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/Transforms/Scalar.h"
89 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
90 #include "llvm/Transforms/Utils/Local.h"
91 #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 class LoopVectorizeHints {
893 FK_Undefined = -1, ///< Not selected.
894 FK_Disabled = 0, ///< Forcing disabled.
895 FK_Enabled = 1, ///< Forcing enabled.
898 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
899 : Width(VectorizationFactor),
900 Unroll(DisableUnrolling),
902 LoopID(L->getLoopID()) {
904 // force-vector-unroll overrides DisableUnrolling.
905 if (VectorizationUnroll.getNumOccurrences() > 0)
906 Unroll = VectorizationUnroll;
908 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
909 << "LV: Unrolling disabled by the pass manager\n");
912 /// Return the loop vectorizer metadata prefix.
913 static StringRef Prefix() { return "llvm.vectorizer."; }
915 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
916 SmallVector<Value*, 2> Vals;
917 Vals.push_back(MDString::get(Context, Name));
918 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
919 return MDNode::get(Context, Vals);
922 /// Mark the loop L as already vectorized by setting the width to 1.
923 void setAlreadyVectorized(Loop *L) {
924 LLVMContext &Context = L->getHeader()->getContext();
928 // Create a new loop id with one more operand for the already_vectorized
929 // hint. If the loop already has a loop id then copy the existing operands.
930 SmallVector<Value*, 4> Vals(1);
932 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
933 Vals.push_back(LoopID->getOperand(i));
935 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
936 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
938 MDNode *NewLoopID = MDNode::get(Context, Vals);
939 // Set operand 0 to refer to the loop id itself.
940 NewLoopID->replaceOperandWith(0, NewLoopID);
942 L->setLoopID(NewLoopID);
944 LoopID->replaceAllUsesWith(NewLoopID);
949 unsigned getWidth() const { return Width; }
950 unsigned getUnroll() const { return Unroll; }
951 enum ForceKind getForce() const { return Force; }
952 MDNode *getLoopID() const { return LoopID; }
955 /// Find hints specified in the loop metadata.
956 void getHints(const Loop *L) {
960 // First operand should refer to the loop id itself.
961 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
962 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
964 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
965 const MDString *S = nullptr;
966 SmallVector<Value*, 4> Args;
968 // The expected hint is either a MDString or a MDNode with the first
969 // operand a MDString.
970 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
971 if (!MD || MD->getNumOperands() == 0)
973 S = dyn_cast<MDString>(MD->getOperand(0));
974 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
975 Args.push_back(MD->getOperand(i));
977 S = dyn_cast<MDString>(LoopID->getOperand(i));
978 assert(Args.size() == 0 && "too many arguments for MDString");
984 // Check if the hint starts with the vectorizer prefix.
985 StringRef Hint = S->getString();
986 if (!Hint.startswith(Prefix()))
988 // Remove the prefix.
989 Hint = Hint.substr(Prefix().size(), StringRef::npos);
991 if (Args.size() == 1)
992 getHint(Hint, Args[0]);
996 // Check string hint with one operand.
997 void getHint(StringRef Hint, Value *Arg) {
998 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1000 unsigned Val = C->getZExtValue();
1002 if (Hint == "width") {
1003 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1006 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1007 } else if (Hint == "unroll") {
1008 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1011 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1012 } else if (Hint == "enable") {
1013 if (C->getBitWidth() == 1)
1014 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1015 : LoopVectorizeHints::FK_Disabled;
1017 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1019 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1023 /// Vectorization width.
1025 /// Vectorization unroll factor.
1027 /// Vectorization forced
1028 enum ForceKind Force;
1033 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1035 return V.push_back(&L);
1037 for (Loop *InnerL : L)
1038 addInnerLoop(*InnerL, V);
1041 /// The LoopVectorize Pass.
1042 struct LoopVectorize : public FunctionPass {
1043 /// Pass identification, replacement for typeid
1046 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1048 DisableUnrolling(NoUnrolling),
1049 AlwaysVectorize(AlwaysVectorize) {
1050 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1053 ScalarEvolution *SE;
1054 const DataLayout *DL;
1056 TargetTransformInfo *TTI;
1058 BlockFrequencyInfo *BFI;
1059 TargetLibraryInfo *TLI;
1060 bool DisableUnrolling;
1061 bool AlwaysVectorize;
1063 BlockFrequency ColdEntryFreq;
1065 bool runOnFunction(Function &F) override {
1066 SE = &getAnalysis<ScalarEvolution>();
1067 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1068 DL = DLP ? &DLP->getDataLayout() : nullptr;
1069 LI = &getAnalysis<LoopInfo>();
1070 TTI = &getAnalysis<TargetTransformInfo>();
1071 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1072 BFI = &getAnalysis<BlockFrequencyInfo>();
1073 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1075 // Compute some weights outside of the loop over the loops. Compute this
1076 // using a BranchProbability to re-use its scaling math.
1077 const BranchProbability ColdProb(1, 5); // 20%
1078 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1080 // If the target claims to have no vector registers don't attempt
1082 if (!TTI->getNumberOfRegisters(true))
1086 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1087 << ": Missing data layout\n");
1091 // Build up a worklist of inner-loops to vectorize. This is necessary as
1092 // the act of vectorizing or partially unrolling a loop creates new loops
1093 // and can invalidate iterators across the loops.
1094 SmallVector<Loop *, 8> Worklist;
1097 addInnerLoop(*L, Worklist);
1099 LoopsAnalyzed += Worklist.size();
1101 // Now walk the identified inner loops.
1102 bool Changed = false;
1103 while (!Worklist.empty())
1104 Changed |= processLoop(Worklist.pop_back_val());
1106 // Process each loop nest in the function.
1110 bool processLoop(Loop *L) {
1111 assert(L->empty() && "Only process inner loops.");
1112 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1113 << L->getHeader()->getParent()->getName() << "\" from "
1114 << getDebugLocString(L->getHeader()->getFirstNonPHIOrDbg())
1117 LoopVectorizeHints Hints(L, DisableUnrolling);
1119 DEBUG(dbgs() << "LV: Loop hints:"
1121 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1123 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1125 : "?")) << " width=" << Hints.getWidth()
1126 << " unroll=" << Hints.getUnroll() << "\n");
1128 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1129 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1133 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1134 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1138 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1139 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1143 // Check the loop for a trip count threshold:
1144 // do not vectorize loops with a tiny trip count.
1145 BasicBlock *Latch = L->getLoopLatch();
1146 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1147 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1148 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1149 << "This loop is not worth vectorizing.");
1150 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1151 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1153 DEBUG(dbgs() << "\n");
1158 // Check if it is legal to vectorize the loop.
1159 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
1160 if (!LVL.canVectorize()) {
1161 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1165 // Use the cost model.
1166 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1168 // Check the function attributes to find out if this function should be
1169 // optimized for size.
1170 Function *F = L->getHeader()->getParent();
1171 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1172 F->hasFnAttribute(Attribute::OptimizeForSize);
1174 // Compute the weighted frequency of this loop being executed and see if it
1175 // is less than 20% of the function entry baseline frequency. Note that we
1176 // always have a canonical loop here because we think we *can* vectoriez.
1177 // FIXME: This is hidden behind a flag due to pervasive problems with
1178 // exactly what block frequency models.
1179 if (LoopVectorizeWithBlockFrequency) {
1180 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1181 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1182 LoopEntryFreq < ColdEntryFreq)
1186 // Check the function attributes to see if implicit floats are allowed.a
1187 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1188 // an integer loop and the vector instructions selected are purely integer
1189 // vector instructions?
1190 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1191 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1192 "attribute is used.\n");
1196 // Select the optimal vectorization factor.
1197 const LoopVectorizationCostModel::VectorizationFactor VF =
1198 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1200 LoopVectorizeHints::FK_Enabled);
1202 // Select the unroll factor.
1204 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1206 DEBUG(dbgs() << "LV: Found a vectorizable loop ("
1207 << VF.Width << ") in "
1208 << getDebugLocString(L->getHeader()->getFirstNonPHIOrDbg())
1210 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1212 if (VF.Width == 1) {
1213 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1216 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1218 // Report the unrolling decision.
1219 F->getContext().emitOptimizationRemark(
1220 DEBUG_TYPE, *F, L->getStartLoc(),
1221 Twine("unrolled with interleaving factor " + Twine(UF) +
1222 " (vectorization not beneficial)"));
1224 // We decided not to vectorize, but we may want to unroll.
1225 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1226 Unroller.vectorize(&LVL);
1228 // If we decided that it is *legal* to vectorize the loop then do it.
1229 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1233 // Report the vectorization decision.
1234 F->getContext().emitOptimizationRemark(
1235 DEBUG_TYPE, *F, L->getStartLoc(),
1236 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1237 ", unrolling interleave factor: " + Twine(UF) + ")");
1240 // Mark the loop as already vectorized to avoid vectorizing again.
1241 Hints.setAlreadyVectorized(L);
1243 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1247 void getAnalysisUsage(AnalysisUsage &AU) const override {
1248 AU.addRequiredID(LoopSimplifyID);
1249 AU.addRequiredID(LCSSAID);
1250 AU.addRequired<BlockFrequencyInfo>();
1251 AU.addRequired<DominatorTreeWrapperPass>();
1252 AU.addRequired<LoopInfo>();
1253 AU.addRequired<ScalarEvolution>();
1254 AU.addRequired<TargetTransformInfo>();
1255 AU.addPreserved<LoopInfo>();
1256 AU.addPreserved<DominatorTreeWrapperPass>();
1261 } // end anonymous namespace
1263 //===----------------------------------------------------------------------===//
1264 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1265 // LoopVectorizationCostModel.
1266 //===----------------------------------------------------------------------===//
1268 static Value *stripIntegerCast(Value *V) {
1269 if (CastInst *CI = dyn_cast<CastInst>(V))
1270 if (CI->getOperand(0)->getType()->isIntegerTy())
1271 return CI->getOperand(0);
1275 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1277 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1279 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1280 ValueToValueMap &PtrToStride,
1281 Value *Ptr, Value *OrigPtr = nullptr) {
1283 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1285 // If there is an entry in the map return the SCEV of the pointer with the
1286 // symbolic stride replaced by one.
1287 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1288 if (SI != PtrToStride.end()) {
1289 Value *StrideVal = SI->second;
1292 StrideVal = stripIntegerCast(StrideVal);
1294 // Replace symbolic stride by one.
1295 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1296 ValueToValueMap RewriteMap;
1297 RewriteMap[StrideVal] = One;
1300 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1301 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1306 // Otherwise, just return the SCEV of the original pointer.
1307 return SE->getSCEV(Ptr);
1310 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1311 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1312 ValueToValueMap &Strides) {
1313 // Get the stride replaced scev.
1314 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1315 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1316 assert(AR && "Invalid addrec expression");
1317 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1318 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1319 Pointers.push_back(Ptr);
1320 Starts.push_back(AR->getStart());
1321 Ends.push_back(ScEnd);
1322 IsWritePtr.push_back(WritePtr);
1323 DependencySetId.push_back(DepSetId);
1326 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1327 // We need to place the broadcast of invariant variables outside the loop.
1328 Instruction *Instr = dyn_cast<Instruction>(V);
1330 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1331 Instr->getParent()) != LoopVectorBody.end());
1332 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1334 // Place the code for broadcasting invariant variables in the new preheader.
1335 IRBuilder<>::InsertPointGuard Guard(Builder);
1337 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1339 // Broadcast the scalar into all locations in the vector.
1340 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1345 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1347 assert(Val->getType()->isVectorTy() && "Must be a vector");
1348 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1349 "Elem must be an integer");
1350 // Create the types.
1351 Type *ITy = Val->getType()->getScalarType();
1352 VectorType *Ty = cast<VectorType>(Val->getType());
1353 int VLen = Ty->getNumElements();
1354 SmallVector<Constant*, 8> Indices;
1356 // Create a vector of consecutive numbers from zero to VF.
1357 for (int i = 0; i < VLen; ++i) {
1358 int64_t Idx = Negate ? (-i) : i;
1359 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1362 // Add the consecutive indices to the vector value.
1363 Constant *Cv = ConstantVector::get(Indices);
1364 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1365 return Builder.CreateAdd(Val, Cv, "induction");
1368 /// \brief Find the operand of the GEP that should be checked for consecutive
1369 /// stores. This ignores trailing indices that have no effect on the final
1371 static unsigned getGEPInductionOperand(const DataLayout *DL,
1372 const GetElementPtrInst *Gep) {
1373 unsigned LastOperand = Gep->getNumOperands() - 1;
1374 unsigned GEPAllocSize = DL->getTypeAllocSize(
1375 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1377 // Walk backwards and try to peel off zeros.
1378 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1379 // Find the type we're currently indexing into.
1380 gep_type_iterator GEPTI = gep_type_begin(Gep);
1381 std::advance(GEPTI, LastOperand - 1);
1383 // If it's a type with the same allocation size as the result of the GEP we
1384 // can peel off the zero index.
1385 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1393 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1394 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1395 // Make sure that the pointer does not point to structs.
1396 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1399 // If this value is a pointer induction variable we know it is consecutive.
1400 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1401 if (Phi && Inductions.count(Phi)) {
1402 InductionInfo II = Inductions[Phi];
1403 if (IK_PtrInduction == II.IK)
1405 else if (IK_ReversePtrInduction == II.IK)
1409 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1413 unsigned NumOperands = Gep->getNumOperands();
1414 Value *GpPtr = Gep->getPointerOperand();
1415 // If this GEP value is a consecutive pointer induction variable and all of
1416 // the indices are constant then we know it is consecutive. We can
1417 Phi = dyn_cast<PHINode>(GpPtr);
1418 if (Phi && Inductions.count(Phi)) {
1420 // Make sure that the pointer does not point to structs.
1421 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1422 if (GepPtrType->getElementType()->isAggregateType())
1425 // Make sure that all of the index operands are loop invariant.
1426 for (unsigned i = 1; i < NumOperands; ++i)
1427 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1430 InductionInfo II = Inductions[Phi];
1431 if (IK_PtrInduction == II.IK)
1433 else if (IK_ReversePtrInduction == II.IK)
1437 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1439 // Check that all of the gep indices are uniform except for our induction
1441 for (unsigned i = 0; i != NumOperands; ++i)
1442 if (i != InductionOperand &&
1443 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1446 // We can emit wide load/stores only if the last non-zero index is the
1447 // induction variable.
1448 const SCEV *Last = nullptr;
1449 if (!Strides.count(Gep))
1450 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1452 // Because of the multiplication by a stride we can have a s/zext cast.
1453 // We are going to replace this stride by 1 so the cast is safe to ignore.
1455 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1456 // %0 = trunc i64 %indvars.iv to i32
1457 // %mul = mul i32 %0, %Stride1
1458 // %idxprom = zext i32 %mul to i64 << Safe cast.
1459 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1461 Last = replaceSymbolicStrideSCEV(SE, Strides,
1462 Gep->getOperand(InductionOperand), Gep);
1463 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1465 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1469 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1470 const SCEV *Step = AR->getStepRecurrence(*SE);
1472 // The memory is consecutive because the last index is consecutive
1473 // and all other indices are loop invariant.
1476 if (Step->isAllOnesValue())
1483 bool LoopVectorizationLegality::isUniform(Value *V) {
1484 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1487 InnerLoopVectorizer::VectorParts&
1488 InnerLoopVectorizer::getVectorValue(Value *V) {
1489 assert(V != Induction && "The new induction variable should not be used.");
1490 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1492 // If we have a stride that is replaced by one, do it here.
1493 if (Legal->hasStride(V))
1494 V = ConstantInt::get(V->getType(), 1);
1496 // If we have this scalar in the map, return it.
1497 if (WidenMap.has(V))
1498 return WidenMap.get(V);
1500 // If this scalar is unknown, assume that it is a constant or that it is
1501 // loop invariant. Broadcast V and save the value for future uses.
1502 Value *B = getBroadcastInstrs(V);
1503 return WidenMap.splat(V, B);
1506 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1507 assert(Vec->getType()->isVectorTy() && "Invalid type");
1508 SmallVector<Constant*, 8> ShuffleMask;
1509 for (unsigned i = 0; i < VF; ++i)
1510 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1512 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1513 ConstantVector::get(ShuffleMask),
1517 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1518 // Attempt to issue a wide load.
1519 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1520 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1522 assert((LI || SI) && "Invalid Load/Store instruction");
1524 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1525 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1526 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1527 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1528 // An alignment of 0 means target abi alignment. We need to use the scalar's
1529 // target abi alignment in such a case.
1531 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1532 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1533 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1534 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1536 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1537 return scalarizeInstruction(Instr, true);
1539 if (ScalarAllocatedSize != VectorElementSize)
1540 return scalarizeInstruction(Instr);
1542 // If the pointer is loop invariant or if it is non-consecutive,
1543 // scalarize the load.
1544 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1545 bool Reverse = ConsecutiveStride < 0;
1546 bool UniformLoad = LI && Legal->isUniform(Ptr);
1547 if (!ConsecutiveStride || UniformLoad)
1548 return scalarizeInstruction(Instr);
1550 Constant *Zero = Builder.getInt32(0);
1551 VectorParts &Entry = WidenMap.get(Instr);
1553 // Handle consecutive loads/stores.
1554 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1555 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1556 setDebugLocFromInst(Builder, Gep);
1557 Value *PtrOperand = Gep->getPointerOperand();
1558 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1559 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1561 // Create the new GEP with the new induction variable.
1562 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1563 Gep2->setOperand(0, FirstBasePtr);
1564 Gep2->setName("gep.indvar.base");
1565 Ptr = Builder.Insert(Gep2);
1567 setDebugLocFromInst(Builder, Gep);
1568 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1569 OrigLoop) && "Base ptr must be invariant");
1571 // The last index does not have to be the induction. It can be
1572 // consecutive and be a function of the index. For example A[I+1];
1573 unsigned NumOperands = Gep->getNumOperands();
1574 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1575 // Create the new GEP with the new induction variable.
1576 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1578 for (unsigned i = 0; i < NumOperands; ++i) {
1579 Value *GepOperand = Gep->getOperand(i);
1580 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1582 // Update last index or loop invariant instruction anchored in loop.
1583 if (i == InductionOperand ||
1584 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1585 assert((i == InductionOperand ||
1586 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1587 "Must be last index or loop invariant");
1589 VectorParts &GEPParts = getVectorValue(GepOperand);
1590 Value *Index = GEPParts[0];
1591 Index = Builder.CreateExtractElement(Index, Zero);
1592 Gep2->setOperand(i, Index);
1593 Gep2->setName("gep.indvar.idx");
1596 Ptr = Builder.Insert(Gep2);
1598 // Use the induction element ptr.
1599 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1600 setDebugLocFromInst(Builder, Ptr);
1601 VectorParts &PtrVal = getVectorValue(Ptr);
1602 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1607 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1608 "We do not allow storing to uniform addresses");
1609 setDebugLocFromInst(Builder, SI);
1610 // We don't want to update the value in the map as it might be used in
1611 // another expression. So don't use a reference type for "StoredVal".
1612 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1614 for (unsigned Part = 0; Part < UF; ++Part) {
1615 // Calculate the pointer for the specific unroll-part.
1616 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1619 // If we store to reverse consecutive memory locations then we need
1620 // to reverse the order of elements in the stored value.
1621 StoredVal[Part] = reverseVector(StoredVal[Part]);
1622 // If the address is consecutive but reversed, then the
1623 // wide store needs to start at the last vector element.
1624 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1625 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1628 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1629 DataTy->getPointerTo(AddressSpace));
1630 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1636 assert(LI && "Must have a load instruction");
1637 setDebugLocFromInst(Builder, LI);
1638 for (unsigned Part = 0; Part < UF; ++Part) {
1639 // Calculate the pointer for the specific unroll-part.
1640 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1643 // If the address is consecutive but reversed, then the
1644 // wide store needs to start at the last vector element.
1645 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1646 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1649 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1650 DataTy->getPointerTo(AddressSpace));
1651 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1652 cast<LoadInst>(LI)->setAlignment(Alignment);
1653 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1657 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1658 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1659 // Holds vector parameters or scalars, in case of uniform vals.
1660 SmallVector<VectorParts, 4> Params;
1662 setDebugLocFromInst(Builder, Instr);
1664 // Find all of the vectorized parameters.
1665 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1666 Value *SrcOp = Instr->getOperand(op);
1668 // If we are accessing the old induction variable, use the new one.
1669 if (SrcOp == OldInduction) {
1670 Params.push_back(getVectorValue(SrcOp));
1674 // Try using previously calculated values.
1675 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1677 // If the src is an instruction that appeared earlier in the basic block
1678 // then it should already be vectorized.
1679 if (SrcInst && OrigLoop->contains(SrcInst)) {
1680 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1681 // The parameter is a vector value from earlier.
1682 Params.push_back(WidenMap.get(SrcInst));
1684 // The parameter is a scalar from outside the loop. Maybe even a constant.
1685 VectorParts Scalars;
1686 Scalars.append(UF, SrcOp);
1687 Params.push_back(Scalars);
1691 assert(Params.size() == Instr->getNumOperands() &&
1692 "Invalid number of operands");
1694 // Does this instruction return a value ?
1695 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1697 Value *UndefVec = IsVoidRetTy ? nullptr :
1698 UndefValue::get(VectorType::get(Instr->getType(), VF));
1699 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1700 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1702 Instruction *InsertPt = Builder.GetInsertPoint();
1703 BasicBlock *IfBlock = Builder.GetInsertBlock();
1704 BasicBlock *CondBlock = nullptr;
1707 Loop *VectorLp = nullptr;
1708 if (IfPredicateStore) {
1709 assert(Instr->getParent()->getSinglePredecessor() &&
1710 "Only support single predecessor blocks");
1711 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1712 Instr->getParent());
1713 VectorLp = LI->getLoopFor(IfBlock);
1714 assert(VectorLp && "Must have a loop for this block");
1717 // For each vector unroll 'part':
1718 for (unsigned Part = 0; Part < UF; ++Part) {
1719 // For each scalar that we create:
1720 for (unsigned Width = 0; Width < VF; ++Width) {
1723 Value *Cmp = nullptr;
1724 if (IfPredicateStore) {
1725 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1726 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1727 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1728 LoopVectorBody.push_back(CondBlock);
1729 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1730 // Update Builder with newly created basic block.
1731 Builder.SetInsertPoint(InsertPt);
1734 Instruction *Cloned = Instr->clone();
1736 Cloned->setName(Instr->getName() + ".cloned");
1737 // Replace the operands of the cloned instructions with extracted scalars.
1738 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1739 Value *Op = Params[op][Part];
1740 // Param is a vector. Need to extract the right lane.
1741 if (Op->getType()->isVectorTy())
1742 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1743 Cloned->setOperand(op, Op);
1746 // Place the cloned scalar in the new loop.
1747 Builder.Insert(Cloned);
1749 // If the original scalar returns a value we need to place it in a vector
1750 // so that future users will be able to use it.
1752 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1753 Builder.getInt32(Width));
1755 if (IfPredicateStore) {
1756 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1757 LoopVectorBody.push_back(NewIfBlock);
1758 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1759 Builder.SetInsertPoint(InsertPt);
1760 Instruction *OldBr = IfBlock->getTerminator();
1761 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1762 OldBr->eraseFromParent();
1763 IfBlock = NewIfBlock;
1769 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1773 if (Instruction *I = dyn_cast<Instruction>(V))
1774 return I->getParent() == Loc->getParent() ? I : nullptr;
1778 std::pair<Instruction *, Instruction *>
1779 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1780 Instruction *tnullptr = nullptr;
1781 if (!Legal->mustCheckStrides())
1782 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1784 IRBuilder<> ChkBuilder(Loc);
1787 Value *Check = nullptr;
1788 Instruction *FirstInst = nullptr;
1789 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1790 SE = Legal->strides_end();
1792 Value *Ptr = stripIntegerCast(*SI);
1793 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1795 // Store the first instruction we create.
1796 FirstInst = getFirstInst(FirstInst, C, Loc);
1798 Check = ChkBuilder.CreateOr(Check, C);
1803 // We have to do this trickery because the IRBuilder might fold the check to a
1804 // constant expression in which case there is no Instruction anchored in a
1806 LLVMContext &Ctx = Loc->getContext();
1807 Instruction *TheCheck =
1808 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1809 ChkBuilder.Insert(TheCheck, "stride.not.one");
1810 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1812 return std::make_pair(FirstInst, TheCheck);
1815 std::pair<Instruction *, Instruction *>
1816 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1817 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1818 Legal->getRuntimePointerCheck();
1820 Instruction *tnullptr = nullptr;
1821 if (!PtrRtCheck->Need)
1822 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1824 unsigned NumPointers = PtrRtCheck->Pointers.size();
1825 SmallVector<TrackingVH<Value> , 2> Starts;
1826 SmallVector<TrackingVH<Value> , 2> Ends;
1828 LLVMContext &Ctx = Loc->getContext();
1829 SCEVExpander Exp(*SE, "induction");
1830 Instruction *FirstInst = nullptr;
1832 for (unsigned i = 0; i < NumPointers; ++i) {
1833 Value *Ptr = PtrRtCheck->Pointers[i];
1834 const SCEV *Sc = SE->getSCEV(Ptr);
1836 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1837 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1839 Starts.push_back(Ptr);
1840 Ends.push_back(Ptr);
1842 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1843 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1845 // Use this type for pointer arithmetic.
1846 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1848 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1849 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1850 Starts.push_back(Start);
1851 Ends.push_back(End);
1855 IRBuilder<> ChkBuilder(Loc);
1856 // Our instructions might fold to a constant.
1857 Value *MemoryRuntimeCheck = nullptr;
1858 for (unsigned i = 0; i < NumPointers; ++i) {
1859 for (unsigned j = i+1; j < NumPointers; ++j) {
1860 // No need to check if two readonly pointers intersect.
1861 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1864 // Only need to check pointers between two different dependency sets.
1865 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1868 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1869 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1871 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1872 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1873 "Trying to bounds check pointers with different address spaces");
1875 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1876 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1878 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1879 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1880 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1881 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1883 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1884 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1885 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1886 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1887 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1888 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1889 if (MemoryRuntimeCheck) {
1890 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1892 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1894 MemoryRuntimeCheck = IsConflict;
1898 // We have to do this trickery because the IRBuilder might fold the check to a
1899 // constant expression in which case there is no Instruction anchored in a
1901 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1902 ConstantInt::getTrue(Ctx));
1903 ChkBuilder.Insert(Check, "memcheck.conflict");
1904 FirstInst = getFirstInst(FirstInst, Check, Loc);
1905 return std::make_pair(FirstInst, Check);
1908 void InnerLoopVectorizer::createEmptyLoop() {
1910 In this function we generate a new loop. The new loop will contain
1911 the vectorized instructions while the old loop will continue to run the
1914 [ ] <-- vector loop bypass (may consist of multiple blocks).
1917 | [ ] <-- vector pre header.
1921 | [ ]_| <-- vector loop.
1924 >[ ] <--- middle-block.
1927 | [ ] <--- new preheader.
1931 | [ ]_| <-- old scalar loop to handle remainder.
1934 >[ ] <-- exit block.
1938 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
1939 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
1940 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
1941 assert(ExitBlock && "Must have an exit block");
1943 // Some loops have a single integer induction variable, while other loops
1944 // don't. One example is c++ iterators that often have multiple pointer
1945 // induction variables. In the code below we also support a case where we
1946 // don't have a single induction variable.
1947 OldInduction = Legal->getInduction();
1948 Type *IdxTy = Legal->getWidestInductionType();
1950 // Find the loop boundaries.
1951 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
1952 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
1954 // The exit count might have the type of i64 while the phi is i32. This can
1955 // happen if we have an induction variable that is sign extended before the
1956 // compare. The only way that we get a backedge taken count is that the
1957 // induction variable was signed and as such will not overflow. In such a case
1958 // truncation is legal.
1959 if (ExitCount->getType()->getPrimitiveSizeInBits() >
1960 IdxTy->getPrimitiveSizeInBits())
1961 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
1963 ExitCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
1964 // Get the total trip count from the count by adding 1.
1965 ExitCount = SE->getAddExpr(ExitCount,
1966 SE->getConstant(ExitCount->getType(), 1));
1968 // Expand the trip count and place the new instructions in the preheader.
1969 // Notice that the pre-header does not change, only the loop body.
1970 SCEVExpander Exp(*SE, "induction");
1972 // Count holds the overall loop count (N).
1973 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
1974 BypassBlock->getTerminator());
1976 // The loop index does not have to start at Zero. Find the original start
1977 // value from the induction PHI node. If we don't have an induction variable
1978 // then we know that it starts at zero.
1979 Builder.SetInsertPoint(BypassBlock->getTerminator());
1980 Value *StartIdx = ExtendedIdx = OldInduction ?
1981 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
1983 ConstantInt::get(IdxTy, 0);
1985 assert(BypassBlock && "Invalid loop structure");
1986 LoopBypassBlocks.push_back(BypassBlock);
1988 // Split the single block loop into the two loop structure described above.
1989 BasicBlock *VectorPH =
1990 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
1991 BasicBlock *VecBody =
1992 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
1993 BasicBlock *MiddleBlock =
1994 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
1995 BasicBlock *ScalarPH =
1996 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
1998 // Create and register the new vector loop.
1999 Loop* Lp = new Loop();
2000 Loop *ParentLoop = OrigLoop->getParentLoop();
2002 // Insert the new loop into the loop nest and register the new basic blocks
2003 // before calling any utilities such as SCEV that require valid LoopInfo.
2005 ParentLoop->addChildLoop(Lp);
2006 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2007 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2008 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2010 LI->addTopLevelLoop(Lp);
2012 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2014 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2016 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2018 // Generate the induction variable.
2019 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2020 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2021 // The loop step is equal to the vectorization factor (num of SIMD elements)
2022 // times the unroll factor (num of SIMD instructions).
2023 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2025 // This is the IR builder that we use to add all of the logic for bypassing
2026 // the new vector loop.
2027 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2028 setDebugLocFromInst(BypassBuilder,
2029 getDebugLocFromInstOrOperands(OldInduction));
2031 // We may need to extend the index in case there is a type mismatch.
2032 // We know that the count starts at zero and does not overflow.
2033 if (Count->getType() != IdxTy) {
2034 // The exit count can be of pointer type. Convert it to the correct
2036 if (ExitCount->getType()->isPointerTy())
2037 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2039 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2042 // Add the start index to the loop count to get the new end index.
2043 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2045 // Now we need to generate the expression for N - (N % VF), which is
2046 // the part that the vectorized body will execute.
2047 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2048 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2049 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2050 "end.idx.rnd.down");
2052 // Now, compare the new count to zero. If it is zero skip the vector loop and
2053 // jump to the scalar loop.
2054 Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
2057 BasicBlock *LastBypassBlock = BypassBlock;
2059 // Generate the code to check that the strides we assumed to be one are really
2060 // one. We want the new basic block to start at the first instruction in a
2061 // sequence of instructions that form a check.
2062 Instruction *StrideCheck;
2063 Instruction *FirstCheckInst;
2064 std::tie(FirstCheckInst, StrideCheck) =
2065 addStrideCheck(BypassBlock->getTerminator());
2067 // Create a new block containing the stride check.
2068 BasicBlock *CheckBlock =
2069 BypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2071 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2072 LoopBypassBlocks.push_back(CheckBlock);
2074 // Replace the branch into the memory check block with a conditional branch
2075 // for the "few elements case".
2076 Instruction *OldTerm = BypassBlock->getTerminator();
2077 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2078 OldTerm->eraseFromParent();
2081 LastBypassBlock = CheckBlock;
2084 // Generate the code that checks in runtime if arrays overlap. We put the
2085 // checks into a separate block to make the more common case of few elements
2087 Instruction *MemRuntimeCheck;
2088 std::tie(FirstCheckInst, MemRuntimeCheck) =
2089 addRuntimeCheck(LastBypassBlock->getTerminator());
2090 if (MemRuntimeCheck) {
2091 // Create a new block containing the memory check.
2092 BasicBlock *CheckBlock =
2093 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2095 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2096 LoopBypassBlocks.push_back(CheckBlock);
2098 // Replace the branch into the memory check block with a conditional branch
2099 // for the "few elements case".
2100 Instruction *OldTerm = LastBypassBlock->getTerminator();
2101 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2102 OldTerm->eraseFromParent();
2104 Cmp = MemRuntimeCheck;
2105 LastBypassBlock = CheckBlock;
2108 LastBypassBlock->getTerminator()->eraseFromParent();
2109 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2112 // We are going to resume the execution of the scalar loop.
2113 // Go over all of the induction variables that we found and fix the
2114 // PHIs that are left in the scalar version of the loop.
2115 // The starting values of PHI nodes depend on the counter of the last
2116 // iteration in the vectorized loop.
2117 // If we come from a bypass edge then we need to start from the original
2120 // This variable saves the new starting index for the scalar loop.
2121 PHINode *ResumeIndex = nullptr;
2122 LoopVectorizationLegality::InductionList::iterator I, E;
2123 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2124 // Set builder to point to last bypass block.
2125 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2126 for (I = List->begin(), E = List->end(); I != E; ++I) {
2127 PHINode *OrigPhi = I->first;
2128 LoopVectorizationLegality::InductionInfo II = I->second;
2130 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2131 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2132 MiddleBlock->getTerminator());
2133 // We might have extended the type of the induction variable but we need a
2134 // truncated version for the scalar loop.
2135 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2136 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2137 MiddleBlock->getTerminator()) : nullptr;
2139 Value *EndValue = nullptr;
2141 case LoopVectorizationLegality::IK_NoInduction:
2142 llvm_unreachable("Unknown induction");
2143 case LoopVectorizationLegality::IK_IntInduction: {
2144 // Handle the integer induction counter.
2145 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2147 // We have the canonical induction variable.
2148 if (OrigPhi == OldInduction) {
2149 // Create a truncated version of the resume value for the scalar loop,
2150 // we might have promoted the type to a larger width.
2152 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2153 // The new PHI merges the original incoming value, in case of a bypass,
2154 // or the value at the end of the vectorized loop.
2155 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2156 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2157 TruncResumeVal->addIncoming(EndValue, VecBody);
2159 // We know what the end value is.
2160 EndValue = IdxEndRoundDown;
2161 // We also know which PHI node holds it.
2162 ResumeIndex = ResumeVal;
2166 // Not the canonical induction variable - add the vector loop count to the
2168 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2169 II.StartValue->getType(),
2171 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2174 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2175 // Convert the CountRoundDown variable to the PHI size.
2176 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2177 II.StartValue->getType(),
2179 // Handle reverse integer induction counter.
2180 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2183 case LoopVectorizationLegality::IK_PtrInduction: {
2184 // For pointer induction variables, calculate the offset using
2186 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2190 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2191 // The value at the end of the loop for the reverse pointer is calculated
2192 // by creating a GEP with a negative index starting from the start value.
2193 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2194 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2196 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2202 // The new PHI merges the original incoming value, in case of a bypass,
2203 // or the value at the end of the vectorized loop.
2204 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
2205 if (OrigPhi == OldInduction)
2206 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2208 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2210 ResumeVal->addIncoming(EndValue, VecBody);
2212 // Fix the scalar body counter (PHI node).
2213 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2214 // The old inductions phi node in the scalar body needs the truncated value.
2215 if (OrigPhi == OldInduction)
2216 OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
2218 OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
2221 // If we are generating a new induction variable then we also need to
2222 // generate the code that calculates the exit value. This value is not
2223 // simply the end of the counter because we may skip the vectorized body
2224 // in case of a runtime check.
2226 assert(!ResumeIndex && "Unexpected resume value found");
2227 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2228 MiddleBlock->getTerminator());
2229 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2230 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2231 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2234 // Make sure that we found the index where scalar loop needs to continue.
2235 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2236 "Invalid resume Index");
2238 // Add a check in the middle block to see if we have completed
2239 // all of the iterations in the first vector loop.
2240 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2241 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2242 ResumeIndex, "cmp.n",
2243 MiddleBlock->getTerminator());
2245 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2246 // Remove the old terminator.
2247 MiddleBlock->getTerminator()->eraseFromParent();
2249 // Create i+1 and fill the PHINode.
2250 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2251 Induction->addIncoming(StartIdx, VectorPH);
2252 Induction->addIncoming(NextIdx, VecBody);
2253 // Create the compare.
2254 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2255 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2257 // Now we have two terminators. Remove the old one from the block.
2258 VecBody->getTerminator()->eraseFromParent();
2260 // Get ready to start creating new instructions into the vectorized body.
2261 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2264 LoopVectorPreHeader = VectorPH;
2265 LoopScalarPreHeader = ScalarPH;
2266 LoopMiddleBlock = MiddleBlock;
2267 LoopExitBlock = ExitBlock;
2268 LoopVectorBody.push_back(VecBody);
2269 LoopScalarBody = OldBasicBlock;
2271 LoopVectorizeHints Hints(Lp, true);
2272 Hints.setAlreadyVectorized(Lp);
2275 /// This function returns the identity element (or neutral element) for
2276 /// the operation K.
2278 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2283 // Adding, Xoring, Oring zero to a number does not change it.
2284 return ConstantInt::get(Tp, 0);
2285 case RK_IntegerMult:
2286 // Multiplying a number by 1 does not change it.
2287 return ConstantInt::get(Tp, 1);
2289 // AND-ing a number with an all-1 value does not change it.
2290 return ConstantInt::get(Tp, -1, true);
2292 // Multiplying a number by 1 does not change it.
2293 return ConstantFP::get(Tp, 1.0L);
2295 // Adding zero to a number does not change it.
2296 return ConstantFP::get(Tp, 0.0L);
2298 llvm_unreachable("Unknown reduction kind");
2302 /// This function translates the reduction kind to an LLVM binary operator.
2304 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2306 case LoopVectorizationLegality::RK_IntegerAdd:
2307 return Instruction::Add;
2308 case LoopVectorizationLegality::RK_IntegerMult:
2309 return Instruction::Mul;
2310 case LoopVectorizationLegality::RK_IntegerOr:
2311 return Instruction::Or;
2312 case LoopVectorizationLegality::RK_IntegerAnd:
2313 return Instruction::And;
2314 case LoopVectorizationLegality::RK_IntegerXor:
2315 return Instruction::Xor;
2316 case LoopVectorizationLegality::RK_FloatMult:
2317 return Instruction::FMul;
2318 case LoopVectorizationLegality::RK_FloatAdd:
2319 return Instruction::FAdd;
2320 case LoopVectorizationLegality::RK_IntegerMinMax:
2321 return Instruction::ICmp;
2322 case LoopVectorizationLegality::RK_FloatMinMax:
2323 return Instruction::FCmp;
2325 llvm_unreachable("Unknown reduction operation");
2329 Value *createMinMaxOp(IRBuilder<> &Builder,
2330 LoopVectorizationLegality::MinMaxReductionKind RK,
2333 CmpInst::Predicate P = CmpInst::ICMP_NE;
2336 llvm_unreachable("Unknown min/max reduction kind");
2337 case LoopVectorizationLegality::MRK_UIntMin:
2338 P = CmpInst::ICMP_ULT;
2340 case LoopVectorizationLegality::MRK_UIntMax:
2341 P = CmpInst::ICMP_UGT;
2343 case LoopVectorizationLegality::MRK_SIntMin:
2344 P = CmpInst::ICMP_SLT;
2346 case LoopVectorizationLegality::MRK_SIntMax:
2347 P = CmpInst::ICMP_SGT;
2349 case LoopVectorizationLegality::MRK_FloatMin:
2350 P = CmpInst::FCMP_OLT;
2352 case LoopVectorizationLegality::MRK_FloatMax:
2353 P = CmpInst::FCMP_OGT;
2358 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2359 RK == LoopVectorizationLegality::MRK_FloatMax)
2360 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2362 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2364 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2369 struct CSEDenseMapInfo {
2370 static bool canHandle(Instruction *I) {
2371 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2372 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2374 static inline Instruction *getEmptyKey() {
2375 return DenseMapInfo<Instruction *>::getEmptyKey();
2377 static inline Instruction *getTombstoneKey() {
2378 return DenseMapInfo<Instruction *>::getTombstoneKey();
2380 static unsigned getHashValue(Instruction *I) {
2381 assert(canHandle(I) && "Unknown instruction!");
2382 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2383 I->value_op_end()));
2385 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2386 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2387 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2389 return LHS->isIdenticalTo(RHS);
2394 /// \brief Check whether this block is a predicated block.
2395 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2396 /// = ...; " blocks. We start with one vectorized basic block. For every
2397 /// conditional block we split this vectorized block. Therefore, every second
2398 /// block will be a predicated one.
2399 static bool isPredicatedBlock(unsigned BlockNum) {
2400 return BlockNum % 2;
2403 ///\brief Perform cse of induction variable instructions.
2404 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2405 // Perform simple cse.
2406 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2407 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2408 BasicBlock *BB = BBs[i];
2409 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2410 Instruction *In = I++;
2412 if (!CSEDenseMapInfo::canHandle(In))
2415 // Check if we can replace this instruction with any of the
2416 // visited instructions.
2417 if (Instruction *V = CSEMap.lookup(In)) {
2418 In->replaceAllUsesWith(V);
2419 In->eraseFromParent();
2422 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2423 // ...;" blocks for predicated stores. Every second block is a predicated
2425 if (isPredicatedBlock(i))
2433 /// \brief Adds a 'fast' flag to floating point operations.
2434 static Value *addFastMathFlag(Value *V) {
2435 if (isa<FPMathOperator>(V)){
2436 FastMathFlags Flags;
2437 Flags.setUnsafeAlgebra();
2438 cast<Instruction>(V)->setFastMathFlags(Flags);
2443 void InnerLoopVectorizer::vectorizeLoop() {
2444 //===------------------------------------------------===//
2446 // Notice: any optimization or new instruction that go
2447 // into the code below should be also be implemented in
2450 //===------------------------------------------------===//
2451 Constant *Zero = Builder.getInt32(0);
2453 // In order to support reduction variables we need to be able to vectorize
2454 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2455 // stages. First, we create a new vector PHI node with no incoming edges.
2456 // We use this value when we vectorize all of the instructions that use the
2457 // PHI. Next, after all of the instructions in the block are complete we
2458 // add the new incoming edges to the PHI. At this point all of the
2459 // instructions in the basic block are vectorized, so we can use them to
2460 // construct the PHI.
2461 PhiVector RdxPHIsToFix;
2463 // Scan the loop in a topological order to ensure that defs are vectorized
2465 LoopBlocksDFS DFS(OrigLoop);
2468 // Vectorize all of the blocks in the original loop.
2469 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2470 be = DFS.endRPO(); bb != be; ++bb)
2471 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2473 // At this point every instruction in the original loop is widened to
2474 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2475 // that we vectorized. The PHI nodes are currently empty because we did
2476 // not want to introduce cycles. Notice that the remaining PHI nodes
2477 // that we need to fix are reduction variables.
2479 // Create the 'reduced' values for each of the induction vars.
2480 // The reduced values are the vector values that we scalarize and combine
2481 // after the loop is finished.
2482 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2484 PHINode *RdxPhi = *it;
2485 assert(RdxPhi && "Unable to recover vectorized PHI");
2487 // Find the reduction variable descriptor.
2488 assert(Legal->getReductionVars()->count(RdxPhi) &&
2489 "Unable to find the reduction variable");
2490 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2491 (*Legal->getReductionVars())[RdxPhi];
2493 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2495 // We need to generate a reduction vector from the incoming scalar.
2496 // To do so, we need to generate the 'identity' vector and override
2497 // one of the elements with the incoming scalar reduction. We need
2498 // to do it in the vector-loop preheader.
2499 Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
2501 // This is the vector-clone of the value that leaves the loop.
2502 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2503 Type *VecTy = VectorExit[0]->getType();
2505 // Find the reduction identity variable. Zero for addition, or, xor,
2506 // one for multiplication, -1 for And.
2509 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2510 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2511 // MinMax reduction have the start value as their identify.
2513 VectorStart = Identity = RdxDesc.StartValue;
2515 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2520 // Handle other reduction kinds:
2522 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2523 VecTy->getScalarType());
2526 // This vector is the Identity vector where the first element is the
2527 // incoming scalar reduction.
2528 VectorStart = RdxDesc.StartValue;
2530 Identity = ConstantVector::getSplat(VF, Iden);
2532 // This vector is the Identity vector where the first element is the
2533 // incoming scalar reduction.
2534 VectorStart = Builder.CreateInsertElement(Identity,
2535 RdxDesc.StartValue, Zero);
2539 // Fix the vector-loop phi.
2540 // We created the induction variable so we know that the
2541 // preheader is the first entry.
2542 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2544 // Reductions do not have to start at zero. They can start with
2545 // any loop invariant values.
2546 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2547 BasicBlock *Latch = OrigLoop->getLoopLatch();
2548 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2549 VectorParts &Val = getVectorValue(LoopVal);
2550 for (unsigned part = 0; part < UF; ++part) {
2551 // Make sure to add the reduction stat value only to the
2552 // first unroll part.
2553 Value *StartVal = (part == 0) ? VectorStart : Identity;
2554 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2555 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2556 LoopVectorBody.back());
2559 // Before each round, move the insertion point right between
2560 // the PHIs and the values we are going to write.
2561 // This allows us to write both PHINodes and the extractelement
2563 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2565 VectorParts RdxParts;
2566 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2567 for (unsigned part = 0; part < UF; ++part) {
2568 // This PHINode contains the vectorized reduction variable, or
2569 // the initial value vector, if we bypass the vector loop.
2570 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2571 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2572 Value *StartVal = (part == 0) ? VectorStart : Identity;
2573 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2574 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2575 NewPhi->addIncoming(RdxExitVal[part],
2576 LoopVectorBody.back());
2577 RdxParts.push_back(NewPhi);
2580 // Reduce all of the unrolled parts into a single vector.
2581 Value *ReducedPartRdx = RdxParts[0];
2582 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2583 setDebugLocFromInst(Builder, ReducedPartRdx);
2584 for (unsigned part = 1; part < UF; ++part) {
2585 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2586 // Floating point operations had to be 'fast' to enable the reduction.
2587 ReducedPartRdx = addFastMathFlag(
2588 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2589 ReducedPartRdx, "bin.rdx"));
2591 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2592 ReducedPartRdx, RdxParts[part]);
2596 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2597 // and vector ops, reducing the set of values being computed by half each
2599 assert(isPowerOf2_32(VF) &&
2600 "Reduction emission only supported for pow2 vectors!");
2601 Value *TmpVec = ReducedPartRdx;
2602 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2603 for (unsigned i = VF; i != 1; i >>= 1) {
2604 // Move the upper half of the vector to the lower half.
2605 for (unsigned j = 0; j != i/2; ++j)
2606 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2608 // Fill the rest of the mask with undef.
2609 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2610 UndefValue::get(Builder.getInt32Ty()));
2613 Builder.CreateShuffleVector(TmpVec,
2614 UndefValue::get(TmpVec->getType()),
2615 ConstantVector::get(ShuffleMask),
2618 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2619 // Floating point operations had to be 'fast' to enable the reduction.
2620 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2621 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2623 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2626 // The result is in the first element of the vector.
2627 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2628 Builder.getInt32(0));
2631 // Now, we need to fix the users of the reduction variable
2632 // inside and outside of the scalar remainder loop.
2633 // We know that the loop is in LCSSA form. We need to update the
2634 // PHI nodes in the exit blocks.
2635 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2636 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2637 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2638 if (!LCSSAPhi) break;
2640 // All PHINodes need to have a single entry edge, or two if
2641 // we already fixed them.
2642 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2644 // We found our reduction value exit-PHI. Update it with the
2645 // incoming bypass edge.
2646 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2647 // Add an edge coming from the bypass.
2648 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2651 }// end of the LCSSA phi scan.
2653 // Fix the scalar loop reduction variable with the incoming reduction sum
2654 // from the vector body and from the backedge value.
2655 int IncomingEdgeBlockIdx =
2656 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2657 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2658 // Pick the other block.
2659 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2660 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
2661 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2662 }// end of for each redux variable.
2666 // Remove redundant induction instructions.
2667 cse(LoopVectorBody);
2670 void InnerLoopVectorizer::fixLCSSAPHIs() {
2671 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2672 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2673 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2674 if (!LCSSAPhi) break;
2675 if (LCSSAPhi->getNumIncomingValues() == 1)
2676 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2681 InnerLoopVectorizer::VectorParts
2682 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2683 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2686 // Look for cached value.
2687 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2688 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2689 if (ECEntryIt != MaskCache.end())
2690 return ECEntryIt->second;
2692 VectorParts SrcMask = createBlockInMask(Src);
2694 // The terminator has to be a branch inst!
2695 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2696 assert(BI && "Unexpected terminator found");
2698 if (BI->isConditional()) {
2699 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2701 if (BI->getSuccessor(0) != Dst)
2702 for (unsigned part = 0; part < UF; ++part)
2703 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2705 for (unsigned part = 0; part < UF; ++part)
2706 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2708 MaskCache[Edge] = EdgeMask;
2712 MaskCache[Edge] = SrcMask;
2716 InnerLoopVectorizer::VectorParts
2717 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2718 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2720 // Loop incoming mask is all-one.
2721 if (OrigLoop->getHeader() == BB) {
2722 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2723 return getVectorValue(C);
2726 // This is the block mask. We OR all incoming edges, and with zero.
2727 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2728 VectorParts BlockMask = getVectorValue(Zero);
2731 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2732 VectorParts EM = createEdgeMask(*it, BB);
2733 for (unsigned part = 0; part < UF; ++part)
2734 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2740 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2741 InnerLoopVectorizer::VectorParts &Entry,
2742 unsigned UF, unsigned VF, PhiVector *PV) {
2743 PHINode* P = cast<PHINode>(PN);
2744 // Handle reduction variables:
2745 if (Legal->getReductionVars()->count(P)) {
2746 for (unsigned part = 0; part < UF; ++part) {
2747 // This is phase one of vectorizing PHIs.
2748 Type *VecTy = (VF == 1) ? PN->getType() :
2749 VectorType::get(PN->getType(), VF);
2750 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2751 LoopVectorBody.back()-> getFirstInsertionPt());
2757 setDebugLocFromInst(Builder, P);
2758 // Check for PHI nodes that are lowered to vector selects.
2759 if (P->getParent() != OrigLoop->getHeader()) {
2760 // We know that all PHIs in non-header blocks are converted into
2761 // selects, so we don't have to worry about the insertion order and we
2762 // can just use the builder.
2763 // At this point we generate the predication tree. There may be
2764 // duplications since this is a simple recursive scan, but future
2765 // optimizations will clean it up.
2767 unsigned NumIncoming = P->getNumIncomingValues();
2769 // Generate a sequence of selects of the form:
2770 // SELECT(Mask3, In3,
2771 // SELECT(Mask2, In2,
2773 for (unsigned In = 0; In < NumIncoming; In++) {
2774 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2776 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2778 for (unsigned part = 0; part < UF; ++part) {
2779 // We might have single edge PHIs (blocks) - use an identity
2780 // 'select' for the first PHI operand.
2782 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2785 // Select between the current value and the previous incoming edge
2786 // based on the incoming mask.
2787 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2788 Entry[part], "predphi");
2794 // This PHINode must be an induction variable.
2795 // Make sure that we know about it.
2796 assert(Legal->getInductionVars()->count(P) &&
2797 "Not an induction variable");
2799 LoopVectorizationLegality::InductionInfo II =
2800 Legal->getInductionVars()->lookup(P);
2803 case LoopVectorizationLegality::IK_NoInduction:
2804 llvm_unreachable("Unknown induction");
2805 case LoopVectorizationLegality::IK_IntInduction: {
2806 assert(P->getType() == II.StartValue->getType() && "Types must match");
2807 Type *PhiTy = P->getType();
2809 if (P == OldInduction) {
2810 // Handle the canonical induction variable. We might have had to
2812 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2814 // Handle other induction variables that are now based on the
2816 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2818 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2819 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2822 Broadcasted = getBroadcastInstrs(Broadcasted);
2823 // After broadcasting the induction variable we need to make the vector
2824 // consecutive by adding 0, 1, 2, etc.
2825 for (unsigned part = 0; part < UF; ++part)
2826 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
2829 case LoopVectorizationLegality::IK_ReverseIntInduction:
2830 case LoopVectorizationLegality::IK_PtrInduction:
2831 case LoopVectorizationLegality::IK_ReversePtrInduction:
2832 // Handle reverse integer and pointer inductions.
2833 Value *StartIdx = ExtendedIdx;
2834 // This is the normalized GEP that starts counting at zero.
2835 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
2838 // Handle the reverse integer induction variable case.
2839 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
2840 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
2841 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
2843 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
2846 // This is a new value so do not hoist it out.
2847 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
2848 // After broadcasting the induction variable we need to make the
2849 // vector consecutive by adding ... -3, -2, -1, 0.
2850 for (unsigned part = 0; part < UF; ++part)
2851 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
2856 // Handle the pointer induction variable case.
2857 assert(P->getType()->isPointerTy() && "Unexpected type.");
2859 // Is this a reverse induction ptr or a consecutive induction ptr.
2860 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
2863 // This is the vector of results. Notice that we don't generate
2864 // vector geps because scalar geps result in better code.
2865 for (unsigned part = 0; part < UF; ++part) {
2867 int EltIndex = (part) * (Reverse ? -1 : 1);
2868 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2871 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2873 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2875 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2877 Entry[part] = SclrGep;
2881 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2882 for (unsigned int i = 0; i < VF; ++i) {
2883 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
2884 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2887 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
2889 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
2891 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
2893 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2894 Builder.getInt32(i),
2897 Entry[part] = VecVal;
2903 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
2904 // For each instruction in the old loop.
2905 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2906 VectorParts &Entry = WidenMap.get(it);
2907 switch (it->getOpcode()) {
2908 case Instruction::Br:
2909 // Nothing to do for PHIs and BR, since we already took care of the
2910 // loop control flow instructions.
2912 case Instruction::PHI:{
2913 // Vectorize PHINodes.
2914 widenPHIInstruction(it, Entry, UF, VF, PV);
2918 case Instruction::Add:
2919 case Instruction::FAdd:
2920 case Instruction::Sub:
2921 case Instruction::FSub:
2922 case Instruction::Mul:
2923 case Instruction::FMul:
2924 case Instruction::UDiv:
2925 case Instruction::SDiv:
2926 case Instruction::FDiv:
2927 case Instruction::URem:
2928 case Instruction::SRem:
2929 case Instruction::FRem:
2930 case Instruction::Shl:
2931 case Instruction::LShr:
2932 case Instruction::AShr:
2933 case Instruction::And:
2934 case Instruction::Or:
2935 case Instruction::Xor: {
2936 // Just widen binops.
2937 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
2938 setDebugLocFromInst(Builder, BinOp);
2939 VectorParts &A = getVectorValue(it->getOperand(0));
2940 VectorParts &B = getVectorValue(it->getOperand(1));
2942 // Use this vector value for all users of the original instruction.
2943 for (unsigned Part = 0; Part < UF; ++Part) {
2944 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
2946 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
2947 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
2948 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
2949 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
2950 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
2952 if (VecOp && isa<PossiblyExactOperator>(VecOp))
2953 VecOp->setIsExact(BinOp->isExact());
2955 // Copy the fast-math flags.
2956 if (VecOp && isa<FPMathOperator>(V))
2957 VecOp->setFastMathFlags(it->getFastMathFlags());
2963 case Instruction::Select: {
2965 // If the selector is loop invariant we can create a select
2966 // instruction with a scalar condition. Otherwise, use vector-select.
2967 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
2969 setDebugLocFromInst(Builder, it);
2971 // The condition can be loop invariant but still defined inside the
2972 // loop. This means that we can't just use the original 'cond' value.
2973 // We have to take the 'vectorized' value and pick the first lane.
2974 // Instcombine will make this a no-op.
2975 VectorParts &Cond = getVectorValue(it->getOperand(0));
2976 VectorParts &Op0 = getVectorValue(it->getOperand(1));
2977 VectorParts &Op1 = getVectorValue(it->getOperand(2));
2979 Value *ScalarCond = (VF == 1) ? Cond[0] :
2980 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
2982 for (unsigned Part = 0; Part < UF; ++Part) {
2983 Entry[Part] = Builder.CreateSelect(
2984 InvariantCond ? ScalarCond : Cond[Part],
2991 case Instruction::ICmp:
2992 case Instruction::FCmp: {
2993 // Widen compares. Generate vector compares.
2994 bool FCmp = (it->getOpcode() == Instruction::FCmp);
2995 CmpInst *Cmp = dyn_cast<CmpInst>(it);
2996 setDebugLocFromInst(Builder, it);
2997 VectorParts &A = getVectorValue(it->getOperand(0));
2998 VectorParts &B = getVectorValue(it->getOperand(1));
2999 for (unsigned Part = 0; Part < UF; ++Part) {
3002 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3004 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3010 case Instruction::Store:
3011 case Instruction::Load:
3012 vectorizeMemoryInstruction(it);
3014 case Instruction::ZExt:
3015 case Instruction::SExt:
3016 case Instruction::FPToUI:
3017 case Instruction::FPToSI:
3018 case Instruction::FPExt:
3019 case Instruction::PtrToInt:
3020 case Instruction::IntToPtr:
3021 case Instruction::SIToFP:
3022 case Instruction::UIToFP:
3023 case Instruction::Trunc:
3024 case Instruction::FPTrunc:
3025 case Instruction::BitCast: {
3026 CastInst *CI = dyn_cast<CastInst>(it);
3027 setDebugLocFromInst(Builder, it);
3028 /// Optimize the special case where the source is the induction
3029 /// variable. Notice that we can only optimize the 'trunc' case
3030 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3031 /// c. other casts depend on pointer size.
3032 if (CI->getOperand(0) == OldInduction &&
3033 it->getOpcode() == Instruction::Trunc) {
3034 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3036 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3037 for (unsigned Part = 0; Part < UF; ++Part)
3038 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3041 /// Vectorize casts.
3042 Type *DestTy = (VF == 1) ? CI->getType() :
3043 VectorType::get(CI->getType(), VF);
3045 VectorParts &A = getVectorValue(it->getOperand(0));
3046 for (unsigned Part = 0; Part < UF; ++Part)
3047 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3051 case Instruction::Call: {
3052 // Ignore dbg intrinsics.
3053 if (isa<DbgInfoIntrinsic>(it))
3055 setDebugLocFromInst(Builder, it);
3057 Module *M = BB->getParent()->getParent();
3058 CallInst *CI = cast<CallInst>(it);
3059 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3060 assert(ID && "Not an intrinsic call!");
3062 case Intrinsic::lifetime_end:
3063 case Intrinsic::lifetime_start:
3064 scalarizeInstruction(it);
3067 for (unsigned Part = 0; Part < UF; ++Part) {
3068 SmallVector<Value *, 4> Args;
3069 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3070 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3071 Args.push_back(Arg[Part]);
3073 Type *Tys[] = {CI->getType()};
3075 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3077 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3078 Entry[Part] = Builder.CreateCall(F, Args);
3086 // All other instructions are unsupported. Scalarize them.
3087 scalarizeInstruction(it);
3090 }// end of for_each instr.
3093 void InnerLoopVectorizer::updateAnalysis() {
3094 // Forget the original basic block.
3095 SE->forgetLoop(OrigLoop);
3097 // Update the dominator tree information.
3098 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3099 "Entry does not dominate exit.");
3101 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3102 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3103 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3105 // Due to if predication of stores we might create a sequence of "if(pred)
3106 // a[i] = ...; " blocks.
3107 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3109 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3110 else if (isPredicatedBlock(i)) {
3111 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3113 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3117 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
3118 DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
3119 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3120 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3122 DEBUG(DT->verifyDomTree());
3125 /// \brief Check whether it is safe to if-convert this phi node.
3127 /// Phi nodes with constant expressions that can trap are not safe to if
3129 static bool canIfConvertPHINodes(BasicBlock *BB) {
3130 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3131 PHINode *Phi = dyn_cast<PHINode>(I);
3134 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3135 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3142 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3143 if (!EnableIfConversion)
3146 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3148 // A list of pointers that we can safely read and write to.
3149 SmallPtrSet<Value *, 8> SafePointes;
3151 // Collect safe addresses.
3152 for (Loop::block_iterator BI = TheLoop->block_begin(),
3153 BE = TheLoop->block_end(); BI != BE; ++BI) {
3154 BasicBlock *BB = *BI;
3156 if (blockNeedsPredication(BB))
3159 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3160 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3161 SafePointes.insert(LI->getPointerOperand());
3162 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3163 SafePointes.insert(SI->getPointerOperand());
3167 // Collect the blocks that need predication.
3168 BasicBlock *Header = TheLoop->getHeader();
3169 for (Loop::block_iterator BI = TheLoop->block_begin(),
3170 BE = TheLoop->block_end(); BI != BE; ++BI) {
3171 BasicBlock *BB = *BI;
3173 // We don't support switch statements inside loops.
3174 if (!isa<BranchInst>(BB->getTerminator()))
3177 // We must be able to predicate all blocks that need to be predicated.
3178 if (blockNeedsPredication(BB)) {
3179 if (!blockCanBePredicated(BB, SafePointes))
3181 } else if (BB != Header && !canIfConvertPHINodes(BB))
3186 // We can if-convert this loop.
3190 bool LoopVectorizationLegality::canVectorize() {
3191 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3192 // be canonicalized.
3193 if (!TheLoop->getLoopPreheader())
3196 // We can only vectorize innermost loops.
3197 if (TheLoop->getSubLoopsVector().size())
3200 // We must have a single backedge.
3201 if (TheLoop->getNumBackEdges() != 1)
3204 // We must have a single exiting block.
3205 if (!TheLoop->getExitingBlock())
3208 // We need to have a loop header.
3209 DEBUG(dbgs() << "LV: Found a loop: " <<
3210 TheLoop->getHeader()->getName() << '\n');
3212 // Check if we can if-convert non-single-bb loops.
3213 unsigned NumBlocks = TheLoop->getNumBlocks();
3214 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3215 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3219 // ScalarEvolution needs to be able to find the exit count.
3220 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3221 if (ExitCount == SE->getCouldNotCompute()) {
3222 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3226 // Check if we can vectorize the instructions and CFG in this loop.
3227 if (!canVectorizeInstrs()) {
3228 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3232 // Go over each instruction and look at memory deps.
3233 if (!canVectorizeMemory()) {
3234 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3238 // Collect all of the variables that remain uniform after vectorization.
3239 collectLoopUniforms();
3241 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3242 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3245 // Okay! We can vectorize. At this point we don't have any other mem analysis
3246 // which may limit our maximum vectorization factor, so just return true with
3251 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3252 if (Ty->isPointerTy())
3253 return DL.getIntPtrType(Ty);
3255 // It is possible that char's or short's overflow when we ask for the loop's
3256 // trip count, work around this by changing the type size.
3257 if (Ty->getScalarSizeInBits() < 32)
3258 return Type::getInt32Ty(Ty->getContext());
3263 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3264 Ty0 = convertPointerToIntegerType(DL, Ty0);
3265 Ty1 = convertPointerToIntegerType(DL, Ty1);
3266 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3271 /// \brief Check that the instruction has outside loop users and is not an
3272 /// identified reduction variable.
3273 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3274 SmallPtrSet<Value *, 4> &Reductions) {
3275 // Reduction instructions are allowed to have exit users. All other
3276 // instructions must not have external users.
3277 if (!Reductions.count(Inst))
3278 //Check that all of the users of the loop are inside the BB.
3279 for (User *U : Inst->users()) {
3280 Instruction *UI = cast<Instruction>(U);
3281 // This user may be a reduction exit value.
3282 if (!TheLoop->contains(UI)) {
3283 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3290 bool LoopVectorizationLegality::canVectorizeInstrs() {
3291 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3292 BasicBlock *Header = TheLoop->getHeader();
3294 // Look for the attribute signaling the absence of NaNs.
3295 Function &F = *Header->getParent();
3296 if (F.hasFnAttribute("no-nans-fp-math"))
3297 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3298 AttributeSet::FunctionIndex,
3299 "no-nans-fp-math").getValueAsString() == "true";
3301 // For each block in the loop.
3302 for (Loop::block_iterator bb = TheLoop->block_begin(),
3303 be = TheLoop->block_end(); bb != be; ++bb) {
3305 // Scan the instructions in the block and look for hazards.
3306 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3309 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3310 Type *PhiTy = Phi->getType();
3311 // Check that this PHI type is allowed.
3312 if (!PhiTy->isIntegerTy() &&
3313 !PhiTy->isFloatingPointTy() &&
3314 !PhiTy->isPointerTy()) {
3315 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3319 // If this PHINode is not in the header block, then we know that we
3320 // can convert it to select during if-conversion. No need to check if
3321 // the PHIs in this block are induction or reduction variables.
3322 if (*bb != Header) {
3323 // Check that this instruction has no outside users or is an
3324 // identified reduction value with an outside user.
3325 if(!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3330 // We only allow if-converted PHIs with more than two incoming values.
3331 if (Phi->getNumIncomingValues() != 2) {
3332 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3336 // This is the value coming from the preheader.
3337 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3338 // Check if this is an induction variable.
3339 InductionKind IK = isInductionVariable(Phi);
3341 if (IK_NoInduction != IK) {
3342 // Get the widest type.
3344 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3346 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3348 // Int inductions are special because we only allow one IV.
3349 if (IK == IK_IntInduction) {
3350 // Use the phi node with the widest type as induction. Use the last
3351 // one if there are multiple (no good reason for doing this other
3352 // than it is expedient).
3353 if (!Induction || PhiTy == WidestIndTy)
3357 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3358 Inductions[Phi] = InductionInfo(StartValue, IK);
3360 // Until we explicitly handle the case of an induction variable with
3361 // an outside loop user we have to give up vectorizing this loop.
3362 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3368 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3369 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3372 if (AddReductionVar(Phi, RK_IntegerMult)) {
3373 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3376 if (AddReductionVar(Phi, RK_IntegerOr)) {
3377 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3380 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3381 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3384 if (AddReductionVar(Phi, RK_IntegerXor)) {
3385 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3388 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3389 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3392 if (AddReductionVar(Phi, RK_FloatMult)) {
3393 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3396 if (AddReductionVar(Phi, RK_FloatAdd)) {
3397 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3400 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3401 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3406 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3408 }// end of PHI handling
3410 // We still don't handle functions. However, we can ignore dbg intrinsic
3411 // calls and we do handle certain intrinsic and libm functions.
3412 CallInst *CI = dyn_cast<CallInst>(it);
3413 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3414 DEBUG(dbgs() << "LV: Found a call site.\n");
3418 // Check that the instruction return type is vectorizable.
3419 // Also, we can't vectorize extractelement instructions.
3420 if ((!VectorType::isValidElementType(it->getType()) &&
3421 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3422 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3426 // Check that the stored type is vectorizable.
3427 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3428 Type *T = ST->getValueOperand()->getType();
3429 if (!VectorType::isValidElementType(T))
3431 if (EnableMemAccessVersioning)
3432 collectStridedAcccess(ST);
3435 if (EnableMemAccessVersioning)
3436 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3437 collectStridedAcccess(LI);
3439 // Reduction instructions are allowed to have exit users.
3440 // All other instructions must not have external users.
3441 if (hasOutsideLoopUser(TheLoop, it, AllowedExit))
3449 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3450 if (Inductions.empty())
3457 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3458 /// return the induction operand of the gep pointer.
3459 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3460 const DataLayout *DL, Loop *Lp) {
3461 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3465 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3467 // Check that all of the gep indices are uniform except for our induction
3469 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3470 if (i != InductionOperand &&
3471 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3473 return GEP->getOperand(InductionOperand);
3476 ///\brief Look for a cast use of the passed value.
3477 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3478 Value *UniqueCast = nullptr;
3479 for (User *U : Ptr->users()) {
3480 CastInst *CI = dyn_cast<CastInst>(U);
3481 if (CI && CI->getType() == Ty) {
3491 ///\brief Get the stride of a pointer access in a loop.
3492 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3493 /// pointer to the Value, or null otherwise.
3494 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3495 const DataLayout *DL, Loop *Lp) {
3496 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3497 if (!PtrTy || PtrTy->isAggregateType())
3500 // Try to remove a gep instruction to make the pointer (actually index at this
3501 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3502 // pointer, otherwise, we are analyzing the index.
3503 Value *OrigPtr = Ptr;
3505 // The size of the pointer access.
3506 int64_t PtrAccessSize = 1;
3508 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3509 const SCEV *V = SE->getSCEV(Ptr);
3513 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3514 V = C->getOperand();
3516 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3520 V = S->getStepRecurrence(*SE);
3524 // Strip off the size of access multiplication if we are still analyzing the
3526 if (OrigPtr == Ptr) {
3527 DL->getTypeAllocSize(PtrTy->getElementType());
3528 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3529 if (M->getOperand(0)->getSCEVType() != scConstant)
3532 const APInt &APStepVal =
3533 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3535 // Huge step value - give up.
3536 if (APStepVal.getBitWidth() > 64)
3539 int64_t StepVal = APStepVal.getSExtValue();
3540 if (PtrAccessSize != StepVal)
3542 V = M->getOperand(1);
3547 Type *StripedOffRecurrenceCast = nullptr;
3548 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3549 StripedOffRecurrenceCast = C->getType();
3550 V = C->getOperand();
3553 // Look for the loop invariant symbolic value.
3554 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3558 Value *Stride = U->getValue();
3559 if (!Lp->isLoopInvariant(Stride))
3562 // If we have stripped off the recurrence cast we have to make sure that we
3563 // return the value that is used in this loop so that we can replace it later.
3564 if (StripedOffRecurrenceCast)
3565 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3570 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3571 Value *Ptr = nullptr;
3572 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3573 Ptr = LI->getPointerOperand();
3574 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3575 Ptr = SI->getPointerOperand();
3579 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3583 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3584 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3585 Strides[Ptr] = Stride;
3586 StrideSet.insert(Stride);
3589 void LoopVectorizationLegality::collectLoopUniforms() {
3590 // We now know that the loop is vectorizable!
3591 // Collect variables that will remain uniform after vectorization.
3592 std::vector<Value*> Worklist;
3593 BasicBlock *Latch = TheLoop->getLoopLatch();
3595 // Start with the conditional branch and walk up the block.
3596 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3598 // Also add all consecutive pointer values; these values will be uniform
3599 // after vectorization (and subsequent cleanup) and, until revectorization is
3600 // supported, all dependencies must also be uniform.
3601 for (Loop::block_iterator B = TheLoop->block_begin(),
3602 BE = TheLoop->block_end(); B != BE; ++B)
3603 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3605 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3606 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3608 while (Worklist.size()) {
3609 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3610 Worklist.pop_back();
3612 // Look at instructions inside this loop.
3613 // Stop when reaching PHI nodes.
3614 // TODO: we need to follow values all over the loop, not only in this block.
3615 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3618 // This is a known uniform.
3621 // Insert all operands.
3622 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3627 /// \brief Analyses memory accesses in a loop.
3629 /// Checks whether run time pointer checks are needed and builds sets for data
3630 /// dependence checking.
3631 class AccessAnalysis {
3633 /// \brief Read or write access location.
3634 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3635 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3637 /// \brief Set of potential dependent memory accesses.
3638 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3640 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3641 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3642 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3644 /// \brief Register a load and whether it is only read from.
3645 void addLoad(Value *Ptr, bool IsReadOnly) {
3646 Accesses.insert(MemAccessInfo(Ptr, false));
3648 ReadOnlyPtr.insert(Ptr);
3651 /// \brief Register a store.
3652 void addStore(Value *Ptr) {
3653 Accesses.insert(MemAccessInfo(Ptr, true));
3656 /// \brief Check whether we can check the pointers at runtime for
3657 /// non-intersection.
3658 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3659 unsigned &NumComparisons, ScalarEvolution *SE,
3660 Loop *TheLoop, ValueToValueMap &Strides,
3661 bool ShouldCheckStride = false);
3663 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3664 /// and builds sets of dependent accesses.
3665 void buildDependenceSets() {
3666 // Process read-write pointers first.
3667 processMemAccesses(false);
3668 // Next, process read pointers.
3669 processMemAccesses(true);
3672 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3674 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3675 void resetDepChecks() { CheckDeps.clear(); }
3677 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3680 typedef SetVector<MemAccessInfo> PtrAccessSet;
3681 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3683 /// \brief Go over all memory access or only the deferred ones if
3684 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3685 /// and build sets of dependency check candidates.
3686 void processMemAccesses(bool UseDeferred);
3688 /// Set of all accesses.
3689 PtrAccessSet Accesses;
3691 /// Set of access to check after all writes have been processed.
3692 PtrAccessSet DeferredAccesses;
3694 /// Map of pointers to last access encountered.
3695 UnderlyingObjToAccessMap ObjToLastAccess;
3697 /// Set of accesses that need a further dependence check.
3698 MemAccessInfoSet CheckDeps;
3700 /// Set of pointers that are read only.
3701 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3703 /// Set of underlying objects already written to.
3704 SmallPtrSet<Value*, 16> WriteObjects;
3706 const DataLayout *DL;
3708 /// Sets of potentially dependent accesses - members of one set share an
3709 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3710 /// dependence check.
3711 DepCandidates &DepCands;
3713 bool AreAllWritesIdentified;
3714 bool AreAllReadsIdentified;
3715 bool IsRTCheckNeeded;
3718 } // end anonymous namespace
3720 /// \brief Check whether a pointer can participate in a runtime bounds check.
3721 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3723 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3724 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3728 return AR->isAffine();
3731 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3732 /// the address space.
3733 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3734 const Loop *Lp, ValueToValueMap &StridesMap);
3736 bool AccessAnalysis::canCheckPtrAtRT(
3737 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3738 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3739 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3740 // Find pointers with computable bounds. We are going to use this information
3741 // to place a runtime bound check.
3742 unsigned NumReadPtrChecks = 0;
3743 unsigned NumWritePtrChecks = 0;
3744 bool CanDoRT = true;
3746 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3747 // We assign consecutive id to access from different dependence sets.
3748 // Accesses within the same set don't need a runtime check.
3749 unsigned RunningDepId = 1;
3750 DenseMap<Value *, unsigned> DepSetId;
3752 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3754 const MemAccessInfo &Access = *AI;
3755 Value *Ptr = Access.getPointer();
3756 bool IsWrite = Access.getInt();
3758 // Just add write checks if we have both.
3759 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3763 ++NumWritePtrChecks;
3767 if (hasComputableBounds(SE, StridesMap, Ptr) &&
3768 // When we run after a failing dependency check we have to make sure we
3769 // don't have wrapping pointers.
3770 (!ShouldCheckStride ||
3771 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
3772 // The id of the dependence set.
3775 if (IsDepCheckNeeded) {
3776 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
3777 unsigned &LeaderId = DepSetId[Leader];
3779 LeaderId = RunningDepId++;
3782 // Each access has its own dependence set.
3783 DepId = RunningDepId++;
3785 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
3787 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
3793 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
3794 NumComparisons = 0; // Only one dependence set.
3796 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
3797 NumWritePtrChecks - 1));
3800 // If the pointers that we would use for the bounds comparison have different
3801 // address spaces, assume the values aren't directly comparable, so we can't
3802 // use them for the runtime check. We also have to assume they could
3803 // overlap. In the future there should be metadata for whether address spaces
3805 unsigned NumPointers = RtCheck.Pointers.size();
3806 for (unsigned i = 0; i < NumPointers; ++i) {
3807 for (unsigned j = i + 1; j < NumPointers; ++j) {
3808 // Only need to check pointers between two different dependency sets.
3809 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
3812 Value *PtrI = RtCheck.Pointers[i];
3813 Value *PtrJ = RtCheck.Pointers[j];
3815 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
3816 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
3818 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
3819 " different address spaces\n");
3828 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
3829 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
3832 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
3833 // We process the set twice: first we process read-write pointers, last we
3834 // process read-only pointers. This allows us to skip dependence tests for
3835 // read-only pointers.
3837 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
3838 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
3839 const MemAccessInfo &Access = *AI;
3840 Value *Ptr = Access.getPointer();
3841 bool IsWrite = Access.getInt();
3843 DepCands.insert(Access);
3845 // Memorize read-only pointers for later processing and skip them in the
3846 // first round (they need to be checked after we have seen all write
3847 // pointers). Note: we also mark pointer that are not consecutive as
3848 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
3849 // second check for "!IsWrite".
3850 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
3851 if (!UseDeferred && IsReadOnlyPtr) {
3852 DeferredAccesses.insert(Access);
3856 bool NeedDepCheck = false;
3857 // Check whether there is the possibility of dependency because of
3858 // underlying objects being the same.
3859 typedef SmallVector<Value*, 16> ValueVector;
3860 ValueVector TempObjects;
3861 GetUnderlyingObjects(Ptr, TempObjects, DL);
3862 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
3864 Value *UnderlyingObj = *UI;
3866 // If this is a write then it needs to be an identified object. If this a
3867 // read and all writes (so far) are identified function scope objects we
3868 // don't need an identified underlying object but only an Argument (the
3869 // next write is going to invalidate this assumption if it is
3871 // This is a micro-optimization for the case where all writes are
3872 // identified and we have one argument pointer.
3873 // Otherwise, we do need a runtime check.
3874 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
3875 (!IsWrite && (!AreAllWritesIdentified ||
3876 !isa<Argument>(UnderlyingObj)) &&
3877 !isIdentifiedObject(UnderlyingObj))) {
3878 DEBUG(dbgs() << "LV: Found an unidentified " <<
3879 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
3881 IsRTCheckNeeded = (IsRTCheckNeeded ||
3882 !isIdentifiedObject(UnderlyingObj) ||
3883 !AreAllReadsIdentified);
3886 AreAllWritesIdentified = false;
3888 AreAllReadsIdentified = false;
3891 // If this is a write - check other reads and writes for conflicts. If
3892 // this is a read only check other writes for conflicts (but only if there
3893 // is no other write to the ptr - this is an optimization to catch "a[i] =
3894 // a[i] + " without having to do a dependence check).
3895 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
3896 NeedDepCheck = true;
3899 WriteObjects.insert(UnderlyingObj);
3901 // Create sets of pointers connected by shared underlying objects.
3902 UnderlyingObjToAccessMap::iterator Prev =
3903 ObjToLastAccess.find(UnderlyingObj);
3904 if (Prev != ObjToLastAccess.end())
3905 DepCands.unionSets(Access, Prev->second);
3907 ObjToLastAccess[UnderlyingObj] = Access;
3911 CheckDeps.insert(Access);
3916 /// \brief Checks memory dependences among accesses to the same underlying
3917 /// object to determine whether there vectorization is legal or not (and at
3918 /// which vectorization factor).
3920 /// This class works under the assumption that we already checked that memory
3921 /// locations with different underlying pointers are "must-not alias".
3922 /// We use the ScalarEvolution framework to symbolically evalutate access
3923 /// functions pairs. Since we currently don't restructure the loop we can rely
3924 /// on the program order of memory accesses to determine their safety.
3925 /// At the moment we will only deem accesses as safe for:
3926 /// * A negative constant distance assuming program order.
3928 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
3929 /// a[i] = tmp; y = a[i];
3931 /// The latter case is safe because later checks guarantuee that there can't
3932 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
3933 /// the same variable: a header phi can only be an induction or a reduction, a
3934 /// reduction can't have a memory sink, an induction can't have a memory
3935 /// source). This is important and must not be violated (or we have to
3936 /// resort to checking for cycles through memory).
3938 /// * A positive constant distance assuming program order that is bigger
3939 /// than the biggest memory access.
3941 /// tmp = a[i] OR b[i] = x
3942 /// a[i+2] = tmp y = b[i+2];
3944 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
3946 /// * Zero distances and all accesses have the same size.
3948 class MemoryDepChecker {
3950 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3951 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3953 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
3954 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
3955 ShouldRetryWithRuntimeCheck(false) {}
3957 /// \brief Register the location (instructions are given increasing numbers)
3958 /// of a write access.
3959 void addAccess(StoreInst *SI) {
3960 Value *Ptr = SI->getPointerOperand();
3961 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
3962 InstMap.push_back(SI);
3966 /// \brief Register the location (instructions are given increasing numbers)
3967 /// of a write access.
3968 void addAccess(LoadInst *LI) {
3969 Value *Ptr = LI->getPointerOperand();
3970 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
3971 InstMap.push_back(LI);
3975 /// \brief Check whether the dependencies between the accesses are safe.
3977 /// Only checks sets with elements in \p CheckDeps.
3978 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
3979 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
3981 /// \brief The maximum number of bytes of a vector register we can vectorize
3982 /// the accesses safely with.
3983 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
3985 /// \brief In same cases when the dependency check fails we can still
3986 /// vectorize the loop with a dynamic array access check.
3987 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
3990 ScalarEvolution *SE;
3991 const DataLayout *DL;
3992 const Loop *InnermostLoop;
3994 /// \brief Maps access locations (ptr, read/write) to program order.
3995 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
3997 /// \brief Memory access instructions in program order.
3998 SmallVector<Instruction *, 16> InstMap;
4000 /// \brief The program order index to be used for the next instruction.
4003 // We can access this many bytes in parallel safely.
4004 unsigned MaxSafeDepDistBytes;
4006 /// \brief If we see a non-constant dependence distance we can still try to
4007 /// vectorize this loop with runtime checks.
4008 bool ShouldRetryWithRuntimeCheck;
4010 /// \brief Check whether there is a plausible dependence between the two
4013 /// Access \p A must happen before \p B in program order. The two indices
4014 /// identify the index into the program order map.
4016 /// This function checks whether there is a plausible dependence (or the
4017 /// absence of such can't be proved) between the two accesses. If there is a
4018 /// plausible dependence but the dependence distance is bigger than one
4019 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4020 /// distance is smaller than any other distance encountered so far).
4021 /// Otherwise, this function returns true signaling a possible dependence.
4022 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4023 const MemAccessInfo &B, unsigned BIdx,
4024 ValueToValueMap &Strides);
4026 /// \brief Check whether the data dependence could prevent store-load
4028 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4031 } // end anonymous namespace
4033 static bool isInBoundsGep(Value *Ptr) {
4034 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4035 return GEP->isInBounds();
4039 /// \brief Check whether the access through \p Ptr has a constant stride.
4040 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4041 const Loop *Lp, ValueToValueMap &StridesMap) {
4042 const Type *Ty = Ptr->getType();
4043 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4045 // Make sure that the pointer does not point to aggregate types.
4046 const PointerType *PtrTy = cast<PointerType>(Ty);
4047 if (PtrTy->getElementType()->isAggregateType()) {
4048 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4053 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4055 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4057 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4058 << *Ptr << " SCEV: " << *PtrScev << "\n");
4062 // The accesss function must stride over the innermost loop.
4063 if (Lp != AR->getLoop()) {
4064 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4065 *Ptr << " SCEV: " << *PtrScev << "\n");
4068 // The address calculation must not wrap. Otherwise, a dependence could be
4070 // An inbounds getelementptr that is a AddRec with a unit stride
4071 // cannot wrap per definition. The unit stride requirement is checked later.
4072 // An getelementptr without an inbounds attribute and unit stride would have
4073 // to access the pointer value "0" which is undefined behavior in address
4074 // space 0, therefore we can also vectorize this case.
4075 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4076 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4077 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4078 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4079 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4080 << *Ptr << " SCEV: " << *PtrScev << "\n");
4084 // Check the step is constant.
4085 const SCEV *Step = AR->getStepRecurrence(*SE);
4087 // Calculate the pointer stride and check if it is consecutive.
4088 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4090 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4091 " SCEV: " << *PtrScev << "\n");
4095 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4096 const APInt &APStepVal = C->getValue()->getValue();
4098 // Huge step value - give up.
4099 if (APStepVal.getBitWidth() > 64)
4102 int64_t StepVal = APStepVal.getSExtValue();
4105 int64_t Stride = StepVal / Size;
4106 int64_t Rem = StepVal % Size;
4110 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4111 // know we can't "wrap around the address space". In case of address space
4112 // zero we know that this won't happen without triggering undefined behavior.
4113 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4114 Stride != 1 && Stride != -1)
4120 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4121 unsigned TypeByteSize) {
4122 // If loads occur at a distance that is not a multiple of a feasible vector
4123 // factor store-load forwarding does not take place.
4124 // Positive dependences might cause troubles because vectorizing them might
4125 // prevent store-load forwarding making vectorized code run a lot slower.
4126 // a[i] = a[i-3] ^ a[i-8];
4127 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4128 // hence on your typical architecture store-load forwarding does not take
4129 // place. Vectorizing in such cases does not make sense.
4130 // Store-load forwarding distance.
4131 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4132 // Maximum vector factor.
4133 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4134 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4135 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4137 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4139 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4140 MaxVFWithoutSLForwardIssues = (vf >>=1);
4145 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4146 DEBUG(dbgs() << "LV: Distance " << Distance <<
4147 " that could cause a store-load forwarding conflict\n");
4151 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4152 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4153 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4157 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4158 const MemAccessInfo &B, unsigned BIdx,
4159 ValueToValueMap &Strides) {
4160 assert (AIdx < BIdx && "Must pass arguments in program order");
4162 Value *APtr = A.getPointer();
4163 Value *BPtr = B.getPointer();
4164 bool AIsWrite = A.getInt();
4165 bool BIsWrite = B.getInt();
4167 // Two reads are independent.
4168 if (!AIsWrite && !BIsWrite)
4171 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4172 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4174 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4175 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4177 const SCEV *Src = AScev;
4178 const SCEV *Sink = BScev;
4180 // If the induction step is negative we have to invert source and sink of the
4182 if (StrideAPtr < 0) {
4185 std::swap(APtr, BPtr);
4186 std::swap(Src, Sink);
4187 std::swap(AIsWrite, BIsWrite);
4188 std::swap(AIdx, BIdx);
4189 std::swap(StrideAPtr, StrideBPtr);
4192 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4194 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4195 << "(Induction step: " << StrideAPtr << ")\n");
4196 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4197 << *InstMap[BIdx] << ": " << *Dist << "\n");
4199 // Need consecutive accesses. We don't want to vectorize
4200 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4201 // the address space.
4202 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4203 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4207 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4209 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4210 ShouldRetryWithRuntimeCheck = true;
4214 Type *ATy = APtr->getType()->getPointerElementType();
4215 Type *BTy = BPtr->getType()->getPointerElementType();
4216 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4218 // Negative distances are not plausible dependencies.
4219 const APInt &Val = C->getValue()->getValue();
4220 if (Val.isNegative()) {
4221 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4222 if (IsTrueDataDependence &&
4223 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4227 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4231 // Write to the same location with the same size.
4232 // Could be improved to assert type sizes are the same (i32 == float, etc).
4236 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4240 assert(Val.isStrictlyPositive() && "Expect a positive value");
4242 // Positive distance bigger than max vectorization factor.
4245 "LV: ReadWrite-Write positive dependency with different types\n");
4249 unsigned Distance = (unsigned) Val.getZExtValue();
4251 // Bail out early if passed-in parameters make vectorization not feasible.
4252 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4253 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4255 // The distance must be bigger than the size needed for a vectorized version
4256 // of the operation and the size of the vectorized operation must not be
4257 // bigger than the currrent maximum size.
4258 if (Distance < 2*TypeByteSize ||
4259 2*TypeByteSize > MaxSafeDepDistBytes ||
4260 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4261 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4262 << Val.getSExtValue() << '\n');
4266 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4267 Distance : MaxSafeDepDistBytes;
4269 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4270 if (IsTrueDataDependence &&
4271 couldPreventStoreLoadForward(Distance, TypeByteSize))
4274 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4275 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4280 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4281 MemAccessInfoSet &CheckDeps,
4282 ValueToValueMap &Strides) {
4284 MaxSafeDepDistBytes = -1U;
4285 while (!CheckDeps.empty()) {
4286 MemAccessInfo CurAccess = *CheckDeps.begin();
4288 // Get the relevant memory access set.
4289 EquivalenceClasses<MemAccessInfo>::iterator I =
4290 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4292 // Check accesses within this set.
4293 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4294 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4296 // Check every access pair.
4298 CheckDeps.erase(*AI);
4299 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4301 // Check every accessing instruction pair in program order.
4302 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4303 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4304 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4305 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4306 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4308 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4319 bool LoopVectorizationLegality::canVectorizeMemory() {
4321 typedef SmallVector<Value*, 16> ValueVector;
4322 typedef SmallPtrSet<Value*, 16> ValueSet;
4324 // Holds the Load and Store *instructions*.
4328 // Holds all the different accesses in the loop.
4329 unsigned NumReads = 0;
4330 unsigned NumReadWrites = 0;
4332 PtrRtCheck.Pointers.clear();
4333 PtrRtCheck.Need = false;
4335 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4336 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4339 for (Loop::block_iterator bb = TheLoop->block_begin(),
4340 be = TheLoop->block_end(); bb != be; ++bb) {
4342 // Scan the BB and collect legal loads and stores.
4343 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4346 // If this is a load, save it. If this instruction can read from memory
4347 // but is not a load, then we quit. Notice that we don't handle function
4348 // calls that read or write.
4349 if (it->mayReadFromMemory()) {
4350 // Many math library functions read the rounding mode. We will only
4351 // vectorize a loop if it contains known function calls that don't set
4352 // the flag. Therefore, it is safe to ignore this read from memory.
4353 CallInst *Call = dyn_cast<CallInst>(it);
4354 if (Call && getIntrinsicIDForCall(Call, TLI))
4357 LoadInst *Ld = dyn_cast<LoadInst>(it);
4358 if (!Ld) return false;
4359 if (!Ld->isSimple() && !IsAnnotatedParallel) {
4360 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4364 Loads.push_back(Ld);
4365 DepChecker.addAccess(Ld);
4369 // Save 'store' instructions. Abort if other instructions write to memory.
4370 if (it->mayWriteToMemory()) {
4371 StoreInst *St = dyn_cast<StoreInst>(it);
4372 if (!St) return false;
4373 if (!St->isSimple() && !IsAnnotatedParallel) {
4374 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4378 Stores.push_back(St);
4379 DepChecker.addAccess(St);
4384 // Now we have two lists that hold the loads and the stores.
4385 // Next, we find the pointers that they use.
4387 // Check if we see any stores. If there are no stores, then we don't
4388 // care if the pointers are *restrict*.
4389 if (!Stores.size()) {
4390 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4394 AccessAnalysis::DepCandidates DependentAccesses;
4395 AccessAnalysis Accesses(DL, DependentAccesses);
4397 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4398 // multiple times on the same object. If the ptr is accessed twice, once
4399 // for read and once for write, it will only appear once (on the write
4400 // list). This is okay, since we are going to check for conflicts between
4401 // writes and between reads and writes, but not between reads and reads.
4404 ValueVector::iterator I, IE;
4405 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4406 StoreInst *ST = cast<StoreInst>(*I);
4407 Value* Ptr = ST->getPointerOperand();
4409 if (isUniform(Ptr)) {
4410 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4414 // If we did *not* see this pointer before, insert it to the read-write
4415 // list. At this phase it is only a 'write' list.
4416 if (Seen.insert(Ptr)) {
4418 Accesses.addStore(Ptr);
4422 if (IsAnnotatedParallel) {
4424 << "LV: A loop annotated parallel, ignore memory dependency "
4429 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4430 LoadInst *LD = cast<LoadInst>(*I);
4431 Value* Ptr = LD->getPointerOperand();
4432 // If we did *not* see this pointer before, insert it to the
4433 // read list. If we *did* see it before, then it is already in
4434 // the read-write list. This allows us to vectorize expressions
4435 // such as A[i] += x; Because the address of A[i] is a read-write
4436 // pointer. This only works if the index of A[i] is consecutive.
4437 // If the address of i is unknown (for example A[B[i]]) then we may
4438 // read a few words, modify, and write a few words, and some of the
4439 // words may be written to the same address.
4440 bool IsReadOnlyPtr = false;
4441 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4443 IsReadOnlyPtr = true;
4445 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4448 // If we write (or read-write) to a single destination and there are no
4449 // other reads in this loop then is it safe to vectorize.
4450 if (NumReadWrites == 1 && NumReads == 0) {
4451 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4455 // Build dependence sets and check whether we need a runtime pointer bounds
4457 Accesses.buildDependenceSets();
4458 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4460 // Find pointers with computable bounds. We are going to use this information
4461 // to place a runtime bound check.
4462 unsigned NumComparisons = 0;
4463 bool CanDoRT = false;
4465 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4468 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4469 " pointer comparisons.\n");
4471 // If we only have one set of dependences to check pointers among we don't
4472 // need a runtime check.
4473 if (NumComparisons == 0 && NeedRTCheck)
4474 NeedRTCheck = false;
4476 // Check that we did not collect too many pointers or found an unsizeable
4478 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4484 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4487 if (NeedRTCheck && !CanDoRT) {
4488 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4489 "the array bounds.\n");
4494 PtrRtCheck.Need = NeedRTCheck;
4496 bool CanVecMem = true;
4497 if (Accesses.isDependencyCheckNeeded()) {
4498 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4499 CanVecMem = DepChecker.areDepsSafe(
4500 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4501 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4503 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4504 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4507 // Clear the dependency checks. We assume they are not needed.
4508 Accesses.resetDepChecks();
4511 PtrRtCheck.Need = true;
4513 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4514 TheLoop, Strides, true);
4515 // Check that we did not collect too many pointers or found an unsizeable
4517 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4518 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4527 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4528 " need a runtime memory check.\n");
4533 static bool hasMultipleUsesOf(Instruction *I,
4534 SmallPtrSet<Instruction *, 8> &Insts) {
4535 unsigned NumUses = 0;
4536 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4537 if (Insts.count(dyn_cast<Instruction>(*Use)))
4546 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4547 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4548 if (!Set.count(dyn_cast<Instruction>(*Use)))
4553 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4554 ReductionKind Kind) {
4555 if (Phi->getNumIncomingValues() != 2)
4558 // Reduction variables are only found in the loop header block.
4559 if (Phi->getParent() != TheLoop->getHeader())
4562 // Obtain the reduction start value from the value that comes from the loop
4564 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4566 // ExitInstruction is the single value which is used outside the loop.
4567 // We only allow for a single reduction value to be used outside the loop.
4568 // This includes users of the reduction, variables (which form a cycle
4569 // which ends in the phi node).
4570 Instruction *ExitInstruction = nullptr;
4571 // Indicates that we found a reduction operation in our scan.
4572 bool FoundReduxOp = false;
4574 // We start with the PHI node and scan for all of the users of this
4575 // instruction. All users must be instructions that can be used as reduction
4576 // variables (such as ADD). We must have a single out-of-block user. The cycle
4577 // must include the original PHI.
4578 bool FoundStartPHI = false;
4580 // To recognize min/max patterns formed by a icmp select sequence, we store
4581 // the number of instruction we saw from the recognized min/max pattern,
4582 // to make sure we only see exactly the two instructions.
4583 unsigned NumCmpSelectPatternInst = 0;
4584 ReductionInstDesc ReduxDesc(false, nullptr);
4586 SmallPtrSet<Instruction *, 8> VisitedInsts;
4587 SmallVector<Instruction *, 8> Worklist;
4588 Worklist.push_back(Phi);
4589 VisitedInsts.insert(Phi);
4591 // A value in the reduction can be used:
4592 // - By the reduction:
4593 // - Reduction operation:
4594 // - One use of reduction value (safe).
4595 // - Multiple use of reduction value (not safe).
4597 // - All uses of the PHI must be the reduction (safe).
4598 // - Otherwise, not safe.
4599 // - By one instruction outside of the loop (safe).
4600 // - By further instructions outside of the loop (not safe).
4601 // - By an instruction that is not part of the reduction (not safe).
4603 // * An instruction type other than PHI or the reduction operation.
4604 // * A PHI in the header other than the initial PHI.
4605 while (!Worklist.empty()) {
4606 Instruction *Cur = Worklist.back();
4607 Worklist.pop_back();
4610 // If the instruction has no users then this is a broken chain and can't be
4611 // a reduction variable.
4612 if (Cur->use_empty())
4615 bool IsAPhi = isa<PHINode>(Cur);
4617 // A header PHI use other than the original PHI.
4618 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4621 // Reductions of instructions such as Div, and Sub is only possible if the
4622 // LHS is the reduction variable.
4623 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4624 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4625 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4628 // Any reduction instruction must be of one of the allowed kinds.
4629 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4630 if (!ReduxDesc.IsReduction)
4633 // A reduction operation must only have one use of the reduction value.
4634 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4635 hasMultipleUsesOf(Cur, VisitedInsts))
4638 // All inputs to a PHI node must be a reduction value.
4639 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4642 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4643 isa<SelectInst>(Cur)))
4644 ++NumCmpSelectPatternInst;
4645 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4646 isa<SelectInst>(Cur)))
4647 ++NumCmpSelectPatternInst;
4649 // Check whether we found a reduction operator.
4650 FoundReduxOp |= !IsAPhi;
4652 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4653 // onto the stack. This way we are going to have seen all inputs to PHI
4654 // nodes once we get to them.
4655 SmallVector<Instruction *, 8> NonPHIs;
4656 SmallVector<Instruction *, 8> PHIs;
4657 for (User *U : Cur->users()) {
4658 Instruction *UI = cast<Instruction>(U);
4660 // Check if we found the exit user.
4661 BasicBlock *Parent = UI->getParent();
4662 if (!TheLoop->contains(Parent)) {
4663 // Exit if you find multiple outside users or if the header phi node is
4664 // being used. In this case the user uses the value of the previous
4665 // iteration, in which case we would loose "VF-1" iterations of the
4666 // reduction operation if we vectorize.
4667 if (ExitInstruction != nullptr || Cur == Phi)
4670 // The instruction used by an outside user must be the last instruction
4671 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4672 // operations on the value.
4673 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4676 ExitInstruction = Cur;
4680 // Process instructions only once (termination). Each reduction cycle
4681 // value must only be used once, except by phi nodes and min/max
4682 // reductions which are represented as a cmp followed by a select.
4683 ReductionInstDesc IgnoredVal(false, nullptr);
4684 if (VisitedInsts.insert(UI)) {
4685 if (isa<PHINode>(UI))
4688 NonPHIs.push_back(UI);
4689 } else if (!isa<PHINode>(UI) &&
4690 ((!isa<FCmpInst>(UI) &&
4691 !isa<ICmpInst>(UI) &&
4692 !isa<SelectInst>(UI)) ||
4693 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4696 // Remember that we completed the cycle.
4698 FoundStartPHI = true;
4700 Worklist.append(PHIs.begin(), PHIs.end());
4701 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4704 // This means we have seen one but not the other instruction of the
4705 // pattern or more than just a select and cmp.
4706 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4707 NumCmpSelectPatternInst != 2)
4710 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4713 // We found a reduction var if we have reached the original phi node and we
4714 // only have a single instruction with out-of-loop users.
4716 // This instruction is allowed to have out-of-loop users.
4717 AllowedExit.insert(ExitInstruction);
4719 // Save the description of this reduction variable.
4720 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4721 ReduxDesc.MinMaxKind);
4722 Reductions[Phi] = RD;
4723 // We've ended the cycle. This is a reduction variable if we have an
4724 // outside user and it has a binary op.
4729 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4730 /// pattern corresponding to a min(X, Y) or max(X, Y).
4731 LoopVectorizationLegality::ReductionInstDesc
4732 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4733 ReductionInstDesc &Prev) {
4735 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4736 "Expect a select instruction");
4737 Instruction *Cmp = nullptr;
4738 SelectInst *Select = nullptr;
4740 // We must handle the select(cmp()) as a single instruction. Advance to the
4742 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4743 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4744 return ReductionInstDesc(false, I);
4745 return ReductionInstDesc(Select, Prev.MinMaxKind);
4748 // Only handle single use cases for now.
4749 if (!(Select = dyn_cast<SelectInst>(I)))
4750 return ReductionInstDesc(false, I);
4751 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4752 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4753 return ReductionInstDesc(false, I);
4754 if (!Cmp->hasOneUse())
4755 return ReductionInstDesc(false, I);
4760 // Look for a min/max pattern.
4761 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4762 return ReductionInstDesc(Select, MRK_UIntMin);
4763 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4764 return ReductionInstDesc(Select, MRK_UIntMax);
4765 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4766 return ReductionInstDesc(Select, MRK_SIntMax);
4767 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4768 return ReductionInstDesc(Select, MRK_SIntMin);
4769 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4770 return ReductionInstDesc(Select, MRK_FloatMin);
4771 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4772 return ReductionInstDesc(Select, MRK_FloatMax);
4773 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4774 return ReductionInstDesc(Select, MRK_FloatMin);
4775 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4776 return ReductionInstDesc(Select, MRK_FloatMax);
4778 return ReductionInstDesc(false, I);
4781 LoopVectorizationLegality::ReductionInstDesc
4782 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4784 ReductionInstDesc &Prev) {
4785 bool FP = I->getType()->isFloatingPointTy();
4786 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
4787 switch (I->getOpcode()) {
4789 return ReductionInstDesc(false, I);
4790 case Instruction::PHI:
4791 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4792 Kind != RK_FloatMinMax))
4793 return ReductionInstDesc(false, I);
4794 return ReductionInstDesc(I, Prev.MinMaxKind);
4795 case Instruction::Sub:
4796 case Instruction::Add:
4797 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4798 case Instruction::Mul:
4799 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4800 case Instruction::And:
4801 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4802 case Instruction::Or:
4803 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4804 case Instruction::Xor:
4805 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4806 case Instruction::FMul:
4807 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4808 case Instruction::FAdd:
4809 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4810 case Instruction::FCmp:
4811 case Instruction::ICmp:
4812 case Instruction::Select:
4813 if (Kind != RK_IntegerMinMax &&
4814 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4815 return ReductionInstDesc(false, I);
4816 return isMinMaxSelectCmpPattern(I, Prev);
4820 LoopVectorizationLegality::InductionKind
4821 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
4822 Type *PhiTy = Phi->getType();
4823 // We only handle integer and pointer inductions variables.
4824 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4825 return IK_NoInduction;
4827 // Check that the PHI is consecutive.
4828 const SCEV *PhiScev = SE->getSCEV(Phi);
4829 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4831 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4832 return IK_NoInduction;
4834 const SCEV *Step = AR->getStepRecurrence(*SE);
4836 // Integer inductions need to have a stride of one.
4837 if (PhiTy->isIntegerTy()) {
4839 return IK_IntInduction;
4840 if (Step->isAllOnesValue())
4841 return IK_ReverseIntInduction;
4842 return IK_NoInduction;
4845 // Calculate the pointer stride and check if it is consecutive.
4846 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4848 return IK_NoInduction;
4850 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4851 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
4852 if (C->getValue()->equalsInt(Size))
4853 return IK_PtrInduction;
4854 else if (C->getValue()->equalsInt(0 - Size))
4855 return IK_ReversePtrInduction;
4857 return IK_NoInduction;
4860 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4861 Value *In0 = const_cast<Value*>(V);
4862 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4866 return Inductions.count(PN);
4869 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4870 assert(TheLoop->contains(BB) && "Unknown block used");
4872 // Blocks that do not dominate the latch need predication.
4873 BasicBlock* Latch = TheLoop->getLoopLatch();
4874 return !DT->dominates(BB, Latch);
4877 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4878 SmallPtrSet<Value *, 8>& SafePtrs) {
4879 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4880 // We might be able to hoist the load.
4881 if (it->mayReadFromMemory()) {
4882 LoadInst *LI = dyn_cast<LoadInst>(it);
4883 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
4887 // We don't predicate stores at the moment.
4888 if (it->mayWriteToMemory()) {
4889 StoreInst *SI = dyn_cast<StoreInst>(it);
4890 // We only support predication of stores in basic blocks with one
4892 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
4893 !SafePtrs.count(SI->getPointerOperand()) ||
4894 !SI->getParent()->getSinglePredecessor())
4900 // Check that we don't have a constant expression that can trap as operand.
4901 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4903 if (Constant *C = dyn_cast<Constant>(*OI))
4908 // The instructions below can trap.
4909 switch (it->getOpcode()) {
4911 case Instruction::UDiv:
4912 case Instruction::SDiv:
4913 case Instruction::URem:
4914 case Instruction::SRem:
4922 LoopVectorizationCostModel::VectorizationFactor
4923 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
4925 bool ForceVectorization) {
4926 // Width 1 means no vectorize
4927 VectorizationFactor Factor = { 1U, 0U };
4928 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4929 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4933 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
4934 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4938 // Find the trip count.
4939 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
4940 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4942 unsigned WidestType = getWidestType();
4943 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4944 unsigned MaxSafeDepDist = -1U;
4945 if (Legal->getMaxSafeDepDistBytes() != -1U)
4946 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4947 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4948 WidestRegister : MaxSafeDepDist);
4949 unsigned MaxVectorSize = WidestRegister / WidestType;
4950 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4951 DEBUG(dbgs() << "LV: The Widest register is: "
4952 << WidestRegister << " bits.\n");
4954 if (MaxVectorSize == 0) {
4955 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4959 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
4960 " into one vector!");
4962 unsigned VF = MaxVectorSize;
4964 // If we optimize the program for size, avoid creating the tail loop.
4966 // If we are unable to calculate the trip count then don't try to vectorize.
4968 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4972 // Find the maximum SIMD width that can fit within the trip count.
4973 VF = TC % MaxVectorSize;
4978 // If the trip count that we found modulo the vectorization factor is not
4979 // zero then we require a tail.
4981 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4987 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4988 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4990 Factor.Width = UserVF;
4994 float Cost = expectedCost(1);
4996 const float ScalarCost = Cost;
4999 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5001 // Ignore scalar width, because the user explicitly wants vectorization.
5002 if (ForceVectorization && VF > 1) {
5004 Cost = expectedCost(Width) / (float)Width;
5007 for (unsigned i=2; i <= VF; i*=2) {
5008 // Notice that the vector loop needs to be executed less times, so
5009 // we need to divide the cost of the vector loops by the width of
5010 // the vector elements.
5011 float VectorCost = expectedCost(i) / (float)i;
5012 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5013 (int)VectorCost << ".\n");
5014 if (VectorCost < Cost) {
5020 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5021 << "LV: Vectorization seems to be not beneficial, "
5022 << "but was forced by a user.\n");
5023 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5024 Factor.Width = Width;
5025 Factor.Cost = Width * Cost;
5029 unsigned LoopVectorizationCostModel::getWidestType() {
5030 unsigned MaxWidth = 8;
5033 for (Loop::block_iterator bb = TheLoop->block_begin(),
5034 be = TheLoop->block_end(); bb != be; ++bb) {
5035 BasicBlock *BB = *bb;
5037 // For each instruction in the loop.
5038 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5039 Type *T = it->getType();
5041 // Only examine Loads, Stores and PHINodes.
5042 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5045 // Examine PHI nodes that are reduction variables.
5046 if (PHINode *PN = dyn_cast<PHINode>(it))
5047 if (!Legal->getReductionVars()->count(PN))
5050 // Examine the stored values.
5051 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5052 T = ST->getValueOperand()->getType();
5054 // Ignore loaded pointer types and stored pointer types that are not
5055 // consecutive. However, we do want to take consecutive stores/loads of
5056 // pointer vectors into account.
5057 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5060 MaxWidth = std::max(MaxWidth,
5061 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5069 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5072 unsigned LoopCost) {
5074 // -- The unroll heuristics --
5075 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5076 // There are many micro-architectural considerations that we can't predict
5077 // at this level. For example frontend pressure (on decode or fetch) due to
5078 // code size, or the number and capabilities of the execution ports.
5080 // We use the following heuristics to select the unroll factor:
5081 // 1. If the code has reductions the we unroll in order to break the cross
5082 // iteration dependency.
5083 // 2. If the loop is really small then we unroll in order to reduce the loop
5085 // 3. We don't unroll if we think that we will spill registers to memory due
5086 // to the increased register pressure.
5088 // Use the user preference, unless 'auto' is selected.
5092 // When we optimize for size we don't unroll.
5096 // We used the distance for the unroll factor.
5097 if (Legal->getMaxSafeDepDistBytes() != -1U)
5100 // Do not unroll loops with a relatively small trip count.
5101 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5102 TheLoop->getLoopLatch());
5103 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5106 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5107 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5111 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5112 TargetNumRegisters = ForceTargetNumScalarRegs;
5114 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5115 TargetNumRegisters = ForceTargetNumVectorRegs;
5118 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5119 // We divide by these constants so assume that we have at least one
5120 // instruction that uses at least one register.
5121 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5122 R.NumInstructions = std::max(R.NumInstructions, 1U);
5124 // We calculate the unroll factor using the following formula.
5125 // Subtract the number of loop invariants from the number of available
5126 // registers. These registers are used by all of the unrolled instances.
5127 // Next, divide the remaining registers by the number of registers that is
5128 // required by the loop, in order to estimate how many parallel instances
5129 // fit without causing spills. All of this is rounded down if necessary to be
5130 // a power of two. We want power of two unroll factors to simplify any
5131 // addressing operations or alignment considerations.
5132 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5135 // Don't count the induction variable as unrolled.
5136 if (EnableIndVarRegisterHeur)
5137 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5138 std::max(1U, (R.MaxLocalUsers - 1)));
5140 // Clamp the unroll factor ranges to reasonable factors.
5141 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5143 // Check if the user has overridden the unroll max.
5145 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5146 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5148 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5149 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5152 // If we did not calculate the cost for VF (because the user selected the VF)
5153 // then we calculate the cost of VF here.
5155 LoopCost = expectedCost(VF);
5157 // Clamp the calculated UF to be between the 1 and the max unroll factor
5158 // that the target allows.
5159 if (UF > MaxUnrollSize)
5164 // Unroll if we vectorized this loop and there is a reduction that could
5165 // benefit from unrolling.
5166 if (VF > 1 && Legal->getReductionVars()->size()) {
5167 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5171 // Note that if we've already vectorized the loop we will have done the
5172 // runtime check and so unrolling won't require further checks.
5173 bool UnrollingRequiresRuntimePointerCheck =
5174 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5176 // We want to unroll small loops in order to reduce the loop overhead and
5177 // potentially expose ILP opportunities.
5178 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5179 if (!UnrollingRequiresRuntimePointerCheck &&
5180 LoopCost < SmallLoopCost) {
5181 // We assume that the cost overhead is 1 and we use the cost model
5182 // to estimate the cost of the loop and unroll until the cost of the
5183 // loop overhead is about 5% of the cost of the loop.
5184 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5186 // Unroll until store/load ports (estimated by max unroll factor) are
5188 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5189 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5191 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5192 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5193 return std::max(StoresUF, LoadsUF);
5196 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5200 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5204 LoopVectorizationCostModel::RegisterUsage
5205 LoopVectorizationCostModel::calculateRegisterUsage() {
5206 // This function calculates the register usage by measuring the highest number
5207 // of values that are alive at a single location. Obviously, this is a very
5208 // rough estimation. We scan the loop in a topological order in order and
5209 // assign a number to each instruction. We use RPO to ensure that defs are
5210 // met before their users. We assume that each instruction that has in-loop
5211 // users starts an interval. We record every time that an in-loop value is
5212 // used, so we have a list of the first and last occurrences of each
5213 // instruction. Next, we transpose this data structure into a multi map that
5214 // holds the list of intervals that *end* at a specific location. This multi
5215 // map allows us to perform a linear search. We scan the instructions linearly
5216 // and record each time that a new interval starts, by placing it in a set.
5217 // If we find this value in the multi-map then we remove it from the set.
5218 // The max register usage is the maximum size of the set.
5219 // We also search for instructions that are defined outside the loop, but are
5220 // used inside the loop. We need this number separately from the max-interval
5221 // usage number because when we unroll, loop-invariant values do not take
5223 LoopBlocksDFS DFS(TheLoop);
5227 R.NumInstructions = 0;
5229 // Each 'key' in the map opens a new interval. The values
5230 // of the map are the index of the 'last seen' usage of the
5231 // instruction that is the key.
5232 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5233 // Maps instruction to its index.
5234 DenseMap<unsigned, Instruction*> IdxToInstr;
5235 // Marks the end of each interval.
5236 IntervalMap EndPoint;
5237 // Saves the list of instruction indices that are used in the loop.
5238 SmallSet<Instruction*, 8> Ends;
5239 // Saves the list of values that are used in the loop but are
5240 // defined outside the loop, such as arguments and constants.
5241 SmallPtrSet<Value*, 8> LoopInvariants;
5244 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5245 be = DFS.endRPO(); bb != be; ++bb) {
5246 R.NumInstructions += (*bb)->size();
5247 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5249 Instruction *I = it;
5250 IdxToInstr[Index++] = I;
5252 // Save the end location of each USE.
5253 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5254 Value *U = I->getOperand(i);
5255 Instruction *Instr = dyn_cast<Instruction>(U);
5257 // Ignore non-instruction values such as arguments, constants, etc.
5258 if (!Instr) continue;
5260 // If this instruction is outside the loop then record it and continue.
5261 if (!TheLoop->contains(Instr)) {
5262 LoopInvariants.insert(Instr);
5266 // Overwrite previous end points.
5267 EndPoint[Instr] = Index;
5273 // Saves the list of intervals that end with the index in 'key'.
5274 typedef SmallVector<Instruction*, 2> InstrList;
5275 DenseMap<unsigned, InstrList> TransposeEnds;
5277 // Transpose the EndPoints to a list of values that end at each index.
5278 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5280 TransposeEnds[it->second].push_back(it->first);
5282 SmallSet<Instruction*, 8> OpenIntervals;
5283 unsigned MaxUsage = 0;
5286 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5287 for (unsigned int i = 0; i < Index; ++i) {
5288 Instruction *I = IdxToInstr[i];
5289 // Ignore instructions that are never used within the loop.
5290 if (!Ends.count(I)) continue;
5292 // Remove all of the instructions that end at this location.
5293 InstrList &List = TransposeEnds[i];
5294 for (unsigned int j=0, e = List.size(); j < e; ++j)
5295 OpenIntervals.erase(List[j]);
5297 // Count the number of live interals.
5298 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5300 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5301 OpenIntervals.size() << '\n');
5303 // Add the current instruction to the list of open intervals.
5304 OpenIntervals.insert(I);
5307 unsigned Invariant = LoopInvariants.size();
5308 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5309 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5310 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5312 R.LoopInvariantRegs = Invariant;
5313 R.MaxLocalUsers = MaxUsage;
5317 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5321 for (Loop::block_iterator bb = TheLoop->block_begin(),
5322 be = TheLoop->block_end(); bb != be; ++bb) {
5323 unsigned BlockCost = 0;
5324 BasicBlock *BB = *bb;
5326 // For each instruction in the old loop.
5327 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5328 // Skip dbg intrinsics.
5329 if (isa<DbgInfoIntrinsic>(it))
5332 unsigned C = getInstructionCost(it, VF);
5334 // Check if we should override the cost.
5335 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5336 C = ForceTargetInstructionCost;
5339 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5340 VF << " For instruction: " << *it << '\n');
5343 // We assume that if-converted blocks have a 50% chance of being executed.
5344 // When the code is scalar then some of the blocks are avoided due to CF.
5345 // When the code is vectorized we execute all code paths.
5346 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5355 /// \brief Check whether the address computation for a non-consecutive memory
5356 /// access looks like an unlikely candidate for being merged into the indexing
5359 /// We look for a GEP which has one index that is an induction variable and all
5360 /// other indices are loop invariant. If the stride of this access is also
5361 /// within a small bound we decide that this address computation can likely be
5362 /// merged into the addressing mode.
5363 /// In all other cases, we identify the address computation as complex.
5364 static bool isLikelyComplexAddressComputation(Value *Ptr,
5365 LoopVectorizationLegality *Legal,
5366 ScalarEvolution *SE,
5367 const Loop *TheLoop) {
5368 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5372 // We are looking for a gep with all loop invariant indices except for one
5373 // which should be an induction variable.
5374 unsigned NumOperands = Gep->getNumOperands();
5375 for (unsigned i = 1; i < NumOperands; ++i) {
5376 Value *Opd = Gep->getOperand(i);
5377 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5378 !Legal->isInductionVariable(Opd))
5382 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5383 // can likely be merged into the address computation.
5384 unsigned MaxMergeDistance = 64;
5386 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5390 // Check the step is constant.
5391 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5392 // Calculate the pointer stride and check if it is consecutive.
5393 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5397 const APInt &APStepVal = C->getValue()->getValue();
5399 // Huge step value - give up.
5400 if (APStepVal.getBitWidth() > 64)
5403 int64_t StepVal = APStepVal.getSExtValue();
5405 return StepVal > MaxMergeDistance;
5408 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5409 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5415 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5416 // If we know that this instruction will remain uniform, check the cost of
5417 // the scalar version.
5418 if (Legal->isUniformAfterVectorization(I))
5421 Type *RetTy = I->getType();
5422 Type *VectorTy = ToVectorTy(RetTy, VF);
5424 // TODO: We need to estimate the cost of intrinsic calls.
5425 switch (I->getOpcode()) {
5426 case Instruction::GetElementPtr:
5427 // We mark this instruction as zero-cost because the cost of GEPs in
5428 // vectorized code depends on whether the corresponding memory instruction
5429 // is scalarized or not. Therefore, we handle GEPs with the memory
5430 // instruction cost.
5432 case Instruction::Br: {
5433 return TTI.getCFInstrCost(I->getOpcode());
5435 case Instruction::PHI:
5436 //TODO: IF-converted IFs become selects.
5438 case Instruction::Add:
5439 case Instruction::FAdd:
5440 case Instruction::Sub:
5441 case Instruction::FSub:
5442 case Instruction::Mul:
5443 case Instruction::FMul:
5444 case Instruction::UDiv:
5445 case Instruction::SDiv:
5446 case Instruction::FDiv:
5447 case Instruction::URem:
5448 case Instruction::SRem:
5449 case Instruction::FRem:
5450 case Instruction::Shl:
5451 case Instruction::LShr:
5452 case Instruction::AShr:
5453 case Instruction::And:
5454 case Instruction::Or:
5455 case Instruction::Xor: {
5456 // Since we will replace the stride by 1 the multiplication should go away.
5457 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5459 // Certain instructions can be cheaper to vectorize if they have a constant
5460 // second vector operand. One example of this are shifts on x86.
5461 TargetTransformInfo::OperandValueKind Op1VK =
5462 TargetTransformInfo::OK_AnyValue;
5463 TargetTransformInfo::OperandValueKind Op2VK =
5464 TargetTransformInfo::OK_AnyValue;
5465 Value *Op2 = I->getOperand(1);
5467 // Check for a splat of a constant or for a non uniform vector of constants.
5468 if (isa<ConstantInt>(Op2))
5469 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5470 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5471 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5472 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5473 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5476 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5478 case Instruction::Select: {
5479 SelectInst *SI = cast<SelectInst>(I);
5480 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5481 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5482 Type *CondTy = SI->getCondition()->getType();
5484 CondTy = VectorType::get(CondTy, VF);
5486 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5488 case Instruction::ICmp:
5489 case Instruction::FCmp: {
5490 Type *ValTy = I->getOperand(0)->getType();
5491 VectorTy = ToVectorTy(ValTy, VF);
5492 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5494 case Instruction::Store:
5495 case Instruction::Load: {
5496 StoreInst *SI = dyn_cast<StoreInst>(I);
5497 LoadInst *LI = dyn_cast<LoadInst>(I);
5498 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5500 VectorTy = ToVectorTy(ValTy, VF);
5502 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5503 unsigned AS = SI ? SI->getPointerAddressSpace() :
5504 LI->getPointerAddressSpace();
5505 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5506 // We add the cost of address computation here instead of with the gep
5507 // instruction because only here we know whether the operation is
5510 return TTI.getAddressComputationCost(VectorTy) +
5511 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5513 // Scalarized loads/stores.
5514 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5515 bool Reverse = ConsecutiveStride < 0;
5516 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5517 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5518 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5519 bool IsComplexComputation =
5520 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5522 // The cost of extracting from the value vector and pointer vector.
5523 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5524 for (unsigned i = 0; i < VF; ++i) {
5525 // The cost of extracting the pointer operand.
5526 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5527 // In case of STORE, the cost of ExtractElement from the vector.
5528 // In case of LOAD, the cost of InsertElement into the returned
5530 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5531 Instruction::InsertElement,
5535 // The cost of the scalar loads/stores.
5536 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5537 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5542 // Wide load/stores.
5543 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5544 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5547 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5551 case Instruction::ZExt:
5552 case Instruction::SExt:
5553 case Instruction::FPToUI:
5554 case Instruction::FPToSI:
5555 case Instruction::FPExt:
5556 case Instruction::PtrToInt:
5557 case Instruction::IntToPtr:
5558 case Instruction::SIToFP:
5559 case Instruction::UIToFP:
5560 case Instruction::Trunc:
5561 case Instruction::FPTrunc:
5562 case Instruction::BitCast: {
5563 // We optimize the truncation of induction variable.
5564 // The cost of these is the same as the scalar operation.
5565 if (I->getOpcode() == Instruction::Trunc &&
5566 Legal->isInductionVariable(I->getOperand(0)))
5567 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5568 I->getOperand(0)->getType());
5570 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5571 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5573 case Instruction::Call: {
5574 CallInst *CI = cast<CallInst>(I);
5575 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5576 assert(ID && "Not an intrinsic call!");
5577 Type *RetTy = ToVectorTy(CI->getType(), VF);
5578 SmallVector<Type*, 4> Tys;
5579 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5580 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5581 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5584 // We are scalarizing the instruction. Return the cost of the scalar
5585 // instruction, plus the cost of insert and extract into vector
5586 // elements, times the vector width.
5589 if (!RetTy->isVoidTy() && VF != 1) {
5590 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5592 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5595 // The cost of inserting the results plus extracting each one of the
5597 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5600 // The cost of executing VF copies of the scalar instruction. This opcode
5601 // is unknown. Assume that it is the same as 'mul'.
5602 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5608 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5609 if (Scalar->isVoidTy() || VF == 1)
5611 return VectorType::get(Scalar, VF);
5614 char LoopVectorize::ID = 0;
5615 static const char lv_name[] = "Loop Vectorization";
5616 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5617 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5618 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5619 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5620 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5621 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5622 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5623 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5624 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5627 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5628 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5632 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5633 // Check for a store.
5634 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5635 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5637 // Check for a load.
5638 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5639 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5645 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5646 bool IfPredicateStore) {
5647 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5648 // Holds vector parameters or scalars, in case of uniform vals.
5649 SmallVector<VectorParts, 4> Params;
5651 setDebugLocFromInst(Builder, Instr);
5653 // Find all of the vectorized parameters.
5654 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5655 Value *SrcOp = Instr->getOperand(op);
5657 // If we are accessing the old induction variable, use the new one.
5658 if (SrcOp == OldInduction) {
5659 Params.push_back(getVectorValue(SrcOp));
5663 // Try using previously calculated values.
5664 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5666 // If the src is an instruction that appeared earlier in the basic block
5667 // then it should already be vectorized.
5668 if (SrcInst && OrigLoop->contains(SrcInst)) {
5669 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5670 // The parameter is a vector value from earlier.
5671 Params.push_back(WidenMap.get(SrcInst));
5673 // The parameter is a scalar from outside the loop. Maybe even a constant.
5674 VectorParts Scalars;
5675 Scalars.append(UF, SrcOp);
5676 Params.push_back(Scalars);
5680 assert(Params.size() == Instr->getNumOperands() &&
5681 "Invalid number of operands");
5683 // Does this instruction return a value ?
5684 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5686 Value *UndefVec = IsVoidRetTy ? nullptr :
5687 UndefValue::get(Instr->getType());
5688 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5689 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5691 Instruction *InsertPt = Builder.GetInsertPoint();
5692 BasicBlock *IfBlock = Builder.GetInsertBlock();
5693 BasicBlock *CondBlock = nullptr;
5696 Loop *VectorLp = nullptr;
5697 if (IfPredicateStore) {
5698 assert(Instr->getParent()->getSinglePredecessor() &&
5699 "Only support single predecessor blocks");
5700 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5701 Instr->getParent());
5702 VectorLp = LI->getLoopFor(IfBlock);
5703 assert(VectorLp && "Must have a loop for this block");
5706 // For each vector unroll 'part':
5707 for (unsigned Part = 0; Part < UF; ++Part) {
5708 // For each scalar that we create:
5710 // Start an "if (pred) a[i] = ..." block.
5711 Value *Cmp = nullptr;
5712 if (IfPredicateStore) {
5713 if (Cond[Part]->getType()->isVectorTy())
5715 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5716 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5717 ConstantInt::get(Cond[Part]->getType(), 1));
5718 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5719 LoopVectorBody.push_back(CondBlock);
5720 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5721 // Update Builder with newly created basic block.
5722 Builder.SetInsertPoint(InsertPt);
5725 Instruction *Cloned = Instr->clone();
5727 Cloned->setName(Instr->getName() + ".cloned");
5728 // Replace the operands of the cloned instructions with extracted scalars.
5729 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5730 Value *Op = Params[op][Part];
5731 Cloned->setOperand(op, Op);
5734 // Place the cloned scalar in the new loop.
5735 Builder.Insert(Cloned);
5737 // If the original scalar returns a value we need to place it in a vector
5738 // so that future users will be able to use it.
5740 VecResults[Part] = Cloned;
5743 if (IfPredicateStore) {
5744 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5745 LoopVectorBody.push_back(NewIfBlock);
5746 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
5747 Builder.SetInsertPoint(InsertPt);
5748 Instruction *OldBr = IfBlock->getTerminator();
5749 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5750 OldBr->eraseFromParent();
5751 IfBlock = NewIfBlock;
5756 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5757 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5758 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5760 return scalarizeInstruction(Instr, IfPredicateStore);
5763 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5767 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5771 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
5773 // When unrolling and the VF is 1, we only need to add a simple scalar.
5774 Type *ITy = Val->getType();
5775 assert(!ITy->isVectorTy() && "Val must be a scalar");
5776 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
5777 return Builder.CreateAdd(Val, C, "induction");