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. SingleBlockLoopVectorizer - 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.
28 //===----------------------------------------------------------------------===//
30 // The reduction-variable vectorization is based on the paper:
31 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33 // Variable uniformity checks are inspired by:
34 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
36 // Other ideas/concepts are from:
37 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
39 //===----------------------------------------------------------------------===//
40 #define LV_NAME "loop-vectorize"
41 #define DEBUG_TYPE LV_NAME
42 #include "llvm/Constants.h"
43 #include "llvm/DerivedTypes.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/LLVMContext.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Analysis/LoopPass.h"
48 #include "llvm/Value.h"
49 #include "llvm/Function.h"
50 #include "llvm/Analysis/Verifier.h"
51 #include "llvm/Module.h"
52 #include "llvm/Type.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/ScalarEvolution.h"
58 #include "llvm/Analysis/Dominators.h"
59 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
60 #include "llvm/Analysis/ScalarEvolutionExpander.h"
61 #include "llvm/Analysis/LoopInfo.h"
62 #include "llvm/Analysis/ValueTracking.h"
63 #include "llvm/Transforms/Scalar.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/TargetTransformInfo.h"
66 #include "llvm/Support/CommandLine.h"
67 #include "llvm/Support/Debug.h"
68 #include "llvm/Support/raw_ostream.h"
69 #include "llvm/DataLayout.h"
70 #include "llvm/Transforms/Utils/Local.h"
74 static cl::opt<unsigned>
75 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
76 cl::desc("Set the default vectorization width. Zero is autoselect."));
78 /// We don't vectorize loops with a known constant trip count below this number.
79 const unsigned TinyTripCountThreshold = 16;
81 /// When performing a runtime memory check, do not check more than this
82 /// number of pointers. Notice that the check is quadratic!
83 const unsigned RuntimeMemoryCheckThreshold = 2;
87 // Forward declarations.
88 class LoopVectorizationLegality;
89 class LoopVectorizationCostModel;
91 /// SingleBlockLoopVectorizer vectorizes loops which contain only one basic
92 /// block to a specified vectorization factor (VF).
93 /// This class performs the widening of scalars into vectors, or multiple
94 /// scalars. This class also implements the following features:
95 /// * It inserts an epilogue loop for handling loops that don't have iteration
96 /// counts that are known to be a multiple of the vectorization factor.
97 /// * It handles the code generation for reduction variables.
98 /// * Scalarization (implementation using scalars) of un-vectorizable
100 /// SingleBlockLoopVectorizer does not perform any vectorization-legality
101 /// checks, and relies on the caller to check for the different legality
102 /// aspects. The SingleBlockLoopVectorizer relies on the
103 /// LoopVectorizationLegality class to provide information about the induction
104 /// and reduction variables that were found to a given vectorization factor.
105 class SingleBlockLoopVectorizer {
108 SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
109 DominatorTree *dt, LPPassManager *Lpm,
111 OrigLoop(Orig), SE(Se), LI(Li), DT(dt), LPM(Lpm), VF(VecWidth),
112 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
114 // Perform the actual loop widening (vectorization).
115 void vectorize(LoopVectorizationLegality *Legal) {
116 ///Create a new empty loop. Unlink the old loop and connect the new one.
117 createEmptyLoop(Legal);
118 /// Widen each instruction in the old loop to a new one in the new loop.
119 /// Use the Legality module to find the induction and reduction variables.
120 vectorizeLoop(Legal);
121 // Register the new loop and update the analysis passes.
126 /// Create an empty loop, based on the loop ranges of the old loop.
127 void createEmptyLoop(LoopVectorizationLegality *Legal);
128 /// Copy and widen the instructions from the old loop.
129 void vectorizeLoop(LoopVectorizationLegality *Legal);
130 /// Insert the new loop to the loop hierarchy and pass manager
131 /// and update the analysis passes.
132 void updateAnalysis();
134 /// This instruction is un-vectorizable. Implement it as a sequence
136 void scalarizeInstruction(Instruction *Instr);
138 /// Create a broadcast instruction. This method generates a broadcast
139 /// instruction (shuffle) for loop invariant values and for the induction
140 /// value. If this is the induction variable then we extend it to N, N+1, ...
141 /// this is needed because each iteration in the loop corresponds to a SIMD
143 Value *getBroadcastInstrs(Value *V);
145 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
146 /// for each element in the vector. Starting from zero.
147 Value *getConsecutiveVector(Value* Val);
149 /// When we go over instructions in the basic block we rely on previous
150 /// values within the current basic block or on loop invariant values.
151 /// When we widen (vectorize) values we place them in the map. If the values
152 /// are not within the map, they have to be loop invariant, so we simply
153 /// broadcast them into a vector.
154 Value *getVectorValue(Value *V);
156 /// Get a uniform vector of constant integers. We use this to get
157 /// vectors of ones and zeros for the reduction code.
158 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
160 typedef DenseMap<Value*, Value*> ValueMap;
162 /// The original loop.
164 // Scev analysis to use.
170 // Loop Pass Manager;
172 // The vectorization factor to use.
175 // The builder that we use
178 // --- Vectorization state ---
180 /// The vector-loop preheader.
181 BasicBlock *LoopVectorPreHeader;
182 /// The scalar-loop preheader.
183 BasicBlock *LoopScalarPreHeader;
184 /// Middle Block between the vector and the scalar.
185 BasicBlock *LoopMiddleBlock;
186 ///The ExitBlock of the scalar loop.
187 BasicBlock *LoopExitBlock;
188 ///The vector loop body.
189 BasicBlock *LoopVectorBody;
190 ///The scalar loop body.
191 BasicBlock *LoopScalarBody;
192 ///The first bypass block.
193 BasicBlock *LoopBypassBlock;
195 /// The new Induction variable which was added to the new block.
197 /// The induction variable of the old basic block.
198 PHINode *OldInduction;
199 // Maps scalars to widened vectors.
203 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
204 /// to what vectorization factor.
205 /// This class does not look at the profitability of vectorization, only the
206 /// legality. This class has two main kinds of checks:
207 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
208 /// will change the order of memory accesses in a way that will change the
209 /// correctness of the program.
210 /// * Scalars checks - The code in canVectorizeBlock checks for a number
211 /// of different conditions, such as the availability of a single induction
212 /// variable, that all types are supported and vectorize-able, etc.
213 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
214 /// This class is also used by SingleBlockLoopVectorizer for identifying
215 /// induction variable and the different reduction variables.
216 class LoopVectorizationLegality {
218 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
219 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
221 /// This represents the kinds of reductions that we support.
223 NoReduction, /// Not a reduction.
224 IntegerAdd, /// Sum of numbers.
225 IntegerMult, /// Product of numbers.
226 IntegerOr, /// Bitwise or logical OR of numbers.
227 IntegerAnd, /// Bitwise or logical AND of numbers.
228 IntegerXor /// Bitwise or logical XOR of numbers.
231 /// This POD struct holds information about reduction variables.
232 struct ReductionDescriptor {
234 ReductionDescriptor():
235 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
238 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
239 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
241 // The starting value of the reduction.
242 // It does not have to be zero!
244 // The instruction who's value is used outside the loop.
245 Instruction *LoopExitInstr;
246 // The kind of the reduction.
250 // This POD struct holds information about the memory runtime legality
251 // check that a group of pointers do not overlap.
252 struct RuntimePointerCheck {
253 /// This flag indicates if we need to add the runtime check.
255 /// Holds the pointers that we need to check.
256 SmallVector<Value*, 2> Pointers;
259 /// ReductionList contains the reduction descriptors for all
260 /// of the reductions that were found in the loop.
261 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
263 /// InductionList saves induction variables and maps them to the initial
264 /// value entring the loop.
265 typedef DenseMap<PHINode*, Value*> InductionList;
267 /// Returns true if it is legal to vectorize this loop.
268 /// This does not mean that it is profitable to vectorize this
269 /// loop, only that it is legal to do so.
272 /// Returns the Induction variable.
273 PHINode *getInduction() {return Induction;}
275 /// Returns the reduction variables found in the loop.
276 ReductionList *getReductionVars() { return &Reductions; }
278 /// Returns the induction variables found in the loop.
279 InductionList *getInductionVars() { return &Inductions; }
281 /// Check if the pointer returned by this GEP is consecutive
282 /// when the index is vectorized. This happens when the last
283 /// index of the GEP is consecutive, like the induction variable.
284 /// This check allows us to vectorize A[idx] into a wide load/store.
285 bool isConsecutiveGep(Value *Ptr);
287 /// Returns true if the value V is uniform within the loop.
288 bool isUniform(Value *V);
290 /// Returns true if this instruction will remain scalar after vectorization.
291 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
293 /// Returns the information that we collected about runtime memory check.
294 RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; }
296 /// Check if a single basic block loop is vectorizable.
297 /// At this point we know that this is a loop with a constant trip count
298 /// and we only need to check individual instructions.
299 bool canVectorizeBlock(BasicBlock &BB);
301 /// When we vectorize loops we may change the order in which
302 /// we read and write from memory. This method checks if it is
303 /// legal to vectorize the code, considering only memory constrains.
304 /// Returns true if BB is vectorizable
305 bool canVectorizeMemory(BasicBlock &BB);
307 /// Returns True, if 'Phi' is the kind of reduction variable for type
308 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
309 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
310 /// Returns true if the instruction I can be a reduction variable of type
312 bool isReductionInstr(Instruction *I, ReductionKind Kind);
313 /// Returns True, if 'Phi' is an induction variable.
314 bool isInductionVariable(PHINode *Phi);
315 /// Return true if can compute the address bounds of Ptr within the loop.
316 bool hasComputableBounds(Value *Ptr);
318 /// The loop that we evaluate.
322 /// DataLayout analysis.
325 // --- vectorization state --- //
327 /// Holds the integer induction variable. This is the counter of the
330 /// Holds the reduction variables.
331 ReductionList Reductions;
332 /// Holds all of the induction variables that we found in the loop.
333 /// Notice that inductions don't need to start at zero and that induction
334 /// variables can be pointers.
335 InductionList Inductions;
337 /// Allowed outside users. This holds the reduction
338 /// vars which can be accessed from outside the loop.
339 SmallPtrSet<Value*, 4> AllowedExit;
340 /// This set holds the variables which are known to be uniform after
342 SmallPtrSet<Instruction*, 4> Uniforms;
343 /// We need to check that all of the pointers in this list are disjoint
345 RuntimePointerCheck PtrRtCheck;
348 /// LoopVectorizationCostModel - estimates the expected speedups due to
350 /// In many cases vectorization is not profitable. This can happen because
351 /// of a number of reasons. In this class we mainly attempt to predict
352 /// the expected speedup/slowdowns due to the supported instruction set.
353 /// We use the VectorTargetTransformInfo to query the different backends
354 /// for the cost of different operations.
355 class LoopVectorizationCostModel {
358 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
359 LoopVectorizationLegality *Leg,
360 const VectorTargetTransformInfo *Vtti):
361 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
363 /// Returns the most profitable vectorization factor for the loop that is
364 /// smaller or equal to the VF argument. This method checks every power
366 unsigned findBestVectorizationFactor(unsigned VF = 8);
369 /// Returns the expected execution cost. The unit of the cost does
370 /// not matter because we use the 'cost' units to compare different
371 /// vector widths. The cost that is returned is *not* normalized by
372 /// the factor width.
373 unsigned expectedCost(unsigned VF);
375 /// Returns the execution time cost of an instruction for a given vector
376 /// width. Vector width of one means scalar.
377 unsigned getInstructionCost(Instruction *I, unsigned VF);
379 /// A helper function for converting Scalar types to vector types.
380 /// If the incoming type is void, we return void. If the VF is 1, we return
382 static Type* ToVectorTy(Type *Scalar, unsigned VF);
384 /// The loop that we evaluate.
389 /// Vectorization legality.
390 LoopVectorizationLegality *Legal;
391 /// Vector target information.
392 const VectorTargetTransformInfo *VTTI;
395 struct LoopVectorize : public LoopPass {
396 static char ID; // Pass identification, replacement for typeid
398 LoopVectorize() : LoopPass(ID) {
399 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
405 TargetTransformInfo *TTI;
408 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
409 // We only vectorize innermost loops.
413 SE = &getAnalysis<ScalarEvolution>();
414 DL = getAnalysisIfAvailable<DataLayout>();
415 LI = &getAnalysis<LoopInfo>();
416 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
417 DT = &getAnalysis<DominatorTree>();
419 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
420 L->getHeader()->getParent()->getName() << "\"\n");
422 // Check if it is legal to vectorize the loop.
423 LoopVectorizationLegality LVL(L, SE, DL);
424 if (!LVL.canVectorize()) {
425 DEBUG(dbgs() << "LV: Not vectorizing.\n");
429 // Select the preffered vectorization factor.
431 if (VectorizationFactor == 0) {
432 const VectorTargetTransformInfo *VTTI = 0;
434 VTTI = TTI->getVectorTargetTransformInfo();
435 // Use the cost model.
436 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
437 VF = CM.findBestVectorizationFactor();
440 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
445 // Use the user command flag.
446 VF = VectorizationFactor;
449 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
450 L->getHeader()->getParent()->getParent()->getModuleIdentifier()<<
453 // If we decided that it is *legal* to vectorizer the loop then do it.
454 SingleBlockLoopVectorizer LB(L, SE, LI, DT, &LPM, VF);
457 DEBUG(verifyFunction(*L->getHeader()->getParent()));
461 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
462 LoopPass::getAnalysisUsage(AU);
463 AU.addRequiredID(LoopSimplifyID);
464 AU.addRequiredID(LCSSAID);
465 AU.addRequired<LoopInfo>();
466 AU.addRequired<ScalarEvolution>();
467 AU.addRequired<DominatorTree>();
468 AU.addPreserved<LoopInfo>();
469 AU.addPreserved<DominatorTree>();
474 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
475 // Instructions that access the old induction variable
476 // actually want to get the new one.
477 if (V == OldInduction)
480 LLVMContext &C = V->getContext();
481 Type *VTy = VectorType::get(V->getType(), VF);
482 Type *I32 = IntegerType::getInt32Ty(C);
483 Constant *Zero = ConstantInt::get(I32, 0);
484 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
485 Value *UndefVal = UndefValue::get(VTy);
486 // Insert the value into a new vector.
487 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
488 // Broadcast the scalar into all locations in the vector.
489 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
491 // We are accessing the induction variable. Make sure to promote the
492 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
494 return getConsecutiveVector(Shuf);
498 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
499 assert(Val->getType()->isVectorTy() && "Must be a vector");
500 assert(Val->getType()->getScalarType()->isIntegerTy() &&
501 "Elem must be an integer");
503 Type *ITy = Val->getType()->getScalarType();
504 VectorType *Ty = cast<VectorType>(Val->getType());
505 unsigned VLen = Ty->getNumElements();
506 SmallVector<Constant*, 8> Indices;
508 // Create a vector of consecutive numbers from zero to VF.
509 for (unsigned i = 0; i < VLen; ++i)
510 Indices.push_back(ConstantInt::get(ITy, i));
512 // Add the consecutive indices to the vector value.
513 Constant *Cv = ConstantVector::get(Indices);
514 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
515 return Builder.CreateAdd(Val, Cv, "induction");
518 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
519 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
523 unsigned NumOperands = Gep->getNumOperands();
524 Value *LastIndex = Gep->getOperand(NumOperands - 1);
526 // Check that all of the gep indices are uniform except for the last.
527 for (unsigned i = 0; i < NumOperands - 1; ++i)
528 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
531 // We can emit wide load/stores only of the last index is the induction
533 const SCEV *Last = SE->getSCEV(LastIndex);
534 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
535 const SCEV *Step = AR->getStepRecurrence(*SE);
537 // The memory is consecutive because the last index is consecutive
538 // and all other indices are loop invariant.
546 bool LoopVectorizationLegality::isUniform(Value *V) {
547 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
550 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
551 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
552 // If we saved a vectorized copy of V, use it.
553 Value *&MapEntry = WidenMap[V];
557 // Broadcast V and save the value for future uses.
558 Value *B = getBroadcastInstrs(V);
564 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
565 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
568 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
569 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
570 // Holds vector parameters or scalars, in case of uniform vals.
571 SmallVector<Value*, 8> Params;
573 // Find all of the vectorized parameters.
574 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
575 Value *SrcOp = Instr->getOperand(op);
577 // If we are accessing the old induction variable, use the new one.
578 if (SrcOp == OldInduction) {
579 Params.push_back(getBroadcastInstrs(Induction));
583 // Try using previously calculated values.
584 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
586 // If the src is an instruction that appeared earlier in the basic block
587 // then it should already be vectorized.
588 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
589 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
590 // The parameter is a vector value from earlier.
591 Params.push_back(WidenMap[SrcInst]);
593 // The parameter is a scalar from outside the loop. Maybe even a constant.
594 Params.push_back(SrcOp);
598 assert(Params.size() == Instr->getNumOperands() &&
599 "Invalid number of operands");
601 // Does this instruction return a value ?
602 bool IsVoidRetTy = Instr->getType()->isVoidTy();
603 Value *VecResults = 0;
605 // If we have a return value, create an empty vector. We place the scalarized
606 // instructions in this vector.
608 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
610 // For each scalar that we create:
611 for (unsigned i = 0; i < VF; ++i) {
612 Instruction *Cloned = Instr->clone();
614 Cloned->setName(Instr->getName() + ".cloned");
615 // Replace the operands of the cloned instrucions with extracted scalars.
616 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
617 Value *Op = Params[op];
618 // Param is a vector. Need to extract the right lane.
619 if (Op->getType()->isVectorTy())
620 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
621 Cloned->setOperand(op, Op);
624 // Place the cloned scalar in the new loop.
625 Builder.Insert(Cloned);
627 // If the original scalar returns a value we need to place it in a vector
628 // so that future users will be able to use it.
630 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
631 Builder.getInt32(i));
635 WidenMap[Instr] = VecResults;
639 SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
641 In this function we generate a new loop. The new loop will contain
642 the vectorized instructions while the old loop will continue to run the
645 [ ] <-- vector loop bypass.
648 | [ ] <-- vector pre header.
652 | [ ]_| <-- vector loop.
655 >[ ] <--- middle-block.
658 | [ ] <--- new preheader.
662 | [ ]_| <-- old scalar loop to handle remainder.
669 OldInduction = Legal->getInduction();
670 assert(OldInduction && "We must have a single phi node.");
671 Type *IdxTy = OldInduction->getType();
673 // Find the loop boundaries.
674 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
675 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
677 // Get the total trip count from the count by adding 1.
678 ExitCount = SE->getAddExpr(ExitCount,
679 SE->getConstant(ExitCount->getType(), 1));
680 // We may need to extend the index in case there is a type mismatch.
681 // We know that the count starts at zero and does not overflow.
682 // We are using Zext because it should be less expensive.
683 if (ExitCount->getType() != IdxTy)
684 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
686 // This is the original scalar-loop preheader.
687 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
688 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
689 assert(ExitBlock && "Must have an exit block");
691 // The loop index does not have to start at Zero. It starts with this value.
692 Value *StartIdx = OldInduction->getIncomingValueForBlock(BypassBlock);
694 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
695 assert(BypassBlock && "Invalid loop structure");
697 BasicBlock *VectorPH =
698 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
699 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
702 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
704 BasicBlock *ScalarPH =
705 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
707 // Find the induction variable.
708 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
710 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
712 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
714 // Generate the induction variable.
715 Induction = Builder.CreatePHI(IdxTy, 2, "index");
716 Constant *Step = ConstantInt::get(IdxTy, VF);
718 // Expand the trip count and place the new instructions in the preheader.
719 // Notice that the pre-header does not change, only the loop body.
720 SCEVExpander Exp(*SE, "induction");
721 Instruction *Loc = BypassBlock->getTerminator();
723 // Count holds the overall loop count (N).
724 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
726 // Add the start index to the loop count to get the new end index.
727 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
729 // Now we need to generate the expression for N - (N % VF), which is
730 // the part that the vectorized body will execute.
731 Constant *CIVF = ConstantInt::get(IdxTy, VF);
732 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
733 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
734 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
735 "end.idx.rnd.down", Loc);
737 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
738 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
743 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
744 Legal->getRuntimePointerCheck();
745 Value *MemoryRuntimeCheck = 0;
746 if (PtrRtCheck->Need) {
747 unsigned NumPointers = PtrRtCheck->Pointers.size();
748 SmallVector<Value* , 2> Starts;
749 SmallVector<Value* , 2> Ends;
751 // Use this type for pointer arithmetic.
752 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
754 for (unsigned i=0; i < NumPointers; ++i) {
755 Value *Ptr = PtrRtCheck->Pointers[i];
756 const SCEV *Sc = SE->getSCEV(Ptr);
758 if (SE->isLoopInvariant(Sc, OrigLoop)) {
759 DEBUG(dbgs() << "LV1: Adding RT check for a loop invariant ptr:" <<
761 Starts.push_back(Ptr);
764 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
765 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
766 Value *Start = Exp.expandCodeFor(AR->getStart(), PtrArithTy, Loc);
767 const SCEV *Ex = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
768 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
769 assert(!isa<SCEVCouldNotCompute>(ScEnd) && "Invalid scev range.");
770 Value *End = Exp.expandCodeFor(ScEnd, PtrArithTy, Loc);
771 Starts.push_back(Start);
776 for (unsigned i=0; i < NumPointers; ++i) {
777 for (unsigned j=i+1; j < NumPointers; ++j) {
778 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
779 Starts[0], Ends[1], "bound0", Loc);
780 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
781 Starts[1], Ends[0], "bound1", Loc);
782 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
783 "found.conflict", Loc);
784 if (MemoryRuntimeCheck) {
785 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
788 "conflict.rdx", Loc);
790 MemoryRuntimeCheck = IsConflict;
794 }// end of need-runtime-check code.
796 // If we are using memory runtime checks, include them in.
797 if (MemoryRuntimeCheck) {
798 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
802 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
803 // Remove the old terminator.
804 Loc->eraseFromParent();
806 // We are going to resume the execution of the scalar loop.
807 // Go over all of the induction variables that we found and fix the
808 // PHIs that are left in the scalar version of the loop.
809 // The starting values of PHI nodes depend on the counter of the last
810 // iteration in the vectorized loop.
811 // If we come from a bypass edge then we need to start from the original start
814 // This variable saves the new starting index for the scalar loop.
815 Value *ResumeIndex = 0;
816 LoopVectorizationLegality::InductionList::iterator I, E;
817 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
818 for (I = List->begin(), E = List->end(); I != E; ++I) {
819 PHINode *OrigPhi = I->first;
820 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
821 MiddleBlock->getTerminator());
823 if (OrigPhi->getType()->isIntegerTy()) {
824 // Handle the integer induction counter:
825 assert(OrigPhi == OldInduction && "Unknown integer PHI");
826 // We know what the end value is.
827 EndValue = IdxEndRoundDown;
828 // We also know which PHI node holds it.
829 ResumeIndex = ResumeVal;
831 // For pointer induction variables, calculate the offset using
833 EndValue = GetElementPtrInst::Create(I->second, IdxEndRoundDown,
835 BypassBlock->getTerminator());
838 // The new PHI merges the original incoming value, in case of a bypass,
839 // or the value at the end of the vectorized loop.
840 ResumeVal->addIncoming(I->second, BypassBlock);
841 ResumeVal->addIncoming(EndValue, VecBody);
843 // Fix the scalar body counter (PHI node).
844 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
845 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
848 // Make sure that we found the index where scalar loop needs to continue.
849 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
850 "Invalid resume Index");
852 // Add a check in the middle block to see if we have completed
853 // all of the iterations in the first vector loop.
854 // If (N - N%VF) == N, then we *don't* need to run the remainder.
855 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
856 ResumeIndex, "cmp.n",
857 MiddleBlock->getTerminator());
859 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
860 // Remove the old terminator.
861 MiddleBlock->getTerminator()->eraseFromParent();
863 // Create i+1 and fill the PHINode.
864 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
865 Induction->addIncoming(StartIdx, VectorPH);
866 Induction->addIncoming(NextIdx, VecBody);
867 // Create the compare.
868 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
869 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
871 // Now we have two terminators. Remove the old one from the block.
872 VecBody->getTerminator()->eraseFromParent();
874 // Get ready to start creating new instructions into the vectorized body.
875 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
877 // Register the new loop.
878 Loop* Lp = new Loop();
879 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
881 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
883 Loop *ParentLoop = OrigLoop->getParentLoop();
885 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
886 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
887 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
891 LoopVectorPreHeader = VectorPH;
892 LoopScalarPreHeader = ScalarPH;
893 LoopMiddleBlock = MiddleBlock;
894 LoopExitBlock = ExitBlock;
895 LoopVectorBody = VecBody;
896 LoopScalarBody = OldBasicBlock;
897 LoopBypassBlock = BypassBlock;
900 /// This function returns the identity element (or neutral element) for
903 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
905 case LoopVectorizationLegality::IntegerXor:
906 case LoopVectorizationLegality::IntegerAdd:
907 case LoopVectorizationLegality::IntegerOr:
908 // Adding, Xoring, Oring zero to a number does not change it.
910 case LoopVectorizationLegality::IntegerMult:
911 // Multiplying a number by 1 does not change it.
913 case LoopVectorizationLegality::IntegerAnd:
914 // AND-ing a number with an all-1 value does not change it.
917 llvm_unreachable("Unknown reduction kind");
922 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
923 //===------------------------------------------------===//
925 // Notice: any optimization or new instruction that go
926 // into the code below should be also be implemented in
929 //===------------------------------------------------===//
930 typedef SmallVector<PHINode*, 4> PhiVector;
931 BasicBlock &BB = *OrigLoop->getHeader();
932 Constant *Zero = ConstantInt::get(
933 IntegerType::getInt32Ty(BB.getContext()), 0);
935 // In order to support reduction variables we need to be able to vectorize
936 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
937 // steages. First, we create a new vector PHI node with no incoming edges.
938 // We use this value when we vectorize all of the instructions that use the
939 // PHI. Next, after all of the instructions in the block are complete we
940 // add the new incoming edges to the PHI. At this point all of the
941 // instructions in the basic block are vectorized, so we can use them to
942 // construct the PHI.
943 PhiVector RdxPHIsToFix;
945 // For each instruction in the old loop.
946 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
947 Instruction *Inst = it;
949 switch (Inst->getOpcode()) {
950 case Instruction::Br:
951 // Nothing to do for PHIs and BR, since we already took care of the
952 // loop control flow instructions.
954 case Instruction::PHI:{
955 PHINode* P = cast<PHINode>(Inst);
956 // Special handling for the induction var.
957 if (OldInduction == Inst)
960 // Handle reduction variables:
961 if (Legal->getReductionVars()->count(P)) {
962 // This is phase one of vectorizing PHIs.
963 Type *VecTy = VectorType::get(Inst->getType(), VF);
964 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
965 RdxPHIsToFix.push_back(P);
969 // Handle pointer inductions:
970 if (Legal->getInductionVars()->count(P)) {
971 Value *StartIdx = Legal->getInductionVars()->lookup(OldInduction);
972 Value *StartPtr = Legal->getInductionVars()->lookup(P);
973 // This is the normalized GEP that starts counting at zero.
974 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
976 // This is the first GEP in the sequence.
977 Value *FirstGep = Builder.CreateGEP(StartPtr, NormalizedIdx,
979 // This is the vector of results. Notice that we don't generate vector
980 // geps because scalar geps result in better code.
981 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
982 for (unsigned int i = 0; i < VF; ++i) {
983 Value *SclrGep = Builder.CreateGEP(FirstGep, Builder.getInt32(i),
985 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
990 WidenMap[Inst] = VecVal;
994 case Instruction::Add:
995 case Instruction::FAdd:
996 case Instruction::Sub:
997 case Instruction::FSub:
998 case Instruction::Mul:
999 case Instruction::FMul:
1000 case Instruction::UDiv:
1001 case Instruction::SDiv:
1002 case Instruction::FDiv:
1003 case Instruction::URem:
1004 case Instruction::SRem:
1005 case Instruction::FRem:
1006 case Instruction::Shl:
1007 case Instruction::LShr:
1008 case Instruction::AShr:
1009 case Instruction::And:
1010 case Instruction::Or:
1011 case Instruction::Xor: {
1012 // Just widen binops.
1013 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1014 Value *A = getVectorValue(Inst->getOperand(0));
1015 Value *B = getVectorValue(Inst->getOperand(1));
1017 // Use this vector value for all users of the original instruction.
1018 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1021 // Update the NSW, NUW and Exact flags.
1022 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1023 if (isa<OverflowingBinaryOperator>(BinOp)) {
1024 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1025 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1027 if (isa<PossiblyExactOperator>(VecOp))
1028 VecOp->setIsExact(BinOp->isExact());
1031 case Instruction::Select: {
1033 // If the selector is loop invariant we can create a select
1034 // instruction with a scalar condition. Otherwise, use vector-select.
1035 Value *Cond = Inst->getOperand(0);
1036 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1038 // The condition can be loop invariant but still defined inside the
1039 // loop. This means that we can't just use the original 'cond' value.
1040 // We have to take the 'vectorized' value and pick the first lane.
1041 // Instcombine will make this a no-op.
1042 Cond = getVectorValue(Cond);
1044 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1046 Value *Op0 = getVectorValue(Inst->getOperand(1));
1047 Value *Op1 = getVectorValue(Inst->getOperand(2));
1048 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
1052 case Instruction::ICmp:
1053 case Instruction::FCmp: {
1054 // Widen compares. Generate vector compares.
1055 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
1056 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
1057 Value *A = getVectorValue(Inst->getOperand(0));
1058 Value *B = getVectorValue(Inst->getOperand(1));
1060 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1062 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1066 case Instruction::Store: {
1067 // Attempt to issue a wide store.
1068 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1069 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1070 Value *Ptr = SI->getPointerOperand();
1071 unsigned Alignment = SI->getAlignment();
1073 assert(!Legal->isUniform(Ptr) &&
1074 "We do not allow storing to uniform addresses");
1076 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1078 // This store does not use GEPs.
1079 if (!Legal->isConsecutiveGep(Gep)) {
1080 scalarizeInstruction(Inst);
1084 // The last index does not have to be the induction. It can be
1085 // consecutive and be a function of the index. For example A[I+1];
1086 unsigned NumOperands = Gep->getNumOperands();
1087 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1088 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1090 // Create the new GEP with the new induction variable.
1091 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1092 Gep2->setOperand(NumOperands - 1, LastIndex);
1093 Ptr = Builder.Insert(Gep2);
1094 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1095 Value *Val = getVectorValue(SI->getValueOperand());
1096 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1099 case Instruction::Load: {
1100 // Attempt to issue a wide load.
1101 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1102 Type *RetTy = VectorType::get(LI->getType(), VF);
1103 Value *Ptr = LI->getPointerOperand();
1104 unsigned Alignment = LI->getAlignment();
1105 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1107 // If we don't have a gep, or that the pointer is loop invariant,
1108 // scalarize the load.
1109 if (!Gep || Legal->isUniform(Gep) || !Legal->isConsecutiveGep(Gep)) {
1110 scalarizeInstruction(Inst);
1114 // The last index does not have to be the induction. It can be
1115 // consecutive and be a function of the index. For example A[I+1];
1116 unsigned NumOperands = Gep->getNumOperands();
1117 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1118 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1120 // Create the new GEP with the new induction variable.
1121 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1122 Gep2->setOperand(NumOperands - 1, LastIndex);
1123 Ptr = Builder.Insert(Gep2);
1124 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1125 LI = Builder.CreateLoad(Ptr);
1126 LI->setAlignment(Alignment);
1127 // Use this vector value for all users of the load.
1128 WidenMap[Inst] = LI;
1131 case Instruction::ZExt:
1132 case Instruction::SExt:
1133 case Instruction::FPToUI:
1134 case Instruction::FPToSI:
1135 case Instruction::FPExt:
1136 case Instruction::PtrToInt:
1137 case Instruction::IntToPtr:
1138 case Instruction::SIToFP:
1139 case Instruction::UIToFP:
1140 case Instruction::Trunc:
1141 case Instruction::FPTrunc:
1142 case Instruction::BitCast: {
1143 /// Vectorize bitcasts.
1144 CastInst *CI = dyn_cast<CastInst>(Inst);
1145 Value *A = getVectorValue(Inst->getOperand(0));
1146 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1147 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1152 /// All other instructions are unsupported. Scalarize them.
1153 scalarizeInstruction(Inst);
1156 }// end of for_each instr.
1158 // At this point every instruction in the original loop is widended to
1159 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1160 // that we vectorized. The PHI nodes are currently empty because we did
1161 // not want to introduce cycles. Notice that the remaining PHI nodes
1162 // that we need to fix are reduction variables.
1164 // Create the 'reduced' values for each of the induction vars.
1165 // The reduced values are the vector values that we scalarize and combine
1166 // after the loop is finished.
1167 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1169 PHINode *RdxPhi = *it;
1170 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1171 assert(RdxPhi && "Unable to recover vectorized PHI");
1173 // Find the reduction variable descriptor.
1174 assert(Legal->getReductionVars()->count(RdxPhi) &&
1175 "Unable to find the reduction variable");
1176 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1177 (*Legal->getReductionVars())[RdxPhi];
1179 // We need to generate a reduction vector from the incoming scalar.
1180 // To do so, we need to generate the 'identity' vector and overide
1181 // one of the elements with the incoming scalar reduction. We need
1182 // to do it in the vector-loop preheader.
1183 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1185 // This is the vector-clone of the value that leaves the loop.
1186 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1187 Type *VecTy = VectorExit->getType();
1189 // Find the reduction identity variable. Zero for addition, or, xor,
1190 // one for multiplication, -1 for And.
1191 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1192 VecTy->getScalarType());
1194 // This vector is the Identity vector where the first element is the
1195 // incoming scalar reduction.
1196 Value *VectorStart = Builder.CreateInsertElement(Identity,
1197 RdxDesc.StartValue, Zero);
1199 // Fix the vector-loop phi.
1200 // We created the induction variable so we know that the
1201 // preheader is the first entry.
1202 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1204 // Reductions do not have to start at zero. They can start with
1205 // any loop invariant values.
1206 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1207 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1208 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1209 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1211 // Before each round, move the insertion point right between
1212 // the PHIs and the values we are going to write.
1213 // This allows us to write both PHINodes and the extractelement
1215 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1217 // This PHINode contains the vectorized reduction variable, or
1218 // the initial value vector, if we bypass the vector loop.
1219 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1220 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1221 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1223 // Extract the first scalar.
1225 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1226 // Extract and reduce the remaining vector elements.
1227 for (unsigned i=1; i < VF; ++i) {
1229 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1230 switch (RdxDesc.Kind) {
1231 case LoopVectorizationLegality::IntegerAdd:
1232 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1234 case LoopVectorizationLegality::IntegerMult:
1235 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1237 case LoopVectorizationLegality::IntegerOr:
1238 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1240 case LoopVectorizationLegality::IntegerAnd:
1241 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1243 case LoopVectorizationLegality::IntegerXor:
1244 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1247 llvm_unreachable("Unknown reduction operation");
1251 // Now, we need to fix the users of the reduction variable
1252 // inside and outside of the scalar remainder loop.
1253 // We know that the loop is in LCSSA form. We need to update the
1254 // PHI nodes in the exit blocks.
1255 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1256 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1257 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1258 if (!LCSSAPhi) continue;
1260 // All PHINodes need to have a single entry edge, or two if
1261 // we already fixed them.
1262 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1264 // We found our reduction value exit-PHI. Update it with the
1265 // incoming bypass edge.
1266 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1267 // Add an edge coming from the bypass.
1268 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1271 }// end of the LCSSA phi scan.
1273 // Fix the scalar loop reduction variable with the incoming reduction sum
1274 // from the vector body and from the backedge value.
1275 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1276 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1277 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1278 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1279 }// end of for each redux variable.
1282 void SingleBlockLoopVectorizer::updateAnalysis() {
1283 // The original basic block.
1284 SE->forgetLoop(OrigLoop);
1286 // Update the dominator tree information.
1287 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1288 "Entry does not dominate exit.");
1290 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1291 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1292 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1293 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1294 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1295 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1297 DEBUG(DT->verifyAnalysis());
1300 bool LoopVectorizationLegality::canVectorize() {
1301 if (!TheLoop->getLoopPreheader()) {
1302 assert(false && "No preheader!!");
1303 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1307 // We can only vectorize single basic block loops.
1308 unsigned NumBlocks = TheLoop->getNumBlocks();
1309 if (NumBlocks != 1) {
1310 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1314 // We need to have a loop header.
1315 BasicBlock *BB = TheLoop->getHeader();
1316 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1318 // ScalarEvolution needs to be able to find the exit count.
1319 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1320 if (ExitCount == SE->getCouldNotCompute()) {
1321 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1325 // Do not loop-vectorize loops with a tiny trip count.
1326 unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB);
1327 if (TC > 0u && TC < TinyTripCountThreshold) {
1328 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1329 "This loop is not worth vectorizing.\n");
1333 // Go over each instruction and look at memory deps.
1334 if (!canVectorizeBlock(*BB)) {
1335 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1339 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1340 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1343 // Okay! We can vectorize. At this point we don't have any other mem analysis
1344 // which may limit our maximum vectorization factor, so just return true with
1349 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1350 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1352 // Scan the instructions in the block and look for hazards.
1353 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1354 Instruction *I = it;
1356 if (PHINode *Phi = dyn_cast<PHINode>(I)) {
1357 // This should not happen because the loop should be normalized.
1358 if (Phi->getNumIncomingValues() != 2) {
1359 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1363 // This is the value coming from the preheader.
1364 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1366 // We only look at integer and pointer phi nodes.
1367 if (Phi->getType()->isPointerTy() && isInductionVariable(Phi)) {
1368 DEBUG(dbgs() << "LV: Found a pointer induction variable.\n");
1369 Inductions[Phi] = StartValue;
1371 } else if (!Phi->getType()->isIntegerTy()) {
1372 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1376 // Handle integer PHIs:
1377 if (isInductionVariable(Phi)) {
1379 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1382 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1384 Inductions[Phi] = StartValue;
1387 if (AddReductionVar(Phi, IntegerAdd)) {
1388 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1391 if (AddReductionVar(Phi, IntegerMult)) {
1392 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1395 if (AddReductionVar(Phi, IntegerOr)) {
1396 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1399 if (AddReductionVar(Phi, IntegerAnd)) {
1400 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1403 if (AddReductionVar(Phi, IntegerXor)) {
1404 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1408 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1410 }// end of PHI handling
1412 // We still don't handle functions.
1413 CallInst *CI = dyn_cast<CallInst>(I);
1415 DEBUG(dbgs() << "LV: Found a call site.\n");
1419 // We do not re-vectorize vectors.
1420 if (!VectorType::isValidElementType(I->getType()) &&
1421 !I->getType()->isVoidTy()) {
1422 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1426 // Reduction instructions are allowed to have exit users.
1427 // All other instructions must not have external users.
1428 if (!AllowedExit.count(I))
1429 //Check that all of the users of the loop are inside the BB.
1430 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1432 Instruction *U = cast<Instruction>(*it);
1433 // This user may be a reduction exit value.
1434 BasicBlock *Parent = U->getParent();
1435 if (Parent != &BB) {
1436 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1443 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1447 // Don't vectorize if the memory dependencies do not allow vectorization.
1448 if (!canVectorizeMemory(BB))
1451 // We now know that the loop is vectorizable!
1452 // Collect variables that will remain uniform after vectorization.
1453 std::vector<Value*> Worklist;
1455 // Start with the conditional branch and walk up the block.
1456 Worklist.push_back(BB.getTerminator()->getOperand(0));
1458 while (Worklist.size()) {
1459 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1460 Worklist.pop_back();
1461 // Look at instructions inside this block.
1463 if (I->getParent() != &BB) continue;
1465 // Stop when reaching PHI nodes.
1466 if (isa<PHINode>(I)) {
1467 assert(I == Induction && "Found a uniform PHI that is not the induction");
1471 // This is a known uniform.
1474 // Insert all operands.
1475 for (int i=0, Op = I->getNumOperands(); i < Op; ++i) {
1476 Worklist.push_back(I->getOperand(i));
1483 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1484 typedef SmallVector<Value*, 16> ValueVector;
1485 typedef SmallPtrSet<Value*, 16> ValueSet;
1486 // Holds the Load and Store *instructions*.
1489 PtrRtCheck.Pointers.clear();
1490 PtrRtCheck.Need = false;
1492 // Scan the BB and collect legal loads and stores.
1493 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1494 Instruction *I = it;
1496 // If this is a load, save it. If this instruction can read from memory
1497 // but is not a load, then we quit. Notice that we don't handle function
1498 // calls that read or write.
1499 if (I->mayReadFromMemory()) {
1500 LoadInst *Ld = dyn_cast<LoadInst>(I);
1501 if (!Ld) return false;
1502 if (!Ld->isSimple()) {
1503 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1506 Loads.push_back(Ld);
1510 // Save store instructions. Abort if other instructions write to memory.
1511 if (I->mayWriteToMemory()) {
1512 StoreInst *St = dyn_cast<StoreInst>(I);
1513 if (!St) return false;
1514 if (!St->isSimple()) {
1515 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1518 Stores.push_back(St);
1522 // Now we have two lists that hold the loads and the stores.
1523 // Next, we find the pointers that they use.
1525 // Check if we see any stores. If there are no stores, then we don't
1526 // care if the pointers are *restrict*.
1527 if (!Stores.size()) {
1528 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1532 // Holds the read and read-write *pointers* that we find.
1534 ValueVector ReadWrites;
1536 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1537 // multiple times on the same object. If the ptr is accessed twice, once
1538 // for read and once for write, it will only appear once (on the write
1539 // list). This is okay, since we are going to check for conflicts between
1540 // writes and between reads and writes, but not between reads and reads.
1543 ValueVector::iterator I, IE;
1544 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1545 StoreInst *ST = dyn_cast<StoreInst>(*I);
1546 assert(ST && "Bad StoreInst");
1547 Value* Ptr = ST->getPointerOperand();
1549 if (isUniform(Ptr)) {
1550 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1554 // If we did *not* see this pointer before, insert it to
1555 // the read-write list. At this phase it is only a 'write' list.
1556 if (Seen.insert(Ptr))
1557 ReadWrites.push_back(Ptr);
1560 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1561 LoadInst *LD = dyn_cast<LoadInst>(*I);
1562 assert(LD && "Bad LoadInst");
1563 Value* Ptr = LD->getPointerOperand();
1564 // If we did *not* see this pointer before, insert it to the
1565 // read list. If we *did* see it before, then it is already in
1566 // the read-write list. This allows us to vectorize expressions
1567 // such as A[i] += x; Because the address of A[i] is a read-write
1568 // pointer. This only works if the index of A[i] is consecutive.
1569 // If the address of i is unknown (for example A[B[i]]) then we may
1570 // read a few words, modify, and write a few words, and some of the
1571 // words may be written to the same address.
1572 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1573 Reads.push_back(Ptr);
1576 // If we write (or read-write) to a single destination and there are no
1577 // other reads in this loop then is it safe to vectorize.
1578 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1579 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1583 // Find pointers with computable bounds. We are going to use this information
1584 // to place a runtime bound check.
1586 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1587 if (hasComputableBounds(*I)) {
1588 PtrRtCheck.Pointers.push_back(*I);
1589 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1594 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1595 if (hasComputableBounds(*I)) {
1596 PtrRtCheck.Pointers.push_back(*I);
1597 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1603 // Check that we did not collect too many pointers or found a
1604 // unsizeable pointer.
1605 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1606 PtrRtCheck.Pointers.clear();
1610 PtrRtCheck.Need = RT;
1613 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1616 // Now that the pointers are in two lists (Reads and ReadWrites), we
1617 // can check that there are no conflicts between each of the writes and
1618 // between the writes to the reads.
1619 ValueSet WriteObjects;
1620 ValueVector TempObjects;
1622 // Check that the read-writes do not conflict with other read-write
1624 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1625 GetUnderlyingObjects(*I, TempObjects, DL);
1626 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1628 if (!isIdentifiedObject(*it)) {
1629 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1632 if (!WriteObjects.insert(*it)) {
1633 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1638 TempObjects.clear();
1641 /// Check that the reads don't conflict with the read-writes.
1642 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1643 GetUnderlyingObjects(*I, TempObjects, DL);
1644 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1646 if (!isIdentifiedObject(*it)) {
1647 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1650 if (WriteObjects.count(*it)) {
1651 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1656 TempObjects.clear();
1659 // It is safe to vectorize and we don't need any runtime checks.
1660 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1661 PtrRtCheck.Pointers.clear();
1662 PtrRtCheck.Need = false;
1666 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1667 ReductionKind Kind) {
1668 if (Phi->getNumIncomingValues() != 2)
1671 // Find the possible incoming reduction variable.
1672 BasicBlock *BB = Phi->getParent();
1673 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1674 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1675 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1677 // ExitInstruction is the single value which is used outside the loop.
1678 // We only allow for a single reduction value to be used outside the loop.
1679 // This includes users of the reduction, variables (which form a cycle
1680 // which ends in the phi node).
1681 Instruction *ExitInstruction = 0;
1683 // Iter is our iterator. We start with the PHI node and scan for all of the
1684 // users of this instruction. All users must be instructions which can be
1685 // used as reduction variables (such as ADD). We may have a single
1686 // out-of-block user. They cycle must end with the original PHI.
1687 // Also, we can't have multiple block-local users.
1688 Instruction *Iter = Phi;
1690 // Any reduction instr must be of one of the allowed kinds.
1691 if (!isReductionInstr(Iter, Kind))
1694 // Did we found a user inside this block ?
1695 bool FoundInBlockUser = false;
1696 // Did we reach the initial PHI node ?
1697 bool FoundStartPHI = false;
1699 // If the instruction has no users then this is a broken
1700 // chain and can't be a reduction variable.
1701 if (Iter->use_empty())
1704 // For each of the *users* of iter.
1705 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1707 Instruction *U = cast<Instruction>(*it);
1708 // We already know that the PHI is a user.
1710 FoundStartPHI = true;
1713 // Check if we found the exit user.
1714 BasicBlock *Parent = U->getParent();
1716 // We must have a single exit instruction.
1717 if (ExitInstruction != 0)
1719 ExitInstruction = Iter;
1721 // We can't have multiple inside users.
1722 if (FoundInBlockUser)
1724 FoundInBlockUser = true;
1728 // We found a reduction var if we have reached the original
1729 // phi node and we only have a single instruction with out-of-loop
1731 if (FoundStartPHI && ExitInstruction) {
1732 // This instruction is allowed to have out-of-loop users.
1733 AllowedExit.insert(ExitInstruction);
1735 // Save the description of this reduction variable.
1736 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1737 Reductions[Phi] = RD;
1744 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1745 ReductionKind Kind) {
1746 switch (I->getOpcode()) {
1749 case Instruction::PHI:
1752 case Instruction::Add:
1753 case Instruction::Sub:
1754 return Kind == IntegerAdd;
1755 case Instruction::Mul:
1756 return Kind == IntegerMult;
1757 case Instruction::And:
1758 return Kind == IntegerAnd;
1759 case Instruction::Or:
1760 return Kind == IntegerOr;
1761 case Instruction::Xor:
1762 return Kind == IntegerXor;
1766 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1767 Type *PhiTy = Phi->getType();
1768 // We only handle integer and pointer inductions variables.
1769 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1772 // Check that the PHI is consecutive and starts at zero.
1773 const SCEV *PhiScev = SE->getSCEV(Phi);
1774 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1776 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1779 const SCEV *Step = AR->getStepRecurrence(*SE);
1781 // Integer inductions need to have a stride of one.
1782 if (PhiTy->isIntegerTy())
1783 return Step->isOne();
1785 // Calculate the pointer stride and check if it is consecutive.
1786 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1787 if (!C) return false;
1789 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1790 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1791 return (C->getValue()->equalsInt(Size));
1794 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1795 const SCEV *PhiScev = SE->getSCEV(Ptr);
1796 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1800 return AR->isAffine();
1804 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1806 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1810 float Cost = expectedCost(1);
1812 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1813 for (unsigned i=2; i <= VF; i*=2) {
1814 // Notice that the vector loop needs to be executed less times, so
1815 // we need to divide the cost of the vector loops by the width of
1816 // the vector elements.
1817 float VectorCost = expectedCost(i) / (float)i;
1818 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1819 (int)VectorCost << ".\n");
1820 if (VectorCost < Cost) {
1826 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1830 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1831 // We can only estimate the cost of single basic block loops.
1832 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1834 BasicBlock *BB = TheLoop->getHeader();
1837 // For each instruction in the old loop.
1838 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1839 Instruction *Inst = it;
1840 unsigned C = getInstructionCost(Inst, VF);
1842 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1843 " For instruction: "<< *Inst << "\n");
1850 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1851 assert(VTTI && "Invalid vector target transformation info");
1853 // If we know that this instruction will remain uniform, check the cost of
1854 // the scalar version.
1855 if (Legal->isUniformAfterVectorization(I))
1858 Type *RetTy = I->getType();
1859 Type *VectorTy = ToVectorTy(RetTy, VF);
1862 // TODO: We need to estimate the cost of intrinsic calls.
1863 switch (I->getOpcode()) {
1864 case Instruction::GetElementPtr:
1865 // We mark this instruction as zero-cost because scalar GEPs are usually
1866 // lowered to the intruction addressing mode. At the moment we don't
1867 // generate vector geps.
1869 case Instruction::Br: {
1870 return VTTI->getCFInstrCost(I->getOpcode());
1872 case Instruction::PHI:
1874 case Instruction::Add:
1875 case Instruction::FAdd:
1876 case Instruction::Sub:
1877 case Instruction::FSub:
1878 case Instruction::Mul:
1879 case Instruction::FMul:
1880 case Instruction::UDiv:
1881 case Instruction::SDiv:
1882 case Instruction::FDiv:
1883 case Instruction::URem:
1884 case Instruction::SRem:
1885 case Instruction::FRem:
1886 case Instruction::Shl:
1887 case Instruction::LShr:
1888 case Instruction::AShr:
1889 case Instruction::And:
1890 case Instruction::Or:
1891 case Instruction::Xor: {
1892 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
1894 case Instruction::Select: {
1895 SelectInst *SI = cast<SelectInst>(I);
1896 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1897 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1898 Type *CondTy = SI->getCondition()->getType();
1900 CondTy = VectorType::get(CondTy, VF);
1902 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
1904 case Instruction::ICmp:
1905 case Instruction::FCmp: {
1906 Type *ValTy = I->getOperand(0)->getType();
1907 VectorTy = ToVectorTy(ValTy, VF);
1908 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
1910 case Instruction::Store: {
1911 StoreInst *SI = cast<StoreInst>(I);
1912 Type *ValTy = SI->getValueOperand()->getType();
1913 VectorTy = ToVectorTy(ValTy, VF);
1916 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
1917 SI->getAlignment(), SI->getPointerAddressSpace());
1919 // Scalarized stores.
1920 if (!Legal->isConsecutiveGep(SI->getPointerOperand())) {
1922 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1924 // The cost of extracting from the value vector.
1925 Cost += VF * (ExtCost);
1926 // The cost of the scalar stores.
1927 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1928 ValTy->getScalarType(),
1930 SI->getPointerAddressSpace());
1935 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
1936 SI->getPointerAddressSpace());
1938 case Instruction::Load: {
1939 LoadInst *LI = cast<LoadInst>(I);
1942 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
1944 LI->getPointerAddressSpace());
1946 // Scalarized loads.
1947 if (!Legal->isConsecutiveGep(LI->getPointerOperand())) {
1949 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
1950 // The cost of inserting the loaded value into the result vector.
1951 Cost += VF * (InCost);
1952 // The cost of the scalar stores.
1953 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1954 RetTy->getScalarType(),
1956 LI->getPointerAddressSpace());
1961 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
1962 LI->getPointerAddressSpace());
1964 case Instruction::ZExt:
1965 case Instruction::SExt:
1966 case Instruction::FPToUI:
1967 case Instruction::FPToSI:
1968 case Instruction::FPExt:
1969 case Instruction::PtrToInt:
1970 case Instruction::IntToPtr:
1971 case Instruction::SIToFP:
1972 case Instruction::UIToFP:
1973 case Instruction::Trunc:
1974 case Instruction::FPTrunc:
1975 case Instruction::BitCast: {
1976 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
1977 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
1980 // We are scalarizing the instruction. Return the cost of the scalar
1981 // instruction, plus the cost of insert and extract into vector
1982 // elements, times the vector width.
1985 bool IsVoid = RetTy->isVoidTy();
1987 unsigned InsCost = (IsVoid ? 0 :
1988 VTTI->getInstrCost(Instruction::InsertElement,
1991 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1994 // The cost of inserting the results plus extracting each one of the
1996 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
1998 // The cost of executing VF copies of the scalar instruction.
1999 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
2005 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2006 if (Scalar->isVoidTy() || VF == 1)
2008 return VectorType::get(Scalar, VF);
2013 char LoopVectorize::ID = 0;
2014 static const char lv_name[] = "Loop Vectorization";
2015 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2016 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2017 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2018 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2019 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2022 Pass *createLoopVectorizePass() {
2023 return new LoopVectorize();