1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR. Legalization of the IR is done
12 // in the codegen. However, the vectorizer uses (will use) the codegen
13 // interfaces to generate IR that is likely to result in an optimal binary.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/MapVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/StringExtras.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/AliasSetTracker.h"
57 #include "llvm/Analysis/Dominators.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopIterator.h"
60 #include "llvm/Analysis/LoopPass.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpander.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/TargetTransformInfo.h"
65 #include "llvm/Analysis/ValueTracking.h"
66 #include "llvm/Analysis/Verifier.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Function.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/LLVMContext.h"
75 #include "llvm/IR/Module.h"
76 #include "llvm/IR/Type.h"
77 #include "llvm/IR/Value.h"
78 #include "llvm/Pass.h"
79 #include "llvm/Support/CommandLine.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/PatternMatch.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/Target/TargetLibraryInfo.h"
84 #include "llvm/Transforms/Scalar.h"
85 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
86 #include "llvm/Transforms/Utils/Local.h"
91 using namespace llvm::PatternMatch;
93 static cl::opt<unsigned>
94 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
95 cl::desc("Sets the SIMD width. Zero is autoselect."));
97 static cl::opt<unsigned>
98 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
99 cl::desc("Sets the vectorization unroll count. "
100 "Zero is autoselect."));
103 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
104 cl::desc("Enable if-conversion during vectorization."));
106 /// We don't vectorize loops with a known constant trip count below this number.
107 static cl::opt<unsigned>
108 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
110 cl::desc("Don't vectorize loops with a constant "
111 "trip count that is smaller than this "
114 /// We don't unroll loops with a known constant trip count below this number.
115 static const unsigned TinyTripCountUnrollThreshold = 128;
117 /// When performing a runtime memory check, do not check more than this
118 /// number of pointers. Notice that the check is quadratic!
119 static const unsigned RuntimeMemoryCheckThreshold = 4;
121 /// We use a metadata with this name to indicate that a scalar loop was
122 /// vectorized and that we don't need to re-vectorize it if we run into it
125 AlreadyVectorizedMDName = "llvm.vectorizer.already_vectorized";
129 // Forward declarations.
130 class LoopVectorizationLegality;
131 class LoopVectorizationCostModel;
133 /// InnerLoopVectorizer vectorizes loops which contain only one basic
134 /// block to a specified vectorization factor (VF).
135 /// This class performs the widening of scalars into vectors, or multiple
136 /// scalars. This class also implements the following features:
137 /// * It inserts an epilogue loop for handling loops that don't have iteration
138 /// counts that are known to be a multiple of the vectorization factor.
139 /// * It handles the code generation for reduction variables.
140 /// * Scalarization (implementation using scalars) of un-vectorizable
142 /// InnerLoopVectorizer does not perform any vectorization-legality
143 /// checks, and relies on the caller to check for the different legality
144 /// aspects. The InnerLoopVectorizer relies on the
145 /// LoopVectorizationLegality class to provide information about the induction
146 /// and reduction variables that were found to a given vectorization factor.
147 class InnerLoopVectorizer {
149 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
150 DominatorTree *DT, DataLayout *DL,
151 const TargetLibraryInfo *TLI, unsigned VecWidth,
152 unsigned UnrollFactor)
153 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
154 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
155 OldInduction(0), WidenMap(UnrollFactor) {}
157 // Perform the actual loop widening (vectorization).
158 void vectorize(LoopVectorizationLegality *Legal) {
159 // Create a new empty loop. Unlink the old loop and connect the new one.
160 createEmptyLoop(Legal);
161 // Widen each instruction in the old loop to a new one in the new loop.
162 // Use the Legality module to find the induction and reduction variables.
163 vectorizeLoop(Legal);
164 // Register the new loop and update the analysis passes.
169 /// A small list of PHINodes.
170 typedef SmallVector<PHINode*, 4> PhiVector;
171 /// When we unroll loops we have multiple vector values for each scalar.
172 /// This data structure holds the unrolled and vectorized values that
173 /// originated from one scalar instruction.
174 typedef SmallVector<Value*, 2> VectorParts;
176 /// Add code that checks at runtime if the accessed arrays overlap.
177 /// Returns the comparator value or NULL if no check is needed.
178 Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
180 /// Create an empty loop, based on the loop ranges of the old loop.
181 void createEmptyLoop(LoopVectorizationLegality *Legal);
182 /// Copy and widen the instructions from the old loop.
183 void vectorizeLoop(LoopVectorizationLegality *Legal);
185 /// A helper function that computes the predicate of the block BB, assuming
186 /// that the header block of the loop is set to True. It returns the *entry*
187 /// mask for the block BB.
188 VectorParts createBlockInMask(BasicBlock *BB);
189 /// A helper function that computes the predicate of the edge between SRC
191 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
193 /// A helper function to vectorize a single BB within the innermost loop.
194 void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
197 /// Insert the new loop to the loop hierarchy and pass manager
198 /// and update the analysis passes.
199 void updateAnalysis();
201 /// This instruction is un-vectorizable. Implement it as a sequence
203 void scalarizeInstruction(Instruction *Instr);
205 /// Vectorize Load and Store instructions,
206 void vectorizeMemoryInstruction(Instruction *Instr,
207 LoopVectorizationLegality *Legal);
209 /// Create a broadcast instruction. This method generates a broadcast
210 /// instruction (shuffle) for loop invariant values and for the induction
211 /// value. If this is the induction variable then we extend it to N, N+1, ...
212 /// this is needed because each iteration in the loop corresponds to a SIMD
214 Value *getBroadcastInstrs(Value *V);
216 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
217 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
218 /// The sequence starts at StartIndex.
219 Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
221 /// When we go over instructions in the basic block we rely on previous
222 /// values within the current basic block or on loop invariant values.
223 /// When we widen (vectorize) values we place them in the map. If the values
224 /// are not within the map, they have to be loop invariant, so we simply
225 /// broadcast them into a vector.
226 VectorParts &getVectorValue(Value *V);
228 /// Generate a shuffle sequence that will reverse the vector Vec.
229 Value *reverseVector(Value *Vec);
231 /// This is a helper class that holds the vectorizer state. It maps scalar
232 /// instructions to vector instructions. When the code is 'unrolled' then
233 /// then a single scalar value is mapped to multiple vector parts. The parts
234 /// are stored in the VectorPart type.
236 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
238 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
240 /// \return True if 'Key' is saved in the Value Map.
241 bool has(Value *Key) const { return MapStorage.count(Key); }
243 /// Initializes a new entry in the map. Sets all of the vector parts to the
244 /// save value in 'Val'.
245 /// \return A reference to a vector with splat values.
246 VectorParts &splat(Value *Key, Value *Val) {
247 VectorParts &Entry = MapStorage[Key];
248 Entry.assign(UF, Val);
252 ///\return A reference to the value that is stored at 'Key'.
253 VectorParts &get(Value *Key) {
254 VectorParts &Entry = MapStorage[Key];
257 assert(Entry.size() == UF);
262 /// The unroll factor. Each entry in the map stores this number of vector
266 /// Map storage. We use std::map and not DenseMap because insertions to a
267 /// dense map invalidates its iterators.
268 std::map<Value *, VectorParts> MapStorage;
271 /// The original loop.
273 /// Scev analysis to use.
281 /// Target Library Info.
282 const TargetLibraryInfo *TLI;
284 /// The vectorization SIMD factor to use. Each vector will have this many
287 /// The vectorization unroll factor to use. Each scalar is vectorized to this
288 /// many different vector instructions.
291 /// The builder that we use
294 // --- Vectorization state ---
296 /// The vector-loop preheader.
297 BasicBlock *LoopVectorPreHeader;
298 /// The scalar-loop preheader.
299 BasicBlock *LoopScalarPreHeader;
300 /// Middle Block between the vector and the scalar.
301 BasicBlock *LoopMiddleBlock;
302 ///The ExitBlock of the scalar loop.
303 BasicBlock *LoopExitBlock;
304 ///The vector loop body.
305 BasicBlock *LoopVectorBody;
306 ///The scalar loop body.
307 BasicBlock *LoopScalarBody;
308 /// A list of all bypass blocks. The first block is the entry of the loop.
309 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
311 /// The new Induction variable which was added to the new block.
313 /// The induction variable of the old basic block.
314 PHINode *OldInduction;
315 /// Maps scalars to widened vectors.
319 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
320 /// to what vectorization factor.
321 /// This class does not look at the profitability of vectorization, only the
322 /// legality. This class has two main kinds of checks:
323 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
324 /// will change the order of memory accesses in a way that will change the
325 /// correctness of the program.
326 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
327 /// checks for a number of different conditions, such as the availability of a
328 /// single induction variable, that all types are supported and vectorize-able,
329 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
330 /// This class is also used by InnerLoopVectorizer for identifying
331 /// induction variable and the different reduction variables.
332 class LoopVectorizationLegality {
334 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
335 DominatorTree *DT, TargetTransformInfo* TTI,
336 AliasAnalysis *AA, TargetLibraryInfo *TLI)
337 : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
340 /// This enum represents the kinds of reductions that we support.
342 RK_NoReduction, ///< Not a reduction.
343 RK_IntegerAdd, ///< Sum of integers.
344 RK_IntegerMult, ///< Product of integers.
345 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
346 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
347 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
348 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
349 RK_FloatAdd, ///< Sum of floats.
350 RK_FloatMult ///< Product of floats.
353 /// This enum represents the kinds of inductions that we support.
355 IK_NoInduction, ///< Not an induction variable.
356 IK_IntInduction, ///< Integer induction variable. Step = 1.
357 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
358 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
359 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
362 // This enum represents the kind of minmax reduction.
363 enum MinMaxReductionKind {
371 /// This POD struct holds information about reduction variables.
372 struct ReductionDescriptor {
373 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
374 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
376 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
377 MinMaxReductionKind MK)
378 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
380 // The starting value of the reduction.
381 // It does not have to be zero!
383 // The instruction who's value is used outside the loop.
384 Instruction *LoopExitInstr;
385 // The kind of the reduction.
387 // If this a min/max reduction the kind of reduction.
388 MinMaxReductionKind MinMaxKind;
391 /// This POD struct holds information about a potential reduction operation.
392 struct ReductionInstDesc {
393 ReductionInstDesc(bool IsRedux, Instruction *I) :
394 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
396 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
397 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
399 // Is this instruction a reduction candidate.
401 // The last instruction in a min/max pattern (select of the select(icmp())
402 // pattern), or the current reduction instruction otherwise.
403 Instruction *PatternLastInst;
404 // If this is a min/max pattern the comparison predicate.
405 MinMaxReductionKind MinMaxKind;
408 // This POD struct holds information about the memory runtime legality
409 // check that a group of pointers do not overlap.
410 struct RuntimePointerCheck {
411 RuntimePointerCheck() : Need(false) {}
413 /// Reset the state of the pointer runtime information.
421 /// Insert a pointer and calculate the start and end SCEVs.
422 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr);
424 /// This flag indicates if we need to add the runtime check.
426 /// Holds the pointers that we need to check.
427 SmallVector<Value*, 2> Pointers;
428 /// Holds the pointer value at the beginning of the loop.
429 SmallVector<const SCEV*, 2> Starts;
430 /// Holds the pointer value at the end of the loop.
431 SmallVector<const SCEV*, 2> Ends;
434 /// A POD for saving information about induction variables.
435 struct InductionInfo {
436 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
437 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
444 /// ReductionList contains the reduction descriptors for all
445 /// of the reductions that were found in the loop.
446 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
448 /// InductionList saves induction variables and maps them to the
449 /// induction descriptor.
450 typedef MapVector<PHINode*, InductionInfo> InductionList;
452 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
453 /// respective Store/Load instruction(s) to calculate aliasing.
454 typedef MapVector<Value*, Instruction* > AliasMap;
455 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
457 /// Returns true if it is legal to vectorize this loop.
458 /// This does not mean that it is profitable to vectorize this
459 /// loop, only that it is legal to do so.
462 /// Returns the Induction variable.
463 PHINode *getInduction() { return Induction; }
465 /// Returns the reduction variables found in the loop.
466 ReductionList *getReductionVars() { return &Reductions; }
468 /// Returns the induction variables found in the loop.
469 InductionList *getInductionVars() { return &Inductions; }
471 /// Returns True if V is an induction variable in this loop.
472 bool isInductionVariable(const Value *V);
474 /// Return true if the block BB needs to be predicated in order for the loop
475 /// to be vectorized.
476 bool blockNeedsPredication(BasicBlock *BB);
478 /// Check if this pointer is consecutive when vectorizing. This happens
479 /// when the last index of the GEP is the induction variable, or that the
480 /// pointer itself is an induction variable.
481 /// This check allows us to vectorize A[idx] into a wide load/store.
483 /// 0 - Stride is unknown or non consecutive.
484 /// 1 - Address is consecutive.
485 /// -1 - Address is consecutive, and decreasing.
486 int isConsecutivePtr(Value *Ptr);
488 /// Returns true if the value V is uniform within the loop.
489 bool isUniform(Value *V);
491 /// Returns true if this instruction will remain scalar after vectorization.
492 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
494 /// Returns the information that we collected about runtime memory check.
495 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
497 /// This function returns the identity element (or neutral element) for
499 static Constant *getReductionIdentity(ReductionKind K, Type *Tp,
500 MinMaxReductionKind MinMaxK);
502 /// Check if a single basic block loop is vectorizable.
503 /// At this point we know that this is a loop with a constant trip count
504 /// and we only need to check individual instructions.
505 bool canVectorizeInstrs();
507 /// When we vectorize loops we may change the order in which
508 /// we read and write from memory. This method checks if it is
509 /// legal to vectorize the code, considering only memory constrains.
510 /// Returns true if the loop is vectorizable
511 bool canVectorizeMemory();
513 /// Return true if we can vectorize this loop using the IF-conversion
515 bool canVectorizeWithIfConvert();
517 /// Collect the variables that need to stay uniform after vectorization.
518 void collectLoopUniforms();
520 /// Return true if all of the instructions in the block can be speculatively
522 bool blockCanBePredicated(BasicBlock *BB);
524 /// Returns True, if 'Phi' is the kind of reduction variable for type
525 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
526 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
527 /// Returns a struct describing if the instruction 'I' can be a reduction
528 /// variable of type 'Kind'. If the reduction is a min/max pattern of
529 /// select(icmp()) this function advances the instruction pointer 'I' from the
530 /// compare instruction to the select instruction and stores this pointer in
531 /// 'PatternLastInst' member of the returned struct.
532 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
533 ReductionInstDesc &Desc);
534 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
535 /// pattern corresponding to a min(X, Y) or max(X, Y).
536 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
537 ReductionInstDesc &Prev);
538 /// Returns the induction kind of Phi. This function may return NoInduction
539 /// if the PHI is not an induction variable.
540 InductionKind isInductionVariable(PHINode *Phi);
541 /// Return true if can compute the address bounds of Ptr within the loop.
542 bool hasComputableBounds(Value *Ptr);
543 /// Return true if there is the chance of write reorder.
544 bool hasPossibleGlobalWriteReorder(Value *Object,
546 AliasMultiMap &WriteObjects,
547 unsigned MaxByteWidth);
548 /// Return the AA location for a load or a store.
549 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
552 /// The loop that we evaluate.
556 /// DataLayout analysis.
561 TargetTransformInfo *TTI;
564 /// Target Library Info.
565 TargetLibraryInfo *TLI;
567 // --- vectorization state --- //
569 /// Holds the integer induction variable. This is the counter of the
572 /// Holds the reduction variables.
573 ReductionList Reductions;
574 /// Holds all of the induction variables that we found in the loop.
575 /// Notice that inductions don't need to start at zero and that induction
576 /// variables can be pointers.
577 InductionList Inductions;
579 /// Allowed outside users. This holds the reduction
580 /// vars which can be accessed from outside the loop.
581 SmallPtrSet<Value*, 4> AllowedExit;
582 /// This set holds the variables which are known to be uniform after
584 SmallPtrSet<Instruction*, 4> Uniforms;
585 /// We need to check that all of the pointers in this list are disjoint
587 RuntimePointerCheck PtrRtCheck;
590 /// LoopVectorizationCostModel - estimates the expected speedups due to
592 /// In many cases vectorization is not profitable. This can happen because of
593 /// a number of reasons. In this class we mainly attempt to predict the
594 /// expected speedup/slowdowns due to the supported instruction set. We use the
595 /// TargetTransformInfo to query the different backends for the cost of
596 /// different operations.
597 class LoopVectorizationCostModel {
599 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
600 LoopVectorizationLegality *Legal,
601 const TargetTransformInfo &TTI,
602 DataLayout *DL, const TargetLibraryInfo *TLI)
603 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
605 /// Information about vectorization costs
606 struct VectorizationFactor {
607 unsigned Width; // Vector width with best cost
608 unsigned Cost; // Cost of the loop with that width
610 /// \return The most profitable vectorization factor and the cost of that VF.
611 /// This method checks every power of two up to VF. If UserVF is not ZERO
612 /// then this vectorization factor will be selected if vectorization is
614 VectorizationFactor selectVectorizationFactor(bool OptForSize,
617 /// \return The size (in bits) of the widest type in the code that
618 /// needs to be vectorized. We ignore values that remain scalar such as
619 /// 64 bit loop indices.
620 unsigned getWidestType();
622 /// \return The most profitable unroll factor.
623 /// If UserUF is non-zero then this method finds the best unroll-factor
624 /// based on register pressure and other parameters.
625 /// VF and LoopCost are the selected vectorization factor and the cost of the
627 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
630 /// \brief A struct that represents some properties of the register usage
632 struct RegisterUsage {
633 /// Holds the number of loop invariant values that are used in the loop.
634 unsigned LoopInvariantRegs;
635 /// Holds the maximum number of concurrent live intervals in the loop.
636 unsigned MaxLocalUsers;
637 /// Holds the number of instructions in the loop.
638 unsigned NumInstructions;
641 /// \return information about the register usage of the loop.
642 RegisterUsage calculateRegisterUsage();
645 /// Returns the expected execution cost. The unit of the cost does
646 /// not matter because we use the 'cost' units to compare different
647 /// vector widths. The cost that is returned is *not* normalized by
648 /// the factor width.
649 unsigned expectedCost(unsigned VF);
651 /// Returns the execution time cost of an instruction for a given vector
652 /// width. Vector width of one means scalar.
653 unsigned getInstructionCost(Instruction *I, unsigned VF);
655 /// A helper function for converting Scalar types to vector types.
656 /// If the incoming type is void, we return void. If the VF is 1, we return
658 static Type* ToVectorTy(Type *Scalar, unsigned VF);
660 /// Returns whether the instruction is a load or store and will be a emitted
661 /// as a vector operation.
662 bool isConsecutiveLoadOrStore(Instruction *I);
664 /// The loop that we evaluate.
668 /// Loop Info analysis.
670 /// Vectorization legality.
671 LoopVectorizationLegality *Legal;
672 /// Vector target information.
673 const TargetTransformInfo &TTI;
674 /// Target data layout information.
676 /// Target Library Info.
677 const TargetLibraryInfo *TLI;
680 /// The LoopVectorize Pass.
681 struct LoopVectorize : public LoopPass {
682 /// Pass identification, replacement for typeid
685 explicit LoopVectorize() : LoopPass(ID) {
686 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
692 TargetTransformInfo *TTI;
695 TargetLibraryInfo *TLI;
697 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
698 // We only vectorize innermost loops.
702 SE = &getAnalysis<ScalarEvolution>();
703 DL = getAnalysisIfAvailable<DataLayout>();
704 LI = &getAnalysis<LoopInfo>();
705 TTI = &getAnalysis<TargetTransformInfo>();
706 DT = &getAnalysis<DominatorTree>();
707 AA = getAnalysisIfAvailable<AliasAnalysis>();
708 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
711 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
715 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
716 L->getHeader()->getParent()->getName() << "\"\n");
718 // Check if it is legal to vectorize the loop.
719 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
720 if (!LVL.canVectorize()) {
721 DEBUG(dbgs() << "LV: Not vectorizing.\n");
725 // Use the cost model.
726 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
728 // Check the function attributes to find out if this function should be
729 // optimized for size.
730 Function *F = L->getHeader()->getParent();
731 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
732 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
733 unsigned FnIndex = AttributeSet::FunctionIndex;
734 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
735 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
738 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
739 "attribute is used.\n");
743 // Select the optimal vectorization factor.
744 LoopVectorizationCostModel::VectorizationFactor VF;
745 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
746 // Select the unroll factor.
747 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
751 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
755 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
756 F->getParent()->getModuleIdentifier()<<"\n");
757 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
759 // If we decided that it is *legal* to vectorize the loop then do it.
760 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
763 DEBUG(verifyFunction(*L->getHeader()->getParent()));
767 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
768 LoopPass::getAnalysisUsage(AU);
769 AU.addRequiredID(LoopSimplifyID);
770 AU.addRequiredID(LCSSAID);
771 AU.addRequired<DominatorTree>();
772 AU.addRequired<LoopInfo>();
773 AU.addRequired<ScalarEvolution>();
774 AU.addRequired<TargetTransformInfo>();
775 AU.addPreserved<LoopInfo>();
776 AU.addPreserved<DominatorTree>();
781 } // end anonymous namespace
783 //===----------------------------------------------------------------------===//
784 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
785 // LoopVectorizationCostModel.
786 //===----------------------------------------------------------------------===//
789 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
790 Loop *Lp, Value *Ptr) {
791 const SCEV *Sc = SE->getSCEV(Ptr);
792 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
793 assert(AR && "Invalid addrec expression");
794 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
795 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
796 Pointers.push_back(Ptr);
797 Starts.push_back(AR->getStart());
798 Ends.push_back(ScEnd);
801 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
802 // Save the current insertion location.
803 Instruction *Loc = Builder.GetInsertPoint();
805 // We need to place the broadcast of invariant variables outside the loop.
806 Instruction *Instr = dyn_cast<Instruction>(V);
807 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
808 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
810 // Place the code for broadcasting invariant variables in the new preheader.
812 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
814 // Broadcast the scalar into all locations in the vector.
815 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
817 // Restore the builder insertion point.
819 Builder.SetInsertPoint(Loc);
824 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
826 assert(Val->getType()->isVectorTy() && "Must be a vector");
827 assert(Val->getType()->getScalarType()->isIntegerTy() &&
828 "Elem must be an integer");
830 Type *ITy = Val->getType()->getScalarType();
831 VectorType *Ty = cast<VectorType>(Val->getType());
832 int VLen = Ty->getNumElements();
833 SmallVector<Constant*, 8> Indices;
835 // Create a vector of consecutive numbers from zero to VF.
836 for (int i = 0; i < VLen; ++i) {
837 int Idx = Negate ? (-i): i;
838 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
841 // Add the consecutive indices to the vector value.
842 Constant *Cv = ConstantVector::get(Indices);
843 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
844 return Builder.CreateAdd(Val, Cv, "induction");
847 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
848 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
849 // Make sure that the pointer does not point to structs.
850 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
853 // If this value is a pointer induction variable we know it is consecutive.
854 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
855 if (Phi && Inductions.count(Phi)) {
856 InductionInfo II = Inductions[Phi];
857 if (IK_PtrInduction == II.IK)
859 else if (IK_ReversePtrInduction == II.IK)
863 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
867 unsigned NumOperands = Gep->getNumOperands();
868 Value *LastIndex = Gep->getOperand(NumOperands - 1);
870 Value *GpPtr = Gep->getPointerOperand();
871 // If this GEP value is a consecutive pointer induction variable and all of
872 // the indices are constant then we know it is consecutive. We can
873 Phi = dyn_cast<PHINode>(GpPtr);
874 if (Phi && Inductions.count(Phi)) {
876 // Make sure that the pointer does not point to structs.
877 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
878 if (GepPtrType->getElementType()->isAggregateType())
881 // Make sure that all of the index operands are loop invariant.
882 for (unsigned i = 1; i < NumOperands; ++i)
883 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
886 InductionInfo II = Inductions[Phi];
887 if (IK_PtrInduction == II.IK)
889 else if (IK_ReversePtrInduction == II.IK)
893 // Check that all of the gep indices are uniform except for the last.
894 for (unsigned i = 0; i < NumOperands - 1; ++i)
895 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
898 // We can emit wide load/stores only if the last index is the induction
900 const SCEV *Last = SE->getSCEV(LastIndex);
901 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
902 const SCEV *Step = AR->getStepRecurrence(*SE);
904 // The memory is consecutive because the last index is consecutive
905 // and all other indices are loop invariant.
908 if (Step->isAllOnesValue())
915 bool LoopVectorizationLegality::isUniform(Value *V) {
916 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
919 InnerLoopVectorizer::VectorParts&
920 InnerLoopVectorizer::getVectorValue(Value *V) {
921 assert(V != Induction && "The new induction variable should not be used.");
922 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
924 // If we have this scalar in the map, return it.
926 return WidenMap.get(V);
928 // If this scalar is unknown, assume that it is a constant or that it is
929 // loop invariant. Broadcast V and save the value for future uses.
930 Value *B = getBroadcastInstrs(V);
931 return WidenMap.splat(V, B);
934 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
935 assert(Vec->getType()->isVectorTy() && "Invalid type");
936 SmallVector<Constant*, 8> ShuffleMask;
937 for (unsigned i = 0; i < VF; ++i)
938 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
940 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
941 ConstantVector::get(ShuffleMask),
946 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
947 LoopVectorizationLegality *Legal) {
948 // Attempt to issue a wide load.
949 LoadInst *LI = dyn_cast<LoadInst>(Instr);
950 StoreInst *SI = dyn_cast<StoreInst>(Instr);
952 assert((LI || SI) && "Invalid Load/Store instruction");
954 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
955 Type *DataTy = VectorType::get(ScalarDataTy, VF);
956 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
957 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
959 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
960 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
962 if (ScalarAllocatedSize != VectorElementSize)
963 return scalarizeInstruction(Instr);
965 // If the pointer is loop invariant or if it is non consecutive,
966 // scalarize the load.
967 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
968 bool Reverse = ConsecutiveStride < 0;
969 bool UniformLoad = LI && Legal->isUniform(Ptr);
970 if (!ConsecutiveStride || UniformLoad)
971 return scalarizeInstruction(Instr);
973 Constant *Zero = Builder.getInt32(0);
974 VectorParts &Entry = WidenMap.get(Instr);
976 // Handle consecutive loads/stores.
977 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
978 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
979 Value *PtrOperand = Gep->getPointerOperand();
980 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
981 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
983 // Create the new GEP with the new induction variable.
984 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
985 Gep2->setOperand(0, FirstBasePtr);
986 Gep2->setName("gep.indvar.base");
987 Ptr = Builder.Insert(Gep2);
989 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
990 OrigLoop) && "Base ptr must be invariant");
992 // The last index does not have to be the induction. It can be
993 // consecutive and be a function of the index. For example A[I+1];
994 unsigned NumOperands = Gep->getNumOperands();
996 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
997 VectorParts &GEPParts = getVectorValue(LastGepOperand);
998 Value *LastIndex = GEPParts[0];
999 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1001 // Create the new GEP with the new induction variable.
1002 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1003 Gep2->setOperand(NumOperands - 1, LastIndex);
1004 Gep2->setName("gep.indvar.idx");
1005 Ptr = Builder.Insert(Gep2);
1007 // Use the induction element ptr.
1008 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1009 VectorParts &PtrVal = getVectorValue(Ptr);
1010 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1015 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1016 "We do not allow storing to uniform addresses");
1018 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1019 for (unsigned Part = 0; Part < UF; ++Part) {
1020 // Calculate the pointer for the specific unroll-part.
1021 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1024 // If we store to reverse consecutive memory locations then we need
1025 // to reverse the order of elements in the stored value.
1026 StoredVal[Part] = reverseVector(StoredVal[Part]);
1027 // If the address is consecutive but reversed, then the
1028 // wide store needs to start at the last vector element.
1029 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1030 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1033 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1034 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1038 for (unsigned Part = 0; Part < UF; ++Part) {
1039 // Calculate the pointer for the specific unroll-part.
1040 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1043 // If the address is consecutive but reversed, then the
1044 // wide store needs to start at the last vector element.
1045 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1046 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1049 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1050 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1051 cast<LoadInst>(LI)->setAlignment(Alignment);
1052 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1056 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1057 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1058 // Holds vector parameters or scalars, in case of uniform vals.
1059 SmallVector<VectorParts, 4> Params;
1061 // Find all of the vectorized parameters.
1062 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1063 Value *SrcOp = Instr->getOperand(op);
1065 // If we are accessing the old induction variable, use the new one.
1066 if (SrcOp == OldInduction) {
1067 Params.push_back(getVectorValue(SrcOp));
1071 // Try using previously calculated values.
1072 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1074 // If the src is an instruction that appeared earlier in the basic block
1075 // then it should already be vectorized.
1076 if (SrcInst && OrigLoop->contains(SrcInst)) {
1077 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1078 // The parameter is a vector value from earlier.
1079 Params.push_back(WidenMap.get(SrcInst));
1081 // The parameter is a scalar from outside the loop. Maybe even a constant.
1082 VectorParts Scalars;
1083 Scalars.append(UF, SrcOp);
1084 Params.push_back(Scalars);
1088 assert(Params.size() == Instr->getNumOperands() &&
1089 "Invalid number of operands");
1091 // Does this instruction return a value ?
1092 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1094 Value *UndefVec = IsVoidRetTy ? 0 :
1095 UndefValue::get(VectorType::get(Instr->getType(), VF));
1096 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1097 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1099 // For each vector unroll 'part':
1100 for (unsigned Part = 0; Part < UF; ++Part) {
1101 // For each scalar that we create:
1102 for (unsigned Width = 0; Width < VF; ++Width) {
1103 Instruction *Cloned = Instr->clone();
1105 Cloned->setName(Instr->getName() + ".cloned");
1106 // Replace the operands of the cloned instrucions with extracted scalars.
1107 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1108 Value *Op = Params[op][Part];
1109 // Param is a vector. Need to extract the right lane.
1110 if (Op->getType()->isVectorTy())
1111 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1112 Cloned->setOperand(op, Op);
1115 // Place the cloned scalar in the new loop.
1116 Builder.Insert(Cloned);
1118 // If the original scalar returns a value we need to place it in a vector
1119 // so that future users will be able to use it.
1121 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1122 Builder.getInt32(Width));
1128 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1130 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1131 Legal->getRuntimePointerCheck();
1133 if (!PtrRtCheck->Need)
1136 Instruction *MemoryRuntimeCheck = 0;
1137 unsigned NumPointers = PtrRtCheck->Pointers.size();
1138 SmallVector<Value* , 2> Starts;
1139 SmallVector<Value* , 2> Ends;
1141 SCEVExpander Exp(*SE, "induction");
1143 // Use this type for pointer arithmetic.
1144 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1146 for (unsigned i = 0; i < NumPointers; ++i) {
1147 Value *Ptr = PtrRtCheck->Pointers[i];
1148 const SCEV *Sc = SE->getSCEV(Ptr);
1150 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1151 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1153 Starts.push_back(Ptr);
1154 Ends.push_back(Ptr);
1156 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1158 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1159 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1160 Starts.push_back(Start);
1161 Ends.push_back(End);
1165 IRBuilder<> ChkBuilder(Loc);
1167 for (unsigned i = 0; i < NumPointers; ++i) {
1168 for (unsigned j = i+1; j < NumPointers; ++j) {
1169 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1170 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1171 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1172 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1174 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1175 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1176 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1177 if (MemoryRuntimeCheck)
1178 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1181 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1185 return MemoryRuntimeCheck;
1189 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1191 In this function we generate a new loop. The new loop will contain
1192 the vectorized instructions while the old loop will continue to run the
1195 [ ] <-- vector loop bypass (may consist of multiple blocks).
1198 | [ ] <-- vector pre header.
1202 | [ ]_| <-- vector loop.
1205 >[ ] <--- middle-block.
1208 | [ ] <--- new preheader.
1212 | [ ]_| <-- old scalar loop to handle remainder.
1215 >[ ] <-- exit block.
1219 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1220 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1221 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1222 assert(ExitBlock && "Must have an exit block");
1224 // Mark the old scalar loop with metadata that tells us not to vectorize this
1225 // loop again if we run into it.
1226 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), ArrayRef<Value*>());
1227 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD);
1229 // Some loops have a single integer induction variable, while other loops
1230 // don't. One example is c++ iterators that often have multiple pointer
1231 // induction variables. In the code below we also support a case where we
1232 // don't have a single induction variable.
1233 OldInduction = Legal->getInduction();
1234 Type *IdxTy = OldInduction ? OldInduction->getType() :
1235 DL->getIntPtrType(SE->getContext());
1237 // Find the loop boundaries.
1238 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1239 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1241 // Get the total trip count from the count by adding 1.
1242 ExitCount = SE->getAddExpr(ExitCount,
1243 SE->getConstant(ExitCount->getType(), 1));
1245 // Expand the trip count and place the new instructions in the preheader.
1246 // Notice that the pre-header does not change, only the loop body.
1247 SCEVExpander Exp(*SE, "induction");
1249 // Count holds the overall loop count (N).
1250 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1251 BypassBlock->getTerminator());
1253 // The loop index does not have to start at Zero. Find the original start
1254 // value from the induction PHI node. If we don't have an induction variable
1255 // then we know that it starts at zero.
1256 Value *StartIdx = OldInduction ?
1257 OldInduction->getIncomingValueForBlock(BypassBlock):
1258 ConstantInt::get(IdxTy, 0);
1260 assert(BypassBlock && "Invalid loop structure");
1261 LoopBypassBlocks.push_back(BypassBlock);
1263 // Split the single block loop into the two loop structure described above.
1264 BasicBlock *VectorPH =
1265 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1266 BasicBlock *VecBody =
1267 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1268 BasicBlock *MiddleBlock =
1269 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1270 BasicBlock *ScalarPH =
1271 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1273 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1275 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1277 // Generate the induction variable.
1278 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1279 // The loop step is equal to the vectorization factor (num of SIMD elements)
1280 // times the unroll factor (num of SIMD instructions).
1281 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1283 // This is the IR builder that we use to add all of the logic for bypassing
1284 // the new vector loop.
1285 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1287 // We may need to extend the index in case there is a type mismatch.
1288 // We know that the count starts at zero and does not overflow.
1289 if (Count->getType() != IdxTy) {
1290 // The exit count can be of pointer type. Convert it to the correct
1292 if (ExitCount->getType()->isPointerTy())
1293 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1295 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1298 // Add the start index to the loop count to get the new end index.
1299 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1301 // Now we need to generate the expression for N - (N % VF), which is
1302 // the part that the vectorized body will execute.
1303 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1304 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1305 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1306 "end.idx.rnd.down");
1308 // Now, compare the new count to zero. If it is zero skip the vector loop and
1309 // jump to the scalar loop.
1310 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1313 BasicBlock *LastBypassBlock = BypassBlock;
1315 // Generate the code that checks in runtime if arrays overlap. We put the
1316 // checks into a separate block to make the more common case of few elements
1318 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1319 BypassBlock->getTerminator());
1320 if (MemRuntimeCheck) {
1321 // Create a new block containing the memory check.
1322 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1324 LoopBypassBlocks.push_back(CheckBlock);
1326 // Replace the branch into the memory check block with a conditional branch
1327 // for the "few elements case".
1328 Instruction *OldTerm = BypassBlock->getTerminator();
1329 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1330 OldTerm->eraseFromParent();
1332 Cmp = MemRuntimeCheck;
1333 LastBypassBlock = CheckBlock;
1336 LastBypassBlock->getTerminator()->eraseFromParent();
1337 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1340 // We are going to resume the execution of the scalar loop.
1341 // Go over all of the induction variables that we found and fix the
1342 // PHIs that are left in the scalar version of the loop.
1343 // The starting values of PHI nodes depend on the counter of the last
1344 // iteration in the vectorized loop.
1345 // If we come from a bypass edge then we need to start from the original
1348 // This variable saves the new starting index for the scalar loop.
1349 PHINode *ResumeIndex = 0;
1350 LoopVectorizationLegality::InductionList::iterator I, E;
1351 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1352 for (I = List->begin(), E = List->end(); I != E; ++I) {
1353 PHINode *OrigPhi = I->first;
1354 LoopVectorizationLegality::InductionInfo II = I->second;
1355 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1356 MiddleBlock->getTerminator());
1357 Value *EndValue = 0;
1359 case LoopVectorizationLegality::IK_NoInduction:
1360 llvm_unreachable("Unknown induction");
1361 case LoopVectorizationLegality::IK_IntInduction: {
1362 // Handle the integer induction counter:
1363 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1364 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1365 // We know what the end value is.
1366 EndValue = IdxEndRoundDown;
1367 // We also know which PHI node holds it.
1368 ResumeIndex = ResumeVal;
1371 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1372 // Convert the CountRoundDown variable to the PHI size.
1373 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1374 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1375 Value *CRD = CountRoundDown;
1376 if (CRDSize > IISize)
1377 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1378 II.StartValue->getType(), "tr.crd",
1379 LoopBypassBlocks.back()->getTerminator());
1380 else if (CRDSize < IISize)
1381 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1382 II.StartValue->getType(),
1384 LoopBypassBlocks.back()->getTerminator());
1385 // Handle reverse integer induction counter:
1387 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1388 LoopBypassBlocks.back()->getTerminator());
1391 case LoopVectorizationLegality::IK_PtrInduction: {
1392 // For pointer induction variables, calculate the offset using
1395 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end",
1396 LoopBypassBlocks.back()->getTerminator());
1399 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1400 // The value at the end of the loop for the reverse pointer is calculated
1401 // by creating a GEP with a negative index starting from the start value.
1402 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1403 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown,
1405 LoopBypassBlocks.back()->getTerminator());
1406 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx,
1408 LoopBypassBlocks.back()->getTerminator());
1413 // The new PHI merges the original incoming value, in case of a bypass,
1414 // or the value at the end of the vectorized loop.
1415 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1416 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1417 ResumeVal->addIncoming(EndValue, VecBody);
1419 // Fix the scalar body counter (PHI node).
1420 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1421 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1424 // If we are generating a new induction variable then we also need to
1425 // generate the code that calculates the exit value. This value is not
1426 // simply the end of the counter because we may skip the vectorized body
1427 // in case of a runtime check.
1429 assert(!ResumeIndex && "Unexpected resume value found");
1430 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1431 MiddleBlock->getTerminator());
1432 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1433 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1434 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1437 // Make sure that we found the index where scalar loop needs to continue.
1438 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1439 "Invalid resume Index");
1441 // Add a check in the middle block to see if we have completed
1442 // all of the iterations in the first vector loop.
1443 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1444 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1445 ResumeIndex, "cmp.n",
1446 MiddleBlock->getTerminator());
1448 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1449 // Remove the old terminator.
1450 MiddleBlock->getTerminator()->eraseFromParent();
1452 // Create i+1 and fill the PHINode.
1453 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1454 Induction->addIncoming(StartIdx, VectorPH);
1455 Induction->addIncoming(NextIdx, VecBody);
1456 // Create the compare.
1457 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1458 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1460 // Now we have two terminators. Remove the old one from the block.
1461 VecBody->getTerminator()->eraseFromParent();
1463 // Get ready to start creating new instructions into the vectorized body.
1464 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1466 // Create and register the new vector loop.
1467 Loop* Lp = new Loop();
1468 Loop *ParentLoop = OrigLoop->getParentLoop();
1470 // Insert the new loop into the loop nest and register the new basic blocks.
1472 ParentLoop->addChildLoop(Lp);
1473 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1474 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1475 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1476 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1477 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1479 LI->addTopLevelLoop(Lp);
1482 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1485 LoopVectorPreHeader = VectorPH;
1486 LoopScalarPreHeader = ScalarPH;
1487 LoopMiddleBlock = MiddleBlock;
1488 LoopExitBlock = ExitBlock;
1489 LoopVectorBody = VecBody;
1490 LoopScalarBody = OldBasicBlock;
1493 /// This function returns the identity element (or neutral element) for
1494 /// the operation K.
1496 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp,
1497 MinMaxReductionKind MinMaxK) {
1502 // Adding, Xoring, Oring zero to a number does not change it.
1503 return ConstantInt::get(Tp, 0);
1504 case RK_IntegerMult:
1505 // Multiplying a number by 1 does not change it.
1506 return ConstantInt::get(Tp, 1);
1508 // AND-ing a number with an all-1 value does not change it.
1509 return ConstantInt::get(Tp, -1, true);
1511 // Multiplying a number by 1 does not change it.
1512 return ConstantFP::get(Tp, 1.0L);
1514 // Adding zero to a number does not change it.
1515 return ConstantFP::get(Tp, 0.0L);
1516 case RK_IntegerMinMax:
1518 default: llvm_unreachable("Unknown min/max predicate");
1520 return ConstantInt::getAllOnesValue(Tp);
1522 return ConstantInt::get(Tp, 0);
1524 unsigned BitWidth = Tp->getPrimitiveSizeInBits();
1525 return ConstantInt::get(Tp->getContext(),
1526 APInt::getSignedMaxValue(BitWidth));
1528 case LoopVectorizationLegality::MRK_SIntMax: {
1529 unsigned BitWidth = Tp->getPrimitiveSizeInBits();
1530 return ConstantInt::get(Tp->getContext(),
1531 APInt::getSignedMinValue(BitWidth));
1535 llvm_unreachable("Unknown reduction kind");
1539 static Intrinsic::ID
1540 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1541 // If we have an intrinsic call, check if it is trivially vectorizable.
1542 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1543 switch (II->getIntrinsicID()) {
1544 case Intrinsic::sqrt:
1545 case Intrinsic::sin:
1546 case Intrinsic::cos:
1547 case Intrinsic::exp:
1548 case Intrinsic::exp2:
1549 case Intrinsic::log:
1550 case Intrinsic::log10:
1551 case Intrinsic::log2:
1552 case Intrinsic::fabs:
1553 case Intrinsic::floor:
1554 case Intrinsic::ceil:
1555 case Intrinsic::trunc:
1556 case Intrinsic::rint:
1557 case Intrinsic::nearbyint:
1558 case Intrinsic::pow:
1559 case Intrinsic::fma:
1560 case Intrinsic::fmuladd:
1561 return II->getIntrinsicID();
1563 return Intrinsic::not_intrinsic;
1568 return Intrinsic::not_intrinsic;
1571 Function *F = CI->getCalledFunction();
1572 // We're going to make assumptions on the semantics of the functions, check
1573 // that the target knows that it's available in this environment.
1574 if (!F || !TLI->getLibFunc(F->getName(), Func))
1575 return Intrinsic::not_intrinsic;
1577 // Otherwise check if we have a call to a function that can be turned into a
1578 // vector intrinsic.
1585 return Intrinsic::sin;
1589 return Intrinsic::cos;
1593 return Intrinsic::exp;
1595 case LibFunc::exp2f:
1596 case LibFunc::exp2l:
1597 return Intrinsic::exp2;
1601 return Intrinsic::log;
1602 case LibFunc::log10:
1603 case LibFunc::log10f:
1604 case LibFunc::log10l:
1605 return Intrinsic::log10;
1607 case LibFunc::log2f:
1608 case LibFunc::log2l:
1609 return Intrinsic::log2;
1611 case LibFunc::fabsf:
1612 case LibFunc::fabsl:
1613 return Intrinsic::fabs;
1614 case LibFunc::floor:
1615 case LibFunc::floorf:
1616 case LibFunc::floorl:
1617 return Intrinsic::floor;
1619 case LibFunc::ceilf:
1620 case LibFunc::ceill:
1621 return Intrinsic::ceil;
1622 case LibFunc::trunc:
1623 case LibFunc::truncf:
1624 case LibFunc::truncl:
1625 return Intrinsic::trunc;
1627 case LibFunc::rintf:
1628 case LibFunc::rintl:
1629 return Intrinsic::rint;
1630 case LibFunc::nearbyint:
1631 case LibFunc::nearbyintf:
1632 case LibFunc::nearbyintl:
1633 return Intrinsic::nearbyint;
1637 return Intrinsic::pow;
1640 return Intrinsic::not_intrinsic;
1643 /// This function translates the reduction kind to an LLVM binary operator.
1645 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1647 case LoopVectorizationLegality::RK_IntegerAdd:
1648 return Instruction::Add;
1649 case LoopVectorizationLegality::RK_IntegerMult:
1650 return Instruction::Mul;
1651 case LoopVectorizationLegality::RK_IntegerOr:
1652 return Instruction::Or;
1653 case LoopVectorizationLegality::RK_IntegerAnd:
1654 return Instruction::And;
1655 case LoopVectorizationLegality::RK_IntegerXor:
1656 return Instruction::Xor;
1657 case LoopVectorizationLegality::RK_FloatMult:
1658 return Instruction::FMul;
1659 case LoopVectorizationLegality::RK_FloatAdd:
1660 return Instruction::FAdd;
1661 case LoopVectorizationLegality::RK_IntegerMinMax:
1662 return Instruction::ICmp;
1664 llvm_unreachable("Unknown reduction operation");
1668 Value *createMinMaxOp(IRBuilder<> &Builder,
1669 LoopVectorizationLegality::MinMaxReductionKind RK,
1672 CmpInst::Predicate P = CmpInst::ICMP_NE;
1675 llvm_unreachable("Unknown min/max reduction kind");
1676 case LoopVectorizationLegality::MRK_UIntMin:
1677 P = CmpInst::ICMP_ULT;
1679 case LoopVectorizationLegality::MRK_UIntMax:
1680 P = CmpInst::ICMP_UGT;
1682 case LoopVectorizationLegality::MRK_SIntMin:
1683 P = CmpInst::ICMP_SLT;
1685 case LoopVectorizationLegality::MRK_SIntMax:
1686 P = CmpInst::ICMP_SGT;
1688 Value *Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1689 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1694 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1695 //===------------------------------------------------===//
1697 // Notice: any optimization or new instruction that go
1698 // into the code below should be also be implemented in
1701 //===------------------------------------------------===//
1702 Constant *Zero = Builder.getInt32(0);
1704 // In order to support reduction variables we need to be able to vectorize
1705 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1706 // stages. First, we create a new vector PHI node with no incoming edges.
1707 // We use this value when we vectorize all of the instructions that use the
1708 // PHI. Next, after all of the instructions in the block are complete we
1709 // add the new incoming edges to the PHI. At this point all of the
1710 // instructions in the basic block are vectorized, so we can use them to
1711 // construct the PHI.
1712 PhiVector RdxPHIsToFix;
1714 // Scan the loop in a topological order to ensure that defs are vectorized
1716 LoopBlocksDFS DFS(OrigLoop);
1719 // Vectorize all of the blocks in the original loop.
1720 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1721 be = DFS.endRPO(); bb != be; ++bb)
1722 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1724 // At this point every instruction in the original loop is widened to
1725 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1726 // that we vectorized. The PHI nodes are currently empty because we did
1727 // not want to introduce cycles. Notice that the remaining PHI nodes
1728 // that we need to fix are reduction variables.
1730 // Create the 'reduced' values for each of the induction vars.
1731 // The reduced values are the vector values that we scalarize and combine
1732 // after the loop is finished.
1733 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1735 PHINode *RdxPhi = *it;
1736 assert(RdxPhi && "Unable to recover vectorized PHI");
1738 // Find the reduction variable descriptor.
1739 assert(Legal->getReductionVars()->count(RdxPhi) &&
1740 "Unable to find the reduction variable");
1741 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1742 (*Legal->getReductionVars())[RdxPhi];
1744 // We need to generate a reduction vector from the incoming scalar.
1745 // To do so, we need to generate the 'identity' vector and overide
1746 // one of the elements with the incoming scalar reduction. We need
1747 // to do it in the vector-loop preheader.
1748 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1750 // This is the vector-clone of the value that leaves the loop.
1751 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1752 Type *VecTy = VectorExit[0]->getType();
1754 // Find the reduction identity variable. Zero for addition, or, xor,
1755 // one for multiplication, -1 for And.
1757 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1758 VecTy->getScalarType(),
1759 RdxDesc.MinMaxKind);
1760 Constant *Identity = ConstantVector::getSplat(VF, Iden);
1762 // This vector is the Identity vector where the first element is the
1763 // incoming scalar reduction.
1764 Value *VectorStart = Builder.CreateInsertElement(Identity,
1765 RdxDesc.StartValue, Zero);
1767 // Fix the vector-loop phi.
1768 // We created the induction variable so we know that the
1769 // preheader is the first entry.
1770 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1772 // Reductions do not have to start at zero. They can start with
1773 // any loop invariant values.
1774 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1775 BasicBlock *Latch = OrigLoop->getLoopLatch();
1776 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1777 VectorParts &Val = getVectorValue(LoopVal);
1778 for (unsigned part = 0; part < UF; ++part) {
1779 // Make sure to add the reduction stat value only to the
1780 // first unroll part.
1781 Value *StartVal = (part == 0) ? VectorStart : Identity;
1782 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1783 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1786 // Before each round, move the insertion point right between
1787 // the PHIs and the values we are going to write.
1788 // This allows us to write both PHINodes and the extractelement
1790 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1792 VectorParts RdxParts;
1793 for (unsigned part = 0; part < UF; ++part) {
1794 // This PHINode contains the vectorized reduction variable, or
1795 // the initial value vector, if we bypass the vector loop.
1796 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1797 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1798 Value *StartVal = (part == 0) ? VectorStart : Identity;
1799 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1800 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1801 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1802 RdxParts.push_back(NewPhi);
1805 // Reduce all of the unrolled parts into a single vector.
1806 Value *ReducedPartRdx = RdxParts[0];
1807 unsigned Op = getReductionBinOp(RdxDesc.Kind);
1808 for (unsigned part = 1; part < UF; ++part) {
1809 if (Op != Instruction::ICmp)
1810 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
1811 RdxParts[part], ReducedPartRdx,
1814 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
1815 ReducedPartRdx, RdxParts[part]);
1818 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1819 // and vector ops, reducing the set of values being computed by half each
1821 assert(isPowerOf2_32(VF) &&
1822 "Reduction emission only supported for pow2 vectors!");
1823 Value *TmpVec = ReducedPartRdx;
1824 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1825 for (unsigned i = VF; i != 1; i >>= 1) {
1826 // Move the upper half of the vector to the lower half.
1827 for (unsigned j = 0; j != i/2; ++j)
1828 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1830 // Fill the rest of the mask with undef.
1831 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1832 UndefValue::get(Builder.getInt32Ty()));
1835 Builder.CreateShuffleVector(TmpVec,
1836 UndefValue::get(TmpVec->getType()),
1837 ConstantVector::get(ShuffleMask),
1840 if (Op != Instruction::ICmp)
1841 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1844 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
1847 // The result is in the first element of the vector.
1848 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1850 // Now, we need to fix the users of the reduction variable
1851 // inside and outside of the scalar remainder loop.
1852 // We know that the loop is in LCSSA form. We need to update the
1853 // PHI nodes in the exit blocks.
1854 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1855 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1856 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1857 if (!LCSSAPhi) continue;
1859 // All PHINodes need to have a single entry edge, or two if
1860 // we already fixed them.
1861 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1863 // We found our reduction value exit-PHI. Update it with the
1864 // incoming bypass edge.
1865 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1866 // Add an edge coming from the bypass.
1867 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1870 }// end of the LCSSA phi scan.
1872 // Fix the scalar loop reduction variable with the incoming reduction sum
1873 // from the vector body and from the backedge value.
1874 int IncomingEdgeBlockIdx =
1875 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1876 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1877 // Pick the other block.
1878 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1879 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1880 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1881 }// end of for each redux variable.
1883 // The Loop exit block may have single value PHI nodes where the incoming
1884 // value is 'undef'. While vectorizing we only handled real values that
1885 // were defined inside the loop. Here we handle the 'undef case'.
1887 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1888 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1889 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1890 if (!LCSSAPhi) continue;
1891 if (LCSSAPhi->getNumIncomingValues() == 1)
1892 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1897 InnerLoopVectorizer::VectorParts
1898 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1899 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1902 VectorParts SrcMask = createBlockInMask(Src);
1904 // The terminator has to be a branch inst!
1905 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1906 assert(BI && "Unexpected terminator found");
1908 if (BI->isConditional()) {
1909 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1911 if (BI->getSuccessor(0) != Dst)
1912 for (unsigned part = 0; part < UF; ++part)
1913 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1915 for (unsigned part = 0; part < UF; ++part)
1916 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1923 InnerLoopVectorizer::VectorParts
1924 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1925 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1927 // Loop incoming mask is all-one.
1928 if (OrigLoop->getHeader() == BB) {
1929 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1930 return getVectorValue(C);
1933 // This is the block mask. We OR all incoming edges, and with zero.
1934 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1935 VectorParts BlockMask = getVectorValue(Zero);
1938 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1939 VectorParts EM = createEdgeMask(*it, BB);
1940 for (unsigned part = 0; part < UF; ++part)
1941 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1948 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1949 BasicBlock *BB, PhiVector *PV) {
1950 // For each instruction in the old loop.
1951 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1952 VectorParts &Entry = WidenMap.get(it);
1953 switch (it->getOpcode()) {
1954 case Instruction::Br:
1955 // Nothing to do for PHIs and BR, since we already took care of the
1956 // loop control flow instructions.
1958 case Instruction::PHI:{
1959 PHINode* P = cast<PHINode>(it);
1960 // Handle reduction variables:
1961 if (Legal->getReductionVars()->count(P)) {
1962 for (unsigned part = 0; part < UF; ++part) {
1963 // This is phase one of vectorizing PHIs.
1964 Type *VecTy = VectorType::get(it->getType(), VF);
1965 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1966 LoopVectorBody-> getFirstInsertionPt());
1972 // Check for PHI nodes that are lowered to vector selects.
1973 if (P->getParent() != OrigLoop->getHeader()) {
1974 // We know that all PHIs in non header blocks are converted into
1975 // selects, so we don't have to worry about the insertion order and we
1976 // can just use the builder.
1978 // At this point we generate the predication tree. There may be
1979 // duplications since this is a simple recursive scan, but future
1980 // optimizations will clean it up.
1981 VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
1984 for (unsigned part = 0; part < UF; ++part) {
1985 VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
1986 VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
1987 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
1993 // This PHINode must be an induction variable.
1994 // Make sure that we know about it.
1995 assert(Legal->getInductionVars()->count(P) &&
1996 "Not an induction variable");
1998 LoopVectorizationLegality::InductionInfo II =
1999 Legal->getInductionVars()->lookup(P);
2002 case LoopVectorizationLegality::IK_NoInduction:
2003 llvm_unreachable("Unknown induction");
2004 case LoopVectorizationLegality::IK_IntInduction: {
2005 assert(P == OldInduction && "Unexpected PHI");
2006 Value *Broadcasted = getBroadcastInstrs(Induction);
2007 // After broadcasting the induction variable we need to make the
2008 // vector consecutive by adding 0, 1, 2 ...
2009 for (unsigned part = 0; part < UF; ++part)
2010 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2013 case LoopVectorizationLegality::IK_ReverseIntInduction:
2014 case LoopVectorizationLegality::IK_PtrInduction:
2015 case LoopVectorizationLegality::IK_ReversePtrInduction:
2016 // Handle reverse integer and pointer inductions.
2017 Value *StartIdx = 0;
2018 // If we have a single integer induction variable then use it.
2019 // Otherwise, start counting at zero.
2021 LoopVectorizationLegality::InductionInfo OldII =
2022 Legal->getInductionVars()->lookup(OldInduction);
2023 StartIdx = OldII.StartValue;
2025 StartIdx = ConstantInt::get(Induction->getType(), 0);
2027 // This is the normalized GEP that starts counting at zero.
2028 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2031 // Handle the reverse integer induction variable case.
2032 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2033 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2034 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2036 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2039 // This is a new value so do not hoist it out.
2040 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2041 // After broadcasting the induction variable we need to make the
2042 // vector consecutive by adding ... -3, -2, -1, 0.
2043 for (unsigned part = 0; part < UF; ++part)
2044 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
2048 // Handle the pointer induction variable case.
2049 assert(P->getType()->isPointerTy() && "Unexpected type.");
2051 // Is this a reverse induction ptr or a consecutive induction ptr.
2052 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2055 // This is the vector of results. Notice that we don't generate
2056 // vector geps because scalar geps result in better code.
2057 for (unsigned part = 0; part < UF; ++part) {
2058 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2059 for (unsigned int i = 0; i < VF; ++i) {
2060 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2061 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2064 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2066 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2068 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2070 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2071 Builder.getInt32(i),
2074 Entry[part] = VecVal;
2081 case Instruction::Add:
2082 case Instruction::FAdd:
2083 case Instruction::Sub:
2084 case Instruction::FSub:
2085 case Instruction::Mul:
2086 case Instruction::FMul:
2087 case Instruction::UDiv:
2088 case Instruction::SDiv:
2089 case Instruction::FDiv:
2090 case Instruction::URem:
2091 case Instruction::SRem:
2092 case Instruction::FRem:
2093 case Instruction::Shl:
2094 case Instruction::LShr:
2095 case Instruction::AShr:
2096 case Instruction::And:
2097 case Instruction::Or:
2098 case Instruction::Xor: {
2099 // Just widen binops.
2100 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2101 VectorParts &A = getVectorValue(it->getOperand(0));
2102 VectorParts &B = getVectorValue(it->getOperand(1));
2104 // Use this vector value for all users of the original instruction.
2105 for (unsigned Part = 0; Part < UF; ++Part) {
2106 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2108 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2109 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2110 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2111 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2112 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2114 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2115 VecOp->setIsExact(BinOp->isExact());
2121 case Instruction::Select: {
2123 // If the selector is loop invariant we can create a select
2124 // instruction with a scalar condition. Otherwise, use vector-select.
2125 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2128 // The condition can be loop invariant but still defined inside the
2129 // loop. This means that we can't just use the original 'cond' value.
2130 // We have to take the 'vectorized' value and pick the first lane.
2131 // Instcombine will make this a no-op.
2132 VectorParts &Cond = getVectorValue(it->getOperand(0));
2133 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2134 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2135 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2136 Builder.getInt32(0));
2137 for (unsigned Part = 0; Part < UF; ++Part) {
2138 Entry[Part] = Builder.CreateSelect(
2139 InvariantCond ? ScalarCond : Cond[Part],
2146 case Instruction::ICmp:
2147 case Instruction::FCmp: {
2148 // Widen compares. Generate vector compares.
2149 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2150 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2151 VectorParts &A = getVectorValue(it->getOperand(0));
2152 VectorParts &B = getVectorValue(it->getOperand(1));
2153 for (unsigned Part = 0; Part < UF; ++Part) {
2156 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2158 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2164 case Instruction::Store:
2165 case Instruction::Load:
2166 vectorizeMemoryInstruction(it, Legal);
2168 case Instruction::ZExt:
2169 case Instruction::SExt:
2170 case Instruction::FPToUI:
2171 case Instruction::FPToSI:
2172 case Instruction::FPExt:
2173 case Instruction::PtrToInt:
2174 case Instruction::IntToPtr:
2175 case Instruction::SIToFP:
2176 case Instruction::UIToFP:
2177 case Instruction::Trunc:
2178 case Instruction::FPTrunc:
2179 case Instruction::BitCast: {
2180 CastInst *CI = dyn_cast<CastInst>(it);
2181 /// Optimize the special case where the source is the induction
2182 /// variable. Notice that we can only optimize the 'trunc' case
2183 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2184 /// c. other casts depend on pointer size.
2185 if (CI->getOperand(0) == OldInduction &&
2186 it->getOpcode() == Instruction::Trunc) {
2187 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2189 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2190 for (unsigned Part = 0; Part < UF; ++Part)
2191 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2194 /// Vectorize casts.
2195 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2197 VectorParts &A = getVectorValue(it->getOperand(0));
2198 for (unsigned Part = 0; Part < UF; ++Part)
2199 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2203 case Instruction::Call: {
2204 // Ignore dbg intrinsics.
2205 if (isa<DbgInfoIntrinsic>(it))
2208 Module *M = BB->getParent()->getParent();
2209 CallInst *CI = cast<CallInst>(it);
2210 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2211 assert(ID && "Not an intrinsic call!");
2212 for (unsigned Part = 0; Part < UF; ++Part) {
2213 SmallVector<Value*, 4> Args;
2214 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2215 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2216 Args.push_back(Arg[Part]);
2218 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2219 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2220 Entry[Part] = Builder.CreateCall(F, Args);
2226 // All other instructions are unsupported. Scalarize them.
2227 scalarizeInstruction(it);
2230 }// end of for_each instr.
2233 void InnerLoopVectorizer::updateAnalysis() {
2234 // Forget the original basic block.
2235 SE->forgetLoop(OrigLoop);
2237 // Update the dominator tree information.
2238 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2239 "Entry does not dominate exit.");
2241 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2242 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2243 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2244 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2245 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2246 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2247 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2248 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2250 DEBUG(DT->verifyAnalysis());
2253 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2254 if (!EnableIfConversion)
2257 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2258 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2260 // Collect the blocks that need predication.
2261 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2262 BasicBlock *BB = LoopBlocks[i];
2264 // We don't support switch statements inside loops.
2265 if (!isa<BranchInst>(BB->getTerminator()))
2268 // We must have at most two predecessors because we need to convert
2269 // all PHIs to selects.
2270 unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
2274 // We must be able to predicate all blocks that need to be predicated.
2275 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2279 // We can if-convert this loop.
2283 bool LoopVectorizationLegality::canVectorize() {
2284 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2286 // We can only vectorize innermost loops.
2287 if (TheLoop->getSubLoopsVector().size())
2290 // We must have a single backedge.
2291 if (TheLoop->getNumBackEdges() != 1)
2294 // We must have a single exiting block.
2295 if (!TheLoop->getExitingBlock())
2298 unsigned NumBlocks = TheLoop->getNumBlocks();
2300 // Check if we can if-convert non single-bb loops.
2301 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2302 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2306 // We need to have a loop header.
2307 BasicBlock *Latch = TheLoop->getLoopLatch();
2308 DEBUG(dbgs() << "LV: Found a loop: " <<
2309 TheLoop->getHeader()->getName() << "\n");
2311 // ScalarEvolution needs to be able to find the exit count.
2312 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2313 if (ExitCount == SE->getCouldNotCompute()) {
2314 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2318 // Do not loop-vectorize loops with a tiny trip count.
2319 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2320 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2321 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2322 "This loop is not worth vectorizing.\n");
2326 // Check if we can vectorize the instructions and CFG in this loop.
2327 if (!canVectorizeInstrs()) {
2328 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2332 // Go over each instruction and look at memory deps.
2333 if (!canVectorizeMemory()) {
2334 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2338 // Collect all of the variables that remain uniform after vectorization.
2339 collectLoopUniforms();
2341 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2342 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2345 // Okay! We can vectorize. At this point we don't have any other mem analysis
2346 // which may limit our maximum vectorization factor, so just return true with
2351 bool LoopVectorizationLegality::canVectorizeInstrs() {
2352 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2353 BasicBlock *Header = TheLoop->getHeader();
2355 // If we marked the scalar loop as "already vectorized" then no need
2356 // to vectorize it again.
2357 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) {
2358 DEBUG(dbgs() << "LV: This loop was vectorized before\n");
2362 // For each block in the loop.
2363 for (Loop::block_iterator bb = TheLoop->block_begin(),
2364 be = TheLoop->block_end(); bb != be; ++bb) {
2366 // Scan the instructions in the block and look for hazards.
2367 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2370 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2371 // This should not happen because the loop should be normalized.
2372 if (Phi->getNumIncomingValues() != 2) {
2373 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2377 // Check that this PHI type is allowed.
2378 if (!Phi->getType()->isIntegerTy() &&
2379 !Phi->getType()->isFloatingPointTy() &&
2380 !Phi->getType()->isPointerTy()) {
2381 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2385 // If this PHINode is not in the header block, then we know that we
2386 // can convert it to select during if-conversion. No need to check if
2387 // the PHIs in this block are induction or reduction variables.
2391 // This is the value coming from the preheader.
2392 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2393 // Check if this is an induction variable.
2394 InductionKind IK = isInductionVariable(Phi);
2396 if (IK_NoInduction != IK) {
2397 // Int inductions are special because we only allow one IV.
2398 if (IK == IK_IntInduction) {
2400 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2406 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2407 Inductions[Phi] = InductionInfo(StartValue, IK);
2411 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2412 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2415 if (AddReductionVar(Phi, RK_IntegerMult)) {
2416 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2419 if (AddReductionVar(Phi, RK_IntegerOr)) {
2420 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2423 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2424 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2427 if (AddReductionVar(Phi, RK_IntegerXor)) {
2428 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2431 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2432 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2435 if (AddReductionVar(Phi, RK_FloatMult)) {
2436 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2439 if (AddReductionVar(Phi, RK_FloatAdd)) {
2440 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2444 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2446 }// end of PHI handling
2448 // We still don't handle functions. However, we can ignore dbg intrinsic
2449 // calls and we do handle certain intrinsic and libm functions.
2450 CallInst *CI = dyn_cast<CallInst>(it);
2451 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2452 DEBUG(dbgs() << "LV: Found a call site.\n");
2456 // Check that the instruction return type is vectorizable.
2457 if (!VectorType::isValidElementType(it->getType()) &&
2458 !it->getType()->isVoidTy()) {
2459 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2463 // Check that the stored type is vectorizable.
2464 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2465 Type *T = ST->getValueOperand()->getType();
2466 if (!VectorType::isValidElementType(T))
2470 // Reduction instructions are allowed to have exit users.
2471 // All other instructions must not have external users.
2472 if (!AllowedExit.count(it))
2473 //Check that all of the users of the loop are inside the BB.
2474 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2476 Instruction *U = cast<Instruction>(*I);
2477 // This user may be a reduction exit value.
2478 if (!TheLoop->contains(U)) {
2479 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2488 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2489 assert(getInductionVars()->size() && "No induction variables");
2495 void LoopVectorizationLegality::collectLoopUniforms() {
2496 // We now know that the loop is vectorizable!
2497 // Collect variables that will remain uniform after vectorization.
2498 std::vector<Value*> Worklist;
2499 BasicBlock *Latch = TheLoop->getLoopLatch();
2501 // Start with the conditional branch and walk up the block.
2502 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2504 while (Worklist.size()) {
2505 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2506 Worklist.pop_back();
2508 // Look at instructions inside this loop.
2509 // Stop when reaching PHI nodes.
2510 // TODO: we need to follow values all over the loop, not only in this block.
2511 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2514 // This is a known uniform.
2517 // Insert all operands.
2518 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2519 Worklist.push_back(I->getOperand(i));
2524 AliasAnalysis::Location
2525 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2526 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2527 return AA->getLocation(Store);
2528 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2529 return AA->getLocation(Load);
2531 llvm_unreachable("Should be either load or store instruction");
2535 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2538 AliasMultiMap& WriteObjects,
2539 unsigned MaxByteWidth) {
2541 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2543 std::vector<Instruction*>::iterator
2544 it = WriteObjects[Object].begin(),
2545 end = WriteObjects[Object].end();
2547 for (; it != end; ++it) {
2548 Instruction* I = *it;
2552 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2553 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2554 ThatLoc.getWithNewSize(MaxByteWidth)))
2560 bool LoopVectorizationLegality::canVectorizeMemory() {
2562 typedef SmallVector<Value*, 16> ValueVector;
2563 typedef SmallPtrSet<Value*, 16> ValueSet;
2564 // Holds the Load and Store *instructions*.
2567 PtrRtCheck.Pointers.clear();
2568 PtrRtCheck.Need = false;
2570 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2573 for (Loop::block_iterator bb = TheLoop->block_begin(),
2574 be = TheLoop->block_end(); bb != be; ++bb) {
2576 // Scan the BB and collect legal loads and stores.
2577 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2580 // If this is a load, save it. If this instruction can read from memory
2581 // but is not a load, then we quit. Notice that we don't handle function
2582 // calls that read or write.
2583 if (it->mayReadFromMemory()) {
2584 LoadInst *Ld = dyn_cast<LoadInst>(it);
2585 if (!Ld) return false;
2586 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2587 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2590 Loads.push_back(Ld);
2594 // Save 'store' instructions. Abort if other instructions write to memory.
2595 if (it->mayWriteToMemory()) {
2596 StoreInst *St = dyn_cast<StoreInst>(it);
2597 if (!St) return false;
2598 if (!St->isSimple() && !IsAnnotatedParallel) {
2599 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2602 Stores.push_back(St);
2607 // Now we have two lists that hold the loads and the stores.
2608 // Next, we find the pointers that they use.
2610 // Check if we see any stores. If there are no stores, then we don't
2611 // care if the pointers are *restrict*.
2612 if (!Stores.size()) {
2613 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2617 // Holds the read and read-write *pointers* that we find. These maps hold
2618 // unique values for pointers (so no need for multi-map).
2620 AliasMap ReadWrites;
2622 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2623 // multiple times on the same object. If the ptr is accessed twice, once
2624 // for read and once for write, it will only appear once (on the write
2625 // list). This is okay, since we are going to check for conflicts between
2626 // writes and between reads and writes, but not between reads and reads.
2629 ValueVector::iterator I, IE;
2630 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2631 StoreInst *ST = cast<StoreInst>(*I);
2632 Value* Ptr = ST->getPointerOperand();
2634 if (isUniform(Ptr)) {
2635 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2639 // If we did *not* see this pointer before, insert it to
2640 // the read-write list. At this phase it is only a 'write' list.
2641 if (Seen.insert(Ptr))
2642 ReadWrites.insert(std::make_pair(Ptr, ST));
2645 if (IsAnnotatedParallel) {
2647 << "LV: A loop annotated parallel, ignore memory dependency "
2652 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2653 LoadInst *LD = cast<LoadInst>(*I);
2654 Value* Ptr = LD->getPointerOperand();
2655 // If we did *not* see this pointer before, insert it to the
2656 // read list. If we *did* see it before, then it is already in
2657 // the read-write list. This allows us to vectorize expressions
2658 // such as A[i] += x; Because the address of A[i] is a read-write
2659 // pointer. This only works if the index of A[i] is consecutive.
2660 // If the address of i is unknown (for example A[B[i]]) then we may
2661 // read a few words, modify, and write a few words, and some of the
2662 // words may be written to the same address.
2663 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2664 Reads.insert(std::make_pair(Ptr, LD));
2667 // If we write (or read-write) to a single destination and there are no
2668 // other reads in this loop then is it safe to vectorize.
2669 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2670 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2674 // Find pointers with computable bounds. We are going to use this information
2675 // to place a runtime bound check.
2676 bool CanDoRT = true;
2677 AliasMap::iterator MI, ME;
2678 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2679 Value *V = (*MI).first;
2680 if (hasComputableBounds(V)) {
2681 PtrRtCheck.insert(SE, TheLoop, V);
2682 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2688 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2689 Value *V = (*MI).first;
2690 if (hasComputableBounds(V)) {
2691 PtrRtCheck.insert(SE, TheLoop, V);
2692 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2699 // Check that we did not collect too many pointers or found a
2700 // unsizeable pointer.
2701 if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
2707 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2710 bool NeedRTCheck = false;
2712 // Biggest vectorized access possible, vector width * unroll factor.
2713 // TODO: We're being very pessimistic here, find a way to know the
2714 // real access width before getting here.
2715 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2716 TTI->getMaximumUnrollFactor();
2717 // Now that the pointers are in two lists (Reads and ReadWrites), we
2718 // can check that there are no conflicts between each of the writes and
2719 // between the writes to the reads.
2720 // Note that WriteObjects duplicates the stores (indexed now by underlying
2721 // objects) to avoid pointing to elements inside ReadWrites.
2722 // TODO: Maybe create a new type where they can interact without duplication.
2723 AliasMultiMap WriteObjects;
2724 ValueVector TempObjects;
2726 // Check that the read-writes do not conflict with other read-write
2728 bool AllWritesIdentified = true;
2729 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2730 Value *Val = (*MI).first;
2731 Instruction *Inst = (*MI).second;
2733 GetUnderlyingObjects(Val, TempObjects, DL);
2734 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2736 if (!isIdentifiedObject(*UI)) {
2737 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2739 AllWritesIdentified = false;
2742 // Never seen it before, can't alias.
2743 if (WriteObjects[*UI].empty()) {
2744 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2745 WriteObjects[*UI].push_back(Inst);
2748 // Direct alias found.
2749 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2750 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2754 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2756 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2757 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2759 // If global alias, make sure they do alias.
2760 if (hasPossibleGlobalWriteReorder(*UI,
2764 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2769 // Didn't alias, insert into map for further reference.
2770 WriteObjects[*UI].push_back(Inst);
2772 TempObjects.clear();
2775 /// Check that the reads don't conflict with the read-writes.
2776 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2777 Value *Val = (*MI).first;
2778 GetUnderlyingObjects(Val, TempObjects, DL);
2779 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2781 // If all of the writes are identified then we don't care if the read
2782 // pointer is identified or not.
2783 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2784 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2788 // Never seen it before, can't alias.
2789 if (WriteObjects[*UI].empty())
2791 // Direct alias found.
2792 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2793 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2797 DEBUG(dbgs() << "LV: Found a global value: "
2799 Instruction *Inst = (*MI).second;
2800 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2801 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2803 // If global alias, make sure they do alias.
2804 if (hasPossibleGlobalWriteReorder(*UI,
2808 DEBUG(dbgs() << "LV: Found a possible read-write reorder:"
2813 TempObjects.clear();
2816 PtrRtCheck.Need = NeedRTCheck;
2817 if (NeedRTCheck && !CanDoRT) {
2818 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2819 "the array bounds.\n");
2824 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2825 " need a runtime memory check.\n");
2829 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2830 ReductionKind Kind) {
2831 if (Phi->getNumIncomingValues() != 2)
2834 // Reduction variables are only found in the loop header block.
2835 if (Phi->getParent() != TheLoop->getHeader())
2838 // Obtain the reduction start value from the value that comes from the loop
2840 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2842 // ExitInstruction is the single value which is used outside the loop.
2843 // We only allow for a single reduction value to be used outside the loop.
2844 // This includes users of the reduction, variables (which form a cycle
2845 // which ends in the phi node).
2846 Instruction *ExitInstruction = 0;
2847 // Indicates that we found a binary operation in our scan.
2848 bool FoundBinOp = false;
2850 // Iter is our iterator. We start with the PHI node and scan for all of the
2851 // users of this instruction. All users must be instructions that can be
2852 // used as reduction variables (such as ADD). We may have a single
2853 // out-of-block user. The cycle must end with the original PHI.
2854 Instruction *Iter = Phi;
2856 // To recognize min/max patterns formed by a icmp select sequence, we store
2857 // the number of instruction we saw from the recognized min/max pattern,
2858 // such that we don't stop when we see the phi has two uses (one by the select
2859 // and one by the icmp) and to make sure we only see exactly the two
2861 unsigned NumICmpSelectPatternInst = 0;
2862 ReductionInstDesc ReduxDesc(false, 0);
2864 // Avoid cycles in the chain.
2865 SmallPtrSet<Instruction *, 8> VisitedInsts;
2866 while (VisitedInsts.insert(Iter)) {
2867 // If the instruction has no users then this is a broken
2868 // chain and can't be a reduction variable.
2869 if (Iter->use_empty())
2872 // Did we find a user inside this loop already ?
2873 bool FoundInBlockUser = false;
2874 // Did we reach the initial PHI node already ?
2875 bool FoundStartPHI = false;
2877 // Is this a bin op ?
2878 FoundBinOp |= !isa<PHINode>(Iter);
2880 // For each of the *users* of iter.
2881 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2883 Instruction *U = cast<Instruction>(*it);
2884 // We already know that the PHI is a user.
2886 FoundStartPHI = true;
2890 // Check if we found the exit user.
2891 BasicBlock *Parent = U->getParent();
2892 if (!TheLoop->contains(Parent)) {
2893 // Exit if you find multiple outside users.
2894 if (ExitInstruction != 0)
2896 ExitInstruction = Iter;
2899 // We allow in-loop PHINodes which are not the original reduction PHI
2900 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2901 // structure) then don't skip this PHI.
2902 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2903 U->getParent() != TheLoop->getHeader() &&
2904 TheLoop->contains(U) &&
2905 Iter->hasNUsesOrMore(2))
2908 // We can't have multiple inside users except for a combination of
2909 // icmp/select both using the phi.
2910 if (FoundInBlockUser && !NumICmpSelectPatternInst)
2912 FoundInBlockUser = true;
2914 // Any reduction instr must be of one of the allowed kinds.
2915 ReduxDesc = isReductionInstr(U, Kind, ReduxDesc);
2916 if (!ReduxDesc.IsReduction)
2919 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(U) ||
2920 isa<SelectInst>(U)))
2921 ++NumICmpSelectPatternInst;
2923 // Reductions of instructions such as Div, and Sub is only
2924 // possible if the LHS is the reduction variable.
2925 if (!U->isCommutative() && !isa<PHINode>(U) && !isa<SelectInst>(U) &&
2926 !isa<ICmpInst>(U) && U->getOperand(0) != Iter)
2929 Iter = ReduxDesc.PatternLastInst;
2932 // This means we have seen one but not the other instruction of the
2933 // pattern or more than just a select and cmp.
2934 if (Kind == RK_IntegerMinMax && NumICmpSelectPatternInst != 2)
2937 // We found a reduction var if we have reached the original
2938 // phi node and we only have a single instruction with out-of-loop
2940 if (FoundStartPHI) {
2941 // This instruction is allowed to have out-of-loop users.
2942 AllowedExit.insert(ExitInstruction);
2944 // Save the description of this reduction variable.
2945 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
2946 ReduxDesc.MinMaxKind);
2947 Reductions[Phi] = RD;
2948 // We've ended the cycle. This is a reduction variable if we have an
2949 // outside user and it has a binary op.
2950 return FoundBinOp && ExitInstruction;
2957 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
2958 /// pattern corresponding to a min(X, Y) or max(X, Y).
2959 LoopVectorizationLegality::ReductionInstDesc
2960 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, ReductionInstDesc &Prev) {
2962 assert((isa<ICmpInst>(I) || isa<SelectInst>(I)) &&
2963 "Expect a select instruction");
2965 SelectInst *Select = 0;
2967 // We must handle the select(cmp()) as a single instruction. Advance to the
2969 if ((Cmp = dyn_cast<ICmpInst>(I))) {
2970 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
2971 return ReductionInstDesc(false, I);
2972 return ReductionInstDesc(Select, Prev.MinMaxKind);
2975 // Only handle single use cases for now.
2976 if (!(Select = dyn_cast<SelectInst>(I)))
2977 return ReductionInstDesc(false, I);
2978 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))))
2979 return ReductionInstDesc(false, I);
2980 if (!Cmp->hasOneUse())
2981 return ReductionInstDesc(false, I);
2983 Value *CmpLeft = Cmp->getOperand(0);
2984 Value *CmpRight = Cmp->getOperand(1);
2986 // Look for a min/max pattern.
2987 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
2988 return ReductionInstDesc(Select, MRK_UIntMin);
2989 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
2990 return ReductionInstDesc(Select, MRK_UIntMax);
2991 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
2992 return ReductionInstDesc(Select, MRK_SIntMax);
2993 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
2994 return ReductionInstDesc(Select, MRK_SIntMin);
2996 return ReductionInstDesc(false, I);
2999 LoopVectorizationLegality::ReductionInstDesc
3000 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3002 ReductionInstDesc &Prev) {
3003 bool FP = I->getType()->isFloatingPointTy();
3004 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3005 switch (I->getOpcode()) {
3007 return ReductionInstDesc(false, I);
3008 case Instruction::PHI:
3009 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd))
3010 return ReductionInstDesc(false, I);
3011 return ReductionInstDesc(I, Prev.MinMaxKind);
3012 case Instruction::Sub:
3013 case Instruction::Add:
3014 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3015 case Instruction::Mul:
3016 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3017 case Instruction::And:
3018 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3019 case Instruction::Or:
3020 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3021 case Instruction::Xor:
3022 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3023 case Instruction::FMul:
3024 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3025 case Instruction::FAdd:
3026 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3027 case Instruction::ICmp:
3028 case Instruction::Select:
3029 if (Kind != RK_IntegerMinMax)
3030 return ReductionInstDesc(false, I);
3031 return isMinMaxSelectCmpPattern(I, Prev);
3035 LoopVectorizationLegality::InductionKind
3036 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3037 Type *PhiTy = Phi->getType();
3038 // We only handle integer and pointer inductions variables.
3039 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3040 return IK_NoInduction;
3042 // Check that the PHI is consecutive.
3043 const SCEV *PhiScev = SE->getSCEV(Phi);
3044 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3046 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3047 return IK_NoInduction;
3049 const SCEV *Step = AR->getStepRecurrence(*SE);
3051 // Integer inductions need to have a stride of one.
3052 if (PhiTy->isIntegerTy()) {
3054 return IK_IntInduction;
3055 if (Step->isAllOnesValue())
3056 return IK_ReverseIntInduction;
3057 return IK_NoInduction;
3060 // Calculate the pointer stride and check if it is consecutive.
3061 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3063 return IK_NoInduction;
3065 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3066 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3067 if (C->getValue()->equalsInt(Size))
3068 return IK_PtrInduction;
3069 else if (C->getValue()->equalsInt(0 - Size))
3070 return IK_ReversePtrInduction;
3072 return IK_NoInduction;
3075 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3076 Value *In0 = const_cast<Value*>(V);
3077 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3081 return Inductions.count(PN);
3084 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3085 assert(TheLoop->contains(BB) && "Unknown block used");
3087 // Blocks that do not dominate the latch need predication.
3088 BasicBlock* Latch = TheLoop->getLoopLatch();
3089 return !DT->dominates(BB, Latch);
3092 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3093 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3094 // We don't predicate loads/stores at the moment.
3095 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
3098 // The instructions below can trap.
3099 switch (it->getOpcode()) {
3101 case Instruction::UDiv:
3102 case Instruction::SDiv:
3103 case Instruction::URem:
3104 case Instruction::SRem:
3112 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3113 const SCEV *PhiScev = SE->getSCEV(Ptr);
3114 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3118 return AR->isAffine();
3121 LoopVectorizationCostModel::VectorizationFactor
3122 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3124 // Width 1 means no vectorize
3125 VectorizationFactor Factor = { 1U, 0U };
3126 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3127 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3131 // Find the trip count.
3132 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3133 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3135 unsigned WidestType = getWidestType();
3136 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3137 unsigned MaxVectorSize = WidestRegister / WidestType;
3138 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3139 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3141 if (MaxVectorSize == 0) {
3142 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3146 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3147 " into one vector!");
3149 unsigned VF = MaxVectorSize;
3151 // If we optimize the program for size, avoid creating the tail loop.
3153 // If we are unable to calculate the trip count then don't try to vectorize.
3155 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3159 // Find the maximum SIMD width that can fit within the trip count.
3160 VF = TC % MaxVectorSize;
3165 // If the trip count that we found modulo the vectorization factor is not
3166 // zero then we require a tail.
3168 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3174 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3175 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3177 Factor.Width = UserVF;
3181 float Cost = expectedCost(1);
3183 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3184 for (unsigned i=2; i <= VF; i*=2) {
3185 // Notice that the vector loop needs to be executed less times, so
3186 // we need to divide the cost of the vector loops by the width of
3187 // the vector elements.
3188 float VectorCost = expectedCost(i) / (float)i;
3189 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3190 (int)VectorCost << ".\n");
3191 if (VectorCost < Cost) {
3197 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3198 Factor.Width = Width;
3199 Factor.Cost = Width * Cost;
3203 unsigned LoopVectorizationCostModel::getWidestType() {
3204 unsigned MaxWidth = 8;
3207 for (Loop::block_iterator bb = TheLoop->block_begin(),
3208 be = TheLoop->block_end(); bb != be; ++bb) {
3209 BasicBlock *BB = *bb;
3211 // For each instruction in the loop.
3212 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3213 Type *T = it->getType();
3215 // Only examine Loads, Stores and PHINodes.
3216 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3219 // Examine PHI nodes that are reduction variables.
3220 if (PHINode *PN = dyn_cast<PHINode>(it))
3221 if (!Legal->getReductionVars()->count(PN))
3224 // Examine the stored values.
3225 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3226 T = ST->getValueOperand()->getType();
3228 // Ignore loaded pointer types and stored pointer types that are not
3229 // consecutive. However, we do want to take consecutive stores/loads of
3230 // pointer vectors into account.
3231 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3234 MaxWidth = std::max(MaxWidth,
3235 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3243 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3246 unsigned LoopCost) {
3248 // -- The unroll heuristics --
3249 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3250 // There are many micro-architectural considerations that we can't predict
3251 // at this level. For example frontend pressure (on decode or fetch) due to
3252 // code size, or the number and capabilities of the execution ports.
3254 // We use the following heuristics to select the unroll factor:
3255 // 1. If the code has reductions the we unroll in order to break the cross
3256 // iteration dependency.
3257 // 2. If the loop is really small then we unroll in order to reduce the loop
3259 // 3. We don't unroll if we think that we will spill registers to memory due
3260 // to the increased register pressure.
3262 // Use the user preference, unless 'auto' is selected.
3266 // When we optimize for size we don't unroll.
3270 // Do not unroll loops with a relatively small trip count.
3271 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3272 TheLoop->getLoopLatch());
3273 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3276 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3277 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3278 " vector registers\n");
3280 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3281 // We divide by these constants so assume that we have at least one
3282 // instruction that uses at least one register.
3283 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3284 R.NumInstructions = std::max(R.NumInstructions, 1U);
3286 // We calculate the unroll factor using the following formula.
3287 // Subtract the number of loop invariants from the number of available
3288 // registers. These registers are used by all of the unrolled instances.
3289 // Next, divide the remaining registers by the number of registers that is
3290 // required by the loop, in order to estimate how many parallel instances
3291 // fit without causing spills.
3292 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3294 // Clamp the unroll factor ranges to reasonable factors.
3295 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3297 // If we did not calculate the cost for VF (because the user selected the VF)
3298 // then we calculate the cost of VF here.
3300 LoopCost = expectedCost(VF);
3302 // Clamp the calculated UF to be between the 1 and the max unroll factor
3303 // that the target allows.
3304 if (UF > MaxUnrollSize)
3309 if (Legal->getReductionVars()->size()) {
3310 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3314 // We want to unroll tiny loops in order to reduce the loop overhead.
3315 // We assume that the cost overhead is 1 and we use the cost model
3316 // to estimate the cost of the loop and unroll until the cost of the
3317 // loop overhead is about 5% of the cost of the loop.
3318 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3319 if (LoopCost < 20) {
3320 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3321 unsigned NewUF = 20/LoopCost + 1;
3322 return std::min(NewUF, UF);
3325 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3329 LoopVectorizationCostModel::RegisterUsage
3330 LoopVectorizationCostModel::calculateRegisterUsage() {
3331 // This function calculates the register usage by measuring the highest number
3332 // of values that are alive at a single location. Obviously, this is a very
3333 // rough estimation. We scan the loop in a topological order in order and
3334 // assign a number to each instruction. We use RPO to ensure that defs are
3335 // met before their users. We assume that each instruction that has in-loop
3336 // users starts an interval. We record every time that an in-loop value is
3337 // used, so we have a list of the first and last occurrences of each
3338 // instruction. Next, we transpose this data structure into a multi map that
3339 // holds the list of intervals that *end* at a specific location. This multi
3340 // map allows us to perform a linear search. We scan the instructions linearly
3341 // and record each time that a new interval starts, by placing it in a set.
3342 // If we find this value in the multi-map then we remove it from the set.
3343 // The max register usage is the maximum size of the set.
3344 // We also search for instructions that are defined outside the loop, but are
3345 // used inside the loop. We need this number separately from the max-interval
3346 // usage number because when we unroll, loop-invariant values do not take
3348 LoopBlocksDFS DFS(TheLoop);
3352 R.NumInstructions = 0;
3354 // Each 'key' in the map opens a new interval. The values
3355 // of the map are the index of the 'last seen' usage of the
3356 // instruction that is the key.
3357 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3358 // Maps instruction to its index.
3359 DenseMap<unsigned, Instruction*> IdxToInstr;
3360 // Marks the end of each interval.
3361 IntervalMap EndPoint;
3362 // Saves the list of instruction indices that are used in the loop.
3363 SmallSet<Instruction*, 8> Ends;
3364 // Saves the list of values that are used in the loop but are
3365 // defined outside the loop, such as arguments and constants.
3366 SmallPtrSet<Value*, 8> LoopInvariants;
3369 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3370 be = DFS.endRPO(); bb != be; ++bb) {
3371 R.NumInstructions += (*bb)->size();
3372 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3374 Instruction *I = it;
3375 IdxToInstr[Index++] = I;
3377 // Save the end location of each USE.
3378 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3379 Value *U = I->getOperand(i);
3380 Instruction *Instr = dyn_cast<Instruction>(U);
3382 // Ignore non-instruction values such as arguments, constants, etc.
3383 if (!Instr) continue;
3385 // If this instruction is outside the loop then record it and continue.
3386 if (!TheLoop->contains(Instr)) {
3387 LoopInvariants.insert(Instr);
3391 // Overwrite previous end points.
3392 EndPoint[Instr] = Index;
3398 // Saves the list of intervals that end with the index in 'key'.
3399 typedef SmallVector<Instruction*, 2> InstrList;
3400 DenseMap<unsigned, InstrList> TransposeEnds;
3402 // Transpose the EndPoints to a list of values that end at each index.
3403 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3405 TransposeEnds[it->second].push_back(it->first);
3407 SmallSet<Instruction*, 8> OpenIntervals;
3408 unsigned MaxUsage = 0;
3411 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3412 for (unsigned int i = 0; i < Index; ++i) {
3413 Instruction *I = IdxToInstr[i];
3414 // Ignore instructions that are never used within the loop.
3415 if (!Ends.count(I)) continue;
3417 // Remove all of the instructions that end at this location.
3418 InstrList &List = TransposeEnds[i];
3419 for (unsigned int j=0, e = List.size(); j < e; ++j)
3420 OpenIntervals.erase(List[j]);
3422 // Count the number of live interals.
3423 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3425 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3426 OpenIntervals.size() <<"\n");
3428 // Add the current instruction to the list of open intervals.
3429 OpenIntervals.insert(I);
3432 unsigned Invariant = LoopInvariants.size();
3433 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3434 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3435 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3437 R.LoopInvariantRegs = Invariant;
3438 R.MaxLocalUsers = MaxUsage;
3442 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3446 for (Loop::block_iterator bb = TheLoop->block_begin(),
3447 be = TheLoop->block_end(); bb != be; ++bb) {
3448 unsigned BlockCost = 0;
3449 BasicBlock *BB = *bb;
3451 // For each instruction in the old loop.
3452 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3453 // Skip dbg intrinsics.
3454 if (isa<DbgInfoIntrinsic>(it))
3457 unsigned C = getInstructionCost(it, VF);
3459 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3460 VF << " For instruction: "<< *it << "\n");
3463 // We assume that if-converted blocks have a 50% chance of being executed.
3464 // When the code is scalar then some of the blocks are avoided due to CF.
3465 // When the code is vectorized we execute all code paths.
3466 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3476 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3477 // If we know that this instruction will remain uniform, check the cost of
3478 // the scalar version.
3479 if (Legal->isUniformAfterVectorization(I))
3482 Type *RetTy = I->getType();
3483 Type *VectorTy = ToVectorTy(RetTy, VF);
3485 // TODO: We need to estimate the cost of intrinsic calls.
3486 switch (I->getOpcode()) {
3487 case Instruction::GetElementPtr:
3488 // We mark this instruction as zero-cost because the cost of GEPs in
3489 // vectorized code depends on whether the corresponding memory instruction
3490 // is scalarized or not. Therefore, we handle GEPs with the memory
3491 // instruction cost.
3493 case Instruction::Br: {
3494 return TTI.getCFInstrCost(I->getOpcode());
3496 case Instruction::PHI:
3497 //TODO: IF-converted IFs become selects.
3499 case Instruction::Add:
3500 case Instruction::FAdd:
3501 case Instruction::Sub:
3502 case Instruction::FSub:
3503 case Instruction::Mul:
3504 case Instruction::FMul:
3505 case Instruction::UDiv:
3506 case Instruction::SDiv:
3507 case Instruction::FDiv:
3508 case Instruction::URem:
3509 case Instruction::SRem:
3510 case Instruction::FRem:
3511 case Instruction::Shl:
3512 case Instruction::LShr:
3513 case Instruction::AShr:
3514 case Instruction::And:
3515 case Instruction::Or:
3516 case Instruction::Xor: {
3517 // Certain instructions can be cheaper to vectorize if they have a constant
3518 // second vector operand. One example of this are shifts on x86.
3519 TargetTransformInfo::OperandValueKind Op1VK =
3520 TargetTransformInfo::OK_AnyValue;
3521 TargetTransformInfo::OperandValueKind Op2VK =
3522 TargetTransformInfo::OK_AnyValue;
3524 if (isa<ConstantInt>(I->getOperand(1)))
3525 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3527 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3529 case Instruction::Select: {
3530 SelectInst *SI = cast<SelectInst>(I);
3531 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3532 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3533 Type *CondTy = SI->getCondition()->getType();
3535 CondTy = VectorType::get(CondTy, VF);
3537 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3539 case Instruction::ICmp:
3540 case Instruction::FCmp: {
3541 Type *ValTy = I->getOperand(0)->getType();
3542 VectorTy = ToVectorTy(ValTy, VF);
3543 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3545 case Instruction::Store:
3546 case Instruction::Load: {
3547 StoreInst *SI = dyn_cast<StoreInst>(I);
3548 LoadInst *LI = dyn_cast<LoadInst>(I);
3549 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3551 VectorTy = ToVectorTy(ValTy, VF);
3553 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3554 unsigned AS = SI ? SI->getPointerAddressSpace() :
3555 LI->getPointerAddressSpace();
3556 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3557 // We add the cost of address computation here instead of with the gep
3558 // instruction because only here we know whether the operation is
3561 return TTI.getAddressComputationCost(VectorTy) +
3562 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3564 // Scalarized loads/stores.
3565 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3566 bool Reverse = ConsecutiveStride < 0;
3567 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3568 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3569 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3571 // The cost of extracting from the value vector and pointer vector.
3572 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3573 for (unsigned i = 0; i < VF; ++i) {
3574 // The cost of extracting the pointer operand.
3575 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3576 // In case of STORE, the cost of ExtractElement from the vector.
3577 // In case of LOAD, the cost of InsertElement into the returned
3579 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3580 Instruction::InsertElement,
3584 // The cost of the scalar loads/stores.
3585 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3586 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3591 // Wide load/stores.
3592 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3593 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3596 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3600 case Instruction::ZExt:
3601 case Instruction::SExt:
3602 case Instruction::FPToUI:
3603 case Instruction::FPToSI:
3604 case Instruction::FPExt:
3605 case Instruction::PtrToInt:
3606 case Instruction::IntToPtr:
3607 case Instruction::SIToFP:
3608 case Instruction::UIToFP:
3609 case Instruction::Trunc:
3610 case Instruction::FPTrunc:
3611 case Instruction::BitCast: {
3612 // We optimize the truncation of induction variable.
3613 // The cost of these is the same as the scalar operation.
3614 if (I->getOpcode() == Instruction::Trunc &&
3615 Legal->isInductionVariable(I->getOperand(0)))
3616 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3617 I->getOperand(0)->getType());
3619 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3620 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3622 case Instruction::Call: {
3623 CallInst *CI = cast<CallInst>(I);
3624 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3625 assert(ID && "Not an intrinsic call!");
3626 Type *RetTy = ToVectorTy(CI->getType(), VF);
3627 SmallVector<Type*, 4> Tys;
3628 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3629 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3630 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3633 // We are scalarizing the instruction. Return the cost of the scalar
3634 // instruction, plus the cost of insert and extract into vector
3635 // elements, times the vector width.
3638 if (!RetTy->isVoidTy() && VF != 1) {
3639 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3641 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3644 // The cost of inserting the results plus extracting each one of the
3646 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3649 // The cost of executing VF copies of the scalar instruction. This opcode
3650 // is unknown. Assume that it is the same as 'mul'.
3651 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3657 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3658 if (Scalar->isVoidTy() || VF == 1)
3660 return VectorType::get(Scalar, VF);
3663 char LoopVectorize::ID = 0;
3664 static const char lv_name[] = "Loop Vectorization";
3665 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3666 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3667 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3668 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3669 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3670 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3673 Pass *createLoopVectorizePass() {
3674 return new LoopVectorize();
3678 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3679 // Check for a store.
3680 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3681 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3683 // Check for a load.
3684 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3685 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;