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. InnerLoopVectorizer - 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/Transforms/Vectorize.h"
47 #include "llvm/ADT/SmallVector.h"
48 #include "llvm/ADT/StringExtras.h"
49 #include "llvm/Analysis/AliasAnalysis.h"
50 #include "llvm/Analysis/AliasSetTracker.h"
51 #include "llvm/Analysis/Dominators.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopPass.h"
54 #include "llvm/Analysis/ScalarEvolution.h"
55 #include "llvm/Analysis/ScalarEvolutionExpander.h"
56 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
57 #include "llvm/Analysis/ValueTracking.h"
58 #include "llvm/Analysis/Verifier.h"
59 #include "llvm/Constants.h"
60 #include "llvm/DataLayout.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Function.h"
63 #include "llvm/Instructions.h"
64 #include "llvm/LLVMContext.h"
65 #include "llvm/Module.h"
66 #include "llvm/Pass.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/raw_ostream.h"
70 #include "llvm/TargetTransformInfo.h"
71 #include "llvm/Transforms/Scalar.h"
72 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
73 #include "llvm/Transforms/Utils/Local.h"
74 #include "llvm/Type.h"
75 #include "llvm/Value.h"
79 static cl::opt<unsigned>
80 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
81 cl::desc("Set the default vectorization width. Zero is autoselect."));
84 EnableIfConversion("enable-if-conversion", cl::init(false), cl::Hidden,
85 cl::desc("Enable if-conversion during vectorization."));
87 /// We don't vectorize loops with a known constant trip count below this number.
88 const unsigned TinyTripCountThreshold = 16;
90 /// When performing a runtime memory check, do not check more than this
91 /// number of pointers. Notice that the check is quadratic!
92 const unsigned RuntimeMemoryCheckThreshold = 2;
94 /// This is the highest vector width that we try to generate.
95 const unsigned MaxVectorSize = 8;
99 // Forward declarations.
100 class LoopVectorizationLegality;
101 class LoopVectorizationCostModel;
103 /// InnerLoopVectorizer vectorizes loops which contain only one basic
104 /// block to a specified vectorization factor (VF).
105 /// This class performs the widening of scalars into vectors, or multiple
106 /// scalars. This class also implements the following features:
107 /// * It inserts an epilogue loop for handling loops that don't have iteration
108 /// counts that are known to be a multiple of the vectorization factor.
109 /// * It handles the code generation for reduction variables.
110 /// * Scalarization (implementation using scalars) of un-vectorizable
112 /// InnerLoopVectorizer does not perform any vectorization-legality
113 /// checks, and relies on the caller to check for the different legality
114 /// aspects. The InnerLoopVectorizer relies on the
115 /// LoopVectorizationLegality class to provide information about the induction
116 /// and reduction variables that were found to a given vectorization factor.
117 class InnerLoopVectorizer {
120 InnerLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
121 DominatorTree *Dt, DataLayout *Dl, unsigned VecWidth):
122 OrigLoop(Orig), SE(Se), LI(Li), DT(Dt), DL(Dl), VF(VecWidth),
123 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
125 // Perform the actual loop widening (vectorization).
126 void vectorize(LoopVectorizationLegality *Legal) {
127 // Create a new empty loop. Unlink the old loop and connect the new one.
128 createEmptyLoop(Legal);
129 // Widen each instruction in the old loop to a new one in the new loop.
130 // Use the Legality module to find the induction and reduction variables.
131 vectorizeLoop(Legal);
132 // Register the new loop and update the analysis passes.
137 /// Add code that checks at runtime if the accessed arrays overlap.
138 /// Returns the comperator value or NULL if no check is needed.
139 Value *addRuntimeCheck(LoopVectorizationLegality *Legal,
141 /// Create an empty loop, based on the loop ranges of the old loop.
142 void createEmptyLoop(LoopVectorizationLegality *Legal);
143 /// Copy and widen the instructions from the old loop.
144 void vectorizeLoop(LoopVectorizationLegality *Legal);
145 /// Insert the new loop to the loop hierarchy and pass manager
146 /// and update the analysis passes.
147 void updateAnalysis();
149 /// This instruction is un-vectorizable. Implement it as a sequence
151 void scalarizeInstruction(Instruction *Instr);
153 /// Create a broadcast instruction. This method generates a broadcast
154 /// instruction (shuffle) for loop invariant values and for the induction
155 /// value. If this is the induction variable then we extend it to N, N+1, ...
156 /// this is needed because each iteration in the loop corresponds to a SIMD
158 Value *getBroadcastInstrs(Value *V);
160 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
161 /// for each element in the vector. Starting from zero.
162 Value *getConsecutiveVector(Value* Val);
164 /// When we go over instructions in the basic block we rely on previous
165 /// values within the current basic block or on loop invariant values.
166 /// When we widen (vectorize) values we place them in the map. If the values
167 /// are not within the map, they have to be loop invariant, so we simply
168 /// broadcast them into a vector.
169 Value *getVectorValue(Value *V);
171 /// Get a uniform vector of constant integers. We use this to get
172 /// vectors of ones and zeros for the reduction code.
173 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
175 typedef DenseMap<Value*, Value*> ValueMap;
177 /// The original loop.
179 // Scev analysis to use.
187 // The vectorization factor to use.
190 // The builder that we use
193 // --- Vectorization state ---
195 /// The vector-loop preheader.
196 BasicBlock *LoopVectorPreHeader;
197 /// The scalar-loop preheader.
198 BasicBlock *LoopScalarPreHeader;
199 /// Middle Block between the vector and the scalar.
200 BasicBlock *LoopMiddleBlock;
201 ///The ExitBlock of the scalar loop.
202 BasicBlock *LoopExitBlock;
203 ///The vector loop body.
204 BasicBlock *LoopVectorBody;
205 ///The scalar loop body.
206 BasicBlock *LoopScalarBody;
207 ///The first bypass block.
208 BasicBlock *LoopBypassBlock;
210 /// The new Induction variable which was added to the new block.
212 /// The induction variable of the old basic block.
213 PHINode *OldInduction;
214 // Maps scalars to widened vectors.
218 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
219 /// to what vectorization factor.
220 /// This class does not look at the profitability of vectorization, only the
221 /// legality. This class has two main kinds of checks:
222 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
223 /// will change the order of memory accesses in a way that will change the
224 /// correctness of the program.
225 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
226 /// checks for a number of different conditions, such as the availability of a
227 /// single induction variable, that all types are supported and vectorize-able,
228 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
229 /// This class is also used by InnerLoopVectorizer for identifying
230 /// induction variable and the different reduction variables.
231 class LoopVectorizationLegality {
233 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl,
235 TheLoop(Lp), SE(Se), DL(Dl), DT(Dt), Induction(0) { }
237 /// This represents the kinds of reductions that we support.
239 NoReduction, /// Not a reduction.
240 IntegerAdd, /// Sum of numbers.
241 IntegerMult, /// Product of numbers.
242 IntegerOr, /// Bitwise or logical OR of numbers.
243 IntegerAnd, /// Bitwise or logical AND of numbers.
244 IntegerXor /// Bitwise or logical XOR of numbers.
247 /// This POD struct holds information about reduction variables.
248 struct ReductionDescriptor {
250 ReductionDescriptor():
251 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
254 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
255 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
257 // The starting value of the reduction.
258 // It does not have to be zero!
260 // The instruction who's value is used outside the loop.
261 Instruction *LoopExitInstr;
262 // The kind of the reduction.
266 // This POD struct holds information about the memory runtime legality
267 // check that a group of pointers do not overlap.
268 struct RuntimePointerCheck {
269 RuntimePointerCheck(): Need(false) {}
271 /// Reset the state of the pointer runtime information.
279 /// Insert a pointer and calculate the start and end SCEVs.
280 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr) {
281 const SCEV *Sc = SE->getSCEV(Ptr);
282 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
283 assert(AR && "Invalid addrec expression");
284 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
285 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
286 Pointers.push_back(Ptr);
287 Starts.push_back(AR->getStart());
288 Ends.push_back(ScEnd);
291 /// This flag indicates if we need to add the runtime check.
293 /// Holds the pointers that we need to check.
294 SmallVector<Value*, 2> Pointers;
295 /// Holds the pointer value at the beginning of the loop.
296 SmallVector<const SCEV*, 2> Starts;
297 /// Holds the pointer value at the end of the loop.
298 SmallVector<const SCEV*, 2> Ends;
301 /// ReductionList contains the reduction descriptors for all
302 /// of the reductions that were found in the loop.
303 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
305 /// InductionList saves induction variables and maps them to the initial
306 /// value entring the loop.
307 typedef DenseMap<PHINode*, Value*> InductionList;
309 /// Returns true if it is legal to vectorize this loop.
310 /// This does not mean that it is profitable to vectorize this
311 /// loop, only that it is legal to do so.
314 /// Returns the Induction variable.
315 PHINode *getInduction() {return Induction;}
317 /// Returns the reduction variables found in the loop.
318 ReductionList *getReductionVars() { return &Reductions; }
320 /// Returns the induction variables found in the loop.
321 InductionList *getInductionVars() { return &Inductions; }
323 /// Return true if the block BB needs to be predicated in order for the loop
324 /// to be vectorized.
325 bool blockNeedsPredication(BasicBlock *BB);
327 /// Check if this pointer is consecutive when vectorizing. This happens
328 /// when the last index of the GEP is the induction variable, or that the
329 /// pointer itself is an induction variable.
330 /// This check allows us to vectorize A[idx] into a wide load/store.
331 bool isConsecutivePtr(Value *Ptr);
333 /// Returns true if the value V is uniform within the loop.
334 bool isUniform(Value *V);
336 /// Returns true if this instruction will remain scalar after vectorization.
337 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
339 /// Returns the information that we collected about runtime memory check.
340 RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; }
342 /// Check if a single basic block loop is vectorizable.
343 /// At this point we know that this is a loop with a constant trip count
344 /// and we only need to check individual instructions.
345 bool canVectorizeInstrs();
347 /// When we vectorize loops we may change the order in which
348 /// we read and write from memory. This method checks if it is
349 /// legal to vectorize the code, considering only memory constrains.
350 /// Returns true if the loop is vectorizable
351 bool canVectorizeMemory();
353 /// Return true if we can vectorize this loop using the IF-conversion
355 bool canVectorizeWithIfConvert();
357 /// Collect the variables that need to stay uniform after vectorization.
358 void collectLoopUniforms();
360 /// Return true if all of the instructions in the block can be speculatively
362 bool blockCanBePredicated(BasicBlock *BB);
364 /// Returns True, if 'Phi' is the kind of reduction variable for type
365 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
366 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
367 /// Returns true if the instruction I can be a reduction variable of type
369 bool isReductionInstr(Instruction *I, ReductionKind Kind);
370 /// Returns True, if 'Phi' is an induction variable.
371 bool isInductionVariable(PHINode *Phi);
372 /// Return true if can compute the address bounds of Ptr within the loop.
373 bool hasComputableBounds(Value *Ptr);
375 /// The loop that we evaluate.
379 /// DataLayout analysis.
384 // --- vectorization state --- //
386 /// Holds the integer induction variable. This is the counter of the
389 /// Holds the reduction variables.
390 ReductionList Reductions;
391 /// Holds all of the induction variables that we found in the loop.
392 /// Notice that inductions don't need to start at zero and that induction
393 /// variables can be pointers.
394 InductionList Inductions;
396 /// Allowed outside users. This holds the reduction
397 /// vars which can be accessed from outside the loop.
398 SmallPtrSet<Value*, 4> AllowedExit;
399 /// This set holds the variables which are known to be uniform after
401 SmallPtrSet<Instruction*, 4> Uniforms;
402 /// We need to check that all of the pointers in this list are disjoint
404 RuntimePointerCheck PtrRtCheck;
407 /// LoopVectorizationCostModel - estimates the expected speedups due to
409 /// In many cases vectorization is not profitable. This can happen because
410 /// of a number of reasons. In this class we mainly attempt to predict
411 /// the expected speedup/slowdowns due to the supported instruction set.
412 /// We use the VectorTargetTransformInfo to query the different backends
413 /// for the cost of different operations.
414 class LoopVectorizationCostModel {
417 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
418 LoopVectorizationLegality *Leg,
419 const VectorTargetTransformInfo *Vtti):
420 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
422 /// Returns the most profitable vectorization factor for the loop that is
423 /// smaller or equal to the VF argument. This method checks every power
425 unsigned findBestVectorizationFactor(unsigned VF = MaxVectorSize);
428 /// Returns the expected execution cost. The unit of the cost does
429 /// not matter because we use the 'cost' units to compare different
430 /// vector widths. The cost that is returned is *not* normalized by
431 /// the factor width.
432 unsigned expectedCost(unsigned VF);
434 /// Returns the execution time cost of an instruction for a given vector
435 /// width. Vector width of one means scalar.
436 unsigned getInstructionCost(Instruction *I, unsigned VF);
438 /// A helper function for converting Scalar types to vector types.
439 /// If the incoming type is void, we return void. If the VF is 1, we return
441 static Type* ToVectorTy(Type *Scalar, unsigned VF);
443 /// The loop that we evaluate.
448 /// Vectorization legality.
449 LoopVectorizationLegality *Legal;
450 /// Vector target information.
451 const VectorTargetTransformInfo *VTTI;
454 struct LoopVectorize : public LoopPass {
455 static char ID; // Pass identification, replacement for typeid
457 LoopVectorize() : LoopPass(ID) {
458 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
464 TargetTransformInfo *TTI;
467 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
468 // We only vectorize innermost loops.
472 SE = &getAnalysis<ScalarEvolution>();
473 DL = getAnalysisIfAvailable<DataLayout>();
474 LI = &getAnalysis<LoopInfo>();
475 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
476 DT = &getAnalysis<DominatorTree>();
478 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
479 L->getHeader()->getParent()->getName() << "\"\n");
481 // Check if it is legal to vectorize the loop.
482 LoopVectorizationLegality LVL(L, SE, DL, DT);
483 if (!LVL.canVectorize()) {
484 DEBUG(dbgs() << "LV: Not vectorizing.\n");
488 // Select the preffered vectorization factor.
490 if (VectorizationFactor == 0) {
491 const VectorTargetTransformInfo *VTTI = 0;
493 VTTI = TTI->getVectorTargetTransformInfo();
494 // Use the cost model.
495 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
496 VF = CM.findBestVectorizationFactor();
499 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
504 // Use the user command flag.
505 VF = VectorizationFactor;
508 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
509 L->getHeader()->getParent()->getParent()->getModuleIdentifier()<<
512 // If we decided that it is *legal* to vectorizer the loop then do it.
513 InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF);
516 DEBUG(verifyFunction(*L->getHeader()->getParent()));
520 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
521 LoopPass::getAnalysisUsage(AU);
522 AU.addRequiredID(LoopSimplifyID);
523 AU.addRequiredID(LCSSAID);
524 AU.addRequired<LoopInfo>();
525 AU.addRequired<ScalarEvolution>();
526 AU.addRequired<DominatorTree>();
527 AU.addPreserved<LoopInfo>();
528 AU.addPreserved<DominatorTree>();
533 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
535 LLVMContext &C = V->getContext();
536 Type *VTy = VectorType::get(V->getType(), VF);
537 Type *I32 = IntegerType::getInt32Ty(C);
539 // Save the current insertion location.
540 Instruction *Loc = Builder.GetInsertPoint();
542 // We need to place the broadcast of invariant variables outside the loop.
543 bool Invariant = (OrigLoop->isLoopInvariant(V) && V != Induction);
545 // Place the code for broadcasting invariant variables in the new preheader.
547 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
549 Constant *Zero = ConstantInt::get(I32, 0);
550 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
551 Value *UndefVal = UndefValue::get(VTy);
552 // Insert the value into a new vector.
553 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
554 // Broadcast the scalar into all locations in the vector.
555 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
558 // Restore the builder insertion point.
560 Builder.SetInsertPoint(Loc);
565 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val) {
566 assert(Val->getType()->isVectorTy() && "Must be a vector");
567 assert(Val->getType()->getScalarType()->isIntegerTy() &&
568 "Elem must be an integer");
570 Type *ITy = Val->getType()->getScalarType();
571 VectorType *Ty = cast<VectorType>(Val->getType());
572 unsigned VLen = Ty->getNumElements();
573 SmallVector<Constant*, 8> Indices;
575 // Create a vector of consecutive numbers from zero to VF.
576 for (unsigned i = 0; i < VLen; ++i)
577 Indices.push_back(ConstantInt::get(ITy, i));
579 // Add the consecutive indices to the vector value.
580 Constant *Cv = ConstantVector::get(Indices);
581 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
582 return Builder.CreateAdd(Val, Cv, "induction");
585 bool LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
586 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
588 // If this pointer is an induction variable, return it.
589 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
590 if (Phi && getInductionVars()->count(Phi))
593 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
597 unsigned NumOperands = Gep->getNumOperands();
598 Value *LastIndex = Gep->getOperand(NumOperands - 1);
600 // Check that all of the gep indices are uniform except for the last.
601 for (unsigned i = 0; i < NumOperands - 1; ++i)
602 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
605 // We can emit wide load/stores only if the last index is the induction
607 const SCEV *Last = SE->getSCEV(LastIndex);
608 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
609 const SCEV *Step = AR->getStepRecurrence(*SE);
611 // The memory is consecutive because the last index is consecutive
612 // and all other indices are loop invariant.
620 bool LoopVectorizationLegality::isUniform(Value *V) {
621 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
624 Value *InnerLoopVectorizer::getVectorValue(Value *V) {
625 assert(V != Induction && "The new induction variable should not be used.");
626 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
627 // If we saved a vectorized copy of V, use it.
628 Value *&MapEntry = WidenMap[V];
632 // Broadcast V and save the value for future uses.
633 Value *B = getBroadcastInstrs(V);
639 InnerLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
640 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
643 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
644 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
645 // Holds vector parameters or scalars, in case of uniform vals.
646 SmallVector<Value*, 8> Params;
648 // Find all of the vectorized parameters.
649 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
650 Value *SrcOp = Instr->getOperand(op);
652 // If we are accessing the old induction variable, use the new one.
653 if (SrcOp == OldInduction) {
654 Params.push_back(getVectorValue(SrcOp));
658 // Try using previously calculated values.
659 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
661 // If the src is an instruction that appeared earlier in the basic block
662 // then it should already be vectorized.
663 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
664 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
665 // The parameter is a vector value from earlier.
666 Params.push_back(WidenMap[SrcInst]);
668 // The parameter is a scalar from outside the loop. Maybe even a constant.
669 Params.push_back(SrcOp);
673 assert(Params.size() == Instr->getNumOperands() &&
674 "Invalid number of operands");
676 // Does this instruction return a value ?
677 bool IsVoidRetTy = Instr->getType()->isVoidTy();
678 Value *VecResults = 0;
680 // If we have a return value, create an empty vector. We place the scalarized
681 // instructions in this vector.
683 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
685 // For each scalar that we create:
686 for (unsigned i = 0; i < VF; ++i) {
687 Instruction *Cloned = Instr->clone();
689 Cloned->setName(Instr->getName() + ".cloned");
690 // Replace the operands of the cloned instrucions with extracted scalars.
691 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
692 Value *Op = Params[op];
693 // Param is a vector. Need to extract the right lane.
694 if (Op->getType()->isVectorTy())
695 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
696 Cloned->setOperand(op, Op);
699 // Place the cloned scalar in the new loop.
700 Builder.Insert(Cloned);
702 // If the original scalar returns a value we need to place it in a vector
703 // so that future users will be able to use it.
705 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
706 Builder.getInt32(i));
710 WidenMap[Instr] = VecResults;
714 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
716 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
717 Legal->getRuntimePointerCheck();
719 if (!PtrRtCheck->Need)
722 Value *MemoryRuntimeCheck = 0;
723 unsigned NumPointers = PtrRtCheck->Pointers.size();
724 SmallVector<Value* , 2> Starts;
725 SmallVector<Value* , 2> Ends;
727 SCEVExpander Exp(*SE, "induction");
729 // Use this type for pointer arithmetic.
730 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
732 for (unsigned i = 0; i < NumPointers; ++i) {
733 Value *Ptr = PtrRtCheck->Pointers[i];
734 const SCEV *Sc = SE->getSCEV(Ptr);
736 if (SE->isLoopInvariant(Sc, OrigLoop)) {
737 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
739 Starts.push_back(Ptr);
742 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
744 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i],
746 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
747 Starts.push_back(Start);
752 for (unsigned i = 0; i < NumPointers; ++i) {
753 for (unsigned j = i+1; j < NumPointers; ++j) {
754 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
755 Starts[i], Ends[j], "bound0", Loc);
756 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
757 Starts[j], Ends[i], "bound1", Loc);
758 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
759 "found.conflict", Loc);
760 if (MemoryRuntimeCheck)
761 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
764 "conflict.rdx", Loc);
766 MemoryRuntimeCheck = IsConflict;
771 return MemoryRuntimeCheck;
775 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
777 In this function we generate a new loop. The new loop will contain
778 the vectorized instructions while the old loop will continue to run the
781 [ ] <-- vector loop bypass.
784 | [ ] <-- vector pre header.
788 | [ ]_| <-- vector loop.
791 >[ ] <--- middle-block.
794 | [ ] <--- new preheader.
798 | [ ]_| <-- old scalar loop to handle remainder.
805 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
806 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
807 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
808 assert(ExitBlock && "Must have an exit block");
810 // Some loops have a single integer induction variable, while other loops
811 // don't. One example is c++ iterators that often have multiple pointer
812 // induction variables. In the code below we also support a case where we
813 // don't have a single induction variable.
814 OldInduction = Legal->getInduction();
815 Type *IdxTy = OldInduction ? OldInduction->getType() :
816 DL->getIntPtrType(SE->getContext());
818 // Find the loop boundaries.
819 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
820 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
822 // Get the total trip count from the count by adding 1.
823 ExitCount = SE->getAddExpr(ExitCount,
824 SE->getConstant(ExitCount->getType(), 1));
826 // Expand the trip count and place the new instructions in the preheader.
827 // Notice that the pre-header does not change, only the loop body.
828 SCEVExpander Exp(*SE, "induction");
830 // Count holds the overall loop count (N).
831 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
832 BypassBlock->getTerminator());
834 // The loop index does not have to start at Zero. Find the original start
835 // value from the induction PHI node. If we don't have an induction variable
836 // then we know that it starts at zero.
837 Value *StartIdx = OldInduction ?
838 OldInduction->getIncomingValueForBlock(BypassBlock):
839 ConstantInt::get(IdxTy, 0);
841 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
842 assert(BypassBlock && "Invalid loop structure");
844 // Generate the code that checks in runtime if arrays overlap.
845 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal,
846 BypassBlock->getTerminator());
848 // Split the single block loop into the two loop structure described above.
849 BasicBlock *VectorPH =
850 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
851 BasicBlock *VecBody =
852 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
853 BasicBlock *MiddleBlock =
854 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
855 BasicBlock *ScalarPH =
856 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
858 // This is the location in which we add all of the logic for bypassing
859 // the new vector loop.
860 Instruction *Loc = BypassBlock->getTerminator();
862 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
864 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
866 // Generate the induction variable.
867 Induction = Builder.CreatePHI(IdxTy, 2, "index");
868 Constant *Step = ConstantInt::get(IdxTy, VF);
870 // We may need to extend the index in case there is a type mismatch.
871 // We know that the count starts at zero and does not overflow.
872 if (Count->getType() != IdxTy) {
873 // The exit count can be of pointer type. Convert it to the correct
875 if (ExitCount->getType()->isPointerTy())
876 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
878 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
881 // Add the start index to the loop count to get the new end index.
882 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
884 // Now we need to generate the expression for N - (N % VF), which is
885 // the part that the vectorized body will execute.
886 Constant *CIVF = ConstantInt::get(IdxTy, VF);
887 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
888 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
889 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
890 "end.idx.rnd.down", Loc);
892 // Now, compare the new count to zero. If it is zero skip the vector loop and
893 // jump to the scalar loop.
894 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
899 // If we are using memory runtime checks, include them in.
900 if (MemoryRuntimeCheck)
901 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
904 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
905 // Remove the old terminator.
906 Loc->eraseFromParent();
908 // We are going to resume the execution of the scalar loop.
909 // Go over all of the induction variables that we found and fix the
910 // PHIs that are left in the scalar version of the loop.
911 // The starting values of PHI nodes depend on the counter of the last
912 // iteration in the vectorized loop.
913 // If we come from a bypass edge then we need to start from the original start
916 // This variable saves the new starting index for the scalar loop.
917 PHINode *ResumeIndex = 0;
918 LoopVectorizationLegality::InductionList::iterator I, E;
919 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
920 for (I = List->begin(), E = List->end(); I != E; ++I) {
921 PHINode *OrigPhi = I->first;
922 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
923 MiddleBlock->getTerminator());
925 if (OrigPhi->getType()->isIntegerTy()) {
926 // Handle the integer induction counter:
927 assert(OrigPhi == OldInduction && "Unknown integer PHI");
928 // We know what the end value is.
929 EndValue = IdxEndRoundDown;
930 // We also know which PHI node holds it.
931 ResumeIndex = ResumeVal;
933 // For pointer induction variables, calculate the offset using
935 EndValue = GetElementPtrInst::Create(I->second, CountRoundDown,
937 BypassBlock->getTerminator());
940 // The new PHI merges the original incoming value, in case of a bypass,
941 // or the value at the end of the vectorized loop.
942 ResumeVal->addIncoming(I->second, BypassBlock);
943 ResumeVal->addIncoming(EndValue, VecBody);
945 // Fix the scalar body counter (PHI node).
946 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
947 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
950 // If we are generating a new induction variable then we also need to
951 // generate the code that calculates the exit value. This value is not
952 // simply the end of the counter because we may skip the vectorized body
953 // in case of a runtime check.
955 assert(!ResumeIndex && "Unexpected resume value found");
956 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
957 MiddleBlock->getTerminator());
958 ResumeIndex->addIncoming(StartIdx, BypassBlock);
959 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
962 // Make sure that we found the index where scalar loop needs to continue.
963 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
964 "Invalid resume Index");
966 // Add a check in the middle block to see if we have completed
967 // all of the iterations in the first vector loop.
968 // If (N - N%VF) == N, then we *don't* need to run the remainder.
969 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
970 ResumeIndex, "cmp.n",
971 MiddleBlock->getTerminator());
973 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
974 // Remove the old terminator.
975 MiddleBlock->getTerminator()->eraseFromParent();
977 // Create i+1 and fill the PHINode.
978 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
979 Induction->addIncoming(StartIdx, VectorPH);
980 Induction->addIncoming(NextIdx, VecBody);
981 // Create the compare.
982 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
983 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
985 // Now we have two terminators. Remove the old one from the block.
986 VecBody->getTerminator()->eraseFromParent();
988 // Get ready to start creating new instructions into the vectorized body.
989 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
991 // Create and register the new vector loop.
992 Loop* Lp = new Loop();
993 Loop *ParentLoop = OrigLoop->getParentLoop();
995 // Insert the new loop into the loop nest and register the new basic blocks.
997 ParentLoop->addChildLoop(Lp);
998 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
999 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1000 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1002 LI->addTopLevelLoop(Lp);
1005 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1008 LoopVectorPreHeader = VectorPH;
1009 LoopScalarPreHeader = ScalarPH;
1010 LoopMiddleBlock = MiddleBlock;
1011 LoopExitBlock = ExitBlock;
1012 LoopVectorBody = VecBody;
1013 LoopScalarBody = OldBasicBlock;
1014 LoopBypassBlock = BypassBlock;
1017 /// This function returns the identity element (or neutral element) for
1018 /// the operation K.
1020 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
1022 case LoopVectorizationLegality::IntegerXor:
1023 case LoopVectorizationLegality::IntegerAdd:
1024 case LoopVectorizationLegality::IntegerOr:
1025 // Adding, Xoring, Oring zero to a number does not change it.
1027 case LoopVectorizationLegality::IntegerMult:
1028 // Multiplying a number by 1 does not change it.
1030 case LoopVectorizationLegality::IntegerAnd:
1031 // AND-ing a number with an all-1 value does not change it.
1034 llvm_unreachable("Unknown reduction kind");
1039 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1040 //===------------------------------------------------===//
1042 // Notice: any optimization or new instruction that go
1043 // into the code below should be also be implemented in
1046 //===------------------------------------------------===//
1047 typedef SmallVector<PHINode*, 4> PhiVector;
1048 BasicBlock &BB = *OrigLoop->getHeader();
1049 Constant *Zero = ConstantInt::get(
1050 IntegerType::getInt32Ty(BB.getContext()), 0);
1052 // In order to support reduction variables we need to be able to vectorize
1053 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1054 // stages. First, we create a new vector PHI node with no incoming edges.
1055 // We use this value when we vectorize all of the instructions that use the
1056 // PHI. Next, after all of the instructions in the block are complete we
1057 // add the new incoming edges to the PHI. At this point all of the
1058 // instructions in the basic block are vectorized, so we can use them to
1059 // construct the PHI.
1060 PhiVector RdxPHIsToFix;
1062 // For each instruction in the old loop.
1063 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1064 Instruction *Inst = it;
1066 switch (Inst->getOpcode()) {
1067 case Instruction::Br:
1068 // Nothing to do for PHIs and BR, since we already took care of the
1069 // loop control flow instructions.
1071 case Instruction::PHI:{
1072 PHINode* P = cast<PHINode>(Inst);
1073 // Handle reduction variables:
1074 if (Legal->getReductionVars()->count(P)) {
1075 // This is phase one of vectorizing PHIs.
1076 Type *VecTy = VectorType::get(Inst->getType(), VF);
1077 WidenMap[Inst] = PHINode::Create(VecTy, 2, "vec.phi",
1078 LoopVectorBody->getFirstInsertionPt());
1079 RdxPHIsToFix.push_back(P);
1083 // This PHINode must be an induction variable.
1084 // Make sure that we know about it.
1085 assert(Legal->getInductionVars()->count(P) &&
1086 "Not an induction variable");
1088 if (P->getType()->isIntegerTy()) {
1089 assert(P == OldInduction && "Unexpected PHI");
1090 Value *Broadcasted = getBroadcastInstrs(Induction);
1091 // After broadcasting the induction variable we need to make the
1092 // vector consecutive by adding 0, 1, 2 ...
1093 Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted);
1095 WidenMap[OldInduction] = ConsecutiveInduction;
1099 // Handle pointer inductions.
1100 assert(P->getType()->isPointerTy() && "Unexpected type.");
1101 Value *StartIdx = OldInduction ?
1102 Legal->getInductionVars()->lookup(OldInduction) :
1103 ConstantInt::get(Induction->getType(), 0);
1105 // This is the pointer value coming into the loop.
1106 Value *StartPtr = Legal->getInductionVars()->lookup(P);
1108 // This is the normalized GEP that starts counting at zero.
1109 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1112 // This is the vector of results. Notice that we don't generate vector
1113 // geps because scalar geps result in better code.
1114 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1115 for (unsigned int i = 0; i < VF; ++i) {
1116 Constant *Idx = ConstantInt::get(Induction->getType(), i);
1117 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1118 Value *SclrGep = Builder.CreateGEP(StartPtr, GlobalIdx, "next.gep");
1119 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1120 Builder.getInt32(i),
1124 WidenMap[Inst] = VecVal;
1127 case Instruction::Add:
1128 case Instruction::FAdd:
1129 case Instruction::Sub:
1130 case Instruction::FSub:
1131 case Instruction::Mul:
1132 case Instruction::FMul:
1133 case Instruction::UDiv:
1134 case Instruction::SDiv:
1135 case Instruction::FDiv:
1136 case Instruction::URem:
1137 case Instruction::SRem:
1138 case Instruction::FRem:
1139 case Instruction::Shl:
1140 case Instruction::LShr:
1141 case Instruction::AShr:
1142 case Instruction::And:
1143 case Instruction::Or:
1144 case Instruction::Xor: {
1145 // Just widen binops.
1146 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
1147 Value *A = getVectorValue(Inst->getOperand(0));
1148 Value *B = getVectorValue(Inst->getOperand(1));
1150 // Use this vector value for all users of the original instruction.
1151 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
1154 // Update the NSW, NUW and Exact flags.
1155 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1156 if (isa<OverflowingBinaryOperator>(BinOp)) {
1157 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1158 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1160 if (isa<PossiblyExactOperator>(VecOp))
1161 VecOp->setIsExact(BinOp->isExact());
1164 case Instruction::Select: {
1166 // If the selector is loop invariant we can create a select
1167 // instruction with a scalar condition. Otherwise, use vector-select.
1168 Value *Cond = Inst->getOperand(0);
1169 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
1171 // The condition can be loop invariant but still defined inside the
1172 // loop. This means that we can't just use the original 'cond' value.
1173 // We have to take the 'vectorized' value and pick the first lane.
1174 // Instcombine will make this a no-op.
1175 Cond = getVectorValue(Cond);
1177 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
1179 Value *Op0 = getVectorValue(Inst->getOperand(1));
1180 Value *Op1 = getVectorValue(Inst->getOperand(2));
1181 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
1185 case Instruction::ICmp:
1186 case Instruction::FCmp: {
1187 // Widen compares. Generate vector compares.
1188 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
1189 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
1190 Value *A = getVectorValue(Inst->getOperand(0));
1191 Value *B = getVectorValue(Inst->getOperand(1));
1193 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
1195 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
1199 case Instruction::Store: {
1200 // Attempt to issue a wide store.
1201 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1202 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1203 Value *Ptr = SI->getPointerOperand();
1204 unsigned Alignment = SI->getAlignment();
1206 assert(!Legal->isUniform(Ptr) &&
1207 "We do not allow storing to uniform addresses");
1209 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1211 // This store does not use GEPs.
1212 if (!Legal->isConsecutivePtr(Ptr)) {
1213 scalarizeInstruction(Inst);
1218 // The last index does not have to be the induction. It can be
1219 // consecutive and be a function of the index. For example A[I+1];
1220 unsigned NumOperands = Gep->getNumOperands();
1221 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1222 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1224 // Create the new GEP with the new induction variable.
1225 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1226 Gep2->setOperand(NumOperands - 1, LastIndex);
1227 Ptr = Builder.Insert(Gep2);
1229 // Use the induction element ptr.
1230 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1231 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1233 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1234 Value *Val = getVectorValue(SI->getValueOperand());
1235 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1238 case Instruction::Load: {
1239 // Attempt to issue a wide load.
1240 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1241 Type *RetTy = VectorType::get(LI->getType(), VF);
1242 Value *Ptr = LI->getPointerOperand();
1243 unsigned Alignment = LI->getAlignment();
1244 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1246 // If the pointer is loop invariant or if it is non consecutive,
1247 // scalarize the load.
1248 bool Con = Legal->isConsecutivePtr(Ptr);
1249 if (Legal->isUniform(Ptr) || !Con) {
1250 scalarizeInstruction(Inst);
1255 // The last index does not have to be the induction. It can be
1256 // consecutive and be a function of the index. For example A[I+1];
1257 unsigned NumOperands = Gep->getNumOperands();
1258 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1259 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1261 // Create the new GEP with the new induction variable.
1262 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1263 Gep2->setOperand(NumOperands - 1, LastIndex);
1264 Ptr = Builder.Insert(Gep2);
1266 // Use the induction element ptr.
1267 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1268 Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero);
1271 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1272 LI = Builder.CreateLoad(Ptr);
1273 LI->setAlignment(Alignment);
1274 // Use this vector value for all users of the load.
1275 WidenMap[Inst] = LI;
1278 case Instruction::ZExt:
1279 case Instruction::SExt:
1280 case Instruction::FPToUI:
1281 case Instruction::FPToSI:
1282 case Instruction::FPExt:
1283 case Instruction::PtrToInt:
1284 case Instruction::IntToPtr:
1285 case Instruction::SIToFP:
1286 case Instruction::UIToFP:
1287 case Instruction::Trunc:
1288 case Instruction::FPTrunc:
1289 case Instruction::BitCast: {
1290 /// Vectorize bitcasts.
1291 CastInst *CI = dyn_cast<CastInst>(Inst);
1292 Value *A = getVectorValue(Inst->getOperand(0));
1293 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1294 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1299 /// All other instructions are unsupported. Scalarize them.
1300 scalarizeInstruction(Inst);
1303 }// end of for_each instr.
1305 // At this point every instruction in the original loop is widended to
1306 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1307 // that we vectorized. The PHI nodes are currently empty because we did
1308 // not want to introduce cycles. Notice that the remaining PHI nodes
1309 // that we need to fix are reduction variables.
1311 // Create the 'reduced' values for each of the induction vars.
1312 // The reduced values are the vector values that we scalarize and combine
1313 // after the loop is finished.
1314 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1316 PHINode *RdxPhi = *it;
1317 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1318 assert(RdxPhi && "Unable to recover vectorized PHI");
1320 // Find the reduction variable descriptor.
1321 assert(Legal->getReductionVars()->count(RdxPhi) &&
1322 "Unable to find the reduction variable");
1323 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1324 (*Legal->getReductionVars())[RdxPhi];
1326 // We need to generate a reduction vector from the incoming scalar.
1327 // To do so, we need to generate the 'identity' vector and overide
1328 // one of the elements with the incoming scalar reduction. We need
1329 // to do it in the vector-loop preheader.
1330 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1332 // This is the vector-clone of the value that leaves the loop.
1333 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1334 Type *VecTy = VectorExit->getType();
1336 // Find the reduction identity variable. Zero for addition, or, xor,
1337 // one for multiplication, -1 for And.
1338 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1339 VecTy->getScalarType());
1341 // This vector is the Identity vector where the first element is the
1342 // incoming scalar reduction.
1343 Value *VectorStart = Builder.CreateInsertElement(Identity,
1344 RdxDesc.StartValue, Zero);
1346 // Fix the vector-loop phi.
1347 // We created the induction variable so we know that the
1348 // preheader is the first entry.
1349 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1351 // Reductions do not have to start at zero. They can start with
1352 // any loop invariant values.
1353 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1354 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1355 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1356 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1358 // Before each round, move the insertion point right between
1359 // the PHIs and the values we are going to write.
1360 // This allows us to write both PHINodes and the extractelement
1362 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1364 // This PHINode contains the vectorized reduction variable, or
1365 // the initial value vector, if we bypass the vector loop.
1366 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1367 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1368 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1370 // Extract the first scalar.
1372 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1373 // Extract and reduce the remaining vector elements.
1374 for (unsigned i=1; i < VF; ++i) {
1376 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1377 switch (RdxDesc.Kind) {
1378 case LoopVectorizationLegality::IntegerAdd:
1379 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1381 case LoopVectorizationLegality::IntegerMult:
1382 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1384 case LoopVectorizationLegality::IntegerOr:
1385 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1387 case LoopVectorizationLegality::IntegerAnd:
1388 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1390 case LoopVectorizationLegality::IntegerXor:
1391 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1394 llvm_unreachable("Unknown reduction operation");
1398 // Now, we need to fix the users of the reduction variable
1399 // inside and outside of the scalar remainder loop.
1400 // We know that the loop is in LCSSA form. We need to update the
1401 // PHI nodes in the exit blocks.
1402 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1403 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1404 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1405 if (!LCSSAPhi) continue;
1407 // All PHINodes need to have a single entry edge, or two if
1408 // we already fixed them.
1409 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1411 // We found our reduction value exit-PHI. Update it with the
1412 // incoming bypass edge.
1413 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1414 // Add an edge coming from the bypass.
1415 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1418 }// end of the LCSSA phi scan.
1420 // Fix the scalar loop reduction variable with the incoming reduction sum
1421 // from the vector body and from the backedge value.
1422 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1423 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1424 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1425 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1426 }// end of for each redux variable.
1429 void InnerLoopVectorizer::updateAnalysis() {
1430 // Forget the original basic block.
1431 SE->forgetLoop(OrigLoop);
1433 // Update the dominator tree information.
1434 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1435 "Entry does not dominate exit.");
1437 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1438 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1439 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1440 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1441 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1442 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1444 DEBUG(DT->verifyAnalysis());
1448 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
1449 if (!EnableIfConversion)
1452 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
1453 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
1455 // Collect the blocks that need predication.
1456 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
1457 BasicBlock *BB = LoopBlocks[i];
1459 // We must have at most two predecessors because we need to convert
1460 // all PHIs to selects.
1461 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
1465 // We must be able to predicate all blocks that need to be predicated.
1466 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
1470 // We can if-convert this loop.
1474 bool LoopVectorizationLegality::canVectorize() {
1475 assert(TheLoop->getLoopPreheader() && "No preheader!!");
1477 // We can only vectorize innermost loops.
1478 if (TheLoop->getSubLoopsVector().size())
1481 // We must have a single backedge.
1482 if (TheLoop->getNumBackEdges() != 1)
1485 // We must have a single exiting block.
1486 if (!TheLoop->getExitingBlock())
1489 unsigned NumBlocks = TheLoop->getNumBlocks();
1491 // Check if we can if-convert non single-bb loops.
1492 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
1493 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
1497 // We need to have a loop header.
1498 BasicBlock *Header = TheLoop->getHeader();
1499 BasicBlock *Latch = TheLoop->getLoopLatch();
1500 DEBUG(dbgs() << "LV: Found a loop: " << Header->getName() << "\n");
1502 // ScalarEvolution needs to be able to find the exit count.
1503 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
1504 if (ExitCount == SE->getCouldNotCompute()) {
1505 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1509 // Do not loop-vectorize loops with a tiny trip count.
1510 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
1511 if (TC > 0u && TC < TinyTripCountThreshold) {
1512 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1513 "This loop is not worth vectorizing.\n");
1517 // Check if we can vectorize the instructions and CFG in this loop.
1518 if (!canVectorizeInstrs()) {
1519 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
1523 // Go over each instruction and look at memory deps.
1524 if (!canVectorizeMemory()) {
1525 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
1529 // Collect all of the variables that remain uniform after vectorization.
1530 collectLoopUniforms();
1532 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1533 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1536 // Okay! We can vectorize. At this point we don't have any other mem analysis
1537 // which may limit our maximum vectorization factor, so just return true with
1542 bool LoopVectorizationLegality::canVectorizeInstrs() {
1543 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
1544 BasicBlock *Header = TheLoop->getHeader();
1546 // For each block in the loop.
1547 for (Loop::block_iterator bb = TheLoop->block_begin(),
1548 be = TheLoop->block_end(); bb != be; ++bb) {
1550 // Scan the instructions in the block and look for hazards.
1551 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1554 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
1555 // This should not happen because the loop should be normalized.
1556 if (Phi->getNumIncomingValues() != 2) {
1557 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1561 // If this PHINode is not in the header block, then we know that we
1562 // can convert it to select during if-conversion.
1566 // This is the value coming from the preheader.
1567 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
1569 // We only look at integer and pointer phi nodes.
1570 if (Phi->getType()->isPointerTy() && isInductionVariable(Phi)) {
1571 DEBUG(dbgs() << "LV: Found a pointer induction variable.\n");
1572 Inductions[Phi] = StartValue;
1574 } else if (!Phi->getType()->isIntegerTy()) {
1575 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
1579 // Handle integer PHIs:
1580 if (isInductionVariable(Phi)) {
1582 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1585 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1587 Inductions[Phi] = StartValue;
1590 if (AddReductionVar(Phi, IntegerAdd)) {
1591 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1594 if (AddReductionVar(Phi, IntegerMult)) {
1595 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1598 if (AddReductionVar(Phi, IntegerOr)) {
1599 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1602 if (AddReductionVar(Phi, IntegerAnd)) {
1603 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1606 if (AddReductionVar(Phi, IntegerXor)) {
1607 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1611 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1613 }// end of PHI handling
1615 // We still don't handle functions.
1616 CallInst *CI = dyn_cast<CallInst>(it);
1618 DEBUG(dbgs() << "LV: Found a call site.\n");
1622 // We do not re-vectorize vectors.
1623 if (!VectorType::isValidElementType(it->getType()) &&
1624 !it->getType()->isVoidTy()) {
1625 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1629 // Reduction instructions are allowed to have exit users.
1630 // All other instructions must not have external users.
1631 if (!AllowedExit.count(it))
1632 //Check that all of the users of the loop are inside the BB.
1633 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
1635 Instruction *U = cast<Instruction>(*I);
1636 // This user may be a reduction exit value.
1637 if (!TheLoop->contains(U)) {
1638 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1647 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
1648 assert(getInductionVars()->size() && "No induction variables");
1654 void LoopVectorizationLegality::collectLoopUniforms() {
1655 // We now know that the loop is vectorizable!
1656 // Collect variables that will remain uniform after vectorization.
1657 std::vector<Value*> Worklist;
1658 BasicBlock *Latch = TheLoop->getLoopLatch();
1660 // Start with the conditional branch and walk up the block.
1661 Worklist.push_back(Latch->getTerminator()->getOperand(0));
1663 while (Worklist.size()) {
1664 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1665 Worklist.pop_back();
1667 // Look at instructions inside this loop.
1668 // Stop when reaching PHI nodes.
1669 // TODO: we need to follow values all over the loop, not only in this block.
1670 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
1673 // This is a known uniform.
1676 // Insert all operands.
1677 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
1678 Worklist.push_back(I->getOperand(i));
1683 bool LoopVectorizationLegality::canVectorizeMemory() {
1684 typedef SmallVector<Value*, 16> ValueVector;
1685 typedef SmallPtrSet<Value*, 16> ValueSet;
1686 // Holds the Load and Store *instructions*.
1689 PtrRtCheck.Pointers.clear();
1690 PtrRtCheck.Need = false;
1693 for (Loop::block_iterator bb = TheLoop->block_begin(),
1694 be = TheLoop->block_end(); bb != be; ++bb) {
1696 // Scan the BB and collect legal loads and stores.
1697 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1700 // If this is a load, save it. If this instruction can read from memory
1701 // but is not a load, then we quit. Notice that we don't handle function
1702 // calls that read or write.
1703 if (it->mayReadFromMemory()) {
1704 LoadInst *Ld = dyn_cast<LoadInst>(it);
1705 if (!Ld) return false;
1706 if (!Ld->isSimple()) {
1707 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1710 Loads.push_back(Ld);
1714 // Save 'store' instructions. Abort if other instructions write to memory.
1715 if (it->mayWriteToMemory()) {
1716 StoreInst *St = dyn_cast<StoreInst>(it);
1717 if (!St) return false;
1718 if (!St->isSimple()) {
1719 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1722 Stores.push_back(St);
1727 // Now we have two lists that hold the loads and the stores.
1728 // Next, we find the pointers that they use.
1730 // Check if we see any stores. If there are no stores, then we don't
1731 // care if the pointers are *restrict*.
1732 if (!Stores.size()) {
1733 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1737 // Holds the read and read-write *pointers* that we find.
1739 ValueVector ReadWrites;
1741 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1742 // multiple times on the same object. If the ptr is accessed twice, once
1743 // for read and once for write, it will only appear once (on the write
1744 // list). This is okay, since we are going to check for conflicts between
1745 // writes and between reads and writes, but not between reads and reads.
1748 ValueVector::iterator I, IE;
1749 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1750 StoreInst *ST = dyn_cast<StoreInst>(*I);
1751 assert(ST && "Bad StoreInst");
1752 Value* Ptr = ST->getPointerOperand();
1754 if (isUniform(Ptr)) {
1755 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1759 // If we did *not* see this pointer before, insert it to
1760 // the read-write list. At this phase it is only a 'write' list.
1761 if (Seen.insert(Ptr))
1762 ReadWrites.push_back(Ptr);
1765 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1766 LoadInst *LD = dyn_cast<LoadInst>(*I);
1767 assert(LD && "Bad LoadInst");
1768 Value* Ptr = LD->getPointerOperand();
1769 // If we did *not* see this pointer before, insert it to the
1770 // read list. If we *did* see it before, then it is already in
1771 // the read-write list. This allows us to vectorize expressions
1772 // such as A[i] += x; Because the address of A[i] is a read-write
1773 // pointer. This only works if the index of A[i] is consecutive.
1774 // If the address of i is unknown (for example A[B[i]]) then we may
1775 // read a few words, modify, and write a few words, and some of the
1776 // words may be written to the same address.
1777 if (Seen.insert(Ptr) || !isConsecutivePtr(Ptr))
1778 Reads.push_back(Ptr);
1781 // If we write (or read-write) to a single destination and there are no
1782 // other reads in this loop then is it safe to vectorize.
1783 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1784 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1788 // Find pointers with computable bounds. We are going to use this information
1789 // to place a runtime bound check.
1791 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1792 if (hasComputableBounds(*I)) {
1793 PtrRtCheck.insert(SE, TheLoop, *I);
1794 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1799 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1800 if (hasComputableBounds(*I)) {
1801 PtrRtCheck.insert(SE, TheLoop, *I);
1802 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1808 // Check that we did not collect too many pointers or found a
1809 // unsizeable pointer.
1810 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1815 PtrRtCheck.Need = RT;
1818 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1821 // Now that the pointers are in two lists (Reads and ReadWrites), we
1822 // can check that there are no conflicts between each of the writes and
1823 // between the writes to the reads.
1824 ValueSet WriteObjects;
1825 ValueVector TempObjects;
1827 // Check that the read-writes do not conflict with other read-write
1829 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1830 GetUnderlyingObjects(*I, TempObjects, DL);
1831 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1833 if (!isIdentifiedObject(*it)) {
1834 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1837 if (!WriteObjects.insert(*it)) {
1838 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1843 TempObjects.clear();
1846 /// Check that the reads don't conflict with the read-writes.
1847 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1848 GetUnderlyingObjects(*I, TempObjects, DL);
1849 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1851 if (!isIdentifiedObject(*it)) {
1852 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1855 if (WriteObjects.count(*it)) {
1856 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1861 TempObjects.clear();
1864 // It is safe to vectorize and we don't need any runtime checks.
1865 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1870 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1871 ReductionKind Kind) {
1872 if (Phi->getNumIncomingValues() != 2)
1875 // Find the possible incoming reduction variable.
1876 BasicBlock *BB = Phi->getParent();
1877 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1878 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1879 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1881 // ExitInstruction is the single value which is used outside the loop.
1882 // We only allow for a single reduction value to be used outside the loop.
1883 // This includes users of the reduction, variables (which form a cycle
1884 // which ends in the phi node).
1885 Instruction *ExitInstruction = 0;
1887 // Iter is our iterator. We start with the PHI node and scan for all of the
1888 // users of this instruction. All users must be instructions which can be
1889 // used as reduction variables (such as ADD). We may have a single
1890 // out-of-block user. They cycle must end with the original PHI.
1891 // Also, we can't have multiple block-local users.
1892 Instruction *Iter = Phi;
1894 // Any reduction instr must be of one of the allowed kinds.
1895 if (!isReductionInstr(Iter, Kind))
1898 // Did we found a user inside this block ?
1899 bool FoundInBlockUser = false;
1900 // Did we reach the initial PHI node ?
1901 bool FoundStartPHI = false;
1903 // If the instruction has no users then this is a broken
1904 // chain and can't be a reduction variable.
1905 if (Iter->use_empty())
1908 // For each of the *users* of iter.
1909 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1911 Instruction *U = cast<Instruction>(*it);
1912 // We already know that the PHI is a user.
1914 FoundStartPHI = true;
1917 // Check if we found the exit user.
1918 BasicBlock *Parent = U->getParent();
1920 // We must have a single exit instruction.
1921 if (ExitInstruction != 0)
1923 ExitInstruction = Iter;
1925 // We can't have multiple inside users.
1926 if (FoundInBlockUser)
1928 FoundInBlockUser = true;
1932 // We found a reduction var if we have reached the original
1933 // phi node and we only have a single instruction with out-of-loop
1935 if (FoundStartPHI && ExitInstruction) {
1936 // This instruction is allowed to have out-of-loop users.
1937 AllowedExit.insert(ExitInstruction);
1939 // Save the description of this reduction variable.
1940 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1941 Reductions[Phi] = RD;
1948 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1949 ReductionKind Kind) {
1950 switch (I->getOpcode()) {
1953 case Instruction::PHI:
1956 case Instruction::Add:
1957 case Instruction::Sub:
1958 return Kind == IntegerAdd;
1959 case Instruction::Mul:
1960 return Kind == IntegerMult;
1961 case Instruction::And:
1962 return Kind == IntegerAnd;
1963 case Instruction::Or:
1964 return Kind == IntegerOr;
1965 case Instruction::Xor:
1966 return Kind == IntegerXor;
1970 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1971 Type *PhiTy = Phi->getType();
1972 // We only handle integer and pointer inductions variables.
1973 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1976 // Check that the PHI is consecutive and starts at zero.
1977 const SCEV *PhiScev = SE->getSCEV(Phi);
1978 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1980 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1983 const SCEV *Step = AR->getStepRecurrence(*SE);
1985 // Integer inductions need to have a stride of one.
1986 if (PhiTy->isIntegerTy())
1987 return Step->isOne();
1989 // Calculate the pointer stride and check if it is consecutive.
1990 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1991 if (!C) return false;
1993 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1994 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
1995 return (C->getValue()->equalsInt(Size));
1998 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
1999 assert(TheLoop->contains(BB) && "Unknown block used");
2001 // Blocks that do not dominate the latch need predication.
2002 BasicBlock* Latch = TheLoop->getLoopLatch();
2003 return !DT->dominates(BB, Latch);
2006 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
2007 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2008 // We don't predicate loads/stores at the moment.
2009 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
2012 // The isntructions below can trap.
2013 switch (it->getOpcode()) {
2015 case Instruction::UDiv:
2016 case Instruction::SDiv:
2017 case Instruction::URem:
2018 case Instruction::SRem:
2026 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
2027 const SCEV *PhiScev = SE->getSCEV(Ptr);
2028 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2032 return AR->isAffine();
2036 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
2038 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
2042 float Cost = expectedCost(1);
2044 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
2045 for (unsigned i=2; i <= VF; i*=2) {
2046 // Notice that the vector loop needs to be executed less times, so
2047 // we need to divide the cost of the vector loops by the width of
2048 // the vector elements.
2049 float VectorCost = expectedCost(i) / (float)i;
2050 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
2051 (int)VectorCost << ".\n");
2052 if (VectorCost < Cost) {
2058 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
2062 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
2066 for (Loop::block_iterator bb = TheLoop->block_begin(),
2067 be = TheLoop->block_end(); bb != be; ++bb) {
2068 unsigned BlockCost = 0;
2069 BasicBlock *BB = *bb;
2071 // For each instruction in the old loop.
2072 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2074 unsigned C = getInstructionCost(it, VF);
2076 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
2077 VF << " For instruction: "<< *it << "\n");
2080 // TODO: if-converted blocks can have a high-nest level. We need to
2081 // calculate the loop nest level and multiply the cost accordingly.
2082 if (Legal->blockNeedsPredication(*bb))
2092 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
2093 assert(VTTI && "Invalid vector target transformation info");
2095 // If we know that this instruction will remain uniform, check the cost of
2096 // the scalar version.
2097 if (Legal->isUniformAfterVectorization(I))
2100 Type *RetTy = I->getType();
2101 Type *VectorTy = ToVectorTy(RetTy, VF);
2104 // TODO: We need to estimate the cost of intrinsic calls.
2105 switch (I->getOpcode()) {
2106 case Instruction::GetElementPtr:
2107 // We mark this instruction as zero-cost because scalar GEPs are usually
2108 // lowered to the intruction addressing mode. At the moment we don't
2109 // generate vector geps.
2111 case Instruction::Br: {
2112 return VTTI->getCFInstrCost(I->getOpcode());
2114 case Instruction::PHI:
2115 //TODO: IF-converted IFs become selects.
2117 case Instruction::Add:
2118 case Instruction::FAdd:
2119 case Instruction::Sub:
2120 case Instruction::FSub:
2121 case Instruction::Mul:
2122 case Instruction::FMul:
2123 case Instruction::UDiv:
2124 case Instruction::SDiv:
2125 case Instruction::FDiv:
2126 case Instruction::URem:
2127 case Instruction::SRem:
2128 case Instruction::FRem:
2129 case Instruction::Shl:
2130 case Instruction::LShr:
2131 case Instruction::AShr:
2132 case Instruction::And:
2133 case Instruction::Or:
2134 case Instruction::Xor:
2135 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
2136 case Instruction::Select: {
2137 SelectInst *SI = cast<SelectInst>(I);
2138 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
2139 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
2140 Type *CondTy = SI->getCondition()->getType();
2142 CondTy = VectorType::get(CondTy, VF);
2144 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
2146 case Instruction::ICmp:
2147 case Instruction::FCmp: {
2148 Type *ValTy = I->getOperand(0)->getType();
2149 VectorTy = ToVectorTy(ValTy, VF);
2150 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
2152 case Instruction::Store: {
2153 StoreInst *SI = cast<StoreInst>(I);
2154 Type *ValTy = SI->getValueOperand()->getType();
2155 VectorTy = ToVectorTy(ValTy, VF);
2158 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
2159 SI->getAlignment(), SI->getPointerAddressSpace());
2161 // Scalarized stores.
2162 if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
2164 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2166 // The cost of extracting from the value vector.
2167 Cost += VF * (ExtCost);
2168 // The cost of the scalar stores.
2169 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2170 ValTy->getScalarType(),
2172 SI->getPointerAddressSpace());
2177 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
2178 SI->getPointerAddressSpace());
2180 case Instruction::Load: {
2181 LoadInst *LI = cast<LoadInst>(I);
2184 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
2186 LI->getPointerAddressSpace());
2188 // Scalarized loads.
2189 if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
2191 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
2192 // The cost of inserting the loaded value into the result vector.
2193 Cost += VF * (InCost);
2194 // The cost of the scalar stores.
2195 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
2196 RetTy->getScalarType(),
2198 LI->getPointerAddressSpace());
2203 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2204 LI->getPointerAddressSpace());
2206 case Instruction::ZExt:
2207 case Instruction::SExt:
2208 case Instruction::FPToUI:
2209 case Instruction::FPToSI:
2210 case Instruction::FPExt:
2211 case Instruction::PtrToInt:
2212 case Instruction::IntToPtr:
2213 case Instruction::SIToFP:
2214 case Instruction::UIToFP:
2215 case Instruction::Trunc:
2216 case Instruction::FPTrunc:
2217 case Instruction::BitCast: {
2218 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2219 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
2222 // We are scalarizing the instruction. Return the cost of the scalar
2223 // instruction, plus the cost of insert and extract into vector
2224 // elements, times the vector width.
2227 bool IsVoid = RetTy->isVoidTy();
2229 unsigned InsCost = (IsVoid ? 0 :
2230 VTTI->getInstrCost(Instruction::InsertElement,
2233 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
2236 // The cost of inserting the results plus extracting each one of the
2238 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
2240 // The cost of executing VF copies of the scalar instruction.
2241 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
2247 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
2248 if (Scalar->isVoidTy() || VF == 1)
2250 return VectorType::get(Scalar, VF);
2255 char LoopVectorize::ID = 0;
2256 static const char lv_name[] = "Loop Vectorization";
2257 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
2258 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2259 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2260 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2261 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
2264 Pass *createLoopVectorizePass() {
2265 return new LoopVectorize();