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. Legalization of the IR is done
12 // in the codegen. However, the vectorizes uses (will use) the codegen
13 // interfaces to generate IR that is likely to result in an optimal binary.
15 // The loop vectorizer combines consecutive loop iteration 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/raw_ostream.h"
82 #include "llvm/Transforms/Scalar.h"
83 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
84 #include "llvm/Transforms/Utils/Local.h"
90 static cl::opt<unsigned>
91 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
92 cl::desc("Sets the SIMD width. Zero is autoselect."));
94 static cl::opt<unsigned>
95 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
96 cl::desc("Sets the vectorization unroll count. "
97 "Zero is autoselect."));
100 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
101 cl::desc("Enable if-conversion during vectorization."));
103 /// We don't vectorize loops with a known constant trip count below this number.
104 static cl::opt<unsigned>
105 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16), cl::Hidden,
106 cl::desc("The minimum trip count in the loops to vectorize."));
108 /// We don't unroll loops with a known constant trip count below this number.
109 static const unsigned TinyTripCountUnrollThreshold = 128;
111 /// When performing a runtime memory check, do not check more than this
112 /// number of pointers. Notice that the check is quadratic!
113 static const unsigned RuntimeMemoryCheckThreshold = 4;
117 // Forward declarations.
118 class LoopVectorizationLegality;
119 class LoopVectorizationCostModel;
121 /// InnerLoopVectorizer vectorizes loops which contain only one basic
122 /// block to a specified vectorization factor (VF).
123 /// This class performs the widening of scalars into vectors, or multiple
124 /// scalars. This class also implements the following features:
125 /// * It inserts an epilogue loop for handling loops that don't have iteration
126 /// counts that are known to be a multiple of the vectorization factor.
127 /// * It handles the code generation for reduction variables.
128 /// * Scalarization (implementation using scalars) of un-vectorizable
130 /// InnerLoopVectorizer does not perform any vectorization-legality
131 /// checks, and relies on the caller to check for the different legality
132 /// aspects. The InnerLoopVectorizer relies on the
133 /// LoopVectorizationLegality class to provide information about the induction
134 /// and reduction variables that were found to a given vectorization factor.
135 class InnerLoopVectorizer {
137 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
138 DominatorTree *DT, DataLayout *DL, unsigned VecWidth,
139 unsigned UnrollFactor)
140 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), VF(VecWidth),
141 UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
142 OldInduction(0), WidenMap(UnrollFactor) {}
144 // Perform the actual loop widening (vectorization).
145 void vectorize(LoopVectorizationLegality *Legal) {
146 // Create a new empty loop. Unlink the old loop and connect the new one.
147 createEmptyLoop(Legal);
148 // Widen each instruction in the old loop to a new one in the new loop.
149 // Use the Legality module to find the induction and reduction variables.
150 vectorizeLoop(Legal);
151 // Register the new loop and update the analysis passes.
156 /// A small list of PHINodes.
157 typedef SmallVector<PHINode*, 4> PhiVector;
158 /// When we unroll loops we have multiple vector values for each scalar.
159 /// This data structure holds the unrolled and vectorized values that
160 /// originated from one scalar instruction.
161 typedef SmallVector<Value*, 2> VectorParts;
163 /// Add code that checks at runtime if the accessed arrays overlap.
164 /// Returns the comparator value or NULL if no check is needed.
165 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
167 /// Create an empty loop, based on the loop ranges of the old loop.
168 void createEmptyLoop(LoopVectorizationLegality *Legal);
169 /// Copy and widen the instructions from the old loop.
170 void vectorizeLoop(LoopVectorizationLegality *Legal);
172 /// A helper function that computes the predicate of the block BB, assuming
173 /// that the header block of the loop is set to True. It returns the *entry*
174 /// mask for the block BB.
175 VectorParts createBlockInMask(BasicBlock *BB);
176 /// A helper function that computes the predicate of the edge between SRC
178 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
180 /// A helper function to vectorize a single BB within the innermost loop.
181 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
184 /// Insert the new loop to the loop hierarchy and pass manager
185 /// and update the analysis passes.
186 void updateAnalysis();
188 /// This instruction is un-vectorizable. Implement it as a sequence
190 void scalarizeInstruction(Instruction *Instr);
192 /// Vectorize Load and Store instructions,
193 void vectorizeMemoryInstruction(Instruction *Instr,
194 LoopVectorizationLegality *Legal);
196 /// Create a broadcast instruction. This method generates a broadcast
197 /// instruction (shuffle) for loop invariant values and for the induction
198 /// value. If this is the induction variable then we extend it to N, N+1, ...
199 /// this is needed because each iteration in the loop corresponds to a SIMD
201 Value *getBroadcastInstrs(Value *V);
203 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
204 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
205 /// The sequence starts at StartIndex.
206 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
208 /// When we go over instructions in the basic block we rely on previous
209 /// values within the current basic block or on loop invariant values.
210 /// When we widen (vectorize) values we place them in the map. If the values
211 /// are not within the map, they have to be loop invariant, so we simply
212 /// broadcast them into a vector.
213 VectorParts &getVectorValue(Value *V);
215 /// Generate a shuffle sequence that will reverse the vector Vec.
216 Value *reverseVector(Value *Vec);
218 /// This is a helper class that holds the vectorizer state. It maps scalar
219 /// instructions to vector instructions. When the code is 'unrolled' then
220 /// then a single scalar value is mapped to multiple vector parts. The parts
221 /// are stored in the VectorPart type.
223 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
225 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
227 /// \return True if 'Key' is saved in the Value Map.
228 bool has(Value *Key) const { return MapStorage.count(Key); }
230 /// Initializes a new entry in the map. Sets all of the vector parts to the
231 /// save value in 'Val'.
232 /// \return A reference to a vector with splat values.
233 VectorParts &splat(Value *Key, Value *Val) {
234 VectorParts &Entry = MapStorage[Key];
235 Entry.assign(UF, Val);
239 ///\return A reference to the value that is stored at 'Key'.
240 VectorParts &get(Value *Key) {
241 VectorParts &Entry = MapStorage[Key];
244 assert(Entry.size() == UF);
249 /// The unroll factor. Each entry in the map stores this number of vector
253 /// Map storage. We use std::map and not DenseMap because insertions to a
254 /// dense map invalidates its iterators.
255 std::map<Value *, VectorParts> MapStorage;
258 /// The original loop.
260 /// Scev analysis to use.
268 /// The vectorization SIMD factor to use. Each vector will have this many
271 /// The vectorization unroll factor to use. Each scalar is vectorized to this
272 /// many different vector instructions.
275 /// The builder that we use
278 // --- Vectorization state ---
280 /// The vector-loop preheader.
281 BasicBlock *LoopVectorPreHeader;
282 /// The scalar-loop preheader.
283 BasicBlock *LoopScalarPreHeader;
284 /// Middle Block between the vector and the scalar.
285 BasicBlock *LoopMiddleBlock;
286 ///The ExitBlock of the scalar loop.
287 BasicBlock *LoopExitBlock;
288 ///The vector loop body.
289 BasicBlock *LoopVectorBody;
290 ///The scalar loop body.
291 BasicBlock *LoopScalarBody;
292 /// A list of all bypass blocks. The first block is the entry of the loop.
293 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
295 /// The new Induction variable which was added to the new block.
297 /// The induction variable of the old basic block.
298 PHINode *OldInduction;
299 /// Maps scalars to widened vectors.
303 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
304 /// to what vectorization factor.
305 /// This class does not look at the profitability of vectorization, only the
306 /// legality. This class has two main kinds of checks:
307 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
308 /// will change the order of memory accesses in a way that will change the
309 /// correctness of the program.
310 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
311 /// checks for a number of different conditions, such as the availability of a
312 /// single induction variable, that all types are supported and vectorize-able,
313 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
314 /// This class is also used by InnerLoopVectorizer for identifying
315 /// induction variable and the different reduction variables.
316 class LoopVectorizationLegality {
318 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
320 : TheLoop(L), SE(SE), DL(DL), DT(DT), Induction(0) {}
322 /// This enum represents the kinds of reductions that we support.
324 RK_NoReduction, ///< Not a reduction.
325 RK_IntegerAdd, ///< Sum of integers.
326 RK_IntegerMult, ///< Product of integers.
327 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
328 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
329 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
330 RK_FloatAdd, ///< Sum of floats.
331 RK_FloatMult ///< Product of floats.
334 /// This enum represents the kinds of inductions that we support.
336 IK_NoInduction, ///< Not an induction variable.
337 IK_IntInduction, ///< Integer induction variable. Step = 1.
338 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
339 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
340 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
343 /// This POD struct holds information about reduction variables.
344 struct ReductionDescriptor {
345 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
346 Kind(RK_NoReduction) {}
348 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K)
349 : StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
351 // The starting value of the reduction.
352 // It does not have to be zero!
354 // The instruction who's value is used outside the loop.
355 Instruction *LoopExitInstr;
356 // The kind of the reduction.
360 // This POD struct holds information about the memory runtime legality
361 // check that a group of pointers do not overlap.
362 struct RuntimePointerCheck {
363 RuntimePointerCheck() : Need(false) {}
365 /// Reset the state of the pointer runtime information.
373 /// Insert a pointer and calculate the start and end SCEVs.
374 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr);
376 /// This flag indicates if we need to add the runtime check.
378 /// Holds the pointers that we need to check.
379 SmallVector<Value*, 2> Pointers;
380 /// Holds the pointer value at the beginning of the loop.
381 SmallVector<const SCEV*, 2> Starts;
382 /// Holds the pointer value at the end of the loop.
383 SmallVector<const SCEV*, 2> Ends;
386 /// A POD for saving information about induction variables.
387 struct InductionInfo {
388 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
389 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
396 /// ReductionList contains the reduction descriptors for all
397 /// of the reductions that were found in the loop.
398 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
400 /// InductionList saves induction variables and maps them to the
401 /// induction descriptor.
402 typedef MapVector<PHINode*, InductionInfo> InductionList;
404 /// Returns true if it is legal to vectorize this loop.
405 /// This does not mean that it is profitable to vectorize this
406 /// loop, only that it is legal to do so.
409 /// Returns the Induction variable.
410 PHINode *getInduction() { return Induction; }
412 /// Returns the reduction variables found in the loop.
413 ReductionList *getReductionVars() { return &Reductions; }
415 /// Returns the induction variables found in the loop.
416 InductionList *getInductionVars() { return &Inductions; }
418 /// Returns True if V is an induction variable in this loop.
419 bool isInductionVariable(const Value *V);
421 /// Return true if the block BB needs to be predicated in order for the loop
422 /// to be vectorized.
423 bool blockNeedsPredication(BasicBlock *BB);
425 /// Check if this pointer is consecutive when vectorizing. This happens
426 /// when the last index of the GEP is the induction variable, or that the
427 /// pointer itself is an induction variable.
428 /// This check allows us to vectorize A[idx] into a wide load/store.
430 /// 0 - Stride is unknown or non consecutive.
431 /// 1 - Address is consecutive.
432 /// -1 - Address is consecutive, and decreasing.
433 int isConsecutivePtr(Value *Ptr);
435 /// Returns true if the value V is uniform within the loop.
436 bool isUniform(Value *V);
438 /// Returns true if this instruction will remain scalar after vectorization.
439 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
441 /// Returns the information that we collected about runtime memory check.
442 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
444 /// Check if a single basic block loop is vectorizable.
445 /// At this point we know that this is a loop with a constant trip count
446 /// and we only need to check individual instructions.
447 bool canVectorizeInstrs();
449 /// When we vectorize loops we may change the order in which
450 /// we read and write from memory. This method checks if it is
451 /// legal to vectorize the code, considering only memory constrains.
452 /// Returns true if the loop is vectorizable
453 bool canVectorizeMemory();
455 /// Return true if we can vectorize this loop using the IF-conversion
457 bool canVectorizeWithIfConvert();
459 /// Collect the variables that need to stay uniform after vectorization.
460 void collectLoopUniforms();
462 /// Return true if all of the instructions in the block can be speculatively
464 bool blockCanBePredicated(BasicBlock *BB);
466 /// Returns True, if 'Phi' is the kind of reduction variable for type
467 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
468 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
469 /// Returns true if the instruction I can be a reduction variable of type
471 bool isReductionInstr(Instruction *I, ReductionKind Kind);
472 /// Returns the induction kind of Phi. This function may return NoInduction
473 /// if the PHI is not an induction variable.
474 InductionKind isInductionVariable(PHINode *Phi);
475 /// Return true if can compute the address bounds of Ptr within the loop.
476 bool hasComputableBounds(Value *Ptr);
478 /// The loop that we evaluate.
482 /// DataLayout analysis.
487 // --- vectorization state --- //
489 /// Holds the integer induction variable. This is the counter of the
492 /// Holds the reduction variables.
493 ReductionList Reductions;
494 /// Holds all of the induction variables that we found in the loop.
495 /// Notice that inductions don't need to start at zero and that induction
496 /// variables can be pointers.
497 InductionList Inductions;
499 /// Allowed outside users. This holds the reduction
500 /// vars which can be accessed from outside the loop.
501 SmallPtrSet<Value*, 4> AllowedExit;
502 /// This set holds the variables which are known to be uniform after
504 SmallPtrSet<Instruction*, 4> Uniforms;
505 /// We need to check that all of the pointers in this list are disjoint
507 RuntimePointerCheck PtrRtCheck;
510 /// LoopVectorizationCostModel - estimates the expected speedups due to
512 /// In many cases vectorization is not profitable. This can happen because of
513 /// a number of reasons. In this class we mainly attempt to predict the
514 /// expected speedup/slowdowns due to the supported instruction set. We use the
515 /// TargetTransformInfo to query the different backends for the cost of
516 /// different operations.
517 class LoopVectorizationCostModel {
519 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
520 LoopVectorizationLegality *Legal,
521 const TargetTransformInfo &TTI)
522 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI) {}
524 /// Information about vectorization costs
525 struct VectorizationFactor {
526 unsigned Width; // Vector width with best cost
527 unsigned Cost; // Cost of the loop with that width
529 /// \return The most profitable vectorization factor and the cost of that VF.
530 /// This method checks every power of two up to VF. If UserVF is not ZERO
531 /// then this vectorization factor will be selected if vectorization is
533 VectorizationFactor selectVectorizationFactor(bool OptForSize, unsigned UserVF);
535 /// \return The size (in bits) of the widest type in the code that
536 /// needs to be vectorized. We ignore values that remain scalar such as
537 /// 64 bit loop indices.
538 unsigned getWidestType();
540 /// \return The most profitable unroll factor.
541 /// If UserUF is non-zero then this method finds the best unroll-factor
542 /// based on register pressure and other parameters.
543 /// VF and LoopCost are the selected vectorization factor and the cost of the
545 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
548 /// \brief A struct that represents some properties of the register usage
550 struct RegisterUsage {
551 /// Holds the number of loop invariant values that are used in the loop.
552 unsigned LoopInvariantRegs;
553 /// Holds the maximum number of concurrent live intervals in the loop.
554 unsigned MaxLocalUsers;
555 /// Holds the number of instructions in the loop.
556 unsigned NumInstructions;
559 /// \return information about the register usage of the loop.
560 RegisterUsage calculateRegisterUsage();
563 /// Returns the expected execution cost. The unit of the cost does
564 /// not matter because we use the 'cost' units to compare different
565 /// vector widths. The cost that is returned is *not* normalized by
566 /// the factor width.
567 unsigned expectedCost(unsigned VF);
569 /// Returns the execution time cost of an instruction for a given vector
570 /// width. Vector width of one means scalar.
571 unsigned getInstructionCost(Instruction *I, unsigned VF);
573 /// A helper function for converting Scalar types to vector types.
574 /// If the incoming type is void, we return void. If the VF is 1, we return
576 static Type* ToVectorTy(Type *Scalar, unsigned VF);
578 /// The loop that we evaluate.
582 /// Loop Info analysis.
584 /// Vectorization legality.
585 LoopVectorizationLegality *Legal;
586 /// Vector target information.
587 const TargetTransformInfo &TTI;
590 /// The LoopVectorize Pass.
591 struct LoopVectorize : public LoopPass {
592 /// Pass identification, replacement for typeid
595 explicit LoopVectorize() : LoopPass(ID) {
596 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
602 TargetTransformInfo *TTI;
605 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
606 // We only vectorize innermost loops.
610 SE = &getAnalysis<ScalarEvolution>();
611 DL = getAnalysisIfAvailable<DataLayout>();
612 LI = &getAnalysis<LoopInfo>();
613 TTI = &getAnalysis<TargetTransformInfo>();
614 DT = &getAnalysis<DominatorTree>();
616 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
617 L->getHeader()->getParent()->getName() << "\"\n");
619 // Check if it is legal to vectorize the loop.
620 LoopVectorizationLegality LVL(L, SE, DL, DT);
621 if (!LVL.canVectorize()) {
622 DEBUG(dbgs() << "LV: Not vectorizing.\n");
626 // Use the cost model.
627 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI);
629 // Check the function attribues to find out if this function should be
630 // optimized for size.
631 Function *F = L->getHeader()->getParent();
632 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
633 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
634 unsigned FnIndex = AttributeSet::FunctionIndex;
635 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
636 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
639 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
640 "attribute is used.\n");
644 // Select the optimal vectorization factor.
645 LoopVectorizationCostModel::VectorizationFactor VF;
646 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
647 // Select the unroll factor.
648 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
652 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
656 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
657 F->getParent()->getModuleIdentifier()<<"\n");
658 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
660 // If we decided that it is *legal* to vectorizer the loop then do it.
661 InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF.Width, UF);
664 DEBUG(verifyFunction(*L->getHeader()->getParent()));
668 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
669 LoopPass::getAnalysisUsage(AU);
670 AU.addRequiredID(LoopSimplifyID);
671 AU.addRequiredID(LCSSAID);
672 AU.addRequired<DominatorTree>();
673 AU.addRequired<LoopInfo>();
674 AU.addRequired<ScalarEvolution>();
675 AU.addRequired<TargetTransformInfo>();
676 AU.addPreserved<LoopInfo>();
677 AU.addPreserved<DominatorTree>();
682 } // end anonymous namespace
684 //===----------------------------------------------------------------------===//
685 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
686 // LoopVectorizationCostModel.
687 //===----------------------------------------------------------------------===//
690 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
691 Loop *Lp, Value *Ptr) {
692 const SCEV *Sc = SE->getSCEV(Ptr);
693 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
694 assert(AR && "Invalid addrec expression");
695 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
696 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
697 Pointers.push_back(Ptr);
698 Starts.push_back(AR->getStart());
699 Ends.push_back(ScEnd);
702 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
703 // Save the current insertion location.
704 Instruction *Loc = Builder.GetInsertPoint();
706 // We need to place the broadcast of invariant variables outside the loop.
707 Instruction *Instr = dyn_cast<Instruction>(V);
708 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
709 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
711 // Place the code for broadcasting invariant variables in the new preheader.
713 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
715 // Broadcast the scalar into all locations in the vector.
716 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
718 // Restore the builder insertion point.
720 Builder.SetInsertPoint(Loc);
725 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
727 assert(Val->getType()->isVectorTy() && "Must be a vector");
728 assert(Val->getType()->getScalarType()->isIntegerTy() &&
729 "Elem must be an integer");
731 Type *ITy = Val->getType()->getScalarType();
732 VectorType *Ty = cast<VectorType>(Val->getType());
733 int VLen = Ty->getNumElements();
734 SmallVector<Constant*, 8> Indices;
736 // Create a vector of consecutive numbers from zero to VF.
737 for (int i = 0; i < VLen; ++i) {
738 int Idx = Negate ? (-i): i;
739 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
742 // Add the consecutive indices to the vector value.
743 Constant *Cv = ConstantVector::get(Indices);
744 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
745 return Builder.CreateAdd(Val, Cv, "induction");
748 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
749 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
750 // Make sure that the pointer does not point to structs.
751 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
754 // If this value is a pointer induction variable we know it is consecutive.
755 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
756 if (Phi && Inductions.count(Phi)) {
757 InductionInfo II = Inductions[Phi];
758 if (IK_PtrInduction == II.IK)
760 else if (IK_ReversePtrInduction == II.IK)
764 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
768 unsigned NumOperands = Gep->getNumOperands();
769 Value *LastIndex = Gep->getOperand(NumOperands - 1);
771 Value *GpPtr = Gep->getPointerOperand();
772 // If this GEP value is a consecutive pointer induction variable and all of
773 // the indices are constant then we know it is consecutive. We can
774 Phi = dyn_cast<PHINode>(GpPtr);
775 if (Phi && Inductions.count(Phi)) {
777 // Make sure that the pointer does not point to structs.
778 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
779 if (GepPtrType->getElementType()->isAggregateType())
782 // Make sure that all of the index operands are loop invariant.
783 for (unsigned i = 1; i < NumOperands; ++i)
784 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
787 InductionInfo II = Inductions[Phi];
788 if (IK_PtrInduction == II.IK)
790 else if (IK_ReversePtrInduction == II.IK)
794 // Check that all of the gep indices are uniform except for the last.
795 for (unsigned i = 0; i < NumOperands - 1; ++i)
796 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
799 // We can emit wide load/stores only if the last index is the induction
801 const SCEV *Last = SE->getSCEV(LastIndex);
802 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
803 const SCEV *Step = AR->getStepRecurrence(*SE);
805 // The memory is consecutive because the last index is consecutive
806 // and all other indices are loop invariant.
809 if (Step->isAllOnesValue())
816 bool LoopVectorizationLegality::isUniform(Value *V) {
817 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
820 InnerLoopVectorizer::VectorParts&
821 InnerLoopVectorizer::getVectorValue(Value *V) {
822 assert(V != Induction && "The new induction variable should not be used.");
823 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
825 // If we have this scalar in the map, return it.
827 return WidenMap.get(V);
829 // If this scalar is unknown, assume that it is a constant or that it is
830 // loop invariant. Broadcast V and save the value for future uses.
831 Value *B = getBroadcastInstrs(V);
832 return WidenMap.splat(V, B);
835 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
836 assert(Vec->getType()->isVectorTy() && "Invalid type");
837 SmallVector<Constant*, 8> ShuffleMask;
838 for (unsigned i = 0; i < VF; ++i)
839 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
841 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
842 ConstantVector::get(ShuffleMask),
847 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
848 LoopVectorizationLegality *Legal) {
849 // Attempt to issue a wide load.
850 LoadInst *LI = dyn_cast<LoadInst>(Instr);
851 StoreInst *SI = dyn_cast<StoreInst>(Instr);
853 assert((LI || SI) && "Invalid Load/Store instruction");
855 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
856 Type *DataTy = VectorType::get(ScalarDataTy, VF);
857 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
858 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
860 // If the pointer is loop invariant or if it is non consecutive,
861 // scalarize the load.
862 int Stride = Legal->isConsecutivePtr(Ptr);
863 bool Reverse = Stride < 0;
864 bool UniformLoad = LI && Legal->isUniform(Ptr);
865 if (Stride == 0 || UniformLoad)
866 return scalarizeInstruction(Instr);
868 Constant *Zero = Builder.getInt32(0);
869 VectorParts &Entry = WidenMap.get(Instr);
871 // Handle consecutive loads/stores.
872 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
873 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
874 Value *PtrOperand = Gep->getPointerOperand();
875 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
876 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
878 // Create the new GEP with the new induction variable.
879 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
880 Gep2->setOperand(0, FirstBasePtr);
881 Gep2->setName("gep.indvar.base");
882 Ptr = Builder.Insert(Gep2);
884 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
885 OrigLoop) && "Base ptr must be invariant");
887 // The last index does not have to be the induction. It can be
888 // consecutive and be a function of the index. For example A[I+1];
889 unsigned NumOperands = Gep->getNumOperands();
891 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
892 VectorParts &GEPParts = getVectorValue(LastGepOperand);
893 Value *LastIndex = GEPParts[0];
894 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
896 // Create the new GEP with the new induction variable.
897 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
898 Gep2->setOperand(NumOperands - 1, LastIndex);
899 Gep2->setName("gep.indvar.idx");
900 Ptr = Builder.Insert(Gep2);
902 // Use the induction element ptr.
903 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
904 VectorParts &PtrVal = getVectorValue(Ptr);
905 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
910 assert(!Legal->isUniform(SI->getPointerOperand()) &&
911 "We do not allow storing to uniform addresses");
913 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
914 for (unsigned Part = 0; Part < UF; ++Part) {
915 // Calculate the pointer for the specific unroll-part.
916 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
919 // If we store to reverse consecutive memory locations then we need
920 // to reverse the order of elements in the stored value.
921 StoredVal[Part] = reverseVector(StoredVal[Part]);
922 // If the address is consecutive but reversed, then the
923 // wide store needs to start at the last vector element.
924 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
925 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
928 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
929 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
933 for (unsigned Part = 0; Part < UF; ++Part) {
934 // Calculate the pointer for the specific unroll-part.
935 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
938 // If the address is consecutive but reversed, then the
939 // wide store needs to start at the last vector element.
940 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
941 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
944 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
945 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
946 cast<LoadInst>(LI)->setAlignment(Alignment);
947 Entry[Part] = Reverse ? reverseVector(LI) : LI;
951 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
952 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
953 // Holds vector parameters or scalars, in case of uniform vals.
954 SmallVector<VectorParts, 4> Params;
956 // Find all of the vectorized parameters.
957 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
958 Value *SrcOp = Instr->getOperand(op);
960 // If we are accessing the old induction variable, use the new one.
961 if (SrcOp == OldInduction) {
962 Params.push_back(getVectorValue(SrcOp));
966 // Try using previously calculated values.
967 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
969 // If the src is an instruction that appeared earlier in the basic block
970 // then it should already be vectorized.
971 if (SrcInst && OrigLoop->contains(SrcInst)) {
972 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
973 // The parameter is a vector value from earlier.
974 Params.push_back(WidenMap.get(SrcInst));
976 // The parameter is a scalar from outside the loop. Maybe even a constant.
978 Scalars.append(UF, SrcOp);
979 Params.push_back(Scalars);
983 assert(Params.size() == Instr->getNumOperands() &&
984 "Invalid number of operands");
986 // Does this instruction return a value ?
987 bool IsVoidRetTy = Instr->getType()->isVoidTy();
989 Value *UndefVec = IsVoidRetTy ? 0 :
990 UndefValue::get(VectorType::get(Instr->getType(), VF));
991 // Create a new entry in the WidenMap and initialize it to Undef or Null.
992 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
994 // For each scalar that we create:
995 for (unsigned Width = 0; Width < VF; ++Width) {
996 // For each vector unroll 'part':
997 for (unsigned Part = 0; Part < UF; ++Part) {
998 Instruction *Cloned = Instr->clone();
1000 Cloned->setName(Instr->getName() + ".cloned");
1001 // Replace the operands of the cloned instrucions with extracted scalars.
1002 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1003 Value *Op = Params[op][Part];
1004 // Param is a vector. Need to extract the right lane.
1005 if (Op->getType()->isVectorTy())
1006 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1007 Cloned->setOperand(op, Op);
1010 // Place the cloned scalar in the new loop.
1011 Builder.Insert(Cloned);
1013 // If the original scalar returns a value we need to place it in a vector
1014 // so that future users will be able to use it.
1016 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1017 Builder.getInt32(Width));
1023 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1025 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1026 Legal->getRuntimePointerCheck();
1028 if (!PtrRtCheck->Need)
1031 Instruction *MemoryRuntimeCheck = 0;
1032 unsigned NumPointers = PtrRtCheck->Pointers.size();
1033 SmallVector<Value* , 2> Starts;
1034 SmallVector<Value* , 2> Ends;
1036 SCEVExpander Exp(*SE, "induction");
1038 // Use this type for pointer arithmetic.
1039 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1041 for (unsigned i = 0; i < NumPointers; ++i) {
1042 Value *Ptr = PtrRtCheck->Pointers[i];
1043 const SCEV *Sc = SE->getSCEV(Ptr);
1045 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1046 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1048 Starts.push_back(Ptr);
1049 Ends.push_back(Ptr);
1051 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1053 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1054 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1055 Starts.push_back(Start);
1056 Ends.push_back(End);
1060 IRBuilder<> ChkBuilder(Loc);
1062 for (unsigned i = 0; i < NumPointers; ++i) {
1063 for (unsigned j = i+1; j < NumPointers; ++j) {
1064 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1065 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1066 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1067 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1069 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1070 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1071 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1072 if (MemoryRuntimeCheck)
1073 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1076 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1080 return MemoryRuntimeCheck;
1084 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1086 In this function we generate a new loop. The new loop will contain
1087 the vectorized instructions while the old loop will continue to run the
1090 [ ] <-- vector loop bypass (may consist of multiple blocks).
1093 | [ ] <-- vector pre header.
1097 | [ ]_| <-- vector loop.
1100 >[ ] <--- middle-block.
1103 | [ ] <--- new preheader.
1107 | [ ]_| <-- old scalar loop to handle remainder.
1110 >[ ] <-- exit block.
1114 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1115 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1116 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1117 assert(ExitBlock && "Must have an exit block");
1119 // Some loops have a single integer induction variable, while other loops
1120 // don't. One example is c++ iterators that often have multiple pointer
1121 // induction variables. In the code below we also support a case where we
1122 // don't have a single induction variable.
1123 OldInduction = Legal->getInduction();
1124 Type *IdxTy = OldInduction ? OldInduction->getType() :
1125 DL->getIntPtrType(SE->getContext());
1127 // Find the loop boundaries.
1128 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1129 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1131 // Get the total trip count from the count by adding 1.
1132 ExitCount = SE->getAddExpr(ExitCount,
1133 SE->getConstant(ExitCount->getType(), 1));
1135 // Expand the trip count and place the new instructions in the preheader.
1136 // Notice that the pre-header does not change, only the loop body.
1137 SCEVExpander Exp(*SE, "induction");
1139 // Count holds the overall loop count (N).
1140 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1141 BypassBlock->getTerminator());
1143 // The loop index does not have to start at Zero. Find the original start
1144 // value from the induction PHI node. If we don't have an induction variable
1145 // then we know that it starts at zero.
1146 Value *StartIdx = OldInduction ?
1147 OldInduction->getIncomingValueForBlock(BypassBlock):
1148 ConstantInt::get(IdxTy, 0);
1150 assert(BypassBlock && "Invalid loop structure");
1151 LoopBypassBlocks.push_back(BypassBlock);
1153 // Split the single block loop into the two loop structure described above.
1154 BasicBlock *VectorPH =
1155 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1156 BasicBlock *VecBody =
1157 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1158 BasicBlock *MiddleBlock =
1159 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1160 BasicBlock *ScalarPH =
1161 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1163 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1165 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1167 // Generate the induction variable.
1168 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1169 // The loop step is equal to the vectorization factor (num of SIMD elements)
1170 // times the unroll factor (num of SIMD instructions).
1171 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1173 // This is the IR builder that we use to add all of the logic for bypassing
1174 // the new vector loop.
1175 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1177 // We may need to extend the index in case there is a type mismatch.
1178 // We know that the count starts at zero and does not overflow.
1179 if (Count->getType() != IdxTy) {
1180 // The exit count can be of pointer type. Convert it to the correct
1182 if (ExitCount->getType()->isPointerTy())
1183 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1185 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1188 // Add the start index to the loop count to get the new end index.
1189 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1191 // Now we need to generate the expression for N - (N % VF), which is
1192 // the part that the vectorized body will execute.
1193 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1194 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1195 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1196 "end.idx.rnd.down");
1198 // Now, compare the new count to zero. If it is zero skip the vector loop and
1199 // jump to the scalar loop.
1200 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1203 BasicBlock *LastBypassBlock = BypassBlock;
1205 // Generate the code that checks in runtime if arrays overlap. We put the
1206 // checks into a separate block to make the more common case of few elements
1208 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1209 BypassBlock->getTerminator());
1210 if (MemRuntimeCheck) {
1211 // Create a new block containing the memory check.
1212 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1214 LoopBypassBlocks.push_back(CheckBlock);
1216 // Replace the branch into the memory check block with a conditional branch
1217 // for the "few elements case".
1218 Instruction *OldTerm = BypassBlock->getTerminator();
1219 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1220 OldTerm->eraseFromParent();
1222 Cmp = MemRuntimeCheck;
1223 LastBypassBlock = CheckBlock;
1226 LastBypassBlock->getTerminator()->eraseFromParent();
1227 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1230 // We are going to resume the execution of the scalar loop.
1231 // Go over all of the induction variables that we found and fix the
1232 // PHIs that are left in the scalar version of the loop.
1233 // The starting values of PHI nodes depend on the counter of the last
1234 // iteration in the vectorized loop.
1235 // If we come from a bypass edge then we need to start from the original
1238 // This variable saves the new starting index for the scalar loop.
1239 PHINode *ResumeIndex = 0;
1240 LoopVectorizationLegality::InductionList::iterator I, E;
1241 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1242 for (I = List->begin(), E = List->end(); I != E; ++I) {
1243 PHINode *OrigPhi = I->first;
1244 LoopVectorizationLegality::InductionInfo II = I->second;
1245 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1246 MiddleBlock->getTerminator());
1247 Value *EndValue = 0;
1249 case LoopVectorizationLegality::IK_NoInduction:
1250 llvm_unreachable("Unknown induction");
1251 case LoopVectorizationLegality::IK_IntInduction: {
1252 // Handle the integer induction counter:
1253 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1254 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1255 // We know what the end value is.
1256 EndValue = IdxEndRoundDown;
1257 // We also know which PHI node holds it.
1258 ResumeIndex = ResumeVal;
1261 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1262 // Convert the CountRoundDown variable to the PHI size.
1263 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1264 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1265 Value *CRD = CountRoundDown;
1266 if (CRDSize > IISize)
1267 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1268 II.StartValue->getType(), "tr.crd",
1269 LoopBypassBlocks.back()->getTerminator());
1270 else if (CRDSize < IISize)
1271 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1272 II.StartValue->getType(),
1274 LoopBypassBlocks.back()->getTerminator());
1275 // Handle reverse integer induction counter:
1277 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1278 LoopBypassBlocks.back()->getTerminator());
1281 case LoopVectorizationLegality::IK_PtrInduction: {
1282 // For pointer induction variables, calculate the offset using
1285 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end",
1286 LoopBypassBlocks.back()->getTerminator());
1289 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1290 // The value at the end of the loop for the reverse pointer is calculated
1291 // by creating a GEP with a negative index starting from the start value.
1292 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1293 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown,
1295 LoopBypassBlocks.back()->getTerminator());
1296 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx,
1298 LoopBypassBlocks.back()->getTerminator());
1303 // The new PHI merges the original incoming value, in case of a bypass,
1304 // or the value at the end of the vectorized loop.
1305 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1306 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1307 ResumeVal->addIncoming(EndValue, VecBody);
1309 // Fix the scalar body counter (PHI node).
1310 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1311 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1314 // If we are generating a new induction variable then we also need to
1315 // generate the code that calculates the exit value. This value is not
1316 // simply the end of the counter because we may skip the vectorized body
1317 // in case of a runtime check.
1319 assert(!ResumeIndex && "Unexpected resume value found");
1320 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1321 MiddleBlock->getTerminator());
1322 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1323 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1324 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1327 // Make sure that we found the index where scalar loop needs to continue.
1328 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1329 "Invalid resume Index");
1331 // Add a check in the middle block to see if we have completed
1332 // all of the iterations in the first vector loop.
1333 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1334 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1335 ResumeIndex, "cmp.n",
1336 MiddleBlock->getTerminator());
1338 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1339 // Remove the old terminator.
1340 MiddleBlock->getTerminator()->eraseFromParent();
1342 // Create i+1 and fill the PHINode.
1343 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1344 Induction->addIncoming(StartIdx, VectorPH);
1345 Induction->addIncoming(NextIdx, VecBody);
1346 // Create the compare.
1347 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1348 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1350 // Now we have two terminators. Remove the old one from the block.
1351 VecBody->getTerminator()->eraseFromParent();
1353 // Get ready to start creating new instructions into the vectorized body.
1354 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1356 // Create and register the new vector loop.
1357 Loop* Lp = new Loop();
1358 Loop *ParentLoop = OrigLoop->getParentLoop();
1360 // Insert the new loop into the loop nest and register the new basic blocks.
1362 ParentLoop->addChildLoop(Lp);
1363 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1364 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1365 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1366 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1367 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1369 LI->addTopLevelLoop(Lp);
1372 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1375 LoopVectorPreHeader = VectorPH;
1376 LoopScalarPreHeader = ScalarPH;
1377 LoopMiddleBlock = MiddleBlock;
1378 LoopExitBlock = ExitBlock;
1379 LoopVectorBody = VecBody;
1380 LoopScalarBody = OldBasicBlock;
1383 /// This function returns the identity element (or neutral element) for
1384 /// the operation K.
1386 getReductionIdentity(LoopVectorizationLegality::ReductionKind K, Type *Tp) {
1388 case LoopVectorizationLegality:: RK_IntegerXor:
1389 case LoopVectorizationLegality:: RK_IntegerAdd:
1390 case LoopVectorizationLegality:: RK_IntegerOr:
1391 // Adding, Xoring, Oring zero to a number does not change it.
1392 return ConstantInt::get(Tp, 0);
1393 case LoopVectorizationLegality:: RK_IntegerMult:
1394 // Multiplying a number by 1 does not change it.
1395 return ConstantInt::get(Tp, 1);
1396 case LoopVectorizationLegality:: RK_IntegerAnd:
1397 // AND-ing a number with an all-1 value does not change it.
1398 return ConstantInt::get(Tp, -1, true);
1399 case LoopVectorizationLegality:: RK_FloatMult:
1400 // Multiplying a number by 1 does not change it.
1401 return ConstantFP::get(Tp, 1.0L);
1402 case LoopVectorizationLegality:: RK_FloatAdd:
1403 // Adding zero to a number does not change it.
1404 return ConstantFP::get(Tp, 0.0L);
1406 llvm_unreachable("Unknown reduction kind");
1411 isTriviallyVectorizableIntrinsic(Instruction *Inst) {
1412 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
1415 switch (II->getIntrinsicID()) {
1416 case Intrinsic::sqrt:
1417 case Intrinsic::sin:
1418 case Intrinsic::cos:
1419 case Intrinsic::exp:
1420 case Intrinsic::exp2:
1421 case Intrinsic::log:
1422 case Intrinsic::log10:
1423 case Intrinsic::log2:
1424 case Intrinsic::fabs:
1425 case Intrinsic::floor:
1426 case Intrinsic::ceil:
1427 case Intrinsic::trunc:
1428 case Intrinsic::rint:
1429 case Intrinsic::nearbyint:
1430 case Intrinsic::pow:
1431 case Intrinsic::fma:
1432 case Intrinsic::fmuladd:
1440 /// This function translates the reduction kind to an LLVM binary operator.
1441 static Instruction::BinaryOps
1442 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1444 case LoopVectorizationLegality::RK_IntegerAdd:
1445 return Instruction::Add;
1446 case LoopVectorizationLegality::RK_IntegerMult:
1447 return Instruction::Mul;
1448 case LoopVectorizationLegality::RK_IntegerOr:
1449 return Instruction::Or;
1450 case LoopVectorizationLegality::RK_IntegerAnd:
1451 return Instruction::And;
1452 case LoopVectorizationLegality::RK_IntegerXor:
1453 return Instruction::Xor;
1454 case LoopVectorizationLegality::RK_FloatMult:
1455 return Instruction::FMul;
1456 case LoopVectorizationLegality::RK_FloatAdd:
1457 return Instruction::FAdd;
1459 llvm_unreachable("Unknown reduction operation");
1464 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1465 //===------------------------------------------------===//
1467 // Notice: any optimization or new instruction that go
1468 // into the code below should be also be implemented in
1471 //===------------------------------------------------===//
1472 Constant *Zero = Builder.getInt32(0);
1474 // In order to support reduction variables we need to be able to vectorize
1475 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1476 // stages. First, we create a new vector PHI node with no incoming edges.
1477 // We use this value when we vectorize all of the instructions that use the
1478 // PHI. Next, after all of the instructions in the block are complete we
1479 // add the new incoming edges to the PHI. At this point all of the
1480 // instructions in the basic block are vectorized, so we can use them to
1481 // construct the PHI.
1482 PhiVector RdxPHIsToFix;
1484 // Scan the loop in a topological order to ensure that defs are vectorized
1486 LoopBlocksDFS DFS(OrigLoop);
1489 // Vectorize all of the blocks in the original loop.
1490 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1491 be = DFS.endRPO(); bb != be; ++bb)
1492 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1494 // At this point every instruction in the original loop is widened to
1495 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1496 // that we vectorized. The PHI nodes are currently empty because we did
1497 // not want to introduce cycles. Notice that the remaining PHI nodes
1498 // that we need to fix are reduction variables.
1500 // Create the 'reduced' values for each of the induction vars.
1501 // The reduced values are the vector values that we scalarize and combine
1502 // after the loop is finished.
1503 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1505 PHINode *RdxPhi = *it;
1506 assert(RdxPhi && "Unable to recover vectorized PHI");
1508 // Find the reduction variable descriptor.
1509 assert(Legal->getReductionVars()->count(RdxPhi) &&
1510 "Unable to find the reduction variable");
1511 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1512 (*Legal->getReductionVars())[RdxPhi];
1514 // We need to generate a reduction vector from the incoming scalar.
1515 // To do so, we need to generate the 'identity' vector and overide
1516 // one of the elements with the incoming scalar reduction. We need
1517 // to do it in the vector-loop preheader.
1518 Builder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1520 // This is the vector-clone of the value that leaves the loop.
1521 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1522 Type *VecTy = VectorExit[0]->getType();
1524 // Find the reduction identity variable. Zero for addition, or, xor,
1525 // one for multiplication, -1 for And.
1526 Constant *Iden = getReductionIdentity(RdxDesc.Kind, VecTy->getScalarType());
1527 Constant *Identity = ConstantVector::getSplat(VF, Iden);
1529 // This vector is the Identity vector where the first element is the
1530 // incoming scalar reduction.
1531 Value *VectorStart = Builder.CreateInsertElement(Identity,
1532 RdxDesc.StartValue, Zero);
1534 // Fix the vector-loop phi.
1535 // We created the induction variable so we know that the
1536 // preheader is the first entry.
1537 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1539 // Reductions do not have to start at zero. They can start with
1540 // any loop invariant values.
1541 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1542 BasicBlock *Latch = OrigLoop->getLoopLatch();
1543 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1544 VectorParts &Val = getVectorValue(LoopVal);
1545 for (unsigned part = 0; part < UF; ++part) {
1546 // Make sure to add the reduction stat value only to the
1547 // first unroll part.
1548 Value *StartVal = (part == 0) ? VectorStart : Identity;
1549 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1550 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1553 // Before each round, move the insertion point right between
1554 // the PHIs and the values we are going to write.
1555 // This allows us to write both PHINodes and the extractelement
1557 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1559 VectorParts RdxParts;
1560 for (unsigned part = 0; part < UF; ++part) {
1561 // This PHINode contains the vectorized reduction variable, or
1562 // the initial value vector, if we bypass the vector loop.
1563 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1564 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1565 Value *StartVal = (part == 0) ? VectorStart : Identity;
1566 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1567 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1568 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1569 RdxParts.push_back(NewPhi);
1572 // Reduce all of the unrolled parts into a single vector.
1573 Value *ReducedPartRdx = RdxParts[0];
1574 for (unsigned part = 1; part < UF; ++part) {
1575 Instruction::BinaryOps Op = getReductionBinOp(RdxDesc.Kind);
1576 ReducedPartRdx = Builder.CreateBinOp(Op, RdxParts[part], ReducedPartRdx,
1580 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1581 // and vector ops, reducing the set of values being computed by half each
1583 assert(isPowerOf2_32(VF) &&
1584 "Reduction emission only supported for pow2 vectors!");
1585 Value *TmpVec = ReducedPartRdx;
1586 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1587 for (unsigned i = VF; i != 1; i >>= 1) {
1588 // Move the upper half of the vector to the lower half.
1589 for (unsigned j = 0; j != i/2; ++j)
1590 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1592 // Fill the rest of the mask with undef.
1593 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1594 UndefValue::get(Builder.getInt32Ty()));
1597 Builder.CreateShuffleVector(TmpVec,
1598 UndefValue::get(TmpVec->getType()),
1599 ConstantVector::get(ShuffleMask),
1602 Instruction::BinaryOps Op = getReductionBinOp(RdxDesc.Kind);
1603 TmpVec = Builder.CreateBinOp(Op, TmpVec, Shuf, "bin.rdx");
1606 // The result is in the first element of the vector.
1607 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1609 // Now, we need to fix the users of the reduction variable
1610 // inside and outside of the scalar remainder loop.
1611 // We know that the loop is in LCSSA form. We need to update the
1612 // PHI nodes in the exit blocks.
1613 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1614 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1615 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1616 if (!LCSSAPhi) continue;
1618 // All PHINodes need to have a single entry edge, or two if
1619 // we already fixed them.
1620 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1622 // We found our reduction value exit-PHI. Update it with the
1623 // incoming bypass edge.
1624 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1625 // Add an edge coming from the bypass.
1626 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1629 }// end of the LCSSA phi scan.
1631 // Fix the scalar loop reduction variable with the incoming reduction sum
1632 // from the vector body and from the backedge value.
1633 int IncomingEdgeBlockIdx =
1634 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1635 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1636 // Pick the other block.
1637 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1638 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1639 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1640 }// end of for each redux variable.
1642 // The Loop exit block may have single value PHI nodes where the incoming
1643 // value is 'undef'. While vectorizing we only handled real values that
1644 // were defined inside the loop. Here we handle the 'undef case'.
1646 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1647 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1648 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1649 if (!LCSSAPhi) continue;
1650 if (LCSSAPhi->getNumIncomingValues() == 1)
1651 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1656 InnerLoopVectorizer::VectorParts
1657 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1658 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1661 VectorParts SrcMask = createBlockInMask(Src);
1663 // The terminator has to be a branch inst!
1664 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1665 assert(BI && "Unexpected terminator found");
1667 if (BI->isConditional()) {
1668 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1670 if (BI->getSuccessor(0) != Dst)
1671 for (unsigned part = 0; part < UF; ++part)
1672 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1674 for (unsigned part = 0; part < UF; ++part)
1675 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1682 InnerLoopVectorizer::VectorParts
1683 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1684 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1686 // Loop incoming mask is all-one.
1687 if (OrigLoop->getHeader() == BB) {
1688 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1689 return getVectorValue(C);
1692 // This is the block mask. We OR all incoming edges, and with zero.
1693 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1694 VectorParts BlockMask = getVectorValue(Zero);
1697 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1698 VectorParts EM = createEdgeMask(*it, BB);
1699 for (unsigned part = 0; part < UF; ++part)
1700 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1707 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1708 BasicBlock *BB, PhiVector *PV) {
1709 // For each instruction in the old loop.
1710 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1711 VectorParts &Entry = WidenMap.get(it);
1712 switch (it->getOpcode()) {
1713 case Instruction::Br:
1714 // Nothing to do for PHIs and BR, since we already took care of the
1715 // loop control flow instructions.
1717 case Instruction::PHI:{
1718 PHINode* P = cast<PHINode>(it);
1719 // Handle reduction variables:
1720 if (Legal->getReductionVars()->count(P)) {
1721 for (unsigned part = 0; part < UF; ++part) {
1722 // This is phase one of vectorizing PHIs.
1723 Type *VecTy = VectorType::get(it->getType(), VF);
1724 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1725 LoopVectorBody-> getFirstInsertionPt());
1731 // Check for PHI nodes that are lowered to vector selects.
1732 if (P->getParent() != OrigLoop->getHeader()) {
1733 // We know that all PHIs in non header blocks are converted into
1734 // selects, so we don't have to worry about the insertion order and we
1735 // can just use the builder.
1737 // At this point we generate the predication tree. There may be
1738 // duplications since this is a simple recursive scan, but future
1739 // optimizations will clean it up.
1740 VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
1743 for (unsigned part = 0; part < UF; ++part) {
1744 VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
1745 VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
1746 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
1752 // This PHINode must be an induction variable.
1753 // Make sure that we know about it.
1754 assert(Legal->getInductionVars()->count(P) &&
1755 "Not an induction variable");
1757 LoopVectorizationLegality::InductionInfo II =
1758 Legal->getInductionVars()->lookup(P);
1761 case LoopVectorizationLegality::IK_NoInduction:
1762 llvm_unreachable("Unknown induction");
1763 case LoopVectorizationLegality::IK_IntInduction: {
1764 assert(P == OldInduction && "Unexpected PHI");
1765 Value *Broadcasted = getBroadcastInstrs(Induction);
1766 // After broadcasting the induction variable we need to make the
1767 // vector consecutive by adding 0, 1, 2 ...
1768 for (unsigned part = 0; part < UF; ++part)
1769 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
1772 case LoopVectorizationLegality::IK_ReverseIntInduction:
1773 case LoopVectorizationLegality::IK_PtrInduction:
1774 case LoopVectorizationLegality::IK_ReversePtrInduction:
1775 // Handle reverse integer and pointer inductions.
1776 Value *StartIdx = 0;
1777 // If we have a single integer induction variable then use it.
1778 // Otherwise, start counting at zero.
1780 LoopVectorizationLegality::InductionInfo OldII =
1781 Legal->getInductionVars()->lookup(OldInduction);
1782 StartIdx = OldII.StartValue;
1784 StartIdx = ConstantInt::get(Induction->getType(), 0);
1786 // This is the normalized GEP that starts counting at zero.
1787 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1790 // Handle the reverse integer induction variable case.
1791 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
1792 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
1793 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
1795 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
1798 // This is a new value so do not hoist it out.
1799 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
1800 // After broadcasting the induction variable we need to make the
1801 // vector consecutive by adding ... -3, -2, -1, 0.
1802 for (unsigned part = 0; part < UF; ++part)
1803 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
1807 // Handle the pointer induction variable case.
1808 assert(P->getType()->isPointerTy() && "Unexpected type.");
1810 // Is this a reverse induction ptr or a consecutive induction ptr.
1811 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
1814 // This is the vector of results. Notice that we don't generate
1815 // vector geps because scalar geps result in better code.
1816 for (unsigned part = 0; part < UF; ++part) {
1817 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1818 for (unsigned int i = 0; i < VF; ++i) {
1819 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
1820 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
1823 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1825 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
1827 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
1829 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1830 Builder.getInt32(i),
1833 Entry[part] = VecVal;
1840 case Instruction::Add:
1841 case Instruction::FAdd:
1842 case Instruction::Sub:
1843 case Instruction::FSub:
1844 case Instruction::Mul:
1845 case Instruction::FMul:
1846 case Instruction::UDiv:
1847 case Instruction::SDiv:
1848 case Instruction::FDiv:
1849 case Instruction::URem:
1850 case Instruction::SRem:
1851 case Instruction::FRem:
1852 case Instruction::Shl:
1853 case Instruction::LShr:
1854 case Instruction::AShr:
1855 case Instruction::And:
1856 case Instruction::Or:
1857 case Instruction::Xor: {
1858 // Just widen binops.
1859 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
1860 VectorParts &A = getVectorValue(it->getOperand(0));
1861 VectorParts &B = getVectorValue(it->getOperand(1));
1863 // Use this vector value for all users of the original instruction.
1864 for (unsigned Part = 0; Part < UF; ++Part) {
1865 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
1867 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
1868 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
1869 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
1870 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1871 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1873 if (VecOp && isa<PossiblyExactOperator>(VecOp))
1874 VecOp->setIsExact(BinOp->isExact());
1880 case Instruction::Select: {
1882 // If the selector is loop invariant we can create a select
1883 // instruction with a scalar condition. Otherwise, use vector-select.
1884 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
1887 // The condition can be loop invariant but still defined inside the
1888 // loop. This means that we can't just use the original 'cond' value.
1889 // We have to take the 'vectorized' value and pick the first lane.
1890 // Instcombine will make this a no-op.
1891 VectorParts &Cond = getVectorValue(it->getOperand(0));
1892 VectorParts &Op0 = getVectorValue(it->getOperand(1));
1893 VectorParts &Op1 = getVectorValue(it->getOperand(2));
1894 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
1895 Builder.getInt32(0));
1896 for (unsigned Part = 0; Part < UF; ++Part) {
1897 Entry[Part] = Builder.CreateSelect(
1898 InvariantCond ? ScalarCond : Cond[Part],
1905 case Instruction::ICmp:
1906 case Instruction::FCmp: {
1907 // Widen compares. Generate vector compares.
1908 bool FCmp = (it->getOpcode() == Instruction::FCmp);
1909 CmpInst *Cmp = dyn_cast<CmpInst>(it);
1910 VectorParts &A = getVectorValue(it->getOperand(0));
1911 VectorParts &B = getVectorValue(it->getOperand(1));
1912 for (unsigned Part = 0; Part < UF; ++Part) {
1915 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
1917 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
1923 case Instruction::Store:
1924 case Instruction::Load:
1925 vectorizeMemoryInstruction(it, Legal);
1927 case Instruction::ZExt:
1928 case Instruction::SExt:
1929 case Instruction::FPToUI:
1930 case Instruction::FPToSI:
1931 case Instruction::FPExt:
1932 case Instruction::PtrToInt:
1933 case Instruction::IntToPtr:
1934 case Instruction::SIToFP:
1935 case Instruction::UIToFP:
1936 case Instruction::Trunc:
1937 case Instruction::FPTrunc:
1938 case Instruction::BitCast: {
1939 CastInst *CI = dyn_cast<CastInst>(it);
1940 /// Optimize the special case where the source is the induction
1941 /// variable. Notice that we can only optimize the 'trunc' case
1942 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
1943 /// c. other casts depend on pointer size.
1944 if (CI->getOperand(0) == OldInduction &&
1945 it->getOpcode() == Instruction::Trunc) {
1946 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
1948 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
1949 for (unsigned Part = 0; Part < UF; ++Part)
1950 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
1953 /// Vectorize casts.
1954 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1956 VectorParts &A = getVectorValue(it->getOperand(0));
1957 for (unsigned Part = 0; Part < UF; ++Part)
1958 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
1962 case Instruction::Call: {
1963 assert(isTriviallyVectorizableIntrinsic(it));
1964 Module *M = BB->getParent()->getParent();
1965 IntrinsicInst *II = cast<IntrinsicInst>(it);
1966 Intrinsic::ID ID = II->getIntrinsicID();
1967 for (unsigned Part = 0; Part < UF; ++Part) {
1968 SmallVector<Value*, 4> Args;
1969 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) {
1970 VectorParts &Arg = getVectorValue(II->getArgOperand(i));
1971 Args.push_back(Arg[Part]);
1973 Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) };
1974 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
1975 Entry[Part] = Builder.CreateCall(F, Args);
1981 // All other instructions are unsupported. Scalarize them.
1982 scalarizeInstruction(it);
1985 }// end of for_each instr.
1988 void InnerLoopVectorizer::updateAnalysis() {
1989 // Forget the original basic block.
1990 SE->forgetLoop(OrigLoop);
1992 // Update the dominator tree information.
1993 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
1994 "Entry does not dominate exit.");
1996 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1997 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
1998 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
1999 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2000 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2001 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2002 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2003 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2005 DEBUG(DT->verifyAnalysis());
2008 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2009 if (!EnableIfConversion)
2012 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2013 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2015 // Collect the blocks that need predication.
2016 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2017 BasicBlock *BB = LoopBlocks[i];
2019 // We don't support switch statements inside loops.
2020 if (!isa<BranchInst>(BB->getTerminator()))
2023 // We must have at most two predecessors because we need to convert
2024 // all PHIs to selects.
2025 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
2029 // We must be able to predicate all blocks that need to be predicated.
2030 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2034 // We can if-convert this loop.
2038 bool LoopVectorizationLegality::canVectorize() {
2039 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2041 // We can only vectorize innermost loops.
2042 if (TheLoop->getSubLoopsVector().size())
2045 // We must have a single backedge.
2046 if (TheLoop->getNumBackEdges() != 1)
2049 // We must have a single exiting block.
2050 if (!TheLoop->getExitingBlock())
2053 unsigned NumBlocks = TheLoop->getNumBlocks();
2055 // Check if we can if-convert non single-bb loops.
2056 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2057 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2061 // We need to have a loop header.
2062 BasicBlock *Latch = TheLoop->getLoopLatch();
2063 DEBUG(dbgs() << "LV: Found a loop: " <<
2064 TheLoop->getHeader()->getName() << "\n");
2066 // ScalarEvolution needs to be able to find the exit count.
2067 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2068 if (ExitCount == SE->getCouldNotCompute()) {
2069 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2073 // Do not loop-vectorize loops with a tiny trip count.
2074 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2075 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2076 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2077 "This loop is not worth vectorizing.\n");
2081 // Check if we can vectorize the instructions and CFG in this loop.
2082 if (!canVectorizeInstrs()) {
2083 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2087 // Go over each instruction and look at memory deps.
2088 if (!canVectorizeMemory()) {
2089 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2093 // Collect all of the variables that remain uniform after vectorization.
2094 collectLoopUniforms();
2096 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2097 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2100 // Okay! We can vectorize. At this point we don't have any other mem analysis
2101 // which may limit our maximum vectorization factor, so just return true with
2106 bool LoopVectorizationLegality::canVectorizeInstrs() {
2107 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2108 BasicBlock *Header = TheLoop->getHeader();
2110 // For each block in the loop.
2111 for (Loop::block_iterator bb = TheLoop->block_begin(),
2112 be = TheLoop->block_end(); bb != be; ++bb) {
2114 // Scan the instructions in the block and look for hazards.
2115 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2118 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2119 // This should not happen because the loop should be normalized.
2120 if (Phi->getNumIncomingValues() != 2) {
2121 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2125 // Check that this PHI type is allowed.
2126 if (!Phi->getType()->isIntegerTy() &&
2127 !Phi->getType()->isFloatingPointTy() &&
2128 !Phi->getType()->isPointerTy()) {
2129 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2133 // If this PHINode is not in the header block, then we know that we
2134 // can convert it to select during if-conversion. No need to check if
2135 // the PHIs in this block are induction or reduction variables.
2139 // This is the value coming from the preheader.
2140 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2141 // Check if this is an induction variable.
2142 InductionKind IK = isInductionVariable(Phi);
2144 if (IK_NoInduction != IK) {
2145 // Int inductions are special because we only allow one IV.
2146 if (IK == IK_IntInduction) {
2148 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2154 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2155 Inductions[Phi] = InductionInfo(StartValue, IK);
2159 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2160 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2163 if (AddReductionVar(Phi, RK_IntegerMult)) {
2164 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2167 if (AddReductionVar(Phi, RK_IntegerOr)) {
2168 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2171 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2172 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2175 if (AddReductionVar(Phi, RK_IntegerXor)) {
2176 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2179 if (AddReductionVar(Phi, RK_FloatMult)) {
2180 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2183 if (AddReductionVar(Phi, RK_FloatAdd)) {
2184 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2188 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2190 }// end of PHI handling
2192 // We still don't handle functions.
2193 CallInst *CI = dyn_cast<CallInst>(it);
2194 if (CI && !isTriviallyVectorizableIntrinsic(it)) {
2195 DEBUG(dbgs() << "LV: Found a call site.\n");
2199 // Check that the instruction return type is vectorizable.
2200 if (!VectorType::isValidElementType(it->getType()) &&
2201 !it->getType()->isVoidTy()) {
2202 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2206 // Check that the stored type is vectorizable.
2207 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2208 Type *T = ST->getValueOperand()->getType();
2209 if (!VectorType::isValidElementType(T))
2213 // Reduction instructions are allowed to have exit users.
2214 // All other instructions must not have external users.
2215 if (!AllowedExit.count(it))
2216 //Check that all of the users of the loop are inside the BB.
2217 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2219 Instruction *U = cast<Instruction>(*I);
2220 // This user may be a reduction exit value.
2221 if (!TheLoop->contains(U)) {
2222 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2231 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2232 assert(getInductionVars()->size() && "No induction variables");
2238 void LoopVectorizationLegality::collectLoopUniforms() {
2239 // We now know that the loop is vectorizable!
2240 // Collect variables that will remain uniform after vectorization.
2241 std::vector<Value*> Worklist;
2242 BasicBlock *Latch = TheLoop->getLoopLatch();
2244 // Start with the conditional branch and walk up the block.
2245 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2247 while (Worklist.size()) {
2248 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2249 Worklist.pop_back();
2251 // Look at instructions inside this loop.
2252 // Stop when reaching PHI nodes.
2253 // TODO: we need to follow values all over the loop, not only in this block.
2254 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2257 // This is a known uniform.
2260 // Insert all operands.
2261 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2262 Worklist.push_back(I->getOperand(i));
2267 bool LoopVectorizationLegality::canVectorizeMemory() {
2268 typedef SmallVector<Value*, 16> ValueVector;
2269 typedef SmallPtrSet<Value*, 16> ValueSet;
2270 // Holds the Load and Store *instructions*.
2273 PtrRtCheck.Pointers.clear();
2274 PtrRtCheck.Need = false;
2277 for (Loop::block_iterator bb = TheLoop->block_begin(),
2278 be = TheLoop->block_end(); bb != be; ++bb) {
2280 // Scan the BB and collect legal loads and stores.
2281 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2284 // If this is a load, save it. If this instruction can read from memory
2285 // but is not a load, then we quit. Notice that we don't handle function
2286 // calls that read or write.
2287 if (it->mayReadFromMemory()) {
2288 LoadInst *Ld = dyn_cast<LoadInst>(it);
2289 if (!Ld) return false;
2290 if (!Ld->isSimple()) {
2291 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2294 Loads.push_back(Ld);
2298 // Save 'store' instructions. Abort if other instructions write to memory.
2299 if (it->mayWriteToMemory()) {
2300 StoreInst *St = dyn_cast<StoreInst>(it);
2301 if (!St) return false;
2302 if (!St->isSimple()) {
2303 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2306 Stores.push_back(St);
2311 // Now we have two lists that hold the loads and the stores.
2312 // Next, we find the pointers that they use.
2314 // Check if we see any stores. If there are no stores, then we don't
2315 // care if the pointers are *restrict*.
2316 if (!Stores.size()) {
2317 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2321 // Holds the read and read-write *pointers* that we find.
2323 ValueVector ReadWrites;
2325 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2326 // multiple times on the same object. If the ptr is accessed twice, once
2327 // for read and once for write, it will only appear once (on the write
2328 // list). This is okay, since we are going to check for conflicts between
2329 // writes and between reads and writes, but not between reads and reads.
2332 ValueVector::iterator I, IE;
2333 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2334 StoreInst *ST = cast<StoreInst>(*I);
2335 Value* Ptr = ST->getPointerOperand();
2337 if (isUniform(Ptr)) {
2338 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2342 // If we did *not* see this pointer before, insert it to
2343 // the read-write list. At this phase it is only a 'write' list.
2344 if (Seen.insert(Ptr))
2345 ReadWrites.push_back(Ptr);
2348 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2349 LoadInst *LD = cast<LoadInst>(*I);
2350 Value* Ptr = LD->getPointerOperand();
2351 // If we did *not* see this pointer before, insert it to the
2352 // read list. If we *did* see it before, then it is already in
2353 // the read-write list. This allows us to vectorize expressions
2354 // such as A[i] += x; Because the address of A[i] is a read-write
2355 // pointer. This only works if the index of A[i] is consecutive.
2356 // If the address of i is unknown (for example A[B[i]]) then we may
2357 // read a few words, modify, and write a few words, and some of the
2358 // words may be written to the same address.
2359 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2360 Reads.push_back(Ptr);
2363 // If we write (or read-write) to a single destination and there are no
2364 // other reads in this loop then is it safe to vectorize.
2365 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2366 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2370 // Find pointers with computable bounds. We are going to use this information
2371 // to place a runtime bound check.
2372 bool CanDoRT = true;
2373 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
2374 if (hasComputableBounds(*I)) {
2375 PtrRtCheck.insert(SE, TheLoop, *I);
2376 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
2381 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
2382 if (hasComputableBounds(*I)) {
2383 PtrRtCheck.insert(SE, TheLoop, *I);
2384 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
2390 // Check that we did not collect too many pointers or found a
2391 // unsizeable pointer.
2392 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
2398 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2401 bool NeedRTCheck = false;
2403 // Now that the pointers are in two lists (Reads and ReadWrites), we
2404 // can check that there are no conflicts between each of the writes and
2405 // between the writes to the reads.
2406 ValueSet WriteObjects;
2407 ValueVector TempObjects;
2409 // Check that the read-writes do not conflict with other read-write
2411 bool AllWritesIdentified = true;
2412 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
2413 GetUnderlyingObjects(*I, TempObjects, DL);
2414 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
2416 if (!isIdentifiedObject(*it)) {
2417 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
2419 AllWritesIdentified = false;
2421 if (!WriteObjects.insert(*it)) {
2422 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2427 TempObjects.clear();
2430 /// Check that the reads don't conflict with the read-writes.
2431 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
2432 GetUnderlyingObjects(*I, TempObjects, DL);
2433 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
2435 // If all of the writes are identified then we don't care if the read
2436 // pointer is identified or not.
2437 if (!AllWritesIdentified && !isIdentifiedObject(*it)) {
2438 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
2441 if (WriteObjects.count(*it)) {
2442 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
2447 TempObjects.clear();
2450 PtrRtCheck.Need = NeedRTCheck;
2451 if (NeedRTCheck && !CanDoRT) {
2452 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2453 "the array bounds.\n");
2458 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2459 " need a runtime memory check.\n");
2463 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2464 ReductionKind Kind) {
2465 if (Phi->getNumIncomingValues() != 2)
2468 // Reduction variables are only found in the loop header block.
2469 if (Phi->getParent() != TheLoop->getHeader())
2472 // Obtain the reduction start value from the value that comes from the loop
2474 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2476 // ExitInstruction is the single value which is used outside the loop.
2477 // We only allow for a single reduction value to be used outside the loop.
2478 // This includes users of the reduction, variables (which form a cycle
2479 // which ends in the phi node).
2480 Instruction *ExitInstruction = 0;
2481 // Indicates that we found a binary operation in our scan.
2482 bool FoundBinOp = false;
2484 // Iter is our iterator. We start with the PHI node and scan for all of the
2485 // users of this instruction. All users must be instructions that can be
2486 // used as reduction variables (such as ADD). We may have a single
2487 // out-of-block user. The cycle must end with the original PHI.
2488 Instruction *Iter = Phi;
2490 // If the instruction has no users then this is a broken
2491 // chain and can't be a reduction variable.
2492 if (Iter->use_empty())
2495 // Did we find a user inside this loop already ?
2496 bool FoundInBlockUser = false;
2497 // Did we reach the initial PHI node already ?
2498 bool FoundStartPHI = false;
2500 // Is this a bin op ?
2501 FoundBinOp |= !isa<PHINode>(Iter);
2503 // For each of the *users* of iter.
2504 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2506 Instruction *U = cast<Instruction>(*it);
2507 // We already know that the PHI is a user.
2509 FoundStartPHI = true;
2513 // Check if we found the exit user.
2514 BasicBlock *Parent = U->getParent();
2515 if (!TheLoop->contains(Parent)) {
2516 // Exit if you find multiple outside users.
2517 if (ExitInstruction != 0)
2519 ExitInstruction = Iter;
2522 // We allow in-loop PHINodes which are not the original reduction PHI
2523 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2524 // structure) then don't skip this PHI.
2525 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2526 U->getParent() != TheLoop->getHeader() &&
2527 TheLoop->contains(U) &&
2528 Iter->getNumUses() > 1)
2531 // We can't have multiple inside users.
2532 if (FoundInBlockUser)
2534 FoundInBlockUser = true;
2536 // Any reduction instr must be of one of the allowed kinds.
2537 if (!isReductionInstr(U, Kind))
2540 // Reductions of instructions such as Div, and Sub is only
2541 // possible if the LHS is the reduction variable.
2542 if (!U->isCommutative() && !isa<PHINode>(U) && U->getOperand(0) != Iter)
2548 // We found a reduction var if we have reached the original
2549 // phi node and we only have a single instruction with out-of-loop
2551 if (FoundStartPHI) {
2552 // This instruction is allowed to have out-of-loop users.
2553 AllowedExit.insert(ExitInstruction);
2555 // Save the description of this reduction variable.
2556 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
2557 Reductions[Phi] = RD;
2558 // We've ended the cycle. This is a reduction variable if we have an
2559 // outside user and it has a binary op.
2560 return FoundBinOp && ExitInstruction;
2566 LoopVectorizationLegality::isReductionInstr(Instruction *I,
2567 ReductionKind Kind) {
2568 bool FP = I->getType()->isFloatingPointTy();
2569 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
2571 switch (I->getOpcode()) {
2574 case Instruction::PHI:
2575 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd))
2579 case Instruction::Sub:
2580 case Instruction::Add:
2581 return Kind == RK_IntegerAdd;
2582 case Instruction::SDiv:
2583 case Instruction::UDiv:
2584 case Instruction::Mul:
2585 return Kind == RK_IntegerMult;
2586 case Instruction::And:
2587 return Kind == RK_IntegerAnd;
2588 case Instruction::Or:
2589 return Kind == RK_IntegerOr;
2590 case Instruction::Xor:
2591 return Kind == RK_IntegerXor;
2592 case Instruction::FMul:
2593 return Kind == RK_FloatMult && FastMath;
2594 case Instruction::FAdd:
2595 return Kind == RK_FloatAdd && FastMath;
2599 LoopVectorizationLegality::InductionKind
2600 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
2601 Type *PhiTy = Phi->getType();
2602 // We only handle integer and pointer inductions variables.
2603 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
2604 return IK_NoInduction;
2606 // Check that the PHI is consecutive.
2607 const SCEV *PhiScev = SE->getSCEV(Phi);
2608 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2610 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
2611 return IK_NoInduction;
2613 const SCEV *Step = AR->getStepRecurrence(*SE);
2615 // Integer inductions need to have a stride of one.
2616 if (PhiTy->isIntegerTy()) {
2618 return IK_IntInduction;
2619 if (Step->isAllOnesValue())
2620 return IK_ReverseIntInduction;
2621 return IK_NoInduction;
2624 // Calculate the pointer stride and check if it is consecutive.
2625 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
2627 return IK_NoInduction;
2629 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
2630 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
2631 if (C->getValue()->equalsInt(Size))
2632 return IK_PtrInduction;
2633 else if (C->getValue()->equalsInt(0 - Size))
2634 return IK_ReversePtrInduction;
2636 return IK_NoInduction;
2639 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
2640 Value *In0 = const_cast<Value*>(V);
2641 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
2645 return Inductions.count(PN);
2648 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
2649 assert(TheLoop->contains(BB) && "Unknown block used");
2651 // Blocks that do not dominate the latch need predication.
2652 BasicBlock* Latch = TheLoop->getLoopLatch();
2653 return !DT->dominates(BB, Latch);
2656 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
2657 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2658 // We don't predicate loads/stores at the moment.
2659 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
2662 // The instructions below can trap.
2663 switch (it->getOpcode()) {
2665 case Instruction::UDiv:
2666 case Instruction::SDiv:
2667 case Instruction::URem:
2668 case Instruction::SRem:
2676 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
2677 const SCEV *PhiScev = SE->getSCEV(Ptr);
2678 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2682 return AR->isAffine();
2685 LoopVectorizationCostModel::VectorizationFactor
2686 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
2688 // Width 1 means no vectorize
2689 VectorizationFactor Factor = { 1U, 0U };
2690 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
2691 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
2695 // Find the trip count.
2696 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
2697 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
2699 unsigned WidestType = getWidestType();
2700 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
2701 unsigned MaxVectorSize = WidestRegister / WidestType;
2702 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
2703 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
2705 if (MaxVectorSize == 0) {
2706 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
2710 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
2711 " into one vector!");
2713 unsigned VF = MaxVectorSize;
2715 // If we optimize the program for size, avoid creating the tail loop.
2717 // If we are unable to calculate the trip count then don't try to vectorize.
2719 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2723 // Find the maximum SIMD width that can fit within the trip count.
2724 VF = TC % MaxVectorSize;
2729 // If the trip count that we found modulo the vectorization factor is not
2730 // zero then we require a tail.
2732 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2738 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
2739 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
2741 Factor.Width = UserVF;
2745 float Cost = expectedCost(1);
2747 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
2748 for (unsigned i=2; i <= VF; i*=2) {
2749 // Notice that the vector loop needs to be executed less times, so
2750 // we need to divide the cost of the vector loops by the width of
2751 // the vector elements.
2752 float VectorCost = expectedCost(i) / (float)i;
2753 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
2754 (int)VectorCost << ".\n");
2755 if (VectorCost < Cost) {
2761 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
2762 Factor.Width = Width;
2763 Factor.Cost = Width * Cost;
2767 unsigned LoopVectorizationCostModel::getWidestType() {
2768 unsigned MaxWidth = 8;
2771 for (Loop::block_iterator bb = TheLoop->block_begin(),
2772 be = TheLoop->block_end(); bb != be; ++bb) {
2773 BasicBlock *BB = *bb;
2775 // For each instruction in the loop.
2776 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2777 Type *T = it->getType();
2779 // Only examine Loads, Stores and PHINodes.
2780 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
2783 // Examine PHI nodes that are reduction variables.
2784 if (PHINode *PN = dyn_cast<PHINode>(it))
2785 if (!Legal->getReductionVars()->count(PN))
2788 // Examine the stored values.
2789 if (StoreInst *ST = dyn_cast<StoreInst>(it))
2790 T = ST->getValueOperand()->getType();
2792 // Ignore stored/loaded pointer types.
2793 if (T->isPointerTy())
2796 MaxWidth = std::max(MaxWidth, T->getScalarSizeInBits());
2804 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
2807 unsigned LoopCost) {
2809 // -- The unroll heuristics --
2810 // We unroll the loop in order to expose ILP and reduce the loop overhead.
2811 // There are many micro-architectural considerations that we can't predict
2812 // at this level. For example frontend pressure (on decode or fetch) due to
2813 // code size, or the number and capabilities of the execution ports.
2815 // We use the following heuristics to select the unroll factor:
2816 // 1. If the code has reductions the we unroll in order to break the cross
2817 // iteration dependency.
2818 // 2. If the loop is really small then we unroll in order to reduce the loop
2820 // 3. We don't unroll if we think that we will spill registers to memory due
2821 // to the increased register pressure.
2823 // Use the user preference, unless 'auto' is selected.
2827 // When we optimize for size we don't unroll.
2831 // Do not unroll loops with a relatively small trip count.
2832 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
2833 TheLoop->getLoopLatch());
2834 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
2837 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
2838 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
2839 " vector registers\n");
2841 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
2842 // We divide by these constants so assume that we have at least one
2843 // instruction that uses at least one register.
2844 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
2845 R.NumInstructions = std::max(R.NumInstructions, 1U);
2847 // We calculate the unroll factor using the following formula.
2848 // Subtract the number of loop invariants from the number of available
2849 // registers. These registers are used by all of the unrolled instances.
2850 // Next, divide the remaining registers by the number of registers that is
2851 // required by the loop, in order to estimate how many parallel instances
2852 // fit without causing spills.
2853 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
2855 // Clamp the unroll factor ranges to reasonable factors.
2856 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
2858 // If we did not calculate the cost for VF (because the user selected the VF)
2859 // then we calculate the cost of VF here.
2861 LoopCost = expectedCost(VF);
2863 // Clamp the calculated UF to be between the 1 and the max unroll factor
2864 // that the target allows.
2865 if (UF > MaxUnrollSize)
2870 if (Legal->getReductionVars()->size()) {
2871 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
2875 // We want to unroll tiny loops in order to reduce the loop overhead.
2876 // We assume that the cost overhead is 1 and we use the cost model
2877 // to estimate the cost of the loop and unroll until the cost of the
2878 // loop overhead is about 5% of the cost of the loop.
2879 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
2880 if (LoopCost < 20) {
2881 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
2882 unsigned NewUF = 20/LoopCost + 1;
2883 return std::min(NewUF, UF);
2886 DEBUG(dbgs() << "LV: Not Unrolling. \n");
2890 LoopVectorizationCostModel::RegisterUsage
2891 LoopVectorizationCostModel::calculateRegisterUsage() {
2892 // This function calculates the register usage by measuring the highest number
2893 // of values that are alive at a single location. Obviously, this is a very
2894 // rough estimation. We scan the loop in a topological order in order and
2895 // assign a number to each instruction. We use RPO to ensure that defs are
2896 // met before their users. We assume that each instruction that has in-loop
2897 // users starts an interval. We record every time that an in-loop value is
2898 // used, so we have a list of the first and last occurrences of each
2899 // instruction. Next, we transpose this data structure into a multi map that
2900 // holds the list of intervals that *end* at a specific location. This multi
2901 // map allows us to perform a linear search. We scan the instructions linearly
2902 // and record each time that a new interval starts, by placing it in a set.
2903 // If we find this value in the multi-map then we remove it from the set.
2904 // The max register usage is the maximum size of the set.
2905 // We also search for instructions that are defined outside the loop, but are
2906 // used inside the loop. We need this number separately from the max-interval
2907 // usage number because when we unroll, loop-invariant values do not take
2909 LoopBlocksDFS DFS(TheLoop);
2913 R.NumInstructions = 0;
2915 // Each 'key' in the map opens a new interval. The values
2916 // of the map are the index of the 'last seen' usage of the
2917 // instruction that is the key.
2918 typedef DenseMap<Instruction*, unsigned> IntervalMap;
2919 // Maps instruction to its index.
2920 DenseMap<unsigned, Instruction*> IdxToInstr;
2921 // Marks the end of each interval.
2922 IntervalMap EndPoint;
2923 // Saves the list of instruction indices that are used in the loop.
2924 SmallSet<Instruction*, 8> Ends;
2925 // Saves the list of values that are used in the loop but are
2926 // defined outside the loop, such as arguments and constants.
2927 SmallPtrSet<Value*, 8> LoopInvariants;
2930 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2931 be = DFS.endRPO(); bb != be; ++bb) {
2932 R.NumInstructions += (*bb)->size();
2933 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2935 Instruction *I = it;
2936 IdxToInstr[Index++] = I;
2938 // Save the end location of each USE.
2939 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
2940 Value *U = I->getOperand(i);
2941 Instruction *Instr = dyn_cast<Instruction>(U);
2943 // Ignore non-instruction values such as arguments, constants, etc.
2944 if (!Instr) continue;
2946 // If this instruction is outside the loop then record it and continue.
2947 if (!TheLoop->contains(Instr)) {
2948 LoopInvariants.insert(Instr);
2952 // Overwrite previous end points.
2953 EndPoint[Instr] = Index;
2959 // Saves the list of intervals that end with the index in 'key'.
2960 typedef SmallVector<Instruction*, 2> InstrList;
2961 DenseMap<unsigned, InstrList> TransposeEnds;
2963 // Transpose the EndPoints to a list of values that end at each index.
2964 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
2966 TransposeEnds[it->second].push_back(it->first);
2968 SmallSet<Instruction*, 8> OpenIntervals;
2969 unsigned MaxUsage = 0;
2972 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
2973 for (unsigned int i = 0; i < Index; ++i) {
2974 Instruction *I = IdxToInstr[i];
2975 // Ignore instructions that are never used within the loop.
2976 if (!Ends.count(I)) continue;
2978 // Remove all of the instructions that end at this location.
2979 InstrList &List = TransposeEnds[i];
2980 for (unsigned int j=0, e = List.size(); j < e; ++j)
2981 OpenIntervals.erase(List[j]);
2983 // Count the number of live interals.
2984 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
2986 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
2987 OpenIntervals.size() <<"\n");
2989 // Add the current instruction to the list of open intervals.
2990 OpenIntervals.insert(I);
2993 unsigned Invariant = LoopInvariants.size();
2994 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
2995 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
2996 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
2998 R.LoopInvariantRegs = Invariant;
2999 R.MaxLocalUsers = MaxUsage;
3003 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3007 for (Loop::block_iterator bb = TheLoop->block_begin(),
3008 be = TheLoop->block_end(); bb != be; ++bb) {
3009 unsigned BlockCost = 0;
3010 BasicBlock *BB = *bb;
3012 // For each instruction in the old loop.
3013 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3014 unsigned C = getInstructionCost(it, VF);
3016 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3017 VF << " For instruction: "<< *it << "\n");
3020 // We assume that if-converted blocks have a 50% chance of being executed.
3021 // When the code is scalar then some of the blocks are avoided due to CF.
3022 // When the code is vectorized we execute all code paths.
3023 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3033 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3034 // If we know that this instruction will remain uniform, check the cost of
3035 // the scalar version.
3036 if (Legal->isUniformAfterVectorization(I))
3039 Type *RetTy = I->getType();
3040 Type *VectorTy = ToVectorTy(RetTy, VF);
3042 // TODO: We need to estimate the cost of intrinsic calls.
3043 switch (I->getOpcode()) {
3044 case Instruction::GetElementPtr:
3045 // We mark this instruction as zero-cost because scalar GEPs are usually
3046 // lowered to the intruction addressing mode. At the moment we don't
3047 // generate vector geps.
3049 case Instruction::Br: {
3050 return TTI.getCFInstrCost(I->getOpcode());
3052 case Instruction::PHI:
3053 //TODO: IF-converted IFs become selects.
3055 case Instruction::Add:
3056 case Instruction::FAdd:
3057 case Instruction::Sub:
3058 case Instruction::FSub:
3059 case Instruction::Mul:
3060 case Instruction::FMul:
3061 case Instruction::UDiv:
3062 case Instruction::SDiv:
3063 case Instruction::FDiv:
3064 case Instruction::URem:
3065 case Instruction::SRem:
3066 case Instruction::FRem:
3067 case Instruction::Shl:
3068 case Instruction::LShr:
3069 case Instruction::AShr:
3070 case Instruction::And:
3071 case Instruction::Or:
3072 case Instruction::Xor:
3073 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy);
3074 case Instruction::Select: {
3075 SelectInst *SI = cast<SelectInst>(I);
3076 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3077 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3078 Type *CondTy = SI->getCondition()->getType();
3080 CondTy = VectorType::get(CondTy, VF);
3082 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3084 case Instruction::ICmp:
3085 case Instruction::FCmp: {
3086 Type *ValTy = I->getOperand(0)->getType();
3087 VectorTy = ToVectorTy(ValTy, VF);
3088 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3090 case Instruction::Store: {
3091 StoreInst *SI = cast<StoreInst>(I);
3092 Type *ValTy = SI->getValueOperand()->getType();
3093 VectorTy = ToVectorTy(ValTy, VF);
3096 return TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
3098 SI->getPointerAddressSpace());
3100 // Scalarized stores.
3101 int Stride = Legal->isConsecutivePtr(SI->getPointerOperand());
3102 bool Reverse = Stride < 0;
3106 // The cost of extracting from the value vector and pointer vector.
3107 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3108 for (unsigned i = 0; i < VF; ++i) {
3109 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy,
3111 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3114 // The cost of the scalar stores.
3115 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3117 SI->getPointerAddressSpace());
3122 unsigned Cost = TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
3124 SI->getPointerAddressSpace());
3126 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3130 case Instruction::Load: {
3131 LoadInst *LI = cast<LoadInst>(I);
3134 return TTI.getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
3135 LI->getPointerAddressSpace());
3137 // Scalarized loads.
3138 int Stride = Legal->isConsecutivePtr(LI->getPointerOperand());
3139 bool Reverse = Stride < 0;
3142 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3144 // The cost of extracting from the pointer vector.
3145 for (unsigned i = 0; i < VF; ++i)
3146 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3148 // The cost of inserting data to the result vector.
3149 for (unsigned i = 0; i < VF; ++i)
3150 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, VectorTy, i);
3152 // The cost of the scalar stores.
3153 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), RetTy->getScalarType(),
3155 LI->getPointerAddressSpace());
3160 unsigned Cost = TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
3162 LI->getPointerAddressSpace());
3164 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
3167 case Instruction::ZExt:
3168 case Instruction::SExt:
3169 case Instruction::FPToUI:
3170 case Instruction::FPToSI:
3171 case Instruction::FPExt:
3172 case Instruction::PtrToInt:
3173 case Instruction::IntToPtr:
3174 case Instruction::SIToFP:
3175 case Instruction::UIToFP:
3176 case Instruction::Trunc:
3177 case Instruction::FPTrunc:
3178 case Instruction::BitCast: {
3179 // We optimize the truncation of induction variable.
3180 // The cost of these is the same as the scalar operation.
3181 if (I->getOpcode() == Instruction::Trunc &&
3182 Legal->isInductionVariable(I->getOperand(0)))
3183 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3184 I->getOperand(0)->getType());
3186 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3187 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3189 case Instruction::Call: {
3190 assert(isTriviallyVectorizableIntrinsic(I));
3191 IntrinsicInst *II = cast<IntrinsicInst>(I);
3192 Type *RetTy = ToVectorTy(II->getType(), VF);
3193 SmallVector<Type*, 4> Tys;
3194 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i)
3195 Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF));
3196 return TTI.getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys);
3199 // We are scalarizing the instruction. Return the cost of the scalar
3200 // instruction, plus the cost of insert and extract into vector
3201 // elements, times the vector width.
3204 if (!RetTy->isVoidTy() && VF != 1) {
3205 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3207 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3210 // The cost of inserting the results plus extracting each one of the
3212 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3215 // The cost of executing VF copies of the scalar instruction. This opcode
3216 // is unknown. Assume that it is the same as 'mul'.
3217 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3223 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3224 if (Scalar->isVoidTy() || VF == 1)
3226 return VectorType::get(Scalar, VF);
3229 char LoopVectorize::ID = 0;
3230 static const char lv_name[] = "Loop Vectorization";
3231 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3232 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3233 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3234 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3235 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3236 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3239 Pass *createLoopVectorizePass() {
3240 return new LoopVectorize();