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/EquivalenceClasses.h"
51 #include "llvm/ADT/MapVector.h"
52 #include "llvm/ADT/SetVector.h"
53 #include "llvm/ADT/SmallPtrSet.h"
54 #include "llvm/ADT/SmallSet.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/StringExtras.h"
57 #include "llvm/Analysis/AliasAnalysis.h"
58 #include "llvm/Analysis/AliasSetTracker.h"
59 #include "llvm/Analysis/Dominators.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/Analysis/Verifier.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DerivedTypes.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.h"
80 #include "llvm/Pass.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/Debug.h"
83 #include "llvm/Support/PatternMatch.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/Support/ValueHandle.h"
86 #include "llvm/Target/TargetLibraryInfo.h"
87 #include "llvm/Transforms/Scalar.h"
88 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
89 #include "llvm/Transforms/Utils/Local.h"
94 using namespace llvm::PatternMatch;
96 static cl::opt<unsigned>
97 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
98 cl::desc("Sets the SIMD width. Zero is autoselect."));
100 static cl::opt<unsigned>
101 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
102 cl::desc("Sets the vectorization unroll count. "
103 "Zero is autoselect."));
106 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
107 cl::desc("Enable if-conversion during vectorization."));
109 /// We don't vectorize loops with a known constant trip count below this number.
110 static cl::opt<unsigned>
111 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
113 cl::desc("Don't vectorize loops with a constant "
114 "trip count that is smaller than this "
117 /// We don't unroll loops with a known constant trip count below this number.
118 static const unsigned TinyTripCountUnrollThreshold = 128;
120 /// When performing memory disambiguation checks at runtime do not make more
121 /// than this number of comparisons.
122 static const unsigned RuntimeMemoryCheckThreshold = 8;
124 /// Maximum simd width.
125 static const unsigned MaxVectorWidth = 64;
127 /// Maximum vectorization unroll count.
128 static const unsigned MaxUnrollFactor = 16;
132 // Forward declarations.
133 class LoopVectorizationLegality;
134 class LoopVectorizationCostModel;
136 /// InnerLoopVectorizer vectorizes loops which contain only one basic
137 /// block to a specified vectorization factor (VF).
138 /// This class performs the widening of scalars into vectors, or multiple
139 /// scalars. This class also implements the following features:
140 /// * It inserts an epilogue loop for handling loops that don't have iteration
141 /// counts that are known to be a multiple of the vectorization factor.
142 /// * It handles the code generation for reduction variables.
143 /// * Scalarization (implementation using scalars) of un-vectorizable
145 /// InnerLoopVectorizer does not perform any vectorization-legality
146 /// checks, and relies on the caller to check for the different legality
147 /// aspects. The InnerLoopVectorizer relies on the
148 /// LoopVectorizationLegality class to provide information about the induction
149 /// and reduction variables that were found to a given vectorization factor.
150 class InnerLoopVectorizer {
152 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
153 DominatorTree *DT, DataLayout *DL,
154 const TargetLibraryInfo *TLI, unsigned VecWidth,
155 unsigned UnrollFactor)
156 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
157 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
158 OldInduction(0), WidenMap(UnrollFactor) {}
160 // Perform the actual loop widening (vectorization).
161 void vectorize(LoopVectorizationLegality *Legal) {
162 // Create a new empty loop. Unlink the old loop and connect the new one.
163 createEmptyLoop(Legal);
164 // Widen each instruction in the old loop to a new one in the new loop.
165 // Use the Legality module to find the induction and reduction variables.
166 vectorizeLoop(Legal);
167 // Register the new loop and update the analysis passes.
172 /// A small list of PHINodes.
173 typedef SmallVector<PHINode*, 4> PhiVector;
174 /// When we unroll loops we have multiple vector values for each scalar.
175 /// This data structure holds the unrolled and vectorized values that
176 /// originated from one scalar instruction.
177 typedef SmallVector<Value*, 2> VectorParts;
179 /// Add code that checks at runtime if the accessed arrays overlap.
180 /// Returns the comparator value or NULL if no check is needed.
181 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
183 /// Create an empty loop, based on the loop ranges of the old loop.
184 void createEmptyLoop(LoopVectorizationLegality *Legal);
185 /// Copy and widen the instructions from the old loop.
186 void vectorizeLoop(LoopVectorizationLegality *Legal);
188 /// A helper function that computes the predicate of the block BB, assuming
189 /// that the header block of the loop is set to True. It returns the *entry*
190 /// mask for the block BB.
191 VectorParts createBlockInMask(BasicBlock *BB);
192 /// A helper function that computes the predicate of the edge between SRC
194 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
196 /// A helper function to vectorize a single BB within the innermost loop.
197 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
200 /// Insert the new loop to the loop hierarchy and pass manager
201 /// and update the analysis passes.
202 void updateAnalysis();
204 /// This instruction is un-vectorizable. Implement it as a sequence
206 void scalarizeInstruction(Instruction *Instr);
208 /// Vectorize Load and Store instructions,
209 void vectorizeMemoryInstruction(Instruction *Instr,
210 LoopVectorizationLegality *Legal);
212 /// Create a broadcast instruction. This method generates a broadcast
213 /// instruction (shuffle) for loop invariant values and for the induction
214 /// value. If this is the induction variable then we extend it to N, N+1, ...
215 /// this is needed because each iteration in the loop corresponds to a SIMD
217 Value *getBroadcastInstrs(Value *V);
219 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
220 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
221 /// The sequence starts at StartIndex.
222 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
224 /// When we go over instructions in the basic block we rely on previous
225 /// values within the current basic block or on loop invariant values.
226 /// When we widen (vectorize) values we place them in the map. If the values
227 /// are not within the map, they have to be loop invariant, so we simply
228 /// broadcast them into a vector.
229 VectorParts &getVectorValue(Value *V);
231 /// Generate a shuffle sequence that will reverse the vector Vec.
232 Value *reverseVector(Value *Vec);
234 /// This is a helper class that holds the vectorizer state. It maps scalar
235 /// instructions to vector instructions. When the code is 'unrolled' then
236 /// then a single scalar value is mapped to multiple vector parts. The parts
237 /// are stored in the VectorPart type.
239 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
241 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
243 /// \return True if 'Key' is saved in the Value Map.
244 bool has(Value *Key) const { return MapStorage.count(Key); }
246 /// Initializes a new entry in the map. Sets all of the vector parts to the
247 /// save value in 'Val'.
248 /// \return A reference to a vector with splat values.
249 VectorParts &splat(Value *Key, Value *Val) {
250 VectorParts &Entry = MapStorage[Key];
251 Entry.assign(UF, Val);
255 ///\return A reference to the value that is stored at 'Key'.
256 VectorParts &get(Value *Key) {
257 VectorParts &Entry = MapStorage[Key];
260 assert(Entry.size() == UF);
265 /// The unroll factor. Each entry in the map stores this number of vector
269 /// Map storage. We use std::map and not DenseMap because insertions to a
270 /// dense map invalidates its iterators.
271 std::map<Value *, VectorParts> MapStorage;
274 /// The original loop.
276 /// Scev analysis to use.
284 /// Target Library Info.
285 const TargetLibraryInfo *TLI;
287 /// The vectorization SIMD factor to use. Each vector will have this many
290 /// The vectorization unroll factor to use. Each scalar is vectorized to this
291 /// many different vector instructions.
294 /// The builder that we use
297 // --- Vectorization state ---
299 /// The vector-loop preheader.
300 BasicBlock *LoopVectorPreHeader;
301 /// The scalar-loop preheader.
302 BasicBlock *LoopScalarPreHeader;
303 /// Middle Block between the vector and the scalar.
304 BasicBlock *LoopMiddleBlock;
305 ///The ExitBlock of the scalar loop.
306 BasicBlock *LoopExitBlock;
307 ///The vector loop body.
308 BasicBlock *LoopVectorBody;
309 ///The scalar loop body.
310 BasicBlock *LoopScalarBody;
311 /// A list of all bypass blocks. The first block is the entry of the loop.
312 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
314 /// The new Induction variable which was added to the new block.
316 /// The induction variable of the old basic block.
317 PHINode *OldInduction;
318 /// Holds the extended (to the widest induction type) start index.
320 /// Maps scalars to widened vectors.
324 /// \brief Check if conditionally executed loads are hoistable.
326 /// This class has two functions: isHoistableLoad and canHoistAllLoads.
327 /// isHoistableLoad should be called on all load instructions that are executed
328 /// conditionally. After all conditional loads are processed, the client should
329 /// call canHoistAllLoads to determine if all of the conditional executed loads
330 /// have an unconditional memory access to the same memory address in the loop.
332 typedef SmallPtrSet<Value *, 8> MemorySet;
336 MemorySet CondLoadAddrSet;
339 LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
341 /// \brief Check if the instruction is a load with a identifiable address.
342 bool isHoistableLoad(Instruction *L);
344 /// \brief Check if all of the conditional loads are hoistable because there
345 /// exists an unconditional memory access to the same address in the loop.
346 bool canHoistAllLoads();
349 bool LoadHoisting::isHoistableLoad(Instruction *L) {
350 LoadInst *LI = dyn_cast<LoadInst>(L);
354 CondLoadAddrSet.insert(LI->getPointerOperand());
358 static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
359 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
360 if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
361 Set.insert(LI->getPointerOperand());
362 else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
363 Set.insert(SI->getPointerOperand());
367 bool LoadHoisting::canHoistAllLoads() {
368 // No conditional loads.
369 if (CondLoadAddrSet.empty())
372 MemorySet UncondMemAccesses;
373 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
374 BasicBlock *LoopLatch = TheLoop->getLoopLatch();
376 // Iterate over the unconditional blocks and collect memory access addresses.
377 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
378 BasicBlock *BB = LoopBlocks[i];
380 // Ignore conditional blocks.
381 if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
384 addMemAccesses(BB, UncondMemAccesses);
387 // And make sure there is a matching unconditional access for every
389 for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
390 ME = CondLoadAddrSet.end(); MI != ME; ++MI)
391 if (!UncondMemAccesses.count(*MI))
397 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
398 /// to what vectorization factor.
399 /// This class does not look at the profitability of vectorization, only the
400 /// legality. This class has two main kinds of checks:
401 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
402 /// will change the order of memory accesses in a way that will change the
403 /// correctness of the program.
404 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
405 /// checks for a number of different conditions, such as the availability of a
406 /// single induction variable, that all types are supported and vectorize-able,
407 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
408 /// This class is also used by InnerLoopVectorizer for identifying
409 /// induction variable and the different reduction variables.
410 class LoopVectorizationLegality {
412 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
413 DominatorTree *DT, TargetTransformInfo* TTI,
414 AliasAnalysis *AA, TargetLibraryInfo *TLI)
415 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
416 Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
417 LoadSpeculation(L, DT) {}
419 /// This enum represents the kinds of reductions that we support.
421 RK_NoReduction, ///< Not a reduction.
422 RK_IntegerAdd, ///< Sum of integers.
423 RK_IntegerMult, ///< Product of integers.
424 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
425 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
426 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
427 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
428 RK_FloatAdd, ///< Sum of floats.
429 RK_FloatMult, ///< Product of floats.
430 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
433 /// This enum represents the kinds of inductions that we support.
435 IK_NoInduction, ///< Not an induction variable.
436 IK_IntInduction, ///< Integer induction variable. Step = 1.
437 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
438 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
439 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
442 // This enum represents the kind of minmax reduction.
443 enum MinMaxReductionKind {
453 /// This POD struct holds information about reduction variables.
454 struct ReductionDescriptor {
455 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
456 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
458 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
459 MinMaxReductionKind MK)
460 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
462 // The starting value of the reduction.
463 // It does not have to be zero!
464 TrackingVH<Value> StartValue;
465 // The instruction who's value is used outside the loop.
466 Instruction *LoopExitInstr;
467 // The kind of the reduction.
469 // If this a min/max reduction the kind of reduction.
470 MinMaxReductionKind MinMaxKind;
473 /// This POD struct holds information about a potential reduction operation.
474 struct ReductionInstDesc {
475 ReductionInstDesc(bool IsRedux, Instruction *I) :
476 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
478 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
479 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
481 // Is this instruction a reduction candidate.
483 // The last instruction in a min/max pattern (select of the select(icmp())
484 // pattern), or the current reduction instruction otherwise.
485 Instruction *PatternLastInst;
486 // If this is a min/max pattern the comparison predicate.
487 MinMaxReductionKind MinMaxKind;
490 // This POD struct holds information about the memory runtime legality
491 // check that a group of pointers do not overlap.
492 struct RuntimePointerCheck {
493 RuntimePointerCheck() : Need(false) {}
495 /// Reset the state of the pointer runtime information.
503 /// Insert a pointer and calculate the start and end SCEVs.
504 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr);
506 /// This flag indicates if we need to add the runtime check.
508 /// Holds the pointers that we need to check.
509 SmallVector<TrackingVH<Value>, 2> Pointers;
510 /// Holds the pointer value at the beginning of the loop.
511 SmallVector<const SCEV*, 2> Starts;
512 /// Holds the pointer value at the end of the loop.
513 SmallVector<const SCEV*, 2> Ends;
514 /// Holds the information if this pointer is used for writing to memory.
515 SmallVector<bool, 2> IsWritePtr;
518 /// A POD for saving information about induction variables.
519 struct InductionInfo {
520 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
521 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
523 TrackingVH<Value> StartValue;
528 /// ReductionList contains the reduction descriptors for all
529 /// of the reductions that were found in the loop.
530 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
532 /// InductionList saves induction variables and maps them to the
533 /// induction descriptor.
534 typedef MapVector<PHINode*, InductionInfo> InductionList;
536 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
537 /// respective Store/Load instruction(s) to calculate aliasing.
538 typedef MapVector<Value*, Instruction* > AliasMap;
539 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
541 /// Returns true if it is legal to vectorize this loop.
542 /// This does not mean that it is profitable to vectorize this
543 /// loop, only that it is legal to do so.
546 /// Returns the Induction variable.
547 PHINode *getInduction() { return Induction; }
549 /// Returns the reduction variables found in the loop.
550 ReductionList *getReductionVars() { return &Reductions; }
552 /// Returns the induction variables found in the loop.
553 InductionList *getInductionVars() { return &Inductions; }
555 /// Returns the widest induction type.
556 Type *getWidestInductionType() { return WidestIndTy; }
558 /// Returns True if V is an induction variable in this loop.
559 bool isInductionVariable(const Value *V);
561 /// Return true if the block BB needs to be predicated in order for the loop
562 /// to be vectorized.
563 bool blockNeedsPredication(BasicBlock *BB);
565 /// Check if this pointer is consecutive when vectorizing. This happens
566 /// when the last index of the GEP is the induction variable, or that the
567 /// pointer itself is an induction variable.
568 /// This check allows us to vectorize A[idx] into a wide load/store.
570 /// 0 - Stride is unknown or non consecutive.
571 /// 1 - Address is consecutive.
572 /// -1 - Address is consecutive, and decreasing.
573 int isConsecutivePtr(Value *Ptr);
575 /// Returns true if the value V is uniform within the loop.
576 bool isUniform(Value *V);
578 /// Returns true if this instruction will remain scalar after vectorization.
579 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
581 /// Returns the information that we collected about runtime memory check.
582 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
584 /// This function returns the identity element (or neutral element) for
586 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
588 /// Check if a single basic block loop is vectorizable.
589 /// At this point we know that this is a loop with a constant trip count
590 /// and we only need to check individual instructions.
591 bool canVectorizeInstrs();
593 /// When we vectorize loops we may change the order in which
594 /// we read and write from memory. This method checks if it is
595 /// legal to vectorize the code, considering only memory constrains.
596 /// Returns true if the loop is vectorizable
597 bool canVectorizeMemory();
599 /// Return true if we can vectorize this loop using the IF-conversion
601 bool canVectorizeWithIfConvert();
603 /// Collect the variables that need to stay uniform after vectorization.
604 void collectLoopUniforms();
606 /// Return true if all of the instructions in the block can be speculatively
608 bool blockCanBePredicated(BasicBlock *BB);
610 /// Returns True, if 'Phi' is the kind of reduction variable for type
611 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
612 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
613 /// Returns a struct describing if the instruction 'I' can be a reduction
614 /// variable of type 'Kind'. If the reduction is a min/max pattern of
615 /// select(icmp()) this function advances the instruction pointer 'I' from the
616 /// compare instruction to the select instruction and stores this pointer in
617 /// 'PatternLastInst' member of the returned struct.
618 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
619 ReductionInstDesc &Desc);
620 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
621 /// pattern corresponding to a min(X, Y) or max(X, Y).
622 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
623 ReductionInstDesc &Prev);
624 /// Returns the induction kind of Phi. This function may return NoInduction
625 /// if the PHI is not an induction variable.
626 InductionKind isInductionVariable(PHINode *Phi);
627 /// Return true if can compute the address bounds of Ptr within the loop.
628 bool hasComputableBounds(Value *Ptr);
629 /// Return true if there is the chance of write reorder.
630 bool hasPossibleGlobalWriteReorder(Value *Object,
632 AliasMultiMap &WriteObjects,
633 unsigned MaxByteWidth);
634 /// Return the AA location for a load or a store.
635 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
638 /// The loop that we evaluate.
642 /// DataLayout analysis.
647 TargetTransformInfo *TTI;
650 /// Target Library Info.
651 TargetLibraryInfo *TLI;
653 // --- vectorization state --- //
655 /// Holds the integer induction variable. This is the counter of the
658 /// Holds the reduction variables.
659 ReductionList Reductions;
660 /// Holds all of the induction variables that we found in the loop.
661 /// Notice that inductions don't need to start at zero and that induction
662 /// variables can be pointers.
663 InductionList Inductions;
664 /// Holds the widest induction type encountered.
667 /// Allowed outside users. This holds the reduction
668 /// vars which can be accessed from outside the loop.
669 SmallPtrSet<Value*, 4> AllowedExit;
670 /// This set holds the variables which are known to be uniform after
672 SmallPtrSet<Instruction*, 4> Uniforms;
673 /// We need to check that all of the pointers in this list are disjoint
675 RuntimePointerCheck PtrRtCheck;
676 /// Can we assume the absence of NaNs.
677 bool HasFunNoNaNAttr;
679 /// Utility to determine whether loads can be speculated.
680 LoadHoisting LoadSpeculation;
683 /// LoopVectorizationCostModel - estimates the expected speedups due to
685 /// In many cases vectorization is not profitable. This can happen because of
686 /// a number of reasons. In this class we mainly attempt to predict the
687 /// expected speedup/slowdowns due to the supported instruction set. We use the
688 /// TargetTransformInfo to query the different backends for the cost of
689 /// different operations.
690 class LoopVectorizationCostModel {
692 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
693 LoopVectorizationLegality *Legal,
694 const TargetTransformInfo &TTI,
695 DataLayout *DL, const TargetLibraryInfo *TLI)
696 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
698 /// Information about vectorization costs
699 struct VectorizationFactor {
700 unsigned Width; // Vector width with best cost
701 unsigned Cost; // Cost of the loop with that width
703 /// \return The most profitable vectorization factor and the cost of that VF.
704 /// This method checks every power of two up to VF. If UserVF is not ZERO
705 /// then this vectorization factor will be selected if vectorization is
707 VectorizationFactor selectVectorizationFactor(bool OptForSize,
710 /// \return The size (in bits) of the widest type in the code that
711 /// needs to be vectorized. We ignore values that remain scalar such as
712 /// 64 bit loop indices.
713 unsigned getWidestType();
715 /// \return The most profitable unroll factor.
716 /// If UserUF is non-zero then this method finds the best unroll-factor
717 /// based on register pressure and other parameters.
718 /// VF and LoopCost are the selected vectorization factor and the cost of the
720 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
723 /// \brief A struct that represents some properties of the register usage
725 struct RegisterUsage {
726 /// Holds the number of loop invariant values that are used in the loop.
727 unsigned LoopInvariantRegs;
728 /// Holds the maximum number of concurrent live intervals in the loop.
729 unsigned MaxLocalUsers;
730 /// Holds the number of instructions in the loop.
731 unsigned NumInstructions;
734 /// \return information about the register usage of the loop.
735 RegisterUsage calculateRegisterUsage();
738 /// Returns the expected execution cost. The unit of the cost does
739 /// not matter because we use the 'cost' units to compare different
740 /// vector widths. The cost that is returned is *not* normalized by
741 /// the factor width.
742 unsigned expectedCost(unsigned VF);
744 /// Returns the execution time cost of an instruction for a given vector
745 /// width. Vector width of one means scalar.
746 unsigned getInstructionCost(Instruction *I, unsigned VF);
748 /// A helper function for converting Scalar types to vector types.
749 /// If the incoming type is void, we return void. If the VF is 1, we return
751 static Type* ToVectorTy(Type *Scalar, unsigned VF);
753 /// Returns whether the instruction is a load or store and will be a emitted
754 /// as a vector operation.
755 bool isConsecutiveLoadOrStore(Instruction *I);
757 /// The loop that we evaluate.
761 /// Loop Info analysis.
763 /// Vectorization legality.
764 LoopVectorizationLegality *Legal;
765 /// Vector target information.
766 const TargetTransformInfo &TTI;
767 /// Target data layout information.
769 /// Target Library Info.
770 const TargetLibraryInfo *TLI;
773 /// Utility class for getting and setting loop vectorizer hints in the form
774 /// of loop metadata.
775 struct LoopVectorizeHints {
776 /// Vectorization width.
778 /// Vectorization unroll factor.
781 LoopVectorizeHints(const Loop *L)
782 : Width(VectorizationFactor)
783 , Unroll(VectorizationUnroll)
784 , LoopID(L->getLoopID()) {
786 // The command line options override any loop metadata except for when
787 // width == 1 which is used to indicate the loop is already vectorized.
788 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
789 Width = VectorizationFactor;
790 if (VectorizationUnroll.getNumOccurrences() > 0)
791 Unroll = VectorizationUnroll;
794 /// Return the loop vectorizer metadata prefix.
795 static StringRef Prefix() { return "llvm.vectorizer."; }
797 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
798 SmallVector<Value*, 2> Vals;
799 Vals.push_back(MDString::get(Context, Name));
800 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
801 return MDNode::get(Context, Vals);
804 /// Mark the loop L as already vectorized by setting the width to 1.
805 void setAlreadyVectorized(Loop *L) {
806 LLVMContext &Context = L->getHeader()->getContext();
810 // Create a new loop id with one more operand for the already_vectorized
811 // hint. If the loop already has a loop id then copy the existing operands.
812 SmallVector<Value*, 4> Vals(1);
814 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
815 Vals.push_back(LoopID->getOperand(i));
817 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
819 MDNode *NewLoopID = MDNode::get(Context, Vals);
820 // Set operand 0 to refer to the loop id itself.
821 NewLoopID->replaceOperandWith(0, NewLoopID);
823 L->setLoopID(NewLoopID);
825 LoopID->replaceAllUsesWith(NewLoopID);
833 /// Find hints specified in the loop metadata.
834 void getHints(const Loop *L) {
838 // First operand should refer to the loop id itself.
839 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
840 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
842 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
843 const MDString *S = 0;
844 SmallVector<Value*, 4> Args;
846 // The expected hint is either a MDString or a MDNode with the first
847 // operand a MDString.
848 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
849 if (!MD || MD->getNumOperands() == 0)
851 S = dyn_cast<MDString>(MD->getOperand(0));
852 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
853 Args.push_back(MD->getOperand(i));
855 S = dyn_cast<MDString>(LoopID->getOperand(i));
856 assert(Args.size() == 0 && "too many arguments for MDString");
862 // Check if the hint starts with the vectorizer prefix.
863 StringRef Hint = S->getString();
864 if (!Hint.startswith(Prefix()))
866 // Remove the prefix.
867 Hint = Hint.substr(Prefix().size(), StringRef::npos);
869 if (Args.size() == 1)
870 getHint(Hint, Args[0]);
874 // Check string hint with one operand.
875 void getHint(StringRef Hint, Value *Arg) {
876 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
878 unsigned Val = C->getZExtValue();
880 if (Hint == "width") {
881 assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
882 "Invalid width metadata");
884 } else if (Hint == "unroll") {
885 assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
886 "Invalid unroll metadata");
889 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
893 /// The LoopVectorize Pass.
894 struct LoopVectorize : public LoopPass {
895 /// Pass identification, replacement for typeid
898 explicit LoopVectorize() : LoopPass(ID) {
899 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
905 TargetTransformInfo *TTI;
908 TargetLibraryInfo *TLI;
910 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
911 // We only vectorize innermost loops.
915 SE = &getAnalysis<ScalarEvolution>();
916 DL = getAnalysisIfAvailable<DataLayout>();
917 LI = &getAnalysis<LoopInfo>();
918 TTI = &getAnalysis<TargetTransformInfo>();
919 DT = &getAnalysis<DominatorTree>();
920 AA = getAnalysisIfAvailable<AliasAnalysis>();
921 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
924 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
928 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
929 L->getHeader()->getParent()->getName() << "\"\n");
931 LoopVectorizeHints Hints(L);
933 if (Hints.Width == 1) {
934 DEBUG(dbgs() << "LV: Not vectorizing.\n");
938 // Check if it is legal to vectorize the loop.
939 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
940 if (!LVL.canVectorize()) {
941 DEBUG(dbgs() << "LV: Not vectorizing.\n");
945 // Use the cost model.
946 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
948 // Check the function attributes to find out if this function should be
949 // optimized for size.
950 Function *F = L->getHeader()->getParent();
951 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
952 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
953 unsigned FnIndex = AttributeSet::FunctionIndex;
954 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
955 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
958 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
959 "attribute is used.\n");
963 // Select the optimal vectorization factor.
964 LoopVectorizationCostModel::VectorizationFactor VF;
965 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
966 // Select the unroll factor.
967 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
971 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
975 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
976 F->getParent()->getModuleIdentifier()<<"\n");
977 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
979 // If we decided that it is *legal* to vectorize the loop then do it.
980 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
983 // Mark the loop as already vectorized to avoid vectorizing again.
984 Hints.setAlreadyVectorized(L);
986 DEBUG(verifyFunction(*L->getHeader()->getParent()));
990 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
991 LoopPass::getAnalysisUsage(AU);
992 AU.addRequiredID(LoopSimplifyID);
993 AU.addRequiredID(LCSSAID);
994 AU.addRequired<DominatorTree>();
995 AU.addRequired<LoopInfo>();
996 AU.addRequired<ScalarEvolution>();
997 AU.addRequired<TargetTransformInfo>();
998 AU.addPreserved<LoopInfo>();
999 AU.addPreserved<DominatorTree>();
1004 } // end anonymous namespace
1006 //===----------------------------------------------------------------------===//
1007 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1008 // LoopVectorizationCostModel.
1009 //===----------------------------------------------------------------------===//
1012 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
1013 Loop *Lp, Value *Ptr,
1015 const SCEV *Sc = SE->getSCEV(Ptr);
1016 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1017 assert(AR && "Invalid addrec expression");
1018 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1019 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1020 Pointers.push_back(Ptr);
1021 Starts.push_back(AR->getStart());
1022 Ends.push_back(ScEnd);
1023 IsWritePtr.push_back(WritePtr);
1026 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1027 // Save the current insertion location.
1028 Instruction *Loc = Builder.GetInsertPoint();
1030 // We need to place the broadcast of invariant variables outside the loop.
1031 Instruction *Instr = dyn_cast<Instruction>(V);
1032 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1033 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1035 // Place the code for broadcasting invariant variables in the new preheader.
1037 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1039 // Broadcast the scalar into all locations in the vector.
1040 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1042 // Restore the builder insertion point.
1044 Builder.SetInsertPoint(Loc);
1049 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1051 assert(Val->getType()->isVectorTy() && "Must be a vector");
1052 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1053 "Elem must be an integer");
1054 // Create the types.
1055 Type *ITy = Val->getType()->getScalarType();
1056 VectorType *Ty = cast<VectorType>(Val->getType());
1057 int VLen = Ty->getNumElements();
1058 SmallVector<Constant*, 8> Indices;
1060 // Create a vector of consecutive numbers from zero to VF.
1061 for (int i = 0; i < VLen; ++i) {
1062 int64_t Idx = Negate ? (-i) : i;
1063 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1066 // Add the consecutive indices to the vector value.
1067 Constant *Cv = ConstantVector::get(Indices);
1068 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1069 return Builder.CreateAdd(Val, Cv, "induction");
1072 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1073 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
1074 // Make sure that the pointer does not point to structs.
1075 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
1078 // If this value is a pointer induction variable we know it is consecutive.
1079 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1080 if (Phi && Inductions.count(Phi)) {
1081 InductionInfo II = Inductions[Phi];
1082 if (IK_PtrInduction == II.IK)
1084 else if (IK_ReversePtrInduction == II.IK)
1088 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1092 unsigned NumOperands = Gep->getNumOperands();
1093 Value *LastIndex = Gep->getOperand(NumOperands - 1);
1095 Value *GpPtr = Gep->getPointerOperand();
1096 // If this GEP value is a consecutive pointer induction variable and all of
1097 // the indices are constant then we know it is consecutive. We can
1098 Phi = dyn_cast<PHINode>(GpPtr);
1099 if (Phi && Inductions.count(Phi)) {
1101 // Make sure that the pointer does not point to structs.
1102 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1103 if (GepPtrType->getElementType()->isAggregateType())
1106 // Make sure that all of the index operands are loop invariant.
1107 for (unsigned i = 1; i < NumOperands; ++i)
1108 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1111 InductionInfo II = Inductions[Phi];
1112 if (IK_PtrInduction == II.IK)
1114 else if (IK_ReversePtrInduction == II.IK)
1118 // Check that all of the gep indices are uniform except for the last.
1119 for (unsigned i = 0; i < NumOperands - 1; ++i)
1120 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1123 // We can emit wide load/stores only if the last index is the induction
1125 const SCEV *Last = SE->getSCEV(LastIndex);
1126 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1127 const SCEV *Step = AR->getStepRecurrence(*SE);
1129 // The memory is consecutive because the last index is consecutive
1130 // and all other indices are loop invariant.
1133 if (Step->isAllOnesValue())
1140 bool LoopVectorizationLegality::isUniform(Value *V) {
1141 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1144 InnerLoopVectorizer::VectorParts&
1145 InnerLoopVectorizer::getVectorValue(Value *V) {
1146 assert(V != Induction && "The new induction variable should not be used.");
1147 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1149 // If we have this scalar in the map, return it.
1150 if (WidenMap.has(V))
1151 return WidenMap.get(V);
1153 // If this scalar is unknown, assume that it is a constant or that it is
1154 // loop invariant. Broadcast V and save the value for future uses.
1155 Value *B = getBroadcastInstrs(V);
1156 return WidenMap.splat(V, B);
1159 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1160 assert(Vec->getType()->isVectorTy() && "Invalid type");
1161 SmallVector<Constant*, 8> ShuffleMask;
1162 for (unsigned i = 0; i < VF; ++i)
1163 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1165 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1166 ConstantVector::get(ShuffleMask),
1171 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
1172 LoopVectorizationLegality *Legal) {
1173 // Attempt to issue a wide load.
1174 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1175 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1177 assert((LI || SI) && "Invalid Load/Store instruction");
1179 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1180 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1181 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1182 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1183 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1184 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1185 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1187 if (ScalarAllocatedSize != VectorElementSize)
1188 return scalarizeInstruction(Instr);
1190 // If the pointer is loop invariant or if it is non consecutive,
1191 // scalarize the load.
1192 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1193 bool Reverse = ConsecutiveStride < 0;
1194 bool UniformLoad = LI && Legal->isUniform(Ptr);
1195 if (!ConsecutiveStride || UniformLoad)
1196 return scalarizeInstruction(Instr);
1198 Constant *Zero = Builder.getInt32(0);
1199 VectorParts &Entry = WidenMap.get(Instr);
1201 // Handle consecutive loads/stores.
1202 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1203 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1204 Value *PtrOperand = Gep->getPointerOperand();
1205 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1206 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1208 // Create the new GEP with the new induction variable.
1209 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1210 Gep2->setOperand(0, FirstBasePtr);
1211 Gep2->setName("gep.indvar.base");
1212 Ptr = Builder.Insert(Gep2);
1214 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1215 OrigLoop) && "Base ptr must be invariant");
1217 // The last index does not have to be the induction. It can be
1218 // consecutive and be a function of the index. For example A[I+1];
1219 unsigned NumOperands = Gep->getNumOperands();
1221 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1222 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1223 Value *LastIndex = GEPParts[0];
1224 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1226 // Create the new GEP with the new induction variable.
1227 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1228 Gep2->setOperand(NumOperands - 1, LastIndex);
1229 Gep2->setName("gep.indvar.idx");
1230 Ptr = Builder.Insert(Gep2);
1232 // Use the induction element ptr.
1233 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1234 VectorParts &PtrVal = getVectorValue(Ptr);
1235 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1240 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1241 "We do not allow storing to uniform addresses");
1243 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1244 for (unsigned Part = 0; Part < UF; ++Part) {
1245 // Calculate the pointer for the specific unroll-part.
1246 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1249 // If we store to reverse consecutive memory locations then we need
1250 // to reverse the order of elements in the stored value.
1251 StoredVal[Part] = reverseVector(StoredVal[Part]);
1252 // If the address is consecutive but reversed, then the
1253 // wide store needs to start at the last vector element.
1254 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1255 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1258 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1259 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1263 for (unsigned Part = 0; Part < UF; ++Part) {
1264 // Calculate the pointer for the specific unroll-part.
1265 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1268 // If the address is consecutive but reversed, then the
1269 // wide store needs to start at the last vector element.
1270 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1271 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1274 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
1275 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1276 cast<LoadInst>(LI)->setAlignment(Alignment);
1277 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1281 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1282 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1283 // Holds vector parameters or scalars, in case of uniform vals.
1284 SmallVector<VectorParts, 4> Params;
1286 // Find all of the vectorized parameters.
1287 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1288 Value *SrcOp = Instr->getOperand(op);
1290 // If we are accessing the old induction variable, use the new one.
1291 if (SrcOp == OldInduction) {
1292 Params.push_back(getVectorValue(SrcOp));
1296 // Try using previously calculated values.
1297 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1299 // If the src is an instruction that appeared earlier in the basic block
1300 // then it should already be vectorized.
1301 if (SrcInst && OrigLoop->contains(SrcInst)) {
1302 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1303 // The parameter is a vector value from earlier.
1304 Params.push_back(WidenMap.get(SrcInst));
1306 // The parameter is a scalar from outside the loop. Maybe even a constant.
1307 VectorParts Scalars;
1308 Scalars.append(UF, SrcOp);
1309 Params.push_back(Scalars);
1313 assert(Params.size() == Instr->getNumOperands() &&
1314 "Invalid number of operands");
1316 // Does this instruction return a value ?
1317 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1319 Value *UndefVec = IsVoidRetTy ? 0 :
1320 UndefValue::get(VectorType::get(Instr->getType(), VF));
1321 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1322 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1324 // For each vector unroll 'part':
1325 for (unsigned Part = 0; Part < UF; ++Part) {
1326 // For each scalar that we create:
1327 for (unsigned Width = 0; Width < VF; ++Width) {
1328 Instruction *Cloned = Instr->clone();
1330 Cloned->setName(Instr->getName() + ".cloned");
1331 // Replace the operands of the cloned instrucions with extracted scalars.
1332 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1333 Value *Op = Params[op][Part];
1334 // Param is a vector. Need to extract the right lane.
1335 if (Op->getType()->isVectorTy())
1336 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1337 Cloned->setOperand(op, Op);
1340 // Place the cloned scalar in the new loop.
1341 Builder.Insert(Cloned);
1343 // If the original scalar returns a value we need to place it in a vector
1344 // so that future users will be able to use it.
1346 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1347 Builder.getInt32(Width));
1353 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1355 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1356 Legal->getRuntimePointerCheck();
1358 if (!PtrRtCheck->Need)
1361 Instruction *MemoryRuntimeCheck = 0;
1362 unsigned NumPointers = PtrRtCheck->Pointers.size();
1363 SmallVector<Value* , 2> Starts;
1364 SmallVector<Value* , 2> Ends;
1366 SCEVExpander Exp(*SE, "induction");
1368 // Use this type for pointer arithmetic.
1369 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1371 for (unsigned i = 0; i < NumPointers; ++i) {
1372 Value *Ptr = PtrRtCheck->Pointers[i];
1373 const SCEV *Sc = SE->getSCEV(Ptr);
1375 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1376 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1378 Starts.push_back(Ptr);
1379 Ends.push_back(Ptr);
1381 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1383 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1384 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1385 Starts.push_back(Start);
1386 Ends.push_back(End);
1390 IRBuilder<> ChkBuilder(Loc);
1392 for (unsigned i = 0; i < NumPointers; ++i) {
1393 for (unsigned j = i+1; j < NumPointers; ++j) {
1394 // No need to check if two readonly pointers intersect.
1395 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1398 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1399 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1400 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1401 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1403 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1404 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1405 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1406 if (MemoryRuntimeCheck)
1407 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1410 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1414 return MemoryRuntimeCheck;
1418 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1420 In this function we generate a new loop. The new loop will contain
1421 the vectorized instructions while the old loop will continue to run the
1424 [ ] <-- vector loop bypass (may consist of multiple blocks).
1427 | [ ] <-- vector pre header.
1431 | [ ]_| <-- vector loop.
1434 >[ ] <--- middle-block.
1437 | [ ] <--- new preheader.
1441 | [ ]_| <-- old scalar loop to handle remainder.
1444 >[ ] <-- exit block.
1448 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1449 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1450 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1451 assert(ExitBlock && "Must have an exit block");
1453 // Some loops have a single integer induction variable, while other loops
1454 // don't. One example is c++ iterators that often have multiple pointer
1455 // induction variables. In the code below we also support a case where we
1456 // don't have a single induction variable.
1457 OldInduction = Legal->getInduction();
1458 Type *IdxTy = Legal->getWidestInductionType();
1460 // Find the loop boundaries.
1461 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1462 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1464 // Get the total trip count from the count by adding 1.
1465 ExitCount = SE->getAddExpr(ExitCount,
1466 SE->getConstant(ExitCount->getType(), 1));
1468 // Expand the trip count and place the new instructions in the preheader.
1469 // Notice that the pre-header does not change, only the loop body.
1470 SCEVExpander Exp(*SE, "induction");
1472 // Count holds the overall loop count (N).
1473 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1474 BypassBlock->getTerminator());
1476 // The loop index does not have to start at Zero. Find the original start
1477 // value from the induction PHI node. If we don't have an induction variable
1478 // then we know that it starts at zero.
1479 Builder.SetInsertPoint(BypassBlock->getTerminator());
1480 Value *StartIdx = ExtendedIdx = OldInduction ?
1481 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1483 ConstantInt::get(IdxTy, 0);
1485 assert(BypassBlock && "Invalid loop structure");
1486 LoopBypassBlocks.push_back(BypassBlock);
1488 // Split the single block loop into the two loop structure described above.
1489 BasicBlock *VectorPH =
1490 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1491 BasicBlock *VecBody =
1492 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1493 BasicBlock *MiddleBlock =
1494 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1495 BasicBlock *ScalarPH =
1496 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1498 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1500 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1502 // Generate the induction variable.
1503 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1504 // The loop step is equal to the vectorization factor (num of SIMD elements)
1505 // times the unroll factor (num of SIMD instructions).
1506 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1508 // This is the IR builder that we use to add all of the logic for bypassing
1509 // the new vector loop.
1510 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1512 // We may need to extend the index in case there is a type mismatch.
1513 // We know that the count starts at zero and does not overflow.
1514 if (Count->getType() != IdxTy) {
1515 // The exit count can be of pointer type. Convert it to the correct
1517 if (ExitCount->getType()->isPointerTy())
1518 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1520 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1523 // Add the start index to the loop count to get the new end index.
1524 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1526 // Now we need to generate the expression for N - (N % VF), which is
1527 // the part that the vectorized body will execute.
1528 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1529 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1530 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1531 "end.idx.rnd.down");
1533 // Now, compare the new count to zero. If it is zero skip the vector loop and
1534 // jump to the scalar loop.
1535 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1538 BasicBlock *LastBypassBlock = BypassBlock;
1540 // Generate the code that checks in runtime if arrays overlap. We put the
1541 // checks into a separate block to make the more common case of few elements
1543 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1544 BypassBlock->getTerminator());
1545 if (MemRuntimeCheck) {
1546 // Create a new block containing the memory check.
1547 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1549 LoopBypassBlocks.push_back(CheckBlock);
1551 // Replace the branch into the memory check block with a conditional branch
1552 // for the "few elements case".
1553 Instruction *OldTerm = BypassBlock->getTerminator();
1554 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1555 OldTerm->eraseFromParent();
1557 Cmp = MemRuntimeCheck;
1558 LastBypassBlock = CheckBlock;
1561 LastBypassBlock->getTerminator()->eraseFromParent();
1562 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1565 // We are going to resume the execution of the scalar loop.
1566 // Go over all of the induction variables that we found and fix the
1567 // PHIs that are left in the scalar version of the loop.
1568 // The starting values of PHI nodes depend on the counter of the last
1569 // iteration in the vectorized loop.
1570 // If we come from a bypass edge then we need to start from the original
1573 // This variable saves the new starting index for the scalar loop.
1574 PHINode *ResumeIndex = 0;
1575 LoopVectorizationLegality::InductionList::iterator I, E;
1576 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1577 // Set builder to point to last bypass block.
1578 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
1579 for (I = List->begin(), E = List->end(); I != E; ++I) {
1580 PHINode *OrigPhi = I->first;
1581 LoopVectorizationLegality::InductionInfo II = I->second;
1583 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
1584 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
1585 MiddleBlock->getTerminator());
1586 // We might have extended the type of the induction variable but we need a
1587 // truncated version for the scalar loop.
1588 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
1589 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
1590 MiddleBlock->getTerminator()) : 0;
1592 Value *EndValue = 0;
1594 case LoopVectorizationLegality::IK_NoInduction:
1595 llvm_unreachable("Unknown induction");
1596 case LoopVectorizationLegality::IK_IntInduction: {
1597 // Handle the integer induction counter.
1598 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1600 // We have the canonical induction variable.
1601 if (OrigPhi == OldInduction) {
1602 // Create a truncated version of the resume value for the scalar loop,
1603 // we might have promoted the type to a larger width.
1605 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
1606 // The new PHI merges the original incoming value, in case of a bypass,
1607 // or the value at the end of the vectorized loop.
1608 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1609 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1610 TruncResumeVal->addIncoming(EndValue, VecBody);
1612 // We know what the end value is.
1613 EndValue = IdxEndRoundDown;
1614 // We also know which PHI node holds it.
1615 ResumeIndex = ResumeVal;
1619 // Not the canonical induction variable - add the vector loop count to the
1621 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1622 II.StartValue->getType(),
1624 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
1627 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1628 // Convert the CountRoundDown variable to the PHI size.
1629 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
1630 II.StartValue->getType(),
1632 // Handle reverse integer induction counter.
1633 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
1636 case LoopVectorizationLegality::IK_PtrInduction: {
1637 // For pointer induction variables, calculate the offset using
1639 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
1643 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1644 // The value at the end of the loop for the reverse pointer is calculated
1645 // by creating a GEP with a negative index starting from the start value.
1646 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1647 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
1649 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
1655 // The new PHI merges the original incoming value, in case of a bypass,
1656 // or the value at the end of the vectorized loop.
1657 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
1658 if (OrigPhi == OldInduction)
1659 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
1661 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1663 ResumeVal->addIncoming(EndValue, VecBody);
1665 // Fix the scalar body counter (PHI node).
1666 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1667 // The old inductions phi node in the scalar body needs the truncated value.
1668 if (OrigPhi == OldInduction)
1669 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
1671 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1674 // If we are generating a new induction variable then we also need to
1675 // generate the code that calculates the exit value. This value is not
1676 // simply the end of the counter because we may skip the vectorized body
1677 // in case of a runtime check.
1679 assert(!ResumeIndex && "Unexpected resume value found");
1680 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1681 MiddleBlock->getTerminator());
1682 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1683 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1684 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1687 // Make sure that we found the index where scalar loop needs to continue.
1688 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1689 "Invalid resume Index");
1691 // Add a check in the middle block to see if we have completed
1692 // all of the iterations in the first vector loop.
1693 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1694 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1695 ResumeIndex, "cmp.n",
1696 MiddleBlock->getTerminator());
1698 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1699 // Remove the old terminator.
1700 MiddleBlock->getTerminator()->eraseFromParent();
1702 // Create i+1 and fill the PHINode.
1703 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1704 Induction->addIncoming(StartIdx, VectorPH);
1705 Induction->addIncoming(NextIdx, VecBody);
1706 // Create the compare.
1707 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1708 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1710 // Now we have two terminators. Remove the old one from the block.
1711 VecBody->getTerminator()->eraseFromParent();
1713 // Get ready to start creating new instructions into the vectorized body.
1714 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1716 // Create and register the new vector loop.
1717 Loop* Lp = new Loop();
1718 Loop *ParentLoop = OrigLoop->getParentLoop();
1720 // Insert the new loop into the loop nest and register the new basic blocks.
1722 ParentLoop->addChildLoop(Lp);
1723 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1724 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1725 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1726 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1727 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1729 LI->addTopLevelLoop(Lp);
1732 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1735 LoopVectorPreHeader = VectorPH;
1736 LoopScalarPreHeader = ScalarPH;
1737 LoopMiddleBlock = MiddleBlock;
1738 LoopExitBlock = ExitBlock;
1739 LoopVectorBody = VecBody;
1740 LoopScalarBody = OldBasicBlock;
1743 /// This function returns the identity element (or neutral element) for
1744 /// the operation K.
1746 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1751 // Adding, Xoring, Oring zero to a number does not change it.
1752 return ConstantInt::get(Tp, 0);
1753 case RK_IntegerMult:
1754 // Multiplying a number by 1 does not change it.
1755 return ConstantInt::get(Tp, 1);
1757 // AND-ing a number with an all-1 value does not change it.
1758 return ConstantInt::get(Tp, -1, true);
1760 // Multiplying a number by 1 does not change it.
1761 return ConstantFP::get(Tp, 1.0L);
1763 // Adding zero to a number does not change it.
1764 return ConstantFP::get(Tp, 0.0L);
1766 llvm_unreachable("Unknown reduction kind");
1770 static Intrinsic::ID
1771 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1772 // If we have an intrinsic call, check if it is trivially vectorizable.
1773 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1774 switch (II->getIntrinsicID()) {
1775 case Intrinsic::sqrt:
1776 case Intrinsic::sin:
1777 case Intrinsic::cos:
1778 case Intrinsic::exp:
1779 case Intrinsic::exp2:
1780 case Intrinsic::log:
1781 case Intrinsic::log10:
1782 case Intrinsic::log2:
1783 case Intrinsic::fabs:
1784 case Intrinsic::floor:
1785 case Intrinsic::ceil:
1786 case Intrinsic::trunc:
1787 case Intrinsic::rint:
1788 case Intrinsic::nearbyint:
1789 case Intrinsic::pow:
1790 case Intrinsic::fma:
1791 case Intrinsic::fmuladd:
1792 return II->getIntrinsicID();
1794 return Intrinsic::not_intrinsic;
1799 return Intrinsic::not_intrinsic;
1802 Function *F = CI->getCalledFunction();
1803 // We're going to make assumptions on the semantics of the functions, check
1804 // that the target knows that it's available in this environment.
1805 if (!F || !TLI->getLibFunc(F->getName(), Func))
1806 return Intrinsic::not_intrinsic;
1808 // Otherwise check if we have a call to a function that can be turned into a
1809 // vector intrinsic.
1816 return Intrinsic::sin;
1820 return Intrinsic::cos;
1824 return Intrinsic::exp;
1826 case LibFunc::exp2f:
1827 case LibFunc::exp2l:
1828 return Intrinsic::exp2;
1832 return Intrinsic::log;
1833 case LibFunc::log10:
1834 case LibFunc::log10f:
1835 case LibFunc::log10l:
1836 return Intrinsic::log10;
1838 case LibFunc::log2f:
1839 case LibFunc::log2l:
1840 return Intrinsic::log2;
1842 case LibFunc::fabsf:
1843 case LibFunc::fabsl:
1844 return Intrinsic::fabs;
1845 case LibFunc::floor:
1846 case LibFunc::floorf:
1847 case LibFunc::floorl:
1848 return Intrinsic::floor;
1850 case LibFunc::ceilf:
1851 case LibFunc::ceill:
1852 return Intrinsic::ceil;
1853 case LibFunc::trunc:
1854 case LibFunc::truncf:
1855 case LibFunc::truncl:
1856 return Intrinsic::trunc;
1858 case LibFunc::rintf:
1859 case LibFunc::rintl:
1860 return Intrinsic::rint;
1861 case LibFunc::nearbyint:
1862 case LibFunc::nearbyintf:
1863 case LibFunc::nearbyintl:
1864 return Intrinsic::nearbyint;
1868 return Intrinsic::pow;
1871 return Intrinsic::not_intrinsic;
1874 /// This function translates the reduction kind to an LLVM binary operator.
1876 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1878 case LoopVectorizationLegality::RK_IntegerAdd:
1879 return Instruction::Add;
1880 case LoopVectorizationLegality::RK_IntegerMult:
1881 return Instruction::Mul;
1882 case LoopVectorizationLegality::RK_IntegerOr:
1883 return Instruction::Or;
1884 case LoopVectorizationLegality::RK_IntegerAnd:
1885 return Instruction::And;
1886 case LoopVectorizationLegality::RK_IntegerXor:
1887 return Instruction::Xor;
1888 case LoopVectorizationLegality::RK_FloatMult:
1889 return Instruction::FMul;
1890 case LoopVectorizationLegality::RK_FloatAdd:
1891 return Instruction::FAdd;
1892 case LoopVectorizationLegality::RK_IntegerMinMax:
1893 return Instruction::ICmp;
1894 case LoopVectorizationLegality::RK_FloatMinMax:
1895 return Instruction::FCmp;
1897 llvm_unreachable("Unknown reduction operation");
1901 Value *createMinMaxOp(IRBuilder<> &Builder,
1902 LoopVectorizationLegality::MinMaxReductionKind RK,
1905 CmpInst::Predicate P = CmpInst::ICMP_NE;
1908 llvm_unreachable("Unknown min/max reduction kind");
1909 case LoopVectorizationLegality::MRK_UIntMin:
1910 P = CmpInst::ICMP_ULT;
1912 case LoopVectorizationLegality::MRK_UIntMax:
1913 P = CmpInst::ICMP_UGT;
1915 case LoopVectorizationLegality::MRK_SIntMin:
1916 P = CmpInst::ICMP_SLT;
1918 case LoopVectorizationLegality::MRK_SIntMax:
1919 P = CmpInst::ICMP_SGT;
1921 case LoopVectorizationLegality::MRK_FloatMin:
1922 P = CmpInst::FCMP_OLT;
1924 case LoopVectorizationLegality::MRK_FloatMax:
1925 P = CmpInst::FCMP_OGT;
1930 if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
1931 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
1933 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1935 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1940 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1941 //===------------------------------------------------===//
1943 // Notice: any optimization or new instruction that go
1944 // into the code below should be also be implemented in
1947 //===------------------------------------------------===//
1948 Constant *Zero = Builder.getInt32(0);
1950 // In order to support reduction variables we need to be able to vectorize
1951 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1952 // stages. First, we create a new vector PHI node with no incoming edges.
1953 // We use this value when we vectorize all of the instructions that use the
1954 // PHI. Next, after all of the instructions in the block are complete we
1955 // add the new incoming edges to the PHI. At this point all of the
1956 // instructions in the basic block are vectorized, so we can use them to
1957 // construct the PHI.
1958 PhiVector RdxPHIsToFix;
1960 // Scan the loop in a topological order to ensure that defs are vectorized
1962 LoopBlocksDFS DFS(OrigLoop);
1965 // Vectorize all of the blocks in the original loop.
1966 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1967 be = DFS.endRPO(); bb != be; ++bb)
1968 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1970 // At this point every instruction in the original loop is widened to
1971 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1972 // that we vectorized. The PHI nodes are currently empty because we did
1973 // not want to introduce cycles. Notice that the remaining PHI nodes
1974 // that we need to fix are reduction variables.
1976 // Create the 'reduced' values for each of the induction vars.
1977 // The reduced values are the vector values that we scalarize and combine
1978 // after the loop is finished.
1979 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1981 PHINode *RdxPhi = *it;
1982 assert(RdxPhi && "Unable to recover vectorized PHI");
1984 // Find the reduction variable descriptor.
1985 assert(Legal->getReductionVars()->count(RdxPhi) &&
1986 "Unable to find the reduction variable");
1987 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1988 (*Legal->getReductionVars())[RdxPhi];
1990 // We need to generate a reduction vector from the incoming scalar.
1991 // To do so, we need to generate the 'identity' vector and overide
1992 // one of the elements with the incoming scalar reduction. We need
1993 // to do it in the vector-loop preheader.
1994 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1996 // This is the vector-clone of the value that leaves the loop.
1997 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1998 Type *VecTy = VectorExit[0]->getType();
2000 // Find the reduction identity variable. Zero for addition, or, xor,
2001 // one for multiplication, -1 for And.
2004 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2005 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2006 // MinMax reduction have the start value as their identify.
2007 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
2011 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2012 VecTy->getScalarType());
2013 Identity = ConstantVector::getSplat(VF, Iden);
2015 // This vector is the Identity vector where the first element is the
2016 // incoming scalar reduction.
2017 VectorStart = Builder.CreateInsertElement(Identity,
2018 RdxDesc.StartValue, Zero);
2021 // Fix the vector-loop phi.
2022 // We created the induction variable so we know that the
2023 // preheader is the first entry.
2024 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2026 // Reductions do not have to start at zero. They can start with
2027 // any loop invariant values.
2028 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2029 BasicBlock *Latch = OrigLoop->getLoopLatch();
2030 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2031 VectorParts &Val = getVectorValue(LoopVal);
2032 for (unsigned part = 0; part < UF; ++part) {
2033 // Make sure to add the reduction stat value only to the
2034 // first unroll part.
2035 Value *StartVal = (part == 0) ? VectorStart : Identity;
2036 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2037 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
2040 // Before each round, move the insertion point right between
2041 // the PHIs and the values we are going to write.
2042 // This allows us to write both PHINodes and the extractelement
2044 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2046 VectorParts RdxParts;
2047 for (unsigned part = 0; part < UF; ++part) {
2048 // This PHINode contains the vectorized reduction variable, or
2049 // the initial value vector, if we bypass the vector loop.
2050 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2051 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2052 Value *StartVal = (part == 0) ? VectorStart : Identity;
2053 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2054 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2055 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
2056 RdxParts.push_back(NewPhi);
2059 // Reduce all of the unrolled parts into a single vector.
2060 Value *ReducedPartRdx = RdxParts[0];
2061 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2062 for (unsigned part = 1; part < UF; ++part) {
2063 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2064 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2065 RdxParts[part], ReducedPartRdx,
2068 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2069 ReducedPartRdx, RdxParts[part]);
2072 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2073 // and vector ops, reducing the set of values being computed by half each
2075 assert(isPowerOf2_32(VF) &&
2076 "Reduction emission only supported for pow2 vectors!");
2077 Value *TmpVec = ReducedPartRdx;
2078 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2079 for (unsigned i = VF; i != 1; i >>= 1) {
2080 // Move the upper half of the vector to the lower half.
2081 for (unsigned j = 0; j != i/2; ++j)
2082 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2084 // Fill the rest of the mask with undef.
2085 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2086 UndefValue::get(Builder.getInt32Ty()));
2089 Builder.CreateShuffleVector(TmpVec,
2090 UndefValue::get(TmpVec->getType()),
2091 ConstantVector::get(ShuffleMask),
2094 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2095 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2098 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2101 // The result is in the first element of the vector.
2102 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2104 // Now, we need to fix the users of the reduction variable
2105 // inside and outside of the scalar remainder loop.
2106 // We know that the loop is in LCSSA form. We need to update the
2107 // PHI nodes in the exit blocks.
2108 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2109 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2110 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2111 if (!LCSSAPhi) continue;
2113 // All PHINodes need to have a single entry edge, or two if
2114 // we already fixed them.
2115 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2117 // We found our reduction value exit-PHI. Update it with the
2118 // incoming bypass edge.
2119 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2120 // Add an edge coming from the bypass.
2121 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
2124 }// end of the LCSSA phi scan.
2126 // Fix the scalar loop reduction variable with the incoming reduction sum
2127 // from the vector body and from the backedge value.
2128 int IncomingEdgeBlockIdx =
2129 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2130 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2131 // Pick the other block.
2132 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2133 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
2134 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2135 }// end of for each redux variable.
2137 // The Loop exit block may have single value PHI nodes where the incoming
2138 // value is 'undef'. While vectorizing we only handled real values that
2139 // were defined inside the loop. Here we handle the 'undef case'.
2141 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2142 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2143 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2144 if (!LCSSAPhi) continue;
2145 if (LCSSAPhi->getNumIncomingValues() == 1)
2146 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2151 InnerLoopVectorizer::VectorParts
2152 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2153 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2156 VectorParts SrcMask = createBlockInMask(Src);
2158 // The terminator has to be a branch inst!
2159 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2160 assert(BI && "Unexpected terminator found");
2162 if (BI->isConditional()) {
2163 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2165 if (BI->getSuccessor(0) != Dst)
2166 for (unsigned part = 0; part < UF; ++part)
2167 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2169 for (unsigned part = 0; part < UF; ++part)
2170 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2177 InnerLoopVectorizer::VectorParts
2178 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2179 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2181 // Loop incoming mask is all-one.
2182 if (OrigLoop->getHeader() == BB) {
2183 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2184 return getVectorValue(C);
2187 // This is the block mask. We OR all incoming edges, and with zero.
2188 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2189 VectorParts BlockMask = getVectorValue(Zero);
2192 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2193 VectorParts EM = createEdgeMask(*it, BB);
2194 for (unsigned part = 0; part < UF; ++part)
2195 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2202 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
2203 BasicBlock *BB, PhiVector *PV) {
2204 // For each instruction in the old loop.
2205 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2206 VectorParts &Entry = WidenMap.get(it);
2207 switch (it->getOpcode()) {
2208 case Instruction::Br:
2209 // Nothing to do for PHIs and BR, since we already took care of the
2210 // loop control flow instructions.
2212 case Instruction::PHI:{
2213 PHINode* P = cast<PHINode>(it);
2214 // Handle reduction variables:
2215 if (Legal->getReductionVars()->count(P)) {
2216 for (unsigned part = 0; part < UF; ++part) {
2217 // This is phase one of vectorizing PHIs.
2218 Type *VecTy = VectorType::get(it->getType(), VF);
2219 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2220 LoopVectorBody-> getFirstInsertionPt());
2226 // Check for PHI nodes that are lowered to vector selects.
2227 if (P->getParent() != OrigLoop->getHeader()) {
2228 // We know that all PHIs in non header blocks are converted into
2229 // selects, so we don't have to worry about the insertion order and we
2230 // can just use the builder.
2231 // At this point we generate the predication tree. There may be
2232 // duplications since this is a simple recursive scan, but future
2233 // optimizations will clean it up.
2235 unsigned NumIncoming = P->getNumIncomingValues();
2237 // Generate a sequence of selects of the form:
2238 // SELECT(Mask3, In3,
2239 // SELECT(Mask2, In2,
2241 for (unsigned In = 0; In < NumIncoming; In++) {
2242 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2244 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2246 for (unsigned part = 0; part < UF; ++part) {
2247 // We might have single edge PHIs (blocks) - use an identity
2248 // 'select' for the first PHI operand.
2250 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2253 // Select between the current value and the previous incoming edge
2254 // based on the incoming mask.
2255 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2256 Entry[part], "predphi");
2262 // This PHINode must be an induction variable.
2263 // Make sure that we know about it.
2264 assert(Legal->getInductionVars()->count(P) &&
2265 "Not an induction variable");
2267 LoopVectorizationLegality::InductionInfo II =
2268 Legal->getInductionVars()->lookup(P);
2271 case LoopVectorizationLegality::IK_NoInduction:
2272 llvm_unreachable("Unknown induction");
2273 case LoopVectorizationLegality::IK_IntInduction: {
2274 assert(P->getType() == II.StartValue->getType() && "Types must match");
2275 Type *PhiTy = P->getType();
2277 if (P == OldInduction) {
2278 // Handle the canonical induction variable. We might have had to
2280 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2282 // Handle other induction variables that are now based on the
2284 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2286 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2287 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2290 Broadcasted = getBroadcastInstrs(Broadcasted);
2291 // After broadcasting the induction variable we need to make the vector
2292 // consecutive by adding 0, 1, 2, etc.
2293 for (unsigned part = 0; part < UF; ++part)
2294 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2297 case LoopVectorizationLegality::IK_ReverseIntInduction:
2298 case LoopVectorizationLegality::IK_PtrInduction:
2299 case LoopVectorizationLegality::IK_ReversePtrInduction:
2300 // Handle reverse integer and pointer inductions.
2301 Value *StartIdx = ExtendedIdx;
2302 // This is the normalized GEP that starts counting at zero.
2303 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2306 // Handle the reverse integer induction variable case.
2307 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2308 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2309 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2311 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2314 // This is a new value so do not hoist it out.
2315 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2316 // After broadcasting the induction variable we need to make the
2317 // vector consecutive by adding ... -3, -2, -1, 0.
2318 for (unsigned part = 0; part < UF; ++part)
2319 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2324 // Handle the pointer induction variable case.
2325 assert(P->getType()->isPointerTy() && "Unexpected type.");
2327 // Is this a reverse induction ptr or a consecutive induction ptr.
2328 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2331 // This is the vector of results. Notice that we don't generate
2332 // vector geps because scalar geps result in better code.
2333 for (unsigned part = 0; part < UF; ++part) {
2334 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2335 for (unsigned int i = 0; i < VF; ++i) {
2336 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2337 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2340 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2342 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2344 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2346 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2347 Builder.getInt32(i),
2350 Entry[part] = VecVal;
2357 case Instruction::Add:
2358 case Instruction::FAdd:
2359 case Instruction::Sub:
2360 case Instruction::FSub:
2361 case Instruction::Mul:
2362 case Instruction::FMul:
2363 case Instruction::UDiv:
2364 case Instruction::SDiv:
2365 case Instruction::FDiv:
2366 case Instruction::URem:
2367 case Instruction::SRem:
2368 case Instruction::FRem:
2369 case Instruction::Shl:
2370 case Instruction::LShr:
2371 case Instruction::AShr:
2372 case Instruction::And:
2373 case Instruction::Or:
2374 case Instruction::Xor: {
2375 // Just widen binops.
2376 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2377 VectorParts &A = getVectorValue(it->getOperand(0));
2378 VectorParts &B = getVectorValue(it->getOperand(1));
2380 // Use this vector value for all users of the original instruction.
2381 for (unsigned Part = 0; Part < UF; ++Part) {
2382 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2384 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2385 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2386 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2387 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2388 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2390 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2391 VecOp->setIsExact(BinOp->isExact());
2397 case Instruction::Select: {
2399 // If the selector is loop invariant we can create a select
2400 // instruction with a scalar condition. Otherwise, use vector-select.
2401 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2404 // The condition can be loop invariant but still defined inside the
2405 // loop. This means that we can't just use the original 'cond' value.
2406 // We have to take the 'vectorized' value and pick the first lane.
2407 // Instcombine will make this a no-op.
2408 VectorParts &Cond = getVectorValue(it->getOperand(0));
2409 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2410 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2411 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2412 Builder.getInt32(0));
2413 for (unsigned Part = 0; Part < UF; ++Part) {
2414 Entry[Part] = Builder.CreateSelect(
2415 InvariantCond ? ScalarCond : Cond[Part],
2422 case Instruction::ICmp:
2423 case Instruction::FCmp: {
2424 // Widen compares. Generate vector compares.
2425 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2426 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2427 VectorParts &A = getVectorValue(it->getOperand(0));
2428 VectorParts &B = getVectorValue(it->getOperand(1));
2429 for (unsigned Part = 0; Part < UF; ++Part) {
2432 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2434 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2440 case Instruction::Store:
2441 case Instruction::Load:
2442 vectorizeMemoryInstruction(it, Legal);
2444 case Instruction::ZExt:
2445 case Instruction::SExt:
2446 case Instruction::FPToUI:
2447 case Instruction::FPToSI:
2448 case Instruction::FPExt:
2449 case Instruction::PtrToInt:
2450 case Instruction::IntToPtr:
2451 case Instruction::SIToFP:
2452 case Instruction::UIToFP:
2453 case Instruction::Trunc:
2454 case Instruction::FPTrunc:
2455 case Instruction::BitCast: {
2456 CastInst *CI = dyn_cast<CastInst>(it);
2457 /// Optimize the special case where the source is the induction
2458 /// variable. Notice that we can only optimize the 'trunc' case
2459 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2460 /// c. other casts depend on pointer size.
2461 if (CI->getOperand(0) == OldInduction &&
2462 it->getOpcode() == Instruction::Trunc) {
2463 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2465 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2466 for (unsigned Part = 0; Part < UF; ++Part)
2467 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2470 /// Vectorize casts.
2471 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2473 VectorParts &A = getVectorValue(it->getOperand(0));
2474 for (unsigned Part = 0; Part < UF; ++Part)
2475 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2479 case Instruction::Call: {
2480 // Ignore dbg intrinsics.
2481 if (isa<DbgInfoIntrinsic>(it))
2484 Module *M = BB->getParent()->getParent();
2485 CallInst *CI = cast<CallInst>(it);
2486 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2487 assert(ID && "Not an intrinsic call!");
2488 for (unsigned Part = 0; Part < UF; ++Part) {
2489 SmallVector<Value*, 4> Args;
2490 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2491 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2492 Args.push_back(Arg[Part]);
2494 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2495 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2496 Entry[Part] = Builder.CreateCall(F, Args);
2502 // All other instructions are unsupported. Scalarize them.
2503 scalarizeInstruction(it);
2506 }// end of for_each instr.
2509 void InnerLoopVectorizer::updateAnalysis() {
2510 // Forget the original basic block.
2511 SE->forgetLoop(OrigLoop);
2513 // Update the dominator tree information.
2514 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2515 "Entry does not dominate exit.");
2517 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2518 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2519 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2520 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2521 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2522 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2523 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2524 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2526 DEBUG(DT->verifyAnalysis());
2529 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2530 if (!EnableIfConversion)
2533 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2534 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2536 // Collect the blocks that need predication.
2537 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2538 BasicBlock *BB = LoopBlocks[i];
2540 // We don't support switch statements inside loops.
2541 if (!isa<BranchInst>(BB->getTerminator()))
2544 // We must be able to predicate all blocks that need to be predicated.
2545 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2549 // Check that we can actually speculate the hoistable loads.
2550 if (!LoadSpeculation.canHoistAllLoads())
2553 // We can if-convert this loop.
2557 bool LoopVectorizationLegality::canVectorize() {
2558 // We must have a loop in canonical form. Loops with indirectbr in them cannot
2559 // be canonicalized.
2560 if (!TheLoop->getLoopPreheader())
2563 // We can only vectorize innermost loops.
2564 if (TheLoop->getSubLoopsVector().size())
2567 // We must have a single backedge.
2568 if (TheLoop->getNumBackEdges() != 1)
2571 // We must have a single exiting block.
2572 if (!TheLoop->getExitingBlock())
2575 unsigned NumBlocks = TheLoop->getNumBlocks();
2577 // Check if we can if-convert non single-bb loops.
2578 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2579 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2583 // We need to have a loop header.
2584 BasicBlock *Latch = TheLoop->getLoopLatch();
2585 DEBUG(dbgs() << "LV: Found a loop: " <<
2586 TheLoop->getHeader()->getName() << "\n");
2588 // ScalarEvolution needs to be able to find the exit count.
2589 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
2590 if (ExitCount == SE->getCouldNotCompute()) {
2591 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2595 // Do not loop-vectorize loops with a tiny trip count.
2596 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2597 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2598 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2599 "This loop is not worth vectorizing.\n");
2603 // Check if we can vectorize the instructions and CFG in this loop.
2604 if (!canVectorizeInstrs()) {
2605 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2609 // Go over each instruction and look at memory deps.
2610 if (!canVectorizeMemory()) {
2611 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2615 // Collect all of the variables that remain uniform after vectorization.
2616 collectLoopUniforms();
2618 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2619 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2622 // Okay! We can vectorize. At this point we don't have any other mem analysis
2623 // which may limit our maximum vectorization factor, so just return true with
2628 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
2629 if (Ty->isPointerTy())
2630 return DL.getIntPtrType(Ty->getContext());
2634 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
2635 Ty0 = convertPointerToIntegerType(DL, Ty0);
2636 Ty1 = convertPointerToIntegerType(DL, Ty1);
2637 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
2642 /// \brief Check that the instruction has outside loop users and is not an
2643 /// identified reduction variable.
2644 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
2645 SmallPtrSet<Value *, 4> &Reductions) {
2646 // Reduction instructions are allowed to have exit users. All other
2647 // instructions must not have external users.
2648 if (!Reductions.count(Inst))
2649 //Check that all of the users of the loop are inside the BB.
2650 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
2652 Instruction *U = cast<Instruction>(*I);
2653 // This user may be a reduction exit value.
2654 if (!TheLoop->contains(U)) {
2655 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2662 bool LoopVectorizationLegality::canVectorizeInstrs() {
2663 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2664 BasicBlock *Header = TheLoop->getHeader();
2666 // Look for the attribute signaling the absence of NaNs.
2667 Function &F = *Header->getParent();
2668 if (F.hasFnAttribute("no-nans-fp-math"))
2669 HasFunNoNaNAttr = F.getAttributes().getAttribute(
2670 AttributeSet::FunctionIndex,
2671 "no-nans-fp-math").getValueAsString() == "true";
2673 // For each block in the loop.
2674 for (Loop::block_iterator bb = TheLoop->block_begin(),
2675 be = TheLoop->block_end(); bb != be; ++bb) {
2677 // Scan the instructions in the block and look for hazards.
2678 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2681 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2682 Type *PhiTy = Phi->getType();
2683 // Check that this PHI type is allowed.
2684 if (!PhiTy->isIntegerTy() &&
2685 !PhiTy->isFloatingPointTy() &&
2686 !PhiTy->isPointerTy()) {
2687 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2691 // If this PHINode is not in the header block, then we know that we
2692 // can convert it to select during if-conversion. No need to check if
2693 // the PHIs in this block are induction or reduction variables.
2694 if (*bb != Header) {
2695 // Check that this instruction has no outside users or is an
2696 // identified reduction value with an outside user.
2697 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
2702 // We only allow if-converted PHIs with more than two incoming values.
2703 if (Phi->getNumIncomingValues() != 2) {
2704 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2708 // This is the value coming from the preheader.
2709 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2710 // Check if this is an induction variable.
2711 InductionKind IK = isInductionVariable(Phi);
2713 if (IK_NoInduction != IK) {
2714 // Get the widest type.
2716 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
2718 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
2720 // Int inductions are special because we only allow one IV.
2721 if (IK == IK_IntInduction) {
2722 // Use the phi node with the widest type as induction. Use the last
2723 // one if there are multiple (no good reason for doing this other
2724 // than it is expedient).
2725 if (!Induction || PhiTy == WidestIndTy)
2729 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2730 Inductions[Phi] = InductionInfo(StartValue, IK);
2734 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2735 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2738 if (AddReductionVar(Phi, RK_IntegerMult)) {
2739 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2742 if (AddReductionVar(Phi, RK_IntegerOr)) {
2743 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2746 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2747 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2750 if (AddReductionVar(Phi, RK_IntegerXor)) {
2751 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2754 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2755 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2758 if (AddReductionVar(Phi, RK_FloatMult)) {
2759 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2762 if (AddReductionVar(Phi, RK_FloatAdd)) {
2763 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2766 if (AddReductionVar(Phi, RK_FloatMinMax)) {
2767 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
2771 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2773 }// end of PHI handling
2775 // We still don't handle functions. However, we can ignore dbg intrinsic
2776 // calls and we do handle certain intrinsic and libm functions.
2777 CallInst *CI = dyn_cast<CallInst>(it);
2778 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2779 DEBUG(dbgs() << "LV: Found a call site.\n");
2783 // Check that the instruction return type is vectorizable.
2784 if (!VectorType::isValidElementType(it->getType()) &&
2785 !it->getType()->isVoidTy()) {
2786 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2790 // Check that the stored type is vectorizable.
2791 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2792 Type *T = ST->getValueOperand()->getType();
2793 if (!VectorType::isValidElementType(T))
2797 // Reduction instructions are allowed to have exit users.
2798 // All other instructions must not have external users.
2799 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
2807 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2808 if (Inductions.empty())
2815 void LoopVectorizationLegality::collectLoopUniforms() {
2816 // We now know that the loop is vectorizable!
2817 // Collect variables that will remain uniform after vectorization.
2818 std::vector<Value*> Worklist;
2819 BasicBlock *Latch = TheLoop->getLoopLatch();
2821 // Start with the conditional branch and walk up the block.
2822 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2824 while (Worklist.size()) {
2825 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2826 Worklist.pop_back();
2828 // Look at instructions inside this loop.
2829 // Stop when reaching PHI nodes.
2830 // TODO: we need to follow values all over the loop, not only in this block.
2831 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2834 // This is a known uniform.
2837 // Insert all operands.
2838 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
2842 /// \brief Analyses memory accesses in a loop.
2844 /// Checks whether run time pointer checks are needed and builds sets for data
2845 /// dependence checking.
2846 class AccessAnalysis {
2848 /// \brief Read or write access location.
2849 typedef std::pair<Value*, char> MemAccessInfo;
2851 /// \brief Set of potential dependent memory accesses.
2852 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
2854 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
2855 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
2856 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
2858 /// \brief Register a load and whether it is only read from.
2859 void addLoad(Value *Ptr, bool IsReadOnly) {
2860 Accesses.insert(std::make_pair(Ptr, false));
2862 ReadOnlyPtr.insert(Ptr);
2865 /// \brief Register a store.
2866 void addStore(Value *Ptr) {
2867 Accesses.insert(std::make_pair(Ptr, true));
2870 /// \brief Check whether we can check the pointers at runtime for
2871 /// non-intersection.
2872 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2873 unsigned &NumComparisons, ScalarEvolution *SE,
2876 /// \brief Goes over all memory accesses, checks whether a RT check is needed
2877 /// and builds sets of dependent accesses.
2878 void buildDependenceSets() {
2879 // Process read-write pointers first.
2880 processMemAccesses(false);
2881 // Next, process read pointers.
2882 processMemAccesses(true);
2885 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
2887 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
2889 DenseSet<MemAccessInfo> &getDependenciesToCheck() { return CheckDeps; }
2892 typedef SetVector<MemAccessInfo> PtrAccessSet;
2893 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
2895 /// \brief Go over all memory access or only the deferred ones if
2896 /// \p UseDeferred is true and check whether runtime pointer checks are needed
2897 /// and build sets of dependency check candidates.
2898 void processMemAccesses(bool UseDeferred);
2900 /// Set of all accesses.
2901 PtrAccessSet Accesses;
2903 /// Set of access to check after all writes have been processed.
2904 PtrAccessSet DeferredAccesses;
2906 /// Map of pointers to last access encountered.
2907 UnderlyingObjToAccessMap ObjToLastAccess;
2909 /// Set of accesses that need a further dependence check.
2910 DenseSet<MemAccessInfo> CheckDeps;
2912 /// Set of pointers that are read only.
2913 SmallPtrSet<Value*, 16> ReadOnlyPtr;
2915 /// Set of underlying objects already written to.
2916 SmallPtrSet<Value*, 16> WriteObjects;
2920 /// Sets of potentially dependent accesses - members of one set share an
2921 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
2922 /// dependence check.
2923 DepCandidates &DepCands;
2925 bool AreAllWritesIdentified;
2926 bool AreAllReadsIdentified;
2927 bool IsRTCheckNeeded;
2930 /// \brief Check whether a pointer can participate in a runtime bounds check.
2931 static bool hasComputableBounds(ScalarEvolution *SE, Value *Ptr) {
2932 const SCEV *PtrScev = SE->getSCEV(Ptr);
2933 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
2937 return AR->isAffine();
2940 bool AccessAnalysis::canCheckPtrAtRT(
2941 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
2942 unsigned &NumComparisons, ScalarEvolution *SE,
2944 // Find pointers with computable bounds. We are going to use this information
2945 // to place a runtime bound check.
2946 unsigned NumReadPtrChecks = 0;
2947 unsigned NumWritePtrChecks = 0;
2948 bool CanDoRT = true;
2950 bool IsDepCheckNeeded = isDependencyCheckNeeded();
2951 // We assign consecutive id to access from different dependence sets.
2952 // Accesses within the same set don't need a runtime check.
2953 unsigned RunningDepId = 1;
2954 DenseMap<Value *, unsigned> DepSetId;
2956 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
2958 const MemAccessInfo &Access = *AI;
2959 Value *Ptr = Access.first;
2960 bool IsWrite = Access.second;
2962 // Just add write checks if we have both.
2963 if (!IsWrite && Accesses.count(std::make_pair(Ptr, true)))
2967 ++NumWritePtrChecks;
2971 if (hasComputableBounds(SE, Ptr)) {
2972 // The id of the dependence set.
2975 if (IsDepCheckNeeded) {
2976 Value *Leader = DepCands.getLeaderValue(Access).first;
2977 unsigned &LeaderId = DepSetId[Leader];
2979 LeaderId = RunningDepId++;
2982 // Each access has its own dependence set.
2983 DepId = RunningDepId++;
2985 //RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId);
2987 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr <<"\n");
2993 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
2994 NumComparisons = 0; // Only one dependence set.
2996 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
2997 NumWritePtrChecks - 1));
3001 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3002 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3005 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3006 // We process the set twice: first we process read-write pointers, last we
3007 // process read-only pointers. This allows us to skip dependence tests for
3008 // read-only pointers.
3010 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3011 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3012 const MemAccessInfo &Access = *AI;
3013 Value *Ptr = Access.first;
3014 bool IsWrite = Access.second;
3016 DepCands.insert(Access);
3018 // Memorize read-only pointers for later processing and skip them in the
3019 // first round (they need to be checked after we have seen all write
3020 // pointers). Note: we also mark pointer that are not consecutive as
3021 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3022 // second check for "!IsWrite".
3023 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3024 if (!UseDeferred && IsReadOnlyPtr) {
3025 DeferredAccesses.insert(Access);
3029 bool NeedDepCheck = false;
3030 // Check whether there is the possiblity of dependency because of underlying
3031 // objects being the same.
3032 typedef SmallVector<Value*, 16> ValueVector;
3033 ValueVector TempObjects;
3034 GetUnderlyingObjects(Ptr, TempObjects, DL);
3035 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3037 Value *UnderlyingObj = *UI;
3039 // If this is a write then it needs to be an identified object. If this a
3040 // read and all writes (so far) are identified function scope objects we
3041 // don't need an identified underlying object but only an Argument (the
3042 // next write is going to invalidate this assumption if it is
3044 // This is a micro-optimization for the case where all writes are
3045 // identified and we have one argument pointer.
3046 // Otherwise, we do need a runtime check.
3047 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3048 (!IsWrite && (!AreAllWritesIdentified ||
3049 !isa<Argument>(UnderlyingObj)) &&
3050 !isIdentifiedObject(UnderlyingObj))) {
3051 DEBUG(dbgs() << "LV: Found an unidentified " <<
3052 (IsWrite ? "write" : "read" ) << " ptr:" << *UnderlyingObj <<
3054 IsRTCheckNeeded = (IsRTCheckNeeded ||
3055 !isIdentifiedObject(UnderlyingObj) ||
3056 !AreAllReadsIdentified);
3059 AreAllWritesIdentified = false;
3061 AreAllReadsIdentified = false;
3064 // If this is a write - check other reads and writes for conflicts. If
3065 // this is a read only check other writes for conflicts (but only if there
3066 // is no other write to the ptr - this is an optimization to catch "a[i] =
3067 // a[i] + " without having to do a dependence check).
3068 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3069 NeedDepCheck = true;
3072 WriteObjects.insert(UnderlyingObj);
3074 // Create sets of pointers connected by shared underlying objects.
3075 UnderlyingObjToAccessMap::iterator Prev =
3076 ObjToLastAccess.find(UnderlyingObj);
3077 if (Prev != ObjToLastAccess.end())
3078 DepCands.unionSets(Access, Prev->second);
3080 ObjToLastAccess[UnderlyingObj] = Access;
3084 CheckDeps.insert(Access);
3088 /// \brief Checks memory dependences among accesses to the same underlying
3089 /// object to determine whether there vectorization is legal or not (and at
3090 /// which vectorization factor).
3092 /// This class works under the assumption that we already checked that memory
3093 /// locations with different underlying pointers are "must-not alias".
3094 /// We use the ScalarEvolution framework to symbolically evalutate access
3095 /// functions pairs. Since we currently don't restructure the loop we can rely
3096 /// on the program order of memory accesses to determine their safety.
3097 /// At the moment we will only deem accesses as safe for:
3098 /// * A negative constant distance assuming program order.
3100 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3101 /// a[i] = tmp; y = a[i];
3103 /// The latter case is safe because later checks guarantuee that there can't
3104 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3105 /// the same variable: a header phi can only be an induction or a reduction, a
3106 /// reduction can't have a memory sink, an induction can't have a memory
3107 /// source). This is important and must not be violated (or we have to
3108 /// resort to checking for cycles through memory).
3110 /// * A positive constant distance assuming program order that is bigger
3111 /// than the biggest memory access.
3113 /// tmp = a[i] OR b[i] = x
3114 /// a[i+2] = tmp y = b[i+2];
3116 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3118 /// * Zero distances and all accesses have the same size.
3120 class MemoryDepChecker {
3122 typedef std::pair<Value*, char> MemAccessInfo;
3124 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L) :
3125 SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0) {}
3127 /// \brief Register the location (instructions are given increasing numbers)
3128 /// of a write access.
3129 void addAccess(StoreInst *SI) {
3130 Value *Ptr = SI->getPointerOperand();
3131 Accesses[std::make_pair(Ptr, true)].push_back(AccessIdx);
3132 InstMap.push_back(SI);
3136 /// \brief Register the location (instructions are given increasing numbers)
3137 /// of a write access.
3138 void addAccess(LoadInst *LI) {
3139 Value *Ptr = LI->getPointerOperand();
3140 Accesses[std::make_pair(Ptr, false)].push_back(AccessIdx);
3141 InstMap.push_back(LI);
3145 /// \brief Check whether the dependencies between the accesses are safe.
3147 /// Only checks sets with elements in \p CheckDeps.
3148 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3149 DenseSet<MemAccessInfo> &CheckDeps);
3151 /// \brief The maximum number of bytes of a vector register we can vectorize
3152 /// the accesses safely with.
3153 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3156 ScalarEvolution *SE;
3158 const Loop *InnermostLoop;
3160 /// \brief Maps access locations (ptr, read/write) to program order.
3161 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3163 /// \brief Memory access instructions in program order.
3164 SmallVector<Instruction *, 16> InstMap;
3166 /// \brief The program order index to be used for the next instruction.
3169 // We can access this many bytes in parallel safely.
3170 unsigned MaxSafeDepDistBytes;
3172 /// \brief Check whether there is a plausible dependence between the two
3175 /// Access \p A must happen before \p B in program order. The two indices
3176 /// identify the index into the program order map.
3178 /// This function checks whether there is a plausible dependence (or the
3179 /// absence of such can't be proved) between the two accesses. If there is a
3180 /// plausible dependence but the dependence distance is bigger than one
3181 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
3182 /// distance is smaller than any other distance encountered so far).
3183 /// Otherwise, this function returns true signaling a possible dependence.
3184 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
3185 const MemAccessInfo &B, unsigned BIdx);
3187 /// \brief Check whether the data dependence could prevent store-load
3189 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
3192 static bool isInBoundsGep(Value *Ptr) {
3193 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
3194 return GEP->isInBounds();
3198 /// \brief Check whether the access through \p Ptr has a constant stride.
3199 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3201 const Type *PtrTy = Ptr->getType();
3202 assert(PtrTy->isPointerTy() && "Unexpected non ptr");
3204 // Make sure that the pointer does not point to aggregate types.
3205 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType()) {
3206 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr
3211 const SCEV *PtrScev = SE->getSCEV(Ptr);
3212 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3214 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
3215 << *Ptr << " SCEV: " << *PtrScev << "\n");
3219 // The accesss function must stride over the innermost loop.
3220 if (Lp != AR->getLoop()) {
3221 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " << *Ptr
3222 << " SCEV: " << *PtrScev << "\n");
3225 // The address calculation must not wrap. Otherwise, a dependence could be
3226 // inverted. An inbounds getelementptr that is a AddRec with a unit stride
3227 // cannot wrap per definition. The unit stride requirement is checked later.
3228 bool IsInBoundsGEP = isInBoundsGep(Ptr);
3229 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
3230 if (!IsNoWrapAddRec && !IsInBoundsGEP) {
3231 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
3232 << *Ptr << " SCEV: " << *PtrScev << "\n");
3236 // Check the step is constant.
3237 const SCEV *Step = AR->getStepRecurrence(*SE);
3239 // Calculate the pointer stride and check if it is consecutive.
3240 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3242 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
3243 " SCEV: " << *PtrScev << "\n");
3247 int64_t Size = DL->getTypeAllocSize(PtrTy->getPointerElementType());
3248 const APInt &APStepVal = C->getValue()->getValue();
3250 // Huge step value - give up.
3251 if (APStepVal.getBitWidth() > 64)
3254 int64_t StepVal = APStepVal.getSExtValue();
3257 int64_t Stride = StepVal / Size;
3258 int64_t Rem = StepVal % Size;
3262 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
3263 // know we can't "wrap around the address space".
3264 if (!IsNoWrapAddRec && IsInBoundsGEP && Stride != 1 && Stride != -1)
3270 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
3271 unsigned TypeByteSize) {
3272 // If loads occur at a distance that is not a multiple of a feasible vector
3273 // factor store-load forwarding does not take place.
3274 // Positive dependences might cause troubles because vectorizing them might
3275 // prevent store-load forwarding making vectorized code run a lot slower.
3276 // a[i] = a[i-3] ^ a[i-8];
3277 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
3278 // hence on your typical architecture store-load forwarding does not take
3279 // place. Vectorizing in such cases does not make sense.
3280 // Store-load forwarding distance.
3281 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
3282 // Maximum vector factor.
3283 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
3284 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
3285 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
3287 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
3289 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
3290 MaxVFWithoutSLForwardIssues = (vf >>=1);
3295 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
3296 DEBUG(dbgs() << "LV: Distance " << Distance <<
3297 " that could cause a store-load forwarding conflict\n");
3301 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
3302 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
3303 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
3307 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
3308 const MemAccessInfo &B, unsigned BIdx) {
3309 assert (AIdx < BIdx && "Must pass arguments in program order");
3311 Value *APtr = A.first;
3312 Value *BPtr = B.first;
3313 bool AIsWrite = A.second;
3314 bool BIsWrite = B.second;
3316 // Two reads are independent.
3317 if (!AIsWrite && !BIsWrite)
3320 const SCEV *AScev = SE->getSCEV(APtr);
3321 const SCEV *BScev = SE->getSCEV(BPtr);
3323 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop);
3324 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop);
3326 const SCEV *Src = AScev;
3327 const SCEV *Sink = BScev;
3329 // If the induction step is negative we have to invert source and sink of the
3331 if (StrideAPtr < 0) {
3334 std::swap(APtr, BPtr);
3335 std::swap(Src, Sink);
3336 std::swap(AIsWrite, BIsWrite);
3337 std::swap(AIdx, BIdx);
3338 std::swap(StrideAPtr, StrideBPtr);
3341 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
3343 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
3344 << "(Induction step: " << StrideAPtr << ")\n");
3345 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
3346 << *InstMap[BIdx] << ": " << *Dist << "\n");
3348 // Need consecutive accesses. We don't want to vectorize
3349 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
3350 // the address space.
3351 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
3352 DEBUG(dbgs() << "Non-consecutive pointer access\n");
3356 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
3358 DEBUG(dbgs() << "LV: Dependence because of non constant distance\n");
3362 Type *ATy = APtr->getType()->getPointerElementType();
3363 Type *BTy = BPtr->getType()->getPointerElementType();
3364 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
3366 // Negative distances are not plausible dependencies.
3367 const APInt &Val = C->getValue()->getValue();
3368 if (Val.isNegative()) {
3369 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
3370 if (IsTrueDataDependence &&
3371 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
3375 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
3379 // Write to the same location with the same size.
3380 // Could be improved to assert type sizes are the same (i32 == float, etc).
3384 DEBUG(dbgs() << "LV: Zero dependence difference but different types");
3388 assert(Val.isStrictlyPositive() && "Expect a positive value");
3390 // Positive distance bigger than max vectorization factor.
3393 "LV: ReadWrite-Write positive dependency with different types");
3397 unsigned Distance = (unsigned) Val.getZExtValue();
3399 // Bail out early if passed-in parameters make vectorization not feasible.
3400 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
3401 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
3403 // The distance must be bigger than the size needed for a vectorized version
3404 // of the operation and the size of the vectorized operation must not be
3405 // bigger than the currrent maximum size.
3406 if (Distance < 2*TypeByteSize ||
3407 2*TypeByteSize > MaxSafeDepDistBytes ||
3408 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
3409 DEBUG(dbgs() << "LV: Failure because of Positive distance "
3410 << Val.getSExtValue() << "\n");
3414 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
3415 Distance : MaxSafeDepDistBytes;
3417 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
3418 if (IsTrueDataDependence &&
3419 couldPreventStoreLoadForward(Distance, TypeByteSize))
3422 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
3423 " with max VF=" << MaxSafeDepDistBytes/TypeByteSize << "\n");
3429 MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3430 DenseSet<MemAccessInfo> &CheckDeps) {
3432 MaxSafeDepDistBytes = -1U;
3433 while (!CheckDeps.empty()) {
3434 MemAccessInfo CurAccess = *CheckDeps.begin();
3436 // Get the relevant memory access set.
3437 EquivalenceClasses<MemAccessInfo>::iterator I =
3438 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
3440 // Check accesses within this set.
3441 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
3442 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
3444 // Check every access pair.
3446 CheckDeps.erase(*AI);
3447 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
3449 // Check every accessing instruction pair in program order.
3450 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
3451 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
3452 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
3453 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
3454 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2))
3456 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1))
3467 AliasAnalysis::Location
3468 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
3469 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
3470 return AA->getLocation(Store);
3471 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
3472 return AA->getLocation(Load);
3474 llvm_unreachable("Should be either load or store instruction");
3478 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
3481 AliasMultiMap& WriteObjects,
3482 unsigned MaxByteWidth) {
3484 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
3486 std::vector<Instruction*>::iterator
3487 it = WriteObjects[Object].begin(),
3488 end = WriteObjects[Object].end();
3490 for (; it != end; ++it) {
3491 Instruction* I = *it;
3495 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
3496 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
3497 ThatLoc.getWithNewSize(MaxByteWidth)))
3503 bool LoopVectorizationLegality::canVectorizeMemory() {
3505 typedef SmallVector<Value*, 16> ValueVector;
3506 typedef SmallPtrSet<Value*, 16> ValueSet;
3507 // Holds the Load and Store *instructions*.
3510 PtrRtCheck.Pointers.clear();
3511 PtrRtCheck.Need = false;
3513 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
3516 for (Loop::block_iterator bb = TheLoop->block_begin(),
3517 be = TheLoop->block_end(); bb != be; ++bb) {
3519 // Scan the BB and collect legal loads and stores.
3520 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3523 // If this is a load, save it. If this instruction can read from memory
3524 // but is not a load, then we quit. Notice that we don't handle function
3525 // calls that read or write.
3526 if (it->mayReadFromMemory()) {
3527 LoadInst *Ld = dyn_cast<LoadInst>(it);
3528 if (!Ld) return false;
3529 if (!Ld->isSimple() && !IsAnnotatedParallel) {
3530 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
3533 Loads.push_back(Ld);
3537 // Save 'store' instructions. Abort if other instructions write to memory.
3538 if (it->mayWriteToMemory()) {
3539 StoreInst *St = dyn_cast<StoreInst>(it);
3540 if (!St) return false;
3541 if (!St->isSimple() && !IsAnnotatedParallel) {
3542 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
3545 Stores.push_back(St);
3550 // Now we have two lists that hold the loads and the stores.
3551 // Next, we find the pointers that they use.
3553 // Check if we see any stores. If there are no stores, then we don't
3554 // care if the pointers are *restrict*.
3555 if (!Stores.size()) {
3556 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
3560 // Holds the read and read-write *pointers* that we find. These maps hold
3561 // unique values for pointers (so no need for multi-map).
3563 AliasMap ReadWrites;
3565 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
3566 // multiple times on the same object. If the ptr is accessed twice, once
3567 // for read and once for write, it will only appear once (on the write
3568 // list). This is okay, since we are going to check for conflicts between
3569 // writes and between reads and writes, but not between reads and reads.
3572 ValueVector::iterator I, IE;
3573 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
3574 StoreInst *ST = cast<StoreInst>(*I);
3575 Value* Ptr = ST->getPointerOperand();
3577 if (isUniform(Ptr)) {
3578 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
3582 // If we did *not* see this pointer before, insert it to
3583 // the read-write list. At this phase it is only a 'write' list.
3584 if (Seen.insert(Ptr))
3585 ReadWrites.insert(std::make_pair(Ptr, ST));
3588 if (IsAnnotatedParallel) {
3590 << "LV: A loop annotated parallel, ignore memory dependency "
3595 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
3596 LoadInst *LD = cast<LoadInst>(*I);
3597 Value* Ptr = LD->getPointerOperand();
3598 // If we did *not* see this pointer before, insert it to the
3599 // read list. If we *did* see it before, then it is already in
3600 // the read-write list. This allows us to vectorize expressions
3601 // such as A[i] += x; Because the address of A[i] is a read-write
3602 // pointer. This only works if the index of A[i] is consecutive.
3603 // If the address of i is unknown (for example A[B[i]]) then we may
3604 // read a few words, modify, and write a few words, and some of the
3605 // words may be written to the same address.
3606 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
3607 Reads.insert(std::make_pair(Ptr, LD));
3610 // If we write (or read-write) to a single destination and there are no
3611 // other reads in this loop then is it safe to vectorize.
3612 if (ReadWrites.size() == 1 && Reads.size() == 0) {
3613 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
3617 unsigned NumReadPtrs = 0;
3618 unsigned NumWritePtrs = 0;
3620 // Find pointers with computable bounds. We are going to use this information
3621 // to place a runtime bound check.
3622 bool CanDoRT = true;
3623 AliasMap::iterator MI, ME;
3624 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
3625 Value *V = (*MI).first;
3626 if (hasComputableBounds(V)) {
3627 PtrRtCheck.insert(SE, TheLoop, V, true);
3629 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
3635 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
3636 Value *V = (*MI).first;
3637 if (hasComputableBounds(V)) {
3638 PtrRtCheck.insert(SE, TheLoop, V, false);
3640 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
3647 // Check that we did not collect too many pointers or found a
3648 // unsizeable pointer.
3649 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
3650 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
3651 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
3657 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
3660 bool NeedRTCheck = false;
3662 // Biggest vectorized access possible, vector width * unroll factor.
3663 // TODO: We're being very pessimistic here, find a way to know the
3664 // real access width before getting here.
3665 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
3666 TTI->getMaximumUnrollFactor();
3667 // Now that the pointers are in two lists (Reads and ReadWrites), we
3668 // can check that there are no conflicts between each of the writes and
3669 // between the writes to the reads.
3670 // Note that WriteObjects duplicates the stores (indexed now by underlying
3671 // objects) to avoid pointing to elements inside ReadWrites.
3672 // TODO: Maybe create a new type where they can interact without duplication.
3673 AliasMultiMap WriteObjects;
3674 ValueVector TempObjects;
3676 // Check that the read-writes do not conflict with other read-write
3678 bool AllWritesIdentified = true;
3679 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
3680 Value *Val = (*MI).first;
3681 Instruction *Inst = (*MI).second;
3683 GetUnderlyingObjects(Val, TempObjects, DL);
3684 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
3686 if (!isIdentifiedObject(*UI)) {
3687 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
3689 AllWritesIdentified = false;
3692 // Never seen it before, can't alias.
3693 if (WriteObjects[*UI].empty()) {
3694 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
3695 WriteObjects[*UI].push_back(Inst);
3698 // Direct alias found.
3699 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
3700 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
3704 DEBUG(dbgs() << "LV: Found a conflicting global value:"
3706 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
3707 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
3709 // If global alias, make sure they do alias.
3710 if (hasPossibleGlobalWriteReorder(*UI,
3714 DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
3719 // Didn't alias, insert into map for further reference.
3720 WriteObjects[*UI].push_back(Inst);
3722 TempObjects.clear();
3725 /// Check that the reads don't conflict with the read-writes.
3726 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
3727 Value *Val = (*MI).first;
3728 GetUnderlyingObjects(Val, TempObjects, DL);
3729 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
3731 // If all of the writes are identified then we don't care if the read
3732 // pointer is identified or not.
3733 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
3734 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
3738 // Never seen it before, can't alias.
3739 if (WriteObjects[*UI].empty())
3741 // Direct alias found.
3742 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
3743 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
3747 DEBUG(dbgs() << "LV: Found a global value: "
3749 Instruction *Inst = (*MI).second;
3750 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
3751 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
3753 // If global alias, make sure they do alias.
3754 if (hasPossibleGlobalWriteReorder(*UI,
3758 DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
3763 TempObjects.clear();
3766 PtrRtCheck.Need = NeedRTCheck;
3767 if (NeedRTCheck && !CanDoRT) {
3768 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
3769 "the array bounds.\n");
3774 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
3775 " need a runtime memory check.\n");
3779 static bool hasMultipleUsesOf(Instruction *I,
3780 SmallPtrSet<Instruction *, 8> &Insts) {
3781 unsigned NumUses = 0;
3782 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3783 if (Insts.count(dyn_cast<Instruction>(*Use)))
3792 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
3793 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3794 if (!Set.count(dyn_cast<Instruction>(*Use)))
3799 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3800 ReductionKind Kind) {
3801 if (Phi->getNumIncomingValues() != 2)
3804 // Reduction variables are only found in the loop header block.
3805 if (Phi->getParent() != TheLoop->getHeader())
3808 // Obtain the reduction start value from the value that comes from the loop
3810 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3812 // ExitInstruction is the single value which is used outside the loop.
3813 // We only allow for a single reduction value to be used outside the loop.
3814 // This includes users of the reduction, variables (which form a cycle
3815 // which ends in the phi node).
3816 Instruction *ExitInstruction = 0;
3817 // Indicates that we found a reduction operation in our scan.
3818 bool FoundReduxOp = false;
3820 // We start with the PHI node and scan for all of the users of this
3821 // instruction. All users must be instructions that can be used as reduction
3822 // variables (such as ADD). We must have a single out-of-block user. The cycle
3823 // must include the original PHI.
3824 bool FoundStartPHI = false;
3826 // To recognize min/max patterns formed by a icmp select sequence, we store
3827 // the number of instruction we saw from the recognized min/max pattern,
3828 // to make sure we only see exactly the two instructions.
3829 unsigned NumCmpSelectPatternInst = 0;
3830 ReductionInstDesc ReduxDesc(false, 0);
3832 SmallPtrSet<Instruction *, 8> VisitedInsts;
3833 SmallVector<Instruction *, 8> Worklist;
3834 Worklist.push_back(Phi);
3835 VisitedInsts.insert(Phi);
3837 // A value in the reduction can be used:
3838 // - By the reduction:
3839 // - Reduction operation:
3840 // - One use of reduction value (safe).
3841 // - Multiple use of reduction value (not safe).
3843 // - All uses of the PHI must be the reduction (safe).
3844 // - Otherwise, not safe.
3845 // - By one instruction outside of the loop (safe).
3846 // - By further instructions outside of the loop (not safe).
3847 // - By an instruction that is not part of the reduction (not safe).
3849 // * An instruction type other than PHI or the reduction operation.
3850 // * A PHI in the header other than the initial PHI.
3851 while (!Worklist.empty()) {
3852 Instruction *Cur = Worklist.back();
3853 Worklist.pop_back();
3856 // If the instruction has no users then this is a broken chain and can't be
3857 // a reduction variable.
3858 if (Cur->use_empty())
3861 bool IsAPhi = isa<PHINode>(Cur);
3863 // A header PHI use other than the original PHI.
3864 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3867 // Reductions of instructions such as Div, and Sub is only possible if the
3868 // LHS is the reduction variable.
3869 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3870 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3871 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3874 // Any reduction instruction must be of one of the allowed kinds.
3875 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3876 if (!ReduxDesc.IsReduction)
3879 // A reduction operation must only have one use of the reduction value.
3880 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3881 hasMultipleUsesOf(Cur, VisitedInsts))
3884 // All inputs to a PHI node must be a reduction value.
3885 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3888 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3889 isa<SelectInst>(Cur)))
3890 ++NumCmpSelectPatternInst;
3891 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3892 isa<SelectInst>(Cur)))
3893 ++NumCmpSelectPatternInst;
3895 // Check whether we found a reduction operator.
3896 FoundReduxOp |= !IsAPhi;
3898 // Process users of current instruction. Push non PHI nodes after PHI nodes
3899 // onto the stack. This way we are going to have seen all inputs to PHI
3900 // nodes once we get to them.
3901 SmallVector<Instruction *, 8> NonPHIs;
3902 SmallVector<Instruction *, 8> PHIs;
3903 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
3905 Instruction *Usr = cast<Instruction>(*UI);
3907 // Check if we found the exit user.
3908 BasicBlock *Parent = Usr->getParent();
3909 if (!TheLoop->contains(Parent)) {
3910 // Exit if you find multiple outside users.
3911 if (ExitInstruction != 0)
3913 ExitInstruction = Cur;
3917 // Process instructions only once (termination).
3918 if (VisitedInsts.insert(Usr)) {
3919 if (isa<PHINode>(Usr))
3920 PHIs.push_back(Usr);
3922 NonPHIs.push_back(Usr);
3924 // Remember that we completed the cycle.
3926 FoundStartPHI = true;
3928 Worklist.append(PHIs.begin(), PHIs.end());
3929 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3932 // This means we have seen one but not the other instruction of the
3933 // pattern or more than just a select and cmp.
3934 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3935 NumCmpSelectPatternInst != 2)
3938 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3941 // We found a reduction var if we have reached the original phi node and we
3942 // only have a single instruction with out-of-loop users.
3944 // This instruction is allowed to have out-of-loop users.
3945 AllowedExit.insert(ExitInstruction);
3947 // Save the description of this reduction variable.
3948 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
3949 ReduxDesc.MinMaxKind);
3950 Reductions[Phi] = RD;
3951 // We've ended the cycle. This is a reduction variable if we have an
3952 // outside user and it has a binary op.
3957 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
3958 /// pattern corresponding to a min(X, Y) or max(X, Y).
3959 LoopVectorizationLegality::ReductionInstDesc
3960 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
3961 ReductionInstDesc &Prev) {
3963 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
3964 "Expect a select instruction");
3965 Instruction *Cmp = 0;
3966 SelectInst *Select = 0;
3968 // We must handle the select(cmp()) as a single instruction. Advance to the
3970 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
3971 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
3972 return ReductionInstDesc(false, I);
3973 return ReductionInstDesc(Select, Prev.MinMaxKind);
3976 // Only handle single use cases for now.
3977 if (!(Select = dyn_cast<SelectInst>(I)))
3978 return ReductionInstDesc(false, I);
3979 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
3980 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
3981 return ReductionInstDesc(false, I);
3982 if (!Cmp->hasOneUse())
3983 return ReductionInstDesc(false, I);
3988 // Look for a min/max pattern.
3989 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3990 return ReductionInstDesc(Select, MRK_UIntMin);
3991 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3992 return ReductionInstDesc(Select, MRK_UIntMax);
3993 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3994 return ReductionInstDesc(Select, MRK_SIntMax);
3995 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3996 return ReductionInstDesc(Select, MRK_SIntMin);
3997 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3998 return ReductionInstDesc(Select, MRK_FloatMin);
3999 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4000 return ReductionInstDesc(Select, MRK_FloatMax);
4001 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4002 return ReductionInstDesc(Select, MRK_FloatMin);
4003 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4004 return ReductionInstDesc(Select, MRK_FloatMax);
4006 return ReductionInstDesc(false, I);
4009 LoopVectorizationLegality::ReductionInstDesc
4010 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4012 ReductionInstDesc &Prev) {
4013 bool FP = I->getType()->isFloatingPointTy();
4014 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4015 switch (I->getOpcode()) {
4017 return ReductionInstDesc(false, I);
4018 case Instruction::PHI:
4019 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4020 Kind != RK_FloatMinMax))
4021 return ReductionInstDesc(false, I);
4022 return ReductionInstDesc(I, Prev.MinMaxKind);
4023 case Instruction::Sub:
4024 case Instruction::Add:
4025 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4026 case Instruction::Mul:
4027 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4028 case Instruction::And:
4029 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4030 case Instruction::Or:
4031 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4032 case Instruction::Xor:
4033 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4034 case Instruction::FMul:
4035 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4036 case Instruction::FAdd:
4037 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4038 case Instruction::FCmp:
4039 case Instruction::ICmp:
4040 case Instruction::Select:
4041 if (Kind != RK_IntegerMinMax &&
4042 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4043 return ReductionInstDesc(false, I);
4044 return isMinMaxSelectCmpPattern(I, Prev);
4048 LoopVectorizationLegality::InductionKind
4049 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4050 Type *PhiTy = Phi->getType();
4051 // We only handle integer and pointer inductions variables.
4052 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4053 return IK_NoInduction;
4055 // Check that the PHI is consecutive.
4056 const SCEV *PhiScev = SE->getSCEV(Phi);
4057 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4059 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4060 return IK_NoInduction;
4062 const SCEV *Step = AR->getStepRecurrence(*SE);
4064 // Integer inductions need to have a stride of one.
4065 if (PhiTy->isIntegerTy()) {
4067 return IK_IntInduction;
4068 if (Step->isAllOnesValue())
4069 return IK_ReverseIntInduction;
4070 return IK_NoInduction;
4073 // Calculate the pointer stride and check if it is consecutive.
4074 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4076 return IK_NoInduction;
4078 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4079 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4080 if (C->getValue()->equalsInt(Size))
4081 return IK_PtrInduction;
4082 else if (C->getValue()->equalsInt(0 - Size))
4083 return IK_ReversePtrInduction;
4085 return IK_NoInduction;
4088 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4089 Value *In0 = const_cast<Value*>(V);
4090 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4094 return Inductions.count(PN);
4097 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4098 assert(TheLoop->contains(BB) && "Unknown block used");
4100 // Blocks that do not dominate the latch need predication.
4101 BasicBlock* Latch = TheLoop->getLoopLatch();
4102 return !DT->dominates(BB, Latch);
4105 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
4106 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4107 // We might be able to hoist the load.
4108 if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
4111 // We don't predicate stores at the moment.
4112 if (it->mayWriteToMemory() || it->mayThrow())
4115 // The instructions below can trap.
4116 switch (it->getOpcode()) {
4118 case Instruction::UDiv:
4119 case Instruction::SDiv:
4120 case Instruction::URem:
4121 case Instruction::SRem:
4129 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
4130 const SCEV *PhiScev = SE->getSCEV(Ptr);
4131 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4135 return AR->isAffine();
4138 LoopVectorizationCostModel::VectorizationFactor
4139 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4141 // Width 1 means no vectorize
4142 VectorizationFactor Factor = { 1U, 0U };
4143 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4144 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4148 // Find the trip count.
4149 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4150 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
4152 unsigned WidestType = getWidestType();
4153 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4154 unsigned MaxVectorSize = WidestRegister / WidestType;
4155 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4156 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
4158 if (MaxVectorSize == 0) {
4159 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4163 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4164 " into one vector!");
4166 unsigned VF = MaxVectorSize;
4168 // If we optimize the program for size, avoid creating the tail loop.
4170 // If we are unable to calculate the trip count then don't try to vectorize.
4172 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4176 // Find the maximum SIMD width that can fit within the trip count.
4177 VF = TC % MaxVectorSize;
4182 // If the trip count that we found modulo the vectorization factor is not
4183 // zero then we require a tail.
4185 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4191 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4192 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
4194 Factor.Width = UserVF;
4198 float Cost = expectedCost(1);
4200 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
4201 for (unsigned i=2; i <= VF; i*=2) {
4202 // Notice that the vector loop needs to be executed less times, so
4203 // we need to divide the cost of the vector loops by the width of
4204 // the vector elements.
4205 float VectorCost = expectedCost(i) / (float)i;
4206 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
4207 (int)VectorCost << ".\n");
4208 if (VectorCost < Cost) {
4214 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
4215 Factor.Width = Width;
4216 Factor.Cost = Width * Cost;
4220 unsigned LoopVectorizationCostModel::getWidestType() {
4221 unsigned MaxWidth = 8;
4224 for (Loop::block_iterator bb = TheLoop->block_begin(),
4225 be = TheLoop->block_end(); bb != be; ++bb) {
4226 BasicBlock *BB = *bb;
4228 // For each instruction in the loop.
4229 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4230 Type *T = it->getType();
4232 // Only examine Loads, Stores and PHINodes.
4233 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4236 // Examine PHI nodes that are reduction variables.
4237 if (PHINode *PN = dyn_cast<PHINode>(it))
4238 if (!Legal->getReductionVars()->count(PN))
4241 // Examine the stored values.
4242 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4243 T = ST->getValueOperand()->getType();
4245 // Ignore loaded pointer types and stored pointer types that are not
4246 // consecutive. However, we do want to take consecutive stores/loads of
4247 // pointer vectors into account.
4248 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4251 MaxWidth = std::max(MaxWidth,
4252 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4260 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4263 unsigned LoopCost) {
4265 // -- The unroll heuristics --
4266 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4267 // There are many micro-architectural considerations that we can't predict
4268 // at this level. For example frontend pressure (on decode or fetch) due to
4269 // code size, or the number and capabilities of the execution ports.
4271 // We use the following heuristics to select the unroll factor:
4272 // 1. If the code has reductions the we unroll in order to break the cross
4273 // iteration dependency.
4274 // 2. If the loop is really small then we unroll in order to reduce the loop
4276 // 3. We don't unroll if we think that we will spill registers to memory due
4277 // to the increased register pressure.
4279 // Use the user preference, unless 'auto' is selected.
4283 // When we optimize for size we don't unroll.
4287 // Do not unroll loops with a relatively small trip count.
4288 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
4289 TheLoop->getLoopLatch());
4290 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4293 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
4294 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
4295 " vector registers\n");
4297 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4298 // We divide by these constants so assume that we have at least one
4299 // instruction that uses at least one register.
4300 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4301 R.NumInstructions = std::max(R.NumInstructions, 1U);
4303 // We calculate the unroll factor using the following formula.
4304 // Subtract the number of loop invariants from the number of available
4305 // registers. These registers are used by all of the unrolled instances.
4306 // Next, divide the remaining registers by the number of registers that is
4307 // required by the loop, in order to estimate how many parallel instances
4308 // fit without causing spills.
4309 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
4311 // Clamp the unroll factor ranges to reasonable factors.
4312 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
4314 // If we did not calculate the cost for VF (because the user selected the VF)
4315 // then we calculate the cost of VF here.
4317 LoopCost = expectedCost(VF);
4319 // Clamp the calculated UF to be between the 1 and the max unroll factor
4320 // that the target allows.
4321 if (UF > MaxUnrollSize)
4326 if (Legal->getReductionVars()->size()) {
4327 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
4331 // We want to unroll tiny loops in order to reduce the loop overhead.
4332 // We assume that the cost overhead is 1 and we use the cost model
4333 // to estimate the cost of the loop and unroll until the cost of the
4334 // loop overhead is about 5% of the cost of the loop.
4335 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
4336 if (LoopCost < 20) {
4337 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
4338 unsigned NewUF = 20/LoopCost + 1;
4339 return std::min(NewUF, UF);
4342 DEBUG(dbgs() << "LV: Not Unrolling. \n");
4346 LoopVectorizationCostModel::RegisterUsage
4347 LoopVectorizationCostModel::calculateRegisterUsage() {
4348 // This function calculates the register usage by measuring the highest number
4349 // of values that are alive at a single location. Obviously, this is a very
4350 // rough estimation. We scan the loop in a topological order in order and
4351 // assign a number to each instruction. We use RPO to ensure that defs are
4352 // met before their users. We assume that each instruction that has in-loop
4353 // users starts an interval. We record every time that an in-loop value is
4354 // used, so we have a list of the first and last occurrences of each
4355 // instruction. Next, we transpose this data structure into a multi map that
4356 // holds the list of intervals that *end* at a specific location. This multi
4357 // map allows us to perform a linear search. We scan the instructions linearly
4358 // and record each time that a new interval starts, by placing it in a set.
4359 // If we find this value in the multi-map then we remove it from the set.
4360 // The max register usage is the maximum size of the set.
4361 // We also search for instructions that are defined outside the loop, but are
4362 // used inside the loop. We need this number separately from the max-interval
4363 // usage number because when we unroll, loop-invariant values do not take
4365 LoopBlocksDFS DFS(TheLoop);
4369 R.NumInstructions = 0;
4371 // Each 'key' in the map opens a new interval. The values
4372 // of the map are the index of the 'last seen' usage of the
4373 // instruction that is the key.
4374 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4375 // Maps instruction to its index.
4376 DenseMap<unsigned, Instruction*> IdxToInstr;
4377 // Marks the end of each interval.
4378 IntervalMap EndPoint;
4379 // Saves the list of instruction indices that are used in the loop.
4380 SmallSet<Instruction*, 8> Ends;
4381 // Saves the list of values that are used in the loop but are
4382 // defined outside the loop, such as arguments and constants.
4383 SmallPtrSet<Value*, 8> LoopInvariants;
4386 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4387 be = DFS.endRPO(); bb != be; ++bb) {
4388 R.NumInstructions += (*bb)->size();
4389 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4391 Instruction *I = it;
4392 IdxToInstr[Index++] = I;
4394 // Save the end location of each USE.
4395 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4396 Value *U = I->getOperand(i);
4397 Instruction *Instr = dyn_cast<Instruction>(U);
4399 // Ignore non-instruction values such as arguments, constants, etc.
4400 if (!Instr) continue;
4402 // If this instruction is outside the loop then record it and continue.
4403 if (!TheLoop->contains(Instr)) {
4404 LoopInvariants.insert(Instr);
4408 // Overwrite previous end points.
4409 EndPoint[Instr] = Index;
4415 // Saves the list of intervals that end with the index in 'key'.
4416 typedef SmallVector<Instruction*, 2> InstrList;
4417 DenseMap<unsigned, InstrList> TransposeEnds;
4419 // Transpose the EndPoints to a list of values that end at each index.
4420 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4422 TransposeEnds[it->second].push_back(it->first);
4424 SmallSet<Instruction*, 8> OpenIntervals;
4425 unsigned MaxUsage = 0;
4428 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4429 for (unsigned int i = 0; i < Index; ++i) {
4430 Instruction *I = IdxToInstr[i];
4431 // Ignore instructions that are never used within the loop.
4432 if (!Ends.count(I)) continue;
4434 // Remove all of the instructions that end at this location.
4435 InstrList &List = TransposeEnds[i];
4436 for (unsigned int j=0, e = List.size(); j < e; ++j)
4437 OpenIntervals.erase(List[j]);
4439 // Count the number of live interals.
4440 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4442 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4443 OpenIntervals.size() <<"\n");
4445 // Add the current instruction to the list of open intervals.
4446 OpenIntervals.insert(I);
4449 unsigned Invariant = LoopInvariants.size();
4450 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
4451 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
4452 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
4454 R.LoopInvariantRegs = Invariant;
4455 R.MaxLocalUsers = MaxUsage;
4459 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4463 for (Loop::block_iterator bb = TheLoop->block_begin(),
4464 be = TheLoop->block_end(); bb != be; ++bb) {
4465 unsigned BlockCost = 0;
4466 BasicBlock *BB = *bb;
4468 // For each instruction in the old loop.
4469 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4470 // Skip dbg intrinsics.
4471 if (isa<DbgInfoIntrinsic>(it))
4474 unsigned C = getInstructionCost(it, VF);
4476 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
4477 VF << " For instruction: "<< *it << "\n");
4480 // We assume that if-converted blocks have a 50% chance of being executed.
4481 // When the code is scalar then some of the blocks are avoided due to CF.
4482 // When the code is vectorized we execute all code paths.
4483 if (Legal->blockNeedsPredication(*bb) && VF == 1)
4493 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4494 // If we know that this instruction will remain uniform, check the cost of
4495 // the scalar version.
4496 if (Legal->isUniformAfterVectorization(I))
4499 Type *RetTy = I->getType();
4500 Type *VectorTy = ToVectorTy(RetTy, VF);
4502 // TODO: We need to estimate the cost of intrinsic calls.
4503 switch (I->getOpcode()) {
4504 case Instruction::GetElementPtr:
4505 // We mark this instruction as zero-cost because the cost of GEPs in
4506 // vectorized code depends on whether the corresponding memory instruction
4507 // is scalarized or not. Therefore, we handle GEPs with the memory
4508 // instruction cost.
4510 case Instruction::Br: {
4511 return TTI.getCFInstrCost(I->getOpcode());
4513 case Instruction::PHI:
4514 //TODO: IF-converted IFs become selects.
4516 case Instruction::Add:
4517 case Instruction::FAdd:
4518 case Instruction::Sub:
4519 case Instruction::FSub:
4520 case Instruction::Mul:
4521 case Instruction::FMul:
4522 case Instruction::UDiv:
4523 case Instruction::SDiv:
4524 case Instruction::FDiv:
4525 case Instruction::URem:
4526 case Instruction::SRem:
4527 case Instruction::FRem:
4528 case Instruction::Shl:
4529 case Instruction::LShr:
4530 case Instruction::AShr:
4531 case Instruction::And:
4532 case Instruction::Or:
4533 case Instruction::Xor: {
4534 // Certain instructions can be cheaper to vectorize if they have a constant
4535 // second vector operand. One example of this are shifts on x86.
4536 TargetTransformInfo::OperandValueKind Op1VK =
4537 TargetTransformInfo::OK_AnyValue;
4538 TargetTransformInfo::OperandValueKind Op2VK =
4539 TargetTransformInfo::OK_AnyValue;
4541 if (isa<ConstantInt>(I->getOperand(1)))
4542 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4544 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
4546 case Instruction::Select: {
4547 SelectInst *SI = cast<SelectInst>(I);
4548 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4549 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4550 Type *CondTy = SI->getCondition()->getType();
4552 CondTy = VectorType::get(CondTy, VF);
4554 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4556 case Instruction::ICmp:
4557 case Instruction::FCmp: {
4558 Type *ValTy = I->getOperand(0)->getType();
4559 VectorTy = ToVectorTy(ValTy, VF);
4560 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4562 case Instruction::Store:
4563 case Instruction::Load: {
4564 StoreInst *SI = dyn_cast<StoreInst>(I);
4565 LoadInst *LI = dyn_cast<LoadInst>(I);
4566 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4568 VectorTy = ToVectorTy(ValTy, VF);
4570 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4571 unsigned AS = SI ? SI->getPointerAddressSpace() :
4572 LI->getPointerAddressSpace();
4573 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4574 // We add the cost of address computation here instead of with the gep
4575 // instruction because only here we know whether the operation is
4578 return TTI.getAddressComputationCost(VectorTy) +
4579 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4581 // Scalarized loads/stores.
4582 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4583 bool Reverse = ConsecutiveStride < 0;
4584 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4585 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4586 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4588 // The cost of extracting from the value vector and pointer vector.
4589 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4590 for (unsigned i = 0; i < VF; ++i) {
4591 // The cost of extracting the pointer operand.
4592 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4593 // In case of STORE, the cost of ExtractElement from the vector.
4594 // In case of LOAD, the cost of InsertElement into the returned
4596 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4597 Instruction::InsertElement,
4601 // The cost of the scalar loads/stores.
4602 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
4603 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4608 // Wide load/stores.
4609 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4610 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4613 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4617 case Instruction::ZExt:
4618 case Instruction::SExt:
4619 case Instruction::FPToUI:
4620 case Instruction::FPToSI:
4621 case Instruction::FPExt:
4622 case Instruction::PtrToInt:
4623 case Instruction::IntToPtr:
4624 case Instruction::SIToFP:
4625 case Instruction::UIToFP:
4626 case Instruction::Trunc:
4627 case Instruction::FPTrunc:
4628 case Instruction::BitCast: {
4629 // We optimize the truncation of induction variable.
4630 // The cost of these is the same as the scalar operation.
4631 if (I->getOpcode() == Instruction::Trunc &&
4632 Legal->isInductionVariable(I->getOperand(0)))
4633 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4634 I->getOperand(0)->getType());
4636 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4637 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4639 case Instruction::Call: {
4640 CallInst *CI = cast<CallInst>(I);
4641 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4642 assert(ID && "Not an intrinsic call!");
4643 Type *RetTy = ToVectorTy(CI->getType(), VF);
4644 SmallVector<Type*, 4> Tys;
4645 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4646 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4647 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4650 // We are scalarizing the instruction. Return the cost of the scalar
4651 // instruction, plus the cost of insert and extract into vector
4652 // elements, times the vector width.
4655 if (!RetTy->isVoidTy() && VF != 1) {
4656 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4658 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4661 // The cost of inserting the results plus extracting each one of the
4663 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4666 // The cost of executing VF copies of the scalar instruction. This opcode
4667 // is unknown. Assume that it is the same as 'mul'.
4668 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4674 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4675 if (Scalar->isVoidTy() || VF == 1)
4677 return VectorType::get(Scalar, VF);
4680 char LoopVectorize::ID = 0;
4681 static const char lv_name[] = "Loop Vectorization";
4682 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4683 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
4684 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4685 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4686 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4687 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4690 Pass *createLoopVectorizePass() {
4691 return new LoopVectorize();
4695 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4696 // Check for a store.
4697 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
4698 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
4700 // Check for a load.
4701 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
4702 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;