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 //===----------------------------------------------------------------------===//
45 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/MapVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.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/Dominators.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopIterator.h"
60 #include "llvm/Analysis/LoopPass.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpander.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/TargetTransformInfo.h"
65 #include "llvm/Analysis/ValueTracking.h"
66 #include "llvm/Analysis/Verifier.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Function.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/LLVMContext.h"
75 #include "llvm/IR/Module.h"
76 #include "llvm/IR/Type.h"
77 #include "llvm/IR/Value.h"
78 #include "llvm/Pass.h"
79 #include "llvm/Support/CommandLine.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/raw_ostream.h"
82 #include "llvm/Transforms/Scalar.h"
83 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
84 #include "llvm/Transforms/Utils/Local.h"
90 static cl::opt<unsigned>
91 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
92 cl::desc("Sets the SIMD width. Zero is autoselect."));
94 static cl::opt<unsigned>
95 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
96 cl::desc("Sets the vectorization unroll count. "
97 "Zero is autoselect."));
100 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
101 cl::desc("Enable if-conversion during vectorization."));
103 /// We don't vectorize loops with a known constant trip count below this number.
104 static const unsigned TinyTripCountThreshold = 16;
106 /// When performing a runtime memory check, do not check more than this
107 /// number of pointers. Notice that the check is quadratic!
108 static const unsigned RuntimeMemoryCheckThreshold = 4;
110 /// This is the highest vector width that we try to generate.
111 static const unsigned MaxVectorSize = 8;
113 /// This is the highest Unroll Factor.
114 static const unsigned MaxUnrollSize = 4;
118 // Forward declarations.
119 class LoopVectorizationLegality;
120 class LoopVectorizationCostModel;
122 /// InnerLoopVectorizer vectorizes loops which contain only one basic
123 /// block to a specified vectorization factor (VF).
124 /// This class performs the widening of scalars into vectors, or multiple
125 /// scalars. This class also implements the following features:
126 /// * It inserts an epilogue loop for handling loops that don't have iteration
127 /// counts that are known to be a multiple of the vectorization factor.
128 /// * It handles the code generation for reduction variables.
129 /// * Scalarization (implementation using scalars) of un-vectorizable
131 /// InnerLoopVectorizer does not perform any vectorization-legality
132 /// checks, and relies on the caller to check for the different legality
133 /// aspects. The InnerLoopVectorizer relies on the
134 /// LoopVectorizationLegality class to provide information about the induction
135 /// and reduction variables that were found to a given vectorization factor.
136 class InnerLoopVectorizer {
138 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
139 DominatorTree *DT, DataLayout *DL, unsigned VecWidth,
140 unsigned UnrollFactor)
141 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), VF(VecWidth),
142 UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
143 OldInduction(0), WidenMap(UnrollFactor) {}
145 // Perform the actual loop widening (vectorization).
146 void vectorize(LoopVectorizationLegality *Legal) {
147 // Create a new empty loop. Unlink the old loop and connect the new one.
148 createEmptyLoop(Legal);
149 // Widen each instruction in the old loop to a new one in the new loop.
150 // Use the Legality module to find the induction and reduction variables.
151 vectorizeLoop(Legal);
152 // Register the new loop and update the analysis passes.
157 /// A small list of PHINodes.
158 typedef SmallVector<PHINode*, 4> PhiVector;
159 /// When we unroll loops we have multiple vector values for each scalar.
160 /// This data structure holds the unrolled and vectorized values that
161 /// originated from one scalar instruction.
162 typedef SmallVector<Value*, 2> VectorParts;
164 /// Add code that checks at runtime if the accessed arrays overlap.
165 /// Returns the comparator value or NULL if no check is needed.
166 Value *addRuntimeCheck(LoopVectorizationLegality *Legal,
168 /// Create an empty loop, based on the loop ranges of the old loop.
169 void createEmptyLoop(LoopVectorizationLegality *Legal);
170 /// Copy and widen the instructions from the old loop.
171 void vectorizeLoop(LoopVectorizationLegality *Legal);
173 /// A helper function that computes the predicate of the block BB, assuming
174 /// that the header block of the loop is set to True. It returns the *entry*
175 /// mask for the block BB.
176 VectorParts createBlockInMask(BasicBlock *BB);
177 /// A helper function that computes the predicate of the edge between SRC
179 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
181 /// A helper function to vectorize a single BB within the innermost loop.
182 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
185 /// Insert the new loop to the loop hierarchy and pass manager
186 /// and update the analysis passes.
187 void updateAnalysis();
189 /// This instruction is un-vectorizable. Implement it as a sequence
191 void scalarizeInstruction(Instruction *Instr);
193 /// Create a broadcast instruction. This method generates a broadcast
194 /// instruction (shuffle) for loop invariant values and for the induction
195 /// value. If this is the induction variable then we extend it to N, N+1, ...
196 /// this is needed because each iteration in the loop corresponds to a SIMD
198 Value *getBroadcastInstrs(Value *V);
200 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
201 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
202 /// The sequence starts at StartIndex.
203 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
205 /// When we go over instructions in the basic block we rely on previous
206 /// values within the current basic block or on loop invariant values.
207 /// When we widen (vectorize) values we place them in the map. If the values
208 /// are not within the map, they have to be loop invariant, so we simply
209 /// broadcast them into a vector.
210 VectorParts &getVectorValue(Value *V);
212 /// Get a uniform vector of constant integers. We use this to get
213 /// vectors of ones and zeros for the reduction code.
214 Constant* getUniformVector(unsigned Val, Type* ScalarTy);
216 /// Generate a shuffle sequence that will reverse the vector Vec.
217 Value *reverseVector(Value *Vec);
219 /// This is a helper class that holds the vectorizer state. It maps scalar
220 /// instructions to vector instructions. When the code is 'unrolled' then
221 /// then a single scalar value is mapped to multiple vector parts. The parts
222 /// are stored in the VectorPart type.
224 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
226 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
228 /// \return True if 'Key' is saved in the Value Map.
229 bool has(Value *Key) { return MapStoreage.count(Key); }
231 /// Initializes a new entry in the map. Sets all of the vector parts to the
232 /// save value in 'Val'.
233 /// \return A reference to a vector with splat values.
234 VectorParts &splat(Value *Key, Value *Val) {
235 MapStoreage[Key].clear();
236 MapStoreage[Key].append(UF, Val);
237 return MapStoreage[Key];
240 ///\return A reference to the value that is stored at 'Key'.
241 VectorParts &get(Value *Key) {
243 MapStoreage[Key].resize(UF);
244 return MapStoreage[Key];
247 /// The unroll factor. Each entry in the map stores this number of vector
251 /// Map storage. We use std::map and not DenseMap because insertions to a
252 /// dense map invalidates its iterators.
253 std::map<Value*, VectorParts> MapStoreage;
256 /// The original loop.
258 /// Scev analysis to use.
266 /// The vectorization SIMD factor to use. Each vector will have this many
269 /// The vectorization unroll factor to use. Each scalar is vectorized to this
270 /// many different vector instructions.
273 /// The builder that we use
276 // --- Vectorization state ---
278 /// The vector-loop preheader.
279 BasicBlock *LoopVectorPreHeader;
280 /// The scalar-loop preheader.
281 BasicBlock *LoopScalarPreHeader;
282 /// Middle Block between the vector and the scalar.
283 BasicBlock *LoopMiddleBlock;
284 ///The ExitBlock of the scalar loop.
285 BasicBlock *LoopExitBlock;
286 ///The vector loop body.
287 BasicBlock *LoopVectorBody;
288 ///The scalar loop body.
289 BasicBlock *LoopScalarBody;
290 ///The first bypass block.
291 BasicBlock *LoopBypassBlock;
293 /// The new Induction variable which was added to the new block.
295 /// The induction variable of the old basic block.
296 PHINode *OldInduction;
297 /// Maps scalars to widened vectors.
301 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
302 /// to what vectorization factor.
303 /// This class does not look at the profitability of vectorization, only the
304 /// legality. This class has two main kinds of checks:
305 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
306 /// will change the order of memory accesses in a way that will change the
307 /// correctness of the program.
308 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
309 /// checks for a number of different conditions, such as the availability of a
310 /// single induction variable, that all types are supported and vectorize-able,
311 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
312 /// This class is also used by InnerLoopVectorizer for identifying
313 /// induction variable and the different reduction variables.
314 class LoopVectorizationLegality {
316 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
318 : TheLoop(L), SE(SE), DL(DL), DT(DT), Induction(0) {}
320 /// This enum represents the kinds of reductions that we support.
322 NoReduction, ///< Not a reduction.
323 IntegerAdd, ///< Sum of numbers.
324 IntegerMult, ///< Product of numbers.
325 IntegerOr, ///< Bitwise or logical OR of numbers.
326 IntegerAnd, ///< Bitwise or logical AND of numbers.
327 IntegerXor ///< Bitwise or logical XOR of numbers.
330 /// This enum represents the kinds of inductions that we support.
332 NoInduction, ///< Not an induction variable.
333 IntInduction, ///< Integer induction variable. Step = 1.
334 ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
335 PtrInduction ///< Pointer induction variable. Step = sizeof(elem).
338 /// This POD struct holds information about reduction variables.
339 struct ReductionDescriptor {
340 ReductionDescriptor() : StartValue(0), LoopExitInstr(0), Kind(NoReduction) {
343 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K)
344 : StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
346 // The starting value of the reduction.
347 // It does not have to be zero!
349 // The instruction who's value is used outside the loop.
350 Instruction *LoopExitInstr;
351 // The kind of the reduction.
355 // This POD struct holds information about the memory runtime legality
356 // check that a group of pointers do not overlap.
357 struct RuntimePointerCheck {
358 RuntimePointerCheck() : Need(false) {}
360 /// Reset the state of the pointer runtime information.
368 /// Insert a pointer and calculate the start and end SCEVs.
369 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr);
371 /// This flag indicates if we need to add the runtime check.
373 /// Holds the pointers that we need to check.
374 SmallVector<Value*, 2> Pointers;
375 /// Holds the pointer value at the beginning of the loop.
376 SmallVector<const SCEV*, 2> Starts;
377 /// Holds the pointer value at the end of the loop.
378 SmallVector<const SCEV*, 2> Ends;
381 /// A POD for saving information about induction variables.
382 struct InductionInfo {
383 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
384 InductionInfo() : StartValue(0), IK(NoInduction) {}
391 /// ReductionList contains the reduction descriptors for all
392 /// of the reductions that were found in the loop.
393 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
395 /// InductionList saves induction variables and maps them to the
396 /// induction descriptor.
397 typedef MapVector<PHINode*, InductionInfo> InductionList;
399 /// Returns true if it is legal to vectorize this loop.
400 /// This does not mean that it is profitable to vectorize this
401 /// loop, only that it is legal to do so.
404 /// Returns the Induction variable.
405 PHINode *getInduction() { return Induction; }
407 /// Returns the reduction variables found in the loop.
408 ReductionList *getReductionVars() { return &Reductions; }
410 /// Returns the induction variables found in the loop.
411 InductionList *getInductionVars() { return &Inductions; }
413 /// Returns True if V is an induction variable in this loop.
414 bool isInductionVariable(const Value *V);
416 /// Return true if the block BB needs to be predicated in order for the loop
417 /// to be vectorized.
418 bool blockNeedsPredication(BasicBlock *BB);
420 /// Check if this pointer is consecutive when vectorizing. This happens
421 /// when the last index of the GEP is the induction variable, or that the
422 /// pointer itself is an induction variable.
423 /// This check allows us to vectorize A[idx] into a wide load/store.
425 /// 0 - Stride is unknown or non consecutive.
426 /// 1 - Address is consecutive.
427 /// -1 - Address is consecutive, and decreasing.
428 int isConsecutivePtr(Value *Ptr);
430 /// Returns true if the value V is uniform within the loop.
431 bool isUniform(Value *V);
433 /// Returns true if this instruction will remain scalar after vectorization.
434 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
436 /// Returns the information that we collected about runtime memory check.
437 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
439 /// Check if a single basic block loop is vectorizable.
440 /// At this point we know that this is a loop with a constant trip count
441 /// and we only need to check individual instructions.
442 bool canVectorizeInstrs();
444 /// When we vectorize loops we may change the order in which
445 /// we read and write from memory. This method checks if it is
446 /// legal to vectorize the code, considering only memory constrains.
447 /// Returns true if the loop is vectorizable
448 bool canVectorizeMemory();
450 /// Return true if we can vectorize this loop using the IF-conversion
452 bool canVectorizeWithIfConvert();
454 /// Collect the variables that need to stay uniform after vectorization.
455 void collectLoopUniforms();
457 /// Return true if all of the instructions in the block can be speculatively
459 bool blockCanBePredicated(BasicBlock *BB);
461 /// Returns True, if 'Phi' is the kind of reduction variable for type
462 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
463 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
464 /// Returns true if the instruction I can be a reduction variable of type
466 bool isReductionInstr(Instruction *I, ReductionKind Kind);
467 /// Returns the induction kind of Phi. This function may return NoInduction
468 /// if the PHI is not an induction variable.
469 InductionKind isInductionVariable(PHINode *Phi);
470 /// Return true if can compute the address bounds of Ptr within the loop.
471 bool hasComputableBounds(Value *Ptr);
473 /// The loop that we evaluate.
477 /// DataLayout analysis.
482 // --- vectorization state --- //
484 /// Holds the integer induction variable. This is the counter of the
487 /// Holds the reduction variables.
488 ReductionList Reductions;
489 /// Holds all of the induction variables that we found in the loop.
490 /// Notice that inductions don't need to start at zero and that induction
491 /// variables can be pointers.
492 InductionList Inductions;
494 /// Allowed outside users. This holds the reduction
495 /// vars which can be accessed from outside the loop.
496 SmallPtrSet<Value*, 4> AllowedExit;
497 /// This set holds the variables which are known to be uniform after
499 SmallPtrSet<Instruction*, 4> Uniforms;
500 /// We need to check that all of the pointers in this list are disjoint
502 RuntimePointerCheck PtrRtCheck;
505 /// LoopVectorizationCostModel - estimates the expected speedups due to
507 /// In many cases vectorization is not profitable. This can happen because of
508 /// a number of reasons. In this class we mainly attempt to predict the
509 /// expected speedup/slowdowns due to the supported instruction set. We use the
510 /// TargetTransformInfo to query the different backends for the cost of
511 /// different operations.
512 class LoopVectorizationCostModel {
514 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
515 LoopVectorizationLegality *Legal,
516 const TargetTransformInfo &TTI)
517 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI) {}
519 /// \return The most profitable vectorization factor.
520 /// This method checks every power of two up to VF. If UserVF is not ZERO
521 /// then this vectorization factor will be selected if vectorization is
523 unsigned selectVectorizationFactor(bool OptForSize, unsigned UserVF);
526 /// \return The most profitable unroll factor.
527 /// If UserUF is non-zero then this method finds the best unroll-factor
528 /// based on register pressure and other parameters.
529 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF);
531 /// \brief A struct that represents some properties of the register usage
533 struct RegisterUsage {
534 /// Holds the number of loop invariant values that are used in the loop.
535 unsigned LoopInvariantRegs;
536 /// Holds the maximum number of concurrent live intervals in the loop.
537 unsigned MaxLocalUsers;
538 /// Holds the number of instructions in the loop.
539 unsigned NumInstructions;
542 /// \return information about the register usage of the loop.
543 RegisterUsage calculateRegisterUsage();
546 /// Returns the expected execution cost. The unit of the cost does
547 /// not matter because we use the 'cost' units to compare different
548 /// vector widths. The cost that is returned is *not* normalized by
549 /// the factor width.
550 unsigned expectedCost(unsigned VF);
552 /// Returns the execution time cost of an instruction for a given vector
553 /// width. Vector width of one means scalar.
554 unsigned getInstructionCost(Instruction *I, unsigned VF);
556 /// A helper function for converting Scalar types to vector types.
557 /// If the incoming type is void, we return void. If the VF is 1, we return
559 static Type* ToVectorTy(Type *Scalar, unsigned VF);
561 /// The loop that we evaluate.
565 /// Loop Info analysis.
567 /// Vectorization legality.
568 LoopVectorizationLegality *Legal;
569 /// Vector target information.
570 const TargetTransformInfo &TTI;
573 /// The LoopVectorize Pass.
574 struct LoopVectorize : public LoopPass {
575 /// Pass identification, replacement for typeid
578 explicit LoopVectorize() : LoopPass(ID) {
579 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
585 TargetTransformInfo *TTI;
588 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
589 // We only vectorize innermost loops.
593 SE = &getAnalysis<ScalarEvolution>();
594 DL = getAnalysisIfAvailable<DataLayout>();
595 LI = &getAnalysis<LoopInfo>();
596 TTI = &getAnalysis<TargetTransformInfo>();
597 DT = &getAnalysis<DominatorTree>();
599 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
600 L->getHeader()->getParent()->getName() << "\"\n");
602 // Check if it is legal to vectorize the loop.
603 LoopVectorizationLegality LVL(L, SE, DL, DT);
604 if (!LVL.canVectorize()) {
605 DEBUG(dbgs() << "LV: Not vectorizing.\n");
609 // Use the cost model.
610 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI);
612 // Check the function attribues to find out if this function should be
613 // optimized for size.
614 Function *F = L->getHeader()->getParent();
615 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
616 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
617 unsigned FnIndex = AttributeSet::FunctionIndex;
618 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
619 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
622 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
623 "attribute is used.\n");
627 unsigned VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
628 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll);
631 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
635 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
636 F->getParent()->getModuleIdentifier()<<"\n");
637 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
639 // If we decided that it is *legal* to vectorizer the loop then do it.
640 InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF, UF);
643 DEBUG(verifyFunction(*L->getHeader()->getParent()));
647 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
648 LoopPass::getAnalysisUsage(AU);
649 AU.addRequiredID(LoopSimplifyID);
650 AU.addRequiredID(LCSSAID);
651 AU.addRequired<DominatorTree>();
652 AU.addRequired<LoopInfo>();
653 AU.addRequired<ScalarEvolution>();
654 AU.addRequired<TargetTransformInfo>();
655 AU.addPreserved<LoopInfo>();
656 AU.addPreserved<DominatorTree>();
661 } // end anonymous namespace
663 //===----------------------------------------------------------------------===//
664 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
665 // LoopVectorizationCostModel.
666 //===----------------------------------------------------------------------===//
669 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
670 Loop *Lp, Value *Ptr) {
671 const SCEV *Sc = SE->getSCEV(Ptr);
672 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
673 assert(AR && "Invalid addrec expression");
674 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
675 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
676 Pointers.push_back(Ptr);
677 Starts.push_back(AR->getStart());
678 Ends.push_back(ScEnd);
681 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
682 // Save the current insertion location.
683 Instruction *Loc = Builder.GetInsertPoint();
685 // We need to place the broadcast of invariant variables outside the loop.
686 Instruction *Instr = dyn_cast<Instruction>(V);
687 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
688 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
690 // Place the code for broadcasting invariant variables in the new preheader.
692 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
694 // Broadcast the scalar into all locations in the vector.
695 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
697 // Restore the builder insertion point.
699 Builder.SetInsertPoint(Loc);
704 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
706 assert(Val->getType()->isVectorTy() && "Must be a vector");
707 assert(Val->getType()->getScalarType()->isIntegerTy() &&
708 "Elem must be an integer");
710 Type *ITy = Val->getType()->getScalarType();
711 VectorType *Ty = cast<VectorType>(Val->getType());
712 int VLen = Ty->getNumElements();
713 SmallVector<Constant*, 8> Indices;
715 // Create a vector of consecutive numbers from zero to VF.
716 for (int i = 0; i < VLen; ++i) {
717 int Idx = Negate ? (-i): i;
718 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
721 // Add the consecutive indices to the vector value.
722 Constant *Cv = ConstantVector::get(Indices);
723 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
724 return Builder.CreateAdd(Val, Cv, "induction");
727 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
728 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
730 // If this value is a pointer induction variable we know it is consecutive.
731 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
732 if (Phi && Inductions.count(Phi)) {
733 InductionInfo II = Inductions[Phi];
734 if (PtrInduction == II.IK)
738 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
742 unsigned NumOperands = Gep->getNumOperands();
743 Value *LastIndex = Gep->getOperand(NumOperands - 1);
745 // Check that all of the gep indices are uniform except for the last.
746 for (unsigned i = 0; i < NumOperands - 1; ++i)
747 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
750 // We can emit wide load/stores only if the last index is the induction
752 const SCEV *Last = SE->getSCEV(LastIndex);
753 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
754 const SCEV *Step = AR->getStepRecurrence(*SE);
756 // The memory is consecutive because the last index is consecutive
757 // and all other indices are loop invariant.
760 if (Step->isAllOnesValue())
767 bool LoopVectorizationLegality::isUniform(Value *V) {
768 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
771 InnerLoopVectorizer::VectorParts&
772 InnerLoopVectorizer::getVectorValue(Value *V) {
773 assert(V != Induction && "The new induction variable should not be used.");
774 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
776 // If we have this scalar in the map, return it.
778 return WidenMap.get(V);
780 // If this scalar is unknown, assume that it is a constant or that it is
781 // loop invariant. Broadcast V and save the value for future uses.
782 Value *B = getBroadcastInstrs(V);
783 WidenMap.splat(V, B);
784 return WidenMap.get(V);
788 InnerLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
789 return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
792 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
793 assert(Vec->getType()->isVectorTy() && "Invalid type");
794 SmallVector<Constant*, 8> ShuffleMask;
795 for (unsigned i = 0; i < VF; ++i)
796 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
798 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
799 ConstantVector::get(ShuffleMask),
803 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
804 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
805 // Holds vector parameters or scalars, in case of uniform vals.
806 SmallVector<VectorParts, 4> Params;
808 // Find all of the vectorized parameters.
809 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
810 Value *SrcOp = Instr->getOperand(op);
812 // If we are accessing the old induction variable, use the new one.
813 if (SrcOp == OldInduction) {
814 Params.push_back(getVectorValue(SrcOp));
818 // Try using previously calculated values.
819 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
821 // If the src is an instruction that appeared earlier in the basic block
822 // then it should already be vectorized.
823 if (SrcInst && OrigLoop->contains(SrcInst)) {
824 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
825 // The parameter is a vector value from earlier.
826 Params.push_back(WidenMap.get(SrcInst));
828 // The parameter is a scalar from outside the loop. Maybe even a constant.
830 Scalars.append(UF, SrcOp);
831 Params.push_back(Scalars);
835 assert(Params.size() == Instr->getNumOperands() &&
836 "Invalid number of operands");
838 // Does this instruction return a value ?
839 bool IsVoidRetTy = Instr->getType()->isVoidTy();
841 Value *UndefVec = IsVoidRetTy ? 0 :
842 UndefValue::get(VectorType::get(Instr->getType(), VF));
843 // Create a new entry in the WidenMap and initialize it to Undef or Null.
844 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
846 // For each scalar that we create:
847 for (unsigned Width = 0; Width < VF; ++Width) {
848 // For each vector unroll 'part':
849 for (unsigned Part = 0; Part < UF; ++Part) {
850 Instruction *Cloned = Instr->clone();
852 Cloned->setName(Instr->getName() + ".cloned");
853 // Replace the operands of the cloned instrucions with extracted scalars.
854 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
855 Value *Op = Params[op][Part];
856 // Param is a vector. Need to extract the right lane.
857 if (Op->getType()->isVectorTy())
858 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
859 Cloned->setOperand(op, Op);
862 // Place the cloned scalar in the new loop.
863 Builder.Insert(Cloned);
865 // If the original scalar returns a value we need to place it in a vector
866 // so that future users will be able to use it.
868 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
869 Builder.getInt32(Width));
875 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
877 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
878 Legal->getRuntimePointerCheck();
880 if (!PtrRtCheck->Need)
883 Value *MemoryRuntimeCheck = 0;
884 unsigned NumPointers = PtrRtCheck->Pointers.size();
885 SmallVector<Value* , 2> Starts;
886 SmallVector<Value* , 2> Ends;
888 SCEVExpander Exp(*SE, "induction");
890 // Use this type for pointer arithmetic.
891 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
893 for (unsigned i = 0; i < NumPointers; ++i) {
894 Value *Ptr = PtrRtCheck->Pointers[i];
895 const SCEV *Sc = SE->getSCEV(Ptr);
897 if (SE->isLoopInvariant(Sc, OrigLoop)) {
898 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
900 Starts.push_back(Ptr);
903 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
905 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
906 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
907 Starts.push_back(Start);
912 for (unsigned i = 0; i < NumPointers; ++i) {
913 for (unsigned j = i+1; j < NumPointers; ++j) {
914 Instruction::CastOps Op = Instruction::BitCast;
915 Value *Start0 = CastInst::Create(Op, Starts[i], PtrArithTy, "bc", Loc);
916 Value *Start1 = CastInst::Create(Op, Starts[j], PtrArithTy, "bc", Loc);
917 Value *End0 = CastInst::Create(Op, Ends[i], PtrArithTy, "bc", Loc);
918 Value *End1 = CastInst::Create(Op, Ends[j], PtrArithTy, "bc", Loc);
920 Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
921 Start0, End1, "bound0", Loc);
922 Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
923 Start1, End0, "bound1", Loc);
924 Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
925 "found.conflict", Loc);
926 if (MemoryRuntimeCheck)
927 MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
930 "conflict.rdx", Loc);
932 MemoryRuntimeCheck = IsConflict;
937 return MemoryRuntimeCheck;
941 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
943 In this function we generate a new loop. The new loop will contain
944 the vectorized instructions while the old loop will continue to run the
947 [ ] <-- vector loop bypass.
950 | [ ] <-- vector pre header.
954 | [ ]_| <-- vector loop.
957 >[ ] <--- middle-block.
960 | [ ] <--- new preheader.
964 | [ ]_| <-- old scalar loop to handle remainder.
971 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
972 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
973 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
974 assert(ExitBlock && "Must have an exit block");
976 // Some loops have a single integer induction variable, while other loops
977 // don't. One example is c++ iterators that often have multiple pointer
978 // induction variables. In the code below we also support a case where we
979 // don't have a single induction variable.
980 OldInduction = Legal->getInduction();
981 Type *IdxTy = OldInduction ? OldInduction->getType() :
982 DL->getIntPtrType(SE->getContext());
984 // Find the loop boundaries.
985 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
986 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
988 // Get the total trip count from the count by adding 1.
989 ExitCount = SE->getAddExpr(ExitCount,
990 SE->getConstant(ExitCount->getType(), 1));
992 // Expand the trip count and place the new instructions in the preheader.
993 // Notice that the pre-header does not change, only the loop body.
994 SCEVExpander Exp(*SE, "induction");
996 // Count holds the overall loop count (N).
997 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
998 BypassBlock->getTerminator());
1000 // The loop index does not have to start at Zero. Find the original start
1001 // value from the induction PHI node. If we don't have an induction variable
1002 // then we know that it starts at zero.
1003 Value *StartIdx = OldInduction ?
1004 OldInduction->getIncomingValueForBlock(BypassBlock):
1005 ConstantInt::get(IdxTy, 0);
1007 assert(BypassBlock && "Invalid loop structure");
1009 // Generate the code that checks in runtime if arrays overlap.
1010 Value *MemoryRuntimeCheck = addRuntimeCheck(Legal,
1011 BypassBlock->getTerminator());
1013 // Split the single block loop into the two loop structure described above.
1014 BasicBlock *VectorPH =
1015 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1016 BasicBlock *VecBody =
1017 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1018 BasicBlock *MiddleBlock =
1019 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1020 BasicBlock *ScalarPH =
1021 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1023 // This is the location in which we add all of the logic for bypassing
1024 // the new vector loop.
1025 Instruction *Loc = BypassBlock->getTerminator();
1027 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1029 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1031 // Generate the induction variable.
1032 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1033 // The loop step is equal to the vectorization factor (num of SIMD elements)
1034 // times the unroll factor (num of SIMD instructions).
1035 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1037 // We may need to extend the index in case there is a type mismatch.
1038 // We know that the count starts at zero and does not overflow.
1039 if (Count->getType() != IdxTy) {
1040 // The exit count can be of pointer type. Convert it to the correct
1042 if (ExitCount->getType()->isPointerTy())
1043 Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
1045 Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
1048 // Add the start index to the loop count to get the new end index.
1049 Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
1051 // Now we need to generate the expression for N - (N % VF), which is
1052 // the part that the vectorized body will execute.
1053 Value *R = BinaryOperator::CreateURem(Count, Step, "n.mod.vf", Loc);
1054 Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
1055 Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
1056 "end.idx.rnd.down", Loc);
1058 // Now, compare the new count to zero. If it is zero skip the vector loop and
1059 // jump to the scalar loop.
1060 Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
1065 // If we are using memory runtime checks, include them in.
1066 if (MemoryRuntimeCheck)
1067 Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
1070 BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
1071 // Remove the old terminator.
1072 Loc->eraseFromParent();
1074 // We are going to resume the execution of the scalar loop.
1075 // Go over all of the induction variables that we found and fix the
1076 // PHIs that are left in the scalar version of the loop.
1077 // The starting values of PHI nodes depend on the counter of the last
1078 // iteration in the vectorized loop.
1079 // If we come from a bypass edge then we need to start from the original
1082 // This variable saves the new starting index for the scalar loop.
1083 PHINode *ResumeIndex = 0;
1084 LoopVectorizationLegality::InductionList::iterator I, E;
1085 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1086 for (I = List->begin(), E = List->end(); I != E; ++I) {
1087 PHINode *OrigPhi = I->first;
1088 LoopVectorizationLegality::InductionInfo II = I->second;
1089 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1090 MiddleBlock->getTerminator());
1091 Value *EndValue = 0;
1093 case LoopVectorizationLegality::NoInduction:
1094 llvm_unreachable("Unknown induction");
1095 case LoopVectorizationLegality::IntInduction: {
1096 // Handle the integer induction counter:
1097 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1098 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1099 // We know what the end value is.
1100 EndValue = IdxEndRoundDown;
1101 // We also know which PHI node holds it.
1102 ResumeIndex = ResumeVal;
1105 case LoopVectorizationLegality::ReverseIntInduction: {
1106 // Convert the CountRoundDown variable to the PHI size.
1107 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1108 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1109 Value *CRD = CountRoundDown;
1110 if (CRDSize > IISize)
1111 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1112 II.StartValue->getType(),
1113 "tr.crd", BypassBlock->getTerminator());
1114 else if (CRDSize < IISize)
1115 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1116 II.StartValue->getType(),
1117 "sext.crd", BypassBlock->getTerminator());
1118 // Handle reverse integer induction counter:
1119 EndValue = BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1120 BypassBlock->getTerminator());
1123 case LoopVectorizationLegality::PtrInduction: {
1124 // For pointer induction variables, calculate the offset using
1126 EndValue = GetElementPtrInst::Create(II.StartValue, CountRoundDown,
1128 BypassBlock->getTerminator());
1133 // The new PHI merges the original incoming value, in case of a bypass,
1134 // or the value at the end of the vectorized loop.
1135 ResumeVal->addIncoming(II.StartValue, BypassBlock);
1136 ResumeVal->addIncoming(EndValue, VecBody);
1138 // Fix the scalar body counter (PHI node).
1139 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1140 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1143 // If we are generating a new induction variable then we also need to
1144 // generate the code that calculates the exit value. This value is not
1145 // simply the end of the counter because we may skip the vectorized body
1146 // in case of a runtime check.
1148 assert(!ResumeIndex && "Unexpected resume value found");
1149 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1150 MiddleBlock->getTerminator());
1151 ResumeIndex->addIncoming(StartIdx, BypassBlock);
1152 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1155 // Make sure that we found the index where scalar loop needs to continue.
1156 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1157 "Invalid resume Index");
1159 // Add a check in the middle block to see if we have completed
1160 // all of the iterations in the first vector loop.
1161 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1162 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1163 ResumeIndex, "cmp.n",
1164 MiddleBlock->getTerminator());
1166 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1167 // Remove the old terminator.
1168 MiddleBlock->getTerminator()->eraseFromParent();
1170 // Create i+1 and fill the PHINode.
1171 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1172 Induction->addIncoming(StartIdx, VectorPH);
1173 Induction->addIncoming(NextIdx, VecBody);
1174 // Create the compare.
1175 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1176 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1178 // Now we have two terminators. Remove the old one from the block.
1179 VecBody->getTerminator()->eraseFromParent();
1181 // Get ready to start creating new instructions into the vectorized body.
1182 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1184 // Create and register the new vector loop.
1185 Loop* Lp = new Loop();
1186 Loop *ParentLoop = OrigLoop->getParentLoop();
1188 // Insert the new loop into the loop nest and register the new basic blocks.
1190 ParentLoop->addChildLoop(Lp);
1191 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1192 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1193 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1195 LI->addTopLevelLoop(Lp);
1198 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1201 LoopVectorPreHeader = VectorPH;
1202 LoopScalarPreHeader = ScalarPH;
1203 LoopMiddleBlock = MiddleBlock;
1204 LoopExitBlock = ExitBlock;
1205 LoopVectorBody = VecBody;
1206 LoopScalarBody = OldBasicBlock;
1207 LoopBypassBlock = BypassBlock;
1210 /// This function returns the identity element (or neutral element) for
1211 /// the operation K.
1213 getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
1215 case LoopVectorizationLegality::IntegerXor:
1216 case LoopVectorizationLegality::IntegerAdd:
1217 case LoopVectorizationLegality::IntegerOr:
1218 // Adding, Xoring, Oring zero to a number does not change it.
1220 case LoopVectorizationLegality::IntegerMult:
1221 // Multiplying a number by 1 does not change it.
1223 case LoopVectorizationLegality::IntegerAnd:
1224 // AND-ing a number with an all-1 value does not change it.
1227 llvm_unreachable("Unknown reduction kind");
1232 isTriviallyVectorizableIntrinsic(Instruction *Inst) {
1233 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
1236 switch (II->getIntrinsicID()) {
1237 case Intrinsic::sqrt:
1238 case Intrinsic::sin:
1239 case Intrinsic::cos:
1240 case Intrinsic::exp:
1241 case Intrinsic::exp2:
1242 case Intrinsic::log:
1243 case Intrinsic::log10:
1244 case Intrinsic::log2:
1245 case Intrinsic::fabs:
1246 case Intrinsic::floor:
1247 case Intrinsic::ceil:
1248 case Intrinsic::trunc:
1249 case Intrinsic::rint:
1250 case Intrinsic::nearbyint:
1251 case Intrinsic::pow:
1252 case Intrinsic::fma:
1253 case Intrinsic::fmuladd:
1262 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1263 //===------------------------------------------------===//
1265 // Notice: any optimization or new instruction that go
1266 // into the code below should be also be implemented in
1269 //===------------------------------------------------===//
1270 BasicBlock &BB = *OrigLoop->getHeader();
1272 ConstantInt::get(IntegerType::getInt32Ty(BB.getContext()), 0);
1274 // In order to support reduction variables we need to be able to vectorize
1275 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1276 // stages. First, we create a new vector PHI node with no incoming edges.
1277 // We use this value when we vectorize all of the instructions that use the
1278 // PHI. Next, after all of the instructions in the block are complete we
1279 // add the new incoming edges to the PHI. At this point all of the
1280 // instructions in the basic block are vectorized, so we can use them to
1281 // construct the PHI.
1282 PhiVector RdxPHIsToFix;
1284 // Scan the loop in a topological order to ensure that defs are vectorized
1286 LoopBlocksDFS DFS(OrigLoop);
1289 // Vectorize all of the blocks in the original loop.
1290 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1291 be = DFS.endRPO(); bb != be; ++bb)
1292 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1294 // At this point every instruction in the original loop is widened to
1295 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1296 // that we vectorized. The PHI nodes are currently empty because we did
1297 // not want to introduce cycles. Notice that the remaining PHI nodes
1298 // that we need to fix are reduction variables.
1300 // Create the 'reduced' values for each of the induction vars.
1301 // The reduced values are the vector values that we scalarize and combine
1302 // after the loop is finished.
1303 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1305 PHINode *RdxPhi = *it;
1306 assert(RdxPhi && "Unable to recover vectorized PHI");
1308 // Find the reduction variable descriptor.
1309 assert(Legal->getReductionVars()->count(RdxPhi) &&
1310 "Unable to find the reduction variable");
1311 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1312 (*Legal->getReductionVars())[RdxPhi];
1314 // We need to generate a reduction vector from the incoming scalar.
1315 // To do so, we need to generate the 'identity' vector and overide
1316 // one of the elements with the incoming scalar reduction. We need
1317 // to do it in the vector-loop preheader.
1318 Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
1320 // This is the vector-clone of the value that leaves the loop.
1321 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1322 Type *VecTy = VectorExit[0]->getType();
1324 // Find the reduction identity variable. Zero for addition, or, xor,
1325 // one for multiplication, -1 for And.
1326 Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
1327 VecTy->getScalarType());
1329 // This vector is the Identity vector where the first element is the
1330 // incoming scalar reduction.
1331 Value *VectorStart = Builder.CreateInsertElement(Identity,
1332 RdxDesc.StartValue, Zero);
1334 // Fix the vector-loop phi.
1335 // We created the induction variable so we know that the
1336 // preheader is the first entry.
1337 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1339 // Reductions do not have to start at zero. They can start with
1340 // any loop invariant values.
1341 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1342 BasicBlock *Latch = OrigLoop->getLoopLatch();
1343 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1344 VectorParts &Val = getVectorValue(LoopVal);
1345 for (unsigned part = 0; part < UF; ++part) {
1346 // Make sure to add the reduction stat value only to the
1347 // first unroll part.
1348 Value *StartVal = (part == 0) ? VectorStart : Identity;
1349 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1350 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1353 // Before each round, move the insertion point right between
1354 // the PHIs and the values we are going to write.
1355 // This allows us to write both PHINodes and the extractelement
1357 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1359 VectorParts RdxParts;
1360 for (unsigned part = 0; part < UF; ++part) {
1361 // This PHINode contains the vectorized reduction variable, or
1362 // the initial value vector, if we bypass the vector loop.
1363 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1364 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1365 Value *StartVal = (part == 0) ? VectorStart : Identity;
1366 NewPhi->addIncoming(StartVal, LoopBypassBlock);
1367 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1368 RdxParts.push_back(NewPhi);
1371 // Reduce all of the unrolled parts into a single vector.
1372 Value *ReducedPartRdx = RdxParts[0];
1373 for (unsigned part = 1; part < UF; ++part) {
1374 switch (RdxDesc.Kind) {
1375 case LoopVectorizationLegality::IntegerAdd:
1377 Builder.CreateAdd(RdxParts[part], ReducedPartRdx, "add.rdx");
1379 case LoopVectorizationLegality::IntegerMult:
1381 Builder.CreateMul(RdxParts[part], ReducedPartRdx, "mul.rdx");
1383 case LoopVectorizationLegality::IntegerOr:
1385 Builder.CreateOr(RdxParts[part], ReducedPartRdx, "or.rdx");
1387 case LoopVectorizationLegality::IntegerAnd:
1389 Builder.CreateAnd(RdxParts[part], ReducedPartRdx, "and.rdx");
1391 case LoopVectorizationLegality::IntegerXor:
1393 Builder.CreateXor(RdxParts[part], ReducedPartRdx, "xor.rdx");
1396 llvm_unreachable("Unknown reduction operation");
1401 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1402 // and vector ops, reducing the set of values being computed by half each
1404 assert(isPowerOf2_32(VF) &&
1405 "Reduction emission only supported for pow2 vectors!");
1406 Value *TmpVec = ReducedPartRdx;
1407 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1408 for (unsigned i = VF; i != 1; i >>= 1) {
1409 // Move the upper half of the vector to the lower half.
1410 for (unsigned j = 0; j != i/2; ++j)
1411 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1413 // Fill the rest of the mask with undef.
1414 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1415 UndefValue::get(Builder.getInt32Ty()));
1418 Builder.CreateShuffleVector(TmpVec,
1419 UndefValue::get(TmpVec->getType()),
1420 ConstantVector::get(ShuffleMask),
1423 // Emit the operation on the shuffled value.
1424 switch (RdxDesc.Kind) {
1425 case LoopVectorizationLegality::IntegerAdd:
1426 TmpVec = Builder.CreateAdd(TmpVec, Shuf, "add.rdx");
1428 case LoopVectorizationLegality::IntegerMult:
1429 TmpVec = Builder.CreateMul(TmpVec, Shuf, "mul.rdx");
1431 case LoopVectorizationLegality::IntegerOr:
1432 TmpVec = Builder.CreateOr(TmpVec, Shuf, "or.rdx");
1434 case LoopVectorizationLegality::IntegerAnd:
1435 TmpVec = Builder.CreateAnd(TmpVec, Shuf, "and.rdx");
1437 case LoopVectorizationLegality::IntegerXor:
1438 TmpVec = Builder.CreateXor(TmpVec, Shuf, "xor.rdx");
1441 llvm_unreachable("Unknown reduction operation");
1445 // The result is in the first element of the vector.
1446 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1448 // Now, we need to fix the users of the reduction variable
1449 // inside and outside of the scalar remainder loop.
1450 // We know that the loop is in LCSSA form. We need to update the
1451 // PHI nodes in the exit blocks.
1452 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1453 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1454 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1455 if (!LCSSAPhi) continue;
1457 // All PHINodes need to have a single entry edge, or two if
1458 // we already fixed them.
1459 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1461 // We found our reduction value exit-PHI. Update it with the
1462 // incoming bypass edge.
1463 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1464 // Add an edge coming from the bypass.
1465 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1468 }// end of the LCSSA phi scan.
1470 // Fix the scalar loop reduction variable with the incoming reduction sum
1471 // from the vector body and from the backedge value.
1472 int IncomingEdgeBlockIdx =
1473 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1474 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1475 // Pick the other block.
1476 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1477 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1478 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1479 }// end of for each redux variable.
1481 // The Loop exit block may have single value PHI nodes where the incoming
1482 // value is 'undef'. While vectorizing we only handled real values that
1483 // were defined inside the loop. Here we handle the 'undef case'.
1485 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1486 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1487 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1488 if (!LCSSAPhi) continue;
1489 if (LCSSAPhi->getNumIncomingValues() == 1)
1490 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1495 InnerLoopVectorizer::VectorParts
1496 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1497 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1500 VectorParts SrcMask = createBlockInMask(Src);
1502 // The terminator has to be a branch inst!
1503 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1504 assert(BI && "Unexpected terminator found");
1506 if (BI->isConditional()) {
1507 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1509 if (BI->getSuccessor(0) != Dst)
1510 for (unsigned part = 0; part < UF; ++part)
1511 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1513 for (unsigned part = 0; part < UF; ++part)
1514 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1521 InnerLoopVectorizer::VectorParts
1522 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1523 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1525 // Loop incoming mask is all-one.
1526 if (OrigLoop->getHeader() == BB) {
1527 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1528 return getVectorValue(C);
1531 // This is the block mask. We OR all incoming edges, and with zero.
1532 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1533 VectorParts BlockMask = getVectorValue(Zero);
1536 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1537 VectorParts EM = createEdgeMask(*it, BB);
1538 for (unsigned part = 0; part < UF; ++part)
1539 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1546 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1547 BasicBlock *BB, PhiVector *PV) {
1548 Constant *Zero = Builder.getInt32(0);
1550 // For each instruction in the old loop.
1551 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1552 VectorParts &Entry = WidenMap.get(it);
1553 switch (it->getOpcode()) {
1554 case Instruction::Br:
1555 // Nothing to do for PHIs and BR, since we already took care of the
1556 // loop control flow instructions.
1558 case Instruction::PHI:{
1559 PHINode* P = cast<PHINode>(it);
1560 // Handle reduction variables:
1561 if (Legal->getReductionVars()->count(P)) {
1562 for (unsigned part = 0; part < UF; ++part) {
1563 // This is phase one of vectorizing PHIs.
1564 Type *VecTy = VectorType::get(it->getType(), VF);
1565 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1566 LoopVectorBody-> getFirstInsertionPt());
1572 // Check for PHI nodes that are lowered to vector selects.
1573 if (P->getParent() != OrigLoop->getHeader()) {
1574 // We know that all PHIs in non header blocks are converted into
1575 // selects, so we don't have to worry about the insertion order and we
1576 // can just use the builder.
1578 // At this point we generate the predication tree. There may be
1579 // duplications since this is a simple recursive scan, but future
1580 // optimizations will clean it up.
1581 VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
1584 for (unsigned part = 0; part < UF; ++part) {
1585 VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
1586 VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
1587 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
1593 // This PHINode must be an induction variable.
1594 // Make sure that we know about it.
1595 assert(Legal->getInductionVars()->count(P) &&
1596 "Not an induction variable");
1598 LoopVectorizationLegality::InductionInfo II =
1599 Legal->getInductionVars()->lookup(P);
1602 case LoopVectorizationLegality::NoInduction:
1603 llvm_unreachable("Unknown induction");
1604 case LoopVectorizationLegality::IntInduction: {
1605 assert(P == OldInduction && "Unexpected PHI");
1606 Value *Broadcasted = getBroadcastInstrs(Induction);
1607 // After broadcasting the induction variable we need to make the
1608 // vector consecutive by adding 0, 1, 2 ...
1609 for (unsigned part = 0; part < UF; ++part)
1610 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
1613 case LoopVectorizationLegality::ReverseIntInduction:
1614 case LoopVectorizationLegality::PtrInduction:
1615 // Handle reverse integer and pointer inductions.
1616 Value *StartIdx = 0;
1617 // If we have a single integer induction variable then use it.
1618 // Otherwise, start counting at zero.
1620 LoopVectorizationLegality::InductionInfo OldII =
1621 Legal->getInductionVars()->lookup(OldInduction);
1622 StartIdx = OldII.StartValue;
1624 StartIdx = ConstantInt::get(Induction->getType(), 0);
1626 // This is the normalized GEP that starts counting at zero.
1627 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1630 // Handle the reverse integer induction variable case.
1631 if (LoopVectorizationLegality::ReverseIntInduction == II.IK) {
1632 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
1633 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
1635 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
1638 // This is a new value so do not hoist it out.
1639 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
1640 // After broadcasting the induction variable we need to make the
1641 // vector consecutive by adding ... -3, -2, -1, 0.
1642 for (unsigned part = 0; part < UF; ++part)
1643 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
1647 // Handle the pointer induction variable case.
1648 assert(P->getType()->isPointerTy() && "Unexpected type.");
1650 // This is the vector of results. Notice that we don't generate
1651 // vector geps because scalar geps result in better code.
1652 for (unsigned part = 0; part < UF; ++part) {
1653 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1654 for (unsigned int i = 0; i < VF; ++i) {
1655 Constant *Idx = ConstantInt::get(Induction->getType(),
1657 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx,
1659 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
1661 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1662 Builder.getInt32(i),
1665 Entry[part] = VecVal;
1672 case Instruction::Add:
1673 case Instruction::FAdd:
1674 case Instruction::Sub:
1675 case Instruction::FSub:
1676 case Instruction::Mul:
1677 case Instruction::FMul:
1678 case Instruction::UDiv:
1679 case Instruction::SDiv:
1680 case Instruction::FDiv:
1681 case Instruction::URem:
1682 case Instruction::SRem:
1683 case Instruction::FRem:
1684 case Instruction::Shl:
1685 case Instruction::LShr:
1686 case Instruction::AShr:
1687 case Instruction::And:
1688 case Instruction::Or:
1689 case Instruction::Xor: {
1690 // Just widen binops.
1691 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
1692 VectorParts &A = getVectorValue(it->getOperand(0));
1693 VectorParts &B = getVectorValue(it->getOperand(1));
1695 // Use this vector value for all users of the original instruction.
1696 for (unsigned Part = 0; Part < UF; ++Part) {
1697 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
1699 // Update the NSW, NUW and Exact flags.
1700 BinaryOperator *VecOp = cast<BinaryOperator>(V);
1701 if (isa<OverflowingBinaryOperator>(BinOp)) {
1702 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1703 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1705 if (isa<PossiblyExactOperator>(VecOp))
1706 VecOp->setIsExact(BinOp->isExact());
1712 case Instruction::Select: {
1714 // If the selector is loop invariant we can create a select
1715 // instruction with a scalar condition. Otherwise, use vector-select.
1716 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
1719 // The condition can be loop invariant but still defined inside the
1720 // loop. This means that we can't just use the original 'cond' value.
1721 // We have to take the 'vectorized' value and pick the first lane.
1722 // Instcombine will make this a no-op.
1723 VectorParts &Cond = getVectorValue(it->getOperand(0));
1724 VectorParts &Op0 = getVectorValue(it->getOperand(1));
1725 VectorParts &Op1 = getVectorValue(it->getOperand(2));
1726 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
1727 Builder.getInt32(0));
1728 for (unsigned Part = 0; Part < UF; ++Part) {
1729 Entry[Part] = Builder.CreateSelect(
1730 InvariantCond ? ScalarCond : Cond[Part],
1737 case Instruction::ICmp:
1738 case Instruction::FCmp: {
1739 // Widen compares. Generate vector compares.
1740 bool FCmp = (it->getOpcode() == Instruction::FCmp);
1741 CmpInst *Cmp = dyn_cast<CmpInst>(it);
1742 VectorParts &A = getVectorValue(it->getOperand(0));
1743 VectorParts &B = getVectorValue(it->getOperand(1));
1744 for (unsigned Part = 0; Part < UF; ++Part) {
1747 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
1749 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
1755 case Instruction::Store: {
1756 // Attempt to issue a wide store.
1757 StoreInst *SI = dyn_cast<StoreInst>(it);
1758 Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
1759 Value *Ptr = SI->getPointerOperand();
1760 unsigned Alignment = SI->getAlignment();
1762 assert(!Legal->isUniform(Ptr) &&
1763 "We do not allow storing to uniform addresses");
1766 int Stride = Legal->isConsecutivePtr(Ptr);
1767 bool Reverse = Stride < 0;
1769 scalarizeInstruction(it);
1773 // Handle consecutive stores.
1775 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1777 // The last index does not have to be the induction. It can be
1778 // consecutive and be a function of the index. For example A[I+1];
1779 unsigned NumOperands = Gep->getNumOperands();
1781 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1782 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1783 Value *LastIndex = GEPParts[0];
1784 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1786 // Create the new GEP with the new induction variable.
1787 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1788 Gep2->setOperand(NumOperands - 1, LastIndex);
1789 Ptr = Builder.Insert(Gep2);
1791 // Use the induction element ptr.
1792 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1793 VectorParts &PtrVal = getVectorValue(Ptr);
1794 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1797 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1798 for (unsigned Part = 0; Part < UF; ++Part) {
1799 // Calculate the pointer for the specific unroll-part.
1800 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1803 // If we store to reverse consecutive memory locations then we need
1804 // to reverse the order of elements in the stored value.
1805 StoredVal[Part] = reverseVector(StoredVal[Part]);
1806 // If the address is consecutive but reversed, then the
1807 // wide store needs to start at the last vector element.
1808 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1809 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1812 Value *VecPtr = Builder.CreateBitCast(PartPtr, StTy->getPointerTo());
1813 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1817 case Instruction::Load: {
1818 // Attempt to issue a wide load.
1819 LoadInst *LI = dyn_cast<LoadInst>(it);
1820 Type *RetTy = VectorType::get(LI->getType(), VF);
1821 Value *Ptr = LI->getPointerOperand();
1822 unsigned Alignment = LI->getAlignment();
1824 // If the pointer is loop invariant or if it is non consecutive,
1825 // scalarize the load.
1826 int Stride = Legal->isConsecutivePtr(Ptr);
1827 bool Reverse = Stride < 0;
1828 if (Legal->isUniform(Ptr) || Stride == 0) {
1829 scalarizeInstruction(it);
1833 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1835 // The last index does not have to be the induction. It can be
1836 // consecutive and be a function of the index. For example A[I+1];
1837 unsigned NumOperands = Gep->getNumOperands();
1839 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1840 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1841 Value *LastIndex = GEPParts[0];
1842 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1844 // Create the new GEP with the new induction variable.
1845 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1846 Gep2->setOperand(NumOperands - 1, LastIndex);
1847 Ptr = Builder.Insert(Gep2);
1849 // Use the induction element ptr.
1850 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1851 VectorParts &PtrVal = getVectorValue(Ptr);
1852 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1855 for (unsigned Part = 0; Part < UF; ++Part) {
1856 // Calculate the pointer for the specific unroll-part.
1857 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1860 // If the address is consecutive but reversed, then the
1861 // wide store needs to start at the last vector element.
1862 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1863 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1866 Value *VecPtr = Builder.CreateBitCast(PartPtr, RetTy->getPointerTo());
1867 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1868 cast<LoadInst>(LI)->setAlignment(Alignment);
1869 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1873 case Instruction::ZExt:
1874 case Instruction::SExt:
1875 case Instruction::FPToUI:
1876 case Instruction::FPToSI:
1877 case Instruction::FPExt:
1878 case Instruction::PtrToInt:
1879 case Instruction::IntToPtr:
1880 case Instruction::SIToFP:
1881 case Instruction::UIToFP:
1882 case Instruction::Trunc:
1883 case Instruction::FPTrunc:
1884 case Instruction::BitCast: {
1885 CastInst *CI = dyn_cast<CastInst>(it);
1886 /// Optimize the special case where the source is the induction
1887 /// variable. Notice that we can only optimize the 'trunc' case
1888 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
1889 /// c. other casts depend on pointer size.
1890 if (CI->getOperand(0) == OldInduction &&
1891 it->getOpcode() == Instruction::Trunc) {
1892 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
1894 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
1895 for (unsigned Part = 0; Part < UF; ++Part)
1896 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
1899 /// Vectorize casts.
1900 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
1902 VectorParts &A = getVectorValue(it->getOperand(0));
1903 for (unsigned Part = 0; Part < UF; ++Part)
1904 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
1908 case Instruction::Call: {
1909 assert(isTriviallyVectorizableIntrinsic(it));
1910 Module *M = BB->getParent()->getParent();
1911 IntrinsicInst *II = cast<IntrinsicInst>(it);
1912 Intrinsic::ID ID = II->getIntrinsicID();
1913 for (unsigned Part = 0; Part < UF; ++Part) {
1914 SmallVector<Value*, 4> Args;
1915 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) {
1916 VectorParts &Arg = getVectorValue(II->getArgOperand(i));
1917 Args.push_back(Arg[Part]);
1919 Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) };
1920 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
1921 Entry[Part] = Builder.CreateCall(F, Args);
1927 // All other instructions are unsupported. Scalarize them.
1928 scalarizeInstruction(it);
1931 }// end of for_each instr.
1934 void InnerLoopVectorizer::updateAnalysis() {
1935 // Forget the original basic block.
1936 SE->forgetLoop(OrigLoop);
1938 // Update the dominator tree information.
1939 assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
1940 "Entry does not dominate exit.");
1942 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
1943 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
1944 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
1945 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
1946 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
1947 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
1949 DEBUG(DT->verifyAnalysis());
1952 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
1953 if (!EnableIfConversion)
1956 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
1957 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
1959 // Collect the blocks that need predication.
1960 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
1961 BasicBlock *BB = LoopBlocks[i];
1963 // We don't support switch statements inside loops.
1964 if (!isa<BranchInst>(BB->getTerminator()))
1967 // We must have at most two predecessors because we need to convert
1968 // all PHIs to selects.
1969 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
1973 // We must be able to predicate all blocks that need to be predicated.
1974 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
1978 // We can if-convert this loop.
1982 bool LoopVectorizationLegality::canVectorize() {
1983 assert(TheLoop->getLoopPreheader() && "No preheader!!");
1985 // We can only vectorize innermost loops.
1986 if (TheLoop->getSubLoopsVector().size())
1989 // We must have a single backedge.
1990 if (TheLoop->getNumBackEdges() != 1)
1993 // We must have a single exiting block.
1994 if (!TheLoop->getExitingBlock())
1997 unsigned NumBlocks = TheLoop->getNumBlocks();
1999 // Check if we can if-convert non single-bb loops.
2000 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2001 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2005 // We need to have a loop header.
2006 BasicBlock *Latch = TheLoop->getLoopLatch();
2007 DEBUG(dbgs() << "LV: Found a loop: " <<
2008 TheLoop->getHeader()->getName() << "\n");
2010 // ScalarEvolution needs to be able to find the exit count.
2011 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2012 if (ExitCount == SE->getCouldNotCompute()) {
2013 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2017 // Do not loop-vectorize loops with a tiny trip count.
2018 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2019 if (TC > 0u && TC < TinyTripCountThreshold) {
2020 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2021 "This loop is not worth vectorizing.\n");
2025 // Check if we can vectorize the instructions and CFG in this loop.
2026 if (!canVectorizeInstrs()) {
2027 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2031 // Go over each instruction and look at memory deps.
2032 if (!canVectorizeMemory()) {
2033 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2037 // Collect all of the variables that remain uniform after vectorization.
2038 collectLoopUniforms();
2040 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2041 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2044 // Okay! We can vectorize. At this point we don't have any other mem analysis
2045 // which may limit our maximum vectorization factor, so just return true with
2050 bool LoopVectorizationLegality::canVectorizeInstrs() {
2051 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2052 BasicBlock *Header = TheLoop->getHeader();
2054 // For each block in the loop.
2055 for (Loop::block_iterator bb = TheLoop->block_begin(),
2056 be = TheLoop->block_end(); bb != be; ++bb) {
2058 // Scan the instructions in the block and look for hazards.
2059 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2062 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2063 // This should not happen because the loop should be normalized.
2064 if (Phi->getNumIncomingValues() != 2) {
2065 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2069 // Check that this PHI type is allowed.
2070 if (!Phi->getType()->isIntegerTy() &&
2071 !Phi->getType()->isPointerTy()) {
2072 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2076 // If this PHINode is not in the header block, then we know that we
2077 // can convert it to select during if-conversion. No need to check if
2078 // the PHIs in this block are induction or reduction variables.
2082 // This is the value coming from the preheader.
2083 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2084 // Check if this is an induction variable.
2085 InductionKind IK = isInductionVariable(Phi);
2087 if (NoInduction != IK) {
2088 // Int inductions are special because we only allow one IV.
2089 if (IK == IntInduction) {
2091 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2097 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2098 Inductions[Phi] = InductionInfo(StartValue, IK);
2102 if (AddReductionVar(Phi, IntegerAdd)) {
2103 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2106 if (AddReductionVar(Phi, IntegerMult)) {
2107 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2110 if (AddReductionVar(Phi, IntegerOr)) {
2111 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2114 if (AddReductionVar(Phi, IntegerAnd)) {
2115 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2118 if (AddReductionVar(Phi, IntegerXor)) {
2119 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2123 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2125 }// end of PHI handling
2127 // We still don't handle functions.
2128 CallInst *CI = dyn_cast<CallInst>(it);
2129 if (CI && !isTriviallyVectorizableIntrinsic(it)) {
2130 DEBUG(dbgs() << "LV: Found a call site.\n");
2134 // Check that the instruction return type is vectorizable.
2135 if (!VectorType::isValidElementType(it->getType()) &&
2136 !it->getType()->isVoidTy()) {
2137 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2141 // Check that the stored type is vectorizable.
2142 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2143 Type *T = ST->getValueOperand()->getType();
2144 if (!VectorType::isValidElementType(T))
2148 // Reduction instructions are allowed to have exit users.
2149 // All other instructions must not have external users.
2150 if (!AllowedExit.count(it))
2151 //Check that all of the users of the loop are inside the BB.
2152 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2154 Instruction *U = cast<Instruction>(*I);
2155 // This user may be a reduction exit value.
2156 if (!TheLoop->contains(U)) {
2157 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2166 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2167 assert(getInductionVars()->size() && "No induction variables");
2173 void LoopVectorizationLegality::collectLoopUniforms() {
2174 // We now know that the loop is vectorizable!
2175 // Collect variables that will remain uniform after vectorization.
2176 std::vector<Value*> Worklist;
2177 BasicBlock *Latch = TheLoop->getLoopLatch();
2179 // Start with the conditional branch and walk up the block.
2180 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2182 while (Worklist.size()) {
2183 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2184 Worklist.pop_back();
2186 // Look at instructions inside this loop.
2187 // Stop when reaching PHI nodes.
2188 // TODO: we need to follow values all over the loop, not only in this block.
2189 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2192 // This is a known uniform.
2195 // Insert all operands.
2196 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2197 Worklist.push_back(I->getOperand(i));
2202 bool LoopVectorizationLegality::canVectorizeMemory() {
2203 typedef SmallVector<Value*, 16> ValueVector;
2204 typedef SmallPtrSet<Value*, 16> ValueSet;
2205 // Holds the Load and Store *instructions*.
2208 PtrRtCheck.Pointers.clear();
2209 PtrRtCheck.Need = false;
2212 for (Loop::block_iterator bb = TheLoop->block_begin(),
2213 be = TheLoop->block_end(); bb != be; ++bb) {
2215 // Scan the BB and collect legal loads and stores.
2216 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2219 // If this is a load, save it. If this instruction can read from memory
2220 // but is not a load, then we quit. Notice that we don't handle function
2221 // calls that read or write.
2222 if (it->mayReadFromMemory()) {
2223 LoadInst *Ld = dyn_cast<LoadInst>(it);
2224 if (!Ld) return false;
2225 if (!Ld->isSimple()) {
2226 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2229 Loads.push_back(Ld);
2233 // Save 'store' instructions. Abort if other instructions write to memory.
2234 if (it->mayWriteToMemory()) {
2235 StoreInst *St = dyn_cast<StoreInst>(it);
2236 if (!St) return false;
2237 if (!St->isSimple()) {
2238 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2241 Stores.push_back(St);
2246 // Now we have two lists that hold the loads and the stores.
2247 // Next, we find the pointers that they use.
2249 // Check if we see any stores. If there are no stores, then we don't
2250 // care if the pointers are *restrict*.
2251 if (!Stores.size()) {
2252 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2256 // Holds the read and read-write *pointers* that we find.
2258 ValueVector ReadWrites;
2260 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2261 // multiple times on the same object. If the ptr is accessed twice, once
2262 // for read and once for write, it will only appear once (on the write
2263 // list). This is okay, since we are going to check for conflicts between
2264 // writes and between reads and writes, but not between reads and reads.
2267 ValueVector::iterator I, IE;
2268 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2269 StoreInst *ST = cast<StoreInst>(*I);
2270 Value* Ptr = ST->getPointerOperand();
2272 if (isUniform(Ptr)) {
2273 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2277 // If we did *not* see this pointer before, insert it to
2278 // the read-write list. At this phase it is only a 'write' list.
2279 if (Seen.insert(Ptr))
2280 ReadWrites.push_back(Ptr);
2283 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2284 LoadInst *LD = cast<LoadInst>(*I);
2285 Value* Ptr = LD->getPointerOperand();
2286 // If we did *not* see this pointer before, insert it to the
2287 // read list. If we *did* see it before, then it is already in
2288 // the read-write list. This allows us to vectorize expressions
2289 // such as A[i] += x; Because the address of A[i] is a read-write
2290 // pointer. This only works if the index of A[i] is consecutive.
2291 // If the address of i is unknown (for example A[B[i]]) then we may
2292 // read a few words, modify, and write a few words, and some of the
2293 // words may be written to the same address.
2294 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2295 Reads.push_back(Ptr);
2298 // If we write (or read-write) to a single destination and there are no
2299 // other reads in this loop then is it safe to vectorize.
2300 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2301 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2305 // Find pointers with computable bounds. We are going to use this information
2306 // to place a runtime bound check.
2307 bool CanDoRT = true;
2308 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
2309 if (hasComputableBounds(*I)) {
2310 PtrRtCheck.insert(SE, TheLoop, *I);
2311 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
2316 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
2317 if (hasComputableBounds(*I)) {
2318 PtrRtCheck.insert(SE, TheLoop, *I);
2319 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
2325 // Check that we did not collect too many pointers or found a
2326 // unsizeable pointer.
2327 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
2333 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2336 bool NeedRTCheck = false;
2338 // Now that the pointers are in two lists (Reads and ReadWrites), we
2339 // can check that there are no conflicts between each of the writes and
2340 // between the writes to the reads.
2341 ValueSet WriteObjects;
2342 ValueVector TempObjects;
2344 // Check that the read-writes do not conflict with other read-write
2346 bool AllWritesIdentified = true;
2347 for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
2348 GetUnderlyingObjects(*I, TempObjects, DL);
2349 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
2351 if (!isIdentifiedObject(*it)) {
2352 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
2354 AllWritesIdentified = false;
2356 if (!WriteObjects.insert(*it)) {
2357 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2362 TempObjects.clear();
2365 /// Check that the reads don't conflict with the read-writes.
2366 for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
2367 GetUnderlyingObjects(*I, TempObjects, DL);
2368 for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
2370 // If all of the writes are identified then we don't care if the read
2371 // pointer is identified or not.
2372 if (!AllWritesIdentified && !isIdentifiedObject(*it)) {
2373 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
2376 if (WriteObjects.count(*it)) {
2377 DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
2382 TempObjects.clear();
2385 PtrRtCheck.Need = NeedRTCheck;
2386 if (NeedRTCheck && !CanDoRT) {
2387 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2388 "the array bounds.\n");
2393 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2394 " need a runtime memory check.\n");
2398 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2399 ReductionKind Kind) {
2400 if (Phi->getNumIncomingValues() != 2)
2403 // Reduction variables are only found in the loop header block.
2404 if (Phi->getParent() != TheLoop->getHeader())
2407 // Obtain the reduction start value from the value that comes from the loop
2409 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2411 // ExitInstruction is the single value which is used outside the loop.
2412 // We only allow for a single reduction value to be used outside the loop.
2413 // This includes users of the reduction, variables (which form a cycle
2414 // which ends in the phi node).
2415 Instruction *ExitInstruction = 0;
2417 // Iter is our iterator. We start with the PHI node and scan for all of the
2418 // users of this instruction. All users must be instructions that can be
2419 // used as reduction variables (such as ADD). We may have a single
2420 // out-of-block user. The cycle must end with the original PHI.
2421 Instruction *Iter = Phi;
2423 // If the instruction has no users then this is a broken
2424 // chain and can't be a reduction variable.
2425 if (Iter->use_empty())
2428 // Did we find a user inside this loop already ?
2429 bool FoundInBlockUser = false;
2430 // Did we reach the initial PHI node already ?
2431 bool FoundStartPHI = false;
2433 // For each of the *users* of iter.
2434 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2436 Instruction *U = cast<Instruction>(*it);
2437 // We already know that the PHI is a user.
2439 FoundStartPHI = true;
2443 // Check if we found the exit user.
2444 BasicBlock *Parent = U->getParent();
2445 if (!TheLoop->contains(Parent)) {
2446 // Exit if you find multiple outside users.
2447 if (ExitInstruction != 0)
2449 ExitInstruction = Iter;
2452 // We allow in-loop PHINodes which are not the original reduction PHI
2453 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2454 // structure) then don't skip this PHI.
2455 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2456 U->getParent() != TheLoop->getHeader() &&
2457 TheLoop->contains(U) &&
2458 Iter->getNumUses() > 1)
2461 // We can't have multiple inside users.
2462 if (FoundInBlockUser)
2464 FoundInBlockUser = true;
2466 // Any reduction instr must be of one of the allowed kinds.
2467 if (!isReductionInstr(U, Kind))
2470 // Reductions of instructions such as Div, and Sub is only
2471 // possible if the LHS is the reduction variable.
2472 if (!U->isCommutative() && U->getOperand(0) != Iter)
2478 // We found a reduction var if we have reached the original
2479 // phi node and we only have a single instruction with out-of-loop
2481 if (FoundStartPHI && ExitInstruction) {
2482 // This instruction is allowed to have out-of-loop users.
2483 AllowedExit.insert(ExitInstruction);
2485 // Save the description of this reduction variable.
2486 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
2487 Reductions[Phi] = RD;
2491 // If we've reached the start PHI but did not find an outside user then
2492 // this is dead code. Abort.
2499 LoopVectorizationLegality::isReductionInstr(Instruction *I,
2500 ReductionKind Kind) {
2501 switch (I->getOpcode()) {
2504 case Instruction::PHI:
2507 case Instruction::Sub:
2508 case Instruction::Add:
2509 return Kind == IntegerAdd;
2510 case Instruction::SDiv:
2511 case Instruction::UDiv:
2512 case Instruction::Mul:
2513 return Kind == IntegerMult;
2514 case Instruction::And:
2515 return Kind == IntegerAnd;
2516 case Instruction::Or:
2517 return Kind == IntegerOr;
2518 case Instruction::Xor:
2519 return Kind == IntegerXor;
2523 LoopVectorizationLegality::InductionKind
2524 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
2525 Type *PhiTy = Phi->getType();
2526 // We only handle integer and pointer inductions variables.
2527 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
2530 // Check that the PHI is consecutive and starts at zero.
2531 const SCEV *PhiScev = SE->getSCEV(Phi);
2532 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2534 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
2537 const SCEV *Step = AR->getStepRecurrence(*SE);
2539 // Integer inductions need to have a stride of one.
2540 if (PhiTy->isIntegerTy()) {
2542 return IntInduction;
2543 if (Step->isAllOnesValue())
2544 return ReverseIntInduction;
2548 // Calculate the pointer stride and check if it is consecutive.
2549 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
2553 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
2554 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
2555 if (C->getValue()->equalsInt(Size))
2556 return PtrInduction;
2561 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
2562 Value *In0 = const_cast<Value*>(V);
2563 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
2567 return Inductions.count(PN);
2570 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
2571 assert(TheLoop->contains(BB) && "Unknown block used");
2573 // Blocks that do not dominate the latch need predication.
2574 BasicBlock* Latch = TheLoop->getLoopLatch();
2575 return !DT->dominates(BB, Latch);
2578 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
2579 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2580 // We don't predicate loads/stores at the moment.
2581 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
2584 // The instructions below can trap.
2585 switch (it->getOpcode()) {
2587 case Instruction::UDiv:
2588 case Instruction::SDiv:
2589 case Instruction::URem:
2590 case Instruction::SRem:
2598 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
2599 const SCEV *PhiScev = SE->getSCEV(Ptr);
2600 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2604 return AR->isAffine();
2608 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
2610 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
2611 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
2615 // Find the trip count.
2616 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
2617 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
2619 unsigned VF = MaxVectorSize;
2621 // If we optimize the program for size, avoid creating the tail loop.
2623 // If we are unable to calculate the trip count then don't try to vectorize.
2625 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2629 // Find the maximum SIMD width that can fit within the trip count.
2630 VF = TC % MaxVectorSize;
2635 // If the trip count that we found modulo the vectorization factor is not
2636 // zero then we require a tail.
2638 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2644 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
2645 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
2650 float Cost = expectedCost(1);
2652 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
2653 for (unsigned i=2; i <= VF; i*=2) {
2654 // Notice that the vector loop needs to be executed less times, so
2655 // we need to divide the cost of the vector loops by the width of
2656 // the vector elements.
2657 float VectorCost = expectedCost(i) / (float)i;
2658 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
2659 (int)VectorCost << ".\n");
2660 if (VectorCost < Cost) {
2666 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
2671 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
2673 // Use the user preference, unless 'auto' is selected.
2677 // When we optimize for size we don't unroll.
2681 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
2682 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
2683 " vector registers\n");
2685 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
2686 // We divide by these constants so assume that we have at least one
2687 // instruction that uses at least one register.
2688 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
2689 R.NumInstructions = std::max(R.NumInstructions, 1U);
2691 // We calculate the unroll factor using the following formula.
2692 // Subtract the number of loop invariants from the number of available
2693 // registers. These registers are used by all of the unrolled instances.
2694 // Next, divide the remaining registers by the number of registers that is
2695 // required by the loop, in order to estimate how many parallel instances
2696 // fit without causing spills.
2697 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
2699 // We don't want to unroll the loops to the point where they do not fit into
2700 // the decoded cache. Assume that we only allow 32 IR instructions.
2701 UF = std::min(UF, (32 / R.NumInstructions));
2703 // Clamp the unroll factor ranges to reasonable factors.
2704 if (UF > MaxUnrollSize)
2712 LoopVectorizationCostModel::RegisterUsage
2713 LoopVectorizationCostModel::calculateRegisterUsage() {
2714 // This function calculates the register usage by measuring the highest number
2715 // of values that are alive at a single location. Obviously, this is a very
2716 // rough estimation. We scan the loop in a topological order in order and
2717 // assign a number to each instruction. We use RPO to ensure that defs are
2718 // met before their users. We assume that each instruction that has in-loop
2719 // users starts an interval. We record every time that an in-loop value is
2720 // used, so we have a list of the first and last occurrences of each
2721 // instruction. Next, we transpose this data structure into a multi map that
2722 // holds the list of intervals that *end* at a specific location. This multi
2723 // map allows us to perform a linear search. We scan the instructions linearly
2724 // and record each time that a new interval starts, by placing it in a set.
2725 // If we find this value in the multi-map then we remove it from the set.
2726 // The max register usage is the maximum size of the set.
2727 // We also search for instructions that are defined outside the loop, but are
2728 // used inside the loop. We need this number separately from the max-interval
2729 // usage number because when we unroll, loop-invariant values do not take
2731 LoopBlocksDFS DFS(TheLoop);
2735 R.NumInstructions = 0;
2737 // Each 'key' in the map opens a new interval. The values
2738 // of the map are the index of the 'last seen' usage of the
2739 // instruction that is the key.
2740 typedef DenseMap<Instruction*, unsigned> IntervalMap;
2741 // Maps instruction to its index.
2742 DenseMap<unsigned, Instruction*> IdxToInstr;
2743 // Marks the end of each interval.
2744 IntervalMap EndPoint;
2745 // Saves the list of instruction indices that are used in the loop.
2746 SmallSet<Instruction*, 8> Ends;
2747 // Saves the list of values that are used in the loop but are
2748 // defined outside the loop, such as arguments and constants.
2749 SmallPtrSet<Value*, 8> LoopInvariants;
2752 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2753 be = DFS.endRPO(); bb != be; ++bb) {
2754 R.NumInstructions += (*bb)->size();
2755 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2757 Instruction *I = it;
2758 IdxToInstr[Index++] = I;
2760 // Save the end location of each USE.
2761 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
2762 Value *U = I->getOperand(i);
2763 Instruction *Instr = dyn_cast<Instruction>(U);
2765 // Ignore non-instruction values such as arguments, constants, etc.
2766 if (!Instr) continue;
2768 // If this instruction is outside the loop then record it and continue.
2769 if (!TheLoop->contains(Instr)) {
2770 LoopInvariants.insert(Instr);
2774 // Overwrite previous end points.
2775 EndPoint[Instr] = Index;
2781 // Saves the list of intervals that end with the index in 'key'.
2782 typedef SmallVector<Instruction*, 2> InstrList;
2783 DenseMap<unsigned, InstrList> TransposeEnds;
2785 // Transpose the EndPoints to a list of values that end at each index.
2786 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
2788 TransposeEnds[it->second].push_back(it->first);
2790 SmallSet<Instruction*, 8> OpenIntervals;
2791 unsigned MaxUsage = 0;
2794 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
2795 for (unsigned int i = 0; i < Index; ++i) {
2796 Instruction *I = IdxToInstr[i];
2797 // Ignore instructions that are never used within the loop.
2798 if (!Ends.count(I)) continue;
2800 // Remove all of the instructions that end at this location.
2801 InstrList &List = TransposeEnds[i];
2802 for (unsigned int j=0, e = List.size(); j < e; ++j)
2803 OpenIntervals.erase(List[j]);
2805 // Count the number of live interals.
2806 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
2808 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
2809 OpenIntervals.size() <<"\n");
2811 // Add the current instruction to the list of open intervals.
2812 OpenIntervals.insert(I);
2815 unsigned Invariant = LoopInvariants.size();
2816 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
2817 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
2818 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
2820 R.LoopInvariantRegs = Invariant;
2821 R.MaxLocalUsers = MaxUsage;
2825 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
2829 for (Loop::block_iterator bb = TheLoop->block_begin(),
2830 be = TheLoop->block_end(); bb != be; ++bb) {
2831 unsigned BlockCost = 0;
2832 BasicBlock *BB = *bb;
2834 // For each instruction in the old loop.
2835 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2836 unsigned C = getInstructionCost(it, VF);
2838 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
2839 VF << " For instruction: "<< *it << "\n");
2842 // We assume that if-converted blocks have a 50% chance of being executed.
2843 // When the code is scalar then some of the blocks are avoided due to CF.
2844 // When the code is vectorized we execute all code paths.
2845 if (Legal->blockNeedsPredication(*bb) && VF == 1)
2855 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
2856 // If we know that this instruction will remain uniform, check the cost of
2857 // the scalar version.
2858 if (Legal->isUniformAfterVectorization(I))
2861 Type *RetTy = I->getType();
2862 Type *VectorTy = ToVectorTy(RetTy, VF);
2864 // TODO: We need to estimate the cost of intrinsic calls.
2865 switch (I->getOpcode()) {
2866 case Instruction::GetElementPtr:
2867 // We mark this instruction as zero-cost because scalar GEPs are usually
2868 // lowered to the intruction addressing mode. At the moment we don't
2869 // generate vector geps.
2871 case Instruction::Br: {
2872 return TTI.getCFInstrCost(I->getOpcode());
2874 case Instruction::PHI:
2875 //TODO: IF-converted IFs become selects.
2877 case Instruction::Add:
2878 case Instruction::FAdd:
2879 case Instruction::Sub:
2880 case Instruction::FSub:
2881 case Instruction::Mul:
2882 case Instruction::FMul:
2883 case Instruction::UDiv:
2884 case Instruction::SDiv:
2885 case Instruction::FDiv:
2886 case Instruction::URem:
2887 case Instruction::SRem:
2888 case Instruction::FRem:
2889 case Instruction::Shl:
2890 case Instruction::LShr:
2891 case Instruction::AShr:
2892 case Instruction::And:
2893 case Instruction::Or:
2894 case Instruction::Xor:
2895 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy);
2896 case Instruction::Select: {
2897 SelectInst *SI = cast<SelectInst>(I);
2898 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
2899 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
2900 Type *CondTy = SI->getCondition()->getType();
2902 CondTy = VectorType::get(CondTy, VF);
2904 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
2906 case Instruction::ICmp:
2907 case Instruction::FCmp: {
2908 Type *ValTy = I->getOperand(0)->getType();
2909 VectorTy = ToVectorTy(ValTy, VF);
2910 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
2912 case Instruction::Store: {
2913 StoreInst *SI = cast<StoreInst>(I);
2914 Type *ValTy = SI->getValueOperand()->getType();
2915 VectorTy = ToVectorTy(ValTy, VF);
2918 return TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
2920 SI->getPointerAddressSpace());
2922 // Scalarized stores.
2923 int Stride = Legal->isConsecutivePtr(SI->getPointerOperand());
2924 bool Reverse = Stride < 0;
2928 // The cost of extracting from the value vector and pointer vector.
2929 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2930 for (unsigned i = 0; i < VF; ++i) {
2931 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy,
2933 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
2936 // The cost of the scalar stores.
2937 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
2939 SI->getPointerAddressSpace());
2944 unsigned Cost = TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
2946 SI->getPointerAddressSpace());
2948 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
2952 case Instruction::Load: {
2953 LoadInst *LI = cast<LoadInst>(I);
2956 return TTI.getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(),
2957 LI->getPointerAddressSpace());
2959 // Scalarized loads.
2960 int Stride = Legal->isConsecutivePtr(LI->getPointerOperand());
2961 bool Reverse = Stride < 0;
2964 Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
2966 // The cost of extracting from the pointer vector.
2967 for (unsigned i = 0; i < VF; ++i)
2968 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
2970 // The cost of inserting data to the result vector.
2971 for (unsigned i = 0; i < VF; ++i)
2972 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, VectorTy, i);
2974 // The cost of the scalar stores.
2975 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), RetTy->getScalarType(),
2977 LI->getPointerAddressSpace());
2982 unsigned Cost = TTI.getMemoryOpCost(I->getOpcode(), VectorTy,
2984 LI->getPointerAddressSpace());
2986 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
2989 case Instruction::ZExt:
2990 case Instruction::SExt:
2991 case Instruction::FPToUI:
2992 case Instruction::FPToSI:
2993 case Instruction::FPExt:
2994 case Instruction::PtrToInt:
2995 case Instruction::IntToPtr:
2996 case Instruction::SIToFP:
2997 case Instruction::UIToFP:
2998 case Instruction::Trunc:
2999 case Instruction::FPTrunc:
3000 case Instruction::BitCast: {
3001 // We optimize the truncation of induction variable.
3002 // The cost of these is the same as the scalar operation.
3003 if (I->getOpcode() == Instruction::Trunc &&
3004 Legal->isInductionVariable(I->getOperand(0)))
3005 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3006 I->getOperand(0)->getType());
3008 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3009 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3011 case Instruction::Call: {
3012 assert(isTriviallyVectorizableIntrinsic(I));
3013 IntrinsicInst *II = cast<IntrinsicInst>(I);
3014 Type *RetTy = ToVectorTy(II->getType(), VF);
3015 SmallVector<Type*, 4> Tys;
3016 for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i)
3017 Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF));
3018 return TTI.getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys);
3021 // We are scalarizing the instruction. Return the cost of the scalar
3022 // instruction, plus the cost of insert and extract into vector
3023 // elements, times the vector width.
3026 if (!RetTy->isVoidTy() && VF != 1) {
3027 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3029 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3032 // The cost of inserting the results plus extracting each one of the
3034 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3037 // The cost of executing VF copies of the scalar instruction. This opcode
3038 // is unknown. Assume that it is the same as 'mul'.
3039 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3045 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3046 if (Scalar->isVoidTy() || VF == 1)
3048 return VectorType::get(Scalar, VF);
3051 char LoopVectorize::ID = 0;
3052 static const char lv_name[] = "Loop Vectorization";
3053 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3054 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3055 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3056 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3057 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3058 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3061 Pass *createLoopVectorizePass() {
3062 return new LoopVectorize();