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, DataLayout *dl,
112 OrigLoop(Orig), SE(Se), LI(Li), DT(dt), DL(dl), LPM(Lpm), VF(VecWidth),
113 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
115 // Perform the actual loop widening (vectorization).
116 void vectorize(LoopVectorizationLegality *Legal) {
117 ///Create a new empty loop. Unlink the old loop and connect the new one.
118 createEmptyLoop(Legal);
119 /// Widen each instruction in the old loop to a new one in the new loop.
120 /// Use the Legality module to find the induction and reduction variables.
121 vectorizeLoop(Legal);
122 // Register the new loop and update the analysis passes.
127 /// Create an empty loop, based on the loop ranges of the old loop.
128 void createEmptyLoop(LoopVectorizationLegality *Legal);
129 /// Copy and widen the instructions from the old loop.
130 void vectorizeLoop(LoopVectorizationLegality *Legal);
131 /// Insert the new loop to the loop hierarchy and pass manager
132 /// and update the analysis passes.
133 void updateAnalysis();
135 /// This instruction is un-vectorizable. Implement it as a sequence
137 void scalarizeInstruction(Instruction *Instr);
139 /// Create a broadcast instruction. This method generates a broadcast
140 /// instruction (shuffle) for loop invariant values and for the induction
141 /// value. If this is the induction variable then we extend it to N, N+1, ...
142 /// this is needed because each iteration in the loop corresponds to a SIMD
144 Value *getBroadcastInstrs(Value *V);
146 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
147 /// for each element in the vector. Starting from zero.
148 Value *getConsecutiveVector(Value* Val);
150 /// When we go over instructions in the basic block we rely on previous
151 /// values within the current basic block or on loop invariant values.
152 /// When we widen (vectorize) values we place them in the map. If the values
153 /// are not within the map, they have to be loop invariant, so we simply
154 /// broadcast them into a vector.
155 Value *getVectorValue(Value *V);
157 /// Get a uniform vector of constant integers. We use this to get
158 /// vectors of ones and zeros for the reduction code.
159 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
161 typedef DenseMap<Value*, Value*> ValueMap;
163 /// The original loop.
165 // Scev analysis to use.
173 // Loop Pass Manager;
175 // The vectorization factor to use.
178 // The builder that we use
181 // --- Vectorization state ---
183 /// The vector-loop preheader.
184 BasicBlock *LoopVectorPreHeader;
185 /// The scalar-loop preheader.
186 BasicBlock *LoopScalarPreHeader;
187 /// Middle Block between the vector and the scalar.
188 BasicBlock *LoopMiddleBlock;
189 ///The ExitBlock of the scalar loop.
190 BasicBlock *LoopExitBlock;
191 ///The vector loop body.
192 BasicBlock *LoopVectorBody;
193 ///The scalar loop body.
194 BasicBlock *LoopScalarBody;
195 ///The first bypass block.
196 BasicBlock *LoopBypassBlock;
198 /// The new Induction variable which was added to the new block.
200 /// The induction variable of the old basic block.
201 PHINode *OldInduction;
202 // Maps scalars to widened vectors.
206 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
207 /// to what vectorization factor.
208 /// This class does not look at the profitability of vectorization, only the
209 /// legality. This class has two main kinds of checks:
210 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
211 /// will change the order of memory accesses in a way that will change the
212 /// correctness of the program.
213 /// * Scalars checks - The code in canVectorizeBlock checks for a number
214 /// of different conditions, such as the availability of a single induction
215 /// variable, that all types are supported and vectorize-able, etc.
216 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
217 /// This class is also used by SingleBlockLoopVectorizer for identifying
218 /// induction variable and the different reduction variables.
219 class LoopVectorizationLegality {
221 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
222 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
224 /// This represents the kinds of reductions that we support.
226 NoReduction, /// Not a reduction.
227 IntegerAdd, /// Sum of numbers.
228 IntegerMult, /// Product of numbers.
229 IntegerOr, /// Bitwise or logical OR of numbers.
230 IntegerAnd, /// Bitwise or logical AND of numbers.
231 IntegerXor /// Bitwise or logical XOR of numbers.
234 /// This POD struct holds information about reduction variables.
235 struct ReductionDescriptor {
237 ReductionDescriptor():
238 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
241 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
242 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
244 // The starting value of the reduction.
245 // It does not have to be zero!
247 // The instruction who's value is used outside the loop.
248 Instruction *LoopExitInstr;
249 // The kind of the reduction.
253 // This POD struct holds information about the memory runtime legality
254 // check that a group of pointers do not overlap.
255 struct RuntimePointerCheck {
256 RuntimePointerCheck(): Need(false) {}
258 /// Reset the state of the pointer runtime information.
266 /// Insert a pointer and calculate the start and end SCEVs.
267 void insert_pointer(ScalarEvolution *SE, Loop *Lp, Value *Ptr) {
268 const SCEV *Sc = SE->getSCEV(Ptr);
269 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
270 assert(AR && "Invalid addrec expression");
271 const SCEV *Ex = SE->getExitCount(Lp, Lp->getHeader());
272 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
273 Pointers.push_back(Ptr);
274 Starts.push_back(AR->getStart());
275 Ends.push_back(ScEnd);
278 /// This flag indicates if we need to add the runtime check.
280 /// Holds the pointers that we need to check.
281 SmallVector<Value*, 2> Pointers;
282 /// Holds the pointer value at the beginning of the loop.
283 SmallVector<const SCEV*, 2> Starts;
284 /// Holds the pointer value at the end of the loop.
285 SmallVector<const SCEV*, 2> Ends;
288 /// ReductionList contains the reduction descriptors for all
289 /// of the reductions that were found in the loop.
290 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
292 /// InductionList saves induction variables and maps them to the initial
293 /// value entring the loop.
294 typedef DenseMap<PHINode*, Value*> InductionList;
296 /// Returns true if it is legal to vectorize this loop.
297 /// This does not mean that it is profitable to vectorize this
298 /// loop, only that it is legal to do so.
301 /// Returns the Induction variable.
302 PHINode *getInduction() {return Induction;}
304 /// Returns the reduction variables found in the loop.
305 ReductionList *getReductionVars() { return &Reductions; }
307 /// Returns the induction variables found in the loop.
308 InductionList *getInductionVars() { return &Inductions; }
310 /// Check if this pointer is consecutive when vectorizing. This happens
311 /// when the last index of the GEP is the induction variable, or that the
312 /// pointer itself is an induction variable.
313 /// This check allows us to vectorize A[idx] into a wide load/store.
314 bool isConsecutivePtr(Value *Ptr);
316 /// Returns true if the value V is uniform within the loop.
317 bool isUniform(Value *V);
319 /// Returns true if this instruction will remain scalar after vectorization.
320 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
322 /// Returns the information that we collected about runtime memory check.
323 RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; }
325 /// Check if a single basic block loop is vectorizable.
326 /// At this point we know that this is a loop with a constant trip count
327 /// and we only need to check individual instructions.
328 bool canVectorizeBlock(BasicBlock &BB);
330 /// When we vectorize loops we may change the order in which
331 /// we read and write from memory. This method checks if it is
332 /// legal to vectorize the code, considering only memory constrains.
333 /// Returns true if BB is vectorizable
334 bool canVectorizeMemory(BasicBlock &BB);
336 /// Returns True, if 'Phi' is the kind of reduction variable for type
337 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
338 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
339 /// Returns true if the instruction I can be a reduction variable of type
341 bool isReductionInstr(Instruction *I, ReductionKind Kind);
342 /// Returns True, if 'Phi' is an induction variable.
343 bool isInductionVariable(PHINode *Phi);
344 /// Return true if can compute the address bounds of Ptr within the loop.
345 bool hasComputableBounds(Value *Ptr);
347 /// The loop that we evaluate.
351 /// DataLayout analysis.
354 // --- vectorization state --- //
356 /// Holds the integer induction variable. This is the counter of the
359 /// Holds the reduction variables.
360 ReductionList Reductions;
361 /// Holds all of the induction variables that we found in the loop.
362 /// Notice that inductions don't need to start at zero and that induction
363 /// variables can be pointers.
364 InductionList Inductions;
366 /// Allowed outside users. This holds the reduction
367 /// vars which can be accessed from outside the loop.
368 SmallPtrSet<Value*, 4> AllowedExit;
369 /// This set holds the variables which are known to be uniform after
371 SmallPtrSet<Instruction*, 4> Uniforms;
372 /// We need to check that all of the pointers in this list are disjoint
374 RuntimePointerCheck PtrRtCheck;
377 /// LoopVectorizationCostModel - estimates the expected speedups due to
379 /// In many cases vectorization is not profitable. This can happen because
380 /// of a number of reasons. In this class we mainly attempt to predict
381 /// the expected speedup/slowdowns due to the supported instruction set.
382 /// We use the VectorTargetTransformInfo to query the different backends
383 /// for the cost of different operations.
384 class LoopVectorizationCostModel {
387 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
388 LoopVectorizationLegality *Leg,
389 const VectorTargetTransformInfo *Vtti):
390 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
392 /// Returns the most profitable vectorization factor for the loop that is
393 /// smaller or equal to the VF argument. This method checks every power
395 unsigned findBestVectorizationFactor(unsigned VF = 8);
398 /// Returns the expected execution cost. The unit of the cost does
399 /// not matter because we use the 'cost' units to compare different
400 /// vector widths. The cost that is returned is *not* normalized by
401 /// the factor width.
402 unsigned expectedCost(unsigned VF);
404 /// Returns the execution time cost of an instruction for a given vector
405 /// width. Vector width of one means scalar.
406 unsigned getInstructionCost(Instruction *I, unsigned VF);
408 /// A helper function for converting Scalar types to vector types.
409 /// If the incoming type is void, we return void. If the VF is 1, we return
411 static Type* ToVectorTy(Type *Scalar, unsigned VF);
413 /// The loop that we evaluate.
418 /// Vectorization legality.
419 LoopVectorizationLegality *Legal;
420 /// Vector target information.
421 const VectorTargetTransformInfo *VTTI;
424 struct LoopVectorize : public LoopPass {
425 static char ID; // Pass identification, replacement for typeid
427 LoopVectorize() : LoopPass(ID) {
428 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
434 TargetTransformInfo *TTI;
437 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
438 // We only vectorize innermost loops.
442 SE = &getAnalysis<ScalarEvolution>();
443 DL = getAnalysisIfAvailable<DataLayout>();
444 LI = &getAnalysis<LoopInfo>();
445 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
446 DT = &getAnalysis<DominatorTree>();
448 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
449 L->getHeader()->getParent()->getName() << "\"\n");
451 // Check if it is legal to vectorize the loop.
452 LoopVectorizationLegality LVL(L, SE, DL);
453 if (!LVL.canVectorize()) {
454 DEBUG(dbgs() << "LV: Not vectorizing.\n");
458 // Select the preffered vectorization factor.
460 if (VectorizationFactor == 0) {
461 const VectorTargetTransformInfo *VTTI = 0;
463 VTTI = TTI->getVectorTargetTransformInfo();
464 // Use the cost model.
465 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
466 VF = CM.findBestVectorizationFactor();
469 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
474 // Use the user command flag.
475 VF = VectorizationFactor;
478 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
479 L->getHeader()->getParent()->getParent()->getModuleIdentifier()<<
482 // If we decided that it is *legal* to vectorizer the loop then do it.
483 SingleBlockLoopVectorizer LB(L, SE, LI, DT, DL, &LPM, VF);
486 DEBUG(verifyFunction(*L->getHeader()->getParent()));
490 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
491 LoopPass::getAnalysisUsage(AU);
492 AU.addRequiredID(LoopSimplifyID);
493 AU.addRequiredID(LCSSAID);
494 AU.addRequired<LoopInfo>();
495 AU.addRequired<ScalarEvolution>();
496 AU.addRequired<DominatorTree>();
497 AU.addPreserved<LoopInfo>();
498 AU.addPreserved<DominatorTree>();
503 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
505 LLVMContext &C = V->getContext();
506 Type *VTy = VectorType::get(V->getType(), VF);
507 Type *I32 = IntegerType::getInt32Ty(C);
508 Constant *Zero = ConstantInt::get(I32, 0);
509 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
510 Value *UndefVal = UndefValue::get(VTy);
511 // Insert the value into a new vector.
512 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
513 // Broadcast the scalar into all locations in the vector.
514 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
516 // We are accessing the induction variable. Make sure to promote the
517 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
519 return getConsecutiveVector(Shuf);
523 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
524 assert(Val->getType()->isVectorTy() && "Must be a vector");
525 assert(Val->getType()->getScalarType()->isIntegerTy() &&
526 "Elem must be an integer");
528 Type *ITy = Val->getType()->getScalarType();
529 VectorType *Ty = cast<VectorType>(Val->getType());
530 unsigned VLen = Ty->getNumElements();
531 SmallVector<Constant*, 8> Indices;
533 // Create a vector of consecutive numbers from zero to VF.
534 for (unsigned i = 0; i < VLen; ++i)
535 Indices.push_back(ConstantInt::get(ITy, i));
537 // Add the consecutive indices to the vector value.
538 Constant *Cv = ConstantVector::get(Indices);
539 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
540 return Builder.CreateAdd(Val, Cv, "induction");
543 bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
544 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
546 // If this pointer is an induction variable, return it.
547 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
548 if (Phi && getInductionVars()->count(Phi))
551 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
555 unsigned NumOperands = Gep->getNumOperands();
556 Value *LastIndex = Gep->getOperand(NumOperands - 1);
558 // Check that all of the gep indices are uniform except for the last.
559 for (unsigned i = 0; i < NumOperands - 1; ++i)
560 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
563 // We can emit wide load/stores only of the last index is the induction
565 const SCEV *Last = SE->getSCEV(LastIndex);
566 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
567 const SCEV *Step = AR->getStepRecurrence(*SE);
569 // The memory is consecutive because the last index is consecutive
570 // and all other indices are loop invariant.
578 bool LoopVectorizationLegality::isUniform(Value *V) {
579 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
582 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
583 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
584 // If we saved a vectorized copy of V, use it.
585 Value *&MapEntry = WidenMap[V];
589 // Broadcast V and save the value for future uses.
590 Value *B = getBroadcastInstrs(V);
596 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
597 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
600 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
601 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
602 // Holds vector parameters or scalars, in case of uniform vals.
603 SmallVector<Value*, 8> Params;
605 // Find all of the vectorized parameters.
606 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
607 Value *SrcOp = Instr->getOperand(op);
609 // If we are accessing the old induction variable, use the new one.
610 if (SrcOp == OldInduction) {
611 Params.push_back(getVectorValue(Induction));
615 // Try using previously calculated values.
616 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
618 // If the src is an instruction that appeared earlier in the basic block
619 // then it should already be vectorized.
620 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
621 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
622 // The parameter is a vector value from earlier.
623 Params.push_back(WidenMap[SrcInst]);
625 // The parameter is a scalar from outside the loop. Maybe even a constant.
626 Params.push_back(SrcOp);
630 assert(Params.size() == Instr->getNumOperands() &&
631 "Invalid number of operands");
633 // Does this instruction return a value ?
634 bool IsVoidRetTy = Instr->getType()->isVoidTy();
635 Value *VecResults = 0;
637 // If we have a return value, create an empty vector. We place the scalarized
638 // instructions in this vector.
640 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
642 // For each scalar that we create:
643 for (unsigned i = 0; i < VF; ++i) {
644 Instruction *Cloned = Instr->clone();
646 Cloned->setName(Instr->getName() + ".cloned");
647 // Replace the operands of the cloned instrucions with extracted scalars.
648 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
649 Value *Op = Params[op];
650 // Param is a vector. Need to extract the right lane.
651 if (Op->getType()->isVectorTy())
652 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
653 Cloned->setOperand(op, Op);
656 // Place the cloned scalar in the new loop.
657 Builder.Insert(Cloned);
659 // If the original scalar returns a value we need to place it in a vector
660 // so that future users will be able to use it.
662 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
663 Builder.getInt32(i));
667 WidenMap[Instr] = VecResults;
671 SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
673 In this function we generate a new loop. The new loop will contain
674 the vectorized instructions while the old loop will continue to run the
677 [ ] <-- vector loop bypass.
680 | [ ] <-- vector pre header.
684 | [ ]_| <-- vector loop.
687 >[ ] <--- middle-block.
690 | [ ] <--- new preheader.
694 | [ ]_| <-- old scalar loop to handle remainder.
701 // Some loops have a single integer induction variable, while other loops
702 // don't. One example is c++ iterators that often have multiple pointer
703 // induction variables. In the code below we also support a case where we
704 // don't have a single induction variable.
705 OldInduction = Legal->getInduction();
706 Type *IdxTy = OldInduction ? OldInduction->getType() :
707 DL->getIntPtrType(SE->getContext());
709 // Find the loop boundaries.
710 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
711 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
713 // Get the total trip count from the count by adding 1.
714 ExitCount = SE->getAddExpr(ExitCount,
715 SE->getConstant(ExitCount->getType(), 1));
717 // This is the original scalar-loop preheader.
718 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
719 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
720 assert(ExitBlock && "Must have an exit block");
722 // The loop index does not have to start at Zero. Find the original start
723 // value from the induction PHI node. If we don't have an induction variable
724 // then we know that it starts at zero.
725 Value *StartIdx = OldInduction ?
726 OldInduction->getIncomingValueForBlock(BypassBlock):
727 ConstantInt::get(IdxTy, 0);
729 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
730 assert(BypassBlock && "Invalid loop structure");
732 BasicBlock *VectorPH =
733 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
734 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
737 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
739 BasicBlock *ScalarPH =
740 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
742 // Find the induction variable.
743 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
745 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
747 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
749 // Generate the induction variable.
750 Induction = Builder.CreatePHI(IdxTy, 2, "index");
751 Constant *Step = ConstantInt::get(IdxTy, VF);
753 // Expand the trip count and place the new instructions in the preheader.
754 // Notice that the pre-header does not change, only the loop body.
755 SCEVExpander Exp(*SE, "induction");
756 Instruction *Loc = BypassBlock->getTerminator();
758 // Count holds the overall loop count (N).
759 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), Loc);
761 // We may need to extend the index in case there is a type mismatch.
762 // We know that the count starts at zero and does not overflow.
763 if (Count->getType() != IdxTy) {
764 // The exit count can be of pointer type. Convert it to the correct
766 if (ExitCount->getType()->isPointerTy())
767 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
769 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
772 // Add the start index to the loop count to get the new end index.
773 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
775 // Now we need to generate the expression for N - (N % VF), which is
776 // the part that the vectorized body will execute.
777 Constant *CIVF = ConstantInt::get(IdxTy, VF);
778 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
779 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
780 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
781 "end.idx.rnd.down", Loc);
783 // Now, compare the new count to zero. If it is zero skip the vector loop and
784 // jump to the scalar loop.
785 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
790 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
791 Legal->getRuntimePointerCheck();
792 Value *MemoryRuntimeCheck = 0;
793 if (PtrRtCheck->Need) {
794 unsigned NumPointers = PtrRtCheck->Pointers.size();
795 SmallVector<Value* , 2> Starts;
796 SmallVector<Value* , 2> Ends;
798 // Use this type for pointer arithmetic.
799 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
801 for (unsigned i=0; i < NumPointers; ++i) {
802 Value *Ptr = PtrRtCheck->Pointers[i];
803 const SCEV *Sc = SE->getSCEV(Ptr);
805 if (SE->isLoopInvariant(Sc, OrigLoop)) {
806 DEBUG(dbgs() << "LV1: Adding RT check for a loop invariant ptr:" <<
808 Starts.push_back(Ptr);
811 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
813 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i],
815 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
816 Starts.push_back(Start);
821 for (unsigned i = 0; i < NumPointers; ++i) {
822 for (unsigned j = i+1; j < NumPointers; ++j) {
823 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
824 Starts[i], Ends[j], "bound0", Loc);
825 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
826 Starts[j], Ends[i], "bound1", Loc);
827 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
828 "found.conflict", Loc);
829 if (MemoryRuntimeCheck) {
830 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
833 "conflict.rdx", Loc);
835 MemoryRuntimeCheck = IsConflict;
839 }// end of need-runtime-check code.
841 // If we are using memory runtime checks, include them in.
842 if (MemoryRuntimeCheck) {
843 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
847 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
848 // Remove the old terminator.
849 Loc->eraseFromParent();
851 // We are going to resume the execution of the scalar loop.
852 // Go over all of the induction variables that we found and fix the
853 // PHIs that are left in the scalar version of the loop.
854 // The starting values of PHI nodes depend on the counter of the last
855 // iteration in the vectorized loop.
856 // If we come from a bypass edge then we need to start from the original start
859 // This variable saves the new starting index for the scalar loop.
860 PHINode *ResumeIndex = 0;
861 LoopVectorizationLegality::InductionList::iterator I, E;
862 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
863 for (I = List->begin(), E = List->end(); I != E; ++I) {
864 PHINode *OrigPhi = I->first;
865 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
866 MiddleBlock->getTerminator());
868 if (OrigPhi->getType()->isIntegerTy()) {
869 // Handle the integer induction counter:
870 assert(OrigPhi == OldInduction && "Unknown integer PHI");
871 // We know what the end value is.
872 EndValue = IdxEndRoundDown;
873 // We also know which PHI node holds it.
874 ResumeIndex = ResumeVal;
876 // For pointer induction variables, calculate the offset using
878 EndValue = GetElementPtrInst::Create(I->second, CountRoundDown,
880 BypassBlock->getTerminator());
883 // The new PHI merges the original incoming value, in case of a bypass,
884 // or the value at the end of the vectorized loop.
885 ResumeVal->addIncoming(I->second, BypassBlock);
886 ResumeVal->addIncoming(EndValue, VecBody);
888 // Fix the scalar body counter (PHI node).
889 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
890 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
893 // If we are generating a new induction variable then we also need to
894 // generate the code that calculates the exit value. This value is not
895 // simply the end of the counter because we may skip the vectorized body
896 // in case of a runtime check.
898 assert(!ResumeIndex && "Unexpected resume value found");
899 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
900 MiddleBlock->getTerminator());
901 ResumeIndex->addIncoming(StartIdx, BypassBlock);
902 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
905 // Make sure that we found the index where scalar loop needs to continue.
906 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
907 "Invalid resume Index");
909 // Add a check in the middle block to see if we have completed
910 // all of the iterations in the first vector loop.
911 // If (N - N%VF) == N, then we *don't* need to run the remainder.
912 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
913 ResumeIndex, "cmp.n",
914 MiddleBlock->getTerminator());
916 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
917 // Remove the old terminator.
918 MiddleBlock->getTerminator()->eraseFromParent();
920 // Create i+1 and fill the PHINode.
921 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
922 Induction->addIncoming(StartIdx, VectorPH);
923 Induction->addIncoming(NextIdx, VecBody);
924 // Create the compare.
925 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
926 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
928 // Now we have two terminators. Remove the old one from the block.
929 VecBody->getTerminator()->eraseFromParent();
931 // Get ready to start creating new instructions into the vectorized body.
932 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
934 // Register the new loop.
935 Loop* Lp = new Loop();
936 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
938 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
940 Loop *ParentLoop = OrigLoop->getParentLoop();
942 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
943 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
944 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
948 LoopVectorPreHeader = VectorPH;
949 LoopScalarPreHeader = ScalarPH;
950 LoopMiddleBlock = MiddleBlock;
951 LoopExitBlock = ExitBlock;
952 LoopVectorBody = VecBody;
953 LoopScalarBody = OldBasicBlock;
954 LoopBypassBlock = BypassBlock;
957 /// This function returns the identity element (or neutral element) for
960 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
962 case LoopVectorizationLegality::IntegerXor:
963 case LoopVectorizationLegality::IntegerAdd:
964 case LoopVectorizationLegality::IntegerOr:
965 // Adding, Xoring, Oring zero to a number does not change it.
967 case LoopVectorizationLegality::IntegerMult:
968 // Multiplying a number by 1 does not change it.
970 case LoopVectorizationLegality::IntegerAnd:
971 // AND-ing a number with an all-1 value does not change it.
974 llvm_unreachable("Unknown reduction kind");
979 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
980 //===------------------------------------------------===//
982 // Notice: any optimization or new instruction that go
983 // into the code below should be also be implemented in
986 //===------------------------------------------------===//
987 typedef SmallVector<PHINode*, 4> PhiVector;
988 BasicBlock &BB = *OrigLoop->getHeader();
989 Constant *Zero = ConstantInt::get(
990 IntegerType::getInt32Ty(BB.getContext()), 0);
992 // In order to support reduction variables we need to be able to vectorize
993 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
994 // steages. First, we create a new vector PHI node with no incoming edges.
995 // We use this value when we vectorize all of the instructions that use the
996 // PHI. Next, after all of the instructions in the block are complete we
997 // add the new incoming edges to the PHI. At this point all of the
998 // instructions in the basic block are vectorized, so we can use them to
999 // construct the PHI.
1000 PhiVector RdxPHIsToFix;
1002 // For each instruction in the old loop.
1003 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1004 Instruction *Inst = it;
1006 switch (Inst->getOpcode()) {
1007 case Instruction::Br:
1008 // Nothing to do for PHIs and BR, since we already took care of the
1009 // loop control flow instructions.
1011 case Instruction::PHI:{
1012 PHINode* P = cast<PHINode>(Inst);
1013 // Handle reduction variables:
1014 if (Legal->getReductionVars()->count(P)) {
1015 // This is phase one of vectorizing PHIs.
1016 Type *VecTy = VectorType::get(Inst->getType(), VF);
1017 WidenMap[Inst] = PHINode::Create(VecTy, 2, "vec.phi",
1018 LoopVectorBody->getFirstInsertionPt());
1019 RdxPHIsToFix.push_back(P);
1023 // This PHINode must be an induction variable.
1024 // Make sure that we know about it.
1025 assert(Legal->getInductionVars()->count(P) &&
1026 "Not an induction variable");
1028 if (P->getType()->isIntegerTy()) {
1029 assert(P == OldInduction && "Unexpected PHI");
1030 WidenMap[Inst] = getBroadcastInstrs(Induction);
1034 // Handle pointer inductions:
1035 assert(P->getType()->isPointerTy() && "Unexpected type.");
1036 Value *StartIdx = OldInduction ?
1037 Legal->getInductionVars()->lookup(OldInduction) :
1038 ConstantInt::get(Induction->getType(), 0);
1040 // This is the pointer value coming into the loop.
1041 Value *StartPtr = Legal->getInductionVars()->lookup(P);
1043 // This is the normalized GEP that starts counting at zero.
1044 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1047 // This is the vector of results. Notice that we don't generate vector
1048 // geps because scalar geps result in better code.
1049 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1050 for (unsigned int i = 0; i < VF; ++i) {
1051 Constant *Idx = ConstantInt::get(Induction->getType(), i);
1052 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1053 Value *SclrGep = Builder.CreateGEP(StartPtr, GlobalIdx, "next.gep");
1054 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1055 Builder.getInt32(i),
1059 WidenMap[Inst] = VecVal;
1062 case Instruction::Add:
1063 case Instruction::FAdd:
1064 case Instruction::Sub:
1065 case Instruction::FSub:
1066 case Instruction::Mul:
1067 case Instruction::FMul:
1068 case Instruction::UDiv:
1069 case Instruction::SDiv:
1070 case Instruction::FDiv:
1071 case Instruction::URem:
1072 case Instruction::SRem:
1073 case Instruction::FRem:
1074 case Instruction::Shl:
1075 case Instruction::LShr:
1076 case Instruction::AShr:
1077 case Instruction::And:
1078 case Instruction::Or:
1079 case Instruction::Xor: {
1080 // Just widen binops.
1081 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1082 Value *A = getVectorValue(Inst->getOperand(0));
1083 Value *B = getVectorValue(Inst->getOperand(1));
1085 // Use this vector value for all users of the original instruction.
1086 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1089 // Update the NSW, NUW and Exact flags.
1090 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1091 if (isa<OverflowingBinaryOperator>(BinOp)) {
1092 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1093 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1095 if (isa<PossiblyExactOperator>(VecOp))
1096 VecOp->setIsExact(BinOp->isExact());
1099 case Instruction::Select: {
1101 // If the selector is loop invariant we can create a select
1102 // instruction with a scalar condition. Otherwise, use vector-select.
1103 Value *Cond = Inst->getOperand(0);
1104 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1106 // The condition can be loop invariant but still defined inside the
1107 // loop. This means that we can't just use the original 'cond' value.
1108 // We have to take the 'vectorized' value and pick the first lane.
1109 // Instcombine will make this a no-op.
1110 Cond = getVectorValue(Cond);
1112 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1114 Value *Op0 = getVectorValue(Inst->getOperand(1));
1115 Value *Op1 = getVectorValue(Inst->getOperand(2));
1116 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
1120 case Instruction::ICmp:
1121 case Instruction::FCmp: {
1122 // Widen compares. Generate vector compares.
1123 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
1124 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
1125 Value *A = getVectorValue(Inst->getOperand(0));
1126 Value *B = getVectorValue(Inst->getOperand(1));
1128 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1130 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1134 case Instruction::Store: {
1135 // Attempt to issue a wide store.
1136 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1137 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1138 Value *Ptr = SI->getPointerOperand();
1139 unsigned Alignment = SI->getAlignment();
1141 assert(!Legal->isUniform(Ptr) &&
1142 "We do not allow storing to uniform addresses");
1144 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1146 // This store does not use GEPs.
1147 if (!Legal->isConsecutivePtr(Ptr)) {
1148 scalarizeInstruction(Inst);
1153 // The last index does not have to be the induction. It can be
1154 // consecutive and be a function of the index. For example A[I+1];
1155 unsigned NumOperands = Gep->getNumOperands();
1156 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1157 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1159 // Create the new GEP with the new induction variable.
1160 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1161 Gep2->setOperand(NumOperands - 1, LastIndex);
1162 Ptr = Builder.Insert(Gep2);
1164 // Use the induction element ptr.
1165 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1166 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1168 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1169 Value *Val = getVectorValue(SI->getValueOperand());
1170 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1173 case Instruction::Load: {
1174 // Attempt to issue a wide load.
1175 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1176 Type *RetTy = VectorType::get(LI->getType(), VF);
1177 Value *Ptr = LI->getPointerOperand();
1178 unsigned Alignment = LI->getAlignment();
1179 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1181 // If the pointer is loop invariant or if it is non consecutive,
1182 // scalarize the load.
1183 bool Con = Legal->isConsecutivePtr(Ptr);
1184 if (Legal->isUniform(Ptr) || !Con) {
1185 scalarizeInstruction(Inst);
1190 // The last index does not have to be the induction. It can be
1191 // consecutive and be a function of the index. For example A[I+1];
1192 unsigned NumOperands = Gep->getNumOperands();
1193 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1194 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1196 // Create the new GEP with the new induction variable.
1197 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1198 Gep2->setOperand(NumOperands - 1, LastIndex);
1199 Ptr = Builder.Insert(Gep2);
1201 // Use the induction element ptr.
1202 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1203 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1206 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1207 LI = Builder.CreateLoad(Ptr);
1208 LI->setAlignment(Alignment);
1209 // Use this vector value for all users of the load.
1210 WidenMap[Inst] = LI;
1213 case Instruction::ZExt:
1214 case Instruction::SExt:
1215 case Instruction::FPToUI:
1216 case Instruction::FPToSI:
1217 case Instruction::FPExt:
1218 case Instruction::PtrToInt:
1219 case Instruction::IntToPtr:
1220 case Instruction::SIToFP:
1221 case Instruction::UIToFP:
1222 case Instruction::Trunc:
1223 case Instruction::FPTrunc:
1224 case Instruction::BitCast: {
1225 /// Vectorize bitcasts.
1226 CastInst *CI = dyn_cast<CastInst>(Inst);
1227 Value *A = getVectorValue(Inst->getOperand(0));
1228 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1229 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1234 /// All other instructions are unsupported. Scalarize them.
1235 scalarizeInstruction(Inst);
1238 }// end of for_each instr.
1240 // At this point every instruction in the original loop is widended to
1241 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1242 // that we vectorized. The PHI nodes are currently empty because we did
1243 // not want to introduce cycles. Notice that the remaining PHI nodes
1244 // that we need to fix are reduction variables.
1246 // Create the 'reduced' values for each of the induction vars.
1247 // The reduced values are the vector values that we scalarize and combine
1248 // after the loop is finished.
1249 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1251 PHINode *RdxPhi = *it;
1252 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1253 assert(RdxPhi && "Unable to recover vectorized PHI");
1255 // Find the reduction variable descriptor.
1256 assert(Legal->getReductionVars()->count(RdxPhi) &&
1257 "Unable to find the reduction variable");
1258 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1259 (*Legal->getReductionVars())[RdxPhi];
1261 // We need to generate a reduction vector from the incoming scalar.
1262 // To do so, we need to generate the 'identity' vector and overide
1263 // one of the elements with the incoming scalar reduction. We need
1264 // to do it in the vector-loop preheader.
1265 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1267 // This is the vector-clone of the value that leaves the loop.
1268 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1269 Type *VecTy = VectorExit->getType();
1271 // Find the reduction identity variable. Zero for addition, or, xor,
1272 // one for multiplication, -1 for And.
1273 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1274 VecTy->getScalarType());
1276 // This vector is the Identity vector where the first element is the
1277 // incoming scalar reduction.
1278 Value *VectorStart = Builder.CreateInsertElement(Identity,
1279 RdxDesc.StartValue, Zero);
1281 // Fix the vector-loop phi.
1282 // We created the induction variable so we know that the
1283 // preheader is the first entry.
1284 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1286 // Reductions do not have to start at zero. They can start with
1287 // any loop invariant values.
1288 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1289 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1290 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1291 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1293 // Before each round, move the insertion point right between
1294 // the PHIs and the values we are going to write.
1295 // This allows us to write both PHINodes and the extractelement
1297 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1299 // This PHINode contains the vectorized reduction variable, or
1300 // the initial value vector, if we bypass the vector loop.
1301 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1302 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1303 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1305 // Extract the first scalar.
1307 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1308 // Extract and reduce the remaining vector elements.
1309 for (unsigned i=1; i < VF; ++i) {
1311 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1312 switch (RdxDesc.Kind) {
1313 case LoopVectorizationLegality::IntegerAdd:
1314 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1316 case LoopVectorizationLegality::IntegerMult:
1317 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1319 case LoopVectorizationLegality::IntegerOr:
1320 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1322 case LoopVectorizationLegality::IntegerAnd:
1323 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1325 case LoopVectorizationLegality::IntegerXor:
1326 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1329 llvm_unreachable("Unknown reduction operation");
1333 // Now, we need to fix the users of the reduction variable
1334 // inside and outside of the scalar remainder loop.
1335 // We know that the loop is in LCSSA form. We need to update the
1336 // PHI nodes in the exit blocks.
1337 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1338 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1339 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1340 if (!LCSSAPhi) continue;
1342 // All PHINodes need to have a single entry edge, or two if
1343 // we already fixed them.
1344 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1346 // We found our reduction value exit-PHI. Update it with the
1347 // incoming bypass edge.
1348 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1349 // Add an edge coming from the bypass.
1350 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1353 }// end of the LCSSA phi scan.
1355 // Fix the scalar loop reduction variable with the incoming reduction sum
1356 // from the vector body and from the backedge value.
1357 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1358 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1359 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1360 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1361 }// end of for each redux variable.
1364 void SingleBlockLoopVectorizer::updateAnalysis() {
1365 // The original basic block.
1366 SE->forgetLoop(OrigLoop);
1368 // Update the dominator tree information.
1369 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1370 "Entry does not dominate exit.");
1372 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1373 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1374 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1375 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1376 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1377 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1379 DEBUG(DT->verifyAnalysis());
1382 bool LoopVectorizationLegality::canVectorize() {
1383 if (!TheLoop->getLoopPreheader()) {
1384 assert(false && "No preheader!!");
1385 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1389 // We can only vectorize single basic block loops.
1390 unsigned NumBlocks = TheLoop->getNumBlocks();
1391 if (NumBlocks != 1) {
1392 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1396 // We need to have a loop header.
1397 BasicBlock *BB = TheLoop->getHeader();
1398 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1400 // ScalarEvolution needs to be able to find the exit count.
1401 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1402 if (ExitCount == SE->getCouldNotCompute()) {
1403 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1407 // Do not loop-vectorize loops with a tiny trip count.
1408 unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB);
1409 if (TC > 0u && TC < TinyTripCountThreshold) {
1410 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1411 "This loop is not worth vectorizing.\n");
1415 // Go over each instruction and look at memory deps.
1416 if (!canVectorizeBlock(*BB)) {
1417 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1421 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1422 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1425 // Okay! We can vectorize. At this point we don't have any other mem analysis
1426 // which may limit our maximum vectorization factor, so just return true with
1431 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1433 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1435 // Scan the instructions in the block and look for hazards.
1436 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1437 Instruction *I = it;
1439 if (PHINode *Phi = dyn_cast<PHINode>(I)) {
1440 // This should not happen because the loop should be normalized.
1441 if (Phi->getNumIncomingValues() != 2) {
1442 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1446 // This is the value coming from the preheader.
1447 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1449 // We only look at integer and pointer phi nodes.
1450 if (Phi->getType()->isPointerTy() && isInductionVariable(Phi)) {
1451 DEBUG(dbgs() << "LV: Found a pointer induction variable.\n");
1452 Inductions[Phi] = StartValue;
1454 } else if (!Phi->getType()->isIntegerTy()) {
1455 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1459 // Handle integer PHIs:
1460 if (isInductionVariable(Phi)) {
1462 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1465 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1467 Inductions[Phi] = StartValue;
1470 if (AddReductionVar(Phi, IntegerAdd)) {
1471 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1474 if (AddReductionVar(Phi, IntegerMult)) {
1475 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1478 if (AddReductionVar(Phi, IntegerOr)) {
1479 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1482 if (AddReductionVar(Phi, IntegerAnd)) {
1483 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1486 if (AddReductionVar(Phi, IntegerXor)) {
1487 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1491 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1493 }// end of PHI handling
1495 // We still don't handle functions.
1496 CallInst *CI = dyn_cast<CallInst>(I);
1498 DEBUG(dbgs() << "LV: Found a call site.\n");
1502 // We do not re-vectorize vectors.
1503 if (!VectorType::isValidElementType(I->getType()) &&
1504 !I->getType()->isVoidTy()) {
1505 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1509 // Reduction instructions are allowed to have exit users.
1510 // All other instructions must not have external users.
1511 if (!AllowedExit.count(I))
1512 //Check that all of the users of the loop are inside the BB.
1513 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1515 Instruction *U = cast<Instruction>(*it);
1516 // This user may be a reduction exit value.
1517 BasicBlock *Parent = U->getParent();
1518 if (Parent != &BB) {
1519 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1526 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
1527 assert(getInductionVars()->size() && "No induction variables");
1530 // Don't vectorize if the memory dependencies do not allow vectorization.
1531 if (!canVectorizeMemory(BB))
1534 // We now know that the loop is vectorizable!
1535 // Collect variables that will remain uniform after vectorization.
1536 std::vector<Value*> Worklist;
1538 // Start with the conditional branch and walk up the block.
1539 Worklist.push_back(BB.getTerminator()->getOperand(0));
1541 while (Worklist.size()) {
1542 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1543 Worklist.pop_back();
1545 // Look at instructions inside this block. Stop when reaching PHI nodes.
1546 if (!I || I->getParent() != &BB || isa<PHINode>(I))
1549 // This is a known uniform.
1552 // Insert all operands.
1553 for (int i=0, Op = I->getNumOperands(); i < Op; ++i) {
1554 Worklist.push_back(I->getOperand(i));
1561 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1562 typedef SmallVector<Value*, 16> ValueVector;
1563 typedef SmallPtrSet<Value*, 16> ValueSet;
1564 // Holds the Load and Store *instructions*.
1567 PtrRtCheck.Pointers.clear();
1568 PtrRtCheck.Need = false;
1570 // Scan the BB and collect legal loads and stores.
1571 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1572 Instruction *I = it;
1574 // If this is a load, save it. If this instruction can read from memory
1575 // but is not a load, then we quit. Notice that we don't handle function
1576 // calls that read or write.
1577 if (I->mayReadFromMemory()) {
1578 LoadInst *Ld = dyn_cast<LoadInst>(I);
1579 if (!Ld) return false;
1580 if (!Ld->isSimple()) {
1581 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1584 Loads.push_back(Ld);
1588 // Save store instructions. Abort if other instructions write to memory.
1589 if (I->mayWriteToMemory()) {
1590 StoreInst *St = dyn_cast<StoreInst>(I);
1591 if (!St) return false;
1592 if (!St->isSimple()) {
1593 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1596 Stores.push_back(St);
1600 // Now we have two lists that hold the loads and the stores.
1601 // Next, we find the pointers that they use.
1603 // Check if we see any stores. If there are no stores, then we don't
1604 // care if the pointers are *restrict*.
1605 if (!Stores.size()) {
1606 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1610 // Holds the read and read-write *pointers* that we find.
1612 ValueVector ReadWrites;
1614 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1615 // multiple times on the same object. If the ptr is accessed twice, once
1616 // for read and once for write, it will only appear once (on the write
1617 // list). This is okay, since we are going to check for conflicts between
1618 // writes and between reads and writes, but not between reads and reads.
1621 ValueVector::iterator I, IE;
1622 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1623 StoreInst *ST = dyn_cast<StoreInst>(*I);
1624 assert(ST && "Bad StoreInst");
1625 Value* Ptr = ST->getPointerOperand();
1627 if (isUniform(Ptr)) {
1628 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1632 // If we did *not* see this pointer before, insert it to
1633 // the read-write list. At this phase it is only a 'write' list.
1634 if (Seen.insert(Ptr))
1635 ReadWrites.push_back(Ptr);
1638 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1639 LoadInst *LD = dyn_cast<LoadInst>(*I);
1640 assert(LD && "Bad LoadInst");
1641 Value* Ptr = LD->getPointerOperand();
1642 // If we did *not* see this pointer before, insert it to the
1643 // read list. If we *did* see it before, then it is already in
1644 // the read-write list. This allows us to vectorize expressions
1645 // such as A[i] += x; Because the address of A[i] is a read-write
1646 // pointer. This only works if the index of A[i] is consecutive.
1647 // If the address of i is unknown (for example A[B[i]]) then we may
1648 // read a few words, modify, and write a few words, and some of the
1649 // words may be written to the same address.
1650 if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr))
1651 Reads.push_back(Ptr);
1654 // If we write (or read-write) to a single destination and there are no
1655 // other reads in this loop then is it safe to vectorize.
1656 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1657 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1661 // Find pointers with computable bounds. We are going to use this information
1662 // to place a runtime bound check.
1664 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1665 if (hasComputableBounds(*I)) {
1666 PtrRtCheck.insert_pointer(SE, TheLoop, *I);
1667 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1672 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1673 if (hasComputableBounds(*I)) {
1674 PtrRtCheck.insert_pointer(SE, TheLoop, *I);
1675 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1681 // Check that we did not collect too many pointers or found a
1682 // unsizeable pointer.
1683 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1688 PtrRtCheck.Need = RT;
1691 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1694 // Now that the pointers are in two lists (Reads and ReadWrites), we
1695 // can check that there are no conflicts between each of the writes and
1696 // between the writes to the reads.
1697 ValueSet WriteObjects;
1698 ValueVector TempObjects;
1700 // Check that the read-writes do not conflict with other read-write
1702 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1703 GetUnderlyingObjects(*I, TempObjects, DL);
1704 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1706 if (!isIdentifiedObject(*it)) {
1707 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1710 if (!WriteObjects.insert(*it)) {
1711 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1716 TempObjects.clear();
1719 /// Check that the reads don't conflict with the read-writes.
1720 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1721 GetUnderlyingObjects(*I, TempObjects, DL);
1722 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1724 if (!isIdentifiedObject(*it)) {
1725 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1728 if (WriteObjects.count(*it)) {
1729 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1734 TempObjects.clear();
1737 // It is safe to vectorize and we don't need any runtime checks.
1738 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1743 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1744 ReductionKind Kind) {
1745 if (Phi->getNumIncomingValues() != 2)
1748 // Find the possible incoming reduction variable.
1749 BasicBlock *BB = Phi->getParent();
1750 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1751 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1752 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1754 // ExitInstruction is the single value which is used outside the loop.
1755 // We only allow for a single reduction value to be used outside the loop.
1756 // This includes users of the reduction, variables (which form a cycle
1757 // which ends in the phi node).
1758 Instruction *ExitInstruction = 0;
1760 // Iter is our iterator. We start with the PHI node and scan for all of the
1761 // users of this instruction. All users must be instructions which can be
1762 // used as reduction variables (such as ADD). We may have a single
1763 // out-of-block user. They cycle must end with the original PHI.
1764 // Also, we can't have multiple block-local users.
1765 Instruction *Iter = Phi;
1767 // Any reduction instr must be of one of the allowed kinds.
1768 if (!isReductionInstr(Iter, Kind))
1771 // Did we found a user inside this block ?
1772 bool FoundInBlockUser = false;
1773 // Did we reach the initial PHI node ?
1774 bool FoundStartPHI = false;
1776 // If the instruction has no users then this is a broken
1777 // chain and can't be a reduction variable.
1778 if (Iter->use_empty())
1781 // For each of the *users* of iter.
1782 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1784 Instruction *U = cast<Instruction>(*it);
1785 // We already know that the PHI is a user.
1787 FoundStartPHI = true;
1790 // Check if we found the exit user.
1791 BasicBlock *Parent = U->getParent();
1793 // We must have a single exit instruction.
1794 if (ExitInstruction != 0)
1796 ExitInstruction = Iter;
1798 // We can't have multiple inside users.
1799 if (FoundInBlockUser)
1801 FoundInBlockUser = true;
1805 // We found a reduction var if we have reached the original
1806 // phi node and we only have a single instruction with out-of-loop
1808 if (FoundStartPHI && ExitInstruction) {
1809 // This instruction is allowed to have out-of-loop users.
1810 AllowedExit.insert(ExitInstruction);
1812 // Save the description of this reduction variable.
1813 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1814 Reductions[Phi] = RD;
1821 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1822 ReductionKind Kind) {
1823 switch (I->getOpcode()) {
1826 case Instruction::PHI:
1829 case Instruction::Add:
1830 case Instruction::Sub:
1831 return Kind == IntegerAdd;
1832 case Instruction::Mul:
1833 return Kind == IntegerMult;
1834 case Instruction::And:
1835 return Kind == IntegerAnd;
1836 case Instruction::Or:
1837 return Kind == IntegerOr;
1838 case Instruction::Xor:
1839 return Kind == IntegerXor;
1843 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1844 Type *PhiTy = Phi->getType();
1845 // We only handle integer and pointer inductions variables.
1846 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1849 // Check that the PHI is consecutive and starts at zero.
1850 const SCEV *PhiScev = SE->getSCEV(Phi);
1851 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1853 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1856 const SCEV *Step = AR->getStepRecurrence(*SE);
1858 // Integer inductions need to have a stride of one.
1859 if (PhiTy->isIntegerTy())
1860 return Step->isOne();
1862 // Calculate the pointer stride and check if it is consecutive.
1863 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1864 if (!C) return false;
1866 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1867 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1868 return (C->getValue()->equalsInt(Size));
1871 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1872 const SCEV *PhiScev = SE->getSCEV(Ptr);
1873 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1877 return AR->isAffine();
1881 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1883 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1887 float Cost = expectedCost(1);
1889 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1890 for (unsigned i=2; i <= VF; i*=2) {
1891 // Notice that the vector loop needs to be executed less times, so
1892 // we need to divide the cost of the vector loops by the width of
1893 // the vector elements.
1894 float VectorCost = expectedCost(i) / (float)i;
1895 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1896 (int)VectorCost << ".\n");
1897 if (VectorCost < Cost) {
1903 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1907 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1908 // We can only estimate the cost of single basic block loops.
1909 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1911 BasicBlock *BB = TheLoop->getHeader();
1914 // For each instruction in the old loop.
1915 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1916 Instruction *Inst = it;
1917 unsigned C = getInstructionCost(Inst, VF);
1919 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1920 " For instruction: "<< *Inst << "\n");
1927 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1928 assert(VTTI && "Invalid vector target transformation info");
1930 // If we know that this instruction will remain uniform, check the cost of
1931 // the scalar version.
1932 if (Legal->isUniformAfterVectorization(I))
1935 Type *RetTy = I->getType();
1936 Type *VectorTy = ToVectorTy(RetTy, VF);
1939 // TODO: We need to estimate the cost of intrinsic calls.
1940 switch (I->getOpcode()) {
1941 case Instruction::GetElementPtr:
1942 // We mark this instruction as zero-cost because scalar GEPs are usually
1943 // lowered to the intruction addressing mode. At the moment we don't
1944 // generate vector geps.
1946 case Instruction::Br: {
1947 return VTTI->getCFInstrCost(I->getOpcode());
1949 case Instruction::PHI:
1951 case Instruction::Add:
1952 case Instruction::FAdd:
1953 case Instruction::Sub:
1954 case Instruction::FSub:
1955 case Instruction::Mul:
1956 case Instruction::FMul:
1957 case Instruction::UDiv:
1958 case Instruction::SDiv:
1959 case Instruction::FDiv:
1960 case Instruction::URem:
1961 case Instruction::SRem:
1962 case Instruction::FRem:
1963 case Instruction::Shl:
1964 case Instruction::LShr:
1965 case Instruction::AShr:
1966 case Instruction::And:
1967 case Instruction::Or:
1968 case Instruction::Xor: {
1969 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
1971 case Instruction::Select: {
1972 SelectInst *SI = cast<SelectInst>(I);
1973 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1974 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1975 Type *CondTy = SI->getCondition()->getType();
1977 CondTy = VectorType::get(CondTy, VF);
1979 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
1981 case Instruction::ICmp:
1982 case Instruction::FCmp: {
1983 Type *ValTy = I->getOperand(0)->getType();
1984 VectorTy = ToVectorTy(ValTy, VF);
1985 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
1987 case Instruction::Store: {
1988 StoreInst *SI = cast<StoreInst>(I);
1989 Type *ValTy = SI->getValueOperand()->getType();
1990 VectorTy = ToVectorTy(ValTy, VF);
1993 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
1994 SI->getAlignment(), SI->getPointerAddressSpace());
1996 // Scalarized stores.
1997 if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
1999 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2001 // The cost of extracting from the value vector.
2002 Cost += VF * (ExtCost);
2003 // The cost of the scalar stores.
2004 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2005 ValTy->getScalarType(),
2007 SI->getPointerAddressSpace());
2012 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
2013 SI->getPointerAddressSpace());
2015 case Instruction::Load: {
2016 LoadInst *LI = cast<LoadInst>(I);
2019 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
2021 LI->getPointerAddressSpace());
2023 // Scalarized loads.
2024 if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
2026 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
2027 // The cost of inserting the loaded value into the result vector.
2028 Cost += VF * (InCost);
2029 // The cost of the scalar stores.
2030 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2031 RetTy->getScalarType(),
2033 LI->getPointerAddressSpace());
2038 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2039 LI->getPointerAddressSpace());
2041 case Instruction::ZExt:
2042 case Instruction::SExt:
2043 case Instruction::FPToUI:
2044 case Instruction::FPToSI:
2045 case Instruction::FPExt:
2046 case Instruction::PtrToInt:
2047 case Instruction::IntToPtr:
2048 case Instruction::SIToFP:
2049 case Instruction::UIToFP:
2050 case Instruction::Trunc:
2051 case Instruction::FPTrunc:
2052 case Instruction::BitCast: {
2053 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2054 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
2057 // We are scalarizing the instruction. Return the cost of the scalar
2058 // instruction, plus the cost of insert and extract into vector
2059 // elements, times the vector width.
2062 bool IsVoid = RetTy->isVoidTy();
2064 unsigned InsCost = (IsVoid ? 0 :
2065 VTTI->getInstrCost(Instruction::InsertElement,
2068 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2071 // The cost of inserting the results plus extracting each one of the
2073 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
2075 // The cost of executing VF copies of the scalar instruction.
2076 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
2082 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2083 if (Scalar->isVoidTy() || VF == 1)
2085 return VectorType::get(Scalar, VF);
2090 char LoopVectorize::ID = 0;
2091 static const char lv_name[] = "Loop Vectorization";
2092 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2093 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2094 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2095 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2096 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2099 Pass *createLoopVectorizePass() {
2100 return new LoopVectorize();