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.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
44 #define LV_NAME "loop-vectorize"
45 #define DEBUG_TYPE LV_NAME
46 #include "llvm/Constants.h"
47 #include "llvm/DerivedTypes.h"
48 #include "llvm/Instructions.h"
49 #include "llvm/LLVMContext.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Analysis/LoopPass.h"
52 #include "llvm/Value.h"
53 #include "llvm/Function.h"
54 #include "llvm/Analysis/Verifier.h"
55 #include "llvm/Module.h"
56 #include "llvm/Type.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/StringExtras.h"
59 #include "llvm/Analysis/AliasAnalysis.h"
60 #include "llvm/Analysis/AliasSetTracker.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/Dominators.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/LoopInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/Transforms/Scalar.h"
68 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
69 #include "llvm/TargetTransformInfo.h"
70 #include "llvm/Support/CommandLine.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/raw_ostream.h"
73 #include "llvm/DataLayout.h"
74 #include "llvm/Transforms/Utils/Local.h"
78 static cl::opt<unsigned>
79 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
80 cl::desc("Set the default vectorization width. Zero is autoselect."));
82 /// We don't vectorize loops with a known constant trip count below this number.
83 const unsigned TinyTripCountThreshold = 16;
85 /// When performing a runtime memory check, do not check more than this
86 /// number of pointers. Notice that the check is quadratic!
87 const unsigned RuntimeMemoryCheckThreshold = 2;
91 // Forward declarations.
92 class LoopVectorizationLegality;
93 class LoopVectorizationCostModel;
95 /// SingleBlockLoopVectorizer vectorizes loops which contain only one basic
96 /// block to a specified vectorization factor (VF).
97 /// This class performs the widening of scalars into vectors, or multiple
98 /// scalars. This class also implements the following features:
99 /// * It inserts an epilogue loop for handling loops that don't have iteration
100 /// counts that are known to be a multiple of the vectorization factor.
101 /// * It handles the code generation for reduction variables.
102 /// * Scalarization (implementation using scalars) of un-vectorizable
104 /// SingleBlockLoopVectorizer does not perform any vectorization-legality
105 /// checks, and relies on the caller to check for the different legality
106 /// aspects. The SingleBlockLoopVectorizer relies on the
107 /// LoopVectorizationLegality class to provide information about the induction
108 /// and reduction variables that were found to a given vectorization factor.
109 class SingleBlockLoopVectorizer {
112 SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
113 DominatorTree *dt, DataLayout *dl,
116 OrigLoop(Orig), SE(Se), LI(Li), DT(dt), DL(dl), LPM(Lpm), VF(VecWidth),
117 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
119 // Perform the actual loop widening (vectorization).
120 void vectorize(LoopVectorizationLegality *Legal) {
121 // Create a new empty loop. Unlink the old loop and connect the new one.
122 createEmptyLoop(Legal);
123 // Widen each instruction in the old loop to a new one in the new loop.
124 // Use the Legality module to find the induction and reduction variables.
125 vectorizeLoop(Legal);
126 // Register the new loop and update the analysis passes.
131 /// Add code that checks at runtime if the accessed arrays overlap.
132 /// Returns the comperator value or NULL if no check is needed.
133 Value* addRuntimeCheck(LoopVectorizationLegality *Legal,
135 /// Create an empty loop, based on the loop ranges of the old loop.
136 void createEmptyLoop(LoopVectorizationLegality *Legal);
137 /// Copy and widen the instructions from the old loop.
138 void vectorizeLoop(LoopVectorizationLegality *Legal);
139 /// Insert the new loop to the loop hierarchy and pass manager
140 /// and update the analysis passes.
141 void updateAnalysis();
143 /// This instruction is un-vectorizable. Implement it as a sequence
145 void scalarizeInstruction(Instruction *Instr);
147 /// Create a broadcast instruction. This method generates a broadcast
148 /// instruction (shuffle) for loop invariant values and for the induction
149 /// value. If this is the induction variable then we extend it to N, N+1, ...
150 /// this is needed because each iteration in the loop corresponds to a SIMD
152 Value *getBroadcastInstrs(Value *V);
154 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
155 /// for each element in the vector. Starting from zero.
156 Value *getConsecutiveVector(Value* Val);
158 /// When we go over instructions in the basic block we rely on previous
159 /// values within the current basic block or on loop invariant values.
160 /// When we widen (vectorize) values we place them in the map. If the values
161 /// are not within the map, they have to be loop invariant, so we simply
162 /// broadcast them into a vector.
163 Value *getVectorValue(Value *V);
165 /// Get a uniform vector of constant integers. We use this to get
166 /// vectors of ones and zeros for the reduction code.
167 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
169 typedef DenseMap<Value*, Value*> ValueMap;
171 /// The original loop.
173 // Scev analysis to use.
181 // Loop Pass Manager;
183 // The vectorization factor to use.
186 // The builder that we use
189 // --- Vectorization state ---
191 /// The vector-loop preheader.
192 BasicBlock *LoopVectorPreHeader;
193 /// The scalar-loop preheader.
194 BasicBlock *LoopScalarPreHeader;
195 /// Middle Block between the vector and the scalar.
196 BasicBlock *LoopMiddleBlock;
197 ///The ExitBlock of the scalar loop.
198 BasicBlock *LoopExitBlock;
199 ///The vector loop body.
200 BasicBlock *LoopVectorBody;
201 ///The scalar loop body.
202 BasicBlock *LoopScalarBody;
203 ///The first bypass block.
204 BasicBlock *LoopBypassBlock;
206 /// The new Induction variable which was added to the new block.
208 /// The induction variable of the old basic block.
209 PHINode *OldInduction;
210 // Maps scalars to widened vectors.
214 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
215 /// to what vectorization factor.
216 /// This class does not look at the profitability of vectorization, only the
217 /// legality. This class has two main kinds of checks:
218 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
219 /// will change the order of memory accesses in a way that will change the
220 /// correctness of the program.
221 /// * Scalars checks - The code in canVectorizeBlock checks for a number
222 /// of different conditions, such as the availability of a single induction
223 /// variable, that all types are supported and vectorize-able, etc.
224 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
225 /// This class is also used by SingleBlockLoopVectorizer for identifying
226 /// induction variable and the different reduction variables.
227 class LoopVectorizationLegality {
229 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
230 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
232 /// This represents the kinds of reductions that we support.
234 NoReduction, /// Not a reduction.
235 IntegerAdd, /// Sum of numbers.
236 IntegerMult, /// Product of numbers.
237 IntegerOr, /// Bitwise or logical OR of numbers.
238 IntegerAnd, /// Bitwise or logical AND of numbers.
239 IntegerXor /// Bitwise or logical XOR of numbers.
242 /// This POD struct holds information about reduction variables.
243 struct ReductionDescriptor {
245 ReductionDescriptor():
246 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
249 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
250 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
252 // The starting value of the reduction.
253 // It does not have to be zero!
255 // The instruction who's value is used outside the loop.
256 Instruction *LoopExitInstr;
257 // The kind of the reduction.
261 // This POD struct holds information about the memory runtime legality
262 // check that a group of pointers do not overlap.
263 struct RuntimePointerCheck {
264 RuntimePointerCheck(): Need(false) {}
266 /// Reset the state of the pointer runtime information.
274 /// Insert a pointer and calculate the start and end SCEVs.
275 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr) {
276 const SCEV *Sc = SE->getSCEV(Ptr);
277 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
278 assert(AR && "Invalid addrec expression");
279 const SCEV *Ex = SE->getExitCount(Lp, Lp->getHeader());
280 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
281 Pointers.push_back(Ptr);
282 Starts.push_back(AR->getStart());
283 Ends.push_back(ScEnd);
286 /// This flag indicates if we need to add the runtime check.
288 /// Holds the pointers that we need to check.
289 SmallVector<Value*, 2> Pointers;
290 /// Holds the pointer value at the beginning of the loop.
291 SmallVector<const SCEV*, 2> Starts;
292 /// Holds the pointer value at the end of the loop.
293 SmallVector<const SCEV*, 2> Ends;
296 /// ReductionList contains the reduction descriptors for all
297 /// of the reductions that were found in the loop.
298 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
300 /// InductionList saves induction variables and maps them to the initial
301 /// value entring the loop.
302 typedef DenseMap<PHINode*, Value*> InductionList;
304 /// Returns true if it is legal to vectorize this loop.
305 /// This does not mean that it is profitable to vectorize this
306 /// loop, only that it is legal to do so.
309 /// Returns the Induction variable.
310 PHINode *getInduction() {return Induction;}
312 /// Returns the reduction variables found in the loop.
313 ReductionList *getReductionVars() { return &Reductions; }
315 /// Returns the induction variables found in the loop.
316 InductionList *getInductionVars() { return &Inductions; }
318 /// Check if this pointer is consecutive when vectorizing. This happens
319 /// when the last index of the GEP is the induction variable, or that the
320 /// pointer itself is an induction variable.
321 /// This check allows us to vectorize A[idx] into a wide load/store.
322 bool isConsecutivePtr(Value *Ptr);
324 /// Returns true if the value V is uniform within the loop.
325 bool isUniform(Value *V);
327 /// Returns true if this instruction will remain scalar after vectorization.
328 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
330 /// Returns the information that we collected about runtime memory check.
331 RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; }
333 /// Check if a single basic block loop is vectorizable.
334 /// At this point we know that this is a loop with a constant trip count
335 /// and we only need to check individual instructions.
336 bool canVectorizeBlock(BasicBlock &BB);
338 /// When we vectorize loops we may change the order in which
339 /// we read and write from memory. This method checks if it is
340 /// legal to vectorize the code, considering only memory constrains.
341 /// Returns true if BB is vectorizable
342 bool canVectorizeMemory(BasicBlock &BB);
344 /// Returns True, if 'Phi' is the kind of reduction variable for type
345 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
346 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
347 /// Returns true if the instruction I can be a reduction variable of type
349 bool isReductionInstr(Instruction *I, ReductionKind Kind);
350 /// Returns True, if 'Phi' is an induction variable.
351 bool isInductionVariable(PHINode *Phi);
352 /// Return true if can compute the address bounds of Ptr within the loop.
353 bool hasComputableBounds(Value *Ptr);
355 /// The loop that we evaluate.
359 /// DataLayout analysis.
362 // --- vectorization state --- //
364 /// Holds the integer induction variable. This is the counter of the
367 /// Holds the reduction variables.
368 ReductionList Reductions;
369 /// Holds all of the induction variables that we found in the loop.
370 /// Notice that inductions don't need to start at zero and that induction
371 /// variables can be pointers.
372 InductionList Inductions;
374 /// Allowed outside users. This holds the reduction
375 /// vars which can be accessed from outside the loop.
376 SmallPtrSet<Value*, 4> AllowedExit;
377 /// This set holds the variables which are known to be uniform after
379 SmallPtrSet<Instruction*, 4> Uniforms;
380 /// We need to check that all of the pointers in this list are disjoint
382 RuntimePointerCheck PtrRtCheck;
385 /// LoopVectorizationCostModel - estimates the expected speedups due to
387 /// In many cases vectorization is not profitable. This can happen because
388 /// of a number of reasons. In this class we mainly attempt to predict
389 /// the expected speedup/slowdowns due to the supported instruction set.
390 /// We use the VectorTargetTransformInfo to query the different backends
391 /// for the cost of different operations.
392 class LoopVectorizationCostModel {
395 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
396 LoopVectorizationLegality *Leg,
397 const VectorTargetTransformInfo *Vtti):
398 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
400 /// Returns the most profitable vectorization factor for the loop that is
401 /// smaller or equal to the VF argument. This method checks every power
403 unsigned findBestVectorizationFactor(unsigned VF = 8);
406 /// Returns the expected execution cost. The unit of the cost does
407 /// not matter because we use the 'cost' units to compare different
408 /// vector widths. The cost that is returned is *not* normalized by
409 /// the factor width.
410 unsigned expectedCost(unsigned VF);
412 /// Returns the execution time cost of an instruction for a given vector
413 /// width. Vector width of one means scalar.
414 unsigned getInstructionCost(Instruction *I, unsigned VF);
416 /// A helper function for converting Scalar types to vector types.
417 /// If the incoming type is void, we return void. If the VF is 1, we return
419 static Type* ToVectorTy(Type *Scalar, unsigned VF);
421 /// The loop that we evaluate.
426 /// Vectorization legality.
427 LoopVectorizationLegality *Legal;
428 /// Vector target information.
429 const VectorTargetTransformInfo *VTTI;
432 struct LoopVectorize : public LoopPass {
433 static char ID; // Pass identification, replacement for typeid
435 LoopVectorize() : LoopPass(ID) {
436 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
442 TargetTransformInfo *TTI;
445 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
446 // We only vectorize innermost loops.
450 SE = &getAnalysis<ScalarEvolution>();
451 DL = getAnalysisIfAvailable<DataLayout>();
452 LI = &getAnalysis<LoopInfo>();
453 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
454 DT = &getAnalysis<DominatorTree>();
456 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
457 L->getHeader()->getParent()->getName() << "\"\n");
459 // Check if it is legal to vectorize the loop.
460 LoopVectorizationLegality LVL(L, SE, DL);
461 if (!LVL.canVectorize()) {
462 DEBUG(dbgs() << "LV: Not vectorizing.\n");
466 // Select the preffered vectorization factor.
468 if (VectorizationFactor == 0) {
469 const VectorTargetTransformInfo *VTTI = 0;
471 VTTI = TTI->getVectorTargetTransformInfo();
472 // Use the cost model.
473 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
474 VF = CM.findBestVectorizationFactor();
477 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
482 // Use the user command flag.
483 VF = VectorizationFactor;
486 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
487 L->getHeader()->getParent()->getParent()->getModuleIdentifier()<<
490 // If we decided that it is *legal* to vectorizer the loop then do it.
491 SingleBlockLoopVectorizer LB(L, SE, LI, DT, DL, &LPM, VF);
494 DEBUG(verifyFunction(*L->getHeader()->getParent()));
498 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
499 LoopPass::getAnalysisUsage(AU);
500 AU.addRequiredID(LoopSimplifyID);
501 AU.addRequiredID(LCSSAID);
502 AU.addRequired<LoopInfo>();
503 AU.addRequired<ScalarEvolution>();
504 AU.addRequired<DominatorTree>();
505 AU.addPreserved<LoopInfo>();
506 AU.addPreserved<DominatorTree>();
511 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
513 LLVMContext &C = V->getContext();
514 Type *VTy = VectorType::get(V->getType(), VF);
515 Type *I32 = IntegerType::getInt32Ty(C);
516 Constant *Zero = ConstantInt::get(I32, 0);
517 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
518 Value *UndefVal = UndefValue::get(VTy);
519 // Insert the value into a new vector.
520 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
521 // Broadcast the scalar into all locations in the vector.
522 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
524 // We are accessing the induction variable. Make sure to promote the
525 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
527 return getConsecutiveVector(Shuf);
531 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
532 assert(Val->getType()->isVectorTy() && "Must be a vector");
533 assert(Val->getType()->getScalarType()->isIntegerTy() &&
534 "Elem must be an integer");
536 Type *ITy = Val->getType()->getScalarType();
537 VectorType *Ty = cast<VectorType>(Val->getType());
538 unsigned VLen = Ty->getNumElements();
539 SmallVector<Constant*, 8> Indices;
541 // Create a vector of consecutive numbers from zero to VF.
542 for (unsigned i = 0; i < VLen; ++i)
543 Indices.push_back(ConstantInt::get(ITy, i));
545 // Add the consecutive indices to the vector value.
546 Constant *Cv = ConstantVector::get(Indices);
547 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
548 return Builder.CreateAdd(Val, Cv, "induction");
551 bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
552 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
554 // If this pointer is an induction variable, return it.
555 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
556 if (Phi && getInductionVars()->count(Phi))
559 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
563 unsigned NumOperands = Gep->getNumOperands();
564 Value *LastIndex = Gep->getOperand(NumOperands - 1);
566 // Check that all of the gep indices are uniform except for the last.
567 for (unsigned i = 0; i < NumOperands - 1; ++i)
568 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
571 // We can emit wide load/stores only of the last index is the induction
573 const SCEV *Last = SE->getSCEV(LastIndex);
574 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
575 const SCEV *Step = AR->getStepRecurrence(*SE);
577 // The memory is consecutive because the last index is consecutive
578 // and all other indices are loop invariant.
586 bool LoopVectorizationLegality::isUniform(Value *V) {
587 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
590 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
591 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
592 // If we saved a vectorized copy of V, use it.
593 Value *&MapEntry = WidenMap[V];
597 // Broadcast V and save the value for future uses.
598 Value *B = getBroadcastInstrs(V);
604 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
605 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
608 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
609 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
610 // Holds vector parameters or scalars, in case of uniform vals.
611 SmallVector<Value*, 8> Params;
613 // Find all of the vectorized parameters.
614 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
615 Value *SrcOp = Instr->getOperand(op);
617 // If we are accessing the old induction variable, use the new one.
618 if (SrcOp == OldInduction) {
619 Params.push_back(getVectorValue(Induction));
623 // Try using previously calculated values.
624 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
626 // If the src is an instruction that appeared earlier in the basic block
627 // then it should already be vectorized.
628 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
629 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
630 // The parameter is a vector value from earlier.
631 Params.push_back(WidenMap[SrcInst]);
633 // The parameter is a scalar from outside the loop. Maybe even a constant.
634 Params.push_back(SrcOp);
638 assert(Params.size() == Instr->getNumOperands() &&
639 "Invalid number of operands");
641 // Does this instruction return a value ?
642 bool IsVoidRetTy = Instr->getType()->isVoidTy();
643 Value *VecResults = 0;
645 // If we have a return value, create an empty vector. We place the scalarized
646 // instructions in this vector.
648 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
650 // For each scalar that we create:
651 for (unsigned i = 0; i < VF; ++i) {
652 Instruction *Cloned = Instr->clone();
654 Cloned->setName(Instr->getName() + ".cloned");
655 // Replace the operands of the cloned instrucions with extracted scalars.
656 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
657 Value *Op = Params[op];
658 // Param is a vector. Need to extract the right lane.
659 if (Op->getType()->isVectorTy())
660 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
661 Cloned->setOperand(op, Op);
664 // Place the cloned scalar in the new loop.
665 Builder.Insert(Cloned);
667 // If the original scalar returns a value we need to place it in a vector
668 // so that future users will be able to use it.
670 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
671 Builder.getInt32(i));
675 WidenMap[Instr] = VecResults;
679 SingleBlockLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
681 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
682 Legal->getRuntimePointerCheck();
684 if (!PtrRtCheck->Need)
687 Value *MemoryRuntimeCheck = 0;
688 unsigned NumPointers = PtrRtCheck->Pointers.size();
689 SmallVector<Value* , 2> Starts;
690 SmallVector<Value* , 2> Ends;
692 SCEVExpander Exp(*SE, "induction");
694 // Use this type for pointer arithmetic.
695 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
697 for (unsigned i=0; i < NumPointers; ++i) {
698 Value *Ptr = PtrRtCheck->Pointers[i];
699 const SCEV *Sc = SE->getSCEV(Ptr);
701 if (SE->isLoopInvariant(Sc, OrigLoop)) {
702 DEBUG(dbgs() << "LV1: Adding RT check for a loop invariant ptr:" <<
704 Starts.push_back(Ptr);
707 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
709 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i],
711 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
712 Starts.push_back(Start);
717 for (unsigned i = 0; i < NumPointers; ++i) {
718 for (unsigned j = i+1; j < NumPointers; ++j) {
719 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
720 Starts[i], Ends[j], "bound0", Loc);
721 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
722 Starts[j], Ends[i], "bound1", Loc);
723 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
724 "found.conflict", Loc);
725 if (MemoryRuntimeCheck) {
726 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
729 "conflict.rdx", Loc);
731 MemoryRuntimeCheck = IsConflict;
736 return MemoryRuntimeCheck;
740 SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
742 In this function we generate a new loop. The new loop will contain
743 the vectorized instructions while the old loop will continue to run the
746 [ ] <-- vector loop bypass.
749 | [ ] <-- vector pre header.
753 | [ ]_| <-- vector loop.
756 >[ ] <--- middle-block.
759 | [ ] <--- new preheader.
763 | [ ]_| <-- old scalar loop to handle remainder.
770 // Some loops have a single integer induction variable, while other loops
771 // don't. One example is c++ iterators that often have multiple pointer
772 // induction variables. In the code below we also support a case where we
773 // don't have a single induction variable.
774 OldInduction = Legal->getInduction();
775 Type *IdxTy = OldInduction ? OldInduction->getType() :
776 DL->getIntPtrType(SE->getContext());
778 // Find the loop boundaries.
779 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
780 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
782 // Get the total trip count from the count by adding 1.
783 ExitCount = SE->getAddExpr(ExitCount,
784 SE->getConstant(ExitCount->getType(), 1));
786 // This is the original scalar-loop preheader.
787 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
788 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
789 assert(ExitBlock && "Must have an exit block");
791 // The loop index does not have to start at Zero. Find the original start
792 // value from the induction PHI node. If we don't have an induction variable
793 // then we know that it starts at zero.
794 Value *StartIdx = OldInduction ?
795 OldInduction->getIncomingValueForBlock(BypassBlock):
796 ConstantInt::get(IdxTy, 0);
798 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
799 assert(BypassBlock && "Invalid loop structure");
801 BasicBlock *VectorPH =
802 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
803 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
806 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
808 BasicBlock *ScalarPH =
809 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
811 // Find the induction variable.
812 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
814 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
816 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
818 // Generate the induction variable.
819 Induction = Builder.CreatePHI(IdxTy, 2, "index");
820 Constant *Step = ConstantInt::get(IdxTy, VF);
822 // Expand the trip count and place the new instructions in the preheader.
823 // Notice that the pre-header does not change, only the loop body.
824 SCEVExpander Exp(*SE, "induction");
825 Instruction *Loc = BypassBlock->getTerminator();
827 // Count holds the overall loop count (N).
828 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), Loc);
830 // We may need to extend the index in case there is a type mismatch.
831 // We know that the count starts at zero and does not overflow.
832 if (Count->getType() != IdxTy) {
833 // The exit count can be of pointer type. Convert it to the correct
835 if (ExitCount->getType()->isPointerTy())
836 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
838 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
841 // Add the start index to the loop count to get the new end index.
842 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
844 // Now we need to generate the expression for N - (N % VF), which is
845 // the part that the vectorized body will execute.
846 Constant *CIVF = ConstantInt::get(IdxTy, VF);
847 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
848 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
849 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
850 "end.idx.rnd.down", Loc);
852 // Now, compare the new count to zero. If it is zero skip the vector loop and
853 // jump to the scalar loop.
854 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
859 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal, Loc);
861 // If we are using memory runtime checks, include them in.
862 if (MemoryRuntimeCheck) {
863 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
867 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
868 // Remove the old terminator.
869 Loc->eraseFromParent();
871 // We are going to resume the execution of the scalar loop.
872 // Go over all of the induction variables that we found and fix the
873 // PHIs that are left in the scalar version of the loop.
874 // The starting values of PHI nodes depend on the counter of the last
875 // iteration in the vectorized loop.
876 // If we come from a bypass edge then we need to start from the original start
879 // This variable saves the new starting index for the scalar loop.
880 PHINode *ResumeIndex = 0;
881 LoopVectorizationLegality::InductionList::iterator I, E;
882 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
883 for (I = List->begin(), E = List->end(); I != E; ++I) {
884 PHINode *OrigPhi = I->first;
885 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
886 MiddleBlock->getTerminator());
888 if (OrigPhi->getType()->isIntegerTy()) {
889 // Handle the integer induction counter:
890 assert(OrigPhi == OldInduction && "Unknown integer PHI");
891 // We know what the end value is.
892 EndValue = IdxEndRoundDown;
893 // We also know which PHI node holds it.
894 ResumeIndex = ResumeVal;
896 // For pointer induction variables, calculate the offset using
898 EndValue = GetElementPtrInst::Create(I->second, CountRoundDown,
900 BypassBlock->getTerminator());
903 // The new PHI merges the original incoming value, in case of a bypass,
904 // or the value at the end of the vectorized loop.
905 ResumeVal->addIncoming(I->second, BypassBlock);
906 ResumeVal->addIncoming(EndValue, VecBody);
908 // Fix the scalar body counter (PHI node).
909 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
910 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
913 // If we are generating a new induction variable then we also need to
914 // generate the code that calculates the exit value. This value is not
915 // simply the end of the counter because we may skip the vectorized body
916 // in case of a runtime check.
918 assert(!ResumeIndex && "Unexpected resume value found");
919 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
920 MiddleBlock->getTerminator());
921 ResumeIndex->addIncoming(StartIdx, BypassBlock);
922 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
925 // Make sure that we found the index where scalar loop needs to continue.
926 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
927 "Invalid resume Index");
929 // Add a check in the middle block to see if we have completed
930 // all of the iterations in the first vector loop.
931 // If (N - N%VF) == N, then we *don't* need to run the remainder.
932 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
933 ResumeIndex, "cmp.n",
934 MiddleBlock->getTerminator());
936 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
937 // Remove the old terminator.
938 MiddleBlock->getTerminator()->eraseFromParent();
940 // Create i+1 and fill the PHINode.
941 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
942 Induction->addIncoming(StartIdx, VectorPH);
943 Induction->addIncoming(NextIdx, VecBody);
944 // Create the compare.
945 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
946 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
948 // Now we have two terminators. Remove the old one from the block.
949 VecBody->getTerminator()->eraseFromParent();
951 // Get ready to start creating new instructions into the vectorized body.
952 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
954 // Register the new loop.
955 Loop* Lp = new Loop();
956 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
958 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
960 Loop *ParentLoop = OrigLoop->getParentLoop();
962 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
963 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
964 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
968 LoopVectorPreHeader = VectorPH;
969 LoopScalarPreHeader = ScalarPH;
970 LoopMiddleBlock = MiddleBlock;
971 LoopExitBlock = ExitBlock;
972 LoopVectorBody = VecBody;
973 LoopScalarBody = OldBasicBlock;
974 LoopBypassBlock = BypassBlock;
977 /// This function returns the identity element (or neutral element) for
980 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
982 case LoopVectorizationLegality::IntegerXor:
983 case LoopVectorizationLegality::IntegerAdd:
984 case LoopVectorizationLegality::IntegerOr:
985 // Adding, Xoring, Oring zero to a number does not change it.
987 case LoopVectorizationLegality::IntegerMult:
988 // Multiplying a number by 1 does not change it.
990 case LoopVectorizationLegality::IntegerAnd:
991 // AND-ing a number with an all-1 value does not change it.
994 llvm_unreachable("Unknown reduction kind");
999 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1000 //===------------------------------------------------===//
1002 // Notice: any optimization or new instruction that go
1003 // into the code below should be also be implemented in
1006 //===------------------------------------------------===//
1007 typedef SmallVector<PHINode*, 4> PhiVector;
1008 BasicBlock &BB = *OrigLoop->getHeader();
1009 Constant *Zero = ConstantInt::get(
1010 IntegerType::getInt32Ty(BB.getContext()), 0);
1012 // In order to support reduction variables we need to be able to vectorize
1013 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1014 // steages. First, we create a new vector PHI node with no incoming edges.
1015 // We use this value when we vectorize all of the instructions that use the
1016 // PHI. Next, after all of the instructions in the block are complete we
1017 // add the new incoming edges to the PHI. At this point all of the
1018 // instructions in the basic block are vectorized, so we can use them to
1019 // construct the PHI.
1020 PhiVector RdxPHIsToFix;
1022 // For each instruction in the old loop.
1023 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1024 Instruction *Inst = it;
1026 switch (Inst->getOpcode()) {
1027 case Instruction::Br:
1028 // Nothing to do for PHIs and BR, since we already took care of the
1029 // loop control flow instructions.
1031 case Instruction::PHI:{
1032 PHINode* P = cast<PHINode>(Inst);
1033 // Handle reduction variables:
1034 if (Legal->getReductionVars()->count(P)) {
1035 // This is phase one of vectorizing PHIs.
1036 Type *VecTy = VectorType::get(Inst->getType(), VF);
1037 WidenMap[Inst] = PHINode::Create(VecTy, 2, "vec.phi",
1038 LoopVectorBody->getFirstInsertionPt());
1039 RdxPHIsToFix.push_back(P);
1043 // This PHINode must be an induction variable.
1044 // Make sure that we know about it.
1045 assert(Legal->getInductionVars()->count(P) &&
1046 "Not an induction variable");
1048 if (P->getType()->isIntegerTy()) {
1049 assert(P == OldInduction && "Unexpected PHI");
1050 WidenMap[Inst] = getBroadcastInstrs(Induction);
1054 // Handle pointer inductions:
1055 assert(P->getType()->isPointerTy() && "Unexpected type.");
1056 Value *StartIdx = OldInduction ?
1057 Legal->getInductionVars()->lookup(OldInduction) :
1058 ConstantInt::get(Induction->getType(), 0);
1060 // This is the pointer value coming into the loop.
1061 Value *StartPtr = Legal->getInductionVars()->lookup(P);
1063 // This is the normalized GEP that starts counting at zero.
1064 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1067 // This is the vector of results. Notice that we don't generate vector
1068 // geps because scalar geps result in better code.
1069 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1070 for (unsigned int i = 0; i < VF; ++i) {
1071 Constant *Idx = ConstantInt::get(Induction->getType(), i);
1072 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1073 Value *SclrGep = Builder.CreateGEP(StartPtr, GlobalIdx, "next.gep");
1074 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1075 Builder.getInt32(i),
1079 WidenMap[Inst] = VecVal;
1082 case Instruction::Add:
1083 case Instruction::FAdd:
1084 case Instruction::Sub:
1085 case Instruction::FSub:
1086 case Instruction::Mul:
1087 case Instruction::FMul:
1088 case Instruction::UDiv:
1089 case Instruction::SDiv:
1090 case Instruction::FDiv:
1091 case Instruction::URem:
1092 case Instruction::SRem:
1093 case Instruction::FRem:
1094 case Instruction::Shl:
1095 case Instruction::LShr:
1096 case Instruction::AShr:
1097 case Instruction::And:
1098 case Instruction::Or:
1099 case Instruction::Xor: {
1100 // Just widen binops.
1101 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1102 Value *A = getVectorValue(Inst->getOperand(0));
1103 Value *B = getVectorValue(Inst->getOperand(1));
1105 // Use this vector value for all users of the original instruction.
1106 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1109 // Update the NSW, NUW and Exact flags.
1110 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1111 if (isa<OverflowingBinaryOperator>(BinOp)) {
1112 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1113 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1115 if (isa<PossiblyExactOperator>(VecOp))
1116 VecOp->setIsExact(BinOp->isExact());
1119 case Instruction::Select: {
1121 // If the selector is loop invariant we can create a select
1122 // instruction with a scalar condition. Otherwise, use vector-select.
1123 Value *Cond = Inst->getOperand(0);
1124 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1126 // The condition can be loop invariant but still defined inside the
1127 // loop. This means that we can't just use the original 'cond' value.
1128 // We have to take the 'vectorized' value and pick the first lane.
1129 // Instcombine will make this a no-op.
1130 Cond = getVectorValue(Cond);
1132 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1134 Value *Op0 = getVectorValue(Inst->getOperand(1));
1135 Value *Op1 = getVectorValue(Inst->getOperand(2));
1136 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
1140 case Instruction::ICmp:
1141 case Instruction::FCmp: {
1142 // Widen compares. Generate vector compares.
1143 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
1144 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
1145 Value *A = getVectorValue(Inst->getOperand(0));
1146 Value *B = getVectorValue(Inst->getOperand(1));
1148 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1150 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1154 case Instruction::Store: {
1155 // Attempt to issue a wide store.
1156 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1157 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1158 Value *Ptr = SI->getPointerOperand();
1159 unsigned Alignment = SI->getAlignment();
1161 assert(!Legal->isUniform(Ptr) &&
1162 "We do not allow storing to uniform addresses");
1164 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1166 // This store does not use GEPs.
1167 if (!Legal->isConsecutivePtr(Ptr)) {
1168 scalarizeInstruction(Inst);
1173 // The last index does not have to be the induction. It can be
1174 // consecutive and be a function of the index. For example A[I+1];
1175 unsigned NumOperands = Gep->getNumOperands();
1176 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1177 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1179 // Create the new GEP with the new induction variable.
1180 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1181 Gep2->setOperand(NumOperands - 1, LastIndex);
1182 Ptr = Builder.Insert(Gep2);
1184 // Use the induction element ptr.
1185 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1186 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1188 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1189 Value *Val = getVectorValue(SI->getValueOperand());
1190 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1193 case Instruction::Load: {
1194 // Attempt to issue a wide load.
1195 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1196 Type *RetTy = VectorType::get(LI->getType(), VF);
1197 Value *Ptr = LI->getPointerOperand();
1198 unsigned Alignment = LI->getAlignment();
1199 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1201 // If the pointer is loop invariant or if it is non consecutive,
1202 // scalarize the load.
1203 bool Con = Legal->isConsecutivePtr(Ptr);
1204 if (Legal->isUniform(Ptr) || !Con) {
1205 scalarizeInstruction(Inst);
1210 // The last index does not have to be the induction. It can be
1211 // consecutive and be a function of the index. For example A[I+1];
1212 unsigned NumOperands = Gep->getNumOperands();
1213 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1214 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1216 // Create the new GEP with the new induction variable.
1217 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1218 Gep2->setOperand(NumOperands - 1, LastIndex);
1219 Ptr = Builder.Insert(Gep2);
1221 // Use the induction element ptr.
1222 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1223 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1226 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1227 LI = Builder.CreateLoad(Ptr);
1228 LI->setAlignment(Alignment);
1229 // Use this vector value for all users of the load.
1230 WidenMap[Inst] = LI;
1233 case Instruction::ZExt:
1234 case Instruction::SExt:
1235 case Instruction::FPToUI:
1236 case Instruction::FPToSI:
1237 case Instruction::FPExt:
1238 case Instruction::PtrToInt:
1239 case Instruction::IntToPtr:
1240 case Instruction::SIToFP:
1241 case Instruction::UIToFP:
1242 case Instruction::Trunc:
1243 case Instruction::FPTrunc:
1244 case Instruction::BitCast: {
1245 /// Vectorize bitcasts.
1246 CastInst *CI = dyn_cast<CastInst>(Inst);
1247 Value *A = getVectorValue(Inst->getOperand(0));
1248 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1249 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1254 /// All other instructions are unsupported. Scalarize them.
1255 scalarizeInstruction(Inst);
1258 }// end of for_each instr.
1260 // At this point every instruction in the original loop is widended to
1261 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1262 // that we vectorized. The PHI nodes are currently empty because we did
1263 // not want to introduce cycles. Notice that the remaining PHI nodes
1264 // that we need to fix are reduction variables.
1266 // Create the 'reduced' values for each of the induction vars.
1267 // The reduced values are the vector values that we scalarize and combine
1268 // after the loop is finished.
1269 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1271 PHINode *RdxPhi = *it;
1272 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1273 assert(RdxPhi && "Unable to recover vectorized PHI");
1275 // Find the reduction variable descriptor.
1276 assert(Legal->getReductionVars()->count(RdxPhi) &&
1277 "Unable to find the reduction variable");
1278 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1279 (*Legal->getReductionVars())[RdxPhi];
1281 // We need to generate a reduction vector from the incoming scalar.
1282 // To do so, we need to generate the 'identity' vector and overide
1283 // one of the elements with the incoming scalar reduction. We need
1284 // to do it in the vector-loop preheader.
1285 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1287 // This is the vector-clone of the value that leaves the loop.
1288 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1289 Type *VecTy = VectorExit->getType();
1291 // Find the reduction identity variable. Zero for addition, or, xor,
1292 // one for multiplication, -1 for And.
1293 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1294 VecTy->getScalarType());
1296 // This vector is the Identity vector where the first element is the
1297 // incoming scalar reduction.
1298 Value *VectorStart = Builder.CreateInsertElement(Identity,
1299 RdxDesc.StartValue, Zero);
1301 // Fix the vector-loop phi.
1302 // We created the induction variable so we know that the
1303 // preheader is the first entry.
1304 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1306 // Reductions do not have to start at zero. They can start with
1307 // any loop invariant values.
1308 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1309 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1310 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1311 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1313 // Before each round, move the insertion point right between
1314 // the PHIs and the values we are going to write.
1315 // This allows us to write both PHINodes and the extractelement
1317 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1319 // This PHINode contains the vectorized reduction variable, or
1320 // the initial value vector, if we bypass the vector loop.
1321 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1322 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1323 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1325 // Extract the first scalar.
1327 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1328 // Extract and reduce the remaining vector elements.
1329 for (unsigned i=1; i < VF; ++i) {
1331 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1332 switch (RdxDesc.Kind) {
1333 case LoopVectorizationLegality::IntegerAdd:
1334 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1336 case LoopVectorizationLegality::IntegerMult:
1337 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1339 case LoopVectorizationLegality::IntegerOr:
1340 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1342 case LoopVectorizationLegality::IntegerAnd:
1343 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1345 case LoopVectorizationLegality::IntegerXor:
1346 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1349 llvm_unreachable("Unknown reduction operation");
1353 // Now, we need to fix the users of the reduction variable
1354 // inside and outside of the scalar remainder loop.
1355 // We know that the loop is in LCSSA form. We need to update the
1356 // PHI nodes in the exit blocks.
1357 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1358 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1359 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1360 if (!LCSSAPhi) continue;
1362 // All PHINodes need to have a single entry edge, or two if
1363 // we already fixed them.
1364 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1366 // We found our reduction value exit-PHI. Update it with the
1367 // incoming bypass edge.
1368 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1369 // Add an edge coming from the bypass.
1370 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1373 }// end of the LCSSA phi scan.
1375 // Fix the scalar loop reduction variable with the incoming reduction sum
1376 // from the vector body and from the backedge value.
1377 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1378 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1379 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1380 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1381 }// end of for each redux variable.
1384 void SingleBlockLoopVectorizer::updateAnalysis() {
1385 // The original basic block.
1386 SE->forgetLoop(OrigLoop);
1388 // Update the dominator tree information.
1389 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1390 "Entry does not dominate exit.");
1392 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1393 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1394 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1395 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1396 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1397 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1399 DEBUG(DT->verifyAnalysis());
1402 bool LoopVectorizationLegality::canVectorize() {
1403 if (!TheLoop->getLoopPreheader()) {
1404 assert(false && "No preheader!!");
1405 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1409 // We can only vectorize single basic block loops.
1410 unsigned NumBlocks = TheLoop->getNumBlocks();
1411 if (NumBlocks != 1) {
1412 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1416 // We need to have a loop header.
1417 BasicBlock *BB = TheLoop->getHeader();
1418 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1420 // ScalarEvolution needs to be able to find the exit count.
1421 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1422 if (ExitCount == SE->getCouldNotCompute()) {
1423 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1427 // Do not loop-vectorize loops with a tiny trip count.
1428 unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB);
1429 if (TC > 0u && TC < TinyTripCountThreshold) {
1430 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1431 "This loop is not worth vectorizing.\n");
1435 // Go over each instruction and look at memory deps.
1436 if (!canVectorizeBlock(*BB)) {
1437 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1441 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1442 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1445 // Okay! We can vectorize. At this point we don't have any other mem analysis
1446 // which may limit our maximum vectorization factor, so just return true with
1451 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1453 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1455 // Scan the instructions in the block and look for hazards.
1456 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1457 Instruction *I = it;
1459 if (PHINode *Phi = dyn_cast<PHINode>(I)) {
1460 // This should not happen because the loop should be normalized.
1461 if (Phi->getNumIncomingValues() != 2) {
1462 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1466 // This is the value coming from the preheader.
1467 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1469 // We only look at integer and pointer phi nodes.
1470 if (Phi->getType()->isPointerTy() && isInductionVariable(Phi)) {
1471 DEBUG(dbgs() << "LV: Found a pointer induction variable.\n");
1472 Inductions[Phi] = StartValue;
1474 } else if (!Phi->getType()->isIntegerTy()) {
1475 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1479 // Handle integer PHIs:
1480 if (isInductionVariable(Phi)) {
1482 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1485 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1487 Inductions[Phi] = StartValue;
1490 if (AddReductionVar(Phi, IntegerAdd)) {
1491 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1494 if (AddReductionVar(Phi, IntegerMult)) {
1495 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1498 if (AddReductionVar(Phi, IntegerOr)) {
1499 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1502 if (AddReductionVar(Phi, IntegerAnd)) {
1503 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1506 if (AddReductionVar(Phi, IntegerXor)) {
1507 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1511 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1513 }// end of PHI handling
1515 // We still don't handle functions.
1516 CallInst *CI = dyn_cast<CallInst>(I);
1518 DEBUG(dbgs() << "LV: Found a call site.\n");
1522 // We do not re-vectorize vectors.
1523 if (!VectorType::isValidElementType(I->getType()) &&
1524 !I->getType()->isVoidTy()) {
1525 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1529 // Reduction instructions are allowed to have exit users.
1530 // All other instructions must not have external users.
1531 if (!AllowedExit.count(I))
1532 //Check that all of the users of the loop are inside the BB.
1533 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1535 Instruction *U = cast<Instruction>(*it);
1536 // This user may be a reduction exit value.
1537 BasicBlock *Parent = U->getParent();
1538 if (Parent != &BB) {
1539 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1546 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
1547 assert(getInductionVars()->size() && "No induction variables");
1550 // Don't vectorize if the memory dependencies do not allow vectorization.
1551 if (!canVectorizeMemory(BB))
1554 // We now know that the loop is vectorizable!
1555 // Collect variables that will remain uniform after vectorization.
1556 std::vector<Value*> Worklist;
1558 // Start with the conditional branch and walk up the block.
1559 Worklist.push_back(BB.getTerminator()->getOperand(0));
1561 while (Worklist.size()) {
1562 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1563 Worklist.pop_back();
1565 // Look at instructions inside this block. Stop when reaching PHI nodes.
1566 if (!I || I->getParent() != &BB || isa<PHINode>(I))
1569 // This is a known uniform.
1572 // Insert all operands.
1573 for (int i=0, Op = I->getNumOperands(); i < Op; ++i) {
1574 Worklist.push_back(I->getOperand(i));
1581 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1582 typedef SmallVector<Value*, 16> ValueVector;
1583 typedef SmallPtrSet<Value*, 16> ValueSet;
1584 // Holds the Load and Store *instructions*.
1587 PtrRtCheck.Pointers.clear();
1588 PtrRtCheck.Need = false;
1590 // Scan the BB and collect legal loads and stores.
1591 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1592 Instruction *I = it;
1594 // If this is a load, save it. If this instruction can read from memory
1595 // but is not a load, then we quit. Notice that we don't handle function
1596 // calls that read or write.
1597 if (I->mayReadFromMemory()) {
1598 LoadInst *Ld = dyn_cast<LoadInst>(I);
1599 if (!Ld) return false;
1600 if (!Ld->isSimple()) {
1601 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1604 Loads.push_back(Ld);
1608 // Save store instructions. Abort if other instructions write to memory.
1609 if (I->mayWriteToMemory()) {
1610 StoreInst *St = dyn_cast<StoreInst>(I);
1611 if (!St) return false;
1612 if (!St->isSimple()) {
1613 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1616 Stores.push_back(St);
1620 // Now we have two lists that hold the loads and the stores.
1621 // Next, we find the pointers that they use.
1623 // Check if we see any stores. If there are no stores, then we don't
1624 // care if the pointers are *restrict*.
1625 if (!Stores.size()) {
1626 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1630 // Holds the read and read-write *pointers* that we find.
1632 ValueVector ReadWrites;
1634 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1635 // multiple times on the same object. If the ptr is accessed twice, once
1636 // for read and once for write, it will only appear once (on the write
1637 // list). This is okay, since we are going to check for conflicts between
1638 // writes and between reads and writes, but not between reads and reads.
1641 ValueVector::iterator I, IE;
1642 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1643 StoreInst *ST = dyn_cast<StoreInst>(*I);
1644 assert(ST && "Bad StoreInst");
1645 Value* Ptr = ST->getPointerOperand();
1647 if (isUniform(Ptr)) {
1648 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1652 // If we did *not* see this pointer before, insert it to
1653 // the read-write list. At this phase it is only a 'write' list.
1654 if (Seen.insert(Ptr))
1655 ReadWrites.push_back(Ptr);
1658 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1659 LoadInst *LD = dyn_cast<LoadInst>(*I);
1660 assert(LD && "Bad LoadInst");
1661 Value* Ptr = LD->getPointerOperand();
1662 // If we did *not* see this pointer before, insert it to the
1663 // read list. If we *did* see it before, then it is already in
1664 // the read-write list. This allows us to vectorize expressions
1665 // such as A[i] += x; Because the address of A[i] is a read-write
1666 // pointer. This only works if the index of A[i] is consecutive.
1667 // If the address of i is unknown (for example A[B[i]]) then we may
1668 // read a few words, modify, and write a few words, and some of the
1669 // words may be written to the same address.
1670 if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr))
1671 Reads.push_back(Ptr);
1674 // If we write (or read-write) to a single destination and there are no
1675 // other reads in this loop then is it safe to vectorize.
1676 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1677 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1681 // Find pointers with computable bounds. We are going to use this information
1682 // to place a runtime bound check.
1684 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1685 if (hasComputableBounds(*I)) {
1686 PtrRtCheck.insert(SE, TheLoop, *I);
1687 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1692 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1693 if (hasComputableBounds(*I)) {
1694 PtrRtCheck.insert(SE, TheLoop, *I);
1695 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1701 // Check that we did not collect too many pointers or found a
1702 // unsizeable pointer.
1703 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1708 PtrRtCheck.Need = RT;
1711 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1714 // Now that the pointers are in two lists (Reads and ReadWrites), we
1715 // can check that there are no conflicts between each of the writes and
1716 // between the writes to the reads.
1717 ValueSet WriteObjects;
1718 ValueVector TempObjects;
1720 // Check that the read-writes do not conflict with other read-write
1722 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1723 GetUnderlyingObjects(*I, TempObjects, DL);
1724 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1726 if (!isIdentifiedObject(*it)) {
1727 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1730 if (!WriteObjects.insert(*it)) {
1731 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1736 TempObjects.clear();
1739 /// Check that the reads don't conflict with the read-writes.
1740 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1741 GetUnderlyingObjects(*I, TempObjects, DL);
1742 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1744 if (!isIdentifiedObject(*it)) {
1745 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1748 if (WriteObjects.count(*it)) {
1749 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1754 TempObjects.clear();
1757 // It is safe to vectorize and we don't need any runtime checks.
1758 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1763 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1764 ReductionKind Kind) {
1765 if (Phi->getNumIncomingValues() != 2)
1768 // Find the possible incoming reduction variable.
1769 BasicBlock *BB = Phi->getParent();
1770 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1771 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1772 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1774 // ExitInstruction is the single value which is used outside the loop.
1775 // We only allow for a single reduction value to be used outside the loop.
1776 // This includes users of the reduction, variables (which form a cycle
1777 // which ends in the phi node).
1778 Instruction *ExitInstruction = 0;
1780 // Iter is our iterator. We start with the PHI node and scan for all of the
1781 // users of this instruction. All users must be instructions which can be
1782 // used as reduction variables (such as ADD). We may have a single
1783 // out-of-block user. They cycle must end with the original PHI.
1784 // Also, we can't have multiple block-local users.
1785 Instruction *Iter = Phi;
1787 // Any reduction instr must be of one of the allowed kinds.
1788 if (!isReductionInstr(Iter, Kind))
1791 // Did we found a user inside this block ?
1792 bool FoundInBlockUser = false;
1793 // Did we reach the initial PHI node ?
1794 bool FoundStartPHI = false;
1796 // If the instruction has no users then this is a broken
1797 // chain and can't be a reduction variable.
1798 if (Iter->use_empty())
1801 // For each of the *users* of iter.
1802 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1804 Instruction *U = cast<Instruction>(*it);
1805 // We already know that the PHI is a user.
1807 FoundStartPHI = true;
1810 // Check if we found the exit user.
1811 BasicBlock *Parent = U->getParent();
1813 // We must have a single exit instruction.
1814 if (ExitInstruction != 0)
1816 ExitInstruction = Iter;
1818 // We can't have multiple inside users.
1819 if (FoundInBlockUser)
1821 FoundInBlockUser = true;
1825 // We found a reduction var if we have reached the original
1826 // phi node and we only have a single instruction with out-of-loop
1828 if (FoundStartPHI && ExitInstruction) {
1829 // This instruction is allowed to have out-of-loop users.
1830 AllowedExit.insert(ExitInstruction);
1832 // Save the description of this reduction variable.
1833 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1834 Reductions[Phi] = RD;
1841 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1842 ReductionKind Kind) {
1843 switch (I->getOpcode()) {
1846 case Instruction::PHI:
1849 case Instruction::Add:
1850 case Instruction::Sub:
1851 return Kind == IntegerAdd;
1852 case Instruction::Mul:
1853 return Kind == IntegerMult;
1854 case Instruction::And:
1855 return Kind == IntegerAnd;
1856 case Instruction::Or:
1857 return Kind == IntegerOr;
1858 case Instruction::Xor:
1859 return Kind == IntegerXor;
1863 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1864 Type *PhiTy = Phi->getType();
1865 // We only handle integer and pointer inductions variables.
1866 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1869 // Check that the PHI is consecutive and starts at zero.
1870 const SCEV *PhiScev = SE->getSCEV(Phi);
1871 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1873 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1876 const SCEV *Step = AR->getStepRecurrence(*SE);
1878 // Integer inductions need to have a stride of one.
1879 if (PhiTy->isIntegerTy())
1880 return Step->isOne();
1882 // Calculate the pointer stride and check if it is consecutive.
1883 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1884 if (!C) return false;
1886 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1887 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1888 return (C->getValue()->equalsInt(Size));
1891 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1892 const SCEV *PhiScev = SE->getSCEV(Ptr);
1893 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1897 return AR->isAffine();
1901 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1903 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1907 float Cost = expectedCost(1);
1909 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1910 for (unsigned i=2; i <= VF; i*=2) {
1911 // Notice that the vector loop needs to be executed less times, so
1912 // we need to divide the cost of the vector loops by the width of
1913 // the vector elements.
1914 float VectorCost = expectedCost(i) / (float)i;
1915 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1916 (int)VectorCost << ".\n");
1917 if (VectorCost < Cost) {
1923 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1927 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1928 // We can only estimate the cost of single basic block loops.
1929 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1931 BasicBlock *BB = TheLoop->getHeader();
1934 // For each instruction in the old loop.
1935 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1936 Instruction *Inst = it;
1937 unsigned C = getInstructionCost(Inst, VF);
1939 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1940 " For instruction: "<< *Inst << "\n");
1947 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1948 assert(VTTI && "Invalid vector target transformation info");
1950 // If we know that this instruction will remain uniform, check the cost of
1951 // the scalar version.
1952 if (Legal->isUniformAfterVectorization(I))
1955 Type *RetTy = I->getType();
1956 Type *VectorTy = ToVectorTy(RetTy, VF);
1959 // TODO: We need to estimate the cost of intrinsic calls.
1960 switch (I->getOpcode()) {
1961 case Instruction::GetElementPtr:
1962 // We mark this instruction as zero-cost because scalar GEPs are usually
1963 // lowered to the intruction addressing mode. At the moment we don't
1964 // generate vector geps.
1966 case Instruction::Br: {
1967 return VTTI->getCFInstrCost(I->getOpcode());
1969 case Instruction::PHI:
1971 case Instruction::Add:
1972 case Instruction::FAdd:
1973 case Instruction::Sub:
1974 case Instruction::FSub:
1975 case Instruction::Mul:
1976 case Instruction::FMul:
1977 case Instruction::UDiv:
1978 case Instruction::SDiv:
1979 case Instruction::FDiv:
1980 case Instruction::URem:
1981 case Instruction::SRem:
1982 case Instruction::FRem:
1983 case Instruction::Shl:
1984 case Instruction::LShr:
1985 case Instruction::AShr:
1986 case Instruction::And:
1987 case Instruction::Or:
1988 case Instruction::Xor: {
1989 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
1991 case Instruction::Select: {
1992 SelectInst *SI = cast<SelectInst>(I);
1993 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1994 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1995 Type *CondTy = SI->getCondition()->getType();
1997 CondTy = VectorType::get(CondTy, VF);
1999 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
2001 case Instruction::ICmp:
2002 case Instruction::FCmp: {
2003 Type *ValTy = I->getOperand(0)->getType();
2004 VectorTy = ToVectorTy(ValTy, VF);
2005 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
2007 case Instruction::Store: {
2008 StoreInst *SI = cast<StoreInst>(I);
2009 Type *ValTy = SI->getValueOperand()->getType();
2010 VectorTy = ToVectorTy(ValTy, VF);
2013 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
2014 SI->getAlignment(), SI->getPointerAddressSpace());
2016 // Scalarized stores.
2017 if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
2019 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2021 // The cost of extracting from the value vector.
2022 Cost += VF * (ExtCost);
2023 // The cost of the scalar stores.
2024 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2025 ValTy->getScalarType(),
2027 SI->getPointerAddressSpace());
2032 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
2033 SI->getPointerAddressSpace());
2035 case Instruction::Load: {
2036 LoadInst *LI = cast<LoadInst>(I);
2039 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
2041 LI->getPointerAddressSpace());
2043 // Scalarized loads.
2044 if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
2046 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
2047 // The cost of inserting the loaded value into the result vector.
2048 Cost += VF * (InCost);
2049 // The cost of the scalar stores.
2050 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2051 RetTy->getScalarType(),
2053 LI->getPointerAddressSpace());
2058 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2059 LI->getPointerAddressSpace());
2061 case Instruction::ZExt:
2062 case Instruction::SExt:
2063 case Instruction::FPToUI:
2064 case Instruction::FPToSI:
2065 case Instruction::FPExt:
2066 case Instruction::PtrToInt:
2067 case Instruction::IntToPtr:
2068 case Instruction::SIToFP:
2069 case Instruction::UIToFP:
2070 case Instruction::Trunc:
2071 case Instruction::FPTrunc:
2072 case Instruction::BitCast: {
2073 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2074 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
2077 // We are scalarizing the instruction. Return the cost of the scalar
2078 // instruction, plus the cost of insert and extract into vector
2079 // elements, times the vector width.
2082 bool IsVoid = RetTy->isVoidTy();
2084 unsigned InsCost = (IsVoid ? 0 :
2085 VTTI->getInstrCost(Instruction::InsertElement,
2088 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2091 // The cost of inserting the results plus extracting each one of the
2093 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
2095 // The cost of executing VF copies of the scalar instruction.
2096 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
2102 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2103 if (Scalar->isVoidTy() || VF == 1)
2105 return VectorType::get(Scalar, VF);
2110 char LoopVectorize::ID = 0;
2111 static const char lv_name[] = "Loop Vectorization";
2112 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2113 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2114 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2115 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2116 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2119 Pass *createLoopVectorizePass() {
2120 return new LoopVectorize();