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,
119 OrigLoop(Orig), SE(Se), LI(Li), DT(dt), DL(dl), LPM(Lpm), VF(VecWidth),
120 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
122 // Perform the actual loop widening (vectorization).
123 void vectorize(LoopVectorizationLegality *Legal) {
124 // Create a new empty loop. Unlink the old loop and connect the new one.
125 createEmptyLoop(Legal);
126 // Widen each instruction in the old loop to a new one in the new loop.
127 // Use the Legality module to find the induction and reduction variables.
128 vectorizeLoop(Legal);
129 // Register the new loop and update the analysis passes.
134 /// Add code that checks at runtime if the accessed arrays overlap.
135 /// Returns the comperator value or NULL if no check is needed.
136 Value* addRuntimeCheck(LoopVectorizationLegality *Legal,
138 /// Create an empty loop, based on the loop ranges of the old loop.
139 void createEmptyLoop(LoopVectorizationLegality *Legal);
140 /// Copy and widen the instructions from the old loop.
141 void vectorizeLoop(LoopVectorizationLegality *Legal);
142 /// Insert the new loop to the loop hierarchy and pass manager
143 /// and update the analysis passes.
144 void updateAnalysis();
146 /// This instruction is un-vectorizable. Implement it as a sequence
148 void scalarizeInstruction(Instruction *Instr);
150 /// Create a broadcast instruction. This method generates a broadcast
151 /// instruction (shuffle) for loop invariant values and for the induction
152 /// value. If this is the induction variable then we extend it to N, N+1, ...
153 /// this is needed because each iteration in the loop corresponds to a SIMD
155 Value *getBroadcastInstrs(Value *V);
157 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
158 /// for each element in the vector. Starting from zero.
159 Value *getConsecutiveVector(Value* Val);
161 /// When we go over instructions in the basic block we rely on previous
162 /// values within the current basic block or on loop invariant values.
163 /// When we widen (vectorize) values we place them in the map. If the values
164 /// are not within the map, they have to be loop invariant, so we simply
165 /// broadcast them into a vector.
166 Value *getVectorValue(Value *V);
168 /// Get a uniform vector of constant integers. We use this to get
169 /// vectors of ones and zeros for the reduction code.
170 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
172 typedef DenseMap<Value*, Value*> ValueMap;
174 /// The original loop.
176 // Scev analysis to use.
184 // Loop Pass Manager;
186 // The vectorization factor to use.
189 // The builder that we use
192 // --- Vectorization state ---
194 /// The vector-loop preheader.
195 BasicBlock *LoopVectorPreHeader;
196 /// The scalar-loop preheader.
197 BasicBlock *LoopScalarPreHeader;
198 /// Middle Block between the vector and the scalar.
199 BasicBlock *LoopMiddleBlock;
200 ///The ExitBlock of the scalar loop.
201 BasicBlock *LoopExitBlock;
202 ///The vector loop body.
203 BasicBlock *LoopVectorBody;
204 ///The scalar loop body.
205 BasicBlock *LoopScalarBody;
206 ///The first bypass block.
207 BasicBlock *LoopBypassBlock;
209 /// The new Induction variable which was added to the new block.
211 /// The induction variable of the old basic block.
212 PHINode *OldInduction;
213 // Maps scalars to widened vectors.
217 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
218 /// to what vectorization factor.
219 /// This class does not look at the profitability of vectorization, only the
220 /// legality. This class has two main kinds of checks:
221 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
222 /// will change the order of memory accesses in a way that will change the
223 /// correctness of the program.
224 /// * Scalars checks - The code in canVectorizeBlock checks for a number
225 /// of different conditions, such as the availability of a single induction
226 /// variable, that all types are supported and vectorize-able, etc.
227 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
228 /// This class is also used by SingleBlockLoopVectorizer for identifying
229 /// induction variable and the different reduction variables.
230 class LoopVectorizationLegality {
232 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
233 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
235 /// This represents the kinds of reductions that we support.
237 NoReduction, /// Not a reduction.
238 IntegerAdd, /// Sum of numbers.
239 IntegerMult, /// Product of numbers.
240 IntegerOr, /// Bitwise or logical OR of numbers.
241 IntegerAnd, /// Bitwise or logical AND of numbers.
242 IntegerXor /// Bitwise or logical XOR of numbers.
245 /// This POD struct holds information about reduction variables.
246 struct ReductionDescriptor {
248 ReductionDescriptor():
249 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
252 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
253 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
255 // The starting value of the reduction.
256 // It does not have to be zero!
258 // The instruction who's value is used outside the loop.
259 Instruction *LoopExitInstr;
260 // The kind of the reduction.
264 // This POD struct holds information about the memory runtime legality
265 // check that a group of pointers do not overlap.
266 struct RuntimePointerCheck {
267 RuntimePointerCheck(): Need(false) {}
269 /// Reset the state of the pointer runtime information.
277 /// Insert a pointer and calculate the start and end SCEVs.
278 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr) {
279 const SCEV *Sc = SE->getSCEV(Ptr);
280 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
281 assert(AR && "Invalid addrec expression");
282 const SCEV *Ex = SE->getExitCount(Lp, Lp->getHeader());
283 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
284 Pointers.push_back(Ptr);
285 Starts.push_back(AR->getStart());
286 Ends.push_back(ScEnd);
289 /// This flag indicates if we need to add the runtime check.
291 /// Holds the pointers that we need to check.
292 SmallVector<Value*, 2> Pointers;
293 /// Holds the pointer value at the beginning of the loop.
294 SmallVector<const SCEV*, 2> Starts;
295 /// Holds the pointer value at the end of the loop.
296 SmallVector<const SCEV*, 2> Ends;
299 /// ReductionList contains the reduction descriptors for all
300 /// of the reductions that were found in the loop.
301 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
303 /// InductionList saves induction variables and maps them to the initial
304 /// value entring the loop.
305 typedef DenseMap<PHINode*, Value*> InductionList;
307 /// Returns true if it is legal to vectorize this loop.
308 /// This does not mean that it is profitable to vectorize this
309 /// loop, only that it is legal to do so.
312 /// Returns the Induction variable.
313 PHINode *getInduction() {return Induction;}
315 /// Returns the reduction variables found in the loop.
316 ReductionList *getReductionVars() { return &Reductions; }
318 /// Returns the induction variables found in the loop.
319 InductionList *getInductionVars() { return &Inductions; }
321 /// Check if this pointer is consecutive when vectorizing. This happens
322 /// when the last index of the GEP is the induction variable, or that the
323 /// pointer itself is an induction variable.
324 /// This check allows us to vectorize A[idx] into a wide load/store.
325 bool isConsecutivePtr(Value *Ptr);
327 /// Returns true if the value V is uniform within the loop.
328 bool isUniform(Value *V);
330 /// Returns true if this instruction will remain scalar after vectorization.
331 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
333 /// Returns the information that we collected about runtime memory check.
334 RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; }
336 /// Check if a single basic block loop is vectorizable.
337 /// At this point we know that this is a loop with a constant trip count
338 /// and we only need to check individual instructions.
339 bool canVectorizeBlock(BasicBlock &BB);
341 /// When we vectorize loops we may change the order in which
342 /// we read and write from memory. This method checks if it is
343 /// legal to vectorize the code, considering only memory constrains.
344 /// Returns true if BB is vectorizable
345 bool canVectorizeMemory(BasicBlock &BB);
347 /// Returns True, if 'Phi' is the kind of reduction variable for type
348 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
349 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
350 /// Returns true if the instruction I can be a reduction variable of type
352 bool isReductionInstr(Instruction *I, ReductionKind Kind);
353 /// Returns True, if 'Phi' is an induction variable.
354 bool isInductionVariable(PHINode *Phi);
355 /// Return true if can compute the address bounds of Ptr within the loop.
356 bool hasComputableBounds(Value *Ptr);
358 /// The loop that we evaluate.
362 /// DataLayout analysis.
365 // --- vectorization state --- //
367 /// Holds the integer induction variable. This is the counter of the
370 /// Holds the reduction variables.
371 ReductionList Reductions;
372 /// Holds all of the induction variables that we found in the loop.
373 /// Notice that inductions don't need to start at zero and that induction
374 /// variables can be pointers.
375 InductionList Inductions;
377 /// Allowed outside users. This holds the reduction
378 /// vars which can be accessed from outside the loop.
379 SmallPtrSet<Value*, 4> AllowedExit;
380 /// This set holds the variables which are known to be uniform after
382 SmallPtrSet<Instruction*, 4> Uniforms;
383 /// We need to check that all of the pointers in this list are disjoint
385 RuntimePointerCheck PtrRtCheck;
388 /// LoopVectorizationCostModel - estimates the expected speedups due to
390 /// In many cases vectorization is not profitable. This can happen because
391 /// of a number of reasons. In this class we mainly attempt to predict
392 /// the expected speedup/slowdowns due to the supported instruction set.
393 /// We use the VectorTargetTransformInfo to query the different backends
394 /// for the cost of different operations.
395 class LoopVectorizationCostModel {
398 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
399 LoopVectorizationLegality *Leg,
400 const VectorTargetTransformInfo *Vtti):
401 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
403 /// Returns the most profitable vectorization factor for the loop that is
404 /// smaller or equal to the VF argument. This method checks every power
406 unsigned findBestVectorizationFactor(unsigned VF = MaxVectorSize);
409 /// Returns the expected execution cost. The unit of the cost does
410 /// not matter because we use the 'cost' units to compare different
411 /// vector widths. The cost that is returned is *not* normalized by
412 /// the factor width.
413 unsigned expectedCost(unsigned VF);
415 /// Returns the execution time cost of an instruction for a given vector
416 /// width. Vector width of one means scalar.
417 unsigned getInstructionCost(Instruction *I, unsigned VF);
419 /// A helper function for converting Scalar types to vector types.
420 /// If the incoming type is void, we return void. If the VF is 1, we return
422 static Type* ToVectorTy(Type *Scalar, unsigned VF);
424 /// The loop that we evaluate.
429 /// Vectorization legality.
430 LoopVectorizationLegality *Legal;
431 /// Vector target information.
432 const VectorTargetTransformInfo *VTTI;
435 struct LoopVectorize : public LoopPass {
436 static char ID; // Pass identification, replacement for typeid
438 LoopVectorize() : LoopPass(ID) {
439 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
445 TargetTransformInfo *TTI;
448 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
449 // We only vectorize innermost loops.
453 SE = &getAnalysis<ScalarEvolution>();
454 DL = getAnalysisIfAvailable<DataLayout>();
455 LI = &getAnalysis<LoopInfo>();
456 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
457 DT = &getAnalysis<DominatorTree>();
459 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
460 L->getHeader()->getParent()->getName() << "\"\n");
462 // Check if it is legal to vectorize the loop.
463 LoopVectorizationLegality LVL(L, SE, DL);
464 if (!LVL.canVectorize()) {
465 DEBUG(dbgs() << "LV: Not vectorizing.\n");
469 // Select the preffered vectorization factor.
471 if (VectorizationFactor == 0) {
472 const VectorTargetTransformInfo *VTTI = 0;
474 VTTI = TTI->getVectorTargetTransformInfo();
475 // Use the cost model.
476 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
477 VF = CM.findBestVectorizationFactor();
480 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
485 // Use the user command flag.
486 VF = VectorizationFactor;
489 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
490 L->getHeader()->getParent()->getParent()->getModuleIdentifier()<<
493 // If we decided that it is *legal* to vectorizer the loop then do it.
494 SingleBlockLoopVectorizer LB(L, SE, LI, DT, DL, &LPM, VF);
497 DEBUG(verifyFunction(*L->getHeader()->getParent()));
501 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
502 LoopPass::getAnalysisUsage(AU);
503 AU.addRequiredID(LoopSimplifyID);
504 AU.addRequiredID(LCSSAID);
505 AU.addRequired<LoopInfo>();
506 AU.addRequired<ScalarEvolution>();
507 AU.addRequired<DominatorTree>();
508 AU.addPreserved<LoopInfo>();
509 AU.addPreserved<DominatorTree>();
514 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
516 LLVMContext &C = V->getContext();
517 Type *VTy = VectorType::get(V->getType(), VF);
518 Type *I32 = IntegerType::getInt32Ty(C);
519 Constant *Zero = ConstantInt::get(I32, 0);
520 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
521 Value *UndefVal = UndefValue::get(VTy);
522 // Insert the value into a new vector.
523 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
524 // Broadcast the scalar into all locations in the vector.
525 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
527 // We are accessing the induction variable. Make sure to promote the
528 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
530 return getConsecutiveVector(Shuf);
534 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
535 assert(Val->getType()->isVectorTy() && "Must be a vector");
536 assert(Val->getType()->getScalarType()->isIntegerTy() &&
537 "Elem must be an integer");
539 Type *ITy = Val->getType()->getScalarType();
540 VectorType *Ty = cast<VectorType>(Val->getType());
541 unsigned VLen = Ty->getNumElements();
542 SmallVector<Constant*, 8> Indices;
544 // Create a vector of consecutive numbers from zero to VF.
545 for (unsigned i = 0; i < VLen; ++i)
546 Indices.push_back(ConstantInt::get(ITy, i));
548 // Add the consecutive indices to the vector value.
549 Constant *Cv = ConstantVector::get(Indices);
550 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
551 return Builder.CreateAdd(Val, Cv, "induction");
554 bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
555 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
557 // If this pointer is an induction variable, return it.
558 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
559 if (Phi && getInductionVars()->count(Phi))
562 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
566 unsigned NumOperands = Gep->getNumOperands();
567 Value *LastIndex = Gep->getOperand(NumOperands - 1);
569 // Check that all of the gep indices are uniform except for the last.
570 for (unsigned i = 0; i < NumOperands - 1; ++i)
571 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
574 // We can emit wide load/stores only of the last index is the induction
576 const SCEV *Last = SE->getSCEV(LastIndex);
577 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
578 const SCEV *Step = AR->getStepRecurrence(*SE);
580 // The memory is consecutive because the last index is consecutive
581 // and all other indices are loop invariant.
589 bool LoopVectorizationLegality::isUniform(Value *V) {
590 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
593 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
594 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
595 // If we saved a vectorized copy of V, use it.
596 Value *&MapEntry = WidenMap[V];
600 // Broadcast V and save the value for future uses.
601 Value *B = getBroadcastInstrs(V);
607 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
608 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
611 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
612 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
613 // Holds vector parameters or scalars, in case of uniform vals.
614 SmallVector<Value*, 8> Params;
616 // Find all of the vectorized parameters.
617 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
618 Value *SrcOp = Instr->getOperand(op);
620 // If we are accessing the old induction variable, use the new one.
621 if (SrcOp == OldInduction) {
622 Params.push_back(getVectorValue(Induction));
626 // Try using previously calculated values.
627 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
629 // If the src is an instruction that appeared earlier in the basic block
630 // then it should already be vectorized.
631 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
632 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
633 // The parameter is a vector value from earlier.
634 Params.push_back(WidenMap[SrcInst]);
636 // The parameter is a scalar from outside the loop. Maybe even a constant.
637 Params.push_back(SrcOp);
641 assert(Params.size() == Instr->getNumOperands() &&
642 "Invalid number of operands");
644 // Does this instruction return a value ?
645 bool IsVoidRetTy = Instr->getType()->isVoidTy();
646 Value *VecResults = 0;
648 // If we have a return value, create an empty vector. We place the scalarized
649 // instructions in this vector.
651 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
653 // For each scalar that we create:
654 for (unsigned i = 0; i < VF; ++i) {
655 Instruction *Cloned = Instr->clone();
657 Cloned->setName(Instr->getName() + ".cloned");
658 // Replace the operands of the cloned instrucions with extracted scalars.
659 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
660 Value *Op = Params[op];
661 // Param is a vector. Need to extract the right lane.
662 if (Op->getType()->isVectorTy())
663 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
664 Cloned->setOperand(op, Op);
667 // Place the cloned scalar in the new loop.
668 Builder.Insert(Cloned);
670 // If the original scalar returns a value we need to place it in a vector
671 // so that future users will be able to use it.
673 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
674 Builder.getInt32(i));
678 WidenMap[Instr] = VecResults;
682 SingleBlockLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
684 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
685 Legal->getRuntimePointerCheck();
687 if (!PtrRtCheck->Need)
690 Value *MemoryRuntimeCheck = 0;
691 unsigned NumPointers = PtrRtCheck->Pointers.size();
692 SmallVector<Value* , 2> Starts;
693 SmallVector<Value* , 2> Ends;
695 SCEVExpander Exp(*SE, "induction");
697 // Use this type for pointer arithmetic.
698 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
700 for (unsigned i=0; i < NumPointers; ++i) {
701 Value *Ptr = PtrRtCheck->Pointers[i];
702 const SCEV *Sc = SE->getSCEV(Ptr);
704 if (SE->isLoopInvariant(Sc, OrigLoop)) {
705 DEBUG(dbgs() << "LV1: Adding RT check for a loop invariant ptr:" <<
707 Starts.push_back(Ptr);
710 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
712 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i],
714 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
715 Starts.push_back(Start);
720 for (unsigned i = 0; i < NumPointers; ++i) {
721 for (unsigned j = i+1; j < NumPointers; ++j) {
722 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
723 Starts[i], Ends[j], "bound0", Loc);
724 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
725 Starts[j], Ends[i], "bound1", Loc);
726 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
727 "found.conflict", Loc);
728 if (MemoryRuntimeCheck) {
729 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
732 "conflict.rdx", Loc);
734 MemoryRuntimeCheck = IsConflict;
739 return MemoryRuntimeCheck;
743 SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
745 In this function we generate a new loop. The new loop will contain
746 the vectorized instructions while the old loop will continue to run the
749 [ ] <-- vector loop bypass.
752 | [ ] <-- vector pre header.
756 | [ ]_| <-- vector loop.
759 >[ ] <--- middle-block.
762 | [ ] <--- new preheader.
766 | [ ]_| <-- old scalar loop to handle remainder.
773 // Some loops have a single integer induction variable, while other loops
774 // don't. One example is c++ iterators that often have multiple pointer
775 // induction variables. In the code below we also support a case where we
776 // don't have a single induction variable.
777 OldInduction = Legal->getInduction();
778 Type *IdxTy = OldInduction ? OldInduction->getType() :
779 DL->getIntPtrType(SE->getContext());
781 // Find the loop boundaries.
782 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
783 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
785 // Get the total trip count from the count by adding 1.
786 ExitCount = SE->getAddExpr(ExitCount,
787 SE->getConstant(ExitCount->getType(), 1));
789 // This is the original scalar-loop preheader.
790 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
791 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
792 assert(ExitBlock && "Must have an exit block");
794 // The loop index does not have to start at Zero. Find the original start
795 // value from the induction PHI node. If we don't have an induction variable
796 // then we know that it starts at zero.
797 Value *StartIdx = OldInduction ?
798 OldInduction->getIncomingValueForBlock(BypassBlock):
799 ConstantInt::get(IdxTy, 0);
801 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
802 assert(BypassBlock && "Invalid loop structure");
804 BasicBlock *VectorPH =
805 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
806 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
809 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
811 BasicBlock *ScalarPH =
812 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
814 // Find the induction variable.
815 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
817 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
819 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
821 // Generate the induction variable.
822 Induction = Builder.CreatePHI(IdxTy, 2, "index");
823 Constant *Step = ConstantInt::get(IdxTy, VF);
825 // Expand the trip count and place the new instructions in the preheader.
826 // Notice that the pre-header does not change, only the loop body.
827 SCEVExpander Exp(*SE, "induction");
828 Instruction *Loc = BypassBlock->getTerminator();
830 // Count holds the overall loop count (N).
831 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(), Loc);
833 // We may need to extend the index in case there is a type mismatch.
834 // We know that the count starts at zero and does not overflow.
835 if (Count->getType() != IdxTy) {
836 // The exit count can be of pointer type. Convert it to the correct
838 if (ExitCount->getType()->isPointerTy())
839 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
841 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
844 // Add the start index to the loop count to get the new end index.
845 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
847 // Now we need to generate the expression for N - (N % VF), which is
848 // the part that the vectorized body will execute.
849 Constant *CIVF = ConstantInt::get(IdxTy, VF);
850 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
851 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
852 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
853 "end.idx.rnd.down", Loc);
855 // Now, compare the new count to zero. If it is zero skip the vector loop and
856 // jump to the scalar loop.
857 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
862 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal, Loc);
864 // If we are using memory runtime checks, include them in.
865 if (MemoryRuntimeCheck)
866 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
869 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
870 // Remove the old terminator.
871 Loc->eraseFromParent();
873 // We are going to resume the execution of the scalar loop.
874 // Go over all of the induction variables that we found and fix the
875 // PHIs that are left in the scalar version of the loop.
876 // The starting values of PHI nodes depend on the counter of the last
877 // iteration in the vectorized loop.
878 // If we come from a bypass edge then we need to start from the original start
881 // This variable saves the new starting index for the scalar loop.
882 PHINode *ResumeIndex = 0;
883 LoopVectorizationLegality::InductionList::iterator I, E;
884 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
885 for (I = List->begin(), E = List->end(); I != E; ++I) {
886 PHINode *OrigPhi = I->first;
887 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
888 MiddleBlock->getTerminator());
890 if (OrigPhi->getType()->isIntegerTy()) {
891 // Handle the integer induction counter:
892 assert(OrigPhi == OldInduction && "Unknown integer PHI");
893 // We know what the end value is.
894 EndValue = IdxEndRoundDown;
895 // We also know which PHI node holds it.
896 ResumeIndex = ResumeVal;
898 // For pointer induction variables, calculate the offset using
900 EndValue = GetElementPtrInst::Create(I->second, CountRoundDown,
902 BypassBlock->getTerminator());
905 // The new PHI merges the original incoming value, in case of a bypass,
906 // or the value at the end of the vectorized loop.
907 ResumeVal->addIncoming(I->second, BypassBlock);
908 ResumeVal->addIncoming(EndValue, VecBody);
910 // Fix the scalar body counter (PHI node).
911 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
912 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
915 // If we are generating a new induction variable then we also need to
916 // generate the code that calculates the exit value. This value is not
917 // simply the end of the counter because we may skip the vectorized body
918 // in case of a runtime check.
920 assert(!ResumeIndex && "Unexpected resume value found");
921 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
922 MiddleBlock->getTerminator());
923 ResumeIndex->addIncoming(StartIdx, BypassBlock);
924 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
927 // Make sure that we found the index where scalar loop needs to continue.
928 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
929 "Invalid resume Index");
931 // Add a check in the middle block to see if we have completed
932 // all of the iterations in the first vector loop.
933 // If (N - N%VF) == N, then we *don't* need to run the remainder.
934 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
935 ResumeIndex, "cmp.n",
936 MiddleBlock->getTerminator());
938 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
939 // Remove the old terminator.
940 MiddleBlock->getTerminator()->eraseFromParent();
942 // Create i+1 and fill the PHINode.
943 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
944 Induction->addIncoming(StartIdx, VectorPH);
945 Induction->addIncoming(NextIdx, VecBody);
946 // Create the compare.
947 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
948 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
950 // Now we have two terminators. Remove the old one from the block.
951 VecBody->getTerminator()->eraseFromParent();
953 // Get ready to start creating new instructions into the vectorized body.
954 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
956 // Register the new loop.
957 Loop* Lp = new Loop();
958 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
960 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
962 Loop *ParentLoop = OrigLoop->getParentLoop();
964 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
965 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
966 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
970 LoopVectorPreHeader = VectorPH;
971 LoopScalarPreHeader = ScalarPH;
972 LoopMiddleBlock = MiddleBlock;
973 LoopExitBlock = ExitBlock;
974 LoopVectorBody = VecBody;
975 LoopScalarBody = OldBasicBlock;
976 LoopBypassBlock = BypassBlock;
979 /// This function returns the identity element (or neutral element) for
982 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
984 case LoopVectorizationLegality::IntegerXor:
985 case LoopVectorizationLegality::IntegerAdd:
986 case LoopVectorizationLegality::IntegerOr:
987 // Adding, Xoring, Oring zero to a number does not change it.
989 case LoopVectorizationLegality::IntegerMult:
990 // Multiplying a number by 1 does not change it.
992 case LoopVectorizationLegality::IntegerAnd:
993 // AND-ing a number with an all-1 value does not change it.
996 llvm_unreachable("Unknown reduction kind");
1001 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1002 //===------------------------------------------------===//
1004 // Notice: any optimization or new instruction that go
1005 // into the code below should be also be implemented in
1008 //===------------------------------------------------===//
1009 typedef SmallVector<PHINode*, 4> PhiVector;
1010 BasicBlock &BB = *OrigLoop->getHeader();
1011 Constant *Zero = ConstantInt::get(
1012 IntegerType::getInt32Ty(BB.getContext()), 0);
1014 // In order to support reduction variables we need to be able to vectorize
1015 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1016 // steages. First, we create a new vector PHI node with no incoming edges.
1017 // We use this value when we vectorize all of the instructions that use the
1018 // PHI. Next, after all of the instructions in the block are complete we
1019 // add the new incoming edges to the PHI. At this point all of the
1020 // instructions in the basic block are vectorized, so we can use them to
1021 // construct the PHI.
1022 PhiVector RdxPHIsToFix;
1024 // For each instruction in the old loop.
1025 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1026 Instruction *Inst = it;
1028 switch (Inst->getOpcode()) {
1029 case Instruction::Br:
1030 // Nothing to do for PHIs and BR, since we already took care of the
1031 // loop control flow instructions.
1033 case Instruction::PHI:{
1034 PHINode* P = cast<PHINode>(Inst);
1035 // Handle reduction variables:
1036 if (Legal->getReductionVars()->count(P)) {
1037 // This is phase one of vectorizing PHIs.
1038 Type *VecTy = VectorType::get(Inst->getType(), VF);
1039 WidenMap[Inst] = PHINode::Create(VecTy, 2, "vec.phi",
1040 LoopVectorBody->getFirstInsertionPt());
1041 RdxPHIsToFix.push_back(P);
1045 // This PHINode must be an induction variable.
1046 // Make sure that we know about it.
1047 assert(Legal->getInductionVars()->count(P) &&
1048 "Not an induction variable");
1050 if (P->getType()->isIntegerTy()) {
1051 assert(P == OldInduction && "Unexpected PHI");
1052 WidenMap[Inst] = getBroadcastInstrs(Induction);
1056 // Handle pointer inductions:
1057 assert(P->getType()->isPointerTy() && "Unexpected type.");
1058 Value *StartIdx = OldInduction ?
1059 Legal->getInductionVars()->lookup(OldInduction) :
1060 ConstantInt::get(Induction->getType(), 0);
1062 // This is the pointer value coming into the loop.
1063 Value *StartPtr = Legal->getInductionVars()->lookup(P);
1065 // This is the normalized GEP that starts counting at zero.
1066 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1069 // This is the vector of results. Notice that we don't generate vector
1070 // geps because scalar geps result in better code.
1071 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1072 for (unsigned int i = 0; i < VF; ++i) {
1073 Constant *Idx = ConstantInt::get(Induction->getType(), i);
1074 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1075 Value *SclrGep = Builder.CreateGEP(StartPtr, GlobalIdx, "next.gep");
1076 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1077 Builder.getInt32(i),
1081 WidenMap[Inst] = VecVal;
1084 case Instruction::Add:
1085 case Instruction::FAdd:
1086 case Instruction::Sub:
1087 case Instruction::FSub:
1088 case Instruction::Mul:
1089 case Instruction::FMul:
1090 case Instruction::UDiv:
1091 case Instruction::SDiv:
1092 case Instruction::FDiv:
1093 case Instruction::URem:
1094 case Instruction::SRem:
1095 case Instruction::FRem:
1096 case Instruction::Shl:
1097 case Instruction::LShr:
1098 case Instruction::AShr:
1099 case Instruction::And:
1100 case Instruction::Or:
1101 case Instruction::Xor: {
1102 // Just widen binops.
1103 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1104 Value *A = getVectorValue(Inst->getOperand(0));
1105 Value *B = getVectorValue(Inst->getOperand(1));
1107 // Use this vector value for all users of the original instruction.
1108 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1111 // Update the NSW, NUW and Exact flags.
1112 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1113 if (isa<OverflowingBinaryOperator>(BinOp)) {
1114 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1115 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1117 if (isa<PossiblyExactOperator>(VecOp))
1118 VecOp->setIsExact(BinOp->isExact());
1121 case Instruction::Select: {
1123 // If the selector is loop invariant we can create a select
1124 // instruction with a scalar condition. Otherwise, use vector-select.
1125 Value *Cond = Inst->getOperand(0);
1126 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1128 // The condition can be loop invariant but still defined inside the
1129 // loop. This means that we can't just use the original 'cond' value.
1130 // We have to take the 'vectorized' value and pick the first lane.
1131 // Instcombine will make this a no-op.
1132 Cond = getVectorValue(Cond);
1134 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1136 Value *Op0 = getVectorValue(Inst->getOperand(1));
1137 Value *Op1 = getVectorValue(Inst->getOperand(2));
1138 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
1142 case Instruction::ICmp:
1143 case Instruction::FCmp: {
1144 // Widen compares. Generate vector compares.
1145 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
1146 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
1147 Value *A = getVectorValue(Inst->getOperand(0));
1148 Value *B = getVectorValue(Inst->getOperand(1));
1150 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1152 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1156 case Instruction::Store: {
1157 // Attempt to issue a wide store.
1158 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1159 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1160 Value *Ptr = SI->getPointerOperand();
1161 unsigned Alignment = SI->getAlignment();
1163 assert(!Legal->isUniform(Ptr) &&
1164 "We do not allow storing to uniform addresses");
1166 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1168 // This store does not use GEPs.
1169 if (!Legal->isConsecutivePtr(Ptr)) {
1170 scalarizeInstruction(Inst);
1175 // The last index does not have to be the induction. It can be
1176 // consecutive and be a function of the index. For example A[I+1];
1177 unsigned NumOperands = Gep->getNumOperands();
1178 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1179 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1181 // Create the new GEP with the new induction variable.
1182 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1183 Gep2->setOperand(NumOperands - 1, LastIndex);
1184 Ptr = Builder.Insert(Gep2);
1186 // Use the induction element ptr.
1187 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1188 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1190 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1191 Value *Val = getVectorValue(SI->getValueOperand());
1192 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1195 case Instruction::Load: {
1196 // Attempt to issue a wide load.
1197 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1198 Type *RetTy = VectorType::get(LI->getType(), VF);
1199 Value *Ptr = LI->getPointerOperand();
1200 unsigned Alignment = LI->getAlignment();
1201 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1203 // If the pointer is loop invariant or if it is non consecutive,
1204 // scalarize the load.
1205 bool Con = Legal->isConsecutivePtr(Ptr);
1206 if (Legal->isUniform(Ptr) || !Con) {
1207 scalarizeInstruction(Inst);
1212 // The last index does not have to be the induction. It can be
1213 // consecutive and be a function of the index. For example A[I+1];
1214 unsigned NumOperands = Gep->getNumOperands();
1215 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1216 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1218 // Create the new GEP with the new induction variable.
1219 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1220 Gep2->setOperand(NumOperands - 1, LastIndex);
1221 Ptr = Builder.Insert(Gep2);
1223 // Use the induction element ptr.
1224 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1225 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1228 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1229 LI = Builder.CreateLoad(Ptr);
1230 LI->setAlignment(Alignment);
1231 // Use this vector value for all users of the load.
1232 WidenMap[Inst] = LI;
1235 case Instruction::ZExt:
1236 case Instruction::SExt:
1237 case Instruction::FPToUI:
1238 case Instruction::FPToSI:
1239 case Instruction::FPExt:
1240 case Instruction::PtrToInt:
1241 case Instruction::IntToPtr:
1242 case Instruction::SIToFP:
1243 case Instruction::UIToFP:
1244 case Instruction::Trunc:
1245 case Instruction::FPTrunc:
1246 case Instruction::BitCast: {
1247 /// Vectorize bitcasts.
1248 CastInst *CI = dyn_cast<CastInst>(Inst);
1249 Value *A = getVectorValue(Inst->getOperand(0));
1250 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1251 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1256 /// All other instructions are unsupported. Scalarize them.
1257 scalarizeInstruction(Inst);
1260 }// end of for_each instr.
1262 // At this point every instruction in the original loop is widended to
1263 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1264 // that we vectorized. The PHI nodes are currently empty because we did
1265 // not want to introduce cycles. Notice that the remaining PHI nodes
1266 // that we need to fix are reduction variables.
1268 // Create the 'reduced' values for each of the induction vars.
1269 // The reduced values are the vector values that we scalarize and combine
1270 // after the loop is finished.
1271 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1273 PHINode *RdxPhi = *it;
1274 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1275 assert(RdxPhi && "Unable to recover vectorized PHI");
1277 // Find the reduction variable descriptor.
1278 assert(Legal->getReductionVars()->count(RdxPhi) &&
1279 "Unable to find the reduction variable");
1280 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1281 (*Legal->getReductionVars())[RdxPhi];
1283 // We need to generate a reduction vector from the incoming scalar.
1284 // To do so, we need to generate the 'identity' vector and overide
1285 // one of the elements with the incoming scalar reduction. We need
1286 // to do it in the vector-loop preheader.
1287 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1289 // This is the vector-clone of the value that leaves the loop.
1290 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1291 Type *VecTy = VectorExit->getType();
1293 // Find the reduction identity variable. Zero for addition, or, xor,
1294 // one for multiplication, -1 for And.
1295 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1296 VecTy->getScalarType());
1298 // This vector is the Identity vector where the first element is the
1299 // incoming scalar reduction.
1300 Value *VectorStart = Builder.CreateInsertElement(Identity,
1301 RdxDesc.StartValue, Zero);
1303 // Fix the vector-loop phi.
1304 // We created the induction variable so we know that the
1305 // preheader is the first entry.
1306 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1308 // Reductions do not have to start at zero. They can start with
1309 // any loop invariant values.
1310 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1311 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1312 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1313 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1315 // Before each round, move the insertion point right between
1316 // the PHIs and the values we are going to write.
1317 // This allows us to write both PHINodes and the extractelement
1319 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1321 // This PHINode contains the vectorized reduction variable, or
1322 // the initial value vector, if we bypass the vector loop.
1323 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1324 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1325 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1327 // Extract the first scalar.
1329 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1330 // Extract and reduce the remaining vector elements.
1331 for (unsigned i=1; i < VF; ++i) {
1333 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1334 switch (RdxDesc.Kind) {
1335 case LoopVectorizationLegality::IntegerAdd:
1336 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1338 case LoopVectorizationLegality::IntegerMult:
1339 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1341 case LoopVectorizationLegality::IntegerOr:
1342 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1344 case LoopVectorizationLegality::IntegerAnd:
1345 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1347 case LoopVectorizationLegality::IntegerXor:
1348 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1351 llvm_unreachable("Unknown reduction operation");
1355 // Now, we need to fix the users of the reduction variable
1356 // inside and outside of the scalar remainder loop.
1357 // We know that the loop is in LCSSA form. We need to update the
1358 // PHI nodes in the exit blocks.
1359 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1360 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1361 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1362 if (!LCSSAPhi) continue;
1364 // All PHINodes need to have a single entry edge, or two if
1365 // we already fixed them.
1366 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1368 // We found our reduction value exit-PHI. Update it with the
1369 // incoming bypass edge.
1370 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1371 // Add an edge coming from the bypass.
1372 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1375 }// end of the LCSSA phi scan.
1377 // Fix the scalar loop reduction variable with the incoming reduction sum
1378 // from the vector body and from the backedge value.
1379 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1380 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1381 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1382 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1383 }// end of for each redux variable.
1386 void SingleBlockLoopVectorizer::updateAnalysis() {
1387 // The original basic block.
1388 SE->forgetLoop(OrigLoop);
1390 // Update the dominator tree information.
1391 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1392 "Entry does not dominate exit.");
1394 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1395 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1396 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1397 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1398 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1399 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1401 DEBUG(DT->verifyAnalysis());
1404 bool LoopVectorizationLegality::canVectorize() {
1405 if (!TheLoop->getLoopPreheader()) {
1406 assert(false && "No preheader!!");
1407 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1411 // We can only vectorize single basic block loops.
1412 unsigned NumBlocks = TheLoop->getNumBlocks();
1413 if (NumBlocks != 1) {
1414 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1418 // We need to have a loop header.
1419 BasicBlock *BB = TheLoop->getHeader();
1420 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1422 // ScalarEvolution needs to be able to find the exit count.
1423 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1424 if (ExitCount == SE->getCouldNotCompute()) {
1425 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1429 // Do not loop-vectorize loops with a tiny trip count.
1430 unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB);
1431 if (TC > 0u && TC < TinyTripCountThreshold) {
1432 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1433 "This loop is not worth vectorizing.\n");
1437 // Go over each instruction and look at memory deps.
1438 if (!canVectorizeBlock(*BB)) {
1439 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1443 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1444 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1447 // Okay! We can vectorize. At this point we don't have any other mem analysis
1448 // which may limit our maximum vectorization factor, so just return true with
1453 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1455 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1457 // Scan the instructions in the block and look for hazards.
1458 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1459 Instruction *I = it;
1461 if (PHINode *Phi = dyn_cast<PHINode>(I)) {
1462 // This should not happen because the loop should be normalized.
1463 if (Phi->getNumIncomingValues() != 2) {
1464 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1468 // This is the value coming from the preheader.
1469 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1471 // We only look at integer and pointer phi nodes.
1472 if (Phi->getType()->isPointerTy() && isInductionVariable(Phi)) {
1473 DEBUG(dbgs() << "LV: Found a pointer induction variable.\n");
1474 Inductions[Phi] = StartValue;
1476 } else if (!Phi->getType()->isIntegerTy()) {
1477 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1481 // Handle integer PHIs:
1482 if (isInductionVariable(Phi)) {
1484 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1487 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1489 Inductions[Phi] = StartValue;
1492 if (AddReductionVar(Phi, IntegerAdd)) {
1493 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1496 if (AddReductionVar(Phi, IntegerMult)) {
1497 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1500 if (AddReductionVar(Phi, IntegerOr)) {
1501 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1504 if (AddReductionVar(Phi, IntegerAnd)) {
1505 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1508 if (AddReductionVar(Phi, IntegerXor)) {
1509 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1513 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1515 }// end of PHI handling
1517 // We still don't handle functions.
1518 CallInst *CI = dyn_cast<CallInst>(I);
1520 DEBUG(dbgs() << "LV: Found a call site.\n");
1524 // We do not re-vectorize vectors.
1525 if (!VectorType::isValidElementType(I->getType()) &&
1526 !I->getType()->isVoidTy()) {
1527 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1531 // Reduction instructions are allowed to have exit users.
1532 // All other instructions must not have external users.
1533 if (!AllowedExit.count(I))
1534 //Check that all of the users of the loop are inside the BB.
1535 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1537 Instruction *U = cast<Instruction>(*it);
1538 // This user may be a reduction exit value.
1539 BasicBlock *Parent = U->getParent();
1540 if (Parent != &BB) {
1541 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1548 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
1549 assert(getInductionVars()->size() && "No induction variables");
1552 // Don't vectorize if the memory dependencies do not allow vectorization.
1553 if (!canVectorizeMemory(BB))
1556 // We now know that the loop is vectorizable!
1557 // Collect variables that will remain uniform after vectorization.
1558 std::vector<Value*> Worklist;
1560 // Start with the conditional branch and walk up the block.
1561 Worklist.push_back(BB.getTerminator()->getOperand(0));
1563 while (Worklist.size()) {
1564 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1565 Worklist.pop_back();
1567 // Look at instructions inside this block. Stop when reaching PHI nodes.
1568 if (!I || I->getParent() != &BB || isa<PHINode>(I))
1571 // This is a known uniform.
1574 // Insert all operands.
1575 for (int i=0, Op = I->getNumOperands(); i < Op; ++i) {
1576 Worklist.push_back(I->getOperand(i));
1583 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1584 typedef SmallVector<Value*, 16> ValueVector;
1585 typedef SmallPtrSet<Value*, 16> ValueSet;
1586 // Holds the Load and Store *instructions*.
1589 PtrRtCheck.Pointers.clear();
1590 PtrRtCheck.Need = false;
1592 // Scan the BB and collect legal loads and stores.
1593 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1594 Instruction *I = it;
1596 // If this is a load, save it. If this instruction can read from memory
1597 // but is not a load, then we quit. Notice that we don't handle function
1598 // calls that read or write.
1599 if (I->mayReadFromMemory()) {
1600 LoadInst *Ld = dyn_cast<LoadInst>(I);
1601 if (!Ld) return false;
1602 if (!Ld->isSimple()) {
1603 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1606 Loads.push_back(Ld);
1610 // Save store instructions. Abort if other instructions write to memory.
1611 if (I->mayWriteToMemory()) {
1612 StoreInst *St = dyn_cast<StoreInst>(I);
1613 if (!St) return false;
1614 if (!St->isSimple()) {
1615 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1618 Stores.push_back(St);
1622 // Now we have two lists that hold the loads and the stores.
1623 // Next, we find the pointers that they use.
1625 // Check if we see any stores. If there are no stores, then we don't
1626 // care if the pointers are *restrict*.
1627 if (!Stores.size()) {
1628 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1632 // Holds the read and read-write *pointers* that we find.
1634 ValueVector ReadWrites;
1636 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1637 // multiple times on the same object. If the ptr is accessed twice, once
1638 // for read and once for write, it will only appear once (on the write
1639 // list). This is okay, since we are going to check for conflicts between
1640 // writes and between reads and writes, but not between reads and reads.
1643 ValueVector::iterator I, IE;
1644 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1645 StoreInst *ST = dyn_cast<StoreInst>(*I);
1646 assert(ST && "Bad StoreInst");
1647 Value* Ptr = ST->getPointerOperand();
1649 if (isUniform(Ptr)) {
1650 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1654 // If we did *not* see this pointer before, insert it to
1655 // the read-write list. At this phase it is only a 'write' list.
1656 if (Seen.insert(Ptr))
1657 ReadWrites.push_back(Ptr);
1660 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1661 LoadInst *LD = dyn_cast<LoadInst>(*I);
1662 assert(LD && "Bad LoadInst");
1663 Value* Ptr = LD->getPointerOperand();
1664 // If we did *not* see this pointer before, insert it to the
1665 // read list. If we *did* see it before, then it is already in
1666 // the read-write list. This allows us to vectorize expressions
1667 // such as A[i] += x; Because the address of A[i] is a read-write
1668 // pointer. This only works if the index of A[i] is consecutive.
1669 // If the address of i is unknown (for example A[B[i]]) then we may
1670 // read a few words, modify, and write a few words, and some of the
1671 // words may be written to the same address.
1672 if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr))
1673 Reads.push_back(Ptr);
1676 // If we write (or read-write) to a single destination and there are no
1677 // other reads in this loop then is it safe to vectorize.
1678 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1679 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1683 // Find pointers with computable bounds. We are going to use this information
1684 // to place a runtime bound check.
1686 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1687 if (hasComputableBounds(*I)) {
1688 PtrRtCheck.insert(SE, TheLoop, *I);
1689 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1694 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1695 if (hasComputableBounds(*I)) {
1696 PtrRtCheck.insert(SE, TheLoop, *I);
1697 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1703 // Check that we did not collect too many pointers or found a
1704 // unsizeable pointer.
1705 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1710 PtrRtCheck.Need = RT;
1713 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1716 // Now that the pointers are in two lists (Reads and ReadWrites), we
1717 // can check that there are no conflicts between each of the writes and
1718 // between the writes to the reads.
1719 ValueSet WriteObjects;
1720 ValueVector TempObjects;
1722 // Check that the read-writes do not conflict with other read-write
1724 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1725 GetUnderlyingObjects(*I, TempObjects, DL);
1726 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1728 if (!isIdentifiedObject(*it)) {
1729 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1732 if (!WriteObjects.insert(*it)) {
1733 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1738 TempObjects.clear();
1741 /// Check that the reads don't conflict with the read-writes.
1742 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1743 GetUnderlyingObjects(*I, TempObjects, DL);
1744 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1746 if (!isIdentifiedObject(*it)) {
1747 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1750 if (WriteObjects.count(*it)) {
1751 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1756 TempObjects.clear();
1759 // It is safe to vectorize and we don't need any runtime checks.
1760 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1765 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1766 ReductionKind Kind) {
1767 if (Phi->getNumIncomingValues() != 2)
1770 // Find the possible incoming reduction variable.
1771 BasicBlock *BB = Phi->getParent();
1772 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1773 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1774 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1776 // ExitInstruction is the single value which is used outside the loop.
1777 // We only allow for a single reduction value to be used outside the loop.
1778 // This includes users of the reduction, variables (which form a cycle
1779 // which ends in the phi node).
1780 Instruction *ExitInstruction = 0;
1782 // Iter is our iterator. We start with the PHI node and scan for all of the
1783 // users of this instruction. All users must be instructions which can be
1784 // used as reduction variables (such as ADD). We may have a single
1785 // out-of-block user. They cycle must end with the original PHI.
1786 // Also, we can't have multiple block-local users.
1787 Instruction *Iter = Phi;
1789 // Any reduction instr must be of one of the allowed kinds.
1790 if (!isReductionInstr(Iter, Kind))
1793 // Did we found a user inside this block ?
1794 bool FoundInBlockUser = false;
1795 // Did we reach the initial PHI node ?
1796 bool FoundStartPHI = false;
1798 // If the instruction has no users then this is a broken
1799 // chain and can't be a reduction variable.
1800 if (Iter->use_empty())
1803 // For each of the *users* of iter.
1804 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1806 Instruction *U = cast<Instruction>(*it);
1807 // We already know that the PHI is a user.
1809 FoundStartPHI = true;
1812 // Check if we found the exit user.
1813 BasicBlock *Parent = U->getParent();
1815 // We must have a single exit instruction.
1816 if (ExitInstruction != 0)
1818 ExitInstruction = Iter;
1820 // We can't have multiple inside users.
1821 if (FoundInBlockUser)
1823 FoundInBlockUser = true;
1827 // We found a reduction var if we have reached the original
1828 // phi node and we only have a single instruction with out-of-loop
1830 if (FoundStartPHI && ExitInstruction) {
1831 // This instruction is allowed to have out-of-loop users.
1832 AllowedExit.insert(ExitInstruction);
1834 // Save the description of this reduction variable.
1835 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1836 Reductions[Phi] = RD;
1843 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1844 ReductionKind Kind) {
1845 switch (I->getOpcode()) {
1848 case Instruction::PHI:
1851 case Instruction::Add:
1852 case Instruction::Sub:
1853 return Kind == IntegerAdd;
1854 case Instruction::Mul:
1855 return Kind == IntegerMult;
1856 case Instruction::And:
1857 return Kind == IntegerAnd;
1858 case Instruction::Or:
1859 return Kind == IntegerOr;
1860 case Instruction::Xor:
1861 return Kind == IntegerXor;
1865 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1866 Type *PhiTy = Phi->getType();
1867 // We only handle integer and pointer inductions variables.
1868 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1871 // Check that the PHI is consecutive and starts at zero.
1872 const SCEV *PhiScev = SE->getSCEV(Phi);
1873 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1875 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1878 const SCEV *Step = AR->getStepRecurrence(*SE);
1880 // Integer inductions need to have a stride of one.
1881 if (PhiTy->isIntegerTy())
1882 return Step->isOne();
1884 // Calculate the pointer stride and check if it is consecutive.
1885 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1886 if (!C) return false;
1888 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1889 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1890 return (C->getValue()->equalsInt(Size));
1893 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1894 const SCEV *PhiScev = SE->getSCEV(Ptr);
1895 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1899 return AR->isAffine();
1903 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1905 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1909 float Cost = expectedCost(1);
1911 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1912 for (unsigned i=2; i <= VF; i*=2) {
1913 // Notice that the vector loop needs to be executed less times, so
1914 // we need to divide the cost of the vector loops by the width of
1915 // the vector elements.
1916 float VectorCost = expectedCost(i) / (float)i;
1917 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1918 (int)VectorCost << ".\n");
1919 if (VectorCost < Cost) {
1925 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1929 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1930 // We can only estimate the cost of single basic block loops.
1931 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1933 BasicBlock *BB = TheLoop->getHeader();
1936 // For each instruction in the old loop.
1937 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1938 Instruction *Inst = it;
1939 unsigned C = getInstructionCost(Inst, VF);
1941 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1942 " For instruction: "<< *Inst << "\n");
1949 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1950 assert(VTTI && "Invalid vector target transformation info");
1952 // If we know that this instruction will remain uniform, check the cost of
1953 // the scalar version.
1954 if (Legal->isUniformAfterVectorization(I))
1957 Type *RetTy = I->getType();
1958 Type *VectorTy = ToVectorTy(RetTy, VF);
1961 // TODO: We need to estimate the cost of intrinsic calls.
1962 switch (I->getOpcode()) {
1963 case Instruction::GetElementPtr:
1964 // We mark this instruction as zero-cost because scalar GEPs are usually
1965 // lowered to the intruction addressing mode. At the moment we don't
1966 // generate vector geps.
1968 case Instruction::Br: {
1969 return VTTI->getCFInstrCost(I->getOpcode());
1971 case Instruction::PHI:
1973 case Instruction::Add:
1974 case Instruction::FAdd:
1975 case Instruction::Sub:
1976 case Instruction::FSub:
1977 case Instruction::Mul:
1978 case Instruction::FMul:
1979 case Instruction::UDiv:
1980 case Instruction::SDiv:
1981 case Instruction::FDiv:
1982 case Instruction::URem:
1983 case Instruction::SRem:
1984 case Instruction::FRem:
1985 case Instruction::Shl:
1986 case Instruction::LShr:
1987 case Instruction::AShr:
1988 case Instruction::And:
1989 case Instruction::Or:
1990 case Instruction::Xor: {
1991 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
1993 case Instruction::Select: {
1994 SelectInst *SI = cast<SelectInst>(I);
1995 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1996 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1997 Type *CondTy = SI->getCondition()->getType();
1999 CondTy = VectorType::get(CondTy, VF);
2001 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
2003 case Instruction::ICmp:
2004 case Instruction::FCmp: {
2005 Type *ValTy = I->getOperand(0)->getType();
2006 VectorTy = ToVectorTy(ValTy, VF);
2007 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
2009 case Instruction::Store: {
2010 StoreInst *SI = cast<StoreInst>(I);
2011 Type *ValTy = SI->getValueOperand()->getType();
2012 VectorTy = ToVectorTy(ValTy, VF);
2015 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
2016 SI->getAlignment(), SI->getPointerAddressSpace());
2018 // Scalarized stores.
2019 if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
2021 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2023 // The cost of extracting from the value vector.
2024 Cost += VF * (ExtCost);
2025 // The cost of the scalar stores.
2026 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2027 ValTy->getScalarType(),
2029 SI->getPointerAddressSpace());
2034 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
2035 SI->getPointerAddressSpace());
2037 case Instruction::Load: {
2038 LoadInst *LI = cast<LoadInst>(I);
2041 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
2043 LI->getPointerAddressSpace());
2045 // Scalarized loads.
2046 if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
2048 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
2049 // The cost of inserting the loaded value into the result vector.
2050 Cost += VF * (InCost);
2051 // The cost of the scalar stores.
2052 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2053 RetTy->getScalarType(),
2055 LI->getPointerAddressSpace());
2060 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2061 LI->getPointerAddressSpace());
2063 case Instruction::ZExt:
2064 case Instruction::SExt:
2065 case Instruction::FPToUI:
2066 case Instruction::FPToSI:
2067 case Instruction::FPExt:
2068 case Instruction::PtrToInt:
2069 case Instruction::IntToPtr:
2070 case Instruction::SIToFP:
2071 case Instruction::UIToFP:
2072 case Instruction::Trunc:
2073 case Instruction::FPTrunc:
2074 case Instruction::BitCast: {
2075 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2076 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
2079 // We are scalarizing the instruction. Return the cost of the scalar
2080 // instruction, plus the cost of insert and extract into vector
2081 // elements, times the vector width.
2084 bool IsVoid = RetTy->isVoidTy();
2086 unsigned InsCost = (IsVoid ? 0 :
2087 VTTI->getInstrCost(Instruction::InsertElement,
2090 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2093 // The cost of inserting the results plus extracting each one of the
2095 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
2097 // The cost of executing VF copies of the scalar instruction.
2098 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
2104 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2105 if (Scalar->isVoidTy() || VF == 1)
2107 return VectorType::get(Scalar, VF);
2112 char LoopVectorize::ID = 0;
2113 static const char lv_name[] = "Loop Vectorization";
2114 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2115 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2116 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2117 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2118 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2121 Pass *createLoopVectorizePass() {
2122 return new LoopVectorize();