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 const unsigned TinyTripCountVectorThreshold = 16;
106 /// We don't unroll loops with a known constant trip count below this number.
107 static const unsigned TinyTripCountUnrollThreshold = 128;
109 /// When performing a runtime memory check, do not check more than this
110 /// number of pointers. Notice that the check is quadratic!
111 static const unsigned RuntimeMemoryCheckThreshold = 4;
115 // Forward declarations.
116 class LoopVectorizationLegality;
117 class LoopVectorizationCostModel;
119 /// InnerLoopVectorizer vectorizes loops which contain only one basic
120 /// block to a specified vectorization factor (VF).
121 /// This class performs the widening of scalars into vectors, or multiple
122 /// scalars. This class also implements the following features:
123 /// * It inserts an epilogue loop for handling loops that don't have iteration
124 /// counts that are known to be a multiple of the vectorization factor.
125 /// * It handles the code generation for reduction variables.
126 /// * Scalarization (implementation using scalars) of un-vectorizable
128 /// InnerLoopVectorizer does not perform any vectorization-legality
129 /// checks, and relies on the caller to check for the different legality
130 /// aspects. The InnerLoopVectorizer relies on the
131 /// LoopVectorizationLegality class to provide information about the induction
132 /// and reduction variables that were found to a given vectorization factor.
133 class InnerLoopVectorizer {
135 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
136 DominatorTree *DT, DataLayout *DL, unsigned VecWidth,
137 unsigned UnrollFactor)
138 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), VF(VecWidth),
139 UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
140 OldInduction(0), WidenMap(UnrollFactor) {}
142 // Perform the actual loop widening (vectorization).
143 void vectorize(LoopVectorizationLegality *Legal) {
144 // Create a new empty loop. Unlink the old loop and connect the new one.
145 createEmptyLoop(Legal);
146 // Widen each instruction in the old loop to a new one in the new loop.
147 // Use the Legality module to find the induction and reduction variables.
148 vectorizeLoop(Legal);
149 // Register the new loop and update the analysis passes.
154 /// A small list of PHINodes.
155 typedef SmallVector<PHINode*, 4> PhiVector;
156 /// When we unroll loops we have multiple vector values for each scalar.
157 /// This data structure holds the unrolled and vectorized values that
158 /// originated from one scalar instruction.
159 typedef SmallVector<Value*, 2> VectorParts;
161 /// Add code that checks at runtime if the accessed arrays overlap.
162 /// Returns the comparator value or NULL if no check is needed.
163 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
165 /// Create an empty loop, based on the loop ranges of the old loop.
166 void createEmptyLoop(LoopVectorizationLegality *Legal);
167 /// Copy and widen the instructions from the old loop.
168 void vectorizeLoop(LoopVectorizationLegality *Legal);
170 /// A helper function that computes the predicate of the block BB, assuming
171 /// that the header block of the loop is set to True. It returns the *entry*
172 /// mask for the block BB.
173 VectorParts createBlockInMask(BasicBlock *BB);
174 /// A helper function that computes the predicate of the edge between SRC
176 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
178 /// A helper function to vectorize a single BB within the innermost loop.
179 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
182 /// Insert the new loop to the loop hierarchy and pass manager
183 /// and update the analysis passes.
184 void updateAnalysis();
186 /// This instruction is un-vectorizable. Implement it as a sequence
188 void scalarizeInstruction(Instruction *Instr);
190 /// Vectorize Load and Store instructions,
191 void vectorizeMemoryInstruction(Instruction *Instr,
192 LoopVectorizationLegality *Legal);
194 /// Create a broadcast instruction. This method generates a broadcast
195 /// instruction (shuffle) for loop invariant values and for the induction
196 /// value. If this is the induction variable then we extend it to N, N+1, ...
197 /// this is needed because each iteration in the loop corresponds to a SIMD
199 Value *getBroadcastInstrs(Value *V);
201 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
202 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
203 /// The sequence starts at StartIndex.
204 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
206 /// When we go over instructions in the basic block we rely on previous
207 /// values within the current basic block or on loop invariant values.
208 /// When we widen (vectorize) values we place them in the map. If the values
209 /// are not within the map, they have to be loop invariant, so we simply
210 /// broadcast them into a vector.
211 VectorParts &getVectorValue(Value *V);
213 /// Generate a shuffle sequence that will reverse the vector Vec.
214 Value *reverseVector(Value *Vec);
216 /// This is a helper class that holds the vectorizer state. It maps scalar
217 /// instructions to vector instructions. When the code is 'unrolled' then
218 /// then a single scalar value is mapped to multiple vector parts. The parts
219 /// are stored in the VectorPart type.
221 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
223 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
225 /// \return True if 'Key' is saved in the Value Map.
226 bool has(Value *Key) { return MapStoreage.count(Key); }
228 /// Initializes a new entry in the map. Sets all of the vector parts to the
229 /// save value in 'Val'.
230 /// \return A reference to a vector with splat values.
231 VectorParts &splat(Value *Key, Value *Val) {
232 MapStoreage[Key].clear();
233 MapStoreage[Key].append(UF, Val);
234 return MapStoreage[Key];
237 ///\return A reference to the value that is stored at 'Key'.
238 VectorParts &get(Value *Key) {
240 MapStoreage[Key].resize(UF);
241 return MapStoreage[Key];
244 /// The unroll factor. Each entry in the map stores this number of vector
248 /// Map storage. We use std::map and not DenseMap because insertions to a
249 /// dense map invalidates its iterators.
250 std::map<Value*, VectorParts> MapStoreage;
253 /// The original loop.
255 /// Scev analysis to use.
263 /// The vectorization SIMD factor to use. Each vector will have this many
266 /// The vectorization unroll factor to use. Each scalar is vectorized to this
267 /// many different vector instructions.
270 /// The builder that we use
273 // --- Vectorization state ---
275 /// The vector-loop preheader.
276 BasicBlock *LoopVectorPreHeader;
277 /// The scalar-loop preheader.
278 BasicBlock *LoopScalarPreHeader;
279 /// Middle Block between the vector and the scalar.
280 BasicBlock *LoopMiddleBlock;
281 ///The ExitBlock of the scalar loop.
282 BasicBlock *LoopExitBlock;
283 ///The vector loop body.
284 BasicBlock *LoopVectorBody;
285 ///The scalar loop body.
286 BasicBlock *LoopScalarBody;
287 /// A list of all bypass blocks. The first block is the entry of the loop.
288 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
290 /// The new Induction variable which was added to the new block.
292 /// The induction variable of the old basic block.
293 PHINode *OldInduction;
294 /// Maps scalars to widened vectors.
298 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
299 /// to what vectorization factor.
300 /// This class does not look at the profitability of vectorization, only the
301 /// legality. This class has two main kinds of checks:
302 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
303 /// will change the order of memory accesses in a way that will change the
304 /// correctness of the program.
305 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
306 /// checks for a number of different conditions, such as the availability of a
307 /// single induction variable, that all types are supported and vectorize-able,
308 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
309 /// This class is also used by InnerLoopVectorizer for identifying
310 /// induction variable and the different reduction variables.
311 class LoopVectorizationLegality {
313 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
315 : TheLoop(L), SE(SE), DL(DL), DT(DT), Induction(0) {}
317 /// This enum represents the kinds of reductions that we support.
319 RK_NoReduction, ///< Not a reduction.
320 RK_IntegerAdd, ///< Sum of integers.
321 RK_IntegerMult, ///< Product of integers.
322 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
323 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
324 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
325 RK_FloatAdd, ///< Sum of floats.
326 RK_FloatMult ///< Product of floats.
329 /// This enum represents the kinds of inductions that we support.
331 IK_NoInduction, ///< Not an induction variable.
332 IK_IntInduction, ///< Integer induction variable. Step = 1.
333 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
334 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
335 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
338 /// This POD struct holds information about reduction variables.
339 struct ReductionDescriptor {
340 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
341 Kind(RK_NoReduction) {}
343 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K)
344 : StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
346 // The starting value of the reduction.
347 // It does not have to be zero!
349 // The instruction who's value is used outside the loop.
350 Instruction *LoopExitInstr;
351 // The kind of the reduction.
355 // This POD struct holds information about the memory runtime legality
356 // check that a group of pointers do not overlap.
357 struct RuntimePointerCheck {
358 RuntimePointerCheck() : Need(false) {}
360 /// Reset the state of the pointer runtime information.
368 /// Insert a pointer and calculate the start and end SCEVs.
369 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr);
371 /// This flag indicates if we need to add the runtime check.
373 /// Holds the pointers that we need to check.
374 SmallVector<Value*, 2> Pointers;
375 /// Holds the pointer value at the beginning of the loop.
376 SmallVector<const SCEV*, 2> Starts;
377 /// Holds the pointer value at the end of the loop.
378 SmallVector<const SCEV*, 2> Ends;
381 /// A POD for saving information about induction variables.
382 struct InductionInfo {
383 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
384 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
391 /// ReductionList contains the reduction descriptors for all
392 /// of the reductions that were found in the loop.
393 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
395 /// InductionList saves induction variables and maps them to the
396 /// induction descriptor.
397 typedef MapVector<PHINode*, InductionInfo> InductionList;
399 /// Returns true if it is legal to vectorize this loop.
400 /// This does not mean that it is profitable to vectorize this
401 /// loop, only that it is legal to do so.
404 /// Returns the Induction variable.
405 PHINode *getInduction() { return Induction; }
407 /// Returns the reduction variables found in the loop.
408 ReductionList *getReductionVars() { return &Reductions; }
410 /// Returns the induction variables found in the loop.
411 InductionList *getInductionVars() { return &Inductions; }
413 /// Returns True if V is an induction variable in this loop.
414 bool isInductionVariable(const Value *V);
416 /// Return true if the block BB needs to be predicated in order for the loop
417 /// to be vectorized.
418 bool blockNeedsPredication(BasicBlock *BB);
420 /// Check if this pointer is consecutive when vectorizing. This happens
421 /// when the last index of the GEP is the induction variable, or that the
422 /// pointer itself is an induction variable.
423 /// This check allows us to vectorize A[idx] into a wide load/store.
425 /// 0 - Stride is unknown or non consecutive.
426 /// 1 - Address is consecutive.
427 /// -1 - Address is consecutive, and decreasing.
428 int isConsecutivePtr(Value *Ptr);
430 /// Returns true if the value V is uniform within the loop.
431 bool isUniform(Value *V);
433 /// Returns true if this instruction will remain scalar after vectorization.
434 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
436 /// Returns the information that we collected about runtime memory check.
437 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
439 /// Check if a single basic block loop is vectorizable.
440 /// At this point we know that this is a loop with a constant trip count
441 /// and we only need to check individual instructions.
442 bool canVectorizeInstrs();
444 /// When we vectorize loops we may change the order in which
445 /// we read and write from memory. This method checks if it is
446 /// legal to vectorize the code, considering only memory constrains.
447 /// Returns true if the loop is vectorizable
448 bool canVectorizeMemory();
450 /// Return true if we can vectorize this loop using the IF-conversion
452 bool canVectorizeWithIfConvert();
454 /// Collect the variables that need to stay uniform after vectorization.
455 void collectLoopUniforms();
457 /// Return true if all of the instructions in the block can be speculatively
459 bool blockCanBePredicated(BasicBlock *BB);
461 /// Returns True, if 'Phi' is the kind of reduction variable for type
462 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
463 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
464 /// Returns true if the instruction I can be a reduction variable of type
466 bool isReductionInstr(Instruction *I, ReductionKind Kind);
467 /// Returns the induction kind of Phi. This function may return NoInduction
468 /// if the PHI is not an induction variable.
469 InductionKind isInductionVariable(PHINode *Phi);
470 /// Return true if can compute the address bounds of Ptr within the loop.
471 bool hasComputableBounds(Value *Ptr);
473 /// The loop that we evaluate.
477 /// DataLayout analysis.
482 // --- vectorization state --- //
484 /// Holds the integer induction variable. This is the counter of the
487 /// Holds the reduction variables.
488 ReductionList Reductions;
489 /// Holds all of the induction variables that we found in the loop.
490 /// Notice that inductions don't need to start at zero and that induction
491 /// variables can be pointers.
492 InductionList Inductions;
494 /// Allowed outside users. This holds the reduction
495 /// vars which can be accessed from outside the loop.
496 SmallPtrSet<Value*, 4> AllowedExit;
497 /// This set holds the variables which are known to be uniform after
499 SmallPtrSet<Instruction*, 4> Uniforms;
500 /// We need to check that all of the pointers in this list are disjoint
502 RuntimePointerCheck PtrRtCheck;
505 /// LoopVectorizationCostModel - estimates the expected speedups due to
507 /// In many cases vectorization is not profitable. This can happen because of
508 /// a number of reasons. In this class we mainly attempt to predict the
509 /// expected speedup/slowdowns due to the supported instruction set. We use the
510 /// TargetTransformInfo to query the different backends for the cost of
511 /// different operations.
512 class LoopVectorizationCostModel {
514 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
515 LoopVectorizationLegality *Legal,
516 const TargetTransformInfo &TTI)
517 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI) {}
519 /// \return The most profitable vectorization factor and the cost of that VF.
520 /// This method checks every power of two up to VF. If UserVF is not ZERO
521 /// then this vectorization factor will be selected if vectorization is
523 std::pair<unsigned, unsigned>
524 selectVectorizationFactor(bool OptForSize, unsigned UserVF);
526 /// \returns The size (in bits) of the widest type in the code that
527 /// needs to be vectorized. We ignore values that remain scalar such as
528 /// 64 bit loop indices.
529 unsigned getWidestType();
531 /// \return The most profitable unroll factor.
532 /// If UserUF is non-zero then this method finds the best unroll-factor
533 /// based on register pressure and other parameters.
534 /// VF and LoopCost are the selected vectorization factor and the cost of the
536 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
539 /// \brief A struct that represents some properties of the register usage
541 struct RegisterUsage {
542 /// Holds the number of loop invariant values that are used in the loop.
543 unsigned LoopInvariantRegs;
544 /// Holds the maximum number of concurrent live intervals in the loop.
545 unsigned MaxLocalUsers;
546 /// Holds the number of instructions in the loop.
547 unsigned NumInstructions;
550 /// \return information about the register usage of the loop.
551 RegisterUsage calculateRegisterUsage();
554 /// Returns the expected execution cost. The unit of the cost does
555 /// not matter because we use the 'cost' units to compare different
556 /// vector widths. The cost that is returned is *not* normalized by
557 /// the factor width.
558 unsigned expectedCost(unsigned VF);
560 /// Returns the execution time cost of an instruction for a given vector
561 /// width. Vector width of one means scalar.
562 unsigned getInstructionCost(Instruction *I, unsigned VF);
564 /// A helper function for converting Scalar types to vector types.
565 /// If the incoming type is void, we return void. If the VF is 1, we return
567 static Type* ToVectorTy(Type *Scalar, unsigned VF);
569 /// The loop that we evaluate.
573 /// Loop Info analysis.
575 /// Vectorization legality.
576 LoopVectorizationLegality *Legal;
577 /// Vector target information.
578 const TargetTransformInfo &TTI;
581 /// The LoopVectorize Pass.
582 struct LoopVectorize : public LoopPass {
583 /// Pass identification, replacement for typeid
586 explicit LoopVectorize() : LoopPass(ID) {
587 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
593 TargetTransformInfo *TTI;
596 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
597 // We only vectorize innermost loops.
601 SE = &getAnalysis<ScalarEvolution>();
602 DL = getAnalysisIfAvailable<DataLayout>();
603 LI = &getAnalysis<LoopInfo>();
604 TTI = &getAnalysis<TargetTransformInfo>();
605 DT = &getAnalysis<DominatorTree>();
607 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
608 L->getHeader()->getParent()->getName() << "\"\n");
610 // Check if it is legal to vectorize the loop.
611 LoopVectorizationLegality LVL(L, SE, DL, DT);
612 if (!LVL.canVectorize()) {
613 DEBUG(dbgs() << "LV: Not vectorizing.\n");
617 // Use the cost model.
618 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI);
620 // Check the function attribues to find out if this function should be
621 // optimized for size.
622 Function *F = L->getHeader()->getParent();
623 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
624 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
625 unsigned FnIndex = AttributeSet::FunctionIndex;
626 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
627 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
630 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
631 "attribute is used.\n");
635 // Select the optimal vectorization factor.
636 std::pair<unsigned, unsigned> VFPair;
637 VFPair = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
638 // Select the unroll factor.
639 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
640 VFPair.first, VFPair.second);
641 unsigned VF = VFPair.first;
644 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
648 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
649 F->getParent()->getModuleIdentifier()<<"\n");
650 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
652 // If we decided that it is *legal* to vectorizer the loop then do it.
653 InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF, UF);
656 DEBUG(verifyFunction(*L->getHeader()->getParent()));
660 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
661 LoopPass::getAnalysisUsage(AU);
662 AU.addRequiredID(LoopSimplifyID);
663 AU.addRequiredID(LCSSAID);
664 AU.addRequired<DominatorTree>();
665 AU.addRequired<LoopInfo>();
666 AU.addRequired<ScalarEvolution>();
667 AU.addRequired<TargetTransformInfo>();
668 AU.addPreserved<LoopInfo>();
669 AU.addPreserved<DominatorTree>();
674 } // end anonymous namespace
676 //===----------------------------------------------------------------------===//
677 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
678 // LoopVectorizationCostModel.
679 //===----------------------------------------------------------------------===//
682 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
683 Loop *Lp, Value *Ptr) {
684 const SCEV *Sc = SE->getSCEV(Ptr);
685 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
686 assert(AR && "Invalid addrec expression");
687 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
688 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
689 Pointers.push_back(Ptr);
690 Starts.push_back(AR->getStart());
691 Ends.push_back(ScEnd);
694 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
695 // Save the current insertion location.
696 Instruction *Loc = Builder.GetInsertPoint();
698 // We need to place the broadcast of invariant variables outside the loop.
699 Instruction *Instr = dyn_cast<Instruction>(V);
700 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
701 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
703 // Place the code for broadcasting invariant variables in the new preheader.
705 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
707 // Broadcast the scalar into all locations in the vector.
708 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
710 // Restore the builder insertion point.
712 Builder.SetInsertPoint(Loc);
717 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
719 assert(Val->getType()->isVectorTy() && "Must be a vector");
720 assert(Val->getType()->getScalarType()->isIntegerTy() &&
721 "Elem must be an integer");
723 Type *ITy = Val->getType()->getScalarType();
724 VectorType *Ty = cast<VectorType>(Val->getType());
725 int VLen = Ty->getNumElements();
726 SmallVector<Constant*, 8> Indices;
728 // Create a vector of consecutive numbers from zero to VF.
729 for (int i = 0; i < VLen; ++i) {
730 int Idx = Negate ? (-i): i;
731 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
734 // Add the consecutive indices to the vector value.
735 Constant *Cv = ConstantVector::get(Indices);
736 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
737 return Builder.CreateAdd(Val, Cv, "induction");
740 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
741 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
742 // Make sure that the pointer does not point to structs.
743 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
746 // If this value is a pointer induction variable we know it is consecutive.
747 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
748 if (Phi && Inductions.count(Phi)) {
749 InductionInfo II = Inductions[Phi];
750 if (IK_PtrInduction == II.IK)
752 else if (IK_ReversePtrInduction == II.IK)
756 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
760 unsigned NumOperands = Gep->getNumOperands();
761 Value *LastIndex = Gep->getOperand(NumOperands - 1);
763 Value *GpPtr = Gep->getPointerOperand();
764 // If this GEP value is a consecutive pointer induction variable and all of
765 // the indices are constant then we know it is consecutive. We can
766 Phi = dyn_cast<PHINode>(GpPtr);
767 if (Phi && Inductions.count(Phi)) {
769 // Make sure that the pointer does not point to structs.
770 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
771 if (GepPtrType->getElementType()->isAggregateType())
774 // Make sure that all of the index operands are loop invariant.
775 for (unsigned i = 1; i < NumOperands; ++i)
776 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
779 InductionInfo II = Inductions[Phi];
780 if (IK_PtrInduction == II.IK)
782 else if (IK_ReversePtrInduction == II.IK)
786 // Check that all of the gep indices are uniform except for the last.
787 for (unsigned i = 0; i < NumOperands - 1; ++i)
788 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
791 // We can emit wide load/stores only if the last index is the induction
793 const SCEV *Last = SE->getSCEV(LastIndex);
794 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
795 const SCEV *Step = AR->getStepRecurrence(*SE);
797 // The memory is consecutive because the last index is consecutive
798 // and all other indices are loop invariant.
801 if (Step->isAllOnesValue())
808 bool LoopVectorizationLegality::isUniform(Value *V) {
809 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
812 InnerLoopVectorizer::VectorParts&
813 InnerLoopVectorizer::getVectorValue(Value *V) {
814 assert(V != Induction && "The new induction variable should not be used.");
815 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
817 // If we have this scalar in the map, return it.
819 return WidenMap.get(V);
821 // If this scalar is unknown, assume that it is a constant or that it is
822 // loop invariant. Broadcast V and save the value for future uses.
823 Value *B = getBroadcastInstrs(V);
824 WidenMap.splat(V, B);
825 return WidenMap.get(V);
828 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
829 assert(Vec->getType()->isVectorTy() && "Invalid type");
830 SmallVector<Constant*, 8> ShuffleMask;
831 for (unsigned i = 0; i < VF; ++i)
832 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
834 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
835 ConstantVector::get(ShuffleMask),
840 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
841 LoopVectorizationLegality *Legal) {
842 // Attempt to issue a wide load.
843 LoadInst *LI = dyn_cast<LoadInst>(Instr);
844 StoreInst *SI = dyn_cast<StoreInst>(Instr);
846 assert((LI || SI) && "Invalid Load/Store instruction");
848 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
849 Type *DataTy = VectorType::get(ScalarDataTy, VF);
850 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
851 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
853 // If the pointer is loop invariant or if it is non consecutive,
854 // scalarize the load.
855 int Stride = Legal->isConsecutivePtr(Ptr);
856 bool Reverse = Stride < 0;
857 bool UniformLoad = LI && Legal->isUniform(Ptr);
858 if (Stride == 0 || UniformLoad)
859 return scalarizeInstruction(Instr);
861 Constant *Zero = Builder.getInt32(0);
862 VectorParts &Entry = WidenMap.get(Instr);
864 // Handle consecutive loads/stores.
865 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
866 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
867 Value *PtrOperand = Gep->getPointerOperand();
868 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
869 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
871 // Create the new GEP with the new induction variable.
872 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
873 Gep2->setOperand(0, FirstBasePtr);
874 Gep2->setName("gep.indvar.base");
875 Ptr = Builder.Insert(Gep2);
877 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
878 OrigLoop) && "Base ptr must be invariant");
880 // The last index does not have to be the induction. It can be
881 // consecutive and be a function of the index. For example A[I+1];
882 unsigned NumOperands = Gep->getNumOperands();
884 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
885 VectorParts &GEPParts = getVectorValue(LastGepOperand);
886 Value *LastIndex = GEPParts[0];
887 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
889 // Create the new GEP with the new induction variable.
890 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
891 Gep2->setOperand(NumOperands - 1, LastIndex);
892 Gep2->setName("gep.indvar.idx");
893 Ptr = Builder.Insert(Gep2);
895 // Use the induction element ptr.
896 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
897 VectorParts &PtrVal = getVectorValue(Ptr);
898 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
903 assert(!Legal->isUniform(SI->getPointerOperand()) &&
904 "We do not allow storing to uniform addresses");
906 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
907 for (unsigned Part = 0; Part < UF; ++Part) {
908 // Calculate the pointer for the specific unroll-part.
909 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
912 // If we store to reverse consecutive memory locations then we need
913 // to reverse the order of elements in the stored value.
914 StoredVal[Part] = reverseVector(StoredVal[Part]);
915 // If the address is consecutive but reversed, then the
916 // wide store needs to start at the last vector element.
917 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
918 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
921 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
922 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
926 for (unsigned Part = 0; Part < UF; ++Part) {
927 // Calculate the pointer for the specific unroll-part.
928 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
931 // If the address is consecutive but reversed, then the
932 // wide store needs to start at the last vector element.
933 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
934 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
937 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
938 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
939 cast<LoadInst>(LI)->setAlignment(Alignment);
940 Entry[Part] = Reverse ? reverseVector(LI) : LI;
944 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
945 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
946 // Holds vector parameters or scalars, in case of uniform vals.
947 SmallVector<VectorParts, 4> Params;
949 // Find all of the vectorized parameters.
950 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
951 Value *SrcOp = Instr->getOperand(op);
953 // If we are accessing the old induction variable, use the new one.
954 if (SrcOp == OldInduction) {
955 Params.push_back(getVectorValue(SrcOp));
959 // Try using previously calculated values.
960 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
962 // If the src is an instruction that appeared earlier in the basic block
963 // then it should already be vectorized.
964 if (SrcInst && OrigLoop->contains(SrcInst)) {
965 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
966 // The parameter is a vector value from earlier.
967 Params.push_back(WidenMap.get(SrcInst));
969 // The parameter is a scalar from outside the loop. Maybe even a constant.
971 Scalars.append(UF, SrcOp);
972 Params.push_back(Scalars);
976 assert(Params.size() == Instr->getNumOperands() &&
977 "Invalid number of operands");
979 // Does this instruction return a value ?
980 bool IsVoidRetTy = Instr->getType()->isVoidTy();
982 Value *UndefVec = IsVoidRetTy ? 0 :
983 UndefValue::get(VectorType::get(Instr->getType(), VF));
984 // Create a new entry in the WidenMap and initialize it to Undef or Null.
985 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
987 // For each scalar that we create:
988 for (unsigned Width = 0; Width < VF; ++Width) {
989 // For each vector unroll 'part':
990 for (unsigned Part = 0; Part < UF; ++Part) {
991 Instruction *Cloned = Instr->clone();
993 Cloned->setName(Instr->getName() + ".cloned");
994 // Replace the operands of the cloned instrucions with extracted scalars.
995 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
996 Value *Op = Params[op][Part];
997 // Param is a vector. Need to extract the right lane.
998 if (Op->getType()->isVectorTy())
999 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1000 Cloned->setOperand(op, Op);
1003 // Place the cloned scalar in the new loop.
1004 Builder.Insert(Cloned);
1006 // If the original scalar returns a value we need to place it in a vector
1007 // so that future users will be able to use it.
1009 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1010 Builder.getInt32(Width));
1016 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1018 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1019 Legal->getRuntimePointerCheck();
1021 if (!PtrRtCheck->Need)
1024 Instruction *MemoryRuntimeCheck = 0;
1025 unsigned NumPointers = PtrRtCheck->Pointers.size();
1026 SmallVector<Value* , 2> Starts;
1027 SmallVector<Value* , 2> Ends;
1029 SCEVExpander Exp(*SE, "induction");
1031 // Use this type for pointer arithmetic.
1032 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1034 for (unsigned i = 0; i < NumPointers; ++i) {
1035 Value *Ptr = PtrRtCheck->Pointers[i];
1036 const SCEV *Sc = SE->getSCEV(Ptr);
1038 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1039 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1041 Starts.push_back(Ptr);
1042 Ends.push_back(Ptr);
1044 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1046 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1047 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1048 Starts.push_back(Start);
1049 Ends.push_back(End);
1053 IRBuilder<> ChkBuilder(Loc);
1055 for (unsigned i = 0; i < NumPointers; ++i) {
1056 for (unsigned j = i+1; j < NumPointers; ++j) {
1057 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1058 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1059 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1060 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1062 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1063 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1064 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1065 if (MemoryRuntimeCheck)
1066 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1069 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1073 return MemoryRuntimeCheck;
1077 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1079 In this function we generate a new loop. The new loop will contain
1080 the vectorized instructions while the old loop will continue to run the
1083 [ ] <-- vector loop bypass (may consist of multiple blocks).
1086 | [ ] <-- vector pre header.
1090 | [ ]_| <-- vector loop.
1093 >[ ] <--- middle-block.
1096 | [ ] <--- new preheader.
1100 | [ ]_| <-- old scalar loop to handle remainder.
1103 >[ ] <-- exit block.
1107 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1108 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1109 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1110 assert(ExitBlock && "Must have an exit block");
1112 // Some loops have a single integer induction variable, while other loops
1113 // don't. One example is c++ iterators that often have multiple pointer
1114 // induction variables. In the code below we also support a case where we
1115 // don't have a single induction variable.
1116 OldInduction = Legal->getInduction();
1117 Type *IdxTy = OldInduction ? OldInduction->getType() :
1118 DL->getIntPtrType(SE->getContext());
1120 // Find the loop boundaries.
1121 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1122 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1124 // Get the total trip count from the count by adding 1.
1125 ExitCount = SE->getAddExpr(ExitCount,
1126 SE->getConstant(ExitCount->getType(), 1));
1128 // Expand the trip count and place the new instructions in the preheader.
1129 // Notice that the pre-header does not change, only the loop body.
1130 SCEVExpander Exp(*SE, "induction");
1132 // Count holds the overall loop count (N).
1133 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1134 BypassBlock->getTerminator());
1136 // The loop index does not have to start at Zero. Find the original start
1137 // value from the induction PHI node. If we don't have an induction variable
1138 // then we know that it starts at zero.
1139 Value *StartIdx = OldInduction ?
1140 OldInduction->getIncomingValueForBlock(BypassBlock):
1141 ConstantInt::get(IdxTy, 0);
1143 assert(BypassBlock && "Invalid loop structure");
1144 LoopBypassBlocks.push_back(BypassBlock);
1146 // Split the single block loop into the two loop structure described above.
1147 BasicBlock *VectorPH =
1148 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1149 BasicBlock *VecBody =
1150 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1151 BasicBlock *MiddleBlock =
1152 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1153 BasicBlock *ScalarPH =
1154 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1156 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1158 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1160 // Generate the induction variable.
1161 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1162 // The loop step is equal to the vectorization factor (num of SIMD elements)
1163 // times the unroll factor (num of SIMD instructions).
1164 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1166 // This is the IR builder that we use to add all of the logic for bypassing
1167 // the new vector loop.
1168 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1170 // We may need to extend the index in case there is a type mismatch.
1171 // We know that the count starts at zero and does not overflow.
1172 if (Count->getType() != IdxTy) {
1173 // The exit count can be of pointer type. Convert it to the correct
1175 if (ExitCount->getType()->isPointerTy())
1176 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1178 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1181 // Add the start index to the loop count to get the new end index.
1182 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1184 // Now we need to generate the expression for N - (N % VF), which is
1185 // the part that the vectorized body will execute.
1186 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1187 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1188 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1189 "end.idx.rnd.down");
1191 // Now, compare the new count to zero. If it is zero skip the vector loop and
1192 // jump to the scalar loop.
1193 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1196 BasicBlock *LastBypassBlock = BypassBlock;
1198 // Generate the code that checks in runtime if arrays overlap. We put the
1199 // checks into a separate block to make the more common case of few elements
1201 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1202 BypassBlock->getTerminator());
1203 if (MemRuntimeCheck) {
1204 // Create a new block containing the memory check.
1205 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1207 LoopBypassBlocks.push_back(CheckBlock);
1209 // Replace the branch into the memory check block with a conditional branch
1210 // for the "few elements case".
1211 Instruction *OldTerm = BypassBlock->getTerminator();
1212 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1213 OldTerm->eraseFromParent();
1215 Cmp = MemRuntimeCheck;
1216 LastBypassBlock = CheckBlock;
1219 LastBypassBlock->getTerminator()->eraseFromParent();
1220 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1223 // We are going to resume the execution of the scalar loop.
1224 // Go over all of the induction variables that we found and fix the
1225 // PHIs that are left in the scalar version of the loop.
1226 // The starting values of PHI nodes depend on the counter of the last
1227 // iteration in the vectorized loop.
1228 // If we come from a bypass edge then we need to start from the original
1231 // This variable saves the new starting index for the scalar loop.
1232 PHINode *ResumeIndex = 0;
1233 LoopVectorizationLegality::InductionList::iterator I, E;
1234 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1235 for (I = List->begin(), E = List->end(); I != E; ++I) {
1236 PHINode *OrigPhi = I->first;
1237 LoopVectorizationLegality::InductionInfo II = I->second;
1238 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1239 MiddleBlock->getTerminator());
1240 Value *EndValue = 0;
1242 case LoopVectorizationLegality::IK_NoInduction:
1243 llvm_unreachable("Unknown induction");
1244 case LoopVectorizationLegality::IK_IntInduction: {
1245 // Handle the integer induction counter:
1246 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1247 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1248 // We know what the end value is.
1249 EndValue = IdxEndRoundDown;
1250 // We also know which PHI node holds it.
1251 ResumeIndex = ResumeVal;
1254 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1255 // Convert the CountRoundDown variable to the PHI size.
1256 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1257 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1258 Value *CRD = CountRoundDown;
1259 if (CRDSize > IISize)
1260 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1261 II.StartValue->getType(), "tr.crd",
1262 LoopBypassBlocks.back()->getTerminator());
1263 else if (CRDSize < IISize)
1264 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1265 II.StartValue->getType(),
1267 LoopBypassBlocks.back()->getTerminator());
1268 // Handle reverse integer induction counter:
1270 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1271 LoopBypassBlocks.back()->getTerminator());
1274 case LoopVectorizationLegality::IK_PtrInduction: {
1275 // For pointer induction variables, calculate the offset using
1278 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end",
1279 LoopBypassBlocks.back()->getTerminator());
1282 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1283 // The value at the end of the loop for the reverse pointer is calculated
1284 // by creating a GEP with a negative index starting from the start value.
1285 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1286 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown,
1288 LoopBypassBlocks.back()->getTerminator());
1289 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx,
1291 LoopBypassBlocks.back()->getTerminator());
1296 // The new PHI merges the original incoming value, in case of a bypass,
1297 // or the value at the end of the vectorized loop.
1298 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1299 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1300 ResumeVal->addIncoming(EndValue, VecBody);
1302 // Fix the scalar body counter (PHI node).
1303 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1304 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1307 // If we are generating a new induction variable then we also need to
1308 // generate the code that calculates the exit value. This value is not
1309 // simply the end of the counter because we may skip the vectorized body
1310 // in case of a runtime check.
1312 assert(!ResumeIndex && "Unexpected resume value found");
1313 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1314 MiddleBlock->getTerminator());
1315 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1316 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1317 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1320 // Make sure that we found the index where scalar loop needs to continue.
1321 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1322 "Invalid resume Index");
1324 // Add a check in the middle block to see if we have completed
1325 // all of the iterations in the first vector loop.
1326 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1327 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1328 ResumeIndex, "cmp.n",
1329 MiddleBlock->getTerminator());
1331 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1332 // Remove the old terminator.
1333 MiddleBlock->getTerminator()->eraseFromParent();
1335 // Create i+1 and fill the PHINode.
1336 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1337 Induction->addIncoming(StartIdx, VectorPH);
1338 Induction->addIncoming(NextIdx, VecBody);
1339 // Create the compare.
1340 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1341 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1343 // Now we have two terminators. Remove the old one from the block.
1344 VecBody->getTerminator()->eraseFromParent();
1346 // Get ready to start creating new instructions into the vectorized body.
1347 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1349 // Create and register the new vector loop.
1350 Loop* Lp = new Loop();
1351 Loop *ParentLoop = OrigLoop->getParentLoop();
1353 // Insert the new loop into the loop nest and register the new basic blocks.
1355 ParentLoop->addChildLoop(Lp);
1356 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1357 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1358 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1359 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1360 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1362 LI->addTopLevelLoop(Lp);
1365 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1368 LoopVectorPreHeader = VectorPH;
1369 LoopScalarPreHeader = ScalarPH;
1370 LoopMiddleBlock = MiddleBlock;
1371 LoopExitBlock = ExitBlock;
1372 LoopVectorBody = VecBody;
1373 LoopScalarBody = OldBasicBlock;
1376 /// This function returns the identity element (or neutral element) for
1377 /// the operation K.
1379 getReductionIdentity(LoopVectorizationLegality::ReductionKind K, Type *Tp) {
1381 case LoopVectorizationLegality:: RK_IntegerXor:
1382 case LoopVectorizationLegality:: RK_IntegerAdd:
1383 case LoopVectorizationLegality:: RK_IntegerOr:
1384 // Adding, Xoring, Oring zero to a number does not change it.
1385 return ConstantInt::get(Tp, 0);
1386 case LoopVectorizationLegality:: RK_IntegerMult:
1387 // Multiplying a number by 1 does not change it.
1388 return ConstantInt::get(Tp, 1);
1389 case LoopVectorizationLegality:: RK_IntegerAnd:
1390 // AND-ing a number with an all-1 value does not change it.
1391 return ConstantInt::get(Tp, -1, true);
1392 case LoopVectorizationLegality:: RK_FloatMult:
1393 // Multiplying a number by 1 does not change it.
1394 return ConstantFP::get(Tp, 1.0L);
1395 case LoopVectorizationLegality:: RK_FloatAdd:
1396 // Adding zero to a number does not change it.
1397 return ConstantFP::get(Tp, 0.0L);
1399 llvm_unreachable("Unknown reduction kind");
1404 isTriviallyVectorizableIntrinsic(Instruction *Inst) {
1405 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
1408 switch (II->getIntrinsicID()) {
1409 case Intrinsic::sqrt:
1410 case Intrinsic::sin:
1411 case Intrinsic::cos:
1412 case Intrinsic::exp:
1413 case Intrinsic::exp2:
1414 case Intrinsic::log:
1415 case Intrinsic::log10:
1416 case Intrinsic::log2:
1417 case Intrinsic::fabs:
1418 case Intrinsic::floor:
1419 case Intrinsic::ceil:
1420 case Intrinsic::trunc:
1421 case Intrinsic::rint:
1422 case Intrinsic::nearbyint:
1423 case Intrinsic::pow:
1424 case Intrinsic::fma:
1425 case Intrinsic::fmuladd:
1433 /// This function translates the reduction kind to an LLVM binary operator.
1434 static Instruction::BinaryOps
1435 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1437 case LoopVectorizationLegality::RK_IntegerAdd:
1438 return Instruction::Add;
1439 case LoopVectorizationLegality::RK_IntegerMult:
1440 return Instruction::Mul;
1441 case LoopVectorizationLegality::RK_IntegerOr:
1442 return Instruction::Or;
1443 case LoopVectorizationLegality::RK_IntegerAnd:
1444 return Instruction::And;
1445 case LoopVectorizationLegality::RK_IntegerXor:
1446 return Instruction::Xor;
1447 case LoopVectorizationLegality::RK_FloatMult:
1448 return Instruction::FMul;
1449 case LoopVectorizationLegality::RK_FloatAdd:
1450 return Instruction::FAdd;
1452 llvm_unreachable("Unknown reduction operation");
1457 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1458 //===------------------------------------------------===//
1460 // Notice: any optimization or new instruction that go
1461 // into the code below should be also be implemented in
1464 //===------------------------------------------------===//
1465 Constant *Zero = Builder.getInt32(0);
1467 // In order to support reduction variables we need to be able to vectorize
1468 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1469 // stages. First, we create a new vector PHI node with no incoming edges.
1470 // We use this value when we vectorize all of the instructions that use the
1471 // PHI. Next, after all of the instructions in the block are complete we
1472 // add the new incoming edges to the PHI. At this point all of the
1473 // instructions in the basic block are vectorized, so we can use them to
1474 // construct the PHI.
1475 PhiVector RdxPHIsToFix;
1477 // Scan the loop in a topological order to ensure that defs are vectorized
1479 LoopBlocksDFS DFS(OrigLoop);
1482 // Vectorize all of the blocks in the original loop.
1483 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1484 be = DFS.endRPO(); bb != be; ++bb)
1485 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1487 // At this point every instruction in the original loop is widened to
1488 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1489 // that we vectorized. The PHI nodes are currently empty because we did
1490 // not want to introduce cycles. Notice that the remaining PHI nodes
1491 // that we need to fix are reduction variables.
1493 // Create the 'reduced' values for each of the induction vars.
1494 // The reduced values are the vector values that we scalarize and combine
1495 // after the loop is finished.
1496 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1498 PHINode *RdxPhi = *it;
1499 assert(RdxPhi && "Unable to recover vectorized PHI");
1501 // Find the reduction variable descriptor.
1502 assert(Legal->getReductionVars()->count(RdxPhi) &&
1503 "Unable to find the reduction variable");
1504 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1505 (*Legal->getReductionVars())[RdxPhi];
1507 // We need to generate a reduction vector from the incoming scalar.
1508 // To do so, we need to generate the 'identity' vector and overide
1509 // one of the elements with the incoming scalar reduction. We need
1510 // to do it in the vector-loop preheader.
1511 Builder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1513 // This is the vector-clone of the value that leaves the loop.
1514 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1515 Type *VecTy = VectorExit[0]->getType();
1517 // Find the reduction identity variable. Zero for addition, or, xor,
1518 // one for multiplication, -1 for And.
1519 Constant *Iden = getReductionIdentity(RdxDesc.Kind, VecTy->getScalarType());
1520 Constant *Identity = ConstantVector::getSplat(VF, Iden);
1522 // This vector is the Identity vector where the first element is the
1523 // incoming scalar reduction.
1524 Value *VectorStart = Builder.CreateInsertElement(Identity,
1525 RdxDesc.StartValue, Zero);
1527 // Fix the vector-loop phi.
1528 // We created the induction variable so we know that the
1529 // preheader is the first entry.
1530 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1532 // Reductions do not have to start at zero. They can start with
1533 // any loop invariant values.
1534 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1535 BasicBlock *Latch = OrigLoop->getLoopLatch();
1536 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1537 VectorParts &Val = getVectorValue(LoopVal);
1538 for (unsigned part = 0; part < UF; ++part) {
1539 // Make sure to add the reduction stat value only to the
1540 // first unroll part.
1541 Value *StartVal = (part == 0) ? VectorStart : Identity;
1542 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1543 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1546 // Before each round, move the insertion point right between
1547 // the PHIs and the values we are going to write.
1548 // This allows us to write both PHINodes and the extractelement
1550 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1552 VectorParts RdxParts;
1553 for (unsigned part = 0; part < UF; ++part) {
1554 // This PHINode contains the vectorized reduction variable, or
1555 // the initial value vector, if we bypass the vector loop.
1556 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1557 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1558 Value *StartVal = (part == 0) ? VectorStart : Identity;
1559 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1560 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1561 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1562 RdxParts.push_back(NewPhi);
1565 // Reduce all of the unrolled parts into a single vector.
1566 Value *ReducedPartRdx = RdxParts[0];
1567 for (unsigned part = 1; part < UF; ++part) {
1568 Instruction::BinaryOps Op = getReductionBinOp(RdxDesc.Kind);
1569 ReducedPartRdx = Builder.CreateBinOp(Op, RdxParts[part], ReducedPartRdx,
1573 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1574 // and vector ops, reducing the set of values being computed by half each
1576 assert(isPowerOf2_32(VF) &&
1577 "Reduction emission only supported for pow2 vectors!");
1578 Value *TmpVec = ReducedPartRdx;
1579 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1580 for (unsigned i = VF; i != 1; i >>= 1) {
1581 // Move the upper half of the vector to the lower half.
1582 for (unsigned j = 0; j != i/2; ++j)
1583 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1585 // Fill the rest of the mask with undef.
1586 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1587 UndefValue::get(Builder.getInt32Ty()));
1590 Builder.CreateShuffleVector(TmpVec,
1591 UndefValue::get(TmpVec->getType()),
1592 ConstantVector::get(ShuffleMask),
1595 Instruction::BinaryOps Op = getReductionBinOp(RdxDesc.Kind);
1596 TmpVec = Builder.CreateBinOp(Op, TmpVec, Shuf, "bin.rdx");
1599 // The result is in the first element of the vector.
1600 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1602 // Now, we need to fix the users of the reduction variable
1603 // inside and outside of the scalar remainder loop.
1604 // We know that the loop is in LCSSA form. We need to update the
1605 // PHI nodes in the exit blocks.
1606 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1607 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1608 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1609 if (!LCSSAPhi) continue;
1611 // All PHINodes need to have a single entry edge, or two if
1612 // we already fixed them.
1613 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1615 // We found our reduction value exit-PHI. Update it with the
1616 // incoming bypass edge.
1617 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1618 // Add an edge coming from the bypass.
1619 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1622 }// end of the LCSSA phi scan.
1624 // Fix the scalar loop reduction variable with the incoming reduction sum
1625 // from the vector body and from the backedge value.
1626 int IncomingEdgeBlockIdx =
1627 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1628 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1629 // Pick the other block.
1630 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1631 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1632 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1633 }// end of for each redux variable.
1635 // The Loop exit block may have single value PHI nodes where the incoming
1636 // value is 'undef'. While vectorizing we only handled real values that
1637 // were defined inside the loop. Here we handle the 'undef case'.
1639 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1640 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1641 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1642 if (!LCSSAPhi) continue;
1643 if (LCSSAPhi->getNumIncomingValues() == 1)
1644 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1649 InnerLoopVectorizer::VectorParts
1650 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1651 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1654 VectorParts SrcMask = createBlockInMask(Src);
1656 // The terminator has to be a branch inst!
1657 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1658 assert(BI && "Unexpected terminator found");
1660 if (BI->isConditional()) {
1661 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1663 if (BI->getSuccessor(0) != Dst)
1664 for (unsigned part = 0; part < UF; ++part)
1665 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1667 for (unsigned part = 0; part < UF; ++part)
1668 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1675 InnerLoopVectorizer::VectorParts
1676 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1677 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1679 // Loop incoming mask is all-one.
1680 if (OrigLoop->getHeader() == BB) {
1681 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1682 return getVectorValue(C);
1685 // This is the block mask. We OR all incoming edges, and with zero.
1686 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1687 VectorParts BlockMask = getVectorValue(Zero);
1690 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1691 VectorParts EM = createEdgeMask(*it, BB);
1692 for (unsigned part = 0; part < UF; ++part)
1693 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1700 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1701 BasicBlock *BB, PhiVector *PV) {
1702 // For each instruction in the old loop.
1703 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1704 VectorParts &Entry = WidenMap.get(it);
1705 switch (it->getOpcode()) {
1706 case Instruction::Br:
1707 // Nothing to do for PHIs and BR, since we already took care of the
1708 // loop control flow instructions.
1710 case Instruction::PHI:{
1711 PHINode* P = cast<PHINode>(it);
1712 // Handle reduction variables:
1713 if (Legal->getReductionVars()->count(P)) {
1714 for (unsigned part = 0; part < UF; ++part) {
1715 // This is phase one of vectorizing PHIs.
1716 Type *VecTy = VectorType::get(it->getType(), VF);
1717 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1718 LoopVectorBody-> getFirstInsertionPt());
1724 // Check for PHI nodes that are lowered to vector selects.
1725 if (P->getParent() != OrigLoop->getHeader()) {
1726 // We know that all PHIs in non header blocks are converted into
1727 // selects, so we don't have to worry about the insertion order and we
1728 // can just use the builder.
1730 // At this point we generate the predication tree. There may be
1731 // duplications since this is a simple recursive scan, but future
1732 // optimizations will clean it up.
1733 VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
1736 for (unsigned part = 0; part < UF; ++part) {
1737 VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
1738 VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
1739 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
1745 // This PHINode must be an induction variable.
1746 // Make sure that we know about it.
1747 assert(Legal->getInductionVars()->count(P) &&
1748 "Not an induction variable");
1750 LoopVectorizationLegality::InductionInfo II =
1751 Legal->getInductionVars()->lookup(P);
1754 case LoopVectorizationLegality::IK_NoInduction:
1755 llvm_unreachable("Unknown induction");
1756 case LoopVectorizationLegality::IK_IntInduction: {
1757 assert(P == OldInduction && "Unexpected PHI");
1758 Value *Broadcasted = getBroadcastInstrs(Induction);
1759 // After broadcasting the induction variable we need to make the
1760 // vector consecutive by adding 0, 1, 2 ...
1761 for (unsigned part = 0; part < UF; ++part)
1762 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
1765 case LoopVectorizationLegality::IK_ReverseIntInduction:
1766 case LoopVectorizationLegality::IK_PtrInduction:
1767 case LoopVectorizationLegality::IK_ReversePtrInduction:
1768 // Handle reverse integer and pointer inductions.
1769 Value *StartIdx = 0;
1770 // If we have a single integer induction variable then use it.
1771 // Otherwise, start counting at zero.
1773 LoopVectorizationLegality::InductionInfo OldII =
1774 Legal->getInductionVars()->lookup(OldInduction);
1775 StartIdx = OldII.StartValue;
1777 StartIdx = ConstantInt::get(Induction->getType(), 0);
1779 // This is the normalized GEP that starts counting at zero.
1780 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1783 // Handle the reverse integer induction variable case.
1784 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
1785 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
1786 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
1788 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
1791 // This is a new value so do not hoist it out.
1792 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
1793 // After broadcasting the induction variable we need to make the
1794 // vector consecutive by adding ... -3, -2, -1, 0.
1795 for (unsigned part = 0; part < UF; ++part)
1796 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
1800 // Handle the pointer induction variable case.
1801 assert(P->getType()->isPointerTy() && "Unexpected type.");
1803 // Is this a reverse induction ptr or a consecutive induction ptr.
1804 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
1807 // This is the vector of results. Notice that we don't generate
1808 // vector geps because scalar geps result in better code.
1809 for (unsigned part = 0; part < UF; ++part) {
1810 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1811 for (unsigned int i = 0; i < VF; ++i) {
1812 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
1813 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
1816 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1818 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
1820 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
1822 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1823 Builder.getInt32(i),
1826 Entry[part] = VecVal;
1833 case Instruction::Add:
1834 case Instruction::FAdd:
1835 case Instruction::Sub:
1836 case Instruction::FSub:
1837 case Instruction::Mul:
1838 case Instruction::FMul:
1839 case Instruction::UDiv:
1840 case Instruction::SDiv:
1841 case Instruction::FDiv:
1842 case Instruction::URem:
1843 case Instruction::SRem:
1844 case Instruction::FRem:
1845 case Instruction::Shl:
1846 case Instruction::LShr:
1847 case Instruction::AShr:
1848 case Instruction::And:
1849 case Instruction::Or:
1850 case Instruction::Xor: {
1851 // Just widen binops.
1852 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
1853 VectorParts &A = getVectorValue(it->getOperand(0));
1854 VectorParts &B = getVectorValue(it->getOperand(1));
1856 // Use this vector value for all users of the original instruction.
1857 for (unsigned Part = 0; Part < UF; ++Part) {
1858 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
1860 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
1861 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
1862 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
1863 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1864 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1866 if (VecOp && isa<PossiblyExactOperator>(VecOp))
1867 VecOp->setIsExact(BinOp->isExact());
1873 case Instruction::Select: {
1875 // If the selector is loop invariant we can create a select
1876 // instruction with a scalar condition. Otherwise, use vector-select.
1877 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
1880 // The condition can be loop invariant but still defined inside the
1881 // loop. This means that we can't just use the original 'cond' value.
1882 // We have to take the 'vectorized' value and pick the first lane.
1883 // Instcombine will make this a no-op.
1884 VectorParts &Cond = getVectorValue(it->getOperand(0));
1885 VectorParts &Op0 = getVectorValue(it->getOperand(1));
1886 VectorParts &Op1 = getVectorValue(it->getOperand(2));
1887 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
1888 Builder.getInt32(0));
1889 for (unsigned Part = 0; Part < UF; ++Part) {
1890 Entry[Part] = Builder.CreateSelect(
1891 InvariantCond ? ScalarCond : Cond[Part],
1898 case Instruction::ICmp:
1899 case Instruction::FCmp: {
1900 // Widen compares. Generate vector compares.
1901 bool FCmp = (it->getOpcode() == Instruction::FCmp);
1902 CmpInst *Cmp = dyn_cast<CmpInst>(it);
1903 VectorParts &A = getVectorValue(it->getOperand(0));
1904 VectorParts &B = getVectorValue(it->getOperand(1));
1905 for (unsigned Part = 0; Part < UF; ++Part) {
1908 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
1910 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
1916 case Instruction::Store:
1917 case Instruction::Load:
1918 vectorizeMemoryInstruction(it, Legal);
1920 case Instruction::ZExt:
1921 case Instruction::SExt:
1922 case Instruction::FPToUI:
1923 case Instruction::FPToSI:
1924 case Instruction::FPExt:
1925 case Instruction::PtrToInt:
1926 case Instruction::IntToPtr:
1927 case Instruction::SIToFP:
1928 case Instruction::UIToFP:
1929 case Instruction::Trunc:
1930 case Instruction::FPTrunc:
1931 case Instruction::BitCast: {
1932 CastInst *CI = dyn_cast<CastInst>(it);
1933 /// Optimize the special case where the source is the induction
1934 /// variable. Notice that we can only optimize the 'trunc' case
1935 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
1936 /// c. other casts depend on pointer size.
1937 if (CI->getOperand(0) == OldInduction &&
1938 it->getOpcode() == Instruction::Trunc) {
1939 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
1941 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
1942 for (unsigned Part = 0; Part < UF; ++Part)
1943 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
1946 /// Vectorize casts.
1947 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1949 VectorParts &A = getVectorValue(it->getOperand(0));
1950 for (unsigned Part = 0; Part < UF; ++Part)
1951 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
1955 case Instruction::Call: {
1956 assert(isTriviallyVectorizableIntrinsic(it));
1957 Module *M = BB->getParent()->getParent();
1958 IntrinsicInst *II = cast<IntrinsicInst>(it);
1959 Intrinsic::ID ID = II->getIntrinsicID();
1960 for (unsigned Part = 0; Part < UF; ++Part) {
1961 SmallVector<Value*, 4> Args;
1962 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) {
1963 VectorParts &Arg = getVectorValue(II->getArgOperand(i));
1964 Args.push_back(Arg[Part]);
1966 Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) };
1967 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
1968 Entry[Part] = Builder.CreateCall(F, Args);
1974 // All other instructions are unsupported. Scalarize them.
1975 scalarizeInstruction(it);
1978 }// end of for_each instr.
1981 void InnerLoopVectorizer::updateAnalysis() {
1982 // Forget the original basic block.
1983 SE->forgetLoop(OrigLoop);
1985 // Update the dominator tree information.
1986 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
1987 "Entry does not dominate exit.");
1989 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1990 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
1991 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
1992 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1993 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
1994 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1995 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1996 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1998 DEBUG(DT->verifyAnalysis());
2001 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2002 if (!EnableIfConversion)
2005 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2006 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2008 // Collect the blocks that need predication.
2009 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2010 BasicBlock *BB = LoopBlocks[i];
2012 // We don't support switch statements inside loops.
2013 if (!isa<BranchInst>(BB->getTerminator()))
2016 // We must have at most two predecessors because we need to convert
2017 // all PHIs to selects.
2018 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
2022 // We must be able to predicate all blocks that need to be predicated.
2023 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2027 // We can if-convert this loop.
2031 bool LoopVectorizationLegality::canVectorize() {
2032 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2034 // We can only vectorize innermost loops.
2035 if (TheLoop->getSubLoopsVector().size())
2038 // We must have a single backedge.
2039 if (TheLoop->getNumBackEdges() != 1)
2042 // We must have a single exiting block.
2043 if (!TheLoop->getExitingBlock())
2046 unsigned NumBlocks = TheLoop->getNumBlocks();
2048 // Check if we can if-convert non single-bb loops.
2049 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2050 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2054 // We need to have a loop header.
2055 BasicBlock *Latch = TheLoop->getLoopLatch();
2056 DEBUG(dbgs() << "LV: Found a loop: " <<
2057 TheLoop->getHeader()->getName() << "\n");
2059 // ScalarEvolution needs to be able to find the exit count.
2060 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2061 if (ExitCount == SE->getCouldNotCompute()) {
2062 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2066 // Do not loop-vectorize loops with a tiny trip count.
2067 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2068 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2069 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2070 "This loop is not worth vectorizing.\n");
2074 // Check if we can vectorize the instructions and CFG in this loop.
2075 if (!canVectorizeInstrs()) {
2076 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2080 // Go over each instruction and look at memory deps.
2081 if (!canVectorizeMemory()) {
2082 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2086 // Collect all of the variables that remain uniform after vectorization.
2087 collectLoopUniforms();
2089 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2090 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2093 // Okay! We can vectorize. At this point we don't have any other mem analysis
2094 // which may limit our maximum vectorization factor, so just return true with
2099 bool LoopVectorizationLegality::canVectorizeInstrs() {
2100 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2101 BasicBlock *Header = TheLoop->getHeader();
2103 // For each block in the loop.
2104 for (Loop::block_iterator bb = TheLoop->block_begin(),
2105 be = TheLoop->block_end(); bb != be; ++bb) {
2107 // Scan the instructions in the block and look for hazards.
2108 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2111 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2112 // This should not happen because the loop should be normalized.
2113 if (Phi->getNumIncomingValues() != 2) {
2114 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2118 // Check that this PHI type is allowed.
2119 if (!Phi->getType()->isIntegerTy() &&
2120 !Phi->getType()->isFloatingPointTy() &&
2121 !Phi->getType()->isPointerTy()) {
2122 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2126 // If this PHINode is not in the header block, then we know that we
2127 // can convert it to select during if-conversion. No need to check if
2128 // the PHIs in this block are induction or reduction variables.
2132 // This is the value coming from the preheader.
2133 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2134 // Check if this is an induction variable.
2135 InductionKind IK = isInductionVariable(Phi);
2137 if (IK_NoInduction != IK) {
2138 // Int inductions are special because we only allow one IV.
2139 if (IK == IK_IntInduction) {
2141 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2147 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2148 Inductions[Phi] = InductionInfo(StartValue, IK);
2152 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2153 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2156 if (AddReductionVar(Phi, RK_IntegerMult)) {
2157 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2160 if (AddReductionVar(Phi, RK_IntegerOr)) {
2161 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2164 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2165 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2168 if (AddReductionVar(Phi, RK_IntegerXor)) {
2169 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2172 if (AddReductionVar(Phi, RK_FloatMult)) {
2173 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2176 if (AddReductionVar(Phi, RK_FloatAdd)) {
2177 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2181 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2183 }// end of PHI handling
2185 // We still don't handle functions.
2186 CallInst *CI = dyn_cast<CallInst>(it);
2187 if (CI && !isTriviallyVectorizableIntrinsic(it)) {
2188 DEBUG(dbgs() << "LV: Found a call site.\n");
2192 // Check that the instruction return type is vectorizable.
2193 if (!VectorType::isValidElementType(it->getType()) &&
2194 !it->getType()->isVoidTy()) {
2195 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2199 // Check that the stored type is vectorizable.
2200 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2201 Type *T = ST->getValueOperand()->getType();
2202 if (!VectorType::isValidElementType(T))
2206 // Reduction instructions are allowed to have exit users.
2207 // All other instructions must not have external users.
2208 if (!AllowedExit.count(it))
2209 //Check that all of the users of the loop are inside the BB.
2210 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2212 Instruction *U = cast<Instruction>(*I);
2213 // This user may be a reduction exit value.
2214 if (!TheLoop->contains(U)) {
2215 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2224 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2225 assert(getInductionVars()->size() && "No induction variables");
2231 void LoopVectorizationLegality::collectLoopUniforms() {
2232 // We now know that the loop is vectorizable!
2233 // Collect variables that will remain uniform after vectorization.
2234 std::vector<Value*> Worklist;
2235 BasicBlock *Latch = TheLoop->getLoopLatch();
2237 // Start with the conditional branch and walk up the block.
2238 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2240 while (Worklist.size()) {
2241 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2242 Worklist.pop_back();
2244 // Look at instructions inside this loop.
2245 // Stop when reaching PHI nodes.
2246 // TODO: we need to follow values all over the loop, not only in this block.
2247 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2250 // This is a known uniform.
2253 // Insert all operands.
2254 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2255 Worklist.push_back(I->getOperand(i));
2260 bool LoopVectorizationLegality::canVectorizeMemory() {
2261 typedef SmallVector<Value*, 16> ValueVector;
2262 typedef SmallPtrSet<Value*, 16> ValueSet;
2263 // Holds the Load and Store *instructions*.
2266 PtrRtCheck.Pointers.clear();
2267 PtrRtCheck.Need = false;
2270 for (Loop::block_iterator bb = TheLoop->block_begin(),
2271 be = TheLoop->block_end(); bb != be; ++bb) {
2273 // Scan the BB and collect legal loads and stores.
2274 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2277 // If this is a load, save it. If this instruction can read from memory
2278 // but is not a load, then we quit. Notice that we don't handle function
2279 // calls that read or write.
2280 if (it->mayReadFromMemory()) {
2281 LoadInst *Ld = dyn_cast<LoadInst>(it);
2282 if (!Ld) return false;
2283 if (!Ld->isSimple()) {
2284 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2287 Loads.push_back(Ld);
2291 // Save 'store' instructions. Abort if other instructions write to memory.
2292 if (it->mayWriteToMemory()) {
2293 StoreInst *St = dyn_cast<StoreInst>(it);
2294 if (!St) return false;
2295 if (!St->isSimple()) {
2296 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2299 Stores.push_back(St);
2304 // Now we have two lists that hold the loads and the stores.
2305 // Next, we find the pointers that they use.
2307 // Check if we see any stores. If there are no stores, then we don't
2308 // care if the pointers are *restrict*.
2309 if (!Stores.size()) {
2310 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2314 // Holds the read and read-write *pointers* that we find.
2316 ValueVector ReadWrites;
2318 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2319 // multiple times on the same object. If the ptr is accessed twice, once
2320 // for read and once for write, it will only appear once (on the write
2321 // list). This is okay, since we are going to check for conflicts between
2322 // writes and between reads and writes, but not between reads and reads.
2325 ValueVector::iterator I, IE;
2326 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2327 StoreInst *ST = cast<StoreInst>(*I);
2328 Value* Ptr = ST->getPointerOperand();
2330 if (isUniform(Ptr)) {
2331 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2335 // If we did *not* see this pointer before, insert it to
2336 // the read-write list. At this phase it is only a 'write' list.
2337 if (Seen.insert(Ptr))
2338 ReadWrites.push_back(Ptr);
2341 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2342 LoadInst *LD = cast<LoadInst>(*I);
2343 Value* Ptr = LD->getPointerOperand();
2344 // If we did *not* see this pointer before, insert it to the
2345 // read list. If we *did* see it before, then it is already in
2346 // the read-write list. This allows us to vectorize expressions
2347 // such as A[i] += x; Because the address of A[i] is a read-write
2348 // pointer. This only works if the index of A[i] is consecutive.
2349 // If the address of i is unknown (for example A[B[i]]) then we may
2350 // read a few words, modify, and write a few words, and some of the
2351 // words may be written to the same address.
2352 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2353 Reads.push_back(Ptr);
2356 // If we write (or read-write) to a single destination and there are no
2357 // other reads in this loop then is it safe to vectorize.
2358 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2359 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2363 // Find pointers with computable bounds. We are going to use this information
2364 // to place a runtime bound check.
2365 bool CanDoRT = true;
2366 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
2367 if (hasComputableBounds(*I)) {
2368 PtrRtCheck.insert(SE, TheLoop, *I);
2369 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
2374 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
2375 if (hasComputableBounds(*I)) {
2376 PtrRtCheck.insert(SE, TheLoop, *I);
2377 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
2383 // Check that we did not collect too many pointers or found a
2384 // unsizeable pointer.
2385 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
2391 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2394 bool NeedRTCheck = false;
2396 // Now that the pointers are in two lists (Reads and ReadWrites), we
2397 // can check that there are no conflicts between each of the writes and
2398 // between the writes to the reads.
2399 ValueSet WriteObjects;
2400 ValueVector TempObjects;
2402 // Check that the read-writes do not conflict with other read-write
2404 bool AllWritesIdentified = true;
2405 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
2406 GetUnderlyingObjects(*I, TempObjects, DL);
2407 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
2409 if (!isIdentifiedObject(*it)) {
2410 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
2412 AllWritesIdentified = false;
2414 if (!WriteObjects.insert(*it)) {
2415 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2420 TempObjects.clear();
2423 /// Check that the reads don't conflict with the read-writes.
2424 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
2425 GetUnderlyingObjects(*I, TempObjects, DL);
2426 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
2428 // If all of the writes are identified then we don't care if the read
2429 // pointer is identified or not.
2430 if (!AllWritesIdentified && !isIdentifiedObject(*it)) {
2431 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
2434 if (WriteObjects.count(*it)) {
2435 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
2440 TempObjects.clear();
2443 PtrRtCheck.Need = NeedRTCheck;
2444 if (NeedRTCheck && !CanDoRT) {
2445 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2446 "the array bounds.\n");
2451 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2452 " need a runtime memory check.\n");
2456 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2457 ReductionKind Kind) {
2458 if (Phi->getNumIncomingValues() != 2)
2461 // Reduction variables are only found in the loop header block.
2462 if (Phi->getParent() != TheLoop->getHeader())
2465 // Obtain the reduction start value from the value that comes from the loop
2467 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2469 // ExitInstruction is the single value which is used outside the loop.
2470 // We only allow for a single reduction value to be used outside the loop.
2471 // This includes users of the reduction, variables (which form a cycle
2472 // which ends in the phi node).
2473 Instruction *ExitInstruction = 0;
2474 // Indicates that we found a binary operation in our scan.
2475 bool FoundBinOp = false;
2477 // Iter is our iterator. We start with the PHI node and scan for all of the
2478 // users of this instruction. All users must be instructions that can be
2479 // used as reduction variables (such as ADD). We may have a single
2480 // out-of-block user. The cycle must end with the original PHI.
2481 Instruction *Iter = Phi;
2483 // If the instruction has no users then this is a broken
2484 // chain and can't be a reduction variable.
2485 if (Iter->use_empty())
2488 // Did we find a user inside this loop already ?
2489 bool FoundInBlockUser = false;
2490 // Did we reach the initial PHI node already ?
2491 bool FoundStartPHI = false;
2493 // Is this a bin op ?
2494 FoundBinOp |= !isa<PHINode>(Iter);
2496 // For each of the *users* of iter.
2497 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2499 Instruction *U = cast<Instruction>(*it);
2500 // We already know that the PHI is a user.
2502 FoundStartPHI = true;
2506 // Check if we found the exit user.
2507 BasicBlock *Parent = U->getParent();
2508 if (!TheLoop->contains(Parent)) {
2509 // Exit if you find multiple outside users.
2510 if (ExitInstruction != 0)
2512 ExitInstruction = Iter;
2515 // We allow in-loop PHINodes which are not the original reduction PHI
2516 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2517 // structure) then don't skip this PHI.
2518 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2519 U->getParent() != TheLoop->getHeader() &&
2520 TheLoop->contains(U) &&
2521 Iter->getNumUses() > 1)
2524 // We can't have multiple inside users.
2525 if (FoundInBlockUser)
2527 FoundInBlockUser = true;
2529 // Any reduction instr must be of one of the allowed kinds.
2530 if (!isReductionInstr(U, Kind))
2533 // Reductions of instructions such as Div, and Sub is only
2534 // possible if the LHS is the reduction variable.
2535 if (!U->isCommutative() && !isa<PHINode>(U) && U->getOperand(0) != Iter)
2541 // We found a reduction var if we have reached the original
2542 // phi node and we only have a single instruction with out-of-loop
2544 if (FoundStartPHI) {
2545 // This instruction is allowed to have out-of-loop users.
2546 AllowedExit.insert(ExitInstruction);
2548 // Save the description of this reduction variable.
2549 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
2550 Reductions[Phi] = RD;
2551 // We've ended the cycle. This is a reduction variable if we have an
2552 // outside user and it has a binary op.
2553 return FoundBinOp && ExitInstruction;
2559 LoopVectorizationLegality::isReductionInstr(Instruction *I,
2560 ReductionKind Kind) {
2561 bool FP = I->getType()->isFloatingPointTy();
2562 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
2564 switch (I->getOpcode()) {
2567 case Instruction::PHI:
2568 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd))
2572 case Instruction::Sub:
2573 case Instruction::Add:
2574 return Kind == RK_IntegerAdd;
2575 case Instruction::SDiv:
2576 case Instruction::UDiv:
2577 case Instruction::Mul:
2578 return Kind == RK_IntegerMult;
2579 case Instruction::And:
2580 return Kind == RK_IntegerAnd;
2581 case Instruction::Or:
2582 return Kind == RK_IntegerOr;
2583 case Instruction::Xor:
2584 return Kind == RK_IntegerXor;
2585 case Instruction::FMul:
2586 return Kind == RK_FloatMult && FastMath;
2587 case Instruction::FAdd:
2588 return Kind == RK_FloatAdd && FastMath;
2592 LoopVectorizationLegality::InductionKind
2593 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
2594 Type *PhiTy = Phi->getType();
2595 // We only handle integer and pointer inductions variables.
2596 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
2597 return IK_NoInduction;
2599 // Check that the PHI is consecutive.
2600 const SCEV *PhiScev = SE->getSCEV(Phi);
2601 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2603 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
2604 return IK_NoInduction;
2606 const SCEV *Step = AR->getStepRecurrence(*SE);
2608 // Integer inductions need to have a stride of one.
2609 if (PhiTy->isIntegerTy()) {
2611 return IK_IntInduction;
2612 if (Step->isAllOnesValue())
2613 return IK_ReverseIntInduction;
2614 return IK_NoInduction;
2617 // Calculate the pointer stride and check if it is consecutive.
2618 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
2620 return IK_NoInduction;
2622 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
2623 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
2624 if (C->getValue()->equalsInt(Size))
2625 return IK_PtrInduction;
2626 else if (C->getValue()->equalsInt(0 - Size))
2627 return IK_ReversePtrInduction;
2629 return IK_NoInduction;
2632 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
2633 Value *In0 = const_cast<Value*>(V);
2634 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
2638 return Inductions.count(PN);
2641 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
2642 assert(TheLoop->contains(BB) && "Unknown block used");
2644 // Blocks that do not dominate the latch need predication.
2645 BasicBlock* Latch = TheLoop->getLoopLatch();
2646 return !DT->dominates(BB, Latch);
2649 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
2650 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2651 // We don't predicate loads/stores at the moment.
2652 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
2655 // The instructions below can trap.
2656 switch (it->getOpcode()) {
2658 case Instruction::UDiv:
2659 case Instruction::SDiv:
2660 case Instruction::URem:
2661 case Instruction::SRem:
2669 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
2670 const SCEV *PhiScev = SE->getSCEV(Ptr);
2671 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2675 return AR->isAffine();
2678 std::pair<unsigned, unsigned>
2679 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
2681 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
2682 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
2683 return std::make_pair(1U, 0U);
2686 // Find the trip count.
2687 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
2688 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
2690 unsigned WidestType = getWidestType();
2691 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
2692 unsigned MaxVectorSize = WidestRegister / WidestType;
2693 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
2694 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
2696 if (MaxVectorSize == 0) {
2697 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
2701 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
2702 " into one vector!");
2704 unsigned VF = MaxVectorSize;
2706 // If we optimize the program for size, avoid creating the tail loop.
2708 // If we are unable to calculate the trip count then don't try to vectorize.
2710 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2711 return std::make_pair(1U, 0U);
2714 // Find the maximum SIMD width that can fit within the trip count.
2715 VF = TC % MaxVectorSize;
2720 // If the trip count that we found modulo the vectorization factor is not
2721 // zero then we require a tail.
2723 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2724 return std::make_pair(1U, 0U);
2729 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
2730 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
2732 return std::make_pair(UserVF, 0U);
2735 float Cost = expectedCost(1);
2737 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
2738 for (unsigned i=2; i <= VF; i*=2) {
2739 // Notice that the vector loop needs to be executed less times, so
2740 // we need to divide the cost of the vector loops by the width of
2741 // the vector elements.
2742 float VectorCost = expectedCost(i) / (float)i;
2743 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
2744 (int)VectorCost << ".\n");
2745 if (VectorCost < Cost) {
2751 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
2752 unsigned LoopCost = VF * Cost;
2753 return std::make_pair(Width, LoopCost);
2756 unsigned LoopVectorizationCostModel::getWidestType() {
2757 unsigned MaxWidth = 8;
2760 for (Loop::block_iterator bb = TheLoop->block_begin(),
2761 be = TheLoop->block_end(); bb != be; ++bb) {
2762 BasicBlock *BB = *bb;
2764 // For each instruction in the loop.
2765 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2766 Type *T = it->getType();
2768 // Only examine Loads, Stores and PHINodes.
2769 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
2772 // Examine PHI nodes that are reduction variables.
2773 if (PHINode *PN = dyn_cast<PHINode>(it))
2774 if (!Legal->getReductionVars()->count(PN))
2777 // Examine the stored values.
2778 if (StoreInst *ST = dyn_cast<StoreInst>(it))
2779 T = ST->getValueOperand()->getType();
2781 // Ignore stored/loaded pointer types.
2782 if (T->isPointerTy())
2785 MaxWidth = std::max(MaxWidth, T->getScalarSizeInBits());
2793 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
2796 unsigned LoopCost) {
2798 // -- The unroll heuristics --
2799 // We unroll the loop in order to expose ILP and reduce the loop overhead.
2800 // There are many micro-architectural considerations that we can't predict
2801 // at this level. For example frontend pressure (on decode or fetch) due to
2802 // code size, or the number and capabilities of the execution ports.
2804 // We use the following heuristics to select the unroll factor:
2805 // 1. If the code has reductions the we unroll in order to break the cross
2806 // iteration dependency.
2807 // 2. If the loop is really small then we unroll in order to reduce the loop
2809 // 3. We don't unroll if we think that we will spill registers to memory due
2810 // to the increased register pressure.
2812 // Use the user preference, unless 'auto' is selected.
2816 // When we optimize for size we don't unroll.
2820 // Do not unroll loops with a relatively small trip count.
2821 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
2822 TheLoop->getLoopLatch());
2823 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
2826 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
2827 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
2828 " vector registers\n");
2830 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
2831 // We divide by these constants so assume that we have at least one
2832 // instruction that uses at least one register.
2833 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
2834 R.NumInstructions = std::max(R.NumInstructions, 1U);
2836 // We calculate the unroll factor using the following formula.
2837 // Subtract the number of loop invariants from the number of available
2838 // registers. These registers are used by all of the unrolled instances.
2839 // Next, divide the remaining registers by the number of registers that is
2840 // required by the loop, in order to estimate how many parallel instances
2841 // fit without causing spills.
2842 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
2844 // Clamp the unroll factor ranges to reasonable factors.
2845 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
2847 // If we did not calculate the cost for VF (because the user selected the VF)
2848 // then we calculate the cost of VF here.
2850 LoopCost = expectedCost(VF);
2852 // Clamp the calculated UF to be between the 1 and the max unroll factor
2853 // that the target allows.
2854 if (UF > MaxUnrollSize)
2859 if (Legal->getReductionVars()->size()) {
2860 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
2864 // We want to unroll tiny loops in order to reduce the loop overhead.
2865 // We assume that the cost overhead is 1 and we use the cost model
2866 // to estimate the cost of the loop and unroll until the cost of the
2867 // loop overhead is about 5% of the cost of the loop.
2868 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
2869 if (LoopCost < 20) {
2870 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
2871 unsigned NewUF = 20/LoopCost + 1;
2872 return std::min(NewUF, UF);
2875 DEBUG(dbgs() << "LV: Not Unrolling. \n");
2879 LoopVectorizationCostModel::RegisterUsage
2880 LoopVectorizationCostModel::calculateRegisterUsage() {
2881 // This function calculates the register usage by measuring the highest number
2882 // of values that are alive at a single location. Obviously, this is a very
2883 // rough estimation. We scan the loop in a topological order in order and
2884 // assign a number to each instruction. We use RPO to ensure that defs are
2885 // met before their users. We assume that each instruction that has in-loop
2886 // users starts an interval. We record every time that an in-loop value is
2887 // used, so we have a list of the first and last occurrences of each
2888 // instruction. Next, we transpose this data structure into a multi map that
2889 // holds the list of intervals that *end* at a specific location. This multi
2890 // map allows us to perform a linear search. We scan the instructions linearly
2891 // and record each time that a new interval starts, by placing it in a set.
2892 // If we find this value in the multi-map then we remove it from the set.
2893 // The max register usage is the maximum size of the set.
2894 // We also search for instructions that are defined outside the loop, but are
2895 // used inside the loop. We need this number separately from the max-interval
2896 // usage number because when we unroll, loop-invariant values do not take
2898 LoopBlocksDFS DFS(TheLoop);
2902 R.NumInstructions = 0;
2904 // Each 'key' in the map opens a new interval. The values
2905 // of the map are the index of the 'last seen' usage of the
2906 // instruction that is the key.
2907 typedef DenseMap<Instruction*, unsigned> IntervalMap;
2908 // Maps instruction to its index.
2909 DenseMap<unsigned, Instruction*> IdxToInstr;
2910 // Marks the end of each interval.
2911 IntervalMap EndPoint;
2912 // Saves the list of instruction indices that are used in the loop.
2913 SmallSet<Instruction*, 8> Ends;
2914 // Saves the list of values that are used in the loop but are
2915 // defined outside the loop, such as arguments and constants.
2916 SmallPtrSet<Value*, 8> LoopInvariants;
2919 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2920 be = DFS.endRPO(); bb != be; ++bb) {
2921 R.NumInstructions += (*bb)->size();
2922 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2924 Instruction *I = it;
2925 IdxToInstr[Index++] = I;
2927 // Save the end location of each USE.
2928 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
2929 Value *U = I->getOperand(i);
2930 Instruction *Instr = dyn_cast<Instruction>(U);
2932 // Ignore non-instruction values such as arguments, constants, etc.
2933 if (!Instr) continue;
2935 // If this instruction is outside the loop then record it and continue.
2936 if (!TheLoop->contains(Instr)) {
2937 LoopInvariants.insert(Instr);
2941 // Overwrite previous end points.
2942 EndPoint[Instr] = Index;
2948 // Saves the list of intervals that end with the index in 'key'.
2949 typedef SmallVector<Instruction*, 2> InstrList;
2950 DenseMap<unsigned, InstrList> TransposeEnds;
2952 // Transpose the EndPoints to a list of values that end at each index.
2953 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
2955 TransposeEnds[it->second].push_back(it->first);
2957 SmallSet<Instruction*, 8> OpenIntervals;
2958 unsigned MaxUsage = 0;
2961 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
2962 for (unsigned int i = 0; i < Index; ++i) {
2963 Instruction *I = IdxToInstr[i];
2964 // Ignore instructions that are never used within the loop.
2965 if (!Ends.count(I)) continue;
2967 // Remove all of the instructions that end at this location.
2968 InstrList &List = TransposeEnds[i];
2969 for (unsigned int j=0, e = List.size(); j < e; ++j)
2970 OpenIntervals.erase(List[j]);
2972 // Count the number of live interals.
2973 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
2975 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
2976 OpenIntervals.size() <<"\n");
2978 // Add the current instruction to the list of open intervals.
2979 OpenIntervals.insert(I);
2982 unsigned Invariant = LoopInvariants.size();
2983 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
2984 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
2985 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
2987 R.LoopInvariantRegs = Invariant;
2988 R.MaxLocalUsers = MaxUsage;
2992 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
2996 for (Loop::block_iterator bb = TheLoop->block_begin(),
2997 be = TheLoop->block_end(); bb != be; ++bb) {
2998 unsigned BlockCost = 0;
2999 BasicBlock *BB = *bb;
3001 // For each instruction in the old loop.
3002 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3003 unsigned C = getInstructionCost(it, VF);
3005 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3006 VF << " For instruction: "<< *it << "\n");
3009 // We assume that if-converted blocks have a 50% chance of being executed.
3010 // When the code is scalar then some of the blocks are avoided due to CF.
3011 // When the code is vectorized we execute all code paths.
3012 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3022 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3023 // If we know that this instruction will remain uniform, check the cost of
3024 // the scalar version.
3025 if (Legal->isUniformAfterVectorization(I))
3028 Type *RetTy = I->getType();
3029 Type *VectorTy = ToVectorTy(RetTy, VF);
3031 // TODO: We need to estimate the cost of intrinsic calls.
3032 switch (I->getOpcode()) {
3033 case Instruction::GetElementPtr:
3034 // We mark this instruction as zero-cost because scalar GEPs are usually
3035 // lowered to the intruction addressing mode. At the moment we don't
3036 // generate vector geps.
3038 case Instruction::Br: {
3039 return TTI.getCFInstrCost(I->getOpcode());
3041 case Instruction::PHI:
3042 //TODO: IF-converted IFs become selects.
3044 case Instruction::Add:
3045 case Instruction::FAdd:
3046 case Instruction::Sub:
3047 case Instruction::FSub:
3048 case Instruction::Mul:
3049 case Instruction::FMul:
3050 case Instruction::UDiv:
3051 case Instruction::SDiv:
3052 case Instruction::FDiv:
3053 case Instruction::URem:
3054 case Instruction::SRem:
3055 case Instruction::FRem:
3056 case Instruction::Shl:
3057 case Instruction::LShr:
3058 case Instruction::AShr:
3059 case Instruction::And:
3060 case Instruction::Or:
3061 case Instruction::Xor:
3062 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy);
3063 case Instruction::Select: {
3064 SelectInst *SI = cast<SelectInst>(I);
3065 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3066 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3067 Type *CondTy = SI->getCondition()->getType();
3069 CondTy = VectorType::get(CondTy, VF);
3071 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3073 case Instruction::ICmp:
3074 case Instruction::FCmp: {
3075 Type *ValTy = I->getOperand(0)->getType();
3076 VectorTy = ToVectorTy(ValTy, VF);
3077 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3079 case Instruction::Store: {
3080 StoreInst *SI = cast<StoreInst>(I);
3081 Type *ValTy = SI->getValueOperand()->getType();
3082 VectorTy = ToVectorTy(ValTy, VF);
3085 return TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
3087 SI->getPointerAddressSpace());
3089 // Scalarized stores.
3090 int Stride = Legal->isConsecutivePtr(SI->getPointerOperand());
3091 bool Reverse = Stride < 0;
3095 // The cost of extracting from the value vector and pointer vector.
3096 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3097 for (unsigned i = 0; i < VF; ++i) {
3098 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy,
3100 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3103 // The cost of the scalar stores.
3104 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3106 SI->getPointerAddressSpace());
3111 unsigned Cost = TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
3113 SI->getPointerAddressSpace());
3115 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3119 case Instruction::Load: {
3120 LoadInst *LI = cast<LoadInst>(I);
3123 return TTI.getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
3124 LI->getPointerAddressSpace());
3126 // Scalarized loads.
3127 int Stride = Legal->isConsecutivePtr(LI->getPointerOperand());
3128 bool Reverse = Stride < 0;
3131 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3133 // The cost of extracting from the pointer vector.
3134 for (unsigned i = 0; i < VF; ++i)
3135 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3137 // The cost of inserting data to the result vector.
3138 for (unsigned i = 0; i < VF; ++i)
3139 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, VectorTy, i);
3141 // The cost of the scalar stores.
3142 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), RetTy->getScalarType(),
3144 LI->getPointerAddressSpace());
3149 unsigned Cost = TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
3151 LI->getPointerAddressSpace());
3153 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
3156 case Instruction::ZExt:
3157 case Instruction::SExt:
3158 case Instruction::FPToUI:
3159 case Instruction::FPToSI:
3160 case Instruction::FPExt:
3161 case Instruction::PtrToInt:
3162 case Instruction::IntToPtr:
3163 case Instruction::SIToFP:
3164 case Instruction::UIToFP:
3165 case Instruction::Trunc:
3166 case Instruction::FPTrunc:
3167 case Instruction::BitCast: {
3168 // We optimize the truncation of induction variable.
3169 // The cost of these is the same as the scalar operation.
3170 if (I->getOpcode() == Instruction::Trunc &&
3171 Legal->isInductionVariable(I->getOperand(0)))
3172 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3173 I->getOperand(0)->getType());
3175 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3176 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3178 case Instruction::Call: {
3179 assert(isTriviallyVectorizableIntrinsic(I));
3180 IntrinsicInst *II = cast<IntrinsicInst>(I);
3181 Type *RetTy = ToVectorTy(II->getType(), VF);
3182 SmallVector<Type*, 4> Tys;
3183 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i)
3184 Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF));
3185 return TTI.getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys);
3188 // We are scalarizing the instruction. Return the cost of the scalar
3189 // instruction, plus the cost of insert and extract into vector
3190 // elements, times the vector width.
3193 if (!RetTy->isVoidTy() && VF != 1) {
3194 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3196 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3199 // The cost of inserting the results plus extracting each one of the
3201 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3204 // The cost of executing VF copies of the scalar instruction. This opcode
3205 // is unknown. Assume that it is the same as 'mul'.
3206 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3212 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3213 if (Scalar->isVoidTy() || VF == 1)
3215 return VectorType::get(Scalar, VF);
3218 char LoopVectorize::ID = 0;
3219 static const char lv_name[] = "Loop Vectorization";
3220 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3221 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3222 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3223 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3224 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3225 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3228 Pass *createLoopVectorizePass() {
3229 return new LoopVectorize();