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 memory disambiguation checks at runtime do not make more
118 /// than this number of comparisons.
119 static const unsigned RuntimeMemoryCheckThreshold = 8;
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, bool WritePtr);
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;
432 /// Holds the information if this pointer is used for writing to memory.
433 SmallVector<bool, 2> IsWritePtr;
436 /// A POD for saving information about induction variables.
437 struct InductionInfo {
438 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
439 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
446 /// ReductionList contains the reduction descriptors for all
447 /// of the reductions that were found in the loop.
448 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
450 /// InductionList saves induction variables and maps them to the
451 /// induction descriptor.
452 typedef MapVector<PHINode*, InductionInfo> InductionList;
454 /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
455 /// respective Store/Load instruction(s) to calculate aliasing.
456 typedef MapVector<Value*, Instruction* > AliasMap;
457 typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
459 /// Returns true if it is legal to vectorize this loop.
460 /// This does not mean that it is profitable to vectorize this
461 /// loop, only that it is legal to do so.
464 /// Returns the Induction variable.
465 PHINode *getInduction() { return Induction; }
467 /// Returns the reduction variables found in the loop.
468 ReductionList *getReductionVars() { return &Reductions; }
470 /// Returns the induction variables found in the loop.
471 InductionList *getInductionVars() { return &Inductions; }
473 /// Returns True if V is an induction variable in this loop.
474 bool isInductionVariable(const Value *V);
476 /// Return true if the block BB needs to be predicated in order for the loop
477 /// to be vectorized.
478 bool blockNeedsPredication(BasicBlock *BB);
480 /// Check if this pointer is consecutive when vectorizing. This happens
481 /// when the last index of the GEP is the induction variable, or that the
482 /// pointer itself is an induction variable.
483 /// This check allows us to vectorize A[idx] into a wide load/store.
485 /// 0 - Stride is unknown or non consecutive.
486 /// 1 - Address is consecutive.
487 /// -1 - Address is consecutive, and decreasing.
488 int isConsecutivePtr(Value *Ptr);
490 /// Returns true if the value V is uniform within the loop.
491 bool isUniform(Value *V);
493 /// Returns true if this instruction will remain scalar after vectorization.
494 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
496 /// Returns the information that we collected about runtime memory check.
497 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
499 /// This function returns the identity element (or neutral element) for
501 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
503 /// Check if a single basic block loop is vectorizable.
504 /// At this point we know that this is a loop with a constant trip count
505 /// and we only need to check individual instructions.
506 bool canVectorizeInstrs();
508 /// When we vectorize loops we may change the order in which
509 /// we read and write from memory. This method checks if it is
510 /// legal to vectorize the code, considering only memory constrains.
511 /// Returns true if the loop is vectorizable
512 bool canVectorizeMemory();
514 /// Return true if we can vectorize this loop using the IF-conversion
516 bool canVectorizeWithIfConvert();
518 /// Collect the variables that need to stay uniform after vectorization.
519 void collectLoopUniforms();
521 /// Return true if all of the instructions in the block can be speculatively
523 bool blockCanBePredicated(BasicBlock *BB);
525 /// Returns True, if 'Phi' is the kind of reduction variable for type
526 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
527 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
528 /// Returns a struct describing if the instruction 'I' can be a reduction
529 /// variable of type 'Kind'. If the reduction is a min/max pattern of
530 /// select(icmp()) this function advances the instruction pointer 'I' from the
531 /// compare instruction to the select instruction and stores this pointer in
532 /// 'PatternLastInst' member of the returned struct.
533 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
534 ReductionInstDesc &Desc);
535 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
536 /// pattern corresponding to a min(X, Y) or max(X, Y).
537 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
538 ReductionInstDesc &Prev);
539 /// Returns the induction kind of Phi. This function may return NoInduction
540 /// if the PHI is not an induction variable.
541 InductionKind isInductionVariable(PHINode *Phi);
542 /// Return true if can compute the address bounds of Ptr within the loop.
543 bool hasComputableBounds(Value *Ptr);
544 /// Return true if there is the chance of write reorder.
545 bool hasPossibleGlobalWriteReorder(Value *Object,
547 AliasMultiMap &WriteObjects,
548 unsigned MaxByteWidth);
549 /// Return the AA location for a load or a store.
550 AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
553 /// The loop that we evaluate.
557 /// DataLayout analysis.
562 TargetTransformInfo *TTI;
565 /// Target Library Info.
566 TargetLibraryInfo *TLI;
568 // --- vectorization state --- //
570 /// Holds the integer induction variable. This is the counter of the
573 /// Holds the reduction variables.
574 ReductionList Reductions;
575 /// Holds all of the induction variables that we found in the loop.
576 /// Notice that inductions don't need to start at zero and that induction
577 /// variables can be pointers.
578 InductionList Inductions;
580 /// Allowed outside users. This holds the reduction
581 /// vars which can be accessed from outside the loop.
582 SmallPtrSet<Value*, 4> AllowedExit;
583 /// This set holds the variables which are known to be uniform after
585 SmallPtrSet<Instruction*, 4> Uniforms;
586 /// We need to check that all of the pointers in this list are disjoint
588 RuntimePointerCheck PtrRtCheck;
591 /// LoopVectorizationCostModel - estimates the expected speedups due to
593 /// In many cases vectorization is not profitable. This can happen because of
594 /// a number of reasons. In this class we mainly attempt to predict the
595 /// expected speedup/slowdowns due to the supported instruction set. We use the
596 /// TargetTransformInfo to query the different backends for the cost of
597 /// different operations.
598 class LoopVectorizationCostModel {
600 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
601 LoopVectorizationLegality *Legal,
602 const TargetTransformInfo &TTI,
603 DataLayout *DL, const TargetLibraryInfo *TLI)
604 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
606 /// Information about vectorization costs
607 struct VectorizationFactor {
608 unsigned Width; // Vector width with best cost
609 unsigned Cost; // Cost of the loop with that width
611 /// \return The most profitable vectorization factor and the cost of that VF.
612 /// This method checks every power of two up to VF. If UserVF is not ZERO
613 /// then this vectorization factor will be selected if vectorization is
615 VectorizationFactor selectVectorizationFactor(bool OptForSize,
618 /// \return The size (in bits) of the widest type in the code that
619 /// needs to be vectorized. We ignore values that remain scalar such as
620 /// 64 bit loop indices.
621 unsigned getWidestType();
623 /// \return The most profitable unroll factor.
624 /// If UserUF is non-zero then this method finds the best unroll-factor
625 /// based on register pressure and other parameters.
626 /// VF and LoopCost are the selected vectorization factor and the cost of the
628 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
631 /// \brief A struct that represents some properties of the register usage
633 struct RegisterUsage {
634 /// Holds the number of loop invariant values that are used in the loop.
635 unsigned LoopInvariantRegs;
636 /// Holds the maximum number of concurrent live intervals in the loop.
637 unsigned MaxLocalUsers;
638 /// Holds the number of instructions in the loop.
639 unsigned NumInstructions;
642 /// \return information about the register usage of the loop.
643 RegisterUsage calculateRegisterUsage();
646 /// Returns the expected execution cost. The unit of the cost does
647 /// not matter because we use the 'cost' units to compare different
648 /// vector widths. The cost that is returned is *not* normalized by
649 /// the factor width.
650 unsigned expectedCost(unsigned VF);
652 /// Returns the execution time cost of an instruction for a given vector
653 /// width. Vector width of one means scalar.
654 unsigned getInstructionCost(Instruction *I, unsigned VF);
656 /// A helper function for converting Scalar types to vector types.
657 /// If the incoming type is void, we return void. If the VF is 1, we return
659 static Type* ToVectorTy(Type *Scalar, unsigned VF);
661 /// Returns whether the instruction is a load or store and will be a emitted
662 /// as a vector operation.
663 bool isConsecutiveLoadOrStore(Instruction *I);
665 /// The loop that we evaluate.
669 /// Loop Info analysis.
671 /// Vectorization legality.
672 LoopVectorizationLegality *Legal;
673 /// Vector target information.
674 const TargetTransformInfo &TTI;
675 /// Target data layout information.
677 /// Target Library Info.
678 const TargetLibraryInfo *TLI;
681 /// The LoopVectorize Pass.
682 struct LoopVectorize : public LoopPass {
683 /// Pass identification, replacement for typeid
686 explicit LoopVectorize() : LoopPass(ID) {
687 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
693 TargetTransformInfo *TTI;
696 TargetLibraryInfo *TLI;
698 virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
699 // We only vectorize innermost loops.
703 SE = &getAnalysis<ScalarEvolution>();
704 DL = getAnalysisIfAvailable<DataLayout>();
705 LI = &getAnalysis<LoopInfo>();
706 TTI = &getAnalysis<TargetTransformInfo>();
707 DT = &getAnalysis<DominatorTree>();
708 AA = getAnalysisIfAvailable<AliasAnalysis>();
709 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
712 DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
716 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
717 L->getHeader()->getParent()->getName() << "\"\n");
719 // Check if it is legal to vectorize the loop.
720 LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
721 if (!LVL.canVectorize()) {
722 DEBUG(dbgs() << "LV: Not vectorizing.\n");
726 // Use the cost model.
727 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
729 // Check the function attributes to find out if this function should be
730 // optimized for size.
731 Function *F = L->getHeader()->getParent();
732 Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
733 Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
734 unsigned FnIndex = AttributeSet::FunctionIndex;
735 bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
736 bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
739 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
740 "attribute is used.\n");
744 // Select the optimal vectorization factor.
745 LoopVectorizationCostModel::VectorizationFactor VF;
746 VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
747 // Select the unroll factor.
748 unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll,
752 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
756 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
757 F->getParent()->getModuleIdentifier()<<"\n");
758 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
760 // If we decided that it is *legal* to vectorize the loop then do it.
761 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
764 DEBUG(verifyFunction(*L->getHeader()->getParent()));
768 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
769 LoopPass::getAnalysisUsage(AU);
770 AU.addRequiredID(LoopSimplifyID);
771 AU.addRequiredID(LCSSAID);
772 AU.addRequired<DominatorTree>();
773 AU.addRequired<LoopInfo>();
774 AU.addRequired<ScalarEvolution>();
775 AU.addRequired<TargetTransformInfo>();
776 AU.addPreserved<LoopInfo>();
777 AU.addPreserved<DominatorTree>();
782 } // end anonymous namespace
784 //===----------------------------------------------------------------------===//
785 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
786 // LoopVectorizationCostModel.
787 //===----------------------------------------------------------------------===//
790 LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
791 Loop *Lp, Value *Ptr,
793 const SCEV *Sc = SE->getSCEV(Ptr);
794 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
795 assert(AR && "Invalid addrec expression");
796 const SCEV *Ex = SE->getExitCount(Lp, Lp->getLoopLatch());
797 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
798 Pointers.push_back(Ptr);
799 Starts.push_back(AR->getStart());
800 Ends.push_back(ScEnd);
801 IsWritePtr.push_back(WritePtr);
804 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
805 // Save the current insertion location.
806 Instruction *Loc = Builder.GetInsertPoint();
808 // We need to place the broadcast of invariant variables outside the loop.
809 Instruction *Instr = dyn_cast<Instruction>(V);
810 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
811 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
813 // Place the code for broadcasting invariant variables in the new preheader.
815 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
817 // Broadcast the scalar into all locations in the vector.
818 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
820 // Restore the builder insertion point.
822 Builder.SetInsertPoint(Loc);
827 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
829 assert(Val->getType()->isVectorTy() && "Must be a vector");
830 assert(Val->getType()->getScalarType()->isIntegerTy() &&
831 "Elem must be an integer");
833 Type *ITy = Val->getType()->getScalarType();
834 VectorType *Ty = cast<VectorType>(Val->getType());
835 int VLen = Ty->getNumElements();
836 SmallVector<Constant*, 8> Indices;
838 // Create a vector of consecutive numbers from zero to VF.
839 for (int i = 0; i < VLen; ++i) {
840 int Idx = Negate ? (-i): i;
841 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
844 // Add the consecutive indices to the vector value.
845 Constant *Cv = ConstantVector::get(Indices);
846 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
847 return Builder.CreateAdd(Val, Cv, "induction");
850 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
851 assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
852 // Make sure that the pointer does not point to structs.
853 if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
856 // If this value is a pointer induction variable we know it is consecutive.
857 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
858 if (Phi && Inductions.count(Phi)) {
859 InductionInfo II = Inductions[Phi];
860 if (IK_PtrInduction == II.IK)
862 else if (IK_ReversePtrInduction == II.IK)
866 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
870 unsigned NumOperands = Gep->getNumOperands();
871 Value *LastIndex = Gep->getOperand(NumOperands - 1);
873 Value *GpPtr = Gep->getPointerOperand();
874 // If this GEP value is a consecutive pointer induction variable and all of
875 // the indices are constant then we know it is consecutive. We can
876 Phi = dyn_cast<PHINode>(GpPtr);
877 if (Phi && Inductions.count(Phi)) {
879 // Make sure that the pointer does not point to structs.
880 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
881 if (GepPtrType->getElementType()->isAggregateType())
884 // Make sure that all of the index operands are loop invariant.
885 for (unsigned i = 1; i < NumOperands; ++i)
886 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
889 InductionInfo II = Inductions[Phi];
890 if (IK_PtrInduction == II.IK)
892 else if (IK_ReversePtrInduction == II.IK)
896 // Check that all of the gep indices are uniform except for the last.
897 for (unsigned i = 0; i < NumOperands - 1; ++i)
898 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
901 // We can emit wide load/stores only if the last index is the induction
903 const SCEV *Last = SE->getSCEV(LastIndex);
904 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
905 const SCEV *Step = AR->getStepRecurrence(*SE);
907 // The memory is consecutive because the last index is consecutive
908 // and all other indices are loop invariant.
911 if (Step->isAllOnesValue())
918 bool LoopVectorizationLegality::isUniform(Value *V) {
919 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
922 InnerLoopVectorizer::VectorParts&
923 InnerLoopVectorizer::getVectorValue(Value *V) {
924 assert(V != Induction && "The new induction variable should not be used.");
925 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
927 // If we have this scalar in the map, return it.
929 return WidenMap.get(V);
931 // If this scalar is unknown, assume that it is a constant or that it is
932 // loop invariant. Broadcast V and save the value for future uses.
933 Value *B = getBroadcastInstrs(V);
934 return WidenMap.splat(V, B);
937 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
938 assert(Vec->getType()->isVectorTy() && "Invalid type");
939 SmallVector<Constant*, 8> ShuffleMask;
940 for (unsigned i = 0; i < VF; ++i)
941 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
943 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
944 ConstantVector::get(ShuffleMask),
949 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
950 LoopVectorizationLegality *Legal) {
951 // Attempt to issue a wide load.
952 LoadInst *LI = dyn_cast<LoadInst>(Instr);
953 StoreInst *SI = dyn_cast<StoreInst>(Instr);
955 assert((LI || SI) && "Invalid Load/Store instruction");
957 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
958 Type *DataTy = VectorType::get(ScalarDataTy, VF);
959 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
960 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
962 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
963 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
965 if (ScalarAllocatedSize != VectorElementSize)
966 return scalarizeInstruction(Instr);
968 // If the pointer is loop invariant or if it is non consecutive,
969 // scalarize the load.
970 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
971 bool Reverse = ConsecutiveStride < 0;
972 bool UniformLoad = LI && Legal->isUniform(Ptr);
973 if (!ConsecutiveStride || UniformLoad)
974 return scalarizeInstruction(Instr);
976 Constant *Zero = Builder.getInt32(0);
977 VectorParts &Entry = WidenMap.get(Instr);
979 // Handle consecutive loads/stores.
980 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
981 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
982 Value *PtrOperand = Gep->getPointerOperand();
983 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
984 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
986 // Create the new GEP with the new induction variable.
987 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
988 Gep2->setOperand(0, FirstBasePtr);
989 Gep2->setName("gep.indvar.base");
990 Ptr = Builder.Insert(Gep2);
992 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
993 OrigLoop) && "Base ptr must be invariant");
995 // The last index does not have to be the induction. It can be
996 // consecutive and be a function of the index. For example A[I+1];
997 unsigned NumOperands = Gep->getNumOperands();
999 Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
1000 VectorParts &GEPParts = getVectorValue(LastGepOperand);
1001 Value *LastIndex = GEPParts[0];
1002 LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
1004 // Create the new GEP with the new induction variable.
1005 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1006 Gep2->setOperand(NumOperands - 1, LastIndex);
1007 Gep2->setName("gep.indvar.idx");
1008 Ptr = Builder.Insert(Gep2);
1010 // Use the induction element ptr.
1011 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1012 VectorParts &PtrVal = getVectorValue(Ptr);
1013 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1018 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1019 "We do not allow storing to uniform addresses");
1021 VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
1022 for (unsigned Part = 0; Part < UF; ++Part) {
1023 // Calculate the pointer for the specific unroll-part.
1024 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1027 // If we store to reverse consecutive memory locations then we need
1028 // to reverse the order of elements in the stored value.
1029 StoredVal[Part] = reverseVector(StoredVal[Part]);
1030 // If the address is consecutive but reversed, then the
1031 // wide store needs to start at the last vector element.
1032 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1033 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1036 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1037 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1041 for (unsigned Part = 0; Part < UF; ++Part) {
1042 // Calculate the pointer for the specific unroll-part.
1043 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1046 // If the address is consecutive but reversed, then the
1047 // wide store needs to start at the last vector element.
1048 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1049 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1052 Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
1053 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1054 cast<LoadInst>(LI)->setAlignment(Alignment);
1055 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1059 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
1060 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1061 // Holds vector parameters or scalars, in case of uniform vals.
1062 SmallVector<VectorParts, 4> Params;
1064 // Find all of the vectorized parameters.
1065 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1066 Value *SrcOp = Instr->getOperand(op);
1068 // If we are accessing the old induction variable, use the new one.
1069 if (SrcOp == OldInduction) {
1070 Params.push_back(getVectorValue(SrcOp));
1074 // Try using previously calculated values.
1075 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1077 // If the src is an instruction that appeared earlier in the basic block
1078 // then it should already be vectorized.
1079 if (SrcInst && OrigLoop->contains(SrcInst)) {
1080 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1081 // The parameter is a vector value from earlier.
1082 Params.push_back(WidenMap.get(SrcInst));
1084 // The parameter is a scalar from outside the loop. Maybe even a constant.
1085 VectorParts Scalars;
1086 Scalars.append(UF, SrcOp);
1087 Params.push_back(Scalars);
1091 assert(Params.size() == Instr->getNumOperands() &&
1092 "Invalid number of operands");
1094 // Does this instruction return a value ?
1095 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1097 Value *UndefVec = IsVoidRetTy ? 0 :
1098 UndefValue::get(VectorType::get(Instr->getType(), VF));
1099 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1100 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1102 // For each vector unroll 'part':
1103 for (unsigned Part = 0; Part < UF; ++Part) {
1104 // For each scalar that we create:
1105 for (unsigned Width = 0; Width < VF; ++Width) {
1106 Instruction *Cloned = Instr->clone();
1108 Cloned->setName(Instr->getName() + ".cloned");
1109 // Replace the operands of the cloned instrucions with extracted scalars.
1110 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1111 Value *Op = Params[op][Part];
1112 // Param is a vector. Need to extract the right lane.
1113 if (Op->getType()->isVectorTy())
1114 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1115 Cloned->setOperand(op, Op);
1118 // Place the cloned scalar in the new loop.
1119 Builder.Insert(Cloned);
1121 // If the original scalar returns a value we need to place it in a vector
1122 // so that future users will be able to use it.
1124 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1125 Builder.getInt32(Width));
1131 InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
1133 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1134 Legal->getRuntimePointerCheck();
1136 if (!PtrRtCheck->Need)
1139 Instruction *MemoryRuntimeCheck = 0;
1140 unsigned NumPointers = PtrRtCheck->Pointers.size();
1141 SmallVector<Value* , 2> Starts;
1142 SmallVector<Value* , 2> Ends;
1144 SCEVExpander Exp(*SE, "induction");
1146 // Use this type for pointer arithmetic.
1147 Type* PtrArithTy = Type::getInt8PtrTy(Loc->getContext(), 0);
1149 for (unsigned i = 0; i < NumPointers; ++i) {
1150 Value *Ptr = PtrRtCheck->Pointers[i];
1151 const SCEV *Sc = SE->getSCEV(Ptr);
1153 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1154 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1156 Starts.push_back(Ptr);
1157 Ends.push_back(Ptr);
1159 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n");
1161 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1162 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1163 Starts.push_back(Start);
1164 Ends.push_back(End);
1168 IRBuilder<> ChkBuilder(Loc);
1170 for (unsigned i = 0; i < NumPointers; ++i) {
1171 for (unsigned j = i+1; j < NumPointers; ++j) {
1172 // No need to check if two readonly pointers intersect.
1173 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1176 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
1177 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
1178 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
1179 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
1181 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1182 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1183 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1184 if (MemoryRuntimeCheck)
1185 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1188 MemoryRuntimeCheck = cast<Instruction>(IsConflict);
1192 return MemoryRuntimeCheck;
1196 InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
1198 In this function we generate a new loop. The new loop will contain
1199 the vectorized instructions while the old loop will continue to run the
1202 [ ] <-- vector loop bypass (may consist of multiple blocks).
1205 | [ ] <-- vector pre header.
1209 | [ ]_| <-- vector loop.
1212 >[ ] <--- middle-block.
1215 | [ ] <--- new preheader.
1219 | [ ]_| <-- old scalar loop to handle remainder.
1222 >[ ] <-- exit block.
1226 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1227 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1228 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1229 assert(ExitBlock && "Must have an exit block");
1231 // Mark the old scalar loop with metadata that tells us not to vectorize this
1232 // loop again if we run into it.
1233 MDNode *MD = MDNode::get(OldBasicBlock->getContext(), None);
1234 OldBasicBlock->getTerminator()->setMetadata(AlreadyVectorizedMDName, MD);
1236 // Some loops have a single integer induction variable, while other loops
1237 // don't. One example is c++ iterators that often have multiple pointer
1238 // induction variables. In the code below we also support a case where we
1239 // don't have a single induction variable.
1240 OldInduction = Legal->getInduction();
1241 Type *IdxTy = OldInduction ? OldInduction->getType() :
1242 DL->getIntPtrType(SE->getContext());
1244 // Find the loop boundaries.
1245 const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
1246 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1248 // Get the total trip count from the count by adding 1.
1249 ExitCount = SE->getAddExpr(ExitCount,
1250 SE->getConstant(ExitCount->getType(), 1));
1252 // Expand the trip count and place the new instructions in the preheader.
1253 // Notice that the pre-header does not change, only the loop body.
1254 SCEVExpander Exp(*SE, "induction");
1256 // Count holds the overall loop count (N).
1257 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1258 BypassBlock->getTerminator());
1260 // The loop index does not have to start at Zero. Find the original start
1261 // value from the induction PHI node. If we don't have an induction variable
1262 // then we know that it starts at zero.
1263 Value *StartIdx = OldInduction ?
1264 OldInduction->getIncomingValueForBlock(BypassBlock):
1265 ConstantInt::get(IdxTy, 0);
1267 assert(BypassBlock && "Invalid loop structure");
1268 LoopBypassBlocks.push_back(BypassBlock);
1270 // Split the single block loop into the two loop structure described above.
1271 BasicBlock *VectorPH =
1272 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1273 BasicBlock *VecBody =
1274 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1275 BasicBlock *MiddleBlock =
1276 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1277 BasicBlock *ScalarPH =
1278 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1280 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1282 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1284 // Generate the induction variable.
1285 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1286 // The loop step is equal to the vectorization factor (num of SIMD elements)
1287 // times the unroll factor (num of SIMD instructions).
1288 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1290 // This is the IR builder that we use to add all of the logic for bypassing
1291 // the new vector loop.
1292 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1294 // We may need to extend the index in case there is a type mismatch.
1295 // We know that the count starts at zero and does not overflow.
1296 if (Count->getType() != IdxTy) {
1297 // The exit count can be of pointer type. Convert it to the correct
1299 if (ExitCount->getType()->isPointerTy())
1300 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1302 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1305 // Add the start index to the loop count to get the new end index.
1306 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1308 // Now we need to generate the expression for N - (N % VF), which is
1309 // the part that the vectorized body will execute.
1310 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1311 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1312 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1313 "end.idx.rnd.down");
1315 // Now, compare the new count to zero. If it is zero skip the vector loop and
1316 // jump to the scalar loop.
1317 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1320 BasicBlock *LastBypassBlock = BypassBlock;
1322 // Generate the code that checks in runtime if arrays overlap. We put the
1323 // checks into a separate block to make the more common case of few elements
1325 Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
1326 BypassBlock->getTerminator());
1327 if (MemRuntimeCheck) {
1328 // Create a new block containing the memory check.
1329 BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
1331 LoopBypassBlocks.push_back(CheckBlock);
1333 // Replace the branch into the memory check block with a conditional branch
1334 // for the "few elements case".
1335 Instruction *OldTerm = BypassBlock->getTerminator();
1336 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1337 OldTerm->eraseFromParent();
1339 Cmp = MemRuntimeCheck;
1340 LastBypassBlock = CheckBlock;
1343 LastBypassBlock->getTerminator()->eraseFromParent();
1344 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
1347 // We are going to resume the execution of the scalar loop.
1348 // Go over all of the induction variables that we found and fix the
1349 // PHIs that are left in the scalar version of the loop.
1350 // The starting values of PHI nodes depend on the counter of the last
1351 // iteration in the vectorized loop.
1352 // If we come from a bypass edge then we need to start from the original
1355 // This variable saves the new starting index for the scalar loop.
1356 PHINode *ResumeIndex = 0;
1357 LoopVectorizationLegality::InductionList::iterator I, E;
1358 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
1359 for (I = List->begin(), E = List->end(); I != E; ++I) {
1360 PHINode *OrigPhi = I->first;
1361 LoopVectorizationLegality::InductionInfo II = I->second;
1362 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
1363 MiddleBlock->getTerminator());
1364 Value *EndValue = 0;
1366 case LoopVectorizationLegality::IK_NoInduction:
1367 llvm_unreachable("Unknown induction");
1368 case LoopVectorizationLegality::IK_IntInduction: {
1369 // Handle the integer induction counter:
1370 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
1371 assert(OrigPhi == OldInduction && "Unknown integer PHI");
1372 // We know what the end value is.
1373 EndValue = IdxEndRoundDown;
1374 // We also know which PHI node holds it.
1375 ResumeIndex = ResumeVal;
1378 case LoopVectorizationLegality::IK_ReverseIntInduction: {
1379 // Convert the CountRoundDown variable to the PHI size.
1380 unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
1381 unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
1382 Value *CRD = CountRoundDown;
1383 if (CRDSize > IISize)
1384 CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
1385 II.StartValue->getType(), "tr.crd",
1386 LoopBypassBlocks.back()->getTerminator());
1387 else if (CRDSize < IISize)
1388 CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
1389 II.StartValue->getType(),
1391 LoopBypassBlocks.back()->getTerminator());
1392 // Handle reverse integer induction counter:
1394 BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
1395 LoopBypassBlocks.back()->getTerminator());
1398 case LoopVectorizationLegality::IK_PtrInduction: {
1399 // For pointer induction variables, calculate the offset using
1402 GetElementPtrInst::Create(II.StartValue, CountRoundDown, "ptr.ind.end",
1403 LoopBypassBlocks.back()->getTerminator());
1406 case LoopVectorizationLegality::IK_ReversePtrInduction: {
1407 // The value at the end of the loop for the reverse pointer is calculated
1408 // by creating a GEP with a negative index starting from the start value.
1409 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
1410 Value *NegIdx = BinaryOperator::CreateSub(Zero, CountRoundDown,
1412 LoopBypassBlocks.back()->getTerminator());
1413 EndValue = GetElementPtrInst::Create(II.StartValue, NegIdx,
1415 LoopBypassBlocks.back()->getTerminator());
1420 // The new PHI merges the original incoming value, in case of a bypass,
1421 // or the value at the end of the vectorized loop.
1422 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1423 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
1424 ResumeVal->addIncoming(EndValue, VecBody);
1426 // Fix the scalar body counter (PHI node).
1427 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
1428 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
1431 // If we are generating a new induction variable then we also need to
1432 // generate the code that calculates the exit value. This value is not
1433 // simply the end of the counter because we may skip the vectorized body
1434 // in case of a runtime check.
1436 assert(!ResumeIndex && "Unexpected resume value found");
1437 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
1438 MiddleBlock->getTerminator());
1439 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1440 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
1441 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
1444 // Make sure that we found the index where scalar loop needs to continue.
1445 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
1446 "Invalid resume Index");
1448 // Add a check in the middle block to see if we have completed
1449 // all of the iterations in the first vector loop.
1450 // If (N - N%VF) == N, then we *don't* need to run the remainder.
1451 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
1452 ResumeIndex, "cmp.n",
1453 MiddleBlock->getTerminator());
1455 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
1456 // Remove the old terminator.
1457 MiddleBlock->getTerminator()->eraseFromParent();
1459 // Create i+1 and fill the PHINode.
1460 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
1461 Induction->addIncoming(StartIdx, VectorPH);
1462 Induction->addIncoming(NextIdx, VecBody);
1463 // Create the compare.
1464 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
1465 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
1467 // Now we have two terminators. Remove the old one from the block.
1468 VecBody->getTerminator()->eraseFromParent();
1470 // Get ready to start creating new instructions into the vectorized body.
1471 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
1473 // Create and register the new vector loop.
1474 Loop* Lp = new Loop();
1475 Loop *ParentLoop = OrigLoop->getParentLoop();
1477 // Insert the new loop into the loop nest and register the new basic blocks.
1479 ParentLoop->addChildLoop(Lp);
1480 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
1481 ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
1482 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1483 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1484 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1486 LI->addTopLevelLoop(Lp);
1489 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1492 LoopVectorPreHeader = VectorPH;
1493 LoopScalarPreHeader = ScalarPH;
1494 LoopMiddleBlock = MiddleBlock;
1495 LoopExitBlock = ExitBlock;
1496 LoopVectorBody = VecBody;
1497 LoopScalarBody = OldBasicBlock;
1500 /// This function returns the identity element (or neutral element) for
1501 /// the operation K.
1503 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
1508 // Adding, Xoring, Oring zero to a number does not change it.
1509 return ConstantInt::get(Tp, 0);
1510 case RK_IntegerMult:
1511 // Multiplying a number by 1 does not change it.
1512 return ConstantInt::get(Tp, 1);
1514 // AND-ing a number with an all-1 value does not change it.
1515 return ConstantInt::get(Tp, -1, true);
1517 // Multiplying a number by 1 does not change it.
1518 return ConstantFP::get(Tp, 1.0L);
1520 // Adding zero to a number does not change it.
1521 return ConstantFP::get(Tp, 0.0L);
1523 llvm_unreachable("Unknown reduction kind");
1527 static Intrinsic::ID
1528 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
1529 // If we have an intrinsic call, check if it is trivially vectorizable.
1530 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1531 switch (II->getIntrinsicID()) {
1532 case Intrinsic::sqrt:
1533 case Intrinsic::sin:
1534 case Intrinsic::cos:
1535 case Intrinsic::exp:
1536 case Intrinsic::exp2:
1537 case Intrinsic::log:
1538 case Intrinsic::log10:
1539 case Intrinsic::log2:
1540 case Intrinsic::fabs:
1541 case Intrinsic::floor:
1542 case Intrinsic::ceil:
1543 case Intrinsic::trunc:
1544 case Intrinsic::rint:
1545 case Intrinsic::nearbyint:
1546 case Intrinsic::pow:
1547 case Intrinsic::fma:
1548 case Intrinsic::fmuladd:
1549 return II->getIntrinsicID();
1551 return Intrinsic::not_intrinsic;
1556 return Intrinsic::not_intrinsic;
1559 Function *F = CI->getCalledFunction();
1560 // We're going to make assumptions on the semantics of the functions, check
1561 // that the target knows that it's available in this environment.
1562 if (!F || !TLI->getLibFunc(F->getName(), Func))
1563 return Intrinsic::not_intrinsic;
1565 // Otherwise check if we have a call to a function that can be turned into a
1566 // vector intrinsic.
1573 return Intrinsic::sin;
1577 return Intrinsic::cos;
1581 return Intrinsic::exp;
1583 case LibFunc::exp2f:
1584 case LibFunc::exp2l:
1585 return Intrinsic::exp2;
1589 return Intrinsic::log;
1590 case LibFunc::log10:
1591 case LibFunc::log10f:
1592 case LibFunc::log10l:
1593 return Intrinsic::log10;
1595 case LibFunc::log2f:
1596 case LibFunc::log2l:
1597 return Intrinsic::log2;
1599 case LibFunc::fabsf:
1600 case LibFunc::fabsl:
1601 return Intrinsic::fabs;
1602 case LibFunc::floor:
1603 case LibFunc::floorf:
1604 case LibFunc::floorl:
1605 return Intrinsic::floor;
1607 case LibFunc::ceilf:
1608 case LibFunc::ceill:
1609 return Intrinsic::ceil;
1610 case LibFunc::trunc:
1611 case LibFunc::truncf:
1612 case LibFunc::truncl:
1613 return Intrinsic::trunc;
1615 case LibFunc::rintf:
1616 case LibFunc::rintl:
1617 return Intrinsic::rint;
1618 case LibFunc::nearbyint:
1619 case LibFunc::nearbyintf:
1620 case LibFunc::nearbyintl:
1621 return Intrinsic::nearbyint;
1625 return Intrinsic::pow;
1628 return Intrinsic::not_intrinsic;
1631 /// This function translates the reduction kind to an LLVM binary operator.
1633 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
1635 case LoopVectorizationLegality::RK_IntegerAdd:
1636 return Instruction::Add;
1637 case LoopVectorizationLegality::RK_IntegerMult:
1638 return Instruction::Mul;
1639 case LoopVectorizationLegality::RK_IntegerOr:
1640 return Instruction::Or;
1641 case LoopVectorizationLegality::RK_IntegerAnd:
1642 return Instruction::And;
1643 case LoopVectorizationLegality::RK_IntegerXor:
1644 return Instruction::Xor;
1645 case LoopVectorizationLegality::RK_FloatMult:
1646 return Instruction::FMul;
1647 case LoopVectorizationLegality::RK_FloatAdd:
1648 return Instruction::FAdd;
1649 case LoopVectorizationLegality::RK_IntegerMinMax:
1650 return Instruction::ICmp;
1652 llvm_unreachable("Unknown reduction operation");
1656 Value *createMinMaxOp(IRBuilder<> &Builder,
1657 LoopVectorizationLegality::MinMaxReductionKind RK,
1660 CmpInst::Predicate P = CmpInst::ICMP_NE;
1663 llvm_unreachable("Unknown min/max reduction kind");
1664 case LoopVectorizationLegality::MRK_UIntMin:
1665 P = CmpInst::ICMP_ULT;
1667 case LoopVectorizationLegality::MRK_UIntMax:
1668 P = CmpInst::ICMP_UGT;
1670 case LoopVectorizationLegality::MRK_SIntMin:
1671 P = CmpInst::ICMP_SLT;
1673 case LoopVectorizationLegality::MRK_SIntMax:
1674 P = CmpInst::ICMP_SGT;
1676 Value *Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
1677 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
1682 InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
1683 //===------------------------------------------------===//
1685 // Notice: any optimization or new instruction that go
1686 // into the code below should be also be implemented in
1689 //===------------------------------------------------===//
1690 Constant *Zero = Builder.getInt32(0);
1692 // In order to support reduction variables we need to be able to vectorize
1693 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
1694 // stages. First, we create a new vector PHI node with no incoming edges.
1695 // We use this value when we vectorize all of the instructions that use the
1696 // PHI. Next, after all of the instructions in the block are complete we
1697 // add the new incoming edges to the PHI. At this point all of the
1698 // instructions in the basic block are vectorized, so we can use them to
1699 // construct the PHI.
1700 PhiVector RdxPHIsToFix;
1702 // Scan the loop in a topological order to ensure that defs are vectorized
1704 LoopBlocksDFS DFS(OrigLoop);
1707 // Vectorize all of the blocks in the original loop.
1708 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
1709 be = DFS.endRPO(); bb != be; ++bb)
1710 vectorizeBlockInLoop(Legal, *bb, &RdxPHIsToFix);
1712 // At this point every instruction in the original loop is widened to
1713 // a vector form. We are almost done. Now, we need to fix the PHI nodes
1714 // that we vectorized. The PHI nodes are currently empty because we did
1715 // not want to introduce cycles. Notice that the remaining PHI nodes
1716 // that we need to fix are reduction variables.
1718 // Create the 'reduced' values for each of the induction vars.
1719 // The reduced values are the vector values that we scalarize and combine
1720 // after the loop is finished.
1721 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
1723 PHINode *RdxPhi = *it;
1724 assert(RdxPhi && "Unable to recover vectorized PHI");
1726 // Find the reduction variable descriptor.
1727 assert(Legal->getReductionVars()->count(RdxPhi) &&
1728 "Unable to find the reduction variable");
1729 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
1730 (*Legal->getReductionVars())[RdxPhi];
1732 // We need to generate a reduction vector from the incoming scalar.
1733 // To do so, we need to generate the 'identity' vector and overide
1734 // one of the elements with the incoming scalar reduction. We need
1735 // to do it in the vector-loop preheader.
1736 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
1738 // This is the vector-clone of the value that leaves the loop.
1739 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
1740 Type *VecTy = VectorExit[0]->getType();
1742 // Find the reduction identity variable. Zero for addition, or, xor,
1743 // one for multiplication, -1 for And.
1746 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax)
1747 // MinMax reduction have the start value as their identify.
1748 VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
1752 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
1753 VecTy->getScalarType());
1754 Identity = ConstantVector::getSplat(VF, Iden);
1756 // This vector is the Identity vector where the first element is the
1757 // incoming scalar reduction.
1758 VectorStart = Builder.CreateInsertElement(Identity,
1759 RdxDesc.StartValue, Zero);
1762 // Fix the vector-loop phi.
1763 // We created the induction variable so we know that the
1764 // preheader is the first entry.
1765 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
1767 // Reductions do not have to start at zero. They can start with
1768 // any loop invariant values.
1769 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
1770 BasicBlock *Latch = OrigLoop->getLoopLatch();
1771 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
1772 VectorParts &Val = getVectorValue(LoopVal);
1773 for (unsigned part = 0; part < UF; ++part) {
1774 // Make sure to add the reduction stat value only to the
1775 // first unroll part.
1776 Value *StartVal = (part == 0) ? VectorStart : Identity;
1777 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
1778 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
1781 // Before each round, move the insertion point right between
1782 // the PHIs and the values we are going to write.
1783 // This allows us to write both PHINodes and the extractelement
1785 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
1787 VectorParts RdxParts;
1788 for (unsigned part = 0; part < UF; ++part) {
1789 // This PHINode contains the vectorized reduction variable, or
1790 // the initial value vector, if we bypass the vector loop.
1791 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
1792 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
1793 Value *StartVal = (part == 0) ? VectorStart : Identity;
1794 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
1795 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
1796 NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
1797 RdxParts.push_back(NewPhi);
1800 // Reduce all of the unrolled parts into a single vector.
1801 Value *ReducedPartRdx = RdxParts[0];
1802 unsigned Op = getReductionBinOp(RdxDesc.Kind);
1803 for (unsigned part = 1; part < UF; ++part) {
1804 if (Op != Instruction::ICmp)
1805 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
1806 RdxParts[part], ReducedPartRdx,
1809 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
1810 ReducedPartRdx, RdxParts[part]);
1813 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1814 // and vector ops, reducing the set of values being computed by half each
1816 assert(isPowerOf2_32(VF) &&
1817 "Reduction emission only supported for pow2 vectors!");
1818 Value *TmpVec = ReducedPartRdx;
1819 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
1820 for (unsigned i = VF; i != 1; i >>= 1) {
1821 // Move the upper half of the vector to the lower half.
1822 for (unsigned j = 0; j != i/2; ++j)
1823 ShuffleMask[j] = Builder.getInt32(i/2 + j);
1825 // Fill the rest of the mask with undef.
1826 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
1827 UndefValue::get(Builder.getInt32Ty()));
1830 Builder.CreateShuffleVector(TmpVec,
1831 UndefValue::get(TmpVec->getType()),
1832 ConstantVector::get(ShuffleMask),
1835 if (Op != Instruction::ICmp)
1836 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
1839 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
1842 // The result is in the first element of the vector.
1843 Value *Scalar0 = Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1845 // Now, we need to fix the users of the reduction variable
1846 // inside and outside of the scalar remainder loop.
1847 // We know that the loop is in LCSSA form. We need to update the
1848 // PHI nodes in the exit blocks.
1849 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1850 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1851 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1852 if (!LCSSAPhi) continue;
1854 // All PHINodes need to have a single entry edge, or two if
1855 // we already fixed them.
1856 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
1858 // We found our reduction value exit-PHI. Update it with the
1859 // incoming bypass edge.
1860 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
1861 // Add an edge coming from the bypass.
1862 LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock);
1865 }// end of the LCSSA phi scan.
1867 // Fix the scalar loop reduction variable with the incoming reduction sum
1868 // from the vector body and from the backedge value.
1869 int IncomingEdgeBlockIdx =
1870 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
1871 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
1872 // Pick the other block.
1873 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
1874 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0);
1875 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
1876 }// end of for each redux variable.
1878 // The Loop exit block may have single value PHI nodes where the incoming
1879 // value is 'undef'. While vectorizing we only handled real values that
1880 // were defined inside the loop. Here we handle the 'undef case'.
1882 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
1883 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
1884 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
1885 if (!LCSSAPhi) continue;
1886 if (LCSSAPhi->getNumIncomingValues() == 1)
1887 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
1892 InnerLoopVectorizer::VectorParts
1893 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
1894 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
1897 VectorParts SrcMask = createBlockInMask(Src);
1899 // The terminator has to be a branch inst!
1900 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
1901 assert(BI && "Unexpected terminator found");
1903 if (BI->isConditional()) {
1904 VectorParts EdgeMask = getVectorValue(BI->getCondition());
1906 if (BI->getSuccessor(0) != Dst)
1907 for (unsigned part = 0; part < UF; ++part)
1908 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
1910 for (unsigned part = 0; part < UF; ++part)
1911 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
1918 InnerLoopVectorizer::VectorParts
1919 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
1920 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
1922 // Loop incoming mask is all-one.
1923 if (OrigLoop->getHeader() == BB) {
1924 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
1925 return getVectorValue(C);
1928 // This is the block mask. We OR all incoming edges, and with zero.
1929 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
1930 VectorParts BlockMask = getVectorValue(Zero);
1933 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
1934 VectorParts EM = createEdgeMask(*it, BB);
1935 for (unsigned part = 0; part < UF; ++part)
1936 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
1943 InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
1944 BasicBlock *BB, PhiVector *PV) {
1945 // For each instruction in the old loop.
1946 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
1947 VectorParts &Entry = WidenMap.get(it);
1948 switch (it->getOpcode()) {
1949 case Instruction::Br:
1950 // Nothing to do for PHIs and BR, since we already took care of the
1951 // loop control flow instructions.
1953 case Instruction::PHI:{
1954 PHINode* P = cast<PHINode>(it);
1955 // Handle reduction variables:
1956 if (Legal->getReductionVars()->count(P)) {
1957 for (unsigned part = 0; part < UF; ++part) {
1958 // This is phase one of vectorizing PHIs.
1959 Type *VecTy = VectorType::get(it->getType(), VF);
1960 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
1961 LoopVectorBody-> getFirstInsertionPt());
1967 // Check for PHI nodes that are lowered to vector selects.
1968 if (P->getParent() != OrigLoop->getHeader()) {
1969 // We know that all PHIs in non header blocks are converted into
1970 // selects, so we don't have to worry about the insertion order and we
1971 // can just use the builder.
1972 // At this point we generate the predication tree. There may be
1973 // duplications since this is a simple recursive scan, but future
1974 // optimizations will clean it up.
1976 unsigned NumIncoming = P->getNumIncomingValues();
1977 assert(NumIncoming > 1 && "Invalid PHI");
1979 // Generate a sequence of selects of the form:
1980 // SELECT(Mask3, In3,
1981 // SELECT(Mask2, In2,
1983 for (unsigned In = 0; In < NumIncoming; In++) {
1984 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
1986 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
1988 for (unsigned part = 0; part < UF; ++part) {
1989 // We don't need to 'select' the first PHI operand because it is
1990 // the default value if all of the other masks don't match.
1992 Entry[part] = In0[part];
1994 // Select between the current value and the previous incoming edge
1995 // based on the incoming mask.
1996 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
1997 Entry[part], "predphi");
2003 // This PHINode must be an induction variable.
2004 // Make sure that we know about it.
2005 assert(Legal->getInductionVars()->count(P) &&
2006 "Not an induction variable");
2008 LoopVectorizationLegality::InductionInfo II =
2009 Legal->getInductionVars()->lookup(P);
2012 case LoopVectorizationLegality::IK_NoInduction:
2013 llvm_unreachable("Unknown induction");
2014 case LoopVectorizationLegality::IK_IntInduction: {
2015 assert(P == OldInduction && "Unexpected PHI");
2016 Value *Broadcasted = getBroadcastInstrs(Induction);
2017 // After broadcasting the induction variable we need to make the
2018 // vector consecutive by adding 0, 1, 2 ...
2019 for (unsigned part = 0; part < UF; ++part)
2020 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2023 case LoopVectorizationLegality::IK_ReverseIntInduction:
2024 case LoopVectorizationLegality::IK_PtrInduction:
2025 case LoopVectorizationLegality::IK_ReversePtrInduction:
2026 // Handle reverse integer and pointer inductions.
2027 Value *StartIdx = 0;
2028 // If we have a single integer induction variable then use it.
2029 // Otherwise, start counting at zero.
2031 LoopVectorizationLegality::InductionInfo OldII =
2032 Legal->getInductionVars()->lookup(OldInduction);
2033 StartIdx = OldII.StartValue;
2035 StartIdx = ConstantInt::get(Induction->getType(), 0);
2037 // This is the normalized GEP that starts counting at zero.
2038 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2041 // Handle the reverse integer induction variable case.
2042 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2043 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2044 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2046 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2049 // This is a new value so do not hoist it out.
2050 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2051 // After broadcasting the induction variable we need to make the
2052 // vector consecutive by adding ... -3, -2, -1, 0.
2053 for (unsigned part = 0; part < UF; ++part)
2054 Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
2058 // Handle the pointer induction variable case.
2059 assert(P->getType()->isPointerTy() && "Unexpected type.");
2061 // Is this a reverse induction ptr or a consecutive induction ptr.
2062 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2065 // This is the vector of results. Notice that we don't generate
2066 // vector geps because scalar geps result in better code.
2067 for (unsigned part = 0; part < UF; ++part) {
2068 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2069 for (unsigned int i = 0; i < VF; ++i) {
2070 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2071 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2074 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2076 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2078 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2080 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2081 Builder.getInt32(i),
2084 Entry[part] = VecVal;
2091 case Instruction::Add:
2092 case Instruction::FAdd:
2093 case Instruction::Sub:
2094 case Instruction::FSub:
2095 case Instruction::Mul:
2096 case Instruction::FMul:
2097 case Instruction::UDiv:
2098 case Instruction::SDiv:
2099 case Instruction::FDiv:
2100 case Instruction::URem:
2101 case Instruction::SRem:
2102 case Instruction::FRem:
2103 case Instruction::Shl:
2104 case Instruction::LShr:
2105 case Instruction::AShr:
2106 case Instruction::And:
2107 case Instruction::Or:
2108 case Instruction::Xor: {
2109 // Just widen binops.
2110 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2111 VectorParts &A = getVectorValue(it->getOperand(0));
2112 VectorParts &B = getVectorValue(it->getOperand(1));
2114 // Use this vector value for all users of the original instruction.
2115 for (unsigned Part = 0; Part < UF; ++Part) {
2116 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2118 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2119 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2120 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2121 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2122 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2124 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2125 VecOp->setIsExact(BinOp->isExact());
2131 case Instruction::Select: {
2133 // If the selector is loop invariant we can create a select
2134 // instruction with a scalar condition. Otherwise, use vector-select.
2135 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2138 // The condition can be loop invariant but still defined inside the
2139 // loop. This means that we can't just use the original 'cond' value.
2140 // We have to take the 'vectorized' value and pick the first lane.
2141 // Instcombine will make this a no-op.
2142 VectorParts &Cond = getVectorValue(it->getOperand(0));
2143 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2144 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2145 Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
2146 Builder.getInt32(0));
2147 for (unsigned Part = 0; Part < UF; ++Part) {
2148 Entry[Part] = Builder.CreateSelect(
2149 InvariantCond ? ScalarCond : Cond[Part],
2156 case Instruction::ICmp:
2157 case Instruction::FCmp: {
2158 // Widen compares. Generate vector compares.
2159 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2160 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2161 VectorParts &A = getVectorValue(it->getOperand(0));
2162 VectorParts &B = getVectorValue(it->getOperand(1));
2163 for (unsigned Part = 0; Part < UF; ++Part) {
2166 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
2168 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
2174 case Instruction::Store:
2175 case Instruction::Load:
2176 vectorizeMemoryInstruction(it, Legal);
2178 case Instruction::ZExt:
2179 case Instruction::SExt:
2180 case Instruction::FPToUI:
2181 case Instruction::FPToSI:
2182 case Instruction::FPExt:
2183 case Instruction::PtrToInt:
2184 case Instruction::IntToPtr:
2185 case Instruction::SIToFP:
2186 case Instruction::UIToFP:
2187 case Instruction::Trunc:
2188 case Instruction::FPTrunc:
2189 case Instruction::BitCast: {
2190 CastInst *CI = dyn_cast<CastInst>(it);
2191 /// Optimize the special case where the source is the induction
2192 /// variable. Notice that we can only optimize the 'trunc' case
2193 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
2194 /// c. other casts depend on pointer size.
2195 if (CI->getOperand(0) == OldInduction &&
2196 it->getOpcode() == Instruction::Trunc) {
2197 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
2199 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
2200 for (unsigned Part = 0; Part < UF; ++Part)
2201 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
2204 /// Vectorize casts.
2205 Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
2207 VectorParts &A = getVectorValue(it->getOperand(0));
2208 for (unsigned Part = 0; Part < UF; ++Part)
2209 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
2213 case Instruction::Call: {
2214 // Ignore dbg intrinsics.
2215 if (isa<DbgInfoIntrinsic>(it))
2218 Module *M = BB->getParent()->getParent();
2219 CallInst *CI = cast<CallInst>(it);
2220 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2221 assert(ID && "Not an intrinsic call!");
2222 for (unsigned Part = 0; Part < UF; ++Part) {
2223 SmallVector<Value*, 4> Args;
2224 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
2225 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
2226 Args.push_back(Arg[Part]);
2228 Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
2229 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
2230 Entry[Part] = Builder.CreateCall(F, Args);
2236 // All other instructions are unsupported. Scalarize them.
2237 scalarizeInstruction(it);
2240 }// end of for_each instr.
2243 void InnerLoopVectorizer::updateAnalysis() {
2244 // Forget the original basic block.
2245 SE->forgetLoop(OrigLoop);
2247 // Update the dominator tree information.
2248 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
2249 "Entry does not dominate exit.");
2251 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2252 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
2253 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
2254 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
2255 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
2256 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
2257 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
2258 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
2260 DEBUG(DT->verifyAnalysis());
2263 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
2264 if (!EnableIfConversion)
2267 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
2268 std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
2270 // Collect the blocks that need predication.
2271 for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
2272 BasicBlock *BB = LoopBlocks[i];
2274 // We don't support switch statements inside loops.
2275 if (!isa<BranchInst>(BB->getTerminator()))
2278 // We must be able to predicate all blocks that need to be predicated.
2279 if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
2283 // We can if-convert this loop.
2287 bool LoopVectorizationLegality::canVectorize() {
2288 assert(TheLoop->getLoopPreheader() && "No preheader!!");
2290 // We can only vectorize innermost loops.
2291 if (TheLoop->getSubLoopsVector().size())
2294 // We must have a single backedge.
2295 if (TheLoop->getNumBackEdges() != 1)
2298 // We must have a single exiting block.
2299 if (!TheLoop->getExitingBlock())
2302 unsigned NumBlocks = TheLoop->getNumBlocks();
2304 // Check if we can if-convert non single-bb loops.
2305 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
2306 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
2310 // We need to have a loop header.
2311 BasicBlock *Latch = TheLoop->getLoopLatch();
2312 DEBUG(dbgs() << "LV: Found a loop: " <<
2313 TheLoop->getHeader()->getName() << "\n");
2315 // ScalarEvolution needs to be able to find the exit count.
2316 const SCEV *ExitCount = SE->getExitCount(TheLoop, Latch);
2317 if (ExitCount == SE->getCouldNotCompute()) {
2318 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
2322 // Do not loop-vectorize loops with a tiny trip count.
2323 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
2324 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
2325 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
2326 "This loop is not worth vectorizing.\n");
2330 // Check if we can vectorize the instructions and CFG in this loop.
2331 if (!canVectorizeInstrs()) {
2332 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
2336 // Go over each instruction and look at memory deps.
2337 if (!canVectorizeMemory()) {
2338 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
2342 // Collect all of the variables that remain uniform after vectorization.
2343 collectLoopUniforms();
2345 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
2346 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
2349 // Okay! We can vectorize. At this point we don't have any other mem analysis
2350 // which may limit our maximum vectorization factor, so just return true with
2355 bool LoopVectorizationLegality::canVectorizeInstrs() {
2356 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
2357 BasicBlock *Header = TheLoop->getHeader();
2359 // If we marked the scalar loop as "already vectorized" then no need
2360 // to vectorize it again.
2361 if (Header->getTerminator()->getMetadata(AlreadyVectorizedMDName)) {
2362 DEBUG(dbgs() << "LV: This loop was vectorized before\n");
2366 // For each block in the loop.
2367 for (Loop::block_iterator bb = TheLoop->block_begin(),
2368 be = TheLoop->block_end(); bb != be; ++bb) {
2370 // Scan the instructions in the block and look for hazards.
2371 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2374 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
2375 // Check that this PHI type is allowed.
2376 if (!Phi->getType()->isIntegerTy() &&
2377 !Phi->getType()->isFloatingPointTy() &&
2378 !Phi->getType()->isPointerTy()) {
2379 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
2383 // If this PHINode is not in the header block, then we know that we
2384 // can convert it to select during if-conversion. No need to check if
2385 // the PHIs in this block are induction or reduction variables.
2389 // We only allow if-converted PHIs with more than two incoming values.
2390 if (Phi->getNumIncomingValues() != 2) {
2391 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
2395 // This is the value coming from the preheader.
2396 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
2397 // Check if this is an induction variable.
2398 InductionKind IK = isInductionVariable(Phi);
2400 if (IK_NoInduction != IK) {
2401 // Int inductions are special because we only allow one IV.
2402 if (IK == IK_IntInduction) {
2404 DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
2410 DEBUG(dbgs() << "LV: Found an induction variable.\n");
2411 Inductions[Phi] = InductionInfo(StartValue, IK);
2415 if (AddReductionVar(Phi, RK_IntegerAdd)) {
2416 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
2419 if (AddReductionVar(Phi, RK_IntegerMult)) {
2420 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
2423 if (AddReductionVar(Phi, RK_IntegerOr)) {
2424 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
2427 if (AddReductionVar(Phi, RK_IntegerAnd)) {
2428 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
2431 if (AddReductionVar(Phi, RK_IntegerXor)) {
2432 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
2435 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
2436 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
2439 if (AddReductionVar(Phi, RK_FloatMult)) {
2440 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
2443 if (AddReductionVar(Phi, RK_FloatAdd)) {
2444 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
2448 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
2450 }// end of PHI handling
2452 // We still don't handle functions. However, we can ignore dbg intrinsic
2453 // calls and we do handle certain intrinsic and libm functions.
2454 CallInst *CI = dyn_cast<CallInst>(it);
2455 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
2456 DEBUG(dbgs() << "LV: Found a call site.\n");
2460 // Check that the instruction return type is vectorizable.
2461 if (!VectorType::isValidElementType(it->getType()) &&
2462 !it->getType()->isVoidTy()) {
2463 DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
2467 // Check that the stored type is vectorizable.
2468 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
2469 Type *T = ST->getValueOperand()->getType();
2470 if (!VectorType::isValidElementType(T))
2474 // Reduction instructions are allowed to have exit users.
2475 // All other instructions must not have external users.
2476 if (!AllowedExit.count(it))
2477 //Check that all of the users of the loop are inside the BB.
2478 for (Value::use_iterator I = it->use_begin(), E = it->use_end();
2480 Instruction *U = cast<Instruction>(*I);
2481 // This user may be a reduction exit value.
2482 if (!TheLoop->contains(U)) {
2483 DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
2492 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
2493 assert(getInductionVars()->size() && "No induction variables");
2499 void LoopVectorizationLegality::collectLoopUniforms() {
2500 // We now know that the loop is vectorizable!
2501 // Collect variables that will remain uniform after vectorization.
2502 std::vector<Value*> Worklist;
2503 BasicBlock *Latch = TheLoop->getLoopLatch();
2505 // Start with the conditional branch and walk up the block.
2506 Worklist.push_back(Latch->getTerminator()->getOperand(0));
2508 while (Worklist.size()) {
2509 Instruction *I = dyn_cast<Instruction>(Worklist.back());
2510 Worklist.pop_back();
2512 // Look at instructions inside this loop.
2513 // Stop when reaching PHI nodes.
2514 // TODO: we need to follow values all over the loop, not only in this block.
2515 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
2518 // This is a known uniform.
2521 // Insert all operands.
2522 for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
2523 Worklist.push_back(I->getOperand(i));
2528 AliasAnalysis::Location
2529 LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
2530 if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
2531 return AA->getLocation(Store);
2532 else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
2533 return AA->getLocation(Load);
2535 llvm_unreachable("Should be either load or store instruction");
2539 LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
2542 AliasMultiMap& WriteObjects,
2543 unsigned MaxByteWidth) {
2545 AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
2547 std::vector<Instruction*>::iterator
2548 it = WriteObjects[Object].begin(),
2549 end = WriteObjects[Object].end();
2551 for (; it != end; ++it) {
2552 Instruction* I = *it;
2556 AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
2557 if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
2558 ThatLoc.getWithNewSize(MaxByteWidth)))
2564 bool LoopVectorizationLegality::canVectorizeMemory() {
2566 typedef SmallVector<Value*, 16> ValueVector;
2567 typedef SmallPtrSet<Value*, 16> ValueSet;
2568 // Holds the Load and Store *instructions*.
2571 PtrRtCheck.Pointers.clear();
2572 PtrRtCheck.Need = false;
2574 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2577 for (Loop::block_iterator bb = TheLoop->block_begin(),
2578 be = TheLoop->block_end(); bb != be; ++bb) {
2580 // Scan the BB and collect legal loads and stores.
2581 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
2584 // If this is a load, save it. If this instruction can read from memory
2585 // but is not a load, then we quit. Notice that we don't handle function
2586 // calls that read or write.
2587 if (it->mayReadFromMemory()) {
2588 LoadInst *Ld = dyn_cast<LoadInst>(it);
2589 if (!Ld) return false;
2590 if (!Ld->isSimple() && !IsAnnotatedParallel) {
2591 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
2594 Loads.push_back(Ld);
2598 // Save 'store' instructions. Abort if other instructions write to memory.
2599 if (it->mayWriteToMemory()) {
2600 StoreInst *St = dyn_cast<StoreInst>(it);
2601 if (!St) return false;
2602 if (!St->isSimple() && !IsAnnotatedParallel) {
2603 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
2606 Stores.push_back(St);
2611 // Now we have two lists that hold the loads and the stores.
2612 // Next, we find the pointers that they use.
2614 // Check if we see any stores. If there are no stores, then we don't
2615 // care if the pointers are *restrict*.
2616 if (!Stores.size()) {
2617 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
2621 // Holds the read and read-write *pointers* that we find. These maps hold
2622 // unique values for pointers (so no need for multi-map).
2624 AliasMap ReadWrites;
2626 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
2627 // multiple times on the same object. If the ptr is accessed twice, once
2628 // for read and once for write, it will only appear once (on the write
2629 // list). This is okay, since we are going to check for conflicts between
2630 // writes and between reads and writes, but not between reads and reads.
2633 ValueVector::iterator I, IE;
2634 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
2635 StoreInst *ST = cast<StoreInst>(*I);
2636 Value* Ptr = ST->getPointerOperand();
2638 if (isUniform(Ptr)) {
2639 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
2643 // If we did *not* see this pointer before, insert it to
2644 // the read-write list. At this phase it is only a 'write' list.
2645 if (Seen.insert(Ptr))
2646 ReadWrites.insert(std::make_pair(Ptr, ST));
2649 if (IsAnnotatedParallel) {
2651 << "LV: A loop annotated parallel, ignore memory dependency "
2656 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
2657 LoadInst *LD = cast<LoadInst>(*I);
2658 Value* Ptr = LD->getPointerOperand();
2659 // If we did *not* see this pointer before, insert it to the
2660 // read list. If we *did* see it before, then it is already in
2661 // the read-write list. This allows us to vectorize expressions
2662 // such as A[i] += x; Because the address of A[i] is a read-write
2663 // pointer. This only works if the index of A[i] is consecutive.
2664 // If the address of i is unknown (for example A[B[i]]) then we may
2665 // read a few words, modify, and write a few words, and some of the
2666 // words may be written to the same address.
2667 if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
2668 Reads.insert(std::make_pair(Ptr, LD));
2671 // If we write (or read-write) to a single destination and there are no
2672 // other reads in this loop then is it safe to vectorize.
2673 if (ReadWrites.size() == 1 && Reads.size() == 0) {
2674 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
2678 unsigned NumReadPtrs = 0;
2679 unsigned NumWritePtrs = 0;
2681 // Find pointers with computable bounds. We are going to use this information
2682 // to place a runtime bound check.
2683 bool CanDoRT = true;
2684 AliasMap::iterator MI, ME;
2685 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2686 Value *V = (*MI).first;
2687 if (hasComputableBounds(V)) {
2688 PtrRtCheck.insert(SE, TheLoop, V, true);
2690 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2696 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2697 Value *V = (*MI).first;
2698 if (hasComputableBounds(V)) {
2699 PtrRtCheck.insert(SE, TheLoop, V, false);
2701 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
2708 // Check that we did not collect too many pointers or found a
2709 // unsizeable pointer.
2710 unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
2711 DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
2712 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
2718 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
2721 bool NeedRTCheck = false;
2723 // Biggest vectorized access possible, vector width * unroll factor.
2724 // TODO: We're being very pessimistic here, find a way to know the
2725 // real access width before getting here.
2726 unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
2727 TTI->getMaximumUnrollFactor();
2728 // Now that the pointers are in two lists (Reads and ReadWrites), we
2729 // can check that there are no conflicts between each of the writes and
2730 // between the writes to the reads.
2731 // Note that WriteObjects duplicates the stores (indexed now by underlying
2732 // objects) to avoid pointing to elements inside ReadWrites.
2733 // TODO: Maybe create a new type where they can interact without duplication.
2734 AliasMultiMap WriteObjects;
2735 ValueVector TempObjects;
2737 // Check that the read-writes do not conflict with other read-write
2739 bool AllWritesIdentified = true;
2740 for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
2741 Value *Val = (*MI).first;
2742 Instruction *Inst = (*MI).second;
2744 GetUnderlyingObjects(Val, TempObjects, DL);
2745 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2747 if (!isIdentifiedObject(*UI)) {
2748 DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
2750 AllWritesIdentified = false;
2753 // Never seen it before, can't alias.
2754 if (WriteObjects[*UI].empty()) {
2755 DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
2756 WriteObjects[*UI].push_back(Inst);
2759 // Direct alias found.
2760 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2761 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2765 DEBUG(dbgs() << "LV: Found a conflicting global value:"
2767 DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
2768 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2770 // If global alias, make sure they do alias.
2771 if (hasPossibleGlobalWriteReorder(*UI,
2775 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2780 // Didn't alias, insert into map for further reference.
2781 WriteObjects[*UI].push_back(Inst);
2783 TempObjects.clear();
2786 /// Check that the reads don't conflict with the read-writes.
2787 for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
2788 Value *Val = (*MI).first;
2789 GetUnderlyingObjects(Val, TempObjects, DL);
2790 for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
2792 // If all of the writes are identified then we don't care if the read
2793 // pointer is identified or not.
2794 if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
2795 DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
2799 // Never seen it before, can't alias.
2800 if (WriteObjects[*UI].empty())
2802 // Direct alias found.
2803 if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
2804 DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
2808 DEBUG(dbgs() << "LV: Found a global value: "
2810 Instruction *Inst = (*MI).second;
2811 DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
2812 DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
2814 // If global alias, make sure they do alias.
2815 if (hasPossibleGlobalWriteReorder(*UI,
2819 DEBUG(dbgs() << "LV: Found a possible read-write reorder:"
2824 TempObjects.clear();
2827 PtrRtCheck.Need = NeedRTCheck;
2828 if (NeedRTCheck && !CanDoRT) {
2829 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
2830 "the array bounds.\n");
2835 DEBUG(dbgs() << "LV: We "<< (NeedRTCheck ? "" : "don't") <<
2836 " need a runtime memory check.\n");
2840 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
2841 ReductionKind Kind) {
2842 if (Phi->getNumIncomingValues() != 2)
2845 // Reduction variables are only found in the loop header block.
2846 if (Phi->getParent() != TheLoop->getHeader())
2849 // Obtain the reduction start value from the value that comes from the loop
2851 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
2853 // ExitInstruction is the single value which is used outside the loop.
2854 // We only allow for a single reduction value to be used outside the loop.
2855 // This includes users of the reduction, variables (which form a cycle
2856 // which ends in the phi node).
2857 Instruction *ExitInstruction = 0;
2858 // Indicates that we found a binary operation in our scan.
2859 bool FoundBinOp = false;
2861 // Iter is our iterator. We start with the PHI node and scan for all of the
2862 // users of this instruction. All users must be instructions that can be
2863 // used as reduction variables (such as ADD). We may have a single
2864 // out-of-block user. The cycle must end with the original PHI.
2865 Instruction *Iter = Phi;
2867 // To recognize min/max patterns formed by a icmp select sequence, we store
2868 // the number of instruction we saw from the recognized min/max pattern,
2869 // such that we don't stop when we see the phi has two uses (one by the select
2870 // and one by the icmp) and to make sure we only see exactly the two
2872 unsigned NumICmpSelectPatternInst = 0;
2873 ReductionInstDesc ReduxDesc(false, 0);
2875 // Avoid cycles in the chain.
2876 SmallPtrSet<Instruction *, 8> VisitedInsts;
2877 while (VisitedInsts.insert(Iter)) {
2878 // If the instruction has no users then this is a broken
2879 // chain and can't be a reduction variable.
2880 if (Iter->use_empty())
2883 // Did we find a user inside this loop already ?
2884 bool FoundInBlockUser = false;
2885 // Did we reach the initial PHI node already ?
2886 bool FoundStartPHI = false;
2888 // Is this a bin op ?
2889 FoundBinOp |= !isa<PHINode>(Iter);
2891 // For each of the *users* of iter.
2892 for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
2894 Instruction *U = cast<Instruction>(*it);
2895 // We already know that the PHI is a user.
2897 FoundStartPHI = true;
2901 // Check if we found the exit user.
2902 BasicBlock *Parent = U->getParent();
2903 if (!TheLoop->contains(Parent)) {
2904 // Exit if you find multiple outside users.
2905 if (ExitInstruction != 0)
2907 ExitInstruction = Iter;
2910 // We allow in-loop PHINodes which are not the original reduction PHI
2911 // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
2912 // structure) then don't skip this PHI.
2913 if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
2914 U->getParent() != TheLoop->getHeader() &&
2915 TheLoop->contains(U) &&
2916 Iter->hasNUsesOrMore(2))
2919 // We can't have multiple inside users except for a combination of
2920 // icmp/select both using the phi.
2921 if (FoundInBlockUser && !NumICmpSelectPatternInst)
2923 FoundInBlockUser = true;
2925 // Any reduction instr must be of one of the allowed kinds.
2926 ReduxDesc = isReductionInstr(U, Kind, ReduxDesc);
2927 if (!ReduxDesc.IsReduction)
2930 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(U) ||
2931 isa<SelectInst>(U)))
2932 ++NumICmpSelectPatternInst;
2934 // Reductions of instructions such as Div, and Sub is only
2935 // possible if the LHS is the reduction variable.
2936 if (!U->isCommutative() && !isa<PHINode>(U) && !isa<SelectInst>(U) &&
2937 !isa<ICmpInst>(U) && U->getOperand(0) != Iter)
2940 Iter = ReduxDesc.PatternLastInst;
2943 // This means we have seen one but not the other instruction of the
2944 // pattern or more than just a select and cmp.
2945 if (Kind == RK_IntegerMinMax && NumICmpSelectPatternInst != 2)
2948 // We found a reduction var if we have reached the original
2949 // phi node and we only have a single instruction with out-of-loop
2951 if (FoundStartPHI) {
2952 // This instruction is allowed to have out-of-loop users.
2953 AllowedExit.insert(ExitInstruction);
2955 // Save the description of this reduction variable.
2956 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
2957 ReduxDesc.MinMaxKind);
2958 Reductions[Phi] = RD;
2959 // We've ended the cycle. This is a reduction variable if we have an
2960 // outside user and it has a binary op.
2961 return FoundBinOp && ExitInstruction;
2968 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
2969 /// pattern corresponding to a min(X, Y) or max(X, Y).
2970 LoopVectorizationLegality::ReductionInstDesc
2971 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I, ReductionInstDesc &Prev) {
2973 assert((isa<ICmpInst>(I) || isa<SelectInst>(I)) &&
2974 "Expect a select instruction");
2976 SelectInst *Select = 0;
2978 // We must handle the select(cmp()) as a single instruction. Advance to the
2980 if ((Cmp = dyn_cast<ICmpInst>(I))) {
2981 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
2982 return ReductionInstDesc(false, I);
2983 return ReductionInstDesc(Select, Prev.MinMaxKind);
2986 // Only handle single use cases for now.
2987 if (!(Select = dyn_cast<SelectInst>(I)))
2988 return ReductionInstDesc(false, I);
2989 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))))
2990 return ReductionInstDesc(false, I);
2991 if (!Cmp->hasOneUse())
2992 return ReductionInstDesc(false, I);
2994 Value *CmpLeft = Cmp->getOperand(0);
2995 Value *CmpRight = Cmp->getOperand(1);
2997 // Look for a min/max pattern.
2998 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
2999 return ReductionInstDesc(Select, MRK_UIntMin);
3000 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3001 return ReductionInstDesc(Select, MRK_UIntMax);
3002 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3003 return ReductionInstDesc(Select, MRK_SIntMax);
3004 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
3005 return ReductionInstDesc(Select, MRK_SIntMin);
3007 return ReductionInstDesc(false, I);
3010 LoopVectorizationLegality::ReductionInstDesc
3011 LoopVectorizationLegality::isReductionInstr(Instruction *I,
3013 ReductionInstDesc &Prev) {
3014 bool FP = I->getType()->isFloatingPointTy();
3015 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
3016 switch (I->getOpcode()) {
3018 return ReductionInstDesc(false, I);
3019 case Instruction::PHI:
3020 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd))
3021 return ReductionInstDesc(false, I);
3022 return ReductionInstDesc(I, Prev.MinMaxKind);
3023 case Instruction::Sub:
3024 case Instruction::Add:
3025 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
3026 case Instruction::Mul:
3027 return ReductionInstDesc(Kind == RK_IntegerMult, I);
3028 case Instruction::And:
3029 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
3030 case Instruction::Or:
3031 return ReductionInstDesc(Kind == RK_IntegerOr, I);
3032 case Instruction::Xor:
3033 return ReductionInstDesc(Kind == RK_IntegerXor, I);
3034 case Instruction::FMul:
3035 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
3036 case Instruction::FAdd:
3037 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
3038 case Instruction::ICmp:
3039 case Instruction::Select:
3040 if (Kind != RK_IntegerMinMax)
3041 return ReductionInstDesc(false, I);
3042 return isMinMaxSelectCmpPattern(I, Prev);
3046 LoopVectorizationLegality::InductionKind
3047 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
3048 Type *PhiTy = Phi->getType();
3049 // We only handle integer and pointer inductions variables.
3050 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
3051 return IK_NoInduction;
3053 // Check that the PHI is consecutive.
3054 const SCEV *PhiScev = SE->getSCEV(Phi);
3055 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3057 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
3058 return IK_NoInduction;
3060 const SCEV *Step = AR->getStepRecurrence(*SE);
3062 // Integer inductions need to have a stride of one.
3063 if (PhiTy->isIntegerTy()) {
3065 return IK_IntInduction;
3066 if (Step->isAllOnesValue())
3067 return IK_ReverseIntInduction;
3068 return IK_NoInduction;
3071 // Calculate the pointer stride and check if it is consecutive.
3072 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
3074 return IK_NoInduction;
3076 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
3077 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
3078 if (C->getValue()->equalsInt(Size))
3079 return IK_PtrInduction;
3080 else if (C->getValue()->equalsInt(0 - Size))
3081 return IK_ReversePtrInduction;
3083 return IK_NoInduction;
3086 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
3087 Value *In0 = const_cast<Value*>(V);
3088 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
3092 return Inductions.count(PN);
3095 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
3096 assert(TheLoop->contains(BB) && "Unknown block used");
3098 // Blocks that do not dominate the latch need predication.
3099 BasicBlock* Latch = TheLoop->getLoopLatch();
3100 return !DT->dominates(BB, Latch);
3103 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
3104 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3105 // We don't predicate loads/stores at the moment.
3106 if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
3109 // The instructions below can trap.
3110 switch (it->getOpcode()) {
3112 case Instruction::UDiv:
3113 case Instruction::SDiv:
3114 case Instruction::URem:
3115 case Instruction::SRem:
3123 bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) {
3124 const SCEV *PhiScev = SE->getSCEV(Ptr);
3125 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
3129 return AR->isAffine();
3132 LoopVectorizationCostModel::VectorizationFactor
3133 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
3135 // Width 1 means no vectorize
3136 VectorizationFactor Factor = { 1U, 0U };
3137 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
3138 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
3142 // Find the trip count.
3143 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
3144 DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
3146 unsigned WidestType = getWidestType();
3147 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
3148 unsigned MaxVectorSize = WidestRegister / WidestType;
3149 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
3150 DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
3152 if (MaxVectorSize == 0) {
3153 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
3157 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
3158 " into one vector!");
3160 unsigned VF = MaxVectorSize;
3162 // If we optimize the program for size, avoid creating the tail loop.
3164 // If we are unable to calculate the trip count then don't try to vectorize.
3166 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3170 // Find the maximum SIMD width that can fit within the trip count.
3171 VF = TC % MaxVectorSize;
3176 // If the trip count that we found modulo the vectorization factor is not
3177 // zero then we require a tail.
3179 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
3185 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
3186 DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
3188 Factor.Width = UserVF;
3192 float Cost = expectedCost(1);
3194 DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n");
3195 for (unsigned i=2; i <= VF; i*=2) {
3196 // Notice that the vector loop needs to be executed less times, so
3197 // we need to divide the cost of the vector loops by the width of
3198 // the vector elements.
3199 float VectorCost = expectedCost(i) / (float)i;
3200 DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " <<
3201 (int)VectorCost << ".\n");
3202 if (VectorCost < Cost) {
3208 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
3209 Factor.Width = Width;
3210 Factor.Cost = Width * Cost;
3214 unsigned LoopVectorizationCostModel::getWidestType() {
3215 unsigned MaxWidth = 8;
3218 for (Loop::block_iterator bb = TheLoop->block_begin(),
3219 be = TheLoop->block_end(); bb != be; ++bb) {
3220 BasicBlock *BB = *bb;
3222 // For each instruction in the loop.
3223 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3224 Type *T = it->getType();
3226 // Only examine Loads, Stores and PHINodes.
3227 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
3230 // Examine PHI nodes that are reduction variables.
3231 if (PHINode *PN = dyn_cast<PHINode>(it))
3232 if (!Legal->getReductionVars()->count(PN))
3235 // Examine the stored values.
3236 if (StoreInst *ST = dyn_cast<StoreInst>(it))
3237 T = ST->getValueOperand()->getType();
3239 // Ignore loaded pointer types and stored pointer types that are not
3240 // consecutive. However, we do want to take consecutive stores/loads of
3241 // pointer vectors into account.
3242 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
3245 MaxWidth = std::max(MaxWidth,
3246 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
3254 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
3257 unsigned LoopCost) {
3259 // -- The unroll heuristics --
3260 // We unroll the loop in order to expose ILP and reduce the loop overhead.
3261 // There are many micro-architectural considerations that we can't predict
3262 // at this level. For example frontend pressure (on decode or fetch) due to
3263 // code size, or the number and capabilities of the execution ports.
3265 // We use the following heuristics to select the unroll factor:
3266 // 1. If the code has reductions the we unroll in order to break the cross
3267 // iteration dependency.
3268 // 2. If the loop is really small then we unroll in order to reduce the loop
3270 // 3. We don't unroll if we think that we will spill registers to memory due
3271 // to the increased register pressure.
3273 // Use the user preference, unless 'auto' is selected.
3277 // When we optimize for size we don't unroll.
3281 // Do not unroll loops with a relatively small trip count.
3282 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
3283 TheLoop->getLoopLatch());
3284 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
3287 unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
3288 DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
3289 " vector registers\n");
3291 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
3292 // We divide by these constants so assume that we have at least one
3293 // instruction that uses at least one register.
3294 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
3295 R.NumInstructions = std::max(R.NumInstructions, 1U);
3297 // We calculate the unroll factor using the following formula.
3298 // Subtract the number of loop invariants from the number of available
3299 // registers. These registers are used by all of the unrolled instances.
3300 // Next, divide the remaining registers by the number of registers that is
3301 // required by the loop, in order to estimate how many parallel instances
3302 // fit without causing spills.
3303 unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
3305 // Clamp the unroll factor ranges to reasonable factors.
3306 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
3308 // If we did not calculate the cost for VF (because the user selected the VF)
3309 // then we calculate the cost of VF here.
3311 LoopCost = expectedCost(VF);
3313 // Clamp the calculated UF to be between the 1 and the max unroll factor
3314 // that the target allows.
3315 if (UF > MaxUnrollSize)
3320 if (Legal->getReductionVars()->size()) {
3321 DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
3325 // We want to unroll tiny loops in order to reduce the loop overhead.
3326 // We assume that the cost overhead is 1 and we use the cost model
3327 // to estimate the cost of the loop and unroll until the cost of the
3328 // loop overhead is about 5% of the cost of the loop.
3329 DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
3330 if (LoopCost < 20) {
3331 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
3332 unsigned NewUF = 20/LoopCost + 1;
3333 return std::min(NewUF, UF);
3336 DEBUG(dbgs() << "LV: Not Unrolling. \n");
3340 LoopVectorizationCostModel::RegisterUsage
3341 LoopVectorizationCostModel::calculateRegisterUsage() {
3342 // This function calculates the register usage by measuring the highest number
3343 // of values that are alive at a single location. Obviously, this is a very
3344 // rough estimation. We scan the loop in a topological order in order and
3345 // assign a number to each instruction. We use RPO to ensure that defs are
3346 // met before their users. We assume that each instruction that has in-loop
3347 // users starts an interval. We record every time that an in-loop value is
3348 // used, so we have a list of the first and last occurrences of each
3349 // instruction. Next, we transpose this data structure into a multi map that
3350 // holds the list of intervals that *end* at a specific location. This multi
3351 // map allows us to perform a linear search. We scan the instructions linearly
3352 // and record each time that a new interval starts, by placing it in a set.
3353 // If we find this value in the multi-map then we remove it from the set.
3354 // The max register usage is the maximum size of the set.
3355 // We also search for instructions that are defined outside the loop, but are
3356 // used inside the loop. We need this number separately from the max-interval
3357 // usage number because when we unroll, loop-invariant values do not take
3359 LoopBlocksDFS DFS(TheLoop);
3363 R.NumInstructions = 0;
3365 // Each 'key' in the map opens a new interval. The values
3366 // of the map are the index of the 'last seen' usage of the
3367 // instruction that is the key.
3368 typedef DenseMap<Instruction*, unsigned> IntervalMap;
3369 // Maps instruction to its index.
3370 DenseMap<unsigned, Instruction*> IdxToInstr;
3371 // Marks the end of each interval.
3372 IntervalMap EndPoint;
3373 // Saves the list of instruction indices that are used in the loop.
3374 SmallSet<Instruction*, 8> Ends;
3375 // Saves the list of values that are used in the loop but are
3376 // defined outside the loop, such as arguments and constants.
3377 SmallPtrSet<Value*, 8> LoopInvariants;
3380 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3381 be = DFS.endRPO(); bb != be; ++bb) {
3382 R.NumInstructions += (*bb)->size();
3383 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3385 Instruction *I = it;
3386 IdxToInstr[Index++] = I;
3388 // Save the end location of each USE.
3389 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
3390 Value *U = I->getOperand(i);
3391 Instruction *Instr = dyn_cast<Instruction>(U);
3393 // Ignore non-instruction values such as arguments, constants, etc.
3394 if (!Instr) continue;
3396 // If this instruction is outside the loop then record it and continue.
3397 if (!TheLoop->contains(Instr)) {
3398 LoopInvariants.insert(Instr);
3402 // Overwrite previous end points.
3403 EndPoint[Instr] = Index;
3409 // Saves the list of intervals that end with the index in 'key'.
3410 typedef SmallVector<Instruction*, 2> InstrList;
3411 DenseMap<unsigned, InstrList> TransposeEnds;
3413 // Transpose the EndPoints to a list of values that end at each index.
3414 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
3416 TransposeEnds[it->second].push_back(it->first);
3418 SmallSet<Instruction*, 8> OpenIntervals;
3419 unsigned MaxUsage = 0;
3422 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
3423 for (unsigned int i = 0; i < Index; ++i) {
3424 Instruction *I = IdxToInstr[i];
3425 // Ignore instructions that are never used within the loop.
3426 if (!Ends.count(I)) continue;
3428 // Remove all of the instructions that end at this location.
3429 InstrList &List = TransposeEnds[i];
3430 for (unsigned int j=0, e = List.size(); j < e; ++j)
3431 OpenIntervals.erase(List[j]);
3433 // Count the number of live interals.
3434 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
3436 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
3437 OpenIntervals.size() <<"\n");
3439 // Add the current instruction to the list of open intervals.
3440 OpenIntervals.insert(I);
3443 unsigned Invariant = LoopInvariants.size();
3444 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << " \n");
3445 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << " \n");
3446 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << " \n");
3448 R.LoopInvariantRegs = Invariant;
3449 R.MaxLocalUsers = MaxUsage;
3453 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
3457 for (Loop::block_iterator bb = TheLoop->block_begin(),
3458 be = TheLoop->block_end(); bb != be; ++bb) {
3459 unsigned BlockCost = 0;
3460 BasicBlock *BB = *bb;
3462 // For each instruction in the old loop.
3463 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3464 // Skip dbg intrinsics.
3465 if (isa<DbgInfoIntrinsic>(it))
3468 unsigned C = getInstructionCost(it, VF);
3470 DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
3471 VF << " For instruction: "<< *it << "\n");
3474 // We assume that if-converted blocks have a 50% chance of being executed.
3475 // When the code is scalar then some of the blocks are avoided due to CF.
3476 // When the code is vectorized we execute all code paths.
3477 if (Legal->blockNeedsPredication(*bb) && VF == 1)
3487 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
3488 // If we know that this instruction will remain uniform, check the cost of
3489 // the scalar version.
3490 if (Legal->isUniformAfterVectorization(I))
3493 Type *RetTy = I->getType();
3494 Type *VectorTy = ToVectorTy(RetTy, VF);
3496 // TODO: We need to estimate the cost of intrinsic calls.
3497 switch (I->getOpcode()) {
3498 case Instruction::GetElementPtr:
3499 // We mark this instruction as zero-cost because the cost of GEPs in
3500 // vectorized code depends on whether the corresponding memory instruction
3501 // is scalarized or not. Therefore, we handle GEPs with the memory
3502 // instruction cost.
3504 case Instruction::Br: {
3505 return TTI.getCFInstrCost(I->getOpcode());
3507 case Instruction::PHI:
3508 //TODO: IF-converted IFs become selects.
3510 case Instruction::Add:
3511 case Instruction::FAdd:
3512 case Instruction::Sub:
3513 case Instruction::FSub:
3514 case Instruction::Mul:
3515 case Instruction::FMul:
3516 case Instruction::UDiv:
3517 case Instruction::SDiv:
3518 case Instruction::FDiv:
3519 case Instruction::URem:
3520 case Instruction::SRem:
3521 case Instruction::FRem:
3522 case Instruction::Shl:
3523 case Instruction::LShr:
3524 case Instruction::AShr:
3525 case Instruction::And:
3526 case Instruction::Or:
3527 case Instruction::Xor: {
3528 // Certain instructions can be cheaper to vectorize if they have a constant
3529 // second vector operand. One example of this are shifts on x86.
3530 TargetTransformInfo::OperandValueKind Op1VK =
3531 TargetTransformInfo::OK_AnyValue;
3532 TargetTransformInfo::OperandValueKind Op2VK =
3533 TargetTransformInfo::OK_AnyValue;
3535 if (isa<ConstantInt>(I->getOperand(1)))
3536 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
3538 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
3540 case Instruction::Select: {
3541 SelectInst *SI = cast<SelectInst>(I);
3542 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
3543 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
3544 Type *CondTy = SI->getCondition()->getType();
3546 CondTy = VectorType::get(CondTy, VF);
3548 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
3550 case Instruction::ICmp:
3551 case Instruction::FCmp: {
3552 Type *ValTy = I->getOperand(0)->getType();
3553 VectorTy = ToVectorTy(ValTy, VF);
3554 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
3556 case Instruction::Store:
3557 case Instruction::Load: {
3558 StoreInst *SI = dyn_cast<StoreInst>(I);
3559 LoadInst *LI = dyn_cast<LoadInst>(I);
3560 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
3562 VectorTy = ToVectorTy(ValTy, VF);
3564 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
3565 unsigned AS = SI ? SI->getPointerAddressSpace() :
3566 LI->getPointerAddressSpace();
3567 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
3568 // We add the cost of address computation here instead of with the gep
3569 // instruction because only here we know whether the operation is
3572 return TTI.getAddressComputationCost(VectorTy) +
3573 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3575 // Scalarized loads/stores.
3576 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
3577 bool Reverse = ConsecutiveStride < 0;
3578 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
3579 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
3580 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
3582 // The cost of extracting from the value vector and pointer vector.
3583 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
3584 for (unsigned i = 0; i < VF; ++i) {
3585 // The cost of extracting the pointer operand.
3586 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
3587 // In case of STORE, the cost of ExtractElement from the vector.
3588 // In case of LOAD, the cost of InsertElement into the returned
3590 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
3591 Instruction::InsertElement,
3595 // The cost of the scalar loads/stores.
3596 Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
3597 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
3602 // Wide load/stores.
3603 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
3604 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
3607 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
3611 case Instruction::ZExt:
3612 case Instruction::SExt:
3613 case Instruction::FPToUI:
3614 case Instruction::FPToSI:
3615 case Instruction::FPExt:
3616 case Instruction::PtrToInt:
3617 case Instruction::IntToPtr:
3618 case Instruction::SIToFP:
3619 case Instruction::UIToFP:
3620 case Instruction::Trunc:
3621 case Instruction::FPTrunc:
3622 case Instruction::BitCast: {
3623 // We optimize the truncation of induction variable.
3624 // The cost of these is the same as the scalar operation.
3625 if (I->getOpcode() == Instruction::Trunc &&
3626 Legal->isInductionVariable(I->getOperand(0)))
3627 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
3628 I->getOperand(0)->getType());
3630 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
3631 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
3633 case Instruction::Call: {
3634 CallInst *CI = cast<CallInst>(I);
3635 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3636 assert(ID && "Not an intrinsic call!");
3637 Type *RetTy = ToVectorTy(CI->getType(), VF);
3638 SmallVector<Type*, 4> Tys;
3639 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3640 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3641 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3644 // We are scalarizing the instruction. Return the cost of the scalar
3645 // instruction, plus the cost of insert and extract into vector
3646 // elements, times the vector width.
3649 if (!RetTy->isVoidTy() && VF != 1) {
3650 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
3652 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
3655 // The cost of inserting the results plus extracting each one of the
3657 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
3660 // The cost of executing VF copies of the scalar instruction. This opcode
3661 // is unknown. Assume that it is the same as 'mul'.
3662 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
3668 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
3669 if (Scalar->isVoidTy() || VF == 1)
3671 return VectorType::get(Scalar, VF);
3674 char LoopVectorize::ID = 0;
3675 static const char lv_name[] = "Loop Vectorization";
3676 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
3677 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3678 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3679 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3680 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3681 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
3684 Pass *createLoopVectorizePass() {
3685 return new LoopVectorize();
3689 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
3690 // Check for a store.
3691 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
3692 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
3694 // Check for a load.
3695 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
3696 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;