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 vectorizer 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 iterations 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/Target/TargetLibraryInfo.h"
83 #include "llvm/Transforms/Scalar.h"
84 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
85 #include "llvm/Transforms/Utils/Local.h"
91 static cl::opt<unsigned>
92 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
93 cl::desc("Sets the SIMD width. Zero is autoselect."));
95 static cl::opt<unsigned>
96 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
97 cl::desc("Sets the vectorization unroll count. "
98 "Zero is autoselect."));
101 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
102 cl::desc("Enable if-conversion during vectorization."));
104 /// We don't vectorize loops with a known constant trip count below this number.
105 static cl::opt<unsigned>
106 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
108 cl::desc("Don't vectorize loops with a constant "
109 "trip count that is smaller than this "
112 /// We don't unroll loops with a known constant trip count below this number.
113 static const unsigned TinyTripCountUnrollThreshold = 128;
115 /// When performing a runtime memory check, do not check more than this
116 /// number of pointers. Notice that the check is quadratic!
117 static const unsigned RuntimeMemoryCheckThreshold = 4;
121 // Forward declarations.
122 class LoopVectorizationLegality;
123 class LoopVectorizationCostModel;
125 /// InnerLoopVectorizer vectorizes loops which contain only one basic
126 /// block to a specified vectorization factor (VF).
127 /// This class performs the widening of scalars into vectors, or multiple
128 /// scalars. This class also implements the following features:
129 /// * It inserts an epilogue loop for handling loops that don't have iteration
130 /// counts that are known to be a multiple of the vectorization factor.
131 /// * It handles the code generation for reduction variables.
132 /// * Scalarization (implementation using scalars) of un-vectorizable
134 /// InnerLoopVectorizer does not perform any vectorization-legality
135 /// checks, and relies on the caller to check for the different legality
136 /// aspects. The InnerLoopVectorizer relies on the
137 /// LoopVectorizationLegality class to provide information about the induction
138 /// and reduction variables that were found to a given vectorization factor.
139 class InnerLoopVectorizer {
141 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
142 DominatorTree *DT, DataLayout *DL,
143 const TargetLibraryInfo *TLI, unsigned VecWidth,
144 unsigned UnrollFactor)
145 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
146 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
147 OldInduction(0), WidenMap(UnrollFactor) {}
149 // Perform the actual loop widening (vectorization).
150 void vectorize(LoopVectorizationLegality *Legal) {
151 // Create a new empty loop. Unlink the old loop and connect the new one.
152 createEmptyLoop(Legal);
153 // Widen each instruction in the old loop to a new one in the new loop.
154 // Use the Legality module to find the induction and reduction variables.
155 vectorizeLoop(Legal);
156 // Register the new loop and update the analysis passes.
161 /// A small list of PHINodes.
162 typedef SmallVector<PHINode*, 4> PhiVector;
163 /// When we unroll loops we have multiple vector values for each scalar.
164 /// This data structure holds the unrolled and vectorized values that
165 /// originated from one scalar instruction.
166 typedef SmallVector<Value*, 2> VectorParts;
168 /// Add code that checks at runtime if the accessed arrays overlap.
169 /// Returns the comparator value or NULL if no check is needed.
170 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
172 /// Create an empty loop, based on the loop ranges of the old loop.
173 void createEmptyLoop(LoopVectorizationLegality *Legal);
174 /// Copy and widen the instructions from the old loop.
175 void vectorizeLoop(LoopVectorizationLegality *Legal);
177 /// A helper function that computes the predicate of the block BB, assuming
178 /// that the header block of the loop is set to True. It returns the *entry*
179 /// mask for the block BB.
180 VectorParts createBlockInMask(BasicBlock *BB);
181 /// A helper function that computes the predicate of the edge between SRC
183 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
185 /// A helper function to vectorize a single BB within the innermost loop.
186 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
189 /// Insert the new loop to the loop hierarchy and pass manager
190 /// and update the analysis passes.
191 void updateAnalysis();
193 /// This instruction is un-vectorizable. Implement it as a sequence
195 void scalarizeInstruction(Instruction *Instr);
197 /// Vectorize Load and Store instructions,
198 void vectorizeMemoryInstruction(Instruction *Instr,
199 LoopVectorizationLegality *Legal);
201 /// Create a broadcast instruction. This method generates a broadcast
202 /// instruction (shuffle) for loop invariant values and for the induction
203 /// value. If this is the induction variable then we extend it to N, N+1, ...
204 /// this is needed because each iteration in the loop corresponds to a SIMD
206 Value *getBroadcastInstrs(Value *V);
208 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
209 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
210 /// The sequence starts at StartIndex.
211 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
213 /// When we go over instructions in the basic block we rely on previous
214 /// values within the current basic block or on loop invariant values.
215 /// When we widen (vectorize) values we place them in the map. If the values
216 /// are not within the map, they have to be loop invariant, so we simply
217 /// broadcast them into a vector.
218 VectorParts &getVectorValue(Value *V);
220 /// Generate a shuffle sequence that will reverse the vector Vec.
221 Value *reverseVector(Value *Vec);
223 /// This is a helper class that holds the vectorizer state. It maps scalar
224 /// instructions to vector instructions. When the code is 'unrolled' then
225 /// then a single scalar value is mapped to multiple vector parts. The parts
226 /// are stored in the VectorPart type.
228 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
230 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
232 /// \return True if 'Key' is saved in the Value Map.
233 bool has(Value *Key) const { return MapStorage.count(Key); }
235 /// Initializes a new entry in the map. Sets all of the vector parts to the
236 /// save value in 'Val'.
237 /// \return A reference to a vector with splat values.
238 VectorParts &splat(Value *Key, Value *Val) {
239 VectorParts &Entry = MapStorage[Key];
240 Entry.assign(UF, Val);
244 ///\return A reference to the value that is stored at 'Key'.
245 VectorParts &get(Value *Key) {
246 VectorParts &Entry = MapStorage[Key];
249 assert(Entry.size() == UF);
254 /// The unroll factor. Each entry in the map stores this number of vector
258 /// Map storage. We use std::map and not DenseMap because insertions to a
259 /// dense map invalidates its iterators.
260 std::map<Value *, VectorParts> MapStorage;
263 /// The original loop.
265 /// Scev analysis to use.
273 /// Target Library Info.
274 const TargetLibraryInfo *TLI;
276 /// The vectorization SIMD factor to use. Each vector will have this many
279 /// The vectorization unroll factor to use. Each scalar is vectorized to this
280 /// many different vector instructions.
283 /// The builder that we use
286 // --- Vectorization state ---
288 /// The vector-loop preheader.
289 BasicBlock *LoopVectorPreHeader;
290 /// The scalar-loop preheader.
291 BasicBlock *LoopScalarPreHeader;
292 /// Middle Block between the vector and the scalar.
293 BasicBlock *LoopMiddleBlock;
294 ///The ExitBlock of the scalar loop.
295 BasicBlock *LoopExitBlock;
296 ///The vector loop body.
297 BasicBlock *LoopVectorBody;
298 ///The scalar loop body.
299 BasicBlock *LoopScalarBody;
300 /// A list of all bypass blocks. The first block is the entry of the loop.
301 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
303 /// The new Induction variable which was added to the new block.
305 /// The induction variable of the old basic block.
306 PHINode *OldInduction;
307 /// Maps scalars to widened vectors.
311 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
312 /// to what vectorization factor.
313 /// This class does not look at the profitability of vectorization, only the
314 /// legality. This class has two main kinds of checks:
315 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
316 /// will change the order of memory accesses in a way that will change the
317 /// correctness of the program.
318 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
319 /// checks for a number of different conditions, such as the availability of a
320 /// single induction variable, that all types are supported and vectorize-able,
321 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
322 /// This class is also used by InnerLoopVectorizer for identifying
323 /// induction variable and the different reduction variables.
324 class LoopVectorizationLegality {
326 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
327 DominatorTree *DT, TargetTransformInfo* TTI,
328 AliasAnalysis *AA, TargetLibraryInfo *TLI)
329 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
332 /// This enum represents the kinds of reductions that we support.
334 RK_NoReduction, ///< Not a reduction.
335 RK_IntegerAdd, ///< Sum of integers.
336 RK_IntegerMult, ///< Product of integers.
337 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
338 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
339 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
340 RK_FloatAdd, ///< Sum of floats.
341 RK_FloatMult ///< Product of floats.
344 /// This enum represents the kinds of inductions that we support.
346 IK_NoInduction, ///< Not an induction variable.
347 IK_IntInduction, ///< Integer induction variable. Step = 1.
348 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
349 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
350 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
353 /// This POD struct holds information about reduction variables.
354 struct ReductionDescriptor {
355 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
356 Kind(RK_NoReduction) {}
358 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K)
359 : StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
361 // The starting value of the reduction.
362 // It does not have to be zero!
364 // The instruction who's value is used outside the loop.
365 Instruction *LoopExitInstr;
366 // The kind of the reduction.
370 // This POD struct holds information about the memory runtime legality
371 // check that a group of pointers do not overlap.
372 struct RuntimePointerCheck {
373 RuntimePointerCheck() : Need(false) {}
375 /// Reset the state of the pointer runtime information.
383 /// Insert a pointer and calculate the start and end SCEVs.
384 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr);
386 /// This flag indicates if we need to add the runtime check.
388 /// Holds the pointers that we need to check.
389 SmallVector<Value*, 2> Pointers;
390 /// Holds the pointer value at the beginning of the loop.
391 SmallVector<const SCEV*, 2> Starts;
392 /// Holds the pointer value at the end of the loop.
393 SmallVector<const SCEV*, 2> Ends;
396 /// A POD for saving information about induction variables.
397 struct InductionInfo {
398 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
399 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
406 /// ReductionList contains the reduction descriptors for all
407 /// of the reductions that were found in the loop.
408 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
410 /// InductionList saves induction variables and maps them to the
411 /// induction descriptor.
412 typedef MapVector<PHINode*, InductionInfo> InductionList;
414 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
415 /// respective Store/Load instruction(s) to calculate aliasing.
416 typedef DenseMap<Value*, Instruction* > AliasMap;
417 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
419 /// Returns true if it is legal to vectorize this loop.
420 /// This does not mean that it is profitable to vectorize this
421 /// loop, only that it is legal to do so.
424 /// Returns the Induction variable.
425 PHINode *getInduction() { return Induction; }
427 /// Returns the reduction variables found in the loop.
428 ReductionList *getReductionVars() { return &Reductions; }
430 /// Returns the induction variables found in the loop.
431 InductionList *getInductionVars() { return &Inductions; }
433 /// Returns True if V is an induction variable in this loop.
434 bool isInductionVariable(const Value *V);
436 /// Return true if the block BB needs to be predicated in order for the loop
437 /// to be vectorized.
438 bool blockNeedsPredication(BasicBlock *BB);
440 /// Check if this pointer is consecutive when vectorizing. This happens
441 /// when the last index of the GEP is the induction variable, or that the
442 /// pointer itself is an induction variable.
443 /// This check allows us to vectorize A[idx] into a wide load/store.
445 /// 0 - Stride is unknown or non consecutive.
446 /// 1 - Address is consecutive.
447 /// -1 - Address is consecutive, and decreasing.
448 int isConsecutivePtr(Value *Ptr);
450 /// Returns true if the value V is uniform within the loop.
451 bool isUniform(Value *V);
453 /// Returns true if this instruction will remain scalar after vectorization.
454 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
456 /// Returns the information that we collected about runtime memory check.
457 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
459 /// Check if a single basic block loop is vectorizable.
460 /// At this point we know that this is a loop with a constant trip count
461 /// and we only need to check individual instructions.
462 bool canVectorizeInstrs();
464 /// When we vectorize loops we may change the order in which
465 /// we read and write from memory. This method checks if it is
466 /// legal to vectorize the code, considering only memory constrains.
467 /// Returns true if the loop is vectorizable
468 bool canVectorizeMemory();
470 /// Return true if we can vectorize this loop using the IF-conversion
472 bool canVectorizeWithIfConvert();
474 /// Collect the variables that need to stay uniform after vectorization.
475 void collectLoopUniforms();
477 /// Return true if all of the instructions in the block can be speculatively
479 bool blockCanBePredicated(BasicBlock *BB);
481 /// Returns True, if 'Phi' is the kind of reduction variable for type
482 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
483 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
484 /// Returns true if the instruction I can be a reduction variable of type
486 bool isReductionInstr(Instruction *I, ReductionKind Kind);
487 /// Returns the induction kind of Phi. This function may return NoInduction
488 /// if the PHI is not an induction variable.
489 InductionKind isInductionVariable(PHINode *Phi);
490 /// Return true if can compute the address bounds of Ptr within the loop.
491 bool hasComputableBounds(Value *Ptr);
492 /// Return true if there is the chance of write reorder.
493 bool hasPossibleGlobalWriteReorder(Value *Object,
495 AliasMultiMap &WriteObjects,
496 unsigned MaxByteWidth);
497 /// Return the AA location for a load or a store.
498 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
501 /// The loop that we evaluate.
505 /// DataLayout analysis.
510 TargetTransformInfo *TTI;
513 /// Target Library Info.
514 TargetLibraryInfo *TLI;
516 // --- vectorization state --- //
518 /// Holds the integer induction variable. This is the counter of the
521 /// Holds the reduction variables.
522 ReductionList Reductions;
523 /// Holds all of the induction variables that we found in the loop.
524 /// Notice that inductions don't need to start at zero and that induction
525 /// variables can be pointers.
526 InductionList Inductions;
528 /// Allowed outside users. This holds the reduction
529 /// vars which can be accessed from outside the loop.
530 SmallPtrSet<Value*, 4> AllowedExit;
531 /// This set holds the variables which are known to be uniform after
533 SmallPtrSet<Instruction*, 4> Uniforms;
534 /// We need to check that all of the pointers in this list are disjoint
536 RuntimePointerCheck PtrRtCheck;
539 /// LoopVectorizationCostModel - estimates the expected speedups due to
541 /// In many cases vectorization is not profitable. This can happen because of
542 /// a number of reasons. In this class we mainly attempt to predict the
543 /// expected speedup/slowdowns due to the supported instruction set. We use the
544 /// TargetTransformInfo to query the different backends for the cost of
545 /// different operations.
546 class LoopVectorizationCostModel {
548 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
549 LoopVectorizationLegality *Legal,
550 const TargetTransformInfo &TTI,
551 DataLayout *DL, const TargetLibraryInfo *TLI)
552 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
554 /// Information about vectorization costs
555 struct VectorizationFactor {
556 unsigned Width; // Vector width with best cost
557 unsigned Cost; // Cost of the loop with that width
559 /// \return The most profitable vectorization factor and the cost of that VF.
560 /// This method checks every power of two up to VF. If UserVF is not ZERO
561 /// then this vectorization factor will be selected if vectorization is
563 VectorizationFactor selectVectorizationFactor(bool OptForSize,
566 /// \return The size (in bits) of the widest type in the code that
567 /// needs to be vectorized. We ignore values that remain scalar such as
568 /// 64 bit loop indices.
569 unsigned getWidestType();
571 /// \return The most profitable unroll factor.
572 /// If UserUF is non-zero then this method finds the best unroll-factor
573 /// based on register pressure and other parameters.
574 /// VF and LoopCost are the selected vectorization factor and the cost of the
576 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
579 /// \brief A struct that represents some properties of the register usage
581 struct RegisterUsage {
582 /// Holds the number of loop invariant values that are used in the loop.
583 unsigned LoopInvariantRegs;
584 /// Holds the maximum number of concurrent live intervals in the loop.
585 unsigned MaxLocalUsers;
586 /// Holds the number of instructions in the loop.
587 unsigned NumInstructions;
590 /// \return information about the register usage of the loop.
591 RegisterUsage calculateRegisterUsage();
594 /// Returns the expected execution cost. The unit of the cost does
595 /// not matter because we use the 'cost' units to compare different
596 /// vector widths. The cost that is returned is *not* normalized by
597 /// the factor width.
598 unsigned expectedCost(unsigned VF);
600 /// Returns the execution time cost of an instruction for a given vector
601 /// width. Vector width of one means scalar.
602 unsigned getInstructionCost(Instruction *I, unsigned VF);
604 /// A helper function for converting Scalar types to vector types.
605 /// If the incoming type is void, we return void. If the VF is 1, we return
607 static Type* ToVectorTy(Type *Scalar, unsigned VF);
609 /// Returns whether the instruction is a load or store and will be a emitted
610 /// as a vector operation.
611 bool isConsecutiveLoadOrStore(Instruction *I);
613 /// The loop that we evaluate.
617 /// Loop Info analysis.
619 /// Vectorization legality.
620 LoopVectorizationLegality *Legal;
621 /// Vector target information.
622 const TargetTransformInfo &TTI;
623 /// Target data layout information.
625 /// Target Library Info.
626 const TargetLibraryInfo *TLI;
629 /// The LoopVectorize Pass.
630 struct LoopVectorize : public LoopPass {
631 /// Pass identification, replacement for typeid
634 explicit LoopVectorize() : LoopPass(ID) {
635 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
641 TargetTransformInfo *TTI;
644 TargetLibraryInfo *TLI;
646 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
647 // We only vectorize innermost loops.
651 SE = &getAnalysis<ScalarEvolution>();
652 DL = getAnalysisIfAvailable<DataLayout>();
653 LI = &getAnalysis<LoopInfo>();
654 TTI = &getAnalysis<TargetTransformInfo>();
655 DT = &getAnalysis<DominatorTree>();
656 AA = getAnalysisIfAvailable<AliasAnalysis>();
657 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
659 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
660 L->getHeader()->getParent()->getName() << "\"\n");
662 // Check if it is legal to vectorize the loop.
663 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
664 if (!LVL.canVectorize()) {
665 DEBUG(dbgs() << "LV: Not vectorizing.\n");
669 // Use the cost model.
670 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
672 // Check the function attributes to find out if this function should be
673 // optimized for size.
674 Function *F = L->getHeader()->getParent();
675 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
676 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
677 unsigned FnIndex = AttributeSet::FunctionIndex;
678 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
679 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
682 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
683 "attribute is used.\n");
687 // Select the optimal vectorization factor.
688 LoopVectorizationCostModel::VectorizationFactor VF;
689 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
690 // Select the unroll factor.
691 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
695 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
699 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
700 F->getParent()->getModuleIdentifier()<<"\n");
701 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
703 // If we decided that it is *legal* to vectorize the loop then do it.
704 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
707 DEBUG(verifyFunction(*L->getHeader()->getParent()));
711 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
712 LoopPass::getAnalysisUsage(AU);
713 AU.addRequiredID(LoopSimplifyID);
714 AU.addRequiredID(LCSSAID);
715 AU.addRequired<DominatorTree>();
716 AU.addRequired<LoopInfo>();
717 AU.addRequired<ScalarEvolution>();
718 AU.addRequired<TargetTransformInfo>();
719 AU.addPreserved<LoopInfo>();
720 AU.addPreserved<DominatorTree>();
725 } // end anonymous namespace
727 //===----------------------------------------------------------------------===//
728 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
729 // LoopVectorizationCostModel.
730 //===----------------------------------------------------------------------===//
733 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
734 Loop *Lp, Value *Ptr) {
735 const SCEV *Sc = SE->getSCEV(Ptr);
736 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
737 assert(AR && "Invalid addrec expression");
738 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
739 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
740 Pointers.push_back(Ptr);
741 Starts.push_back(AR->getStart());
742 Ends.push_back(ScEnd);
745 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
746 // Save the current insertion location.
747 Instruction *Loc = Builder.GetInsertPoint();
749 // We need to place the broadcast of invariant variables outside the loop.
750 Instruction *Instr = dyn_cast<Instruction>(V);
751 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
752 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
754 // Place the code for broadcasting invariant variables in the new preheader.
756 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
758 // Broadcast the scalar into all locations in the vector.
759 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
761 // Restore the builder insertion point.
763 Builder.SetInsertPoint(Loc);
768 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
770 assert(Val->getType()->isVectorTy() && "Must be a vector");
771 assert(Val->getType()->getScalarType()->isIntegerTy() &&
772 "Elem must be an integer");
774 Type *ITy = Val->getType()->getScalarType();
775 VectorType *Ty = cast<VectorType>(Val->getType());
776 int VLen = Ty->getNumElements();
777 SmallVector<Constant*, 8> Indices;
779 // Create a vector of consecutive numbers from zero to VF.
780 for (int i = 0; i < VLen; ++i) {
781 int Idx = Negate ? (-i): i;
782 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
785 // Add the consecutive indices to the vector value.
786 Constant *Cv = ConstantVector::get(Indices);
787 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
788 return Builder.CreateAdd(Val, Cv, "induction");
791 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
792 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
793 // Make sure that the pointer does not point to structs.
794 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
797 // If this value is a pointer induction variable we know it is consecutive.
798 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
799 if (Phi && Inductions.count(Phi)) {
800 InductionInfo II = Inductions[Phi];
801 if (IK_PtrInduction == II.IK)
803 else if (IK_ReversePtrInduction == II.IK)
807 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
811 unsigned NumOperands = Gep->getNumOperands();
812 Value *LastIndex = Gep->getOperand(NumOperands - 1);
814 Value *GpPtr = Gep->getPointerOperand();
815 // If this GEP value is a consecutive pointer induction variable and all of
816 // the indices are constant then we know it is consecutive. We can
817 Phi = dyn_cast<PHINode>(GpPtr);
818 if (Phi && Inductions.count(Phi)) {
820 // Make sure that the pointer does not point to structs.
821 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
822 if (GepPtrType->getElementType()->isAggregateType())
825 // Make sure that all of the index operands are loop invariant.
826 for (unsigned i = 1; i < NumOperands; ++i)
827 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
830 InductionInfo II = Inductions[Phi];
831 if (IK_PtrInduction == II.IK)
833 else if (IK_ReversePtrInduction == II.IK)
837 // Check that all of the gep indices are uniform except for the last.
838 for (unsigned i = 0; i < NumOperands - 1; ++i)
839 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
842 // We can emit wide load/stores only if the last index is the induction
844 const SCEV *Last = SE->getSCEV(LastIndex);
845 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
846 const SCEV *Step = AR->getStepRecurrence(*SE);
848 // The memory is consecutive because the last index is consecutive
849 // and all other indices are loop invariant.
852 if (Step->isAllOnesValue())
859 bool LoopVectorizationLegality::isUniform(Value *V) {
860 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
863 InnerLoopVectorizer::VectorParts&
864 InnerLoopVectorizer::getVectorValue(Value *V) {
865 assert(V != Induction && "The new induction variable should not be used.");
866 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
868 // If we have this scalar in the map, return it.
870 return WidenMap.get(V);
872 // If this scalar is unknown, assume that it is a constant or that it is
873 // loop invariant. Broadcast V and save the value for future uses.
874 Value *B = getBroadcastInstrs(V);
875 return WidenMap.splat(V, B);
878 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
879 assert(Vec->getType()->isVectorTy() && "Invalid type");
880 SmallVector<Constant*, 8> ShuffleMask;
881 for (unsigned i = 0; i < VF; ++i)
882 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
884 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
885 ConstantVector::get(ShuffleMask),
890 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
891 LoopVectorizationLegality *Legal) {
892 // Attempt to issue a wide load.
893 LoadInst *LI = dyn_cast<LoadInst>(Instr);
894 StoreInst *SI = dyn_cast<StoreInst>(Instr);
896 assert((LI || SI) && "Invalid Load/Store instruction");
898 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
899 Type *DataTy = VectorType::get(ScalarDataTy, VF);
900 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
901 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
903 // If the pointer is loop invariant or if it is non consecutive,
904 // scalarize the load.
905 int Stride = Legal->isConsecutivePtr(Ptr);
906 bool Reverse = Stride < 0;
907 bool UniformLoad = LI && Legal->isUniform(Ptr);
908 if (Stride == 0 || UniformLoad)
909 return scalarizeInstruction(Instr);
911 Constant *Zero = Builder.getInt32(0);
912 VectorParts &Entry = WidenMap.get(Instr);
914 // Handle consecutive loads/stores.
915 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
916 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
917 Value *PtrOperand = Gep->getPointerOperand();
918 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
919 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
921 // Create the new GEP with the new induction variable.
922 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
923 Gep2->setOperand(0, FirstBasePtr);
924 Gep2->setName("gep.indvar.base");
925 Ptr = Builder.Insert(Gep2);
927 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
928 OrigLoop) && "Base ptr must be invariant");
930 // The last index does not have to be the induction. It can be
931 // consecutive and be a function of the index. For example A[I+1];
932 unsigned NumOperands = Gep->getNumOperands();
934 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
935 VectorParts &GEPParts = getVectorValue(LastGepOperand);
936 Value *LastIndex = GEPParts[0];
937 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
939 // Create the new GEP with the new induction variable.
940 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
941 Gep2->setOperand(NumOperands - 1, LastIndex);
942 Gep2->setName("gep.indvar.idx");
943 Ptr = Builder.Insert(Gep2);
945 // Use the induction element ptr.
946 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
947 VectorParts &PtrVal = getVectorValue(Ptr);
948 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
953 assert(!Legal->isUniform(SI->getPointerOperand()) &&
954 "We do not allow storing to uniform addresses");
956 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
957 for (unsigned Part = 0; Part < UF; ++Part) {
958 // Calculate the pointer for the specific unroll-part.
959 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
962 // If we store to reverse consecutive memory locations then we need
963 // to reverse the order of elements in the stored value.
964 StoredVal[Part] = reverseVector(StoredVal[Part]);
965 // If the address is consecutive but reversed, then the
966 // wide store needs to start at the last vector element.
967 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
968 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
971 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
972 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
976 for (unsigned Part = 0; Part < UF; ++Part) {
977 // Calculate the pointer for the specific unroll-part.
978 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
981 // If the address is consecutive but reversed, then the
982 // wide store needs to start at the last vector element.
983 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
984 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
987 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
988 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
989 cast<LoadInst>(LI)->setAlignment(Alignment);
990 Entry[Part] = Reverse ? reverseVector(LI) : LI;
994 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
995 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
996 // Holds vector parameters or scalars, in case of uniform vals.
997 SmallVector<VectorParts, 4> Params;
999 // Find all of the vectorized parameters.
1000 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1001 Value *SrcOp = Instr->getOperand(op);
1003 // If we are accessing the old induction variable, use the new one.
1004 if (SrcOp == OldInduction) {
1005 Params.push_back(getVectorValue(SrcOp));
1009 // Try using previously calculated values.
1010 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1012 // If the src is an instruction that appeared earlier in the basic block
1013 // then it should already be vectorized.
1014 if (SrcInst && OrigLoop->contains(SrcInst)) {
1015 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1016 // The parameter is a vector value from earlier.
1017 Params.push_back(WidenMap.get(SrcInst));
1019 // The parameter is a scalar from outside the loop. Maybe even a constant.
1020 VectorParts Scalars;
1021 Scalars.append(UF, SrcOp);
1022 Params.push_back(Scalars);
1026 assert(Params.size() == Instr->getNumOperands() &&
1027 "Invalid number of operands");
1029 // Does this instruction return a value ?
1030 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1032 Value *UndefVec = IsVoidRetTy ? 0 :
1033 UndefValue::get(VectorType::get(Instr->getType(), VF));
1034 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1035 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1037 // For each scalar that we create:
1038 for (unsigned Width = 0; Width < VF; ++Width) {
1039 // For each vector unroll 'part':
1040 for (unsigned Part = 0; Part < UF; ++Part) {
1041 Instruction *Cloned = Instr->clone();
1043 Cloned->setName(Instr->getName() + ".cloned");
1044 // Replace the operands of the cloned instrucions with extracted scalars.
1045 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1046 Value *Op = Params[op][Part];
1047 // Param is a vector. Need to extract the right lane.
1048 if (Op->getType()->isVectorTy())
1049 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1050 Cloned->setOperand(op, Op);
1053 // Place the cloned scalar in the new loop.
1054 Builder.Insert(Cloned);
1056 // If the original scalar returns a value we need to place it in a vector
1057 // so that future users will be able to use it.
1059 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1060 Builder.getInt32(Width));
1066 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1068 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1069 Legal->getRuntimePointerCheck();
1071 if (!PtrRtCheck->Need)
1074 Instruction *MemoryRuntimeCheck = 0;
1075 unsigned NumPointers = PtrRtCheck->Pointers.size();
1076 SmallVector<Value* , 2> Starts;
1077 SmallVector<Value* , 2> Ends;
1079 SCEVExpander Exp(*SE, "induction");
1081 // Use this type for pointer arithmetic.
1082 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1084 for (unsigned i = 0; i < NumPointers; ++i) {
1085 Value *Ptr = PtrRtCheck->Pointers[i];
1086 const SCEV *Sc = SE->getSCEV(Ptr);
1088 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1089 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1091 Starts.push_back(Ptr);
1092 Ends.push_back(Ptr);
1094 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1096 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1097 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1098 Starts.push_back(Start);
1099 Ends.push_back(End);
1103 IRBuilder<> ChkBuilder(Loc);
1105 for (unsigned i = 0; i < NumPointers; ++i) {
1106 for (unsigned j = i+1; j < NumPointers; ++j) {
1107 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1108 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1109 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1110 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1112 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1113 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1114 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1115 if (MemoryRuntimeCheck)
1116 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1119 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1123 return MemoryRuntimeCheck;
1127 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1129 In this function we generate a new loop. The new loop will contain
1130 the vectorized instructions while the old loop will continue to run the
1133 [ ] <-- vector loop bypass (may consist of multiple blocks).
1136 | [ ] <-- vector pre header.
1140 | [ ]_| <-- vector loop.
1143 >[ ] <--- middle-block.
1146 | [ ] <--- new preheader.
1150 | [ ]_| <-- old scalar loop to handle remainder.
1153 >[ ] <-- exit block.
1157 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1158 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1159 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1160 assert(ExitBlock && "Must have an exit block");
1162 // Some loops have a single integer induction variable, while other loops
1163 // don't. One example is c++ iterators that often have multiple pointer
1164 // induction variables. In the code below we also support a case where we
1165 // don't have a single induction variable.
1166 OldInduction = Legal->getInduction();
1167 Type *IdxTy = OldInduction ? OldInduction->getType() :
1168 DL->getIntPtrType(SE->getContext());
1170 // Find the loop boundaries.
1171 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1172 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1174 // Get the total trip count from the count by adding 1.
1175 ExitCount = SE->getAddExpr(ExitCount,
1176 SE->getConstant(ExitCount->getType(), 1));
1178 // Expand the trip count and place the new instructions in the preheader.
1179 // Notice that the pre-header does not change, only the loop body.
1180 SCEVExpander Exp(*SE, "induction");
1182 // Count holds the overall loop count (N).
1183 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1184 BypassBlock->getTerminator());
1186 // The loop index does not have to start at Zero. Find the original start
1187 // value from the induction PHI node. If we don't have an induction variable
1188 // then we know that it starts at zero.
1189 Value *StartIdx = OldInduction ?
1190 OldInduction->getIncomingValueForBlock(BypassBlock):
1191 ConstantInt::get(IdxTy, 0);
1193 assert(BypassBlock && "Invalid loop structure");
1194 LoopBypassBlocks.push_back(BypassBlock);
1196 // Split the single block loop into the two loop structure described above.
1197 BasicBlock *VectorPH =
1198 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1199 BasicBlock *VecBody =
1200 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1201 BasicBlock *MiddleBlock =
1202 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1203 BasicBlock *ScalarPH =
1204 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1206 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1208 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1210 // Generate the induction variable.
1211 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1212 // The loop step is equal to the vectorization factor (num of SIMD elements)
1213 // times the unroll factor (num of SIMD instructions).
1214 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1216 // This is the IR builder that we use to add all of the logic for bypassing
1217 // the new vector loop.
1218 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1220 // We may need to extend the index in case there is a type mismatch.
1221 // We know that the count starts at zero and does not overflow.
1222 if (Count->getType() != IdxTy) {
1223 // The exit count can be of pointer type. Convert it to the correct
1225 if (ExitCount->getType()->isPointerTy())
1226 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1228 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1231 // Add the start index to the loop count to get the new end index.
1232 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1234 // Now we need to generate the expression for N - (N % VF), which is
1235 // the part that the vectorized body will execute.
1236 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1237 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1238 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1239 "end.idx.rnd.down");
1241 // Now, compare the new count to zero. If it is zero skip the vector loop and
1242 // jump to the scalar loop.
1243 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1246 BasicBlock *LastBypassBlock = BypassBlock;
1248 // Generate the code that checks in runtime if arrays overlap. We put the
1249 // checks into a separate block to make the more common case of few elements
1251 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1252 BypassBlock->getTerminator());
1253 if (MemRuntimeCheck) {
1254 // Create a new block containing the memory check.
1255 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1257 LoopBypassBlocks.push_back(CheckBlock);
1259 // Replace the branch into the memory check block with a conditional branch
1260 // for the "few elements case".
1261 Instruction *OldTerm = BypassBlock->getTerminator();
1262 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1263 OldTerm->eraseFromParent();
1265 Cmp = MemRuntimeCheck;
1266 LastBypassBlock = CheckBlock;
1269 LastBypassBlock->getTerminator()->eraseFromParent();
1270 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1273 // We are going to resume the execution of the scalar loop.
1274 // Go over all of the induction variables that we found and fix the
1275 // PHIs that are left in the scalar version of the loop.
1276 // The starting values of PHI nodes depend on the counter of the last
1277 // iteration in the vectorized loop.
1278 // If we come from a bypass edge then we need to start from the original
1281 // This variable saves the new starting index for the scalar loop.
1282 PHINode *ResumeIndex = 0;
1283 LoopVectorizationLegality::InductionList::iterator I, E;
1284 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1285 for (I = List->begin(), E = List->end(); I != E; ++I) {
1286 PHINode *OrigPhi = I->first;
1287 LoopVectorizationLegality::InductionInfo II = I->second;
1288 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1289 MiddleBlock->getTerminator());
1290 Value *EndValue = 0;
1292 case LoopVectorizationLegality::IK_NoInduction:
1293 llvm_unreachable("Unknown induction");
1294 case LoopVectorizationLegality::IK_IntInduction: {
1295 // Handle the integer induction counter:
1296 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1297 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1298 // We know what the end value is.
1299 EndValue = IdxEndRoundDown;
1300 // We also know which PHI node holds it.
1301 ResumeIndex = ResumeVal;
1304 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1305 // Convert the CountRoundDown variable to the PHI size.
1306 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1307 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1308 Value *CRD = CountRoundDown;
1309 if (CRDSize > IISize)
1310 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1311 II.StartValue->getType(), "tr.crd",
1312 LoopBypassBlocks.back()->getTerminator());
1313 else if (CRDSize < IISize)
1314 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1315 II.StartValue->getType(),
1317 LoopBypassBlocks.back()->getTerminator());
1318 // Handle reverse integer induction counter:
1320 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1321 LoopBypassBlocks.back()->getTerminator());
1324 case LoopVectorizationLegality::IK_PtrInduction: {
1325 // For pointer induction variables, calculate the offset using
1328 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end",
1329 LoopBypassBlocks.back()->getTerminator());
1332 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1333 // The value at the end of the loop for the reverse pointer is calculated
1334 // by creating a GEP with a negative index starting from the start value.
1335 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1336 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown,
1338 LoopBypassBlocks.back()->getTerminator());
1339 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx,
1341 LoopBypassBlocks.back()->getTerminator());
1346 // The new PHI merges the original incoming value, in case of a bypass,
1347 // or the value at the end of the vectorized loop.
1348 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1349 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1350 ResumeVal->addIncoming(EndValue, VecBody);
1352 // Fix the scalar body counter (PHI node).
1353 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1354 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1357 // If we are generating a new induction variable then we also need to
1358 // generate the code that calculates the exit value. This value is not
1359 // simply the end of the counter because we may skip the vectorized body
1360 // in case of a runtime check.
1362 assert(!ResumeIndex && "Unexpected resume value found");
1363 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1364 MiddleBlock->getTerminator());
1365 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1366 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1367 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1370 // Make sure that we found the index where scalar loop needs to continue.
1371 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1372 "Invalid resume Index");
1374 // Add a check in the middle block to see if we have completed
1375 // all of the iterations in the first vector loop.
1376 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1377 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1378 ResumeIndex, "cmp.n",
1379 MiddleBlock->getTerminator());
1381 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1382 // Remove the old terminator.
1383 MiddleBlock->getTerminator()->eraseFromParent();
1385 // Create i+1 and fill the PHINode.
1386 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1387 Induction->addIncoming(StartIdx, VectorPH);
1388 Induction->addIncoming(NextIdx, VecBody);
1389 // Create the compare.
1390 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1391 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1393 // Now we have two terminators. Remove the old one from the block.
1394 VecBody->getTerminator()->eraseFromParent();
1396 // Get ready to start creating new instructions into the vectorized body.
1397 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1399 // Create and register the new vector loop.
1400 Loop* Lp = new Loop();
1401 Loop *ParentLoop = OrigLoop->getParentLoop();
1403 // Insert the new loop into the loop nest and register the new basic blocks.
1405 ParentLoop->addChildLoop(Lp);
1406 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1407 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1408 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1409 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1410 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1412 LI->addTopLevelLoop(Lp);
1415 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1418 LoopVectorPreHeader = VectorPH;
1419 LoopScalarPreHeader = ScalarPH;
1420 LoopMiddleBlock = MiddleBlock;
1421 LoopExitBlock = ExitBlock;
1422 LoopVectorBody = VecBody;
1423 LoopScalarBody = OldBasicBlock;
1426 /// This function returns the identity element (or neutral element) for
1427 /// the operation K.
1429 getReductionIdentity(LoopVectorizationLegality::ReductionKind K, Type *Tp) {
1431 case LoopVectorizationLegality:: RK_IntegerXor:
1432 case LoopVectorizationLegality:: RK_IntegerAdd:
1433 case LoopVectorizationLegality:: RK_IntegerOr:
1434 // Adding, Xoring, Oring zero to a number does not change it.
1435 return ConstantInt::get(Tp, 0);
1436 case LoopVectorizationLegality:: RK_IntegerMult:
1437 // Multiplying a number by 1 does not change it.
1438 return ConstantInt::get(Tp, 1);
1439 case LoopVectorizationLegality:: RK_IntegerAnd:
1440 // AND-ing a number with an all-1 value does not change it.
1441 return ConstantInt::get(Tp, -1, true);
1442 case LoopVectorizationLegality:: RK_FloatMult:
1443 // Multiplying a number by 1 does not change it.
1444 return ConstantFP::get(Tp, 1.0L);
1445 case LoopVectorizationLegality:: RK_FloatAdd:
1446 // Adding zero to a number does not change it.
1447 return ConstantFP::get(Tp, 0.0L);
1449 llvm_unreachable("Unknown reduction kind");
1453 static Intrinsic::ID
1454 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1455 // If we have an intrinsic call, check if it is trivially vectorizable.
1456 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1457 switch (II->getIntrinsicID()) {
1458 case Intrinsic::sqrt:
1459 case Intrinsic::sin:
1460 case Intrinsic::cos:
1461 case Intrinsic::exp:
1462 case Intrinsic::exp2:
1463 case Intrinsic::log:
1464 case Intrinsic::log10:
1465 case Intrinsic::log2:
1466 case Intrinsic::fabs:
1467 case Intrinsic::floor:
1468 case Intrinsic::ceil:
1469 case Intrinsic::trunc:
1470 case Intrinsic::rint:
1471 case Intrinsic::nearbyint:
1472 case Intrinsic::pow:
1473 case Intrinsic::fma:
1474 case Intrinsic::fmuladd:
1475 return II->getIntrinsicID();
1477 return Intrinsic::not_intrinsic;
1482 return Intrinsic::not_intrinsic;
1485 Function *F = CI->getCalledFunction();
1486 // We're going to make assumptions on the semantics of the functions, check
1487 // that the target knows that it's available in this environment.
1488 if (!F || !TLI->getLibFunc(F->getName(), Func))
1489 return Intrinsic::not_intrinsic;
1491 // Otherwise check if we have a call to a function that can be turned into a
1492 // vector intrinsic.
1499 return Intrinsic::sin;
1503 return Intrinsic::cos;
1507 return Intrinsic::exp;
1509 case LibFunc::exp2f:
1510 case LibFunc::exp2l:
1511 return Intrinsic::exp2;
1515 return Intrinsic::log;
1516 case LibFunc::log10:
1517 case LibFunc::log10f:
1518 case LibFunc::log10l:
1519 return Intrinsic::log10;
1521 case LibFunc::log2f:
1522 case LibFunc::log2l:
1523 return Intrinsic::log2;
1525 case LibFunc::fabsf:
1526 case LibFunc::fabsl:
1527 return Intrinsic::fabs;
1528 case LibFunc::floor:
1529 case LibFunc::floorf:
1530 case LibFunc::floorl:
1531 return Intrinsic::floor;
1533 case LibFunc::ceilf:
1534 case LibFunc::ceill:
1535 return Intrinsic::ceil;
1536 case LibFunc::trunc:
1537 case LibFunc::truncf:
1538 case LibFunc::truncl:
1539 return Intrinsic::trunc;
1541 case LibFunc::rintf:
1542 case LibFunc::rintl:
1543 return Intrinsic::rint;
1544 case LibFunc::nearbyint:
1545 case LibFunc::nearbyintf:
1546 case LibFunc::nearbyintl:
1547 return Intrinsic::nearbyint;
1551 return Intrinsic::pow;
1554 return Intrinsic::not_intrinsic;
1557 /// This function translates the reduction kind to an LLVM binary operator.
1558 static Instruction::BinaryOps
1559 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1561 case LoopVectorizationLegality::RK_IntegerAdd:
1562 return Instruction::Add;
1563 case LoopVectorizationLegality::RK_IntegerMult:
1564 return Instruction::Mul;
1565 case LoopVectorizationLegality::RK_IntegerOr:
1566 return Instruction::Or;
1567 case LoopVectorizationLegality::RK_IntegerAnd:
1568 return Instruction::And;
1569 case LoopVectorizationLegality::RK_IntegerXor:
1570 return Instruction::Xor;
1571 case LoopVectorizationLegality::RK_FloatMult:
1572 return Instruction::FMul;
1573 case LoopVectorizationLegality::RK_FloatAdd:
1574 return Instruction::FAdd;
1576 llvm_unreachable("Unknown reduction operation");
1581 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1582 //===------------------------------------------------===//
1584 // Notice: any optimization or new instruction that go
1585 // into the code below should be also be implemented in
1588 //===------------------------------------------------===//
1589 Constant *Zero = Builder.getInt32(0);
1591 // In order to support reduction variables we need to be able to vectorize
1592 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1593 // stages. First, we create a new vector PHI node with no incoming edges.
1594 // We use this value when we vectorize all of the instructions that use the
1595 // PHI. Next, after all of the instructions in the block are complete we
1596 // add the new incoming edges to the PHI. At this point all of the
1597 // instructions in the basic block are vectorized, so we can use them to
1598 // construct the PHI.
1599 PhiVector RdxPHIsToFix;
1601 // Scan the loop in a topological order to ensure that defs are vectorized
1603 LoopBlocksDFS DFS(OrigLoop);
1606 // Vectorize all of the blocks in the original loop.
1607 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1608 be = DFS.endRPO(); bb != be; ++bb)
1609 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1611 // At this point every instruction in the original loop is widened to
1612 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1613 // that we vectorized. The PHI nodes are currently empty because we did
1614 // not want to introduce cycles. Notice that the remaining PHI nodes
1615 // that we need to fix are reduction variables.
1617 // Create the 'reduced' values for each of the induction vars.
1618 // The reduced values are the vector values that we scalarize and combine
1619 // after the loop is finished.
1620 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1622 PHINode *RdxPhi = *it;
1623 assert(RdxPhi && "Unable to recover vectorized PHI");
1625 // Find the reduction variable descriptor.
1626 assert(Legal->getReductionVars()->count(RdxPhi) &&
1627 "Unable to find the reduction variable");
1628 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1629 (*Legal->getReductionVars())[RdxPhi];
1631 // We need to generate a reduction vector from the incoming scalar.
1632 // To do so, we need to generate the 'identity' vector and overide
1633 // one of the elements with the incoming scalar reduction. We need
1634 // to do it in the vector-loop preheader.
1635 Builder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1637 // This is the vector-clone of the value that leaves the loop.
1638 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1639 Type *VecTy = VectorExit[0]->getType();
1641 // Find the reduction identity variable. Zero for addition, or, xor,
1642 // one for multiplication, -1 for And.
1643 Constant *Iden = getReductionIdentity(RdxDesc.Kind, VecTy->getScalarType());
1644 Constant *Identity = ConstantVector::getSplat(VF, Iden);
1646 // This vector is the Identity vector where the first element is the
1647 // incoming scalar reduction.
1648 Value *VectorStart = Builder.CreateInsertElement(Identity,
1649 RdxDesc.StartValue, Zero);
1651 // Fix the vector-loop phi.
1652 // We created the induction variable so we know that the
1653 // preheader is the first entry.
1654 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1656 // Reductions do not have to start at zero. They can start with
1657 // any loop invariant values.
1658 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1659 BasicBlock *Latch = OrigLoop->getLoopLatch();
1660 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1661 VectorParts &Val = getVectorValue(LoopVal);
1662 for (unsigned part = 0; part < UF; ++part) {
1663 // Make sure to add the reduction stat value only to the
1664 // first unroll part.
1665 Value *StartVal = (part == 0) ? VectorStart : Identity;
1666 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1667 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1670 // Before each round, move the insertion point right between
1671 // the PHIs and the values we are going to write.
1672 // This allows us to write both PHINodes and the extractelement
1674 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1676 VectorParts RdxParts;
1677 for (unsigned part = 0; part < UF; ++part) {
1678 // This PHINode contains the vectorized reduction variable, or
1679 // the initial value vector, if we bypass the vector loop.
1680 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1681 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1682 Value *StartVal = (part == 0) ? VectorStart : Identity;
1683 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1684 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1685 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1686 RdxParts.push_back(NewPhi);
1689 // Reduce all of the unrolled parts into a single vector.
1690 Value *ReducedPartRdx = RdxParts[0];
1691 for (unsigned part = 1; part < UF; ++part) {
1692 Instruction::BinaryOps Op = getReductionBinOp(RdxDesc.Kind);
1693 ReducedPartRdx = Builder.CreateBinOp(Op, RdxParts[part], ReducedPartRdx,
1697 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1698 // and vector ops, reducing the set of values being computed by half each
1700 assert(isPowerOf2_32(VF) &&
1701 "Reduction emission only supported for pow2 vectors!");
1702 Value *TmpVec = ReducedPartRdx;
1703 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1704 for (unsigned i = VF; i != 1; i >>= 1) {
1705 // Move the upper half of the vector to the lower half.
1706 for (unsigned j = 0; j != i/2; ++j)
1707 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1709 // Fill the rest of the mask with undef.
1710 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1711 UndefValue::get(Builder.getInt32Ty()));
1714 Builder.CreateShuffleVector(TmpVec,
1715 UndefValue::get(TmpVec->getType()),
1716 ConstantVector::get(ShuffleMask),
1719 Instruction::BinaryOps Op = getReductionBinOp(RdxDesc.Kind);
1720 TmpVec = Builder.CreateBinOp(Op, TmpVec, Shuf, "bin.rdx");
1723 // The result is in the first element of the vector.
1724 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1726 // Now, we need to fix the users of the reduction variable
1727 // inside and outside of the scalar remainder loop.
1728 // We know that the loop is in LCSSA form. We need to update the
1729 // PHI nodes in the exit blocks.
1730 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1731 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1732 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1733 if (!LCSSAPhi) continue;
1735 // All PHINodes need to have a single entry edge, or two if
1736 // we already fixed them.
1737 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1739 // We found our reduction value exit-PHI. Update it with the
1740 // incoming bypass edge.
1741 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1742 // Add an edge coming from the bypass.
1743 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1746 }// end of the LCSSA phi scan.
1748 // Fix the scalar loop reduction variable with the incoming reduction sum
1749 // from the vector body and from the backedge value.
1750 int IncomingEdgeBlockIdx =
1751 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1752 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1753 // Pick the other block.
1754 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1755 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1756 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1757 }// end of for each redux variable.
1759 // The Loop exit block may have single value PHI nodes where the incoming
1760 // value is 'undef'. While vectorizing we only handled real values that
1761 // were defined inside the loop. Here we handle the 'undef case'.
1763 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1764 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1765 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1766 if (!LCSSAPhi) continue;
1767 if (LCSSAPhi->getNumIncomingValues() == 1)
1768 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1773 InnerLoopVectorizer::VectorParts
1774 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1775 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1778 VectorParts SrcMask = createBlockInMask(Src);
1780 // The terminator has to be a branch inst!
1781 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1782 assert(BI && "Unexpected terminator found");
1784 if (BI->isConditional()) {
1785 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1787 if (BI->getSuccessor(0) != Dst)
1788 for (unsigned part = 0; part < UF; ++part)
1789 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1791 for (unsigned part = 0; part < UF; ++part)
1792 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1799 InnerLoopVectorizer::VectorParts
1800 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1801 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1803 // Loop incoming mask is all-one.
1804 if (OrigLoop->getHeader() == BB) {
1805 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1806 return getVectorValue(C);
1809 // This is the block mask. We OR all incoming edges, and with zero.
1810 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1811 VectorParts BlockMask = getVectorValue(Zero);
1814 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1815 VectorParts EM = createEdgeMask(*it, BB);
1816 for (unsigned part = 0; part < UF; ++part)
1817 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1824 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1825 BasicBlock *BB, PhiVector *PV) {
1826 // For each instruction in the old loop.
1827 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1828 VectorParts &Entry = WidenMap.get(it);
1829 switch (it->getOpcode()) {
1830 case Instruction::Br:
1831 // Nothing to do for PHIs and BR, since we already took care of the
1832 // loop control flow instructions.
1834 case Instruction::PHI:{
1835 PHINode* P = cast<PHINode>(it);
1836 // Handle reduction variables:
1837 if (Legal->getReductionVars()->count(P)) {
1838 for (unsigned part = 0; part < UF; ++part) {
1839 // This is phase one of vectorizing PHIs.
1840 Type *VecTy = VectorType::get(it->getType(), VF);
1841 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1842 LoopVectorBody-> getFirstInsertionPt());
1848 // Check for PHI nodes that are lowered to vector selects.
1849 if (P->getParent() != OrigLoop->getHeader()) {
1850 // We know that all PHIs in non header blocks are converted into
1851 // selects, so we don't have to worry about the insertion order and we
1852 // can just use the builder.
1854 // At this point we generate the predication tree. There may be
1855 // duplications since this is a simple recursive scan, but future
1856 // optimizations will clean it up.
1857 VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
1860 for (unsigned part = 0; part < UF; ++part) {
1861 VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
1862 VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
1863 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
1869 // This PHINode must be an induction variable.
1870 // Make sure that we know about it.
1871 assert(Legal->getInductionVars()->count(P) &&
1872 "Not an induction variable");
1874 LoopVectorizationLegality::InductionInfo II =
1875 Legal->getInductionVars()->lookup(P);
1878 case LoopVectorizationLegality::IK_NoInduction:
1879 llvm_unreachable("Unknown induction");
1880 case LoopVectorizationLegality::IK_IntInduction: {
1881 assert(P == OldInduction && "Unexpected PHI");
1882 Value *Broadcasted = getBroadcastInstrs(Induction);
1883 // After broadcasting the induction variable we need to make the
1884 // vector consecutive by adding 0, 1, 2 ...
1885 for (unsigned part = 0; part < UF; ++part)
1886 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
1889 case LoopVectorizationLegality::IK_ReverseIntInduction:
1890 case LoopVectorizationLegality::IK_PtrInduction:
1891 case LoopVectorizationLegality::IK_ReversePtrInduction:
1892 // Handle reverse integer and pointer inductions.
1893 Value *StartIdx = 0;
1894 // If we have a single integer induction variable then use it.
1895 // Otherwise, start counting at zero.
1897 LoopVectorizationLegality::InductionInfo OldII =
1898 Legal->getInductionVars()->lookup(OldInduction);
1899 StartIdx = OldII.StartValue;
1901 StartIdx = ConstantInt::get(Induction->getType(), 0);
1903 // This is the normalized GEP that starts counting at zero.
1904 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
1907 // Handle the reverse integer induction variable case.
1908 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
1909 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
1910 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
1912 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
1915 // This is a new value so do not hoist it out.
1916 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
1917 // After broadcasting the induction variable we need to make the
1918 // vector consecutive by adding ... -3, -2, -1, 0.
1919 for (unsigned part = 0; part < UF; ++part)
1920 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
1924 // Handle the pointer induction variable case.
1925 assert(P->getType()->isPointerTy() && "Unexpected type.");
1927 // Is this a reverse induction ptr or a consecutive induction ptr.
1928 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
1931 // This is the vector of results. Notice that we don't generate
1932 // vector geps because scalar geps result in better code.
1933 for (unsigned part = 0; part < UF; ++part) {
1934 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
1935 for (unsigned int i = 0; i < VF; ++i) {
1936 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
1937 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
1940 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
1942 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
1944 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
1946 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
1947 Builder.getInt32(i),
1950 Entry[part] = VecVal;
1957 case Instruction::Add:
1958 case Instruction::FAdd:
1959 case Instruction::Sub:
1960 case Instruction::FSub:
1961 case Instruction::Mul:
1962 case Instruction::FMul:
1963 case Instruction::UDiv:
1964 case Instruction::SDiv:
1965 case Instruction::FDiv:
1966 case Instruction::URem:
1967 case Instruction::SRem:
1968 case Instruction::FRem:
1969 case Instruction::Shl:
1970 case Instruction::LShr:
1971 case Instruction::AShr:
1972 case Instruction::And:
1973 case Instruction::Or:
1974 case Instruction::Xor: {
1975 // Just widen binops.
1976 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
1977 VectorParts &A = getVectorValue(it->getOperand(0));
1978 VectorParts &B = getVectorValue(it->getOperand(1));
1980 // Use this vector value for all users of the original instruction.
1981 for (unsigned Part = 0; Part < UF; ++Part) {
1982 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
1984 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
1985 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
1986 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
1987 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
1988 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
1990 if (VecOp && isa<PossiblyExactOperator>(VecOp))
1991 VecOp->setIsExact(BinOp->isExact());
1997 case Instruction::Select: {
1999 // If the selector is loop invariant we can create a select
2000 // instruction with a scalar condition. Otherwise, use vector-select.
2001 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2004 // The condition can be loop invariant but still defined inside the
2005 // loop. This means that we can't just use the original 'cond' value.
2006 // We have to take the 'vectorized' value and pick the first lane.
2007 // Instcombine will make this a no-op.
2008 VectorParts &Cond = getVectorValue(it->getOperand(0));
2009 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2010 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2011 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2012 Builder.getInt32(0));
2013 for (unsigned Part = 0; Part < UF; ++Part) {
2014 Entry[Part] = Builder.CreateSelect(
2015 InvariantCond ? ScalarCond : Cond[Part],
2022 case Instruction::ICmp:
2023 case Instruction::FCmp: {
2024 // Widen compares. Generate vector compares.
2025 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2026 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2027 VectorParts &A = getVectorValue(it->getOperand(0));
2028 VectorParts &B = getVectorValue(it->getOperand(1));
2029 for (unsigned Part = 0; Part < UF; ++Part) {
2032 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2034 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2040 case Instruction::Store:
2041 case Instruction::Load:
2042 vectorizeMemoryInstruction(it, Legal);
2044 case Instruction::ZExt:
2045 case Instruction::SExt:
2046 case Instruction::FPToUI:
2047 case Instruction::FPToSI:
2048 case Instruction::FPExt:
2049 case Instruction::PtrToInt:
2050 case Instruction::IntToPtr:
2051 case Instruction::SIToFP:
2052 case Instruction::UIToFP:
2053 case Instruction::Trunc:
2054 case Instruction::FPTrunc:
2055 case Instruction::BitCast: {
2056 CastInst *CI = dyn_cast<CastInst>(it);
2057 /// Optimize the special case where the source is the induction
2058 /// variable. Notice that we can only optimize the 'trunc' case
2059 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2060 /// c. other casts depend on pointer size.
2061 if (CI->getOperand(0) == OldInduction &&
2062 it->getOpcode() == Instruction::Trunc) {
2063 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2065 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2066 for (unsigned Part = 0; Part < UF; ++Part)
2067 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2070 /// Vectorize casts.
2071 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2073 VectorParts &A = getVectorValue(it->getOperand(0));
2074 for (unsigned Part = 0; Part < UF; ++Part)
2075 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2079 case Instruction::Call: {
2080 Module *M = BB->getParent()->getParent();
2081 CallInst *CI = cast<CallInst>(it);
2082 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2083 assert(ID && "Not an intrinsic call!");
2084 for (unsigned Part = 0; Part < UF; ++Part) {
2085 SmallVector<Value*, 4> Args;
2086 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2087 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2088 Args.push_back(Arg[Part]);
2090 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2091 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2092 Entry[Part] = Builder.CreateCall(F, Args);
2098 // All other instructions are unsupported. Scalarize them.
2099 scalarizeInstruction(it);
2102 }// end of for_each instr.
2105 void InnerLoopVectorizer::updateAnalysis() {
2106 // Forget the original basic block.
2107 SE->forgetLoop(OrigLoop);
2109 // Update the dominator tree information.
2110 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2111 "Entry does not dominate exit.");
2113 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2114 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2115 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2116 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2117 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2118 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2119 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2120 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2122 DEBUG(DT->verifyAnalysis());
2125 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2126 if (!EnableIfConversion)
2129 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2130 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2132 // Collect the blocks that need predication.
2133 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2134 BasicBlock *BB = LoopBlocks[i];
2136 // We don't support switch statements inside loops.
2137 if (!isa<BranchInst>(BB->getTerminator()))
2140 // We must have at most two predecessors because we need to convert
2141 // all PHIs to selects.
2142 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
2146 // We must be able to predicate all blocks that need to be predicated.
2147 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2151 // We can if-convert this loop.
2155 bool LoopVectorizationLegality::canVectorize() {
2156 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2158 // We can only vectorize innermost loops.
2159 if (TheLoop->getSubLoopsVector().size())
2162 // We must have a single backedge.
2163 if (TheLoop->getNumBackEdges() != 1)
2166 // We must have a single exiting block.
2167 if (!TheLoop->getExitingBlock())
2170 unsigned NumBlocks = TheLoop->getNumBlocks();
2172 // Check if we can if-convert non single-bb loops.
2173 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2174 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2178 // We need to have a loop header.
2179 BasicBlock *Latch = TheLoop->getLoopLatch();
2180 DEBUG(dbgs() << "LV: Found a loop: " <<
2181 TheLoop->getHeader()->getName() << "\n");
2183 // ScalarEvolution needs to be able to find the exit count.
2184 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2185 if (ExitCount == SE->getCouldNotCompute()) {
2186 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2190 // Do not loop-vectorize loops with a tiny trip count.
2191 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2192 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2193 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2194 "This loop is not worth vectorizing.\n");
2198 // Check if we can vectorize the instructions and CFG in this loop.
2199 if (!canVectorizeInstrs()) {
2200 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2204 // Go over each instruction and look at memory deps.
2205 if (!canVectorizeMemory()) {
2206 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2210 // Collect all of the variables that remain uniform after vectorization.
2211 collectLoopUniforms();
2213 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2214 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2217 // Okay! We can vectorize. At this point we don't have any other mem analysis
2218 // which may limit our maximum vectorization factor, so just return true with
2223 bool LoopVectorizationLegality::canVectorizeInstrs() {
2224 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2225 BasicBlock *Header = TheLoop->getHeader();
2227 // For each block in the loop.
2228 for (Loop::block_iterator bb = TheLoop->block_begin(),
2229 be = TheLoop->block_end(); bb != be; ++bb) {
2231 // Scan the instructions in the block and look for hazards.
2232 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2235 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2236 // This should not happen because the loop should be normalized.
2237 if (Phi->getNumIncomingValues() != 2) {
2238 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2242 // Check that this PHI type is allowed.
2243 if (!Phi->getType()->isIntegerTy() &&
2244 !Phi->getType()->isFloatingPointTy() &&
2245 !Phi->getType()->isPointerTy()) {
2246 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2250 // If this PHINode is not in the header block, then we know that we
2251 // can convert it to select during if-conversion. No need to check if
2252 // the PHIs in this block are induction or reduction variables.
2256 // This is the value coming from the preheader.
2257 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2258 // Check if this is an induction variable.
2259 InductionKind IK = isInductionVariable(Phi);
2261 if (IK_NoInduction != IK) {
2262 // Int inductions are special because we only allow one IV.
2263 if (IK == IK_IntInduction) {
2265 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2271 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2272 Inductions[Phi] = InductionInfo(StartValue, IK);
2276 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2277 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2280 if (AddReductionVar(Phi, RK_IntegerMult)) {
2281 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2284 if (AddReductionVar(Phi, RK_IntegerOr)) {
2285 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2288 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2289 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2292 if (AddReductionVar(Phi, RK_IntegerXor)) {
2293 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2296 if (AddReductionVar(Phi, RK_FloatMult)) {
2297 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2300 if (AddReductionVar(Phi, RK_FloatAdd)) {
2301 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2305 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2307 }// end of PHI handling
2309 // We still don't handle functions.
2310 CallInst *CI = dyn_cast<CallInst>(it);
2311 if (CI && !getIntrinsicIDForCall(CI, TLI)) {
2312 DEBUG(dbgs() << "LV: Found a call site.\n");
2316 // Check that the instruction return type is vectorizable.
2317 if (!VectorType::isValidElementType(it->getType()) &&
2318 !it->getType()->isVoidTy()) {
2319 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2323 // Check that the stored type is vectorizable.
2324 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2325 Type *T = ST->getValueOperand()->getType();
2326 if (!VectorType::isValidElementType(T))
2330 // Reduction instructions are allowed to have exit users.
2331 // All other instructions must not have external users.
2332 if (!AllowedExit.count(it))
2333 //Check that all of the users of the loop are inside the BB.
2334 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2336 Instruction *U = cast<Instruction>(*I);
2337 // This user may be a reduction exit value.
2338 if (!TheLoop->contains(U)) {
2339 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2348 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2349 assert(getInductionVars()->size() && "No induction variables");
2355 void LoopVectorizationLegality::collectLoopUniforms() {
2356 // We now know that the loop is vectorizable!
2357 // Collect variables that will remain uniform after vectorization.
2358 std::vector<Value*> Worklist;
2359 BasicBlock *Latch = TheLoop->getLoopLatch();
2361 // Start with the conditional branch and walk up the block.
2362 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2364 while (Worklist.size()) {
2365 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2366 Worklist.pop_back();
2368 // Look at instructions inside this loop.
2369 // Stop when reaching PHI nodes.
2370 // TODO: we need to follow values all over the loop, not only in this block.
2371 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2374 // This is a known uniform.
2377 // Insert all operands.
2378 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2379 Worklist.push_back(I->getOperand(i));
2384 AliasAnalysis::Location
2385 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2386 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2387 return AA->getLocation(Store);
2388 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2389 return AA->getLocation(Load);
2391 llvm_unreachable("Should be either load or store instruction");
2395 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2398 AliasMultiMap& WriteObjects,
2399 unsigned MaxByteWidth) {
2401 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2403 std::vector<Instruction*>::iterator
2404 it = WriteObjects[Object].begin(),
2405 end = WriteObjects[Object].end();
2407 for (; it != end; ++it) {
2408 Instruction* I = *it;
2412 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2413 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2414 ThatLoc.getWithNewSize(MaxByteWidth)))
2420 bool LoopVectorizationLegality::canVectorizeMemory() {
2422 if (TheLoop->isAnnotatedParallel()) {
2424 << "LV: A loop annotated parallel, ignore memory dependency "
2429 typedef SmallVector<Value*, 16> ValueVector;
2430 typedef SmallPtrSet<Value*, 16> ValueSet;
2431 // Holds the Load and Store *instructions*.
2434 PtrRtCheck.Pointers.clear();
2435 PtrRtCheck.Need = false;
2438 for (Loop::block_iterator bb = TheLoop->block_begin(),
2439 be = TheLoop->block_end(); bb != be; ++bb) {
2441 // Scan the BB and collect legal loads and stores.
2442 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2445 // If this is a load, save it. If this instruction can read from memory
2446 // but is not a load, then we quit. Notice that we don't handle function
2447 // calls that read or write.
2448 if (it->mayReadFromMemory()) {
2449 LoadInst *Ld = dyn_cast<LoadInst>(it);
2450 if (!Ld) return false;
2451 if (!Ld->isSimple()) {
2452 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2455 Loads.push_back(Ld);
2459 // Save 'store' instructions. Abort if other instructions write to memory.
2460 if (it->mayWriteToMemory()) {
2461 StoreInst *St = dyn_cast<StoreInst>(it);
2462 if (!St) return false;
2463 if (!St->isSimple()) {
2464 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2467 Stores.push_back(St);
2472 // Now we have two lists that hold the loads and the stores.
2473 // Next, we find the pointers that they use.
2475 // Check if we see any stores. If there are no stores, then we don't
2476 // care if the pointers are *restrict*.
2477 if (!Stores.size()) {
2478 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2482 // Holds the read and read-write *pointers* that we find. These maps hold
2483 // unique values for pointers (so no need for multi-map).
2485 AliasMap ReadWrites;
2487 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2488 // multiple times on the same object. If the ptr is accessed twice, once
2489 // for read and once for write, it will only appear once (on the write
2490 // list). This is okay, since we are going to check for conflicts between
2491 // writes and between reads and writes, but not between reads and reads.
2494 ValueVector::iterator I, IE;
2495 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2496 StoreInst *ST = cast<StoreInst>(*I);
2497 Value* Ptr = ST->getPointerOperand();
2499 if (isUniform(Ptr)) {
2500 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2504 // If we did *not* see this pointer before, insert it to
2505 // the read-write list. At this phase it is only a 'write' list.
2506 if (Seen.insert(Ptr))
2507 ReadWrites.insert(std::make_pair(Ptr, ST));
2510 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2511 LoadInst *LD = cast<LoadInst>(*I);
2512 Value* Ptr = LD->getPointerOperand();
2513 // If we did *not* see this pointer before, insert it to the
2514 // read list. If we *did* see it before, then it is already in
2515 // the read-write list. This allows us to vectorize expressions
2516 // such as A[i] += x; Because the address of A[i] is a read-write
2517 // pointer. This only works if the index of A[i] is consecutive.
2518 // If the address of i is unknown (for example A[B[i]]) then we may
2519 // read a few words, modify, and write a few words, and some of the
2520 // words may be written to the same address.
2521 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2522 Reads.insert(std::make_pair(Ptr, LD));
2525 // If we write (or read-write) to a single destination and there are no
2526 // other reads in this loop then is it safe to vectorize.
2527 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2528 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2532 // Find pointers with computable bounds. We are going to use this information
2533 // to place a runtime bound check.
2534 bool CanDoRT = true;
2535 AliasMap::iterator MI, ME;
2536 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2537 Value *V = (*MI).first;
2538 if (hasComputableBounds(V)) {
2539 PtrRtCheck.insert(SE, TheLoop, V);
2540 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2546 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2547 Value *V = (*MI).first;
2548 if (hasComputableBounds(V)) {
2549 PtrRtCheck.insert(SE, TheLoop, V);
2550 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2557 // Check that we did not collect too many pointers or found a
2558 // unsizeable pointer.
2559 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
2565 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2568 bool NeedRTCheck = false;
2570 // Biggest vectorized access possible, vector width * unroll factor.
2571 // TODO: We're being very pessimistic here, find a way to know the
2572 // real access width before getting here.
2573 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2574 TTI->getMaximumUnrollFactor();
2575 // Now that the pointers are in two lists (Reads and ReadWrites), we
2576 // can check that there are no conflicts between each of the writes and
2577 // between the writes to the reads.
2578 // Note that WriteObjects duplicates the stores (indexed now by underlying
2579 // objects) to avoid pointing to elements inside ReadWrites.
2580 // TODO: Maybe create a new type where they can interact without duplication.
2581 AliasMultiMap WriteObjects;
2582 ValueVector TempObjects;
2584 // Check that the read-writes do not conflict with other read-write
2586 bool AllWritesIdentified = true;
2587 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2588 Value *Val = (*MI).first;
2589 Instruction *Inst = (*MI).second;
2591 GetUnderlyingObjects(Val, TempObjects, DL);
2592 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2594 if (!isIdentifiedObject(*UI)) {
2595 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2597 AllWritesIdentified = false;
2600 // Never seen it before, can't alias.
2601 if (WriteObjects[*UI].empty()) {
2602 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2603 WriteObjects[*UI].push_back(Inst);
2606 // Direct alias found.
2607 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2608 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2612 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2614 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2615 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2617 // If global alias, make sure they do alias.
2618 if (hasPossibleGlobalWriteReorder(*UI,
2622 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2627 // Didn't alias, insert into map for further reference.
2628 WriteObjects[*UI].push_back(Inst);
2630 TempObjects.clear();
2633 /// Check that the reads don't conflict with the read-writes.
2634 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2635 Value *Val = (*MI).first;
2636 GetUnderlyingObjects(Val, TempObjects, DL);
2637 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2639 // If all of the writes are identified then we don't care if the read
2640 // pointer is identified or not.
2641 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2642 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2646 // Never seen it before, can't alias.
2647 if (WriteObjects[*UI].empty())
2649 // Direct alias found.
2650 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2651 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2655 DEBUG(dbgs() << "LV: Found a global value: "
2657 Instruction *Inst = (*MI).second;
2658 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2659 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2661 // If global alias, make sure they do alias.
2662 if (hasPossibleGlobalWriteReorder(*UI,
2666 DEBUG(dbgs() << "LV: Found a possible read-write reorder:"
2671 TempObjects.clear();
2674 PtrRtCheck.Need = NeedRTCheck;
2675 if (NeedRTCheck && !CanDoRT) {
2676 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2677 "the array bounds.\n");
2682 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2683 " need a runtime memory check.\n");
2687 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2688 ReductionKind Kind) {
2689 if (Phi->getNumIncomingValues() != 2)
2692 // Reduction variables are only found in the loop header block.
2693 if (Phi->getParent() != TheLoop->getHeader())
2696 // Obtain the reduction start value from the value that comes from the loop
2698 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2700 // ExitInstruction is the single value which is used outside the loop.
2701 // We only allow for a single reduction value to be used outside the loop.
2702 // This includes users of the reduction, variables (which form a cycle
2703 // which ends in the phi node).
2704 Instruction *ExitInstruction = 0;
2705 // Indicates that we found a binary operation in our scan.
2706 bool FoundBinOp = false;
2708 // Iter is our iterator. We start with the PHI node and scan for all of the
2709 // users of this instruction. All users must be instructions that can be
2710 // used as reduction variables (such as ADD). We may have a single
2711 // out-of-block user. The cycle must end with the original PHI.
2712 Instruction *Iter = Phi;
2714 // If the instruction has no users then this is a broken
2715 // chain and can't be a reduction variable.
2716 if (Iter->use_empty())
2719 // Did we find a user inside this loop already ?
2720 bool FoundInBlockUser = false;
2721 // Did we reach the initial PHI node already ?
2722 bool FoundStartPHI = false;
2724 // Is this a bin op ?
2725 FoundBinOp |= !isa<PHINode>(Iter);
2727 // Remember the current instruction.
2728 Instruction *OldIter = Iter;
2730 // For each of the *users* of iter.
2731 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2733 Instruction *U = cast<Instruction>(*it);
2734 // We already know that the PHI is a user.
2736 FoundStartPHI = true;
2740 // Check if we found the exit user.
2741 BasicBlock *Parent = U->getParent();
2742 if (!TheLoop->contains(Parent)) {
2743 // Exit if you find multiple outside users.
2744 if (ExitInstruction != 0)
2746 ExitInstruction = Iter;
2749 // We allow in-loop PHINodes which are not the original reduction PHI
2750 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2751 // structure) then don't skip this PHI.
2752 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2753 U->getParent() != TheLoop->getHeader() &&
2754 TheLoop->contains(U) &&
2755 Iter->hasNUsesOrMore(2))
2758 // We can't have multiple inside users.
2759 if (FoundInBlockUser)
2761 FoundInBlockUser = true;
2763 // Any reduction instr must be of one of the allowed kinds.
2764 if (!isReductionInstr(U, Kind))
2767 // Reductions of instructions such as Div, and Sub is only
2768 // possible if the LHS is the reduction variable.
2769 if (!U->isCommutative() && !isa<PHINode>(U) && U->getOperand(0) != Iter)
2775 // If all uses were skipped this can't be a reduction variable.
2776 if (Iter == OldIter)
2779 // We found a reduction var if we have reached the original
2780 // phi node and we only have a single instruction with out-of-loop
2782 if (FoundStartPHI) {
2783 // This instruction is allowed to have out-of-loop users.
2784 AllowedExit.insert(ExitInstruction);
2786 // Save the description of this reduction variable.
2787 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
2788 Reductions[Phi] = RD;
2789 // We've ended the cycle. This is a reduction variable if we have an
2790 // outside user and it has a binary op.
2791 return FoundBinOp && ExitInstruction;
2797 LoopVectorizationLegality::isReductionInstr(Instruction *I,
2798 ReductionKind Kind) {
2799 bool FP = I->getType()->isFloatingPointTy();
2800 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
2802 switch (I->getOpcode()) {
2805 case Instruction::PHI:
2806 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd))
2810 case Instruction::Sub:
2811 case Instruction::Add:
2812 return Kind == RK_IntegerAdd;
2813 case Instruction::SDiv:
2814 case Instruction::UDiv:
2815 case Instruction::Mul:
2816 return Kind == RK_IntegerMult;
2817 case Instruction::And:
2818 return Kind == RK_IntegerAnd;
2819 case Instruction::Or:
2820 return Kind == RK_IntegerOr;
2821 case Instruction::Xor:
2822 return Kind == RK_IntegerXor;
2823 case Instruction::FMul:
2824 return Kind == RK_FloatMult && FastMath;
2825 case Instruction::FAdd:
2826 return Kind == RK_FloatAdd && FastMath;
2830 LoopVectorizationLegality::InductionKind
2831 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
2832 Type *PhiTy = Phi->getType();
2833 // We only handle integer and pointer inductions variables.
2834 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
2835 return IK_NoInduction;
2837 // Check that the PHI is consecutive.
2838 const SCEV *PhiScev = SE->getSCEV(Phi);
2839 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2841 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
2842 return IK_NoInduction;
2844 const SCEV *Step = AR->getStepRecurrence(*SE);
2846 // Integer inductions need to have a stride of one.
2847 if (PhiTy->isIntegerTy()) {
2849 return IK_IntInduction;
2850 if (Step->isAllOnesValue())
2851 return IK_ReverseIntInduction;
2852 return IK_NoInduction;
2855 // Calculate the pointer stride and check if it is consecutive.
2856 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
2858 return IK_NoInduction;
2860 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
2861 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
2862 if (C->getValue()->equalsInt(Size))
2863 return IK_PtrInduction;
2864 else if (C->getValue()->equalsInt(0 - Size))
2865 return IK_ReversePtrInduction;
2867 return IK_NoInduction;
2870 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
2871 Value *In0 = const_cast<Value*>(V);
2872 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
2876 return Inductions.count(PN);
2879 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
2880 assert(TheLoop->contains(BB) && "Unknown block used");
2882 // Blocks that do not dominate the latch need predication.
2883 BasicBlock* Latch = TheLoop->getLoopLatch();
2884 return !DT->dominates(BB, Latch);
2887 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
2888 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2889 // We don't predicate loads/stores at the moment.
2890 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
2893 // The instructions below can trap.
2894 switch (it->getOpcode()) {
2896 case Instruction::UDiv:
2897 case Instruction::SDiv:
2898 case Instruction::URem:
2899 case Instruction::SRem:
2907 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
2908 const SCEV *PhiScev = SE->getSCEV(Ptr);
2909 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
2913 return AR->isAffine();
2916 LoopVectorizationCostModel::VectorizationFactor
2917 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
2919 // Width 1 means no vectorize
2920 VectorizationFactor Factor = { 1U, 0U };
2921 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
2922 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
2926 // Find the trip count.
2927 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
2928 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
2930 unsigned WidestType = getWidestType();
2931 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
2932 unsigned MaxVectorSize = WidestRegister / WidestType;
2933 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
2934 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
2936 if (MaxVectorSize == 0) {
2937 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
2941 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
2942 " into one vector!");
2944 unsigned VF = MaxVectorSize;
2946 // If we optimize the program for size, avoid creating the tail loop.
2948 // If we are unable to calculate the trip count then don't try to vectorize.
2950 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2954 // Find the maximum SIMD width that can fit within the trip count.
2955 VF = TC % MaxVectorSize;
2960 // If the trip count that we found modulo the vectorization factor is not
2961 // zero then we require a tail.
2963 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
2969 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
2970 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
2972 Factor.Width = UserVF;
2976 float Cost = expectedCost(1);
2978 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
2979 for (unsigned i=2; i <= VF; i*=2) {
2980 // Notice that the vector loop needs to be executed less times, so
2981 // we need to divide the cost of the vector loops by the width of
2982 // the vector elements.
2983 float VectorCost = expectedCost(i) / (float)i;
2984 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
2985 (int)VectorCost << ".\n");
2986 if (VectorCost < Cost) {
2992 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
2993 Factor.Width = Width;
2994 Factor.Cost = Width * Cost;
2998 unsigned LoopVectorizationCostModel::getWidestType() {
2999 unsigned MaxWidth = 8;
3002 for (Loop::block_iterator bb = TheLoop->block_begin(),
3003 be = TheLoop->block_end(); bb != be; ++bb) {
3004 BasicBlock *BB = *bb;
3006 // For each instruction in the loop.
3007 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3008 Type *T = it->getType();
3010 // Only examine Loads, Stores and PHINodes.
3011 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3014 // Examine PHI nodes that are reduction variables.
3015 if (PHINode *PN = dyn_cast<PHINode>(it))
3016 if (!Legal->getReductionVars()->count(PN))
3019 // Examine the stored values.
3020 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3021 T = ST->getValueOperand()->getType();
3023 // Ignore loaded pointer types and stored pointer types that are not
3024 // consecutive. However, we do want to take consecutive stores/loads of
3025 // pointer vectors into account.
3026 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3029 MaxWidth = std::max(MaxWidth,
3030 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3038 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3041 unsigned LoopCost) {
3043 // -- The unroll heuristics --
3044 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3045 // There are many micro-architectural considerations that we can't predict
3046 // at this level. For example frontend pressure (on decode or fetch) due to
3047 // code size, or the number and capabilities of the execution ports.
3049 // We use the following heuristics to select the unroll factor:
3050 // 1. If the code has reductions the we unroll in order to break the cross
3051 // iteration dependency.
3052 // 2. If the loop is really small then we unroll in order to reduce the loop
3054 // 3. We don't unroll if we think that we will spill registers to memory due
3055 // to the increased register pressure.
3057 // Use the user preference, unless 'auto' is selected.
3061 // When we optimize for size we don't unroll.
3065 // Do not unroll loops with a relatively small trip count.
3066 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3067 TheLoop->getLoopLatch());
3068 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3071 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3072 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3073 " vector registers\n");
3075 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3076 // We divide by these constants so assume that we have at least one
3077 // instruction that uses at least one register.
3078 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3079 R.NumInstructions = std::max(R.NumInstructions, 1U);
3081 // We calculate the unroll factor using the following formula.
3082 // Subtract the number of loop invariants from the number of available
3083 // registers. These registers are used by all of the unrolled instances.
3084 // Next, divide the remaining registers by the number of registers that is
3085 // required by the loop, in order to estimate how many parallel instances
3086 // fit without causing spills.
3087 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3089 // Clamp the unroll factor ranges to reasonable factors.
3090 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3092 // If we did not calculate the cost for VF (because the user selected the VF)
3093 // then we calculate the cost of VF here.
3095 LoopCost = expectedCost(VF);
3097 // Clamp the calculated UF to be between the 1 and the max unroll factor
3098 // that the target allows.
3099 if (UF > MaxUnrollSize)
3104 if (Legal->getReductionVars()->size()) {
3105 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3109 // We want to unroll tiny loops in order to reduce the loop overhead.
3110 // We assume that the cost overhead is 1 and we use the cost model
3111 // to estimate the cost of the loop and unroll until the cost of the
3112 // loop overhead is about 5% of the cost of the loop.
3113 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3114 if (LoopCost < 20) {
3115 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3116 unsigned NewUF = 20/LoopCost + 1;
3117 return std::min(NewUF, UF);
3120 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3124 LoopVectorizationCostModel::RegisterUsage
3125 LoopVectorizationCostModel::calculateRegisterUsage() {
3126 // This function calculates the register usage by measuring the highest number
3127 // of values that are alive at a single location. Obviously, this is a very
3128 // rough estimation. We scan the loop in a topological order in order and
3129 // assign a number to each instruction. We use RPO to ensure that defs are
3130 // met before their users. We assume that each instruction that has in-loop
3131 // users starts an interval. We record every time that an in-loop value is
3132 // used, so we have a list of the first and last occurrences of each
3133 // instruction. Next, we transpose this data structure into a multi map that
3134 // holds the list of intervals that *end* at a specific location. This multi
3135 // map allows us to perform a linear search. We scan the instructions linearly
3136 // and record each time that a new interval starts, by placing it in a set.
3137 // If we find this value in the multi-map then we remove it from the set.
3138 // The max register usage is the maximum size of the set.
3139 // We also search for instructions that are defined outside the loop, but are
3140 // used inside the loop. We need this number separately from the max-interval
3141 // usage number because when we unroll, loop-invariant values do not take
3143 LoopBlocksDFS DFS(TheLoop);
3147 R.NumInstructions = 0;
3149 // Each 'key' in the map opens a new interval. The values
3150 // of the map are the index of the 'last seen' usage of the
3151 // instruction that is the key.
3152 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3153 // Maps instruction to its index.
3154 DenseMap<unsigned, Instruction*> IdxToInstr;
3155 // Marks the end of each interval.
3156 IntervalMap EndPoint;
3157 // Saves the list of instruction indices that are used in the loop.
3158 SmallSet<Instruction*, 8> Ends;
3159 // Saves the list of values that are used in the loop but are
3160 // defined outside the loop, such as arguments and constants.
3161 SmallPtrSet<Value*, 8> LoopInvariants;
3164 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3165 be = DFS.endRPO(); bb != be; ++bb) {
3166 R.NumInstructions += (*bb)->size();
3167 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3169 Instruction *I = it;
3170 IdxToInstr[Index++] = I;
3172 // Save the end location of each USE.
3173 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3174 Value *U = I->getOperand(i);
3175 Instruction *Instr = dyn_cast<Instruction>(U);
3177 // Ignore non-instruction values such as arguments, constants, etc.
3178 if (!Instr) continue;
3180 // If this instruction is outside the loop then record it and continue.
3181 if (!TheLoop->contains(Instr)) {
3182 LoopInvariants.insert(Instr);
3186 // Overwrite previous end points.
3187 EndPoint[Instr] = Index;
3193 // Saves the list of intervals that end with the index in 'key'.
3194 typedef SmallVector<Instruction*, 2> InstrList;
3195 DenseMap<unsigned, InstrList> TransposeEnds;
3197 // Transpose the EndPoints to a list of values that end at each index.
3198 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3200 TransposeEnds[it->second].push_back(it->first);
3202 SmallSet<Instruction*, 8> OpenIntervals;
3203 unsigned MaxUsage = 0;
3206 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3207 for (unsigned int i = 0; i < Index; ++i) {
3208 Instruction *I = IdxToInstr[i];
3209 // Ignore instructions that are never used within the loop.
3210 if (!Ends.count(I)) continue;
3212 // Remove all of the instructions that end at this location.
3213 InstrList &List = TransposeEnds[i];
3214 for (unsigned int j=0, e = List.size(); j < e; ++j)
3215 OpenIntervals.erase(List[j]);
3217 // Count the number of live interals.
3218 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3220 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3221 OpenIntervals.size() <<"\n");
3223 // Add the current instruction to the list of open intervals.
3224 OpenIntervals.insert(I);
3227 unsigned Invariant = LoopInvariants.size();
3228 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3229 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3230 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3232 R.LoopInvariantRegs = Invariant;
3233 R.MaxLocalUsers = MaxUsage;
3237 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3241 for (Loop::block_iterator bb = TheLoop->block_begin(),
3242 be = TheLoop->block_end(); bb != be; ++bb) {
3243 unsigned BlockCost = 0;
3244 BasicBlock *BB = *bb;
3246 // For each instruction in the old loop.
3247 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3248 unsigned C = getInstructionCost(it, VF);
3250 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3251 VF << " For instruction: "<< *it << "\n");
3254 // We assume that if-converted blocks have a 50% chance of being executed.
3255 // When the code is scalar then some of the blocks are avoided due to CF.
3256 // When the code is vectorized we execute all code paths.
3257 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3267 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3268 // If we know that this instruction will remain uniform, check the cost of
3269 // the scalar version.
3270 if (Legal->isUniformAfterVectorization(I))
3273 Type *RetTy = I->getType();
3274 Type *VectorTy = ToVectorTy(RetTy, VF);
3276 // TODO: We need to estimate the cost of intrinsic calls.
3277 switch (I->getOpcode()) {
3278 case Instruction::GetElementPtr:
3279 // We mark this instruction as zero-cost because the cost of GEPs in
3280 // vectorized code depends on whether the corresponding memory instruction
3281 // is scalarized or not. Therefore, we handle GEPs with the memory
3282 // instruction cost.
3284 case Instruction::Br: {
3285 return TTI.getCFInstrCost(I->getOpcode());
3287 case Instruction::PHI:
3288 //TODO: IF-converted IFs become selects.
3290 case Instruction::Add:
3291 case Instruction::FAdd:
3292 case Instruction::Sub:
3293 case Instruction::FSub:
3294 case Instruction::Mul:
3295 case Instruction::FMul:
3296 case Instruction::UDiv:
3297 case Instruction::SDiv:
3298 case Instruction::FDiv:
3299 case Instruction::URem:
3300 case Instruction::SRem:
3301 case Instruction::FRem:
3302 case Instruction::Shl:
3303 case Instruction::LShr:
3304 case Instruction::AShr:
3305 case Instruction::And:
3306 case Instruction::Or:
3307 case Instruction::Xor:
3308 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy);
3309 case Instruction::Select: {
3310 SelectInst *SI = cast<SelectInst>(I);
3311 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3312 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3313 Type *CondTy = SI->getCondition()->getType();
3315 CondTy = VectorType::get(CondTy, VF);
3317 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3319 case Instruction::ICmp:
3320 case Instruction::FCmp: {
3321 Type *ValTy = I->getOperand(0)->getType();
3322 VectorTy = ToVectorTy(ValTy, VF);
3323 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3325 case Instruction::Store:
3326 case Instruction::Load: {
3327 StoreInst *SI = dyn_cast<StoreInst>(I);
3328 LoadInst *LI = dyn_cast<LoadInst>(I);
3329 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3331 VectorTy = ToVectorTy(ValTy, VF);
3333 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3334 unsigned AS = SI ? SI->getPointerAddressSpace() :
3335 LI->getPointerAddressSpace();
3336 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3337 // We add the cost of address computation here instead of with the gep
3338 // instruction because only here we know whether the operation is
3341 return TTI.getAddressComputationCost(VectorTy) +
3342 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3344 // Scalarized loads/stores.
3345 int Stride = Legal->isConsecutivePtr(Ptr);
3346 bool Reverse = Stride < 0;
3349 // The cost of extracting from the value vector and pointer vector.
3350 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3351 for (unsigned i = 0; i < VF; ++i) {
3352 // The cost of extracting the pointer operand.
3353 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3354 // In case of STORE, the cost of ExtractElement from the vector.
3355 // In case of LOAD, the cost of InsertElement into the returned
3357 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3358 Instruction::InsertElement,
3362 // The cost of the scalar loads/stores.
3363 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3364 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3369 // Wide load/stores.
3370 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3371 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3374 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3378 case Instruction::ZExt:
3379 case Instruction::SExt:
3380 case Instruction::FPToUI:
3381 case Instruction::FPToSI:
3382 case Instruction::FPExt:
3383 case Instruction::PtrToInt:
3384 case Instruction::IntToPtr:
3385 case Instruction::SIToFP:
3386 case Instruction::UIToFP:
3387 case Instruction::Trunc:
3388 case Instruction::FPTrunc:
3389 case Instruction::BitCast: {
3390 // We optimize the truncation of induction variable.
3391 // The cost of these is the same as the scalar operation.
3392 if (I->getOpcode() == Instruction::Trunc &&
3393 Legal->isInductionVariable(I->getOperand(0)))
3394 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3395 I->getOperand(0)->getType());
3397 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3398 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3400 case Instruction::Call: {
3401 CallInst *CI = cast<CallInst>(I);
3402 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3403 assert(ID && "Not an intrinsic call!");
3404 Type *RetTy = ToVectorTy(CI->getType(), VF);
3405 SmallVector<Type*, 4> Tys;
3406 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3407 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3408 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3411 // We are scalarizing the instruction. Return the cost of the scalar
3412 // instruction, plus the cost of insert and extract into vector
3413 // elements, times the vector width.
3416 if (!RetTy->isVoidTy() && VF != 1) {
3417 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3419 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3422 // The cost of inserting the results plus extracting each one of the
3424 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3427 // The cost of executing VF copies of the scalar instruction. This opcode
3428 // is unknown. Assume that it is the same as 'mul'.
3429 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3435 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3436 if (Scalar->isVoidTy() || VF == 1)
3438 return VectorType::get(Scalar, VF);
3441 char LoopVectorize::ID = 0;
3442 static const char lv_name[] = "Loop Vectorization";
3443 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3444 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3445 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3446 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3447 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3448 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3451 Pass *createLoopVectorizePass() {
3452 return new LoopVectorize();
3456 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3457 // Check for a store.
3458 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3459 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3461 // Check for a load.
3462 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3463 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;