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.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
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/PatternMatch.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/Support/ValueHandle.h"
84 #include "llvm/Target/TargetLibraryInfo.h"
85 #include "llvm/Transforms/Scalar.h"
86 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
87 #include "llvm/Transforms/Utils/Local.h"
92 using namespace llvm::PatternMatch;
94 static cl::opt<unsigned>
95 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
96 cl::desc("Sets the SIMD width. Zero is autoselect."));
98 static cl::opt<unsigned>
99 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
100 cl::desc("Sets the vectorization unroll count. "
101 "Zero is autoselect."));
104 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
105 cl::desc("Enable if-conversion during vectorization."));
107 /// We don't vectorize loops with a known constant trip count below this number.
108 static cl::opt<unsigned>
109 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
111 cl::desc("Don't vectorize loops with a constant "
112 "trip count that is smaller than this "
115 /// We don't unroll loops with a known constant trip count below this number.
116 static const unsigned TinyTripCountUnrollThreshold = 128;
118 /// When performing memory disambiguation checks at runtime do not make more
119 /// than this number of comparisons.
120 static const unsigned RuntimeMemoryCheckThreshold = 8;
122 /// Maximum simd width.
123 static const unsigned MaxVectorWidth = 64;
125 /// Maximum vectorization unroll count.
126 static const unsigned MaxUnrollFactor = 16;
130 // Forward declarations.
131 class LoopVectorizationLegality;
132 class LoopVectorizationCostModel;
134 /// InnerLoopVectorizer vectorizes loops which contain only one basic
135 /// block to a specified vectorization factor (VF).
136 /// This class performs the widening of scalars into vectors, or multiple
137 /// scalars. This class also implements the following features:
138 /// * It inserts an epilogue loop for handling loops that don't have iteration
139 /// counts that are known to be a multiple of the vectorization factor.
140 /// * It handles the code generation for reduction variables.
141 /// * Scalarization (implementation using scalars) of un-vectorizable
143 /// InnerLoopVectorizer does not perform any vectorization-legality
144 /// checks, and relies on the caller to check for the different legality
145 /// aspects. The InnerLoopVectorizer relies on the
146 /// LoopVectorizationLegality class to provide information about the induction
147 /// and reduction variables that were found to a given vectorization factor.
148 class InnerLoopVectorizer {
150 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
151 DominatorTree *DT, DataLayout *DL,
152 const TargetLibraryInfo *TLI, unsigned VecWidth,
153 unsigned UnrollFactor)
154 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
155 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
156 OldInduction(0), WidenMap(UnrollFactor) {}
158 // Perform the actual loop widening (vectorization).
159 void vectorize(LoopVectorizationLegality *Legal) {
160 // Create a new empty loop. Unlink the old loop and connect the new one.
161 createEmptyLoop(Legal);
162 // Widen each instruction in the old loop to a new one in the new loop.
163 // Use the Legality module to find the induction and reduction variables.
164 vectorizeLoop(Legal);
165 // Register the new loop and update the analysis passes.
170 /// A small list of PHINodes.
171 typedef SmallVector<PHINode*, 4> PhiVector;
172 /// When we unroll loops we have multiple vector values for each scalar.
173 /// This data structure holds the unrolled and vectorized values that
174 /// originated from one scalar instruction.
175 typedef SmallVector<Value*, 2> VectorParts;
177 /// Add code that checks at runtime if the accessed arrays overlap.
178 /// Returns the comparator value or NULL if no check is needed.
179 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
181 /// Create an empty loop, based on the loop ranges of the old loop.
182 void createEmptyLoop(LoopVectorizationLegality *Legal);
183 /// Copy and widen the instructions from the old loop.
184 void vectorizeLoop(LoopVectorizationLegality *Legal);
186 /// A helper function that computes the predicate of the block BB, assuming
187 /// that the header block of the loop is set to True. It returns the *entry*
188 /// mask for the block BB.
189 VectorParts createBlockInMask(BasicBlock *BB);
190 /// A helper function that computes the predicate of the edge between SRC
192 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
194 /// A helper function to vectorize a single BB within the innermost loop.
195 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
198 /// Insert the new loop to the loop hierarchy and pass manager
199 /// and update the analysis passes.
200 void updateAnalysis();
202 /// This instruction is un-vectorizable. Implement it as a sequence
204 void scalarizeInstruction(Instruction *Instr);
206 /// Vectorize Load and Store instructions,
207 void vectorizeMemoryInstruction(Instruction *Instr,
208 LoopVectorizationLegality *Legal);
210 /// Create a broadcast instruction. This method generates a broadcast
211 /// instruction (shuffle) for loop invariant values and for the induction
212 /// value. If this is the induction variable then we extend it to N, N+1, ...
213 /// this is needed because each iteration in the loop corresponds to a SIMD
215 Value *getBroadcastInstrs(Value *V);
217 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
218 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
219 /// The sequence starts at StartIndex.
220 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
222 /// When we go over instructions in the basic block we rely on previous
223 /// values within the current basic block or on loop invariant values.
224 /// When we widen (vectorize) values we place them in the map. If the values
225 /// are not within the map, they have to be loop invariant, so we simply
226 /// broadcast them into a vector.
227 VectorParts &getVectorValue(Value *V);
229 /// Generate a shuffle sequence that will reverse the vector Vec.
230 Value *reverseVector(Value *Vec);
232 /// This is a helper class that holds the vectorizer state. It maps scalar
233 /// instructions to vector instructions. When the code is 'unrolled' then
234 /// then a single scalar value is mapped to multiple vector parts. The parts
235 /// are stored in the VectorPart type.
237 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
239 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
241 /// \return True if 'Key' is saved in the Value Map.
242 bool has(Value *Key) const { return MapStorage.count(Key); }
244 /// Initializes a new entry in the map. Sets all of the vector parts to the
245 /// save value in 'Val'.
246 /// \return A reference to a vector with splat values.
247 VectorParts &splat(Value *Key, Value *Val) {
248 VectorParts &Entry = MapStorage[Key];
249 Entry.assign(UF, Val);
253 ///\return A reference to the value that is stored at 'Key'.
254 VectorParts &get(Value *Key) {
255 VectorParts &Entry = MapStorage[Key];
258 assert(Entry.size() == UF);
263 /// The unroll factor. Each entry in the map stores this number of vector
267 /// Map storage. We use std::map and not DenseMap because insertions to a
268 /// dense map invalidates its iterators.
269 std::map<Value *, VectorParts> MapStorage;
272 /// The original loop.
274 /// Scev analysis to use.
282 /// Target Library Info.
283 const TargetLibraryInfo *TLI;
285 /// The vectorization SIMD factor to use. Each vector will have this many
288 /// The vectorization unroll factor to use. Each scalar is vectorized to this
289 /// many different vector instructions.
292 /// The builder that we use
295 // --- Vectorization state ---
297 /// The vector-loop preheader.
298 BasicBlock *LoopVectorPreHeader;
299 /// The scalar-loop preheader.
300 BasicBlock *LoopScalarPreHeader;
301 /// Middle Block between the vector and the scalar.
302 BasicBlock *LoopMiddleBlock;
303 ///The ExitBlock of the scalar loop.
304 BasicBlock *LoopExitBlock;
305 ///The vector loop body.
306 BasicBlock *LoopVectorBody;
307 ///The scalar loop body.
308 BasicBlock *LoopScalarBody;
309 /// A list of all bypass blocks. The first block is the entry of the loop.
310 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
312 /// The new Induction variable which was added to the new block.
314 /// The induction variable of the old basic block.
315 PHINode *OldInduction;
316 /// Holds the extended (to the widest induction type) start index.
318 /// Maps scalars to widened vectors.
322 /// \brief Check if conditionally executed loads are hoistable.
324 /// This class has two functions: isHoistableLoad and canHoistAllLoads.
325 /// isHoistableLoad should be called on all load instructions that are executed
326 /// conditionally. After all conditional loads are processed, the client should
327 /// call canHoistAllLoads to determine if all of the conditional executed loads
328 /// have an unconditional memory access to the same memory address in the loop.
330 typedef SmallPtrSet<Value *, 8> MemorySet;
334 MemorySet CondLoadAddrSet;
337 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
339 /// \brief Check if the instruction is a load with a identifiable address.
340 bool isHoistableLoad(Instruction *L);
342 /// \brief Check if all of the conditional loads are hoistable because there
343 /// exists an unconditional memory access to the same address in the loop.
344 bool canHoistAllLoads();
347 bool LoadHoisting::isHoistableLoad(Instruction *L) {
348 LoadInst *LI = dyn_cast<LoadInst>(L);
352 CondLoadAddrSet.insert(LI->getPointerOperand());
356 static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
357 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
358 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
359 Set.insert(LI->getPointerOperand());
360 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
361 Set.insert(SI->getPointerOperand());
365 bool LoadHoisting::canHoistAllLoads() {
366 // No conditional loads.
367 if (CondLoadAddrSet.empty())
370 MemorySet UncondMemAccesses;
371 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
372 BasicBlock *LoopLatch = TheLoop->getLoopLatch();
374 // Iterate over the unconditional blocks and collect memory access addresses.
375 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
376 BasicBlock *BB = LoopBlocks[i];
378 // Ignore conditional blocks.
379 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
382 addMemAccesses(BB, UncondMemAccesses);
385 // And make sure there is a matching unconditional access for every
387 for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
388 ME = CondLoadAddrSet.end(); MI != ME; ++MI)
389 if (!UncondMemAccesses.count(*MI))
395 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
396 /// to what vectorization factor.
397 /// This class does not look at the profitability of vectorization, only the
398 /// legality. This class has two main kinds of checks:
399 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
400 /// will change the order of memory accesses in a way that will change the
401 /// correctness of the program.
402 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
403 /// checks for a number of different conditions, such as the availability of a
404 /// single induction variable, that all types are supported and vectorize-able,
405 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
406 /// This class is also used by InnerLoopVectorizer for identifying
407 /// induction variable and the different reduction variables.
408 class LoopVectorizationLegality {
410 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
411 DominatorTree *DT, TargetTransformInfo* TTI,
412 AliasAnalysis *AA, TargetLibraryInfo *TLI)
413 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
414 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
415 LoadSpeculation(L, DT) {}
417 /// This enum represents the kinds of reductions that we support.
419 RK_NoReduction, ///< Not a reduction.
420 RK_IntegerAdd, ///< Sum of integers.
421 RK_IntegerMult, ///< Product of integers.
422 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
423 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
424 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
425 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
426 RK_FloatAdd, ///< Sum of floats.
427 RK_FloatMult, ///< Product of floats.
428 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
431 /// This enum represents the kinds of inductions that we support.
433 IK_NoInduction, ///< Not an induction variable.
434 IK_IntInduction, ///< Integer induction variable. Step = 1.
435 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
436 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
437 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
440 // This enum represents the kind of minmax reduction.
441 enum MinMaxReductionKind {
451 /// This POD struct holds information about reduction variables.
452 struct ReductionDescriptor {
453 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
454 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
456 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
457 MinMaxReductionKind MK)
458 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
460 // The starting value of the reduction.
461 // It does not have to be zero!
462 TrackingVH<Value> StartValue;
463 // The instruction who's value is used outside the loop.
464 Instruction *LoopExitInstr;
465 // The kind of the reduction.
467 // If this a min/max reduction the kind of reduction.
468 MinMaxReductionKind MinMaxKind;
471 /// This POD struct holds information about a potential reduction operation.
472 struct ReductionInstDesc {
473 ReductionInstDesc(bool IsRedux, Instruction *I) :
474 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
476 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
477 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
479 // Is this instruction a reduction candidate.
481 // The last instruction in a min/max pattern (select of the select(icmp())
482 // pattern), or the current reduction instruction otherwise.
483 Instruction *PatternLastInst;
484 // If this is a min/max pattern the comparison predicate.
485 MinMaxReductionKind MinMaxKind;
488 // This POD struct holds information about the memory runtime legality
489 // check that a group of pointers do not overlap.
490 struct RuntimePointerCheck {
491 RuntimePointerCheck() : Need(false) {}
493 /// Reset the state of the pointer runtime information.
501 /// Insert a pointer and calculate the start and end SCEVs.
502 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr);
504 /// This flag indicates if we need to add the runtime check.
506 /// Holds the pointers that we need to check.
507 SmallVector<TrackingVH<Value>, 2> Pointers;
508 /// Holds the pointer value at the beginning of the loop.
509 SmallVector<const SCEV*, 2> Starts;
510 /// Holds the pointer value at the end of the loop.
511 SmallVector<const SCEV*, 2> Ends;
512 /// Holds the information if this pointer is used for writing to memory.
513 SmallVector<bool, 2> IsWritePtr;
516 /// A POD for saving information about induction variables.
517 struct InductionInfo {
518 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
519 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
521 TrackingVH<Value> StartValue;
526 /// ReductionList contains the reduction descriptors for all
527 /// of the reductions that were found in the loop.
528 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
530 /// InductionList saves induction variables and maps them to the
531 /// induction descriptor.
532 typedef MapVector<PHINode*, InductionInfo> InductionList;
534 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
535 /// respective Store/Load instruction(s) to calculate aliasing.
536 typedef MapVector<Value*, Instruction* > AliasMap;
537 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
539 /// Returns true if it is legal to vectorize this loop.
540 /// This does not mean that it is profitable to vectorize this
541 /// loop, only that it is legal to do so.
544 /// Returns the Induction variable.
545 PHINode *getInduction() { return Induction; }
547 /// Returns the reduction variables found in the loop.
548 ReductionList *getReductionVars() { return &Reductions; }
550 /// Returns the induction variables found in the loop.
551 InductionList *getInductionVars() { return &Inductions; }
553 /// Returns the widest induction type.
554 Type *getWidestInductionType() { return WidestIndTy; }
556 /// Returns True if V is an induction variable in this loop.
557 bool isInductionVariable(const Value *V);
559 /// Return true if the block BB needs to be predicated in order for the loop
560 /// to be vectorized.
561 bool blockNeedsPredication(BasicBlock *BB);
563 /// Check if this pointer is consecutive when vectorizing. This happens
564 /// when the last index of the GEP is the induction variable, or that the
565 /// pointer itself is an induction variable.
566 /// This check allows us to vectorize A[idx] into a wide load/store.
568 /// 0 - Stride is unknown or non consecutive.
569 /// 1 - Address is consecutive.
570 /// -1 - Address is consecutive, and decreasing.
571 int isConsecutivePtr(Value *Ptr);
573 /// Returns true if the value V is uniform within the loop.
574 bool isUniform(Value *V);
576 /// Returns true if this instruction will remain scalar after vectorization.
577 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
579 /// Returns the information that we collected about runtime memory check.
580 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
582 /// This function returns the identity element (or neutral element) for
584 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
586 /// Check if a single basic block loop is vectorizable.
587 /// At this point we know that this is a loop with a constant trip count
588 /// and we only need to check individual instructions.
589 bool canVectorizeInstrs();
591 /// When we vectorize loops we may change the order in which
592 /// we read and write from memory. This method checks if it is
593 /// legal to vectorize the code, considering only memory constrains.
594 /// Returns true if the loop is vectorizable
595 bool canVectorizeMemory();
597 /// Return true if we can vectorize this loop using the IF-conversion
599 bool canVectorizeWithIfConvert();
601 /// Collect the variables that need to stay uniform after vectorization.
602 void collectLoopUniforms();
604 /// Return true if all of the instructions in the block can be speculatively
606 bool blockCanBePredicated(BasicBlock *BB);
608 /// Returns True, if 'Phi' is the kind of reduction variable for type
609 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
610 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
611 /// Returns a struct describing if the instruction 'I' can be a reduction
612 /// variable of type 'Kind'. If the reduction is a min/max pattern of
613 /// select(icmp()) this function advances the instruction pointer 'I' from the
614 /// compare instruction to the select instruction and stores this pointer in
615 /// 'PatternLastInst' member of the returned struct.
616 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
617 ReductionInstDesc &Desc);
618 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
619 /// pattern corresponding to a min(X, Y) or max(X, Y).
620 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
621 ReductionInstDesc &Prev);
622 /// Returns the induction kind of Phi. This function may return NoInduction
623 /// if the PHI is not an induction variable.
624 InductionKind isInductionVariable(PHINode *Phi);
625 /// Return true if can compute the address bounds of Ptr within the loop.
626 bool hasComputableBounds(Value *Ptr);
627 /// Return true if there is the chance of write reorder.
628 bool hasPossibleGlobalWriteReorder(Value *Object,
630 AliasMultiMap &WriteObjects,
631 unsigned MaxByteWidth);
632 /// Return the AA location for a load or a store.
633 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
636 /// The loop that we evaluate.
640 /// DataLayout analysis.
645 TargetTransformInfo *TTI;
648 /// Target Library Info.
649 TargetLibraryInfo *TLI;
651 // --- vectorization state --- //
653 /// Holds the integer induction variable. This is the counter of the
656 /// Holds the reduction variables.
657 ReductionList Reductions;
658 /// Holds all of the induction variables that we found in the loop.
659 /// Notice that inductions don't need to start at zero and that induction
660 /// variables can be pointers.
661 InductionList Inductions;
662 /// Holds the widest induction type encountered.
665 /// Allowed outside users. This holds the reduction
666 /// vars which can be accessed from outside the loop.
667 SmallPtrSet<Value*, 4> AllowedExit;
668 /// This set holds the variables which are known to be uniform after
670 SmallPtrSet<Instruction*, 4> Uniforms;
671 /// We need to check that all of the pointers in this list are disjoint
673 RuntimePointerCheck PtrRtCheck;
674 /// Can we assume the absence of NaNs.
675 bool HasFunNoNaNAttr;
677 /// Utility to determine whether loads can be speculated.
678 LoadHoisting LoadSpeculation;
681 /// LoopVectorizationCostModel - estimates the expected speedups due to
683 /// In many cases vectorization is not profitable. This can happen because of
684 /// a number of reasons. In this class we mainly attempt to predict the
685 /// expected speedup/slowdowns due to the supported instruction set. We use the
686 /// TargetTransformInfo to query the different backends for the cost of
687 /// different operations.
688 class LoopVectorizationCostModel {
690 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
691 LoopVectorizationLegality *Legal,
692 const TargetTransformInfo &TTI,
693 DataLayout *DL, const TargetLibraryInfo *TLI)
694 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
696 /// Information about vectorization costs
697 struct VectorizationFactor {
698 unsigned Width; // Vector width with best cost
699 unsigned Cost; // Cost of the loop with that width
701 /// \return The most profitable vectorization factor and the cost of that VF.
702 /// This method checks every power of two up to VF. If UserVF is not ZERO
703 /// then this vectorization factor will be selected if vectorization is
705 VectorizationFactor selectVectorizationFactor(bool OptForSize,
708 /// \return The size (in bits) of the widest type in the code that
709 /// needs to be vectorized. We ignore values that remain scalar such as
710 /// 64 bit loop indices.
711 unsigned getWidestType();
713 /// \return The most profitable unroll factor.
714 /// If UserUF is non-zero then this method finds the best unroll-factor
715 /// based on register pressure and other parameters.
716 /// VF and LoopCost are the selected vectorization factor and the cost of the
718 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
721 /// \brief A struct that represents some properties of the register usage
723 struct RegisterUsage {
724 /// Holds the number of loop invariant values that are used in the loop.
725 unsigned LoopInvariantRegs;
726 /// Holds the maximum number of concurrent live intervals in the loop.
727 unsigned MaxLocalUsers;
728 /// Holds the number of instructions in the loop.
729 unsigned NumInstructions;
732 /// \return information about the register usage of the loop.
733 RegisterUsage calculateRegisterUsage();
736 /// Returns the expected execution cost. The unit of the cost does
737 /// not matter because we use the 'cost' units to compare different
738 /// vector widths. The cost that is returned is *not* normalized by
739 /// the factor width.
740 unsigned expectedCost(unsigned VF);
742 /// Returns the execution time cost of an instruction for a given vector
743 /// width. Vector width of one means scalar.
744 unsigned getInstructionCost(Instruction *I, unsigned VF);
746 /// A helper function for converting Scalar types to vector types.
747 /// If the incoming type is void, we return void. If the VF is 1, we return
749 static Type* ToVectorTy(Type *Scalar, unsigned VF);
751 /// Returns whether the instruction is a load or store and will be a emitted
752 /// as a vector operation.
753 bool isConsecutiveLoadOrStore(Instruction *I);
755 /// The loop that we evaluate.
759 /// Loop Info analysis.
761 /// Vectorization legality.
762 LoopVectorizationLegality *Legal;
763 /// Vector target information.
764 const TargetTransformInfo &TTI;
765 /// Target data layout information.
767 /// Target Library Info.
768 const TargetLibraryInfo *TLI;
771 /// Utility class for getting and setting loop vectorizer hints in the form
772 /// of loop metadata.
773 struct LoopVectorizeHints {
774 /// Vectorization width.
776 /// Vectorization unroll factor.
779 LoopVectorizeHints(const Loop *L)
780 : Width(VectorizationFactor)
781 , Unroll(VectorizationUnroll)
782 , LoopID(L->getLoopID()) {
784 // The command line options override any loop metadata except for when
785 // width == 1 which is used to indicate the loop is already vectorized.
786 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
787 Width = VectorizationFactor;
788 if (VectorizationUnroll.getNumOccurrences() > 0)
789 Unroll = VectorizationUnroll;
792 /// Return the loop vectorizer metadata prefix.
793 static StringRef Prefix() { return "llvm.vectorizer."; }
795 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
796 SmallVector<Value*, 2> Vals;
797 Vals.push_back(MDString::get(Context, Name));
798 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
799 return MDNode::get(Context, Vals);
802 /// Mark the loop L as already vectorized by setting the width to 1.
803 void setAlreadyVectorized(Loop *L) {
804 LLVMContext &Context = L->getHeader()->getContext();
808 // Create a new loop id with one more operand for the already_vectorized
809 // hint. If the loop already has a loop id then copy the existing operands.
810 SmallVector<Value*, 4> Vals(1);
812 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
813 Vals.push_back(LoopID->getOperand(i));
815 Twine Name = Prefix() + "width";
816 Vals.push_back(createHint(Context, Name.str(), Width));
818 MDNode *NewLoopID = MDNode::get(Context, Vals);
819 // Set operand 0 to refer to the loop id itself.
820 NewLoopID->replaceOperandWith(0, NewLoopID);
822 L->setLoopID(NewLoopID);
824 LoopID->replaceAllUsesWith(NewLoopID);
832 /// Find hints specified in the loop metadata.
833 void getHints(const Loop *L) {
837 // First operand should refer to the loop id itself.
838 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
839 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
841 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
842 const MDString *S = 0;
843 SmallVector<Value*, 4> Args;
845 // The expected hint is either a MDString or a MDNode with the first
846 // operand a MDString.
847 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
848 if (!MD || MD->getNumOperands() == 0)
850 S = dyn_cast<MDString>(MD->getOperand(0));
851 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
852 Args.push_back(MD->getOperand(i));
854 S = dyn_cast<MDString>(LoopID->getOperand(i));
855 assert(Args.size() == 0 && "too many arguments for MDString");
861 // Check if the hint starts with the vectorizer prefix.
862 StringRef Hint = S->getString();
863 if (!Hint.startswith(Prefix()))
865 // Remove the prefix.
866 Hint = Hint.substr(Prefix().size(), StringRef::npos);
868 if (Args.size() == 1)
869 getHint(Hint, Args[0]);
873 // Check string hint with one operand.
874 void getHint(StringRef Hint, Value *Arg) {
875 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
877 unsigned Val = C->getZExtValue();
879 if (Hint == "width") {
880 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
881 "Invalid width metadata");
883 } else if (Hint == "unroll") {
884 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
885 "Invalid unroll metadata");
888 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
892 /// The LoopVectorize Pass.
893 struct LoopVectorize : public LoopPass {
894 /// Pass identification, replacement for typeid
897 explicit LoopVectorize() : LoopPass(ID) {
898 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
904 TargetTransformInfo *TTI;
907 TargetLibraryInfo *TLI;
909 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
910 // We only vectorize innermost loops.
914 SE = &getAnalysis<ScalarEvolution>();
915 DL = getAnalysisIfAvailable<DataLayout>();
916 LI = &getAnalysis<LoopInfo>();
917 TTI = &getAnalysis<TargetTransformInfo>();
918 DT = &getAnalysis<DominatorTree>();
919 AA = getAnalysisIfAvailable<AliasAnalysis>();
920 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
923 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
927 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
928 L->getHeader()->getParent()->getName() << "\"\n");
930 LoopVectorizeHints Hints(L);
932 if (Hints.Width == 1) {
933 DEBUG(dbgs() << "LV: Not vectorizing.\n");
937 // Check if it is legal to vectorize the loop.
938 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
939 if (!LVL.canVectorize()) {
940 DEBUG(dbgs() << "LV: Not vectorizing.\n");
944 // Use the cost model.
945 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
947 // Check the function attributes to find out if this function should be
948 // optimized for size.
949 Function *F = L->getHeader()->getParent();
950 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
951 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
952 unsigned FnIndex = AttributeSet::FunctionIndex;
953 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
954 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
957 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
958 "attribute is used.\n");
962 // Select the optimal vectorization factor.
963 LoopVectorizationCostModel::VectorizationFactor VF;
964 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
965 // Select the unroll factor.
966 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
970 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
974 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
975 F->getParent()->getModuleIdentifier()<<"\n");
976 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
978 // If we decided that it is *legal* to vectorize the loop then do it.
979 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
982 // Mark the loop as already vectorized to avoid vectorizing again.
983 Hints.setAlreadyVectorized(L);
985 DEBUG(verifyFunction(*L->getHeader()->getParent()));
989 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
990 LoopPass::getAnalysisUsage(AU);
991 AU.addRequiredID(LoopSimplifyID);
992 AU.addRequiredID(LCSSAID);
993 AU.addRequired<DominatorTree>();
994 AU.addRequired<LoopInfo>();
995 AU.addRequired<ScalarEvolution>();
996 AU.addRequired<TargetTransformInfo>();
997 AU.addPreserved<LoopInfo>();
998 AU.addPreserved<DominatorTree>();
1003 } // end anonymous namespace
1005 //===----------------------------------------------------------------------===//
1006 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1007 // LoopVectorizationCostModel.
1008 //===----------------------------------------------------------------------===//
1011 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1012 Loop *Lp, Value *Ptr,
1014 const SCEV *Sc = SE->getSCEV(Ptr);
1015 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1016 assert(AR && "Invalid addrec expression");
1017 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
1018 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1019 Pointers.push_back(Ptr);
1020 Starts.push_back(AR->getStart());
1021 Ends.push_back(ScEnd);
1022 IsWritePtr.push_back(WritePtr);
1025 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1026 // Save the current insertion location.
1027 Instruction *Loc = Builder.GetInsertPoint();
1029 // We need to place the broadcast of invariant variables outside the loop.
1030 Instruction *Instr = dyn_cast<Instruction>(V);
1031 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1032 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1034 // Place the code for broadcasting invariant variables in the new preheader.
1036 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1038 // Broadcast the scalar into all locations in the vector.
1039 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1041 // Restore the builder insertion point.
1043 Builder.SetInsertPoint(Loc);
1048 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1050 assert(Val->getType()->isVectorTy() && "Must be a vector");
1051 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1052 "Elem must be an integer");
1053 // Create the types.
1054 Type *ITy = Val->getType()->getScalarType();
1055 VectorType *Ty = cast<VectorType>(Val->getType());
1056 int VLen = Ty->getNumElements();
1057 SmallVector<Constant*, 8> Indices;
1059 // Create a vector of consecutive numbers from zero to VF.
1060 for (int i = 0; i < VLen; ++i) {
1061 int64_t Idx = Negate ? (-i) : i;
1062 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1065 // Add the consecutive indices to the vector value.
1066 Constant *Cv = ConstantVector::get(Indices);
1067 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1068 return Builder.CreateAdd(Val, Cv, "induction");
1071 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1072 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1073 // Make sure that the pointer does not point to structs.
1074 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1077 // If this value is a pointer induction variable we know it is consecutive.
1078 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1079 if (Phi && Inductions.count(Phi)) {
1080 InductionInfo II = Inductions[Phi];
1081 if (IK_PtrInduction == II.IK)
1083 else if (IK_ReversePtrInduction == II.IK)
1087 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1091 unsigned NumOperands = Gep->getNumOperands();
1092 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1094 Value *GpPtr = Gep->getPointerOperand();
1095 // If this GEP value is a consecutive pointer induction variable and all of
1096 // the indices are constant then we know it is consecutive. We can
1097 Phi = dyn_cast<PHINode>(GpPtr);
1098 if (Phi && Inductions.count(Phi)) {
1100 // Make sure that the pointer does not point to structs.
1101 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1102 if (GepPtrType->getElementType()->isAggregateType())
1105 // Make sure that all of the index operands are loop invariant.
1106 for (unsigned i = 1; i < NumOperands; ++i)
1107 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1110 InductionInfo II = Inductions[Phi];
1111 if (IK_PtrInduction == II.IK)
1113 else if (IK_ReversePtrInduction == II.IK)
1117 // Check that all of the gep indices are uniform except for the last.
1118 for (unsigned i = 0; i < NumOperands - 1; ++i)
1119 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1122 // We can emit wide load/stores only if the last index is the induction
1124 const SCEV *Last = SE->getSCEV(LastIndex);
1125 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1126 const SCEV *Step = AR->getStepRecurrence(*SE);
1128 // The memory is consecutive because the last index is consecutive
1129 // and all other indices are loop invariant.
1132 if (Step->isAllOnesValue())
1139 bool LoopVectorizationLegality::isUniform(Value *V) {
1140 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1143 InnerLoopVectorizer::VectorParts&
1144 InnerLoopVectorizer::getVectorValue(Value *V) {
1145 assert(V != Induction && "The new induction variable should not be used.");
1146 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1148 // If we have this scalar in the map, return it.
1149 if (WidenMap.has(V))
1150 return WidenMap.get(V);
1152 // If this scalar is unknown, assume that it is a constant or that it is
1153 // loop invariant. Broadcast V and save the value for future uses.
1154 Value *B = getBroadcastInstrs(V);
1155 return WidenMap.splat(V, B);
1158 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1159 assert(Vec->getType()->isVectorTy() && "Invalid type");
1160 SmallVector<Constant*, 8> ShuffleMask;
1161 for (unsigned i = 0; i < VF; ++i)
1162 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1164 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1165 ConstantVector::get(ShuffleMask),
1170 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1171 LoopVectorizationLegality *Legal) {
1172 // Attempt to issue a wide load.
1173 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1174 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1176 assert((LI || SI) && "Invalid Load/Store instruction");
1178 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1179 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1180 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1181 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1183 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1184 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1186 if (ScalarAllocatedSize != VectorElementSize)
1187 return scalarizeInstruction(Instr);
1189 // If the pointer is loop invariant or if it is non consecutive,
1190 // scalarize the load.
1191 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1192 bool Reverse = ConsecutiveStride < 0;
1193 bool UniformLoad = LI && Legal->isUniform(Ptr);
1194 if (!ConsecutiveStride || UniformLoad)
1195 return scalarizeInstruction(Instr);
1197 Constant *Zero = Builder.getInt32(0);
1198 VectorParts &Entry = WidenMap.get(Instr);
1200 // Handle consecutive loads/stores.
1201 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1202 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1203 Value *PtrOperand = Gep->getPointerOperand();
1204 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1205 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1207 // Create the new GEP with the new induction variable.
1208 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1209 Gep2->setOperand(0, FirstBasePtr);
1210 Gep2->setName("gep.indvar.base");
1211 Ptr = Builder.Insert(Gep2);
1213 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1214 OrigLoop) && "Base ptr must be invariant");
1216 // The last index does not have to be the induction. It can be
1217 // consecutive and be a function of the index. For example A[I+1];
1218 unsigned NumOperands = Gep->getNumOperands();
1220 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1221 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1222 Value *LastIndex = GEPParts[0];
1223 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1225 // Create the new GEP with the new induction variable.
1226 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1227 Gep2->setOperand(NumOperands - 1, LastIndex);
1228 Gep2->setName("gep.indvar.idx");
1229 Ptr = Builder.Insert(Gep2);
1231 // Use the induction element ptr.
1232 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1233 VectorParts &PtrVal = getVectorValue(Ptr);
1234 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1239 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1240 "We do not allow storing to uniform addresses");
1242 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1243 for (unsigned Part = 0; Part < UF; ++Part) {
1244 // Calculate the pointer for the specific unroll-part.
1245 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1248 // If we store to reverse consecutive memory locations then we need
1249 // to reverse the order of elements in the stored value.
1250 StoredVal[Part] = reverseVector(StoredVal[Part]);
1251 // If the address is consecutive but reversed, then the
1252 // wide store needs to start at the last vector element.
1253 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1254 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1257 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1258 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1262 for (unsigned Part = 0; Part < UF; ++Part) {
1263 // Calculate the pointer for the specific unroll-part.
1264 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1267 // If the address is consecutive but reversed, then the
1268 // wide store needs to start at the last vector element.
1269 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1270 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1273 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1274 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1275 cast<LoadInst>(LI)->setAlignment(Alignment);
1276 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1280 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1281 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1282 // Holds vector parameters or scalars, in case of uniform vals.
1283 SmallVector<VectorParts, 4> Params;
1285 // Find all of the vectorized parameters.
1286 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1287 Value *SrcOp = Instr->getOperand(op);
1289 // If we are accessing the old induction variable, use the new one.
1290 if (SrcOp == OldInduction) {
1291 Params.push_back(getVectorValue(SrcOp));
1295 // Try using previously calculated values.
1296 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1298 // If the src is an instruction that appeared earlier in the basic block
1299 // then it should already be vectorized.
1300 if (SrcInst && OrigLoop->contains(SrcInst)) {
1301 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1302 // The parameter is a vector value from earlier.
1303 Params.push_back(WidenMap.get(SrcInst));
1305 // The parameter is a scalar from outside the loop. Maybe even a constant.
1306 VectorParts Scalars;
1307 Scalars.append(UF, SrcOp);
1308 Params.push_back(Scalars);
1312 assert(Params.size() == Instr->getNumOperands() &&
1313 "Invalid number of operands");
1315 // Does this instruction return a value ?
1316 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1318 Value *UndefVec = IsVoidRetTy ? 0 :
1319 UndefValue::get(VectorType::get(Instr->getType(), VF));
1320 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1321 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1323 // For each vector unroll 'part':
1324 for (unsigned Part = 0; Part < UF; ++Part) {
1325 // For each scalar that we create:
1326 for (unsigned Width = 0; Width < VF; ++Width) {
1327 Instruction *Cloned = Instr->clone();
1329 Cloned->setName(Instr->getName() + ".cloned");
1330 // Replace the operands of the cloned instrucions with extracted scalars.
1331 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1332 Value *Op = Params[op][Part];
1333 // Param is a vector. Need to extract the right lane.
1334 if (Op->getType()->isVectorTy())
1335 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1336 Cloned->setOperand(op, Op);
1339 // Place the cloned scalar in the new loop.
1340 Builder.Insert(Cloned);
1342 // If the original scalar returns a value we need to place it in a vector
1343 // so that future users will be able to use it.
1345 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1346 Builder.getInt32(Width));
1352 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1354 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1355 Legal->getRuntimePointerCheck();
1357 if (!PtrRtCheck->Need)
1360 Instruction *MemoryRuntimeCheck = 0;
1361 unsigned NumPointers = PtrRtCheck->Pointers.size();
1362 SmallVector<Value* , 2> Starts;
1363 SmallVector<Value* , 2> Ends;
1365 SCEVExpander Exp(*SE, "induction");
1367 // Use this type for pointer arithmetic.
1368 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1370 for (unsigned i = 0; i < NumPointers; ++i) {
1371 Value *Ptr = PtrRtCheck->Pointers[i];
1372 const SCEV *Sc = SE->getSCEV(Ptr);
1374 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1375 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1377 Starts.push_back(Ptr);
1378 Ends.push_back(Ptr);
1380 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1382 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1383 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1384 Starts.push_back(Start);
1385 Ends.push_back(End);
1389 IRBuilder<> ChkBuilder(Loc);
1391 for (unsigned i = 0; i < NumPointers; ++i) {
1392 for (unsigned j = i+1; j < NumPointers; ++j) {
1393 // No need to check if two readonly pointers intersect.
1394 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1397 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1398 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1399 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1400 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1402 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1403 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1404 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1405 if (MemoryRuntimeCheck)
1406 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1409 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1413 return MemoryRuntimeCheck;
1417 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1419 In this function we generate a new loop. The new loop will contain
1420 the vectorized instructions while the old loop will continue to run the
1423 [ ] <-- vector loop bypass (may consist of multiple blocks).
1426 | [ ] <-- vector pre header.
1430 | [ ]_| <-- vector loop.
1433 >[ ] <--- middle-block.
1436 | [ ] <--- new preheader.
1440 | [ ]_| <-- old scalar loop to handle remainder.
1443 >[ ] <-- exit block.
1447 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1448 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1449 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1450 assert(ExitBlock && "Must have an exit block");
1452 // Some loops have a single integer induction variable, while other loops
1453 // don't. One example is c++ iterators that often have multiple pointer
1454 // induction variables. In the code below we also support a case where we
1455 // don't have a single induction variable.
1456 OldInduction = Legal->getInduction();
1457 Type *IdxTy = Legal->getWidestInductionType();
1459 // Find the loop boundaries.
1460 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1461 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1463 // Get the total trip count from the count by adding 1.
1464 ExitCount = SE->getAddExpr(ExitCount,
1465 SE->getConstant(ExitCount->getType(), 1));
1467 // Expand the trip count and place the new instructions in the preheader.
1468 // Notice that the pre-header does not change, only the loop body.
1469 SCEVExpander Exp(*SE, "induction");
1471 // Count holds the overall loop count (N).
1472 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1473 BypassBlock->getTerminator());
1475 // The loop index does not have to start at Zero. Find the original start
1476 // value from the induction PHI node. If we don't have an induction variable
1477 // then we know that it starts at zero.
1478 Builder.SetInsertPoint(BypassBlock->getTerminator());
1479 Value *StartIdx = ExtendedIdx = OldInduction ?
1480 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1482 ConstantInt::get(IdxTy, 0);
1484 assert(BypassBlock && "Invalid loop structure");
1485 LoopBypassBlocks.push_back(BypassBlock);
1487 // Split the single block loop into the two loop structure described above.
1488 BasicBlock *VectorPH =
1489 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1490 BasicBlock *VecBody =
1491 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1492 BasicBlock *MiddleBlock =
1493 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1494 BasicBlock *ScalarPH =
1495 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1497 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1499 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1501 // Generate the induction variable.
1502 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1503 // The loop step is equal to the vectorization factor (num of SIMD elements)
1504 // times the unroll factor (num of SIMD instructions).
1505 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1507 // This is the IR builder that we use to add all of the logic for bypassing
1508 // the new vector loop.
1509 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1511 // We may need to extend the index in case there is a type mismatch.
1512 // We know that the count starts at zero and does not overflow.
1513 if (Count->getType() != IdxTy) {
1514 // The exit count can be of pointer type. Convert it to the correct
1516 if (ExitCount->getType()->isPointerTy())
1517 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1519 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1522 // Add the start index to the loop count to get the new end index.
1523 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1525 // Now we need to generate the expression for N - (N % VF), which is
1526 // the part that the vectorized body will execute.
1527 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1528 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1529 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1530 "end.idx.rnd.down");
1532 // Now, compare the new count to zero. If it is zero skip the vector loop and
1533 // jump to the scalar loop.
1534 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1537 BasicBlock *LastBypassBlock = BypassBlock;
1539 // Generate the code that checks in runtime if arrays overlap. We put the
1540 // checks into a separate block to make the more common case of few elements
1542 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1543 BypassBlock->getTerminator());
1544 if (MemRuntimeCheck) {
1545 // Create a new block containing the memory check.
1546 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1548 LoopBypassBlocks.push_back(CheckBlock);
1550 // Replace the branch into the memory check block with a conditional branch
1551 // for the "few elements case".
1552 Instruction *OldTerm = BypassBlock->getTerminator();
1553 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1554 OldTerm->eraseFromParent();
1556 Cmp = MemRuntimeCheck;
1557 LastBypassBlock = CheckBlock;
1560 LastBypassBlock->getTerminator()->eraseFromParent();
1561 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1564 // We are going to resume the execution of the scalar loop.
1565 // Go over all of the induction variables that we found and fix the
1566 // PHIs that are left in the scalar version of the loop.
1567 // The starting values of PHI nodes depend on the counter of the last
1568 // iteration in the vectorized loop.
1569 // If we come from a bypass edge then we need to start from the original
1572 // This variable saves the new starting index for the scalar loop.
1573 PHINode *ResumeIndex = 0;
1574 LoopVectorizationLegality::InductionList::iterator I, E;
1575 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1576 // Set builder to point to last bypass block.
1577 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1578 for (I = List->begin(), E = List->end(); I != E; ++I) {
1579 PHINode *OrigPhi = I->first;
1580 LoopVectorizationLegality::InductionInfo II = I->second;
1582 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1583 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1584 MiddleBlock->getTerminator());
1585 // We might have extended the type of the induction variable but we need a
1586 // truncated version for the scalar loop.
1587 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1588 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1589 MiddleBlock->getTerminator()) : 0;
1591 Value *EndValue = 0;
1593 case LoopVectorizationLegality::IK_NoInduction:
1594 llvm_unreachable("Unknown induction");
1595 case LoopVectorizationLegality::IK_IntInduction: {
1596 // Handle the integer induction counter.
1597 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1599 // We have the canonical induction variable.
1600 if (OrigPhi == OldInduction) {
1601 // Create a truncated version of the resume value for the scalar loop,
1602 // we might have promoted the type to a larger width.
1604 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1605 // The new PHI merges the original incoming value, in case of a bypass,
1606 // or the value at the end of the vectorized loop.
1607 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1608 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1609 TruncResumeVal->addIncoming(EndValue, VecBody);
1611 // We know what the end value is.
1612 EndValue = IdxEndRoundDown;
1613 // We also know which PHI node holds it.
1614 ResumeIndex = ResumeVal;
1618 // Not the canonical induction variable - add the vector loop count to the
1620 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1621 II.StartValue->getType(),
1623 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1626 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1627 // Convert the CountRoundDown variable to the PHI size.
1628 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1629 II.StartValue->getType(),
1631 // Handle reverse integer induction counter.
1632 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1635 case LoopVectorizationLegality::IK_PtrInduction: {
1636 // For pointer induction variables, calculate the offset using
1638 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1642 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1643 // The value at the end of the loop for the reverse pointer is calculated
1644 // by creating a GEP with a negative index starting from the start value.
1645 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1646 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1648 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1654 // The new PHI merges the original incoming value, in case of a bypass,
1655 // or the value at the end of the vectorized loop.
1656 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1657 if (OrigPhi == OldInduction)
1658 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1660 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1662 ResumeVal->addIncoming(EndValue, VecBody);
1664 // Fix the scalar body counter (PHI node).
1665 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1666 // The old inductions phi node in the scalar body needs the truncated value.
1667 if (OrigPhi == OldInduction)
1668 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1670 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1673 // If we are generating a new induction variable then we also need to
1674 // generate the code that calculates the exit value. This value is not
1675 // simply the end of the counter because we may skip the vectorized body
1676 // in case of a runtime check.
1678 assert(!ResumeIndex && "Unexpected resume value found");
1679 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1680 MiddleBlock->getTerminator());
1681 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1682 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1683 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1686 // Make sure that we found the index where scalar loop needs to continue.
1687 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1688 "Invalid resume Index");
1690 // Add a check in the middle block to see if we have completed
1691 // all of the iterations in the first vector loop.
1692 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1693 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1694 ResumeIndex, "cmp.n",
1695 MiddleBlock->getTerminator());
1697 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1698 // Remove the old terminator.
1699 MiddleBlock->getTerminator()->eraseFromParent();
1701 // Create i+1 and fill the PHINode.
1702 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1703 Induction->addIncoming(StartIdx, VectorPH);
1704 Induction->addIncoming(NextIdx, VecBody);
1705 // Create the compare.
1706 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1707 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1709 // Now we have two terminators. Remove the old one from the block.
1710 VecBody->getTerminator()->eraseFromParent();
1712 // Get ready to start creating new instructions into the vectorized body.
1713 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1715 // Create and register the new vector loop.
1716 Loop* Lp = new Loop();
1717 Loop *ParentLoop = OrigLoop->getParentLoop();
1719 // Insert the new loop into the loop nest and register the new basic blocks.
1721 ParentLoop->addChildLoop(Lp);
1722 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1723 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1724 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1725 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1726 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1728 LI->addTopLevelLoop(Lp);
1731 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1734 LoopVectorPreHeader = VectorPH;
1735 LoopScalarPreHeader = ScalarPH;
1736 LoopMiddleBlock = MiddleBlock;
1737 LoopExitBlock = ExitBlock;
1738 LoopVectorBody = VecBody;
1739 LoopScalarBody = OldBasicBlock;
1742 /// This function returns the identity element (or neutral element) for
1743 /// the operation K.
1745 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1750 // Adding, Xoring, Oring zero to a number does not change it.
1751 return ConstantInt::get(Tp, 0);
1752 case RK_IntegerMult:
1753 // Multiplying a number by 1 does not change it.
1754 return ConstantInt::get(Tp, 1);
1756 // AND-ing a number with an all-1 value does not change it.
1757 return ConstantInt::get(Tp, -1, true);
1759 // Multiplying a number by 1 does not change it.
1760 return ConstantFP::get(Tp, 1.0L);
1762 // Adding zero to a number does not change it.
1763 return ConstantFP::get(Tp, 0.0L);
1765 llvm_unreachable("Unknown reduction kind");
1769 static Intrinsic::ID
1770 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1771 // If we have an intrinsic call, check if it is trivially vectorizable.
1772 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1773 switch (II->getIntrinsicID()) {
1774 case Intrinsic::sqrt:
1775 case Intrinsic::sin:
1776 case Intrinsic::cos:
1777 case Intrinsic::exp:
1778 case Intrinsic::exp2:
1779 case Intrinsic::log:
1780 case Intrinsic::log10:
1781 case Intrinsic::log2:
1782 case Intrinsic::fabs:
1783 case Intrinsic::floor:
1784 case Intrinsic::ceil:
1785 case Intrinsic::trunc:
1786 case Intrinsic::rint:
1787 case Intrinsic::nearbyint:
1788 case Intrinsic::pow:
1789 case Intrinsic::fma:
1790 case Intrinsic::fmuladd:
1791 return II->getIntrinsicID();
1793 return Intrinsic::not_intrinsic;
1798 return Intrinsic::not_intrinsic;
1801 Function *F = CI->getCalledFunction();
1802 // We're going to make assumptions on the semantics of the functions, check
1803 // that the target knows that it's available in this environment.
1804 if (!F || !TLI->getLibFunc(F->getName(), Func))
1805 return Intrinsic::not_intrinsic;
1807 // Otherwise check if we have a call to a function that can be turned into a
1808 // vector intrinsic.
1815 return Intrinsic::sin;
1819 return Intrinsic::cos;
1823 return Intrinsic::exp;
1825 case LibFunc::exp2f:
1826 case LibFunc::exp2l:
1827 return Intrinsic::exp2;
1831 return Intrinsic::log;
1832 case LibFunc::log10:
1833 case LibFunc::log10f:
1834 case LibFunc::log10l:
1835 return Intrinsic::log10;
1837 case LibFunc::log2f:
1838 case LibFunc::log2l:
1839 return Intrinsic::log2;
1841 case LibFunc::fabsf:
1842 case LibFunc::fabsl:
1843 return Intrinsic::fabs;
1844 case LibFunc::floor:
1845 case LibFunc::floorf:
1846 case LibFunc::floorl:
1847 return Intrinsic::floor;
1849 case LibFunc::ceilf:
1850 case LibFunc::ceill:
1851 return Intrinsic::ceil;
1852 case LibFunc::trunc:
1853 case LibFunc::truncf:
1854 case LibFunc::truncl:
1855 return Intrinsic::trunc;
1857 case LibFunc::rintf:
1858 case LibFunc::rintl:
1859 return Intrinsic::rint;
1860 case LibFunc::nearbyint:
1861 case LibFunc::nearbyintf:
1862 case LibFunc::nearbyintl:
1863 return Intrinsic::nearbyint;
1867 return Intrinsic::pow;
1870 return Intrinsic::not_intrinsic;
1873 /// This function translates the reduction kind to an LLVM binary operator.
1875 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1877 case LoopVectorizationLegality::RK_IntegerAdd:
1878 return Instruction::Add;
1879 case LoopVectorizationLegality::RK_IntegerMult:
1880 return Instruction::Mul;
1881 case LoopVectorizationLegality::RK_IntegerOr:
1882 return Instruction::Or;
1883 case LoopVectorizationLegality::RK_IntegerAnd:
1884 return Instruction::And;
1885 case LoopVectorizationLegality::RK_IntegerXor:
1886 return Instruction::Xor;
1887 case LoopVectorizationLegality::RK_FloatMult:
1888 return Instruction::FMul;
1889 case LoopVectorizationLegality::RK_FloatAdd:
1890 return Instruction::FAdd;
1891 case LoopVectorizationLegality::RK_IntegerMinMax:
1892 return Instruction::ICmp;
1893 case LoopVectorizationLegality::RK_FloatMinMax:
1894 return Instruction::FCmp;
1896 llvm_unreachable("Unknown reduction operation");
1900 Value *createMinMaxOp(IRBuilder<> &Builder,
1901 LoopVectorizationLegality::MinMaxReductionKind RK,
1904 CmpInst::Predicate P = CmpInst::ICMP_NE;
1907 llvm_unreachable("Unknown min/max reduction kind");
1908 case LoopVectorizationLegality::MRK_UIntMin:
1909 P = CmpInst::ICMP_ULT;
1911 case LoopVectorizationLegality::MRK_UIntMax:
1912 P = CmpInst::ICMP_UGT;
1914 case LoopVectorizationLegality::MRK_SIntMin:
1915 P = CmpInst::ICMP_SLT;
1917 case LoopVectorizationLegality::MRK_SIntMax:
1918 P = CmpInst::ICMP_SGT;
1920 case LoopVectorizationLegality::MRK_FloatMin:
1921 P = CmpInst::FCMP_OLT;
1923 case LoopVectorizationLegality::MRK_FloatMax:
1924 P = CmpInst::FCMP_OGT;
1929 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1930 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1932 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1934 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1939 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1940 //===------------------------------------------------===//
1942 // Notice: any optimization or new instruction that go
1943 // into the code below should be also be implemented in
1946 //===------------------------------------------------===//
1947 Constant *Zero = Builder.getInt32(0);
1949 // In order to support reduction variables we need to be able to vectorize
1950 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1951 // stages. First, we create a new vector PHI node with no incoming edges.
1952 // We use this value when we vectorize all of the instructions that use the
1953 // PHI. Next, after all of the instructions in the block are complete we
1954 // add the new incoming edges to the PHI. At this point all of the
1955 // instructions in the basic block are vectorized, so we can use them to
1956 // construct the PHI.
1957 PhiVector RdxPHIsToFix;
1959 // Scan the loop in a topological order to ensure that defs are vectorized
1961 LoopBlocksDFS DFS(OrigLoop);
1964 // Vectorize all of the blocks in the original loop.
1965 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1966 be = DFS.endRPO(); bb != be; ++bb)
1967 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1969 // At this point every instruction in the original loop is widened to
1970 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1971 // that we vectorized. The PHI nodes are currently empty because we did
1972 // not want to introduce cycles. Notice that the remaining PHI nodes
1973 // that we need to fix are reduction variables.
1975 // Create the 'reduced' values for each of the induction vars.
1976 // The reduced values are the vector values that we scalarize and combine
1977 // after the loop is finished.
1978 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1980 PHINode *RdxPhi = *it;
1981 assert(RdxPhi && "Unable to recover vectorized PHI");
1983 // Find the reduction variable descriptor.
1984 assert(Legal->getReductionVars()->count(RdxPhi) &&
1985 "Unable to find the reduction variable");
1986 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1987 (*Legal->getReductionVars())[RdxPhi];
1989 // We need to generate a reduction vector from the incoming scalar.
1990 // To do so, we need to generate the 'identity' vector and overide
1991 // one of the elements with the incoming scalar reduction. We need
1992 // to do it in the vector-loop preheader.
1993 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1995 // This is the vector-clone of the value that leaves the loop.
1996 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1997 Type *VecTy = VectorExit[0]->getType();
1999 // Find the reduction identity variable. Zero for addition, or, xor,
2000 // one for multiplication, -1 for And.
2003 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2004 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2005 // MinMax reduction have the start value as their identify.
2006 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2010 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2011 VecTy->getScalarType());
2012 Identity = ConstantVector::getSplat(VF, Iden);
2014 // This vector is the Identity vector where the first element is the
2015 // incoming scalar reduction.
2016 VectorStart = Builder.CreateInsertElement(Identity,
2017 RdxDesc.StartValue, Zero);
2020 // Fix the vector-loop phi.
2021 // We created the induction variable so we know that the
2022 // preheader is the first entry.
2023 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2025 // Reductions do not have to start at zero. They can start with
2026 // any loop invariant values.
2027 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2028 BasicBlock *Latch = OrigLoop->getLoopLatch();
2029 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2030 VectorParts &Val = getVectorValue(LoopVal);
2031 for (unsigned part = 0; part < UF; ++part) {
2032 // Make sure to add the reduction stat value only to the
2033 // first unroll part.
2034 Value *StartVal = (part == 0) ? VectorStart : Identity;
2035 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2036 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2039 // Before each round, move the insertion point right between
2040 // the PHIs and the values we are going to write.
2041 // This allows us to write both PHINodes and the extractelement
2043 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2045 VectorParts RdxParts;
2046 for (unsigned part = 0; part < UF; ++part) {
2047 // This PHINode contains the vectorized reduction variable, or
2048 // the initial value vector, if we bypass the vector loop.
2049 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2050 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2051 Value *StartVal = (part == 0) ? VectorStart : Identity;
2052 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2053 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2054 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2055 RdxParts.push_back(NewPhi);
2058 // Reduce all of the unrolled parts into a single vector.
2059 Value *ReducedPartRdx = RdxParts[0];
2060 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2061 for (unsigned part = 1; part < UF; ++part) {
2062 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2063 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2064 RdxParts[part], ReducedPartRdx,
2067 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2068 ReducedPartRdx, RdxParts[part]);
2071 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2072 // and vector ops, reducing the set of values being computed by half each
2074 assert(isPowerOf2_32(VF) &&
2075 "Reduction emission only supported for pow2 vectors!");
2076 Value *TmpVec = ReducedPartRdx;
2077 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2078 for (unsigned i = VF; i != 1; i >>= 1) {
2079 // Move the upper half of the vector to the lower half.
2080 for (unsigned j = 0; j != i/2; ++j)
2081 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2083 // Fill the rest of the mask with undef.
2084 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2085 UndefValue::get(Builder.getInt32Ty()));
2088 Builder.CreateShuffleVector(TmpVec,
2089 UndefValue::get(TmpVec->getType()),
2090 ConstantVector::get(ShuffleMask),
2093 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2094 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2097 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2100 // The result is in the first element of the vector.
2101 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2103 // Now, we need to fix the users of the reduction variable
2104 // inside and outside of the scalar remainder loop.
2105 // We know that the loop is in LCSSA form. We need to update the
2106 // PHI nodes in the exit blocks.
2107 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2108 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2109 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2110 if (!LCSSAPhi) continue;
2112 // All PHINodes need to have a single entry edge, or two if
2113 // we already fixed them.
2114 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2116 // We found our reduction value exit-PHI. Update it with the
2117 // incoming bypass edge.
2118 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2119 // Add an edge coming from the bypass.
2120 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2123 }// end of the LCSSA phi scan.
2125 // Fix the scalar loop reduction variable with the incoming reduction sum
2126 // from the vector body and from the backedge value.
2127 int IncomingEdgeBlockIdx =
2128 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2129 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2130 // Pick the other block.
2131 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2132 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2133 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2134 }// end of for each redux variable.
2136 // The Loop exit block may have single value PHI nodes where the incoming
2137 // value is 'undef'. While vectorizing we only handled real values that
2138 // were defined inside the loop. Here we handle the 'undef case'.
2140 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2141 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2142 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2143 if (!LCSSAPhi) continue;
2144 if (LCSSAPhi->getNumIncomingValues() == 1)
2145 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2150 InnerLoopVectorizer::VectorParts
2151 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2152 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2155 VectorParts SrcMask = createBlockInMask(Src);
2157 // The terminator has to be a branch inst!
2158 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2159 assert(BI && "Unexpected terminator found");
2161 if (BI->isConditional()) {
2162 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2164 if (BI->getSuccessor(0) != Dst)
2165 for (unsigned part = 0; part < UF; ++part)
2166 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2168 for (unsigned part = 0; part < UF; ++part)
2169 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2176 InnerLoopVectorizer::VectorParts
2177 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2178 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2180 // Loop incoming mask is all-one.
2181 if (OrigLoop->getHeader() == BB) {
2182 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2183 return getVectorValue(C);
2186 // This is the block mask. We OR all incoming edges, and with zero.
2187 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2188 VectorParts BlockMask = getVectorValue(Zero);
2191 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2192 VectorParts EM = createEdgeMask(*it, BB);
2193 for (unsigned part = 0; part < UF; ++part)
2194 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2201 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2202 BasicBlock *BB, PhiVector *PV) {
2203 // For each instruction in the old loop.
2204 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2205 VectorParts &Entry = WidenMap.get(it);
2206 switch (it->getOpcode()) {
2207 case Instruction::Br:
2208 // Nothing to do for PHIs and BR, since we already took care of the
2209 // loop control flow instructions.
2211 case Instruction::PHI:{
2212 PHINode* P = cast<PHINode>(it);
2213 // Handle reduction variables:
2214 if (Legal->getReductionVars()->count(P)) {
2215 for (unsigned part = 0; part < UF; ++part) {
2216 // This is phase one of vectorizing PHIs.
2217 Type *VecTy = VectorType::get(it->getType(), VF);
2218 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2219 LoopVectorBody-> getFirstInsertionPt());
2225 // Check for PHI nodes that are lowered to vector selects.
2226 if (P->getParent() != OrigLoop->getHeader()) {
2227 // We know that all PHIs in non header blocks are converted into
2228 // selects, so we don't have to worry about the insertion order and we
2229 // can just use the builder.
2230 // At this point we generate the predication tree. There may be
2231 // duplications since this is a simple recursive scan, but future
2232 // optimizations will clean it up.
2234 unsigned NumIncoming = P->getNumIncomingValues();
2236 // Generate a sequence of selects of the form:
2237 // SELECT(Mask3, In3,
2238 // SELECT(Mask2, In2,
2240 for (unsigned In = 0; In < NumIncoming; In++) {
2241 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2243 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2245 for (unsigned part = 0; part < UF; ++part) {
2246 // We might have single edge PHIs (blocks) - use an identity
2247 // 'select' for the first PHI operand.
2249 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2252 // Select between the current value and the previous incoming edge
2253 // based on the incoming mask.
2254 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2255 Entry[part], "predphi");
2261 // This PHINode must be an induction variable.
2262 // Make sure that we know about it.
2263 assert(Legal->getInductionVars()->count(P) &&
2264 "Not an induction variable");
2266 LoopVectorizationLegality::InductionInfo II =
2267 Legal->getInductionVars()->lookup(P);
2270 case LoopVectorizationLegality::IK_NoInduction:
2271 llvm_unreachable("Unknown induction");
2272 case LoopVectorizationLegality::IK_IntInduction: {
2273 assert(P->getType() == II.StartValue->getType() && "Types must match");
2274 Type *PhiTy = P->getType();
2276 if (P == OldInduction) {
2277 // Handle the canonical induction variable. We might have had to
2279 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2281 // Handle other induction variables that are now based on the
2283 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2285 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2286 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2289 Broadcasted = getBroadcastInstrs(Broadcasted);
2290 // After broadcasting the induction variable we need to make the vector
2291 // consecutive by adding 0, 1, 2, etc.
2292 for (unsigned part = 0; part < UF; ++part)
2293 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2296 case LoopVectorizationLegality::IK_ReverseIntInduction:
2297 case LoopVectorizationLegality::IK_PtrInduction:
2298 case LoopVectorizationLegality::IK_ReversePtrInduction:
2299 // Handle reverse integer and pointer inductions.
2300 Value *StartIdx = ExtendedIdx;
2301 // This is the normalized GEP that starts counting at zero.
2302 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2305 // Handle the reverse integer induction variable case.
2306 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2307 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2308 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2310 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2313 // This is a new value so do not hoist it out.
2314 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2315 // After broadcasting the induction variable we need to make the
2316 // vector consecutive by adding ... -3, -2, -1, 0.
2317 for (unsigned part = 0; part < UF; ++part)
2318 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2323 // Handle the pointer induction variable case.
2324 assert(P->getType()->isPointerTy() && "Unexpected type.");
2326 // Is this a reverse induction ptr or a consecutive induction ptr.
2327 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2330 // This is the vector of results. Notice that we don't generate
2331 // vector geps because scalar geps result in better code.
2332 for (unsigned part = 0; part < UF; ++part) {
2333 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2334 for (unsigned int i = 0; i < VF; ++i) {
2335 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2336 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2339 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2341 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2343 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2345 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2346 Builder.getInt32(i),
2349 Entry[part] = VecVal;
2356 case Instruction::Add:
2357 case Instruction::FAdd:
2358 case Instruction::Sub:
2359 case Instruction::FSub:
2360 case Instruction::Mul:
2361 case Instruction::FMul:
2362 case Instruction::UDiv:
2363 case Instruction::SDiv:
2364 case Instruction::FDiv:
2365 case Instruction::URem:
2366 case Instruction::SRem:
2367 case Instruction::FRem:
2368 case Instruction::Shl:
2369 case Instruction::LShr:
2370 case Instruction::AShr:
2371 case Instruction::And:
2372 case Instruction::Or:
2373 case Instruction::Xor: {
2374 // Just widen binops.
2375 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2376 VectorParts &A = getVectorValue(it->getOperand(0));
2377 VectorParts &B = getVectorValue(it->getOperand(1));
2379 // Use this vector value for all users of the original instruction.
2380 for (unsigned Part = 0; Part < UF; ++Part) {
2381 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2383 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2384 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2385 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2386 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2387 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2389 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2390 VecOp->setIsExact(BinOp->isExact());
2396 case Instruction::Select: {
2398 // If the selector is loop invariant we can create a select
2399 // instruction with a scalar condition. Otherwise, use vector-select.
2400 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2403 // The condition can be loop invariant but still defined inside the
2404 // loop. This means that we can't just use the original 'cond' value.
2405 // We have to take the 'vectorized' value and pick the first lane.
2406 // Instcombine will make this a no-op.
2407 VectorParts &Cond = getVectorValue(it->getOperand(0));
2408 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2409 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2410 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2411 Builder.getInt32(0));
2412 for (unsigned Part = 0; Part < UF; ++Part) {
2413 Entry[Part] = Builder.CreateSelect(
2414 InvariantCond ? ScalarCond : Cond[Part],
2421 case Instruction::ICmp:
2422 case Instruction::FCmp: {
2423 // Widen compares. Generate vector compares.
2424 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2425 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2426 VectorParts &A = getVectorValue(it->getOperand(0));
2427 VectorParts &B = getVectorValue(it->getOperand(1));
2428 for (unsigned Part = 0; Part < UF; ++Part) {
2431 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2433 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2439 case Instruction::Store:
2440 case Instruction::Load:
2441 vectorizeMemoryInstruction(it, Legal);
2443 case Instruction::ZExt:
2444 case Instruction::SExt:
2445 case Instruction::FPToUI:
2446 case Instruction::FPToSI:
2447 case Instruction::FPExt:
2448 case Instruction::PtrToInt:
2449 case Instruction::IntToPtr:
2450 case Instruction::SIToFP:
2451 case Instruction::UIToFP:
2452 case Instruction::Trunc:
2453 case Instruction::FPTrunc:
2454 case Instruction::BitCast: {
2455 CastInst *CI = dyn_cast<CastInst>(it);
2456 /// Optimize the special case where the source is the induction
2457 /// variable. Notice that we can only optimize the 'trunc' case
2458 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2459 /// c. other casts depend on pointer size.
2460 if (CI->getOperand(0) == OldInduction &&
2461 it->getOpcode() == Instruction::Trunc) {
2462 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2464 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2465 for (unsigned Part = 0; Part < UF; ++Part)
2466 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2469 /// Vectorize casts.
2470 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2472 VectorParts &A = getVectorValue(it->getOperand(0));
2473 for (unsigned Part = 0; Part < UF; ++Part)
2474 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2478 case Instruction::Call: {
2479 // Ignore dbg intrinsics.
2480 if (isa<DbgInfoIntrinsic>(it))
2483 Module *M = BB->getParent()->getParent();
2484 CallInst *CI = cast<CallInst>(it);
2485 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2486 assert(ID && "Not an intrinsic call!");
2487 for (unsigned Part = 0; Part < UF; ++Part) {
2488 SmallVector<Value*, 4> Args;
2489 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2490 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2491 Args.push_back(Arg[Part]);
2493 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2494 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2495 Entry[Part] = Builder.CreateCall(F, Args);
2501 // All other instructions are unsupported. Scalarize them.
2502 scalarizeInstruction(it);
2505 }// end of for_each instr.
2508 void InnerLoopVectorizer::updateAnalysis() {
2509 // Forget the original basic block.
2510 SE->forgetLoop(OrigLoop);
2512 // Update the dominator tree information.
2513 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2514 "Entry does not dominate exit.");
2516 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2517 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2518 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2519 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2520 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2521 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2522 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2523 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2525 DEBUG(DT->verifyAnalysis());
2528 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2529 if (!EnableIfConversion)
2532 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2533 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2535 // Collect the blocks that need predication.
2536 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2537 BasicBlock *BB = LoopBlocks[i];
2539 // We don't support switch statements inside loops.
2540 if (!isa<BranchInst>(BB->getTerminator()))
2543 // We must be able to predicate all blocks that need to be predicated.
2544 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2548 // Check that we can actually speculate the hoistable loads.
2549 if (!LoadSpeculation.canHoistAllLoads())
2552 // We can if-convert this loop.
2556 bool LoopVectorizationLegality::canVectorize() {
2557 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2558 // be canonicalized.
2559 if (!TheLoop->getLoopPreheader())
2562 // We can only vectorize innermost loops.
2563 if (TheLoop->getSubLoopsVector().size())
2566 // We must have a single backedge.
2567 if (TheLoop->getNumBackEdges() != 1)
2570 // We must have a single exiting block.
2571 if (!TheLoop->getExitingBlock())
2574 unsigned NumBlocks = TheLoop->getNumBlocks();
2576 // Check if we can if-convert non single-bb loops.
2577 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2578 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2582 // We need to have a loop header.
2583 BasicBlock *Latch = TheLoop->getLoopLatch();
2584 DEBUG(dbgs() << "LV: Found a loop: " <<
2585 TheLoop->getHeader()->getName() << "\n");
2587 // ScalarEvolution needs to be able to find the exit count.
2588 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2589 if (ExitCount == SE->getCouldNotCompute()) {
2590 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2594 // Do not loop-vectorize loops with a tiny trip count.
2595 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2596 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2597 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2598 "This loop is not worth vectorizing.\n");
2602 // Check if we can vectorize the instructions and CFG in this loop.
2603 if (!canVectorizeInstrs()) {
2604 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2608 // Go over each instruction and look at memory deps.
2609 if (!canVectorizeMemory()) {
2610 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2614 // Collect all of the variables that remain uniform after vectorization.
2615 collectLoopUniforms();
2617 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2618 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2621 // Okay! We can vectorize. At this point we don't have any other mem analysis
2622 // which may limit our maximum vectorization factor, so just return true with
2627 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2628 if (Ty->isPointerTy())
2629 return DL.getIntPtrType(Ty->getContext());
2633 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2634 Ty0 = convertPointerToIntegerType(DL, Ty0);
2635 Ty1 = convertPointerToIntegerType(DL, Ty1);
2636 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2641 bool LoopVectorizationLegality::canVectorizeInstrs() {
2642 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2643 BasicBlock *Header = TheLoop->getHeader();
2645 // Look for the attribute signaling the absence of NaNs.
2646 Function &F = *Header->getParent();
2647 if (F.hasFnAttribute("no-nans-fp-math"))
2648 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2649 AttributeSet::FunctionIndex,
2650 "no-nans-fp-math").getValueAsString() == "true";
2652 // For each block in the loop.
2653 for (Loop::block_iterator bb = TheLoop->block_begin(),
2654 be = TheLoop->block_end(); bb != be; ++bb) {
2656 // Scan the instructions in the block and look for hazards.
2657 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2660 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2661 Type *PhiTy = Phi->getType();
2662 // Check that this PHI type is allowed.
2663 if (!PhiTy->isIntegerTy() &&
2664 !PhiTy->isFloatingPointTy() &&
2665 !PhiTy->isPointerTy()) {
2666 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2670 // If this PHINode is not in the header block, then we know that we
2671 // can convert it to select during if-conversion. No need to check if
2672 // the PHIs in this block are induction or reduction variables.
2676 // We only allow if-converted PHIs with more than two incoming values.
2677 if (Phi->getNumIncomingValues() != 2) {
2678 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2682 // This is the value coming from the preheader.
2683 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2684 // Check if this is an induction variable.
2685 InductionKind IK = isInductionVariable(Phi);
2687 if (IK_NoInduction != IK) {
2688 // Get the widest type.
2690 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2692 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2694 // Int inductions are special because we only allow one IV.
2695 if (IK == IK_IntInduction) {
2696 // Use the phi node with the widest type as induction. Use the last
2697 // one if there are multiple (no good reason for doing this other
2698 // than it is expedient).
2699 if (!Induction || PhiTy == WidestIndTy)
2703 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2704 Inductions[Phi] = InductionInfo(StartValue, IK);
2708 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2709 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2712 if (AddReductionVar(Phi, RK_IntegerMult)) {
2713 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2716 if (AddReductionVar(Phi, RK_IntegerOr)) {
2717 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2720 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2721 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2724 if (AddReductionVar(Phi, RK_IntegerXor)) {
2725 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2728 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2729 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2732 if (AddReductionVar(Phi, RK_FloatMult)) {
2733 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2736 if (AddReductionVar(Phi, RK_FloatAdd)) {
2737 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2740 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2741 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2745 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2747 }// end of PHI handling
2749 // We still don't handle functions. However, we can ignore dbg intrinsic
2750 // calls and we do handle certain intrinsic and libm functions.
2751 CallInst *CI = dyn_cast<CallInst>(it);
2752 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2753 DEBUG(dbgs() << "LV: Found a call site.\n");
2757 // Check that the instruction return type is vectorizable.
2758 if (!VectorType::isValidElementType(it->getType()) &&
2759 !it->getType()->isVoidTy()) {
2760 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2764 // Check that the stored type is vectorizable.
2765 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2766 Type *T = ST->getValueOperand()->getType();
2767 if (!VectorType::isValidElementType(T))
2771 // Reduction instructions are allowed to have exit users.
2772 // All other instructions must not have external users.
2773 if (!AllowedExit.count(it))
2774 //Check that all of the users of the loop are inside the BB.
2775 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2777 Instruction *U = cast<Instruction>(*I);
2778 // This user may be a reduction exit value.
2779 if (!TheLoop->contains(U)) {
2780 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2789 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2790 if (Inductions.empty())
2797 void LoopVectorizationLegality::collectLoopUniforms() {
2798 // We now know that the loop is vectorizable!
2799 // Collect variables that will remain uniform after vectorization.
2800 std::vector<Value*> Worklist;
2801 BasicBlock *Latch = TheLoop->getLoopLatch();
2803 // Start with the conditional branch and walk up the block.
2804 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2806 while (Worklist.size()) {
2807 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2808 Worklist.pop_back();
2810 // Look at instructions inside this loop.
2811 // Stop when reaching PHI nodes.
2812 // TODO: we need to follow values all over the loop, not only in this block.
2813 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2816 // This is a known uniform.
2819 // Insert all operands.
2820 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2824 AliasAnalysis::Location
2825 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2826 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2827 return AA->getLocation(Store);
2828 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2829 return AA->getLocation(Load);
2831 llvm_unreachable("Should be either load or store instruction");
2835 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2838 AliasMultiMap& WriteObjects,
2839 unsigned MaxByteWidth) {
2841 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2843 std::vector<Instruction*>::iterator
2844 it = WriteObjects[Object].begin(),
2845 end = WriteObjects[Object].end();
2847 for (; it != end; ++it) {
2848 Instruction* I = *it;
2852 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2853 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2854 ThatLoc.getWithNewSize(MaxByteWidth)))
2860 bool LoopVectorizationLegality::canVectorizeMemory() {
2862 typedef SmallVector<Value*, 16> ValueVector;
2863 typedef SmallPtrSet<Value*, 16> ValueSet;
2864 // Holds the Load and Store *instructions*.
2867 PtrRtCheck.Pointers.clear();
2868 PtrRtCheck.Need = false;
2870 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2873 for (Loop::block_iterator bb = TheLoop->block_begin(),
2874 be = TheLoop->block_end(); bb != be; ++bb) {
2876 // Scan the BB and collect legal loads and stores.
2877 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2880 // If this is a load, save it. If this instruction can read from memory
2881 // but is not a load, then we quit. Notice that we don't handle function
2882 // calls that read or write.
2883 if (it->mayReadFromMemory()) {
2884 LoadInst *Ld = dyn_cast<LoadInst>(it);
2885 if (!Ld) return false;
2886 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2887 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2890 Loads.push_back(Ld);
2894 // Save 'store' instructions. Abort if other instructions write to memory.
2895 if (it->mayWriteToMemory()) {
2896 StoreInst *St = dyn_cast<StoreInst>(it);
2897 if (!St) return false;
2898 if (!St->isSimple() && !IsAnnotatedParallel) {
2899 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2902 Stores.push_back(St);
2907 // Now we have two lists that hold the loads and the stores.
2908 // Next, we find the pointers that they use.
2910 // Check if we see any stores. If there are no stores, then we don't
2911 // care if the pointers are *restrict*.
2912 if (!Stores.size()) {
2913 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2917 // Holds the read and read-write *pointers* that we find. These maps hold
2918 // unique values for pointers (so no need for multi-map).
2920 AliasMap ReadWrites;
2922 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2923 // multiple times on the same object. If the ptr is accessed twice, once
2924 // for read and once for write, it will only appear once (on the write
2925 // list). This is okay, since we are going to check for conflicts between
2926 // writes and between reads and writes, but not between reads and reads.
2929 ValueVector::iterator I, IE;
2930 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2931 StoreInst *ST = cast<StoreInst>(*I);
2932 Value* Ptr = ST->getPointerOperand();
2934 if (isUniform(Ptr)) {
2935 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2939 // If we did *not* see this pointer before, insert it to
2940 // the read-write list. At this phase it is only a 'write' list.
2941 if (Seen.insert(Ptr))
2942 ReadWrites.insert(std::make_pair(Ptr, ST));
2945 if (IsAnnotatedParallel) {
2947 << "LV: A loop annotated parallel, ignore memory dependency "
2952 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2953 LoadInst *LD = cast<LoadInst>(*I);
2954 Value* Ptr = LD->getPointerOperand();
2955 // If we did *not* see this pointer before, insert it to the
2956 // read list. If we *did* see it before, then it is already in
2957 // the read-write list. This allows us to vectorize expressions
2958 // such as A[i] += x; Because the address of A[i] is a read-write
2959 // pointer. This only works if the index of A[i] is consecutive.
2960 // If the address of i is unknown (for example A[B[i]]) then we may
2961 // read a few words, modify, and write a few words, and some of the
2962 // words may be written to the same address.
2963 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2964 Reads.insert(std::make_pair(Ptr, LD));
2967 // If we write (or read-write) to a single destination and there are no
2968 // other reads in this loop then is it safe to vectorize.
2969 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2970 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2974 unsigned NumReadPtrs = 0;
2975 unsigned NumWritePtrs = 0;
2977 // Find pointers with computable bounds. We are going to use this information
2978 // to place a runtime bound check.
2979 bool CanDoRT = true;
2980 AliasMap::iterator MI, ME;
2981 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2982 Value *V = (*MI).first;
2983 if (hasComputableBounds(V)) {
2984 PtrRtCheck.insert(SE, TheLoop, V, true);
2986 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2992 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2993 Value *V = (*MI).first;
2994 if (hasComputableBounds(V)) {
2995 PtrRtCheck.insert(SE, TheLoop, V, false);
2997 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
3004 // Check that we did not collect too many pointers or found a
3005 // unsizeable pointer.
3006 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
3007 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
3008 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3014 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3017 bool NeedRTCheck = false;
3019 // Biggest vectorized access possible, vector width * unroll factor.
3020 // TODO: We're being very pessimistic here, find a way to know the
3021 // real access width before getting here.
3022 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
3023 TTI->getMaximumUnrollFactor();
3024 // Now that the pointers are in two lists (Reads and ReadWrites), we
3025 // can check that there are no conflicts between each of the writes and
3026 // between the writes to the reads.
3027 // Note that WriteObjects duplicates the stores (indexed now by underlying
3028 // objects) to avoid pointing to elements inside ReadWrites.
3029 // TODO: Maybe create a new type where they can interact without duplication.
3030 AliasMultiMap WriteObjects;
3031 ValueVector TempObjects;
3033 // Check that the read-writes do not conflict with other read-write
3035 bool AllWritesIdentified = true;
3036 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
3037 Value *Val = (*MI).first;
3038 Instruction *Inst = (*MI).second;
3040 GetUnderlyingObjects(Val, TempObjects, DL);
3041 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
3043 if (!isIdentifiedObject(*UI)) {
3044 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
3046 AllWritesIdentified = false;
3049 // Never seen it before, can't alias.
3050 if (WriteObjects[*UI].empty()) {
3051 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
3052 WriteObjects[*UI].push_back(Inst);
3055 // Direct alias found.
3056 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
3057 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
3061 DEBUG(dbgs() << "LV: Found a conflicting global value:"
3063 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
3064 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
3066 // If global alias, make sure they do alias.
3067 if (hasPossibleGlobalWriteReorder(*UI,
3071 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
3076 // Didn't alias, insert into map for further reference.
3077 WriteObjects[*UI].push_back(Inst);
3079 TempObjects.clear();
3082 /// Check that the reads don't conflict with the read-writes.
3083 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
3084 Value *Val = (*MI).first;
3085 GetUnderlyingObjects(Val, TempObjects, DL);
3086 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
3088 // If all of the writes are identified then we don't care if the read
3089 // pointer is identified or not.
3090 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
3091 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
3095 // Never seen it before, can't alias.
3096 if (WriteObjects[*UI].empty())
3098 // Direct alias found.
3099 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
3100 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
3104 DEBUG(dbgs() << "LV: Found a global value: "
3106 Instruction *Inst = (*MI).second;
3107 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
3108 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
3110 // If global alias, make sure they do alias.
3111 if (hasPossibleGlobalWriteReorder(*UI,
3115 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
3120 TempObjects.clear();
3123 PtrRtCheck.Need = NeedRTCheck;
3124 if (NeedRTCheck && !CanDoRT) {
3125 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3126 "the array bounds.\n");
3131 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3132 " need a runtime memory check.\n");
3136 static bool hasMultipleUsesOf(Instruction *I,
3137 SmallPtrSet<Instruction *, 8> &Insts) {
3138 unsigned NumUses = 0;
3139 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3140 if (Insts.count(dyn_cast<Instruction>(*Use)))
3149 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3150 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3151 if (!Set.count(dyn_cast<Instruction>(*Use)))
3156 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3157 ReductionKind Kind) {
3158 if (Phi->getNumIncomingValues() != 2)
3161 // Reduction variables are only found in the loop header block.
3162 if (Phi->getParent() != TheLoop->getHeader())
3165 // Obtain the reduction start value from the value that comes from the loop
3167 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3169 // ExitInstruction is the single value which is used outside the loop.
3170 // We only allow for a single reduction value to be used outside the loop.
3171 // This includes users of the reduction, variables (which form a cycle
3172 // which ends in the phi node).
3173 Instruction *ExitInstruction = 0;
3174 // Indicates that we found a reduction operation in our scan.
3175 bool FoundReduxOp = false;
3177 // We start with the PHI node and scan for all of the users of this
3178 // instruction. All users must be instructions that can be used as reduction
3179 // variables (such as ADD). We must have a single out-of-block user. The cycle
3180 // must include the original PHI.
3181 bool FoundStartPHI = false;
3183 // To recognize min/max patterns formed by a icmp select sequence, we store
3184 // the number of instruction we saw from the recognized min/max pattern,
3185 // to make sure we only see exactly the two instructions.
3186 unsigned NumCmpSelectPatternInst = 0;
3187 ReductionInstDesc ReduxDesc(false, 0);
3189 SmallPtrSet<Instruction *, 8> VisitedInsts;
3190 SmallVector<Instruction *, 8> Worklist;
3191 Worklist.push_back(Phi);
3192 VisitedInsts.insert(Phi);
3194 // A value in the reduction can be used:
3195 // - By the reduction:
3196 // - Reduction operation:
3197 // - One use of reduction value (safe).
3198 // - Multiple use of reduction value (not safe).
3200 // - All uses of the PHI must be the reduction (safe).
3201 // - Otherwise, not safe.
3202 // - By one instruction outside of the loop (safe).
3203 // - By further instructions outside of the loop (not safe).
3204 // - By an instruction that is not part of the reduction (not safe).
3206 // * An instruction type other than PHI or the reduction operation.
3207 // * A PHI in the header other than the initial PHI.
3208 while (!Worklist.empty()) {
3209 Instruction *Cur = Worklist.back();
3210 Worklist.pop_back();
3213 // If the instruction has no users then this is a broken chain and can't be
3214 // a reduction variable.
3215 if (Cur->use_empty())
3218 bool IsAPhi = isa<PHINode>(Cur);
3220 // A header PHI use other than the original PHI.
3221 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3224 // Reductions of instructions such as Div, and Sub is only possible if the
3225 // LHS is the reduction variable.
3226 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3227 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3228 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3231 // Any reduction instruction must be of one of the allowed kinds.
3232 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3233 if (!ReduxDesc.IsReduction)
3236 // A reduction operation must only have one use of the reduction value.
3237 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3238 hasMultipleUsesOf(Cur, VisitedInsts))
3241 // All inputs to a PHI node must be a reduction value.
3242 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3245 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3246 isa<SelectInst>(Cur)))
3247 ++NumCmpSelectPatternInst;
3248 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3249 isa<SelectInst>(Cur)))
3250 ++NumCmpSelectPatternInst;
3252 // Check whether we found a reduction operator.
3253 FoundReduxOp |= !IsAPhi;
3255 // Process users of current instruction. Push non PHI nodes after PHI nodes
3256 // onto the stack. This way we are going to have seen all inputs to PHI
3257 // nodes once we get to them.
3258 SmallVector<Instruction *, 8> NonPHIs;
3259 SmallVector<Instruction *, 8> PHIs;
3260 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3262 Instruction *Usr = cast<Instruction>(*UI);
3264 // Check if we found the exit user.
3265 BasicBlock *Parent = Usr->getParent();
3266 if (!TheLoop->contains(Parent)) {
3267 // Exit if you find multiple outside users.
3268 if (ExitInstruction != 0)
3270 ExitInstruction = Cur;
3274 // Process instructions only once (termination).
3275 if (VisitedInsts.insert(Usr)) {
3276 if (isa<PHINode>(Usr))
3277 PHIs.push_back(Usr);
3279 NonPHIs.push_back(Usr);
3281 // Remember that we completed the cycle.
3283 FoundStartPHI = true;
3285 Worklist.append(PHIs.begin(), PHIs.end());
3286 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3289 // This means we have seen one but not the other instruction of the
3290 // pattern or more than just a select and cmp.
3291 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3292 NumCmpSelectPatternInst != 2)
3295 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3298 // We found a reduction var if we have reached the original phi node and we
3299 // only have a single instruction with out-of-loop users.
3301 // This instruction is allowed to have out-of-loop users.
3302 AllowedExit.insert(ExitInstruction);
3304 // Save the description of this reduction variable.
3305 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3306 ReduxDesc.MinMaxKind);
3307 Reductions[Phi] = RD;
3308 // We've ended the cycle. This is a reduction variable if we have an
3309 // outside user and it has a binary op.
3314 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3315 /// pattern corresponding to a min(X, Y) or max(X, Y).
3316 LoopVectorizationLegality::ReductionInstDesc
3317 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3318 ReductionInstDesc &Prev) {
3320 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3321 "Expect a select instruction");
3322 Instruction *Cmp = 0;
3323 SelectInst *Select = 0;
3325 // We must handle the select(cmp()) as a single instruction. Advance to the
3327 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3328 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3329 return ReductionInstDesc(false, I);
3330 return ReductionInstDesc(Select, Prev.MinMaxKind);
3333 // Only handle single use cases for now.
3334 if (!(Select = dyn_cast<SelectInst>(I)))
3335 return ReductionInstDesc(false, I);
3336 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3337 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3338 return ReductionInstDesc(false, I);
3339 if (!Cmp->hasOneUse())
3340 return ReductionInstDesc(false, I);
3345 // Look for a min/max pattern.
3346 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3347 return ReductionInstDesc(Select, MRK_UIntMin);
3348 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3349 return ReductionInstDesc(Select, MRK_UIntMax);
3350 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3351 return ReductionInstDesc(Select, MRK_SIntMax);
3352 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3353 return ReductionInstDesc(Select, MRK_SIntMin);
3354 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3355 return ReductionInstDesc(Select, MRK_FloatMin);
3356 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3357 return ReductionInstDesc(Select, MRK_FloatMax);
3358 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3359 return ReductionInstDesc(Select, MRK_FloatMin);
3360 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3361 return ReductionInstDesc(Select, MRK_FloatMax);
3363 return ReductionInstDesc(false, I);
3366 LoopVectorizationLegality::ReductionInstDesc
3367 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3369 ReductionInstDesc &Prev) {
3370 bool FP = I->getType()->isFloatingPointTy();
3371 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3372 switch (I->getOpcode()) {
3374 return ReductionInstDesc(false, I);
3375 case Instruction::PHI:
3376 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
3377 Kind != RK_FloatMinMax))
3378 return ReductionInstDesc(false, I);
3379 return ReductionInstDesc(I, Prev.MinMaxKind);
3380 case Instruction::Sub:
3381 case Instruction::Add:
3382 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3383 case Instruction::Mul:
3384 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3385 case Instruction::And:
3386 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3387 case Instruction::Or:
3388 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3389 case Instruction::Xor:
3390 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3391 case Instruction::FMul:
3392 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3393 case Instruction::FAdd:
3394 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3395 case Instruction::FCmp:
3396 case Instruction::ICmp:
3397 case Instruction::Select:
3398 if (Kind != RK_IntegerMinMax &&
3399 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
3400 return ReductionInstDesc(false, I);
3401 return isMinMaxSelectCmpPattern(I, Prev);
3405 LoopVectorizationLegality::InductionKind
3406 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3407 Type *PhiTy = Phi->getType();
3408 // We only handle integer and pointer inductions variables.
3409 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3410 return IK_NoInduction;
3412 // Check that the PHI is consecutive.
3413 const SCEV *PhiScev = SE->getSCEV(Phi);
3414 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3416 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3417 return IK_NoInduction;
3419 const SCEV *Step = AR->getStepRecurrence(*SE);
3421 // Integer inductions need to have a stride of one.
3422 if (PhiTy->isIntegerTy()) {
3424 return IK_IntInduction;
3425 if (Step->isAllOnesValue())
3426 return IK_ReverseIntInduction;
3427 return IK_NoInduction;
3430 // Calculate the pointer stride and check if it is consecutive.
3431 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3433 return IK_NoInduction;
3435 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3436 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3437 if (C->getValue()->equalsInt(Size))
3438 return IK_PtrInduction;
3439 else if (C->getValue()->equalsInt(0 - Size))
3440 return IK_ReversePtrInduction;
3442 return IK_NoInduction;
3445 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3446 Value *In0 = const_cast<Value*>(V);
3447 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3451 return Inductions.count(PN);
3454 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3455 assert(TheLoop->contains(BB) && "Unknown block used");
3457 // Blocks that do not dominate the latch need predication.
3458 BasicBlock* Latch = TheLoop->getLoopLatch();
3459 return !DT->dominates(BB, Latch);
3462 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3463 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3464 // We might be able to hoist the load.
3465 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
3468 // We don't predicate stores at the moment.
3469 if (it->mayWriteToMemory() || it->mayThrow())
3472 // The instructions below can trap.
3473 switch (it->getOpcode()) {
3475 case Instruction::UDiv:
3476 case Instruction::SDiv:
3477 case Instruction::URem:
3478 case Instruction::SRem:
3486 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3487 const SCEV *PhiScev = SE->getSCEV(Ptr);
3488 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3492 return AR->isAffine();
3495 LoopVectorizationCostModel::VectorizationFactor
3496 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3498 // Width 1 means no vectorize
3499 VectorizationFactor Factor = { 1U, 0U };
3500 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3501 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3505 // Find the trip count.
3506 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3507 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3509 unsigned WidestType = getWidestType();
3510 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3511 unsigned MaxVectorSize = WidestRegister / WidestType;
3512 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3513 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3515 if (MaxVectorSize == 0) {
3516 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3520 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3521 " into one vector!");
3523 unsigned VF = MaxVectorSize;
3525 // If we optimize the program for size, avoid creating the tail loop.
3527 // If we are unable to calculate the trip count then don't try to vectorize.
3529 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3533 // Find the maximum SIMD width that can fit within the trip count.
3534 VF = TC % MaxVectorSize;
3539 // If the trip count that we found modulo the vectorization factor is not
3540 // zero then we require a tail.
3542 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3548 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3549 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3551 Factor.Width = UserVF;
3555 float Cost = expectedCost(1);
3557 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3558 for (unsigned i=2; i <= VF; i*=2) {
3559 // Notice that the vector loop needs to be executed less times, so
3560 // we need to divide the cost of the vector loops by the width of
3561 // the vector elements.
3562 float VectorCost = expectedCost(i) / (float)i;
3563 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3564 (int)VectorCost << ".\n");
3565 if (VectorCost < Cost) {
3571 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3572 Factor.Width = Width;
3573 Factor.Cost = Width * Cost;
3577 unsigned LoopVectorizationCostModel::getWidestType() {
3578 unsigned MaxWidth = 8;
3581 for (Loop::block_iterator bb = TheLoop->block_begin(),
3582 be = TheLoop->block_end(); bb != be; ++bb) {
3583 BasicBlock *BB = *bb;
3585 // For each instruction in the loop.
3586 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3587 Type *T = it->getType();
3589 // Only examine Loads, Stores and PHINodes.
3590 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3593 // Examine PHI nodes that are reduction variables.
3594 if (PHINode *PN = dyn_cast<PHINode>(it))
3595 if (!Legal->getReductionVars()->count(PN))
3598 // Examine the stored values.
3599 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3600 T = ST->getValueOperand()->getType();
3602 // Ignore loaded pointer types and stored pointer types that are not
3603 // consecutive. However, we do want to take consecutive stores/loads of
3604 // pointer vectors into account.
3605 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3608 MaxWidth = std::max(MaxWidth,
3609 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3617 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3620 unsigned LoopCost) {
3622 // -- The unroll heuristics --
3623 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3624 // There are many micro-architectural considerations that we can't predict
3625 // at this level. For example frontend pressure (on decode or fetch) due to
3626 // code size, or the number and capabilities of the execution ports.
3628 // We use the following heuristics to select the unroll factor:
3629 // 1. If the code has reductions the we unroll in order to break the cross
3630 // iteration dependency.
3631 // 2. If the loop is really small then we unroll in order to reduce the loop
3633 // 3. We don't unroll if we think that we will spill registers to memory due
3634 // to the increased register pressure.
3636 // Use the user preference, unless 'auto' is selected.
3640 // When we optimize for size we don't unroll.
3644 // Do not unroll loops with a relatively small trip count.
3645 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3646 TheLoop->getLoopLatch());
3647 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3650 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3651 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3652 " vector registers\n");
3654 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3655 // We divide by these constants so assume that we have at least one
3656 // instruction that uses at least one register.
3657 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3658 R.NumInstructions = std::max(R.NumInstructions, 1U);
3660 // We calculate the unroll factor using the following formula.
3661 // Subtract the number of loop invariants from the number of available
3662 // registers. These registers are used by all of the unrolled instances.
3663 // Next, divide the remaining registers by the number of registers that is
3664 // required by the loop, in order to estimate how many parallel instances
3665 // fit without causing spills.
3666 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3668 // Clamp the unroll factor ranges to reasonable factors.
3669 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3671 // If we did not calculate the cost for VF (because the user selected the VF)
3672 // then we calculate the cost of VF here.
3674 LoopCost = expectedCost(VF);
3676 // Clamp the calculated UF to be between the 1 and the max unroll factor
3677 // that the target allows.
3678 if (UF > MaxUnrollSize)
3683 if (Legal->getReductionVars()->size()) {
3684 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3688 // We want to unroll tiny loops in order to reduce the loop overhead.
3689 // We assume that the cost overhead is 1 and we use the cost model
3690 // to estimate the cost of the loop and unroll until the cost of the
3691 // loop overhead is about 5% of the cost of the loop.
3692 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3693 if (LoopCost < 20) {
3694 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3695 unsigned NewUF = 20/LoopCost + 1;
3696 return std::min(NewUF, UF);
3699 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3703 LoopVectorizationCostModel::RegisterUsage
3704 LoopVectorizationCostModel::calculateRegisterUsage() {
3705 // This function calculates the register usage by measuring the highest number
3706 // of values that are alive at a single location. Obviously, this is a very
3707 // rough estimation. We scan the loop in a topological order in order and
3708 // assign a number to each instruction. We use RPO to ensure that defs are
3709 // met before their users. We assume that each instruction that has in-loop
3710 // users starts an interval. We record every time that an in-loop value is
3711 // used, so we have a list of the first and last occurrences of each
3712 // instruction. Next, we transpose this data structure into a multi map that
3713 // holds the list of intervals that *end* at a specific location. This multi
3714 // map allows us to perform a linear search. We scan the instructions linearly
3715 // and record each time that a new interval starts, by placing it in a set.
3716 // If we find this value in the multi-map then we remove it from the set.
3717 // The max register usage is the maximum size of the set.
3718 // We also search for instructions that are defined outside the loop, but are
3719 // used inside the loop. We need this number separately from the max-interval
3720 // usage number because when we unroll, loop-invariant values do not take
3722 LoopBlocksDFS DFS(TheLoop);
3726 R.NumInstructions = 0;
3728 // Each 'key' in the map opens a new interval. The values
3729 // of the map are the index of the 'last seen' usage of the
3730 // instruction that is the key.
3731 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3732 // Maps instruction to its index.
3733 DenseMap<unsigned, Instruction*> IdxToInstr;
3734 // Marks the end of each interval.
3735 IntervalMap EndPoint;
3736 // Saves the list of instruction indices that are used in the loop.
3737 SmallSet<Instruction*, 8> Ends;
3738 // Saves the list of values that are used in the loop but are
3739 // defined outside the loop, such as arguments and constants.
3740 SmallPtrSet<Value*, 8> LoopInvariants;
3743 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3744 be = DFS.endRPO(); bb != be; ++bb) {
3745 R.NumInstructions += (*bb)->size();
3746 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3748 Instruction *I = it;
3749 IdxToInstr[Index++] = I;
3751 // Save the end location of each USE.
3752 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3753 Value *U = I->getOperand(i);
3754 Instruction *Instr = dyn_cast<Instruction>(U);
3756 // Ignore non-instruction values such as arguments, constants, etc.
3757 if (!Instr) continue;
3759 // If this instruction is outside the loop then record it and continue.
3760 if (!TheLoop->contains(Instr)) {
3761 LoopInvariants.insert(Instr);
3765 // Overwrite previous end points.
3766 EndPoint[Instr] = Index;
3772 // Saves the list of intervals that end with the index in 'key'.
3773 typedef SmallVector<Instruction*, 2> InstrList;
3774 DenseMap<unsigned, InstrList> TransposeEnds;
3776 // Transpose the EndPoints to a list of values that end at each index.
3777 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3779 TransposeEnds[it->second].push_back(it->first);
3781 SmallSet<Instruction*, 8> OpenIntervals;
3782 unsigned MaxUsage = 0;
3785 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3786 for (unsigned int i = 0; i < Index; ++i) {
3787 Instruction *I = IdxToInstr[i];
3788 // Ignore instructions that are never used within the loop.
3789 if (!Ends.count(I)) continue;
3791 // Remove all of the instructions that end at this location.
3792 InstrList &List = TransposeEnds[i];
3793 for (unsigned int j=0, e = List.size(); j < e; ++j)
3794 OpenIntervals.erase(List[j]);
3796 // Count the number of live interals.
3797 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3799 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3800 OpenIntervals.size() <<"\n");
3802 // Add the current instruction to the list of open intervals.
3803 OpenIntervals.insert(I);
3806 unsigned Invariant = LoopInvariants.size();
3807 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3808 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3809 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3811 R.LoopInvariantRegs = Invariant;
3812 R.MaxLocalUsers = MaxUsage;
3816 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3820 for (Loop::block_iterator bb = TheLoop->block_begin(),
3821 be = TheLoop->block_end(); bb != be; ++bb) {
3822 unsigned BlockCost = 0;
3823 BasicBlock *BB = *bb;
3825 // For each instruction in the old loop.
3826 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3827 // Skip dbg intrinsics.
3828 if (isa<DbgInfoIntrinsic>(it))
3831 unsigned C = getInstructionCost(it, VF);
3833 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3834 VF << " For instruction: "<< *it << "\n");
3837 // We assume that if-converted blocks have a 50% chance of being executed.
3838 // When the code is scalar then some of the blocks are avoided due to CF.
3839 // When the code is vectorized we execute all code paths.
3840 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3850 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3851 // If we know that this instruction will remain uniform, check the cost of
3852 // the scalar version.
3853 if (Legal->isUniformAfterVectorization(I))
3856 Type *RetTy = I->getType();
3857 Type *VectorTy = ToVectorTy(RetTy, VF);
3859 // TODO: We need to estimate the cost of intrinsic calls.
3860 switch (I->getOpcode()) {
3861 case Instruction::GetElementPtr:
3862 // We mark this instruction as zero-cost because the cost of GEPs in
3863 // vectorized code depends on whether the corresponding memory instruction
3864 // is scalarized or not. Therefore, we handle GEPs with the memory
3865 // instruction cost.
3867 case Instruction::Br: {
3868 return TTI.getCFInstrCost(I->getOpcode());
3870 case Instruction::PHI:
3871 //TODO: IF-converted IFs become selects.
3873 case Instruction::Add:
3874 case Instruction::FAdd:
3875 case Instruction::Sub:
3876 case Instruction::FSub:
3877 case Instruction::Mul:
3878 case Instruction::FMul:
3879 case Instruction::UDiv:
3880 case Instruction::SDiv:
3881 case Instruction::FDiv:
3882 case Instruction::URem:
3883 case Instruction::SRem:
3884 case Instruction::FRem:
3885 case Instruction::Shl:
3886 case Instruction::LShr:
3887 case Instruction::AShr:
3888 case Instruction::And:
3889 case Instruction::Or:
3890 case Instruction::Xor: {
3891 // Certain instructions can be cheaper to vectorize if they have a constant
3892 // second vector operand. One example of this are shifts on x86.
3893 TargetTransformInfo::OperandValueKind Op1VK =
3894 TargetTransformInfo::OK_AnyValue;
3895 TargetTransformInfo::OperandValueKind Op2VK =
3896 TargetTransformInfo::OK_AnyValue;
3898 if (isa<ConstantInt>(I->getOperand(1)))
3899 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3901 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3903 case Instruction::Select: {
3904 SelectInst *SI = cast<SelectInst>(I);
3905 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3906 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3907 Type *CondTy = SI->getCondition()->getType();
3909 CondTy = VectorType::get(CondTy, VF);
3911 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3913 case Instruction::ICmp:
3914 case Instruction::FCmp: {
3915 Type *ValTy = I->getOperand(0)->getType();
3916 VectorTy = ToVectorTy(ValTy, VF);
3917 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3919 case Instruction::Store:
3920 case Instruction::Load: {
3921 StoreInst *SI = dyn_cast<StoreInst>(I);
3922 LoadInst *LI = dyn_cast<LoadInst>(I);
3923 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3925 VectorTy = ToVectorTy(ValTy, VF);
3927 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3928 unsigned AS = SI ? SI->getPointerAddressSpace() :
3929 LI->getPointerAddressSpace();
3930 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3931 // We add the cost of address computation here instead of with the gep
3932 // instruction because only here we know whether the operation is
3935 return TTI.getAddressComputationCost(VectorTy) +
3936 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3938 // Scalarized loads/stores.
3939 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3940 bool Reverse = ConsecutiveStride < 0;
3941 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3942 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3943 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3945 // The cost of extracting from the value vector and pointer vector.
3946 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3947 for (unsigned i = 0; i < VF; ++i) {
3948 // The cost of extracting the pointer operand.
3949 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3950 // In case of STORE, the cost of ExtractElement from the vector.
3951 // In case of LOAD, the cost of InsertElement into the returned
3953 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3954 Instruction::InsertElement,
3958 // The cost of the scalar loads/stores.
3959 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3960 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3965 // Wide load/stores.
3966 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3967 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3970 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3974 case Instruction::ZExt:
3975 case Instruction::SExt:
3976 case Instruction::FPToUI:
3977 case Instruction::FPToSI:
3978 case Instruction::FPExt:
3979 case Instruction::PtrToInt:
3980 case Instruction::IntToPtr:
3981 case Instruction::SIToFP:
3982 case Instruction::UIToFP:
3983 case Instruction::Trunc:
3984 case Instruction::FPTrunc:
3985 case Instruction::BitCast: {
3986 // We optimize the truncation of induction variable.
3987 // The cost of these is the same as the scalar operation.
3988 if (I->getOpcode() == Instruction::Trunc &&
3989 Legal->isInductionVariable(I->getOperand(0)))
3990 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3991 I->getOperand(0)->getType());
3993 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3994 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3996 case Instruction::Call: {
3997 CallInst *CI = cast<CallInst>(I);
3998 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3999 assert(ID && "Not an intrinsic call!");
4000 Type *RetTy = ToVectorTy(CI->getType(), VF);
4001 SmallVector<Type*, 4> Tys;
4002 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4003 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4004 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4007 // We are scalarizing the instruction. Return the cost of the scalar
4008 // instruction, plus the cost of insert and extract into vector
4009 // elements, times the vector width.
4012 if (!RetTy->isVoidTy() && VF != 1) {
4013 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4015 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4018 // The cost of inserting the results plus extracting each one of the
4020 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4023 // The cost of executing VF copies of the scalar instruction. This opcode
4024 // is unknown. Assume that it is the same as 'mul'.
4025 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4031 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4032 if (Scalar->isVoidTy() || VF == 1)
4034 return VectorType::get(Scalar, VF);
4037 char LoopVectorize::ID = 0;
4038 static const char lv_name[] = "Loop Vectorization";
4039 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4040 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
4041 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4042 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4043 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4044 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4047 Pass *createLoopVectorizePass() {
4048 return new LoopVectorize();
4052 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4053 // Check for a store.
4054 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4055 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4057 // Check for a load.
4058 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4059 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;