Restore the builder insertion point.
912 Builder.SetInsertPoint(Loc);
917 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
919 assert(Val->getType()->isVectorTy() && "Must be a vector");
920 assert(Val->getType()->getScalarType()->isIntegerTy() &&
921 "Elem must be an integer");
923 Type *ITy = Val->getType()->getScalarType();
924 VectorType *Ty = cast<VectorType>(Val->getType());
925 int VLen = Ty->getNumElements();
926 SmallVector<Constant*, 8> Indices;
928 // Create a vector of consecutive numbers from zero to VF.
929 for (int i = 0; i < VLen; ++i) {
930 int64_t Idx = Negate ? (-i) : i;
931 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
934 // Add the consecutive indices to the vector value.
935 Constant *Cv = ConstantVector::get(Indices);
936 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
937 return Builder.CreateAdd(Val, Cv, "induction");
940 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
941 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
942 // Make sure that the pointer does not point to structs.
943 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
946 // If this value is a pointer induction variable we know it is consecutive.
947 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
948 if (Phi && Inductions.count(Phi)) {
949 InductionInfo II = Inductions[Phi];
950 if (IK_PtrInduction == II.IK)
952 else if (IK_ReversePtrInduction == II.IK)
956 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
960 unsigned NumOperands = Gep->getNumOperands();
961 Value *LastIndex = Gep->getOperand(NumOperands - 1);
963 Value *GpPtr = Gep->getPointerOperand();
964 // If this GEP value is a consecutive pointer induction variable and all of
965 // the indices are constant then we know it is consecutive. We can
966 Phi = dyn_cast<PHINode>(GpPtr);
967 if (Phi && Inductions.count(Phi)) {
969 // Make sure that the pointer does not point to structs.
970 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
971 if (GepPtrType->getElementType()->isAggregateType())
974 // Make sure that all of the index operands are loop invariant.
975 for (unsigned i = 1; i < NumOperands; ++i)
976 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
979 InductionInfo II = Inductions[Phi];
980 if (IK_PtrInduction == II.IK)
982 else if (IK_ReversePtrInduction == II.IK)
986 // Check that all of the gep indices are uniform except for the last.
987 for (unsigned i = 0; i < NumOperands - 1; ++i)
988 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
991 // We can emit wide load/stores only if the last index is the induction
993 const SCEV *Last = SE->getSCEV(LastIndex);
994 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
995 const SCEV *Step = AR->getStepRecurrence(*SE);
997 // The memory is consecutive because the last index is consecutive
998 // and all other indices are loop invariant.
1001 if (Step->isAllOnesValue())
1008 bool LoopVectorizationLegality::isUniform(Value *V) {
1009 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1012 InnerLoopVectorizer::VectorParts&
1013 InnerLoopVectorizer::getVectorValue(Value *V) {
1014 assert(V != Induction && "The new induction variable should not be used.");
1015 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1017 // If we have this scalar in the map, return it.
1018 if (WidenMap.has(V))
1019 return WidenMap.get(V);
1021 // If this scalar is unknown, assume that it is a constant or that it is
1022 // loop invariant. Broadcast V and save the value for future uses.
1023 Value *B = getBroadcastInstrs(V);
1024 return WidenMap.splat(V, B);
1027 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1028 assert(Vec->getType()->isVectorTy() && "Invalid type");
1029 SmallVector<Constant*, 8> ShuffleMask;
1030 for (unsigned i = 0; i < VF; ++i)
1031 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1033 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1034 ConstantVector::get(ShuffleMask),
1039 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1040 LoopVectorizationLegality *Legal) {
1041 // Attempt to issue a wide load.
1042 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1043 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1045 assert((LI || SI) && "Invalid Load/Store instruction");
1047 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1048 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1049 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1050 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1052 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1053 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1055 if (ScalarAllocatedSize != VectorElementSize)
1056 return scalarizeInstruction(Instr);
1058 // If the pointer is loop invariant or if it is non consecutive,
1059 // scalarize the load.
1060 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1061 bool Reverse = ConsecutiveStride < 0;
1062 bool UniformLoad = LI && Legal->isUniform(Ptr);
1063 if (!ConsecutiveStride || UniformLoad)
1064 return scalarizeInstruction(Instr);
1066 Constant *Zero = Builder.getInt32(0);
1067 VectorParts &Entry = WidenMap.get(Instr);
1069 // Handle consecutive loads/stores.
1070 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1071 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1072 Value *PtrOperand = Gep->getPointerOperand();
1073 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1074 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1076 // Create the new GEP with the new induction variable.
1077 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1078 Gep2->setOperand(0, FirstBasePtr);
1079 Gep2->setName("gep.indvar.base");
1080 Ptr = Builder.Insert(Gep2);
1082 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1083 OrigLoop) && "Base ptr must be invariant");
1085 // The last index does not have to be the induction. It can be
1086 // consecutive and be a function of the index. For example A[I+1];
1087 unsigned NumOperands = Gep->getNumOperands();
1089 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1090 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1091 Value *LastIndex = GEPParts[0];
1092 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1094 // Create the new GEP with the new induction variable.
1095 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1096 Gep2->setOperand(NumOperands - 1, LastIndex);
1097 Gep2->setName("gep.indvar.idx");
1098 Ptr = Builder.Insert(Gep2);
1100 // Use the induction element ptr.
1101 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1102 VectorParts &PtrVal = getVectorValue(Ptr);
1103 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1108 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1109 "We do not allow storing to uniform addresses");
1111 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1112 for (unsigned Part = 0; Part < UF; ++Part) {
1113 // Calculate the pointer for the specific unroll-part.
1114 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1117 // If we store to reverse consecutive memory locations then we need
1118 // to reverse the order of elements in the stored value.
1119 StoredVal[Part] = reverseVector(StoredVal[Part]);
1120 // If the address is consecutive but reversed, then the
1121 // wide store needs to start at the last vector element.
1122 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1123 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1126 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1127 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1131 for (unsigned Part = 0; Part < UF; ++Part) {
1132 // Calculate the pointer for the specific unroll-part.
1133 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1136 // If the address is consecutive but reversed, then the
1137 // wide store needs to start at the last vector element.
1138 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1139 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1142 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1143 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1144 cast<LoadInst>(LI)->setAlignment(Alignment);
1145 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1149 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1150 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1151 // Holds vector parameters or scalars, in case of uniform vals.
1152 SmallVector<VectorParts, 4> Params;
1154 // Find all of the vectorized parameters.
1155 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1156 Value *SrcOp = Instr->getOperand(op);
1158 // If we are accessing the old induction variable, use the new one.
1159 if (SrcOp == OldInduction) {
1160 Params.push_back(getVectorValue(SrcOp));
1164 // Try using previously calculated values.
1165 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1167 // If the src is an instruction that appeared earlier in the basic block
1168 // then it should already be vectorized.
1169 if (SrcInst && OrigLoop->contains(SrcInst)) {
1170 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1171 // The parameter is a vector value from earlier.
1172 Params.push_back(WidenMap.get(SrcInst));
1174 // The parameter is a scalar from outside the loop. Maybe even a constant.
1175 VectorParts Scalars;
1176 Scalars.append(UF, SrcOp);
1177 Params.push_back(Scalars);
1181 assert(Params.size() == Instr->getNumOperands() &&
1182 "Invalid number of operands");
1184 // Does this instruction return a value ?
1185 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1187 Value *UndefVec = IsVoidRetTy ? 0 :
1188 UndefValue::get(VectorType::get(Instr->getType(), VF));
1189 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1190 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1192 // For each vector unroll 'part':
1193 for (unsigned Part = 0; Part < UF; ++Part) {
1194 // For each scalar that we create:
1195 for (unsigned Width = 0; Width < VF; ++Width) {
1196 Instruction *Cloned = Instr->clone();
1198 Cloned->setName(Instr->getName() + ".cloned");
1199 // Replace the operands of the cloned instrucions with extracted scalars.
1200 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1201 Value *Op = Params[op][Part];
1202 // Param is a vector. Need to extract the right lane.
1203 if (Op->getType()->isVectorTy())
1204 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1205 Cloned->setOperand(op, Op);
1208 // Place the cloned scalar in the new loop.
1209 Builder.Insert(Cloned);
1211 // If the original scalar returns a value we need to place it in a vector
1212 // so that future users will be able to use it.
1214 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1215 Builder.getInt32(Width));
1221 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1223 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1224 Legal->getRuntimePointerCheck();
1226 if (!PtrRtCheck->Need)
1229 Instruction *MemoryRuntimeCheck = 0;
1230 unsigned NumPointers = PtrRtCheck->Pointers.size();
1231 SmallVector<Value* , 2> Starts;
1232 SmallVector<Value* , 2> Ends;
1234 SCEVExpander Exp(*SE, "induction");
1236 // Use this type for pointer arithmetic.
1237 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1239 for (unsigned i = 0; i < NumPointers; ++i) {
1240 Value *Ptr = PtrRtCheck->Pointers[i];
1241 const SCEV *Sc = SE->getSCEV(Ptr);
1243 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1244 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1246 Starts.push_back(Ptr);
1247 Ends.push_back(Ptr);
1249 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1251 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1252 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1253 Starts.push_back(Start);
1254 Ends.push_back(End);
1258 IRBuilder<> ChkBuilder(Loc);
1260 for (unsigned i = 0; i < NumPointers; ++i) {
1261 for (unsigned j = i+1; j < NumPointers; ++j) {
1262 // No need to check if two readonly pointers intersect.
1263 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1266 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1267 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1268 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1269 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1271 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1272 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1273 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1274 if (MemoryRuntimeCheck)
1275 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1278 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1282 return MemoryRuntimeCheck;
1286 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1288 In this function we generate a new loop. The new loop will contain
1289 the vectorized instructions while the old loop will continue to run the
1292 [ ] <-- vector loop bypass (may consist of multiple blocks).
1295 | [ ] <-- vector pre header.
1299 | [ ]_| <-- vector loop.
1302 >[ ] <--- middle-block.
1305 | [ ] <--- new preheader.
1309 | [ ]_| <-- old scalar loop to handle remainder.
1312 >[ ] <-- exit block.
1316 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1317 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1318 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1319 assert(ExitBlock && "Must have an exit block");
1321 // Mark the old scalar loop with metadata that tells us not to vectorize this
1322 // loop again if we run into it.
1323 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None);
1324 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD);
1326 // Some loops have a single integer induction variable, while other loops
1327 // don't. One example is c++ iterators that often have multiple pointer
1328 // induction variables. In the code below we also support a case where we
1329 // don't have a single induction variable.
1330 OldInduction = Legal->getInduction();
1331 Type *IdxTy = Legal->getWidestInductionType();
1333 // Find the loop boundaries.
1334 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1335 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1337 // Get the total trip count from the count by adding 1.
1338 ExitCount = SE->getAddExpr(ExitCount,
1339 SE->getConstant(ExitCount->getType(), 1));
1341 // Expand the trip count and place the new instructions in the preheader.
1342 // Notice that the pre-header does not change, only the loop body.
1343 SCEVExpander Exp(*SE, "induction");
1345 // Count holds the overall loop count (N).
1346 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1347 BypassBlock->getTerminator());
1349 // The loop index does not have to start at Zero. Find the original start
1350 // value from the induction PHI node. If we don't have an induction variable
1351 // then we know that it starts at zero.
1352 Builder.SetInsertPoint(BypassBlock->getTerminator());
1353 Value *StartIdx = ExtendedIdx = OldInduction ?
1354 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1356 ConstantInt::get(IdxTy, 0);
1358 assert(BypassBlock && "Invalid loop structure");
1359 LoopBypassBlocks.push_back(BypassBlock);
1361 // Split the single block loop into the two loop structure described above.
1362 BasicBlock *VectorPH =
1363 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1364 BasicBlock *VecBody =
1365 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1366 BasicBlock *MiddleBlock =
1367 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1368 BasicBlock *ScalarPH =
1369 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1371 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1373 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1375 // Generate the induction variable.
1376 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1377 // The loop step is equal to the vectorization factor (num of SIMD elements)
1378 // times the unroll factor (num of SIMD instructions).
1379 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1381 // This is the IR builder that we use to add all of the logic for bypassing
1382 // the new vector loop.
1383 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1385 // We may need to extend the index in case there is a type mismatch.
1386 // We know that the count starts at zero and does not overflow.
1387 if (Count->getType() != IdxTy) {
1388 // The exit count can be of pointer type. Convert it to the correct
1390 if (ExitCount->getType()->isPointerTy())
1391 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1393 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1396 // Add the start index to the loop count to get the new end index.
1397 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1399 // Now we need to generate the expression for N - (N % VF), which is
1400 // the part that the vectorized body will execute.
1401 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1402 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1403 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1404 "end.idx.rnd.down");
1406 // Now, compare the new count to zero. If it is zero skip the vector loop and
1407 // jump to the scalar loop.
1408 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1411 BasicBlock *LastBypassBlock = BypassBlock;
1413 // Generate the code that checks in runtime if arrays overlap. We put the
1414 // checks into a separate block to make the more common case of few elements
1416 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1417 BypassBlock->getTerminator());
1418 if (MemRuntimeCheck) {
1419 // Create a new block containing the memory check.
1420 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1422 LoopBypassBlocks.push_back(CheckBlock);
1424 // Replace the branch into the memory check block with a conditional branch
1425 // for the "few elements case".
1426 Instruction *OldTerm = BypassBlock->getTerminator();
1427 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1428 OldTerm->eraseFromParent();
1430 Cmp = MemRuntimeCheck;
1431 LastBypassBlock = CheckBlock;
1434 LastBypassBlock->getTerminator()->eraseFromParent();
1435 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1438 // We are going to resume the execution of the scalar loop.
1439 // Go over all of the induction variables that we found and fix the
1440 // PHIs that are left in the scalar version of the loop.
1441 // The starting values of PHI nodes depend on the counter of the last
1442 // iteration in the vectorized loop.
1443 // If we come from a bypass edge then we need to start from the original
1446 // This variable saves the new starting index for the scalar loop.
1447 PHINode *ResumeIndex = 0;
1448 LoopVectorizationLegality::InductionList::iterator I, E;
1449 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1450 // Set builder to point to last bypass block.
1451 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1452 for (I = List->begin(), E = List->end(); I != E; ++I) {
1453 PHINode *OrigPhi = I->first;
1454 LoopVectorizationLegality::InductionInfo II = I->second;
1456 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1457 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1458 MiddleBlock->getTerminator());
1459 // We might have extended the type of the induction variable but we need a
1460 // truncated version for the scalar loop.
1461 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1462 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1463 MiddleBlock->getTerminator()) : 0;
1465 Value *EndValue = 0;
1467 case LoopVectorizationLegality::IK_NoInduction:
1468 llvm_unreachable("Unknown induction");
1469 case LoopVectorizationLegality::IK_IntInduction: {
1470 // Handle the integer induction counter.
1471 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1473 // We have the canonical induction variable.
1474 if (OrigPhi == OldInduction) {
1475 // Create a truncated version of the resume value for the scalar loop,
1476 // we might have promoted the type to a larger width.
1478 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1479 // The new PHI merges the original incoming value, in case of a bypass,
1480 // or the value at the end of the vectorized loop.
1481 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1482 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1483 TruncResumeVal->addIncoming(EndValue, VecBody);
1485 // We know what the end value is.
1486 EndValue = IdxEndRoundDown;
1487 // We also know which PHI node holds it.
1488 ResumeIndex = ResumeVal;
1492 // Not the canonical induction variable - add the vector loop count to the
1494 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1495 II.StartValue->getType(),
1497 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1500 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1501 // Convert the CountRoundDown variable to the PHI size.
1502 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1503 II.StartValue->getType(),
1505 // Handle reverse integer induction counter.
1506 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1509 case LoopVectorizationLegality::IK_PtrInduction: {
1510 // For pointer induction variables, calculate the offset using
1512 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1516 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1517 // The value at the end of the loop for the reverse pointer is calculated
1518 // by creating a GEP with a negative index starting from the start value.
1519 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1520 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1522 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1528 // The new PHI merges the original incoming value, in case of a bypass,
1529 // or the value at the end of the vectorized loop.
1530 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1531 if (OrigPhi == OldInduction)
1532 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1534 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1536 ResumeVal->addIncoming(EndValue, VecBody);
1538 // Fix the scalar body counter (PHI node).
1539 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1540 // The old inductions phi node in the scalar body needs the truncated value.
1541 if (OrigPhi == OldInduction)
1542 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1544 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1547 // If we are generating a new induction variable then we also need to
1548 // generate the code that calculates the exit value. This value is not
1549 // simply the end of the counter because we may skip the vectorized body
1550 // in case of a runtime check.
1552 assert(!ResumeIndex && "Unexpected resume value found");
1553 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1554 MiddleBlock->getTerminator());
1555 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1556 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1557 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1560 // Make sure that we found the index where scalar loop needs to continue.
1561 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1562 "Invalid resume Index");
1564 // Add a check in the middle block to see if we have completed
1565 // all of the iterations in the first vector loop.
1566 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1567 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1568 ResumeIndex, "cmp.n",
1569 MiddleBlock->getTerminator());
1571 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1572 // Remove the old terminator.
1573 MiddleBlock->getTerminator()->eraseFromParent();
1575 // Create i+1 and fill the PHINode.
1576 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1577 Induction->addIncoming(StartIdx, VectorPH);
1578 Induction->addIncoming(NextIdx, VecBody);
1579 // Create the compare.
1580 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1581 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1583 // Now we have two terminators. Remove the old one from the block.
1584 VecBody->getTerminator()->eraseFromParent();
1586 // Get ready to start creating new instructions into the vectorized body.
1587 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1589 // Create and register the new vector loop.
1590 Loop* Lp = new Loop();
1591 Loop *ParentLoop = OrigLoop->getParentLoop();
1593 // Insert the new loop into the loop nest and register the new basic blocks.
1595 ParentLoop->addChildLoop(Lp);
1596 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1597 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1598 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1599 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1600 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1602 LI->addTopLevelLoop(Lp);
1605 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1608 LoopVectorPreHeader = VectorPH;
1609 LoopScalarPreHeader = ScalarPH;
1610 LoopMiddleBlock = MiddleBlock;
1611 LoopExitBlock = ExitBlock;
1612 LoopVectorBody = VecBody;
1613 LoopScalarBody = OldBasicBlock;
1616 /// This function returns the identity element (or neutral element) for
1617 /// the operation K.
1619 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1624 // Adding, Xoring, Oring zero to a number does not change it.
1625 return ConstantInt::get(Tp, 0);
1626 case RK_IntegerMult:
1627 // Multiplying a number by 1 does not change it.
1628 return ConstantInt::get(Tp, 1);
1630 // AND-ing a number with an all-1 value does not change it.
1631 return ConstantInt::get(Tp, -1, true);
1633 // Multiplying a number by 1 does not change it.
1634 return ConstantFP::get(Tp, 1.0L);
1636 // Adding zero to a number does not change it.
1637 return ConstantFP::get(Tp, 0.0L);
1639 llvm_unreachable("Unknown reduction kind");
1643 static Intrinsic::ID
1644 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1645 // If we have an intrinsic call, check if it is trivially vectorizable.
1646 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1647 switch (II->getIntrinsicID()) {
1648 case Intrinsic::sqrt:
1649 case Intrinsic::sin:
1650 case Intrinsic::cos:
1651 case Intrinsic::exp:
1652 case Intrinsic::exp2:
1653 case Intrinsic::log:
1654 case Intrinsic::log10:
1655 case Intrinsic::log2:
1656 case Intrinsic::fabs:
1657 case Intrinsic::floor:
1658 case Intrinsic::ceil:
1659 case Intrinsic::trunc:
1660 case Intrinsic::rint:
1661 case Intrinsic::nearbyint:
1662 case Intrinsic::pow:
1663 case Intrinsic::fma:
1664 case Intrinsic::fmuladd:
1665 return II->getIntrinsicID();
1667 return Intrinsic::not_intrinsic;
1672 return Intrinsic::not_intrinsic;
1675 Function *F = CI->getCalledFunction();
1676 // We're going to make assumptions on the semantics of the functions, check
1677 // that the target knows that it's available in this environment.
1678 if (!F || !TLI->getLibFunc(F->getName(), Func))
1679 return Intrinsic::not_intrinsic;
1681 // Otherwise check if we have a call to a function that can be turned into a
1682 // vector intrinsic.
1689 return Intrinsic::sin;
1693 return Intrinsic::cos;
1697 return Intrinsic::exp;
1699 case LibFunc::exp2f:
1700 case LibFunc::exp2l:
1701 return Intrinsic::exp2;
1705 return Intrinsic::log;
1706 case LibFunc::log10:
1707 case LibFunc::log10f:
1708 case LibFunc::log10l:
1709 return Intrinsic::log10;
1711 case LibFunc::log2f:
1712 case LibFunc::log2l:
1713 return Intrinsic::log2;
1715 case LibFunc::fabsf:
1716 case LibFunc::fabsl:
1717 return Intrinsic::fabs;
1718 case LibFunc::floor:
1719 case LibFunc::floorf:
1720 case LibFunc::floorl:
1721 return Intrinsic::floor;
1723 case LibFunc::ceilf:
1724 case LibFunc::ceill:
1725 return Intrinsic::ceil;
1726 case LibFunc::trunc:
1727 case LibFunc::truncf:
1728 case LibFunc::truncl:
1729 return Intrinsic::trunc;
1731 case LibFunc::rintf:
1732 case LibFunc::rintl:
1733 return Intrinsic::rint;
1734 case LibFunc::nearbyint:
1735 case LibFunc::nearbyintf:
1736 case LibFunc::nearbyintl:
1737 return Intrinsic::nearbyint;
1741 return Intrinsic::pow;
1744 return Intrinsic::not_intrinsic;
1747 /// This function translates the reduction kind to an LLVM binary operator.
1749 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1751 case LoopVectorizationLegality::RK_IntegerAdd:
1752 return Instruction::Add;
1753 case LoopVectorizationLegality::RK_IntegerMult:
1754 return Instruction::Mul;
1755 case LoopVectorizationLegality::RK_IntegerOr:
1756 return Instruction::Or;
1757 case LoopVectorizationLegality::RK_IntegerAnd:
1758 return Instruction::And;
1759 case LoopVectorizationLegality::RK_IntegerXor:
1760 return Instruction::Xor;
1761 case LoopVectorizationLegality::RK_FloatMult:
1762 return Instruction::FMul;
1763 case LoopVectorizationLegality::RK_FloatAdd:
1764 return Instruction::FAdd;
1765 case LoopVectorizationLegality::RK_IntegerMinMax:
1766 return Instruction::ICmp;
1767 case LoopVectorizationLegality::RK_FloatMinMax:
1768 return Instruction::FCmp;
1770 llvm_unreachable("Unknown reduction operation");
1774 Value *createMinMaxOp(IRBuilder<> &Builder,
1775 LoopVectorizationLegality::MinMaxReductionKind RK,
1778 CmpInst::Predicate P = CmpInst::ICMP_NE;
1781 llvm_unreachable("Unknown min/max reduction kind");
1782 case LoopVectorizationLegality::MRK_UIntMin:
1783 P = CmpInst::ICMP_ULT;
1785 case LoopVectorizationLegality::MRK_UIntMax:
1786 P = CmpInst::ICMP_UGT;
1788 case LoopVectorizationLegality::MRK_SIntMin:
1789 P = CmpInst::ICMP_SLT;
1791 case LoopVectorizationLegality::MRK_SIntMax:
1792 P = CmpInst::ICMP_SGT;
1794 case LoopVectorizationLegality::MRK_FloatMin:
1795 P = CmpInst::FCMP_OLT;
1797 case LoopVectorizationLegality::MRK_FloatMax:
1798 P = CmpInst::FCMP_OGT;
1803 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1804 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1806 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1808 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1813 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1814 //===------------------------------------------------===//
1816 // Notice: any optimization or new instruction that go
1817 // into the code below should be also be implemented in
1820 //===------------------------------------------------===//
1821 Constant *Zero = Builder.getInt32(0);
1823 // In order to support reduction variables we need to be able to vectorize
1824 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1825 // stages. First, we create a new vector PHI node with no incoming edges.
1826 // We use this value when we vectorize all of the instructions that use the
1827 // PHI. Next, after all of the instructions in the block are complete we
1828 // add the new incoming edges to the PHI. At this point all of the
1829 // instructions in the basic block are vectorized, so we can use them to
1830 // construct the PHI.
1831 PhiVector RdxPHIsToFix;
1833 // Scan the loop in a topological order to ensure that defs are vectorized
1835 LoopBlocksDFS DFS(OrigLoop);
1838 // Vectorize all of the blocks in the original loop.
1839 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1840 be = DFS.endRPO(); bb != be; ++bb)
1841 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1843 // At this point every instruction in the original loop is widened to
1844 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1845 // that we vectorized. The PHI nodes are currently empty because we did
1846 // not want to introduce cycles. Notice that the remaining PHI nodes
1847 // that we need to fix are reduction variables.
1849 // Create the 'reduced' values for each of the induction vars.
1850 // The reduced values are the vector values that we scalarize and combine
1851 // after the loop is finished.
1852 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1854 PHINode *RdxPhi = *it;
1855 assert(RdxPhi && "Unable to recover vectorized PHI");
1857 // Find the reduction variable descriptor.
1858 assert(Legal->getReductionVars()->count(RdxPhi) &&
1859 "Unable to find the reduction variable");
1860 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1861 (*Legal->getReductionVars())[RdxPhi];
1863 // We need to generate a reduction vector from the incoming scalar.
1864 // To do so, we need to generate the 'identity' vector and overide
1865 // one of the elements with the incoming scalar reduction. We need
1866 // to do it in the vector-loop preheader.
1867 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1869 // This is the vector-clone of the value that leaves the loop.
1870 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1871 Type *VecTy = VectorExit[0]->getType();
1873 // Find the reduction identity variable. Zero for addition, or, xor,
1874 // one for multiplication, -1 for And.
1877 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
1878 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
1879 // MinMax reduction have the start value as their identify.
1880 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
1884 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1885 VecTy->getScalarType());
1886 Identity = ConstantVector::getSplat(VF, Iden);
1888 // This vector is the Identity vector where the first element is the
1889 // incoming scalar reduction.
1890 VectorStart = Builder.CreateInsertElement(Identity,
1891 RdxDesc.StartValue, Zero);
1894 // Fix the vector-loop phi.
1895 // We created the induction variable so we know that the
1896 // preheader is the first entry.
1897 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1899 // Reductions do not have to start at zero. They can start with
1900 // any loop invariant values.
1901 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1902 BasicBlock *Latch = OrigLoop->getLoopLatch();
1903 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1904 VectorParts &Val = getVectorValue(LoopVal);
1905 for (unsigned part = 0; part < UF; ++part) {
1906 // Make sure to add the reduction stat value only to the
1907 // first unroll part.
1908 Value *StartVal = (part == 0) ? VectorStart : Identity;
1909 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1910 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1913 // Before each round, move the insertion point right between
1914 // the PHIs and the values we are going to write.
1915 // This allows us to write both PHINodes and the extractelement
1917 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1919 VectorParts RdxParts;
1920 for (unsigned part = 0; part < UF; ++part) {
1921 // This PHINode contains the vectorized reduction variable, or
1922 // the initial value vector, if we bypass the vector loop.
1923 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1924 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1925 Value *StartVal = (part == 0) ? VectorStart : Identity;
1926 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1927 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1928 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1929 RdxParts.push_back(NewPhi);
1932 // Reduce all of the unrolled parts into a single vector.
1933 Value *ReducedPartRdx = RdxParts[0];
1934 unsigned Op = getReductionBinOp(RdxDesc.Kind);
1935 for (unsigned part = 1; part < UF; ++part) {
1936 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1937 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
1938 RdxParts[part], ReducedPartRdx,
1941 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
1942 ReducedPartRdx, RdxParts[part]);
1945 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1946 // and vector ops, reducing the set of values being computed by half each
1948 assert(isPowerOf2_32(VF) &&
1949 "Reduction emission only supported for pow2 vectors!");
1950 Value *TmpVec = ReducedPartRdx;
1951 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1952 for (unsigned i = VF; i != 1; i >>= 1) {
1953 // Move the upper half of the vector to the lower half.
1954 for (unsigned j = 0; j != i/2; ++j)
1955 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1957 // Fill the rest of the mask with undef.
1958 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1959 UndefValue::get(Builder.getInt32Ty()));
1962 Builder.CreateShuffleVector(TmpVec,
1963 UndefValue::get(TmpVec->getType()),
1964 ConstantVector::get(ShuffleMask),
1967 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
1968 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1971 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
1974 // The result is in the first element of the vector.
1975 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1977 // Now, we need to fix the users of the reduction variable
1978 // inside and outside of the scalar remainder loop.
1979 // We know that the loop is in LCSSA form. We need to update the
1980 // PHI nodes in the exit blocks.
1981 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1982 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1983 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1984 if (!LCSSAPhi) continue;
1986 // All PHINodes need to have a single entry edge, or two if
1987 // we already fixed them.
1988 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1990 // We found our reduction value exit-PHI. Update it with the
1991 // incoming bypass edge.
1992 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1993 // Add an edge coming from the bypass.
1994 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1997 }// end of the LCSSA phi scan.
1999 // Fix the scalar loop reduction variable with the incoming reduction sum
2000 // from the vector body and from the backedge value.
2001 int IncomingEdgeBlockIdx =
2002 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2003 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2004 // Pick the other block.
2005 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2006 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2007 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2008 }// end of for each redux variable.
2010 // The Loop exit block may have single value PHI nodes where the incoming
2011 // value is 'undef'. While vectorizing we only handled real values that
2012 // were defined inside the loop. Here we handle the 'undef case'.
2014 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2015 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2016 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2017 if (!LCSSAPhi) continue;
2018 if (LCSSAPhi->getNumIncomingValues() == 1)
2019 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2024 InnerLoopVectorizer::VectorParts
2025 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2026 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2029 VectorParts SrcMask = createBlockInMask(Src);
2031 // The terminator has to be a branch inst!
2032 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2033 assert(BI && "Unexpected terminator found");
2035 if (BI->isConditional()) {
2036 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2038 if (BI->getSuccessor(0) != Dst)
2039 for (unsigned part = 0; part < UF; ++part)
2040 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2042 for (unsigned part = 0; part < UF; ++part)
2043 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2050 InnerLoopVectorizer::VectorParts
2051 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2052 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2054 // Loop incoming mask is all-one.
2055 if (OrigLoop->getHeader() == BB) {
2056 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2057 return getVectorValue(C);
2060 // This is the block mask. We OR all incoming edges, and with zero.
2061 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2062 VectorParts BlockMask = getVectorValue(Zero);
2065 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2066 VectorParts EM = createEdgeMask(*it, BB);
2067 for (unsigned part = 0; part < UF; ++part)
2068 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2075 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2076 BasicBlock *BB, PhiVector *PV) {
2077 // For each instruction in the old loop.
2078 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2079 VectorParts &Entry = WidenMap.get(it);
2080 switch (it->getOpcode()) {
2081 case Instruction::Br:
2082 // Nothing to do for PHIs and BR, since we already took care of the
2083 // loop control flow instructions.
2085 case Instruction::PHI:{
2086 PHINode* P = cast<PHINode>(it);
2087 // Handle reduction variables:
2088 if (Legal->getReductionVars()->count(P)) {
2089 for (unsigned part = 0; part < UF; ++part) {
2090 // This is phase one of vectorizing PHIs.
2091 Type *VecTy = VectorType::get(it->getType(), VF);
2092 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2093 LoopVectorBody-> getFirstInsertionPt());
2099 // Check for PHI nodes that are lowered to vector selects.
2100 if (P->getParent() != OrigLoop->getHeader()) {
2101 // We know that all PHIs in non header blocks are converted into
2102 // selects, so we don't have to worry about the insertion order and we
2103 // can just use the builder.
2104 // At this point we generate the predication tree. There may be
2105 // duplications since this is a simple recursive scan, but future
2106 // optimizations will clean it up.
2108 unsigned NumIncoming = P->getNumIncomingValues();
2110 // Generate a sequence of selects of the form:
2111 // SELECT(Mask3, In3,
2112 // SELECT(Mask2, In2,
2114 for (unsigned In = 0; In < NumIncoming; In++) {
2115 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2117 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2119 for (unsigned part = 0; part < UF; ++part) {
2120 // We might have single edge PHIs (blocks) - use an identity
2121 // 'select' for the first PHI operand.
2123 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2126 // Select between the current value and the previous incoming edge
2127 // based on the incoming mask.
2128 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2129 Entry[part], "predphi");
2135 // This PHINode must be an induction variable.
2136 // Make sure that we know about it.
2137 assert(Legal->getInductionVars()->count(P) &&
2138 "Not an induction variable");
2140 LoopVectorizationLegality::InductionInfo II =
2141 Legal->getInductionVars()->lookup(P);
2144 case LoopVectorizationLegality::IK_NoInduction:
2145 llvm_unreachable("Unknown induction");
2146 case LoopVectorizationLegality::IK_IntInduction: {
2147 assert(P->getType() == II.StartValue->getType() && "Types must match");
2148 Type *PhiTy = P->getType();
2150 if (P == OldInduction) {
2151 // Handle the canonical induction variable. We might have had to
2153 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2155 // Handle other induction variables that are now based on the
2157 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2159 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2160 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2163 Broadcasted = getBroadcastInstrs(Broadcasted);
2164 // After broadcasting the induction variable we need to make the vector
2165 // consecutive by adding 0, 1, 2, etc.
2166 for (unsigned part = 0; part < UF; ++part)
2167 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2170 case LoopVectorizationLegality::IK_ReverseIntInduction:
2171 case LoopVectorizationLegality::IK_PtrInduction:
2172 case LoopVectorizationLegality::IK_ReversePtrInduction:
2173 // Handle reverse integer and pointer inductions.
2174 Value *StartIdx = ExtendedIdx;
2175 // This is the normalized GEP that starts counting at zero.
2176 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2179 // Handle the reverse integer induction variable case.
2180 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2181 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2182 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2184 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2187 // This is a new value so do not hoist it out.
2188 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2189 // After broadcasting the induction variable we need to make the
2190 // vector consecutive by adding ... -3, -2, -1, 0.
2191 for (unsigned part = 0; part < UF; ++part)
2192 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2197 // Handle the pointer induction variable case.
2198 assert(P->getType()->isPointerTy() && "Unexpected type.");
2200 // Is this a reverse induction ptr or a consecutive induction ptr.
2201 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2204 // This is the vector of results. Notice that we don't generate
2205 // vector geps because scalar geps result in better code.
2206 for (unsigned part = 0; part < UF; ++part) {
2207 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2208 for (unsigned int i = 0; i < VF; ++i) {
2209 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2210 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2213 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2215 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2217 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2219 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2220 Builder.getInt32(i),
2223 Entry[part] = VecVal;
2230 case Instruction::Add:
2231 case Instruction::FAdd:
2232 case Instruction::Sub:
2233 case Instruction::FSub:
2234 case Instruction::Mul:
2235 case Instruction::FMul:
2236 case Instruction::UDiv:
2237 case Instruction::SDiv:
2238 case Instruction::FDiv:
2239 case Instruction::URem:
2240 case Instruction::SRem:
2241 case Instruction::FRem:
2242 case Instruction::Shl:
2243 case Instruction::LShr:
2244 case Instruction::AShr:
2245 case Instruction::And:
2246 case Instruction::Or:
2247 case Instruction::Xor: {
2248 // Just widen binops.
2249 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2250 VectorParts &A = getVectorValue(it->getOperand(0));
2251 VectorParts &B = getVectorValue(it->getOperand(1));
2253 // Use this vector value for all users of the original instruction.
2254 for (unsigned Part = 0; Part < UF; ++Part) {
2255 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2257 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2258 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2259 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2260 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2261 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2263 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2264 VecOp->setIsExact(BinOp->isExact());
2270 case Instruction::Select: {
2272 // If the selector is loop invariant we can create a select
2273 // instruction with a scalar condition. Otherwise, use vector-select.
2274 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2277 // The condition can be loop invariant but still defined inside the
2278 // loop. This means that we can't just use the original 'cond' value.
2279 // We have to take the 'vectorized' value and pick the first lane.
2280 // Instcombine will make this a no-op.
2281 VectorParts &Cond = getVectorValue(it->getOperand(0));
2282 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2283 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2284 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2285 Builder.getInt32(0));
2286 for (unsigned Part = 0; Part < UF; ++Part) {
2287 Entry[Part] = Builder.CreateSelect(
2288 InvariantCond ? ScalarCond : Cond[Part],
2295 case Instruction::ICmp:
2296 case Instruction::FCmp: {
2297 // Widen compares. Generate vector compares.
2298 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2299 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2300 VectorParts &A = getVectorValue(it->getOperand(0));
2301 VectorParts &B = getVectorValue(it->getOperand(1));
2302 for (unsigned Part = 0; Part < UF; ++Part) {
2305 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2307 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2313 case Instruction::Store:
2314 case Instruction::Load:
2315 vectorizeMemoryInstruction(it, Legal);
2317 case Instruction::ZExt:
2318 case Instruction::SExt:
2319 case Instruction::FPToUI:
2320 case Instruction::FPToSI:
2321 case Instruction::FPExt:
2322 case Instruction::PtrToInt:
2323 case Instruction::IntToPtr:
2324 case Instruction::SIToFP:
2325 case Instruction::UIToFP:
2326 case Instruction::Trunc:
2327 case Instruction::FPTrunc:
2328 case Instruction::BitCast: {
2329 CastInst *CI = dyn_cast<CastInst>(it);
2330 /// Optimize the special case where the source is the induction
2331 /// variable. Notice that we can only optimize the 'trunc' case
2332 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2333 /// c. other casts depend on pointer size.
2334 if (CI->getOperand(0) == OldInduction &&
2335 it->getOpcode() == Instruction::Trunc) {
2336 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2338 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2339 for (unsigned Part = 0; Part < UF; ++Part)
2340 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2343 /// Vectorize casts.
2344 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2346 VectorParts &A = getVectorValue(it->getOperand(0));
2347 for (unsigned Part = 0; Part < UF; ++Part)
2348 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2352 case Instruction::Call: {
2353 // Ignore dbg intrinsics.
2354 if (isa<DbgInfoIntrinsic>(it))
2357 Module *M = BB->getParent()->getParent();
2358 CallInst *CI = cast<CallInst>(it);
2359 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2360 assert(ID && "Not an intrinsic call!");
2361 for (unsigned Part = 0; Part < UF; ++Part) {
2362 SmallVector<Value*, 4> Args;
2363 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2364 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2365 Args.push_back(Arg[Part]);
2367 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2368 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2369 Entry[Part] = Builder.CreateCall(F, Args);
2375 // All other instructions are unsupported. Scalarize them.
2376 scalarizeInstruction(it);
2379 }// end of for_each instr.
2382 void InnerLoopVectorizer::updateAnalysis() {
2383 // Forget the original basic block.
2384 SE->forgetLoop(OrigLoop);
2386 // Update the dominator tree information.
2387 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2388 "Entry does not dominate exit.");
2390 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2391 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2392 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2393 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2394 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2395 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2396 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2397 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2399 DEBUG(DT->verifyAnalysis());
2402 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2403 if (!EnableIfConversion)
2406 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2407 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2409 // Collect the blocks that need predication.
2410 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2411 BasicBlock *BB = LoopBlocks[i];
2413 // We don't support switch statements inside loops.
2414 if (!isa<BranchInst>(BB->getTerminator()))
2417 // We must be able to predicate all blocks that need to be predicated.
2418 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2422 // Check that we can actually speculate the hoistable loads.
2423 if (!LoadSpeculation.canHoistAllLoads())
2426 // We can if-convert this loop.
2430 bool LoopVectorizationLegality::canVectorize() {
2431 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2432 // be canonicalized.
2433 if (!TheLoop->getLoopPreheader())
2436 // We can only vectorize innermost loops.
2437 if (TheLoop->getSubLoopsVector().size())
2440 // We must have a single backedge.
2441 if (TheLoop->getNumBackEdges() != 1)
2444 // We must have a single exiting block.
2445 if (!TheLoop->getExitingBlock())
2448 unsigned NumBlocks = TheLoop->getNumBlocks();
2450 // Check if we can if-convert non single-bb loops.
2451 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2452 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2456 // We need to have a loop header.
2457 BasicBlock *Latch = TheLoop->getLoopLatch();
2458 DEBUG(dbgs() << "LV: Found a loop: " <<
2459 TheLoop->getHeader()->getName() << "\n");
2461 // ScalarEvolution needs to be able to find the exit count.
2462 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2463 if (ExitCount == SE->getCouldNotCompute()) {
2464 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2468 // Do not loop-vectorize loops with a tiny trip count.
2469 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2470 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2471 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2472 "This loop is not worth vectorizing.\n");
2476 // Check if we can vectorize the instructions and CFG in this loop.
2477 if (!canVectorizeInstrs()) {
2478 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2482 // Go over each instruction and look at memory deps.
2483 if (!canVectorizeMemory()) {
2484 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2488 // Collect all of the variables that remain uniform after vectorization.
2489 collectLoopUniforms();
2491 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2492 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2495 // Okay! We can vectorize. At this point we don't have any other mem analysis
2496 // which may limit our maximum vectorization factor, so just return true with
2501 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2502 if (Ty->isPointerTy())
2503 return DL.getIntPtrType(Ty->getContext());
2507 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2508 Ty0 = convertPointerToIntegerType(DL, Ty0);
2509 Ty1 = convertPointerToIntegerType(DL, Ty1);
2510 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2515 bool LoopVectorizationLegality::canVectorizeInstrs() {
2516 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2517 BasicBlock *Header = TheLoop->getHeader();
2519 // If we marked the scalar loop as "already vectorized" then no need
2520 // to vectorize it again.
2521 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) {
2522 DEBUG(dbgs() << "LV: This loop was vectorized before\n");
2526 // Look for the attribute signaling the absence of NaNs.
2527 Function &F = *Header->getParent();
2528 if (F.hasFnAttribute("no-nans-fp-math"))
2529 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2530 AttributeSet::FunctionIndex,
2531 "no-nans-fp-math").getValueAsString() == "true";
2533 // For each block in the loop.
2534 for (Loop::block_iterator bb = TheLoop->block_begin(),
2535 be = TheLoop->block_end(); bb != be; ++bb) {
2537 // Scan the instructions in the block and look for hazards.
2538 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2541 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2542 Type *PhiTy = Phi->getType();
2543 // Check that this PHI type is allowed.
2544 if (!PhiTy->isIntegerTy() &&
2545 !PhiTy->isFloatingPointTy() &&
2546 !PhiTy->isPointerTy()) {
2547 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2551 // If this PHINode is not in the header block, then we know that we
2552 // can convert it to select during if-conversion. No need to check if
2553 // the PHIs in this block are induction or reduction variables.
2557 // We only allow if-converted PHIs with more than two incoming values.
2558 if (Phi->getNumIncomingValues() != 2) {
2559 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2563 // This is the value coming from the preheader.
2564 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2565 // Check if this is an induction variable.
2566 InductionKind IK = isInductionVariable(Phi);
2568 if (IK_NoInduction != IK) {
2569 // Get the widest type.
2571 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2573 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2575 // Int inductions are special because we only allow one IV.
2576 if (IK == IK_IntInduction) {
2577 // Use the phi node with the widest type as induction. Use the last
2578 // one if there are multiple (no good reason for doing this other
2579 // than it is expedient).
2580 if (!Induction || PhiTy == WidestIndTy)
2584 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2585 Inductions[Phi] = InductionInfo(StartValue, IK);
2589 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2590 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2593 if (AddReductionVar(Phi, RK_IntegerMult)) {
2594 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2597 if (AddReductionVar(Phi, RK_IntegerOr)) {
2598 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2601 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2602 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2605 if (AddReductionVar(Phi, RK_IntegerXor)) {
2606 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2609 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2610 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2613 if (AddReductionVar(Phi, RK_FloatMult)) {
2614 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2617 if (AddReductionVar(Phi, RK_FloatAdd)) {
2618 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2621 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2622 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2626 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2628 }// end of PHI handling
2630 // We still don't handle functions. However, we can ignore dbg intrinsic
2631 // calls and we do handle certain intrinsic and libm functions.
2632 CallInst *CI = dyn_cast<CallInst>(it);
2633 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2634 DEBUG(dbgs() << "LV: Found a call site.\n");
2638 // Check that the instruction return type is vectorizable.
2639 if (!VectorType::isValidElementType(it->getType()) &&
2640 !it->getType()->isVoidTy()) {
2641 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2645 // Check that the stored type is vectorizable.
2646 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2647 Type *T = ST->getValueOperand()->getType();
2648 if (!VectorType::isValidElementType(T))
2652 // Reduction instructions are allowed to have exit users.
2653 // All other instructions must not have external users.
2654 if (!AllowedExit.count(it))
2655 //Check that all of the users of the loop are inside the BB.
2656 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2658 Instruction *U = cast<Instruction>(*I);
2659 // This user may be a reduction exit value.
2660 if (!TheLoop->contains(U)) {
2661 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2670 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2671 if (Inductions.empty())
2678 void LoopVectorizationLegality::collectLoopUniforms() {
2679 // We now know that the loop is vectorizable!
2680 // Collect variables that will remain uniform after vectorization.
2681 std::vector<Value*> Worklist;
2682 BasicBlock *Latch = TheLoop->getLoopLatch();
2684 // Start with the conditional branch and walk up the block.
2685 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2687 while (Worklist.size()) {
2688 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2689 Worklist.pop_back();
2691 // Look at instructions inside this loop.
2692 // Stop when reaching PHI nodes.
2693 // TODO: we need to follow values all over the loop, not only in this block.
2694 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2697 // This is a known uniform.
2700 // Insert all operands.
2701 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2705 AliasAnalysis::Location
2706 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2707 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2708 return AA->getLocation(Store);
2709 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2710 return AA->getLocation(Load);
2712 llvm_unreachable("Should be either load or store instruction");
2716 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2719 AliasMultiMap& WriteObjects,
2720 unsigned MaxByteWidth) {
2722 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2724 std::vector<Instruction*>::iterator
2725 it = WriteObjects[Object].begin(),
2726 end = WriteObjects[Object].end();
2728 for (; it != end; ++it) {
2729 Instruction* I = *it;
2733 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2734 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2735 ThatLoc.getWithNewSize(MaxByteWidth)))
2741 bool LoopVectorizationLegality::canVectorizeMemory() {
2743 typedef SmallVector<Value*, 16> ValueVector;
2744 typedef SmallPtrSet<Value*, 16> ValueSet;
2745 // Holds the Load and Store *instructions*.
2748 PtrRtCheck.Pointers.clear();
2749 PtrRtCheck.Need = false;
2751 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2754 for (Loop::block_iterator bb = TheLoop->block_begin(),
2755 be = TheLoop->block_end(); bb != be; ++bb) {
2757 // Scan the BB and collect legal loads and stores.
2758 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2761 // If this is a load, save it. If this instruction can read from memory
2762 // but is not a load, then we quit. Notice that we don't handle function
2763 // calls that read or write.
2764 if (it->mayReadFromMemory()) {
2765 LoadInst *Ld = dyn_cast<LoadInst>(it);
2766 if (!Ld) return false;
2767 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2768 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2771 Loads.push_back(Ld);
2775 // Save 'store' instructions. Abort if other instructions write to memory.
2776 if (it->mayWriteToMemory()) {
2777 StoreInst *St = dyn_cast<StoreInst>(it);
2778 if (!St) return false;
2779 if (!St->isSimple() && !IsAnnotatedParallel) {
2780 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2783 Stores.push_back(St);
2788 // Now we have two lists that hold the loads and the stores.
2789 // Next, we find the pointers that they use.
2791 // Check if we see any stores. If there are no stores, then we don't
2792 // care if the pointers are *restrict*.
2793 if (!Stores.size()) {
2794 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2798 // Holds the read and read-write *pointers* that we find. These maps hold
2799 // unique values for pointers (so no need for multi-map).
2801 AliasMap ReadWrites;
2803 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2804 // multiple times on the same object. If the ptr is accessed twice, once
2805 // for read and once for write, it will only appear once (on the write
2806 // list). This is okay, since we are going to check for conflicts between
2807 // writes and between reads and writes, but not between reads and reads.
2810 ValueVector::iterator I, IE;
2811 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2812 StoreInst *ST = cast<StoreInst>(*I);
2813 Value* Ptr = ST->getPointerOperand();
2815 if (isUniform(Ptr)) {
2816 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2820 // If we did *not* see this pointer before, insert it to
2821 // the read-write list. At this phase it is only a 'write' list.
2822 if (Seen.insert(Ptr))
2823 ReadWrites.insert(std::make_pair(Ptr, ST));
2826 if (IsAnnotatedParallel) {
2828 << "LV: A loop annotated parallel, ignore memory dependency "
2833 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2834 LoadInst *LD = cast<LoadInst>(*I);
2835 Value* Ptr = LD->getPointerOperand();
2836 // If we did *not* see this pointer before, insert it to the
2837 // read list. If we *did* see it before, then it is already in
2838 // the read-write list. This allows us to vectorize expressions
2839 // such as A[i] += x; Because the address of A[i] is a read-write
2840 // pointer. This only works if the index of A[i] is consecutive.
2841 // If the address of i is unknown (for example A[B[i]]) then we may
2842 // read a few words, modify, and write a few words, and some of the
2843 // words may be written to the same address.
2844 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2845 Reads.insert(std::make_pair(Ptr, LD));
2848 // If we write (or read-write) to a single destination and there are no
2849 // other reads in this loop then is it safe to vectorize.
2850 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2851 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2855 unsigned NumReadPtrs = 0;
2856 unsigned NumWritePtrs = 0;
2858 // Find pointers with computable bounds. We are going to use this information
2859 // to place a runtime bound check.
2860 bool CanDoRT = true;
2861 AliasMap::iterator MI, ME;
2862 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2863 Value *V = (*MI).first;
2864 if (hasComputableBounds(V)) {
2865 PtrRtCheck.insert(SE, TheLoop, V, true);
2867 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2873 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2874 Value *V = (*MI).first;
2875 if (hasComputableBounds(V)) {
2876 PtrRtCheck.insert(SE, TheLoop, V, false);
2878 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2885 // Check that we did not collect too many pointers or found a
2886 // unsizeable pointer.
2887 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
2888 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
2889 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
2895 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2898 bool NeedRTCheck = false;
2900 // Biggest vectorized access possible, vector width * unroll factor.
2901 // TODO: We're being very pessimistic here, find a way to know the
2902 // real access width before getting here.
2903 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2904 TTI->getMaximumUnrollFactor();
2905 // Now that the pointers are in two lists (Reads and ReadWrites), we
2906 // can check that there are no conflicts between each of the writes and
2907 // between the writes to the reads.
2908 // Note that WriteObjects duplicates the stores (indexed now by underlying
2909 // objects) to avoid pointing to elements inside ReadWrites.
2910 // TODO: Maybe create a new type where they can interact without duplication.
2911 AliasMultiMap WriteObjects;
2912 ValueVector TempObjects;
2914 // Check that the read-writes do not conflict with other read-write
2916 bool AllWritesIdentified = true;
2917 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2918 Value *Val = (*MI).first;
2919 Instruction *Inst = (*MI).second;
2921 GetUnderlyingObjects(Val, TempObjects, DL);
2922 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2924 if (!isIdentifiedObject(*UI)) {
2925 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2927 AllWritesIdentified = false;
2930 // Never seen it before, can't alias.
2931 if (WriteObjects[*UI].empty()) {
2932 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2933 WriteObjects[*UI].push_back(Inst);
2936 // Direct alias found.
2937 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2938 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2942 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2944 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2945 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2947 // If global alias, make sure they do alias.
2948 if (hasPossibleGlobalWriteReorder(*UI,
2952 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
2957 // Didn't alias, insert into map for further reference.
2958 WriteObjects[*UI].push_back(Inst);
2960 TempObjects.clear();
2963 /// Check that the reads don't conflict with the read-writes.
2964 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2965 Value *Val = (*MI).first;
2966 GetUnderlyingObjects(Val, TempObjects, DL);
2967 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2969 // If all of the writes are identified then we don't care if the read
2970 // pointer is identified or not.
2971 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2972 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2976 // Never seen it before, can't alias.
2977 if (WriteObjects[*UI].empty())
2979 // Direct alias found.
2980 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2981 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2985 DEBUG(dbgs() << "LV: Found a global value: "
2987 Instruction *Inst = (*MI).second;
2988 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2989 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2991 // If global alias, make sure they do alias.
2992 if (hasPossibleGlobalWriteReorder(*UI,
2996 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
3001 TempObjects.clear();
3004 PtrRtCheck.Need = NeedRTCheck;
3005 if (NeedRTCheck && !CanDoRT) {
3006 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3007 "the array bounds.\n");
3012 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3013 " need a runtime memory check.\n");
3017 static bool hasMultipleUsesOf(Instruction *I,
3018 SmallPtrSet<Instruction *, 8> &Insts) {
3019 unsigned NumUses = 0;
3020 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3021 if (Insts.count(dyn_cast<Instruction>(*Use)))
3030 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3031 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3032 if (!Set.count(dyn_cast<Instruction>(*Use)))
3037 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3038 ReductionKind Kind) {
3039 if (Phi->getNumIncomingValues() != 2)
3042 // Reduction variables are only found in the loop header block.
3043 if (Phi->getParent() != TheLoop->getHeader())
3046 // Obtain the reduction start value from the value that comes from the loop
3048 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3050 // ExitInstruction is the single value which is used outside the loop.
3051 // We only allow for a single reduction value to be used outside the loop.
3052 // This includes users of the reduction, variables (which form a cycle
3053 // which ends in the phi node).
3054 Instruction *ExitInstruction = 0;
3055 // Indicates that we found a reduction operation in our scan.
3056 bool FoundReduxOp = false;
3058 // We start with the PHI node and scan for all of the users of this
3059 // instruction. All users must be instructions that can be used as reduction
3060 // variables (such as ADD). We must have a single out-of-block user. The cycle
3061 // must include the original PHI.
3062 bool FoundStartPHI = false;
3064 // To recognize min/max patterns formed by a icmp select sequence, we store
3065 // the number of instruction we saw from the recognized min/max pattern,
3066 // to make sure we only see exactly the two instructions.
3067 unsigned NumCmpSelectPatternInst = 0;
3068 ReductionInstDesc ReduxDesc(false, 0);
3070 SmallPtrSet<Instruction *, 8> VisitedInsts;
3071 SmallVector<Instruction *, 8> Worklist;
3072 Worklist.push_back(Phi);
3073 VisitedInsts.insert(Phi);
3075 // A value in the reduction can be used:
3076 // - By the reduction:
3077 // - Reduction operation:
3078 // - One use of reduction value (safe).
3079 // - Multiple use of reduction value (not safe).
3081 // - All uses of the PHI must be the reduction (safe).
3082 // - Otherwise, not safe.
3083 // - By one instruction outside of the loop (safe).
3084 // - By further instructions outside of the loop (not safe).
3085 // - By an instruction that is not part of the reduction (not safe).
3087 // * An instruction type other than PHI or the reduction operation.
3088 // * A PHI in the header other than the initial PHI.
3089 while (!Worklist.empty()) {
3090 Instruction *Cur = Worklist.back();
3091 Worklist.pop_back();
3094 // If the instruction has no users then this is a broken chain and can't be
3095 // a reduction variable.
3096 if (Cur->use_empty())
3099 bool IsAPhi = isa<PHINode>(Cur);
3101 // A header PHI use other than the original PHI.
3102 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3105 // Reductions of instructions such as Div, and Sub is only possible if the
3106 // LHS is the reduction variable.
3107 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3108 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3109 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3112 // Any reduction instruction must be of one of the allowed kinds.
3113 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3114 if (!ReduxDesc.IsReduction)
3117 // A reduction operation must only have one use of the reduction value.
3118 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3119 hasMultipleUsesOf(Cur, VisitedInsts))
3122 // All inputs to a PHI node must be a reduction value.
3123 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3126 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3127 isa<SelectInst>(Cur)))
3128 ++NumCmpSelectPatternInst;
3129 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3130 isa<SelectInst>(Cur)))
3131 ++NumCmpSelectPatternInst;
3133 // Check whether we found a reduction operator.
3134 FoundReduxOp |= !IsAPhi;
3136 // Process users of current instruction. Push non PHI nodes after PHI nodes
3137 // onto the stack. This way we are going to have seen all inputs to PHI
3138 // nodes once we get to them.
3139 SmallVector<Instruction *, 8> NonPHIs;
3140 SmallVector<Instruction *, 8> PHIs;
3141 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3143 Instruction *Usr = cast<Instruction>(*UI);
3145 // Check if we found the exit user.
3146 BasicBlock *Parent = Usr->getParent();
3147 if (!TheLoop->contains(Parent)) {
3148 // Exit if you find multiple outside users.
3149 if (ExitInstruction != 0)
3151 ExitInstruction = Cur;
3155 // Process instructions only once (termination).
3156 if (VisitedInsts.insert(Usr)) {
3157 if (isa<PHINode>(Usr))
3158 PHIs.push_back(Usr);
3160 NonPHIs.push_back(Usr);
3162 // Remember that we completed the cycle.
3164 FoundStartPHI = true;
3166 Worklist.append(PHIs.begin(), PHIs.end());
3167 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3170 // This means we have seen one but not the other instruction of the
3171 // pattern or more than just a select and cmp.
3172 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3173 NumCmpSelectPatternInst != 2)
3176 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3179 // We found a reduction var if we have reached the original phi node and we
3180 // only have a single instruction with out-of-loop users.
3182 // This instruction is allowed to have out-of-loop users.
3183 AllowedExit.insert(ExitInstruction);
3185 // Save the description of this reduction variable.
3186 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3187 ReduxDesc.MinMaxKind);
3188 Reductions[Phi] = RD;
3189 // We've ended the cycle. This is a reduction variable if we have an
3190 // outside user and it has a binary op.
3195 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3196 /// pattern corresponding to a min(X, Y) or max(X, Y).
3197 LoopVectorizationLegality::ReductionInstDesc
3198 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3199 ReductionInstDesc &Prev) {
3201 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3202 "Expect a select instruction");
3203 Instruction *Cmp = 0;
3204 SelectInst *Select = 0;
3206 // We must handle the select(cmp()) as a single instruction. Advance to the
3208 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3209 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3210 return ReductionInstDesc(false, I);
3211 return ReductionInstDesc(Select, Prev.MinMaxKind);
3214 // Only handle single use cases for now.
3215 if (!(Select = dyn_cast<SelectInst>(I)))
3216 return ReductionInstDesc(false, I);
3217 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3218 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3219 return ReductionInstDesc(false, I);
3220 if (!Cmp->hasOneUse())
3221 return ReductionInstDesc(false, I);
3226 // Look for a min/max pattern.
3227 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3228 return ReductionInstDesc(Select, MRK_UIntMin);
3229 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3230 return ReductionInstDesc(Select, MRK_UIntMax);
3231 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3232 return ReductionInstDesc(Select, MRK_SIntMax);
3233 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3234 return ReductionInstDesc(Select, MRK_SIntMin);
3235 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3236 return ReductionInstDesc(Select, MRK_FloatMin);
3237 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3238 return ReductionInstDesc(Select, MRK_FloatMax);
3239 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3240 return ReductionInstDesc(Select, MRK_FloatMin);
3241 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3242 return ReductionInstDesc(Select, MRK_FloatMax);
3244 return ReductionInstDesc(false, I);
3247 LoopVectorizationLegality::ReductionInstDesc
3248 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3250 ReductionInstDesc &Prev) {
3251 bool FP = I->getType()->isFloatingPointTy();
3252 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3253 switch (I->getOpcode()) {
3255 return ReductionInstDesc(false, I);
3256 case Instruction::PHI:
3257 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3258 Kind != RK_FloatMinMax))
3259 return ReductionInstDesc(false, I);
3260 return ReductionInstDesc(I, Prev.MinMaxKind);
3261 case Instruction::Sub:
3262 case Instruction::Add:
3263 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3264 case Instruction::Mul:
3265 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3266 case Instruction::And:
3267 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3268 case Instruction::Or:
3269 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3270 case Instruction::Xor:
3271 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3272 case Instruction::FMul:
3273 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3274 case Instruction::FAdd:
3275 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3276 case Instruction::FCmp:
3277 case Instruction::ICmp:
3278 case Instruction::Select:
3279 if (Kind != RK_IntegerMinMax &&
3280 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3281 return ReductionInstDesc(false, I);
3282 return isMinMaxSelectCmpPattern(I, Prev);
3286 LoopVectorizationLegality::InductionKind
3287 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3288 Type *PhiTy = Phi->getType();
3289 // We only handle integer and pointer inductions variables.
3290 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3291 return IK_NoInduction;
3293 // Check that the PHI is consecutive.
3294 const SCEV *PhiScev = SE->getSCEV(Phi);
3295 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3297 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3298 return IK_NoInduction;
3300 const SCEV *Step = AR->getStepRecurrence(*SE);
3302 // Integer inductions need to have a stride of one.
3303 if (PhiTy->isIntegerTy()) {
3305 return IK_IntInduction;
3306 if (Step->isAllOnesValue())
3307 return IK_ReverseIntInduction;
3308 return IK_NoInduction;
3311 // Calculate the pointer stride and check if it is consecutive.
3312 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3314 return IK_NoInduction;
3316 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3317 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3318 if (C->getValue()->equalsInt(Size))
3319 return IK_PtrInduction;
3320 else if (C->getValue()->equalsInt(0 - Size))
3321 return IK_ReversePtrInduction;
3323 return IK_NoInduction;
3326 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3327 Value *In0 = const_cast<Value*>(V);
3328 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3332 return Inductions.count(PN);
3335 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3336 assert(TheLoop->contains(BB) && "Unknown block used");
3338 // Blocks that do not dominate the latch need predication.
3339 BasicBlock* Latch = TheLoop->getLoopLatch();
3340 return !DT->dominates(BB, Latch);
3343 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3344 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3345 // We might be able to hoist the load.
3346 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
3349 // We don't predicate stores at the moment.
3350 if (it->mayWriteToMemory() || it->mayThrow())
3353 // The instructions below can trap.
3354 switch (it->getOpcode()) {
3356 case Instruction::UDiv:
3357 case Instruction::SDiv:
3358 case Instruction::URem:
3359 case Instruction::SRem:
3367 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3368 const SCEV *PhiScev = SE->getSCEV(Ptr);
3369 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3373 return AR->isAffine();
3376 LoopVectorizationCostModel::VectorizationFactor
3377 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3379 // Width 1 means no vectorize
3380 VectorizationFactor Factor = { 1U, 0U };
3381 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3382 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3386 // Find the trip count.
3387 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3388 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3390 unsigned WidestType = getWidestType();
3391 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3392 unsigned MaxVectorSize = WidestRegister / WidestType;
3393 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3394 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3396 if (MaxVectorSize == 0) {
3397 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3401 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3402 " into one vector!");
3404 unsigned VF = MaxVectorSize;
3406 // If we optimize the program for size, avoid creating the tail loop.
3408 // If we are unable to calculate the trip count then don't try to vectorize.
3410 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3414 // Find the maximum SIMD width that can fit within the trip count.
3415 VF = TC % MaxVectorSize;
3420 // If the trip count that we found modulo the vectorization factor is not
3421 // zero then we require a tail.
3423 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3429 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3430 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3432 Factor.Width = UserVF;
3436 float Cost = expectedCost(1);
3438 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3439 for (unsigned i=2; i <= VF; i*=2) {
3440 // Notice that the vector loop needs to be executed less times, so
3441 // we need to divide the cost of the vector loops by the width of
3442 // the vector elements.
3443 float VectorCost = expectedCost(i) / (float)i;
3444 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3445 (int)VectorCost << ".\n");
3446 if (VectorCost < Cost) {
3452 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3453 Factor.Width = Width;
3454 Factor.Cost = Width * Cost;
3458 unsigned LoopVectorizationCostModel::getWidestType() {
3459 unsigned MaxWidth = 8;
3462 for (Loop::block_iterator bb = TheLoop->block_begin(),
3463 be = TheLoop->block_end(); bb != be; ++bb) {
3464 BasicBlock *BB = *bb;
3466 // For each instruction in the loop.
3467 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3468 Type *T = it->getType();
3470 // Only examine Loads, Stores and PHINodes.
3471 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3474 // Examine PHI nodes that are reduction variables.
3475 if (PHINode *PN = dyn_cast<PHINode>(it))
3476 if (!Legal->getReductionVars()->count(PN))
3479 // Examine the stored values.
3480 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3481 T = ST->getValueOperand()->getType();
3483 // Ignore loaded pointer types and stored pointer types that are not
3484 // consecutive. However, we do want to take consecutive stores/loads of
3485 // pointer vectors into account.
3486 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3489 MaxWidth = std::max(MaxWidth,
3490 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3498 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3501 unsigned LoopCost) {
3503 // -- The unroll heuristics --
3504 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3505 // There are many micro-architectural considerations that we can't predict
3506 // at this level. For example frontend pressure (on decode or fetch) due to
3507 // code size, or the number and capabilities of the execution ports.
3509 // We use the following heuristics to select the unroll factor:
3510 // 1. If the code has reductions the we unroll in order to break the cross
3511 // iteration dependency.
3512 // 2. If the loop is really small then we unroll in order to reduce the loop
3514 // 3. We don't unroll if we think that we will spill registers to memory due
3515 // to the increased register pressure.
3517 // Use the user preference, unless 'auto' is selected.
3521 // When we optimize for size we don't unroll.
3525 // Do not unroll loops with a relatively small trip count.
3526 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3527 TheLoop->getLoopLatch());
3528 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3531 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3532 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3533 " vector registers\n");
3535 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3536 // We divide by these constants so assume that we have at least one
3537 // instruction that uses at least one register.
3538 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3539 R.NumInstructions = std::max(R.NumInstructions, 1U);
3541 // We calculate the unroll factor using the following formula.
3542 // Subtract the number of loop invariants from the number of available
3543 // registers. These registers are used by all of the unrolled instances.
3544 // Next, divide the remaining registers by the number of registers that is
3545 // required by the loop, in order to estimate how many parallel instances
3546 // fit without causing spills.
3547 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3549 // Clamp the unroll factor ranges to reasonable factors.
3550 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3552 // If we did not calculate the cost for VF (because the user selected the VF)
3553 // then we calculate the cost of VF here.
3555 LoopCost = expectedCost(VF);
3557 // Clamp the calculated UF to be between the 1 and the max unroll factor
3558 // that the target allows.
3559 if (UF > MaxUnrollSize)
3564 if (Legal->getReductionVars()->size()) {
3565 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3569 // We want to unroll tiny loops in order to reduce the loop overhead.
3570 // We assume that the cost overhead is 1 and we use the cost model
3571 // to estimate the cost of the loop and unroll until the cost of the
3572 // loop overhead is about 5% of the cost of the loop.
3573 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3574 if (LoopCost < 20) {
3575 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3576 unsigned NewUF = 20/LoopCost + 1;
3577 return std::min(NewUF, UF);
3580 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3584 LoopVectorizationCostModel::RegisterUsage
3585 LoopVectorizationCostModel::calculateRegisterUsage() {
3586 // This function calculates the register usage by measuring the highest number
3587 // of values that are alive at a single location. Obviously, this is a very
3588 // rough estimation. We scan the loop in a topological order in order and
3589 // assign a number to each instruction. We use RPO to ensure that defs are
3590 // met before their users. We assume that each instruction that has in-loop
3591 // users starts an interval. We record every time that an in-loop value is
3592 // used, so we have a list of the first and last occurrences of each
3593 // instruction. Next, we transpose this data structure into a multi map that
3594 // holds the list of intervals that *end* at a specific location. This multi
3595 // map allows us to perform a linear search. We scan the instructions linearly
3596 // and record each time that a new interval starts, by placing it in a set.
3597 // If we find this value in the multi-map then we remove it from the set.
3598 // The max register usage is the maximum size of the set.
3599 // We also search for instructions that are defined outside the loop, but are
3600 // used inside the loop. We need this number separately from the max-interval
3601 // usage number because when we unroll, loop-invariant values do not take
3603 LoopBlocksDFS DFS(TheLoop);
3607 R.NumInstructions = 0;
3609 // Each 'key' in the map opens a new interval. The values
3610 // of the map are the index of the 'last seen' usage of the
3611 // instruction that is the key.
3612 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3613 // Maps instruction to its index.
3614 DenseMap<unsigned, Instruction*> IdxToInstr;
3615 // Marks the end of each interval.
3616 IntervalMap EndPoint;
3617 // Saves the list of instruction indices that are used in the loop.
3618 SmallSet<Instruction*, 8> Ends;
3619 // Saves the list of values that are used in the loop but are
3620 // defined outside the loop, such as arguments and constants.
3621 SmallPtrSet<Value*, 8> LoopInvariants;
3624 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3625 be = DFS.endRPO(); bb != be; ++bb) {
3626 R.NumInstructions += (*bb)->size();
3627 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3629 Instruction *I = it;
3630 IdxToInstr[Index++] = I;
3632 // Save the end location of each USE.
3633 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3634 Value *U = I->getOperand(i);
3635 Instruction *Instr = dyn_cast<Instruction>(U);
3637 // Ignore non-instruction values such as arguments, constants, etc.
3638 if (!Instr) continue;
3640 // If this instruction is outside the loop then record it and continue.
3641 if (!TheLoop->contains(Instr)) {
3642 LoopInvariants.insert(Instr);
3646 // Overwrite previous end points.
3647 EndPoint[Instr] = Index;
3653 // Saves the list of intervals that end with the index in 'key'.
3654 typedef SmallVector<Instruction*, 2> InstrList;
3655 DenseMap<unsigned, InstrList> TransposeEnds;
3657 // Transpose the EndPoints to a list of values that end at each index.
3658 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3660 TransposeEnds[it->second].push_back(it->first);
3662 SmallSet<Instruction*, 8> OpenIntervals;
3663 unsigned MaxUsage = 0;
3666 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3667 for (unsigned int i = 0; i < Index; ++i) {
3668 Instruction *I = IdxToInstr[i];
3669 // Ignore instructions that are never used within the loop.
3670 if (!Ends.count(I)) continue;
3672 // Remove all of the instructions that end at this location.
3673 InstrList &List = TransposeEnds[i];
3674 for (unsigned int j=0, e = List.size(); j < e; ++j)
3675 OpenIntervals.erase(List[j]);
3677 // Count the number of live interals.
3678 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3680 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3681 OpenIntervals.size() <<"\n");
3683 // Add the current instruction to the list of open intervals.
3684 OpenIntervals.insert(I);
3687 unsigned Invariant = LoopInvariants.size();
3688 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3689 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3690 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3692 R.LoopInvariantRegs = Invariant;
3693 R.MaxLocalUsers = MaxUsage;
3697 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3701 for (Loop::block_iterator bb = TheLoop->block_begin(),
3702 be = TheLoop->block_end(); bb != be; ++bb) {
3703 unsigned BlockCost = 0;
3704 BasicBlock *BB = *bb;
3706 // For each instruction in the old loop.
3707 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3708 // Skip dbg intrinsics.
3709 if (isa<DbgInfoIntrinsic>(it))
3712 unsigned C = getInstructionCost(it, VF);
3714 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3715 VF << " For instruction: "<< *it << "\n");
3718 // We assume that if-converted blocks have a 50% chance of being executed.
3719 // When the code is scalar then some of the blocks are avoided due to CF.
3720 // When the code is vectorized we execute all code paths.
3721 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3731 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3732 // If we know that this instruction will remain uniform, check the cost of
3733 // the scalar version.
3734 if (Legal->isUniformAfterVectorization(I))
3737 Type *RetTy = I->getType();
3738 Type *VectorTy = ToVectorTy(RetTy, VF);
3740 // TODO: We need to estimate the cost of intrinsic calls.
3741 switch (I->getOpcode()) {
3742 case Instruction::GetElementPtr:
3743 // We mark this instruction as zero-cost because the cost of GEPs in
3744 // vectorized code depends on whether the corresponding memory instruction
3745 // is scalarized or not. Therefore, we handle GEPs with the memory
3746 // instruction cost.
3748 case Instruction::Br: {
3749 return TTI.getCFInstrCost(I->getOpcode());
3751 case Instruction::PHI:
3752 //TODO: IF-converted IFs become selects.
3754 case Instruction::Add:
3755 case Instruction::FAdd:
3756 case Instruction::Sub:
3757 case Instruction::FSub:
3758 case Instruction::Mul:
3759 case Instruction::FMul:
3760 case Instruction::UDiv:
3761 case Instruction::SDiv:
3762 case Instruction::FDiv:
3763 case Instruction::URem:
3764 case Instruction::SRem:
3765 case Instruction::FRem:
3766 case Instruction::Shl:
3767 case Instruction::LShr:
3768 case Instruction::AShr:
3769 case Instruction::And:
3770 case Instruction::Or:
3771 case Instruction::Xor: {
3772 // Certain instructions can be cheaper to vectorize if they have a constant
3773 // second vector operand. One example of this are shifts on x86.
3774 TargetTransformInfo::OperandValueKind Op1VK =
3775 TargetTransformInfo::OK_AnyValue;
3776 TargetTransformInfo::OperandValueKind Op2VK =
3777 TargetTransformInfo::OK_AnyValue;
3779 if (isa<ConstantInt>(I->getOperand(1)))
3780 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3782 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3784 case Instruction::Select: {
3785 SelectInst *SI = cast<SelectInst>(I);
3786 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3787 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3788 Type *CondTy = SI->getCondition()->getType();
3790 CondTy = VectorType::get(CondTy, VF);
3792 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3794 case Instruction::ICmp:
3795 case Instruction::FCmp: {
3796 Type *ValTy = I->getOperand(0)->getType();
3797 VectorTy = ToVectorTy(ValTy, VF);
3798 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3800 case Instruction::Store:
3801 case Instruction::Load: {
3802 StoreInst *SI = dyn_cast<StoreInst>(I);
3803 LoadInst *LI = dyn_cast<LoadInst>(I);
3804 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3806 VectorTy = ToVectorTy(ValTy, VF);
3808 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3809 unsigned AS = SI ? SI->getPointerAddressSpace() :
3810 LI->getPointerAddressSpace();
3811 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3812 // We add the cost of address computation here instead of with the gep
3813 // instruction because only here we know whether the operation is
3816 return TTI.getAddressComputationCost(VectorTy) +
3817 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3819 // Scalarized loads/stores.
3820 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3821 bool Reverse = ConsecutiveStride < 0;
3822 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3823 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3824 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3826 // The cost of extracting from the value vector and pointer vector.
3827 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3828 for (unsigned i = 0; i < VF; ++i) {
3829 // The cost of extracting the pointer operand.
3830 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3831 // In case of STORE, the cost of ExtractElement from the vector.
3832 // In case of LOAD, the cost of InsertElement into the returned
3834 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3835 Instruction::InsertElement,
3839 // The cost of the scalar loads/stores.
3840 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3841 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3846 // Wide load/stores.
3847 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3848 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3851 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3855 case Instruction::ZExt:
3856 case Instruction::SExt:
3857 case Instruction::FPToUI:
3858 case Instruction::FPToSI:
3859 case Instruction::FPExt:
3860 case Instruction::PtrToInt:
3861 case Instruction::IntToPtr:
3862 case Instruction::SIToFP:
3863 case Instruction::UIToFP:
3864 case Instruction::Trunc:
3865 case Instruction::FPTrunc:
3866 case Instruction::BitCast: {
3867 // We optimize the truncation of induction variable.
3868 // The cost of these is the same as the scalar operation.
3869 if (I->getOpcode() == Instruction::Trunc &&
3870 Legal->isInductionVariable(I->getOperand(0)))
3871 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3872 I->getOperand(0)->getType());
3874 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3875 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3877 case Instruction::Call: {
3878 CallInst *CI = cast<CallInst>(I);
3879 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3880 assert(ID && "Not an intrinsic call!");
3881 Type *RetTy = ToVectorTy(CI->getType(), VF);
3882 SmallVector<Type*, 4> Tys;
3883 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3884 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3885 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3888 // We are scalarizing the instruction. Return the cost of the scalar
3889 // instruction, plus the cost of insert and extract into vector
3890 // elements, times the vector width.
3893 if (!RetTy->isVoidTy() && VF != 1) {
3894 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3896 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3899 // The cost of inserting the results plus extracting each one of the
3901 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3904 // The cost of executing VF copies of the scalar instruction. This opcode
3905 // is unknown. Assume that it is the same as 'mul'.
3906 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3912 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3913 if (Scalar->isVoidTy() || VF == 1)
3915 return VectorType::get(Scalar, VF);
3918 char LoopVectorize::ID = 0;
3919 static const char lv_name[] = "Loop Vectorization";
3920 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3921 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3922 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3923 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3924 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3925 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3928 Pass *createLoopVectorizePass() {
3929 return new LoopVectorize();
3933 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3934 // Check for a store.
3935 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3936 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3938 // Check for a load.
3939 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3940 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;