1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #define LV_NAME "loop-vectorize"
46 #define DEBUG_TYPE LV_NAME
48 #include "llvm/Transforms/Vectorize.h"
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/EquivalenceClasses.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/StringExtras.h"
58 #include "llvm/Analysis/AliasAnalysis.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/Analysis/LoopIterator.h"
62 #include "llvm/Analysis/LoopPass.h"
63 #include "llvm/Analysis/ScalarEvolution.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.h"
80 #include "llvm/IR/Verifier.h"
81 #include "llvm/Pass.h"
82 #include "llvm/Support/BranchProbability.h"
83 #include "llvm/Support/CommandLine.h"
84 #include "llvm/Support/Debug.h"
85 #include "llvm/Support/PatternMatch.h"
86 #include "llvm/Support/ValueHandle.h"
87 #include "llvm/Support/raw_ostream.h"
88 #include "llvm/Target/TargetLibraryInfo.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/Local.h"
96 using namespace llvm::PatternMatch;
98 static cl::opt<unsigned>
99 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
100 cl::desc("Sets the SIMD width. Zero is autoselect."));
102 static cl::opt<unsigned>
103 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
104 cl::desc("Sets the vectorization unroll count. "
105 "Zero is autoselect."));
108 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
109 cl::desc("Enable if-conversion during vectorization."));
111 /// We don't vectorize loops with a known constant trip count below this number.
112 static cl::opt<unsigned>
113 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
115 cl::desc("Don't vectorize loops with a constant "
116 "trip count that is smaller than this "
119 /// This enables versioning on the strides of symbolically striding memory
120 /// accesses in code like the following.
121 /// for (i = 0; i < N; ++i)
122 /// A[i * Stride1] += B[i * Stride2] ...
124 /// Will be roughly translated to
125 /// if (Stride1 == 1 && Stride2 == 1) {
126 /// for (i = 0; i < N; i+=4)
130 static cl::opt<bool> EnableMemAccessVersioning(
131 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
132 cl::desc("Enable symblic stride memory access versioning"));
134 /// We don't unroll loops with a known constant trip count below this number.
135 static const unsigned TinyTripCountUnrollThreshold = 128;
137 /// When performing memory disambiguation checks at runtime do not make more
138 /// than this number of comparisons.
139 static const unsigned RuntimeMemoryCheckThreshold = 8;
141 /// Maximum simd width.
142 static const unsigned MaxVectorWidth = 64;
144 static cl::opt<unsigned> ForceTargetNumScalarRegs(
145 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
146 cl::desc("A flag that overrides the target's number of scalar registers."));
148 static cl::opt<unsigned> ForceTargetNumVectorRegs(
149 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
150 cl::desc("A flag that overrides the target's number of vector registers."));
152 /// Maximum vectorization unroll count.
153 static const unsigned MaxUnrollFactor = 16;
155 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
156 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's max unroll factor for scalar "
160 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
161 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's max unroll factor for "
163 "vectorized loops."));
165 static cl::opt<unsigned> ForceTargetInstructionCost(
166 "force-target-instruction-cost", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's expected cost for "
168 "an instruction to a single constant value. Mostly "
169 "useful for getting consistent testing."));
171 static cl::opt<unsigned> SmallLoopCost(
172 "small-loop-cost", cl::init(20), cl::Hidden,
173 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
175 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
176 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
177 cl::desc("Enable the use of the block frequency analysis to access PGO "
178 "heuristics minimizing code growth in cold regions and being more "
179 "aggressive in hot regions."));
181 // Runtime unroll loops for load/store throughput.
182 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
183 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
184 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
186 /// The number of stores in a loop that are allowed to need predication.
187 static cl::opt<unsigned> NumberOfStoresToPredicate(
188 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
189 cl::desc("Max number of stores to be predicated behind an if."));
191 static cl::opt<bool> EnableIndVarRegisterHeur(
192 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
193 cl::desc("Count the induction variable only once when unrolling"));
195 static cl::opt<bool> EnableCondStoresVectorization(
196 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
197 cl::desc("Enable if predication of stores during vectorization."));
201 // Forward declarations.
202 class LoopVectorizationLegality;
203 class LoopVectorizationCostModel;
205 /// InnerLoopVectorizer vectorizes loops which contain only one basic
206 /// block to a specified vectorization factor (VF).
207 /// This class performs the widening of scalars into vectors, or multiple
208 /// scalars. This class also implements the following features:
209 /// * It inserts an epilogue loop for handling loops that don't have iteration
210 /// counts that are known to be a multiple of the vectorization factor.
211 /// * It handles the code generation for reduction variables.
212 /// * Scalarization (implementation using scalars) of un-vectorizable
214 /// InnerLoopVectorizer does not perform any vectorization-legality
215 /// checks, and relies on the caller to check for the different legality
216 /// aspects. The InnerLoopVectorizer relies on the
217 /// LoopVectorizationLegality class to provide information about the induction
218 /// and reduction variables that were found to a given vectorization factor.
219 class InnerLoopVectorizer {
221 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
222 DominatorTree *DT, DataLayout *DL,
223 const TargetLibraryInfo *TLI, unsigned VecWidth,
224 unsigned UnrollFactor)
225 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
226 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
227 OldInduction(0), WidenMap(UnrollFactor), Legal(0) {}
229 // Perform the actual loop widening (vectorization).
230 void vectorize(LoopVectorizationLegality *L) {
232 // Create a new empty loop. Unlink the old loop and connect the new one.
234 // Widen each instruction in the old loop to a new one in the new loop.
235 // Use the Legality module to find the induction and reduction variables.
237 // Register the new loop and update the analysis passes.
241 virtual ~InnerLoopVectorizer() {}
244 /// A small list of PHINodes.
245 typedef SmallVector<PHINode*, 4> PhiVector;
246 /// When we unroll loops we have multiple vector values for each scalar.
247 /// This data structure holds the unrolled and vectorized values that
248 /// originated from one scalar instruction.
249 typedef SmallVector<Value*, 2> VectorParts;
251 // When we if-convert we need create edge masks. We have to cache values so
252 // that we don't end up with exponential recursion/IR.
253 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
254 VectorParts> EdgeMaskCache;
256 /// \brief Add code that checks at runtime if the accessed arrays overlap.
258 /// Returns a pair of instructions where the first element is the first
259 /// instruction generated in possibly a sequence of instructions and the
260 /// second value is the final comparator value or NULL if no check is needed.
261 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
263 /// \brief Add checks for strides that where assumed to be 1.
265 /// Returns the last check instruction and the first check instruction in the
266 /// pair as (first, last).
267 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
269 /// Create an empty loop, based on the loop ranges of the old loop.
270 void createEmptyLoop();
271 /// Copy and widen the instructions from the old loop.
272 virtual void vectorizeLoop();
274 /// \brief The Loop exit block may have single value PHI nodes where the
275 /// incoming value is 'Undef'. While vectorizing we only handled real values
276 /// that were defined inside the loop. Here we fix the 'undef case'.
280 /// A helper function that computes the predicate of the block BB, assuming
281 /// that the header block of the loop is set to True. It returns the *entry*
282 /// mask for the block BB.
283 VectorParts createBlockInMask(BasicBlock *BB);
284 /// A helper function that computes the predicate of the edge between SRC
286 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
288 /// A helper function to vectorize a single BB within the innermost loop.
289 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
291 /// Vectorize a single PHINode in a block. This method handles the induction
292 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
293 /// arbitrary length vectors.
294 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
295 unsigned UF, unsigned VF, PhiVector *PV);
297 /// Insert the new loop to the loop hierarchy and pass manager
298 /// and update the analysis passes.
299 void updateAnalysis();
301 /// This instruction is un-vectorizable. Implement it as a sequence
302 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
303 /// scalarized instruction behind an if block predicated on the control
304 /// dependence of the instruction.
305 virtual void scalarizeInstruction(Instruction *Instr,
306 bool IfPredicateStore=false);
308 /// Vectorize Load and Store instructions,
309 virtual void vectorizeMemoryInstruction(Instruction *Instr);
311 /// Create a broadcast instruction. This method generates a broadcast
312 /// instruction (shuffle) for loop invariant values and for the induction
313 /// value. If this is the induction variable then we extend it to N, N+1, ...
314 /// this is needed because each iteration in the loop corresponds to a SIMD
316 virtual Value *getBroadcastInstrs(Value *V);
318 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
319 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
320 /// The sequence starts at StartIndex.
321 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
323 /// When we go over instructions in the basic block we rely on previous
324 /// values within the current basic block or on loop invariant values.
325 /// When we widen (vectorize) values we place them in the map. If the values
326 /// are not within the map, they have to be loop invariant, so we simply
327 /// broadcast them into a vector.
328 VectorParts &getVectorValue(Value *V);
330 /// Generate a shuffle sequence that will reverse the vector Vec.
331 virtual Value *reverseVector(Value *Vec);
333 /// This is a helper class that holds the vectorizer state. It maps scalar
334 /// instructions to vector instructions. When the code is 'unrolled' then
335 /// then a single scalar value is mapped to multiple vector parts. The parts
336 /// are stored in the VectorPart type.
338 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
340 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
342 /// \return True if 'Key' is saved in the Value Map.
343 bool has(Value *Key) const { return MapStorage.count(Key); }
345 /// Initializes a new entry in the map. Sets all of the vector parts to the
346 /// save value in 'Val'.
347 /// \return A reference to a vector with splat values.
348 VectorParts &splat(Value *Key, Value *Val) {
349 VectorParts &Entry = MapStorage[Key];
350 Entry.assign(UF, Val);
354 ///\return A reference to the value that is stored at 'Key'.
355 VectorParts &get(Value *Key) {
356 VectorParts &Entry = MapStorage[Key];
359 assert(Entry.size() == UF);
364 /// The unroll factor. Each entry in the map stores this number of vector
368 /// Map storage. We use std::map and not DenseMap because insertions to a
369 /// dense map invalidates its iterators.
370 std::map<Value *, VectorParts> MapStorage;
373 /// The original loop.
375 /// Scev analysis to use.
383 /// Target Library Info.
384 const TargetLibraryInfo *TLI;
386 /// The vectorization SIMD factor to use. Each vector will have this many
391 /// The vectorization unroll factor to use. Each scalar is vectorized to this
392 /// many different vector instructions.
395 /// The builder that we use
398 // --- Vectorization state ---
400 /// The vector-loop preheader.
401 BasicBlock *LoopVectorPreHeader;
402 /// The scalar-loop preheader.
403 BasicBlock *LoopScalarPreHeader;
404 /// Middle Block between the vector and the scalar.
405 BasicBlock *LoopMiddleBlock;
406 ///The ExitBlock of the scalar loop.
407 BasicBlock *LoopExitBlock;
408 ///The vector loop body.
409 SmallVector<BasicBlock *, 4> LoopVectorBody;
410 ///The scalar loop body.
411 BasicBlock *LoopScalarBody;
412 /// A list of all bypass blocks. The first block is the entry of the loop.
413 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
415 /// The new Induction variable which was added to the new block.
417 /// The induction variable of the old basic block.
418 PHINode *OldInduction;
419 /// Holds the extended (to the widest induction type) start index.
421 /// Maps scalars to widened vectors.
423 EdgeMaskCache MaskCache;
425 LoopVectorizationLegality *Legal;
428 class InnerLoopUnroller : public InnerLoopVectorizer {
430 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
431 DominatorTree *DT, DataLayout *DL,
432 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
433 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
436 virtual void scalarizeInstruction(Instruction *Instr, bool IfPredicateStore = false);
437 virtual void vectorizeMemoryInstruction(Instruction *Instr);
438 virtual Value *getBroadcastInstrs(Value *V);
439 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
440 virtual Value *reverseVector(Value *Vec);
443 /// \brief Look for a meaningful debug location on the instruction or it's
445 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
450 if (I->getDebugLoc() != Empty)
453 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
454 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
455 if (OpInst->getDebugLoc() != Empty)
462 /// \brief Set the debug location in the builder using the debug location in the
464 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
465 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
466 B.SetCurrentDebugLocation(Inst->getDebugLoc());
468 B.SetCurrentDebugLocation(DebugLoc());
471 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
472 /// to what vectorization factor.
473 /// This class does not look at the profitability of vectorization, only the
474 /// legality. This class has two main kinds of checks:
475 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
476 /// will change the order of memory accesses in a way that will change the
477 /// correctness of the program.
478 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
479 /// checks for a number of different conditions, such as the availability of a
480 /// single induction variable, that all types are supported and vectorize-able,
481 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
482 /// This class is also used by InnerLoopVectorizer for identifying
483 /// induction variable and the different reduction variables.
484 class LoopVectorizationLegality {
488 unsigned NumPredStores;
490 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
491 DominatorTree *DT, TargetLibraryInfo *TLI)
492 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
493 DT(DT), TLI(TLI), Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
494 MaxSafeDepDistBytes(-1U) {}
496 /// This enum represents the kinds of reductions that we support.
498 RK_NoReduction, ///< Not a reduction.
499 RK_IntegerAdd, ///< Sum of integers.
500 RK_IntegerMult, ///< Product of integers.
501 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
502 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
503 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
504 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
505 RK_FloatAdd, ///< Sum of floats.
506 RK_FloatMult, ///< Product of floats.
507 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
510 /// This enum represents the kinds of inductions that we support.
512 IK_NoInduction, ///< Not an induction variable.
513 IK_IntInduction, ///< Integer induction variable. Step = 1.
514 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
515 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
516 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
519 // This enum represents the kind of minmax reduction.
520 enum MinMaxReductionKind {
530 /// This struct holds information about reduction variables.
531 struct ReductionDescriptor {
532 ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
533 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
535 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
536 MinMaxReductionKind MK)
537 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
539 // The starting value of the reduction.
540 // It does not have to be zero!
541 TrackingVH<Value> StartValue;
542 // The instruction who's value is used outside the loop.
543 Instruction *LoopExitInstr;
544 // The kind of the reduction.
546 // If this a min/max reduction the kind of reduction.
547 MinMaxReductionKind MinMaxKind;
550 /// This POD struct holds information about a potential reduction operation.
551 struct ReductionInstDesc {
552 ReductionInstDesc(bool IsRedux, Instruction *I) :
553 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
555 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
556 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
558 // Is this instruction a reduction candidate.
560 // The last instruction in a min/max pattern (select of the select(icmp())
561 // pattern), or the current reduction instruction otherwise.
562 Instruction *PatternLastInst;
563 // If this is a min/max pattern the comparison predicate.
564 MinMaxReductionKind MinMaxKind;
567 /// This struct holds information about the memory runtime legality
568 /// check that a group of pointers do not overlap.
569 struct RuntimePointerCheck {
570 RuntimePointerCheck() : Need(false) {}
572 /// Reset the state of the pointer runtime information.
579 DependencySetId.clear();
582 /// Insert a pointer and calculate the start and end SCEVs.
583 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
584 unsigned DepSetId, ValueToValueMap &Strides);
586 /// This flag indicates if we need to add the runtime check.
588 /// Holds the pointers that we need to check.
589 SmallVector<TrackingVH<Value>, 2> Pointers;
590 /// Holds the pointer value at the beginning of the loop.
591 SmallVector<const SCEV*, 2> Starts;
592 /// Holds the pointer value at the end of the loop.
593 SmallVector<const SCEV*, 2> Ends;
594 /// Holds the information if this pointer is used for writing to memory.
595 SmallVector<bool, 2> IsWritePtr;
596 /// Holds the id of the set of pointers that could be dependent because of a
597 /// shared underlying object.
598 SmallVector<unsigned, 2> DependencySetId;
601 /// A struct for saving information about induction variables.
602 struct InductionInfo {
603 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
604 InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
606 TrackingVH<Value> StartValue;
611 /// ReductionList contains the reduction descriptors for all
612 /// of the reductions that were found in the loop.
613 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
615 /// InductionList saves induction variables and maps them to the
616 /// induction descriptor.
617 typedef MapVector<PHINode*, InductionInfo> InductionList;
619 /// Returns true if it is legal to vectorize this loop.
620 /// This does not mean that it is profitable to vectorize this
621 /// loop, only that it is legal to do so.
624 /// Returns the Induction variable.
625 PHINode *getInduction() { return Induction; }
627 /// Returns the reduction variables found in the loop.
628 ReductionList *getReductionVars() { return &Reductions; }
630 /// Returns the induction variables found in the loop.
631 InductionList *getInductionVars() { return &Inductions; }
633 /// Returns the widest induction type.
634 Type *getWidestInductionType() { return WidestIndTy; }
636 /// Returns True if V is an induction variable in this loop.
637 bool isInductionVariable(const Value *V);
639 /// Return true if the block BB needs to be predicated in order for the loop
640 /// to be vectorized.
641 bool blockNeedsPredication(BasicBlock *BB);
643 /// Check if this pointer is consecutive when vectorizing. This happens
644 /// when the last index of the GEP is the induction variable, or that the
645 /// pointer itself is an induction variable.
646 /// This check allows us to vectorize A[idx] into a wide load/store.
648 /// 0 - Stride is unknown or non-consecutive.
649 /// 1 - Address is consecutive.
650 /// -1 - Address is consecutive, and decreasing.
651 int isConsecutivePtr(Value *Ptr);
653 /// Returns true if the value V is uniform within the loop.
654 bool isUniform(Value *V);
656 /// Returns true if this instruction will remain scalar after vectorization.
657 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
659 /// Returns the information that we collected about runtime memory check.
660 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
662 /// This function returns the identity element (or neutral element) for
664 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
666 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
668 bool hasStride(Value *V) { return StrideSet.count(V); }
669 bool mustCheckStrides() { return !StrideSet.empty(); }
670 SmallPtrSet<Value *, 8>::iterator strides_begin() {
671 return StrideSet.begin();
673 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
676 /// Check if a single basic block loop is vectorizable.
677 /// At this point we know that this is a loop with a constant trip count
678 /// and we only need to check individual instructions.
679 bool canVectorizeInstrs();
681 /// When we vectorize loops we may change the order in which
682 /// we read and write from memory. This method checks if it is
683 /// legal to vectorize the code, considering only memory constrains.
684 /// Returns true if the loop is vectorizable
685 bool canVectorizeMemory();
687 /// Return true if we can vectorize this loop using the IF-conversion
689 bool canVectorizeWithIfConvert();
691 /// Collect the variables that need to stay uniform after vectorization.
692 void collectLoopUniforms();
694 /// Return true if all of the instructions in the block can be speculatively
695 /// executed. \p SafePtrs is a list of addresses that are known to be legal
696 /// and we know that we can read from them without segfault.
697 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
699 /// Returns True, if 'Phi' is the kind of reduction variable for type
700 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
701 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
702 /// Returns a struct describing if the instruction 'I' can be a reduction
703 /// variable of type 'Kind'. If the reduction is a min/max pattern of
704 /// select(icmp()) this function advances the instruction pointer 'I' from the
705 /// compare instruction to the select instruction and stores this pointer in
706 /// 'PatternLastInst' member of the returned struct.
707 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
708 ReductionInstDesc &Desc);
709 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
710 /// pattern corresponding to a min(X, Y) or max(X, Y).
711 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
712 ReductionInstDesc &Prev);
713 /// Returns the induction kind of Phi. This function may return NoInduction
714 /// if the PHI is not an induction variable.
715 InductionKind isInductionVariable(PHINode *Phi);
717 /// \brief Collect memory access with loop invariant strides.
719 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
721 void collectStridedAcccess(Value *LoadOrStoreInst);
723 /// The loop that we evaluate.
727 /// DataLayout analysis.
731 /// Target Library Info.
732 TargetLibraryInfo *TLI;
734 // --- vectorization state --- //
736 /// Holds the integer induction variable. This is the counter of the
739 /// Holds the reduction variables.
740 ReductionList Reductions;
741 /// Holds all of the induction variables that we found in the loop.
742 /// Notice that inductions don't need to start at zero and that induction
743 /// variables can be pointers.
744 InductionList Inductions;
745 /// Holds the widest induction type encountered.
748 /// Allowed outside users. This holds the reduction
749 /// vars which can be accessed from outside the loop.
750 SmallPtrSet<Value*, 4> AllowedExit;
751 /// This set holds the variables which are known to be uniform after
753 SmallPtrSet<Instruction*, 4> Uniforms;
754 /// We need to check that all of the pointers in this list are disjoint
756 RuntimePointerCheck PtrRtCheck;
757 /// Can we assume the absence of NaNs.
758 bool HasFunNoNaNAttr;
760 unsigned MaxSafeDepDistBytes;
762 ValueToValueMap Strides;
763 SmallPtrSet<Value *, 8> StrideSet;
766 /// LoopVectorizationCostModel - estimates the expected speedups due to
768 /// In many cases vectorization is not profitable. This can happen because of
769 /// a number of reasons. In this class we mainly attempt to predict the
770 /// expected speedup/slowdowns due to the supported instruction set. We use the
771 /// TargetTransformInfo to query the different backends for the cost of
772 /// different operations.
773 class LoopVectorizationCostModel {
775 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
776 LoopVectorizationLegality *Legal,
777 const TargetTransformInfo &TTI,
778 DataLayout *DL, const TargetLibraryInfo *TLI)
779 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
781 /// Information about vectorization costs
782 struct VectorizationFactor {
783 unsigned Width; // Vector width with best cost
784 unsigned Cost; // Cost of the loop with that width
786 /// \return The most profitable vectorization factor and the cost of that VF.
787 /// This method checks every power of two up to VF. If UserVF is not ZERO
788 /// then this vectorization factor will be selected if vectorization is
790 VectorizationFactor selectVectorizationFactor(bool OptForSize,
793 /// \return The size (in bits) of the widest type in the code that
794 /// needs to be vectorized. We ignore values that remain scalar such as
795 /// 64 bit loop indices.
796 unsigned getWidestType();
798 /// \return The most profitable unroll factor.
799 /// If UserUF is non-zero then this method finds the best unroll-factor
800 /// based on register pressure and other parameters.
801 /// VF and LoopCost are the selected vectorization factor and the cost of the
803 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
806 /// \brief A struct that represents some properties of the register usage
808 struct RegisterUsage {
809 /// Holds the number of loop invariant values that are used in the loop.
810 unsigned LoopInvariantRegs;
811 /// Holds the maximum number of concurrent live intervals in the loop.
812 unsigned MaxLocalUsers;
813 /// Holds the number of instructions in the loop.
814 unsigned NumInstructions;
817 /// \return information about the register usage of the loop.
818 RegisterUsage calculateRegisterUsage();
821 /// Returns the expected execution cost. The unit of the cost does
822 /// not matter because we use the 'cost' units to compare different
823 /// vector widths. The cost that is returned is *not* normalized by
824 /// the factor width.
825 unsigned expectedCost(unsigned VF);
827 /// Returns the execution time cost of an instruction for a given vector
828 /// width. Vector width of one means scalar.
829 unsigned getInstructionCost(Instruction *I, unsigned VF);
831 /// A helper function for converting Scalar types to vector types.
832 /// If the incoming type is void, we return void. If the VF is 1, we return
834 static Type* ToVectorTy(Type *Scalar, unsigned VF);
836 /// Returns whether the instruction is a load or store and will be a emitted
837 /// as a vector operation.
838 bool isConsecutiveLoadOrStore(Instruction *I);
840 /// The loop that we evaluate.
844 /// Loop Info analysis.
846 /// Vectorization legality.
847 LoopVectorizationLegality *Legal;
848 /// Vector target information.
849 const TargetTransformInfo &TTI;
850 /// Target data layout information.
852 /// Target Library Info.
853 const TargetLibraryInfo *TLI;
856 /// Utility class for getting and setting loop vectorizer hints in the form
857 /// of loop metadata.
858 struct LoopVectorizeHints {
859 /// Vectorization width.
861 /// Vectorization unroll factor.
863 /// Vectorization forced (-1 not selected, 0 force disabled, 1 force enabled)
866 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
867 : Width(VectorizationFactor)
868 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
870 , LoopID(L->getLoopID()) {
872 // The command line options override any loop metadata except for when
873 // width == 1 which is used to indicate the loop is already vectorized.
874 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
875 Width = VectorizationFactor;
876 if (VectorizationUnroll.getNumOccurrences() > 0)
877 Unroll = VectorizationUnroll;
879 DEBUG(if (DisableUnrolling && Unroll == 1)
880 dbgs() << "LV: Unrolling disabled by the pass manager\n");
883 /// Return the loop vectorizer metadata prefix.
884 static StringRef Prefix() { return "llvm.vectorizer."; }
886 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
887 SmallVector<Value*, 2> Vals;
888 Vals.push_back(MDString::get(Context, Name));
889 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
890 return MDNode::get(Context, Vals);
893 /// Mark the loop L as already vectorized by setting the width to 1.
894 void setAlreadyVectorized(Loop *L) {
895 LLVMContext &Context = L->getHeader()->getContext();
899 // Create a new loop id with one more operand for the already_vectorized
900 // hint. If the loop already has a loop id then copy the existing operands.
901 SmallVector<Value*, 4> Vals(1);
903 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
904 Vals.push_back(LoopID->getOperand(i));
906 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
907 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
909 MDNode *NewLoopID = MDNode::get(Context, Vals);
910 // Set operand 0 to refer to the loop id itself.
911 NewLoopID->replaceOperandWith(0, NewLoopID);
913 L->setLoopID(NewLoopID);
915 LoopID->replaceAllUsesWith(NewLoopID);
923 /// Find hints specified in the loop metadata.
924 void getHints(const Loop *L) {
928 // First operand should refer to the loop id itself.
929 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
930 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
932 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
933 const MDString *S = 0;
934 SmallVector<Value*, 4> Args;
936 // The expected hint is either a MDString or a MDNode with the first
937 // operand a MDString.
938 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
939 if (!MD || MD->getNumOperands() == 0)
941 S = dyn_cast<MDString>(MD->getOperand(0));
942 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
943 Args.push_back(MD->getOperand(i));
945 S = dyn_cast<MDString>(LoopID->getOperand(i));
946 assert(Args.size() == 0 && "too many arguments for MDString");
952 // Check if the hint starts with the vectorizer prefix.
953 StringRef Hint = S->getString();
954 if (!Hint.startswith(Prefix()))
956 // Remove the prefix.
957 Hint = Hint.substr(Prefix().size(), StringRef::npos);
959 if (Args.size() == 1)
960 getHint(Hint, Args[0]);
964 // Check string hint with one operand.
965 void getHint(StringRef Hint, Value *Arg) {
966 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
968 unsigned Val = C->getZExtValue();
970 if (Hint == "width") {
971 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
974 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
975 } else if (Hint == "unroll") {
976 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
979 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
980 } else if (Hint == "enable") {
981 if (C->getBitWidth() == 1)
984 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
986 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
991 static void addInnerLoop(Loop *L, SmallVectorImpl<Loop *> &V) {
993 return V.push_back(L);
995 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
999 /// The LoopVectorize Pass.
1000 struct LoopVectorize : public FunctionPass {
1001 /// Pass identification, replacement for typeid
1004 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1006 DisableUnrolling(NoUnrolling),
1007 AlwaysVectorize(AlwaysVectorize) {
1008 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1011 ScalarEvolution *SE;
1014 TargetTransformInfo *TTI;
1016 BlockFrequencyInfo *BFI;
1017 TargetLibraryInfo *TLI;
1018 bool DisableUnrolling;
1019 bool AlwaysVectorize;
1021 BlockFrequency ColdEntryFreq;
1023 virtual bool runOnFunction(Function &F) {
1024 SE = &getAnalysis<ScalarEvolution>();
1025 DL = getAnalysisIfAvailable<DataLayout>();
1026 LI = &getAnalysis<LoopInfo>();
1027 TTI = &getAnalysis<TargetTransformInfo>();
1028 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1029 BFI = &getAnalysis<BlockFrequencyInfo>();
1030 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1032 // Compute some weights outside of the loop over the loops. Compute this
1033 // using a BranchProbability to re-use its scaling math.
1034 const BranchProbability ColdProb(1, 5); // 20%
1035 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1037 // If the target claims to have no vector registers don't attempt
1039 if (!TTI->getNumberOfRegisters(true))
1043 DEBUG(dbgs() << "LV: Not vectorizing: Missing data layout\n");
1047 // Build up a worklist of inner-loops to vectorize. This is necessary as
1048 // the act of vectorizing or partially unrolling a loop creates new loops
1049 // and can invalidate iterators across the loops.
1050 SmallVector<Loop *, 8> Worklist;
1052 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
1053 addInnerLoop(*I, Worklist);
1055 // Now walk the identified inner loops.
1056 bool Changed = false;
1057 while (!Worklist.empty())
1058 Changed |= processLoop(Worklist.pop_back_val());
1060 // Process each loop nest in the function.
1064 bool processLoop(Loop *L) {
1065 // We only handle inner loops, so if there are children just recurse.
1067 bool Changed = false;
1068 for (Loop::iterator I = L->begin(), E = L->begin(); I != E; ++I)
1069 Changed |= processLoop(*I);
1073 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
1074 L->getHeader()->getParent()->getName() << "\"\n");
1076 LoopVectorizeHints Hints(L, DisableUnrolling);
1078 if (Hints.Force == 0) {
1079 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1083 if (!AlwaysVectorize && Hints.Force != 1) {
1084 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1088 if (Hints.Width == 1 && Hints.Unroll == 1) {
1089 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1093 // Check if it is legal to vectorize the loop.
1094 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1095 if (!LVL.canVectorize()) {
1096 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1100 // Use the cost model.
1101 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1103 // Check the function attributes to find out if this function should be
1104 // optimized for size.
1105 Function *F = L->getHeader()->getParent();
1107 Hints.Force != 1 && F->hasFnAttribute(Attribute::OptimizeForSize);
1109 // Compute the weighted frequency of this loop being executed and see if it
1110 // is less than 20% of the function entry baseline frequency. Note that we
1111 // always have a canonical loop here because we think we *can* vectoriez.
1112 // FIXME: This is hidden behind a flag due to pervasive problems with
1113 // exactly what block frequency models.
1114 if (LoopVectorizeWithBlockFrequency) {
1115 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1116 if (Hints.Force != 1 && LoopEntryFreq < ColdEntryFreq)
1120 // Check the function attributes to see if implicit floats are allowed.a
1121 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1122 // an integer loop and the vector instructions selected are purely integer
1123 // vector instructions?
1124 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1125 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1126 "attribute is used.\n");
1130 // Select the optimal vectorization factor.
1131 LoopVectorizationCostModel::VectorizationFactor VF;
1132 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
1133 // Select the unroll factor.
1134 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1137 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
1138 F->getParent()->getModuleIdentifier() << '\n');
1139 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1141 if (VF.Width == 1) {
1142 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1145 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1146 // We decided not to vectorize, but we may want to unroll.
1147 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1148 Unroller.vectorize(&LVL);
1150 // If we decided that it is *legal* to vectorize the loop then do it.
1151 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1155 // Mark the loop as already vectorized to avoid vectorizing again.
1156 Hints.setAlreadyVectorized(L);
1158 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1162 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1163 AU.addRequiredID(LoopSimplifyID);
1164 AU.addRequiredID(LCSSAID);
1165 AU.addRequired<BlockFrequencyInfo>();
1166 AU.addRequired<DominatorTreeWrapperPass>();
1167 AU.addRequired<LoopInfo>();
1168 AU.addRequired<ScalarEvolution>();
1169 AU.addRequired<TargetTransformInfo>();
1170 AU.addPreserved<LoopInfo>();
1171 AU.addPreserved<DominatorTreeWrapperPass>();
1176 } // end anonymous namespace
1178 //===----------------------------------------------------------------------===//
1179 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1180 // LoopVectorizationCostModel.
1181 //===----------------------------------------------------------------------===//
1183 static Value *stripIntegerCast(Value *V) {
1184 if (CastInst *CI = dyn_cast<CastInst>(V))
1185 if (CI->getOperand(0)->getType()->isIntegerTy())
1186 return CI->getOperand(0);
1190 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1192 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1194 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1195 ValueToValueMap &PtrToStride,
1196 Value *Ptr, Value *OrigPtr = 0) {
1198 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1200 // If there is an entry in the map return the SCEV of the pointer with the
1201 // symbolic stride replaced by one.
1202 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1203 if (SI != PtrToStride.end()) {
1204 Value *StrideVal = SI->second;
1207 StrideVal = stripIntegerCast(StrideVal);
1209 // Replace symbolic stride by one.
1210 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1211 ValueToValueMap RewriteMap;
1212 RewriteMap[StrideVal] = One;
1215 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1216 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1221 // Otherwise, just return the SCEV of the original pointer.
1222 return SE->getSCEV(Ptr);
1225 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1226 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1227 ValueToValueMap &Strides) {
1228 // Get the stride replaced scev.
1229 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1230 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1231 assert(AR && "Invalid addrec expression");
1232 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1233 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1234 Pointers.push_back(Ptr);
1235 Starts.push_back(AR->getStart());
1236 Ends.push_back(ScEnd);
1237 IsWritePtr.push_back(WritePtr);
1238 DependencySetId.push_back(DepSetId);
1241 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1242 // We need to place the broadcast of invariant variables outside the loop.
1243 Instruction *Instr = dyn_cast<Instruction>(V);
1245 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1246 Instr->getParent()) != LoopVectorBody.end());
1247 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1249 // Place the code for broadcasting invariant variables in the new preheader.
1250 IRBuilder<>::InsertPointGuard Guard(Builder);
1252 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1254 // Broadcast the scalar into all locations in the vector.
1255 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1260 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1262 assert(Val->getType()->isVectorTy() && "Must be a vector");
1263 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1264 "Elem must be an integer");
1265 // Create the types.
1266 Type *ITy = Val->getType()->getScalarType();
1267 VectorType *Ty = cast<VectorType>(Val->getType());
1268 int VLen = Ty->getNumElements();
1269 SmallVector<Constant*, 8> Indices;
1271 // Create a vector of consecutive numbers from zero to VF.
1272 for (int i = 0; i < VLen; ++i) {
1273 int64_t Idx = Negate ? (-i) : i;
1274 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1277 // Add the consecutive indices to the vector value.
1278 Constant *Cv = ConstantVector::get(Indices);
1279 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1280 return Builder.CreateAdd(Val, Cv, "induction");
1283 /// \brief Find the operand of the GEP that should be checked for consecutive
1284 /// stores. This ignores trailing indices that have no effect on the final
1286 static unsigned getGEPInductionOperand(DataLayout *DL,
1287 const GetElementPtrInst *Gep) {
1288 unsigned LastOperand = Gep->getNumOperands() - 1;
1289 unsigned GEPAllocSize = DL->getTypeAllocSize(
1290 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1292 // Walk backwards and try to peel off zeros.
1293 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1294 // Find the type we're currently indexing into.
1295 gep_type_iterator GEPTI = gep_type_begin(Gep);
1296 std::advance(GEPTI, LastOperand - 1);
1298 // If it's a type with the same allocation size as the result of the GEP we
1299 // can peel off the zero index.
1300 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1308 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1309 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1310 // Make sure that the pointer does not point to structs.
1311 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1314 // If this value is a pointer induction variable we know it is consecutive.
1315 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1316 if (Phi && Inductions.count(Phi)) {
1317 InductionInfo II = Inductions[Phi];
1318 if (IK_PtrInduction == II.IK)
1320 else if (IK_ReversePtrInduction == II.IK)
1324 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1328 unsigned NumOperands = Gep->getNumOperands();
1329 Value *GpPtr = Gep->getPointerOperand();
1330 // If this GEP value is a consecutive pointer induction variable and all of
1331 // the indices are constant then we know it is consecutive. We can
1332 Phi = dyn_cast<PHINode>(GpPtr);
1333 if (Phi && Inductions.count(Phi)) {
1335 // Make sure that the pointer does not point to structs.
1336 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1337 if (GepPtrType->getElementType()->isAggregateType())
1340 // Make sure that all of the index operands are loop invariant.
1341 for (unsigned i = 1; i < NumOperands; ++i)
1342 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1345 InductionInfo II = Inductions[Phi];
1346 if (IK_PtrInduction == II.IK)
1348 else if (IK_ReversePtrInduction == II.IK)
1352 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1354 // Check that all of the gep indices are uniform except for our induction
1356 for (unsigned i = 0; i != NumOperands; ++i)
1357 if (i != InductionOperand &&
1358 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1361 // We can emit wide load/stores only if the last non-zero index is the
1362 // induction variable.
1363 const SCEV *Last = 0;
1364 if (!Strides.count(Gep))
1365 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1367 // Because of the multiplication by a stride we can have a s/zext cast.
1368 // We are going to replace this stride by 1 so the cast is safe to ignore.
1370 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1371 // %0 = trunc i64 %indvars.iv to i32
1372 // %mul = mul i32 %0, %Stride1
1373 // %idxprom = zext i32 %mul to i64 << Safe cast.
1374 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1376 Last = replaceSymbolicStrideSCEV(SE, Strides,
1377 Gep->getOperand(InductionOperand), Gep);
1378 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1380 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1384 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1385 const SCEV *Step = AR->getStepRecurrence(*SE);
1387 // The memory is consecutive because the last index is consecutive
1388 // and all other indices are loop invariant.
1391 if (Step->isAllOnesValue())
1398 bool LoopVectorizationLegality::isUniform(Value *V) {
1399 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1402 InnerLoopVectorizer::VectorParts&
1403 InnerLoopVectorizer::getVectorValue(Value *V) {
1404 assert(V != Induction && "The new induction variable should not be used.");
1405 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1407 // If we have a stride that is replaced by one, do it here.
1408 if (Legal->hasStride(V))
1409 V = ConstantInt::get(V->getType(), 1);
1411 // If we have this scalar in the map, return it.
1412 if (WidenMap.has(V))
1413 return WidenMap.get(V);
1415 // If this scalar is unknown, assume that it is a constant or that it is
1416 // loop invariant. Broadcast V and save the value for future uses.
1417 Value *B = getBroadcastInstrs(V);
1418 return WidenMap.splat(V, B);
1421 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1422 assert(Vec->getType()->isVectorTy() && "Invalid type");
1423 SmallVector<Constant*, 8> ShuffleMask;
1424 for (unsigned i = 0; i < VF; ++i)
1425 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1427 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1428 ConstantVector::get(ShuffleMask),
1432 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1433 // Attempt to issue a wide load.
1434 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1435 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1437 assert((LI || SI) && "Invalid Load/Store instruction");
1439 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1440 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1441 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1442 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1443 // An alignment of 0 means target abi alignment. We need to use the scalar's
1444 // target abi alignment in such a case.
1446 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1447 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1448 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1449 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1451 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1452 return scalarizeInstruction(Instr, true);
1454 if (ScalarAllocatedSize != VectorElementSize)
1455 return scalarizeInstruction(Instr);
1457 // If the pointer is loop invariant or if it is non-consecutive,
1458 // scalarize the load.
1459 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1460 bool Reverse = ConsecutiveStride < 0;
1461 bool UniformLoad = LI && Legal->isUniform(Ptr);
1462 if (!ConsecutiveStride || UniformLoad)
1463 return scalarizeInstruction(Instr);
1465 Constant *Zero = Builder.getInt32(0);
1466 VectorParts &Entry = WidenMap.get(Instr);
1468 // Handle consecutive loads/stores.
1469 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1470 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1471 setDebugLocFromInst(Builder, Gep);
1472 Value *PtrOperand = Gep->getPointerOperand();
1473 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1474 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1476 // Create the new GEP with the new induction variable.
1477 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1478 Gep2->setOperand(0, FirstBasePtr);
1479 Gep2->setName("gep.indvar.base");
1480 Ptr = Builder.Insert(Gep2);
1482 setDebugLocFromInst(Builder, Gep);
1483 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1484 OrigLoop) && "Base ptr must be invariant");
1486 // The last index does not have to be the induction. It can be
1487 // consecutive and be a function of the index. For example A[I+1];
1488 unsigned NumOperands = Gep->getNumOperands();
1489 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1490 // Create the new GEP with the new induction variable.
1491 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1493 for (unsigned i = 0; i < NumOperands; ++i) {
1494 Value *GepOperand = Gep->getOperand(i);
1495 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1497 // Update last index or loop invariant instruction anchored in loop.
1498 if (i == InductionOperand ||
1499 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1500 assert((i == InductionOperand ||
1501 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1502 "Must be last index or loop invariant");
1504 VectorParts &GEPParts = getVectorValue(GepOperand);
1505 Value *Index = GEPParts[0];
1506 Index = Builder.CreateExtractElement(Index, Zero);
1507 Gep2->setOperand(i, Index);
1508 Gep2->setName("gep.indvar.idx");
1511 Ptr = Builder.Insert(Gep2);
1513 // Use the induction element ptr.
1514 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1515 setDebugLocFromInst(Builder, Ptr);
1516 VectorParts &PtrVal = getVectorValue(Ptr);
1517 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1522 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1523 "We do not allow storing to uniform addresses");
1524 setDebugLocFromInst(Builder, SI);
1525 // We don't want to update the value in the map as it might be used in
1526 // another expression. So don't use a reference type for "StoredVal".
1527 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1529 for (unsigned Part = 0; Part < UF; ++Part) {
1530 // Calculate the pointer for the specific unroll-part.
1531 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1534 // If we store to reverse consecutive memory locations then we need
1535 // to reverse the order of elements in the stored value.
1536 StoredVal[Part] = reverseVector(StoredVal[Part]);
1537 // If the address is consecutive but reversed, then the
1538 // wide store needs to start at the last vector element.
1539 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1540 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1543 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1544 DataTy->getPointerTo(AddressSpace));
1545 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1551 assert(LI && "Must have a load instruction");
1552 setDebugLocFromInst(Builder, LI);
1553 for (unsigned Part = 0; Part < UF; ++Part) {
1554 // Calculate the pointer for the specific unroll-part.
1555 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1558 // If the address is consecutive but reversed, then the
1559 // wide store needs to start at the last vector element.
1560 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1561 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1564 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1565 DataTy->getPointerTo(AddressSpace));
1566 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1567 cast<LoadInst>(LI)->setAlignment(Alignment);
1568 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1572 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1573 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1574 // Holds vector parameters or scalars, in case of uniform vals.
1575 SmallVector<VectorParts, 4> Params;
1577 setDebugLocFromInst(Builder, Instr);
1579 // Find all of the vectorized parameters.
1580 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1581 Value *SrcOp = Instr->getOperand(op);
1583 // If we are accessing the old induction variable, use the new one.
1584 if (SrcOp == OldInduction) {
1585 Params.push_back(getVectorValue(SrcOp));
1589 // Try using previously calculated values.
1590 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1592 // If the src is an instruction that appeared earlier in the basic block
1593 // then it should already be vectorized.
1594 if (SrcInst && OrigLoop->contains(SrcInst)) {
1595 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1596 // The parameter is a vector value from earlier.
1597 Params.push_back(WidenMap.get(SrcInst));
1599 // The parameter is a scalar from outside the loop. Maybe even a constant.
1600 VectorParts Scalars;
1601 Scalars.append(UF, SrcOp);
1602 Params.push_back(Scalars);
1606 assert(Params.size() == Instr->getNumOperands() &&
1607 "Invalid number of operands");
1609 // Does this instruction return a value ?
1610 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1612 Value *UndefVec = IsVoidRetTy ? 0 :
1613 UndefValue::get(VectorType::get(Instr->getType(), VF));
1614 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1615 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1617 Instruction *InsertPt = Builder.GetInsertPoint();
1618 BasicBlock *IfBlock = Builder.GetInsertBlock();
1619 BasicBlock *CondBlock = 0;
1623 if (IfPredicateStore) {
1624 assert(Instr->getParent()->getSinglePredecessor() &&
1625 "Only support single predecessor blocks");
1626 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1627 Instr->getParent());
1628 VectorLp = LI->getLoopFor(IfBlock);
1629 assert(VectorLp && "Must have a loop for this block");
1632 // For each vector unroll 'part':
1633 for (unsigned Part = 0; Part < UF; ++Part) {
1634 // For each scalar that we create:
1635 for (unsigned Width = 0; Width < VF; ++Width) {
1639 if (IfPredicateStore) {
1640 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1641 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1642 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1643 LoopVectorBody.push_back(CondBlock);
1644 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1645 // Update Builder with newly created basic block.
1646 Builder.SetInsertPoint(InsertPt);
1649 Instruction *Cloned = Instr->clone();
1651 Cloned->setName(Instr->getName() + ".cloned");
1652 // Replace the operands of the cloned instructions with extracted scalars.
1653 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1654 Value *Op = Params[op][Part];
1655 // Param is a vector. Need to extract the right lane.
1656 if (Op->getType()->isVectorTy())
1657 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1658 Cloned->setOperand(op, Op);
1661 // Place the cloned scalar in the new loop.
1662 Builder.Insert(Cloned);
1664 // If the original scalar returns a value we need to place it in a vector
1665 // so that future users will be able to use it.
1667 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1668 Builder.getInt32(Width));
1670 if (IfPredicateStore) {
1671 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1672 LoopVectorBody.push_back(NewIfBlock);
1673 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1674 Builder.SetInsertPoint(InsertPt);
1675 Instruction *OldBr = IfBlock->getTerminator();
1676 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1677 OldBr->eraseFromParent();
1678 IfBlock = NewIfBlock;
1684 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1688 if (Instruction *I = dyn_cast<Instruction>(V))
1689 return I->getParent() == Loc->getParent() ? I : 0;
1693 std::pair<Instruction *, Instruction *>
1694 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1695 Instruction *tnullptr = 0;
1696 if (!Legal->mustCheckStrides())
1697 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1699 IRBuilder<> ChkBuilder(Loc);
1703 Instruction *FirstInst = 0;
1704 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1705 SE = Legal->strides_end();
1707 Value *Ptr = stripIntegerCast(*SI);
1708 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1710 // Store the first instruction we create.
1711 FirstInst = getFirstInst(FirstInst, C, Loc);
1713 Check = ChkBuilder.CreateOr(Check, C);
1718 // We have to do this trickery because the IRBuilder might fold the check to a
1719 // constant expression in which case there is no Instruction anchored in a
1721 LLVMContext &Ctx = Loc->getContext();
1722 Instruction *TheCheck =
1723 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1724 ChkBuilder.Insert(TheCheck, "stride.not.one");
1725 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1727 return std::make_pair(FirstInst, TheCheck);
1730 std::pair<Instruction *, Instruction *>
1731 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1732 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1733 Legal->getRuntimePointerCheck();
1735 Instruction *tnullptr = 0;
1736 if (!PtrRtCheck->Need)
1737 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1739 unsigned NumPointers = PtrRtCheck->Pointers.size();
1740 SmallVector<TrackingVH<Value> , 2> Starts;
1741 SmallVector<TrackingVH<Value> , 2> Ends;
1743 LLVMContext &Ctx = Loc->getContext();
1744 SCEVExpander Exp(*SE, "induction");
1745 Instruction *FirstInst = 0;
1747 for (unsigned i = 0; i < NumPointers; ++i) {
1748 Value *Ptr = PtrRtCheck->Pointers[i];
1749 const SCEV *Sc = SE->getSCEV(Ptr);
1751 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1752 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1754 Starts.push_back(Ptr);
1755 Ends.push_back(Ptr);
1757 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1758 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1760 // Use this type for pointer arithmetic.
1761 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1763 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1764 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1765 Starts.push_back(Start);
1766 Ends.push_back(End);
1770 IRBuilder<> ChkBuilder(Loc);
1771 // Our instructions might fold to a constant.
1772 Value *MemoryRuntimeCheck = 0;
1773 for (unsigned i = 0; i < NumPointers; ++i) {
1774 for (unsigned j = i+1; j < NumPointers; ++j) {
1775 // No need to check if two readonly pointers intersect.
1776 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1779 // Only need to check pointers between two different dependency sets.
1780 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1783 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1784 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1786 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1787 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1788 "Trying to bounds check pointers with different address spaces");
1790 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1791 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1793 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1794 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1795 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1796 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1798 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1799 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1800 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1801 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1802 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1803 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1804 if (MemoryRuntimeCheck) {
1805 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1807 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1809 MemoryRuntimeCheck = IsConflict;
1813 // We have to do this trickery because the IRBuilder might fold the check to a
1814 // constant expression in which case there is no Instruction anchored in a
1816 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1817 ConstantInt::getTrue(Ctx));
1818 ChkBuilder.Insert(Check, "memcheck.conflict");
1819 FirstInst = getFirstInst(FirstInst, Check, Loc);
1820 return std::make_pair(FirstInst, Check);
1823 void InnerLoopVectorizer::createEmptyLoop() {
1825 In this function we generate a new loop. The new loop will contain
1826 the vectorized instructions while the old loop will continue to run the
1829 [ ] <-- vector loop bypass (may consist of multiple blocks).
1832 | [ ] <-- vector pre header.
1836 | [ ]_| <-- vector loop.
1839 >[ ] <--- middle-block.
1842 | [ ] <--- new preheader.
1846 | [ ]_| <-- old scalar loop to handle remainder.
1849 >[ ] <-- exit block.
1853 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1854 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1855 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1856 assert(ExitBlock && "Must have an exit block");
1858 // Some loops have a single integer induction variable, while other loops
1859 // don't. One example is c++ iterators that often have multiple pointer
1860 // induction variables. In the code below we also support a case where we
1861 // don't have a single induction variable.
1862 OldInduction = Legal->getInduction();
1863 Type *IdxTy = Legal->getWidestInductionType();
1865 // Find the loop boundaries.
1866 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1867 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1869 // The exit count might have the type of i64 while the phi is i32. This can
1870 // happen if we have an induction variable that is sign extended before the
1871 // compare. The only way that we get a backedge taken count is that the
1872 // induction variable was signed and as such will not overflow. In such a case
1873 // truncation is legal.
1874 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1875 IdxTy->getPrimitiveSizeInBits())
1876 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1878 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1879 // Get the total trip count from the count by adding 1.
1880 ExitCount = SE->getAddExpr(ExitCount,
1881 SE->getConstant(ExitCount->getType(), 1));
1883 // Expand the trip count and place the new instructions in the preheader.
1884 // Notice that the pre-header does not change, only the loop body.
1885 SCEVExpander Exp(*SE, "induction");
1887 // Count holds the overall loop count (N).
1888 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1889 BypassBlock->getTerminator());
1891 // The loop index does not have to start at Zero. Find the original start
1892 // value from the induction PHI node. If we don't have an induction variable
1893 // then we know that it starts at zero.
1894 Builder.SetInsertPoint(BypassBlock->getTerminator());
1895 Value *StartIdx = ExtendedIdx = OldInduction ?
1896 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1898 ConstantInt::get(IdxTy, 0);
1900 assert(BypassBlock && "Invalid loop structure");
1901 LoopBypassBlocks.push_back(BypassBlock);
1903 // Split the single block loop into the two loop structure described above.
1904 BasicBlock *VectorPH =
1905 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1906 BasicBlock *VecBody =
1907 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1908 BasicBlock *MiddleBlock =
1909 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1910 BasicBlock *ScalarPH =
1911 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1913 // Create and register the new vector loop.
1914 Loop* Lp = new Loop();
1915 Loop *ParentLoop = OrigLoop->getParentLoop();
1917 // Insert the new loop into the loop nest and register the new basic blocks
1918 // before calling any utilities such as SCEV that require valid LoopInfo.
1920 ParentLoop->addChildLoop(Lp);
1921 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1922 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1923 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1925 LI->addTopLevelLoop(Lp);
1927 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1929 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1931 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1933 // Generate the induction variable.
1934 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1935 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1936 // The loop step is equal to the vectorization factor (num of SIMD elements)
1937 // times the unroll factor (num of SIMD instructions).
1938 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1940 // This is the IR builder that we use to add all of the logic for bypassing
1941 // the new vector loop.
1942 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1943 setDebugLocFromInst(BypassBuilder,
1944 getDebugLocFromInstOrOperands(OldInduction));
1946 // We may need to extend the index in case there is a type mismatch.
1947 // We know that the count starts at zero and does not overflow.
1948 if (Count->getType() != IdxTy) {
1949 // The exit count can be of pointer type. Convert it to the correct
1951 if (ExitCount->getType()->isPointerTy())
1952 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1954 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1957 // Add the start index to the loop count to get the new end index.
1958 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1960 // Now we need to generate the expression for N - (N % VF), which is
1961 // the part that the vectorized body will execute.
1962 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1963 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1964 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1965 "end.idx.rnd.down");
1967 // Now, compare the new count to zero. If it is zero skip the vector loop and
1968 // jump to the scalar loop.
1969 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1972 BasicBlock *LastBypassBlock = BypassBlock;
1974 // Generate the code to check that the strides we assumed to be one are really
1975 // one. We want the new basic block to start at the first instruction in a
1976 // sequence of instructions that form a check.
1977 Instruction *StrideCheck;
1978 Instruction *FirstCheckInst;
1979 tie(FirstCheckInst, StrideCheck) =
1980 addStrideCheck(BypassBlock->getTerminator());
1982 // Create a new block containing the stride check.
1983 BasicBlock *CheckBlock =
1984 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
1986 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1987 LoopBypassBlocks.push_back(CheckBlock);
1989 // Replace the branch into the memory check block with a conditional branch
1990 // for the "few elements case".
1991 Instruction *OldTerm = BypassBlock->getTerminator();
1992 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1993 OldTerm->eraseFromParent();
1996 LastBypassBlock = CheckBlock;
1999 // Generate the code that checks in runtime if arrays overlap. We put the
2000 // checks into a separate block to make the more common case of few elements
2002 Instruction *MemRuntimeCheck;
2003 tie(FirstCheckInst, MemRuntimeCheck) =
2004 addRuntimeCheck(LastBypassBlock->getTerminator());
2005 if (MemRuntimeCheck) {
2006 // Create a new block containing the memory check.
2007 BasicBlock *CheckBlock =
2008 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2010 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2011 LoopBypassBlocks.push_back(CheckBlock);
2013 // Replace the branch into the memory check block with a conditional branch
2014 // for the "few elements case".
2015 Instruction *OldTerm = LastBypassBlock->getTerminator();
2016 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2017 OldTerm->eraseFromParent();
2019 Cmp = MemRuntimeCheck;
2020 LastBypassBlock = CheckBlock;
2023 LastBypassBlock->getTerminator()->eraseFromParent();
2024 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2027 // We are going to resume the execution of the scalar loop.
2028 // Go over all of the induction variables that we found and fix the
2029 // PHIs that are left in the scalar version of the loop.
2030 // The starting values of PHI nodes depend on the counter of the last
2031 // iteration in the vectorized loop.
2032 // If we come from a bypass edge then we need to start from the original
2035 // This variable saves the new starting index for the scalar loop.
2036 PHINode *ResumeIndex = 0;
2037 LoopVectorizationLegality::InductionList::iterator I, E;
2038 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2039 // Set builder to point to last bypass block.
2040 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2041 for (I = List->begin(), E = List->end(); I != E; ++I) {
2042 PHINode *OrigPhi = I->first;
2043 LoopVectorizationLegality::InductionInfo II = I->second;
2045 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2046 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2047 MiddleBlock->getTerminator());
2048 // We might have extended the type of the induction variable but we need a
2049 // truncated version for the scalar loop.
2050 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2051 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2052 MiddleBlock->getTerminator()) : 0;
2054 Value *EndValue = 0;
2056 case LoopVectorizationLegality::IK_NoInduction:
2057 llvm_unreachable("Unknown induction");
2058 case LoopVectorizationLegality::IK_IntInduction: {
2059 // Handle the integer induction counter.
2060 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2062 // We have the canonical induction variable.
2063 if (OrigPhi == OldInduction) {
2064 // Create a truncated version of the resume value for the scalar loop,
2065 // we might have promoted the type to a larger width.
2067 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2068 // The new PHI merges the original incoming value, in case of a bypass,
2069 // or the value at the end of the vectorized loop.
2070 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2071 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2072 TruncResumeVal->addIncoming(EndValue, VecBody);
2074 // We know what the end value is.
2075 EndValue = IdxEndRoundDown;
2076 // We also know which PHI node holds it.
2077 ResumeIndex = ResumeVal;
2081 // Not the canonical induction variable - add the vector loop count to the
2083 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2084 II.StartValue->getType(),
2086 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2089 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2090 // Convert the CountRoundDown variable to the PHI size.
2091 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2092 II.StartValue->getType(),
2094 // Handle reverse integer induction counter.
2095 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2098 case LoopVectorizationLegality::IK_PtrInduction: {
2099 // For pointer induction variables, calculate the offset using
2101 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2105 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2106 // The value at the end of the loop for the reverse pointer is calculated
2107 // by creating a GEP with a negative index starting from the start value.
2108 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2109 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2111 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2117 // The new PHI merges the original incoming value, in case of a bypass,
2118 // or the value at the end of the vectorized loop.
2119 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2120 if (OrigPhi == OldInduction)
2121 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2123 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2125 ResumeVal->addIncoming(EndValue, VecBody);
2127 // Fix the scalar body counter (PHI node).
2128 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2129 // The old inductions phi node in the scalar body needs the truncated value.
2130 if (OrigPhi == OldInduction)
2131 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2133 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2136 // If we are generating a new induction variable then we also need to
2137 // generate the code that calculates the exit value. This value is not
2138 // simply the end of the counter because we may skip the vectorized body
2139 // in case of a runtime check.
2141 assert(!ResumeIndex && "Unexpected resume value found");
2142 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2143 MiddleBlock->getTerminator());
2144 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2145 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2146 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2149 // Make sure that we found the index where scalar loop needs to continue.
2150 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2151 "Invalid resume Index");
2153 // Add a check in the middle block to see if we have completed
2154 // all of the iterations in the first vector loop.
2155 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2156 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2157 ResumeIndex, "cmp.n",
2158 MiddleBlock->getTerminator());
2160 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2161 // Remove the old terminator.
2162 MiddleBlock->getTerminator()->eraseFromParent();
2164 // Create i+1 and fill the PHINode.
2165 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2166 Induction->addIncoming(StartIdx, VectorPH);
2167 Induction->addIncoming(NextIdx, VecBody);
2168 // Create the compare.
2169 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2170 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2172 // Now we have two terminators. Remove the old one from the block.
2173 VecBody->getTerminator()->eraseFromParent();
2175 // Get ready to start creating new instructions into the vectorized body.
2176 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2179 LoopVectorPreHeader = VectorPH;
2180 LoopScalarPreHeader = ScalarPH;
2181 LoopMiddleBlock = MiddleBlock;
2182 LoopExitBlock = ExitBlock;
2183 LoopVectorBody.push_back(VecBody);
2184 LoopScalarBody = OldBasicBlock;
2186 LoopVectorizeHints Hints(Lp, true);
2187 Hints.setAlreadyVectorized(Lp);
2190 /// This function returns the identity element (or neutral element) for
2191 /// the operation K.
2193 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2198 // Adding, Xoring, Oring zero to a number does not change it.
2199 return ConstantInt::get(Tp, 0);
2200 case RK_IntegerMult:
2201 // Multiplying a number by 1 does not change it.
2202 return ConstantInt::get(Tp, 1);
2204 // AND-ing a number with an all-1 value does not change it.
2205 return ConstantInt::get(Tp, -1, true);
2207 // Multiplying a number by 1 does not change it.
2208 return ConstantFP::get(Tp, 1.0L);
2210 // Adding zero to a number does not change it.
2211 return ConstantFP::get(Tp, 0.0L);
2213 llvm_unreachable("Unknown reduction kind");
2217 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2218 Intrinsic::ID ValidIntrinsicID) {
2219 if (I.getNumArgOperands() != 1 ||
2220 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2221 I.getType() != I.getArgOperand(0)->getType() ||
2222 !I.onlyReadsMemory())
2223 return Intrinsic::not_intrinsic;
2225 return ValidIntrinsicID;
2228 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2229 Intrinsic::ID ValidIntrinsicID) {
2230 if (I.getNumArgOperands() != 2 ||
2231 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2232 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2233 I.getType() != I.getArgOperand(0)->getType() ||
2234 I.getType() != I.getArgOperand(1)->getType() ||
2235 !I.onlyReadsMemory())
2236 return Intrinsic::not_intrinsic;
2238 return ValidIntrinsicID;
2242 static Intrinsic::ID
2243 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2244 // If we have an intrinsic call, check if it is trivially vectorizable.
2245 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2246 switch (II->getIntrinsicID()) {
2247 case Intrinsic::sqrt:
2248 case Intrinsic::sin:
2249 case Intrinsic::cos:
2250 case Intrinsic::exp:
2251 case Intrinsic::exp2:
2252 case Intrinsic::log:
2253 case Intrinsic::log10:
2254 case Intrinsic::log2:
2255 case Intrinsic::fabs:
2256 case Intrinsic::copysign:
2257 case Intrinsic::floor:
2258 case Intrinsic::ceil:
2259 case Intrinsic::trunc:
2260 case Intrinsic::rint:
2261 case Intrinsic::nearbyint:
2262 case Intrinsic::round:
2263 case Intrinsic::pow:
2264 case Intrinsic::fma:
2265 case Intrinsic::fmuladd:
2266 case Intrinsic::lifetime_start:
2267 case Intrinsic::lifetime_end:
2268 return II->getIntrinsicID();
2270 return Intrinsic::not_intrinsic;
2275 return Intrinsic::not_intrinsic;
2278 Function *F = CI->getCalledFunction();
2279 // We're going to make assumptions on the semantics of the functions, check
2280 // that the target knows that it's available in this environment and it does
2281 // not have local linkage.
2282 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2283 return Intrinsic::not_intrinsic;
2285 // Otherwise check if we have a call to a function that can be turned into a
2286 // vector intrinsic.
2293 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2297 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2301 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2303 case LibFunc::exp2f:
2304 case LibFunc::exp2l:
2305 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2309 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2310 case LibFunc::log10:
2311 case LibFunc::log10f:
2312 case LibFunc::log10l:
2313 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2315 case LibFunc::log2f:
2316 case LibFunc::log2l:
2317 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2319 case LibFunc::fabsf:
2320 case LibFunc::fabsl:
2321 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2322 case LibFunc::copysign:
2323 case LibFunc::copysignf:
2324 case LibFunc::copysignl:
2325 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2326 case LibFunc::floor:
2327 case LibFunc::floorf:
2328 case LibFunc::floorl:
2329 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2331 case LibFunc::ceilf:
2332 case LibFunc::ceill:
2333 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2334 case LibFunc::trunc:
2335 case LibFunc::truncf:
2336 case LibFunc::truncl:
2337 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2339 case LibFunc::rintf:
2340 case LibFunc::rintl:
2341 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2342 case LibFunc::nearbyint:
2343 case LibFunc::nearbyintf:
2344 case LibFunc::nearbyintl:
2345 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2346 case LibFunc::round:
2347 case LibFunc::roundf:
2348 case LibFunc::roundl:
2349 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2353 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2356 return Intrinsic::not_intrinsic;
2359 /// This function translates the reduction kind to an LLVM binary operator.
2361 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2363 case LoopVectorizationLegality::RK_IntegerAdd:
2364 return Instruction::Add;
2365 case LoopVectorizationLegality::RK_IntegerMult:
2366 return Instruction::Mul;
2367 case LoopVectorizationLegality::RK_IntegerOr:
2368 return Instruction::Or;
2369 case LoopVectorizationLegality::RK_IntegerAnd:
2370 return Instruction::And;
2371 case LoopVectorizationLegality::RK_IntegerXor:
2372 return Instruction::Xor;
2373 case LoopVectorizationLegality::RK_FloatMult:
2374 return Instruction::FMul;
2375 case LoopVectorizationLegality::RK_FloatAdd:
2376 return Instruction::FAdd;
2377 case LoopVectorizationLegality::RK_IntegerMinMax:
2378 return Instruction::ICmp;
2379 case LoopVectorizationLegality::RK_FloatMinMax:
2380 return Instruction::FCmp;
2382 llvm_unreachable("Unknown reduction operation");
2386 Value *createMinMaxOp(IRBuilder<> &Builder,
2387 LoopVectorizationLegality::MinMaxReductionKind RK,
2390 CmpInst::Predicate P = CmpInst::ICMP_NE;
2393 llvm_unreachable("Unknown min/max reduction kind");
2394 case LoopVectorizationLegality::MRK_UIntMin:
2395 P = CmpInst::ICMP_ULT;
2397 case LoopVectorizationLegality::MRK_UIntMax:
2398 P = CmpInst::ICMP_UGT;
2400 case LoopVectorizationLegality::MRK_SIntMin:
2401 P = CmpInst::ICMP_SLT;
2403 case LoopVectorizationLegality::MRK_SIntMax:
2404 P = CmpInst::ICMP_SGT;
2406 case LoopVectorizationLegality::MRK_FloatMin:
2407 P = CmpInst::FCMP_OLT;
2409 case LoopVectorizationLegality::MRK_FloatMax:
2410 P = CmpInst::FCMP_OGT;
2415 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2416 RK == LoopVectorizationLegality::MRK_FloatMax)
2417 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2419 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2421 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2426 struct CSEDenseMapInfo {
2427 static bool canHandle(Instruction *I) {
2428 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2429 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2431 static inline Instruction *getEmptyKey() {
2432 return DenseMapInfo<Instruction *>::getEmptyKey();
2434 static inline Instruction *getTombstoneKey() {
2435 return DenseMapInfo<Instruction *>::getTombstoneKey();
2437 static unsigned getHashValue(Instruction *I) {
2438 assert(canHandle(I) && "Unknown instruction!");
2439 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2440 I->value_op_end()));
2442 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2443 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2444 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2446 return LHS->isIdenticalTo(RHS);
2451 /// \brief Check whether this block is a predicated block.
2452 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2453 /// = ...; " blocks. We start with one vectorized basic block. For every
2454 /// conditional block we split this vectorized block. Therefore, every second
2455 /// block will be a predicated one.
2456 static bool isPredicatedBlock(unsigned BlockNum) {
2457 return BlockNum % 2;
2460 ///\brief Perform cse of induction variable instructions.
2461 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2462 // Perform simple cse.
2463 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2464 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2465 BasicBlock *BB = BBs[i];
2466 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2467 Instruction *In = I++;
2469 if (!CSEDenseMapInfo::canHandle(In))
2472 // Check if we can replace this instruction with any of the
2473 // visited instructions.
2474 if (Instruction *V = CSEMap.lookup(In)) {
2475 In->replaceAllUsesWith(V);
2476 In->eraseFromParent();
2479 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2480 // ...;" blocks for predicated stores. Every second block is a predicated
2482 if (isPredicatedBlock(i))
2490 void InnerLoopVectorizer::vectorizeLoop() {
2491 //===------------------------------------------------===//
2493 // Notice: any optimization or new instruction that go
2494 // into the code below should be also be implemented in
2497 //===------------------------------------------------===//
2498 Constant *Zero = Builder.getInt32(0);
2500 // In order to support reduction variables we need to be able to vectorize
2501 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2502 // stages. First, we create a new vector PHI node with no incoming edges.
2503 // We use this value when we vectorize all of the instructions that use the
2504 // PHI. Next, after all of the instructions in the block are complete we
2505 // add the new incoming edges to the PHI. At this point all of the
2506 // instructions in the basic block are vectorized, so we can use them to
2507 // construct the PHI.
2508 PhiVector RdxPHIsToFix;
2510 // Scan the loop in a topological order to ensure that defs are vectorized
2512 LoopBlocksDFS DFS(OrigLoop);
2515 // Vectorize all of the blocks in the original loop.
2516 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2517 be = DFS.endRPO(); bb != be; ++bb)
2518 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2520 // At this point every instruction in the original loop is widened to
2521 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2522 // that we vectorized. The PHI nodes are currently empty because we did
2523 // not want to introduce cycles. Notice that the remaining PHI nodes
2524 // that we need to fix are reduction variables.
2526 // Create the 'reduced' values for each of the induction vars.
2527 // The reduced values are the vector values that we scalarize and combine
2528 // after the loop is finished.
2529 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2531 PHINode *RdxPhi = *it;
2532 assert(RdxPhi && "Unable to recover vectorized PHI");
2534 // Find the reduction variable descriptor.
2535 assert(Legal->getReductionVars()->count(RdxPhi) &&
2536 "Unable to find the reduction variable");
2537 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2538 (*Legal->getReductionVars())[RdxPhi];
2540 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2542 // We need to generate a reduction vector from the incoming scalar.
2543 // To do so, we need to generate the 'identity' vector and override
2544 // one of the elements with the incoming scalar reduction. We need
2545 // to do it in the vector-loop preheader.
2546 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2548 // This is the vector-clone of the value that leaves the loop.
2549 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2550 Type *VecTy = VectorExit[0]->getType();
2552 // Find the reduction identity variable. Zero for addition, or, xor,
2553 // one for multiplication, -1 for And.
2556 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2557 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2558 // MinMax reduction have the start value as their identify.
2560 VectorStart = Identity = RdxDesc.StartValue;
2562 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2567 // Handle other reduction kinds:
2569 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2570 VecTy->getScalarType());
2573 // This vector is the Identity vector where the first element is the
2574 // incoming scalar reduction.
2575 VectorStart = RdxDesc.StartValue;
2577 Identity = ConstantVector::getSplat(VF, Iden);
2579 // This vector is the Identity vector where the first element is the
2580 // incoming scalar reduction.
2581 VectorStart = Builder.CreateInsertElement(Identity,
2582 RdxDesc.StartValue, Zero);
2586 // Fix the vector-loop phi.
2587 // We created the induction variable so we know that the
2588 // preheader is the first entry.
2589 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2591 // Reductions do not have to start at zero. They can start with
2592 // any loop invariant values.
2593 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2594 BasicBlock *Latch = OrigLoop->getLoopLatch();
2595 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2596 VectorParts &Val = getVectorValue(LoopVal);
2597 for (unsigned part = 0; part < UF; ++part) {
2598 // Make sure to add the reduction stat value only to the
2599 // first unroll part.
2600 Value *StartVal = (part == 0) ? VectorStart : Identity;
2601 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2602 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2603 LoopVectorBody.back());
2606 // Before each round, move the insertion point right between
2607 // the PHIs and the values we are going to write.
2608 // This allows us to write both PHINodes and the extractelement
2610 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2612 VectorParts RdxParts;
2613 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2614 for (unsigned part = 0; part < UF; ++part) {
2615 // This PHINode contains the vectorized reduction variable, or
2616 // the initial value vector, if we bypass the vector loop.
2617 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2618 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2619 Value *StartVal = (part == 0) ? VectorStart : Identity;
2620 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2621 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2622 NewPhi->addIncoming(RdxExitVal[part],
2623 LoopVectorBody.back());
2624 RdxParts.push_back(NewPhi);
2627 // Reduce all of the unrolled parts into a single vector.
2628 Value *ReducedPartRdx = RdxParts[0];
2629 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2630 setDebugLocFromInst(Builder, ReducedPartRdx);
2631 for (unsigned part = 1; part < UF; ++part) {
2632 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2633 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2634 RdxParts[part], ReducedPartRdx,
2637 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2638 ReducedPartRdx, RdxParts[part]);
2642 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2643 // and vector ops, reducing the set of values being computed by half each
2645 assert(isPowerOf2_32(VF) &&
2646 "Reduction emission only supported for pow2 vectors!");
2647 Value *TmpVec = ReducedPartRdx;
2648 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2649 for (unsigned i = VF; i != 1; i >>= 1) {
2650 // Move the upper half of the vector to the lower half.
2651 for (unsigned j = 0; j != i/2; ++j)
2652 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2654 // Fill the rest of the mask with undef.
2655 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2656 UndefValue::get(Builder.getInt32Ty()));
2659 Builder.CreateShuffleVector(TmpVec,
2660 UndefValue::get(TmpVec->getType()),
2661 ConstantVector::get(ShuffleMask),
2664 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2665 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2668 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2671 // The result is in the first element of the vector.
2672 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2673 Builder.getInt32(0));
2676 // Now, we need to fix the users of the reduction variable
2677 // inside and outside of the scalar remainder loop.
2678 // We know that the loop is in LCSSA form. We need to update the
2679 // PHI nodes in the exit blocks.
2680 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2681 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2682 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2683 if (!LCSSAPhi) break;
2685 // All PHINodes need to have a single entry edge, or two if
2686 // we already fixed them.
2687 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2689 // We found our reduction value exit-PHI. Update it with the
2690 // incoming bypass edge.
2691 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2692 // Add an edge coming from the bypass.
2693 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2696 }// end of the LCSSA phi scan.
2698 // Fix the scalar loop reduction variable with the incoming reduction sum
2699 // from the vector body and from the backedge value.
2700 int IncomingEdgeBlockIdx =
2701 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2702 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2703 // Pick the other block.
2704 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2705 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2706 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2707 }// end of for each redux variable.
2711 // Remove redundant induction instructions.
2712 cse(LoopVectorBody);
2715 void InnerLoopVectorizer::fixLCSSAPHIs() {
2716 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2717 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2718 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2719 if (!LCSSAPhi) break;
2720 if (LCSSAPhi->getNumIncomingValues() == 1)
2721 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2726 InnerLoopVectorizer::VectorParts
2727 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2728 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2731 // Look for cached value.
2732 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2733 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2734 if (ECEntryIt != MaskCache.end())
2735 return ECEntryIt->second;
2737 VectorParts SrcMask = createBlockInMask(Src);
2739 // The terminator has to be a branch inst!
2740 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2741 assert(BI && "Unexpected terminator found");
2743 if (BI->isConditional()) {
2744 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2746 if (BI->getSuccessor(0) != Dst)
2747 for (unsigned part = 0; part < UF; ++part)
2748 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2750 for (unsigned part = 0; part < UF; ++part)
2751 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2753 MaskCache[Edge] = EdgeMask;
2757 MaskCache[Edge] = SrcMask;
2761 InnerLoopVectorizer::VectorParts
2762 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2763 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2765 // Loop incoming mask is all-one.
2766 if (OrigLoop->getHeader() == BB) {
2767 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2768 return getVectorValue(C);
2771 // This is the block mask. We OR all incoming edges, and with zero.
2772 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2773 VectorParts BlockMask = getVectorValue(Zero);
2776 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2777 VectorParts EM = createEdgeMask(*it, BB);
2778 for (unsigned part = 0; part < UF; ++part)
2779 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2785 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2786 InnerLoopVectorizer::VectorParts &Entry,
2787 unsigned UF, unsigned VF, PhiVector *PV) {
2788 PHINode* P = cast<PHINode>(PN);
2789 // Handle reduction variables:
2790 if (Legal->getReductionVars()->count(P)) {
2791 for (unsigned part = 0; part < UF; ++part) {
2792 // This is phase one of vectorizing PHIs.
2793 Type *VecTy = (VF == 1) ? PN->getType() :
2794 VectorType::get(PN->getType(), VF);
2795 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2796 LoopVectorBody.back()-> getFirstInsertionPt());
2802 setDebugLocFromInst(Builder, P);
2803 // Check for PHI nodes that are lowered to vector selects.
2804 if (P->getParent() != OrigLoop->getHeader()) {
2805 // We know that all PHIs in non-header blocks are converted into
2806 // selects, so we don't have to worry about the insertion order and we
2807 // can just use the builder.
2808 // At this point we generate the predication tree. There may be
2809 // duplications since this is a simple recursive scan, but future
2810 // optimizations will clean it up.
2812 unsigned NumIncoming = P->getNumIncomingValues();
2814 // Generate a sequence of selects of the form:
2815 // SELECT(Mask3, In3,
2816 // SELECT(Mask2, In2,
2818 for (unsigned In = 0; In < NumIncoming; In++) {
2819 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2821 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2823 for (unsigned part = 0; part < UF; ++part) {
2824 // We might have single edge PHIs (blocks) - use an identity
2825 // 'select' for the first PHI operand.
2827 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2830 // Select between the current value and the previous incoming edge
2831 // based on the incoming mask.
2832 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2833 Entry[part], "predphi");
2839 // This PHINode must be an induction variable.
2840 // Make sure that we know about it.
2841 assert(Legal->getInductionVars()->count(P) &&
2842 "Not an induction variable");
2844 LoopVectorizationLegality::InductionInfo II =
2845 Legal->getInductionVars()->lookup(P);
2848 case LoopVectorizationLegality::IK_NoInduction:
2849 llvm_unreachable("Unknown induction");
2850 case LoopVectorizationLegality::IK_IntInduction: {
2851 assert(P->getType() == II.StartValue->getType() && "Types must match");
2852 Type *PhiTy = P->getType();
2854 if (P == OldInduction) {
2855 // Handle the canonical induction variable. We might have had to
2857 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2859 // Handle other induction variables that are now based on the
2861 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2863 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2864 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2867 Broadcasted = getBroadcastInstrs(Broadcasted);
2868 // After broadcasting the induction variable we need to make the vector
2869 // consecutive by adding 0, 1, 2, etc.
2870 for (unsigned part = 0; part < UF; ++part)
2871 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2874 case LoopVectorizationLegality::IK_ReverseIntInduction:
2875 case LoopVectorizationLegality::IK_PtrInduction:
2876 case LoopVectorizationLegality::IK_ReversePtrInduction:
2877 // Handle reverse integer and pointer inductions.
2878 Value *StartIdx = ExtendedIdx;
2879 // This is the normalized GEP that starts counting at zero.
2880 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2883 // Handle the reverse integer induction variable case.
2884 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2885 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2886 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2888 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2891 // This is a new value so do not hoist it out.
2892 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2893 // After broadcasting the induction variable we need to make the
2894 // vector consecutive by adding ... -3, -2, -1, 0.
2895 for (unsigned part = 0; part < UF; ++part)
2896 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2901 // Handle the pointer induction variable case.
2902 assert(P->getType()->isPointerTy() && "Unexpected type.");
2904 // Is this a reverse induction ptr or a consecutive induction ptr.
2905 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2908 // This is the vector of results. Notice that we don't generate
2909 // vector geps because scalar geps result in better code.
2910 for (unsigned part = 0; part < UF; ++part) {
2912 int EltIndex = (part) * (Reverse ? -1 : 1);
2913 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2916 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2918 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2920 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2922 Entry[part] = SclrGep;
2926 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2927 for (unsigned int i = 0; i < VF; ++i) {
2928 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2929 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2932 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2934 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2936 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2938 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2939 Builder.getInt32(i),
2942 Entry[part] = VecVal;
2948 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2949 // For each instruction in the old loop.
2950 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2951 VectorParts &Entry = WidenMap.get(it);
2952 switch (it->getOpcode()) {
2953 case Instruction::Br:
2954 // Nothing to do for PHIs and BR, since we already took care of the
2955 // loop control flow instructions.
2957 case Instruction::PHI:{
2958 // Vectorize PHINodes.
2959 widenPHIInstruction(it, Entry, UF, VF, PV);
2963 case Instruction::Add:
2964 case Instruction::FAdd:
2965 case Instruction::Sub:
2966 case Instruction::FSub:
2967 case Instruction::Mul:
2968 case Instruction::FMul:
2969 case Instruction::UDiv:
2970 case Instruction::SDiv:
2971 case Instruction::FDiv:
2972 case Instruction::URem:
2973 case Instruction::SRem:
2974 case Instruction::FRem:
2975 case Instruction::Shl:
2976 case Instruction::LShr:
2977 case Instruction::AShr:
2978 case Instruction::And:
2979 case Instruction::Or:
2980 case Instruction::Xor: {
2981 // Just widen binops.
2982 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2983 setDebugLocFromInst(Builder, BinOp);
2984 VectorParts &A = getVectorValue(it->getOperand(0));
2985 VectorParts &B = getVectorValue(it->getOperand(1));
2987 // Use this vector value for all users of the original instruction.
2988 for (unsigned Part = 0; Part < UF; ++Part) {
2989 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2991 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2992 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2993 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2994 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2995 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2997 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2998 VecOp->setIsExact(BinOp->isExact());
3004 case Instruction::Select: {
3006 // If the selector is loop invariant we can create a select
3007 // instruction with a scalar condition. Otherwise, use vector-select.
3008 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3010 setDebugLocFromInst(Builder, it);
3012 // The condition can be loop invariant but still defined inside the
3013 // loop. This means that we can't just use the original 'cond' value.
3014 // We have to take the 'vectorized' value and pick the first lane.
3015 // Instcombine will make this a no-op.
3016 VectorParts &Cond = getVectorValue(it->getOperand(0));
3017 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3018 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3020 Value *ScalarCond = (VF == 1) ? Cond[0] :
3021 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3023 for (unsigned Part = 0; Part < UF; ++Part) {
3024 Entry[Part] = Builder.CreateSelect(
3025 InvariantCond ? ScalarCond : Cond[Part],
3032 case Instruction::ICmp:
3033 case Instruction::FCmp: {
3034 // Widen compares. Generate vector compares.
3035 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3036 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3037 setDebugLocFromInst(Builder, it);
3038 VectorParts &A = getVectorValue(it->getOperand(0));
3039 VectorParts &B = getVectorValue(it->getOperand(1));
3040 for (unsigned Part = 0; Part < UF; ++Part) {
3043 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3045 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3051 case Instruction::Store:
3052 case Instruction::Load:
3053 vectorizeMemoryInstruction(it);
3055 case Instruction::ZExt:
3056 case Instruction::SExt:
3057 case Instruction::FPToUI:
3058 case Instruction::FPToSI:
3059 case Instruction::FPExt:
3060 case Instruction::PtrToInt:
3061 case Instruction::IntToPtr:
3062 case Instruction::SIToFP:
3063 case Instruction::UIToFP:
3064 case Instruction::Trunc:
3065 case Instruction::FPTrunc:
3066 case Instruction::BitCast: {
3067 CastInst *CI = dyn_cast<CastInst>(it);
3068 setDebugLocFromInst(Builder, it);
3069 /// Optimize the special case where the source is the induction
3070 /// variable. Notice that we can only optimize the 'trunc' case
3071 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3072 /// c. other casts depend on pointer size.
3073 if (CI->getOperand(0) == OldInduction &&
3074 it->getOpcode() == Instruction::Trunc) {
3075 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3077 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3078 for (unsigned Part = 0; Part < UF; ++Part)
3079 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3082 /// Vectorize casts.
3083 Type *DestTy = (VF == 1) ? CI->getType() :
3084 VectorType::get(CI->getType(), VF);
3086 VectorParts &A = getVectorValue(it->getOperand(0));
3087 for (unsigned Part = 0; Part < UF; ++Part)
3088 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3092 case Instruction::Call: {
3093 // Ignore dbg intrinsics.
3094 if (isa<DbgInfoIntrinsic>(it))
3096 setDebugLocFromInst(Builder, it);
3098 Module *M = BB->getParent()->getParent();
3099 CallInst *CI = cast<CallInst>(it);
3100 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3101 assert(ID && "Not an intrinsic call!");
3103 case Intrinsic::lifetime_end:
3104 case Intrinsic::lifetime_start:
3105 scalarizeInstruction(it);
3108 for (unsigned Part = 0; Part < UF; ++Part) {
3109 SmallVector<Value *, 4> Args;
3110 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3111 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3112 Args.push_back(Arg[Part]);
3114 Type *Tys[] = {CI->getType()};
3116 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3118 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3119 Entry[Part] = Builder.CreateCall(F, Args);
3127 // All other instructions are unsupported. Scalarize them.
3128 scalarizeInstruction(it);
3131 }// end of for_each instr.
3134 void InnerLoopVectorizer::updateAnalysis() {
3135 // Forget the original basic block.
3136 SE->forgetLoop(OrigLoop);
3138 // Update the dominator tree information.
3139 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3140 "Entry does not dominate exit.");
3142 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3143 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3144 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3146 // Due to if predication of stores we might create a sequence of "if(pred)
3147 // a[i] = ...; " blocks.
3148 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3150 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3151 else if (isPredicatedBlock(i)) {
3152 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3154 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3158 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3159 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3160 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3161 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3163 DEBUG(DT->verifyDomTree());
3166 /// \brief Check whether it is safe to if-convert this phi node.
3168 /// Phi nodes with constant expressions that can trap are not safe to if
3170 static bool canIfConvertPHINodes(BasicBlock *BB) {
3171 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3172 PHINode *Phi = dyn_cast<PHINode>(I);
3175 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3176 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3183 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3184 if (!EnableIfConversion)
3187 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3189 // A list of pointers that we can safely read and write to.
3190 SmallPtrSet<Value *, 8> SafePointes;
3192 // Collect safe addresses.
3193 for (Loop::block_iterator BI = TheLoop->block_begin(),
3194 BE = TheLoop->block_end(); BI != BE; ++BI) {
3195 BasicBlock *BB = *BI;
3197 if (blockNeedsPredication(BB))
3200 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3201 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3202 SafePointes.insert(LI->getPointerOperand());
3203 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3204 SafePointes.insert(SI->getPointerOperand());
3208 // Collect the blocks that need predication.
3209 BasicBlock *Header = TheLoop->getHeader();
3210 for (Loop::block_iterator BI = TheLoop->block_begin(),
3211 BE = TheLoop->block_end(); BI != BE; ++BI) {
3212 BasicBlock *BB = *BI;
3214 // We don't support switch statements inside loops.
3215 if (!isa<BranchInst>(BB->getTerminator()))
3218 // We must be able to predicate all blocks that need to be predicated.
3219 if (blockNeedsPredication(BB)) {
3220 if (!blockCanBePredicated(BB, SafePointes))
3222 } else if (BB != Header && !canIfConvertPHINodes(BB))
3227 // We can if-convert this loop.
3231 bool LoopVectorizationLegality::canVectorize() {
3232 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3233 // be canonicalized.
3234 if (!TheLoop->getLoopPreheader())
3237 // We can only vectorize innermost loops.
3238 if (TheLoop->getSubLoopsVector().size())
3241 // We must have a single backedge.
3242 if (TheLoop->getNumBackEdges() != 1)
3245 // We must have a single exiting block.
3246 if (!TheLoop->getExitingBlock())
3249 // We need to have a loop header.
3250 DEBUG(dbgs() << "LV: Found a loop: " <<
3251 TheLoop->getHeader()->getName() << '\n');
3253 // Check if we can if-convert non-single-bb loops.
3254 unsigned NumBlocks = TheLoop->getNumBlocks();
3255 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3256 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3260 // ScalarEvolution needs to be able to find the exit count.
3261 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3262 if (ExitCount == SE->getCouldNotCompute()) {
3263 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3267 // Do not loop-vectorize loops with a tiny trip count.
3268 BasicBlock *Latch = TheLoop->getLoopLatch();
3269 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
3270 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
3271 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
3272 "This loop is not worth vectorizing.\n");
3276 // Check if we can vectorize the instructions and CFG in this loop.
3277 if (!canVectorizeInstrs()) {
3278 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3282 // Go over each instruction and look at memory deps.
3283 if (!canVectorizeMemory()) {
3284 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3288 // Collect all of the variables that remain uniform after vectorization.
3289 collectLoopUniforms();
3291 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3292 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3295 // Okay! We can vectorize. At this point we don't have any other mem analysis
3296 // which may limit our maximum vectorization factor, so just return true with
3301 static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
3302 if (Ty->isPointerTy())
3303 return DL.getIntPtrType(Ty);
3305 // It is possible that char's or short's overflow when we ask for the loop's
3306 // trip count, work around this by changing the type size.
3307 if (Ty->getScalarSizeInBits() < 32)
3308 return Type::getInt32Ty(Ty->getContext());
3313 static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
3314 Ty0 = convertPointerToIntegerType(DL, Ty0);
3315 Ty1 = convertPointerToIntegerType(DL, Ty1);
3316 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3321 /// \brief Check that the instruction has outside loop users and is not an
3322 /// identified reduction variable.
3323 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3324 SmallPtrSet<Value *, 4> &Reductions) {
3325 // Reduction instructions are allowed to have exit users. All other
3326 // instructions must not have external users.
3327 if (!Reductions.count(Inst))
3328 //Check that all of the users of the loop are inside the BB.
3329 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
3331 Instruction *U = cast<Instruction>(*I);
3332 // This user may be a reduction exit value.
3333 if (!TheLoop->contains(U)) {
3334 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
3341 bool LoopVectorizationLegality::canVectorizeInstrs() {
3342 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3343 BasicBlock *Header = TheLoop->getHeader();
3345 // Look for the attribute signaling the absence of NaNs.
3346 Function &F = *Header->getParent();
3347 if (F.hasFnAttribute("no-nans-fp-math"))
3348 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3349 AttributeSet::FunctionIndex,
3350 "no-nans-fp-math").getValueAsString() == "true";
3352 // For each block in the loop.
3353 for (Loop::block_iterator bb = TheLoop->block_begin(),
3354 be = TheLoop->block_end(); bb != be; ++bb) {
3356 // Scan the instructions in the block and look for hazards.
3357 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3360 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3361 Type *PhiTy = Phi->getType();
3362 // Check that this PHI type is allowed.
3363 if (!PhiTy->isIntegerTy() &&
3364 !PhiTy->isFloatingPointTy() &&
3365 !PhiTy->isPointerTy()) {
3366 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3370 // If this PHINode is not in the header block, then we know that we
3371 // can convert it to select during if-conversion. No need to check if
3372 // the PHIs in this block are induction or reduction variables.
3373 if (*bb != Header) {
3374 // Check that this instruction has no outside users or is an
3375 // identified reduction value with an outside user.
3376 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3381 // We only allow if-converted PHIs with more than two incoming values.
3382 if (Phi->getNumIncomingValues() != 2) {
3383 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3387 // This is the value coming from the preheader.
3388 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3389 // Check if this is an induction variable.
3390 InductionKind IK = isInductionVariable(Phi);
3392 if (IK_NoInduction != IK) {
3393 // Get the widest type.
3395 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3397 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3399 // Int inductions are special because we only allow one IV.
3400 if (IK == IK_IntInduction) {
3401 // Use the phi node with the widest type as induction. Use the last
3402 // one if there are multiple (no good reason for doing this other
3403 // than it is expedient).
3404 if (!Induction || PhiTy == WidestIndTy)
3408 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3409 Inductions[Phi] = InductionInfo(StartValue, IK);
3411 // Until we explicitly handle the case of an induction variable with
3412 // an outside loop user we have to give up vectorizing this loop.
3413 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3419 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3420 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3423 if (AddReductionVar(Phi, RK_IntegerMult)) {
3424 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3427 if (AddReductionVar(Phi, RK_IntegerOr)) {
3428 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3431 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3432 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3435 if (AddReductionVar(Phi, RK_IntegerXor)) {
3436 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3439 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3440 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3443 if (AddReductionVar(Phi, RK_FloatMult)) {
3444 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3447 if (AddReductionVar(Phi, RK_FloatAdd)) {
3448 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3451 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3452 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3457 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3459 }// end of PHI handling
3461 // We still don't handle functions. However, we can ignore dbg intrinsic
3462 // calls and we do handle certain intrinsic and libm functions.
3463 CallInst *CI = dyn_cast<CallInst>(it);
3464 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3465 DEBUG(dbgs() << "LV: Found a call site.\n");
3469 // Check that the instruction return type is vectorizable.
3470 // Also, we can't vectorize extractelement instructions.
3471 if ((!VectorType::isValidElementType(it->getType()) &&
3472 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3473 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3477 // Check that the stored type is vectorizable.
3478 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3479 Type *T = ST->getValueOperand()->getType();
3480 if (!VectorType::isValidElementType(T))
3482 if (EnableMemAccessVersioning)
3483 collectStridedAcccess(ST);
3486 if (EnableMemAccessVersioning)
3487 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3488 collectStridedAcccess(LI);
3490 // Reduction instructions are allowed to have exit users.
3491 // All other instructions must not have external users.
3492 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3500 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3501 if (Inductions.empty())
3508 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3509 /// return the induction operand of the gep pointer.
3510 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3511 DataLayout *DL, Loop *Lp) {
3512 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3516 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3518 // Check that all of the gep indices are uniform except for our induction
3520 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3521 if (i != InductionOperand &&
3522 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3524 return GEP->getOperand(InductionOperand);
3527 ///\brief Look for a cast use of the passed value.
3528 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3529 Value *UniqueCast = 0;
3530 for (Value::use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end(); UI != UE;
3532 CastInst *CI = dyn_cast<CastInst>(*UI);
3533 if (CI && CI->getType() == Ty) {
3543 ///\brief Get the stride of a pointer access in a loop.
3544 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3545 /// pointer to the Value, or null otherwise.
3546 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3547 DataLayout *DL, Loop *Lp) {
3548 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3549 if (!PtrTy || PtrTy->isAggregateType())
3552 // Try to remove a gep instruction to make the pointer (actually index at this
3553 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3554 // pointer, otherwise, we are analyzing the index.
3555 Value *OrigPtr = Ptr;
3557 // The size of the pointer access.
3558 int64_t PtrAccessSize = 1;
3560 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3561 const SCEV *V = SE->getSCEV(Ptr);
3565 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3566 V = C->getOperand();
3568 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3572 V = S->getStepRecurrence(*SE);
3576 // Strip off the size of access multiplication if we are still analyzing the
3578 if (OrigPtr == Ptr) {
3579 DL->getTypeAllocSize(PtrTy->getElementType());
3580 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3581 if (M->getOperand(0)->getSCEVType() != scConstant)
3584 const APInt &APStepVal =
3585 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3587 // Huge step value - give up.
3588 if (APStepVal.getBitWidth() > 64)
3591 int64_t StepVal = APStepVal.getSExtValue();
3592 if (PtrAccessSize != StepVal)
3594 V = M->getOperand(1);
3599 Type *StripedOffRecurrenceCast = 0;
3600 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3601 StripedOffRecurrenceCast = C->getType();
3602 V = C->getOperand();
3605 // Look for the loop invariant symbolic value.
3606 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3610 Value *Stride = U->getValue();
3611 if (!Lp->isLoopInvariant(Stride))
3614 // If we have stripped off the recurrence cast we have to make sure that we
3615 // return the value that is used in this loop so that we can replace it later.
3616 if (StripedOffRecurrenceCast)
3617 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3622 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3624 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3625 Ptr = LI->getPointerOperand();
3626 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3627 Ptr = SI->getPointerOperand();
3631 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3635 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3636 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3637 Strides[Ptr] = Stride;
3638 StrideSet.insert(Stride);
3641 void LoopVectorizationLegality::collectLoopUniforms() {
3642 // We now know that the loop is vectorizable!
3643 // Collect variables that will remain uniform after vectorization.
3644 std::vector<Value*> Worklist;
3645 BasicBlock *Latch = TheLoop->getLoopLatch();
3647 // Start with the conditional branch and walk up the block.
3648 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3650 while (Worklist.size()) {
3651 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3652 Worklist.pop_back();
3654 // Look at instructions inside this loop.
3655 // Stop when reaching PHI nodes.
3656 // TODO: we need to follow values all over the loop, not only in this block.
3657 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3660 // This is a known uniform.
3663 // Insert all operands.
3664 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3669 /// \brief Analyses memory accesses in a loop.
3671 /// Checks whether run time pointer checks are needed and builds sets for data
3672 /// dependence checking.
3673 class AccessAnalysis {
3675 /// \brief Read or write access location.
3676 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3677 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3679 /// \brief Set of potential dependent memory accesses.
3680 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3682 AccessAnalysis(DataLayout *Dl, DepCandidates &DA) :
3683 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3684 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3686 /// \brief Register a load and whether it is only read from.
3687 void addLoad(Value *Ptr, bool IsReadOnly) {
3688 Accesses.insert(MemAccessInfo(Ptr, false));
3690 ReadOnlyPtr.insert(Ptr);
3693 /// \brief Register a store.
3694 void addStore(Value *Ptr) {
3695 Accesses.insert(MemAccessInfo(Ptr, true));
3698 /// \brief Check whether we can check the pointers at runtime for
3699 /// non-intersection.
3700 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3701 unsigned &NumComparisons, ScalarEvolution *SE,
3702 Loop *TheLoop, ValueToValueMap &Strides,
3703 bool ShouldCheckStride = false);
3705 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3706 /// and builds sets of dependent accesses.
3707 void buildDependenceSets() {
3708 // Process read-write pointers first.
3709 processMemAccesses(false);
3710 // Next, process read pointers.
3711 processMemAccesses(true);
3714 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3716 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3717 void resetDepChecks() { CheckDeps.clear(); }
3719 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3722 typedef SetVector<MemAccessInfo> PtrAccessSet;
3723 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3725 /// \brief Go over all memory access or only the deferred ones if
3726 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3727 /// and build sets of dependency check candidates.
3728 void processMemAccesses(bool UseDeferred);
3730 /// Set of all accesses.
3731 PtrAccessSet Accesses;
3733 /// Set of access to check after all writes have been processed.
3734 PtrAccessSet DeferredAccesses;
3736 /// Map of pointers to last access encountered.
3737 UnderlyingObjToAccessMap ObjToLastAccess;
3739 /// Set of accesses that need a further dependence check.
3740 MemAccessInfoSet CheckDeps;
3742 /// Set of pointers that are read only.
3743 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3745 /// Set of underlying objects already written to.
3746 SmallPtrSet<Value*, 16> WriteObjects;
3750 /// Sets of potentially dependent accesses - members of one set share an
3751 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3752 /// dependence check.
3753 DepCandidates &DepCands;
3755 bool AreAllWritesIdentified;
3756 bool AreAllReadsIdentified;
3757 bool IsRTCheckNeeded;
3760 } // end anonymous namespace
3762 /// \brief Check whether a pointer can participate in a runtime bounds check.
3763 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3765 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3766 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3770 return AR->isAffine();
3773 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3774 /// the address space.
3775 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
3776 const Loop *Lp, ValueToValueMap &StridesMap);
3778 bool AccessAnalysis::canCheckPtrAtRT(
3779 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3780 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3781 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3782 // Find pointers with computable bounds. We are going to use this information
3783 // to place a runtime bound check.
3784 unsigned NumReadPtrChecks = 0;
3785 unsigned NumWritePtrChecks = 0;
3786 bool CanDoRT = true;
3788 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3789 // We assign consecutive id to access from different dependence sets.
3790 // Accesses within the same set don't need a runtime check.
3791 unsigned RunningDepId = 1;
3792 DenseMap<Value *, unsigned> DepSetId;
3794 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3796 const MemAccessInfo &Access = *AI;
3797 Value *Ptr = Access.getPointer();
3798 bool IsWrite = Access.getInt();
3800 // Just add write checks if we have both.
3801 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3805 ++NumWritePtrChecks;
3809 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3810 // When we run after a failing dependency check we have to make sure we
3811 // don't have wrapping pointers.
3812 (!ShouldCheckStride ||
3813 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3814 // The id of the dependence set.
3817 if (IsDepCheckNeeded) {
3818 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3819 unsigned &LeaderId = DepSetId[Leader];
3821 LeaderId = RunningDepId++;
3824 // Each access has its own dependence set.
3825 DepId = RunningDepId++;
3827 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3829 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3835 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3836 NumComparisons = 0; // Only one dependence set.
3838 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3839 NumWritePtrChecks - 1));
3842 // If the pointers that we would use for the bounds comparison have different
3843 // address spaces, assume the values aren't directly comparable, so we can't
3844 // use them for the runtime check. We also have to assume they could
3845 // overlap. In the future there should be metadata for whether address spaces
3847 unsigned NumPointers = RtCheck.Pointers.size();
3848 for (unsigned i = 0; i < NumPointers; ++i) {
3849 for (unsigned j = i + 1; j < NumPointers; ++j) {
3850 // Only need to check pointers between two different dependency sets.
3851 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3854 Value *PtrI = RtCheck.Pointers[i];
3855 Value *PtrJ = RtCheck.Pointers[j];
3857 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3858 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3860 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3861 " different address spaces\n");
3870 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3871 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3874 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3875 // We process the set twice: first we process read-write pointers, last we
3876 // process read-only pointers. This allows us to skip dependence tests for
3877 // read-only pointers.
3879 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3880 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3881 const MemAccessInfo &Access = *AI;
3882 Value *Ptr = Access.getPointer();
3883 bool IsWrite = Access.getInt();
3885 DepCands.insert(Access);
3887 // Memorize read-only pointers for later processing and skip them in the
3888 // first round (they need to be checked after we have seen all write
3889 // pointers). Note: we also mark pointer that are not consecutive as
3890 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3891 // second check for "!IsWrite".
3892 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3893 if (!UseDeferred && IsReadOnlyPtr) {
3894 DeferredAccesses.insert(Access);
3898 bool NeedDepCheck = false;
3899 // Check whether there is the possibility of dependency because of
3900 // underlying objects being the same.
3901 typedef SmallVector<Value*, 16> ValueVector;
3902 ValueVector TempObjects;
3903 GetUnderlyingObjects(Ptr, TempObjects, DL);
3904 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3906 Value *UnderlyingObj = *UI;
3908 // If this is a write then it needs to be an identified object. If this a
3909 // read and all writes (so far) are identified function scope objects we
3910 // don't need an identified underlying object but only an Argument (the
3911 // next write is going to invalidate this assumption if it is
3913 // This is a micro-optimization for the case where all writes are
3914 // identified and we have one argument pointer.
3915 // Otherwise, we do need a runtime check.
3916 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3917 (!IsWrite && (!AreAllWritesIdentified ||
3918 !isa<Argument>(UnderlyingObj)) &&
3919 !isIdentifiedObject(UnderlyingObj))) {
3920 DEBUG(dbgs() << "LV: Found an unidentified " <<
3921 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3923 IsRTCheckNeeded = (IsRTCheckNeeded ||
3924 !isIdentifiedObject(UnderlyingObj) ||
3925 !AreAllReadsIdentified);
3928 AreAllWritesIdentified = false;
3930 AreAllReadsIdentified = false;
3933 // If this is a write - check other reads and writes for conflicts. If
3934 // this is a read only check other writes for conflicts (but only if there
3935 // is no other write to the ptr - this is an optimization to catch "a[i] =
3936 // a[i] + " without having to do a dependence check).
3937 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3938 NeedDepCheck = true;
3941 WriteObjects.insert(UnderlyingObj);
3943 // Create sets of pointers connected by shared underlying objects.
3944 UnderlyingObjToAccessMap::iterator Prev =
3945 ObjToLastAccess.find(UnderlyingObj);
3946 if (Prev != ObjToLastAccess.end())
3947 DepCands.unionSets(Access, Prev->second);
3949 ObjToLastAccess[UnderlyingObj] = Access;
3953 CheckDeps.insert(Access);
3958 /// \brief Checks memory dependences among accesses to the same underlying
3959 /// object to determine whether there vectorization is legal or not (and at
3960 /// which vectorization factor).
3962 /// This class works under the assumption that we already checked that memory
3963 /// locations with different underlying pointers are "must-not alias".
3964 /// We use the ScalarEvolution framework to symbolically evalutate access
3965 /// functions pairs. Since we currently don't restructure the loop we can rely
3966 /// on the program order of memory accesses to determine their safety.
3967 /// At the moment we will only deem accesses as safe for:
3968 /// * A negative constant distance assuming program order.
3970 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3971 /// a[i] = tmp; y = a[i];
3973 /// The latter case is safe because later checks guarantuee that there can't
3974 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3975 /// the same variable: a header phi can only be an induction or a reduction, a
3976 /// reduction can't have a memory sink, an induction can't have a memory
3977 /// source). This is important and must not be violated (or we have to
3978 /// resort to checking for cycles through memory).
3980 /// * A positive constant distance assuming program order that is bigger
3981 /// than the biggest memory access.
3983 /// tmp = a[i] OR b[i] = x
3984 /// a[i+2] = tmp y = b[i+2];
3986 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3988 /// * Zero distances and all accesses have the same size.
3990 class MemoryDepChecker {
3992 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3993 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3995 MemoryDepChecker(ScalarEvolution *Se, DataLayout *Dl, const Loop *L)
3996 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3997 ShouldRetryWithRuntimeCheck(false) {}
3999 /// \brief Register the location (instructions are given increasing numbers)
4000 /// of a write access.
4001 void addAccess(StoreInst *SI) {
4002 Value *Ptr = SI->getPointerOperand();
4003 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4004 InstMap.push_back(SI);
4008 /// \brief Register the location (instructions are given increasing numbers)
4009 /// of a write access.
4010 void addAccess(LoadInst *LI) {
4011 Value *Ptr = LI->getPointerOperand();
4012 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4013 InstMap.push_back(LI);
4017 /// \brief Check whether the dependencies between the accesses are safe.
4019 /// Only checks sets with elements in \p CheckDeps.
4020 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4021 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4023 /// \brief The maximum number of bytes of a vector register we can vectorize
4024 /// the accesses safely with.
4025 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4027 /// \brief In same cases when the dependency check fails we can still
4028 /// vectorize the loop with a dynamic array access check.
4029 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4032 ScalarEvolution *SE;
4034 const Loop *InnermostLoop;
4036 /// \brief Maps access locations (ptr, read/write) to program order.
4037 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4039 /// \brief Memory access instructions in program order.
4040 SmallVector<Instruction *, 16> InstMap;
4042 /// \brief The program order index to be used for the next instruction.
4045 // We can access this many bytes in parallel safely.
4046 unsigned MaxSafeDepDistBytes;
4048 /// \brief If we see a non-constant dependence distance we can still try to
4049 /// vectorize this loop with runtime checks.
4050 bool ShouldRetryWithRuntimeCheck;
4052 /// \brief Check whether there is a plausible dependence between the two
4055 /// Access \p A must happen before \p B in program order. The two indices
4056 /// identify the index into the program order map.
4058 /// This function checks whether there is a plausible dependence (or the
4059 /// absence of such can't be proved) between the two accesses. If there is a
4060 /// plausible dependence but the dependence distance is bigger than one
4061 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4062 /// distance is smaller than any other distance encountered so far).
4063 /// Otherwise, this function returns true signaling a possible dependence.
4064 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4065 const MemAccessInfo &B, unsigned BIdx,
4066 ValueToValueMap &Strides);
4068 /// \brief Check whether the data dependence could prevent store-load
4070 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4073 } // end anonymous namespace
4075 static bool isInBoundsGep(Value *Ptr) {
4076 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4077 return GEP->isInBounds();
4081 /// \brief Check whether the access through \p Ptr has a constant stride.
4082 static int isStridedPtr(ScalarEvolution *SE, DataLayout *DL, Value *Ptr,
4083 const Loop *Lp, ValueToValueMap &StridesMap) {
4084 const Type *Ty = Ptr->getType();
4085 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4087 // Make sure that the pointer does not point to aggregate types.
4088 const PointerType *PtrTy = cast<PointerType>(Ty);
4089 if (PtrTy->getElementType()->isAggregateType()) {
4090 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4095 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4097 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4099 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4100 << *Ptr << " SCEV: " << *PtrScev << "\n");
4104 // The accesss function must stride over the innermost loop.
4105 if (Lp != AR->getLoop()) {
4106 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4107 *Ptr << " SCEV: " << *PtrScev << "\n");
4110 // The address calculation must not wrap. Otherwise, a dependence could be
4112 // An inbounds getelementptr that is a AddRec with a unit stride
4113 // cannot wrap per definition. The unit stride requirement is checked later.
4114 // An getelementptr without an inbounds attribute and unit stride would have
4115 // to access the pointer value "0" which is undefined behavior in address
4116 // space 0, therefore we can also vectorize this case.
4117 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4118 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4119 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4120 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4121 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4122 << *Ptr << " SCEV: " << *PtrScev << "\n");
4126 // Check the step is constant.
4127 const SCEV *Step = AR->getStepRecurrence(*SE);
4129 // Calculate the pointer stride and check if it is consecutive.
4130 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4132 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4133 " SCEV: " << *PtrScev << "\n");
4137 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4138 const APInt &APStepVal = C->getValue()->getValue();
4140 // Huge step value - give up.
4141 if (APStepVal.getBitWidth() > 64)
4144 int64_t StepVal = APStepVal.getSExtValue();
4147 int64_t Stride = StepVal / Size;
4148 int64_t Rem = StepVal % Size;
4152 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4153 // know we can't "wrap around the address space". In case of address space
4154 // zero we know that this won't happen without triggering undefined behavior.
4155 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4156 Stride != 1 && Stride != -1)
4162 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4163 unsigned TypeByteSize) {
4164 // If loads occur at a distance that is not a multiple of a feasible vector
4165 // factor store-load forwarding does not take place.
4166 // Positive dependences might cause troubles because vectorizing them might
4167 // prevent store-load forwarding making vectorized code run a lot slower.
4168 // a[i] = a[i-3] ^ a[i-8];
4169 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4170 // hence on your typical architecture store-load forwarding does not take
4171 // place. Vectorizing in such cases does not make sense.
4172 // Store-load forwarding distance.
4173 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4174 // Maximum vector factor.
4175 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4176 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4177 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4179 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4181 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4182 MaxVFWithoutSLForwardIssues = (vf >>=1);
4187 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4188 DEBUG(dbgs() << "LV: Distance " << Distance <<
4189 " that could cause a store-load forwarding conflict\n");
4193 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4194 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4195 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4199 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4200 const MemAccessInfo &B, unsigned BIdx,
4201 ValueToValueMap &Strides) {
4202 assert (AIdx < BIdx && "Must pass arguments in program order");
4204 Value *APtr = A.getPointer();
4205 Value *BPtr = B.getPointer();
4206 bool AIsWrite = A.getInt();
4207 bool BIsWrite = B.getInt();
4209 // Two reads are independent.
4210 if (!AIsWrite && !BIsWrite)
4213 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4214 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4216 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4217 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4219 const SCEV *Src = AScev;
4220 const SCEV *Sink = BScev;
4222 // If the induction step is negative we have to invert source and sink of the
4224 if (StrideAPtr < 0) {
4227 std::swap(APtr, BPtr);
4228 std::swap(Src, Sink);
4229 std::swap(AIsWrite, BIsWrite);
4230 std::swap(AIdx, BIdx);
4231 std::swap(StrideAPtr, StrideBPtr);
4234 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4236 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4237 << "(Induction step: " << StrideAPtr << ")\n");
4238 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4239 << *InstMap[BIdx] << ": " << *Dist << "\n");
4241 // Need consecutive accesses. We don't want to vectorize
4242 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4243 // the address space.
4244 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4245 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4249 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4251 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4252 ShouldRetryWithRuntimeCheck = true;
4256 Type *ATy = APtr->getType()->getPointerElementType();
4257 Type *BTy = BPtr->getType()->getPointerElementType();
4258 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4260 // Negative distances are not plausible dependencies.
4261 const APInt &Val = C->getValue()->getValue();
4262 if (Val.isNegative()) {
4263 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4264 if (IsTrueDataDependence &&
4265 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4269 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4273 // Write to the same location with the same size.
4274 // Could be improved to assert type sizes are the same (i32 == float, etc).
4278 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4282 assert(Val.isStrictlyPositive() && "Expect a positive value");
4284 // Positive distance bigger than max vectorization factor.
4287 "LV: ReadWrite-Write positive dependency with different types\n");
4291 unsigned Distance = (unsigned) Val.getZExtValue();
4293 // Bail out early if passed-in parameters make vectorization not feasible.
4294 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4295 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4297 // The distance must be bigger than the size needed for a vectorized version
4298 // of the operation and the size of the vectorized operation must not be
4299 // bigger than the currrent maximum size.
4300 if (Distance < 2*TypeByteSize ||
4301 2*TypeByteSize > MaxSafeDepDistBytes ||
4302 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4303 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4304 << Val.getSExtValue() << '\n');
4308 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4309 Distance : MaxSafeDepDistBytes;
4311 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4312 if (IsTrueDataDependence &&
4313 couldPreventStoreLoadForward(Distance, TypeByteSize))
4316 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4317 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4322 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4323 MemAccessInfoSet &CheckDeps,
4324 ValueToValueMap &Strides) {
4326 MaxSafeDepDistBytes = -1U;
4327 while (!CheckDeps.empty()) {
4328 MemAccessInfo CurAccess = *CheckDeps.begin();
4330 // Get the relevant memory access set.
4331 EquivalenceClasses<MemAccessInfo>::iterator I =
4332 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4334 // Check accesses within this set.
4335 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4336 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4338 // Check every access pair.
4340 CheckDeps.erase(*AI);
4341 EquivalenceClasses<MemAccessInfo>::member_iterator OI = llvm::next(AI);
4343 // Check every accessing instruction pair in program order.
4344 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4345 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4346 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4347 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4348 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4350 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4361 bool LoopVectorizationLegality::canVectorizeMemory() {
4363 typedef SmallVector<Value*, 16> ValueVector;
4364 typedef SmallPtrSet<Value*, 16> ValueSet;
4366 // Holds the Load and Store *instructions*.
4370 // Holds all the different accesses in the loop.
4371 unsigned NumReads = 0;
4372 unsigned NumReadWrites = 0;
4374 PtrRtCheck.Pointers.clear();
4375 PtrRtCheck.Need = false;
4377 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4378 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4381 for (Loop::block_iterator bb = TheLoop->block_begin(),
4382 be = TheLoop->block_end(); bb != be; ++bb) {
4384 // Scan the BB and collect legal loads and stores.
4385 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4388 // If this is a load, save it. If this instruction can read from memory
4389 // but is not a load, then we quit. Notice that we don't handle function
4390 // calls that read or write.
4391 if (it->mayReadFromMemory()) {
4392 // Many math library functions read the rounding mode. We will only
4393 // vectorize a loop if it contains known function calls that don't set
4394 // the flag. Therefore, it is safe to ignore this read from memory.
4395 CallInst *Call = dyn_cast<CallInst>(it);
4396 if (Call && getIntrinsicIDForCall(Call, TLI))
4399 LoadInst *Ld = dyn_cast<LoadInst>(it);
4400 if (!Ld) return false;
4401 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4402 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4406 Loads.push_back(Ld);
4407 DepChecker.addAccess(Ld);
4411 // Save 'store' instructions. Abort if other instructions write to memory.
4412 if (it->mayWriteToMemory()) {
4413 StoreInst *St = dyn_cast<StoreInst>(it);
4414 if (!St) return false;
4415 if (!St->isSimple() && !IsAnnotatedParallel) {
4416 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4420 Stores.push_back(St);
4421 DepChecker.addAccess(St);
4426 // Now we have two lists that hold the loads and the stores.
4427 // Next, we find the pointers that they use.
4429 // Check if we see any stores. If there are no stores, then we don't
4430 // care if the pointers are *restrict*.
4431 if (!Stores.size()) {
4432 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4436 AccessAnalysis::DepCandidates DependentAccesses;
4437 AccessAnalysis Accesses(DL, DependentAccesses);
4439 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4440 // multiple times on the same object. If the ptr is accessed twice, once
4441 // for read and once for write, it will only appear once (on the write
4442 // list). This is okay, since we are going to check for conflicts between
4443 // writes and between reads and writes, but not between reads and reads.
4446 ValueVector::iterator I, IE;
4447 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4448 StoreInst *ST = cast<StoreInst>(*I);
4449 Value* Ptr = ST->getPointerOperand();
4451 if (isUniform(Ptr)) {
4452 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4456 // If we did *not* see this pointer before, insert it to the read-write
4457 // list. At this phase it is only a 'write' list.
4458 if (Seen.insert(Ptr)) {
4460 Accesses.addStore(Ptr);
4464 if (IsAnnotatedParallel) {
4466 << "LV: A loop annotated parallel, ignore memory dependency "
4471 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4472 LoadInst *LD = cast<LoadInst>(*I);
4473 Value* Ptr = LD->getPointerOperand();
4474 // If we did *not* see this pointer before, insert it to the
4475 // read list. If we *did* see it before, then it is already in
4476 // the read-write list. This allows us to vectorize expressions
4477 // such as A[i] += x; Because the address of A[i] is a read-write
4478 // pointer. This only works if the index of A[i] is consecutive.
4479 // If the address of i is unknown (for example A[B[i]]) then we may
4480 // read a few words, modify, and write a few words, and some of the
4481 // words may be written to the same address.
4482 bool IsReadOnlyPtr = false;
4483 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4485 IsReadOnlyPtr = true;
4487 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4490 // If we write (or read-write) to a single destination and there are no
4491 // other reads in this loop then is it safe to vectorize.
4492 if (NumReadWrites == 1 && NumReads == 0) {
4493 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4497 // Build dependence sets and check whether we need a runtime pointer bounds
4499 Accesses.buildDependenceSets();
4500 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4502 // Find pointers with computable bounds. We are going to use this information
4503 // to place a runtime bound check.
4504 unsigned NumComparisons = 0;
4505 bool CanDoRT = false;
4507 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4510 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4511 " pointer comparisons.\n");
4513 // If we only have one set of dependences to check pointers among we don't
4514 // need a runtime check.
4515 if (NumComparisons == 0 && NeedRTCheck)
4516 NeedRTCheck = false;
4518 // Check that we did not collect too many pointers or found an unsizeable
4520 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4526 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4529 if (NeedRTCheck && !CanDoRT) {
4530 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4531 "the array bounds.\n");
4536 PtrRtCheck.Need = NeedRTCheck;
4538 bool CanVecMem = true;
4539 if (Accesses.isDependencyCheckNeeded()) {
4540 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4541 CanVecMem = DepChecker.areDepsSafe(
4542 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4543 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4545 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4546 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4549 // Clear the dependency checks. We assume they are not needed.
4550 Accesses.resetDepChecks();
4553 PtrRtCheck.Need = true;
4555 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4556 TheLoop, Strides, true);
4557 // Check that we did not collect too many pointers or found an unsizeable
4559 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4560 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4569 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4570 " need a runtime memory check.\n");
4575 static bool hasMultipleUsesOf(Instruction *I,
4576 SmallPtrSet<Instruction *, 8> &Insts) {
4577 unsigned NumUses = 0;
4578 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4579 if (Insts.count(dyn_cast<Instruction>(*Use)))
4588 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4589 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4590 if (!Set.count(dyn_cast<Instruction>(*Use)))
4595 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4596 ReductionKind Kind) {
4597 if (Phi->getNumIncomingValues() != 2)
4600 // Reduction variables are only found in the loop header block.
4601 if (Phi->getParent() != TheLoop->getHeader())
4604 // Obtain the reduction start value from the value that comes from the loop
4606 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4608 // ExitInstruction is the single value which is used outside the loop.
4609 // We only allow for a single reduction value to be used outside the loop.
4610 // This includes users of the reduction, variables (which form a cycle
4611 // which ends in the phi node).
4612 Instruction *ExitInstruction = 0;
4613 // Indicates that we found a reduction operation in our scan.
4614 bool FoundReduxOp = false;
4616 // We start with the PHI node and scan for all of the users of this
4617 // instruction. All users must be instructions that can be used as reduction
4618 // variables (such as ADD). We must have a single out-of-block user. The cycle
4619 // must include the original PHI.
4620 bool FoundStartPHI = false;
4622 // To recognize min/max patterns formed by a icmp select sequence, we store
4623 // the number of instruction we saw from the recognized min/max pattern,
4624 // to make sure we only see exactly the two instructions.
4625 unsigned NumCmpSelectPatternInst = 0;
4626 ReductionInstDesc ReduxDesc(false, 0);
4628 SmallPtrSet<Instruction *, 8> VisitedInsts;
4629 SmallVector<Instruction *, 8> Worklist;
4630 Worklist.push_back(Phi);
4631 VisitedInsts.insert(Phi);
4633 // A value in the reduction can be used:
4634 // - By the reduction:
4635 // - Reduction operation:
4636 // - One use of reduction value (safe).
4637 // - Multiple use of reduction value (not safe).
4639 // - All uses of the PHI must be the reduction (safe).
4640 // - Otherwise, not safe.
4641 // - By one instruction outside of the loop (safe).
4642 // - By further instructions outside of the loop (not safe).
4643 // - By an instruction that is not part of the reduction (not safe).
4645 // * An instruction type other than PHI or the reduction operation.
4646 // * A PHI in the header other than the initial PHI.
4647 while (!Worklist.empty()) {
4648 Instruction *Cur = Worklist.back();
4649 Worklist.pop_back();
4652 // If the instruction has no users then this is a broken chain and can't be
4653 // a reduction variable.
4654 if (Cur->use_empty())
4657 bool IsAPhi = isa<PHINode>(Cur);
4659 // A header PHI use other than the original PHI.
4660 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4663 // Reductions of instructions such as Div, and Sub is only possible if the
4664 // LHS is the reduction variable.
4665 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4666 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4667 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4670 // Any reduction instruction must be of one of the allowed kinds.
4671 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4672 if (!ReduxDesc.IsReduction)
4675 // A reduction operation must only have one use of the reduction value.
4676 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4677 hasMultipleUsesOf(Cur, VisitedInsts))
4680 // All inputs to a PHI node must be a reduction value.
4681 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4684 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4685 isa<SelectInst>(Cur)))
4686 ++NumCmpSelectPatternInst;
4687 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4688 isa<SelectInst>(Cur)))
4689 ++NumCmpSelectPatternInst;
4691 // Check whether we found a reduction operator.
4692 FoundReduxOp |= !IsAPhi;
4694 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4695 // onto the stack. This way we are going to have seen all inputs to PHI
4696 // nodes once we get to them.
4697 SmallVector<Instruction *, 8> NonPHIs;
4698 SmallVector<Instruction *, 8> PHIs;
4699 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4701 Instruction *Usr = cast<Instruction>(*UI);
4703 // Check if we found the exit user.
4704 BasicBlock *Parent = Usr->getParent();
4705 if (!TheLoop->contains(Parent)) {
4706 // Exit if you find multiple outside users or if the header phi node is
4707 // being used. In this case the user uses the value of the previous
4708 // iteration, in which case we would loose "VF-1" iterations of the
4709 // reduction operation if we vectorize.
4710 if (ExitInstruction != 0 || Cur == Phi)
4713 // The instruction used by an outside user must be the last instruction
4714 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4715 // operations on the value.
4716 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4719 ExitInstruction = Cur;
4723 // Process instructions only once (termination). Each reduction cycle
4724 // value must only be used once, except by phi nodes and min/max
4725 // reductions which are represented as a cmp followed by a select.
4726 ReductionInstDesc IgnoredVal(false, 0);
4727 if (VisitedInsts.insert(Usr)) {
4728 if (isa<PHINode>(Usr))
4729 PHIs.push_back(Usr);
4731 NonPHIs.push_back(Usr);
4732 } else if (!isa<PHINode>(Usr) &&
4733 ((!isa<FCmpInst>(Usr) &&
4734 !isa<ICmpInst>(Usr) &&
4735 !isa<SelectInst>(Usr)) ||
4736 !isMinMaxSelectCmpPattern(Usr, IgnoredVal).IsReduction))
4739 // Remember that we completed the cycle.
4741 FoundStartPHI = true;
4743 Worklist.append(PHIs.begin(), PHIs.end());
4744 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4747 // This means we have seen one but not the other instruction of the
4748 // pattern or more than just a select and cmp.
4749 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4750 NumCmpSelectPatternInst != 2)
4753 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4756 // We found a reduction var if we have reached the original phi node and we
4757 // only have a single instruction with out-of-loop users.
4759 // This instruction is allowed to have out-of-loop users.
4760 AllowedExit.insert(ExitInstruction);
4762 // Save the description of this reduction variable.
4763 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4764 ReduxDesc.MinMaxKind);
4765 Reductions[Phi] = RD;
4766 // We've ended the cycle. This is a reduction variable if we have an
4767 // outside user and it has a binary op.
4772 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4773 /// pattern corresponding to a min(X, Y) or max(X, Y).
4774 LoopVectorizationLegality::ReductionInstDesc
4775 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4776 ReductionInstDesc &Prev) {
4778 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4779 "Expect a select instruction");
4780 Instruction *Cmp = 0;
4781 SelectInst *Select = 0;
4783 // We must handle the select(cmp()) as a single instruction. Advance to the
4785 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4786 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4787 return ReductionInstDesc(false, I);
4788 return ReductionInstDesc(Select, Prev.MinMaxKind);
4791 // Only handle single use cases for now.
4792 if (!(Select = dyn_cast<SelectInst>(I)))
4793 return ReductionInstDesc(false, I);
4794 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4795 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4796 return ReductionInstDesc(false, I);
4797 if (!Cmp->hasOneUse())
4798 return ReductionInstDesc(false, I);
4803 // Look for a min/max pattern.
4804 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4805 return ReductionInstDesc(Select, MRK_UIntMin);
4806 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4807 return ReductionInstDesc(Select, MRK_UIntMax);
4808 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4809 return ReductionInstDesc(Select, MRK_SIntMax);
4810 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4811 return ReductionInstDesc(Select, MRK_SIntMin);
4812 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4813 return ReductionInstDesc(Select, MRK_FloatMin);
4814 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4815 return ReductionInstDesc(Select, MRK_FloatMax);
4816 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4817 return ReductionInstDesc(Select, MRK_FloatMin);
4818 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4819 return ReductionInstDesc(Select, MRK_FloatMax);
4821 return ReductionInstDesc(false, I);
4824 LoopVectorizationLegality::ReductionInstDesc
4825 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4827 ReductionInstDesc &Prev) {
4828 bool FP = I->getType()->isFloatingPointTy();
4829 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4830 switch (I->getOpcode()) {
4832 return ReductionInstDesc(false, I);
4833 case Instruction::PHI:
4834 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4835 Kind != RK_FloatMinMax))
4836 return ReductionInstDesc(false, I);
4837 return ReductionInstDesc(I, Prev.MinMaxKind);
4838 case Instruction::Sub:
4839 case Instruction::Add:
4840 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4841 case Instruction::Mul:
4842 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4843 case Instruction::And:
4844 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4845 case Instruction::Or:
4846 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4847 case Instruction::Xor:
4848 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4849 case Instruction::FMul:
4850 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4851 case Instruction::FAdd:
4852 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4853 case Instruction::FCmp:
4854 case Instruction::ICmp:
4855 case Instruction::Select:
4856 if (Kind != RK_IntegerMinMax &&
4857 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4858 return ReductionInstDesc(false, I);
4859 return isMinMaxSelectCmpPattern(I, Prev);
4863 LoopVectorizationLegality::InductionKind
4864 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4865 Type *PhiTy = Phi->getType();
4866 // We only handle integer and pointer inductions variables.
4867 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4868 return IK_NoInduction;
4870 // Check that the PHI is consecutive.
4871 const SCEV *PhiScev = SE->getSCEV(Phi);
4872 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4874 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4875 return IK_NoInduction;
4877 const SCEV *Step = AR->getStepRecurrence(*SE);
4879 // Integer inductions need to have a stride of one.
4880 if (PhiTy->isIntegerTy()) {
4882 return IK_IntInduction;
4883 if (Step->isAllOnesValue())
4884 return IK_ReverseIntInduction;
4885 return IK_NoInduction;
4888 // Calculate the pointer stride and check if it is consecutive.
4889 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4891 return IK_NoInduction;
4893 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4894 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4895 if (C->getValue()->equalsInt(Size))
4896 return IK_PtrInduction;
4897 else if (C->getValue()->equalsInt(0 - Size))
4898 return IK_ReversePtrInduction;
4900 return IK_NoInduction;
4903 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4904 Value *In0 = const_cast<Value*>(V);
4905 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4909 return Inductions.count(PN);
4912 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4913 assert(TheLoop->contains(BB) && "Unknown block used");
4915 // Blocks that do not dominate the latch need predication.
4916 BasicBlock* Latch = TheLoop->getLoopLatch();
4917 return !DT->dominates(BB, Latch);
4920 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4921 SmallPtrSet<Value *, 8>& SafePtrs) {
4922 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4923 // We might be able to hoist the load.
4924 if (it->mayReadFromMemory()) {
4925 LoadInst *LI = dyn_cast<LoadInst>(it);
4926 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4930 // We don't predicate stores at the moment.
4931 if (it->mayWriteToMemory()) {
4932 StoreInst *SI = dyn_cast<StoreInst>(it);
4933 // We only support predication of stores in basic blocks with one
4935 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
4936 !SafePtrs.count(SI->getPointerOperand()) ||
4937 !SI->getParent()->getSinglePredecessor())
4943 // Check that we don't have a constant expression that can trap as operand.
4944 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4946 if (Constant *C = dyn_cast<Constant>(*OI))
4951 // The instructions below can trap.
4952 switch (it->getOpcode()) {
4954 case Instruction::UDiv:
4955 case Instruction::SDiv:
4956 case Instruction::URem:
4957 case Instruction::SRem:
4965 LoopVectorizationCostModel::VectorizationFactor
4966 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4968 // Width 1 means no vectorize
4969 VectorizationFactor Factor = { 1U, 0U };
4970 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4971 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4975 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
4976 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4980 // Find the trip count.
4981 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4982 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4984 unsigned WidestType = getWidestType();
4985 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4986 unsigned MaxSafeDepDist = -1U;
4987 if (Legal->getMaxSafeDepDistBytes() != -1U)
4988 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4989 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4990 WidestRegister : MaxSafeDepDist);
4991 unsigned MaxVectorSize = WidestRegister / WidestType;
4992 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4993 DEBUG(dbgs() << "LV: The Widest register is: "
4994 << WidestRegister << " bits.\n");
4996 if (MaxVectorSize == 0) {
4997 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5001 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5002 " into one vector!");
5004 unsigned VF = MaxVectorSize;
5006 // If we optimize the program for size, avoid creating the tail loop.
5008 // If we are unable to calculate the trip count then don't try to vectorize.
5010 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5014 // Find the maximum SIMD width that can fit within the trip count.
5015 VF = TC % MaxVectorSize;
5020 // If the trip count that we found modulo the vectorization factor is not
5021 // zero then we require a tail.
5023 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5029 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5030 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5032 Factor.Width = UserVF;
5036 float Cost = expectedCost(1);
5038 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
5039 for (unsigned i=2; i <= VF; i*=2) {
5040 // Notice that the vector loop needs to be executed less times, so
5041 // we need to divide the cost of the vector loops by the width of
5042 // the vector elements.
5043 float VectorCost = expectedCost(i) / (float)i;
5044 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5045 (int)VectorCost << ".\n");
5046 if (VectorCost < Cost) {
5052 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
5053 Factor.Width = Width;
5054 Factor.Cost = Width * Cost;
5058 unsigned LoopVectorizationCostModel::getWidestType() {
5059 unsigned MaxWidth = 8;
5062 for (Loop::block_iterator bb = TheLoop->block_begin(),
5063 be = TheLoop->block_end(); bb != be; ++bb) {
5064 BasicBlock *BB = *bb;
5066 // For each instruction in the loop.
5067 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5068 Type *T = it->getType();
5070 // Only examine Loads, Stores and PHINodes.
5071 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5074 // Examine PHI nodes that are reduction variables.
5075 if (PHINode *PN = dyn_cast<PHINode>(it))
5076 if (!Legal->getReductionVars()->count(PN))
5079 // Examine the stored values.
5080 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5081 T = ST->getValueOperand()->getType();
5083 // Ignore loaded pointer types and stored pointer types that are not
5084 // consecutive. However, we do want to take consecutive stores/loads of
5085 // pointer vectors into account.
5086 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5089 MaxWidth = std::max(MaxWidth,
5090 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5098 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5101 unsigned LoopCost) {
5103 // -- The unroll heuristics --
5104 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5105 // There are many micro-architectural considerations that we can't predict
5106 // at this level. For example frontend pressure (on decode or fetch) due to
5107 // code size, or the number and capabilities of the execution ports.
5109 // We use the following heuristics to select the unroll factor:
5110 // 1. If the code has reductions the we unroll in order to break the cross
5111 // iteration dependency.
5112 // 2. If the loop is really small then we unroll in order to reduce the loop
5114 // 3. We don't unroll if we think that we will spill registers to memory due
5115 // to the increased register pressure.
5117 // Use the user preference, unless 'auto' is selected.
5121 // When we optimize for size we don't unroll.
5125 // We used the distance for the unroll factor.
5126 if (Legal->getMaxSafeDepDistBytes() != -1U)
5129 // Do not unroll loops with a relatively small trip count.
5130 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5131 TheLoop->getLoopLatch());
5132 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5135 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5136 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5140 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5141 TargetNumRegisters = ForceTargetNumScalarRegs;
5143 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5144 TargetNumRegisters = ForceTargetNumVectorRegs;
5147 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5148 // We divide by these constants so assume that we have at least one
5149 // instruction that uses at least one register.
5150 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5151 R.NumInstructions = std::max(R.NumInstructions, 1U);
5153 // We calculate the unroll factor using the following formula.
5154 // Subtract the number of loop invariants from the number of available
5155 // registers. These registers are used by all of the unrolled instances.
5156 // Next, divide the remaining registers by the number of registers that is
5157 // required by the loop, in order to estimate how many parallel instances
5158 // fit without causing spills. All of this is rounded down if necessary to be
5159 // a power of two. We want power of two unroll factors to simplify any
5160 // addressing operations or alignment considerations.
5161 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5164 // Don't count the induction variable as unrolled.
5165 if (EnableIndVarRegisterHeur)
5166 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5167 std::max(1U, (R.MaxLocalUsers - 1)));
5169 // Clamp the unroll factor ranges to reasonable factors.
5170 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5172 // Check if the user has overridden the unroll max.
5174 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5175 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5177 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5178 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5181 // If we did not calculate the cost for VF (because the user selected the VF)
5182 // then we calculate the cost of VF here.
5184 LoopCost = expectedCost(VF);
5186 // Clamp the calculated UF to be between the 1 and the max unroll factor
5187 // that the target allows.
5188 if (UF > MaxUnrollSize)
5193 // Unroll if we vectorized this loop and there is a reduction that could
5194 // benefit from unrolling.
5195 if (VF > 1 && Legal->getReductionVars()->size()) {
5196 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5200 // Note that if we've already vectorized the loop we will have done the
5201 // runtime check and so unrolling won't require further checks.
5202 bool UnrollingRequiresRuntimePointerCheck =
5203 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5205 // We want to unroll small loops in order to reduce the loop overhead and
5206 // potentially expose ILP opportunities.
5207 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5208 if (!UnrollingRequiresRuntimePointerCheck &&
5209 LoopCost < SmallLoopCost) {
5210 // We assume that the cost overhead is 1 and we use the cost model
5211 // to estimate the cost of the loop and unroll until the cost of the
5212 // loop overhead is about 5% of the cost of the loop.
5213 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5215 // Unroll until store/load ports (estimated by max unroll factor) are
5217 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5218 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5220 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5221 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5222 return std::max(StoresUF, LoadsUF);
5225 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5229 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5233 LoopVectorizationCostModel::RegisterUsage
5234 LoopVectorizationCostModel::calculateRegisterUsage() {
5235 // This function calculates the register usage by measuring the highest number
5236 // of values that are alive at a single location. Obviously, this is a very
5237 // rough estimation. We scan the loop in a topological order in order and
5238 // assign a number to each instruction. We use RPO to ensure that defs are
5239 // met before their users. We assume that each instruction that has in-loop
5240 // users starts an interval. We record every time that an in-loop value is
5241 // used, so we have a list of the first and last occurrences of each
5242 // instruction. Next, we transpose this data structure into a multi map that
5243 // holds the list of intervals that *end* at a specific location. This multi
5244 // map allows us to perform a linear search. We scan the instructions linearly
5245 // and record each time that a new interval starts, by placing it in a set.
5246 // If we find this value in the multi-map then we remove it from the set.
5247 // The max register usage is the maximum size of the set.
5248 // We also search for instructions that are defined outside the loop, but are
5249 // used inside the loop. We need this number separately from the max-interval
5250 // usage number because when we unroll, loop-invariant values do not take
5252 LoopBlocksDFS DFS(TheLoop);
5256 R.NumInstructions = 0;
5258 // Each 'key' in the map opens a new interval. The values
5259 // of the map are the index of the 'last seen' usage of the
5260 // instruction that is the key.
5261 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5262 // Maps instruction to its index.
5263 DenseMap<unsigned, Instruction*> IdxToInstr;
5264 // Marks the end of each interval.
5265 IntervalMap EndPoint;
5266 // Saves the list of instruction indices that are used in the loop.
5267 SmallSet<Instruction*, 8> Ends;
5268 // Saves the list of values that are used in the loop but are
5269 // defined outside the loop, such as arguments and constants.
5270 SmallPtrSet<Value*, 8> LoopInvariants;
5273 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5274 be = DFS.endRPO(); bb != be; ++bb) {
5275 R.NumInstructions += (*bb)->size();
5276 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5278 Instruction *I = it;
5279 IdxToInstr[Index++] = I;
5281 // Save the end location of each USE.
5282 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5283 Value *U = I->getOperand(i);
5284 Instruction *Instr = dyn_cast<Instruction>(U);
5286 // Ignore non-instruction values such as arguments, constants, etc.
5287 if (!Instr) continue;
5289 // If this instruction is outside the loop then record it and continue.
5290 if (!TheLoop->contains(Instr)) {
5291 LoopInvariants.insert(Instr);
5295 // Overwrite previous end points.
5296 EndPoint[Instr] = Index;
5302 // Saves the list of intervals that end with the index in 'key'.
5303 typedef SmallVector<Instruction*, 2> InstrList;
5304 DenseMap<unsigned, InstrList> TransposeEnds;
5306 // Transpose the EndPoints to a list of values that end at each index.
5307 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5309 TransposeEnds[it->second].push_back(it->first);
5311 SmallSet<Instruction*, 8> OpenIntervals;
5312 unsigned MaxUsage = 0;
5315 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5316 for (unsigned int i = 0; i < Index; ++i) {
5317 Instruction *I = IdxToInstr[i];
5318 // Ignore instructions that are never used within the loop.
5319 if (!Ends.count(I)) continue;
5321 // Remove all of the instructions that end at this location.
5322 InstrList &List = TransposeEnds[i];
5323 for (unsigned int j=0, e = List.size(); j < e; ++j)
5324 OpenIntervals.erase(List[j]);
5326 // Count the number of live interals.
5327 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5329 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5330 OpenIntervals.size() << '\n');
5332 // Add the current instruction to the list of open intervals.
5333 OpenIntervals.insert(I);
5336 unsigned Invariant = LoopInvariants.size();
5337 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5338 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5339 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5341 R.LoopInvariantRegs = Invariant;
5342 R.MaxLocalUsers = MaxUsage;
5346 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5350 for (Loop::block_iterator bb = TheLoop->block_begin(),
5351 be = TheLoop->block_end(); bb != be; ++bb) {
5352 unsigned BlockCost = 0;
5353 BasicBlock *BB = *bb;
5355 // For each instruction in the old loop.
5356 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5357 // Skip dbg intrinsics.
5358 if (isa<DbgInfoIntrinsic>(it))
5361 unsigned C = getInstructionCost(it, VF);
5363 // Check if we should override the cost.
5364 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5365 C = ForceTargetInstructionCost;
5368 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5369 VF << " For instruction: " << *it << '\n');
5372 // We assume that if-converted blocks have a 50% chance of being executed.
5373 // When the code is scalar then some of the blocks are avoided due to CF.
5374 // When the code is vectorized we execute all code paths.
5375 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5384 /// \brief Check whether the address computation for a non-consecutive memory
5385 /// access looks like an unlikely candidate for being merged into the indexing
5388 /// We look for a GEP which has one index that is an induction variable and all
5389 /// other indices are loop invariant. If the stride of this access is also
5390 /// within a small bound we decide that this address computation can likely be
5391 /// merged into the addressing mode.
5392 /// In all other cases, we identify the address computation as complex.
5393 static bool isLikelyComplexAddressComputation(Value *Ptr,
5394 LoopVectorizationLegality *Legal,
5395 ScalarEvolution *SE,
5396 const Loop *TheLoop) {
5397 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5401 // We are looking for a gep with all loop invariant indices except for one
5402 // which should be an induction variable.
5403 unsigned NumOperands = Gep->getNumOperands();
5404 for (unsigned i = 1; i < NumOperands; ++i) {
5405 Value *Opd = Gep->getOperand(i);
5406 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5407 !Legal->isInductionVariable(Opd))
5411 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5412 // can likely be merged into the address computation.
5413 unsigned MaxMergeDistance = 64;
5415 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5419 // Check the step is constant.
5420 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5421 // Calculate the pointer stride and check if it is consecutive.
5422 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5426 const APInt &APStepVal = C->getValue()->getValue();
5428 // Huge step value - give up.
5429 if (APStepVal.getBitWidth() > 64)
5432 int64_t StepVal = APStepVal.getSExtValue();
5434 return StepVal > MaxMergeDistance;
5437 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5438 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5444 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5445 // If we know that this instruction will remain uniform, check the cost of
5446 // the scalar version.
5447 if (Legal->isUniformAfterVectorization(I))
5450 Type *RetTy = I->getType();
5451 Type *VectorTy = ToVectorTy(RetTy, VF);
5453 // TODO: We need to estimate the cost of intrinsic calls.
5454 switch (I->getOpcode()) {
5455 case Instruction::GetElementPtr:
5456 // We mark this instruction as zero-cost because the cost of GEPs in
5457 // vectorized code depends on whether the corresponding memory instruction
5458 // is scalarized or not. Therefore, we handle GEPs with the memory
5459 // instruction cost.
5461 case Instruction::Br: {
5462 return TTI.getCFInstrCost(I->getOpcode());
5464 case Instruction::PHI:
5465 //TODO: IF-converted IFs become selects.
5467 case Instruction::Add:
5468 case Instruction::FAdd:
5469 case Instruction::Sub:
5470 case Instruction::FSub:
5471 case Instruction::Mul:
5472 case Instruction::FMul:
5473 case Instruction::UDiv:
5474 case Instruction::SDiv:
5475 case Instruction::FDiv:
5476 case Instruction::URem:
5477 case Instruction::SRem:
5478 case Instruction::FRem:
5479 case Instruction::Shl:
5480 case Instruction::LShr:
5481 case Instruction::AShr:
5482 case Instruction::And:
5483 case Instruction::Or:
5484 case Instruction::Xor: {
5485 // Since we will replace the stride by 1 the multiplication should go away.
5486 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5488 // Certain instructions can be cheaper to vectorize if they have a constant
5489 // second vector operand. One example of this are shifts on x86.
5490 TargetTransformInfo::OperandValueKind Op1VK =
5491 TargetTransformInfo::OK_AnyValue;
5492 TargetTransformInfo::OperandValueKind Op2VK =
5493 TargetTransformInfo::OK_AnyValue;
5494 Value *Op2 = I->getOperand(1);
5496 // Check for a splat of a constant or for a non uniform vector of constants.
5497 if (isa<ConstantInt>(Op2))
5498 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5499 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5500 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5501 if (cast<Constant>(Op2)->getSplatValue() != NULL)
5502 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5505 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5507 case Instruction::Select: {
5508 SelectInst *SI = cast<SelectInst>(I);
5509 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5510 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5511 Type *CondTy = SI->getCondition()->getType();
5513 CondTy = VectorType::get(CondTy, VF);
5515 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5517 case Instruction::ICmp:
5518 case Instruction::FCmp: {
5519 Type *ValTy = I->getOperand(0)->getType();
5520 VectorTy = ToVectorTy(ValTy, VF);
5521 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5523 case Instruction::Store:
5524 case Instruction::Load: {
5525 StoreInst *SI = dyn_cast<StoreInst>(I);
5526 LoadInst *LI = dyn_cast<LoadInst>(I);
5527 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5529 VectorTy = ToVectorTy(ValTy, VF);
5531 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5532 unsigned AS = SI ? SI->getPointerAddressSpace() :
5533 LI->getPointerAddressSpace();
5534 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5535 // We add the cost of address computation here instead of with the gep
5536 // instruction because only here we know whether the operation is
5539 return TTI.getAddressComputationCost(VectorTy) +
5540 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5542 // Scalarized loads/stores.
5543 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5544 bool Reverse = ConsecutiveStride < 0;
5545 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5546 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5547 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5548 bool IsComplexComputation =
5549 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5551 // The cost of extracting from the value vector and pointer vector.
5552 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5553 for (unsigned i = 0; i < VF; ++i) {
5554 // The cost of extracting the pointer operand.
5555 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5556 // In case of STORE, the cost of ExtractElement from the vector.
5557 // In case of LOAD, the cost of InsertElement into the returned
5559 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5560 Instruction::InsertElement,
5564 // The cost of the scalar loads/stores.
5565 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5566 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5571 // Wide load/stores.
5572 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5573 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5576 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5580 case Instruction::ZExt:
5581 case Instruction::SExt:
5582 case Instruction::FPToUI:
5583 case Instruction::FPToSI:
5584 case Instruction::FPExt:
5585 case Instruction::PtrToInt:
5586 case Instruction::IntToPtr:
5587 case Instruction::SIToFP:
5588 case Instruction::UIToFP:
5589 case Instruction::Trunc:
5590 case Instruction::FPTrunc:
5591 case Instruction::BitCast: {
5592 // We optimize the truncation of induction variable.
5593 // The cost of these is the same as the scalar operation.
5594 if (I->getOpcode() == Instruction::Trunc &&
5595 Legal->isInductionVariable(I->getOperand(0)))
5596 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5597 I->getOperand(0)->getType());
5599 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5600 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5602 case Instruction::Call: {
5603 CallInst *CI = cast<CallInst>(I);
5604 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5605 assert(ID && "Not an intrinsic call!");
5606 Type *RetTy = ToVectorTy(CI->getType(), VF);
5607 SmallVector<Type*, 4> Tys;
5608 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5609 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5610 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5613 // We are scalarizing the instruction. Return the cost of the scalar
5614 // instruction, plus the cost of insert and extract into vector
5615 // elements, times the vector width.
5618 if (!RetTy->isVoidTy() && VF != 1) {
5619 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5621 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5624 // The cost of inserting the results plus extracting each one of the
5626 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5629 // The cost of executing VF copies of the scalar instruction. This opcode
5630 // is unknown. Assume that it is the same as 'mul'.
5631 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5637 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5638 if (Scalar->isVoidTy() || VF == 1)
5640 return VectorType::get(Scalar, VF);
5643 char LoopVectorize::ID = 0;
5644 static const char lv_name[] = "Loop Vectorization";
5645 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5646 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5647 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5648 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5649 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5650 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5651 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5652 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5653 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5656 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5657 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5661 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5662 // Check for a store.
5663 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5664 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5666 // Check for a load.
5667 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5668 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5674 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5675 bool IfPredicateStore) {
5676 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5677 // Holds vector parameters or scalars, in case of uniform vals.
5678 SmallVector<VectorParts, 4> Params;
5680 setDebugLocFromInst(Builder, Instr);
5682 // Find all of the vectorized parameters.
5683 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5684 Value *SrcOp = Instr->getOperand(op);
5686 // If we are accessing the old induction variable, use the new one.
5687 if (SrcOp == OldInduction) {
5688 Params.push_back(getVectorValue(SrcOp));
5692 // Try using previously calculated values.
5693 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5695 // If the src is an instruction that appeared earlier in the basic block
5696 // then it should already be vectorized.
5697 if (SrcInst && OrigLoop->contains(SrcInst)) {
5698 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5699 // The parameter is a vector value from earlier.
5700 Params.push_back(WidenMap.get(SrcInst));
5702 // The parameter is a scalar from outside the loop. Maybe even a constant.
5703 VectorParts Scalars;
5704 Scalars.append(UF, SrcOp);
5705 Params.push_back(Scalars);
5709 assert(Params.size() == Instr->getNumOperands() &&
5710 "Invalid number of operands");
5712 // Does this instruction return a value ?
5713 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5715 Value *UndefVec = IsVoidRetTy ? 0 :
5716 UndefValue::get(Instr->getType());
5717 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5718 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5720 Instruction *InsertPt = Builder.GetInsertPoint();
5721 BasicBlock *IfBlock = Builder.GetInsertBlock();
5722 BasicBlock *CondBlock = 0;
5726 if (IfPredicateStore) {
5727 assert(Instr->getParent()->getSinglePredecessor() &&
5728 "Only support single predecessor blocks");
5729 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5730 Instr->getParent());
5731 VectorLp = LI->getLoopFor(IfBlock);
5732 assert(VectorLp && "Must have a loop for this block");
5735 // For each vector unroll 'part':
5736 for (unsigned Part = 0; Part < UF; ++Part) {
5737 // For each scalar that we create:
5739 // Start an "if (pred) a[i] = ..." block.
5741 if (IfPredicateStore) {
5742 if (Cond[Part]->getType()->isVectorTy())
5744 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5745 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5746 ConstantInt::get(Cond[Part]->getType(), 1));
5747 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5748 LoopVectorBody.push_back(CondBlock);
5749 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5750 // Update Builder with newly created basic block.
5751 Builder.SetInsertPoint(InsertPt);
5754 Instruction *Cloned = Instr->clone();
5756 Cloned->setName(Instr->getName() + ".cloned");
5757 // Replace the operands of the cloned instructions with extracted scalars.
5758 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5759 Value *Op = Params[op][Part];
5760 Cloned->setOperand(op, Op);
5763 // Place the cloned scalar in the new loop.
5764 Builder.Insert(Cloned);
5766 // If the original scalar returns a value we need to place it in a vector
5767 // so that future users will be able to use it.
5769 VecResults[Part] = Cloned;
5772 if (IfPredicateStore) {
5773 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5774 LoopVectorBody.push_back(NewIfBlock);
5775 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5776 Builder.SetInsertPoint(InsertPt);
5777 Instruction *OldBr = IfBlock->getTerminator();
5778 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5779 OldBr->eraseFromParent();
5780 IfBlock = NewIfBlock;
5785 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5786 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5787 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5789 return scalarizeInstruction(Instr, IfPredicateStore);
5792 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5796 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5800 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5802 // When unrolling and the VF is 1, we only need to add a simple scalar.
5803 Type *ITy = Val->getType();
5804 assert(!ITy->isVectorTy() && "Val must be a scalar");
5805 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5806 return Builder.CreateAdd(Val, C, "induction");