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/PatternMatch.h"
79 #include "llvm/IR/Type.h"
80 #include "llvm/IR/Value.h"
81 #include "llvm/IR/Verifier.h"
82 #include "llvm/Pass.h"
83 #include "llvm/Support/BranchProbability.h"
84 #include "llvm/Support/CommandLine.h"
85 #include "llvm/Support/Debug.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, const 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.
382 const DataLayout *DL;
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, const 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, const 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.
728 const DataLayout *DL;
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 const 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.
851 const DataLayout *DL;
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;
1012 const DataLayout *DL;
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 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1026 DL = DLP ? &DLP->getDataLayout() : 0;
1027 LI = &getAnalysis<LoopInfo>();
1028 TTI = &getAnalysis<TargetTransformInfo>();
1029 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1030 BFI = &getAnalysis<BlockFrequencyInfo>();
1031 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1033 // Compute some weights outside of the loop over the loops. Compute this
1034 // using a BranchProbability to re-use its scaling math.
1035 const BranchProbability ColdProb(1, 5); // 20%
1036 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1038 // If the target claims to have no vector registers don't attempt
1040 if (!TTI->getNumberOfRegisters(true))
1044 DEBUG(dbgs() << "LV: Not vectorizing: Missing data layout\n");
1048 // Build up a worklist of inner-loops to vectorize. This is necessary as
1049 // the act of vectorizing or partially unrolling a loop creates new loops
1050 // and can invalidate iterators across the loops.
1051 SmallVector<Loop *, 8> Worklist;
1053 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
1054 addInnerLoop(*I, Worklist);
1056 // Now walk the identified inner loops.
1057 bool Changed = false;
1058 while (!Worklist.empty())
1059 Changed |= processLoop(Worklist.pop_back_val());
1061 // Process each loop nest in the function.
1065 bool processLoop(Loop *L) {
1066 // We only handle inner loops, so if there are children just recurse.
1068 bool Changed = false;
1069 for (Loop::iterator I = L->begin(), E = L->begin(); I != E; ++I)
1070 Changed |= processLoop(*I);
1074 DEBUG(dbgs() << "LV: Checking a loop in \"" <<
1075 L->getHeader()->getParent()->getName() << "\"\n");
1077 LoopVectorizeHints Hints(L, DisableUnrolling);
1079 if (Hints.Force == 0) {
1080 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1084 if (!AlwaysVectorize && Hints.Force != 1) {
1085 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1089 if (Hints.Width == 1 && Hints.Unroll == 1) {
1090 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1094 // Check if it is legal to vectorize the loop.
1095 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1096 if (!LVL.canVectorize()) {
1097 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1101 // Use the cost model.
1102 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1104 // Check the function attributes to find out if this function should be
1105 // optimized for size.
1106 Function *F = L->getHeader()->getParent();
1108 Hints.Force != 1 && F->hasFnAttribute(Attribute::OptimizeForSize);
1110 // Compute the weighted frequency of this loop being executed and see if it
1111 // is less than 20% of the function entry baseline frequency. Note that we
1112 // always have a canonical loop here because we think we *can* vectoriez.
1113 // FIXME: This is hidden behind a flag due to pervasive problems with
1114 // exactly what block frequency models.
1115 if (LoopVectorizeWithBlockFrequency) {
1116 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1117 if (Hints.Force != 1 && LoopEntryFreq < ColdEntryFreq)
1121 // Check the function attributes to see if implicit floats are allowed.a
1122 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1123 // an integer loop and the vector instructions selected are purely integer
1124 // vector instructions?
1125 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1126 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1127 "attribute is used.\n");
1131 // Select the optimal vectorization factor.
1132 LoopVectorizationCostModel::VectorizationFactor VF;
1133 VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
1134 // Select the unroll factor.
1135 unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1138 DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
1139 F->getParent()->getModuleIdentifier() << '\n');
1140 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1142 if (VF.Width == 1) {
1143 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1146 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1147 // We decided not to vectorize, but we may want to unroll.
1148 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1149 Unroller.vectorize(&LVL);
1151 // If we decided that it is *legal* to vectorize the loop then do it.
1152 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1156 // Mark the loop as already vectorized to avoid vectorizing again.
1157 Hints.setAlreadyVectorized(L);
1159 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1163 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1164 AU.addRequiredID(LoopSimplifyID);
1165 AU.addRequiredID(LCSSAID);
1166 AU.addRequired<BlockFrequencyInfo>();
1167 AU.addRequired<DominatorTreeWrapperPass>();
1168 AU.addRequired<LoopInfo>();
1169 AU.addRequired<ScalarEvolution>();
1170 AU.addRequired<TargetTransformInfo>();
1171 AU.addPreserved<LoopInfo>();
1172 AU.addPreserved<DominatorTreeWrapperPass>();
1177 } // end anonymous namespace
1179 //===----------------------------------------------------------------------===//
1180 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1181 // LoopVectorizationCostModel.
1182 //===----------------------------------------------------------------------===//
1184 static Value *stripIntegerCast(Value *V) {
1185 if (CastInst *CI = dyn_cast<CastInst>(V))
1186 if (CI->getOperand(0)->getType()->isIntegerTy())
1187 return CI->getOperand(0);
1191 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1193 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1195 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1196 ValueToValueMap &PtrToStride,
1197 Value *Ptr, Value *OrigPtr = 0) {
1199 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1201 // If there is an entry in the map return the SCEV of the pointer with the
1202 // symbolic stride replaced by one.
1203 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1204 if (SI != PtrToStride.end()) {
1205 Value *StrideVal = SI->second;
1208 StrideVal = stripIntegerCast(StrideVal);
1210 // Replace symbolic stride by one.
1211 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1212 ValueToValueMap RewriteMap;
1213 RewriteMap[StrideVal] = One;
1216 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1217 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1222 // Otherwise, just return the SCEV of the original pointer.
1223 return SE->getSCEV(Ptr);
1226 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1227 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1228 ValueToValueMap &Strides) {
1229 // Get the stride replaced scev.
1230 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1231 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1232 assert(AR && "Invalid addrec expression");
1233 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1234 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1235 Pointers.push_back(Ptr);
1236 Starts.push_back(AR->getStart());
1237 Ends.push_back(ScEnd);
1238 IsWritePtr.push_back(WritePtr);
1239 DependencySetId.push_back(DepSetId);
1242 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1243 // We need to place the broadcast of invariant variables outside the loop.
1244 Instruction *Instr = dyn_cast<Instruction>(V);
1246 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1247 Instr->getParent()) != LoopVectorBody.end());
1248 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1250 // Place the code for broadcasting invariant variables in the new preheader.
1251 IRBuilder<>::InsertPointGuard Guard(Builder);
1253 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1255 // Broadcast the scalar into all locations in the vector.
1256 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1261 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1263 assert(Val->getType()->isVectorTy() && "Must be a vector");
1264 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1265 "Elem must be an integer");
1266 // Create the types.
1267 Type *ITy = Val->getType()->getScalarType();
1268 VectorType *Ty = cast<VectorType>(Val->getType());
1269 int VLen = Ty->getNumElements();
1270 SmallVector<Constant*, 8> Indices;
1272 // Create a vector of consecutive numbers from zero to VF.
1273 for (int i = 0; i < VLen; ++i) {
1274 int64_t Idx = Negate ? (-i) : i;
1275 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1278 // Add the consecutive indices to the vector value.
1279 Constant *Cv = ConstantVector::get(Indices);
1280 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1281 return Builder.CreateAdd(Val, Cv, "induction");
1284 /// \brief Find the operand of the GEP that should be checked for consecutive
1285 /// stores. This ignores trailing indices that have no effect on the final
1287 static unsigned getGEPInductionOperand(const DataLayout *DL,
1288 const GetElementPtrInst *Gep) {
1289 unsigned LastOperand = Gep->getNumOperands() - 1;
1290 unsigned GEPAllocSize = DL->getTypeAllocSize(
1291 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1293 // Walk backwards and try to peel off zeros.
1294 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1295 // Find the type we're currently indexing into.
1296 gep_type_iterator GEPTI = gep_type_begin(Gep);
1297 std::advance(GEPTI, LastOperand - 1);
1299 // If it's a type with the same allocation size as the result of the GEP we
1300 // can peel off the zero index.
1301 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1309 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1310 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1311 // Make sure that the pointer does not point to structs.
1312 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1315 // If this value is a pointer induction variable we know it is consecutive.
1316 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1317 if (Phi && Inductions.count(Phi)) {
1318 InductionInfo II = Inductions[Phi];
1319 if (IK_PtrInduction == II.IK)
1321 else if (IK_ReversePtrInduction == II.IK)
1325 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1329 unsigned NumOperands = Gep->getNumOperands();
1330 Value *GpPtr = Gep->getPointerOperand();
1331 // If this GEP value is a consecutive pointer induction variable and all of
1332 // the indices are constant then we know it is consecutive. We can
1333 Phi = dyn_cast<PHINode>(GpPtr);
1334 if (Phi && Inductions.count(Phi)) {
1336 // Make sure that the pointer does not point to structs.
1337 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1338 if (GepPtrType->getElementType()->isAggregateType())
1341 // Make sure that all of the index operands are loop invariant.
1342 for (unsigned i = 1; i < NumOperands; ++i)
1343 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1346 InductionInfo II = Inductions[Phi];
1347 if (IK_PtrInduction == II.IK)
1349 else if (IK_ReversePtrInduction == II.IK)
1353 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1355 // Check that all of the gep indices are uniform except for our induction
1357 for (unsigned i = 0; i != NumOperands; ++i)
1358 if (i != InductionOperand &&
1359 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1362 // We can emit wide load/stores only if the last non-zero index is the
1363 // induction variable.
1364 const SCEV *Last = 0;
1365 if (!Strides.count(Gep))
1366 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1368 // Because of the multiplication by a stride we can have a s/zext cast.
1369 // We are going to replace this stride by 1 so the cast is safe to ignore.
1371 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1372 // %0 = trunc i64 %indvars.iv to i32
1373 // %mul = mul i32 %0, %Stride1
1374 // %idxprom = zext i32 %mul to i64 << Safe cast.
1375 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1377 Last = replaceSymbolicStrideSCEV(SE, Strides,
1378 Gep->getOperand(InductionOperand), Gep);
1379 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1381 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1385 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1386 const SCEV *Step = AR->getStepRecurrence(*SE);
1388 // The memory is consecutive because the last index is consecutive
1389 // and all other indices are loop invariant.
1392 if (Step->isAllOnesValue())
1399 bool LoopVectorizationLegality::isUniform(Value *V) {
1400 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1403 InnerLoopVectorizer::VectorParts&
1404 InnerLoopVectorizer::getVectorValue(Value *V) {
1405 assert(V != Induction && "The new induction variable should not be used.");
1406 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1408 // If we have a stride that is replaced by one, do it here.
1409 if (Legal->hasStride(V))
1410 V = ConstantInt::get(V->getType(), 1);
1412 // If we have this scalar in the map, return it.
1413 if (WidenMap.has(V))
1414 return WidenMap.get(V);
1416 // If this scalar is unknown, assume that it is a constant or that it is
1417 // loop invariant. Broadcast V and save the value for future uses.
1418 Value *B = getBroadcastInstrs(V);
1419 return WidenMap.splat(V, B);
1422 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1423 assert(Vec->getType()->isVectorTy() && "Invalid type");
1424 SmallVector<Constant*, 8> ShuffleMask;
1425 for (unsigned i = 0; i < VF; ++i)
1426 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1428 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1429 ConstantVector::get(ShuffleMask),
1433 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1434 // Attempt to issue a wide load.
1435 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1436 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1438 assert((LI || SI) && "Invalid Load/Store instruction");
1440 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1441 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1442 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1443 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1444 // An alignment of 0 means target abi alignment. We need to use the scalar's
1445 // target abi alignment in such a case.
1447 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1448 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1449 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1450 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1452 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1453 return scalarizeInstruction(Instr, true);
1455 if (ScalarAllocatedSize != VectorElementSize)
1456 return scalarizeInstruction(Instr);
1458 // If the pointer is loop invariant or if it is non-consecutive,
1459 // scalarize the load.
1460 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1461 bool Reverse = ConsecutiveStride < 0;
1462 bool UniformLoad = LI && Legal->isUniform(Ptr);
1463 if (!ConsecutiveStride || UniformLoad)
1464 return scalarizeInstruction(Instr);
1466 Constant *Zero = Builder.getInt32(0);
1467 VectorParts &Entry = WidenMap.get(Instr);
1469 // Handle consecutive loads/stores.
1470 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1471 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1472 setDebugLocFromInst(Builder, Gep);
1473 Value *PtrOperand = Gep->getPointerOperand();
1474 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1475 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1477 // Create the new GEP with the new induction variable.
1478 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1479 Gep2->setOperand(0, FirstBasePtr);
1480 Gep2->setName("gep.indvar.base");
1481 Ptr = Builder.Insert(Gep2);
1483 setDebugLocFromInst(Builder, Gep);
1484 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1485 OrigLoop) && "Base ptr must be invariant");
1487 // The last index does not have to be the induction. It can be
1488 // consecutive and be a function of the index. For example A[I+1];
1489 unsigned NumOperands = Gep->getNumOperands();
1490 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1491 // Create the new GEP with the new induction variable.
1492 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1494 for (unsigned i = 0; i < NumOperands; ++i) {
1495 Value *GepOperand = Gep->getOperand(i);
1496 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1498 // Update last index or loop invariant instruction anchored in loop.
1499 if (i == InductionOperand ||
1500 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1501 assert((i == InductionOperand ||
1502 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1503 "Must be last index or loop invariant");
1505 VectorParts &GEPParts = getVectorValue(GepOperand);
1506 Value *Index = GEPParts[0];
1507 Index = Builder.CreateExtractElement(Index, Zero);
1508 Gep2->setOperand(i, Index);
1509 Gep2->setName("gep.indvar.idx");
1512 Ptr = Builder.Insert(Gep2);
1514 // Use the induction element ptr.
1515 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1516 setDebugLocFromInst(Builder, Ptr);
1517 VectorParts &PtrVal = getVectorValue(Ptr);
1518 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1523 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1524 "We do not allow storing to uniform addresses");
1525 setDebugLocFromInst(Builder, SI);
1526 // We don't want to update the value in the map as it might be used in
1527 // another expression. So don't use a reference type for "StoredVal".
1528 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1530 for (unsigned Part = 0; Part < UF; ++Part) {
1531 // Calculate the pointer for the specific unroll-part.
1532 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1535 // If we store to reverse consecutive memory locations then we need
1536 // to reverse the order of elements in the stored value.
1537 StoredVal[Part] = reverseVector(StoredVal[Part]);
1538 // If the address is consecutive but reversed, then the
1539 // wide store needs to start at the last vector element.
1540 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1541 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1544 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1545 DataTy->getPointerTo(AddressSpace));
1546 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1552 assert(LI && "Must have a load instruction");
1553 setDebugLocFromInst(Builder, LI);
1554 for (unsigned Part = 0; Part < UF; ++Part) {
1555 // Calculate the pointer for the specific unroll-part.
1556 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1559 // If the address is consecutive but reversed, then the
1560 // wide store needs to start at the last vector element.
1561 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1562 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1565 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1566 DataTy->getPointerTo(AddressSpace));
1567 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1568 cast<LoadInst>(LI)->setAlignment(Alignment);
1569 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1573 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1574 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1575 // Holds vector parameters or scalars, in case of uniform vals.
1576 SmallVector<VectorParts, 4> Params;
1578 setDebugLocFromInst(Builder, Instr);
1580 // Find all of the vectorized parameters.
1581 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1582 Value *SrcOp = Instr->getOperand(op);
1584 // If we are accessing the old induction variable, use the new one.
1585 if (SrcOp == OldInduction) {
1586 Params.push_back(getVectorValue(SrcOp));
1590 // Try using previously calculated values.
1591 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1593 // If the src is an instruction that appeared earlier in the basic block
1594 // then it should already be vectorized.
1595 if (SrcInst && OrigLoop->contains(SrcInst)) {
1596 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1597 // The parameter is a vector value from earlier.
1598 Params.push_back(WidenMap.get(SrcInst));
1600 // The parameter is a scalar from outside the loop. Maybe even a constant.
1601 VectorParts Scalars;
1602 Scalars.append(UF, SrcOp);
1603 Params.push_back(Scalars);
1607 assert(Params.size() == Instr->getNumOperands() &&
1608 "Invalid number of operands");
1610 // Does this instruction return a value ?
1611 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1613 Value *UndefVec = IsVoidRetTy ? 0 :
1614 UndefValue::get(VectorType::get(Instr->getType(), VF));
1615 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1616 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1618 Instruction *InsertPt = Builder.GetInsertPoint();
1619 BasicBlock *IfBlock = Builder.GetInsertBlock();
1620 BasicBlock *CondBlock = 0;
1624 if (IfPredicateStore) {
1625 assert(Instr->getParent()->getSinglePredecessor() &&
1626 "Only support single predecessor blocks");
1627 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1628 Instr->getParent());
1629 VectorLp = LI->getLoopFor(IfBlock);
1630 assert(VectorLp && "Must have a loop for this block");
1633 // For each vector unroll 'part':
1634 for (unsigned Part = 0; Part < UF; ++Part) {
1635 // For each scalar that we create:
1636 for (unsigned Width = 0; Width < VF; ++Width) {
1640 if (IfPredicateStore) {
1641 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1642 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1643 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1644 LoopVectorBody.push_back(CondBlock);
1645 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1646 // Update Builder with newly created basic block.
1647 Builder.SetInsertPoint(InsertPt);
1650 Instruction *Cloned = Instr->clone();
1652 Cloned->setName(Instr->getName() + ".cloned");
1653 // Replace the operands of the cloned instructions with extracted scalars.
1654 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1655 Value *Op = Params[op][Part];
1656 // Param is a vector. Need to extract the right lane.
1657 if (Op->getType()->isVectorTy())
1658 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1659 Cloned->setOperand(op, Op);
1662 // Place the cloned scalar in the new loop.
1663 Builder.Insert(Cloned);
1665 // If the original scalar returns a value we need to place it in a vector
1666 // so that future users will be able to use it.
1668 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1669 Builder.getInt32(Width));
1671 if (IfPredicateStore) {
1672 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1673 LoopVectorBody.push_back(NewIfBlock);
1674 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1675 Builder.SetInsertPoint(InsertPt);
1676 Instruction *OldBr = IfBlock->getTerminator();
1677 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1678 OldBr->eraseFromParent();
1679 IfBlock = NewIfBlock;
1685 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1689 if (Instruction *I = dyn_cast<Instruction>(V))
1690 return I->getParent() == Loc->getParent() ? I : 0;
1694 std::pair<Instruction *, Instruction *>
1695 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1696 Instruction *tnullptr = 0;
1697 if (!Legal->mustCheckStrides())
1698 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1700 IRBuilder<> ChkBuilder(Loc);
1704 Instruction *FirstInst = 0;
1705 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1706 SE = Legal->strides_end();
1708 Value *Ptr = stripIntegerCast(*SI);
1709 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1711 // Store the first instruction we create.
1712 FirstInst = getFirstInst(FirstInst, C, Loc);
1714 Check = ChkBuilder.CreateOr(Check, C);
1719 // We have to do this trickery because the IRBuilder might fold the check to a
1720 // constant expression in which case there is no Instruction anchored in a
1722 LLVMContext &Ctx = Loc->getContext();
1723 Instruction *TheCheck =
1724 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1725 ChkBuilder.Insert(TheCheck, "stride.not.one");
1726 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1728 return std::make_pair(FirstInst, TheCheck);
1731 std::pair<Instruction *, Instruction *>
1732 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1733 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1734 Legal->getRuntimePointerCheck();
1736 Instruction *tnullptr = 0;
1737 if (!PtrRtCheck->Need)
1738 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1740 unsigned NumPointers = PtrRtCheck->Pointers.size();
1741 SmallVector<TrackingVH<Value> , 2> Starts;
1742 SmallVector<TrackingVH<Value> , 2> Ends;
1744 LLVMContext &Ctx = Loc->getContext();
1745 SCEVExpander Exp(*SE, "induction");
1746 Instruction *FirstInst = 0;
1748 for (unsigned i = 0; i < NumPointers; ++i) {
1749 Value *Ptr = PtrRtCheck->Pointers[i];
1750 const SCEV *Sc = SE->getSCEV(Ptr);
1752 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1753 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1755 Starts.push_back(Ptr);
1756 Ends.push_back(Ptr);
1758 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1759 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1761 // Use this type for pointer arithmetic.
1762 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1764 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1765 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1766 Starts.push_back(Start);
1767 Ends.push_back(End);
1771 IRBuilder<> ChkBuilder(Loc);
1772 // Our instructions might fold to a constant.
1773 Value *MemoryRuntimeCheck = 0;
1774 for (unsigned i = 0; i < NumPointers; ++i) {
1775 for (unsigned j = i+1; j < NumPointers; ++j) {
1776 // No need to check if two readonly pointers intersect.
1777 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1780 // Only need to check pointers between two different dependency sets.
1781 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1784 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1785 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1787 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1788 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1789 "Trying to bounds check pointers with different address spaces");
1791 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1792 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1794 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1795 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1796 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1797 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1799 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1800 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1801 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1802 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1803 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1804 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1805 if (MemoryRuntimeCheck) {
1806 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1808 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1810 MemoryRuntimeCheck = IsConflict;
1814 // We have to do this trickery because the IRBuilder might fold the check to a
1815 // constant expression in which case there is no Instruction anchored in a
1817 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1818 ConstantInt::getTrue(Ctx));
1819 ChkBuilder.Insert(Check, "memcheck.conflict");
1820 FirstInst = getFirstInst(FirstInst, Check, Loc);
1821 return std::make_pair(FirstInst, Check);
1824 void InnerLoopVectorizer::createEmptyLoop() {
1826 In this function we generate a new loop. The new loop will contain
1827 the vectorized instructions while the old loop will continue to run the
1830 [ ] <-- vector loop bypass (may consist of multiple blocks).
1833 | [ ] <-- vector pre header.
1837 | [ ]_| <-- vector loop.
1840 >[ ] <--- middle-block.
1843 | [ ] <--- new preheader.
1847 | [ ]_| <-- old scalar loop to handle remainder.
1850 >[ ] <-- exit block.
1854 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1855 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1856 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1857 assert(ExitBlock && "Must have an exit block");
1859 // Some loops have a single integer induction variable, while other loops
1860 // don't. One example is c++ iterators that often have multiple pointer
1861 // induction variables. In the code below we also support a case where we
1862 // don't have a single induction variable.
1863 OldInduction = Legal->getInduction();
1864 Type *IdxTy = Legal->getWidestInductionType();
1866 // Find the loop boundaries.
1867 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1868 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1870 // The exit count might have the type of i64 while the phi is i32. This can
1871 // happen if we have an induction variable that is sign extended before the
1872 // compare. The only way that we get a backedge taken count is that the
1873 // induction variable was signed and as such will not overflow. In such a case
1874 // truncation is legal.
1875 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1876 IdxTy->getPrimitiveSizeInBits())
1877 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1879 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1880 // Get the total trip count from the count by adding 1.
1881 ExitCount = SE->getAddExpr(ExitCount,
1882 SE->getConstant(ExitCount->getType(), 1));
1884 // Expand the trip count and place the new instructions in the preheader.
1885 // Notice that the pre-header does not change, only the loop body.
1886 SCEVExpander Exp(*SE, "induction");
1888 // Count holds the overall loop count (N).
1889 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1890 BypassBlock->getTerminator());
1892 // The loop index does not have to start at Zero. Find the original start
1893 // value from the induction PHI node. If we don't have an induction variable
1894 // then we know that it starts at zero.
1895 Builder.SetInsertPoint(BypassBlock->getTerminator());
1896 Value *StartIdx = ExtendedIdx = OldInduction ?
1897 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1899 ConstantInt::get(IdxTy, 0);
1901 assert(BypassBlock && "Invalid loop structure");
1902 LoopBypassBlocks.push_back(BypassBlock);
1904 // Split the single block loop into the two loop structure described above.
1905 BasicBlock *VectorPH =
1906 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1907 BasicBlock *VecBody =
1908 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1909 BasicBlock *MiddleBlock =
1910 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1911 BasicBlock *ScalarPH =
1912 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1914 // Create and register the new vector loop.
1915 Loop* Lp = new Loop();
1916 Loop *ParentLoop = OrigLoop->getParentLoop();
1918 // Insert the new loop into the loop nest and register the new basic blocks
1919 // before calling any utilities such as SCEV that require valid LoopInfo.
1921 ParentLoop->addChildLoop(Lp);
1922 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
1923 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
1924 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
1926 LI->addTopLevelLoop(Lp);
1928 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
1930 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
1932 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
1934 // Generate the induction variable.
1935 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
1936 Induction = Builder.CreatePHI(IdxTy, 2, "index");
1937 // The loop step is equal to the vectorization factor (num of SIMD elements)
1938 // times the unroll factor (num of SIMD instructions).
1939 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
1941 // This is the IR builder that we use to add all of the logic for bypassing
1942 // the new vector loop.
1943 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
1944 setDebugLocFromInst(BypassBuilder,
1945 getDebugLocFromInstOrOperands(OldInduction));
1947 // We may need to extend the index in case there is a type mismatch.
1948 // We know that the count starts at zero and does not overflow.
1949 if (Count->getType() != IdxTy) {
1950 // The exit count can be of pointer type. Convert it to the correct
1952 if (ExitCount->getType()->isPointerTy())
1953 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
1955 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
1958 // Add the start index to the loop count to get the new end index.
1959 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
1961 // Now we need to generate the expression for N - (N % VF), which is
1962 // the part that the vectorized body will execute.
1963 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
1964 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
1965 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
1966 "end.idx.rnd.down");
1968 // Now, compare the new count to zero. If it is zero skip the vector loop and
1969 // jump to the scalar loop.
1970 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
1973 BasicBlock *LastBypassBlock = BypassBlock;
1975 // Generate the code to check that the strides we assumed to be one are really
1976 // one. We want the new basic block to start at the first instruction in a
1977 // sequence of instructions that form a check.
1978 Instruction *StrideCheck;
1979 Instruction *FirstCheckInst;
1980 std::tie(FirstCheckInst, StrideCheck) =
1981 addStrideCheck(BypassBlock->getTerminator());
1983 // Create a new block containing the stride check.
1984 BasicBlock *CheckBlock =
1985 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
1987 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
1988 LoopBypassBlocks.push_back(CheckBlock);
1990 // Replace the branch into the memory check block with a conditional branch
1991 // for the "few elements case".
1992 Instruction *OldTerm = BypassBlock->getTerminator();
1993 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
1994 OldTerm->eraseFromParent();
1997 LastBypassBlock = CheckBlock;
2000 // Generate the code that checks in runtime if arrays overlap. We put the
2001 // checks into a separate block to make the more common case of few elements
2003 Instruction *MemRuntimeCheck;
2004 std::tie(FirstCheckInst, MemRuntimeCheck) =
2005 addRuntimeCheck(LastBypassBlock->getTerminator());
2006 if (MemRuntimeCheck) {
2007 // Create a new block containing the memory check.
2008 BasicBlock *CheckBlock =
2009 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2011 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2012 LoopBypassBlocks.push_back(CheckBlock);
2014 // Replace the branch into the memory check block with a conditional branch
2015 // for the "few elements case".
2016 Instruction *OldTerm = LastBypassBlock->getTerminator();
2017 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2018 OldTerm->eraseFromParent();
2020 Cmp = MemRuntimeCheck;
2021 LastBypassBlock = CheckBlock;
2024 LastBypassBlock->getTerminator()->eraseFromParent();
2025 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2028 // We are going to resume the execution of the scalar loop.
2029 // Go over all of the induction variables that we found and fix the
2030 // PHIs that are left in the scalar version of the loop.
2031 // The starting values of PHI nodes depend on the counter of the last
2032 // iteration in the vectorized loop.
2033 // If we come from a bypass edge then we need to start from the original
2036 // This variable saves the new starting index for the scalar loop.
2037 PHINode *ResumeIndex = 0;
2038 LoopVectorizationLegality::InductionList::iterator I, E;
2039 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2040 // Set builder to point to last bypass block.
2041 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2042 for (I = List->begin(), E = List->end(); I != E; ++I) {
2043 PHINode *OrigPhi = I->first;
2044 LoopVectorizationLegality::InductionInfo II = I->second;
2046 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2047 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2048 MiddleBlock->getTerminator());
2049 // We might have extended the type of the induction variable but we need a
2050 // truncated version for the scalar loop.
2051 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2052 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2053 MiddleBlock->getTerminator()) : 0;
2055 Value *EndValue = 0;
2057 case LoopVectorizationLegality::IK_NoInduction:
2058 llvm_unreachable("Unknown induction");
2059 case LoopVectorizationLegality::IK_IntInduction: {
2060 // Handle the integer induction counter.
2061 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2063 // We have the canonical induction variable.
2064 if (OrigPhi == OldInduction) {
2065 // Create a truncated version of the resume value for the scalar loop,
2066 // we might have promoted the type to a larger width.
2068 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2069 // The new PHI merges the original incoming value, in case of a bypass,
2070 // or the value at the end of the vectorized loop.
2071 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2072 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2073 TruncResumeVal->addIncoming(EndValue, VecBody);
2075 // We know what the end value is.
2076 EndValue = IdxEndRoundDown;
2077 // We also know which PHI node holds it.
2078 ResumeIndex = ResumeVal;
2082 // Not the canonical induction variable - add the vector loop count to the
2084 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2085 II.StartValue->getType(),
2087 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2090 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2091 // Convert the CountRoundDown variable to the PHI size.
2092 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2093 II.StartValue->getType(),
2095 // Handle reverse integer induction counter.
2096 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2099 case LoopVectorizationLegality::IK_PtrInduction: {
2100 // For pointer induction variables, calculate the offset using
2102 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2106 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2107 // The value at the end of the loop for the reverse pointer is calculated
2108 // by creating a GEP with a negative index starting from the start value.
2109 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2110 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2112 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2118 // The new PHI merges the original incoming value, in case of a bypass,
2119 // or the value at the end of the vectorized loop.
2120 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2121 if (OrigPhi == OldInduction)
2122 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2124 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2126 ResumeVal->addIncoming(EndValue, VecBody);
2128 // Fix the scalar body counter (PHI node).
2129 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2130 // The old inductions phi node in the scalar body needs the truncated value.
2131 if (OrigPhi == OldInduction)
2132 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2134 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2137 // If we are generating a new induction variable then we also need to
2138 // generate the code that calculates the exit value. This value is not
2139 // simply the end of the counter because we may skip the vectorized body
2140 // in case of a runtime check.
2142 assert(!ResumeIndex && "Unexpected resume value found");
2143 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2144 MiddleBlock->getTerminator());
2145 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2146 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2147 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2150 // Make sure that we found the index where scalar loop needs to continue.
2151 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2152 "Invalid resume Index");
2154 // Add a check in the middle block to see if we have completed
2155 // all of the iterations in the first vector loop.
2156 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2157 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2158 ResumeIndex, "cmp.n",
2159 MiddleBlock->getTerminator());
2161 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2162 // Remove the old terminator.
2163 MiddleBlock->getTerminator()->eraseFromParent();
2165 // Create i+1 and fill the PHINode.
2166 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2167 Induction->addIncoming(StartIdx, VectorPH);
2168 Induction->addIncoming(NextIdx, VecBody);
2169 // Create the compare.
2170 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2171 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2173 // Now we have two terminators. Remove the old one from the block.
2174 VecBody->getTerminator()->eraseFromParent();
2176 // Get ready to start creating new instructions into the vectorized body.
2177 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2180 LoopVectorPreHeader = VectorPH;
2181 LoopScalarPreHeader = ScalarPH;
2182 LoopMiddleBlock = MiddleBlock;
2183 LoopExitBlock = ExitBlock;
2184 LoopVectorBody.push_back(VecBody);
2185 LoopScalarBody = OldBasicBlock;
2187 LoopVectorizeHints Hints(Lp, true);
2188 Hints.setAlreadyVectorized(Lp);
2191 /// This function returns the identity element (or neutral element) for
2192 /// the operation K.
2194 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2199 // Adding, Xoring, Oring zero to a number does not change it.
2200 return ConstantInt::get(Tp, 0);
2201 case RK_IntegerMult:
2202 // Multiplying a number by 1 does not change it.
2203 return ConstantInt::get(Tp, 1);
2205 // AND-ing a number with an all-1 value does not change it.
2206 return ConstantInt::get(Tp, -1, true);
2208 // Multiplying a number by 1 does not change it.
2209 return ConstantFP::get(Tp, 1.0L);
2211 // Adding zero to a number does not change it.
2212 return ConstantFP::get(Tp, 0.0L);
2214 llvm_unreachable("Unknown reduction kind");
2218 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2219 Intrinsic::ID ValidIntrinsicID) {
2220 if (I.getNumArgOperands() != 1 ||
2221 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2222 I.getType() != I.getArgOperand(0)->getType() ||
2223 !I.onlyReadsMemory())
2224 return Intrinsic::not_intrinsic;
2226 return ValidIntrinsicID;
2229 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2230 Intrinsic::ID ValidIntrinsicID) {
2231 if (I.getNumArgOperands() != 2 ||
2232 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2233 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2234 I.getType() != I.getArgOperand(0)->getType() ||
2235 I.getType() != I.getArgOperand(1)->getType() ||
2236 !I.onlyReadsMemory())
2237 return Intrinsic::not_intrinsic;
2239 return ValidIntrinsicID;
2243 static Intrinsic::ID
2244 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2245 // If we have an intrinsic call, check if it is trivially vectorizable.
2246 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2247 switch (II->getIntrinsicID()) {
2248 case Intrinsic::sqrt:
2249 case Intrinsic::sin:
2250 case Intrinsic::cos:
2251 case Intrinsic::exp:
2252 case Intrinsic::exp2:
2253 case Intrinsic::log:
2254 case Intrinsic::log10:
2255 case Intrinsic::log2:
2256 case Intrinsic::fabs:
2257 case Intrinsic::copysign:
2258 case Intrinsic::floor:
2259 case Intrinsic::ceil:
2260 case Intrinsic::trunc:
2261 case Intrinsic::rint:
2262 case Intrinsic::nearbyint:
2263 case Intrinsic::round:
2264 case Intrinsic::pow:
2265 case Intrinsic::fma:
2266 case Intrinsic::fmuladd:
2267 case Intrinsic::lifetime_start:
2268 case Intrinsic::lifetime_end:
2269 return II->getIntrinsicID();
2271 return Intrinsic::not_intrinsic;
2276 return Intrinsic::not_intrinsic;
2279 Function *F = CI->getCalledFunction();
2280 // We're going to make assumptions on the semantics of the functions, check
2281 // that the target knows that it's available in this environment and it does
2282 // not have local linkage.
2283 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2284 return Intrinsic::not_intrinsic;
2286 // Otherwise check if we have a call to a function that can be turned into a
2287 // vector intrinsic.
2294 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2298 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2302 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2304 case LibFunc::exp2f:
2305 case LibFunc::exp2l:
2306 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2310 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2311 case LibFunc::log10:
2312 case LibFunc::log10f:
2313 case LibFunc::log10l:
2314 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2316 case LibFunc::log2f:
2317 case LibFunc::log2l:
2318 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2320 case LibFunc::fabsf:
2321 case LibFunc::fabsl:
2322 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2323 case LibFunc::copysign:
2324 case LibFunc::copysignf:
2325 case LibFunc::copysignl:
2326 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2327 case LibFunc::floor:
2328 case LibFunc::floorf:
2329 case LibFunc::floorl:
2330 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2332 case LibFunc::ceilf:
2333 case LibFunc::ceill:
2334 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2335 case LibFunc::trunc:
2336 case LibFunc::truncf:
2337 case LibFunc::truncl:
2338 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2340 case LibFunc::rintf:
2341 case LibFunc::rintl:
2342 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2343 case LibFunc::nearbyint:
2344 case LibFunc::nearbyintf:
2345 case LibFunc::nearbyintl:
2346 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2347 case LibFunc::round:
2348 case LibFunc::roundf:
2349 case LibFunc::roundl:
2350 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2354 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2357 return Intrinsic::not_intrinsic;
2360 /// This function translates the reduction kind to an LLVM binary operator.
2362 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2364 case LoopVectorizationLegality::RK_IntegerAdd:
2365 return Instruction::Add;
2366 case LoopVectorizationLegality::RK_IntegerMult:
2367 return Instruction::Mul;
2368 case LoopVectorizationLegality::RK_IntegerOr:
2369 return Instruction::Or;
2370 case LoopVectorizationLegality::RK_IntegerAnd:
2371 return Instruction::And;
2372 case LoopVectorizationLegality::RK_IntegerXor:
2373 return Instruction::Xor;
2374 case LoopVectorizationLegality::RK_FloatMult:
2375 return Instruction::FMul;
2376 case LoopVectorizationLegality::RK_FloatAdd:
2377 return Instruction::FAdd;
2378 case LoopVectorizationLegality::RK_IntegerMinMax:
2379 return Instruction::ICmp;
2380 case LoopVectorizationLegality::RK_FloatMinMax:
2381 return Instruction::FCmp;
2383 llvm_unreachable("Unknown reduction operation");
2387 Value *createMinMaxOp(IRBuilder<> &Builder,
2388 LoopVectorizationLegality::MinMaxReductionKind RK,
2391 CmpInst::Predicate P = CmpInst::ICMP_NE;
2394 llvm_unreachable("Unknown min/max reduction kind");
2395 case LoopVectorizationLegality::MRK_UIntMin:
2396 P = CmpInst::ICMP_ULT;
2398 case LoopVectorizationLegality::MRK_UIntMax:
2399 P = CmpInst::ICMP_UGT;
2401 case LoopVectorizationLegality::MRK_SIntMin:
2402 P = CmpInst::ICMP_SLT;
2404 case LoopVectorizationLegality::MRK_SIntMax:
2405 P = CmpInst::ICMP_SGT;
2407 case LoopVectorizationLegality::MRK_FloatMin:
2408 P = CmpInst::FCMP_OLT;
2410 case LoopVectorizationLegality::MRK_FloatMax:
2411 P = CmpInst::FCMP_OGT;
2416 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2417 RK == LoopVectorizationLegality::MRK_FloatMax)
2418 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2420 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2422 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2427 struct CSEDenseMapInfo {
2428 static bool canHandle(Instruction *I) {
2429 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2430 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2432 static inline Instruction *getEmptyKey() {
2433 return DenseMapInfo<Instruction *>::getEmptyKey();
2435 static inline Instruction *getTombstoneKey() {
2436 return DenseMapInfo<Instruction *>::getTombstoneKey();
2438 static unsigned getHashValue(Instruction *I) {
2439 assert(canHandle(I) && "Unknown instruction!");
2440 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2441 I->value_op_end()));
2443 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2444 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2445 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2447 return LHS->isIdenticalTo(RHS);
2452 /// \brief Check whether this block is a predicated block.
2453 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2454 /// = ...; " blocks. We start with one vectorized basic block. For every
2455 /// conditional block we split this vectorized block. Therefore, every second
2456 /// block will be a predicated one.
2457 static bool isPredicatedBlock(unsigned BlockNum) {
2458 return BlockNum % 2;
2461 ///\brief Perform cse of induction variable instructions.
2462 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2463 // Perform simple cse.
2464 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2465 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2466 BasicBlock *BB = BBs[i];
2467 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2468 Instruction *In = I++;
2470 if (!CSEDenseMapInfo::canHandle(In))
2473 // Check if we can replace this instruction with any of the
2474 // visited instructions.
2475 if (Instruction *V = CSEMap.lookup(In)) {
2476 In->replaceAllUsesWith(V);
2477 In->eraseFromParent();
2480 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2481 // ...;" blocks for predicated stores. Every second block is a predicated
2483 if (isPredicatedBlock(i))
2491 void InnerLoopVectorizer::vectorizeLoop() {
2492 //===------------------------------------------------===//
2494 // Notice: any optimization or new instruction that go
2495 // into the code below should be also be implemented in
2498 //===------------------------------------------------===//
2499 Constant *Zero = Builder.getInt32(0);
2501 // In order to support reduction variables we need to be able to vectorize
2502 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2503 // stages. First, we create a new vector PHI node with no incoming edges.
2504 // We use this value when we vectorize all of the instructions that use the
2505 // PHI. Next, after all of the instructions in the block are complete we
2506 // add the new incoming edges to the PHI. At this point all of the
2507 // instructions in the basic block are vectorized, so we can use them to
2508 // construct the PHI.
2509 PhiVector RdxPHIsToFix;
2511 // Scan the loop in a topological order to ensure that defs are vectorized
2513 LoopBlocksDFS DFS(OrigLoop);
2516 // Vectorize all of the blocks in the original loop.
2517 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2518 be = DFS.endRPO(); bb != be; ++bb)
2519 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2521 // At this point every instruction in the original loop is widened to
2522 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2523 // that we vectorized. The PHI nodes are currently empty because we did
2524 // not want to introduce cycles. Notice that the remaining PHI nodes
2525 // that we need to fix are reduction variables.
2527 // Create the 'reduced' values for each of the induction vars.
2528 // The reduced values are the vector values that we scalarize and combine
2529 // after the loop is finished.
2530 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2532 PHINode *RdxPhi = *it;
2533 assert(RdxPhi && "Unable to recover vectorized PHI");
2535 // Find the reduction variable descriptor.
2536 assert(Legal->getReductionVars()->count(RdxPhi) &&
2537 "Unable to find the reduction variable");
2538 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2539 (*Legal->getReductionVars())[RdxPhi];
2541 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2543 // We need to generate a reduction vector from the incoming scalar.
2544 // To do so, we need to generate the 'identity' vector and override
2545 // one of the elements with the incoming scalar reduction. We need
2546 // to do it in the vector-loop preheader.
2547 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2549 // This is the vector-clone of the value that leaves the loop.
2550 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2551 Type *VecTy = VectorExit[0]->getType();
2553 // Find the reduction identity variable. Zero for addition, or, xor,
2554 // one for multiplication, -1 for And.
2557 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2558 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2559 // MinMax reduction have the start value as their identify.
2561 VectorStart = Identity = RdxDesc.StartValue;
2563 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2568 // Handle other reduction kinds:
2570 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2571 VecTy->getScalarType());
2574 // This vector is the Identity vector where the first element is the
2575 // incoming scalar reduction.
2576 VectorStart = RdxDesc.StartValue;
2578 Identity = ConstantVector::getSplat(VF, Iden);
2580 // This vector is the Identity vector where the first element is the
2581 // incoming scalar reduction.
2582 VectorStart = Builder.CreateInsertElement(Identity,
2583 RdxDesc.StartValue, Zero);
2587 // Fix the vector-loop phi.
2588 // We created the induction variable so we know that the
2589 // preheader is the first entry.
2590 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2592 // Reductions do not have to start at zero. They can start with
2593 // any loop invariant values.
2594 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2595 BasicBlock *Latch = OrigLoop->getLoopLatch();
2596 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2597 VectorParts &Val = getVectorValue(LoopVal);
2598 for (unsigned part = 0; part < UF; ++part) {
2599 // Make sure to add the reduction stat value only to the
2600 // first unroll part.
2601 Value *StartVal = (part == 0) ? VectorStart : Identity;
2602 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2603 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2604 LoopVectorBody.back());
2607 // Before each round, move the insertion point right between
2608 // the PHIs and the values we are going to write.
2609 // This allows us to write both PHINodes and the extractelement
2611 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2613 VectorParts RdxParts;
2614 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2615 for (unsigned part = 0; part < UF; ++part) {
2616 // This PHINode contains the vectorized reduction variable, or
2617 // the initial value vector, if we bypass the vector loop.
2618 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2619 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2620 Value *StartVal = (part == 0) ? VectorStart : Identity;
2621 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2622 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2623 NewPhi->addIncoming(RdxExitVal[part],
2624 LoopVectorBody.back());
2625 RdxParts.push_back(NewPhi);
2628 // Reduce all of the unrolled parts into a single vector.
2629 Value *ReducedPartRdx = RdxParts[0];
2630 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2631 setDebugLocFromInst(Builder, ReducedPartRdx);
2632 for (unsigned part = 1; part < UF; ++part) {
2633 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2634 ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
2635 RdxParts[part], ReducedPartRdx,
2638 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2639 ReducedPartRdx, RdxParts[part]);
2643 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2644 // and vector ops, reducing the set of values being computed by half each
2646 assert(isPowerOf2_32(VF) &&
2647 "Reduction emission only supported for pow2 vectors!");
2648 Value *TmpVec = ReducedPartRdx;
2649 SmallVector<Constant*, 32> ShuffleMask(VF, 0);
2650 for (unsigned i = VF; i != 1; i >>= 1) {
2651 // Move the upper half of the vector to the lower half.
2652 for (unsigned j = 0; j != i/2; ++j)
2653 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2655 // Fill the rest of the mask with undef.
2656 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2657 UndefValue::get(Builder.getInt32Ty()));
2660 Builder.CreateShuffleVector(TmpVec,
2661 UndefValue::get(TmpVec->getType()),
2662 ConstantVector::get(ShuffleMask),
2665 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2666 TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
2669 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2672 // The result is in the first element of the vector.
2673 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2674 Builder.getInt32(0));
2677 // Now, we need to fix the users of the reduction variable
2678 // inside and outside of the scalar remainder loop.
2679 // We know that the loop is in LCSSA form. We need to update the
2680 // PHI nodes in the exit blocks.
2681 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2682 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2683 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2684 if (!LCSSAPhi) break;
2686 // All PHINodes need to have a single entry edge, or two if
2687 // we already fixed them.
2688 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2690 // We found our reduction value exit-PHI. Update it with the
2691 // incoming bypass edge.
2692 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2693 // Add an edge coming from the bypass.
2694 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2697 }// end of the LCSSA phi scan.
2699 // Fix the scalar loop reduction variable with the incoming reduction sum
2700 // from the vector body and from the backedge value.
2701 int IncomingEdgeBlockIdx =
2702 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2703 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2704 // Pick the other block.
2705 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2706 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2707 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2708 }// end of for each redux variable.
2712 // Remove redundant induction instructions.
2713 cse(LoopVectorBody);
2716 void InnerLoopVectorizer::fixLCSSAPHIs() {
2717 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2718 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2719 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2720 if (!LCSSAPhi) break;
2721 if (LCSSAPhi->getNumIncomingValues() == 1)
2722 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2727 InnerLoopVectorizer::VectorParts
2728 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2729 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2732 // Look for cached value.
2733 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2734 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2735 if (ECEntryIt != MaskCache.end())
2736 return ECEntryIt->second;
2738 VectorParts SrcMask = createBlockInMask(Src);
2740 // The terminator has to be a branch inst!
2741 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2742 assert(BI && "Unexpected terminator found");
2744 if (BI->isConditional()) {
2745 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2747 if (BI->getSuccessor(0) != Dst)
2748 for (unsigned part = 0; part < UF; ++part)
2749 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2751 for (unsigned part = 0; part < UF; ++part)
2752 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2754 MaskCache[Edge] = EdgeMask;
2758 MaskCache[Edge] = SrcMask;
2762 InnerLoopVectorizer::VectorParts
2763 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2764 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2766 // Loop incoming mask is all-one.
2767 if (OrigLoop->getHeader() == BB) {
2768 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2769 return getVectorValue(C);
2772 // This is the block mask. We OR all incoming edges, and with zero.
2773 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2774 VectorParts BlockMask = getVectorValue(Zero);
2777 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2778 VectorParts EM = createEdgeMask(*it, BB);
2779 for (unsigned part = 0; part < UF; ++part)
2780 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2786 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2787 InnerLoopVectorizer::VectorParts &Entry,
2788 unsigned UF, unsigned VF, PhiVector *PV) {
2789 PHINode* P = cast<PHINode>(PN);
2790 // Handle reduction variables:
2791 if (Legal->getReductionVars()->count(P)) {
2792 for (unsigned part = 0; part < UF; ++part) {
2793 // This is phase one of vectorizing PHIs.
2794 Type *VecTy = (VF == 1) ? PN->getType() :
2795 VectorType::get(PN->getType(), VF);
2796 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2797 LoopVectorBody.back()-> getFirstInsertionPt());
2803 setDebugLocFromInst(Builder, P);
2804 // Check for PHI nodes that are lowered to vector selects.
2805 if (P->getParent() != OrigLoop->getHeader()) {
2806 // We know that all PHIs in non-header blocks are converted into
2807 // selects, so we don't have to worry about the insertion order and we
2808 // can just use the builder.
2809 // At this point we generate the predication tree. There may be
2810 // duplications since this is a simple recursive scan, but future
2811 // optimizations will clean it up.
2813 unsigned NumIncoming = P->getNumIncomingValues();
2815 // Generate a sequence of selects of the form:
2816 // SELECT(Mask3, In3,
2817 // SELECT(Mask2, In2,
2819 for (unsigned In = 0; In < NumIncoming; In++) {
2820 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2822 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2824 for (unsigned part = 0; part < UF; ++part) {
2825 // We might have single edge PHIs (blocks) - use an identity
2826 // 'select' for the first PHI operand.
2828 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2831 // Select between the current value and the previous incoming edge
2832 // based on the incoming mask.
2833 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2834 Entry[part], "predphi");
2840 // This PHINode must be an induction variable.
2841 // Make sure that we know about it.
2842 assert(Legal->getInductionVars()->count(P) &&
2843 "Not an induction variable");
2845 LoopVectorizationLegality::InductionInfo II =
2846 Legal->getInductionVars()->lookup(P);
2849 case LoopVectorizationLegality::IK_NoInduction:
2850 llvm_unreachable("Unknown induction");
2851 case LoopVectorizationLegality::IK_IntInduction: {
2852 assert(P->getType() == II.StartValue->getType() && "Types must match");
2853 Type *PhiTy = P->getType();
2855 if (P == OldInduction) {
2856 // Handle the canonical induction variable. We might have had to
2858 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2860 // Handle other induction variables that are now based on the
2862 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2864 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2865 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2868 Broadcasted = getBroadcastInstrs(Broadcasted);
2869 // After broadcasting the induction variable we need to make the vector
2870 // consecutive by adding 0, 1, 2, etc.
2871 for (unsigned part = 0; part < UF; ++part)
2872 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2875 case LoopVectorizationLegality::IK_ReverseIntInduction:
2876 case LoopVectorizationLegality::IK_PtrInduction:
2877 case LoopVectorizationLegality::IK_ReversePtrInduction:
2878 // Handle reverse integer and pointer inductions.
2879 Value *StartIdx = ExtendedIdx;
2880 // This is the normalized GEP that starts counting at zero.
2881 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2884 // Handle the reverse integer induction variable case.
2885 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2886 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2887 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2889 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2892 // This is a new value so do not hoist it out.
2893 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2894 // After broadcasting the induction variable we need to make the
2895 // vector consecutive by adding ... -3, -2, -1, 0.
2896 for (unsigned part = 0; part < UF; ++part)
2897 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2902 // Handle the pointer induction variable case.
2903 assert(P->getType()->isPointerTy() && "Unexpected type.");
2905 // Is this a reverse induction ptr or a consecutive induction ptr.
2906 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2909 // This is the vector of results. Notice that we don't generate
2910 // vector geps because scalar geps result in better code.
2911 for (unsigned part = 0; part < UF; ++part) {
2913 int EltIndex = (part) * (Reverse ? -1 : 1);
2914 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2917 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2919 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2921 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2923 Entry[part] = SclrGep;
2927 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2928 for (unsigned int i = 0; i < VF; ++i) {
2929 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2930 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2933 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2935 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2937 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2939 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2940 Builder.getInt32(i),
2943 Entry[part] = VecVal;
2949 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2950 // For each instruction in the old loop.
2951 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2952 VectorParts &Entry = WidenMap.get(it);
2953 switch (it->getOpcode()) {
2954 case Instruction::Br:
2955 // Nothing to do for PHIs and BR, since we already took care of the
2956 // loop control flow instructions.
2958 case Instruction::PHI:{
2959 // Vectorize PHINodes.
2960 widenPHIInstruction(it, Entry, UF, VF, PV);
2964 case Instruction::Add:
2965 case Instruction::FAdd:
2966 case Instruction::Sub:
2967 case Instruction::FSub:
2968 case Instruction::Mul:
2969 case Instruction::FMul:
2970 case Instruction::UDiv:
2971 case Instruction::SDiv:
2972 case Instruction::FDiv:
2973 case Instruction::URem:
2974 case Instruction::SRem:
2975 case Instruction::FRem:
2976 case Instruction::Shl:
2977 case Instruction::LShr:
2978 case Instruction::AShr:
2979 case Instruction::And:
2980 case Instruction::Or:
2981 case Instruction::Xor: {
2982 // Just widen binops.
2983 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2984 setDebugLocFromInst(Builder, BinOp);
2985 VectorParts &A = getVectorValue(it->getOperand(0));
2986 VectorParts &B = getVectorValue(it->getOperand(1));
2988 // Use this vector value for all users of the original instruction.
2989 for (unsigned Part = 0; Part < UF; ++Part) {
2990 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2992 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2993 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2994 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2995 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2996 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2998 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2999 VecOp->setIsExact(BinOp->isExact());
3005 case Instruction::Select: {
3007 // If the selector is loop invariant we can create a select
3008 // instruction with a scalar condition. Otherwise, use vector-select.
3009 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3011 setDebugLocFromInst(Builder, it);
3013 // The condition can be loop invariant but still defined inside the
3014 // loop. This means that we can't just use the original 'cond' value.
3015 // We have to take the 'vectorized' value and pick the first lane.
3016 // Instcombine will make this a no-op.
3017 VectorParts &Cond = getVectorValue(it->getOperand(0));
3018 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3019 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3021 Value *ScalarCond = (VF == 1) ? Cond[0] :
3022 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3024 for (unsigned Part = 0; Part < UF; ++Part) {
3025 Entry[Part] = Builder.CreateSelect(
3026 InvariantCond ? ScalarCond : Cond[Part],
3033 case Instruction::ICmp:
3034 case Instruction::FCmp: {
3035 // Widen compares. Generate vector compares.
3036 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3037 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3038 setDebugLocFromInst(Builder, it);
3039 VectorParts &A = getVectorValue(it->getOperand(0));
3040 VectorParts &B = getVectorValue(it->getOperand(1));
3041 for (unsigned Part = 0; Part < UF; ++Part) {
3044 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3046 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3052 case Instruction::Store:
3053 case Instruction::Load:
3054 vectorizeMemoryInstruction(it);
3056 case Instruction::ZExt:
3057 case Instruction::SExt:
3058 case Instruction::FPToUI:
3059 case Instruction::FPToSI:
3060 case Instruction::FPExt:
3061 case Instruction::PtrToInt:
3062 case Instruction::IntToPtr:
3063 case Instruction::SIToFP:
3064 case Instruction::UIToFP:
3065 case Instruction::Trunc:
3066 case Instruction::FPTrunc:
3067 case Instruction::BitCast: {
3068 CastInst *CI = dyn_cast<CastInst>(it);
3069 setDebugLocFromInst(Builder, it);
3070 /// Optimize the special case where the source is the induction
3071 /// variable. Notice that we can only optimize the 'trunc' case
3072 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3073 /// c. other casts depend on pointer size.
3074 if (CI->getOperand(0) == OldInduction &&
3075 it->getOpcode() == Instruction::Trunc) {
3076 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3078 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3079 for (unsigned Part = 0; Part < UF; ++Part)
3080 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3083 /// Vectorize casts.
3084 Type *DestTy = (VF == 1) ? CI->getType() :
3085 VectorType::get(CI->getType(), VF);
3087 VectorParts &A = getVectorValue(it->getOperand(0));
3088 for (unsigned Part = 0; Part < UF; ++Part)
3089 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3093 case Instruction::Call: {
3094 // Ignore dbg intrinsics.
3095 if (isa<DbgInfoIntrinsic>(it))
3097 setDebugLocFromInst(Builder, it);
3099 Module *M = BB->getParent()->getParent();
3100 CallInst *CI = cast<CallInst>(it);
3101 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3102 assert(ID && "Not an intrinsic call!");
3104 case Intrinsic::lifetime_end:
3105 case Intrinsic::lifetime_start:
3106 scalarizeInstruction(it);
3109 for (unsigned Part = 0; Part < UF; ++Part) {
3110 SmallVector<Value *, 4> Args;
3111 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3112 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3113 Args.push_back(Arg[Part]);
3115 Type *Tys[] = {CI->getType()};
3117 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3119 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3120 Entry[Part] = Builder.CreateCall(F, Args);
3128 // All other instructions are unsupported. Scalarize them.
3129 scalarizeInstruction(it);
3132 }// end of for_each instr.
3135 void InnerLoopVectorizer::updateAnalysis() {
3136 // Forget the original basic block.
3137 SE->forgetLoop(OrigLoop);
3139 // Update the dominator tree information.
3140 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3141 "Entry does not dominate exit.");
3143 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3144 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3145 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3147 // Due to if predication of stores we might create a sequence of "if(pred)
3148 // a[i] = ...; " blocks.
3149 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3151 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3152 else if (isPredicatedBlock(i)) {
3153 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3155 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3159 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3160 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3161 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3162 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3164 DEBUG(DT->verifyDomTree());
3167 /// \brief Check whether it is safe to if-convert this phi node.
3169 /// Phi nodes with constant expressions that can trap are not safe to if
3171 static bool canIfConvertPHINodes(BasicBlock *BB) {
3172 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3173 PHINode *Phi = dyn_cast<PHINode>(I);
3176 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3177 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3184 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3185 if (!EnableIfConversion)
3188 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3190 // A list of pointers that we can safely read and write to.
3191 SmallPtrSet<Value *, 8> SafePointes;
3193 // Collect safe addresses.
3194 for (Loop::block_iterator BI = TheLoop->block_begin(),
3195 BE = TheLoop->block_end(); BI != BE; ++BI) {
3196 BasicBlock *BB = *BI;
3198 if (blockNeedsPredication(BB))
3201 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3202 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3203 SafePointes.insert(LI->getPointerOperand());
3204 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3205 SafePointes.insert(SI->getPointerOperand());
3209 // Collect the blocks that need predication.
3210 BasicBlock *Header = TheLoop->getHeader();
3211 for (Loop::block_iterator BI = TheLoop->block_begin(),
3212 BE = TheLoop->block_end(); BI != BE; ++BI) {
3213 BasicBlock *BB = *BI;
3215 // We don't support switch statements inside loops.
3216 if (!isa<BranchInst>(BB->getTerminator()))
3219 // We must be able to predicate all blocks that need to be predicated.
3220 if (blockNeedsPredication(BB)) {
3221 if (!blockCanBePredicated(BB, SafePointes))
3223 } else if (BB != Header && !canIfConvertPHINodes(BB))
3228 // We can if-convert this loop.
3232 bool LoopVectorizationLegality::canVectorize() {
3233 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3234 // be canonicalized.
3235 if (!TheLoop->getLoopPreheader())
3238 // We can only vectorize innermost loops.
3239 if (TheLoop->getSubLoopsVector().size())
3242 // We must have a single backedge.
3243 if (TheLoop->getNumBackEdges() != 1)
3246 // We must have a single exiting block.
3247 if (!TheLoop->getExitingBlock())
3250 // We need to have a loop header.
3251 DEBUG(dbgs() << "LV: Found a loop: " <<
3252 TheLoop->getHeader()->getName() << '\n');
3254 // Check if we can if-convert non-single-bb loops.
3255 unsigned NumBlocks = TheLoop->getNumBlocks();
3256 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3257 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3261 // ScalarEvolution needs to be able to find the exit count.
3262 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3263 if (ExitCount == SE->getCouldNotCompute()) {
3264 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3268 // Do not loop-vectorize loops with a tiny trip count.
3269 BasicBlock *Latch = TheLoop->getLoopLatch();
3270 unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
3271 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
3272 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
3273 "This loop is not worth vectorizing.\n");
3277 // Check if we can vectorize the instructions and CFG in this loop.
3278 if (!canVectorizeInstrs()) {
3279 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3283 // Go over each instruction and look at memory deps.
3284 if (!canVectorizeMemory()) {
3285 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3289 // Collect all of the variables that remain uniform after vectorization.
3290 collectLoopUniforms();
3292 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3293 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3296 // Okay! We can vectorize. At this point we don't have any other mem analysis
3297 // which may limit our maximum vectorization factor, so just return true with
3302 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3303 if (Ty->isPointerTy())
3304 return DL.getIntPtrType(Ty);
3306 // It is possible that char's or short's overflow when we ask for the loop's
3307 // trip count, work around this by changing the type size.
3308 if (Ty->getScalarSizeInBits() < 32)
3309 return Type::getInt32Ty(Ty->getContext());
3314 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3315 Ty0 = convertPointerToIntegerType(DL, Ty0);
3316 Ty1 = convertPointerToIntegerType(DL, Ty1);
3317 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3322 /// \brief Check that the instruction has outside loop users and is not an
3323 /// identified reduction variable.
3324 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3325 SmallPtrSet<Value *, 4> &Reductions) {
3326 // Reduction instructions are allowed to have exit users. All other
3327 // instructions must not have external users.
3328 if (!Reductions.count(Inst))
3329 //Check that all of the users of the loop are inside the BB.
3330 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
3332 Instruction *U = cast<Instruction>(*I);
3333 // This user may be a reduction exit value.
3334 if (!TheLoop->contains(U)) {
3335 DEBUG(dbgs() << "LV: Found an outside user for : " << *U << '\n');
3342 bool LoopVectorizationLegality::canVectorizeInstrs() {
3343 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3344 BasicBlock *Header = TheLoop->getHeader();
3346 // Look for the attribute signaling the absence of NaNs.
3347 Function &F = *Header->getParent();
3348 if (F.hasFnAttribute("no-nans-fp-math"))
3349 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3350 AttributeSet::FunctionIndex,
3351 "no-nans-fp-math").getValueAsString() == "true";
3353 // For each block in the loop.
3354 for (Loop::block_iterator bb = TheLoop->block_begin(),
3355 be = TheLoop->block_end(); bb != be; ++bb) {
3357 // Scan the instructions in the block and look for hazards.
3358 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3361 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3362 Type *PhiTy = Phi->getType();
3363 // Check that this PHI type is allowed.
3364 if (!PhiTy->isIntegerTy() &&
3365 !PhiTy->isFloatingPointTy() &&
3366 !PhiTy->isPointerTy()) {
3367 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3371 // If this PHINode is not in the header block, then we know that we
3372 // can convert it to select during if-conversion. No need to check if
3373 // the PHIs in this block are induction or reduction variables.
3374 if (*bb != Header) {
3375 // Check that this instruction has no outside users or is an
3376 // identified reduction value with an outside user.
3377 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3382 // We only allow if-converted PHIs with more than two incoming values.
3383 if (Phi->getNumIncomingValues() != 2) {
3384 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3388 // This is the value coming from the preheader.
3389 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3390 // Check if this is an induction variable.
3391 InductionKind IK = isInductionVariable(Phi);
3393 if (IK_NoInduction != IK) {
3394 // Get the widest type.
3396 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3398 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3400 // Int inductions are special because we only allow one IV.
3401 if (IK == IK_IntInduction) {
3402 // Use the phi node with the widest type as induction. Use the last
3403 // one if there are multiple (no good reason for doing this other
3404 // than it is expedient).
3405 if (!Induction || PhiTy == WidestIndTy)
3409 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3410 Inductions[Phi] = InductionInfo(StartValue, IK);
3412 // Until we explicitly handle the case of an induction variable with
3413 // an outside loop user we have to give up vectorizing this loop.
3414 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3420 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3421 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3424 if (AddReductionVar(Phi, RK_IntegerMult)) {
3425 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3428 if (AddReductionVar(Phi, RK_IntegerOr)) {
3429 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3432 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3433 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3436 if (AddReductionVar(Phi, RK_IntegerXor)) {
3437 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3440 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3441 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3444 if (AddReductionVar(Phi, RK_FloatMult)) {
3445 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3448 if (AddReductionVar(Phi, RK_FloatAdd)) {
3449 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3452 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3453 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3458 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3460 }// end of PHI handling
3462 // We still don't handle functions. However, we can ignore dbg intrinsic
3463 // calls and we do handle certain intrinsic and libm functions.
3464 CallInst *CI = dyn_cast<CallInst>(it);
3465 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3466 DEBUG(dbgs() << "LV: Found a call site.\n");
3470 // Check that the instruction return type is vectorizable.
3471 // Also, we can't vectorize extractelement instructions.
3472 if ((!VectorType::isValidElementType(it->getType()) &&
3473 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3474 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3478 // Check that the stored type is vectorizable.
3479 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3480 Type *T = ST->getValueOperand()->getType();
3481 if (!VectorType::isValidElementType(T))
3483 if (EnableMemAccessVersioning)
3484 collectStridedAcccess(ST);
3487 if (EnableMemAccessVersioning)
3488 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3489 collectStridedAcccess(LI);
3491 // Reduction instructions are allowed to have exit users.
3492 // All other instructions must not have external users.
3493 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3501 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3502 if (Inductions.empty())
3509 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3510 /// return the induction operand of the gep pointer.
3511 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3512 const DataLayout *DL, Loop *Lp) {
3513 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3517 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3519 // Check that all of the gep indices are uniform except for our induction
3521 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3522 if (i != InductionOperand &&
3523 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3525 return GEP->getOperand(InductionOperand);
3528 ///\brief Look for a cast use of the passed value.
3529 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3530 Value *UniqueCast = 0;
3531 for (Value::use_iterator UI = Ptr->use_begin(), UE = Ptr->use_end(); UI != UE;
3533 CastInst *CI = dyn_cast<CastInst>(*UI);
3534 if (CI && CI->getType() == Ty) {
3544 ///\brief Get the stride of a pointer access in a loop.
3545 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3546 /// pointer to the Value, or null otherwise.
3547 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3548 const DataLayout *DL, Loop *Lp) {
3549 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3550 if (!PtrTy || PtrTy->isAggregateType())
3553 // Try to remove a gep instruction to make the pointer (actually index at this
3554 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3555 // pointer, otherwise, we are analyzing the index.
3556 Value *OrigPtr = Ptr;
3558 // The size of the pointer access.
3559 int64_t PtrAccessSize = 1;
3561 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3562 const SCEV *V = SE->getSCEV(Ptr);
3566 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3567 V = C->getOperand();
3569 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3573 V = S->getStepRecurrence(*SE);
3577 // Strip off the size of access multiplication if we are still analyzing the
3579 if (OrigPtr == Ptr) {
3580 DL->getTypeAllocSize(PtrTy->getElementType());
3581 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3582 if (M->getOperand(0)->getSCEVType() != scConstant)
3585 const APInt &APStepVal =
3586 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3588 // Huge step value - give up.
3589 if (APStepVal.getBitWidth() > 64)
3592 int64_t StepVal = APStepVal.getSExtValue();
3593 if (PtrAccessSize != StepVal)
3595 V = M->getOperand(1);
3600 Type *StripedOffRecurrenceCast = 0;
3601 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3602 StripedOffRecurrenceCast = C->getType();
3603 V = C->getOperand();
3606 // Look for the loop invariant symbolic value.
3607 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3611 Value *Stride = U->getValue();
3612 if (!Lp->isLoopInvariant(Stride))
3615 // If we have stripped off the recurrence cast we have to make sure that we
3616 // return the value that is used in this loop so that we can replace it later.
3617 if (StripedOffRecurrenceCast)
3618 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3623 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3625 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3626 Ptr = LI->getPointerOperand();
3627 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3628 Ptr = SI->getPointerOperand();
3632 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3636 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3637 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3638 Strides[Ptr] = Stride;
3639 StrideSet.insert(Stride);
3642 void LoopVectorizationLegality::collectLoopUniforms() {
3643 // We now know that the loop is vectorizable!
3644 // Collect variables that will remain uniform after vectorization.
3645 std::vector<Value*> Worklist;
3646 BasicBlock *Latch = TheLoop->getLoopLatch();
3648 // Start with the conditional branch and walk up the block.
3649 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3651 while (Worklist.size()) {
3652 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3653 Worklist.pop_back();
3655 // Look at instructions inside this loop.
3656 // Stop when reaching PHI nodes.
3657 // TODO: we need to follow values all over the loop, not only in this block.
3658 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3661 // This is a known uniform.
3664 // Insert all operands.
3665 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3670 /// \brief Analyses memory accesses in a loop.
3672 /// Checks whether run time pointer checks are needed and builds sets for data
3673 /// dependence checking.
3674 class AccessAnalysis {
3676 /// \brief Read or write access location.
3677 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3678 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3680 /// \brief Set of potential dependent memory accesses.
3681 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3683 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3684 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3685 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3687 /// \brief Register a load and whether it is only read from.
3688 void addLoad(Value *Ptr, bool IsReadOnly) {
3689 Accesses.insert(MemAccessInfo(Ptr, false));
3691 ReadOnlyPtr.insert(Ptr);
3694 /// \brief Register a store.
3695 void addStore(Value *Ptr) {
3696 Accesses.insert(MemAccessInfo(Ptr, true));
3699 /// \brief Check whether we can check the pointers at runtime for
3700 /// non-intersection.
3701 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3702 unsigned &NumComparisons, ScalarEvolution *SE,
3703 Loop *TheLoop, ValueToValueMap &Strides,
3704 bool ShouldCheckStride = false);
3706 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3707 /// and builds sets of dependent accesses.
3708 void buildDependenceSets() {
3709 // Process read-write pointers first.
3710 processMemAccesses(false);
3711 // Next, process read pointers.
3712 processMemAccesses(true);
3715 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3717 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3718 void resetDepChecks() { CheckDeps.clear(); }
3720 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3723 typedef SetVector<MemAccessInfo> PtrAccessSet;
3724 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3726 /// \brief Go over all memory access or only the deferred ones if
3727 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3728 /// and build sets of dependency check candidates.
3729 void processMemAccesses(bool UseDeferred);
3731 /// Set of all accesses.
3732 PtrAccessSet Accesses;
3734 /// Set of access to check after all writes have been processed.
3735 PtrAccessSet DeferredAccesses;
3737 /// Map of pointers to last access encountered.
3738 UnderlyingObjToAccessMap ObjToLastAccess;
3740 /// Set of accesses that need a further dependence check.
3741 MemAccessInfoSet CheckDeps;
3743 /// Set of pointers that are read only.
3744 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3746 /// Set of underlying objects already written to.
3747 SmallPtrSet<Value*, 16> WriteObjects;
3749 const DataLayout *DL;
3751 /// Sets of potentially dependent accesses - members of one set share an
3752 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3753 /// dependence check.
3754 DepCandidates &DepCands;
3756 bool AreAllWritesIdentified;
3757 bool AreAllReadsIdentified;
3758 bool IsRTCheckNeeded;
3761 } // end anonymous namespace
3763 /// \brief Check whether a pointer can participate in a runtime bounds check.
3764 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3766 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3767 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3771 return AR->isAffine();
3774 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3775 /// the address space.
3776 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3777 const Loop *Lp, ValueToValueMap &StridesMap);
3779 bool AccessAnalysis::canCheckPtrAtRT(
3780 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3781 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3782 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3783 // Find pointers with computable bounds. We are going to use this information
3784 // to place a runtime bound check.
3785 unsigned NumReadPtrChecks = 0;
3786 unsigned NumWritePtrChecks = 0;
3787 bool CanDoRT = true;
3789 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3790 // We assign consecutive id to access from different dependence sets.
3791 // Accesses within the same set don't need a runtime check.
3792 unsigned RunningDepId = 1;
3793 DenseMap<Value *, unsigned> DepSetId;
3795 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3797 const MemAccessInfo &Access = *AI;
3798 Value *Ptr = Access.getPointer();
3799 bool IsWrite = Access.getInt();
3801 // Just add write checks if we have both.
3802 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3806 ++NumWritePtrChecks;
3810 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3811 // When we run after a failing dependency check we have to make sure we
3812 // don't have wrapping pointers.
3813 (!ShouldCheckStride ||
3814 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3815 // The id of the dependence set.
3818 if (IsDepCheckNeeded) {
3819 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3820 unsigned &LeaderId = DepSetId[Leader];
3822 LeaderId = RunningDepId++;
3825 // Each access has its own dependence set.
3826 DepId = RunningDepId++;
3828 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3830 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3836 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3837 NumComparisons = 0; // Only one dependence set.
3839 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3840 NumWritePtrChecks - 1));
3843 // If the pointers that we would use for the bounds comparison have different
3844 // address spaces, assume the values aren't directly comparable, so we can't
3845 // use them for the runtime check. We also have to assume they could
3846 // overlap. In the future there should be metadata for whether address spaces
3848 unsigned NumPointers = RtCheck.Pointers.size();
3849 for (unsigned i = 0; i < NumPointers; ++i) {
3850 for (unsigned j = i + 1; j < NumPointers; ++j) {
3851 // Only need to check pointers between two different dependency sets.
3852 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3855 Value *PtrI = RtCheck.Pointers[i];
3856 Value *PtrJ = RtCheck.Pointers[j];
3858 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3859 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3861 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3862 " different address spaces\n");
3871 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3872 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3875 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3876 // We process the set twice: first we process read-write pointers, last we
3877 // process read-only pointers. This allows us to skip dependence tests for
3878 // read-only pointers.
3880 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3881 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3882 const MemAccessInfo &Access = *AI;
3883 Value *Ptr = Access.getPointer();
3884 bool IsWrite = Access.getInt();
3886 DepCands.insert(Access);
3888 // Memorize read-only pointers for later processing and skip them in the
3889 // first round (they need to be checked after we have seen all write
3890 // pointers). Note: we also mark pointer that are not consecutive as
3891 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3892 // second check for "!IsWrite".
3893 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3894 if (!UseDeferred && IsReadOnlyPtr) {
3895 DeferredAccesses.insert(Access);
3899 bool NeedDepCheck = false;
3900 // Check whether there is the possibility of dependency because of
3901 // underlying objects being the same.
3902 typedef SmallVector<Value*, 16> ValueVector;
3903 ValueVector TempObjects;
3904 GetUnderlyingObjects(Ptr, TempObjects, DL);
3905 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3907 Value *UnderlyingObj = *UI;
3909 // If this is a write then it needs to be an identified object. If this a
3910 // read and all writes (so far) are identified function scope objects we
3911 // don't need an identified underlying object but only an Argument (the
3912 // next write is going to invalidate this assumption if it is
3914 // This is a micro-optimization for the case where all writes are
3915 // identified and we have one argument pointer.
3916 // Otherwise, we do need a runtime check.
3917 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3918 (!IsWrite && (!AreAllWritesIdentified ||
3919 !isa<Argument>(UnderlyingObj)) &&
3920 !isIdentifiedObject(UnderlyingObj))) {
3921 DEBUG(dbgs() << "LV: Found an unidentified " <<
3922 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3924 IsRTCheckNeeded = (IsRTCheckNeeded ||
3925 !isIdentifiedObject(UnderlyingObj) ||
3926 !AreAllReadsIdentified);
3929 AreAllWritesIdentified = false;
3931 AreAllReadsIdentified = false;
3934 // If this is a write - check other reads and writes for conflicts. If
3935 // this is a read only check other writes for conflicts (but only if there
3936 // is no other write to the ptr - this is an optimization to catch "a[i] =
3937 // a[i] + " without having to do a dependence check).
3938 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3939 NeedDepCheck = true;
3942 WriteObjects.insert(UnderlyingObj);
3944 // Create sets of pointers connected by shared underlying objects.
3945 UnderlyingObjToAccessMap::iterator Prev =
3946 ObjToLastAccess.find(UnderlyingObj);
3947 if (Prev != ObjToLastAccess.end())
3948 DepCands.unionSets(Access, Prev->second);
3950 ObjToLastAccess[UnderlyingObj] = Access;
3954 CheckDeps.insert(Access);
3959 /// \brief Checks memory dependences among accesses to the same underlying
3960 /// object to determine whether there vectorization is legal or not (and at
3961 /// which vectorization factor).
3963 /// This class works under the assumption that we already checked that memory
3964 /// locations with different underlying pointers are "must-not alias".
3965 /// We use the ScalarEvolution framework to symbolically evalutate access
3966 /// functions pairs. Since we currently don't restructure the loop we can rely
3967 /// on the program order of memory accesses to determine their safety.
3968 /// At the moment we will only deem accesses as safe for:
3969 /// * A negative constant distance assuming program order.
3971 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3972 /// a[i] = tmp; y = a[i];
3974 /// The latter case is safe because later checks guarantuee that there can't
3975 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3976 /// the same variable: a header phi can only be an induction or a reduction, a
3977 /// reduction can't have a memory sink, an induction can't have a memory
3978 /// source). This is important and must not be violated (or we have to
3979 /// resort to checking for cycles through memory).
3981 /// * A positive constant distance assuming program order that is bigger
3982 /// than the biggest memory access.
3984 /// tmp = a[i] OR b[i] = x
3985 /// a[i+2] = tmp y = b[i+2];
3987 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3989 /// * Zero distances and all accesses have the same size.
3991 class MemoryDepChecker {
3993 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3994 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3996 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
3997 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3998 ShouldRetryWithRuntimeCheck(false) {}
4000 /// \brief Register the location (instructions are given increasing numbers)
4001 /// of a write access.
4002 void addAccess(StoreInst *SI) {
4003 Value *Ptr = SI->getPointerOperand();
4004 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4005 InstMap.push_back(SI);
4009 /// \brief Register the location (instructions are given increasing numbers)
4010 /// of a write access.
4011 void addAccess(LoadInst *LI) {
4012 Value *Ptr = LI->getPointerOperand();
4013 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4014 InstMap.push_back(LI);
4018 /// \brief Check whether the dependencies between the accesses are safe.
4020 /// Only checks sets with elements in \p CheckDeps.
4021 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4022 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4024 /// \brief The maximum number of bytes of a vector register we can vectorize
4025 /// the accesses safely with.
4026 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4028 /// \brief In same cases when the dependency check fails we can still
4029 /// vectorize the loop with a dynamic array access check.
4030 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4033 ScalarEvolution *SE;
4034 const DataLayout *DL;
4035 const Loop *InnermostLoop;
4037 /// \brief Maps access locations (ptr, read/write) to program order.
4038 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4040 /// \brief Memory access instructions in program order.
4041 SmallVector<Instruction *, 16> InstMap;
4043 /// \brief The program order index to be used for the next instruction.
4046 // We can access this many bytes in parallel safely.
4047 unsigned MaxSafeDepDistBytes;
4049 /// \brief If we see a non-constant dependence distance we can still try to
4050 /// vectorize this loop with runtime checks.
4051 bool ShouldRetryWithRuntimeCheck;
4053 /// \brief Check whether there is a plausible dependence between the two
4056 /// Access \p A must happen before \p B in program order. The two indices
4057 /// identify the index into the program order map.
4059 /// This function checks whether there is a plausible dependence (or the
4060 /// absence of such can't be proved) between the two accesses. If there is a
4061 /// plausible dependence but the dependence distance is bigger than one
4062 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4063 /// distance is smaller than any other distance encountered so far).
4064 /// Otherwise, this function returns true signaling a possible dependence.
4065 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4066 const MemAccessInfo &B, unsigned BIdx,
4067 ValueToValueMap &Strides);
4069 /// \brief Check whether the data dependence could prevent store-load
4071 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4074 } // end anonymous namespace
4076 static bool isInBoundsGep(Value *Ptr) {
4077 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4078 return GEP->isInBounds();
4082 /// \brief Check whether the access through \p Ptr has a constant stride.
4083 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4084 const Loop *Lp, ValueToValueMap &StridesMap) {
4085 const Type *Ty = Ptr->getType();
4086 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4088 // Make sure that the pointer does not point to aggregate types.
4089 const PointerType *PtrTy = cast<PointerType>(Ty);
4090 if (PtrTy->getElementType()->isAggregateType()) {
4091 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4096 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4098 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4100 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4101 << *Ptr << " SCEV: " << *PtrScev << "\n");
4105 // The accesss function must stride over the innermost loop.
4106 if (Lp != AR->getLoop()) {
4107 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4108 *Ptr << " SCEV: " << *PtrScev << "\n");
4111 // The address calculation must not wrap. Otherwise, a dependence could be
4113 // An inbounds getelementptr that is a AddRec with a unit stride
4114 // cannot wrap per definition. The unit stride requirement is checked later.
4115 // An getelementptr without an inbounds attribute and unit stride would have
4116 // to access the pointer value "0" which is undefined behavior in address
4117 // space 0, therefore we can also vectorize this case.
4118 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4119 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4120 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4121 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4122 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4123 << *Ptr << " SCEV: " << *PtrScev << "\n");
4127 // Check the step is constant.
4128 const SCEV *Step = AR->getStepRecurrence(*SE);
4130 // Calculate the pointer stride and check if it is consecutive.
4131 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4133 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4134 " SCEV: " << *PtrScev << "\n");
4138 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4139 const APInt &APStepVal = C->getValue()->getValue();
4141 // Huge step value - give up.
4142 if (APStepVal.getBitWidth() > 64)
4145 int64_t StepVal = APStepVal.getSExtValue();
4148 int64_t Stride = StepVal / Size;
4149 int64_t Rem = StepVal % Size;
4153 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4154 // know we can't "wrap around the address space". In case of address space
4155 // zero we know that this won't happen without triggering undefined behavior.
4156 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4157 Stride != 1 && Stride != -1)
4163 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4164 unsigned TypeByteSize) {
4165 // If loads occur at a distance that is not a multiple of a feasible vector
4166 // factor store-load forwarding does not take place.
4167 // Positive dependences might cause troubles because vectorizing them might
4168 // prevent store-load forwarding making vectorized code run a lot slower.
4169 // a[i] = a[i-3] ^ a[i-8];
4170 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4171 // hence on your typical architecture store-load forwarding does not take
4172 // place. Vectorizing in such cases does not make sense.
4173 // Store-load forwarding distance.
4174 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4175 // Maximum vector factor.
4176 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4177 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4178 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4180 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4182 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4183 MaxVFWithoutSLForwardIssues = (vf >>=1);
4188 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4189 DEBUG(dbgs() << "LV: Distance " << Distance <<
4190 " that could cause a store-load forwarding conflict\n");
4194 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4195 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4196 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4200 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4201 const MemAccessInfo &B, unsigned BIdx,
4202 ValueToValueMap &Strides) {
4203 assert (AIdx < BIdx && "Must pass arguments in program order");
4205 Value *APtr = A.getPointer();
4206 Value *BPtr = B.getPointer();
4207 bool AIsWrite = A.getInt();
4208 bool BIsWrite = B.getInt();
4210 // Two reads are independent.
4211 if (!AIsWrite && !BIsWrite)
4214 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4215 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4217 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4218 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4220 const SCEV *Src = AScev;
4221 const SCEV *Sink = BScev;
4223 // If the induction step is negative we have to invert source and sink of the
4225 if (StrideAPtr < 0) {
4228 std::swap(APtr, BPtr);
4229 std::swap(Src, Sink);
4230 std::swap(AIsWrite, BIsWrite);
4231 std::swap(AIdx, BIdx);
4232 std::swap(StrideAPtr, StrideBPtr);
4235 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4237 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4238 << "(Induction step: " << StrideAPtr << ")\n");
4239 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4240 << *InstMap[BIdx] << ": " << *Dist << "\n");
4242 // Need consecutive accesses. We don't want to vectorize
4243 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4244 // the address space.
4245 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4246 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4250 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4252 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4253 ShouldRetryWithRuntimeCheck = true;
4257 Type *ATy = APtr->getType()->getPointerElementType();
4258 Type *BTy = BPtr->getType()->getPointerElementType();
4259 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4261 // Negative distances are not plausible dependencies.
4262 const APInt &Val = C->getValue()->getValue();
4263 if (Val.isNegative()) {
4264 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4265 if (IsTrueDataDependence &&
4266 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4270 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4274 // Write to the same location with the same size.
4275 // Could be improved to assert type sizes are the same (i32 == float, etc).
4279 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4283 assert(Val.isStrictlyPositive() && "Expect a positive value");
4285 // Positive distance bigger than max vectorization factor.
4288 "LV: ReadWrite-Write positive dependency with different types\n");
4292 unsigned Distance = (unsigned) Val.getZExtValue();
4294 // Bail out early if passed-in parameters make vectorization not feasible.
4295 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4296 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4298 // The distance must be bigger than the size needed for a vectorized version
4299 // of the operation and the size of the vectorized operation must not be
4300 // bigger than the currrent maximum size.
4301 if (Distance < 2*TypeByteSize ||
4302 2*TypeByteSize > MaxSafeDepDistBytes ||
4303 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4304 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4305 << Val.getSExtValue() << '\n');
4309 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4310 Distance : MaxSafeDepDistBytes;
4312 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4313 if (IsTrueDataDependence &&
4314 couldPreventStoreLoadForward(Distance, TypeByteSize))
4317 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4318 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4323 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4324 MemAccessInfoSet &CheckDeps,
4325 ValueToValueMap &Strides) {
4327 MaxSafeDepDistBytes = -1U;
4328 while (!CheckDeps.empty()) {
4329 MemAccessInfo CurAccess = *CheckDeps.begin();
4331 // Get the relevant memory access set.
4332 EquivalenceClasses<MemAccessInfo>::iterator I =
4333 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4335 // Check accesses within this set.
4336 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4337 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4339 // Check every access pair.
4341 CheckDeps.erase(*AI);
4342 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4344 // Check every accessing instruction pair in program order.
4345 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4346 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4347 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4348 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4349 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4351 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4362 bool LoopVectorizationLegality::canVectorizeMemory() {
4364 typedef SmallVector<Value*, 16> ValueVector;
4365 typedef SmallPtrSet<Value*, 16> ValueSet;
4367 // Holds the Load and Store *instructions*.
4371 // Holds all the different accesses in the loop.
4372 unsigned NumReads = 0;
4373 unsigned NumReadWrites = 0;
4375 PtrRtCheck.Pointers.clear();
4376 PtrRtCheck.Need = false;
4378 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4379 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4382 for (Loop::block_iterator bb = TheLoop->block_begin(),
4383 be = TheLoop->block_end(); bb != be; ++bb) {
4385 // Scan the BB and collect legal loads and stores.
4386 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4389 // If this is a load, save it. If this instruction can read from memory
4390 // but is not a load, then we quit. Notice that we don't handle function
4391 // calls that read or write.
4392 if (it->mayReadFromMemory()) {
4393 // Many math library functions read the rounding mode. We will only
4394 // vectorize a loop if it contains known function calls that don't set
4395 // the flag. Therefore, it is safe to ignore this read from memory.
4396 CallInst *Call = dyn_cast<CallInst>(it);
4397 if (Call && getIntrinsicIDForCall(Call, TLI))
4400 LoadInst *Ld = dyn_cast<LoadInst>(it);
4401 if (!Ld) return false;
4402 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4403 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4407 Loads.push_back(Ld);
4408 DepChecker.addAccess(Ld);
4412 // Save 'store' instructions. Abort if other instructions write to memory.
4413 if (it->mayWriteToMemory()) {
4414 StoreInst *St = dyn_cast<StoreInst>(it);
4415 if (!St) return false;
4416 if (!St->isSimple() && !IsAnnotatedParallel) {
4417 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4421 Stores.push_back(St);
4422 DepChecker.addAccess(St);
4427 // Now we have two lists that hold the loads and the stores.
4428 // Next, we find the pointers that they use.
4430 // Check if we see any stores. If there are no stores, then we don't
4431 // care if the pointers are *restrict*.
4432 if (!Stores.size()) {
4433 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4437 AccessAnalysis::DepCandidates DependentAccesses;
4438 AccessAnalysis Accesses(DL, DependentAccesses);
4440 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4441 // multiple times on the same object. If the ptr is accessed twice, once
4442 // for read and once for write, it will only appear once (on the write
4443 // list). This is okay, since we are going to check for conflicts between
4444 // writes and between reads and writes, but not between reads and reads.
4447 ValueVector::iterator I, IE;
4448 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4449 StoreInst *ST = cast<StoreInst>(*I);
4450 Value* Ptr = ST->getPointerOperand();
4452 if (isUniform(Ptr)) {
4453 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4457 // If we did *not* see this pointer before, insert it to the read-write
4458 // list. At this phase it is only a 'write' list.
4459 if (Seen.insert(Ptr)) {
4461 Accesses.addStore(Ptr);
4465 if (IsAnnotatedParallel) {
4467 << "LV: A loop annotated parallel, ignore memory dependency "
4472 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4473 LoadInst *LD = cast<LoadInst>(*I);
4474 Value* Ptr = LD->getPointerOperand();
4475 // If we did *not* see this pointer before, insert it to the
4476 // read list. If we *did* see it before, then it is already in
4477 // the read-write list. This allows us to vectorize expressions
4478 // such as A[i] += x; Because the address of A[i] is a read-write
4479 // pointer. This only works if the index of A[i] is consecutive.
4480 // If the address of i is unknown (for example A[B[i]]) then we may
4481 // read a few words, modify, and write a few words, and some of the
4482 // words may be written to the same address.
4483 bool IsReadOnlyPtr = false;
4484 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4486 IsReadOnlyPtr = true;
4488 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4491 // If we write (or read-write) to a single destination and there are no
4492 // other reads in this loop then is it safe to vectorize.
4493 if (NumReadWrites == 1 && NumReads == 0) {
4494 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4498 // Build dependence sets and check whether we need a runtime pointer bounds
4500 Accesses.buildDependenceSets();
4501 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4503 // Find pointers with computable bounds. We are going to use this information
4504 // to place a runtime bound check.
4505 unsigned NumComparisons = 0;
4506 bool CanDoRT = false;
4508 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4511 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4512 " pointer comparisons.\n");
4514 // If we only have one set of dependences to check pointers among we don't
4515 // need a runtime check.
4516 if (NumComparisons == 0 && NeedRTCheck)
4517 NeedRTCheck = false;
4519 // Check that we did not collect too many pointers or found an unsizeable
4521 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4527 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4530 if (NeedRTCheck && !CanDoRT) {
4531 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4532 "the array bounds.\n");
4537 PtrRtCheck.Need = NeedRTCheck;
4539 bool CanVecMem = true;
4540 if (Accesses.isDependencyCheckNeeded()) {
4541 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4542 CanVecMem = DepChecker.areDepsSafe(
4543 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4544 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4546 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4547 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4550 // Clear the dependency checks. We assume they are not needed.
4551 Accesses.resetDepChecks();
4554 PtrRtCheck.Need = true;
4556 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4557 TheLoop, Strides, true);
4558 // Check that we did not collect too many pointers or found an unsizeable
4560 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4561 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4570 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4571 " need a runtime memory check.\n");
4576 static bool hasMultipleUsesOf(Instruction *I,
4577 SmallPtrSet<Instruction *, 8> &Insts) {
4578 unsigned NumUses = 0;
4579 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4580 if (Insts.count(dyn_cast<Instruction>(*Use)))
4589 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4590 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4591 if (!Set.count(dyn_cast<Instruction>(*Use)))
4596 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4597 ReductionKind Kind) {
4598 if (Phi->getNumIncomingValues() != 2)
4601 // Reduction variables are only found in the loop header block.
4602 if (Phi->getParent() != TheLoop->getHeader())
4605 // Obtain the reduction start value from the value that comes from the loop
4607 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4609 // ExitInstruction is the single value which is used outside the loop.
4610 // We only allow for a single reduction value to be used outside the loop.
4611 // This includes users of the reduction, variables (which form a cycle
4612 // which ends in the phi node).
4613 Instruction *ExitInstruction = 0;
4614 // Indicates that we found a reduction operation in our scan.
4615 bool FoundReduxOp = false;
4617 // We start with the PHI node and scan for all of the users of this
4618 // instruction. All users must be instructions that can be used as reduction
4619 // variables (such as ADD). We must have a single out-of-block user. The cycle
4620 // must include the original PHI.
4621 bool FoundStartPHI = false;
4623 // To recognize min/max patterns formed by a icmp select sequence, we store
4624 // the number of instruction we saw from the recognized min/max pattern,
4625 // to make sure we only see exactly the two instructions.
4626 unsigned NumCmpSelectPatternInst = 0;
4627 ReductionInstDesc ReduxDesc(false, 0);
4629 SmallPtrSet<Instruction *, 8> VisitedInsts;
4630 SmallVector<Instruction *, 8> Worklist;
4631 Worklist.push_back(Phi);
4632 VisitedInsts.insert(Phi);
4634 // A value in the reduction can be used:
4635 // - By the reduction:
4636 // - Reduction operation:
4637 // - One use of reduction value (safe).
4638 // - Multiple use of reduction value (not safe).
4640 // - All uses of the PHI must be the reduction (safe).
4641 // - Otherwise, not safe.
4642 // - By one instruction outside of the loop (safe).
4643 // - By further instructions outside of the loop (not safe).
4644 // - By an instruction that is not part of the reduction (not safe).
4646 // * An instruction type other than PHI or the reduction operation.
4647 // * A PHI in the header other than the initial PHI.
4648 while (!Worklist.empty()) {
4649 Instruction *Cur = Worklist.back();
4650 Worklist.pop_back();
4653 // If the instruction has no users then this is a broken chain and can't be
4654 // a reduction variable.
4655 if (Cur->use_empty())
4658 bool IsAPhi = isa<PHINode>(Cur);
4660 // A header PHI use other than the original PHI.
4661 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4664 // Reductions of instructions such as Div, and Sub is only possible if the
4665 // LHS is the reduction variable.
4666 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4667 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4668 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4671 // Any reduction instruction must be of one of the allowed kinds.
4672 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4673 if (!ReduxDesc.IsReduction)
4676 // A reduction operation must only have one use of the reduction value.
4677 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4678 hasMultipleUsesOf(Cur, VisitedInsts))
4681 // All inputs to a PHI node must be a reduction value.
4682 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4685 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4686 isa<SelectInst>(Cur)))
4687 ++NumCmpSelectPatternInst;
4688 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4689 isa<SelectInst>(Cur)))
4690 ++NumCmpSelectPatternInst;
4692 // Check whether we found a reduction operator.
4693 FoundReduxOp |= !IsAPhi;
4695 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4696 // onto the stack. This way we are going to have seen all inputs to PHI
4697 // nodes once we get to them.
4698 SmallVector<Instruction *, 8> NonPHIs;
4699 SmallVector<Instruction *, 8> PHIs;
4700 for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
4702 Instruction *Usr = cast<Instruction>(*UI);
4704 // Check if we found the exit user.
4705 BasicBlock *Parent = Usr->getParent();
4706 if (!TheLoop->contains(Parent)) {
4707 // Exit if you find multiple outside users or if the header phi node is
4708 // being used. In this case the user uses the value of the previous
4709 // iteration, in which case we would loose "VF-1" iterations of the
4710 // reduction operation if we vectorize.
4711 if (ExitInstruction != 0 || Cur == Phi)
4714 // The instruction used by an outside user must be the last instruction
4715 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4716 // operations on the value.
4717 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4720 ExitInstruction = Cur;
4724 // Process instructions only once (termination). Each reduction cycle
4725 // value must only be used once, except by phi nodes and min/max
4726 // reductions which are represented as a cmp followed by a select.
4727 ReductionInstDesc IgnoredVal(false, 0);
4728 if (VisitedInsts.insert(Usr)) {
4729 if (isa<PHINode>(Usr))
4730 PHIs.push_back(Usr);
4732 NonPHIs.push_back(Usr);
4733 } else if (!isa<PHINode>(Usr) &&
4734 ((!isa<FCmpInst>(Usr) &&
4735 !isa<ICmpInst>(Usr) &&
4736 !isa<SelectInst>(Usr)) ||
4737 !isMinMaxSelectCmpPattern(Usr, IgnoredVal).IsReduction))
4740 // Remember that we completed the cycle.
4742 FoundStartPHI = true;
4744 Worklist.append(PHIs.begin(), PHIs.end());
4745 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4748 // This means we have seen one but not the other instruction of the
4749 // pattern or more than just a select and cmp.
4750 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4751 NumCmpSelectPatternInst != 2)
4754 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4757 // We found a reduction var if we have reached the original phi node and we
4758 // only have a single instruction with out-of-loop users.
4760 // This instruction is allowed to have out-of-loop users.
4761 AllowedExit.insert(ExitInstruction);
4763 // Save the description of this reduction variable.
4764 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4765 ReduxDesc.MinMaxKind);
4766 Reductions[Phi] = RD;
4767 // We've ended the cycle. This is a reduction variable if we have an
4768 // outside user and it has a binary op.
4773 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4774 /// pattern corresponding to a min(X, Y) or max(X, Y).
4775 LoopVectorizationLegality::ReductionInstDesc
4776 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4777 ReductionInstDesc &Prev) {
4779 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4780 "Expect a select instruction");
4781 Instruction *Cmp = 0;
4782 SelectInst *Select = 0;
4784 // We must handle the select(cmp()) as a single instruction. Advance to the
4786 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4787 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
4788 return ReductionInstDesc(false, I);
4789 return ReductionInstDesc(Select, Prev.MinMaxKind);
4792 // Only handle single use cases for now.
4793 if (!(Select = dyn_cast<SelectInst>(I)))
4794 return ReductionInstDesc(false, I);
4795 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4796 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4797 return ReductionInstDesc(false, I);
4798 if (!Cmp->hasOneUse())
4799 return ReductionInstDesc(false, I);
4804 // Look for a min/max pattern.
4805 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4806 return ReductionInstDesc(Select, MRK_UIntMin);
4807 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4808 return ReductionInstDesc(Select, MRK_UIntMax);
4809 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4810 return ReductionInstDesc(Select, MRK_SIntMax);
4811 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4812 return ReductionInstDesc(Select, MRK_SIntMin);
4813 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4814 return ReductionInstDesc(Select, MRK_FloatMin);
4815 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4816 return ReductionInstDesc(Select, MRK_FloatMax);
4817 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4818 return ReductionInstDesc(Select, MRK_FloatMin);
4819 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4820 return ReductionInstDesc(Select, MRK_FloatMax);
4822 return ReductionInstDesc(false, I);
4825 LoopVectorizationLegality::ReductionInstDesc
4826 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4828 ReductionInstDesc &Prev) {
4829 bool FP = I->getType()->isFloatingPointTy();
4830 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4831 switch (I->getOpcode()) {
4833 return ReductionInstDesc(false, I);
4834 case Instruction::PHI:
4835 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4836 Kind != RK_FloatMinMax))
4837 return ReductionInstDesc(false, I);
4838 return ReductionInstDesc(I, Prev.MinMaxKind);
4839 case Instruction::Sub:
4840 case Instruction::Add:
4841 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4842 case Instruction::Mul:
4843 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4844 case Instruction::And:
4845 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4846 case Instruction::Or:
4847 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4848 case Instruction::Xor:
4849 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4850 case Instruction::FMul:
4851 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4852 case Instruction::FAdd:
4853 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4854 case Instruction::FCmp:
4855 case Instruction::ICmp:
4856 case Instruction::Select:
4857 if (Kind != RK_IntegerMinMax &&
4858 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4859 return ReductionInstDesc(false, I);
4860 return isMinMaxSelectCmpPattern(I, Prev);
4864 LoopVectorizationLegality::InductionKind
4865 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4866 Type *PhiTy = Phi->getType();
4867 // We only handle integer and pointer inductions variables.
4868 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4869 return IK_NoInduction;
4871 // Check that the PHI is consecutive.
4872 const SCEV *PhiScev = SE->getSCEV(Phi);
4873 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4875 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4876 return IK_NoInduction;
4878 const SCEV *Step = AR->getStepRecurrence(*SE);
4880 // Integer inductions need to have a stride of one.
4881 if (PhiTy->isIntegerTy()) {
4883 return IK_IntInduction;
4884 if (Step->isAllOnesValue())
4885 return IK_ReverseIntInduction;
4886 return IK_NoInduction;
4889 // Calculate the pointer stride and check if it is consecutive.
4890 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4892 return IK_NoInduction;
4894 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4895 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4896 if (C->getValue()->equalsInt(Size))
4897 return IK_PtrInduction;
4898 else if (C->getValue()->equalsInt(0 - Size))
4899 return IK_ReversePtrInduction;
4901 return IK_NoInduction;
4904 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4905 Value *In0 = const_cast<Value*>(V);
4906 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4910 return Inductions.count(PN);
4913 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4914 assert(TheLoop->contains(BB) && "Unknown block used");
4916 // Blocks that do not dominate the latch need predication.
4917 BasicBlock* Latch = TheLoop->getLoopLatch();
4918 return !DT->dominates(BB, Latch);
4921 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4922 SmallPtrSet<Value *, 8>& SafePtrs) {
4923 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4924 // We might be able to hoist the load.
4925 if (it->mayReadFromMemory()) {
4926 LoadInst *LI = dyn_cast<LoadInst>(it);
4927 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4931 // We don't predicate stores at the moment.
4932 if (it->mayWriteToMemory()) {
4933 StoreInst *SI = dyn_cast<StoreInst>(it);
4934 // We only support predication of stores in basic blocks with one
4936 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
4937 !SafePtrs.count(SI->getPointerOperand()) ||
4938 !SI->getParent()->getSinglePredecessor())
4944 // Check that we don't have a constant expression that can trap as operand.
4945 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4947 if (Constant *C = dyn_cast<Constant>(*OI))
4952 // The instructions below can trap.
4953 switch (it->getOpcode()) {
4955 case Instruction::UDiv:
4956 case Instruction::SDiv:
4957 case Instruction::URem:
4958 case Instruction::SRem:
4966 LoopVectorizationCostModel::VectorizationFactor
4967 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4969 // Width 1 means no vectorize
4970 VectorizationFactor Factor = { 1U, 0U };
4971 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4972 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4976 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
4977 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4981 // Find the trip count.
4982 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4983 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4985 unsigned WidestType = getWidestType();
4986 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4987 unsigned MaxSafeDepDist = -1U;
4988 if (Legal->getMaxSafeDepDistBytes() != -1U)
4989 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4990 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4991 WidestRegister : MaxSafeDepDist);
4992 unsigned MaxVectorSize = WidestRegister / WidestType;
4993 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4994 DEBUG(dbgs() << "LV: The Widest register is: "
4995 << WidestRegister << " bits.\n");
4997 if (MaxVectorSize == 0) {
4998 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5002 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5003 " into one vector!");
5005 unsigned VF = MaxVectorSize;
5007 // If we optimize the program for size, avoid creating the tail loop.
5009 // If we are unable to calculate the trip count then don't try to vectorize.
5011 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5015 // Find the maximum SIMD width that can fit within the trip count.
5016 VF = TC % MaxVectorSize;
5021 // If the trip count that we found modulo the vectorization factor is not
5022 // zero then we require a tail.
5024 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5030 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5031 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5033 Factor.Width = UserVF;
5037 float Cost = expectedCost(1);
5039 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)Cost << ".\n");
5040 for (unsigned i=2; i <= VF; i*=2) {
5041 // Notice that the vector loop needs to be executed less times, so
5042 // we need to divide the cost of the vector loops by the width of
5043 // the vector elements.
5044 float VectorCost = expectedCost(i) / (float)i;
5045 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5046 (int)VectorCost << ".\n");
5047 if (VectorCost < Cost) {
5053 DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
5054 Factor.Width = Width;
5055 Factor.Cost = Width * Cost;
5059 unsigned LoopVectorizationCostModel::getWidestType() {
5060 unsigned MaxWidth = 8;
5063 for (Loop::block_iterator bb = TheLoop->block_begin(),
5064 be = TheLoop->block_end(); bb != be; ++bb) {
5065 BasicBlock *BB = *bb;
5067 // For each instruction in the loop.
5068 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5069 Type *T = it->getType();
5071 // Only examine Loads, Stores and PHINodes.
5072 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5075 // Examine PHI nodes that are reduction variables.
5076 if (PHINode *PN = dyn_cast<PHINode>(it))
5077 if (!Legal->getReductionVars()->count(PN))
5080 // Examine the stored values.
5081 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5082 T = ST->getValueOperand()->getType();
5084 // Ignore loaded pointer types and stored pointer types that are not
5085 // consecutive. However, we do want to take consecutive stores/loads of
5086 // pointer vectors into account.
5087 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5090 MaxWidth = std::max(MaxWidth,
5091 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5099 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5102 unsigned LoopCost) {
5104 // -- The unroll heuristics --
5105 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5106 // There are many micro-architectural considerations that we can't predict
5107 // at this level. For example frontend pressure (on decode or fetch) due to
5108 // code size, or the number and capabilities of the execution ports.
5110 // We use the following heuristics to select the unroll factor:
5111 // 1. If the code has reductions the we unroll in order to break the cross
5112 // iteration dependency.
5113 // 2. If the loop is really small then we unroll in order to reduce the loop
5115 // 3. We don't unroll if we think that we will spill registers to memory due
5116 // to the increased register pressure.
5118 // Use the user preference, unless 'auto' is selected.
5122 // When we optimize for size we don't unroll.
5126 // We used the distance for the unroll factor.
5127 if (Legal->getMaxSafeDepDistBytes() != -1U)
5130 // Do not unroll loops with a relatively small trip count.
5131 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5132 TheLoop->getLoopLatch());
5133 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5136 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5137 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5141 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5142 TargetNumRegisters = ForceTargetNumScalarRegs;
5144 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5145 TargetNumRegisters = ForceTargetNumVectorRegs;
5148 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5149 // We divide by these constants so assume that we have at least one
5150 // instruction that uses at least one register.
5151 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5152 R.NumInstructions = std::max(R.NumInstructions, 1U);
5154 // We calculate the unroll factor using the following formula.
5155 // Subtract the number of loop invariants from the number of available
5156 // registers. These registers are used by all of the unrolled instances.
5157 // Next, divide the remaining registers by the number of registers that is
5158 // required by the loop, in order to estimate how many parallel instances
5159 // fit without causing spills. All of this is rounded down if necessary to be
5160 // a power of two. We want power of two unroll factors to simplify any
5161 // addressing operations or alignment considerations.
5162 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5165 // Don't count the induction variable as unrolled.
5166 if (EnableIndVarRegisterHeur)
5167 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5168 std::max(1U, (R.MaxLocalUsers - 1)));
5170 // Clamp the unroll factor ranges to reasonable factors.
5171 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5173 // Check if the user has overridden the unroll max.
5175 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5176 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5178 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5179 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5182 // If we did not calculate the cost for VF (because the user selected the VF)
5183 // then we calculate the cost of VF here.
5185 LoopCost = expectedCost(VF);
5187 // Clamp the calculated UF to be between the 1 and the max unroll factor
5188 // that the target allows.
5189 if (UF > MaxUnrollSize)
5194 // Unroll if we vectorized this loop and there is a reduction that could
5195 // benefit from unrolling.
5196 if (VF > 1 && Legal->getReductionVars()->size()) {
5197 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5201 // Note that if we've already vectorized the loop we will have done the
5202 // runtime check and so unrolling won't require further checks.
5203 bool UnrollingRequiresRuntimePointerCheck =
5204 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5206 // We want to unroll small loops in order to reduce the loop overhead and
5207 // potentially expose ILP opportunities.
5208 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5209 if (!UnrollingRequiresRuntimePointerCheck &&
5210 LoopCost < SmallLoopCost) {
5211 // We assume that the cost overhead is 1 and we use the cost model
5212 // to estimate the cost of the loop and unroll until the cost of the
5213 // loop overhead is about 5% of the cost of the loop.
5214 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5216 // Unroll until store/load ports (estimated by max unroll factor) are
5218 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5219 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5221 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5222 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5223 return std::max(StoresUF, LoadsUF);
5226 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5230 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5234 LoopVectorizationCostModel::RegisterUsage
5235 LoopVectorizationCostModel::calculateRegisterUsage() {
5236 // This function calculates the register usage by measuring the highest number
5237 // of values that are alive at a single location. Obviously, this is a very
5238 // rough estimation. We scan the loop in a topological order in order and
5239 // assign a number to each instruction. We use RPO to ensure that defs are
5240 // met before their users. We assume that each instruction that has in-loop
5241 // users starts an interval. We record every time that an in-loop value is
5242 // used, so we have a list of the first and last occurrences of each
5243 // instruction. Next, we transpose this data structure into a multi map that
5244 // holds the list of intervals that *end* at a specific location. This multi
5245 // map allows us to perform a linear search. We scan the instructions linearly
5246 // and record each time that a new interval starts, by placing it in a set.
5247 // If we find this value in the multi-map then we remove it from the set.
5248 // The max register usage is the maximum size of the set.
5249 // We also search for instructions that are defined outside the loop, but are
5250 // used inside the loop. We need this number separately from the max-interval
5251 // usage number because when we unroll, loop-invariant values do not take
5253 LoopBlocksDFS DFS(TheLoop);
5257 R.NumInstructions = 0;
5259 // Each 'key' in the map opens a new interval. The values
5260 // of the map are the index of the 'last seen' usage of the
5261 // instruction that is the key.
5262 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5263 // Maps instruction to its index.
5264 DenseMap<unsigned, Instruction*> IdxToInstr;
5265 // Marks the end of each interval.
5266 IntervalMap EndPoint;
5267 // Saves the list of instruction indices that are used in the loop.
5268 SmallSet<Instruction*, 8> Ends;
5269 // Saves the list of values that are used in the loop but are
5270 // defined outside the loop, such as arguments and constants.
5271 SmallPtrSet<Value*, 8> LoopInvariants;
5274 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5275 be = DFS.endRPO(); bb != be; ++bb) {
5276 R.NumInstructions += (*bb)->size();
5277 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5279 Instruction *I = it;
5280 IdxToInstr[Index++] = I;
5282 // Save the end location of each USE.
5283 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5284 Value *U = I->getOperand(i);
5285 Instruction *Instr = dyn_cast<Instruction>(U);
5287 // Ignore non-instruction values such as arguments, constants, etc.
5288 if (!Instr) continue;
5290 // If this instruction is outside the loop then record it and continue.
5291 if (!TheLoop->contains(Instr)) {
5292 LoopInvariants.insert(Instr);
5296 // Overwrite previous end points.
5297 EndPoint[Instr] = Index;
5303 // Saves the list of intervals that end with the index in 'key'.
5304 typedef SmallVector<Instruction*, 2> InstrList;
5305 DenseMap<unsigned, InstrList> TransposeEnds;
5307 // Transpose the EndPoints to a list of values that end at each index.
5308 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5310 TransposeEnds[it->second].push_back(it->first);
5312 SmallSet<Instruction*, 8> OpenIntervals;
5313 unsigned MaxUsage = 0;
5316 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5317 for (unsigned int i = 0; i < Index; ++i) {
5318 Instruction *I = IdxToInstr[i];
5319 // Ignore instructions that are never used within the loop.
5320 if (!Ends.count(I)) continue;
5322 // Remove all of the instructions that end at this location.
5323 InstrList &List = TransposeEnds[i];
5324 for (unsigned int j=0, e = List.size(); j < e; ++j)
5325 OpenIntervals.erase(List[j]);
5327 // Count the number of live interals.
5328 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5330 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5331 OpenIntervals.size() << '\n');
5333 // Add the current instruction to the list of open intervals.
5334 OpenIntervals.insert(I);
5337 unsigned Invariant = LoopInvariants.size();
5338 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5339 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5340 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5342 R.LoopInvariantRegs = Invariant;
5343 R.MaxLocalUsers = MaxUsage;
5347 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5351 for (Loop::block_iterator bb = TheLoop->block_begin(),
5352 be = TheLoop->block_end(); bb != be; ++bb) {
5353 unsigned BlockCost = 0;
5354 BasicBlock *BB = *bb;
5356 // For each instruction in the old loop.
5357 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5358 // Skip dbg intrinsics.
5359 if (isa<DbgInfoIntrinsic>(it))
5362 unsigned C = getInstructionCost(it, VF);
5364 // Check if we should override the cost.
5365 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5366 C = ForceTargetInstructionCost;
5369 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5370 VF << " For instruction: " << *it << '\n');
5373 // We assume that if-converted blocks have a 50% chance of being executed.
5374 // When the code is scalar then some of the blocks are avoided due to CF.
5375 // When the code is vectorized we execute all code paths.
5376 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5385 /// \brief Check whether the address computation for a non-consecutive memory
5386 /// access looks like an unlikely candidate for being merged into the indexing
5389 /// We look for a GEP which has one index that is an induction variable and all
5390 /// other indices are loop invariant. If the stride of this access is also
5391 /// within a small bound we decide that this address computation can likely be
5392 /// merged into the addressing mode.
5393 /// In all other cases, we identify the address computation as complex.
5394 static bool isLikelyComplexAddressComputation(Value *Ptr,
5395 LoopVectorizationLegality *Legal,
5396 ScalarEvolution *SE,
5397 const Loop *TheLoop) {
5398 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5402 // We are looking for a gep with all loop invariant indices except for one
5403 // which should be an induction variable.
5404 unsigned NumOperands = Gep->getNumOperands();
5405 for (unsigned i = 1; i < NumOperands; ++i) {
5406 Value *Opd = Gep->getOperand(i);
5407 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5408 !Legal->isInductionVariable(Opd))
5412 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5413 // can likely be merged into the address computation.
5414 unsigned MaxMergeDistance = 64;
5416 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5420 // Check the step is constant.
5421 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5422 // Calculate the pointer stride and check if it is consecutive.
5423 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5427 const APInt &APStepVal = C->getValue()->getValue();
5429 // Huge step value - give up.
5430 if (APStepVal.getBitWidth() > 64)
5433 int64_t StepVal = APStepVal.getSExtValue();
5435 return StepVal > MaxMergeDistance;
5438 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5439 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5445 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5446 // If we know that this instruction will remain uniform, check the cost of
5447 // the scalar version.
5448 if (Legal->isUniformAfterVectorization(I))
5451 Type *RetTy = I->getType();
5452 Type *VectorTy = ToVectorTy(RetTy, VF);
5454 // TODO: We need to estimate the cost of intrinsic calls.
5455 switch (I->getOpcode()) {
5456 case Instruction::GetElementPtr:
5457 // We mark this instruction as zero-cost because the cost of GEPs in
5458 // vectorized code depends on whether the corresponding memory instruction
5459 // is scalarized or not. Therefore, we handle GEPs with the memory
5460 // instruction cost.
5462 case Instruction::Br: {
5463 return TTI.getCFInstrCost(I->getOpcode());
5465 case Instruction::PHI:
5466 //TODO: IF-converted IFs become selects.
5468 case Instruction::Add:
5469 case Instruction::FAdd:
5470 case Instruction::Sub:
5471 case Instruction::FSub:
5472 case Instruction::Mul:
5473 case Instruction::FMul:
5474 case Instruction::UDiv:
5475 case Instruction::SDiv:
5476 case Instruction::FDiv:
5477 case Instruction::URem:
5478 case Instruction::SRem:
5479 case Instruction::FRem:
5480 case Instruction::Shl:
5481 case Instruction::LShr:
5482 case Instruction::AShr:
5483 case Instruction::And:
5484 case Instruction::Or:
5485 case Instruction::Xor: {
5486 // Since we will replace the stride by 1 the multiplication should go away.
5487 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5489 // Certain instructions can be cheaper to vectorize if they have a constant
5490 // second vector operand. One example of this are shifts on x86.
5491 TargetTransformInfo::OperandValueKind Op1VK =
5492 TargetTransformInfo::OK_AnyValue;
5493 TargetTransformInfo::OperandValueKind Op2VK =
5494 TargetTransformInfo::OK_AnyValue;
5495 Value *Op2 = I->getOperand(1);
5497 // Check for a splat of a constant or for a non uniform vector of constants.
5498 if (isa<ConstantInt>(Op2))
5499 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5500 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5501 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5502 if (cast<Constant>(Op2)->getSplatValue() != NULL)
5503 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5506 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5508 case Instruction::Select: {
5509 SelectInst *SI = cast<SelectInst>(I);
5510 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5511 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5512 Type *CondTy = SI->getCondition()->getType();
5514 CondTy = VectorType::get(CondTy, VF);
5516 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5518 case Instruction::ICmp:
5519 case Instruction::FCmp: {
5520 Type *ValTy = I->getOperand(0)->getType();
5521 VectorTy = ToVectorTy(ValTy, VF);
5522 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5524 case Instruction::Store:
5525 case Instruction::Load: {
5526 StoreInst *SI = dyn_cast<StoreInst>(I);
5527 LoadInst *LI = dyn_cast<LoadInst>(I);
5528 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5530 VectorTy = ToVectorTy(ValTy, VF);
5532 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5533 unsigned AS = SI ? SI->getPointerAddressSpace() :
5534 LI->getPointerAddressSpace();
5535 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5536 // We add the cost of address computation here instead of with the gep
5537 // instruction because only here we know whether the operation is
5540 return TTI.getAddressComputationCost(VectorTy) +
5541 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5543 // Scalarized loads/stores.
5544 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5545 bool Reverse = ConsecutiveStride < 0;
5546 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5547 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5548 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5549 bool IsComplexComputation =
5550 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5552 // The cost of extracting from the value vector and pointer vector.
5553 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5554 for (unsigned i = 0; i < VF; ++i) {
5555 // The cost of extracting the pointer operand.
5556 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5557 // In case of STORE, the cost of ExtractElement from the vector.
5558 // In case of LOAD, the cost of InsertElement into the returned
5560 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5561 Instruction::InsertElement,
5565 // The cost of the scalar loads/stores.
5566 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5567 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5572 // Wide load/stores.
5573 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5574 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5577 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5581 case Instruction::ZExt:
5582 case Instruction::SExt:
5583 case Instruction::FPToUI:
5584 case Instruction::FPToSI:
5585 case Instruction::FPExt:
5586 case Instruction::PtrToInt:
5587 case Instruction::IntToPtr:
5588 case Instruction::SIToFP:
5589 case Instruction::UIToFP:
5590 case Instruction::Trunc:
5591 case Instruction::FPTrunc:
5592 case Instruction::BitCast: {
5593 // We optimize the truncation of induction variable.
5594 // The cost of these is the same as the scalar operation.
5595 if (I->getOpcode() == Instruction::Trunc &&
5596 Legal->isInductionVariable(I->getOperand(0)))
5597 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5598 I->getOperand(0)->getType());
5600 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5601 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5603 case Instruction::Call: {
5604 CallInst *CI = cast<CallInst>(I);
5605 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5606 assert(ID && "Not an intrinsic call!");
5607 Type *RetTy = ToVectorTy(CI->getType(), VF);
5608 SmallVector<Type*, 4> Tys;
5609 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5610 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5611 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5614 // We are scalarizing the instruction. Return the cost of the scalar
5615 // instruction, plus the cost of insert and extract into vector
5616 // elements, times the vector width.
5619 if (!RetTy->isVoidTy() && VF != 1) {
5620 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5622 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5625 // The cost of inserting the results plus extracting each one of the
5627 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5630 // The cost of executing VF copies of the scalar instruction. This opcode
5631 // is unknown. Assume that it is the same as 'mul'.
5632 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5638 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5639 if (Scalar->isVoidTy() || VF == 1)
5641 return VectorType::get(Scalar, VF);
5644 char LoopVectorize::ID = 0;
5645 static const char lv_name[] = "Loop Vectorization";
5646 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5647 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5648 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5649 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5650 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5651 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5652 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5653 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5654 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5657 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5658 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5662 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5663 // Check for a store.
5664 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5665 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5667 // Check for a load.
5668 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5669 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5675 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5676 bool IfPredicateStore) {
5677 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5678 // Holds vector parameters or scalars, in case of uniform vals.
5679 SmallVector<VectorParts, 4> Params;
5681 setDebugLocFromInst(Builder, Instr);
5683 // Find all of the vectorized parameters.
5684 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5685 Value *SrcOp = Instr->getOperand(op);
5687 // If we are accessing the old induction variable, use the new one.
5688 if (SrcOp == OldInduction) {
5689 Params.push_back(getVectorValue(SrcOp));
5693 // Try using previously calculated values.
5694 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5696 // If the src is an instruction that appeared earlier in the basic block
5697 // then it should already be vectorized.
5698 if (SrcInst && OrigLoop->contains(SrcInst)) {
5699 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5700 // The parameter is a vector value from earlier.
5701 Params.push_back(WidenMap.get(SrcInst));
5703 // The parameter is a scalar from outside the loop. Maybe even a constant.
5704 VectorParts Scalars;
5705 Scalars.append(UF, SrcOp);
5706 Params.push_back(Scalars);
5710 assert(Params.size() == Instr->getNumOperands() &&
5711 "Invalid number of operands");
5713 // Does this instruction return a value ?
5714 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5716 Value *UndefVec = IsVoidRetTy ? 0 :
5717 UndefValue::get(Instr->getType());
5718 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5719 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5721 Instruction *InsertPt = Builder.GetInsertPoint();
5722 BasicBlock *IfBlock = Builder.GetInsertBlock();
5723 BasicBlock *CondBlock = 0;
5727 if (IfPredicateStore) {
5728 assert(Instr->getParent()->getSinglePredecessor() &&
5729 "Only support single predecessor blocks");
5730 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5731 Instr->getParent());
5732 VectorLp = LI->getLoopFor(IfBlock);
5733 assert(VectorLp && "Must have a loop for this block");
5736 // For each vector unroll 'part':
5737 for (unsigned Part = 0; Part < UF; ++Part) {
5738 // For each scalar that we create:
5740 // Start an "if (pred) a[i] = ..." block.
5742 if (IfPredicateStore) {
5743 if (Cond[Part]->getType()->isVectorTy())
5745 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5746 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5747 ConstantInt::get(Cond[Part]->getType(), 1));
5748 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5749 LoopVectorBody.push_back(CondBlock);
5750 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5751 // Update Builder with newly created basic block.
5752 Builder.SetInsertPoint(InsertPt);
5755 Instruction *Cloned = Instr->clone();
5757 Cloned->setName(Instr->getName() + ".cloned");
5758 // Replace the operands of the cloned instructions with extracted scalars.
5759 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5760 Value *Op = Params[op][Part];
5761 Cloned->setOperand(op, Op);
5764 // Place the cloned scalar in the new loop.
5765 Builder.Insert(Cloned);
5767 // If the original scalar returns a value we need to place it in a vector
5768 // so that future users will be able to use it.
5770 VecResults[Part] = Cloned;
5773 if (IfPredicateStore) {
5774 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5775 LoopVectorBody.push_back(NewIfBlock);
5776 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5777 Builder.SetInsertPoint(InsertPt);
5778 Instruction *OldBr = IfBlock->getTerminator();
5779 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5780 OldBr->eraseFromParent();
5781 IfBlock = NewIfBlock;
5786 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5787 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5788 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5790 return scalarizeInstruction(Instr, IfPredicateStore);
5793 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5797 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5801 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5803 // When unrolling and the VF is 1, we only need to add a simple scalar.
5804 Type *ITy = Val->getType();
5805 assert(!ITy->isVectorTy() && "Val must be a scalar");
5806 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5807 return Builder.CreateAdd(Val, C, "induction");