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;
89 /// This is the highest vector width that we try to generate.
90 const unsigned MaxVectorSize = 8;
94 // Forward declarations.
95 class LoopVectorizationLegality;
96 class LoopVectorizationCostModel;
98 /// SingleBlockLoopVectorizer vectorizes loops which contain only one basic
99 /// block to a specified vectorization factor (VF).
100 /// This class performs the widening of scalars into vectors, or multiple
101 /// scalars. This class also implements the following features:
102 /// * It inserts an epilogue loop for handling loops that don't have iteration
103 /// counts that are known to be a multiple of the vectorization factor.
104 /// * It handles the code generation for reduction variables.
105 /// * Scalarization (implementation using scalars) of un-vectorizable
107 /// SingleBlockLoopVectorizer does not perform any vectorization-legality
108 /// checks, and relies on the caller to check for the different legality
109 /// aspects. The SingleBlockLoopVectorizer relies on the
110 /// LoopVectorizationLegality class to provide information about the induction
111 /// and reduction variables that were found to a given vectorization factor.
112 class SingleBlockLoopVectorizer {
115 SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
116 DominatorTree *Dt, DataLayout *Dl,
118 OrigLoop(Orig), SE(Se), LI(Li), DT(Dt), DL(Dl), VF(VecWidth),
119 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
121 // Perform the actual loop widening (vectorization).
122 void vectorize(LoopVectorizationLegality *Legal) {
123 // Create a new empty loop. Unlink the old loop and connect the new one.
124 createEmptyLoop(Legal);
125 // Widen each instruction in the old loop to a new one in the new loop.
126 // Use the Legality module to find the induction and reduction variables.
127 vectorizeLoop(Legal);
128 // Register the new loop and update the analysis passes.
133 /// Add code that checks at runtime if the accessed arrays overlap.
134 /// Returns the comperator value or NULL if no check is needed.
135 Value *addRuntimeCheck(LoopVectorizationLegality *Legal,
137 /// Create an empty loop, based on the loop ranges of the old loop.
138 void createEmptyLoop(LoopVectorizationLegality *Legal);
139 /// Copy and widen the instructions from the old loop.
140 void vectorizeLoop(LoopVectorizationLegality *Legal);
141 /// Insert the new loop to the loop hierarchy and pass manager
142 /// and update the analysis passes.
143 void updateAnalysis();
145 /// This instruction is un-vectorizable. Implement it as a sequence
147 void scalarizeInstruction(Instruction *Instr);
149 /// Create a broadcast instruction. This method generates a broadcast
150 /// instruction (shuffle) for loop invariant values and for the induction
151 /// value. If this is the induction variable then we extend it to N, N+1, ...
152 /// this is needed because each iteration in the loop corresponds to a SIMD
154 Value *getBroadcastInstrs(Value *V);
156 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
157 /// for each element in the vector. Starting from zero.
158 Value *getConsecutiveVector(Value* Val);
160 /// When we go over instructions in the basic block we rely on previous
161 /// values within the current basic block or on loop invariant values.
162 /// When we widen (vectorize) values we place them in the map. If the values
163 /// are not within the map, they have to be loop invariant, so we simply
164 /// broadcast them into a vector.
165 Value *getVectorValue(Value *V);
167 /// Get a uniform vector of constant integers. We use this to get
168 /// vectors of ones and zeros for the reduction code.
169 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
171 typedef DenseMap<Value*, Value*> ValueMap;
173 /// The original loop.
175 // Scev analysis to use.
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 = MaxVectorSize);
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, 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);
517 // Save the current insertion location.
518 Instruction *Loc = Builder.GetInsertPoint();
520 // We need to place the broadcast of invariant variables outside the loop.
521 bool Invariant = (OrigLoop->isLoopInvariant(V) && V != Induction);
523 // Place the code for broadcasting invariant variables in the new preheader.
525 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
527 Constant *Zero = ConstantInt::get(I32, 0);
528 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
529 Value *UndefVal = UndefValue::get(VTy);
530 // Insert the value into a new vector.
531 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
532 // Broadcast the scalar into all locations in the vector.
533 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
536 // Restore the builder insertion point.
538 Builder.SetInsertPoint(Loc);
543 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
544 assert(Val->getType()->isVectorTy() && "Must be a vector");
545 assert(Val->getType()->getScalarType()->isIntegerTy() &&
546 "Elem must be an integer");
548 Type *ITy = Val->getType()->getScalarType();
549 VectorType *Ty = cast<VectorType>(Val->getType());
550 unsigned VLen = Ty->getNumElements();
551 SmallVector<Constant*, 8> Indices;
553 // Create a vector of consecutive numbers from zero to VF.
554 for (unsigned i = 0; i < VLen; ++i)
555 Indices.push_back(ConstantInt::get(ITy, i));
557 // Add the consecutive indices to the vector value.
558 Constant *Cv = ConstantVector::get(Indices);
559 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
560 return Builder.CreateAdd(Val, Cv, "induction");
563 bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
564 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
566 // If this pointer is an induction variable, return it.
567 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
568 if (Phi && getInductionVars()->count(Phi))
571 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
575 unsigned NumOperands = Gep->getNumOperands();
576 Value *LastIndex = Gep->getOperand(NumOperands - 1);
578 // Check that all of the gep indices are uniform except for the last.
579 for (unsigned i = 0; i < NumOperands - 1; ++i)
580 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
583 // We can emit wide load/stores only if the last index is the induction
585 const SCEV *Last = SE->getSCEV(LastIndex);
586 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
587 const SCEV *Step = AR->getStepRecurrence(*SE);
589 // The memory is consecutive because the last index is consecutive
590 // and all other indices are loop invariant.
598 bool LoopVectorizationLegality::isUniform(Value *V) {
599 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
602 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
603 assert(V != Induction && "The new induction variable should not be used.");
604 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
605 // If we saved a vectorized copy of V, use it.
606 Value *&MapEntry = WidenMap[V];
610 // Broadcast V and save the value for future uses.
611 Value *B = getBroadcastInstrs(V);
617 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
618 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
621 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
622 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
623 // Holds vector parameters or scalars, in case of uniform vals.
624 SmallVector<Value*, 8> Params;
626 // Find all of the vectorized parameters.
627 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
628 Value *SrcOp = Instr->getOperand(op);
630 // If we are accessing the old induction variable, use the new one.
631 if (SrcOp == OldInduction) {
632 Params.push_back(getVectorValue(SrcOp));
636 // Try using previously calculated values.
637 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
639 // If the src is an instruction that appeared earlier in the basic block
640 // then it should already be vectorized.
641 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
642 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
643 // The parameter is a vector value from earlier.
644 Params.push_back(WidenMap[SrcInst]);
646 // The parameter is a scalar from outside the loop. Maybe even a constant.
647 Params.push_back(SrcOp);
651 assert(Params.size() == Instr->getNumOperands() &&
652 "Invalid number of operands");
654 // Does this instruction return a value ?
655 bool IsVoidRetTy = Instr->getType()->isVoidTy();
656 Value *VecResults = 0;
658 // If we have a return value, create an empty vector. We place the scalarized
659 // instructions in this vector.
661 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
663 // For each scalar that we create:
664 for (unsigned i = 0; i < VF; ++i) {
665 Instruction *Cloned = Instr->clone();
667 Cloned->setName(Instr->getName() + ".cloned");
668 // Replace the operands of the cloned instrucions with extracted scalars.
669 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
670 Value *Op = Params[op];
671 // Param is a vector. Need to extract the right lane.
672 if (Op->getType()->isVectorTy())
673 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
674 Cloned->setOperand(op, Op);
677 // Place the cloned scalar in the new loop.
678 Builder.Insert(Cloned);
680 // If the original scalar returns a value we need to place it in a vector
681 // so that future users will be able to use it.
683 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
684 Builder.getInt32(i));
688 WidenMap[Instr] = VecResults;
692 SingleBlockLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
694 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
695 Legal->getRuntimePointerCheck();
697 if (!PtrRtCheck->Need)
700 Value *MemoryRuntimeCheck = 0;
701 unsigned NumPointers = PtrRtCheck->Pointers.size();
702 SmallVector<Value* , 2> Starts;
703 SmallVector<Value* , 2> Ends;
705 SCEVExpander Exp(*SE, "induction");
707 // Use this type for pointer arithmetic.
708 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
710 for (unsigned i = 0; i < NumPointers; ++i) {
711 Value *Ptr = PtrRtCheck->Pointers[i];
712 const SCEV *Sc = SE->getSCEV(Ptr);
714 if (SE->isLoopInvariant(Sc, OrigLoop)) {
715 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
717 Starts.push_back(Ptr);
720 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
722 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i],
724 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
725 Starts.push_back(Start);
730 for (unsigned i = 0; i < NumPointers; ++i) {
731 for (unsigned j = i+1; j < NumPointers; ++j) {
732 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
733 Starts[i], Ends[j], "bound0", Loc);
734 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
735 Starts[j], Ends[i], "bound1", Loc);
736 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
737 "found.conflict", Loc);
738 if (MemoryRuntimeCheck)
739 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
742 "conflict.rdx", Loc);
744 MemoryRuntimeCheck = IsConflict;
749 return MemoryRuntimeCheck;
753 SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
755 In this function we generate a new loop. The new loop will contain
756 the vectorized instructions while the old loop will continue to run the
759 [ ] <-- vector loop bypass.
762 | [ ] <-- vector pre header.
766 | [ ]_| <-- vector loop.
769 >[ ] <--- middle-block.
772 | [ ] <--- new preheader.
776 | [ ]_| <-- old scalar loop to handle remainder.
783 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
784 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
785 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
786 assert(ExitBlock && "Must have an exit block");
788 // Some loops have a single integer induction variable, while other loops
789 // don't. One example is c++ iterators that often have multiple pointer
790 // induction variables. In the code below we also support a case where we
791 // don't have a single induction variable.
792 OldInduction = Legal->getInduction();
793 Type *IdxTy = OldInduction ? OldInduction->getType() :
794 DL->getIntPtrType(SE->getContext());
796 // Find the loop boundaries.
797 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
798 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
800 // Get the total trip count from the count by adding 1.
801 ExitCount = SE->getAddExpr(ExitCount,
802 SE->getConstant(ExitCount->getType(), 1));
804 // Expand the trip count and place the new instructions in the preheader.
805 // Notice that the pre-header does not change, only the loop body.
806 SCEVExpander Exp(*SE, "induction");
808 // Count holds the overall loop count (N).
809 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
810 BypassBlock->getTerminator());
812 // The loop index does not have to start at Zero. Find the original start
813 // value from the induction PHI node. If we don't have an induction variable
814 // then we know that it starts at zero.
815 Value *StartIdx = OldInduction ?
816 OldInduction->getIncomingValueForBlock(BypassBlock):
817 ConstantInt::get(IdxTy, 0);
819 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
820 assert(BypassBlock && "Invalid loop structure");
822 // Generate the code that checks in runtime if arrays overlap.
823 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal,
824 BypassBlock->getTerminator());
826 // Split the single block loop into the two loop structure described above.
827 BasicBlock *VectorPH =
828 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
829 BasicBlock *VecBody =
830 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
831 BasicBlock *MiddleBlock =
832 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
833 BasicBlock *ScalarPH =
834 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
836 // This is the location in which we add all of the logic for bypassing
837 // the new vector loop.
838 Instruction *Loc = BypassBlock->getTerminator();
840 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
842 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
844 // Generate the induction variable.
845 Induction = Builder.CreatePHI(IdxTy, 2, "index");
846 Constant *Step = ConstantInt::get(IdxTy, VF);
848 // We may need to extend the index in case there is a type mismatch.
849 // We know that the count starts at zero and does not overflow.
850 if (Count->getType() != IdxTy) {
851 // The exit count can be of pointer type. Convert it to the correct
853 if (ExitCount->getType()->isPointerTy())
854 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
856 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
859 // Add the start index to the loop count to get the new end index.
860 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
862 // Now we need to generate the expression for N - (N % VF), which is
863 // the part that the vectorized body will execute.
864 Constant *CIVF = ConstantInt::get(IdxTy, VF);
865 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
866 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
867 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
868 "end.idx.rnd.down", Loc);
870 // Now, compare the new count to zero. If it is zero skip the vector loop and
871 // jump to the scalar loop.
872 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
877 // If we are using memory runtime checks, include them in.
878 if (MemoryRuntimeCheck)
879 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
882 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
883 // Remove the old terminator.
884 Loc->eraseFromParent();
886 // We are going to resume the execution of the scalar loop.
887 // Go over all of the induction variables that we found and fix the
888 // PHIs that are left in the scalar version of the loop.
889 // The starting values of PHI nodes depend on the counter of the last
890 // iteration in the vectorized loop.
891 // If we come from a bypass edge then we need to start from the original start
894 // This variable saves the new starting index for the scalar loop.
895 PHINode *ResumeIndex = 0;
896 LoopVectorizationLegality::InductionList::iterator I, E;
897 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
898 for (I = List->begin(), E = List->end(); I != E; ++I) {
899 PHINode *OrigPhi = I->first;
900 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
901 MiddleBlock->getTerminator());
903 if (OrigPhi->getType()->isIntegerTy()) {
904 // Handle the integer induction counter:
905 assert(OrigPhi == OldInduction && "Unknown integer PHI");
906 // We know what the end value is.
907 EndValue = IdxEndRoundDown;
908 // We also know which PHI node holds it.
909 ResumeIndex = ResumeVal;
911 // For pointer induction variables, calculate the offset using
913 EndValue = GetElementPtrInst::Create(I->second, CountRoundDown,
915 BypassBlock->getTerminator());
918 // The new PHI merges the original incoming value, in case of a bypass,
919 // or the value at the end of the vectorized loop.
920 ResumeVal->addIncoming(I->second, BypassBlock);
921 ResumeVal->addIncoming(EndValue, VecBody);
923 // Fix the scalar body counter (PHI node).
924 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
925 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
928 // If we are generating a new induction variable then we also need to
929 // generate the code that calculates the exit value. This value is not
930 // simply the end of the counter because we may skip the vectorized body
931 // in case of a runtime check.
933 assert(!ResumeIndex && "Unexpected resume value found");
934 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
935 MiddleBlock->getTerminator());
936 ResumeIndex->addIncoming(StartIdx, BypassBlock);
937 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
940 // Make sure that we found the index where scalar loop needs to continue.
941 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
942 "Invalid resume Index");
944 // Add a check in the middle block to see if we have completed
945 // all of the iterations in the first vector loop.
946 // If (N - N%VF) == N, then we *don't* need to run the remainder.
947 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
948 ResumeIndex, "cmp.n",
949 MiddleBlock->getTerminator());
951 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
952 // Remove the old terminator.
953 MiddleBlock->getTerminator()->eraseFromParent();
955 // Create i+1 and fill the PHINode.
956 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
957 Induction->addIncoming(StartIdx, VectorPH);
958 Induction->addIncoming(NextIdx, VecBody);
959 // Create the compare.
960 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
961 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
963 // Now we have two terminators. Remove the old one from the block.
964 VecBody->getTerminator()->eraseFromParent();
966 // Get ready to start creating new instructions into the vectorized body.
967 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
969 // Create and register the new vector loop.
970 Loop* Lp = new Loop();
971 Loop *ParentLoop = OrigLoop->getParentLoop();
973 // Insert the new loop into the loop nest and register the new basic blocks.
975 ParentLoop->addChildLoop(Lp);
976 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
977 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
978 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
980 LI->addTopLevelLoop(Lp);
983 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
986 LoopVectorPreHeader = VectorPH;
987 LoopScalarPreHeader = ScalarPH;
988 LoopMiddleBlock = MiddleBlock;
989 LoopExitBlock = ExitBlock;
990 LoopVectorBody = VecBody;
991 LoopScalarBody = OldBasicBlock;
992 LoopBypassBlock = BypassBlock;
995 /// This function returns the identity element (or neutral element) for
998 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
1000 case LoopVectorizationLegality::IntegerXor:
1001 case LoopVectorizationLegality::IntegerAdd:
1002 case LoopVectorizationLegality::IntegerOr:
1003 // Adding, Xoring, Oring zero to a number does not change it.
1005 case LoopVectorizationLegality::IntegerMult:
1006 // Multiplying a number by 1 does not change it.
1008 case LoopVectorizationLegality::IntegerAnd:
1009 // AND-ing a number with an all-1 value does not change it.
1012 llvm_unreachable("Unknown reduction kind");
1017 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1018 //===------------------------------------------------===//
1020 // Notice: any optimization or new instruction that go
1021 // into the code below should be also be implemented in
1024 //===------------------------------------------------===//
1025 typedef SmallVector<PHINode*, 4> PhiVector;
1026 BasicBlock &BB = *OrigLoop->getHeader();
1027 Constant *Zero = ConstantInt::get(
1028 IntegerType::getInt32Ty(BB.getContext()), 0);
1030 // In order to support reduction variables we need to be able to vectorize
1031 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1032 // stages. First, we create a new vector PHI node with no incoming edges.
1033 // We use this value when we vectorize all of the instructions that use the
1034 // PHI. Next, after all of the instructions in the block are complete we
1035 // add the new incoming edges to the PHI. At this point all of the
1036 // instructions in the basic block are vectorized, so we can use them to
1037 // construct the PHI.
1038 PhiVector RdxPHIsToFix;
1040 // For each instruction in the old loop.
1041 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1042 Instruction *Inst = it;
1044 switch (Inst->getOpcode()) {
1045 case Instruction::Br:
1046 // Nothing to do for PHIs and BR, since we already took care of the
1047 // loop control flow instructions.
1049 case Instruction::PHI:{
1050 PHINode* P = cast<PHINode>(Inst);
1051 // Handle reduction variables:
1052 if (Legal->getReductionVars()->count(P)) {
1053 // This is phase one of vectorizing PHIs.
1054 Type *VecTy = VectorType::get(Inst->getType(), VF);
1055 WidenMap[Inst] = PHINode::Create(VecTy, 2, "vec.phi",
1056 LoopVectorBody->getFirstInsertionPt());
1057 RdxPHIsToFix.push_back(P);
1061 // This PHINode must be an induction variable.
1062 // Make sure that we know about it.
1063 assert(Legal->getInductionVars()->count(P) &&
1064 "Not an induction variable");
1066 if (P->getType()->isIntegerTy()) {
1067 assert(P == OldInduction && "Unexpected PHI");
1068 Value *Broadcasted = getBroadcastInstrs(Induction);
1069 // After broadcasting the induction variable we need to make the
1070 // vector consecutive by adding 0, 1, 2 ...
1071 Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted);
1073 WidenMap[OldInduction] = ConsecutiveInduction;
1077 // Handle pointer inductions.
1078 assert(P->getType()->isPointerTy() && "Unexpected type.");
1079 Value *StartIdx = OldInduction ?
1080 Legal->getInductionVars()->lookup(OldInduction) :
1081 ConstantInt::get(Induction->getType(), 0);
1083 // This is the pointer value coming into the loop.
1084 Value *StartPtr = Legal->getInductionVars()->lookup(P);
1086 // This is the normalized GEP that starts counting at zero.
1087 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1090 // This is the vector of results. Notice that we don't generate vector
1091 // geps because scalar geps result in better code.
1092 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1093 for (unsigned int i = 0; i < VF; ++i) {
1094 Constant *Idx = ConstantInt::get(Induction->getType(), i);
1095 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1096 Value *SclrGep = Builder.CreateGEP(StartPtr, GlobalIdx, "next.gep");
1097 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1098 Builder.getInt32(i),
1102 WidenMap[Inst] = VecVal;
1105 case Instruction::Add:
1106 case Instruction::FAdd:
1107 case Instruction::Sub:
1108 case Instruction::FSub:
1109 case Instruction::Mul:
1110 case Instruction::FMul:
1111 case Instruction::UDiv:
1112 case Instruction::SDiv:
1113 case Instruction::FDiv:
1114 case Instruction::URem:
1115 case Instruction::SRem:
1116 case Instruction::FRem:
1117 case Instruction::Shl:
1118 case Instruction::LShr:
1119 case Instruction::AShr:
1120 case Instruction::And:
1121 case Instruction::Or:
1122 case Instruction::Xor: {
1123 // Just widen binops.
1124 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1125 Value *A = getVectorValue(Inst->getOperand(0));
1126 Value *B = getVectorValue(Inst->getOperand(1));
1128 // Use this vector value for all users of the original instruction.
1129 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1132 // Update the NSW, NUW and Exact flags.
1133 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1134 if (isa<OverflowingBinaryOperator>(BinOp)) {
1135 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1136 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1138 if (isa<PossiblyExactOperator>(VecOp))
1139 VecOp->setIsExact(BinOp->isExact());
1142 case Instruction::Select: {
1144 // If the selector is loop invariant we can create a select
1145 // instruction with a scalar condition. Otherwise, use vector-select.
1146 Value *Cond = Inst->getOperand(0);
1147 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1149 // The condition can be loop invariant but still defined inside the
1150 // loop. This means that we can't just use the original 'cond' value.
1151 // We have to take the 'vectorized' value and pick the first lane.
1152 // Instcombine will make this a no-op.
1153 Cond = getVectorValue(Cond);
1155 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1157 Value *Op0 = getVectorValue(Inst->getOperand(1));
1158 Value *Op1 = getVectorValue(Inst->getOperand(2));
1159 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
1163 case Instruction::ICmp:
1164 case Instruction::FCmp: {
1165 // Widen compares. Generate vector compares.
1166 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
1167 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
1168 Value *A = getVectorValue(Inst->getOperand(0));
1169 Value *B = getVectorValue(Inst->getOperand(1));
1171 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1173 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1177 case Instruction::Store: {
1178 // Attempt to issue a wide store.
1179 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1180 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1181 Value *Ptr = SI->getPointerOperand();
1182 unsigned Alignment = SI->getAlignment();
1184 assert(!Legal->isUniform(Ptr) &&
1185 "We do not allow storing to uniform addresses");
1187 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1189 // This store does not use GEPs.
1190 if (!Legal->isConsecutivePtr(Ptr)) {
1191 scalarizeInstruction(Inst);
1196 // The last index does not have to be the induction. It can be
1197 // consecutive and be a function of the index. For example A[I+1];
1198 unsigned NumOperands = Gep->getNumOperands();
1199 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1200 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1202 // Create the new GEP with the new induction variable.
1203 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1204 Gep2->setOperand(NumOperands - 1, LastIndex);
1205 Ptr = Builder.Insert(Gep2);
1207 // Use the induction element ptr.
1208 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1209 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1211 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1212 Value *Val = getVectorValue(SI->getValueOperand());
1213 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1216 case Instruction::Load: {
1217 // Attempt to issue a wide load.
1218 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1219 Type *RetTy = VectorType::get(LI->getType(), VF);
1220 Value *Ptr = LI->getPointerOperand();
1221 unsigned Alignment = LI->getAlignment();
1222 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1224 // If the pointer is loop invariant or if it is non consecutive,
1225 // scalarize the load.
1226 bool Con = Legal->isConsecutivePtr(Ptr);
1227 if (Legal->isUniform(Ptr) || !Con) {
1228 scalarizeInstruction(Inst);
1233 // The last index does not have to be the induction. It can be
1234 // consecutive and be a function of the index. For example A[I+1];
1235 unsigned NumOperands = Gep->getNumOperands();
1236 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1237 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1239 // Create the new GEP with the new induction variable.
1240 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1241 Gep2->setOperand(NumOperands - 1, LastIndex);
1242 Ptr = Builder.Insert(Gep2);
1244 // Use the induction element ptr.
1245 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1246 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1249 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1250 LI = Builder.CreateLoad(Ptr);
1251 LI->setAlignment(Alignment);
1252 // Use this vector value for all users of the load.
1253 WidenMap[Inst] = LI;
1256 case Instruction::ZExt:
1257 case Instruction::SExt:
1258 case Instruction::FPToUI:
1259 case Instruction::FPToSI:
1260 case Instruction::FPExt:
1261 case Instruction::PtrToInt:
1262 case Instruction::IntToPtr:
1263 case Instruction::SIToFP:
1264 case Instruction::UIToFP:
1265 case Instruction::Trunc:
1266 case Instruction::FPTrunc:
1267 case Instruction::BitCast: {
1268 /// Vectorize bitcasts.
1269 CastInst *CI = dyn_cast<CastInst>(Inst);
1270 Value *A = getVectorValue(Inst->getOperand(0));
1271 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1272 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1277 /// All other instructions are unsupported. Scalarize them.
1278 scalarizeInstruction(Inst);
1281 }// end of for_each instr.
1283 // At this point every instruction in the original loop is widended to
1284 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1285 // that we vectorized. The PHI nodes are currently empty because we did
1286 // not want to introduce cycles. Notice that the remaining PHI nodes
1287 // that we need to fix are reduction variables.
1289 // Create the 'reduced' values for each of the induction vars.
1290 // The reduced values are the vector values that we scalarize and combine
1291 // after the loop is finished.
1292 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1294 PHINode *RdxPhi = *it;
1295 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1296 assert(RdxPhi && "Unable to recover vectorized PHI");
1298 // Find the reduction variable descriptor.
1299 assert(Legal->getReductionVars()->count(RdxPhi) &&
1300 "Unable to find the reduction variable");
1301 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1302 (*Legal->getReductionVars())[RdxPhi];
1304 // We need to generate a reduction vector from the incoming scalar.
1305 // To do so, we need to generate the 'identity' vector and overide
1306 // one of the elements with the incoming scalar reduction. We need
1307 // to do it in the vector-loop preheader.
1308 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1310 // This is the vector-clone of the value that leaves the loop.
1311 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1312 Type *VecTy = VectorExit->getType();
1314 // Find the reduction identity variable. Zero for addition, or, xor,
1315 // one for multiplication, -1 for And.
1316 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1317 VecTy->getScalarType());
1319 // This vector is the Identity vector where the first element is the
1320 // incoming scalar reduction.
1321 Value *VectorStart = Builder.CreateInsertElement(Identity,
1322 RdxDesc.StartValue, Zero);
1324 // Fix the vector-loop phi.
1325 // We created the induction variable so we know that the
1326 // preheader is the first entry.
1327 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1329 // Reductions do not have to start at zero. They can start with
1330 // any loop invariant values.
1331 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1332 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1333 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1334 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1336 // Before each round, move the insertion point right between
1337 // the PHIs and the values we are going to write.
1338 // This allows us to write both PHINodes and the extractelement
1340 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1342 // This PHINode contains the vectorized reduction variable, or
1343 // the initial value vector, if we bypass the vector loop.
1344 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1345 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1346 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1348 // Extract the first scalar.
1350 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1351 // Extract and reduce the remaining vector elements.
1352 for (unsigned i=1; i < VF; ++i) {
1354 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1355 switch (RdxDesc.Kind) {
1356 case LoopVectorizationLegality::IntegerAdd:
1357 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1359 case LoopVectorizationLegality::IntegerMult:
1360 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1362 case LoopVectorizationLegality::IntegerOr:
1363 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1365 case LoopVectorizationLegality::IntegerAnd:
1366 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1368 case LoopVectorizationLegality::IntegerXor:
1369 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1372 llvm_unreachable("Unknown reduction operation");
1376 // Now, we need to fix the users of the reduction variable
1377 // inside and outside of the scalar remainder loop.
1378 // We know that the loop is in LCSSA form. We need to update the
1379 // PHI nodes in the exit blocks.
1380 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1381 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1382 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1383 if (!LCSSAPhi) continue;
1385 // All PHINodes need to have a single entry edge, or two if
1386 // we already fixed them.
1387 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1389 // We found our reduction value exit-PHI. Update it with the
1390 // incoming bypass edge.
1391 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1392 // Add an edge coming from the bypass.
1393 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1396 }// end of the LCSSA phi scan.
1398 // Fix the scalar loop reduction variable with the incoming reduction sum
1399 // from the vector body and from the backedge value.
1400 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1401 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1402 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1403 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1404 }// end of for each redux variable.
1407 void SingleBlockLoopVectorizer::updateAnalysis() {
1408 // Forget the original basic block.
1409 SE->forgetLoop(OrigLoop);
1411 // Update the dominator tree information.
1412 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1413 "Entry does not dominate exit.");
1415 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1416 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1417 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1418 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1419 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1420 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1422 DEBUG(DT->verifyAnalysis());
1425 bool LoopVectorizationLegality::canVectorize() {
1426 assert(TheLoop->getLoopPreheader() && "No preheader!!");
1428 // We can only vectorize single basic block loops.
1429 unsigned NumBlocks = TheLoop->getNumBlocks();
1430 if (NumBlocks != 1) {
1431 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1435 // We need to have a loop header.
1436 BasicBlock *BB = TheLoop->getHeader();
1437 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1439 // ScalarEvolution needs to be able to find the exit count.
1440 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1441 if (ExitCount == SE->getCouldNotCompute()) {
1442 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1446 // Do not loop-vectorize loops with a tiny trip count.
1447 unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB);
1448 if (TC > 0u && TC < TinyTripCountThreshold) {
1449 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1450 "This loop is not worth vectorizing.\n");
1454 // Go over each instruction and look at memory deps.
1455 if (!canVectorizeBlock(*BB)) {
1456 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1460 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1461 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1464 // Okay! We can vectorize. At this point we don't have any other mem analysis
1465 // which may limit our maximum vectorization factor, so just return true with
1470 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1472 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1474 // Scan the instructions in the block and look for hazards.
1475 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1476 Instruction *I = it;
1478 if (PHINode *Phi = dyn_cast<PHINode>(I)) {
1479 // This should not happen because the loop should be normalized.
1480 if (Phi->getNumIncomingValues() != 2) {
1481 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1485 // This is the value coming from the preheader.
1486 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1488 // We only look at integer and pointer phi nodes.
1489 if (Phi->getType()->isPointerTy() && isInductionVariable(Phi)) {
1490 DEBUG(dbgs() << "LV: Found a pointer induction variable.\n");
1491 Inductions[Phi] = StartValue;
1493 } else if (!Phi->getType()->isIntegerTy()) {
1494 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1498 // Handle integer PHIs:
1499 if (isInductionVariable(Phi)) {
1501 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1504 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1506 Inductions[Phi] = StartValue;
1509 if (AddReductionVar(Phi, IntegerAdd)) {
1510 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1513 if (AddReductionVar(Phi, IntegerMult)) {
1514 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1517 if (AddReductionVar(Phi, IntegerOr)) {
1518 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1521 if (AddReductionVar(Phi, IntegerAnd)) {
1522 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1525 if (AddReductionVar(Phi, IntegerXor)) {
1526 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1530 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1532 }// end of PHI handling
1534 // We still don't handle functions.
1535 CallInst *CI = dyn_cast<CallInst>(I);
1537 DEBUG(dbgs() << "LV: Found a call site.\n");
1541 // We do not re-vectorize vectors.
1542 if (!VectorType::isValidElementType(I->getType()) &&
1543 !I->getType()->isVoidTy()) {
1544 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1548 // Reduction instructions are allowed to have exit users.
1549 // All other instructions must not have external users.
1550 if (!AllowedExit.count(I))
1551 //Check that all of the users of the loop are inside the BB.
1552 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1554 Instruction *U = cast<Instruction>(*it);
1555 // This user may be a reduction exit value.
1556 BasicBlock *Parent = U->getParent();
1557 if (Parent != &BB) {
1558 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1565 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
1566 assert(getInductionVars()->size() && "No induction variables");
1569 // Don't vectorize if the memory dependencies do not allow vectorization.
1570 if (!canVectorizeMemory(BB))
1573 // We now know that the loop is vectorizable!
1574 // Collect variables that will remain uniform after vectorization.
1575 std::vector<Value*> Worklist;
1577 // Start with the conditional branch and walk up the block.
1578 Worklist.push_back(BB.getTerminator()->getOperand(0));
1580 while (Worklist.size()) {
1581 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1582 Worklist.pop_back();
1584 // Look at instructions inside this block. Stop when reaching PHI nodes.
1585 if (!I || I->getParent() != &BB || isa<PHINode>(I))
1588 // This is a known uniform.
1591 // Insert all operands.
1592 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
1593 Worklist.push_back(I->getOperand(i));
1600 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1601 typedef SmallVector<Value*, 16> ValueVector;
1602 typedef SmallPtrSet<Value*, 16> ValueSet;
1603 // Holds the Load and Store *instructions*.
1606 PtrRtCheck.Pointers.clear();
1607 PtrRtCheck.Need = false;
1609 // Scan the BB and collect legal loads and stores.
1610 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1611 Instruction *I = it;
1613 // If this is a load, save it. If this instruction can read from memory
1614 // but is not a load, then we quit. Notice that we don't handle function
1615 // calls that read or write.
1616 if (I->mayReadFromMemory()) {
1617 LoadInst *Ld = dyn_cast<LoadInst>(I);
1618 if (!Ld) return false;
1619 if (!Ld->isSimple()) {
1620 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1623 Loads.push_back(Ld);
1627 // Save store instructions. Abort if other instructions write to memory.
1628 if (I->mayWriteToMemory()) {
1629 StoreInst *St = dyn_cast<StoreInst>(I);
1630 if (!St) return false;
1631 if (!St->isSimple()) {
1632 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1635 Stores.push_back(St);
1639 // Now we have two lists that hold the loads and the stores.
1640 // Next, we find the pointers that they use.
1642 // Check if we see any stores. If there are no stores, then we don't
1643 // care if the pointers are *restrict*.
1644 if (!Stores.size()) {
1645 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1649 // Holds the read and read-write *pointers* that we find.
1651 ValueVector ReadWrites;
1653 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1654 // multiple times on the same object. If the ptr is accessed twice, once
1655 // for read and once for write, it will only appear once (on the write
1656 // list). This is okay, since we are going to check for conflicts between
1657 // writes and between reads and writes, but not between reads and reads.
1660 ValueVector::iterator I, IE;
1661 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1662 StoreInst *ST = dyn_cast<StoreInst>(*I);
1663 assert(ST && "Bad StoreInst");
1664 Value* Ptr = ST->getPointerOperand();
1666 if (isUniform(Ptr)) {
1667 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1671 // If we did *not* see this pointer before, insert it to
1672 // the read-write list. At this phase it is only a 'write' list.
1673 if (Seen.insert(Ptr))
1674 ReadWrites.push_back(Ptr);
1677 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1678 LoadInst *LD = dyn_cast<LoadInst>(*I);
1679 assert(LD && "Bad LoadInst");
1680 Value* Ptr = LD->getPointerOperand();
1681 // If we did *not* see this pointer before, insert it to the
1682 // read list. If we *did* see it before, then it is already in
1683 // the read-write list. This allows us to vectorize expressions
1684 // such as A[i] += x; Because the address of A[i] is a read-write
1685 // pointer. This only works if the index of A[i] is consecutive.
1686 // If the address of i is unknown (for example A[B[i]]) then we may
1687 // read a few words, modify, and write a few words, and some of the
1688 // words may be written to the same address.
1689 if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr))
1690 Reads.push_back(Ptr);
1693 // If we write (or read-write) to a single destination and there are no
1694 // other reads in this loop then is it safe to vectorize.
1695 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1696 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1700 // Find pointers with computable bounds. We are going to use this information
1701 // to place a runtime bound check.
1703 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1704 if (hasComputableBounds(*I)) {
1705 PtrRtCheck.insert(SE, TheLoop, *I);
1706 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1711 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1712 if (hasComputableBounds(*I)) {
1713 PtrRtCheck.insert(SE, TheLoop, *I);
1714 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1720 // Check that we did not collect too many pointers or found a
1721 // unsizeable pointer.
1722 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1727 PtrRtCheck.Need = RT;
1730 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1733 // Now that the pointers are in two lists (Reads and ReadWrites), we
1734 // can check that there are no conflicts between each of the writes and
1735 // between the writes to the reads.
1736 ValueSet WriteObjects;
1737 ValueVector TempObjects;
1739 // Check that the read-writes do not conflict with other read-write
1741 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1742 GetUnderlyingObjects(*I, TempObjects, DL);
1743 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1745 if (!isIdentifiedObject(*it)) {
1746 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1749 if (!WriteObjects.insert(*it)) {
1750 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1755 TempObjects.clear();
1758 /// Check that the reads don't conflict with the read-writes.
1759 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1760 GetUnderlyingObjects(*I, TempObjects, DL);
1761 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1763 if (!isIdentifiedObject(*it)) {
1764 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1767 if (WriteObjects.count(*it)) {
1768 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1773 TempObjects.clear();
1776 // It is safe to vectorize and we don't need any runtime checks.
1777 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1782 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1783 ReductionKind Kind) {
1784 if (Phi->getNumIncomingValues() != 2)
1787 // Find the possible incoming reduction variable.
1788 BasicBlock *BB = Phi->getParent();
1789 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1790 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1791 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1793 // ExitInstruction is the single value which is used outside the loop.
1794 // We only allow for a single reduction value to be used outside the loop.
1795 // This includes users of the reduction, variables (which form a cycle
1796 // which ends in the phi node).
1797 Instruction *ExitInstruction = 0;
1799 // Iter is our iterator. We start with the PHI node and scan for all of the
1800 // users of this instruction. All users must be instructions which can be
1801 // used as reduction variables (such as ADD). We may have a single
1802 // out-of-block user. They cycle must end with the original PHI.
1803 // Also, we can't have multiple block-local users.
1804 Instruction *Iter = Phi;
1806 // Any reduction instr must be of one of the allowed kinds.
1807 if (!isReductionInstr(Iter, Kind))
1810 // Did we found a user inside this block ?
1811 bool FoundInBlockUser = false;
1812 // Did we reach the initial PHI node ?
1813 bool FoundStartPHI = false;
1815 // If the instruction has no users then this is a broken
1816 // chain and can't be a reduction variable.
1817 if (Iter->use_empty())
1820 // For each of the *users* of iter.
1821 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1823 Instruction *U = cast<Instruction>(*it);
1824 // We already know that the PHI is a user.
1826 FoundStartPHI = true;
1829 // Check if we found the exit user.
1830 BasicBlock *Parent = U->getParent();
1832 // We must have a single exit instruction.
1833 if (ExitInstruction != 0)
1835 ExitInstruction = Iter;
1837 // We can't have multiple inside users.
1838 if (FoundInBlockUser)
1840 FoundInBlockUser = true;
1844 // We found a reduction var if we have reached the original
1845 // phi node and we only have a single instruction with out-of-loop
1847 if (FoundStartPHI && ExitInstruction) {
1848 // This instruction is allowed to have out-of-loop users.
1849 AllowedExit.insert(ExitInstruction);
1851 // Save the description of this reduction variable.
1852 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1853 Reductions[Phi] = RD;
1860 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1861 ReductionKind Kind) {
1862 switch (I->getOpcode()) {
1865 case Instruction::PHI:
1868 case Instruction::Add:
1869 case Instruction::Sub:
1870 return Kind == IntegerAdd;
1871 case Instruction::Mul:
1872 return Kind == IntegerMult;
1873 case Instruction::And:
1874 return Kind == IntegerAnd;
1875 case Instruction::Or:
1876 return Kind == IntegerOr;
1877 case Instruction::Xor:
1878 return Kind == IntegerXor;
1882 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1883 Type *PhiTy = Phi->getType();
1884 // We only handle integer and pointer inductions variables.
1885 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1888 // Check that the PHI is consecutive and starts at zero.
1889 const SCEV *PhiScev = SE->getSCEV(Phi);
1890 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1892 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1895 const SCEV *Step = AR->getStepRecurrence(*SE);
1897 // Integer inductions need to have a stride of one.
1898 if (PhiTy->isIntegerTy())
1899 return Step->isOne();
1901 // Calculate the pointer stride and check if it is consecutive.
1902 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1903 if (!C) return false;
1905 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1906 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1907 return (C->getValue()->equalsInt(Size));
1910 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1911 const SCEV *PhiScev = SE->getSCEV(Ptr);
1912 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1916 return AR->isAffine();
1920 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1922 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1926 float Cost = expectedCost(1);
1928 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1929 for (unsigned i=2; i <= VF; i*=2) {
1930 // Notice that the vector loop needs to be executed less times, so
1931 // we need to divide the cost of the vector loops by the width of
1932 // the vector elements.
1933 float VectorCost = expectedCost(i) / (float)i;
1934 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1935 (int)VectorCost << ".\n");
1936 if (VectorCost < Cost) {
1942 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1946 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1947 // We can only estimate the cost of single basic block loops.
1948 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1950 BasicBlock *BB = TheLoop->getHeader();
1953 // For each instruction in the old loop.
1954 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1955 Instruction *Inst = it;
1956 unsigned C = getInstructionCost(Inst, VF);
1958 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1959 " For instruction: "<< *Inst << "\n");
1966 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1967 assert(VTTI && "Invalid vector target transformation info");
1969 // If we know that this instruction will remain uniform, check the cost of
1970 // the scalar version.
1971 if (Legal->isUniformAfterVectorization(I))
1974 Type *RetTy = I->getType();
1975 Type *VectorTy = ToVectorTy(RetTy, VF);
1978 // TODO: We need to estimate the cost of intrinsic calls.
1979 switch (I->getOpcode()) {
1980 case Instruction::GetElementPtr:
1981 // We mark this instruction as zero-cost because scalar GEPs are usually
1982 // lowered to the intruction addressing mode. At the moment we don't
1983 // generate vector geps.
1985 case Instruction::Br: {
1986 return VTTI->getCFInstrCost(I->getOpcode());
1988 case Instruction::PHI:
1990 case Instruction::Add:
1991 case Instruction::FAdd:
1992 case Instruction::Sub:
1993 case Instruction::FSub:
1994 case Instruction::Mul:
1995 case Instruction::FMul:
1996 case Instruction::UDiv:
1997 case Instruction::SDiv:
1998 case Instruction::FDiv:
1999 case Instruction::URem:
2000 case Instruction::SRem:
2001 case Instruction::FRem:
2002 case Instruction::Shl:
2003 case Instruction::LShr:
2004 case Instruction::AShr:
2005 case Instruction::And:
2006 case Instruction::Or:
2007 case Instruction::Xor:
2008 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
2009 case Instruction::Select: {
2010 SelectInst *SI = cast<SelectInst>(I);
2011 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
2012 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
2013 Type *CondTy = SI->getCondition()->getType();
2015 CondTy = VectorType::get(CondTy, VF);
2017 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
2019 case Instruction::ICmp:
2020 case Instruction::FCmp: {
2021 Type *ValTy = I->getOperand(0)->getType();
2022 VectorTy = ToVectorTy(ValTy, VF);
2023 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
2025 case Instruction::Store: {
2026 StoreInst *SI = cast<StoreInst>(I);
2027 Type *ValTy = SI->getValueOperand()->getType();
2028 VectorTy = ToVectorTy(ValTy, VF);
2031 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
2032 SI->getAlignment(), SI->getPointerAddressSpace());
2034 // Scalarized stores.
2035 if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
2037 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2039 // The cost of extracting from the value vector.
2040 Cost += VF * (ExtCost);
2041 // The cost of the scalar stores.
2042 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2043 ValTy->getScalarType(),
2045 SI->getPointerAddressSpace());
2050 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
2051 SI->getPointerAddressSpace());
2053 case Instruction::Load: {
2054 LoadInst *LI = cast<LoadInst>(I);
2057 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
2059 LI->getPointerAddressSpace());
2061 // Scalarized loads.
2062 if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
2064 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
2065 // The cost of inserting the loaded value into the result vector.
2066 Cost += VF * (InCost);
2067 // The cost of the scalar stores.
2068 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2069 RetTy->getScalarType(),
2071 LI->getPointerAddressSpace());
2076 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2077 LI->getPointerAddressSpace());
2079 case Instruction::ZExt:
2080 case Instruction::SExt:
2081 case Instruction::FPToUI:
2082 case Instruction::FPToSI:
2083 case Instruction::FPExt:
2084 case Instruction::PtrToInt:
2085 case Instruction::IntToPtr:
2086 case Instruction::SIToFP:
2087 case Instruction::UIToFP:
2088 case Instruction::Trunc:
2089 case Instruction::FPTrunc:
2090 case Instruction::BitCast: {
2091 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2092 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
2095 // We are scalarizing the instruction. Return the cost of the scalar
2096 // instruction, plus the cost of insert and extract into vector
2097 // elements, times the vector width.
2100 bool IsVoid = RetTy->isVoidTy();
2102 unsigned InsCost = (IsVoid ? 0 :
2103 VTTI->getInstrCost(Instruction::InsertElement,
2106 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2109 // The cost of inserting the results plus extracting each one of the
2111 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
2113 // The cost of executing VF copies of the scalar instruction.
2114 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
2120 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2121 if (Scalar->isVoidTy() || VF == 1)
2123 return VectorType::get(Scalar, VF);
2128 char LoopVectorize::ID = 0;
2129 static const char lv_name[] = "Loop Vectorization";
2130 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2131 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2132 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2133 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2134 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2137 Pass *createLoopVectorizePass() {
2138 return new LoopVectorize();