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/DiagnosticInfo.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/LLVMContext.h"
77 #include "llvm/IR/Module.h"
78 #include "llvm/IR/PatternMatch.h"
79 #include "llvm/IR/Type.h"
80 #include "llvm/IR/Value.h"
81 #include "llvm/IR/ValueHandle.h"
82 #include "llvm/IR/Verifier.h"
83 #include "llvm/Pass.h"
84 #include "llvm/Support/BranchProbability.h"
85 #include "llvm/Support/CommandLine.h"
86 #include "llvm/Support/Debug.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 /// Optimization analysis message produced during vectorization. Messages inform
213 /// the user why vectorization did not occur.
216 raw_string_ostream Out;
220 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
221 Out << "loop not vectorized: ";
224 template <typename A> Report &operator<<(const A &Value) {
229 Instruction *getInstr() { return Instr; }
231 std::string &str() { return Out.str(); }
232 operator Twine() { return Out.str(); }
235 /// InnerLoopVectorizer vectorizes loops which contain only one basic
236 /// block to a specified vectorization factor (VF).
237 /// This class performs the widening of scalars into vectors, or multiple
238 /// scalars. This class also implements the following features:
239 /// * It inserts an epilogue loop for handling loops that don't have iteration
240 /// counts that are known to be a multiple of the vectorization factor.
241 /// * It handles the code generation for reduction variables.
242 /// * Scalarization (implementation using scalars) of un-vectorizable
244 /// InnerLoopVectorizer does not perform any vectorization-legality
245 /// checks, and relies on the caller to check for the different legality
246 /// aspects. The InnerLoopVectorizer relies on the
247 /// LoopVectorizationLegality class to provide information about the induction
248 /// and reduction variables that were found to a given vectorization factor.
249 class InnerLoopVectorizer {
251 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
252 DominatorTree *DT, const DataLayout *DL,
253 const TargetLibraryInfo *TLI, unsigned VecWidth,
254 unsigned UnrollFactor)
255 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
256 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
257 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
260 // Perform the actual loop widening (vectorization).
261 void vectorize(LoopVectorizationLegality *L) {
263 // Create a new empty loop. Unlink the old loop and connect the new one.
265 // Widen each instruction in the old loop to a new one in the new loop.
266 // Use the Legality module to find the induction and reduction variables.
268 // Register the new loop and update the analysis passes.
272 virtual ~InnerLoopVectorizer() {}
275 /// A small list of PHINodes.
276 typedef SmallVector<PHINode*, 4> PhiVector;
277 /// When we unroll loops we have multiple vector values for each scalar.
278 /// This data structure holds the unrolled and vectorized values that
279 /// originated from one scalar instruction.
280 typedef SmallVector<Value*, 2> VectorParts;
282 // When we if-convert we need create edge masks. We have to cache values so
283 // that we don't end up with exponential recursion/IR.
284 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
285 VectorParts> EdgeMaskCache;
287 /// \brief Add code that checks at runtime if the accessed arrays overlap.
289 /// Returns a pair of instructions where the first element is the first
290 /// instruction generated in possibly a sequence of instructions and the
291 /// second value is the final comparator value or NULL if no check is needed.
292 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
294 /// \brief Add checks for strides that where assumed to be 1.
296 /// Returns the last check instruction and the first check instruction in the
297 /// pair as (first, last).
298 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
300 /// Create an empty loop, based on the loop ranges of the old loop.
301 void createEmptyLoop();
302 /// Copy and widen the instructions from the old loop.
303 virtual void vectorizeLoop();
305 /// \brief The Loop exit block may have single value PHI nodes where the
306 /// incoming value is 'Undef'. While vectorizing we only handled real values
307 /// that were defined inside the loop. Here we fix the 'undef case'.
311 /// A helper function that computes the predicate of the block BB, assuming
312 /// that the header block of the loop is set to True. It returns the *entry*
313 /// mask for the block BB.
314 VectorParts createBlockInMask(BasicBlock *BB);
315 /// A helper function that computes the predicate of the edge between SRC
317 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
319 /// A helper function to vectorize a single BB within the innermost loop.
320 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
322 /// Vectorize a single PHINode in a block. This method handles the induction
323 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
324 /// arbitrary length vectors.
325 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
326 unsigned UF, unsigned VF, PhiVector *PV);
328 /// Insert the new loop to the loop hierarchy and pass manager
329 /// and update the analysis passes.
330 void updateAnalysis();
332 /// This instruction is un-vectorizable. Implement it as a sequence
333 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
334 /// scalarized instruction behind an if block predicated on the control
335 /// dependence of the instruction.
336 virtual void scalarizeInstruction(Instruction *Instr,
337 bool IfPredicateStore=false);
339 /// Vectorize Load and Store instructions,
340 virtual void vectorizeMemoryInstruction(Instruction *Instr);
342 /// Create a broadcast instruction. This method generates a broadcast
343 /// instruction (shuffle) for loop invariant values and for the induction
344 /// value. If this is the induction variable then we extend it to N, N+1, ...
345 /// this is needed because each iteration in the loop corresponds to a SIMD
347 virtual Value *getBroadcastInstrs(Value *V);
349 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
350 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
351 /// The sequence starts at StartIndex.
352 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
354 /// When we go over instructions in the basic block we rely on previous
355 /// values within the current basic block or on loop invariant values.
356 /// When we widen (vectorize) values we place them in the map. If the values
357 /// are not within the map, they have to be loop invariant, so we simply
358 /// broadcast them into a vector.
359 VectorParts &getVectorValue(Value *V);
361 /// Generate a shuffle sequence that will reverse the vector Vec.
362 virtual Value *reverseVector(Value *Vec);
364 /// This is a helper class that holds the vectorizer state. It maps scalar
365 /// instructions to vector instructions. When the code is 'unrolled' then
366 /// then a single scalar value is mapped to multiple vector parts. The parts
367 /// are stored in the VectorPart type.
369 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
371 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
373 /// \return True if 'Key' is saved in the Value Map.
374 bool has(Value *Key) const { return MapStorage.count(Key); }
376 /// Initializes a new entry in the map. Sets all of the vector parts to the
377 /// save value in 'Val'.
378 /// \return A reference to a vector with splat values.
379 VectorParts &splat(Value *Key, Value *Val) {
380 VectorParts &Entry = MapStorage[Key];
381 Entry.assign(UF, Val);
385 ///\return A reference to the value that is stored at 'Key'.
386 VectorParts &get(Value *Key) {
387 VectorParts &Entry = MapStorage[Key];
390 assert(Entry.size() == UF);
395 /// The unroll factor. Each entry in the map stores this number of vector
399 /// Map storage. We use std::map and not DenseMap because insertions to a
400 /// dense map invalidates its iterators.
401 std::map<Value *, VectorParts> MapStorage;
404 /// The original loop.
406 /// Scev analysis to use.
413 const DataLayout *DL;
414 /// Target Library Info.
415 const TargetLibraryInfo *TLI;
417 /// The vectorization SIMD factor to use. Each vector will have this many
422 /// The vectorization unroll factor to use. Each scalar is vectorized to this
423 /// many different vector instructions.
426 /// The builder that we use
429 // --- Vectorization state ---
431 /// The vector-loop preheader.
432 BasicBlock *LoopVectorPreHeader;
433 /// The scalar-loop preheader.
434 BasicBlock *LoopScalarPreHeader;
435 /// Middle Block between the vector and the scalar.
436 BasicBlock *LoopMiddleBlock;
437 ///The ExitBlock of the scalar loop.
438 BasicBlock *LoopExitBlock;
439 ///The vector loop body.
440 SmallVector<BasicBlock *, 4> LoopVectorBody;
441 ///The scalar loop body.
442 BasicBlock *LoopScalarBody;
443 /// A list of all bypass blocks. The first block is the entry of the loop.
444 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
446 /// The new Induction variable which was added to the new block.
448 /// The induction variable of the old basic block.
449 PHINode *OldInduction;
450 /// Holds the extended (to the widest induction type) start index.
452 /// Maps scalars to widened vectors.
454 EdgeMaskCache MaskCache;
456 LoopVectorizationLegality *Legal;
459 class InnerLoopUnroller : public InnerLoopVectorizer {
461 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
462 DominatorTree *DT, const DataLayout *DL,
463 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
464 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
467 void scalarizeInstruction(Instruction *Instr,
468 bool IfPredicateStore = false) override;
469 void vectorizeMemoryInstruction(Instruction *Instr) override;
470 Value *getBroadcastInstrs(Value *V) override;
471 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
472 Value *reverseVector(Value *Vec) override;
475 /// \brief Look for a meaningful debug location on the instruction or it's
477 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
482 if (I->getDebugLoc() != Empty)
485 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
486 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
487 if (OpInst->getDebugLoc() != Empty)
494 /// \brief Set the debug location in the builder using the debug location in the
496 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
497 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
498 B.SetCurrentDebugLocation(Inst->getDebugLoc());
500 B.SetCurrentDebugLocation(DebugLoc());
504 /// \return string containing a file name and a line # for the given loop.
505 static std::string getDebugLocString(const Loop *L) {
508 raw_string_ostream OS(Result);
509 const DebugLoc LoopDbgLoc = L->getStartLoc();
510 if (!LoopDbgLoc.isUnknown())
511 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
513 // Just print the module name.
514 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
521 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
522 /// to what vectorization factor.
523 /// This class does not look at the profitability of vectorization, only the
524 /// legality. This class has two main kinds of checks:
525 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
526 /// will change the order of memory accesses in a way that will change the
527 /// correctness of the program.
528 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
529 /// checks for a number of different conditions, such as the availability of a
530 /// single induction variable, that all types are supported and vectorize-able,
531 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
532 /// This class is also used by InnerLoopVectorizer for identifying
533 /// induction variable and the different reduction variables.
534 class LoopVectorizationLegality {
538 unsigned NumPredStores;
540 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
541 DominatorTree *DT, TargetLibraryInfo *TLI,
543 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
544 DT(DT), TLI(TLI), TheFunction(F), Induction(nullptr),
545 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
548 /// This enum represents the kinds of reductions that we support.
550 RK_NoReduction, ///< Not a reduction.
551 RK_IntegerAdd, ///< Sum of integers.
552 RK_IntegerMult, ///< Product of integers.
553 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
554 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
555 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
556 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
557 RK_FloatAdd, ///< Sum of floats.
558 RK_FloatMult, ///< Product of floats.
559 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
562 /// This enum represents the kinds of inductions that we support.
564 IK_NoInduction, ///< Not an induction variable.
565 IK_IntInduction, ///< Integer induction variable. Step = 1.
566 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
567 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
568 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
571 // This enum represents the kind of minmax reduction.
572 enum MinMaxReductionKind {
582 /// This struct holds information about reduction variables.
583 struct ReductionDescriptor {
584 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
585 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
587 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
588 MinMaxReductionKind MK)
589 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
591 // The starting value of the reduction.
592 // It does not have to be zero!
593 TrackingVH<Value> StartValue;
594 // The instruction who's value is used outside the loop.
595 Instruction *LoopExitInstr;
596 // The kind of the reduction.
598 // If this a min/max reduction the kind of reduction.
599 MinMaxReductionKind MinMaxKind;
602 /// This POD struct holds information about a potential reduction operation.
603 struct ReductionInstDesc {
604 ReductionInstDesc(bool IsRedux, Instruction *I) :
605 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
607 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
608 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
610 // Is this instruction a reduction candidate.
612 // The last instruction in a min/max pattern (select of the select(icmp())
613 // pattern), or the current reduction instruction otherwise.
614 Instruction *PatternLastInst;
615 // If this is a min/max pattern the comparison predicate.
616 MinMaxReductionKind MinMaxKind;
619 /// This struct holds information about the memory runtime legality
620 /// check that a group of pointers do not overlap.
621 struct RuntimePointerCheck {
622 RuntimePointerCheck() : Need(false) {}
624 /// Reset the state of the pointer runtime information.
631 DependencySetId.clear();
634 /// Insert a pointer and calculate the start and end SCEVs.
635 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
636 unsigned DepSetId, ValueToValueMap &Strides);
638 /// This flag indicates if we need to add the runtime check.
640 /// Holds the pointers that we need to check.
641 SmallVector<TrackingVH<Value>, 2> Pointers;
642 /// Holds the pointer value at the beginning of the loop.
643 SmallVector<const SCEV*, 2> Starts;
644 /// Holds the pointer value at the end of the loop.
645 SmallVector<const SCEV*, 2> Ends;
646 /// Holds the information if this pointer is used for writing to memory.
647 SmallVector<bool, 2> IsWritePtr;
648 /// Holds the id of the set of pointers that could be dependent because of a
649 /// shared underlying object.
650 SmallVector<unsigned, 2> DependencySetId;
653 /// A struct for saving information about induction variables.
654 struct InductionInfo {
655 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
656 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
658 TrackingVH<Value> StartValue;
663 /// ReductionList contains the reduction descriptors for all
664 /// of the reductions that were found in the loop.
665 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
667 /// InductionList saves induction variables and maps them to the
668 /// induction descriptor.
669 typedef MapVector<PHINode*, InductionInfo> InductionList;
671 /// Returns true if it is legal to vectorize this loop.
672 /// This does not mean that it is profitable to vectorize this
673 /// loop, only that it is legal to do so.
676 /// Returns the Induction variable.
677 PHINode *getInduction() { return Induction; }
679 /// Returns the reduction variables found in the loop.
680 ReductionList *getReductionVars() { return &Reductions; }
682 /// Returns the induction variables found in the loop.
683 InductionList *getInductionVars() { return &Inductions; }
685 /// Returns the widest induction type.
686 Type *getWidestInductionType() { return WidestIndTy; }
688 /// Returns True if V is an induction variable in this loop.
689 bool isInductionVariable(const Value *V);
691 /// Return true if the block BB needs to be predicated in order for the loop
692 /// to be vectorized.
693 bool blockNeedsPredication(BasicBlock *BB);
695 /// Check if this pointer is consecutive when vectorizing. This happens
696 /// when the last index of the GEP is the induction variable, or that the
697 /// pointer itself is an induction variable.
698 /// This check allows us to vectorize A[idx] into a wide load/store.
700 /// 0 - Stride is unknown or non-consecutive.
701 /// 1 - Address is consecutive.
702 /// -1 - Address is consecutive, and decreasing.
703 int isConsecutivePtr(Value *Ptr);
705 /// Returns true if the value V is uniform within the loop.
706 bool isUniform(Value *V);
708 /// Returns true if this instruction will remain scalar after vectorization.
709 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
711 /// Returns the information that we collected about runtime memory check.
712 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
714 /// This function returns the identity element (or neutral element) for
716 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
718 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
720 bool hasStride(Value *V) { return StrideSet.count(V); }
721 bool mustCheckStrides() { return !StrideSet.empty(); }
722 SmallPtrSet<Value *, 8>::iterator strides_begin() {
723 return StrideSet.begin();
725 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
728 /// Check if a single basic block loop is vectorizable.
729 /// At this point we know that this is a loop with a constant trip count
730 /// and we only need to check individual instructions.
731 bool canVectorizeInstrs();
733 /// When we vectorize loops we may change the order in which
734 /// we read and write from memory. This method checks if it is
735 /// legal to vectorize the code, considering only memory constrains.
736 /// Returns true if the loop is vectorizable
737 bool canVectorizeMemory();
739 /// Return true if we can vectorize this loop using the IF-conversion
741 bool canVectorizeWithIfConvert();
743 /// Collect the variables that need to stay uniform after vectorization.
744 void collectLoopUniforms();
746 /// Return true if all of the instructions in the block can be speculatively
747 /// executed. \p SafePtrs is a list of addresses that are known to be legal
748 /// and we know that we can read from them without segfault.
749 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
751 /// Returns True, if 'Phi' is the kind of reduction variable for type
752 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
753 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
754 /// Returns a struct describing if the instruction 'I' can be a reduction
755 /// variable of type 'Kind'. If the reduction is a min/max pattern of
756 /// select(icmp()) this function advances the instruction pointer 'I' from the
757 /// compare instruction to the select instruction and stores this pointer in
758 /// 'PatternLastInst' member of the returned struct.
759 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
760 ReductionInstDesc &Desc);
761 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
762 /// pattern corresponding to a min(X, Y) or max(X, Y).
763 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
764 ReductionInstDesc &Prev);
765 /// Returns the induction kind of Phi. This function may return NoInduction
766 /// if the PHI is not an induction variable.
767 InductionKind isInductionVariable(PHINode *Phi);
769 /// \brief Collect memory access with loop invariant strides.
771 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
773 void collectStridedAcccess(Value *LoadOrStoreInst);
775 /// Report an analysis message to assist the user in diagnosing loops that are
777 void emitAnalysis(Report &Message) {
778 DebugLoc DL = TheLoop->getStartLoc();
779 if (Instruction *I = Message.getInstr())
780 DL = I->getDebugLoc();
781 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
782 *TheFunction, DL, Message.str());
785 /// The loop that we evaluate.
789 /// DataLayout analysis.
790 const DataLayout *DL;
793 /// Target Library Info.
794 TargetLibraryInfo *TLI;
796 Function *TheFunction;
798 // --- vectorization state --- //
800 /// Holds the integer induction variable. This is the counter of the
803 /// Holds the reduction variables.
804 ReductionList Reductions;
805 /// Holds all of the induction variables that we found in the loop.
806 /// Notice that inductions don't need to start at zero and that induction
807 /// variables can be pointers.
808 InductionList Inductions;
809 /// Holds the widest induction type encountered.
812 /// Allowed outside users. This holds the reduction
813 /// vars which can be accessed from outside the loop.
814 SmallPtrSet<Value*, 4> AllowedExit;
815 /// This set holds the variables which are known to be uniform after
817 SmallPtrSet<Instruction*, 4> Uniforms;
818 /// We need to check that all of the pointers in this list are disjoint
820 RuntimePointerCheck PtrRtCheck;
821 /// Can we assume the absence of NaNs.
822 bool HasFunNoNaNAttr;
824 unsigned MaxSafeDepDistBytes;
826 ValueToValueMap Strides;
827 SmallPtrSet<Value *, 8> StrideSet;
830 /// LoopVectorizationCostModel - estimates the expected speedups due to
832 /// In many cases vectorization is not profitable. This can happen because of
833 /// a number of reasons. In this class we mainly attempt to predict the
834 /// expected speedup/slowdowns due to the supported instruction set. We use the
835 /// TargetTransformInfo to query the different backends for the cost of
836 /// different operations.
837 class LoopVectorizationCostModel {
839 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
840 LoopVectorizationLegality *Legal,
841 const TargetTransformInfo &TTI,
842 const DataLayout *DL, const TargetLibraryInfo *TLI)
843 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
845 /// Information about vectorization costs
846 struct VectorizationFactor {
847 unsigned Width; // Vector width with best cost
848 unsigned Cost; // Cost of the loop with that width
850 /// \return The most profitable vectorization factor and the cost of that VF.
851 /// This method checks every power of two up to VF. If UserVF is not ZERO
852 /// then this vectorization factor will be selected if vectorization is
854 VectorizationFactor selectVectorizationFactor(bool OptForSize,
856 bool ForceVectorization);
858 /// \return The size (in bits) of the widest type in the code that
859 /// needs to be vectorized. We ignore values that remain scalar such as
860 /// 64 bit loop indices.
861 unsigned getWidestType();
863 /// \return The most profitable unroll factor.
864 /// If UserUF is non-zero then this method finds the best unroll-factor
865 /// based on register pressure and other parameters.
866 /// VF and LoopCost are the selected vectorization factor and the cost of the
868 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
871 /// \brief A struct that represents some properties of the register usage
873 struct RegisterUsage {
874 /// Holds the number of loop invariant values that are used in the loop.
875 unsigned LoopInvariantRegs;
876 /// Holds the maximum number of concurrent live intervals in the loop.
877 unsigned MaxLocalUsers;
878 /// Holds the number of instructions in the loop.
879 unsigned NumInstructions;
882 /// \return information about the register usage of the loop.
883 RegisterUsage calculateRegisterUsage();
886 /// Returns the expected execution cost. The unit of the cost does
887 /// not matter because we use the 'cost' units to compare different
888 /// vector widths. The cost that is returned is *not* normalized by
889 /// the factor width.
890 unsigned expectedCost(unsigned VF);
892 /// Returns the execution time cost of an instruction for a given vector
893 /// width. Vector width of one means scalar.
894 unsigned getInstructionCost(Instruction *I, unsigned VF);
896 /// A helper function for converting Scalar types to vector types.
897 /// If the incoming type is void, we return void. If the VF is 1, we return
899 static Type* ToVectorTy(Type *Scalar, unsigned VF);
901 /// Returns whether the instruction is a load or store and will be a emitted
902 /// as a vector operation.
903 bool isConsecutiveLoadOrStore(Instruction *I);
905 /// The loop that we evaluate.
909 /// Loop Info analysis.
911 /// Vectorization legality.
912 LoopVectorizationLegality *Legal;
913 /// Vector target information.
914 const TargetTransformInfo &TTI;
915 /// Target data layout information.
916 const DataLayout *DL;
917 /// Target Library Info.
918 const TargetLibraryInfo *TLI;
921 /// Utility class for getting and setting loop vectorizer hints in the form
922 /// of loop metadata.
923 class LoopVectorizeHints {
926 FK_Undefined = -1, ///< Not selected.
927 FK_Disabled = 0, ///< Forcing disabled.
928 FK_Enabled = 1, ///< Forcing enabled.
931 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
932 : Width(VectorizationFactor),
933 Unroll(DisableUnrolling),
935 LoopID(L->getLoopID()) {
937 // force-vector-unroll overrides DisableUnrolling.
938 if (VectorizationUnroll.getNumOccurrences() > 0)
939 Unroll = VectorizationUnroll;
941 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
942 << "LV: Unrolling disabled by the pass manager\n");
945 /// Return the loop vectorizer metadata prefix.
946 static StringRef Prefix() { return "llvm.loop.vectorize."; }
948 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
949 SmallVector<Value*, 2> Vals;
950 Vals.push_back(MDString::get(Context, Name));
951 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
952 return MDNode::get(Context, Vals);
955 /// Mark the loop L as already vectorized by setting the width to 1.
956 void setAlreadyVectorized(Loop *L) {
957 LLVMContext &Context = L->getHeader()->getContext();
961 // Create a new loop id with one more operand for the already_vectorized
962 // hint. If the loop already has a loop id then copy the existing operands.
963 SmallVector<Value*, 4> Vals(1);
965 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
966 Vals.push_back(LoopID->getOperand(i));
968 Vals.push_back(createHint(Context, Twine(Prefix(), "width").str(), Width));
969 Vals.push_back(createHint(Context, Twine(Prefix(), "unroll").str(), 1));
971 MDNode *NewLoopID = MDNode::get(Context, Vals);
972 // Set operand 0 to refer to the loop id itself.
973 NewLoopID->replaceOperandWith(0, NewLoopID);
975 L->setLoopID(NewLoopID);
977 LoopID->replaceAllUsesWith(NewLoopID);
982 std::string emitRemark() const {
984 R << "vectorization ";
986 case LoopVectorizeHints::FK_Disabled:
987 R << "is explicitly disabled";
989 case LoopVectorizeHints::FK_Enabled:
990 R << "is explicitly enabled";
991 if (Width != 0 && Unroll != 0)
992 R << " with width " << Width << " and interleave count " << Unroll;
994 R << " with width " << Width;
995 else if (Unroll != 0)
996 R << " with interleave count " << Unroll;
998 case LoopVectorizeHints::FK_Undefined:
999 R << "was not specified";
1005 unsigned getWidth() const { return Width; }
1006 unsigned getUnroll() const { return Unroll; }
1007 enum ForceKind getForce() const { return Force; }
1008 MDNode *getLoopID() const { return LoopID; }
1011 /// Find hints specified in the loop metadata.
1012 void getHints(const Loop *L) {
1016 // First operand should refer to the loop id itself.
1017 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1018 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1020 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1021 const MDString *S = nullptr;
1022 SmallVector<Value*, 4> Args;
1024 // The expected hint is either a MDString or a MDNode with the first
1025 // operand a MDString.
1026 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1027 if (!MD || MD->getNumOperands() == 0)
1029 S = dyn_cast<MDString>(MD->getOperand(0));
1030 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1031 Args.push_back(MD->getOperand(i));
1033 S = dyn_cast<MDString>(LoopID->getOperand(i));
1034 assert(Args.size() == 0 && "too many arguments for MDString");
1040 // Check if the hint starts with the vectorizer prefix.
1041 StringRef Hint = S->getString();
1042 if (!Hint.startswith(Prefix()))
1044 // Remove the prefix.
1045 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1047 if (Args.size() == 1)
1048 getHint(Hint, Args[0]);
1052 // Check string hint with one operand.
1053 void getHint(StringRef Hint, Value *Arg) {
1054 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1056 unsigned Val = C->getZExtValue();
1058 if (Hint == "width") {
1059 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1062 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1063 } else if (Hint == "unroll") {
1064 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1067 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1068 } else if (Hint == "enable") {
1069 if (C->getBitWidth() == 1)
1070 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1071 : LoopVectorizeHints::FK_Disabled;
1073 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1075 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1079 /// Vectorization width.
1081 /// Vectorization unroll factor.
1083 /// Vectorization forced
1084 enum ForceKind Force;
1089 static void emitMissedWarning(Function *F, Loop *L,
1090 const LoopVectorizeHints &LH) {
1091 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1092 L->getStartLoc(), LH.emitRemark());
1094 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1095 if (LH.getWidth() != 1)
1096 emitLoopVectorizeWarning(
1097 F->getContext(), *F, L->getStartLoc(),
1098 "failed explicitly specified loop vectorization");
1099 else if (LH.getUnroll() != 1)
1100 emitLoopInterleaveWarning(
1101 F->getContext(), *F, L->getStartLoc(),
1102 "failed explicitly specified loop interleaving");
1106 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1108 return V.push_back(&L);
1110 for (Loop *InnerL : L)
1111 addInnerLoop(*InnerL, V);
1114 /// The LoopVectorize Pass.
1115 struct LoopVectorize : public FunctionPass {
1116 /// Pass identification, replacement for typeid
1119 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1121 DisableUnrolling(NoUnrolling),
1122 AlwaysVectorize(AlwaysVectorize) {
1123 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1126 ScalarEvolution *SE;
1127 const DataLayout *DL;
1129 TargetTransformInfo *TTI;
1131 BlockFrequencyInfo *BFI;
1132 TargetLibraryInfo *TLI;
1133 bool DisableUnrolling;
1134 bool AlwaysVectorize;
1136 BlockFrequency ColdEntryFreq;
1138 bool runOnFunction(Function &F) override {
1139 SE = &getAnalysis<ScalarEvolution>();
1140 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1141 DL = DLP ? &DLP->getDataLayout() : nullptr;
1142 LI = &getAnalysis<LoopInfo>();
1143 TTI = &getAnalysis<TargetTransformInfo>();
1144 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1145 BFI = &getAnalysis<BlockFrequencyInfo>();
1146 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1148 // Compute some weights outside of the loop over the loops. Compute this
1149 // using a BranchProbability to re-use its scaling math.
1150 const BranchProbability ColdProb(1, 5); // 20%
1151 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1153 // If the target claims to have no vector registers don't attempt
1155 if (!TTI->getNumberOfRegisters(true))
1159 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1160 << ": Missing data layout\n");
1164 // Build up a worklist of inner-loops to vectorize. This is necessary as
1165 // the act of vectorizing or partially unrolling a loop creates new loops
1166 // and can invalidate iterators across the loops.
1167 SmallVector<Loop *, 8> Worklist;
1170 addInnerLoop(*L, Worklist);
1172 LoopsAnalyzed += Worklist.size();
1174 // Now walk the identified inner loops.
1175 bool Changed = false;
1176 while (!Worklist.empty())
1177 Changed |= processLoop(Worklist.pop_back_val());
1179 // Process each loop nest in the function.
1183 bool processLoop(Loop *L) {
1184 assert(L->empty() && "Only process inner loops.");
1187 const std::string DebugLocStr = getDebugLocString(L);
1190 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1191 << L->getHeader()->getParent()->getName() << "\" from "
1192 << DebugLocStr << "\n");
1194 LoopVectorizeHints Hints(L, DisableUnrolling);
1196 DEBUG(dbgs() << "LV: Loop hints:"
1198 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1200 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1202 : "?")) << " width=" << Hints.getWidth()
1203 << " unroll=" << Hints.getUnroll() << "\n");
1205 // Function containing loop
1206 Function *F = L->getHeader()->getParent();
1208 // Looking at the diagnostic output is the only way to determine if a loop
1209 // was vectorized (other than looking at the IR or machine code), so it
1210 // is important to generate an optimization remark for each loop. Most of
1211 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1212 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1213 // less verbose reporting vectorized loops and unvectorized loops that may
1214 // benefit from vectorization, respectively.
1216 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1217 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1218 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1219 L->getStartLoc(), Hints.emitRemark());
1223 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1224 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1225 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1226 L->getStartLoc(), Hints.emitRemark());
1230 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1231 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1232 emitOptimizationRemarkAnalysis(
1233 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1234 "loop not vectorized: vector width and interleave count are "
1235 "explicitly set to 1");
1239 // Check the loop for a trip count threshold:
1240 // do not vectorize loops with a tiny trip count.
1241 BasicBlock *Latch = L->getLoopLatch();
1242 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1243 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1244 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1245 << "This loop is not worth vectorizing.");
1246 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1247 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1249 DEBUG(dbgs() << "\n");
1250 emitOptimizationRemarkAnalysis(
1251 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1252 "vectorization is not beneficial and is not explicitly forced");
1257 // Check if it is legal to vectorize the loop.
1258 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, F);
1259 if (!LVL.canVectorize()) {
1260 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1261 emitMissedWarning(F, L, Hints);
1265 // Use the cost model.
1266 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1268 // Check the function attributes to find out if this function should be
1269 // optimized for size.
1270 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1271 F->hasFnAttribute(Attribute::OptimizeForSize);
1273 // Compute the weighted frequency of this loop being executed and see if it
1274 // is less than 20% of the function entry baseline frequency. Note that we
1275 // always have a canonical loop here because we think we *can* vectoriez.
1276 // FIXME: This is hidden behind a flag due to pervasive problems with
1277 // exactly what block frequency models.
1278 if (LoopVectorizeWithBlockFrequency) {
1279 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1280 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1281 LoopEntryFreq < ColdEntryFreq)
1285 // Check the function attributes to see if implicit floats are allowed.a
1286 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1287 // an integer loop and the vector instructions selected are purely integer
1288 // vector instructions?
1289 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1290 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1291 "attribute is used.\n");
1292 emitOptimizationRemarkAnalysis(
1293 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1294 "loop not vectorized due to NoImplicitFloat attribute");
1295 emitMissedWarning(F, L, Hints);
1299 // Select the optimal vectorization factor.
1300 const LoopVectorizationCostModel::VectorizationFactor VF =
1301 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1303 LoopVectorizeHints::FK_Enabled);
1305 // Select the unroll factor.
1307 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1309 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1310 << DebugLocStr << '\n');
1311 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1313 if (VF.Width == 1) {
1314 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1317 emitOptimizationRemarkAnalysis(
1318 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1319 "not beneficial to vectorize and user disabled interleaving");
1322 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1324 // Report the unrolling decision.
1325 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1326 Twine("unrolled with interleaving factor " +
1328 " (vectorization not beneficial)"));
1330 // We decided not to vectorize, but we may want to unroll.
1332 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1333 Unroller.vectorize(&LVL);
1335 // If we decided that it is *legal* to vectorize the loop then do it.
1336 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1340 // Report the vectorization decision.
1341 emitOptimizationRemark(
1342 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1343 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1344 ", unrolling interleave factor: " + Twine(UF) + ")");
1347 // Mark the loop as already vectorized to avoid vectorizing again.
1348 Hints.setAlreadyVectorized(L);
1350 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1354 void getAnalysisUsage(AnalysisUsage &AU) const override {
1355 AU.addRequiredID(LoopSimplifyID);
1356 AU.addRequiredID(LCSSAID);
1357 AU.addRequired<BlockFrequencyInfo>();
1358 AU.addRequired<DominatorTreeWrapperPass>();
1359 AU.addRequired<LoopInfo>();
1360 AU.addRequired<ScalarEvolution>();
1361 AU.addRequired<TargetTransformInfo>();
1362 AU.addPreserved<LoopInfo>();
1363 AU.addPreserved<DominatorTreeWrapperPass>();
1368 } // end anonymous namespace
1370 //===----------------------------------------------------------------------===//
1371 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1372 // LoopVectorizationCostModel.
1373 //===----------------------------------------------------------------------===//
1375 static Value *stripIntegerCast(Value *V) {
1376 if (CastInst *CI = dyn_cast<CastInst>(V))
1377 if (CI->getOperand(0)->getType()->isIntegerTy())
1378 return CI->getOperand(0);
1382 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1384 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1386 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1387 ValueToValueMap &PtrToStride,
1388 Value *Ptr, Value *OrigPtr = nullptr) {
1390 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1392 // If there is an entry in the map return the SCEV of the pointer with the
1393 // symbolic stride replaced by one.
1394 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1395 if (SI != PtrToStride.end()) {
1396 Value *StrideVal = SI->second;
1399 StrideVal = stripIntegerCast(StrideVal);
1401 // Replace symbolic stride by one.
1402 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1403 ValueToValueMap RewriteMap;
1404 RewriteMap[StrideVal] = One;
1407 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1408 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1413 // Otherwise, just return the SCEV of the original pointer.
1414 return SE->getSCEV(Ptr);
1417 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1418 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1419 ValueToValueMap &Strides) {
1420 // Get the stride replaced scev.
1421 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1422 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1423 assert(AR && "Invalid addrec expression");
1424 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1425 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1426 Pointers.push_back(Ptr);
1427 Starts.push_back(AR->getStart());
1428 Ends.push_back(ScEnd);
1429 IsWritePtr.push_back(WritePtr);
1430 DependencySetId.push_back(DepSetId);
1433 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1434 // We need to place the broadcast of invariant variables outside the loop.
1435 Instruction *Instr = dyn_cast<Instruction>(V);
1437 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1438 Instr->getParent()) != LoopVectorBody.end());
1439 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1441 // Place the code for broadcasting invariant variables in the new preheader.
1442 IRBuilder<>::InsertPointGuard Guard(Builder);
1444 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1446 // Broadcast the scalar into all locations in the vector.
1447 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1452 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1454 assert(Val->getType()->isVectorTy() && "Must be a vector");
1455 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1456 "Elem must be an integer");
1457 // Create the types.
1458 Type *ITy = Val->getType()->getScalarType();
1459 VectorType *Ty = cast<VectorType>(Val->getType());
1460 int VLen = Ty->getNumElements();
1461 SmallVector<Constant*, 8> Indices;
1463 // Create a vector of consecutive numbers from zero to VF.
1464 for (int i = 0; i < VLen; ++i) {
1465 int64_t Idx = Negate ? (-i) : i;
1466 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1469 // Add the consecutive indices to the vector value.
1470 Constant *Cv = ConstantVector::get(Indices);
1471 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1472 return Builder.CreateAdd(Val, Cv, "induction");
1475 /// \brief Find the operand of the GEP that should be checked for consecutive
1476 /// stores. This ignores trailing indices that have no effect on the final
1478 static unsigned getGEPInductionOperand(const DataLayout *DL,
1479 const GetElementPtrInst *Gep) {
1480 unsigned LastOperand = Gep->getNumOperands() - 1;
1481 unsigned GEPAllocSize = DL->getTypeAllocSize(
1482 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1484 // Walk backwards and try to peel off zeros.
1485 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1486 // Find the type we're currently indexing into.
1487 gep_type_iterator GEPTI = gep_type_begin(Gep);
1488 std::advance(GEPTI, LastOperand - 1);
1490 // If it's a type with the same allocation size as the result of the GEP we
1491 // can peel off the zero index.
1492 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1500 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1501 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1502 // Make sure that the pointer does not point to structs.
1503 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1506 // If this value is a pointer induction variable we know it is consecutive.
1507 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1508 if (Phi && Inductions.count(Phi)) {
1509 InductionInfo II = Inductions[Phi];
1510 if (IK_PtrInduction == II.IK)
1512 else if (IK_ReversePtrInduction == II.IK)
1516 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1520 unsigned NumOperands = Gep->getNumOperands();
1521 Value *GpPtr = Gep->getPointerOperand();
1522 // If this GEP value is a consecutive pointer induction variable and all of
1523 // the indices are constant then we know it is consecutive. We can
1524 Phi = dyn_cast<PHINode>(GpPtr);
1525 if (Phi && Inductions.count(Phi)) {
1527 // Make sure that the pointer does not point to structs.
1528 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1529 if (GepPtrType->getElementType()->isAggregateType())
1532 // Make sure that all of the index operands are loop invariant.
1533 for (unsigned i = 1; i < NumOperands; ++i)
1534 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1537 InductionInfo II = Inductions[Phi];
1538 if (IK_PtrInduction == II.IK)
1540 else if (IK_ReversePtrInduction == II.IK)
1544 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1546 // Check that all of the gep indices are uniform except for our induction
1548 for (unsigned i = 0; i != NumOperands; ++i)
1549 if (i != InductionOperand &&
1550 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1553 // We can emit wide load/stores only if the last non-zero index is the
1554 // induction variable.
1555 const SCEV *Last = nullptr;
1556 if (!Strides.count(Gep))
1557 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1559 // Because of the multiplication by a stride we can have a s/zext cast.
1560 // We are going to replace this stride by 1 so the cast is safe to ignore.
1562 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1563 // %0 = trunc i64 %indvars.iv to i32
1564 // %mul = mul i32 %0, %Stride1
1565 // %idxprom = zext i32 %mul to i64 << Safe cast.
1566 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1568 Last = replaceSymbolicStrideSCEV(SE, Strides,
1569 Gep->getOperand(InductionOperand), Gep);
1570 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1572 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1576 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1577 const SCEV *Step = AR->getStepRecurrence(*SE);
1579 // The memory is consecutive because the last index is consecutive
1580 // and all other indices are loop invariant.
1583 if (Step->isAllOnesValue())
1590 bool LoopVectorizationLegality::isUniform(Value *V) {
1591 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1594 InnerLoopVectorizer::VectorParts&
1595 InnerLoopVectorizer::getVectorValue(Value *V) {
1596 assert(V != Induction && "The new induction variable should not be used.");
1597 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1599 // If we have a stride that is replaced by one, do it here.
1600 if (Legal->hasStride(V))
1601 V = ConstantInt::get(V->getType(), 1);
1603 // If we have this scalar in the map, return it.
1604 if (WidenMap.has(V))
1605 return WidenMap.get(V);
1607 // If this scalar is unknown, assume that it is a constant or that it is
1608 // loop invariant. Broadcast V and save the value for future uses.
1609 Value *B = getBroadcastInstrs(V);
1610 return WidenMap.splat(V, B);
1613 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1614 assert(Vec->getType()->isVectorTy() && "Invalid type");
1615 SmallVector<Constant*, 8> ShuffleMask;
1616 for (unsigned i = 0; i < VF; ++i)
1617 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1619 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1620 ConstantVector::get(ShuffleMask),
1624 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1625 // Attempt to issue a wide load.
1626 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1627 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1629 assert((LI || SI) && "Invalid Load/Store instruction");
1631 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1632 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1633 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1634 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1635 // An alignment of 0 means target abi alignment. We need to use the scalar's
1636 // target abi alignment in such a case.
1638 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1639 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1640 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1641 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1643 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1644 return scalarizeInstruction(Instr, true);
1646 if (ScalarAllocatedSize != VectorElementSize)
1647 return scalarizeInstruction(Instr);
1649 // If the pointer is loop invariant or if it is non-consecutive,
1650 // scalarize the load.
1651 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1652 bool Reverse = ConsecutiveStride < 0;
1653 bool UniformLoad = LI && Legal->isUniform(Ptr);
1654 if (!ConsecutiveStride || UniformLoad)
1655 return scalarizeInstruction(Instr);
1657 Constant *Zero = Builder.getInt32(0);
1658 VectorParts &Entry = WidenMap.get(Instr);
1660 // Handle consecutive loads/stores.
1661 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1662 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1663 setDebugLocFromInst(Builder, Gep);
1664 Value *PtrOperand = Gep->getPointerOperand();
1665 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1666 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1668 // Create the new GEP with the new induction variable.
1669 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1670 Gep2->setOperand(0, FirstBasePtr);
1671 Gep2->setName("gep.indvar.base");
1672 Ptr = Builder.Insert(Gep2);
1674 setDebugLocFromInst(Builder, Gep);
1675 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1676 OrigLoop) && "Base ptr must be invariant");
1678 // The last index does not have to be the induction. It can be
1679 // consecutive and be a function of the index. For example A[I+1];
1680 unsigned NumOperands = Gep->getNumOperands();
1681 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1682 // Create the new GEP with the new induction variable.
1683 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1685 for (unsigned i = 0; i < NumOperands; ++i) {
1686 Value *GepOperand = Gep->getOperand(i);
1687 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1689 // Update last index or loop invariant instruction anchored in loop.
1690 if (i == InductionOperand ||
1691 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1692 assert((i == InductionOperand ||
1693 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1694 "Must be last index or loop invariant");
1696 VectorParts &GEPParts = getVectorValue(GepOperand);
1697 Value *Index = GEPParts[0];
1698 Index = Builder.CreateExtractElement(Index, Zero);
1699 Gep2->setOperand(i, Index);
1700 Gep2->setName("gep.indvar.idx");
1703 Ptr = Builder.Insert(Gep2);
1705 // Use the induction element ptr.
1706 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1707 setDebugLocFromInst(Builder, Ptr);
1708 VectorParts &PtrVal = getVectorValue(Ptr);
1709 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1714 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1715 "We do not allow storing to uniform addresses");
1716 setDebugLocFromInst(Builder, SI);
1717 // We don't want to update the value in the map as it might be used in
1718 // another expression. So don't use a reference type for "StoredVal".
1719 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1721 for (unsigned Part = 0; Part < UF; ++Part) {
1722 // Calculate the pointer for the specific unroll-part.
1723 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1726 // If we store to reverse consecutive memory locations then we need
1727 // to reverse the order of elements in the stored value.
1728 StoredVal[Part] = reverseVector(StoredVal[Part]);
1729 // If the address is consecutive but reversed, then the
1730 // wide store needs to start at the last vector element.
1731 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1732 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1735 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1736 DataTy->getPointerTo(AddressSpace));
1737 Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
1743 assert(LI && "Must have a load instruction");
1744 setDebugLocFromInst(Builder, LI);
1745 for (unsigned Part = 0; Part < UF; ++Part) {
1746 // Calculate the pointer for the specific unroll-part.
1747 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1750 // If the address is consecutive but reversed, then the
1751 // wide store needs to start at the last vector element.
1752 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1753 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1756 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1757 DataTy->getPointerTo(AddressSpace));
1758 Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
1759 cast<LoadInst>(LI)->setAlignment(Alignment);
1760 Entry[Part] = Reverse ? reverseVector(LI) : LI;
1764 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1765 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1766 // Holds vector parameters or scalars, in case of uniform vals.
1767 SmallVector<VectorParts, 4> Params;
1769 setDebugLocFromInst(Builder, Instr);
1771 // Find all of the vectorized parameters.
1772 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1773 Value *SrcOp = Instr->getOperand(op);
1775 // If we are accessing the old induction variable, use the new one.
1776 if (SrcOp == OldInduction) {
1777 Params.push_back(getVectorValue(SrcOp));
1781 // Try using previously calculated values.
1782 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1784 // If the src is an instruction that appeared earlier in the basic block
1785 // then it should already be vectorized.
1786 if (SrcInst && OrigLoop->contains(SrcInst)) {
1787 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1788 // The parameter is a vector value from earlier.
1789 Params.push_back(WidenMap.get(SrcInst));
1791 // The parameter is a scalar from outside the loop. Maybe even a constant.
1792 VectorParts Scalars;
1793 Scalars.append(UF, SrcOp);
1794 Params.push_back(Scalars);
1798 assert(Params.size() == Instr->getNumOperands() &&
1799 "Invalid number of operands");
1801 // Does this instruction return a value ?
1802 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1804 Value *UndefVec = IsVoidRetTy ? nullptr :
1805 UndefValue::get(VectorType::get(Instr->getType(), VF));
1806 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1807 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1809 Instruction *InsertPt = Builder.GetInsertPoint();
1810 BasicBlock *IfBlock = Builder.GetInsertBlock();
1811 BasicBlock *CondBlock = nullptr;
1814 Loop *VectorLp = nullptr;
1815 if (IfPredicateStore) {
1816 assert(Instr->getParent()->getSinglePredecessor() &&
1817 "Only support single predecessor blocks");
1818 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1819 Instr->getParent());
1820 VectorLp = LI->getLoopFor(IfBlock);
1821 assert(VectorLp && "Must have a loop for this block");
1824 // For each vector unroll 'part':
1825 for (unsigned Part = 0; Part < UF; ++Part) {
1826 // For each scalar that we create:
1827 for (unsigned Width = 0; Width < VF; ++Width) {
1830 Value *Cmp = nullptr;
1831 if (IfPredicateStore) {
1832 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1833 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1834 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1835 LoopVectorBody.push_back(CondBlock);
1836 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1837 // Update Builder with newly created basic block.
1838 Builder.SetInsertPoint(InsertPt);
1841 Instruction *Cloned = Instr->clone();
1843 Cloned->setName(Instr->getName() + ".cloned");
1844 // Replace the operands of the cloned instructions with extracted scalars.
1845 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1846 Value *Op = Params[op][Part];
1847 // Param is a vector. Need to extract the right lane.
1848 if (Op->getType()->isVectorTy())
1849 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1850 Cloned->setOperand(op, Op);
1853 // Place the cloned scalar in the new loop.
1854 Builder.Insert(Cloned);
1856 // If the original scalar returns a value we need to place it in a vector
1857 // so that future users will be able to use it.
1859 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1860 Builder.getInt32(Width));
1862 if (IfPredicateStore) {
1863 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1864 LoopVectorBody.push_back(NewIfBlock);
1865 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1866 Builder.SetInsertPoint(InsertPt);
1867 Instruction *OldBr = IfBlock->getTerminator();
1868 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1869 OldBr->eraseFromParent();
1870 IfBlock = NewIfBlock;
1876 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1880 if (Instruction *I = dyn_cast<Instruction>(V))
1881 return I->getParent() == Loc->getParent() ? I : nullptr;
1885 std::pair<Instruction *, Instruction *>
1886 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1887 Instruction *tnullptr = nullptr;
1888 if (!Legal->mustCheckStrides())
1889 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1891 IRBuilder<> ChkBuilder(Loc);
1894 Value *Check = nullptr;
1895 Instruction *FirstInst = nullptr;
1896 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1897 SE = Legal->strides_end();
1899 Value *Ptr = stripIntegerCast(*SI);
1900 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1902 // Store the first instruction we create.
1903 FirstInst = getFirstInst(FirstInst, C, Loc);
1905 Check = ChkBuilder.CreateOr(Check, C);
1910 // We have to do this trickery because the IRBuilder might fold the check to a
1911 // constant expression in which case there is no Instruction anchored in a
1913 LLVMContext &Ctx = Loc->getContext();
1914 Instruction *TheCheck =
1915 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1916 ChkBuilder.Insert(TheCheck, "stride.not.one");
1917 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1919 return std::make_pair(FirstInst, TheCheck);
1922 std::pair<Instruction *, Instruction *>
1923 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1924 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1925 Legal->getRuntimePointerCheck();
1927 Instruction *tnullptr = nullptr;
1928 if (!PtrRtCheck->Need)
1929 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1931 unsigned NumPointers = PtrRtCheck->Pointers.size();
1932 SmallVector<TrackingVH<Value> , 2> Starts;
1933 SmallVector<TrackingVH<Value> , 2> Ends;
1935 LLVMContext &Ctx = Loc->getContext();
1936 SCEVExpander Exp(*SE, "induction");
1937 Instruction *FirstInst = nullptr;
1939 for (unsigned i = 0; i < NumPointers; ++i) {
1940 Value *Ptr = PtrRtCheck->Pointers[i];
1941 const SCEV *Sc = SE->getSCEV(Ptr);
1943 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1944 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1946 Starts.push_back(Ptr);
1947 Ends.push_back(Ptr);
1949 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1950 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1952 // Use this type for pointer arithmetic.
1953 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1955 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1956 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
1957 Starts.push_back(Start);
1958 Ends.push_back(End);
1962 IRBuilder<> ChkBuilder(Loc);
1963 // Our instructions might fold to a constant.
1964 Value *MemoryRuntimeCheck = nullptr;
1965 for (unsigned i = 0; i < NumPointers; ++i) {
1966 for (unsigned j = i+1; j < NumPointers; ++j) {
1967 // No need to check if two readonly pointers intersect.
1968 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
1971 // Only need to check pointers between two different dependency sets.
1972 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
1975 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1976 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1978 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1979 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1980 "Trying to bounds check pointers with different address spaces");
1982 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1983 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1985 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1986 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1987 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1988 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1990 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1991 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1992 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1993 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1994 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1995 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1996 if (MemoryRuntimeCheck) {
1997 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1999 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2001 MemoryRuntimeCheck = IsConflict;
2005 // We have to do this trickery because the IRBuilder might fold the check to a
2006 // constant expression in which case there is no Instruction anchored in a
2008 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2009 ConstantInt::getTrue(Ctx));
2010 ChkBuilder.Insert(Check, "memcheck.conflict");
2011 FirstInst = getFirstInst(FirstInst, Check, Loc);
2012 return std::make_pair(FirstInst, Check);
2015 void InnerLoopVectorizer::createEmptyLoop() {
2017 In this function we generate a new loop. The new loop will contain
2018 the vectorized instructions while the old loop will continue to run the
2021 [ ] <-- Back-edge taken count overflow check.
2024 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2027 || [ ] <-- vector pre header.
2031 || [ ]_| <-- vector loop.
2034 | >[ ] <--- middle-block.
2037 -|- >[ ] <--- new preheader.
2041 | [ ]_| <-- old scalar loop to handle remainder.
2044 >[ ] <-- exit block.
2048 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2049 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2050 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2051 assert(BypassBlock && "Invalid loop structure");
2052 assert(ExitBlock && "Must have an exit block");
2054 // Some loops have a single integer induction variable, while other loops
2055 // don't. One example is c++ iterators that often have multiple pointer
2056 // induction variables. In the code below we also support a case where we
2057 // don't have a single induction variable.
2058 OldInduction = Legal->getInduction();
2059 Type *IdxTy = Legal->getWidestInductionType();
2061 // Find the loop boundaries.
2062 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2063 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2065 // The exit count might have the type of i64 while the phi is i32. This can
2066 // happen if we have an induction variable that is sign extended before the
2067 // compare. The only way that we get a backedge taken count is that the
2068 // induction variable was signed and as such will not overflow. In such a case
2069 // truncation is legal.
2070 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2071 IdxTy->getPrimitiveSizeInBits())
2072 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2074 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2075 // Get the total trip count from the count by adding 1.
2076 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2077 SE->getConstant(BackedgeTakeCount->getType(), 1));
2079 // Expand the trip count and place the new instructions in the preheader.
2080 // Notice that the pre-header does not change, only the loop body.
2081 SCEVExpander Exp(*SE, "induction");
2083 // We need to test whether the backedge-taken count is uint##_max. Adding one
2084 // to it will cause overflow and an incorrect loop trip count in the vector
2085 // body. In case of overflow we want to directly jump to the scalar remainder
2087 Value *BackedgeCount =
2088 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2089 BypassBlock->getTerminator());
2090 if (BackedgeCount->getType()->isPointerTy())
2091 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2092 "backedge.ptrcnt.to.int",
2093 BypassBlock->getTerminator());
2094 Instruction *CheckBCOverflow =
2095 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2096 Constant::getAllOnesValue(BackedgeCount->getType()),
2097 "backedge.overflow", BypassBlock->getTerminator());
2099 // The loop index does not have to start at Zero. Find the original start
2100 // value from the induction PHI node. If we don't have an induction variable
2101 // then we know that it starts at zero.
2102 Builder.SetInsertPoint(BypassBlock->getTerminator());
2103 Value *StartIdx = ExtendedIdx = OldInduction ?
2104 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2106 ConstantInt::get(IdxTy, 0);
2108 // We need an instruction to anchor the overflow check on. StartIdx needs to
2109 // be defined before the overflow check branch. Because the scalar preheader
2110 // is going to merge the start index and so the overflow branch block needs to
2111 // contain a definition of the start index.
2112 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2113 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2114 BypassBlock->getTerminator());
2116 // Count holds the overall loop count (N).
2117 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2118 BypassBlock->getTerminator());
2120 LoopBypassBlocks.push_back(BypassBlock);
2122 // Split the single block loop into the two loop structure described above.
2123 BasicBlock *VectorPH =
2124 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2125 BasicBlock *VecBody =
2126 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2127 BasicBlock *MiddleBlock =
2128 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2129 BasicBlock *ScalarPH =
2130 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2132 // Create and register the new vector loop.
2133 Loop* Lp = new Loop();
2134 Loop *ParentLoop = OrigLoop->getParentLoop();
2136 // Insert the new loop into the loop nest and register the new basic blocks
2137 // before calling any utilities such as SCEV that require valid LoopInfo.
2139 ParentLoop->addChildLoop(Lp);
2140 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2141 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2142 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2144 LI->addTopLevelLoop(Lp);
2146 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2148 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2150 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2152 // Generate the induction variable.
2153 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2154 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2155 // The loop step is equal to the vectorization factor (num of SIMD elements)
2156 // times the unroll factor (num of SIMD instructions).
2157 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2159 // This is the IR builder that we use to add all of the logic for bypassing
2160 // the new vector loop.
2161 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2162 setDebugLocFromInst(BypassBuilder,
2163 getDebugLocFromInstOrOperands(OldInduction));
2165 // We may need to extend the index in case there is a type mismatch.
2166 // We know that the count starts at zero and does not overflow.
2167 if (Count->getType() != IdxTy) {
2168 // The exit count can be of pointer type. Convert it to the correct
2170 if (ExitCount->getType()->isPointerTy())
2171 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2173 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2176 // Add the start index to the loop count to get the new end index.
2177 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2179 // Now we need to generate the expression for N - (N % VF), which is
2180 // the part that the vectorized body will execute.
2181 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2182 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2183 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2184 "end.idx.rnd.down");
2186 // Now, compare the new count to zero. If it is zero skip the vector loop and
2187 // jump to the scalar loop.
2189 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2191 BasicBlock *LastBypassBlock = BypassBlock;
2193 // Generate code to check that the loops trip count that we computed by adding
2194 // one to the backedge-taken count will not overflow.
2196 auto PastOverflowCheck =
2197 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2198 BasicBlock *CheckBlock =
2199 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2201 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2202 LoopBypassBlocks.push_back(CheckBlock);
2203 Instruction *OldTerm = LastBypassBlock->getTerminator();
2204 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2205 OldTerm->eraseFromParent();
2206 LastBypassBlock = CheckBlock;
2209 // Generate the code to check that the strides we assumed to be one are really
2210 // one. We want the new basic block to start at the first instruction in a
2211 // sequence of instructions that form a check.
2212 Instruction *StrideCheck;
2213 Instruction *FirstCheckInst;
2214 std::tie(FirstCheckInst, StrideCheck) =
2215 addStrideCheck(LastBypassBlock->getTerminator());
2217 // Create a new block containing the stride check.
2218 BasicBlock *CheckBlock =
2219 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2221 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2222 LoopBypassBlocks.push_back(CheckBlock);
2224 // Replace the branch into the memory check block with a conditional branch
2225 // for the "few elements case".
2226 Instruction *OldTerm = LastBypassBlock->getTerminator();
2227 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2228 OldTerm->eraseFromParent();
2231 LastBypassBlock = CheckBlock;
2234 // Generate the code that checks in runtime if arrays overlap. We put the
2235 // checks into a separate block to make the more common case of few elements
2237 Instruction *MemRuntimeCheck;
2238 std::tie(FirstCheckInst, MemRuntimeCheck) =
2239 addRuntimeCheck(LastBypassBlock->getTerminator());
2240 if (MemRuntimeCheck) {
2241 // Create a new block containing the memory check.
2242 BasicBlock *CheckBlock =
2243 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2245 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2246 LoopBypassBlocks.push_back(CheckBlock);
2248 // Replace the branch into the memory check block with a conditional branch
2249 // for the "few elements case".
2250 Instruction *OldTerm = LastBypassBlock->getTerminator();
2251 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2252 OldTerm->eraseFromParent();
2254 Cmp = MemRuntimeCheck;
2255 LastBypassBlock = CheckBlock;
2258 LastBypassBlock->getTerminator()->eraseFromParent();
2259 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2262 // We are going to resume the execution of the scalar loop.
2263 // Go over all of the induction variables that we found and fix the
2264 // PHIs that are left in the scalar version of the loop.
2265 // The starting values of PHI nodes depend on the counter of the last
2266 // iteration in the vectorized loop.
2267 // If we come from a bypass edge then we need to start from the original
2270 // This variable saves the new starting index for the scalar loop.
2271 PHINode *ResumeIndex = nullptr;
2272 LoopVectorizationLegality::InductionList::iterator I, E;
2273 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2274 // Set builder to point to last bypass block.
2275 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2276 for (I = List->begin(), E = List->end(); I != E; ++I) {
2277 PHINode *OrigPhi = I->first;
2278 LoopVectorizationLegality::InductionInfo II = I->second;
2280 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2281 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2282 MiddleBlock->getTerminator());
2283 // We might have extended the type of the induction variable but we need a
2284 // truncated version for the scalar loop.
2285 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2286 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2287 MiddleBlock->getTerminator()) : nullptr;
2289 // Create phi nodes to merge from the backedge-taken check block.
2290 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2291 ScalarPH->getTerminator());
2292 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2294 PHINode *BCTruncResumeVal = nullptr;
2295 if (OrigPhi == OldInduction) {
2297 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2298 ScalarPH->getTerminator());
2299 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2302 Value *EndValue = nullptr;
2304 case LoopVectorizationLegality::IK_NoInduction:
2305 llvm_unreachable("Unknown induction");
2306 case LoopVectorizationLegality::IK_IntInduction: {
2307 // Handle the integer induction counter.
2308 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2310 // We have the canonical induction variable.
2311 if (OrigPhi == OldInduction) {
2312 // Create a truncated version of the resume value for the scalar loop,
2313 // we might have promoted the type to a larger width.
2315 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2316 // The new PHI merges the original incoming value, in case of a bypass,
2317 // or the value at the end of the vectorized loop.
2318 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2319 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2320 TruncResumeVal->addIncoming(EndValue, VecBody);
2322 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2324 // We know what the end value is.
2325 EndValue = IdxEndRoundDown;
2326 // We also know which PHI node holds it.
2327 ResumeIndex = ResumeVal;
2331 // Not the canonical induction variable - add the vector loop count to the
2333 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2334 II.StartValue->getType(),
2336 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2339 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2340 // Convert the CountRoundDown variable to the PHI size.
2341 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2342 II.StartValue->getType(),
2344 // Handle reverse integer induction counter.
2345 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2348 case LoopVectorizationLegality::IK_PtrInduction: {
2349 // For pointer induction variables, calculate the offset using
2351 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2355 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2356 // The value at the end of the loop for the reverse pointer is calculated
2357 // by creating a GEP with a negative index starting from the start value.
2358 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2359 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2361 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2367 // The new PHI merges the original incoming value, in case of a bypass,
2368 // or the value at the end of the vectorized loop.
2369 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2370 if (OrigPhi == OldInduction)
2371 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2373 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2375 ResumeVal->addIncoming(EndValue, VecBody);
2377 // Fix the scalar body counter (PHI node).
2378 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2380 // The old induction's phi node in the scalar body needs the truncated
2382 if (OrigPhi == OldInduction) {
2383 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2384 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2386 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2387 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2391 // If we are generating a new induction variable then we also need to
2392 // generate the code that calculates the exit value. This value is not
2393 // simply the end of the counter because we may skip the vectorized body
2394 // in case of a runtime check.
2396 assert(!ResumeIndex && "Unexpected resume value found");
2397 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2398 MiddleBlock->getTerminator());
2399 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2400 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2401 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2404 // Make sure that we found the index where scalar loop needs to continue.
2405 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2406 "Invalid resume Index");
2408 // Add a check in the middle block to see if we have completed
2409 // all of the iterations in the first vector loop.
2410 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2411 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2412 ResumeIndex, "cmp.n",
2413 MiddleBlock->getTerminator());
2415 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2416 // Remove the old terminator.
2417 MiddleBlock->getTerminator()->eraseFromParent();
2419 // Create i+1 and fill the PHINode.
2420 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2421 Induction->addIncoming(StartIdx, VectorPH);
2422 Induction->addIncoming(NextIdx, VecBody);
2423 // Create the compare.
2424 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2425 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2427 // Now we have two terminators. Remove the old one from the block.
2428 VecBody->getTerminator()->eraseFromParent();
2430 // Get ready to start creating new instructions into the vectorized body.
2431 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2434 LoopVectorPreHeader = VectorPH;
2435 LoopScalarPreHeader = ScalarPH;
2436 LoopMiddleBlock = MiddleBlock;
2437 LoopExitBlock = ExitBlock;
2438 LoopVectorBody.push_back(VecBody);
2439 LoopScalarBody = OldBasicBlock;
2441 LoopVectorizeHints Hints(Lp, true);
2442 Hints.setAlreadyVectorized(Lp);
2445 /// This function returns the identity element (or neutral element) for
2446 /// the operation K.
2448 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2453 // Adding, Xoring, Oring zero to a number does not change it.
2454 return ConstantInt::get(Tp, 0);
2455 case RK_IntegerMult:
2456 // Multiplying a number by 1 does not change it.
2457 return ConstantInt::get(Tp, 1);
2459 // AND-ing a number with an all-1 value does not change it.
2460 return ConstantInt::get(Tp, -1, true);
2462 // Multiplying a number by 1 does not change it.
2463 return ConstantFP::get(Tp, 1.0L);
2465 // Adding zero to a number does not change it.
2466 return ConstantFP::get(Tp, 0.0L);
2468 llvm_unreachable("Unknown reduction kind");
2472 /// This function translates the reduction kind to an LLVM binary operator.
2474 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2476 case LoopVectorizationLegality::RK_IntegerAdd:
2477 return Instruction::Add;
2478 case LoopVectorizationLegality::RK_IntegerMult:
2479 return Instruction::Mul;
2480 case LoopVectorizationLegality::RK_IntegerOr:
2481 return Instruction::Or;
2482 case LoopVectorizationLegality::RK_IntegerAnd:
2483 return Instruction::And;
2484 case LoopVectorizationLegality::RK_IntegerXor:
2485 return Instruction::Xor;
2486 case LoopVectorizationLegality::RK_FloatMult:
2487 return Instruction::FMul;
2488 case LoopVectorizationLegality::RK_FloatAdd:
2489 return Instruction::FAdd;
2490 case LoopVectorizationLegality::RK_IntegerMinMax:
2491 return Instruction::ICmp;
2492 case LoopVectorizationLegality::RK_FloatMinMax:
2493 return Instruction::FCmp;
2495 llvm_unreachable("Unknown reduction operation");
2499 Value *createMinMaxOp(IRBuilder<> &Builder,
2500 LoopVectorizationLegality::MinMaxReductionKind RK,
2503 CmpInst::Predicate P = CmpInst::ICMP_NE;
2506 llvm_unreachable("Unknown min/max reduction kind");
2507 case LoopVectorizationLegality::MRK_UIntMin:
2508 P = CmpInst::ICMP_ULT;
2510 case LoopVectorizationLegality::MRK_UIntMax:
2511 P = CmpInst::ICMP_UGT;
2513 case LoopVectorizationLegality::MRK_SIntMin:
2514 P = CmpInst::ICMP_SLT;
2516 case LoopVectorizationLegality::MRK_SIntMax:
2517 P = CmpInst::ICMP_SGT;
2519 case LoopVectorizationLegality::MRK_FloatMin:
2520 P = CmpInst::FCMP_OLT;
2522 case LoopVectorizationLegality::MRK_FloatMax:
2523 P = CmpInst::FCMP_OGT;
2528 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2529 RK == LoopVectorizationLegality::MRK_FloatMax)
2530 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2532 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2534 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2539 struct CSEDenseMapInfo {
2540 static bool canHandle(Instruction *I) {
2541 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2542 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2544 static inline Instruction *getEmptyKey() {
2545 return DenseMapInfo<Instruction *>::getEmptyKey();
2547 static inline Instruction *getTombstoneKey() {
2548 return DenseMapInfo<Instruction *>::getTombstoneKey();
2550 static unsigned getHashValue(Instruction *I) {
2551 assert(canHandle(I) && "Unknown instruction!");
2552 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2553 I->value_op_end()));
2555 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2556 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2557 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2559 return LHS->isIdenticalTo(RHS);
2564 /// \brief Check whether this block is a predicated block.
2565 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2566 /// = ...; " blocks. We start with one vectorized basic block. For every
2567 /// conditional block we split this vectorized block. Therefore, every second
2568 /// block will be a predicated one.
2569 static bool isPredicatedBlock(unsigned BlockNum) {
2570 return BlockNum % 2;
2573 ///\brief Perform cse of induction variable instructions.
2574 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2575 // Perform simple cse.
2576 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2577 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2578 BasicBlock *BB = BBs[i];
2579 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2580 Instruction *In = I++;
2582 if (!CSEDenseMapInfo::canHandle(In))
2585 // Check if we can replace this instruction with any of the
2586 // visited instructions.
2587 if (Instruction *V = CSEMap.lookup(In)) {
2588 In->replaceAllUsesWith(V);
2589 In->eraseFromParent();
2592 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2593 // ...;" blocks for predicated stores. Every second block is a predicated
2595 if (isPredicatedBlock(i))
2603 /// \brief Adds a 'fast' flag to floating point operations.
2604 static Value *addFastMathFlag(Value *V) {
2605 if (isa<FPMathOperator>(V)){
2606 FastMathFlags Flags;
2607 Flags.setUnsafeAlgebra();
2608 cast<Instruction>(V)->setFastMathFlags(Flags);
2613 void InnerLoopVectorizer::vectorizeLoop() {
2614 //===------------------------------------------------===//
2616 // Notice: any optimization or new instruction that go
2617 // into the code below should be also be implemented in
2620 //===------------------------------------------------===//
2621 Constant *Zero = Builder.getInt32(0);
2623 // In order to support reduction variables we need to be able to vectorize
2624 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2625 // stages. First, we create a new vector PHI node with no incoming edges.
2626 // We use this value when we vectorize all of the instructions that use the
2627 // PHI. Next, after all of the instructions in the block are complete we
2628 // add the new incoming edges to the PHI. At this point all of the
2629 // instructions in the basic block are vectorized, so we can use them to
2630 // construct the PHI.
2631 PhiVector RdxPHIsToFix;
2633 // Scan the loop in a topological order to ensure that defs are vectorized
2635 LoopBlocksDFS DFS(OrigLoop);
2638 // Vectorize all of the blocks in the original loop.
2639 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2640 be = DFS.endRPO(); bb != be; ++bb)
2641 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2643 // At this point every instruction in the original loop is widened to
2644 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2645 // that we vectorized. The PHI nodes are currently empty because we did
2646 // not want to introduce cycles. Notice that the remaining PHI nodes
2647 // that we need to fix are reduction variables.
2649 // Create the 'reduced' values for each of the induction vars.
2650 // The reduced values are the vector values that we scalarize and combine
2651 // after the loop is finished.
2652 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2654 PHINode *RdxPhi = *it;
2655 assert(RdxPhi && "Unable to recover vectorized PHI");
2657 // Find the reduction variable descriptor.
2658 assert(Legal->getReductionVars()->count(RdxPhi) &&
2659 "Unable to find the reduction variable");
2660 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2661 (*Legal->getReductionVars())[RdxPhi];
2663 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2665 // We need to generate a reduction vector from the incoming scalar.
2666 // To do so, we need to generate the 'identity' vector and override
2667 // one of the elements with the incoming scalar reduction. We need
2668 // to do it in the vector-loop preheader.
2669 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2671 // This is the vector-clone of the value that leaves the loop.
2672 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2673 Type *VecTy = VectorExit[0]->getType();
2675 // Find the reduction identity variable. Zero for addition, or, xor,
2676 // one for multiplication, -1 for And.
2679 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2680 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2681 // MinMax reduction have the start value as their identify.
2683 VectorStart = Identity = RdxDesc.StartValue;
2685 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2690 // Handle other reduction kinds:
2692 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2693 VecTy->getScalarType());
2696 // This vector is the Identity vector where the first element is the
2697 // incoming scalar reduction.
2698 VectorStart = RdxDesc.StartValue;
2700 Identity = ConstantVector::getSplat(VF, Iden);
2702 // This vector is the Identity vector where the first element is the
2703 // incoming scalar reduction.
2704 VectorStart = Builder.CreateInsertElement(Identity,
2705 RdxDesc.StartValue, Zero);
2709 // Fix the vector-loop phi.
2710 // We created the induction variable so we know that the
2711 // preheader is the first entry.
2712 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2714 // Reductions do not have to start at zero. They can start with
2715 // any loop invariant values.
2716 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2717 BasicBlock *Latch = OrigLoop->getLoopLatch();
2718 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2719 VectorParts &Val = getVectorValue(LoopVal);
2720 for (unsigned part = 0; part < UF; ++part) {
2721 // Make sure to add the reduction stat value only to the
2722 // first unroll part.
2723 Value *StartVal = (part == 0) ? VectorStart : Identity;
2724 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2725 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2726 LoopVectorBody.back());
2729 // Before each round, move the insertion point right between
2730 // the PHIs and the values we are going to write.
2731 // This allows us to write both PHINodes and the extractelement
2733 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2735 VectorParts RdxParts;
2736 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2737 for (unsigned part = 0; part < UF; ++part) {
2738 // This PHINode contains the vectorized reduction variable, or
2739 // the initial value vector, if we bypass the vector loop.
2740 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2741 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2742 Value *StartVal = (part == 0) ? VectorStart : Identity;
2743 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2744 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2745 NewPhi->addIncoming(RdxExitVal[part],
2746 LoopVectorBody.back());
2747 RdxParts.push_back(NewPhi);
2750 // Reduce all of the unrolled parts into a single vector.
2751 Value *ReducedPartRdx = RdxParts[0];
2752 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2753 setDebugLocFromInst(Builder, ReducedPartRdx);
2754 for (unsigned part = 1; part < UF; ++part) {
2755 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2756 // Floating point operations had to be 'fast' to enable the reduction.
2757 ReducedPartRdx = addFastMathFlag(
2758 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2759 ReducedPartRdx, "bin.rdx"));
2761 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2762 ReducedPartRdx, RdxParts[part]);
2766 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2767 // and vector ops, reducing the set of values being computed by half each
2769 assert(isPowerOf2_32(VF) &&
2770 "Reduction emission only supported for pow2 vectors!");
2771 Value *TmpVec = ReducedPartRdx;
2772 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2773 for (unsigned i = VF; i != 1; i >>= 1) {
2774 // Move the upper half of the vector to the lower half.
2775 for (unsigned j = 0; j != i/2; ++j)
2776 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2778 // Fill the rest of the mask with undef.
2779 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2780 UndefValue::get(Builder.getInt32Ty()));
2783 Builder.CreateShuffleVector(TmpVec,
2784 UndefValue::get(TmpVec->getType()),
2785 ConstantVector::get(ShuffleMask),
2788 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2789 // Floating point operations had to be 'fast' to enable the reduction.
2790 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2791 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2793 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2796 // The result is in the first element of the vector.
2797 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2798 Builder.getInt32(0));
2801 // Create a phi node that merges control-flow from the backedge-taken check
2802 // block and the middle block.
2803 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2804 LoopScalarPreHeader->getTerminator());
2805 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2806 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2808 // Now, we need to fix the users of the reduction variable
2809 // inside and outside of the scalar remainder loop.
2810 // We know that the loop is in LCSSA form. We need to update the
2811 // PHI nodes in the exit blocks.
2812 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2813 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2814 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2815 if (!LCSSAPhi) break;
2817 // All PHINodes need to have a single entry edge, or two if
2818 // we already fixed them.
2819 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2821 // We found our reduction value exit-PHI. Update it with the
2822 // incoming bypass edge.
2823 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2824 // Add an edge coming from the bypass.
2825 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2828 }// end of the LCSSA phi scan.
2830 // Fix the scalar loop reduction variable with the incoming reduction sum
2831 // from the vector body and from the backedge value.
2832 int IncomingEdgeBlockIdx =
2833 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2834 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2835 // Pick the other block.
2836 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2837 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2838 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2839 }// end of for each redux variable.
2843 // Remove redundant induction instructions.
2844 cse(LoopVectorBody);
2847 void InnerLoopVectorizer::fixLCSSAPHIs() {
2848 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2849 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2850 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2851 if (!LCSSAPhi) break;
2852 if (LCSSAPhi->getNumIncomingValues() == 1)
2853 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2858 InnerLoopVectorizer::VectorParts
2859 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2860 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2863 // Look for cached value.
2864 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2865 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2866 if (ECEntryIt != MaskCache.end())
2867 return ECEntryIt->second;
2869 VectorParts SrcMask = createBlockInMask(Src);
2871 // The terminator has to be a branch inst!
2872 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2873 assert(BI && "Unexpected terminator found");
2875 if (BI->isConditional()) {
2876 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2878 if (BI->getSuccessor(0) != Dst)
2879 for (unsigned part = 0; part < UF; ++part)
2880 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2882 for (unsigned part = 0; part < UF; ++part)
2883 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2885 MaskCache[Edge] = EdgeMask;
2889 MaskCache[Edge] = SrcMask;
2893 InnerLoopVectorizer::VectorParts
2894 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2895 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2897 // Loop incoming mask is all-one.
2898 if (OrigLoop->getHeader() == BB) {
2899 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2900 return getVectorValue(C);
2903 // This is the block mask. We OR all incoming edges, and with zero.
2904 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2905 VectorParts BlockMask = getVectorValue(Zero);
2908 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2909 VectorParts EM = createEdgeMask(*it, BB);
2910 for (unsigned part = 0; part < UF; ++part)
2911 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2917 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2918 InnerLoopVectorizer::VectorParts &Entry,
2919 unsigned UF, unsigned VF, PhiVector *PV) {
2920 PHINode* P = cast<PHINode>(PN);
2921 // Handle reduction variables:
2922 if (Legal->getReductionVars()->count(P)) {
2923 for (unsigned part = 0; part < UF; ++part) {
2924 // This is phase one of vectorizing PHIs.
2925 Type *VecTy = (VF == 1) ? PN->getType() :
2926 VectorType::get(PN->getType(), VF);
2927 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2928 LoopVectorBody.back()-> getFirstInsertionPt());
2934 setDebugLocFromInst(Builder, P);
2935 // Check for PHI nodes that are lowered to vector selects.
2936 if (P->getParent() != OrigLoop->getHeader()) {
2937 // We know that all PHIs in non-header blocks are converted into
2938 // selects, so we don't have to worry about the insertion order and we
2939 // can just use the builder.
2940 // At this point we generate the predication tree. There may be
2941 // duplications since this is a simple recursive scan, but future
2942 // optimizations will clean it up.
2944 unsigned NumIncoming = P->getNumIncomingValues();
2946 // Generate a sequence of selects of the form:
2947 // SELECT(Mask3, In3,
2948 // SELECT(Mask2, In2,
2950 for (unsigned In = 0; In < NumIncoming; In++) {
2951 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2953 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2955 for (unsigned part = 0; part < UF; ++part) {
2956 // We might have single edge PHIs (blocks) - use an identity
2957 // 'select' for the first PHI operand.
2959 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2962 // Select between the current value and the previous incoming edge
2963 // based on the incoming mask.
2964 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2965 Entry[part], "predphi");
2971 // This PHINode must be an induction variable.
2972 // Make sure that we know about it.
2973 assert(Legal->getInductionVars()->count(P) &&
2974 "Not an induction variable");
2976 LoopVectorizationLegality::InductionInfo II =
2977 Legal->getInductionVars()->lookup(P);
2980 case LoopVectorizationLegality::IK_NoInduction:
2981 llvm_unreachable("Unknown induction");
2982 case LoopVectorizationLegality::IK_IntInduction: {
2983 assert(P->getType() == II.StartValue->getType() && "Types must match");
2984 Type *PhiTy = P->getType();
2986 if (P == OldInduction) {
2987 // Handle the canonical induction variable. We might have had to
2989 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2991 // Handle other induction variables that are now based on the
2993 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2995 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2996 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
2999 Broadcasted = getBroadcastInstrs(Broadcasted);
3000 // After broadcasting the induction variable we need to make the vector
3001 // consecutive by adding 0, 1, 2, etc.
3002 for (unsigned part = 0; part < UF; ++part)
3003 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3006 case LoopVectorizationLegality::IK_ReverseIntInduction:
3007 case LoopVectorizationLegality::IK_PtrInduction:
3008 case LoopVectorizationLegality::IK_ReversePtrInduction:
3009 // Handle reverse integer and pointer inductions.
3010 Value *StartIdx = ExtendedIdx;
3011 // This is the normalized GEP that starts counting at zero.
3012 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3015 // Handle the reverse integer induction variable case.
3016 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3017 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3018 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3020 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3023 // This is a new value so do not hoist it out.
3024 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3025 // After broadcasting the induction variable we need to make the
3026 // vector consecutive by adding ... -3, -2, -1, 0.
3027 for (unsigned part = 0; part < UF; ++part)
3028 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3033 // Handle the pointer induction variable case.
3034 assert(P->getType()->isPointerTy() && "Unexpected type.");
3036 // Is this a reverse induction ptr or a consecutive induction ptr.
3037 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3040 // This is the vector of results. Notice that we don't generate
3041 // vector geps because scalar geps result in better code.
3042 for (unsigned part = 0; part < UF; ++part) {
3044 int EltIndex = (part) * (Reverse ? -1 : 1);
3045 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3048 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3050 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3052 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3054 Entry[part] = SclrGep;
3058 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3059 for (unsigned int i = 0; i < VF; ++i) {
3060 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3061 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3064 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3066 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3068 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3070 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3071 Builder.getInt32(i),
3074 Entry[part] = VecVal;
3080 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3081 // For each instruction in the old loop.
3082 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3083 VectorParts &Entry = WidenMap.get(it);
3084 switch (it->getOpcode()) {
3085 case Instruction::Br:
3086 // Nothing to do for PHIs and BR, since we already took care of the
3087 // loop control flow instructions.
3089 case Instruction::PHI:{
3090 // Vectorize PHINodes.
3091 widenPHIInstruction(it, Entry, UF, VF, PV);
3095 case Instruction::Add:
3096 case Instruction::FAdd:
3097 case Instruction::Sub:
3098 case Instruction::FSub:
3099 case Instruction::Mul:
3100 case Instruction::FMul:
3101 case Instruction::UDiv:
3102 case Instruction::SDiv:
3103 case Instruction::FDiv:
3104 case Instruction::URem:
3105 case Instruction::SRem:
3106 case Instruction::FRem:
3107 case Instruction::Shl:
3108 case Instruction::LShr:
3109 case Instruction::AShr:
3110 case Instruction::And:
3111 case Instruction::Or:
3112 case Instruction::Xor: {
3113 // Just widen binops.
3114 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3115 setDebugLocFromInst(Builder, BinOp);
3116 VectorParts &A = getVectorValue(it->getOperand(0));
3117 VectorParts &B = getVectorValue(it->getOperand(1));
3119 // Use this vector value for all users of the original instruction.
3120 for (unsigned Part = 0; Part < UF; ++Part) {
3121 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3123 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3124 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3125 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3126 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3127 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3129 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3130 VecOp->setIsExact(BinOp->isExact());
3132 // Copy the fast-math flags.
3133 if (VecOp && isa<FPMathOperator>(V))
3134 VecOp->setFastMathFlags(it->getFastMathFlags());
3140 case Instruction::Select: {
3142 // If the selector is loop invariant we can create a select
3143 // instruction with a scalar condition. Otherwise, use vector-select.
3144 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3146 setDebugLocFromInst(Builder, it);
3148 // The condition can be loop invariant but still defined inside the
3149 // loop. This means that we can't just use the original 'cond' value.
3150 // We have to take the 'vectorized' value and pick the first lane.
3151 // Instcombine will make this a no-op.
3152 VectorParts &Cond = getVectorValue(it->getOperand(0));
3153 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3154 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3156 Value *ScalarCond = (VF == 1) ? Cond[0] :
3157 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3159 for (unsigned Part = 0; Part < UF; ++Part) {
3160 Entry[Part] = Builder.CreateSelect(
3161 InvariantCond ? ScalarCond : Cond[Part],
3168 case Instruction::ICmp:
3169 case Instruction::FCmp: {
3170 // Widen compares. Generate vector compares.
3171 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3172 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3173 setDebugLocFromInst(Builder, it);
3174 VectorParts &A = getVectorValue(it->getOperand(0));
3175 VectorParts &B = getVectorValue(it->getOperand(1));
3176 for (unsigned Part = 0; Part < UF; ++Part) {
3179 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3181 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3187 case Instruction::Store:
3188 case Instruction::Load:
3189 vectorizeMemoryInstruction(it);
3191 case Instruction::ZExt:
3192 case Instruction::SExt:
3193 case Instruction::FPToUI:
3194 case Instruction::FPToSI:
3195 case Instruction::FPExt:
3196 case Instruction::PtrToInt:
3197 case Instruction::IntToPtr:
3198 case Instruction::SIToFP:
3199 case Instruction::UIToFP:
3200 case Instruction::Trunc:
3201 case Instruction::FPTrunc:
3202 case Instruction::BitCast: {
3203 CastInst *CI = dyn_cast<CastInst>(it);
3204 setDebugLocFromInst(Builder, it);
3205 /// Optimize the special case where the source is the induction
3206 /// variable. Notice that we can only optimize the 'trunc' case
3207 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3208 /// c. other casts depend on pointer size.
3209 if (CI->getOperand(0) == OldInduction &&
3210 it->getOpcode() == Instruction::Trunc) {
3211 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3213 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3214 for (unsigned Part = 0; Part < UF; ++Part)
3215 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3218 /// Vectorize casts.
3219 Type *DestTy = (VF == 1) ? CI->getType() :
3220 VectorType::get(CI->getType(), VF);
3222 VectorParts &A = getVectorValue(it->getOperand(0));
3223 for (unsigned Part = 0; Part < UF; ++Part)
3224 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3228 case Instruction::Call: {
3229 // Ignore dbg intrinsics.
3230 if (isa<DbgInfoIntrinsic>(it))
3232 setDebugLocFromInst(Builder, it);
3234 Module *M = BB->getParent()->getParent();
3235 CallInst *CI = cast<CallInst>(it);
3236 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3237 assert(ID && "Not an intrinsic call!");
3239 case Intrinsic::lifetime_end:
3240 case Intrinsic::lifetime_start:
3241 scalarizeInstruction(it);
3244 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3245 for (unsigned Part = 0; Part < UF; ++Part) {
3246 SmallVector<Value *, 4> Args;
3247 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3248 if (HasScalarOpd && i == 1) {
3249 Args.push_back(CI->getArgOperand(i));
3252 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3253 Args.push_back(Arg[Part]);
3255 Type *Tys[] = {CI->getType()};
3257 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3259 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3260 Entry[Part] = Builder.CreateCall(F, Args);
3268 // All other instructions are unsupported. Scalarize them.
3269 scalarizeInstruction(it);
3272 }// end of for_each instr.
3275 void InnerLoopVectorizer::updateAnalysis() {
3276 // Forget the original basic block.
3277 SE->forgetLoop(OrigLoop);
3279 // Update the dominator tree information.
3280 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3281 "Entry does not dominate exit.");
3283 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3284 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3285 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3287 // Due to if predication of stores we might create a sequence of "if(pred)
3288 // a[i] = ...; " blocks.
3289 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3291 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3292 else if (isPredicatedBlock(i)) {
3293 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3295 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3299 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3300 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3301 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3302 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3304 DEBUG(DT->verifyDomTree());
3307 /// \brief Check whether it is safe to if-convert this phi node.
3309 /// Phi nodes with constant expressions that can trap are not safe to if
3311 static bool canIfConvertPHINodes(BasicBlock *BB) {
3312 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3313 PHINode *Phi = dyn_cast<PHINode>(I);
3316 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3317 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3324 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3325 if (!EnableIfConversion) {
3326 emitAnalysis(Report() << "if-conversion is disabled");
3330 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3332 // A list of pointers that we can safely read and write to.
3333 SmallPtrSet<Value *, 8> SafePointes;
3335 // Collect safe addresses.
3336 for (Loop::block_iterator BI = TheLoop->block_begin(),
3337 BE = TheLoop->block_end(); BI != BE; ++BI) {
3338 BasicBlock *BB = *BI;
3340 if (blockNeedsPredication(BB))
3343 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3344 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3345 SafePointes.insert(LI->getPointerOperand());
3346 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3347 SafePointes.insert(SI->getPointerOperand());
3351 // Collect the blocks that need predication.
3352 BasicBlock *Header = TheLoop->getHeader();
3353 for (Loop::block_iterator BI = TheLoop->block_begin(),
3354 BE = TheLoop->block_end(); BI != BE; ++BI) {
3355 BasicBlock *BB = *BI;
3357 // We don't support switch statements inside loops.
3358 if (!isa<BranchInst>(BB->getTerminator())) {
3359 emitAnalysis(Report(BB->getTerminator())
3360 << "loop contains a switch statement");
3364 // We must be able to predicate all blocks that need to be predicated.
3365 if (blockNeedsPredication(BB)) {
3366 if (!blockCanBePredicated(BB, SafePointes)) {
3367 emitAnalysis(Report(BB->getTerminator())
3368 << "control flow cannot be substituted for a select");
3371 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3372 emitAnalysis(Report(BB->getTerminator())
3373 << "control flow cannot be substituted for a select");
3378 // We can if-convert this loop.
3382 bool LoopVectorizationLegality::canVectorize() {
3383 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3384 // be canonicalized.
3385 if (!TheLoop->getLoopPreheader()) {
3387 Report() << "loop control flow is not understood by vectorizer");
3391 // We can only vectorize innermost loops.
3392 if (TheLoop->getSubLoopsVector().size()) {
3393 emitAnalysis(Report() << "loop is not the innermost loop");
3397 // We must have a single backedge.
3398 if (TheLoop->getNumBackEdges() != 1) {
3400 Report() << "loop control flow is not understood by vectorizer");
3404 // We must have a single exiting block.
3405 if (!TheLoop->getExitingBlock()) {
3407 Report() << "loop control flow is not understood by vectorizer");
3411 // We need to have a loop header.
3412 DEBUG(dbgs() << "LV: Found a loop: " <<
3413 TheLoop->getHeader()->getName() << '\n');
3415 // Check if we can if-convert non-single-bb loops.
3416 unsigned NumBlocks = TheLoop->getNumBlocks();
3417 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3418 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3422 // ScalarEvolution needs to be able to find the exit count.
3423 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3424 if (ExitCount == SE->getCouldNotCompute()) {
3425 emitAnalysis(Report() << "could not determine number of loop iterations");
3426 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3430 // Check if we can vectorize the instructions and CFG in this loop.
3431 if (!canVectorizeInstrs()) {
3432 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3436 // Go over each instruction and look at memory deps.
3437 if (!canVectorizeMemory()) {
3438 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3442 // Collect all of the variables that remain uniform after vectorization.
3443 collectLoopUniforms();
3445 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3446 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3449 // Okay! We can vectorize. At this point we don't have any other mem analysis
3450 // which may limit our maximum vectorization factor, so just return true with
3455 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3456 if (Ty->isPointerTy())
3457 return DL.getIntPtrType(Ty);
3459 // It is possible that char's or short's overflow when we ask for the loop's
3460 // trip count, work around this by changing the type size.
3461 if (Ty->getScalarSizeInBits() < 32)
3462 return Type::getInt32Ty(Ty->getContext());
3467 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3468 Ty0 = convertPointerToIntegerType(DL, Ty0);
3469 Ty1 = convertPointerToIntegerType(DL, Ty1);
3470 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3475 /// \brief Check that the instruction has outside loop users and is not an
3476 /// identified reduction variable.
3477 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3478 SmallPtrSet<Value *, 4> &Reductions) {
3479 // Reduction instructions are allowed to have exit users. All other
3480 // instructions must not have external users.
3481 if (!Reductions.count(Inst))
3482 //Check that all of the users of the loop are inside the BB.
3483 for (User *U : Inst->users()) {
3484 Instruction *UI = cast<Instruction>(U);
3485 // This user may be a reduction exit value.
3486 if (!TheLoop->contains(UI)) {
3487 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3494 bool LoopVectorizationLegality::canVectorizeInstrs() {
3495 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3496 BasicBlock *Header = TheLoop->getHeader();
3498 // Look for the attribute signaling the absence of NaNs.
3499 Function &F = *Header->getParent();
3500 if (F.hasFnAttribute("no-nans-fp-math"))
3501 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3502 AttributeSet::FunctionIndex,
3503 "no-nans-fp-math").getValueAsString() == "true";
3505 // For each block in the loop.
3506 for (Loop::block_iterator bb = TheLoop->block_begin(),
3507 be = TheLoop->block_end(); bb != be; ++bb) {
3509 // Scan the instructions in the block and look for hazards.
3510 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3513 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3514 Type *PhiTy = Phi->getType();
3515 // Check that this PHI type is allowed.
3516 if (!PhiTy->isIntegerTy() &&
3517 !PhiTy->isFloatingPointTy() &&
3518 !PhiTy->isPointerTy()) {
3519 emitAnalysis(Report(it)
3520 << "loop control flow is not understood by vectorizer");
3521 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3525 // If this PHINode is not in the header block, then we know that we
3526 // can convert it to select during if-conversion. No need to check if
3527 // the PHIs in this block are induction or reduction variables.
3528 if (*bb != Header) {
3529 // Check that this instruction has no outside users or is an
3530 // identified reduction value with an outside user.
3531 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3533 emitAnalysis(Report(it) << "value that could not be identified as "
3534 "reduction is used outside the loop");
3538 // We only allow if-converted PHIs with more than two incoming values.
3539 if (Phi->getNumIncomingValues() != 2) {
3540 emitAnalysis(Report(it)
3541 << "control flow not understood by vectorizer");
3542 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3546 // This is the value coming from the preheader.
3547 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3548 // Check if this is an induction variable.
3549 InductionKind IK = isInductionVariable(Phi);
3551 if (IK_NoInduction != IK) {
3552 // Get the widest type.
3554 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3556 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3558 // Int inductions are special because we only allow one IV.
3559 if (IK == IK_IntInduction) {
3560 // Use the phi node with the widest type as induction. Use the last
3561 // one if there are multiple (no good reason for doing this other
3562 // than it is expedient).
3563 if (!Induction || PhiTy == WidestIndTy)
3567 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3568 Inductions[Phi] = InductionInfo(StartValue, IK);
3570 // Until we explicitly handle the case of an induction variable with
3571 // an outside loop user we have to give up vectorizing this loop.
3572 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3573 emitAnalysis(Report(it) << "use of induction value outside of the "
3574 "loop is not handled by vectorizer");
3581 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3582 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3585 if (AddReductionVar(Phi, RK_IntegerMult)) {
3586 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3589 if (AddReductionVar(Phi, RK_IntegerOr)) {
3590 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3593 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3594 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3597 if (AddReductionVar(Phi, RK_IntegerXor)) {
3598 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3601 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3602 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3605 if (AddReductionVar(Phi, RK_FloatMult)) {
3606 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3609 if (AddReductionVar(Phi, RK_FloatAdd)) {
3610 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3613 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3614 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3619 emitAnalysis(Report(it) << "unvectorizable operation");
3620 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3622 }// end of PHI handling
3624 // We still don't handle functions. However, we can ignore dbg intrinsic
3625 // calls and we do handle certain intrinsic and libm functions.
3626 CallInst *CI = dyn_cast<CallInst>(it);
3627 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3628 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3629 DEBUG(dbgs() << "LV: Found a call site.\n");
3633 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3634 // second argument is the same (i.e. loop invariant)
3636 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3637 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3638 emitAnalysis(Report(it)
3639 << "intrinsic instruction cannot be vectorized");
3640 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3645 // Check that the instruction return type is vectorizable.
3646 // Also, we can't vectorize extractelement instructions.
3647 if ((!VectorType::isValidElementType(it->getType()) &&
3648 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3649 emitAnalysis(Report(it)
3650 << "instruction return type cannot be vectorized");
3651 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3655 // Check that the stored type is vectorizable.
3656 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3657 Type *T = ST->getValueOperand()->getType();
3658 if (!VectorType::isValidElementType(T)) {
3659 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3662 if (EnableMemAccessVersioning)
3663 collectStridedAcccess(ST);
3666 if (EnableMemAccessVersioning)
3667 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3668 collectStridedAcccess(LI);
3670 // Reduction instructions are allowed to have exit users.
3671 // All other instructions must not have external users.
3672 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3673 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3682 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3683 if (Inductions.empty()) {
3684 emitAnalysis(Report()
3685 << "loop induction variable could not be identified");
3693 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3694 /// return the induction operand of the gep pointer.
3695 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3696 const DataLayout *DL, Loop *Lp) {
3697 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3701 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3703 // Check that all of the gep indices are uniform except for our induction
3705 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3706 if (i != InductionOperand &&
3707 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3709 return GEP->getOperand(InductionOperand);
3712 ///\brief Look for a cast use of the passed value.
3713 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3714 Value *UniqueCast = nullptr;
3715 for (User *U : Ptr->users()) {
3716 CastInst *CI = dyn_cast<CastInst>(U);
3717 if (CI && CI->getType() == Ty) {
3727 ///\brief Get the stride of a pointer access in a loop.
3728 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3729 /// pointer to the Value, or null otherwise.
3730 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3731 const DataLayout *DL, Loop *Lp) {
3732 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3733 if (!PtrTy || PtrTy->isAggregateType())
3736 // Try to remove a gep instruction to make the pointer (actually index at this
3737 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3738 // pointer, otherwise, we are analyzing the index.
3739 Value *OrigPtr = Ptr;
3741 // The size of the pointer access.
3742 int64_t PtrAccessSize = 1;
3744 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3745 const SCEV *V = SE->getSCEV(Ptr);
3749 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3750 V = C->getOperand();
3752 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3756 V = S->getStepRecurrence(*SE);
3760 // Strip off the size of access multiplication if we are still analyzing the
3762 if (OrigPtr == Ptr) {
3763 DL->getTypeAllocSize(PtrTy->getElementType());
3764 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3765 if (M->getOperand(0)->getSCEVType() != scConstant)
3768 const APInt &APStepVal =
3769 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3771 // Huge step value - give up.
3772 if (APStepVal.getBitWidth() > 64)
3775 int64_t StepVal = APStepVal.getSExtValue();
3776 if (PtrAccessSize != StepVal)
3778 V = M->getOperand(1);
3783 Type *StripedOffRecurrenceCast = nullptr;
3784 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3785 StripedOffRecurrenceCast = C->getType();
3786 V = C->getOperand();
3789 // Look for the loop invariant symbolic value.
3790 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3794 Value *Stride = U->getValue();
3795 if (!Lp->isLoopInvariant(Stride))
3798 // If we have stripped off the recurrence cast we have to make sure that we
3799 // return the value that is used in this loop so that we can replace it later.
3800 if (StripedOffRecurrenceCast)
3801 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3806 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3807 Value *Ptr = nullptr;
3808 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3809 Ptr = LI->getPointerOperand();
3810 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3811 Ptr = SI->getPointerOperand();
3815 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3819 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3820 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3821 Strides[Ptr] = Stride;
3822 StrideSet.insert(Stride);
3825 void LoopVectorizationLegality::collectLoopUniforms() {
3826 // We now know that the loop is vectorizable!
3827 // Collect variables that will remain uniform after vectorization.
3828 std::vector<Value*> Worklist;
3829 BasicBlock *Latch = TheLoop->getLoopLatch();
3831 // Start with the conditional branch and walk up the block.
3832 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3834 // Also add all consecutive pointer values; these values will be uniform
3835 // after vectorization (and subsequent cleanup) and, until revectorization is
3836 // supported, all dependencies must also be uniform.
3837 for (Loop::block_iterator B = TheLoop->block_begin(),
3838 BE = TheLoop->block_end(); B != BE; ++B)
3839 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3841 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3842 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3844 while (Worklist.size()) {
3845 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3846 Worklist.pop_back();
3848 // Look at instructions inside this loop.
3849 // Stop when reaching PHI nodes.
3850 // TODO: we need to follow values all over the loop, not only in this block.
3851 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3854 // This is a known uniform.
3857 // Insert all operands.
3858 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3863 /// \brief Analyses memory accesses in a loop.
3865 /// Checks whether run time pointer checks are needed and builds sets for data
3866 /// dependence checking.
3867 class AccessAnalysis {
3869 /// \brief Read or write access location.
3870 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3871 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3873 /// \brief Set of potential dependent memory accesses.
3874 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3876 AccessAnalysis(const DataLayout *Dl, DepCandidates &DA) :
3877 DL(Dl), DepCands(DA), AreAllWritesIdentified(true),
3878 AreAllReadsIdentified(true), IsRTCheckNeeded(false) {}
3880 /// \brief Register a load and whether it is only read from.
3881 void addLoad(Value *Ptr, bool IsReadOnly) {
3882 Accesses.insert(MemAccessInfo(Ptr, false));
3884 ReadOnlyPtr.insert(Ptr);
3887 /// \brief Register a store.
3888 void addStore(Value *Ptr) {
3889 Accesses.insert(MemAccessInfo(Ptr, true));
3892 /// \brief Check whether we can check the pointers at runtime for
3893 /// non-intersection.
3894 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3895 unsigned &NumComparisons, ScalarEvolution *SE,
3896 Loop *TheLoop, ValueToValueMap &Strides,
3897 bool ShouldCheckStride = false);
3899 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3900 /// and builds sets of dependent accesses.
3901 void buildDependenceSets() {
3902 // Process read-write pointers first.
3903 processMemAccesses(false);
3904 // Next, process read pointers.
3905 processMemAccesses(true);
3908 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3910 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3911 void resetDepChecks() { CheckDeps.clear(); }
3913 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3916 typedef SetVector<MemAccessInfo> PtrAccessSet;
3917 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
3919 /// \brief Go over all memory access or only the deferred ones if
3920 /// \p UseDeferred is true and check whether runtime pointer checks are needed
3921 /// and build sets of dependency check candidates.
3922 void processMemAccesses(bool UseDeferred);
3924 /// Set of all accesses.
3925 PtrAccessSet Accesses;
3927 /// Set of access to check after all writes have been processed.
3928 PtrAccessSet DeferredAccesses;
3930 /// Map of pointers to last access encountered.
3931 UnderlyingObjToAccessMap ObjToLastAccess;
3933 /// Set of accesses that need a further dependence check.
3934 MemAccessInfoSet CheckDeps;
3936 /// Set of pointers that are read only.
3937 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3939 /// Set of underlying objects already written to.
3940 SmallPtrSet<Value*, 16> WriteObjects;
3942 const DataLayout *DL;
3944 /// Sets of potentially dependent accesses - members of one set share an
3945 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3946 /// dependence check.
3947 DepCandidates &DepCands;
3949 bool AreAllWritesIdentified;
3950 bool AreAllReadsIdentified;
3951 bool IsRTCheckNeeded;
3954 } // end anonymous namespace
3956 /// \brief Check whether a pointer can participate in a runtime bounds check.
3957 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
3959 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
3960 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
3964 return AR->isAffine();
3967 /// \brief Check the stride of the pointer and ensure that it does not wrap in
3968 /// the address space.
3969 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
3970 const Loop *Lp, ValueToValueMap &StridesMap);
3972 bool AccessAnalysis::canCheckPtrAtRT(
3973 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3974 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
3975 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
3976 // Find pointers with computable bounds. We are going to use this information
3977 // to place a runtime bound check.
3978 unsigned NumReadPtrChecks = 0;
3979 unsigned NumWritePtrChecks = 0;
3980 bool CanDoRT = true;
3982 bool IsDepCheckNeeded = isDependencyCheckNeeded();
3983 // We assign consecutive id to access from different dependence sets.
3984 // Accesses within the same set don't need a runtime check.
3985 unsigned RunningDepId = 1;
3986 DenseMap<Value *, unsigned> DepSetId;
3988 for (PtrAccessSet::iterator AI = Accesses.begin(), AE = Accesses.end();
3990 const MemAccessInfo &Access = *AI;
3991 Value *Ptr = Access.getPointer();
3992 bool IsWrite = Access.getInt();
3994 // Just add write checks if we have both.
3995 if (!IsWrite && Accesses.count(MemAccessInfo(Ptr, true)))
3999 ++NumWritePtrChecks;
4003 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4004 // When we run after a failing dependency check we have to make sure we
4005 // don't have wrapping pointers.
4006 (!ShouldCheckStride ||
4007 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4008 // The id of the dependence set.
4011 if (IsDepCheckNeeded) {
4012 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4013 unsigned &LeaderId = DepSetId[Leader];
4015 LeaderId = RunningDepId++;
4018 // Each access has its own dependence set.
4019 DepId = RunningDepId++;
4021 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, StridesMap);
4023 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4029 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4030 NumComparisons = 0; // Only one dependence set.
4032 NumComparisons = (NumWritePtrChecks * (NumReadPtrChecks +
4033 NumWritePtrChecks - 1));
4036 // If the pointers that we would use for the bounds comparison have different
4037 // address spaces, assume the values aren't directly comparable, so we can't
4038 // use them for the runtime check. We also have to assume they could
4039 // overlap. In the future there should be metadata for whether address spaces
4041 unsigned NumPointers = RtCheck.Pointers.size();
4042 for (unsigned i = 0; i < NumPointers; ++i) {
4043 for (unsigned j = i + 1; j < NumPointers; ++j) {
4044 // Only need to check pointers between two different dependency sets.
4045 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4048 Value *PtrI = RtCheck.Pointers[i];
4049 Value *PtrJ = RtCheck.Pointers[j];
4051 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4052 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4054 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4055 " different address spaces\n");
4064 static bool isFunctionScopeIdentifiedObject(Value *Ptr) {
4065 return isNoAliasArgument(Ptr) || isNoAliasCall(Ptr) || isa<AllocaInst>(Ptr);
4068 void AccessAnalysis::processMemAccesses(bool UseDeferred) {
4069 // We process the set twice: first we process read-write pointers, last we
4070 // process read-only pointers. This allows us to skip dependence tests for
4071 // read-only pointers.
4073 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4074 for (PtrAccessSet::iterator AI = S.begin(), AE = S.end(); AI != AE; ++AI) {
4075 const MemAccessInfo &Access = *AI;
4076 Value *Ptr = Access.getPointer();
4077 bool IsWrite = Access.getInt();
4079 DepCands.insert(Access);
4081 // Memorize read-only pointers for later processing and skip them in the
4082 // first round (they need to be checked after we have seen all write
4083 // pointers). Note: we also mark pointer that are not consecutive as
4084 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need the
4085 // second check for "!IsWrite".
4086 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4087 if (!UseDeferred && IsReadOnlyPtr) {
4088 DeferredAccesses.insert(Access);
4092 bool NeedDepCheck = false;
4093 // Check whether there is the possibility of dependency because of
4094 // underlying objects being the same.
4095 typedef SmallVector<Value*, 16> ValueVector;
4096 ValueVector TempObjects;
4097 GetUnderlyingObjects(Ptr, TempObjects, DL);
4098 for (ValueVector::iterator UI = TempObjects.begin(), UE = TempObjects.end();
4100 Value *UnderlyingObj = *UI;
4102 // If this is a write then it needs to be an identified object. If this a
4103 // read and all writes (so far) are identified function scope objects we
4104 // don't need an identified underlying object but only an Argument (the
4105 // next write is going to invalidate this assumption if it is
4107 // This is a micro-optimization for the case where all writes are
4108 // identified and we have one argument pointer.
4109 // Otherwise, we do need a runtime check.
4110 if ((IsWrite && !isFunctionScopeIdentifiedObject(UnderlyingObj)) ||
4111 (!IsWrite && (!AreAllWritesIdentified ||
4112 !isa<Argument>(UnderlyingObj)) &&
4113 !isIdentifiedObject(UnderlyingObj))) {
4114 DEBUG(dbgs() << "LV: Found an unidentified " <<
4115 (IsWrite ? "write" : "read" ) << " ptr: " << *UnderlyingObj <<
4117 IsRTCheckNeeded = (IsRTCheckNeeded ||
4118 !isIdentifiedObject(UnderlyingObj) ||
4119 !AreAllReadsIdentified);
4122 AreAllWritesIdentified = false;
4124 AreAllReadsIdentified = false;
4127 // If this is a write - check other reads and writes for conflicts. If
4128 // this is a read only check other writes for conflicts (but only if there
4129 // is no other write to the ptr - this is an optimization to catch "a[i] =
4130 // a[i] + " without having to do a dependence check).
4131 if ((IsWrite || IsReadOnlyPtr) && WriteObjects.count(UnderlyingObj))
4132 NeedDepCheck = true;
4135 WriteObjects.insert(UnderlyingObj);
4137 // Create sets of pointers connected by shared underlying objects.
4138 UnderlyingObjToAccessMap::iterator Prev =
4139 ObjToLastAccess.find(UnderlyingObj);
4140 if (Prev != ObjToLastAccess.end())
4141 DepCands.unionSets(Access, Prev->second);
4143 ObjToLastAccess[UnderlyingObj] = Access;
4147 CheckDeps.insert(Access);
4152 /// \brief Checks memory dependences among accesses to the same underlying
4153 /// object to determine whether there vectorization is legal or not (and at
4154 /// which vectorization factor).
4156 /// This class works under the assumption that we already checked that memory
4157 /// locations with different underlying pointers are "must-not alias".
4158 /// We use the ScalarEvolution framework to symbolically evalutate access
4159 /// functions pairs. Since we currently don't restructure the loop we can rely
4160 /// on the program order of memory accesses to determine their safety.
4161 /// At the moment we will only deem accesses as safe for:
4162 /// * A negative constant distance assuming program order.
4164 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4165 /// a[i] = tmp; y = a[i];
4167 /// The latter case is safe because later checks guarantuee that there can't
4168 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4169 /// the same variable: a header phi can only be an induction or a reduction, a
4170 /// reduction can't have a memory sink, an induction can't have a memory
4171 /// source). This is important and must not be violated (or we have to
4172 /// resort to checking for cycles through memory).
4174 /// * A positive constant distance assuming program order that is bigger
4175 /// than the biggest memory access.
4177 /// tmp = a[i] OR b[i] = x
4178 /// a[i+2] = tmp y = b[i+2];
4180 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4182 /// * Zero distances and all accesses have the same size.
4184 class MemoryDepChecker {
4186 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4187 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4189 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4190 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4191 ShouldRetryWithRuntimeCheck(false) {}
4193 /// \brief Register the location (instructions are given increasing numbers)
4194 /// of a write access.
4195 void addAccess(StoreInst *SI) {
4196 Value *Ptr = SI->getPointerOperand();
4197 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4198 InstMap.push_back(SI);
4202 /// \brief Register the location (instructions are given increasing numbers)
4203 /// of a write access.
4204 void addAccess(LoadInst *LI) {
4205 Value *Ptr = LI->getPointerOperand();
4206 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4207 InstMap.push_back(LI);
4211 /// \brief Check whether the dependencies between the accesses are safe.
4213 /// Only checks sets with elements in \p CheckDeps.
4214 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4215 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4217 /// \brief The maximum number of bytes of a vector register we can vectorize
4218 /// the accesses safely with.
4219 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4221 /// \brief In same cases when the dependency check fails we can still
4222 /// vectorize the loop with a dynamic array access check.
4223 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4226 ScalarEvolution *SE;
4227 const DataLayout *DL;
4228 const Loop *InnermostLoop;
4230 /// \brief Maps access locations (ptr, read/write) to program order.
4231 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4233 /// \brief Memory access instructions in program order.
4234 SmallVector<Instruction *, 16> InstMap;
4236 /// \brief The program order index to be used for the next instruction.
4239 // We can access this many bytes in parallel safely.
4240 unsigned MaxSafeDepDistBytes;
4242 /// \brief If we see a non-constant dependence distance we can still try to
4243 /// vectorize this loop with runtime checks.
4244 bool ShouldRetryWithRuntimeCheck;
4246 /// \brief Check whether there is a plausible dependence between the two
4249 /// Access \p A must happen before \p B in program order. The two indices
4250 /// identify the index into the program order map.
4252 /// This function checks whether there is a plausible dependence (or the
4253 /// absence of such can't be proved) between the two accesses. If there is a
4254 /// plausible dependence but the dependence distance is bigger than one
4255 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4256 /// distance is smaller than any other distance encountered so far).
4257 /// Otherwise, this function returns true signaling a possible dependence.
4258 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4259 const MemAccessInfo &B, unsigned BIdx,
4260 ValueToValueMap &Strides);
4262 /// \brief Check whether the data dependence could prevent store-load
4264 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4267 } // end anonymous namespace
4269 static bool isInBoundsGep(Value *Ptr) {
4270 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4271 return GEP->isInBounds();
4275 /// \brief Check whether the access through \p Ptr has a constant stride.
4276 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4277 const Loop *Lp, ValueToValueMap &StridesMap) {
4278 const Type *Ty = Ptr->getType();
4279 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4281 // Make sure that the pointer does not point to aggregate types.
4282 const PointerType *PtrTy = cast<PointerType>(Ty);
4283 if (PtrTy->getElementType()->isAggregateType()) {
4284 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4289 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4291 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4293 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4294 << *Ptr << " SCEV: " << *PtrScev << "\n");
4298 // The accesss function must stride over the innermost loop.
4299 if (Lp != AR->getLoop()) {
4300 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4301 *Ptr << " SCEV: " << *PtrScev << "\n");
4304 // The address calculation must not wrap. Otherwise, a dependence could be
4306 // An inbounds getelementptr that is a AddRec with a unit stride
4307 // cannot wrap per definition. The unit stride requirement is checked later.
4308 // An getelementptr without an inbounds attribute and unit stride would have
4309 // to access the pointer value "0" which is undefined behavior in address
4310 // space 0, therefore we can also vectorize this case.
4311 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4312 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4313 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4314 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4315 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4316 << *Ptr << " SCEV: " << *PtrScev << "\n");
4320 // Check the step is constant.
4321 const SCEV *Step = AR->getStepRecurrence(*SE);
4323 // Calculate the pointer stride and check if it is consecutive.
4324 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4326 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4327 " SCEV: " << *PtrScev << "\n");
4331 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4332 const APInt &APStepVal = C->getValue()->getValue();
4334 // Huge step value - give up.
4335 if (APStepVal.getBitWidth() > 64)
4338 int64_t StepVal = APStepVal.getSExtValue();
4341 int64_t Stride = StepVal / Size;
4342 int64_t Rem = StepVal % Size;
4346 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4347 // know we can't "wrap around the address space". In case of address space
4348 // zero we know that this won't happen without triggering undefined behavior.
4349 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4350 Stride != 1 && Stride != -1)
4356 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4357 unsigned TypeByteSize) {
4358 // If loads occur at a distance that is not a multiple of a feasible vector
4359 // factor store-load forwarding does not take place.
4360 // Positive dependences might cause troubles because vectorizing them might
4361 // prevent store-load forwarding making vectorized code run a lot slower.
4362 // a[i] = a[i-3] ^ a[i-8];
4363 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4364 // hence on your typical architecture store-load forwarding does not take
4365 // place. Vectorizing in such cases does not make sense.
4366 // Store-load forwarding distance.
4367 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4368 // Maximum vector factor.
4369 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4370 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4371 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4373 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4375 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4376 MaxVFWithoutSLForwardIssues = (vf >>=1);
4381 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4382 DEBUG(dbgs() << "LV: Distance " << Distance <<
4383 " that could cause a store-load forwarding conflict\n");
4387 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4388 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4389 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4393 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4394 const MemAccessInfo &B, unsigned BIdx,
4395 ValueToValueMap &Strides) {
4396 assert (AIdx < BIdx && "Must pass arguments in program order");
4398 Value *APtr = A.getPointer();
4399 Value *BPtr = B.getPointer();
4400 bool AIsWrite = A.getInt();
4401 bool BIsWrite = B.getInt();
4403 // Two reads are independent.
4404 if (!AIsWrite && !BIsWrite)
4407 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4408 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4410 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4411 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4413 const SCEV *Src = AScev;
4414 const SCEV *Sink = BScev;
4416 // If the induction step is negative we have to invert source and sink of the
4418 if (StrideAPtr < 0) {
4421 std::swap(APtr, BPtr);
4422 std::swap(Src, Sink);
4423 std::swap(AIsWrite, BIsWrite);
4424 std::swap(AIdx, BIdx);
4425 std::swap(StrideAPtr, StrideBPtr);
4428 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4430 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4431 << "(Induction step: " << StrideAPtr << ")\n");
4432 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4433 << *InstMap[BIdx] << ": " << *Dist << "\n");
4435 // Need consecutive accesses. We don't want to vectorize
4436 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4437 // the address space.
4438 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4439 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4443 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4445 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4446 ShouldRetryWithRuntimeCheck = true;
4450 Type *ATy = APtr->getType()->getPointerElementType();
4451 Type *BTy = BPtr->getType()->getPointerElementType();
4452 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4454 // Negative distances are not plausible dependencies.
4455 const APInt &Val = C->getValue()->getValue();
4456 if (Val.isNegative()) {
4457 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4458 if (IsTrueDataDependence &&
4459 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4463 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4467 // Write to the same location with the same size.
4468 // Could be improved to assert type sizes are the same (i32 == float, etc).
4472 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4476 assert(Val.isStrictlyPositive() && "Expect a positive value");
4478 // Positive distance bigger than max vectorization factor.
4481 "LV: ReadWrite-Write positive dependency with different types\n");
4485 unsigned Distance = (unsigned) Val.getZExtValue();
4487 // Bail out early if passed-in parameters make vectorization not feasible.
4488 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4489 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4491 // The distance must be bigger than the size needed for a vectorized version
4492 // of the operation and the size of the vectorized operation must not be
4493 // bigger than the currrent maximum size.
4494 if (Distance < 2*TypeByteSize ||
4495 2*TypeByteSize > MaxSafeDepDistBytes ||
4496 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4497 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4498 << Val.getSExtValue() << '\n');
4502 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4503 Distance : MaxSafeDepDistBytes;
4505 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4506 if (IsTrueDataDependence &&
4507 couldPreventStoreLoadForward(Distance, TypeByteSize))
4510 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4511 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4516 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4517 MemAccessInfoSet &CheckDeps,
4518 ValueToValueMap &Strides) {
4520 MaxSafeDepDistBytes = -1U;
4521 while (!CheckDeps.empty()) {
4522 MemAccessInfo CurAccess = *CheckDeps.begin();
4524 // Get the relevant memory access set.
4525 EquivalenceClasses<MemAccessInfo>::iterator I =
4526 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4528 // Check accesses within this set.
4529 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4530 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4532 // Check every access pair.
4534 CheckDeps.erase(*AI);
4535 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4537 // Check every accessing instruction pair in program order.
4538 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4539 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4540 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4541 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4542 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4544 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4555 bool LoopVectorizationLegality::canVectorizeMemory() {
4557 typedef SmallVector<Value*, 16> ValueVector;
4558 typedef SmallPtrSet<Value*, 16> ValueSet;
4560 // Holds the Load and Store *instructions*.
4564 // Holds all the different accesses in the loop.
4565 unsigned NumReads = 0;
4566 unsigned NumReadWrites = 0;
4568 PtrRtCheck.Pointers.clear();
4569 PtrRtCheck.Need = false;
4571 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4572 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4575 for (Loop::block_iterator bb = TheLoop->block_begin(),
4576 be = TheLoop->block_end(); bb != be; ++bb) {
4578 // Scan the BB and collect legal loads and stores.
4579 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4582 // If this is a load, save it. If this instruction can read from memory
4583 // but is not a load, then we quit. Notice that we don't handle function
4584 // calls that read or write.
4585 if (it->mayReadFromMemory()) {
4586 // Many math library functions read the rounding mode. We will only
4587 // vectorize a loop if it contains known function calls that don't set
4588 // the flag. Therefore, it is safe to ignore this read from memory.
4589 CallInst *Call = dyn_cast<CallInst>(it);
4590 if (Call && getIntrinsicIDForCall(Call, TLI))
4593 LoadInst *Ld = dyn_cast<LoadInst>(it);
4594 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4595 emitAnalysis(Report(Ld)
4596 << "read with atomic ordering or volatile read");
4597 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4601 Loads.push_back(Ld);
4602 DepChecker.addAccess(Ld);
4606 // Save 'store' instructions. Abort if other instructions write to memory.
4607 if (it->mayWriteToMemory()) {
4608 StoreInst *St = dyn_cast<StoreInst>(it);
4610 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4613 if (!St->isSimple() && !IsAnnotatedParallel) {
4614 emitAnalysis(Report(St)
4615 << "write with atomic ordering or volatile write");
4616 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4620 Stores.push_back(St);
4621 DepChecker.addAccess(St);
4626 // Now we have two lists that hold the loads and the stores.
4627 // Next, we find the pointers that they use.
4629 // Check if we see any stores. If there are no stores, then we don't
4630 // care if the pointers are *restrict*.
4631 if (!Stores.size()) {
4632 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4636 AccessAnalysis::DepCandidates DependentAccesses;
4637 AccessAnalysis Accesses(DL, DependentAccesses);
4639 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4640 // multiple times on the same object. If the ptr is accessed twice, once
4641 // for read and once for write, it will only appear once (on the write
4642 // list). This is okay, since we are going to check for conflicts between
4643 // writes and between reads and writes, but not between reads and reads.
4646 ValueVector::iterator I, IE;
4647 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4648 StoreInst *ST = cast<StoreInst>(*I);
4649 Value* Ptr = ST->getPointerOperand();
4651 if (isUniform(Ptr)) {
4654 << "write to a loop invariant address could not be vectorized");
4655 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4659 // If we did *not* see this pointer before, insert it to the read-write
4660 // list. At this phase it is only a 'write' list.
4661 if (Seen.insert(Ptr)) {
4663 Accesses.addStore(Ptr);
4667 if (IsAnnotatedParallel) {
4669 << "LV: A loop annotated parallel, ignore memory dependency "
4674 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4675 LoadInst *LD = cast<LoadInst>(*I);
4676 Value* Ptr = LD->getPointerOperand();
4677 // If we did *not* see this pointer before, insert it to the
4678 // read list. If we *did* see it before, then it is already in
4679 // the read-write list. This allows us to vectorize expressions
4680 // such as A[i] += x; Because the address of A[i] is a read-write
4681 // pointer. This only works if the index of A[i] is consecutive.
4682 // If the address of i is unknown (for example A[B[i]]) then we may
4683 // read a few words, modify, and write a few words, and some of the
4684 // words may be written to the same address.
4685 bool IsReadOnlyPtr = false;
4686 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4688 IsReadOnlyPtr = true;
4690 Accesses.addLoad(Ptr, IsReadOnlyPtr);
4693 // If we write (or read-write) to a single destination and there are no
4694 // other reads in this loop then is it safe to vectorize.
4695 if (NumReadWrites == 1 && NumReads == 0) {
4696 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4700 // Build dependence sets and check whether we need a runtime pointer bounds
4702 Accesses.buildDependenceSets();
4703 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4705 // Find pointers with computable bounds. We are going to use this information
4706 // to place a runtime bound check.
4707 unsigned NumComparisons = 0;
4708 bool CanDoRT = false;
4710 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4713 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4714 " pointer comparisons.\n");
4716 // If we only have one set of dependences to check pointers among we don't
4717 // need a runtime check.
4718 if (NumComparisons == 0 && NeedRTCheck)
4719 NeedRTCheck = false;
4721 // Check that we did not collect too many pointers or found an unsizeable
4723 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4729 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4732 if (NeedRTCheck && !CanDoRT) {
4733 emitAnalysis(Report() << "cannot identify array bounds");
4734 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4735 "the array bounds.\n");
4740 PtrRtCheck.Need = NeedRTCheck;
4742 bool CanVecMem = true;
4743 if (Accesses.isDependencyCheckNeeded()) {
4744 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4745 CanVecMem = DepChecker.areDepsSafe(
4746 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4747 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4749 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4750 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4753 // Clear the dependency checks. We assume they are not needed.
4754 Accesses.resetDepChecks();
4757 PtrRtCheck.Need = true;
4759 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4760 TheLoop, Strides, true);
4761 // Check that we did not collect too many pointers or found an unsizeable
4763 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4764 if (!CanDoRT && NumComparisons > 0)
4765 emitAnalysis(Report()
4766 << "cannot check memory dependencies at runtime");
4768 emitAnalysis(Report()
4769 << NumComparisons << " exceeds limit of "
4770 << RuntimeMemoryCheckThreshold
4771 << " dependent memory operations checked at runtime");
4772 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4782 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4784 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4785 " need a runtime memory check.\n");
4790 static bool hasMultipleUsesOf(Instruction *I,
4791 SmallPtrSet<Instruction *, 8> &Insts) {
4792 unsigned NumUses = 0;
4793 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4794 if (Insts.count(dyn_cast<Instruction>(*Use)))
4803 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4804 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4805 if (!Set.count(dyn_cast<Instruction>(*Use)))
4810 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4811 ReductionKind Kind) {
4812 if (Phi->getNumIncomingValues() != 2)
4815 // Reduction variables are only found in the loop header block.
4816 if (Phi->getParent() != TheLoop->getHeader())
4819 // Obtain the reduction start value from the value that comes from the loop
4821 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4823 // ExitInstruction is the single value which is used outside the loop.
4824 // We only allow for a single reduction value to be used outside the loop.
4825 // This includes users of the reduction, variables (which form a cycle
4826 // which ends in the phi node).
4827 Instruction *ExitInstruction = nullptr;
4828 // Indicates that we found a reduction operation in our scan.
4829 bool FoundReduxOp = false;
4831 // We start with the PHI node and scan for all of the users of this
4832 // instruction. All users must be instructions that can be used as reduction
4833 // variables (such as ADD). We must have a single out-of-block user. The cycle
4834 // must include the original PHI.
4835 bool FoundStartPHI = false;
4837 // To recognize min/max patterns formed by a icmp select sequence, we store
4838 // the number of instruction we saw from the recognized min/max pattern,
4839 // to make sure we only see exactly the two instructions.
4840 unsigned NumCmpSelectPatternInst = 0;
4841 ReductionInstDesc ReduxDesc(false, nullptr);
4843 SmallPtrSet<Instruction *, 8> VisitedInsts;
4844 SmallVector<Instruction *, 8> Worklist;
4845 Worklist.push_back(Phi);
4846 VisitedInsts.insert(Phi);
4848 // A value in the reduction can be used:
4849 // - By the reduction:
4850 // - Reduction operation:
4851 // - One use of reduction value (safe).
4852 // - Multiple use of reduction value (not safe).
4854 // - All uses of the PHI must be the reduction (safe).
4855 // - Otherwise, not safe.
4856 // - By one instruction outside of the loop (safe).
4857 // - By further instructions outside of the loop (not safe).
4858 // - By an instruction that is not part of the reduction (not safe).
4860 // * An instruction type other than PHI or the reduction operation.
4861 // * A PHI in the header other than the initial PHI.
4862 while (!Worklist.empty()) {
4863 Instruction *Cur = Worklist.back();
4864 Worklist.pop_back();
4867 // If the instruction has no users then this is a broken chain and can't be
4868 // a reduction variable.
4869 if (Cur->use_empty())
4872 bool IsAPhi = isa<PHINode>(Cur);
4874 // A header PHI use other than the original PHI.
4875 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4878 // Reductions of instructions such as Div, and Sub is only possible if the
4879 // LHS is the reduction variable.
4880 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4881 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4882 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4885 // Any reduction instruction must be of one of the allowed kinds.
4886 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4887 if (!ReduxDesc.IsReduction)
4890 // A reduction operation must only have one use of the reduction value.
4891 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4892 hasMultipleUsesOf(Cur, VisitedInsts))
4895 // All inputs to a PHI node must be a reduction value.
4896 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4899 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4900 isa<SelectInst>(Cur)))
4901 ++NumCmpSelectPatternInst;
4902 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4903 isa<SelectInst>(Cur)))
4904 ++NumCmpSelectPatternInst;
4906 // Check whether we found a reduction operator.
4907 FoundReduxOp |= !IsAPhi;
4909 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4910 // onto the stack. This way we are going to have seen all inputs to PHI
4911 // nodes once we get to them.
4912 SmallVector<Instruction *, 8> NonPHIs;
4913 SmallVector<Instruction *, 8> PHIs;
4914 for (User *U : Cur->users()) {
4915 Instruction *UI = cast<Instruction>(U);
4917 // Check if we found the exit user.
4918 BasicBlock *Parent = UI->getParent();
4919 if (!TheLoop->contains(Parent)) {
4920 // Exit if you find multiple outside users or if the header phi node is
4921 // being used. In this case the user uses the value of the previous
4922 // iteration, in which case we would loose "VF-1" iterations of the
4923 // reduction operation if we vectorize.
4924 if (ExitInstruction != nullptr || Cur == Phi)
4927 // The instruction used by an outside user must be the last instruction
4928 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4929 // operations on the value.
4930 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4933 ExitInstruction = Cur;
4937 // Process instructions only once (termination). Each reduction cycle
4938 // value must only be used once, except by phi nodes and min/max
4939 // reductions which are represented as a cmp followed by a select.
4940 ReductionInstDesc IgnoredVal(false, nullptr);
4941 if (VisitedInsts.insert(UI)) {
4942 if (isa<PHINode>(UI))
4945 NonPHIs.push_back(UI);
4946 } else if (!isa<PHINode>(UI) &&
4947 ((!isa<FCmpInst>(UI) &&
4948 !isa<ICmpInst>(UI) &&
4949 !isa<SelectInst>(UI)) ||
4950 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4953 // Remember that we completed the cycle.
4955 FoundStartPHI = true;
4957 Worklist.append(PHIs.begin(), PHIs.end());
4958 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4961 // This means we have seen one but not the other instruction of the
4962 // pattern or more than just a select and cmp.
4963 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4964 NumCmpSelectPatternInst != 2)
4967 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4970 // We found a reduction var if we have reached the original phi node and we
4971 // only have a single instruction with out-of-loop users.
4973 // This instruction is allowed to have out-of-loop users.
4974 AllowedExit.insert(ExitInstruction);
4976 // Save the description of this reduction variable.
4977 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4978 ReduxDesc.MinMaxKind);
4979 Reductions[Phi] = RD;
4980 // We've ended the cycle. This is a reduction variable if we have an
4981 // outside user and it has a binary op.
4986 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4987 /// pattern corresponding to a min(X, Y) or max(X, Y).
4988 LoopVectorizationLegality::ReductionInstDesc
4989 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4990 ReductionInstDesc &Prev) {
4992 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4993 "Expect a select instruction");
4994 Instruction *Cmp = nullptr;
4995 SelectInst *Select = nullptr;
4997 // We must handle the select(cmp()) as a single instruction. Advance to the
4999 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5000 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5001 return ReductionInstDesc(false, I);
5002 return ReductionInstDesc(Select, Prev.MinMaxKind);
5005 // Only handle single use cases for now.
5006 if (!(Select = dyn_cast<SelectInst>(I)))
5007 return ReductionInstDesc(false, I);
5008 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5009 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5010 return ReductionInstDesc(false, I);
5011 if (!Cmp->hasOneUse())
5012 return ReductionInstDesc(false, I);
5017 // Look for a min/max pattern.
5018 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5019 return ReductionInstDesc(Select, MRK_UIntMin);
5020 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5021 return ReductionInstDesc(Select, MRK_UIntMax);
5022 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5023 return ReductionInstDesc(Select, MRK_SIntMax);
5024 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5025 return ReductionInstDesc(Select, MRK_SIntMin);
5026 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5027 return ReductionInstDesc(Select, MRK_FloatMin);
5028 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5029 return ReductionInstDesc(Select, MRK_FloatMax);
5030 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5031 return ReductionInstDesc(Select, MRK_FloatMin);
5032 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5033 return ReductionInstDesc(Select, MRK_FloatMax);
5035 return ReductionInstDesc(false, I);
5038 LoopVectorizationLegality::ReductionInstDesc
5039 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5041 ReductionInstDesc &Prev) {
5042 bool FP = I->getType()->isFloatingPointTy();
5043 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
5044 switch (I->getOpcode()) {
5046 return ReductionInstDesc(false, I);
5047 case Instruction::PHI:
5048 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5049 Kind != RK_FloatMinMax))
5050 return ReductionInstDesc(false, I);
5051 return ReductionInstDesc(I, Prev.MinMaxKind);
5052 case Instruction::Sub:
5053 case Instruction::Add:
5054 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5055 case Instruction::Mul:
5056 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5057 case Instruction::And:
5058 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5059 case Instruction::Or:
5060 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5061 case Instruction::Xor:
5062 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5063 case Instruction::FMul:
5064 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5065 case Instruction::FAdd:
5066 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5067 case Instruction::FCmp:
5068 case Instruction::ICmp:
5069 case Instruction::Select:
5070 if (Kind != RK_IntegerMinMax &&
5071 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5072 return ReductionInstDesc(false, I);
5073 return isMinMaxSelectCmpPattern(I, Prev);
5077 LoopVectorizationLegality::InductionKind
5078 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5079 Type *PhiTy = Phi->getType();
5080 // We only handle integer and pointer inductions variables.
5081 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5082 return IK_NoInduction;
5084 // Check that the PHI is consecutive.
5085 const SCEV *PhiScev = SE->getSCEV(Phi);
5086 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5088 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5089 return IK_NoInduction;
5091 const SCEV *Step = AR->getStepRecurrence(*SE);
5093 // Integer inductions need to have a stride of one.
5094 if (PhiTy->isIntegerTy()) {
5096 return IK_IntInduction;
5097 if (Step->isAllOnesValue())
5098 return IK_ReverseIntInduction;
5099 return IK_NoInduction;
5102 // Calculate the pointer stride and check if it is consecutive.
5103 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5105 return IK_NoInduction;
5107 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5108 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5109 if (C->getValue()->equalsInt(Size))
5110 return IK_PtrInduction;
5111 else if (C->getValue()->equalsInt(0 - Size))
5112 return IK_ReversePtrInduction;
5114 return IK_NoInduction;
5117 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5118 Value *In0 = const_cast<Value*>(V);
5119 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5123 return Inductions.count(PN);
5126 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5127 assert(TheLoop->contains(BB) && "Unknown block used");
5129 // Blocks that do not dominate the latch need predication.
5130 BasicBlock* Latch = TheLoop->getLoopLatch();
5131 return !DT->dominates(BB, Latch);
5134 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5135 SmallPtrSet<Value *, 8>& SafePtrs) {
5136 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5137 // We might be able to hoist the load.
5138 if (it->mayReadFromMemory()) {
5139 LoadInst *LI = dyn_cast<LoadInst>(it);
5140 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5144 // We don't predicate stores at the moment.
5145 if (it->mayWriteToMemory()) {
5146 StoreInst *SI = dyn_cast<StoreInst>(it);
5147 // We only support predication of stores in basic blocks with one
5149 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5150 !SafePtrs.count(SI->getPointerOperand()) ||
5151 !SI->getParent()->getSinglePredecessor())
5157 // Check that we don't have a constant expression that can trap as operand.
5158 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5160 if (Constant *C = dyn_cast<Constant>(*OI))
5165 // The instructions below can trap.
5166 switch (it->getOpcode()) {
5168 case Instruction::UDiv:
5169 case Instruction::SDiv:
5170 case Instruction::URem:
5171 case Instruction::SRem:
5179 LoopVectorizationCostModel::VectorizationFactor
5180 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5182 bool ForceVectorization) {
5183 // Width 1 means no vectorize
5184 VectorizationFactor Factor = { 1U, 0U };
5185 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5186 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5190 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5191 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5195 // Find the trip count.
5196 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5197 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5199 unsigned WidestType = getWidestType();
5200 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5201 unsigned MaxSafeDepDist = -1U;
5202 if (Legal->getMaxSafeDepDistBytes() != -1U)
5203 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5204 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5205 WidestRegister : MaxSafeDepDist);
5206 unsigned MaxVectorSize = WidestRegister / WidestType;
5207 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5208 DEBUG(dbgs() << "LV: The Widest register is: "
5209 << WidestRegister << " bits.\n");
5211 if (MaxVectorSize == 0) {
5212 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5216 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5217 " into one vector!");
5219 unsigned VF = MaxVectorSize;
5221 // If we optimize the program for size, avoid creating the tail loop.
5223 // If we are unable to calculate the trip count then don't try to vectorize.
5225 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5229 // Find the maximum SIMD width that can fit within the trip count.
5230 VF = TC % MaxVectorSize;
5235 // If the trip count that we found modulo the vectorization factor is not
5236 // zero then we require a tail.
5238 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5244 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5245 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5247 Factor.Width = UserVF;
5251 float Cost = expectedCost(1);
5253 const float ScalarCost = Cost;
5256 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5258 // Ignore scalar width, because the user explicitly wants vectorization.
5259 if (ForceVectorization && VF > 1) {
5261 Cost = expectedCost(Width) / (float)Width;
5264 for (unsigned i=2; i <= VF; i*=2) {
5265 // Notice that the vector loop needs to be executed less times, so
5266 // we need to divide the cost of the vector loops by the width of
5267 // the vector elements.
5268 float VectorCost = expectedCost(i) / (float)i;
5269 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5270 (int)VectorCost << ".\n");
5271 if (VectorCost < Cost) {
5277 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5278 << "LV: Vectorization seems to be not beneficial, "
5279 << "but was forced by a user.\n");
5280 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5281 Factor.Width = Width;
5282 Factor.Cost = Width * Cost;
5286 unsigned LoopVectorizationCostModel::getWidestType() {
5287 unsigned MaxWidth = 8;
5290 for (Loop::block_iterator bb = TheLoop->block_begin(),
5291 be = TheLoop->block_end(); bb != be; ++bb) {
5292 BasicBlock *BB = *bb;
5294 // For each instruction in the loop.
5295 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5296 Type *T = it->getType();
5298 // Only examine Loads, Stores and PHINodes.
5299 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5302 // Examine PHI nodes that are reduction variables.
5303 if (PHINode *PN = dyn_cast<PHINode>(it))
5304 if (!Legal->getReductionVars()->count(PN))
5307 // Examine the stored values.
5308 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5309 T = ST->getValueOperand()->getType();
5311 // Ignore loaded pointer types and stored pointer types that are not
5312 // consecutive. However, we do want to take consecutive stores/loads of
5313 // pointer vectors into account.
5314 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5317 MaxWidth = std::max(MaxWidth,
5318 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5326 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5329 unsigned LoopCost) {
5331 // -- The unroll heuristics --
5332 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5333 // There are many micro-architectural considerations that we can't predict
5334 // at this level. For example frontend pressure (on decode or fetch) due to
5335 // code size, or the number and capabilities of the execution ports.
5337 // We use the following heuristics to select the unroll factor:
5338 // 1. If the code has reductions the we unroll in order to break the cross
5339 // iteration dependency.
5340 // 2. If the loop is really small then we unroll in order to reduce the loop
5342 // 3. We don't unroll if we think that we will spill registers to memory due
5343 // to the increased register pressure.
5345 // Use the user preference, unless 'auto' is selected.
5349 // When we optimize for size we don't unroll.
5353 // We used the distance for the unroll factor.
5354 if (Legal->getMaxSafeDepDistBytes() != -1U)
5357 // Do not unroll loops with a relatively small trip count.
5358 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5359 TheLoop->getLoopLatch());
5360 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5363 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5364 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5368 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5369 TargetNumRegisters = ForceTargetNumScalarRegs;
5371 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5372 TargetNumRegisters = ForceTargetNumVectorRegs;
5375 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5376 // We divide by these constants so assume that we have at least one
5377 // instruction that uses at least one register.
5378 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5379 R.NumInstructions = std::max(R.NumInstructions, 1U);
5381 // We calculate the unroll factor using the following formula.
5382 // Subtract the number of loop invariants from the number of available
5383 // registers. These registers are used by all of the unrolled instances.
5384 // Next, divide the remaining registers by the number of registers that is
5385 // required by the loop, in order to estimate how many parallel instances
5386 // fit without causing spills. All of this is rounded down if necessary to be
5387 // a power of two. We want power of two unroll factors to simplify any
5388 // addressing operations or alignment considerations.
5389 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5392 // Don't count the induction variable as unrolled.
5393 if (EnableIndVarRegisterHeur)
5394 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5395 std::max(1U, (R.MaxLocalUsers - 1)));
5397 // Clamp the unroll factor ranges to reasonable factors.
5398 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5400 // Check if the user has overridden the unroll max.
5402 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5403 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5405 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5406 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5409 // If we did not calculate the cost for VF (because the user selected the VF)
5410 // then we calculate the cost of VF here.
5412 LoopCost = expectedCost(VF);
5414 // Clamp the calculated UF to be between the 1 and the max unroll factor
5415 // that the target allows.
5416 if (UF > MaxUnrollSize)
5421 // Unroll if we vectorized this loop and there is a reduction that could
5422 // benefit from unrolling.
5423 if (VF > 1 && Legal->getReductionVars()->size()) {
5424 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5428 // Note that if we've already vectorized the loop we will have done the
5429 // runtime check and so unrolling won't require further checks.
5430 bool UnrollingRequiresRuntimePointerCheck =
5431 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5433 // We want to unroll small loops in order to reduce the loop overhead and
5434 // potentially expose ILP opportunities.
5435 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5436 if (!UnrollingRequiresRuntimePointerCheck &&
5437 LoopCost < SmallLoopCost) {
5438 // We assume that the cost overhead is 1 and we use the cost model
5439 // to estimate the cost of the loop and unroll until the cost of the
5440 // loop overhead is about 5% of the cost of the loop.
5441 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5443 // Unroll until store/load ports (estimated by max unroll factor) are
5445 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5446 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5448 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5449 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5450 return std::max(StoresUF, LoadsUF);
5453 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5457 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5461 LoopVectorizationCostModel::RegisterUsage
5462 LoopVectorizationCostModel::calculateRegisterUsage() {
5463 // This function calculates the register usage by measuring the highest number
5464 // of values that are alive at a single location. Obviously, this is a very
5465 // rough estimation. We scan the loop in a topological order in order and
5466 // assign a number to each instruction. We use RPO to ensure that defs are
5467 // met before their users. We assume that each instruction that has in-loop
5468 // users starts an interval. We record every time that an in-loop value is
5469 // used, so we have a list of the first and last occurrences of each
5470 // instruction. Next, we transpose this data structure into a multi map that
5471 // holds the list of intervals that *end* at a specific location. This multi
5472 // map allows us to perform a linear search. We scan the instructions linearly
5473 // and record each time that a new interval starts, by placing it in a set.
5474 // If we find this value in the multi-map then we remove it from the set.
5475 // The max register usage is the maximum size of the set.
5476 // We also search for instructions that are defined outside the loop, but are
5477 // used inside the loop. We need this number separately from the max-interval
5478 // usage number because when we unroll, loop-invariant values do not take
5480 LoopBlocksDFS DFS(TheLoop);
5484 R.NumInstructions = 0;
5486 // Each 'key' in the map opens a new interval. The values
5487 // of the map are the index of the 'last seen' usage of the
5488 // instruction that is the key.
5489 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5490 // Maps instruction to its index.
5491 DenseMap<unsigned, Instruction*> IdxToInstr;
5492 // Marks the end of each interval.
5493 IntervalMap EndPoint;
5494 // Saves the list of instruction indices that are used in the loop.
5495 SmallSet<Instruction*, 8> Ends;
5496 // Saves the list of values that are used in the loop but are
5497 // defined outside the loop, such as arguments and constants.
5498 SmallPtrSet<Value*, 8> LoopInvariants;
5501 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5502 be = DFS.endRPO(); bb != be; ++bb) {
5503 R.NumInstructions += (*bb)->size();
5504 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5506 Instruction *I = it;
5507 IdxToInstr[Index++] = I;
5509 // Save the end location of each USE.
5510 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5511 Value *U = I->getOperand(i);
5512 Instruction *Instr = dyn_cast<Instruction>(U);
5514 // Ignore non-instruction values such as arguments, constants, etc.
5515 if (!Instr) continue;
5517 // If this instruction is outside the loop then record it and continue.
5518 if (!TheLoop->contains(Instr)) {
5519 LoopInvariants.insert(Instr);
5523 // Overwrite previous end points.
5524 EndPoint[Instr] = Index;
5530 // Saves the list of intervals that end with the index in 'key'.
5531 typedef SmallVector<Instruction*, 2> InstrList;
5532 DenseMap<unsigned, InstrList> TransposeEnds;
5534 // Transpose the EndPoints to a list of values that end at each index.
5535 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5537 TransposeEnds[it->second].push_back(it->first);
5539 SmallSet<Instruction*, 8> OpenIntervals;
5540 unsigned MaxUsage = 0;
5543 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5544 for (unsigned int i = 0; i < Index; ++i) {
5545 Instruction *I = IdxToInstr[i];
5546 // Ignore instructions that are never used within the loop.
5547 if (!Ends.count(I)) continue;
5549 // Remove all of the instructions that end at this location.
5550 InstrList &List = TransposeEnds[i];
5551 for (unsigned int j=0, e = List.size(); j < e; ++j)
5552 OpenIntervals.erase(List[j]);
5554 // Count the number of live interals.
5555 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5557 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5558 OpenIntervals.size() << '\n');
5560 // Add the current instruction to the list of open intervals.
5561 OpenIntervals.insert(I);
5564 unsigned Invariant = LoopInvariants.size();
5565 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5566 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5567 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5569 R.LoopInvariantRegs = Invariant;
5570 R.MaxLocalUsers = MaxUsage;
5574 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5578 for (Loop::block_iterator bb = TheLoop->block_begin(),
5579 be = TheLoop->block_end(); bb != be; ++bb) {
5580 unsigned BlockCost = 0;
5581 BasicBlock *BB = *bb;
5583 // For each instruction in the old loop.
5584 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5585 // Skip dbg intrinsics.
5586 if (isa<DbgInfoIntrinsic>(it))
5589 unsigned C = getInstructionCost(it, VF);
5591 // Check if we should override the cost.
5592 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5593 C = ForceTargetInstructionCost;
5596 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5597 VF << " For instruction: " << *it << '\n');
5600 // We assume that if-converted blocks have a 50% chance of being executed.
5601 // When the code is scalar then some of the blocks are avoided due to CF.
5602 // When the code is vectorized we execute all code paths.
5603 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5612 /// \brief Check whether the address computation for a non-consecutive memory
5613 /// access looks like an unlikely candidate for being merged into the indexing
5616 /// We look for a GEP which has one index that is an induction variable and all
5617 /// other indices are loop invariant. If the stride of this access is also
5618 /// within a small bound we decide that this address computation can likely be
5619 /// merged into the addressing mode.
5620 /// In all other cases, we identify the address computation as complex.
5621 static bool isLikelyComplexAddressComputation(Value *Ptr,
5622 LoopVectorizationLegality *Legal,
5623 ScalarEvolution *SE,
5624 const Loop *TheLoop) {
5625 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5629 // We are looking for a gep with all loop invariant indices except for one
5630 // which should be an induction variable.
5631 unsigned NumOperands = Gep->getNumOperands();
5632 for (unsigned i = 1; i < NumOperands; ++i) {
5633 Value *Opd = Gep->getOperand(i);
5634 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5635 !Legal->isInductionVariable(Opd))
5639 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5640 // can likely be merged into the address computation.
5641 unsigned MaxMergeDistance = 64;
5643 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5647 // Check the step is constant.
5648 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5649 // Calculate the pointer stride and check if it is consecutive.
5650 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5654 const APInt &APStepVal = C->getValue()->getValue();
5656 // Huge step value - give up.
5657 if (APStepVal.getBitWidth() > 64)
5660 int64_t StepVal = APStepVal.getSExtValue();
5662 return StepVal > MaxMergeDistance;
5665 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5666 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5672 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5673 // If we know that this instruction will remain uniform, check the cost of
5674 // the scalar version.
5675 if (Legal->isUniformAfterVectorization(I))
5678 Type *RetTy = I->getType();
5679 Type *VectorTy = ToVectorTy(RetTy, VF);
5681 // TODO: We need to estimate the cost of intrinsic calls.
5682 switch (I->getOpcode()) {
5683 case Instruction::GetElementPtr:
5684 // We mark this instruction as zero-cost because the cost of GEPs in
5685 // vectorized code depends on whether the corresponding memory instruction
5686 // is scalarized or not. Therefore, we handle GEPs with the memory
5687 // instruction cost.
5689 case Instruction::Br: {
5690 return TTI.getCFInstrCost(I->getOpcode());
5692 case Instruction::PHI:
5693 //TODO: IF-converted IFs become selects.
5695 case Instruction::Add:
5696 case Instruction::FAdd:
5697 case Instruction::Sub:
5698 case Instruction::FSub:
5699 case Instruction::Mul:
5700 case Instruction::FMul:
5701 case Instruction::UDiv:
5702 case Instruction::SDiv:
5703 case Instruction::FDiv:
5704 case Instruction::URem:
5705 case Instruction::SRem:
5706 case Instruction::FRem:
5707 case Instruction::Shl:
5708 case Instruction::LShr:
5709 case Instruction::AShr:
5710 case Instruction::And:
5711 case Instruction::Or:
5712 case Instruction::Xor: {
5713 // Since we will replace the stride by 1 the multiplication should go away.
5714 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5716 // Certain instructions can be cheaper to vectorize if they have a constant
5717 // second vector operand. One example of this are shifts on x86.
5718 TargetTransformInfo::OperandValueKind Op1VK =
5719 TargetTransformInfo::OK_AnyValue;
5720 TargetTransformInfo::OperandValueKind Op2VK =
5721 TargetTransformInfo::OK_AnyValue;
5722 Value *Op2 = I->getOperand(1);
5724 // Check for a splat of a constant or for a non uniform vector of constants.
5725 if (isa<ConstantInt>(Op2))
5726 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5727 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5728 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5729 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5730 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5733 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5735 case Instruction::Select: {
5736 SelectInst *SI = cast<SelectInst>(I);
5737 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5738 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5739 Type *CondTy = SI->getCondition()->getType();
5741 CondTy = VectorType::get(CondTy, VF);
5743 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5745 case Instruction::ICmp:
5746 case Instruction::FCmp: {
5747 Type *ValTy = I->getOperand(0)->getType();
5748 VectorTy = ToVectorTy(ValTy, VF);
5749 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5751 case Instruction::Store:
5752 case Instruction::Load: {
5753 StoreInst *SI = dyn_cast<StoreInst>(I);
5754 LoadInst *LI = dyn_cast<LoadInst>(I);
5755 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5757 VectorTy = ToVectorTy(ValTy, VF);
5759 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5760 unsigned AS = SI ? SI->getPointerAddressSpace() :
5761 LI->getPointerAddressSpace();
5762 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5763 // We add the cost of address computation here instead of with the gep
5764 // instruction because only here we know whether the operation is
5767 return TTI.getAddressComputationCost(VectorTy) +
5768 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5770 // Scalarized loads/stores.
5771 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5772 bool Reverse = ConsecutiveStride < 0;
5773 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5774 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5775 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5776 bool IsComplexComputation =
5777 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5779 // The cost of extracting from the value vector and pointer vector.
5780 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5781 for (unsigned i = 0; i < VF; ++i) {
5782 // The cost of extracting the pointer operand.
5783 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5784 // In case of STORE, the cost of ExtractElement from the vector.
5785 // In case of LOAD, the cost of InsertElement into the returned
5787 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5788 Instruction::InsertElement,
5792 // The cost of the scalar loads/stores.
5793 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5794 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5799 // Wide load/stores.
5800 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5801 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5804 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5808 case Instruction::ZExt:
5809 case Instruction::SExt:
5810 case Instruction::FPToUI:
5811 case Instruction::FPToSI:
5812 case Instruction::FPExt:
5813 case Instruction::PtrToInt:
5814 case Instruction::IntToPtr:
5815 case Instruction::SIToFP:
5816 case Instruction::UIToFP:
5817 case Instruction::Trunc:
5818 case Instruction::FPTrunc:
5819 case Instruction::BitCast: {
5820 // We optimize the truncation of induction variable.
5821 // The cost of these is the same as the scalar operation.
5822 if (I->getOpcode() == Instruction::Trunc &&
5823 Legal->isInductionVariable(I->getOperand(0)))
5824 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5825 I->getOperand(0)->getType());
5827 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5828 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5830 case Instruction::Call: {
5831 CallInst *CI = cast<CallInst>(I);
5832 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5833 assert(ID && "Not an intrinsic call!");
5834 Type *RetTy = ToVectorTy(CI->getType(), VF);
5835 SmallVector<Type*, 4> Tys;
5836 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5837 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5838 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5841 // We are scalarizing the instruction. Return the cost of the scalar
5842 // instruction, plus the cost of insert and extract into vector
5843 // elements, times the vector width.
5846 if (!RetTy->isVoidTy() && VF != 1) {
5847 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5849 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5852 // The cost of inserting the results plus extracting each one of the
5854 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5857 // The cost of executing VF copies of the scalar instruction. This opcode
5858 // is unknown. Assume that it is the same as 'mul'.
5859 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5865 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5866 if (Scalar->isVoidTy() || VF == 1)
5868 return VectorType::get(Scalar, VF);
5871 char LoopVectorize::ID = 0;
5872 static const char lv_name[] = "Loop Vectorization";
5873 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5874 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5875 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5876 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5877 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5878 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5879 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5880 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5881 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5884 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5885 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5889 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5890 // Check for a store.
5891 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5892 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5894 // Check for a load.
5895 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5896 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5902 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5903 bool IfPredicateStore) {
5904 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5905 // Holds vector parameters or scalars, in case of uniform vals.
5906 SmallVector<VectorParts, 4> Params;
5908 setDebugLocFromInst(Builder, Instr);
5910 // Find all of the vectorized parameters.
5911 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5912 Value *SrcOp = Instr->getOperand(op);
5914 // If we are accessing the old induction variable, use the new one.
5915 if (SrcOp == OldInduction) {
5916 Params.push_back(getVectorValue(SrcOp));
5920 // Try using previously calculated values.
5921 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5923 // If the src is an instruction that appeared earlier in the basic block
5924 // then it should already be vectorized.
5925 if (SrcInst && OrigLoop->contains(SrcInst)) {
5926 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5927 // The parameter is a vector value from earlier.
5928 Params.push_back(WidenMap.get(SrcInst));
5930 // The parameter is a scalar from outside the loop. Maybe even a constant.
5931 VectorParts Scalars;
5932 Scalars.append(UF, SrcOp);
5933 Params.push_back(Scalars);
5937 assert(Params.size() == Instr->getNumOperands() &&
5938 "Invalid number of operands");
5940 // Does this instruction return a value ?
5941 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5943 Value *UndefVec = IsVoidRetTy ? nullptr :
5944 UndefValue::get(Instr->getType());
5945 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5946 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5948 Instruction *InsertPt = Builder.GetInsertPoint();
5949 BasicBlock *IfBlock = Builder.GetInsertBlock();
5950 BasicBlock *CondBlock = nullptr;
5953 Loop *VectorLp = nullptr;
5954 if (IfPredicateStore) {
5955 assert(Instr->getParent()->getSinglePredecessor() &&
5956 "Only support single predecessor blocks");
5957 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5958 Instr->getParent());
5959 VectorLp = LI->getLoopFor(IfBlock);
5960 assert(VectorLp && "Must have a loop for this block");
5963 // For each vector unroll 'part':
5964 for (unsigned Part = 0; Part < UF; ++Part) {
5965 // For each scalar that we create:
5967 // Start an "if (pred) a[i] = ..." block.
5968 Value *Cmp = nullptr;
5969 if (IfPredicateStore) {
5970 if (Cond[Part]->getType()->isVectorTy())
5972 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5973 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5974 ConstantInt::get(Cond[Part]->getType(), 1));
5975 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5976 LoopVectorBody.push_back(CondBlock);
5977 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
5978 // Update Builder with newly created basic block.
5979 Builder.SetInsertPoint(InsertPt);
5982 Instruction *Cloned = Instr->clone();
5984 Cloned->setName(Instr->getName() + ".cloned");
5985 // Replace the operands of the cloned instructions with extracted scalars.
5986 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5987 Value *Op = Params[op][Part];
5988 Cloned->setOperand(op, Op);
5991 // Place the cloned scalar in the new loop.
5992 Builder.Insert(Cloned);
5994 // If the original scalar returns a value we need to place it in a vector
5995 // so that future users will be able to use it.
5997 VecResults[Part] = Cloned;
6000 if (IfPredicateStore) {
6001 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6002 LoopVectorBody.push_back(NewIfBlock);
6003 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6004 Builder.SetInsertPoint(InsertPt);
6005 Instruction *OldBr = IfBlock->getTerminator();
6006 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6007 OldBr->eraseFromParent();
6008 IfBlock = NewIfBlock;
6013 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6014 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6015 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6017 return scalarizeInstruction(Instr, IfPredicateStore);
6020 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6024 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6028 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6030 // When unrolling and the VF is 1, we only need to add a simple scalar.
6031 Type *ITy = Val->getType();
6032 assert(!ITy->isVectorTy() && "Val must be a scalar");
6033 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6034 return Builder.CreateAdd(Val, C, "induction");