1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR. Legalization of the IR is done
12 // in the codegen. However, the vectorizes uses (will use) the codegen
13 // interfaces to generate IR that is likely to result in an optimal binary.
15 // The loop vectorizer combines consecutive loop iteration into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. SingleBlockLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
28 //===----------------------------------------------------------------------===//
30 // The reduction-variable vectorization is based on the paper:
31 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33 // Variable uniformity checks are inspired by:
34 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
36 // Other ideas/concepts are from:
37 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
39 //===----------------------------------------------------------------------===//
40 #define LV_NAME "loop-vectorize"
41 #define DEBUG_TYPE LV_NAME
42 #include "llvm/Constants.h"
43 #include "llvm/DerivedTypes.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/LLVMContext.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Analysis/LoopPass.h"
48 #include "llvm/Value.h"
49 #include "llvm/Function.h"
50 #include "llvm/Analysis/Verifier.h"
51 #include "llvm/Module.h"
52 #include "llvm/Type.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/StringExtras.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/AliasSetTracker.h"
57 #include "llvm/Analysis/ScalarEvolution.h"
58 #include "llvm/Analysis/Dominators.h"
59 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
60 #include "llvm/Analysis/ScalarEvolutionExpander.h"
61 #include "llvm/Analysis/LoopInfo.h"
62 #include "llvm/Analysis/ValueTracking.h"
63 #include "llvm/Transforms/Scalar.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/TargetTransformInfo.h"
66 #include "llvm/Support/CommandLine.h"
67 #include "llvm/Support/Debug.h"
68 #include "llvm/Support/raw_ostream.h"
69 #include "llvm/DataLayout.h"
70 #include "llvm/Transforms/Utils/Local.h"
74 static cl::opt<unsigned>
75 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
76 cl::desc("Set the default vectorization width. Zero is autoselect."));
78 /// We don't vectorize loops with a known constant trip count below this number.
79 const unsigned TinyTripCountThreshold = 16;
81 /// When performing a runtime memory check, do not check more than this
82 /// numner of pointers. Notice that the check is quadratic!
83 const unsigned RuntimeMemoryCheckThreshold = 2;
87 // Forward declarations.
88 class LoopVectorizationLegality;
89 class LoopVectorizationCostModel;
91 /// SingleBlockLoopVectorizer vectorizes loops which contain only one basic
92 /// block to a specified vectorization factor (VF).
93 /// This class performs the widening of scalars into vectors, or multiple
94 /// scalars. This class also implements the following features:
95 /// * It inserts an epilogue loop for handling loops that don't have iteration
96 /// counts that are known to be a multiple of the vectorization factor.
97 /// * It handles the code generation for reduction variables.
98 /// * Scalarization (implementation using scalars) of un-vectorizable
100 /// SingleBlockLoopVectorizer does not perform any vectorization-legality
101 /// checks, and relies on the caller to check for the different legality
102 /// aspects. The SingleBlockLoopVectorizer relies on the
103 /// LoopVectorizationLegality class to provide information about the induction
104 /// and reduction variables that were found to a given vectorization factor.
105 class SingleBlockLoopVectorizer {
108 SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
109 DominatorTree *dt, LPPassManager *Lpm,
111 OrigLoop(Orig), SE(Se), LI(Li), DT(dt), LPM(Lpm), VF(VecWidth),
112 Builder(Se->getContext()), Induction(0), OldInduction(0) { }
114 // Perform the actual loop widening (vectorization).
115 void vectorize(LoopVectorizationLegality *Legal) {
116 ///Create a new empty loop. Unlink the old loop and connect the new one.
117 createEmptyLoop(Legal);
118 /// Widen each instruction in the old loop to a new one in the new loop.
119 /// Use the Legality module to find the induction and reduction variables.
120 vectorizeLoop(Legal);
121 // register the new loop.
126 /// Create an empty loop, based on the loop ranges of the old loop.
127 void createEmptyLoop(LoopVectorizationLegality *Legal);
128 /// Copy and widen the instructions from the old loop.
129 void vectorizeLoop(LoopVectorizationLegality *Legal);
130 /// Insert the new loop to the loop hierarchy and pass manager.
131 void updateAnalysis();
133 /// This instruction is un-vectorizable. Implement it as a sequence
135 void scalarizeInstruction(Instruction *Instr);
137 /// Create a broadcast instruction. This method generates a broadcast
138 /// instruction (shuffle) for loop invariant values and for the induction
139 /// value. If this is the induction variable then we extend it to N, N+1, ...
140 /// this is needed because each iteration in the loop corresponds to a SIMD
142 Value *getBroadcastInstrs(Value *V);
144 /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
145 /// for each element in the vector. Starting from zero.
146 Value *getConsecutiveVector(Value* Val);
148 /// When we go over instructions in the basic block we rely on previous
149 /// values within the current basic block or on loop invariant values.
150 /// When we widen (vectorize) values we place them in the map. If the values
151 /// are not within the map, they have to be loop invariant, so we simply
152 /// broadcast them into a vector.
153 Value *getVectorValue(Value *V);
155 /// Get a uniform vector of constant integers. We use this to get
156 /// vectors of ones and zeros for the reduction code.
157 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
159 typedef DenseMap<Value*, Value*> ValueMap;
161 /// The original loop.
163 // Scev analysis to use.
169 // Loop Pass Manager;
171 // The vectorization factor to use.
174 // The builder that we use
177 // --- Vectorization state ---
179 /// The vector-loop preheader.
180 BasicBlock *LoopVectorPreHeader;
181 /// The scalar-loop preheader.
182 BasicBlock *LoopScalarPreHeader;
183 /// Middle Block between the vector and the scalar.
184 BasicBlock *LoopMiddleBlock;
185 ///The ExitBlock of the scalar loop.
186 BasicBlock *LoopExitBlock;
187 ///The vector loop body.
188 BasicBlock *LoopVectorBody;
189 ///The scalar loop body.
190 BasicBlock *LoopScalarBody;
191 ///The first bypass block.
192 BasicBlock *LoopBypassBlock;
194 /// The new Induction variable which was added to the new block.
196 /// The induction variable of the old basic block.
197 PHINode *OldInduction;
198 // Maps scalars to widened vectors.
202 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
203 /// to what vectorization factor.
204 /// This class does not look at the profitability of vectorization, only the
205 /// legality. This class has two main kinds of checks:
206 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
207 /// will change the order of memory accesses in a way that will change the
208 /// correctness of the program.
209 /// * Scalars checks - The code in canVectorizeBlock checks for a number
210 /// of different conditions, such as the availability of a single induction
211 /// variable, that all types are supported and vectorize-able, etc.
212 /// This code reflects the capabilities of SingleBlockLoopVectorizer.
213 /// This class is also used by SingleBlockLoopVectorizer for identifying
214 /// induction variable and the different reduction variables.
215 class LoopVectorizationLegality {
217 LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
218 TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { }
220 /// This represents the kinds of reductions that we support.
222 NoReduction, /// Not a reduction.
223 IntegerAdd, /// Sum of numbers.
224 IntegerMult, /// Product of numbers.
225 IntegerOr, /// Bitwise or logical OR of numbers.
226 IntegerAnd, /// Bitwise or logical AND of numbers.
227 IntegerXor /// Bitwise or logical XOR of numbers.
230 /// This POD struct holds information about reduction variables.
231 struct ReductionDescriptor {
233 ReductionDescriptor():
234 StartValue(0), LoopExitInstr(0), Kind(NoReduction) {}
237 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K):
238 StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
240 // The starting value of the reduction.
241 // It does not have to be zero!
243 // The instruction who's value is used outside the loop.
244 Instruction *LoopExitInstr;
245 // The kind of the reduction.
249 // This POD struct holds information about the memory runtime legality
250 // check that a group of pointers do not overlap.
251 struct RuntimePointerCheck {
252 /// This flag indicates if we need to add the runtime check.
254 /// Holds the pointers that we need to check.
255 SmallVector<Value*, 2> Pointers;
258 /// ReductionList contains the reduction descriptors for all
259 /// of the reductions that were found in the loop.
260 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
262 /// Returns true if it is legal to vectorize this loop.
263 /// This does not mean that it is profitable to vectorize this
264 /// loop, only that it is legal to do so.
267 /// Returns the Induction variable.
268 PHINode *getInduction() {return Induction;}
270 /// Returns the reduction variables found in the loop.
271 ReductionList *getReductionVars() { return &Reductions; }
273 /// Check if the pointer returned by this GEP is consecutive
274 /// when the index is vectorized. This happens when the last
275 /// index of the GEP is consecutive, like the induction variable.
276 /// This check allows us to vectorize A[idx] into a wide load/store.
277 bool isConsecutiveGep(Value *Ptr);
279 /// Returns true if the value V is uniform within the loop.
280 bool isUniform(Value *V);
282 /// Returns true if this instruction will remain scalar after vectorization.
283 bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);}
285 /// Returns the information that we collected about runtime memory check.
286 RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; }
288 /// Check if a single basic block loop is vectorizable.
289 /// At this point we know that this is a loop with a constant trip count
290 /// and we only need to check individual instructions.
291 bool canVectorizeBlock(BasicBlock &BB);
293 /// When we vectorize loops we may change the order in which
294 /// we read and write from memory. This method checks if it is
295 /// legal to vectorize the code, considering only memory constrains.
296 /// Returns true if BB is vectorizable
297 bool canVectorizeMemory(BasicBlock &BB);
299 /// Returns True, if 'Phi' is the kind of reduction variable for type
300 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
301 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
302 /// Returns true if the instruction I can be a reduction variable of type
304 bool isReductionInstr(Instruction *I, ReductionKind Kind);
305 /// Returns True, if 'Phi' is an induction variable.
306 bool isInductionVariable(PHINode *Phi);
307 /// Return true if we
308 bool hasComputableBounds(Value *Ptr);
310 /// The loop that we evaluate.
314 /// DataLayout analysis.
317 // --- vectorization state --- //
319 /// Holds the induction variable.
321 /// Holds the reduction variables.
322 ReductionList Reductions;
323 /// Allowed outside users. This holds the reduction
324 /// vars which can be accessed from outside the loop.
325 SmallPtrSet<Value*, 4> AllowedExit;
326 /// This set holds the variables which are known to be uniform after
328 SmallPtrSet<Instruction*, 4> Uniforms;
329 /// We need to check that all of the pointers in this list are disjoint
331 RuntimePointerCheck PtrRtCheck;
334 /// LoopVectorizationCostModel - estimates the expected speedups due to
336 /// In many cases vectorization is not profitable. This can happen because
337 /// of a number of reasons. In this class we mainly attempt to predict
338 /// the expected speedup/slowdowns due to the supported instruction set.
339 /// We use the VectorTargetTransformInfo to query the different backends
340 /// for the cost of different operations.
341 class LoopVectorizationCostModel {
344 LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se,
345 LoopVectorizationLegality *Leg,
346 const VectorTargetTransformInfo *Vtti):
347 TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { }
349 /// Returns the most profitable vectorization factor for the loop that is
350 /// smaller or equal to the VF argument. This method checks every power
352 unsigned findBestVectorizationFactor(unsigned VF = 8);
355 /// Returns the expected execution cost. The unit of the cost does
356 /// not matter because we use the 'cost' units to compare different
357 /// vector widths. The cost that is returned is *not* normalized by
358 /// the factor width.
359 unsigned expectedCost(unsigned VF);
361 /// Returns the execution time cost of an instruction for a given vector
362 /// width. Vector width of one means scalar.
363 unsigned getInstructionCost(Instruction *I, unsigned VF);
365 /// A helper function for converting Scalar types to vector types.
366 /// If the incoming type is void, we return void. If the VF is 1, we return
368 static Type* ToVectorTy(Type *Scalar, unsigned VF);
370 /// The loop that we evaluate.
375 /// Vectorization legality.
376 LoopVectorizationLegality *Legal;
377 /// Vector target information.
378 const VectorTargetTransformInfo *VTTI;
381 struct LoopVectorize : public LoopPass {
382 static char ID; // Pass identification, replacement for typeid
384 LoopVectorize() : LoopPass(ID) {
385 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
391 TargetTransformInfo *TTI;
394 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
395 // We only vectorize innermost loops.
399 SE = &getAnalysis<ScalarEvolution>();
400 DL = getAnalysisIfAvailable<DataLayout>();
401 LI = &getAnalysis<LoopInfo>();
402 TTI = getAnalysisIfAvailable<TargetTransformInfo>();
403 DT = &getAnalysis<DominatorTree>();
405 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
406 L->getHeader()->getParent()->getName() << "\"\n");
408 // Check if it is legal to vectorize the loop.
409 LoopVectorizationLegality LVL(L, SE, DL);
410 if (!LVL.canVectorize()) {
411 DEBUG(dbgs() << "LV: Not vectorizing.\n");
415 // Select the preffered vectorization factor.
417 if (VectorizationFactor == 0) {
418 const VectorTargetTransformInfo *VTTI = 0;
420 VTTI = TTI->getVectorTargetTransformInfo();
421 // Use the cost model.
422 LoopVectorizationCostModel CM(L, SE, &LVL, VTTI);
423 VF = CM.findBestVectorizationFactor();
426 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
431 // Use the user command flag.
432 VF = VectorizationFactor;
435 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
436 L->getHeader()->getParent()->getParent()->getModuleIdentifier()<<
439 // If we decided that it is *legal* to vectorizer the loop then do it.
440 SingleBlockLoopVectorizer LB(L, SE, LI, DT, &LPM, VF);
443 DEBUG(verifyFunction(*L->getHeader()->getParent()));
447 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
448 LoopPass::getAnalysisUsage(AU);
449 AU.addRequiredID(LoopSimplifyID);
450 AU.addRequiredID(LCSSAID);
451 AU.addRequired<LoopInfo>();
452 AU.addRequired<ScalarEvolution>();
453 AU.addRequired<DominatorTree>();
454 AU.addPreserved<LoopInfo>();
455 AU.addPreserved<DominatorTree>();
460 Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
461 // Instructions that access the old induction variable
462 // actually want to get the new one.
463 if (V == OldInduction)
466 LLVMContext &C = V->getContext();
467 Type *VTy = VectorType::get(V->getType(), VF);
468 Type *I32 = IntegerType::getInt32Ty(C);
469 Constant *Zero = ConstantInt::get(I32, 0);
470 Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
471 Value *UndefVal = UndefValue::get(VTy);
472 // Insert the value into a new vector.
473 Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero);
474 // Broadcast the scalar into all locations in the vector.
475 Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros,
477 // We are accessing the induction variable. Make sure to promote the
478 // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
480 return getConsecutiveVector(Shuf);
484 Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
485 assert(Val->getType()->isVectorTy() && "Must be a vector");
486 assert(Val->getType()->getScalarType()->isIntegerTy() &&
487 "Elem must be an integer");
489 Type *ITy = Val->getType()->getScalarType();
490 VectorType *Ty = cast<VectorType>(Val->getType());
491 unsigned VLen = Ty->getNumElements();
492 SmallVector<Constant*, 8> Indices;
494 // Create a vector of consecutive numbers from zero to VF.
495 for (unsigned i = 0; i < VLen; ++i)
496 Indices.push_back(ConstantInt::get(ITy, i));
498 // Add the consecutive indices to the vector value.
499 Constant *Cv = ConstantVector::get(Indices);
500 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
501 return Builder.CreateAdd(Val, Cv, "induction");
504 bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) {
505 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
509 unsigned NumOperands = Gep->getNumOperands();
510 Value *LastIndex = Gep->getOperand(NumOperands - 1);
512 // Check that all of the gep indices are uniform except for the last.
513 for (unsigned i = 0; i < NumOperands - 1; ++i)
514 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
517 // We can emit wide load/stores only of the last index is the induction
519 const SCEV *Last = SE->getSCEV(LastIndex);
520 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
521 const SCEV *Step = AR->getStepRecurrence(*SE);
523 // The memory is consecutive because the last index is consecutive
524 // and all other indices are loop invariant.
532 bool LoopVectorizationLegality::isUniform(Value *V) {
533 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
536 Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
537 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
538 // If we saved a vectorized copy of V, use it.
539 Value *&MapEntry = WidenMap[V];
543 // Broadcast V and save the value for future uses.
544 Value *B = getBroadcastInstrs(V);
550 SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
551 SmallVector<Constant*, 8> Indices;
552 // Create a vector of consecutive numbers from zero to VF.
553 for (unsigned i = 0; i < VF; ++i)
554 Indices.push_back(ConstantInt::get(ScalarTy, Val, true));
556 // Add the consecutive indices to the vector value.
557 return ConstantVector::get(Indices);
560 void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
561 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
562 // Holds vector parameters or scalars, in case of uniform vals.
563 SmallVector<Value*, 8> Params;
565 // Find all of the vectorized parameters.
566 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
567 Value *SrcOp = Instr->getOperand(op);
569 // If we are accessing the old induction variable, use the new one.
570 if (SrcOp == OldInduction) {
571 Params.push_back(getBroadcastInstrs(Induction));
575 // Try using previously calculated values.
576 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
578 // If the src is an instruction that appeared earlier in the basic block
579 // then it should already be vectorized.
580 if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
581 assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
582 // The parameter is a vector value from earlier.
583 Params.push_back(WidenMap[SrcInst]);
585 // The parameter is a scalar from outside the loop. Maybe even a constant.
586 Params.push_back(SrcOp);
590 assert(Params.size() == Instr->getNumOperands() &&
591 "Invalid number of operands");
593 // Does this instruction return a value ?
594 bool IsVoidRetTy = Instr->getType()->isVoidTy();
595 Value *VecResults = 0;
597 // If we have a return value, create an empty vector. We place the scalarized
598 // instructions in this vector.
600 VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
602 // For each scalar that we create:
603 for (unsigned i = 0; i < VF; ++i) {
604 Instruction *Cloned = Instr->clone();
606 Cloned->setName(Instr->getName() + ".cloned");
607 // Replace the operands of the cloned instrucions with extracted scalars.
608 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
609 Value *Op = Params[op];
610 // Param is a vector. Need to extract the right lane.
611 if (Op->getType()->isVectorTy())
612 Op = Builder.CreateExtractElement(Op, Builder.getInt32(i));
613 Cloned->setOperand(op, Op);
616 // Place the cloned scalar in the new loop.
617 Builder.Insert(Cloned);
619 // If the original scalar returns a value we need to place it in a vector
620 // so that future users will be able to use it.
622 VecResults = Builder.CreateInsertElement(VecResults, Cloned,
623 Builder.getInt32(i));
627 WidenMap[Instr] = VecResults;
631 SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
633 In this function we generate a new loop. The new loop will contain
634 the vectorized instructions while the old loop will continue to run the
637 [ ] <-- vector loop bypass.
640 | [ ] <-- vector pre header.
644 | [ ]_| <-- vector loop.
647 >[ ] <--- middle-block.
650 | [ ] <--- new preheader.
654 | [ ]_| <-- old scalar loop to handle remainder.
661 OldInduction = Legal->getInduction();
662 assert(OldInduction && "We must have a single phi node.");
663 Type *IdxTy = OldInduction->getType();
665 // Find the loop boundaries.
666 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
667 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
669 // Get the total trip count from the count by adding 1.
670 ExitCount = SE->getAddExpr(ExitCount,
671 SE->getConstant(ExitCount->getType(), 1));
672 // We may need to extend the index in case there is a type mismatch.
673 // We know that the count starts at zero and does not overflow.
674 // We are using Zext because it should be less expensive.
675 if (ExitCount->getType() != IdxTy)
676 ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy);
678 // This is the original scalar-loop preheader.
679 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
680 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
681 assert(ExitBlock && "Must have an exit block");
683 // The loop index does not have to start at Zero. It starts with this value.
684 Value *StartIdx = OldInduction->getIncomingValueForBlock(BypassBlock);
686 assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop");
687 assert(BypassBlock && "Invalid loop structure");
689 BasicBlock *VectorPH =
690 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
691 BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(),
694 BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(),
696 BasicBlock *ScalarPH =
697 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(),
699 // Find the induction variable.
700 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
702 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
704 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
706 // Generate the induction variable.
707 Induction = Builder.CreatePHI(IdxTy, 2, "index");
708 Constant *Step = ConstantInt::get(IdxTy, VF);
710 // Expand the trip count and place the new instructions in the preheader.
711 // Notice that the pre-header does not change, only the loop body.
712 SCEVExpander Exp(*SE, "induction");
713 Instruction *Loc = BypassBlock->getTerminator();
715 // Count holds the overall loop count (N).
716 Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc);
718 // Add the start index to the loop count to get the new end index.
719 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
721 // Now we need to generate the expression for N - (N % VF), which is
722 // the part that the vectorized body will execute.
723 Constant *CIVF = ConstantInt::get(IdxTy, VF);
724 Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
725 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
726 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
727 "end.idx.rnd.down", Loc);
729 // Now, compare the new count to zero. If it is zero, jump to the scalar part.
730 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
735 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
736 Legal->getRuntimePointerCheck();
737 Value *MemoryRuntimeCheck = 0;
738 if (PtrRtCheck->Need) {
739 unsigned NumPointers = PtrRtCheck->Pointers.size();
740 SmallVector<Value* , 2> Starts;
741 SmallVector<Value* , 2> Ends;
743 // Use this type for pointer arithmetic.
744 Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType();
746 for (unsigned i=0; i < NumPointers; ++i) {
747 Value *Ptr = PtrRtCheck->Pointers[i];
748 const SCEV *Sc = SE->getSCEV(Ptr);
750 if (SE->isLoopInvariant(Sc, OrigLoop)) {
751 DEBUG(dbgs() << "LV1: Adding RT check for a loop invariant ptr:" <<
753 Starts.push_back(Ptr);
756 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
757 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
758 Value *Start = Exp.expandCodeFor(AR->getStart(), PtrArithTy, Loc);
759 const SCEV *Ex = SE->getExitCount(OrigLoop, OrigLoop->getHeader());
760 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
761 assert(!isa<SCEVCouldNotCompute>(ScEnd) && "Invalid scev range.");
762 Value *End = Exp.expandCodeFor(ScEnd, PtrArithTy, Loc);
763 Starts.push_back(Start);
768 for (unsigned i=0; i < NumPointers; ++i) {
769 for (unsigned j=i+1; j < NumPointers; ++j) {
770 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
771 Starts[0], Ends[1], "bound0", Loc);
772 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
773 Starts[1], Ends[0], "bound1", Loc);
774 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
775 "found.conflict", Loc);
776 if (MemoryRuntimeCheck) {
777 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
780 "conflict.rdx", Loc);
782 MemoryRuntimeCheck = IsConflict;
786 }// end of need-runtime-check code.
788 // If we are using memory runtime checks, include them in.
789 if (MemoryRuntimeCheck) {
790 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
794 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
795 // Remove the old terminator.
796 Loc->eraseFromParent();
798 // We are going to resume the execution of the scalar loop.
799 // This PHI decides on what number to start. If we come from the
800 // vector loop then we need to start with the end index minus the
801 // index modulo VF. If we come from a bypass edge then we need to start
802 // from the real start.
803 PHINode* ResumeIndex = PHINode::Create(IdxTy, 2, "resume.idx",
804 MiddleBlock->getTerminator());
805 ResumeIndex->addIncoming(StartIdx, BypassBlock);
806 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
808 // Add a check in the middle block to see if we have completed
809 // all of the iterations in the first vector loop.
810 // If (N - N%VF) == N, then we *don't* need to run the remainder.
811 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
812 ResumeIndex, "cmp.n",
813 MiddleBlock->getTerminator());
815 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
816 // Remove the old terminator.
817 MiddleBlock->getTerminator()->eraseFromParent();
819 // Create i+1 and fill the PHINode.
820 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
821 Induction->addIncoming(StartIdx, VectorPH);
822 Induction->addIncoming(NextIdx, VecBody);
823 // Create the compare.
824 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
825 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
827 // Now we have two terminators. Remove the old one from the block.
828 VecBody->getTerminator()->eraseFromParent();
830 // Fix the scalar body iteration count.
831 unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH);
832 OldInduction->setIncomingValue(BlockIdx, ResumeIndex);
834 // Get ready to start creating new instructions into the vectorized body.
835 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
837 // Register the new loop.
838 Loop* Lp = new Loop();
839 LPM->insertLoop(Lp, OrigLoop->getParentLoop());
841 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
843 Loop *ParentLoop = OrigLoop->getParentLoop();
845 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
846 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
847 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
851 LoopVectorPreHeader = VectorPH;
852 LoopScalarPreHeader = ScalarPH;
853 LoopMiddleBlock = MiddleBlock;
854 LoopExitBlock = ExitBlock;
855 LoopVectorBody = VecBody;
856 LoopScalarBody = OldBasicBlock;
857 LoopBypassBlock = BypassBlock;
860 /// This function returns the identity element (or neutral element) for
863 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
865 case LoopVectorizationLegality::IntegerXor:
866 case LoopVectorizationLegality::IntegerAdd:
867 case LoopVectorizationLegality::IntegerOr:
868 // Adding, Xoring, Oring zero to a number does not change it.
870 case LoopVectorizationLegality::IntegerMult:
871 // Multiplying a number by 1 does not change it.
873 case LoopVectorizationLegality::IntegerAnd:
874 // AND-ing a number with an all-1 value does not change it.
877 llvm_unreachable("Unknown reduction kind");
882 SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
883 //===------------------------------------------------===//
885 // Notice: any optimization or new instruction that go
886 // into the code below should be also be implemented in
889 //===------------------------------------------------===//
890 typedef SmallVector<PHINode*, 4> PhiVector;
891 BasicBlock &BB = *OrigLoop->getHeader();
892 Constant *Zero = ConstantInt::get(
893 IntegerType::getInt32Ty(BB.getContext()), 0);
895 // In order to support reduction variables we need to be able to vectorize
896 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
897 // steages. First, we create a new vector PHI node with no incoming edges.
898 // We use this value when we vectorize all of the instructions that use the
899 // PHI. Next, after all of the instructions in the block are complete we
900 // add the new incoming edges to the PHI. At this point all of the
901 // instructions in the basic block are vectorized, so we can use them to
902 // construct the PHI.
905 // For each instruction in the old loop.
906 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
907 Instruction *Inst = it;
909 switch (Inst->getOpcode()) {
910 case Instruction::Br:
911 // Nothing to do for PHIs and BR, since we already took care of the
912 // loop control flow instructions.
914 case Instruction::PHI:{
915 PHINode* P = cast<PHINode>(Inst);
916 // Special handling for the induction var.
917 if (OldInduction == Inst)
919 // This is phase one of vectorizing PHIs.
920 // This has to be a reduction variable.
921 assert(Legal->getReductionVars()->count(P) && "Not a Reduction");
922 Type *VecTy = VectorType::get(Inst->getType(), VF);
923 WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi");
924 PHIsToFix.push_back(P);
927 case Instruction::Add:
928 case Instruction::FAdd:
929 case Instruction::Sub:
930 case Instruction::FSub:
931 case Instruction::Mul:
932 case Instruction::FMul:
933 case Instruction::UDiv:
934 case Instruction::SDiv:
935 case Instruction::FDiv:
936 case Instruction::URem:
937 case Instruction::SRem:
938 case Instruction::FRem:
939 case Instruction::Shl:
940 case Instruction::LShr:
941 case Instruction::AShr:
942 case Instruction::And:
943 case Instruction::Or:
944 case Instruction::Xor: {
945 // Just widen binops.
946 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
947 Value *A = getVectorValue(Inst->getOperand(0));
948 Value *B = getVectorValue(Inst->getOperand(1));
950 // Use this vector value for all users of the original instruction.
951 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
954 // Update the NSW, NUW and Exact flags.
955 BinaryOperator *VecOp = cast<BinaryOperator>(V);
956 if (isa<OverflowingBinaryOperator>(BinOp)) {
957 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
958 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
960 if (isa<PossiblyExactOperator>(VecOp))
961 VecOp->setIsExact(BinOp->isExact());
964 case Instruction::Select: {
966 // If the selector is loop invariant we can create a select
967 // instruction with a scalar condition. Otherwise, use vector-select.
968 Value *Cond = Inst->getOperand(0);
969 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop);
971 // The condition can be loop invariant but still defined inside the
972 // loop. This means that we can't just use the original 'cond' value.
973 // We have to take the 'vectorized' value and pick the first lane.
974 // Instcombine will make this a no-op.
975 Cond = getVectorValue(Cond);
977 Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0));
979 Value *Op0 = getVectorValue(Inst->getOperand(1));
980 Value *Op1 = getVectorValue(Inst->getOperand(2));
981 WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1);
985 case Instruction::ICmp:
986 case Instruction::FCmp: {
987 // Widen compares. Generate vector compares.
988 bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
989 CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
990 Value *A = getVectorValue(Inst->getOperand(0));
991 Value *B = getVectorValue(Inst->getOperand(1));
993 WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
995 WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B);
999 case Instruction::Store: {
1000 // Attempt to issue a wide store.
1001 StoreInst *SI = dyn_cast<StoreInst>(Inst);
1002 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1003 Value *Ptr = SI->getPointerOperand();
1004 unsigned Alignment = SI->getAlignment();
1006 assert(!Legal->isUniform(Ptr) &&
1007 "We do not allow storing to uniform addresses");
1009 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1011 // This store does not use GEPs.
1012 if (!Legal->isConsecutiveGep(Gep)) {
1013 scalarizeInstruction(Inst);
1017 // The last index does not have to be the induction. It can be
1018 // consecutive and be a function of the index. For example A[I+1];
1019 unsigned NumOperands = Gep->getNumOperands();
1020 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
1021 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1023 // Create the new GEP with the new induction variable.
1024 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1025 Gep2->setOperand(NumOperands - 1, LastIndex);
1026 Ptr = Builder.Insert(Gep2);
1027 Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo());
1028 Value *Val = getVectorValue(SI->getValueOperand());
1029 Builder.CreateStore(Val, Ptr)->setAlignment(Alignment);
1032 case Instruction::Load: {
1033 // Attempt to issue a wide load.
1034 LoadInst *LI = dyn_cast<LoadInst>(Inst);
1035 Type *RetTy = VectorType::get(LI->getType(), VF);
1036 Value *Ptr = LI->getPointerOperand();
1037 unsigned Alignment = LI->getAlignment();
1038 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1040 // If we don't have a gep, or that the pointer is loop invariant,
1041 // scalarize the load.
1042 if (!Gep || Legal->isUniform(Gep) || !Legal->isConsecutiveGep(Gep)) {
1043 scalarizeInstruction(Inst);
1047 // The last index does not have to be the induction. It can be
1048 // consecutive and be a function of the index. For example A[I+1];
1049 unsigned NumOperands = Gep->getNumOperands();
1050 Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
1051 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1053 // Create the new GEP with the new induction variable.
1054 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1055 Gep2->setOperand(NumOperands - 1, LastIndex);
1056 Ptr = Builder.Insert(Gep2);
1057 Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo());
1058 LI = Builder.CreateLoad(Ptr);
1059 LI->setAlignment(Alignment);
1060 // Use this vector value for all users of the load.
1061 WidenMap[Inst] = LI;
1064 case Instruction::ZExt:
1065 case Instruction::SExt:
1066 case Instruction::FPToUI:
1067 case Instruction::FPToSI:
1068 case Instruction::FPExt:
1069 case Instruction::PtrToInt:
1070 case Instruction::IntToPtr:
1071 case Instruction::SIToFP:
1072 case Instruction::UIToFP:
1073 case Instruction::Trunc:
1074 case Instruction::FPTrunc:
1075 case Instruction::BitCast: {
1076 /// Vectorize bitcasts.
1077 CastInst *CI = dyn_cast<CastInst>(Inst);
1078 Value *A = getVectorValue(Inst->getOperand(0));
1079 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1080 WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
1085 /// All other instructions are unsupported. Scalarize them.
1086 scalarizeInstruction(Inst);
1089 }// end of for_each instr.
1091 // At this point every instruction in the original loop is widended to
1092 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1093 // that we vectorized. The PHI nodes are currently empty because we did
1094 // not want to introduce cycles. Notice that the remaining PHI nodes
1095 // that we need to fix are reduction variables.
1097 // Create the 'reduced' values for each of the induction vars.
1098 // The reduced values are the vector values that we scalarize and combine
1099 // after the loop is finished.
1100 for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end();
1102 PHINode *RdxPhi = *it;
1103 PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
1104 assert(RdxPhi && "Unable to recover vectorized PHI");
1106 // Find the reduction variable descriptor.
1107 assert(Legal->getReductionVars()->count(RdxPhi) &&
1108 "Unable to find the reduction variable");
1109 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1110 (*Legal->getReductionVars())[RdxPhi];
1112 // We need to generate a reduction vector from the incoming scalar.
1113 // To do so, we need to generate the 'identity' vector and overide
1114 // one of the elements with the incoming scalar reduction. We need
1115 // to do it in the vector-loop preheader.
1116 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1118 // This is the vector-clone of the value that leaves the loop.
1119 Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1120 Type *VecTy = VectorExit->getType();
1122 // Find the reduction identity variable. Zero for addition, or, xor,
1123 // one for multiplication, -1 for And.
1124 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1125 VecTy->getScalarType());
1127 // This vector is the Identity vector where the first element is the
1128 // incoming scalar reduction.
1129 Value *VectorStart = Builder.CreateInsertElement(Identity,
1130 RdxDesc.StartValue, Zero);
1133 // Fix the vector-loop phi.
1134 // We created the induction variable so we know that the
1135 // preheader is the first entry.
1136 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1138 // Reductions do not have to start at zero. They can start with
1139 // any loop invariant values.
1140 VecRdxPhi->addIncoming(VectorStart, VecPreheader);
1141 unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1142 Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx));
1143 VecRdxPhi->addIncoming(Val, LoopVectorBody);
1145 // Before each round, move the insertion point right between
1146 // the PHIs and the values we are going to write.
1147 // This allows us to write both PHINodes and the extractelement
1149 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1151 // This PHINode contains the vectorized reduction variable, or
1152 // the initial value vector, if we bypass the vector loop.
1153 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1154 NewPhi->addIncoming(VectorStart, LoopBypassBlock);
1155 NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody);
1157 // Extract the first scalar.
1159 Builder.CreateExtractElement(NewPhi, Builder.getInt32(0));
1160 // Extract and reduce the remaining vector elements.
1161 for (unsigned i=1; i < VF; ++i) {
1163 Builder.CreateExtractElement(NewPhi, Builder.getInt32(i));
1164 switch (RdxDesc.Kind) {
1165 case LoopVectorizationLegality::IntegerAdd:
1166 Scalar0 = Builder.CreateAdd(Scalar0, Scalar1);
1168 case LoopVectorizationLegality::IntegerMult:
1169 Scalar0 = Builder.CreateMul(Scalar0, Scalar1);
1171 case LoopVectorizationLegality::IntegerOr:
1172 Scalar0 = Builder.CreateOr(Scalar0, Scalar1);
1174 case LoopVectorizationLegality::IntegerAnd:
1175 Scalar0 = Builder.CreateAnd(Scalar0, Scalar1);
1177 case LoopVectorizationLegality::IntegerXor:
1178 Scalar0 = Builder.CreateXor(Scalar0, Scalar1);
1181 llvm_unreachable("Unknown reduction operation");
1185 // Now, we need to fix the users of the reduction variable
1186 // inside and outside of the scalar remainder loop.
1187 // We know that the loop is in LCSSA form. We need to update the
1188 // PHI nodes in the exit blocks.
1189 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1190 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1191 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1192 if (!LCSSAPhi) continue;
1194 // All PHINodes need to have a single entry edge, or two if
1195 // we already fixed them.
1196 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1198 // We found our reduction value exit-PHI. Update it with the
1199 // incoming bypass edge.
1200 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1201 // Add an edge coming from the bypass.
1202 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1205 }// end of the LCSSA phi scan.
1207 // Fix the scalar loop reduction variable with the incoming reduction sum
1208 // from the vector body and from the backedge value.
1209 int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody);
1210 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block.
1211 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1212 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1213 }// end of for each redux variable.
1216 void SingleBlockLoopVectorizer::updateAnalysis() {
1217 // The original basic block.
1218 SE->forgetLoop(OrigLoop);
1220 // Update the dominator tree information.
1221 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1222 "Entry does not dominate exit.");
1224 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1225 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1226 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1227 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1228 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1229 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1231 DEBUG(DT->verifyAnalysis());
1234 bool LoopVectorizationLegality::canVectorize() {
1235 if (!TheLoop->getLoopPreheader()) {
1236 assert(false && "No preheader!!");
1237 DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
1241 // We can only vectorize single basic block loops.
1242 unsigned NumBlocks = TheLoop->getNumBlocks();
1243 if (NumBlocks != 1) {
1244 DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
1248 // We need to have a loop header.
1249 BasicBlock *BB = TheLoop->getHeader();
1250 DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
1252 // ScalarEvolution needs to be able to find the exit count.
1253 const SCEV *ExitCount = SE->getExitCount(TheLoop, BB);
1254 if (ExitCount == SE->getCouldNotCompute()) {
1255 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
1259 // Do not loop-vectorize loops with a tiny trip count.
1260 unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB);
1261 if (TC > 0u && TC < TinyTripCountThreshold) {
1262 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
1263 "This loop is not worth vectorizing.\n");
1267 // Go over each instruction and look at memory deps.
1268 if (!canVectorizeBlock(*BB)) {
1269 DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
1273 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
1274 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
1277 // Okay! We can vectorize. At this point we don't have any other mem analysis
1278 // which may limit our maximum vectorization factor, so just return true with
1283 bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
1284 // Scan the instructions in the block and look for hazards.
1285 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1286 Instruction *I = it;
1288 PHINode *Phi = dyn_cast<PHINode>(I);
1290 // This should not happen because the loop should be normalized.
1291 if (Phi->getNumIncomingValues() != 2) {
1292 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
1295 // We only look at integer phi nodes.
1296 if (!Phi->getType()->isIntegerTy()) {
1297 DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
1301 if (isInductionVariable(Phi)) {
1303 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
1306 DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n");
1310 if (AddReductionVar(Phi, IntegerAdd)) {
1311 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
1314 if (AddReductionVar(Phi, IntegerMult)) {
1315 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
1318 if (AddReductionVar(Phi, IntegerOr)) {
1319 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
1322 if (AddReductionVar(Phi, IntegerAnd)) {
1323 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
1326 if (AddReductionVar(Phi, IntegerXor)) {
1327 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
1331 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
1333 }// end of PHI handling
1335 // We still don't handle functions.
1336 CallInst *CI = dyn_cast<CallInst>(I);
1338 DEBUG(dbgs() << "LV: Found a call site.\n");
1342 // We do not re-vectorize vectors.
1343 if (!VectorType::isValidElementType(I->getType()) &&
1344 !I->getType()->isVoidTy()) {
1345 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
1349 // Reduction instructions are allowed to have exit users.
1350 // All other instructions must not have external users.
1351 if (!AllowedExit.count(I))
1352 //Check that all of the users of the loop are inside the BB.
1353 for (Value::use_iterator it = I->use_begin(), e = I->use_end();
1355 Instruction *U = cast<Instruction>(*it);
1356 // This user may be a reduction exit value.
1357 BasicBlock *Parent = U->getParent();
1358 if (Parent != &BB) {
1359 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
1366 DEBUG(dbgs() << "LV: Did not find an induction var.\n");
1370 // Don't vectorize if the memory dependencies do not allow vectorization.
1371 if (!canVectorizeMemory(BB))
1374 // We now know that the loop is vectorizable!
1375 // Collect variables that will remain uniform after vectorization.
1376 std::vector<Value*> Worklist;
1378 // Start with the conditional branch and walk up the block.
1379 Worklist.push_back(BB.getTerminator()->getOperand(0));
1381 while (Worklist.size()) {
1382 Instruction *I = dyn_cast<Instruction>(Worklist.back());
1383 Worklist.pop_back();
1384 // Look at instructions inside this block.
1386 if (I->getParent() != &BB) continue;
1388 // Stop when reaching PHI nodes.
1389 if (isa<PHINode>(I)) {
1390 assert(I == Induction && "Found a uniform PHI that is not the induction");
1394 // This is a known uniform.
1397 // Insert all operands.
1398 for (int i=0, Op = I->getNumOperands(); i < Op; ++i) {
1399 Worklist.push_back(I->getOperand(i));
1406 bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) {
1407 typedef SmallVector<Value*, 16> ValueVector;
1408 typedef SmallPtrSet<Value*, 16> ValueSet;
1409 // Holds the Load and Store *instructions*.
1412 PtrRtCheck.Pointers.clear();
1413 PtrRtCheck.Need = false;
1415 // Scan the BB and collect legal loads and stores.
1416 for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
1417 Instruction *I = it;
1419 // If this is a load, save it. If this instruction can read from memory
1420 // but is not a load, then we quit. Notice that we don't handle function
1421 // calls that read or write.
1422 if (I->mayReadFromMemory()) {
1423 LoadInst *Ld = dyn_cast<LoadInst>(I);
1424 if (!Ld) return false;
1425 if (!Ld->isSimple()) {
1426 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
1429 Loads.push_back(Ld);
1433 // Save store instructions. Abort if other instructions write to memory.
1434 if (I->mayWriteToMemory()) {
1435 StoreInst *St = dyn_cast<StoreInst>(I);
1436 if (!St) return false;
1437 if (!St->isSimple()) {
1438 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
1441 Stores.push_back(St);
1445 // Now we have two lists that hold the loads and the stores.
1446 // Next, we find the pointers that they use.
1448 // Check if we see any stores. If there are no stores, then we don't
1449 // care if the pointers are *restrict*.
1450 if (!Stores.size()) {
1451 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
1455 // Holds the read and read-write *pointers* that we find.
1457 ValueVector ReadWrites;
1459 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1460 // multiple times on the same object. If the ptr is accessed twice, once
1461 // for read and once for write, it will only appear once (on the write
1462 // list). This is okay, since we are going to check for conflicts between
1463 // writes and between reads and writes, but not between reads and reads.
1466 ValueVector::iterator I, IE;
1467 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1468 StoreInst *ST = dyn_cast<StoreInst>(*I);
1469 assert(ST && "Bad StoreInst");
1470 Value* Ptr = ST->getPointerOperand();
1472 if (isUniform(Ptr)) {
1473 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
1477 // If we did *not* see this pointer before, insert it to
1478 // the read-write list. At this phase it is only a 'write' list.
1479 if (Seen.insert(Ptr))
1480 ReadWrites.push_back(Ptr);
1483 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1484 LoadInst *LD = dyn_cast<LoadInst>(*I);
1485 assert(LD && "Bad LoadInst");
1486 Value* Ptr = LD->getPointerOperand();
1487 // If we did *not* see this pointer before, insert it to the
1488 // read list. If we *did* see it before, then it is already in
1489 // the read-write list. This allows us to vectorize expressions
1490 // such as A[i] += x; Because the address of A[i] is a read-write
1491 // pointer. This only works if the index of A[i] is consecutive.
1492 // If the address of i is unknown (for example A[B[i]]) then we may
1493 // read a few words, modify, and write a few words, and some of the
1494 // words may be written to the same address.
1495 if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr))
1496 Reads.push_back(Ptr);
1499 // If we write (or read-write) to a single destination and there are no
1500 // other reads in this loop then is it safe to vectorize.
1501 if (ReadWrites.size() == 1 && Reads.size() == 0) {
1502 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1506 // Find pointers with computable bounds. We are going to use this information
1507 // to place a runtime bound check.
1509 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
1510 if (hasComputableBounds(*I)) {
1511 PtrRtCheck.Pointers.push_back(*I);
1512 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1517 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
1518 if (hasComputableBounds(*I)) {
1519 PtrRtCheck.Pointers.push_back(*I);
1520 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
1526 // Check that we did not collect too many pointers or found a
1527 // unsizeable pointer.
1528 if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
1529 PtrRtCheck.Pointers.clear();
1533 PtrRtCheck.Need = RT;
1536 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1539 // Now that the pointers are in two lists (Reads and ReadWrites), we
1540 // can check that there are no conflicts between each of the writes and
1541 // between the writes to the reads.
1542 ValueSet WriteObjects;
1543 ValueVector TempObjects;
1545 // Check that the read-writes do not conflict with other read-write
1547 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
1548 GetUnderlyingObjects(*I, TempObjects, DL);
1549 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1551 if (!isIdentifiedObject(*it)) {
1552 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
1555 if (!WriteObjects.insert(*it)) {
1556 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
1561 TempObjects.clear();
1564 /// Check that the reads don't conflict with the read-writes.
1565 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
1566 GetUnderlyingObjects(*I, TempObjects, DL);
1567 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
1569 if (!isIdentifiedObject(*it)) {
1570 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
1573 if (WriteObjects.count(*it)) {
1574 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
1579 TempObjects.clear();
1582 // It is safe to vectorize and we don't need any runtime checks.
1583 DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n");
1584 PtrRtCheck.Pointers.clear();
1585 PtrRtCheck.Need = false;
1589 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
1590 ReductionKind Kind) {
1591 if (Phi->getNumIncomingValues() != 2)
1594 // Find the possible incoming reduction variable.
1595 BasicBlock *BB = Phi->getParent();
1596 int SelfEdgeIdx = Phi->getBasicBlockIndex(BB);
1597 int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry.
1598 Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx);
1600 // ExitInstruction is the single value which is used outside the loop.
1601 // We only allow for a single reduction value to be used outside the loop.
1602 // This includes users of the reduction, variables (which form a cycle
1603 // which ends in the phi node).
1604 Instruction *ExitInstruction = 0;
1606 // Iter is our iterator. We start with the PHI node and scan for all of the
1607 // users of this instruction. All users must be instructions which can be
1608 // used as reduction variables (such as ADD). We may have a single
1609 // out-of-block user. They cycle must end with the original PHI.
1610 // Also, we can't have multiple block-local users.
1611 Instruction *Iter = Phi;
1613 // Any reduction instr must be of one of the allowed kinds.
1614 if (!isReductionInstr(Iter, Kind))
1617 // Did we found a user inside this block ?
1618 bool FoundInBlockUser = false;
1619 // Did we reach the initial PHI node ?
1620 bool FoundStartPHI = false;
1622 // If the instruction has no users then this is a broken
1623 // chain and can't be a reduction variable.
1624 if (Iter->use_empty())
1627 // For each of the *users* of iter.
1628 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
1630 Instruction *U = cast<Instruction>(*it);
1631 // We already know that the PHI is a user.
1633 FoundStartPHI = true;
1636 // Check if we found the exit user.
1637 BasicBlock *Parent = U->getParent();
1639 // We must have a single exit instruction.
1640 if (ExitInstruction != 0)
1642 ExitInstruction = Iter;
1644 // We can't have multiple inside users.
1645 if (FoundInBlockUser)
1647 FoundInBlockUser = true;
1651 // We found a reduction var if we have reached the original
1652 // phi node and we only have a single instruction with out-of-loop
1654 if (FoundStartPHI && ExitInstruction) {
1655 // This instruction is allowed to have out-of-loop users.
1656 AllowedExit.insert(ExitInstruction);
1658 // Save the description of this reduction variable.
1659 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
1660 Reductions[Phi] = RD;
1667 LoopVectorizationLegality::isReductionInstr(Instruction *I,
1668 ReductionKind Kind) {
1669 switch (I->getOpcode()) {
1672 case Instruction::PHI:
1675 case Instruction::Add:
1676 case Instruction::Sub:
1677 return Kind == IntegerAdd;
1678 case Instruction::Mul:
1679 case Instruction::UDiv:
1680 case Instruction::SDiv:
1681 return Kind == IntegerMult;
1682 case Instruction::And:
1683 return Kind == IntegerAnd;
1684 case Instruction::Or:
1685 return Kind == IntegerOr;
1686 case Instruction::Xor:
1687 return Kind == IntegerXor;
1691 bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
1692 // Check that the PHI is consecutive and starts at zero.
1693 const SCEV *PhiScev = SE->getSCEV(Phi);
1694 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1696 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1699 const SCEV *Step = AR->getStepRecurrence(*SE);
1701 if (!Step->isOne()) {
1702 DEBUG(dbgs() << "LV: PHI stride does not equal one.\n");
1708 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
1709 const SCEV *PhiScev = SE->getSCEV(Ptr);
1710 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1714 return AR->isAffine();
1718 LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) {
1720 DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
1724 float Cost = expectedCost(1);
1726 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
1727 for (unsigned i=2; i <= VF; i*=2) {
1728 // Notice that the vector loop needs to be executed less times, so
1729 // we need to divide the cost of the vector loops by the width of
1730 // the vector elements.
1731 float VectorCost = expectedCost(i) / (float)i;
1732 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
1733 (int)VectorCost << ".\n");
1734 if (VectorCost < Cost) {
1740 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
1744 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
1745 // We can only estimate the cost of single basic block loops.
1746 assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop");
1748 BasicBlock *BB = TheLoop->getHeader();
1751 // For each instruction in the old loop.
1752 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1753 Instruction *Inst = it;
1754 unsigned C = getInstructionCost(Inst, VF);
1756 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF <<
1757 " For instruction: "<< *Inst << "\n");
1764 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
1765 assert(VTTI && "Invalid vector target transformation info");
1767 // If we know that this instruction will remain uniform, check the cost of
1768 // the scalar version.
1769 if (Legal->isUniformAfterVectorization(I))
1772 Type *RetTy = I->getType();
1773 Type *VectorTy = ToVectorTy(RetTy, VF);
1776 // TODO: We need to estimate the cost of intrinsic calls.
1777 switch (I->getOpcode()) {
1778 case Instruction::GetElementPtr:
1779 // We mark this instruction as zero-cost because scalar GEPs are usually
1780 // lowered to the intruction addressing mode. At the moment we don't
1781 // generate vector geps.
1783 case Instruction::Br: {
1784 return VTTI->getCFInstrCost(I->getOpcode());
1786 case Instruction::PHI:
1788 case Instruction::Add:
1789 case Instruction::FAdd:
1790 case Instruction::Sub:
1791 case Instruction::FSub:
1792 case Instruction::Mul:
1793 case Instruction::FMul:
1794 case Instruction::UDiv:
1795 case Instruction::SDiv:
1796 case Instruction::FDiv:
1797 case Instruction::URem:
1798 case Instruction::SRem:
1799 case Instruction::FRem:
1800 case Instruction::Shl:
1801 case Instruction::LShr:
1802 case Instruction::AShr:
1803 case Instruction::And:
1804 case Instruction::Or:
1805 case Instruction::Xor: {
1806 return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
1808 case Instruction::Select: {
1809 SelectInst *SI = cast<SelectInst>(I);
1810 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
1811 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
1812 Type *CondTy = SI->getCondition()->getType();
1814 CondTy = VectorType::get(CondTy, VF);
1816 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
1818 case Instruction::ICmp:
1819 case Instruction::FCmp: {
1820 Type *ValTy = I->getOperand(0)->getType();
1821 VectorTy = ToVectorTy(ValTy, VF);
1822 return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
1824 case Instruction::Store: {
1825 StoreInst *SI = cast<StoreInst>(I);
1826 Type *ValTy = SI->getValueOperand()->getType();
1827 VectorTy = ToVectorTy(ValTy, VF);
1830 return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
1831 SI->getAlignment(), SI->getPointerAddressSpace());
1833 // Scalarized stores.
1834 if (!Legal->isConsecutiveGep(SI->getPointerOperand())) {
1836 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1838 // The cost of extracting from the value vector.
1839 Cost += VF * (ExtCost);
1840 // The cost of the scalar stores.
1841 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1842 ValTy->getScalarType(),
1844 SI->getPointerAddressSpace());
1849 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(),
1850 SI->getPointerAddressSpace());
1852 case Instruction::Load: {
1853 LoadInst *LI = cast<LoadInst>(I);
1856 return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
1858 LI->getPointerAddressSpace());
1860 // Scalarized loads.
1861 if (!Legal->isConsecutiveGep(LI->getPointerOperand())) {
1863 unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
1864 // The cost of inserting the loaded value into the result vector.
1865 Cost += VF * (InCost);
1866 // The cost of the scalar stores.
1867 Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
1868 RetTy->getScalarType(),
1870 LI->getPointerAddressSpace());
1875 return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
1876 LI->getPointerAddressSpace());
1878 case Instruction::ZExt:
1879 case Instruction::SExt:
1880 case Instruction::FPToUI:
1881 case Instruction::FPToSI:
1882 case Instruction::FPExt:
1883 case Instruction::PtrToInt:
1884 case Instruction::IntToPtr:
1885 case Instruction::SIToFP:
1886 case Instruction::UIToFP:
1887 case Instruction::Trunc:
1888 case Instruction::FPTrunc:
1889 case Instruction::BitCast: {
1890 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
1891 return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
1894 // We are scalarizing the instruction. Return the cost of the scalar
1895 // instruction, plus the cost of insert and extract into vector
1896 // elements, times the vector width.
1899 bool IsVoid = RetTy->isVoidTy();
1901 unsigned InsCost = (IsVoid ? 0 :
1902 VTTI->getInstrCost(Instruction::InsertElement,
1905 unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
1908 // The cost of inserting the results plus extracting each one of the
1910 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
1912 // The cost of executing VF copies of the scalar instruction.
1913 Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
1919 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
1920 if (Scalar->isVoidTy() || VF == 1)
1922 return VectorType::get(Scalar, VF);
1927 char LoopVectorize::ID = 0;
1928 static const char lv_name[] = "Loop Vectorization";
1929 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
1930 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1931 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1932 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1933 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
1936 Pass *createLoopVectorizePass() {
1937 return new LoopVectorize();