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 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/BlockFrequencyInfo.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopIterator.h"
60 #include "llvm/Analysis/LoopPass.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpander.h"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Analysis/TargetTransformInfo.h"
65 #include "llvm/Analysis/ValueTracking.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DataLayout.h"
68 #include "llvm/IR/DebugInfo.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Dominators.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/LLVMContext.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/IR/PatternMatch.h"
78 #include "llvm/IR/Type.h"
79 #include "llvm/IR/Value.h"
80 #include "llvm/IR/ValueHandle.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/Format.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"
92 #include "llvm/Transforms/Utils/VectorUtils.h"
97 using namespace llvm::PatternMatch;
99 #define LV_NAME "loop-vectorize"
100 #define DEBUG_TYPE LV_NAME
102 STATISTIC(LoopsVectorized, "Number of loops vectorized");
103 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
105 static cl::opt<unsigned>
106 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
107 cl::desc("Sets the SIMD width. Zero is autoselect."));
109 static cl::opt<unsigned>
110 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
111 cl::desc("Sets the vectorization unroll count. "
112 "Zero is autoselect."));
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 /// We don't unroll loops with a known constant trip count below this number.
142 static const unsigned TinyTripCountUnrollThreshold = 128;
144 /// When performing memory disambiguation checks at runtime do not make more
145 /// than this number of comparisons.
146 static const unsigned RuntimeMemoryCheckThreshold = 8;
148 /// Maximum simd width.
149 static const unsigned MaxVectorWidth = 64;
151 static cl::opt<unsigned> ForceTargetNumScalarRegs(
152 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
153 cl::desc("A flag that overrides the target's number of scalar registers."));
155 static cl::opt<unsigned> ForceTargetNumVectorRegs(
156 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of vector registers."));
159 /// Maximum vectorization unroll count.
160 static const unsigned MaxUnrollFactor = 16;
162 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
163 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
164 cl::desc("A flag that overrides the target's max unroll factor for scalar "
167 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
168 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
169 cl::desc("A flag that overrides the target's max unroll factor for "
170 "vectorized loops."));
172 static cl::opt<unsigned> ForceTargetInstructionCost(
173 "force-target-instruction-cost", cl::init(0), cl::Hidden,
174 cl::desc("A flag that overrides the target's expected cost for "
175 "an instruction to a single constant value. Mostly "
176 "useful for getting consistent testing."));
178 static cl::opt<unsigned> SmallLoopCost(
179 "small-loop-cost", cl::init(20), cl::Hidden,
180 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
182 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
183 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
184 cl::desc("Enable the use of the block frequency analysis to access PGO "
185 "heuristics minimizing code growth in cold regions and being more "
186 "aggressive in hot regions."));
188 // Runtime unroll loops for load/store throughput.
189 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
190 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
191 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
193 /// The number of stores in a loop that are allowed to need predication.
194 static cl::opt<unsigned> NumberOfStoresToPredicate(
195 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
196 cl::desc("Max number of stores to be predicated behind an if."));
198 static cl::opt<bool> EnableIndVarRegisterHeur(
199 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
200 cl::desc("Count the induction variable only once when unrolling"));
202 static cl::opt<bool> EnableCondStoresVectorization(
203 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
204 cl::desc("Enable if predication of stores during vectorization."));
208 // Forward declarations.
209 class LoopVectorizationLegality;
210 class LoopVectorizationCostModel;
212 /// InnerLoopVectorizer vectorizes loops which contain only one basic
213 /// block to a specified vectorization factor (VF).
214 /// This class performs the widening of scalars into vectors, or multiple
215 /// scalars. This class also implements the following features:
216 /// * It inserts an epilogue loop for handling loops that don't have iteration
217 /// counts that are known to be a multiple of the vectorization factor.
218 /// * It handles the code generation for reduction variables.
219 /// * Scalarization (implementation using scalars) of un-vectorizable
221 /// InnerLoopVectorizer does not perform any vectorization-legality
222 /// checks, and relies on the caller to check for the different legality
223 /// aspects. The InnerLoopVectorizer relies on the
224 /// LoopVectorizationLegality class to provide information about the induction
225 /// and reduction variables that were found to a given vectorization factor.
226 class InnerLoopVectorizer {
228 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
229 DominatorTree *DT, const DataLayout *DL,
230 const TargetLibraryInfo *TLI, unsigned VecWidth,
231 unsigned UnrollFactor)
232 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
233 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
234 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
237 // Perform the actual loop widening (vectorization).
238 void vectorize(LoopVectorizationLegality *L) {
240 // Create a new empty loop. Unlink the old loop and connect the new one.
242 // Widen each instruction in the old loop to a new one in the new loop.
243 // Use the Legality module to find the induction and reduction variables.
245 // Register the new loop and update the analysis passes.
249 virtual ~InnerLoopVectorizer() {}
252 /// A small list of PHINodes.
253 typedef SmallVector<PHINode*, 4> PhiVector;
254 /// When we unroll loops we have multiple vector values for each scalar.
255 /// This data structure holds the unrolled and vectorized values that
256 /// originated from one scalar instruction.
257 typedef SmallVector<Value*, 2> VectorParts;
259 // When we if-convert we need create edge masks. We have to cache values so
260 // that we don't end up with exponential recursion/IR.
261 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
262 VectorParts> EdgeMaskCache;
264 /// \brief Add code that checks at runtime if the accessed arrays overlap.
266 /// Returns a pair of instructions where the first element is the first
267 /// instruction generated in possibly a sequence of instructions and the
268 /// second value is the final comparator value or NULL if no check is needed.
269 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
271 /// \brief Add checks for strides that where assumed to be 1.
273 /// Returns the last check instruction and the first check instruction in the
274 /// pair as (first, last).
275 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
277 /// Create an empty loop, based on the loop ranges of the old loop.
278 void createEmptyLoop();
279 /// Copy and widen the instructions from the old loop.
280 virtual void vectorizeLoop();
282 /// \brief The Loop exit block may have single value PHI nodes where the
283 /// incoming value is 'Undef'. While vectorizing we only handled real values
284 /// that were defined inside the loop. Here we fix the 'undef case'.
288 /// A helper function that computes the predicate of the block BB, assuming
289 /// that the header block of the loop is set to True. It returns the *entry*
290 /// mask for the block BB.
291 VectorParts createBlockInMask(BasicBlock *BB);
292 /// A helper function that computes the predicate of the edge between SRC
294 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
296 /// A helper function to vectorize a single BB within the innermost loop.
297 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
299 /// Vectorize a single PHINode in a block. This method handles the induction
300 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
301 /// arbitrary length vectors.
302 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
303 unsigned UF, unsigned VF, PhiVector *PV);
305 /// Insert the new loop to the loop hierarchy and pass manager
306 /// and update the analysis passes.
307 void updateAnalysis();
309 /// This instruction is un-vectorizable. Implement it as a sequence
310 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
311 /// scalarized instruction behind an if block predicated on the control
312 /// dependence of the instruction.
313 virtual void scalarizeInstruction(Instruction *Instr,
314 bool IfPredicateStore=false);
316 /// Vectorize Load and Store instructions,
317 virtual void vectorizeMemoryInstruction(Instruction *Instr);
319 /// Create a broadcast instruction. This method generates a broadcast
320 /// instruction (shuffle) for loop invariant values and for the induction
321 /// value. If this is the induction variable then we extend it to N, N+1, ...
322 /// this is needed because each iteration in the loop corresponds to a SIMD
324 virtual Value *getBroadcastInstrs(Value *V);
326 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
327 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
328 /// The sequence starts at StartIndex.
329 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
331 /// When we go over instructions in the basic block we rely on previous
332 /// values within the current basic block or on loop invariant values.
333 /// When we widen (vectorize) values we place them in the map. If the values
334 /// are not within the map, they have to be loop invariant, so we simply
335 /// broadcast them into a vector.
336 VectorParts &getVectorValue(Value *V);
338 /// Generate a shuffle sequence that will reverse the vector Vec.
339 virtual Value *reverseVector(Value *Vec);
341 /// This is a helper class that holds the vectorizer state. It maps scalar
342 /// instructions to vector instructions. When the code is 'unrolled' then
343 /// then a single scalar value is mapped to multiple vector parts. The parts
344 /// are stored in the VectorPart type.
346 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
348 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
350 /// \return True if 'Key' is saved in the Value Map.
351 bool has(Value *Key) const { return MapStorage.count(Key); }
353 /// Initializes a new entry in the map. Sets all of the vector parts to the
354 /// save value in 'Val'.
355 /// \return A reference to a vector with splat values.
356 VectorParts &splat(Value *Key, Value *Val) {
357 VectorParts &Entry = MapStorage[Key];
358 Entry.assign(UF, Val);
362 ///\return A reference to the value that is stored at 'Key'.
363 VectorParts &get(Value *Key) {
364 VectorParts &Entry = MapStorage[Key];
367 assert(Entry.size() == UF);
372 /// The unroll factor. Each entry in the map stores this number of vector
376 /// Map storage. We use std::map and not DenseMap because insertions to a
377 /// dense map invalidates its iterators.
378 std::map<Value *, VectorParts> MapStorage;
381 /// The original loop.
383 /// Scev analysis to use.
390 const DataLayout *DL;
391 /// Target Library Info.
392 const TargetLibraryInfo *TLI;
394 /// The vectorization SIMD factor to use. Each vector will have this many
399 /// The vectorization unroll factor to use. Each scalar is vectorized to this
400 /// many different vector instructions.
403 /// The builder that we use
406 // --- Vectorization state ---
408 /// The vector-loop preheader.
409 BasicBlock *LoopVectorPreHeader;
410 /// The scalar-loop preheader.
411 BasicBlock *LoopScalarPreHeader;
412 /// Middle Block between the vector and the scalar.
413 BasicBlock *LoopMiddleBlock;
414 ///The ExitBlock of the scalar loop.
415 BasicBlock *LoopExitBlock;
416 ///The vector loop body.
417 SmallVector<BasicBlock *, 4> LoopVectorBody;
418 ///The scalar loop body.
419 BasicBlock *LoopScalarBody;
420 /// A list of all bypass blocks. The first block is the entry of the loop.
421 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
423 /// The new Induction variable which was added to the new block.
425 /// The induction variable of the old basic block.
426 PHINode *OldInduction;
427 /// Holds the extended (to the widest induction type) start index.
429 /// Maps scalars to widened vectors.
431 EdgeMaskCache MaskCache;
433 LoopVectorizationLegality *Legal;
436 class InnerLoopUnroller : public InnerLoopVectorizer {
438 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
439 DominatorTree *DT, const DataLayout *DL,
440 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
441 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
444 void scalarizeInstruction(Instruction *Instr,
445 bool IfPredicateStore = false) override;
446 void vectorizeMemoryInstruction(Instruction *Instr) override;
447 Value *getBroadcastInstrs(Value *V) override;
448 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
449 Value *reverseVector(Value *Vec) override;
452 /// \brief Look for a meaningful debug location on the instruction or it's
454 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
459 if (I->getDebugLoc() != Empty)
462 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
463 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
464 if (OpInst->getDebugLoc() != Empty)
471 /// \brief Set the debug location in the builder using the debug location in the
473 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
474 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
475 B.SetCurrentDebugLocation(Inst->getDebugLoc());
477 B.SetCurrentDebugLocation(DebugLoc());
481 /// \return string containing a file name and a line # for the given
483 static format_object3<const char *, const char *, unsigned>
484 getDebugLocString(const Instruction *I) {
486 return format<const char *, const char *, unsigned>("", "", "", 0U);
487 MDNode *N = I->getMetadata("dbg");
489 const StringRef ModuleName =
490 I->getParent()->getParent()->getParent()->getModuleIdentifier();
491 return format<const char *, const char *, unsigned>("%s", ModuleName.data(),
494 const DILocation Loc(N);
495 const unsigned LineNo = Loc.getLineNumber();
496 const char *DirName = Loc.getDirectory().data();
497 const char *FileName = Loc.getFilename().data();
498 return format("%s/%s:%u", DirName, FileName, LineNo);
502 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
503 /// to what vectorization factor.
504 /// This class does not look at the profitability of vectorization, only the
505 /// legality. This class has two main kinds of checks:
506 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
507 /// will change the order of memory accesses in a way that will change the
508 /// correctness of the program.
509 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
510 /// checks for a number of different conditions, such as the availability of a
511 /// single induction variable, that all types are supported and vectorize-able,
512 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
513 /// This class is also used by InnerLoopVectorizer for identifying
514 /// induction variable and the different reduction variables.
515 class LoopVectorizationLegality {
519 unsigned NumPredStores;
521 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
522 DominatorTree *DT, TargetLibraryInfo *TLI)
523 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
524 DT(DT), TLI(TLI), Induction(nullptr), WidestIndTy(nullptr),
525 HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {}
527 /// This enum represents the kinds of reductions that we support.
529 RK_NoReduction, ///< Not a reduction.
530 RK_IntegerAdd, ///< Sum of integers.
531 RK_IntegerMult, ///< Product of integers.
532 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
533 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
534 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
535 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
536 RK_FloatAdd, ///< Sum of floats.
537 RK_FloatMult, ///< Product of floats.
538 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
541 /// This enum represents the kinds of inductions that we support.
543 IK_NoInduction, ///< Not an induction variable.
544 IK_IntInduction, ///< Integer induction variable. Step = 1.
545 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
546 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
547 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
550 // This enum represents the kind of minmax reduction.
551 enum MinMaxReductionKind {
561 /// This struct holds information about reduction variables.
562 struct ReductionDescriptor {
563 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
564 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
566 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
567 MinMaxReductionKind MK)
568 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
570 // The starting value of the reduction.
571 // It does not have to be zero!
572 TrackingVH<Value> StartValue;
573 // The instruction who's value is used outside the loop.
574 Instruction *LoopExitInstr;
575 // The kind of the reduction.
577 // If this a min/max reduction the kind of reduction.
578 MinMaxReductionKind MinMaxKind;
581 /// This POD struct holds information about a potential reduction operation.
582 struct ReductionInstDesc {
583 ReductionInstDesc(bool IsRedux, Instruction *I) :
584 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
586 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
587 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
589 // Is this instruction a reduction candidate.
591 // The last instruction in a min/max pattern (select of the select(icmp())
592 // pattern), or the current reduction instruction otherwise.
593 Instruction *PatternLastInst;
594 // If this is a min/max pattern the comparison predicate.
595 MinMaxReductionKind MinMaxKind;
598 /// This struct holds information about the memory runtime legality
599 /// check that a group of pointers do not overlap.
600 struct RuntimePointerCheck {
601 RuntimePointerCheck() : Need(false) {}
603 /// Reset the state of the pointer runtime information.
610 DependencySetId.clear();
613 /// Insert a pointer and calculate the start and end SCEVs.
614 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
615 unsigned DepSetId, ValueToValueMap &Strides);
617 /// This flag indicates if we need to add the runtime check.
619 /// Holds the pointers that we need to check.
620 SmallVector<TrackingVH<Value>, 2> Pointers;
621 /// Holds the pointer value at the beginning of the loop.
622 SmallVector<const SCEV*, 2> Starts;
623 /// Holds the pointer value at the end of the loop.
624 SmallVector<const SCEV*, 2> Ends;
625 /// Holds the information if this pointer is used for writing to memory.
626 SmallVector<bool, 2> IsWritePtr;
627 /// Holds the id of the set of pointers that could be dependent because of a
628 /// shared underlying object.
629 SmallVector<unsigned, 2> DependencySetId;
632 /// A struct for saving information about induction variables.
633 struct InductionInfo {
634 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
635 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
637 TrackingVH<Value> StartValue;
642 /// ReductionList contains the reduction descriptors for all
643 /// of the reductions that were found in the loop.
644 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
646 /// InductionList saves induction variables and maps them to the
647 /// induction descriptor.
648 typedef MapVector<PHINode*, InductionInfo> InductionList;
650 /// Returns true if it is legal to vectorize this loop.
651 /// This does not mean that it is profitable to vectorize this
652 /// loop, only that it is legal to do so.
655 /// Returns the Induction variable.
656 PHINode *getInduction() { return Induction; }
658 /// Returns the reduction variables found in the loop.
659 ReductionList *getReductionVars() { return &Reductions; }
661 /// Returns the induction variables found in the loop.
662 InductionList *getInductionVars() { return &Inductions; }
664 /// Returns the widest induction type.
665 Type *getWidestInductionType() { return WidestIndTy; }
667 /// Returns True if V is an induction variable in this loop.
668 bool isInductionVariable(const Value *V);
670 /// Return true if the block BB needs to be predicated in order for the loop
671 /// to be vectorized.
672 bool blockNeedsPredication(BasicBlock *BB);
674 /// Check if this pointer is consecutive when vectorizing. This happens
675 /// when the last index of the GEP is the induction variable, or that the
676 /// pointer itself is an induction variable.
677 /// This check allows us to vectorize A[idx] into a wide load/store.
679 /// 0 - Stride is unknown or non-consecutive.
680 /// 1 - Address is consecutive.
681 /// -1 - Address is consecutive, and decreasing.
682 int isConsecutivePtr(Value *Ptr);
684 /// Returns true if the value V is uniform within the loop.
685 bool isUniform(Value *V);
687 /// Returns true if this instruction will remain scalar after vectorization.
688 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
690 /// Returns the information that we collected about runtime memory check.
691 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
693 /// This function returns the identity element (or neutral element) for
695 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
697 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
699 bool hasStride(Value *V) { return StrideSet.count(V); }
700 bool mustCheckStrides() { return !StrideSet.empty(); }
701 SmallPtrSet<Value *, 8>::iterator strides_begin() {
702 return StrideSet.begin();
704 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
707 /// Check if a single basic block loop is vectorizable.
708 /// At this point we know that this is a loop with a constant trip count
709 /// and we only need to check individual instructions.
710 bool canVectorizeInstrs();
712 /// When we vectorize loops we may change the order in which
713 /// we read and write from memory. This method checks if it is
714 /// legal to vectorize the code, considering only memory constrains.
715 /// Returns true if the loop is vectorizable
716 bool canVectorizeMemory();
718 /// Return true if we can vectorize this loop using the IF-conversion
720 bool canVectorizeWithIfConvert();
722 /// Collect the variables that need to stay uniform after vectorization.
723 void collectLoopUniforms();
725 /// Return true if all of the instructions in the block can be speculatively
726 /// executed. \p SafePtrs is a list of addresses that are known to be legal
727 /// and we know that we can read from them without segfault.
728 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
730 /// Returns True, if 'Phi' is the kind of reduction variable for type
731 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
732 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
733 /// Returns a struct describing if the instruction 'I' can be a reduction
734 /// variable of type 'Kind'. If the reduction is a min/max pattern of
735 /// select(icmp()) this function advances the instruction pointer 'I' from the
736 /// compare instruction to the select instruction and stores this pointer in
737 /// 'PatternLastInst' member of the returned struct.
738 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
739 ReductionInstDesc &Desc);
740 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
741 /// pattern corresponding to a min(X, Y) or max(X, Y).
742 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
743 ReductionInstDesc &Prev);
744 /// Returns the induction kind of Phi. This function may return NoInduction
745 /// if the PHI is not an induction variable.
746 InductionKind isInductionVariable(PHINode *Phi);
748 /// \brief Collect memory access with loop invariant strides.
750 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
752 void collectStridedAcccess(Value *LoadOrStoreInst);
754 /// The loop that we evaluate.
758 /// DataLayout analysis.
759 const DataLayout *DL;
762 /// Target Library Info.
763 TargetLibraryInfo *TLI;
765 // --- vectorization state --- //
767 /// Holds the integer induction variable. This is the counter of the
770 /// Holds the reduction variables.
771 ReductionList Reductions;
772 /// Holds all of the induction variables that we found in the loop.
773 /// Notice that inductions don't need to start at zero and that induction
774 /// variables can be pointers.
775 InductionList Inductions;
776 /// Holds the widest induction type encountered.
779 /// Allowed outside users. This holds the reduction
780 /// vars which can be accessed from outside the loop.
781 SmallPtrSet<Value*, 4> AllowedExit;
782 /// This set holds the variables which are known to be uniform after
784 SmallPtrSet<Instruction*, 4> Uniforms;
785 /// We need to check that all of the pointers in this list are disjoint
787 RuntimePointerCheck PtrRtCheck;
788 /// Can we assume the absence of NaNs.
789 bool HasFunNoNaNAttr;
791 unsigned MaxSafeDepDistBytes;
793 ValueToValueMap Strides;
794 SmallPtrSet<Value *, 8> StrideSet;
797 /// LoopVectorizationCostModel - estimates the expected speedups due to
799 /// In many cases vectorization is not profitable. This can happen because of
800 /// a number of reasons. In this class we mainly attempt to predict the
801 /// expected speedup/slowdowns due to the supported instruction set. We use the
802 /// TargetTransformInfo to query the different backends for the cost of
803 /// different operations.
804 class LoopVectorizationCostModel {
806 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
807 LoopVectorizationLegality *Legal,
808 const TargetTransformInfo &TTI,
809 const DataLayout *DL, const TargetLibraryInfo *TLI)
810 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
812 /// Information about vectorization costs
813 struct VectorizationFactor {
814 unsigned Width; // Vector width with best cost
815 unsigned Cost; // Cost of the loop with that width
817 /// \return The most profitable vectorization factor and the cost of that VF.
818 /// This method checks every power of two up to VF. If UserVF is not ZERO
819 /// then this vectorization factor will be selected if vectorization is
821 VectorizationFactor selectVectorizationFactor(bool OptForSize,
823 bool ForceVectorization);
825 /// \return The size (in bits) of the widest type in the code that
826 /// needs to be vectorized. We ignore values that remain scalar such as
827 /// 64 bit loop indices.
828 unsigned getWidestType();
830 /// \return The most profitable unroll factor.
831 /// If UserUF is non-zero then this method finds the best unroll-factor
832 /// based on register pressure and other parameters.
833 /// VF and LoopCost are the selected vectorization factor and the cost of the
835 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
838 /// \brief A struct that represents some properties of the register usage
840 struct RegisterUsage {
841 /// Holds the number of loop invariant values that are used in the loop.
842 unsigned LoopInvariantRegs;
843 /// Holds the maximum number of concurrent live intervals in the loop.
844 unsigned MaxLocalUsers;
845 /// Holds the number of instructions in the loop.
846 unsigned NumInstructions;
849 /// \return information about the register usage of the loop.
850 RegisterUsage calculateRegisterUsage();
853 /// Returns the expected execution cost. The unit of the cost does
854 /// not matter because we use the 'cost' units to compare different
855 /// vector widths. The cost that is returned is *not* normalized by
856 /// the factor width.
857 unsigned expectedCost(unsigned VF);
859 /// Returns the execution time cost of an instruction for a given vector
860 /// width. Vector width of one means scalar.
861 unsigned getInstructionCost(Instruction *I, unsigned VF);
863 /// A helper function for converting Scalar types to vector types.
864 /// If the incoming type is void, we return void. If the VF is 1, we return
866 static Type* ToVectorTy(Type *Scalar, unsigned VF);
868 /// Returns whether the instruction is a load or store and will be a emitted
869 /// as a vector operation.
870 bool isConsecutiveLoadOrStore(Instruction *I);
872 /// The loop that we evaluate.
876 /// Loop Info analysis.
878 /// Vectorization legality.
879 LoopVectorizationLegality *Legal;
880 /// Vector target information.
881 const TargetTransformInfo &TTI;
882 /// Target data layout information.
883 const DataLayout *DL;
884 /// Target Library Info.
885 const TargetLibraryInfo *TLI;
888 /// Utility class for getting and setting loop vectorizer hints in the form
889 /// of loop metadata.
890 struct LoopVectorizeHints {
891 /// Vectorization width.
893 /// Vectorization unroll factor.
895 /// Vectorization forced
897 FK_Undefined = -1, ///< Not selected.
898 FK_Disabled = 0, ///< Forcing disabled.
899 FK_Enabled = 1, ///< Forcing enabled.
902 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
903 : Width(VectorizationFactor)
904 , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
905 , Force(FK_Undefined)
906 , LoopID(L->getLoopID()) {
908 // The command line options override any loop metadata except for when
909 // width == 1 which is used to indicate the loop is already vectorized.
910 if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
911 Width = VectorizationFactor;
912 if (VectorizationUnroll.getNumOccurrences() > 0)
913 Unroll = VectorizationUnroll;
915 DEBUG(if (DisableUnrolling && Unroll == 1)
916 dbgs() << "LV: Unrolling disabled by the pass manager\n");
919 /// Return the loop vectorizer metadata prefix.
920 static StringRef Prefix() { return "llvm.vectorizer."; }
922 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
923 SmallVector<Value*, 2> Vals;
924 Vals.push_back(MDString::get(Context, Name));
925 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
926 return MDNode::get(Context, Vals);
929 /// Mark the loop L as already vectorized by setting the width to 1.
930 void setAlreadyVectorized(Loop *L) {
931 LLVMContext &Context = L->getHeader()->getContext();
935 // Create a new loop id with one more operand for the already_vectorized
936 // hint. If the loop already has a loop id then copy the existing operands.
937 SmallVector<Value*, 4> Vals(1);
939 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
940 Vals.push_back(LoopID->getOperand(i));
942 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
943 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
945 MDNode *NewLoopID = MDNode::get(Context, Vals);
946 // Set operand 0 to refer to the loop id itself.
947 NewLoopID->replaceOperandWith(0, NewLoopID);
949 L->setLoopID(NewLoopID);
951 LoopID->replaceAllUsesWith(NewLoopID);
959 /// Find hints specified in the loop metadata.
960 void getHints(const Loop *L) {
964 // First operand should refer to the loop id itself.
965 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
966 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
968 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
969 const MDString *S = nullptr;
970 SmallVector<Value*, 4> Args;
972 // The expected hint is either a MDString or a MDNode with the first
973 // operand a MDString.
974 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
975 if (!MD || MD->getNumOperands() == 0)
977 S = dyn_cast<MDString>(MD->getOperand(0));
978 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
979 Args.push_back(MD->getOperand(i));
981 S = dyn_cast<MDString>(LoopID->getOperand(i));
982 assert(Args.size() == 0 && "too many arguments for MDString");
988 // Check if the hint starts with the vectorizer prefix.
989 StringRef Hint = S->getString();
990 if (!Hint.startswith(Prefix()))
992 // Remove the prefix.
993 Hint = Hint.substr(Prefix().size(), StringRef::npos);
995 if (Args.size() == 1)
996 getHint(Hint, Args[0]);
1000 // Check string hint with one operand.
1001 void getHint(StringRef Hint, Value *Arg) {
1002 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1004 unsigned Val = C->getZExtValue();
1006 if (Hint == "width") {
1007 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1010 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1011 } else if (Hint == "unroll") {
1012 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1015 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1016 } else if (Hint == "enable") {
1017 if (C->getBitWidth() == 1)
1018 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1019 : LoopVectorizeHints::FK_Disabled;
1021 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1023 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1028 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1030 return V.push_back(&L);
1032 for (Loop *InnerL : L)
1033 addInnerLoop(*InnerL, V);
1036 /// The LoopVectorize Pass.
1037 struct LoopVectorize : public FunctionPass {
1038 /// Pass identification, replacement for typeid
1041 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1043 DisableUnrolling(NoUnrolling),
1044 AlwaysVectorize(AlwaysVectorize) {
1045 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1048 ScalarEvolution *SE;
1049 const DataLayout *DL;
1051 TargetTransformInfo *TTI;
1053 BlockFrequencyInfo *BFI;
1054 TargetLibraryInfo *TLI;
1055 bool DisableUnrolling;
1056 bool AlwaysVectorize;
1058 BlockFrequency ColdEntryFreq;
1060 bool runOnFunction(Function &F) override {
1061 SE = &getAnalysis<ScalarEvolution>();
1062 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1063 DL = DLP ? &DLP->getDataLayout() : nullptr;
1064 LI = &getAnalysis<LoopInfo>();
1065 TTI = &getAnalysis<TargetTransformInfo>();
1066 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1067 BFI = &getAnalysis<BlockFrequencyInfo>();
1068 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1070 // Compute some weights outside of the loop over the loops. Compute this
1071 // using a BranchProbability to re-use its scaling math.
1072 const BranchProbability ColdProb(1, 5); // 20%
1073 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1075 // If the target claims to have no vector registers don't attempt
1077 if (!TTI->getNumberOfRegisters(true))
1081 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1082 << ": Missing data layout\n");
1086 // Build up a worklist of inner-loops to vectorize. This is necessary as
1087 // the act of vectorizing or partially unrolling a loop creates new loops
1088 // and can invalidate iterators across the loops.
1089 SmallVector<Loop *, 8> Worklist;
1092 addInnerLoop(*L, Worklist);
1094 LoopsAnalyzed += Worklist.size();
1096 // Now walk the identified inner loops.
1097 bool Changed = false;
1098 while (!Worklist.empty())
1099 Changed |= processLoop(Worklist.pop_back_val());
1101 // Process each loop nest in the function.
1105 bool processLoop(Loop *L) {
1106 assert(L->empty() && "Only process inner loops.");
1107 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1108 << L->getHeader()->getParent()->getName() << "\" from "
1109 << getDebugLocString(L->getHeader()->getFirstNonPHIOrDbg())
1112 LoopVectorizeHints Hints(L, DisableUnrolling);
1114 DEBUG(dbgs() << "LV: Loop hints:"
1116 << (Hints.Force == LoopVectorizeHints::FK_Disabled
1118 : (Hints.Force == LoopVectorizeHints::FK_Enabled
1120 : "?")) << " width=" << Hints.Width
1121 << " unroll=" << Hints.Unroll << "\n");
1123 if (Hints.Force == LoopVectorizeHints::FK_Disabled) {
1124 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1128 if (!AlwaysVectorize && Hints.Force != LoopVectorizeHints::FK_Enabled) {
1129 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1133 if (Hints.Width == 1 && Hints.Unroll == 1) {
1134 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1138 // Check the loop for a trip count threshold:
1139 // do not vectorize loops with a tiny trip count.
1140 BasicBlock *Latch = L->getLoopLatch();
1141 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1142 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1143 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1144 << "This loop is not worth vectorizing.");
1145 if (Hints.Force == LoopVectorizeHints::FK_Enabled)
1146 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1148 DEBUG(dbgs() << "\n");
1153 // Check if it is legal to vectorize the loop.
1154 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1155 if (!LVL.canVectorize()) {
1156 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1160 // Use the cost model.
1161 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1163 // Check the function attributes to find out if this function should be
1164 // optimized for size.
1165 Function *F = L->getHeader()->getParent();
1166 bool OptForSize = Hints.Force != LoopVectorizeHints::FK_Enabled &&
1167 F->hasFnAttribute(Attribute::OptimizeForSize);
1169 // Compute the weighted frequency of this loop being executed and see if it
1170 // is less than 20% of the function entry baseline frequency. Note that we
1171 // always have a canonical loop here because we think we *can* vectoriez.
1172 // FIXME: This is hidden behind a flag due to pervasive problems with
1173 // exactly what block frequency models.
1174 if (LoopVectorizeWithBlockFrequency) {
1175 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1176 if (Hints.Force != LoopVectorizeHints::FK_Enabled &&
1177 LoopEntryFreq < ColdEntryFreq)
1181 // Check the function attributes to see if implicit floats are allowed.a
1182 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1183 // an integer loop and the vector instructions selected are purely integer
1184 // vector instructions?
1185 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1186 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1187 "attribute is used.\n");
1191 // Select the optimal vectorization factor.
1192 const LoopVectorizationCostModel::VectorizationFactor VF =
1193 CM.selectVectorizationFactor(OptForSize, Hints.Width,
1195 LoopVectorizeHints::FK_Enabled);
1197 // Select the unroll factor.
1198 const unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
1201 DEBUG(dbgs() << "LV: Found a vectorizable loop ("
1202 << VF.Width << ") in "
1203 << getDebugLocString(L->getHeader()->getFirstNonPHIOrDbg())
1205 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1207 if (VF.Width == 1) {
1208 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1211 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1213 // Report the unrolling decision.
1214 F->getContext().emitOptimizationRemark(
1215 DEBUG_TYPE, *F, L->getStartLoc(),
1216 Twine("unrolled with interleaving factor " + Twine(UF) +
1217 " (vectorization not beneficial)"));
1219 // We decided not to vectorize, but we may want to unroll.
1220 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1221 Unroller.vectorize(&LVL);
1223 // If we decided that it is *legal* to vectorize the loop then do it.
1224 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1228 // Report the vectorization decision.
1229 F->getContext().emitOptimizationRemark(
1230 DEBUG_TYPE, *F, L->getStartLoc(),
1231 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1232 ", unrolling interleave factor: " + Twine(UF) + ")");
1235 // Mark the loop as already vectorized to avoid vectorizing again.
1236 Hints.setAlreadyVectorized(L);
1238 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1242 void getAnalysisUsage(AnalysisUsage &AU) const override {
1243 AU.addRequiredID(LoopSimplifyID);
1244 AU.addRequiredID(LCSSAID);
1245 AU.addRequired<BlockFrequencyInfo>();
1246 AU.addRequired<DominatorTreeWrapperPass>();
1247 AU.addRequired<LoopInfo>();
1248 AU.addRequired<ScalarEvolution>();
1249 AU.addRequired<TargetTransformInfo>();
1250 AU.addPreserved<LoopInfo>();
1251 AU.addPreserved<DominatorTreeWrapperPass>();
1256 } // end anonymous namespace
1258 //===----------------------------------------------------------------------===//
1259 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1260 // LoopVectorizationCostModel.
1261 //===----------------------------------------------------------------------===//
1263 static Value *stripIntegerCast(Value *V) {
1264 if (CastInst *CI = dyn_cast<CastInst>(V))
1265 if (CI->getOperand(0)->getType()->isIntegerTy())
1266 return CI->getOperand(0);
1270 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1272 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1274 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1275 ValueToValueMap &PtrToStride,
1276 Value *Ptr, Value *OrigPtr = nullptr) {
1278 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1280 // If there is an entry in the map return the SCEV of the pointer with the
1281 // symbolic stride replaced by one.
1282 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1283 if (SI != PtrToStride.end()) {
1284 Value *StrideVal = SI->second;
1287 StrideVal = stripIntegerCast(StrideVal);
1289 // Replace symbolic stride by one.
1290 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1291 ValueToValueMap RewriteMap;
1292 RewriteMap[StrideVal] = One;
1295 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1296 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1301 // Otherwise, just return the SCEV of the original pointer.
1302 return SE->getSCEV(Ptr);
1305 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1306 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1307 ValueToValueMap &Strides) {
1308 // Get the stride replaced scev.
1309 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1310 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1311 assert(AR && "Invalid addrec expression");
1312 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1313 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1314 Pointers.push_back(Ptr);
1315 Starts.push_back(AR->getStart());
1316 Ends.push_back(ScEnd);
1317 IsWritePtr.push_back(WritePtr);
1318 DependencySetId.push_back(DepSetId);
1321 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1322 // We need to place the broadcast of invariant variables outside the loop.
1323 Instruction *Instr = dyn_cast<Instruction>(V);
1325 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1326 Instr->getParent()) != LoopVectorBody.end());
1327 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1329 // Place the code for broadcasting invariant variables in the new preheader.
1330 IRBuilder<>::InsertPointGuard Guard(Builder);
1332 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1334 // Broadcast the scalar into all locations in the vector.
1335 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1340 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1342 assert(Val->getType()->isVectorTy() && "Must be a vector");
1343 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1344 "Elem must be an integer");
1345 // Create the types.
1346 Type *ITy = Val->getType()->getScalarType();
1347 VectorType *Ty = cast<VectorType>(Val->getType());
1348 int VLen = Ty->getNumElements();
1349 SmallVector<Constant*, 8> Indices;
1351 // Create a vector of consecutive numbers from zero to VF.
1352 for (int i = 0; i < VLen; ++i) {
1353 int64_t Idx = Negate ? (-i) : i;
1354 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1357 // Add the consecutive indices to the vector value.
1358 Constant *Cv = ConstantVector::get(Indices);
1359 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1360 return Builder.CreateAdd(Val, Cv, "induction");
1363 /// \brief Find the operand of the GEP that should be checked for consecutive
1364 /// stores. This ignores trailing indices that have no effect on the final
1366 static unsigned getGEPInductionOperand(const DataLayout *DL,
1367 const GetElementPtrInst *Gep) {
1368 unsigned LastOperand = Gep->getNumOperands() - 1;
1369 unsigned GEPAllocSize = DL->getTypeAllocSize(
1370 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1372 // Walk backwards and try to peel off zeros.
1373 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1374 // Find the type we're currently indexing into.
1375 gep_type_iterator GEPTI = gep_type_begin(Gep);
1376 std::advance(GEPTI, LastOperand - 1);
1378 // If it's a type with the same allocation size as the result of the GEP we
1379 // can peel off the zero index.
1380 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1388 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1389 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1390 // Make sure that the pointer does not point to structs.
1391 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1394 // If this value is a pointer induction variable we know it is consecutive.
1395 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1396 if (Phi && Inductions.count(Phi)) {
1397 InductionInfo II = Inductions[Phi];
1398 if (IK_PtrInduction == II.IK)
1400 else if (IK_ReversePtrInduction == II.IK)
1404 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1408 unsigned NumOperands = Gep->getNumOperands();
1409 Value *GpPtr = Gep->getPointerOperand();
1410 // If this GEP value is a consecutive pointer induction variable and all of
1411 // the indices are constant then we know it is consecutive. We can
1412 Phi = dyn_cast<PHINode>(GpPtr);
1413 if (Phi && Inductions.count(Phi)) {
1415 // Make sure that the pointer does not point to structs.
1416 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1417 if (GepPtrType->getElementType()->isAggregateType())
1420 // Make sure that all of the index operands are loop invariant.
1421 for (unsigned i = 1; i < NumOperands; ++i)
1422 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1425 InductionInfo II = Inductions[Phi];
1426 if (IK_PtrInduction == II.IK)
1428 else if (IK_ReversePtrInduction == II.IK)
1432 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1434 // Check that all of the gep indices are uniform except for our induction
1436 for (unsigned i = 0; i != NumOperands; ++i)
1437 if (i != InductionOperand &&
1438 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1441 // We can emit wide load/stores only if the last non-zero index is the
1442 // induction variable.
1443 const SCEV *Last = nullptr;
1444 if (!Strides.count(Gep))
1445 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1447 // Because of the multiplication by a stride we can have a s/zext cast.
1448 // We are going to replace this stride by 1 so the cast is safe to ignore.
1450 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1451 // %0 = trunc i64 %indvars.iv to i32
1452 // %mul = mul i32 %0, %Stride1
1453 // %idxprom = zext i32 %mul to i64 << Safe cast.
1454 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1456 Last = replaceSymbolicStrideSCEV(SE, Strides,
1457 Gep->getOperand(InductionOperand), Gep);
1458 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1460 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1464 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1465 const SCEV *Step = AR->getStepRecurrence(*SE);
1467 // The memory is consecutive because the last index is consecutive
1468 // and all other indices are loop invariant.
1471 if (Step->isAllOnesValue())
1478 bool LoopVectorizationLegality::isUniform(Value *V) {
1479 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1482 InnerLoopVectorizer::VectorParts&
1483 InnerLoopVectorizer::getVectorValue(Value *V) {
1484 assert(V != Induction && "The new induction variable should not be used.");
1485 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1487 // If we have a stride that is replaced by one, do it here.
1488 if (Legal->hasStride(V))
1489 V = ConstantInt::get(V->getType(), 1);
1491 // If we have this scalar in the map, return it.
1492 if (WidenMap.has(V))
1493 return WidenMap.get(V);
1495 // If this scalar is unknown, assume that it is a constant or that it is
1496 // loop invariant. Broadcast V and save the value for future uses.
1497 Value *B = getBroadcastInstrs(V);
1498 return WidenMap.splat(V, B);
1501 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1502 assert(Vec->getType()->isVectorTy() && "Invalid type");
1503 SmallVector<Constant*, 8> ShuffleMask;
1504 for (unsigned i = 0; i < VF; ++i)
1505 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1507 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1508 ConstantVector::get(ShuffleMask),
1512 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1513 // Attempt to issue a wide load.
1514 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1515 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1517 assert((LI || SI) && "Invalid Load/Store instruction");
1519 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1520 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1521 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1522 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1523 // An alignment of 0 means target abi alignment. We need to use the scalar's
1524 // target abi alignment in such a case.
1526 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1527 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1528 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1529 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1531 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1532 return scalarizeInstruction(Instr, true);
1534 if (ScalarAllocatedSize != VectorElementSize)
1535 return scalarizeInstruction(Instr);
1537 // If the pointer is loop invariant or if it is non-consecutive,
1538 // scalarize the load.
1539 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1540 bool Reverse = ConsecutiveStride < 0;
1541 bool UniformLoad = LI && Legal->isUniform(Ptr);
1542 if (!ConsecutiveStride || UniformLoad)
1543 return scalarizeInstruction(Instr);
1545 Constant *Zero = Builder.getInt32(0);
1546 VectorParts &Entry = WidenMap.get(Instr);
1548 // Handle consecutive loads/stores.
1549 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1550 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1551 setDebugLocFromInst(Builder, Gep);
1552 Value *PtrOperand = Gep->getPointerOperand();
1553 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1554 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1556 // Create the new GEP with the new induction variable.
1557 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1558 Gep2->setOperand(0, FirstBasePtr);
1559 Gep2->setName("gep.indvar.base");
1560 Ptr = Builder.Insert(Gep2);
1562 setDebugLocFromInst(Builder, Gep);
1563 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1564 OrigLoop) && "Base ptr must be invariant");
1566 // The last index does not have to be the induction. It can be
1567 // consecutive and be a function of the index. For example A[I+1];
1568 unsigned NumOperands = Gep->getNumOperands();
1569 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1570 // Create the new GEP with the new induction variable.
1571 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1573 for (unsigned i = 0; i < NumOperands; ++i) {
1574 Value *GepOperand = Gep->getOperand(i);
1575 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1577 // Update last index or loop invariant instruction anchored in loop.
1578 if (i == InductionOperand ||
1579 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1580 assert((i == InductionOperand ||
1581 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1582 "Must be last index or loop invariant");
1584 VectorParts &GEPParts = getVectorValue(GepOperand);
1585 Value *Index = GEPParts[0];
1586 Index = Builder.CreateExtractElement(Index, Zero);
1587 Gep2->setOperand(i, Index);
1588 Gep2->setName("gep.indvar.idx");
1591 Ptr = Builder.Insert(Gep2);
1593 // Use the induction element ptr.
1594 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1595 setDebugLocFromInst(Builder, Ptr);
1596 VectorParts &PtrVal = getVectorValue(Ptr);
1597 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1602 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1603 "We do not allow storing to uniform addresses");
1604 setDebugLocFromInst(Builder, SI);
1605 // We don't want to update the value in the map as it might be used in
1606 // another expression. So don't use a reference type for "StoredVal".
1607 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1609 for (unsigned Part = 0; Part < UF; ++Part) {
1610 // Calculate the pointer for the specific unroll-part.
1611 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1614 // If we store to reverse consecutive memory locations then we need
1615 // to reverse the order of elements in the stored value.
1616 StoredVal[Part] = reverseVector(StoredVal[Part]);
1617 // If the address is consecutive but reversed, then the
1618 // wide store needs to start at the last vector element.
1619 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1620 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1623 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1624 DataTy->getPointerTo(AddressSpace));
1625 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1631 assert(LI && "Must have a load instruction");
1632 setDebugLocFromInst(Builder, LI);
1633 for (unsigned Part = 0; Part < UF; ++Part) {
1634 // Calculate the pointer for the specific unroll-part.
1635 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1638 // If the address is consecutive but reversed, then the
1639 // wide store needs to start at the last vector element.
1640 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1641 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1644 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1645 DataTy->getPointerTo(AddressSpace));
1646 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1647 cast<LoadInst>(LI)->setAlignment(Alignment);
1648 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1652 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1653 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1654 // Holds vector parameters or scalars, in case of uniform vals.
1655 SmallVector<VectorParts, 4> Params;
1657 setDebugLocFromInst(Builder, Instr);
1659 // Find all of the vectorized parameters.
1660 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1661 Value *SrcOp = Instr->getOperand(op);
1663 // If we are accessing the old induction variable, use the new one.
1664 if (SrcOp == OldInduction) {
1665 Params.push_back(getVectorValue(SrcOp));
1669 // Try using previously calculated values.
1670 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1672 // If the src is an instruction that appeared earlier in the basic block
1673 // then it should already be vectorized.
1674 if (SrcInst && OrigLoop->contains(SrcInst)) {
1675 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1676 // The parameter is a vector value from earlier.
1677 Params.push_back(WidenMap.get(SrcInst));
1679 // The parameter is a scalar from outside the loop. Maybe even a constant.
1680 VectorParts Scalars;
1681 Scalars.append(UF, SrcOp);
1682 Params.push_back(Scalars);
1686 assert(Params.size() == Instr->getNumOperands() &&
1687 "Invalid number of operands");
1689 // Does this instruction return a value ?
1690 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1692 Value *UndefVec = IsVoidRetTy ? nullptr :
1693 UndefValue::get(VectorType::get(Instr->getType(), VF));
1694 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1695 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1697 Instruction *InsertPt = Builder.GetInsertPoint();
1698 BasicBlock *IfBlock = Builder.GetInsertBlock();
1699 BasicBlock *CondBlock = nullptr;
1702 Loop *VectorLp = nullptr;
1703 if (IfPredicateStore) {
1704 assert(Instr->getParent()->getSinglePredecessor() &&
1705 "Only support single predecessor blocks");
1706 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1707 Instr->getParent());
1708 VectorLp = LI->getLoopFor(IfBlock);
1709 assert(VectorLp && "Must have a loop for this block");
1712 // For each vector unroll 'part':
1713 for (unsigned Part = 0; Part < UF; ++Part) {
1714 // For each scalar that we create:
1715 for (unsigned Width = 0; Width < VF; ++Width) {
1718 Value *Cmp = nullptr;
1719 if (IfPredicateStore) {
1720 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1721 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1722 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1723 LoopVectorBody.push_back(CondBlock);
1724 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1725 // Update Builder with newly created basic block.
1726 Builder.SetInsertPoint(InsertPt);
1729 Instruction *Cloned = Instr->clone();
1731 Cloned->setName(Instr->getName() + ".cloned");
1732 // Replace the operands of the cloned instructions with extracted scalars.
1733 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1734 Value *Op = Params[op][Part];
1735 // Param is a vector. Need to extract the right lane.
1736 if (Op->getType()->isVectorTy())
1737 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1738 Cloned->setOperand(op, Op);
1741 // Place the cloned scalar in the new loop.
1742 Builder.Insert(Cloned);
1744 // If the original scalar returns a value we need to place it in a vector
1745 // so that future users will be able to use it.
1747 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1748 Builder.getInt32(Width));
1750 if (IfPredicateStore) {
1751 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1752 LoopVectorBody.push_back(NewIfBlock);
1753 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1754 Builder.SetInsertPoint(InsertPt);
1755 Instruction *OldBr = IfBlock->getTerminator();
1756 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1757 OldBr->eraseFromParent();
1758 IfBlock = NewIfBlock;
1764 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1768 if (Instruction *I = dyn_cast<Instruction>(V))
1769 return I->getParent() == Loc->getParent() ? I : nullptr;
1773 std::pair<Instruction *, Instruction *>
1774 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1775 Instruction *tnullptr = nullptr;
1776 if (!Legal->mustCheckStrides())
1777 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1779 IRBuilder<> ChkBuilder(Loc);
1782 Value *Check = nullptr;
1783 Instruction *FirstInst = nullptr;
1784 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1785 SE = Legal->strides_end();
1787 Value *Ptr = stripIntegerCast(*SI);
1788 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1790 // Store the first instruction we create.
1791 FirstInst = getFirstInst(FirstInst, C, Loc);
1793 Check = ChkBuilder.CreateOr(Check, C);
1798 // We have to do this trickery because the IRBuilder might fold the check to a
1799 // constant expression in which case there is no Instruction anchored in a
1801 LLVMContext &Ctx = Loc->getContext();
1802 Instruction *TheCheck =
1803 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1804 ChkBuilder.Insert(TheCheck, "stride.not.one");
1805 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1807 return std::make_pair(FirstInst, TheCheck);
1810 std::pair<Instruction *, Instruction *>
1811 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1812 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1813 Legal->getRuntimePointerCheck();
1815 Instruction *tnullptr = nullptr;
1816 if (!PtrRtCheck->Need)
1817 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1819 unsigned NumPointers = PtrRtCheck->Pointers.size();
1820 SmallVector<TrackingVH<Value> , 2> Starts;
1821 SmallVector<TrackingVH<Value> , 2> Ends;
1823 LLVMContext &Ctx = Loc->getContext();
1824 SCEVExpander Exp(*SE, "induction");
1825 Instruction *FirstInst = nullptr;
1827 for (unsigned i = 0; i < NumPointers; ++i) {
1828 Value *Ptr = PtrRtCheck->Pointers[i];
1829 const SCEV *Sc = SE->getSCEV(Ptr);
1831 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1832 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1834 Starts.push_back(Ptr);
1835 Ends.push_back(Ptr);
1837 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1838 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1840 // Use this type for pointer arithmetic.
1841 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1843 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1844 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1845 Starts.push_back(Start);
1846 Ends.push_back(End);
1850 IRBuilder<> ChkBuilder(Loc);
1851 // Our instructions might fold to a constant.
1852 Value *MemoryRuntimeCheck = nullptr;
1853 for (unsigned i = 0; i < NumPointers; ++i) {
1854 for (unsigned j = i+1; j < NumPointers; ++j) {
1855 // No need to check if two readonly pointers intersect.
1856 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1859 // Only need to check pointers between two different dependency sets.
1860 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1863 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1864 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1866 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1867 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1868 "Trying to bounds check pointers with different address spaces");
1870 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1871 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1873 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1874 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1875 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1876 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1878 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1879 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1880 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1881 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1882 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1883 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1884 if (MemoryRuntimeCheck) {
1885 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1887 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1889 MemoryRuntimeCheck = IsConflict;
1893 // We have to do this trickery because the IRBuilder might fold the check to a
1894 // constant expression in which case there is no Instruction anchored in a
1896 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1897 ConstantInt::getTrue(Ctx));
1898 ChkBuilder.Insert(Check, "memcheck.conflict");
1899 FirstInst = getFirstInst(FirstInst, Check, Loc);
1900 return std::make_pair(FirstInst, Check);
1903 void InnerLoopVectorizer::createEmptyLoop() {
1905 In this function we generate a new loop. The new loop will contain
1906 the vectorized instructions while the old loop will continue to run the
1909 [ ] <-- vector loop bypass (may consist of multiple blocks).
1912 | [ ] <-- vector pre header.
1916 | [ ]_| <-- vector loop.
1919 >[ ] <--- middle-block.
1922 | [ ] <--- new preheader.
1926 | [ ]_| <-- old scalar loop to handle remainder.
1929 >[ ] <-- exit block.
1933 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1934 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1935 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1936 assert(ExitBlock && "Must have an exit block");
1938 // Some loops have a single integer induction variable, while other loops
1939 // don't. One example is c++ iterators that often have multiple pointer
1940 // induction variables. In the code below we also support a case where we
1941 // don't have a single induction variable.
1942 OldInduction = Legal->getInduction();
1943 Type *IdxTy = Legal->getWidestInductionType();
1945 // Find the loop boundaries.
1946 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1947 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1949 // The exit count might have the type of i64 while the phi is i32. This can
1950 // happen if we have an induction variable that is sign extended before the
1951 // compare. The only way that we get a backedge taken count is that the
1952 // induction variable was signed and as such will not overflow. In such a case
1953 // truncation is legal.
1954 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1955 IdxTy->getPrimitiveSizeInBits())
1956 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1958 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1959 // Get the total trip count from the count by adding 1.
1960 ExitCount = SE->getAddExpr(ExitCount,
1961 SE->getConstant(ExitCount->getType(), 1));
1963 // Expand the trip count and place the new instructions in the preheader.
1964 // Notice that the pre-header does not change, only the loop body.
1965 SCEVExpander Exp(*SE, "induction");
1967 // Count holds the overall loop count (N).
1968 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1969 BypassBlock->getTerminator());
1971 // The loop index does not have to start at Zero. Find the original start
1972 // value from the induction PHI node. If we don't have an induction variable
1973 // then we know that it starts at zero.
1974 Builder.SetInsertPoint(BypassBlock->getTerminator());
1975 Value *StartIdx = ExtendedIdx = OldInduction ?
1976 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1978 ConstantInt::get(IdxTy, 0);
1980 assert(BypassBlock && "Invalid loop structure");
1981 LoopBypassBlocks.push_back(BypassBlock);
1983 // Split the single block loop into the two loop structure described above.
1984 BasicBlock *VectorPH =
1985 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1986 BasicBlock *VecBody =
1987 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1988 BasicBlock *MiddleBlock =
1989 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1990 BasicBlock *ScalarPH =
1991 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1993 // Create and register the new vector loop.
1994 Loop* Lp = new Loop();
1995 Loop *ParentLoop = OrigLoop->getParentLoop();
1997 // Insert the new loop into the loop nest and register the new basic blocks
1998 // before calling any utilities such as SCEV that require valid LoopInfo.
2000 ParentLoop->addChildLoop(Lp);
2001 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2002 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2003 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2005 LI->addTopLevelLoop(Lp);
2007 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2009 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2011 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2013 // Generate the induction variable.
2014 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2015 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2016 // The loop step is equal to the vectorization factor (num of SIMD elements)
2017 // times the unroll factor (num of SIMD instructions).
2018 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2020 // This is the IR builder that we use to add all of the logic for bypassing
2021 // the new vector loop.
2022 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2023 setDebugLocFromInst(BypassBuilder,
2024 getDebugLocFromInstOrOperands(OldInduction));
2026 // We may need to extend the index in case there is a type mismatch.
2027 // We know that the count starts at zero and does not overflow.
2028 if (Count->getType() != IdxTy) {
2029 // The exit count can be of pointer type. Convert it to the correct
2031 if (ExitCount->getType()->isPointerTy())
2032 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2034 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2037 // Add the start index to the loop count to get the new end index.
2038 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2040 // Now we need to generate the expression for N - (N % VF), which is
2041 // the part that the vectorized body will execute.
2042 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2043 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2044 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2045 "end.idx.rnd.down");
2047 // Now, compare the new count to zero. If it is zero skip the vector loop and
2048 // jump to the scalar loop.
2049 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
2052 BasicBlock *LastBypassBlock = BypassBlock;
2054 // Generate the code to check that the strides we assumed to be one are really
2055 // one. We want the new basic block to start at the first instruction in a
2056 // sequence of instructions that form a check.
2057 Instruction *StrideCheck;
2058 Instruction *FirstCheckInst;
2059 std::tie(FirstCheckInst, StrideCheck) =
2060 addStrideCheck(BypassBlock->getTerminator());
2062 // Create a new block containing the stride check.
2063 BasicBlock *CheckBlock =
2064 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2066 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2067 LoopBypassBlocks.push_back(CheckBlock);
2069 // Replace the branch into the memory check block with a conditional branch
2070 // for the "few elements case".
2071 Instruction *OldTerm = BypassBlock->getTerminator();
2072 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2073 OldTerm->eraseFromParent();
2076 LastBypassBlock = CheckBlock;
2079 // Generate the code that checks in runtime if arrays overlap. We put the
2080 // checks into a separate block to make the more common case of few elements
2082 Instruction *MemRuntimeCheck;
2083 std::tie(FirstCheckInst, MemRuntimeCheck) =
2084 addRuntimeCheck(LastBypassBlock->getTerminator());
2085 if (MemRuntimeCheck) {
2086 // Create a new block containing the memory check.
2087 BasicBlock *CheckBlock =
2088 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2090 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2091 LoopBypassBlocks.push_back(CheckBlock);
2093 // Replace the branch into the memory check block with a conditional branch
2094 // for the "few elements case".
2095 Instruction *OldTerm = LastBypassBlock->getTerminator();
2096 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2097 OldTerm->eraseFromParent();
2099 Cmp = MemRuntimeCheck;
2100 LastBypassBlock = CheckBlock;
2103 LastBypassBlock->getTerminator()->eraseFromParent();
2104 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2107 // We are going to resume the execution of the scalar loop.
2108 // Go over all of the induction variables that we found and fix the
2109 // PHIs that are left in the scalar version of the loop.
2110 // The starting values of PHI nodes depend on the counter of the last
2111 // iteration in the vectorized loop.
2112 // If we come from a bypass edge then we need to start from the original
2115 // This variable saves the new starting index for the scalar loop.
2116 PHINode *ResumeIndex = nullptr;
2117 LoopVectorizationLegality::InductionList::iterator I, E;
2118 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2119 // Set builder to point to last bypass block.
2120 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2121 for (I = List->begin(), E = List->end(); I != E; ++I) {
2122 PHINode *OrigPhi = I->first;
2123 LoopVectorizationLegality::InductionInfo II = I->second;
2125 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2126 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2127 MiddleBlock->getTerminator());
2128 // We might have extended the type of the induction variable but we need a
2129 // truncated version for the scalar loop.
2130 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2131 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2132 MiddleBlock->getTerminator()) : nullptr;
2134 Value *EndValue = nullptr;
2136 case LoopVectorizationLegality::IK_NoInduction:
2137 llvm_unreachable("Unknown induction");
2138 case LoopVectorizationLegality::IK_IntInduction: {
2139 // Handle the integer induction counter.
2140 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2142 // We have the canonical induction variable.
2143 if (OrigPhi == OldInduction) {
2144 // Create a truncated version of the resume value for the scalar loop,
2145 // we might have promoted the type to a larger width.
2147 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2148 // The new PHI merges the original incoming value, in case of a bypass,
2149 // or the value at the end of the vectorized loop.
2150 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2151 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2152 TruncResumeVal->addIncoming(EndValue, VecBody);
2154 // We know what the end value is.
2155 EndValue = IdxEndRoundDown;
2156 // We also know which PHI node holds it.
2157 ResumeIndex = ResumeVal;
2161 // Not the canonical induction variable - add the vector loop count to the
2163 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2164 II.StartValue->getType(),
2166 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2169 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2170 // Convert the CountRoundDown variable to the PHI size.
2171 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2172 II.StartValue->getType(),
2174 // Handle reverse integer induction counter.
2175 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2178 case LoopVectorizationLegality::IK_PtrInduction: {
2179 // For pointer induction variables, calculate the offset using
2181 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2185 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2186 // The value at the end of the loop for the reverse pointer is calculated
2187 // by creating a GEP with a negative index starting from the start value.
2188 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2189 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2191 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2197 // The new PHI merges the original incoming value, in case of a bypass,
2198 // or the value at the end of the vectorized loop.
2199 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2200 if (OrigPhi == OldInduction)
2201 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2203 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2205 ResumeVal->addIncoming(EndValue, VecBody);
2207 // Fix the scalar body counter (PHI node).
2208 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2209 // The old inductions phi node in the scalar body needs the truncated value.
2210 if (OrigPhi == OldInduction)
2211 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2213 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2216 // If we are generating a new induction variable then we also need to
2217 // generate the code that calculates the exit value. This value is not
2218 // simply the end of the counter because we may skip the vectorized body
2219 // in case of a runtime check.
2221 assert(!ResumeIndex && "Unexpected resume value found");
2222 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2223 MiddleBlock->getTerminator());
2224 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2225 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2226 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2229 // Make sure that we found the index where scalar loop needs to continue.
2230 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2231 "Invalid resume Index");
2233 // Add a check in the middle block to see if we have completed
2234 // all of the iterations in the first vector loop.
2235 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2236 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2237 ResumeIndex, "cmp.n",
2238 MiddleBlock->getTerminator());
2240 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2241 // Remove the old terminator.
2242 MiddleBlock->getTerminator()->eraseFromParent();
2244 // Create i+1 and fill the PHINode.
2245 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2246 Induction->addIncoming(StartIdx, VectorPH);
2247 Induction->addIncoming(NextIdx, VecBody);
2248 // Create the compare.
2249 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2250 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2252 // Now we have two terminators. Remove the old one from the block.
2253 VecBody->getTerminator()->eraseFromParent();
2255 // Get ready to start creating new instructions into the vectorized body.
2256 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2259 LoopVectorPreHeader = VectorPH;
2260 LoopScalarPreHeader = ScalarPH;
2261 LoopMiddleBlock = MiddleBlock;
2262 LoopExitBlock = ExitBlock;
2263 LoopVectorBody.push_back(VecBody);
2264 LoopScalarBody = OldBasicBlock;
2266 LoopVectorizeHints Hints(Lp, true);
2267 Hints.setAlreadyVectorized(Lp);
2270 /// This function returns the identity element (or neutral element) for
2271 /// the operation K.
2273 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2278 // Adding, Xoring, Oring zero to a number does not change it.
2279 return ConstantInt::get(Tp, 0);
2280 case RK_IntegerMult:
2281 // Multiplying a number by 1 does not change it.
2282 return ConstantInt::get(Tp, 1);
2284 // AND-ing a number with an all-1 value does not change it.
2285 return ConstantInt::get(Tp, -1, true);
2287 // Multiplying a number by 1 does not change it.
2288 return ConstantFP::get(Tp, 1.0L);
2290 // Adding zero to a number does not change it.
2291 return ConstantFP::get(Tp, 0.0L);
2293 llvm_unreachable("Unknown reduction kind");
2297 static Intrinsic::ID checkUnaryFloatSignature(const CallInst &I,
2298 Intrinsic::ID ValidIntrinsicID) {
2299 if (I.getNumArgOperands() != 1 ||
2300 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2301 I.getType() != I.getArgOperand(0)->getType() ||
2302 !I.onlyReadsMemory())
2303 return Intrinsic::not_intrinsic;
2305 return ValidIntrinsicID;
2308 static Intrinsic::ID checkBinaryFloatSignature(const CallInst &I,
2309 Intrinsic::ID ValidIntrinsicID) {
2310 if (I.getNumArgOperands() != 2 ||
2311 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
2312 !I.getArgOperand(1)->getType()->isFloatingPointTy() ||
2313 I.getType() != I.getArgOperand(0)->getType() ||
2314 I.getType() != I.getArgOperand(1)->getType() ||
2315 !I.onlyReadsMemory())
2316 return Intrinsic::not_intrinsic;
2318 return ValidIntrinsicID;
2322 static Intrinsic::ID
2323 getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
2324 // If we have an intrinsic call, check if it is trivially vectorizable.
2325 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2326 Intrinsic::ID ID = II->getIntrinsicID();
2327 if (isTriviallyVectorizable(ID) || ID == Intrinsic::lifetime_start ||
2328 ID == Intrinsic::lifetime_end)
2331 return Intrinsic::not_intrinsic;
2335 return Intrinsic::not_intrinsic;
2338 Function *F = CI->getCalledFunction();
2339 // We're going to make assumptions on the semantics of the functions, check
2340 // that the target knows that it's available in this environment and it does
2341 // not have local linkage.
2342 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(F->getName(), Func))
2343 return Intrinsic::not_intrinsic;
2345 // Otherwise check if we have a call to a function that can be turned into a
2346 // vector intrinsic.
2353 return checkUnaryFloatSignature(*CI, Intrinsic::sin);
2357 return checkUnaryFloatSignature(*CI, Intrinsic::cos);
2361 return checkUnaryFloatSignature(*CI, Intrinsic::exp);
2363 case LibFunc::exp2f:
2364 case LibFunc::exp2l:
2365 return checkUnaryFloatSignature(*CI, Intrinsic::exp2);
2369 return checkUnaryFloatSignature(*CI, Intrinsic::log);
2370 case LibFunc::log10:
2371 case LibFunc::log10f:
2372 case LibFunc::log10l:
2373 return checkUnaryFloatSignature(*CI, Intrinsic::log10);
2375 case LibFunc::log2f:
2376 case LibFunc::log2l:
2377 return checkUnaryFloatSignature(*CI, Intrinsic::log2);
2379 case LibFunc::fabsf:
2380 case LibFunc::fabsl:
2381 return checkUnaryFloatSignature(*CI, Intrinsic::fabs);
2382 case LibFunc::copysign:
2383 case LibFunc::copysignf:
2384 case LibFunc::copysignl:
2385 return checkBinaryFloatSignature(*CI, Intrinsic::copysign);
2386 case LibFunc::floor:
2387 case LibFunc::floorf:
2388 case LibFunc::floorl:
2389 return checkUnaryFloatSignature(*CI, Intrinsic::floor);
2391 case LibFunc::ceilf:
2392 case LibFunc::ceill:
2393 return checkUnaryFloatSignature(*CI, Intrinsic::ceil);
2394 case LibFunc::trunc:
2395 case LibFunc::truncf:
2396 case LibFunc::truncl:
2397 return checkUnaryFloatSignature(*CI, Intrinsic::trunc);
2399 case LibFunc::rintf:
2400 case LibFunc::rintl:
2401 return checkUnaryFloatSignature(*CI, Intrinsic::rint);
2402 case LibFunc::nearbyint:
2403 case LibFunc::nearbyintf:
2404 case LibFunc::nearbyintl:
2405 return checkUnaryFloatSignature(*CI, Intrinsic::nearbyint);
2406 case LibFunc::round:
2407 case LibFunc::roundf:
2408 case LibFunc::roundl:
2409 return checkUnaryFloatSignature(*CI, Intrinsic::round);
2413 return checkBinaryFloatSignature(*CI, Intrinsic::pow);
2416 return Intrinsic::not_intrinsic;
2419 /// This function translates the reduction kind to an LLVM binary operator.
2421 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2423 case LoopVectorizationLegality::RK_IntegerAdd:
2424 return Instruction::Add;
2425 case LoopVectorizationLegality::RK_IntegerMult:
2426 return Instruction::Mul;
2427 case LoopVectorizationLegality::RK_IntegerOr:
2428 return Instruction::Or;
2429 case LoopVectorizationLegality::RK_IntegerAnd:
2430 return Instruction::And;
2431 case LoopVectorizationLegality::RK_IntegerXor:
2432 return Instruction::Xor;
2433 case LoopVectorizationLegality::RK_FloatMult:
2434 return Instruction::FMul;
2435 case LoopVectorizationLegality::RK_FloatAdd:
2436 return Instruction::FAdd;
2437 case LoopVectorizationLegality::RK_IntegerMinMax:
2438 return Instruction::ICmp;
2439 case LoopVectorizationLegality::RK_FloatMinMax:
2440 return Instruction::FCmp;
2442 llvm_unreachable("Unknown reduction operation");
2446 Value *createMinMaxOp(IRBuilder<> &Builder,
2447 LoopVectorizationLegality::MinMaxReductionKind RK,
2450 CmpInst::Predicate P = CmpInst::ICMP_NE;
2453 llvm_unreachable("Unknown min/max reduction kind");
2454 case LoopVectorizationLegality::MRK_UIntMin:
2455 P = CmpInst::ICMP_ULT;
2457 case LoopVectorizationLegality::MRK_UIntMax:
2458 P = CmpInst::ICMP_UGT;
2460 case LoopVectorizationLegality::MRK_SIntMin:
2461 P = CmpInst::ICMP_SLT;
2463 case LoopVectorizationLegality::MRK_SIntMax:
2464 P = CmpInst::ICMP_SGT;
2466 case LoopVectorizationLegality::MRK_FloatMin:
2467 P = CmpInst::FCMP_OLT;
2469 case LoopVectorizationLegality::MRK_FloatMax:
2470 P = CmpInst::FCMP_OGT;
2475 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2476 RK == LoopVectorizationLegality::MRK_FloatMax)
2477 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2479 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2481 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2486 struct CSEDenseMapInfo {
2487 static bool canHandle(Instruction *I) {
2488 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2489 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2491 static inline Instruction *getEmptyKey() {
2492 return DenseMapInfo<Instruction *>::getEmptyKey();
2494 static inline Instruction *getTombstoneKey() {
2495 return DenseMapInfo<Instruction *>::getTombstoneKey();
2497 static unsigned getHashValue(Instruction *I) {
2498 assert(canHandle(I) && "Unknown instruction!");
2499 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2500 I->value_op_end()));
2502 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2503 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2504 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2506 return LHS->isIdenticalTo(RHS);
2511 /// \brief Check whether this block is a predicated block.
2512 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2513 /// = ...; " blocks. We start with one vectorized basic block. For every
2514 /// conditional block we split this vectorized block. Therefore, every second
2515 /// block will be a predicated one.
2516 static bool isPredicatedBlock(unsigned BlockNum) {
2517 return BlockNum % 2;
2520 ///\brief Perform cse of induction variable instructions.
2521 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2522 // Perform simple cse.
2523 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2524 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2525 BasicBlock *BB = BBs[i];
2526 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2527 Instruction *In = I++;
2529 if (!CSEDenseMapInfo::canHandle(In))
2532 // Check if we can replace this instruction with any of the
2533 // visited instructions.
2534 if (Instruction *V = CSEMap.lookup(In)) {
2535 In->replaceAllUsesWith(V);
2536 In->eraseFromParent();
2539 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2540 // ...;" blocks for predicated stores. Every second block is a predicated
2542 if (isPredicatedBlock(i))
2550 /// \brief Adds a 'fast' flag to floating point operations.
2551 static Value *addFastMathFlag(Value *V) {
2552 if (isa<FPMathOperator>(V)){
2553 FastMathFlags Flags;
2554 Flags.setUnsafeAlgebra();
2555 cast<Instruction>(V)->setFastMathFlags(Flags);
2560 void InnerLoopVectorizer::vectorizeLoop() {
2561 //===------------------------------------------------===//
2563 // Notice: any optimization or new instruction that go
2564 // into the code below should be also be implemented in
2567 //===------------------------------------------------===//
2568 Constant *Zero = Builder.getInt32(0);
2570 // In order to support reduction variables we need to be able to vectorize
2571 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2572 // stages. First, we create a new vector PHI node with no incoming edges.
2573 // We use this value when we vectorize all of the instructions that use the
2574 // PHI. Next, after all of the instructions in the block are complete we
2575 // add the new incoming edges to the PHI. At this point all of the
2576 // instructions in the basic block are vectorized, so we can use them to
2577 // construct the PHI.
2578 PhiVector RdxPHIsToFix;
2580 // Scan the loop in a topological order to ensure that defs are vectorized
2582 LoopBlocksDFS DFS(OrigLoop);
2585 // Vectorize all of the blocks in the original loop.
2586 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2587 be = DFS.endRPO(); bb != be; ++bb)
2588 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2590 // At this point every instruction in the original loop is widened to
2591 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2592 // that we vectorized. The PHI nodes are currently empty because we did
2593 // not want to introduce cycles. Notice that the remaining PHI nodes
2594 // that we need to fix are reduction variables.
2596 // Create the 'reduced' values for each of the induction vars.
2597 // The reduced values are the vector values that we scalarize and combine
2598 // after the loop is finished.
2599 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2601 PHINode *RdxPhi = *it;
2602 assert(RdxPhi && "Unable to recover vectorized PHI");
2604 // Find the reduction variable descriptor.
2605 assert(Legal->getReductionVars()->count(RdxPhi) &&
2606 "Unable to find the reduction variable");
2607 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2608 (*Legal->getReductionVars())[RdxPhi];
2610 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2612 // We need to generate a reduction vector from the incoming scalar.
2613 // To do so, we need to generate the 'identity' vector and override
2614 // one of the elements with the incoming scalar reduction. We need
2615 // to do it in the vector-loop preheader.
2616 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2618 // This is the vector-clone of the value that leaves the loop.
2619 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2620 Type *VecTy = VectorExit[0]->getType();
2622 // Find the reduction identity variable. Zero for addition, or, xor,
2623 // one for multiplication, -1 for And.
2626 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2627 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2628 // MinMax reduction have the start value as their identify.
2630 VectorStart = Identity = RdxDesc.StartValue;
2632 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2637 // Handle other reduction kinds:
2639 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2640 VecTy->getScalarType());
2643 // This vector is the Identity vector where the first element is the
2644 // incoming scalar reduction.
2645 VectorStart = RdxDesc.StartValue;
2647 Identity = ConstantVector::getSplat(VF, Iden);
2649 // This vector is the Identity vector where the first element is the
2650 // incoming scalar reduction.
2651 VectorStart = Builder.CreateInsertElement(Identity,
2652 RdxDesc.StartValue, Zero);
2656 // Fix the vector-loop phi.
2657 // We created the induction variable so we know that the
2658 // preheader is the first entry.
2659 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2661 // Reductions do not have to start at zero. They can start with
2662 // any loop invariant values.
2663 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2664 BasicBlock *Latch = OrigLoop->getLoopLatch();
2665 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2666 VectorParts &Val = getVectorValue(LoopVal);
2667 for (unsigned part = 0; part < UF; ++part) {
2668 // Make sure to add the reduction stat value only to the
2669 // first unroll part.
2670 Value *StartVal = (part == 0) ? VectorStart : Identity;
2671 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2672 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2673 LoopVectorBody.back());
2676 // Before each round, move the insertion point right between
2677 // the PHIs and the values we are going to write.
2678 // This allows us to write both PHINodes and the extractelement
2680 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2682 VectorParts RdxParts;
2683 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2684 for (unsigned part = 0; part < UF; ++part) {
2685 // This PHINode contains the vectorized reduction variable, or
2686 // the initial value vector, if we bypass the vector loop.
2687 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2688 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2689 Value *StartVal = (part == 0) ? VectorStart : Identity;
2690 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2691 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2692 NewPhi->addIncoming(RdxExitVal[part],
2693 LoopVectorBody.back());
2694 RdxParts.push_back(NewPhi);
2697 // Reduce all of the unrolled parts into a single vector.
2698 Value *ReducedPartRdx = RdxParts[0];
2699 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2700 setDebugLocFromInst(Builder, ReducedPartRdx);
2701 for (unsigned part = 1; part < UF; ++part) {
2702 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2703 // Floating point operations had to be 'fast' to enable the reduction.
2704 ReducedPartRdx = addFastMathFlag(
2705 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2706 ReducedPartRdx, "bin.rdx"));
2708 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2709 ReducedPartRdx, RdxParts[part]);
2713 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2714 // and vector ops, reducing the set of values being computed by half each
2716 assert(isPowerOf2_32(VF) &&
2717 "Reduction emission only supported for pow2 vectors!");
2718 Value *TmpVec = ReducedPartRdx;
2719 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2720 for (unsigned i = VF; i != 1; i >>= 1) {
2721 // Move the upper half of the vector to the lower half.
2722 for (unsigned j = 0; j != i/2; ++j)
2723 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2725 // Fill the rest of the mask with undef.
2726 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2727 UndefValue::get(Builder.getInt32Ty()));
2730 Builder.CreateShuffleVector(TmpVec,
2731 UndefValue::get(TmpVec->getType()),
2732 ConstantVector::get(ShuffleMask),
2735 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2736 // Floating point operations had to be 'fast' to enable the reduction.
2737 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2738 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2740 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2743 // The result is in the first element of the vector.
2744 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2745 Builder.getInt32(0));
2748 // Now, we need to fix the users of the reduction variable
2749 // inside and outside of the scalar remainder loop.
2750 // We know that the loop is in LCSSA form. We need to update the
2751 // PHI nodes in the exit blocks.
2752 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2753 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2754 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2755 if (!LCSSAPhi) break;
2757 // All PHINodes need to have a single entry edge, or two if
2758 // we already fixed them.
2759 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2761 // We found our reduction value exit-PHI. Update it with the
2762 // incoming bypass edge.
2763 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2764 // Add an edge coming from the bypass.
2765 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2768 }// end of the LCSSA phi scan.
2770 // Fix the scalar loop reduction variable with the incoming reduction sum
2771 // from the vector body and from the backedge value.
2772 int IncomingEdgeBlockIdx =
2773 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2774 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2775 // Pick the other block.
2776 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2777 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2778 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2779 }// end of for each redux variable.
2783 // Remove redundant induction instructions.
2784 cse(LoopVectorBody);
2787 void InnerLoopVectorizer::fixLCSSAPHIs() {
2788 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2789 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2790 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2791 if (!LCSSAPhi) break;
2792 if (LCSSAPhi->getNumIncomingValues() == 1)
2793 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2798 InnerLoopVectorizer::VectorParts
2799 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2800 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2803 // Look for cached value.
2804 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2805 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2806 if (ECEntryIt != MaskCache.end())
2807 return ECEntryIt->second;
2809 VectorParts SrcMask = createBlockInMask(Src);
2811 // The terminator has to be a branch inst!
2812 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2813 assert(BI && "Unexpected terminator found");
2815 if (BI->isConditional()) {
2816 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2818 if (BI->getSuccessor(0) != Dst)
2819 for (unsigned part = 0; part < UF; ++part)
2820 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2822 for (unsigned part = 0; part < UF; ++part)
2823 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2825 MaskCache[Edge] = EdgeMask;
2829 MaskCache[Edge] = SrcMask;
2833 InnerLoopVectorizer::VectorParts
2834 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2835 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2837 // Loop incoming mask is all-one.
2838 if (OrigLoop->getHeader() == BB) {
2839 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2840 return getVectorValue(C);
2843 // This is the block mask. We OR all incoming edges, and with zero.
2844 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2845 VectorParts BlockMask = getVectorValue(Zero);
2848 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2849 VectorParts EM = createEdgeMask(*it, BB);
2850 for (unsigned part = 0; part < UF; ++part)
2851 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2857 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2858 InnerLoopVectorizer::VectorParts &Entry,
2859 unsigned UF, unsigned VF, PhiVector *PV) {
2860 PHINode* P = cast<PHINode>(PN);
2861 // Handle reduction variables:
2862 if (Legal->getReductionVars()->count(P)) {
2863 for (unsigned part = 0; part < UF; ++part) {
2864 // This is phase one of vectorizing PHIs.
2865 Type *VecTy = (VF == 1) ? PN->getType() :
2866 VectorType::get(PN->getType(), VF);
2867 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2868 LoopVectorBody.back()-> getFirstInsertionPt());
2874 setDebugLocFromInst(Builder, P);
2875 // Check for PHI nodes that are lowered to vector selects.
2876 if (P->getParent() != OrigLoop->getHeader()) {
2877 // We know that all PHIs in non-header blocks are converted into
2878 // selects, so we don't have to worry about the insertion order and we
2879 // can just use the builder.
2880 // At this point we generate the predication tree. There may be
2881 // duplications since this is a simple recursive scan, but future
2882 // optimizations will clean it up.
2884 unsigned NumIncoming = P->getNumIncomingValues();
2886 // Generate a sequence of selects of the form:
2887 // SELECT(Mask3, In3,
2888 // SELECT(Mask2, In2,
2890 for (unsigned In = 0; In < NumIncoming; In++) {
2891 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2893 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2895 for (unsigned part = 0; part < UF; ++part) {
2896 // We might have single edge PHIs (blocks) - use an identity
2897 // 'select' for the first PHI operand.
2899 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2902 // Select between the current value and the previous incoming edge
2903 // based on the incoming mask.
2904 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2905 Entry[part], "predphi");
2911 // This PHINode must be an induction variable.
2912 // Make sure that we know about it.
2913 assert(Legal->getInductionVars()->count(P) &&
2914 "Not an induction variable");
2916 LoopVectorizationLegality::InductionInfo II =
2917 Legal->getInductionVars()->lookup(P);
2920 case LoopVectorizationLegality::IK_NoInduction:
2921 llvm_unreachable("Unknown induction");
2922 case LoopVectorizationLegality::IK_IntInduction: {
2923 assert(P->getType() == II.StartValue->getType() && "Types must match");
2924 Type *PhiTy = P->getType();
2926 if (P == OldInduction) {
2927 // Handle the canonical induction variable. We might have had to
2929 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2931 // Handle other induction variables that are now based on the
2933 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2935 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2936 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2939 Broadcasted = getBroadcastInstrs(Broadcasted);
2940 // After broadcasting the induction variable we need to make the vector
2941 // consecutive by adding 0, 1, 2, etc.
2942 for (unsigned part = 0; part < UF; ++part)
2943 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2946 case LoopVectorizationLegality::IK_ReverseIntInduction:
2947 case LoopVectorizationLegality::IK_PtrInduction:
2948 case LoopVectorizationLegality::IK_ReversePtrInduction:
2949 // Handle reverse integer and pointer inductions.
2950 Value *StartIdx = ExtendedIdx;
2951 // This is the normalized GEP that starts counting at zero.
2952 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2955 // Handle the reverse integer induction variable case.
2956 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2957 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2958 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2960 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2963 // This is a new value so do not hoist it out.
2964 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2965 // After broadcasting the induction variable we need to make the
2966 // vector consecutive by adding ... -3, -2, -1, 0.
2967 for (unsigned part = 0; part < UF; ++part)
2968 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2973 // Handle the pointer induction variable case.
2974 assert(P->getType()->isPointerTy() && "Unexpected type.");
2976 // Is this a reverse induction ptr or a consecutive induction ptr.
2977 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2980 // This is the vector of results. Notice that we don't generate
2981 // vector geps because scalar geps result in better code.
2982 for (unsigned part = 0; part < UF; ++part) {
2984 int EltIndex = (part) * (Reverse ? -1 : 1);
2985 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2988 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2990 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2992 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2994 Entry[part] = SclrGep;
2998 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2999 for (unsigned int i = 0; i < VF; ++i) {
3000 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3001 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3004 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3006 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3008 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3010 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3011 Builder.getInt32(i),
3014 Entry[part] = VecVal;
3020 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3021 // For each instruction in the old loop.
3022 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3023 VectorParts &Entry = WidenMap.get(it);
3024 switch (it->getOpcode()) {
3025 case Instruction::Br:
3026 // Nothing to do for PHIs and BR, since we already took care of the
3027 // loop control flow instructions.
3029 case Instruction::PHI:{
3030 // Vectorize PHINodes.
3031 widenPHIInstruction(it, Entry, UF, VF, PV);
3035 case Instruction::Add:
3036 case Instruction::FAdd:
3037 case Instruction::Sub:
3038 case Instruction::FSub:
3039 case Instruction::Mul:
3040 case Instruction::FMul:
3041 case Instruction::UDiv:
3042 case Instruction::SDiv:
3043 case Instruction::FDiv:
3044 case Instruction::URem:
3045 case Instruction::SRem:
3046 case Instruction::FRem:
3047 case Instruction::Shl:
3048 case Instruction::LShr:
3049 case Instruction::AShr:
3050 case Instruction::And:
3051 case Instruction::Or:
3052 case Instruction::Xor: {
3053 // Just widen binops.
3054 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3055 setDebugLocFromInst(Builder, BinOp);
3056 VectorParts &A = getVectorValue(it->getOperand(0));
3057 VectorParts &B = getVectorValue(it->getOperand(1));
3059 // Use this vector value for all users of the original instruction.
3060 for (unsigned Part = 0; Part < UF; ++Part) {
3061 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3063 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3064 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3065 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3066 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3067 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3069 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3070 VecOp->setIsExact(BinOp->isExact());
3072 // Copy the fast-math flags.
3073 if (VecOp && isa<FPMathOperator>(V))
3074 VecOp->setFastMathFlags(it->getFastMathFlags());
3080 case Instruction::Select: {
3082 // If the selector is loop invariant we can create a select
3083 // instruction with a scalar condition. Otherwise, use vector-select.
3084 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3086 setDebugLocFromInst(Builder, it);
3088 // The condition can be loop invariant but still defined inside the
3089 // loop. This means that we can't just use the original 'cond' value.
3090 // We have to take the 'vectorized' value and pick the first lane.
3091 // Instcombine will make this a no-op.
3092 VectorParts &Cond = getVectorValue(it->getOperand(0));
3093 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3094 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3096 Value *ScalarCond = (VF == 1) ? Cond[0] :
3097 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3099 for (unsigned Part = 0; Part < UF; ++Part) {
3100 Entry[Part] = Builder.CreateSelect(
3101 InvariantCond ? ScalarCond : Cond[Part],
3108 case Instruction::ICmp:
3109 case Instruction::FCmp: {
3110 // Widen compares. Generate vector compares.
3111 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3112 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3113 setDebugLocFromInst(Builder, it);
3114 VectorParts &A = getVectorValue(it->getOperand(0));
3115 VectorParts &B = getVectorValue(it->getOperand(1));
3116 for (unsigned Part = 0; Part < UF; ++Part) {
3119 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3121 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3127 case Instruction::Store:
3128 case Instruction::Load:
3129 vectorizeMemoryInstruction(it);
3131 case Instruction::ZExt:
3132 case Instruction::SExt:
3133 case Instruction::FPToUI:
3134 case Instruction::FPToSI:
3135 case Instruction::FPExt:
3136 case Instruction::PtrToInt:
3137 case Instruction::IntToPtr:
3138 case Instruction::SIToFP:
3139 case Instruction::UIToFP:
3140 case Instruction::Trunc:
3141 case Instruction::FPTrunc:
3142 case Instruction::BitCast: {
3143 CastInst *CI = dyn_cast<CastInst>(it);
3144 setDebugLocFromInst(Builder, it);
3145 /// Optimize the special case where the source is the induction
3146 /// variable. Notice that we can only optimize the 'trunc' case
3147 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3148 /// c. other casts depend on pointer size.
3149 if (CI->getOperand(0) == OldInduction &&
3150 it->getOpcode() == Instruction::Trunc) {
3151 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3153 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3154 for (unsigned Part = 0; Part < UF; ++Part)
3155 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3158 /// Vectorize casts.
3159 Type *DestTy = (VF == 1) ? CI->getType() :
3160 VectorType::get(CI->getType(), VF);
3162 VectorParts &A = getVectorValue(it->getOperand(0));
3163 for (unsigned Part = 0; Part < UF; ++Part)
3164 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3168 case Instruction::Call: {
3169 // Ignore dbg intrinsics.
3170 if (isa<DbgInfoIntrinsic>(it))
3172 setDebugLocFromInst(Builder, it);
3174 Module *M = BB->getParent()->getParent();
3175 CallInst *CI = cast<CallInst>(it);
3176 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3177 assert(ID && "Not an intrinsic call!");
3179 case Intrinsic::lifetime_end:
3180 case Intrinsic::lifetime_start:
3181 scalarizeInstruction(it);
3184 for (unsigned Part = 0; Part < UF; ++Part) {
3185 SmallVector<Value *, 4> Args;
3186 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3187 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3188 Args.push_back(Arg[Part]);
3190 Type *Tys[] = {CI->getType()};
3192 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3194 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3195 Entry[Part] = Builder.CreateCall(F, Args);
3203 // All other instructions are unsupported. Scalarize them.
3204 scalarizeInstruction(it);
3207 }// end of for_each instr.
3210 void InnerLoopVectorizer::updateAnalysis() {
3211 // Forget the original basic block.
3212 SE->forgetLoop(OrigLoop);
3214 // Update the dominator tree information.
3215 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3216 "Entry does not dominate exit.");
3218 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3219 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3220 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3222 // Due to if predication of stores we might create a sequence of "if(pred)
3223 // a[i] = ...; " blocks.
3224 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3226 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3227 else if (isPredicatedBlock(i)) {
3228 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3230 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3234 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3235 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3236 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3237 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3239 DEBUG(DT->verifyDomTree());
3242 /// \brief Check whether it is safe to if-convert this phi node.
3244 /// Phi nodes with constant expressions that can trap are not safe to if
3246 static bool canIfConvertPHINodes(BasicBlock *BB) {
3247 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3248 PHINode *Phi = dyn_cast<PHINode>(I);
3251 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3252 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3259 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3260 if (!EnableIfConversion)
3263 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3265 // A list of pointers that we can safely read and write to.
3266 SmallPtrSet<Value *, 8> SafePointes;
3268 // Collect safe addresses.
3269 for (Loop::block_iterator BI = TheLoop->block_begin(),
3270 BE = TheLoop->block_end(); BI != BE; ++BI) {
3271 BasicBlock *BB = *BI;
3273 if (blockNeedsPredication(BB))
3276 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3277 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3278 SafePointes.insert(LI->getPointerOperand());
3279 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3280 SafePointes.insert(SI->getPointerOperand());
3284 // Collect the blocks that need predication.
3285 BasicBlock *Header = TheLoop->getHeader();
3286 for (Loop::block_iterator BI = TheLoop->block_begin(),
3287 BE = TheLoop->block_end(); BI != BE; ++BI) {
3288 BasicBlock *BB = *BI;
3290 // We don't support switch statements inside loops.
3291 if (!isa<BranchInst>(BB->getTerminator()))
3294 // We must be able to predicate all blocks that need to be predicated.
3295 if (blockNeedsPredication(BB)) {
3296 if (!blockCanBePredicated(BB, SafePointes))
3298 } else if (BB != Header && !canIfConvertPHINodes(BB))
3303 // We can if-convert this loop.
3307 bool LoopVectorizationLegality::canVectorize() {
3308 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3309 // be canonicalized.
3310 if (!TheLoop->getLoopPreheader())
3313 // We can only vectorize innermost loops.
3314 if (TheLoop->getSubLoopsVector().size())
3317 // We must have a single backedge.
3318 if (TheLoop->getNumBackEdges() != 1)
3321 // We must have a single exiting block.
3322 if (!TheLoop->getExitingBlock())
3325 // We need to have a loop header.
3326 DEBUG(dbgs() << "LV: Found a loop: " <<
3327 TheLoop->getHeader()->getName() << '\n');
3329 // Check if we can if-convert non-single-bb loops.
3330 unsigned NumBlocks = TheLoop->getNumBlocks();
3331 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3332 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3336 // ScalarEvolution needs to be able to find the exit count.
3337 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3338 if (ExitCount == SE->getCouldNotCompute()) {
3339 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3343 // Check if we can vectorize the instructions and CFG in this loop.
3344 if (!canVectorizeInstrs()) {
3345 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3349 // Go over each instruction and look at memory deps.
3350 if (!canVectorizeMemory()) {
3351 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3355 // Collect all of the variables that remain uniform after vectorization.
3356 collectLoopUniforms();
3358 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3359 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3362 // Okay! We can vectorize. At this point we don't have any other mem analysis
3363 // which may limit our maximum vectorization factor, so just return true with
3368 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3369 if (Ty->isPointerTy())
3370 return DL.getIntPtrType(Ty);
3372 // It is possible that char's or short's overflow when we ask for the loop's
3373 // trip count, work around this by changing the type size.
3374 if (Ty->getScalarSizeInBits() < 32)
3375 return Type::getInt32Ty(Ty->getContext());
3380 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3381 Ty0 = convertPointerToIntegerType(DL, Ty0);
3382 Ty1 = convertPointerToIntegerType(DL, Ty1);
3383 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3388 /// \brief Check that the instruction has outside loop users and is not an
3389 /// identified reduction variable.
3390 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3391 SmallPtrSet<Value *, 4> &Reductions) {
3392 // Reduction instructions are allowed to have exit users. All other
3393 // instructions must not have external users.
3394 if (!Reductions.count(Inst))
3395 //Check that all of the users of the loop are inside the BB.
3396 for (User *U : Inst->users()) {
3397 Instruction *UI = cast<Instruction>(U);
3398 // This user may be a reduction exit value.
3399 if (!TheLoop->contains(UI)) {
3400 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3407 bool LoopVectorizationLegality::canVectorizeInstrs() {
3408 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3409 BasicBlock *Header = TheLoop->getHeader();
3411 // Look for the attribute signaling the absence of NaNs.
3412 Function &F = *Header->getParent();
3413 if (F.hasFnAttribute("no-nans-fp-math"))
3414 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3415 AttributeSet::FunctionIndex,
3416 "no-nans-fp-math").getValueAsString() == "true";
3418 // For each block in the loop.
3419 for (Loop::block_iterator bb = TheLoop->block_begin(),
3420 be = TheLoop->block_end(); bb != be; ++bb) {
3422 // Scan the instructions in the block and look for hazards.
3423 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3426 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3427 Type *PhiTy = Phi->getType();
3428 // Check that this PHI type is allowed.
3429 if (!PhiTy->isIntegerTy() &&
3430 !PhiTy->isFloatingPointTy() &&
3431 !PhiTy->isPointerTy()) {
3432 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3436 // If this PHINode is not in the header block, then we know that we
3437 // can convert it to select during if-conversion. No need to check if
3438 // the PHIs in this block are induction or reduction variables.
3439 if (*bb != Header) {
3440 // Check that this instruction has no outside users or is an
3441 // identified reduction value with an outside user.
3442 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3447 // We only allow if-converted PHIs with more than two incoming values.
3448 if (Phi->getNumIncomingValues() != 2) {
3449 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3453 // This is the value coming from the preheader.
3454 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3455 // Check if this is an induction variable.
3456 InductionKind IK = isInductionVariable(Phi);
3458 if (IK_NoInduction != IK) {
3459 // Get the widest type.
3461 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3463 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3465 // Int inductions are special because we only allow one IV.
3466 if (IK == IK_IntInduction) {
3467 // Use the phi node with the widest type as induction. Use the last
3468 // one if there are multiple (no good reason for doing this other
3469 // than it is expedient).
3470 if (!Induction || PhiTy == WidestIndTy)
3474 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3475 Inductions[Phi] = InductionInfo(StartValue, IK);
3477 // Until we explicitly handle the case of an induction variable with
3478 // an outside loop user we have to give up vectorizing this loop.
3479 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3485 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3486 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3489 if (AddReductionVar(Phi, RK_IntegerMult)) {
3490 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3493 if (AddReductionVar(Phi, RK_IntegerOr)) {
3494 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3497 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3498 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3501 if (AddReductionVar(Phi, RK_IntegerXor)) {
3502 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3505 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3506 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3509 if (AddReductionVar(Phi, RK_FloatMult)) {
3510 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3513 if (AddReductionVar(Phi, RK_FloatAdd)) {
3514 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3517 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3518 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3523 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3525 }// end of PHI handling
3527 // We still don't handle functions. However, we can ignore dbg intrinsic
3528 // calls and we do handle certain intrinsic and libm functions.
3529 CallInst *CI = dyn_cast<CallInst>(it);
3530 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3531 DEBUG(dbgs() << "LV: Found a call site.\n");
3535 // Check that the instruction return type is vectorizable.
3536 // Also, we can't vectorize extractelement instructions.
3537 if ((!VectorType::isValidElementType(it->getType()) &&
3538 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3539 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3543 // Check that the stored type is vectorizable.
3544 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3545 Type *T = ST->getValueOperand()->getType();
3546 if (!VectorType::isValidElementType(T))
3548 if (EnableMemAccessVersioning)
3549 collectStridedAcccess(ST);
3552 if (EnableMemAccessVersioning)
3553 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3554 collectStridedAcccess(LI);
3556 // Reduction instructions are allowed to have exit users.
3557 // All other instructions must not have external users.
3558 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3566 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3567 if (Inductions.empty())
3574 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3575 /// return the induction operand of the gep pointer.
3576 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3577 const DataLayout *DL, Loop *Lp) {
3578 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3582 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3584 // Check that all of the gep indices are uniform except for our induction
3586 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3587 if (i != InductionOperand &&
3588 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3590 return GEP->getOperand(InductionOperand);
3593 ///\brief Look for a cast use of the passed value.
3594 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3595 Value *UniqueCast = nullptr;
3596 for (User *U : Ptr->users()) {
3597 CastInst *CI = dyn_cast<CastInst>(U);
3598 if (CI && CI->getType() == Ty) {
3608 ///\brief Get the stride of a pointer access in a loop.
3609 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3610 /// pointer to the Value, or null otherwise.
3611 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3612 const DataLayout *DL, Loop *Lp) {
3613 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3614 if (!PtrTy || PtrTy->isAggregateType())
3617 // Try to remove a gep instruction to make the pointer (actually index at this
3618 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3619 // pointer, otherwise, we are analyzing the index.
3620 Value *OrigPtr = Ptr;
3622 // The size of the pointer access.
3623 int64_t PtrAccessSize = 1;
3625 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3626 const SCEV *V = SE->getSCEV(Ptr);
3630 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3631 V = C->getOperand();
3633 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3637 V = S->getStepRecurrence(*SE);
3641 // Strip off the size of access multiplication if we are still analyzing the
3643 if (OrigPtr == Ptr) {
3644 DL->getTypeAllocSize(PtrTy->getElementType());
3645 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3646 if (M->getOperand(0)->getSCEVType() != scConstant)
3649 const APInt &APStepVal =
3650 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3652 // Huge step value - give up.
3653 if (APStepVal.getBitWidth() > 64)
3656 int64_t StepVal = APStepVal.getSExtValue();
3657 if (PtrAccessSize != StepVal)
3659 V = M->getOperand(1);
3664 Type *StripedOffRecurrenceCast = nullptr;
3665 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3666 StripedOffRecurrenceCast = C->getType();
3667 V = C->getOperand();
3670 // Look for the loop invariant symbolic value.
3671 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3675 Value *Stride = U->getValue();
3676 if (!Lp->isLoopInvariant(Stride))
3679 // If we have stripped off the recurrence cast we have to make sure that we
3680 // return the value that is used in this loop so that we can replace it later.
3681 if (StripedOffRecurrenceCast)
3682 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3687 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3688 Value *Ptr = nullptr;
3689 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3690 Ptr = LI->getPointerOperand();
3691 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3692 Ptr = SI->getPointerOperand();
3696 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3700 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3701 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3702 Strides[Ptr] = Stride;
3703 StrideSet.insert(Stride);
3706 void LoopVectorizationLegality::collectLoopUniforms() {
3707 // We now know that the loop is vectorizable!
3708 // Collect variables that will remain uniform after vectorization.
3709 std::vector<Value*> Worklist;
3710 BasicBlock *Latch = TheLoop->getLoopLatch();
3712 // Start with the conditional branch and walk up the block.
3713 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3715 // Also add all consecutive pointer values; these values will be uniform
3716 // after vectorization (and subsequent cleanup) and, until revectorization is
3717 // supported, all dependencies must also be uniform.
3718 for (Loop::block_iterator B = TheLoop->block_begin(),
3719 BE = TheLoop->block_end(); B != BE; ++B)
3720 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3722 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3723 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3725 while (Worklist.size()) {
3726 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3727 Worklist.pop_back();
3729 // Look at instructions inside this loop.
3730 // Stop when reaching PHI nodes.
3731 // TODO: we need to follow values all over the loop, not only in this block.
3732 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3735 // This is a known uniform.
3738 // Insert all operands.
3739 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3744 /// \brief Analyses memory accesses in a loop.
3746 /// Checks whether run time pointer checks are needed and builds sets for data
3747 /// dependence checking.
3748 class AccessAnalysis {
3750 /// \brief Read or write access location.
3751 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3752 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3754 /// \brief Set of potential dependent memory accesses.
3755 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3757 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3758 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3759 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3761 /// \brief Register a load and whether it is only read from.
3762 void addLoad(Value *Ptr, bool IsReadOnly) {
3763 Accesses.insert(MemAccessInfo(Ptr, false));
3765 ReadOnlyPtr.insert(Ptr);
3768 /// \brief Register a store.
3769 void addStore(Value *Ptr) {
3770 Accesses.insert(MemAccessInfo(Ptr, true));
3773 /// \brief Check whether we can check the pointers at runtime for
3774 /// non-intersection.
3775 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3776 unsigned &NumComparisons, ScalarEvolution *SE,
3777 Loop *TheLoop, ValueToValueMap &Strides,
3778 bool ShouldCheckStride = false);
3780 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3781 /// and builds sets of dependent accesses.
3782 void buildDependenceSets() {
3783 // Process read-write pointers first.
3784 processMemAccesses(false);
3785 // Next, process read pointers.
3786 processMemAccesses(true);
3789 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3791 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3792 void resetDepChecks() { CheckDeps.clear(); }
3794 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3797 typedef SetVector<MemAccessInfo> PtrAccessSet;
3798 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3800 /// \brief Go over all memory access or only the deferred ones if
3801 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3802 /// and build sets of dependency check candidates.
3803 void processMemAccesses(bool UseDeferred);
3805 /// Set of all accesses.
3806 PtrAccessSet Accesses;
3808 /// Set of access to check after all writes have been processed.
3809 PtrAccessSet DeferredAccesses;
3811 /// Map of pointers to last access encountered.
3812 UnderlyingObjToAccessMap ObjToLastAccess;
3814 /// Set of accesses that need a further dependence check.
3815 MemAccessInfoSet CheckDeps;
3817 /// Set of pointers that are read only.
3818 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3820 /// Set of underlying objects already written to.
3821 SmallPtrSet<Value*, 16> WriteObjects;
3823 const DataLayout *DL;
3825 /// Sets of potentially dependent accesses - members of one set share an
3826 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3827 /// dependence check.
3828 DepCandidates &DepCands;
3830 bool AreAllWritesIdentified;
3831 bool AreAllReadsIdentified;
3832 bool IsRTCheckNeeded;
3835 } // end anonymous namespace
3837 /// \brief Check whether a pointer can participate in a runtime bounds check.
3838 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3840 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3841 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3845 return AR->isAffine();
3848 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3849 /// the address space.
3850 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3851 const Loop *Lp, ValueToValueMap &StridesMap);
3853 bool AccessAnalysis::canCheckPtrAtRT(
3854 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3855 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3856 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3857 // Find pointers with computable bounds. We are going to use this information
3858 // to place a runtime bound check.
3859 unsigned NumReadPtrChecks = 0;
3860 unsigned NumWritePtrChecks = 0;
3861 bool CanDoRT = true;
3863 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3864 // We assign consecutive id to access from different dependence sets.
3865 // Accesses within the same set don't need a runtime check.
3866 unsigned RunningDepId = 1;
3867 DenseMap<Value *, unsigned> DepSetId;
3869 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3871 const MemAccessInfo &Access = *AI;
3872 Value *Ptr = Access.getPointer();
3873 bool IsWrite = Access.getInt();
3875 // Just add write checks if we have both.
3876 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3880 ++NumWritePtrChecks;
3884 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3885 // When we run after a failing dependency check we have to make sure we
3886 // don't have wrapping pointers.
3887 (!ShouldCheckStride ||
3888 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3889 // The id of the dependence set.
3892 if (IsDepCheckNeeded) {
3893 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3894 unsigned &LeaderId = DepSetId[Leader];
3896 LeaderId = RunningDepId++;
3899 // Each access has its own dependence set.
3900 DepId = RunningDepId++;
3902 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3904 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3910 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3911 NumComparisons = 0; // Only one dependence set.
3913 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3914 NumWritePtrChecks - 1));
3917 // If the pointers that we would use for the bounds comparison have different
3918 // address spaces, assume the values aren't directly comparable, so we can't
3919 // use them for the runtime check. We also have to assume they could
3920 // overlap. In the future there should be metadata for whether address spaces
3922 unsigned NumPointers = RtCheck.Pointers.size();
3923 for (unsigned i = 0; i < NumPointers; ++i) {
3924 for (unsigned j = i + 1; j < NumPointers; ++j) {
3925 // Only need to check pointers between two different dependency sets.
3926 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3929 Value *PtrI = RtCheck.Pointers[i];
3930 Value *PtrJ = RtCheck.Pointers[j];
3932 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3933 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3935 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3936 " different address spaces\n");
3945 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3946 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3949 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3950 // We process the set twice: first we process read-write pointers, last we
3951 // process read-only pointers. This allows us to skip dependence tests for
3952 // read-only pointers.
3954 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3955 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3956 const MemAccessInfo &Access = *AI;
3957 Value *Ptr = Access.getPointer();
3958 bool IsWrite = Access.getInt();
3960 DepCands.insert(Access);
3962 // Memorize read-only pointers for later processing and skip them in the
3963 // first round (they need to be checked after we have seen all write
3964 // pointers). Note: we also mark pointer that are not consecutive as
3965 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3966 // second check for "!IsWrite".
3967 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3968 if (!UseDeferred && IsReadOnlyPtr) {
3969 DeferredAccesses.insert(Access);
3973 bool NeedDepCheck = false;
3974 // Check whether there is the possibility of dependency because of
3975 // underlying objects being the same.
3976 typedef SmallVector<Value*, 16> ValueVector;
3977 ValueVector TempObjects;
3978 GetUnderlyingObjects(Ptr, TempObjects, DL);
3979 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3981 Value *UnderlyingObj = *UI;
3983 // If this is a write then it needs to be an identified object. If this a
3984 // read and all writes (so far) are identified function scope objects we
3985 // don't need an identified underlying object but only an Argument (the
3986 // next write is going to invalidate this assumption if it is
3988 // This is a micro-optimization for the case where all writes are
3989 // identified and we have one argument pointer.
3990 // Otherwise, we do need a runtime check.
3991 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3992 (!IsWrite && (!AreAllWritesIdentified ||
3993 !isa<Argument>(UnderlyingObj)) &&
3994 !isIdentifiedObject(UnderlyingObj))) {
3995 DEBUG(dbgs() << "LV: Found an unidentified " <<
3996 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3998 IsRTCheckNeeded = (IsRTCheckNeeded ||
3999 !isIdentifiedObject(UnderlyingObj) ||
4000 !AreAllReadsIdentified);
4003 AreAllWritesIdentified = false;
4005 AreAllReadsIdentified = false;
4008 // If this is a write - check other reads and writes for conflicts. If
4009 // this is a read only check other writes for conflicts (but only if there
4010 // is no other write to the ptr - this is an optimization to catch "a[i] =
4011 // a[i] + " without having to do a dependence check).
4012 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
4013 NeedDepCheck = true;
4016 WriteObjects.insert(UnderlyingObj);
4018 // Create sets of pointers connected by shared underlying objects.
4019 UnderlyingObjToAccessMap::iterator Prev =
4020 ObjToLastAccess.find(UnderlyingObj);
4021 if (Prev != ObjToLastAccess.end())
4022 DepCands.unionSets(Access, Prev->second);
4024 ObjToLastAccess[UnderlyingObj] = Access;
4028 CheckDeps.insert(Access);
4033 /// \brief Checks memory dependences among accesses to the same underlying
4034 /// object to determine whether there vectorization is legal or not (and at
4035 /// which vectorization factor).
4037 /// This class works under the assumption that we already checked that memory
4038 /// locations with different underlying pointers are "must-not alias".
4039 /// We use the ScalarEvolution framework to symbolically evalutate access
4040 /// functions pairs. Since we currently don't restructure the loop we can rely
4041 /// on the program order of memory accesses to determine their safety.
4042 /// At the moment we will only deem accesses as safe for:
4043 /// * A negative constant distance assuming program order.
4045 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4046 /// a[i] = tmp; y = a[i];
4048 /// The latter case is safe because later checks guarantuee that there can't
4049 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4050 /// the same variable: a header phi can only be an induction or a reduction, a
4051 /// reduction can't have a memory sink, an induction can't have a memory
4052 /// source). This is important and must not be violated (or we have to
4053 /// resort to checking for cycles through memory).
4055 /// * A positive constant distance assuming program order that is bigger
4056 /// than the biggest memory access.
4058 /// tmp = a[i] OR b[i] = x
4059 /// a[i+2] = tmp y = b[i+2];
4061 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4063 /// * Zero distances and all accesses have the same size.
4065 class MemoryDepChecker {
4067 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4068 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4070 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4071 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4072 ShouldRetryWithRuntimeCheck(false) {}
4074 /// \brief Register the location (instructions are given increasing numbers)
4075 /// of a write access.
4076 void addAccess(StoreInst *SI) {
4077 Value *Ptr = SI->getPointerOperand();
4078 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4079 InstMap.push_back(SI);
4083 /// \brief Register the location (instructions are given increasing numbers)
4084 /// of a write access.
4085 void addAccess(LoadInst *LI) {
4086 Value *Ptr = LI->getPointerOperand();
4087 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4088 InstMap.push_back(LI);
4092 /// \brief Check whether the dependencies between the accesses are safe.
4094 /// Only checks sets with elements in \p CheckDeps.
4095 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4096 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4098 /// \brief The maximum number of bytes of a vector register we can vectorize
4099 /// the accesses safely with.
4100 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4102 /// \brief In same cases when the dependency check fails we can still
4103 /// vectorize the loop with a dynamic array access check.
4104 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4107 ScalarEvolution *SE;
4108 const DataLayout *DL;
4109 const Loop *InnermostLoop;
4111 /// \brief Maps access locations (ptr, read/write) to program order.
4112 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4114 /// \brief Memory access instructions in program order.
4115 SmallVector<Instruction *, 16> InstMap;
4117 /// \brief The program order index to be used for the next instruction.
4120 // We can access this many bytes in parallel safely.
4121 unsigned MaxSafeDepDistBytes;
4123 /// \brief If we see a non-constant dependence distance we can still try to
4124 /// vectorize this loop with runtime checks.
4125 bool ShouldRetryWithRuntimeCheck;
4127 /// \brief Check whether there is a plausible dependence between the two
4130 /// Access \p A must happen before \p B in program order. The two indices
4131 /// identify the index into the program order map.
4133 /// This function checks whether there is a plausible dependence (or the
4134 /// absence of such can't be proved) between the two accesses. If there is a
4135 /// plausible dependence but the dependence distance is bigger than one
4136 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4137 /// distance is smaller than any other distance encountered so far).
4138 /// Otherwise, this function returns true signaling a possible dependence.
4139 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4140 const MemAccessInfo &B, unsigned BIdx,
4141 ValueToValueMap &Strides);
4143 /// \brief Check whether the data dependence could prevent store-load
4145 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4148 } // end anonymous namespace
4150 static bool isInBoundsGep(Value *Ptr) {
4151 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4152 return GEP->isInBounds();
4156 /// \brief Check whether the access through \p Ptr has a constant stride.
4157 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4158 const Loop *Lp, ValueToValueMap &StridesMap) {
4159 const Type *Ty = Ptr->getType();
4160 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4162 // Make sure that the pointer does not point to aggregate types.
4163 const PointerType *PtrTy = cast<PointerType>(Ty);
4164 if (PtrTy->getElementType()->isAggregateType()) {
4165 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4170 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4172 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4174 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4175 << *Ptr << " SCEV: " << *PtrScev << "\n");
4179 // The accesss function must stride over the innermost loop.
4180 if (Lp != AR->getLoop()) {
4181 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4182 *Ptr << " SCEV: " << *PtrScev << "\n");
4185 // The address calculation must not wrap. Otherwise, a dependence could be
4187 // An inbounds getelementptr that is a AddRec with a unit stride
4188 // cannot wrap per definition. The unit stride requirement is checked later.
4189 // An getelementptr without an inbounds attribute and unit stride would have
4190 // to access the pointer value "0" which is undefined behavior in address
4191 // space 0, therefore we can also vectorize this case.
4192 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4193 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4194 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4195 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4196 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4197 << *Ptr << " SCEV: " << *PtrScev << "\n");
4201 // Check the step is constant.
4202 const SCEV *Step = AR->getStepRecurrence(*SE);
4204 // Calculate the pointer stride and check if it is consecutive.
4205 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4207 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4208 " SCEV: " << *PtrScev << "\n");
4212 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4213 const APInt &APStepVal = C->getValue()->getValue();
4215 // Huge step value - give up.
4216 if (APStepVal.getBitWidth() > 64)
4219 int64_t StepVal = APStepVal.getSExtValue();
4222 int64_t Stride = StepVal / Size;
4223 int64_t Rem = StepVal % Size;
4227 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4228 // know we can't "wrap around the address space". In case of address space
4229 // zero we know that this won't happen without triggering undefined behavior.
4230 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4231 Stride != 1 && Stride != -1)
4237 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4238 unsigned TypeByteSize) {
4239 // If loads occur at a distance that is not a multiple of a feasible vector
4240 // factor store-load forwarding does not take place.
4241 // Positive dependences might cause troubles because vectorizing them might
4242 // prevent store-load forwarding making vectorized code run a lot slower.
4243 // a[i] = a[i-3] ^ a[i-8];
4244 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4245 // hence on your typical architecture store-load forwarding does not take
4246 // place. Vectorizing in such cases does not make sense.
4247 // Store-load forwarding distance.
4248 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4249 // Maximum vector factor.
4250 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4251 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4252 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4254 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4256 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4257 MaxVFWithoutSLForwardIssues = (vf >>=1);
4262 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4263 DEBUG(dbgs() << "LV: Distance " << Distance <<
4264 " that could cause a store-load forwarding conflict\n");
4268 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4269 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4270 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4274 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4275 const MemAccessInfo &B, unsigned BIdx,
4276 ValueToValueMap &Strides) {
4277 assert (AIdx < BIdx && "Must pass arguments in program order");
4279 Value *APtr = A.getPointer();
4280 Value *BPtr = B.getPointer();
4281 bool AIsWrite = A.getInt();
4282 bool BIsWrite = B.getInt();
4284 // Two reads are independent.
4285 if (!AIsWrite && !BIsWrite)
4288 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4289 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4291 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4292 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4294 const SCEV *Src = AScev;
4295 const SCEV *Sink = BScev;
4297 // If the induction step is negative we have to invert source and sink of the
4299 if (StrideAPtr < 0) {
4302 std::swap(APtr, BPtr);
4303 std::swap(Src, Sink);
4304 std::swap(AIsWrite, BIsWrite);
4305 std::swap(AIdx, BIdx);
4306 std::swap(StrideAPtr, StrideBPtr);
4309 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4311 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4312 << "(Induction step: " << StrideAPtr << ")\n");
4313 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4314 << *InstMap[BIdx] << ": " << *Dist << "\n");
4316 // Need consecutive accesses. We don't want to vectorize
4317 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4318 // the address space.
4319 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4320 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4324 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4326 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4327 ShouldRetryWithRuntimeCheck = true;
4331 Type *ATy = APtr->getType()->getPointerElementType();
4332 Type *BTy = BPtr->getType()->getPointerElementType();
4333 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4335 // Negative distances are not plausible dependencies.
4336 const APInt &Val = C->getValue()->getValue();
4337 if (Val.isNegative()) {
4338 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4339 if (IsTrueDataDependence &&
4340 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4344 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4348 // Write to the same location with the same size.
4349 // Could be improved to assert type sizes are the same (i32 == float, etc).
4353 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4357 assert(Val.isStrictlyPositive() && "Expect a positive value");
4359 // Positive distance bigger than max vectorization factor.
4362 "LV: ReadWrite-Write positive dependency with different types\n");
4366 unsigned Distance = (unsigned) Val.getZExtValue();
4368 // Bail out early if passed-in parameters make vectorization not feasible.
4369 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4370 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4372 // The distance must be bigger than the size needed for a vectorized version
4373 // of the operation and the size of the vectorized operation must not be
4374 // bigger than the currrent maximum size.
4375 if (Distance < 2*TypeByteSize ||
4376 2*TypeByteSize > MaxSafeDepDistBytes ||
4377 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4378 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4379 << Val.getSExtValue() << '\n');
4383 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4384 Distance : MaxSafeDepDistBytes;
4386 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4387 if (IsTrueDataDependence &&
4388 couldPreventStoreLoadForward(Distance, TypeByteSize))
4391 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4392 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4397 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4398 MemAccessInfoSet &CheckDeps,
4399 ValueToValueMap &Strides) {
4401 MaxSafeDepDistBytes = -1U;
4402 while (!CheckDeps.empty()) {
4403 MemAccessInfo CurAccess = *CheckDeps.begin();
4405 // Get the relevant memory access set.
4406 EquivalenceClasses<MemAccessInfo>::iterator I =
4407 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4409 // Check accesses within this set.
4410 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4411 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4413 // Check every access pair.
4415 CheckDeps.erase(*AI);
4416 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4418 // Check every accessing instruction pair in program order.
4419 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4420 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4421 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4422 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4423 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4425 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4436 bool LoopVectorizationLegality::canVectorizeMemory() {
4438 typedef SmallVector<Value*, 16> ValueVector;
4439 typedef SmallPtrSet<Value*, 16> ValueSet;
4441 // Holds the Load and Store *instructions*.
4445 // Holds all the different accesses in the loop.
4446 unsigned NumReads = 0;
4447 unsigned NumReadWrites = 0;
4449 PtrRtCheck.Pointers.clear();
4450 PtrRtCheck.Need = false;
4452 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4453 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4456 for (Loop::block_iterator bb = TheLoop->block_begin(),
4457 be = TheLoop->block_end(); bb != be; ++bb) {
4459 // Scan the BB and collect legal loads and stores.
4460 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4463 // If this is a load, save it. If this instruction can read from memory
4464 // but is not a load, then we quit. Notice that we don't handle function
4465 // calls that read or write.
4466 if (it->mayReadFromMemory()) {
4467 // Many math library functions read the rounding mode. We will only
4468 // vectorize a loop if it contains known function calls that don't set
4469 // the flag. Therefore, it is safe to ignore this read from memory.
4470 CallInst *Call = dyn_cast<CallInst>(it);
4471 if (Call && getIntrinsicIDForCall(Call, TLI))
4474 LoadInst *Ld = dyn_cast<LoadInst>(it);
4475 if (!Ld) return false;
4476 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4477 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4481 Loads.push_back(Ld);
4482 DepChecker.addAccess(Ld);
4486 // Save 'store' instructions. Abort if other instructions write to memory.
4487 if (it->mayWriteToMemory()) {
4488 StoreInst *St = dyn_cast<StoreInst>(it);
4489 if (!St) return false;
4490 if (!St->isSimple() && !IsAnnotatedParallel) {
4491 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4495 Stores.push_back(St);
4496 DepChecker.addAccess(St);
4501 // Now we have two lists that hold the loads and the stores.
4502 // Next, we find the pointers that they use.
4504 // Check if we see any stores. If there are no stores, then we don't
4505 // care if the pointers are *restrict*.
4506 if (!Stores.size()) {
4507 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4511 AccessAnalysis::DepCandidates DependentAccesses;
4512 AccessAnalysis Accesses(DL, DependentAccesses);
4514 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4515 // multiple times on the same object. If the ptr is accessed twice, once
4516 // for read and once for write, it will only appear once (on the write
4517 // list). This is okay, since we are going to check for conflicts between
4518 // writes and between reads and writes, but not between reads and reads.
4521 ValueVector::iterator I, IE;
4522 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4523 StoreInst *ST = cast<StoreInst>(*I);
4524 Value* Ptr = ST->getPointerOperand();
4526 if (isUniform(Ptr)) {
4527 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4531 // If we did *not* see this pointer before, insert it to the read-write
4532 // list. At this phase it is only a 'write' list.
4533 if (Seen.insert(Ptr)) {
4535 Accesses.addStore(Ptr);
4539 if (IsAnnotatedParallel) {
4541 << "LV: A loop annotated parallel, ignore memory dependency "
4546 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4547 LoadInst *LD = cast<LoadInst>(*I);
4548 Value* Ptr = LD->getPointerOperand();
4549 // If we did *not* see this pointer before, insert it to the
4550 // read list. If we *did* see it before, then it is already in
4551 // the read-write list. This allows us to vectorize expressions
4552 // such as A[i] += x; Because the address of A[i] is a read-write
4553 // pointer. This only works if the index of A[i] is consecutive.
4554 // If the address of i is unknown (for example A[B[i]]) then we may
4555 // read a few words, modify, and write a few words, and some of the
4556 // words may be written to the same address.
4557 bool IsReadOnlyPtr = false;
4558 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4560 IsReadOnlyPtr = true;
4562 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4565 // If we write (or read-write) to a single destination and there are no
4566 // other reads in this loop then is it safe to vectorize.
4567 if (NumReadWrites == 1 && NumReads == 0) {
4568 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4572 // Build dependence sets and check whether we need a runtime pointer bounds
4574 Accesses.buildDependenceSets();
4575 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4577 // Find pointers with computable bounds. We are going to use this information
4578 // to place a runtime bound check.
4579 unsigned NumComparisons = 0;
4580 bool CanDoRT = false;
4582 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4585 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4586 " pointer comparisons.\n");
4588 // If we only have one set of dependences to check pointers among we don't
4589 // need a runtime check.
4590 if (NumComparisons == 0 && NeedRTCheck)
4591 NeedRTCheck = false;
4593 // Check that we did not collect too many pointers or found an unsizeable
4595 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4601 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4604 if (NeedRTCheck && !CanDoRT) {
4605 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4606 "the array bounds.\n");
4611 PtrRtCheck.Need = NeedRTCheck;
4613 bool CanVecMem = true;
4614 if (Accesses.isDependencyCheckNeeded()) {
4615 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4616 CanVecMem = DepChecker.areDepsSafe(
4617 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4618 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4620 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4621 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4624 // Clear the dependency checks. We assume they are not needed.
4625 Accesses.resetDepChecks();
4628 PtrRtCheck.Need = true;
4630 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4631 TheLoop, Strides, true);
4632 // Check that we did not collect too many pointers or found an unsizeable
4634 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4635 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4644 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4645 " need a runtime memory check.\n");
4650 static bool hasMultipleUsesOf(Instruction *I,
4651 SmallPtrSet<Instruction *, 8> &Insts) {
4652 unsigned NumUses = 0;
4653 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4654 if (Insts.count(dyn_cast<Instruction>(*Use)))
4663 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4664 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4665 if (!Set.count(dyn_cast<Instruction>(*Use)))
4670 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4671 ReductionKind Kind) {
4672 if (Phi->getNumIncomingValues() != 2)
4675 // Reduction variables are only found in the loop header block.
4676 if (Phi->getParent() != TheLoop->getHeader())
4679 // Obtain the reduction start value from the value that comes from the loop
4681 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4683 // ExitInstruction is the single value which is used outside the loop.
4684 // We only allow for a single reduction value to be used outside the loop.
4685 // This includes users of the reduction, variables (which form a cycle
4686 // which ends in the phi node).
4687 Instruction *ExitInstruction = nullptr;
4688 // Indicates that we found a reduction operation in our scan.
4689 bool FoundReduxOp = false;
4691 // We start with the PHI node and scan for all of the users of this
4692 // instruction. All users must be instructions that can be used as reduction
4693 // variables (such as ADD). We must have a single out-of-block user. The cycle
4694 // must include the original PHI.
4695 bool FoundStartPHI = false;
4697 // To recognize min/max patterns formed by a icmp select sequence, we store
4698 // the number of instruction we saw from the recognized min/max pattern,
4699 // to make sure we only see exactly the two instructions.
4700 unsigned NumCmpSelectPatternInst = 0;
4701 ReductionInstDesc ReduxDesc(false, nullptr);
4703 SmallPtrSet<Instruction *, 8> VisitedInsts;
4704 SmallVector<Instruction *, 8> Worklist;
4705 Worklist.push_back(Phi);
4706 VisitedInsts.insert(Phi);
4708 // A value in the reduction can be used:
4709 // - By the reduction:
4710 // - Reduction operation:
4711 // - One use of reduction value (safe).
4712 // - Multiple use of reduction value (not safe).
4714 // - All uses of the PHI must be the reduction (safe).
4715 // - Otherwise, not safe.
4716 // - By one instruction outside of the loop (safe).
4717 // - By further instructions outside of the loop (not safe).
4718 // - By an instruction that is not part of the reduction (not safe).
4720 // * An instruction type other than PHI or the reduction operation.
4721 // * A PHI in the header other than the initial PHI.
4722 while (!Worklist.empty()) {
4723 Instruction *Cur = Worklist.back();
4724 Worklist.pop_back();
4727 // If the instruction has no users then this is a broken chain and can't be
4728 // a reduction variable.
4729 if (Cur->use_empty())
4732 bool IsAPhi = isa<PHINode>(Cur);
4734 // A header PHI use other than the original PHI.
4735 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4738 // Reductions of instructions such as Div, and Sub is only possible if the
4739 // LHS is the reduction variable.
4740 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4741 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4742 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4745 // Any reduction instruction must be of one of the allowed kinds.
4746 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4747 if (!ReduxDesc.IsReduction)
4750 // A reduction operation must only have one use of the reduction value.
4751 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4752 hasMultipleUsesOf(Cur, VisitedInsts))
4755 // All inputs to a PHI node must be a reduction value.
4756 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4759 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4760 isa<SelectInst>(Cur)))
4761 ++NumCmpSelectPatternInst;
4762 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4763 isa<SelectInst>(Cur)))
4764 ++NumCmpSelectPatternInst;
4766 // Check whether we found a reduction operator.
4767 FoundReduxOp |= !IsAPhi;
4769 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4770 // onto the stack. This way we are going to have seen all inputs to PHI
4771 // nodes once we get to them.
4772 SmallVector<Instruction *, 8> NonPHIs;
4773 SmallVector<Instruction *, 8> PHIs;
4774 for (User *U : Cur->users()) {
4775 Instruction *UI = cast<Instruction>(U);
4777 // Check if we found the exit user.
4778 BasicBlock *Parent = UI->getParent();
4779 if (!TheLoop->contains(Parent)) {
4780 // Exit if you find multiple outside users or if the header phi node is
4781 // being used. In this case the user uses the value of the previous
4782 // iteration, in which case we would loose "VF-1" iterations of the
4783 // reduction operation if we vectorize.
4784 if (ExitInstruction != nullptr || Cur == Phi)
4787 // The instruction used by an outside user must be the last instruction
4788 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4789 // operations on the value.
4790 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4793 ExitInstruction = Cur;
4797 // Process instructions only once (termination). Each reduction cycle
4798 // value must only be used once, except by phi nodes and min/max
4799 // reductions which are represented as a cmp followed by a select.
4800 ReductionInstDesc IgnoredVal(false, nullptr);
4801 if (VisitedInsts.insert(UI)) {
4802 if (isa<PHINode>(UI))
4805 NonPHIs.push_back(UI);
4806 } else if (!isa<PHINode>(UI) &&
4807 ((!isa<FCmpInst>(UI) &&
4808 !isa<ICmpInst>(UI) &&
4809 !isa<SelectInst>(UI)) ||
4810 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4813 // Remember that we completed the cycle.
4815 FoundStartPHI = true;
4817 Worklist.append(PHIs.begin(), PHIs.end());
4818 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4821 // This means we have seen one but not the other instruction of the
4822 // pattern or more than just a select and cmp.
4823 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4824 NumCmpSelectPatternInst != 2)
4827 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4830 // We found a reduction var if we have reached the original phi node and we
4831 // only have a single instruction with out-of-loop users.
4833 // This instruction is allowed to have out-of-loop users.
4834 AllowedExit.insert(ExitInstruction);
4836 // Save the description of this reduction variable.
4837 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4838 ReduxDesc.MinMaxKind);
4839 Reductions[Phi] = RD;
4840 // We've ended the cycle. This is a reduction variable if we have an
4841 // outside user and it has a binary op.
4846 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4847 /// pattern corresponding to a min(X, Y) or max(X, Y).
4848 LoopVectorizationLegality::ReductionInstDesc
4849 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4850 ReductionInstDesc &Prev) {
4852 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4853 "Expect a select instruction");
4854 Instruction *Cmp = nullptr;
4855 SelectInst *Select = nullptr;
4857 // We must handle the select(cmp()) as a single instruction. Advance to the
4859 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4860 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4861 return ReductionInstDesc(false, I);
4862 return ReductionInstDesc(Select, Prev.MinMaxKind);
4865 // Only handle single use cases for now.
4866 if (!(Select = dyn_cast<SelectInst>(I)))
4867 return ReductionInstDesc(false, I);
4868 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4869 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4870 return ReductionInstDesc(false, I);
4871 if (!Cmp->hasOneUse())
4872 return ReductionInstDesc(false, I);
4877 // Look for a min/max pattern.
4878 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4879 return ReductionInstDesc(Select, MRK_UIntMin);
4880 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4881 return ReductionInstDesc(Select, MRK_UIntMax);
4882 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4883 return ReductionInstDesc(Select, MRK_SIntMax);
4884 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4885 return ReductionInstDesc(Select, MRK_SIntMin);
4886 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4887 return ReductionInstDesc(Select, MRK_FloatMin);
4888 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4889 return ReductionInstDesc(Select, MRK_FloatMax);
4890 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4891 return ReductionInstDesc(Select, MRK_FloatMin);
4892 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4893 return ReductionInstDesc(Select, MRK_FloatMax);
4895 return ReductionInstDesc(false, I);
4898 LoopVectorizationLegality::ReductionInstDesc
4899 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4901 ReductionInstDesc &Prev) {
4902 bool FP = I->getType()->isFloatingPointTy();
4903 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4904 switch (I->getOpcode()) {
4906 return ReductionInstDesc(false, I);
4907 case Instruction::PHI:
4908 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4909 Kind != RK_FloatMinMax))
4910 return ReductionInstDesc(false, I);
4911 return ReductionInstDesc(I, Prev.MinMaxKind);
4912 case Instruction::Sub:
4913 case Instruction::Add:
4914 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4915 case Instruction::Mul:
4916 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4917 case Instruction::And:
4918 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4919 case Instruction::Or:
4920 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4921 case Instruction::Xor:
4922 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4923 case Instruction::FMul:
4924 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4925 case Instruction::FAdd:
4926 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4927 case Instruction::FCmp:
4928 case Instruction::ICmp:
4929 case Instruction::Select:
4930 if (Kind != RK_IntegerMinMax &&
4931 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4932 return ReductionInstDesc(false, I);
4933 return isMinMaxSelectCmpPattern(I, Prev);
4937 LoopVectorizationLegality::InductionKind
4938 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4939 Type *PhiTy = Phi->getType();
4940 // We only handle integer and pointer inductions variables.
4941 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4942 return IK_NoInduction;
4944 // Check that the PHI is consecutive.
4945 const SCEV *PhiScev = SE->getSCEV(Phi);
4946 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4948 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4949 return IK_NoInduction;
4951 const SCEV *Step = AR->getStepRecurrence(*SE);
4953 // Integer inductions need to have a stride of one.
4954 if (PhiTy->isIntegerTy()) {
4956 return IK_IntInduction;
4957 if (Step->isAllOnesValue())
4958 return IK_ReverseIntInduction;
4959 return IK_NoInduction;
4962 // Calculate the pointer stride and check if it is consecutive.
4963 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4965 return IK_NoInduction;
4967 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4968 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4969 if (C->getValue()->equalsInt(Size))
4970 return IK_PtrInduction;
4971 else if (C->getValue()->equalsInt(0 - Size))
4972 return IK_ReversePtrInduction;
4974 return IK_NoInduction;
4977 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4978 Value *In0 = const_cast<Value*>(V);
4979 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4983 return Inductions.count(PN);
4986 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4987 assert(TheLoop->contains(BB) && "Unknown block used");
4989 // Blocks that do not dominate the latch need predication.
4990 BasicBlock* Latch = TheLoop->getLoopLatch();
4991 return !DT->dominates(BB, Latch);
4994 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4995 SmallPtrSet<Value *, 8>& SafePtrs) {
4996 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4997 // We might be able to hoist the load.
4998 if (it->mayReadFromMemory()) {
4999 LoadInst *LI = dyn_cast<LoadInst>(it);
5000 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5004 // We don't predicate stores at the moment.
5005 if (it->mayWriteToMemory()) {
5006 StoreInst *SI = dyn_cast<StoreInst>(it);
5007 // We only support predication of stores in basic blocks with one
5009 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5010 !SafePtrs.count(SI->getPointerOperand()) ||
5011 !SI->getParent()->getSinglePredecessor())
5017 // Check that we don't have a constant expression that can trap as operand.
5018 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5020 if (Constant *C = dyn_cast<Constant>(*OI))
5025 // The instructions below can trap.
5026 switch (it->getOpcode()) {
5028 case Instruction::UDiv:
5029 case Instruction::SDiv:
5030 case Instruction::URem:
5031 case Instruction::SRem:
5039 LoopVectorizationCostModel::VectorizationFactor
5040 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5042 bool ForceVectorization) {
5043 // Width 1 means no vectorize
5044 VectorizationFactor Factor = { 1U, 0U };
5045 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5046 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5050 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5051 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5055 // Find the trip count.
5056 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5057 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5059 unsigned WidestType = getWidestType();
5060 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5061 unsigned MaxSafeDepDist = -1U;
5062 if (Legal->getMaxSafeDepDistBytes() != -1U)
5063 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5064 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5065 WidestRegister : MaxSafeDepDist);
5066 unsigned MaxVectorSize = WidestRegister / WidestType;
5067 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5068 DEBUG(dbgs() << "LV: The Widest register is: "
5069 << WidestRegister << " bits.\n");
5071 if (MaxVectorSize == 0) {
5072 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5076 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5077 " into one vector!");
5079 unsigned VF = MaxVectorSize;
5081 // If we optimize the program for size, avoid creating the tail loop.
5083 // If we are unable to calculate the trip count then don't try to vectorize.
5085 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5089 // Find the maximum SIMD width that can fit within the trip count.
5090 VF = TC % MaxVectorSize;
5095 // If the trip count that we found modulo the vectorization factor is not
5096 // zero then we require a tail.
5098 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5104 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5105 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5107 Factor.Width = UserVF;
5111 float Cost = expectedCost(1);
5113 const float ScalarCost = Cost;
5116 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5118 // Ignore scalar width, because the user explicitly wants vectorization.
5119 if (ForceVectorization && VF > 1) {
5121 Cost = expectedCost(Width) / (float)Width;
5124 for (unsigned i=2; i <= VF; i*=2) {
5125 // Notice that the vector loop needs to be executed less times, so
5126 // we need to divide the cost of the vector loops by the width of
5127 // the vector elements.
5128 float VectorCost = expectedCost(i) / (float)i;
5129 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5130 (int)VectorCost << ".\n");
5131 if (VectorCost < Cost) {
5137 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5138 << "LV: Vectorization seems to be not beneficial, "
5139 << "but was forced by a user.\n");
5140 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5141 Factor.Width = Width;
5142 Factor.Cost = Width * Cost;
5146 unsigned LoopVectorizationCostModel::getWidestType() {
5147 unsigned MaxWidth = 8;
5150 for (Loop::block_iterator bb = TheLoop->block_begin(),
5151 be = TheLoop->block_end(); bb != be; ++bb) {
5152 BasicBlock *BB = *bb;
5154 // For each instruction in the loop.
5155 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5156 Type *T = it->getType();
5158 // Only examine Loads, Stores and PHINodes.
5159 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5162 // Examine PHI nodes that are reduction variables.
5163 if (PHINode *PN = dyn_cast<PHINode>(it))
5164 if (!Legal->getReductionVars()->count(PN))
5167 // Examine the stored values.
5168 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5169 T = ST->getValueOperand()->getType();
5171 // Ignore loaded pointer types and stored pointer types that are not
5172 // consecutive. However, we do want to take consecutive stores/loads of
5173 // pointer vectors into account.
5174 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5177 MaxWidth = std::max(MaxWidth,
5178 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5186 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5189 unsigned LoopCost) {
5191 // -- The unroll heuristics --
5192 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5193 // There are many micro-architectural considerations that we can't predict
5194 // at this level. For example frontend pressure (on decode or fetch) due to
5195 // code size, or the number and capabilities of the execution ports.
5197 // We use the following heuristics to select the unroll factor:
5198 // 1. If the code has reductions the we unroll in order to break the cross
5199 // iteration dependency.
5200 // 2. If the loop is really small then we unroll in order to reduce the loop
5202 // 3. We don't unroll if we think that we will spill registers to memory due
5203 // to the increased register pressure.
5205 // Use the user preference, unless 'auto' is selected.
5209 // When we optimize for size we don't unroll.
5213 // We used the distance for the unroll factor.
5214 if (Legal->getMaxSafeDepDistBytes() != -1U)
5217 // Do not unroll loops with a relatively small trip count.
5218 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5219 TheLoop->getLoopLatch());
5220 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5223 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5224 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5228 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5229 TargetNumRegisters = ForceTargetNumScalarRegs;
5231 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5232 TargetNumRegisters = ForceTargetNumVectorRegs;
5235 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5236 // We divide by these constants so assume that we have at least one
5237 // instruction that uses at least one register.
5238 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5239 R.NumInstructions = std::max(R.NumInstructions, 1U);
5241 // We calculate the unroll factor using the following formula.
5242 // Subtract the number of loop invariants from the number of available
5243 // registers. These registers are used by all of the unrolled instances.
5244 // Next, divide the remaining registers by the number of registers that is
5245 // required by the loop, in order to estimate how many parallel instances
5246 // fit without causing spills. All of this is rounded down if necessary to be
5247 // a power of two. We want power of two unroll factors to simplify any
5248 // addressing operations or alignment considerations.
5249 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5252 // Don't count the induction variable as unrolled.
5253 if (EnableIndVarRegisterHeur)
5254 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5255 std::max(1U, (R.MaxLocalUsers - 1)));
5257 // Clamp the unroll factor ranges to reasonable factors.
5258 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5260 // Check if the user has overridden the unroll max.
5262 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5263 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5265 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5266 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5269 // If we did not calculate the cost for VF (because the user selected the VF)
5270 // then we calculate the cost of VF here.
5272 LoopCost = expectedCost(VF);
5274 // Clamp the calculated UF to be between the 1 and the max unroll factor
5275 // that the target allows.
5276 if (UF > MaxUnrollSize)
5281 // Unroll if we vectorized this loop and there is a reduction that could
5282 // benefit from unrolling.
5283 if (VF > 1 && Legal->getReductionVars()->size()) {
5284 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5288 // Note that if we've already vectorized the loop we will have done the
5289 // runtime check and so unrolling won't require further checks.
5290 bool UnrollingRequiresRuntimePointerCheck =
5291 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5293 // We want to unroll small loops in order to reduce the loop overhead and
5294 // potentially expose ILP opportunities.
5295 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5296 if (!UnrollingRequiresRuntimePointerCheck &&
5297 LoopCost < SmallLoopCost) {
5298 // We assume that the cost overhead is 1 and we use the cost model
5299 // to estimate the cost of the loop and unroll until the cost of the
5300 // loop overhead is about 5% of the cost of the loop.
5301 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5303 // Unroll until store/load ports (estimated by max unroll factor) are
5305 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5306 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5308 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5309 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5310 return std::max(StoresUF, LoadsUF);
5313 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5317 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5321 LoopVectorizationCostModel::RegisterUsage
5322 LoopVectorizationCostModel::calculateRegisterUsage() {
5323 // This function calculates the register usage by measuring the highest number
5324 // of values that are alive at a single location. Obviously, this is a very
5325 // rough estimation. We scan the loop in a topological order in order and
5326 // assign a number to each instruction. We use RPO to ensure that defs are
5327 // met before their users. We assume that each instruction that has in-loop
5328 // users starts an interval. We record every time that an in-loop value is
5329 // used, so we have a list of the first and last occurrences of each
5330 // instruction. Next, we transpose this data structure into a multi map that
5331 // holds the list of intervals that *end* at a specific location. This multi
5332 // map allows us to perform a linear search. We scan the instructions linearly
5333 // and record each time that a new interval starts, by placing it in a set.
5334 // If we find this value in the multi-map then we remove it from the set.
5335 // The max register usage is the maximum size of the set.
5336 // We also search for instructions that are defined outside the loop, but are
5337 // used inside the loop. We need this number separately from the max-interval
5338 // usage number because when we unroll, loop-invariant values do not take
5340 LoopBlocksDFS DFS(TheLoop);
5344 R.NumInstructions = 0;
5346 // Each 'key' in the map opens a new interval. The values
5347 // of the map are the index of the 'last seen' usage of the
5348 // instruction that is the key.
5349 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5350 // Maps instruction to its index.
5351 DenseMap<unsigned, Instruction*> IdxToInstr;
5352 // Marks the end of each interval.
5353 IntervalMap EndPoint;
5354 // Saves the list of instruction indices that are used in the loop.
5355 SmallSet<Instruction*, 8> Ends;
5356 // Saves the list of values that are used in the loop but are
5357 // defined outside the loop, such as arguments and constants.
5358 SmallPtrSet<Value*, 8> LoopInvariants;
5361 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5362 be = DFS.endRPO(); bb != be; ++bb) {
5363 R.NumInstructions += (*bb)->size();
5364 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5366 Instruction *I = it;
5367 IdxToInstr[Index++] = I;
5369 // Save the end location of each USE.
5370 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5371 Value *U = I->getOperand(i);
5372 Instruction *Instr = dyn_cast<Instruction>(U);
5374 // Ignore non-instruction values such as arguments, constants, etc.
5375 if (!Instr) continue;
5377 // If this instruction is outside the loop then record it and continue.
5378 if (!TheLoop->contains(Instr)) {
5379 LoopInvariants.insert(Instr);
5383 // Overwrite previous end points.
5384 EndPoint[Instr] = Index;
5390 // Saves the list of intervals that end with the index in 'key'.
5391 typedef SmallVector<Instruction*, 2> InstrList;
5392 DenseMap<unsigned, InstrList> TransposeEnds;
5394 // Transpose the EndPoints to a list of values that end at each index.
5395 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5397 TransposeEnds[it->second].push_back(it->first);
5399 SmallSet<Instruction*, 8> OpenIntervals;
5400 unsigned MaxUsage = 0;
5403 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5404 for (unsigned int i = 0; i < Index; ++i) {
5405 Instruction *I = IdxToInstr[i];
5406 // Ignore instructions that are never used within the loop.
5407 if (!Ends.count(I)) continue;
5409 // Remove all of the instructions that end at this location.
5410 InstrList &List = TransposeEnds[i];
5411 for (unsigned int j=0, e = List.size(); j < e; ++j)
5412 OpenIntervals.erase(List[j]);
5414 // Count the number of live interals.
5415 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5417 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5418 OpenIntervals.size() << '\n');
5420 // Add the current instruction to the list of open intervals.
5421 OpenIntervals.insert(I);
5424 unsigned Invariant = LoopInvariants.size();
5425 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5426 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5427 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5429 R.LoopInvariantRegs = Invariant;
5430 R.MaxLocalUsers = MaxUsage;
5434 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5438 for (Loop::block_iterator bb = TheLoop->block_begin(),
5439 be = TheLoop->block_end(); bb != be; ++bb) {
5440 unsigned BlockCost = 0;
5441 BasicBlock *BB = *bb;
5443 // For each instruction in the old loop.
5444 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5445 // Skip dbg intrinsics.
5446 if (isa<DbgInfoIntrinsic>(it))
5449 unsigned C = getInstructionCost(it, VF);
5451 // Check if we should override the cost.
5452 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5453 C = ForceTargetInstructionCost;
5456 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5457 VF << " For instruction: " << *it << '\n');
5460 // We assume that if-converted blocks have a 50% chance of being executed.
5461 // When the code is scalar then some of the blocks are avoided due to CF.
5462 // When the code is vectorized we execute all code paths.
5463 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5472 /// \brief Check whether the address computation for a non-consecutive memory
5473 /// access looks like an unlikely candidate for being merged into the indexing
5476 /// We look for a GEP which has one index that is an induction variable and all
5477 /// other indices are loop invariant. If the stride of this access is also
5478 /// within a small bound we decide that this address computation can likely be
5479 /// merged into the addressing mode.
5480 /// In all other cases, we identify the address computation as complex.
5481 static bool isLikelyComplexAddressComputation(Value *Ptr,
5482 LoopVectorizationLegality *Legal,
5483 ScalarEvolution *SE,
5484 const Loop *TheLoop) {
5485 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5489 // We are looking for a gep with all loop invariant indices except for one
5490 // which should be an induction variable.
5491 unsigned NumOperands = Gep->getNumOperands();
5492 for (unsigned i = 1; i < NumOperands; ++i) {
5493 Value *Opd = Gep->getOperand(i);
5494 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5495 !Legal->isInductionVariable(Opd))
5499 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5500 // can likely be merged into the address computation.
5501 unsigned MaxMergeDistance = 64;
5503 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5507 // Check the step is constant.
5508 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5509 // Calculate the pointer stride and check if it is consecutive.
5510 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5514 const APInt &APStepVal = C->getValue()->getValue();
5516 // Huge step value - give up.
5517 if (APStepVal.getBitWidth() > 64)
5520 int64_t StepVal = APStepVal.getSExtValue();
5522 return StepVal > MaxMergeDistance;
5525 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5526 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5532 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5533 // If we know that this instruction will remain uniform, check the cost of
5534 // the scalar version.
5535 if (Legal->isUniformAfterVectorization(I))
5538 Type *RetTy = I->getType();
5539 Type *VectorTy = ToVectorTy(RetTy, VF);
5541 // TODO: We need to estimate the cost of intrinsic calls.
5542 switch (I->getOpcode()) {
5543 case Instruction::GetElementPtr:
5544 // We mark this instruction as zero-cost because the cost of GEPs in
5545 // vectorized code depends on whether the corresponding memory instruction
5546 // is scalarized or not. Therefore, we handle GEPs with the memory
5547 // instruction cost.
5549 case Instruction::Br: {
5550 return TTI.getCFInstrCost(I->getOpcode());
5552 case Instruction::PHI:
5553 //TODO: IF-converted IFs become selects.
5555 case Instruction::Add:
5556 case Instruction::FAdd:
5557 case Instruction::Sub:
5558 case Instruction::FSub:
5559 case Instruction::Mul:
5560 case Instruction::FMul:
5561 case Instruction::UDiv:
5562 case Instruction::SDiv:
5563 case Instruction::FDiv:
5564 case Instruction::URem:
5565 case Instruction::SRem:
5566 case Instruction::FRem:
5567 case Instruction::Shl:
5568 case Instruction::LShr:
5569 case Instruction::AShr:
5570 case Instruction::And:
5571 case Instruction::Or:
5572 case Instruction::Xor: {
5573 // Since we will replace the stride by 1 the multiplication should go away.
5574 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5576 // Certain instructions can be cheaper to vectorize if they have a constant
5577 // second vector operand. One example of this are shifts on x86.
5578 TargetTransformInfo::OperandValueKind Op1VK =
5579 TargetTransformInfo::OK_AnyValue;
5580 TargetTransformInfo::OperandValueKind Op2VK =
5581 TargetTransformInfo::OK_AnyValue;
5582 Value *Op2 = I->getOperand(1);
5584 // Check for a splat of a constant or for a non uniform vector of constants.
5585 if (isa<ConstantInt>(Op2))
5586 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5587 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5588 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5589 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5590 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5593 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5595 case Instruction::Select: {
5596 SelectInst *SI = cast<SelectInst>(I);
5597 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5598 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5599 Type *CondTy = SI->getCondition()->getType();
5601 CondTy = VectorType::get(CondTy, VF);
5603 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5605 case Instruction::ICmp:
5606 case Instruction::FCmp: {
5607 Type *ValTy = I->getOperand(0)->getType();
5608 VectorTy = ToVectorTy(ValTy, VF);
5609 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5611 case Instruction::Store:
5612 case Instruction::Load: {
5613 StoreInst *SI = dyn_cast<StoreInst>(I);
5614 LoadInst *LI = dyn_cast<LoadInst>(I);
5615 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5617 VectorTy = ToVectorTy(ValTy, VF);
5619 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5620 unsigned AS = SI ? SI->getPointerAddressSpace() :
5621 LI->getPointerAddressSpace();
5622 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5623 // We add the cost of address computation here instead of with the gep
5624 // instruction because only here we know whether the operation is
5627 return TTI.getAddressComputationCost(VectorTy) +
5628 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5630 // Scalarized loads/stores.
5631 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5632 bool Reverse = ConsecutiveStride < 0;
5633 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5634 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5635 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5636 bool IsComplexComputation =
5637 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5639 // The cost of extracting from the value vector and pointer vector.
5640 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5641 for (unsigned i = 0; i < VF; ++i) {
5642 // The cost of extracting the pointer operand.
5643 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5644 // In case of STORE, the cost of ExtractElement from the vector.
5645 // In case of LOAD, the cost of InsertElement into the returned
5647 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5648 Instruction::InsertElement,
5652 // The cost of the scalar loads/stores.
5653 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5654 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5659 // Wide load/stores.
5660 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5661 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5664 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5668 case Instruction::ZExt:
5669 case Instruction::SExt:
5670 case Instruction::FPToUI:
5671 case Instruction::FPToSI:
5672 case Instruction::FPExt:
5673 case Instruction::PtrToInt:
5674 case Instruction::IntToPtr:
5675 case Instruction::SIToFP:
5676 case Instruction::UIToFP:
5677 case Instruction::Trunc:
5678 case Instruction::FPTrunc:
5679 case Instruction::BitCast: {
5680 // We optimize the truncation of induction variable.
5681 // The cost of these is the same as the scalar operation.
5682 if (I->getOpcode() == Instruction::Trunc &&
5683 Legal->isInductionVariable(I->getOperand(0)))
5684 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5685 I->getOperand(0)->getType());
5687 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5688 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5690 case Instruction::Call: {
5691 CallInst *CI = cast<CallInst>(I);
5692 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5693 assert(ID && "Not an intrinsic call!");
5694 Type *RetTy = ToVectorTy(CI->getType(), VF);
5695 SmallVector<Type*, 4> Tys;
5696 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5697 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5698 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5701 // We are scalarizing the instruction. Return the cost of the scalar
5702 // instruction, plus the cost of insert and extract into vector
5703 // elements, times the vector width.
5706 if (!RetTy->isVoidTy() && VF != 1) {
5707 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5709 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5712 // The cost of inserting the results plus extracting each one of the
5714 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5717 // The cost of executing VF copies of the scalar instruction. This opcode
5718 // is unknown. Assume that it is the same as 'mul'.
5719 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5725 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5726 if (Scalar->isVoidTy() || VF == 1)
5728 return VectorType::get(Scalar, VF);
5731 char LoopVectorize::ID = 0;
5732 static const char lv_name[] = "Loop Vectorization";
5733 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5734 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5735 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5736 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5737 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5738 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5739 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5740 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5741 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5744 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5745 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5749 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5750 // Check for a store.
5751 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5752 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5754 // Check for a load.
5755 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5756 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5762 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5763 bool IfPredicateStore) {
5764 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5765 // Holds vector parameters or scalars, in case of uniform vals.
5766 SmallVector<VectorParts, 4> Params;
5768 setDebugLocFromInst(Builder, Instr);
5770 // Find all of the vectorized parameters.
5771 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5772 Value *SrcOp = Instr->getOperand(op);
5774 // If we are accessing the old induction variable, use the new one.
5775 if (SrcOp == OldInduction) {
5776 Params.push_back(getVectorValue(SrcOp));
5780 // Try using previously calculated values.
5781 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5783 // If the src is an instruction that appeared earlier in the basic block
5784 // then it should already be vectorized.
5785 if (SrcInst && OrigLoop->contains(SrcInst)) {
5786 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5787 // The parameter is a vector value from earlier.
5788 Params.push_back(WidenMap.get(SrcInst));
5790 // The parameter is a scalar from outside the loop. Maybe even a constant.
5791 VectorParts Scalars;
5792 Scalars.append(UF, SrcOp);
5793 Params.push_back(Scalars);
5797 assert(Params.size() == Instr->getNumOperands() &&
5798 "Invalid number of operands");
5800 // Does this instruction return a value ?
5801 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5803 Value *UndefVec = IsVoidRetTy ? nullptr :
5804 UndefValue::get(Instr->getType());
5805 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5806 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5808 Instruction *InsertPt = Builder.GetInsertPoint();
5809 BasicBlock *IfBlock = Builder.GetInsertBlock();
5810 BasicBlock *CondBlock = nullptr;
5813 Loop *VectorLp = nullptr;
5814 if (IfPredicateStore) {
5815 assert(Instr->getParent()->getSinglePredecessor() &&
5816 "Only support single predecessor blocks");
5817 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5818 Instr->getParent());
5819 VectorLp = LI->getLoopFor(IfBlock);
5820 assert(VectorLp && "Must have a loop for this block");
5823 // For each vector unroll 'part':
5824 for (unsigned Part = 0; Part < UF; ++Part) {
5825 // For each scalar that we create:
5827 // Start an "if (pred) a[i] = ..." block.
5828 Value *Cmp = nullptr;
5829 if (IfPredicateStore) {
5830 if (Cond[Part]->getType()->isVectorTy())
5832 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5833 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5834 ConstantInt::get(Cond[Part]->getType(), 1));
5835 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5836 LoopVectorBody.push_back(CondBlock);
5837 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5838 // Update Builder with newly created basic block.
5839 Builder.SetInsertPoint(InsertPt);
5842 Instruction *Cloned = Instr->clone();
5844 Cloned->setName(Instr->getName() + ".cloned");
5845 // Replace the operands of the cloned instructions with extracted scalars.
5846 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5847 Value *Op = Params[op][Part];
5848 Cloned->setOperand(op, Op);
5851 // Place the cloned scalar in the new loop.
5852 Builder.Insert(Cloned);
5854 // If the original scalar returns a value we need to place it in a vector
5855 // so that future users will be able to use it.
5857 VecResults[Part] = Cloned;
5860 if (IfPredicateStore) {
5861 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5862 LoopVectorBody.push_back(NewIfBlock);
5863 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5864 Builder.SetInsertPoint(InsertPt);
5865 Instruction *OldBr = IfBlock->getTerminator();
5866 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5867 OldBr->eraseFromParent();
5868 IfBlock = NewIfBlock;
5873 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5874 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5875 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5877 return scalarizeInstruction(Instr, IfPredicateStore);
5880 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5884 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5888 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5890 // When unrolling and the VF is 1, we only need to add a simple scalar.
5891 Type *ITy = Val->getType();
5892 assert(!ITy->isVectorTy() && "Val must be a scalar");
5893 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5894 return Builder.CreateAdd(Val, C, "induction");