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/AliasSetTracker.h"
58 #include "llvm/Analysis/BlockFrequencyInfo.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugInfo.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/DiagnosticInfo.h"
72 #include "llvm/IR/Dominators.h"
73 #include "llvm/IR/Function.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/Instructions.h"
76 #include "llvm/IR/IntrinsicInst.h"
77 #include "llvm/IR/LLVMContext.h"
78 #include "llvm/IR/Module.h"
79 #include "llvm/IR/PatternMatch.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/Value.h"
82 #include "llvm/IR/ValueHandle.h"
83 #include "llvm/IR/Verifier.h"
84 #include "llvm/Pass.h"
85 #include "llvm/Support/BranchProbability.h"
86 #include "llvm/Support/CommandLine.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/Local.h"
92 #include "llvm/Transforms/Utils/VectorUtils.h"
98 using namespace llvm::PatternMatch;
100 #define LV_NAME "loop-vectorize"
101 #define DEBUG_TYPE LV_NAME
103 STATISTIC(LoopsVectorized, "Number of loops vectorized");
104 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
106 static cl::opt<unsigned>
107 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
108 cl::desc("Sets the SIMD width. Zero is autoselect."));
110 static cl::opt<unsigned>
111 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
112 cl::desc("Sets the vectorization unroll count. "
113 "Zero is autoselect."));
116 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
117 cl::desc("Enable if-conversion during vectorization."));
119 /// We don't vectorize loops with a known constant trip count below this number.
120 static cl::opt<unsigned>
121 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
123 cl::desc("Don't vectorize loops with a constant "
124 "trip count that is smaller than this "
127 /// This enables versioning on the strides of symbolically striding memory
128 /// accesses in code like the following.
129 /// for (i = 0; i < N; ++i)
130 /// A[i * Stride1] += B[i * Stride2] ...
132 /// Will be roughly translated to
133 /// if (Stride1 == 1 && Stride2 == 1) {
134 /// for (i = 0; i < N; i+=4)
138 static cl::opt<bool> EnableMemAccessVersioning(
139 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
140 cl::desc("Enable symblic stride memory access versioning"));
142 /// We don't unroll loops with a known constant trip count below this number.
143 static const unsigned TinyTripCountUnrollThreshold = 128;
145 /// When performing memory disambiguation checks at runtime do not make more
146 /// than this number of comparisons.
147 static const unsigned RuntimeMemoryCheckThreshold = 8;
149 /// Maximum simd width.
150 static const unsigned MaxVectorWidth = 64;
152 static cl::opt<unsigned> ForceTargetNumScalarRegs(
153 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
154 cl::desc("A flag that overrides the target's number of scalar registers."));
156 static cl::opt<unsigned> ForceTargetNumVectorRegs(
157 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
158 cl::desc("A flag that overrides the target's number of vector registers."));
160 /// Maximum vectorization unroll count.
161 static const unsigned MaxUnrollFactor = 16;
163 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
164 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's max unroll factor for scalar "
168 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
169 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
170 cl::desc("A flag that overrides the target's max unroll factor for "
171 "vectorized loops."));
173 static cl::opt<unsigned> ForceTargetInstructionCost(
174 "force-target-instruction-cost", cl::init(0), cl::Hidden,
175 cl::desc("A flag that overrides the target's expected cost for "
176 "an instruction to a single constant value. Mostly "
177 "useful for getting consistent testing."));
179 static cl::opt<unsigned> SmallLoopCost(
180 "small-loop-cost", cl::init(20), cl::Hidden,
181 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
183 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
184 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
185 cl::desc("Enable the use of the block frequency analysis to access PGO "
186 "heuristics minimizing code growth in cold regions and being more "
187 "aggressive in hot regions."));
189 // Runtime unroll loops for load/store throughput.
190 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
191 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
192 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
194 /// The number of stores in a loop that are allowed to need predication.
195 static cl::opt<unsigned> NumberOfStoresToPredicate(
196 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
197 cl::desc("Max number of stores to be predicated behind an if."));
199 static cl::opt<bool> EnableIndVarRegisterHeur(
200 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
201 cl::desc("Count the induction variable only once when unrolling"));
203 static cl::opt<bool> EnableCondStoresVectorization(
204 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
205 cl::desc("Enable if predication of stores during vectorization."));
207 static cl::opt<unsigned> MaxNestedScalarReductionUF(
208 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
209 cl::desc("The maximum unroll factor to use when unrolling a scalar "
210 "reduction in a nested loop."));
214 // Forward declarations.
215 class LoopVectorizationLegality;
216 class LoopVectorizationCostModel;
217 class LoopVectorizeHints;
219 /// Optimization analysis message produced during vectorization. Messages inform
220 /// the user why vectorization did not occur.
223 raw_string_ostream Out;
227 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
228 Out << "loop not vectorized: ";
231 template <typename A> Report &operator<<(const A &Value) {
236 Instruction *getInstr() { return Instr; }
238 std::string &str() { return Out.str(); }
239 operator Twine() { return Out.str(); }
242 /// InnerLoopVectorizer vectorizes loops which contain only one basic
243 /// block to a specified vectorization factor (VF).
244 /// This class performs the widening of scalars into vectors, or multiple
245 /// scalars. This class also implements the following features:
246 /// * It inserts an epilogue loop for handling loops that don't have iteration
247 /// counts that are known to be a multiple of the vectorization factor.
248 /// * It handles the code generation for reduction variables.
249 /// * Scalarization (implementation using scalars) of un-vectorizable
251 /// InnerLoopVectorizer does not perform any vectorization-legality
252 /// checks, and relies on the caller to check for the different legality
253 /// aspects. The InnerLoopVectorizer relies on the
254 /// LoopVectorizationLegality class to provide information about the induction
255 /// and reduction variables that were found to a given vectorization factor.
256 class InnerLoopVectorizer {
258 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
259 DominatorTree *DT, const DataLayout *DL,
260 const TargetLibraryInfo *TLI, unsigned VecWidth,
261 unsigned UnrollFactor)
262 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
263 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
264 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
267 // Perform the actual loop widening (vectorization).
268 void vectorize(LoopVectorizationLegality *L) {
270 // Create a new empty loop. Unlink the old loop and connect the new one.
272 // Widen each instruction in the old loop to a new one in the new loop.
273 // Use the Legality module to find the induction and reduction variables.
275 // Register the new loop and update the analysis passes.
279 virtual ~InnerLoopVectorizer() {}
282 /// A small list of PHINodes.
283 typedef SmallVector<PHINode*, 4> PhiVector;
284 /// When we unroll loops we have multiple vector values for each scalar.
285 /// This data structure holds the unrolled and vectorized values that
286 /// originated from one scalar instruction.
287 typedef SmallVector<Value*, 2> VectorParts;
289 // When we if-convert we need create edge masks. We have to cache values so
290 // that we don't end up with exponential recursion/IR.
291 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
292 VectorParts> EdgeMaskCache;
294 /// \brief Add code that checks at runtime if the accessed arrays overlap.
296 /// Returns a pair of instructions where the first element is the first
297 /// instruction generated in possibly a sequence of instructions and the
298 /// second value is the final comparator value or NULL if no check is needed.
299 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
301 /// \brief Add checks for strides that where assumed to be 1.
303 /// Returns the last check instruction and the first check instruction in the
304 /// pair as (first, last).
305 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
307 /// Create an empty loop, based on the loop ranges of the old loop.
308 void createEmptyLoop();
309 /// Copy and widen the instructions from the old loop.
310 virtual void vectorizeLoop();
312 /// \brief The Loop exit block may have single value PHI nodes where the
313 /// incoming value is 'Undef'. While vectorizing we only handled real values
314 /// that were defined inside the loop. Here we fix the 'undef case'.
318 /// A helper function that computes the predicate of the block BB, assuming
319 /// that the header block of the loop is set to True. It returns the *entry*
320 /// mask for the block BB.
321 VectorParts createBlockInMask(BasicBlock *BB);
322 /// A helper function that computes the predicate of the edge between SRC
324 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
326 /// A helper function to vectorize a single BB within the innermost loop.
327 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
329 /// Vectorize a single PHINode in a block. This method handles the induction
330 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
331 /// arbitrary length vectors.
332 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
333 unsigned UF, unsigned VF, PhiVector *PV);
335 /// Insert the new loop to the loop hierarchy and pass manager
336 /// and update the analysis passes.
337 void updateAnalysis();
339 /// This instruction is un-vectorizable. Implement it as a sequence
340 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
341 /// scalarized instruction behind an if block predicated on the control
342 /// dependence of the instruction.
343 virtual void scalarizeInstruction(Instruction *Instr,
344 bool IfPredicateStore=false);
346 /// Vectorize Load and Store instructions,
347 virtual void vectorizeMemoryInstruction(Instruction *Instr);
349 /// Create a broadcast instruction. This method generates a broadcast
350 /// instruction (shuffle) for loop invariant values and for the induction
351 /// value. If this is the induction variable then we extend it to N, N+1, ...
352 /// this is needed because each iteration in the loop corresponds to a SIMD
354 virtual Value *getBroadcastInstrs(Value *V);
356 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
357 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
358 /// The sequence starts at StartIndex.
359 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
361 /// When we go over instructions in the basic block we rely on previous
362 /// values within the current basic block or on loop invariant values.
363 /// When we widen (vectorize) values we place them in the map. If the values
364 /// are not within the map, they have to be loop invariant, so we simply
365 /// broadcast them into a vector.
366 VectorParts &getVectorValue(Value *V);
368 /// Generate a shuffle sequence that will reverse the vector Vec.
369 virtual Value *reverseVector(Value *Vec);
371 /// This is a helper class that holds the vectorizer state. It maps scalar
372 /// instructions to vector instructions. When the code is 'unrolled' then
373 /// then a single scalar value is mapped to multiple vector parts. The parts
374 /// are stored in the VectorPart type.
376 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
378 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
380 /// \return True if 'Key' is saved in the Value Map.
381 bool has(Value *Key) const { return MapStorage.count(Key); }
383 /// Initializes a new entry in the map. Sets all of the vector parts to the
384 /// save value in 'Val'.
385 /// \return A reference to a vector with splat values.
386 VectorParts &splat(Value *Key, Value *Val) {
387 VectorParts &Entry = MapStorage[Key];
388 Entry.assign(UF, Val);
392 ///\return A reference to the value that is stored at 'Key'.
393 VectorParts &get(Value *Key) {
394 VectorParts &Entry = MapStorage[Key];
397 assert(Entry.size() == UF);
402 /// The unroll factor. Each entry in the map stores this number of vector
406 /// Map storage. We use std::map and not DenseMap because insertions to a
407 /// dense map invalidates its iterators.
408 std::map<Value *, VectorParts> MapStorage;
411 /// The original loop.
413 /// Scev analysis to use.
422 const DataLayout *DL;
423 /// Target Library Info.
424 const TargetLibraryInfo *TLI;
426 /// The vectorization SIMD factor to use. Each vector will have this many
431 /// The vectorization unroll factor to use. Each scalar is vectorized to this
432 /// many different vector instructions.
435 /// The builder that we use
438 // --- Vectorization state ---
440 /// The vector-loop preheader.
441 BasicBlock *LoopVectorPreHeader;
442 /// The scalar-loop preheader.
443 BasicBlock *LoopScalarPreHeader;
444 /// Middle Block between the vector and the scalar.
445 BasicBlock *LoopMiddleBlock;
446 ///The ExitBlock of the scalar loop.
447 BasicBlock *LoopExitBlock;
448 ///The vector loop body.
449 SmallVector<BasicBlock *, 4> LoopVectorBody;
450 ///The scalar loop body.
451 BasicBlock *LoopScalarBody;
452 /// A list of all bypass blocks. The first block is the entry of the loop.
453 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
455 /// The new Induction variable which was added to the new block.
457 /// The induction variable of the old basic block.
458 PHINode *OldInduction;
459 /// Holds the extended (to the widest induction type) start index.
461 /// Maps scalars to widened vectors.
463 EdgeMaskCache MaskCache;
465 LoopVectorizationLegality *Legal;
468 class InnerLoopUnroller : public InnerLoopVectorizer {
470 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
471 DominatorTree *DT, const DataLayout *DL,
472 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
473 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
476 void scalarizeInstruction(Instruction *Instr,
477 bool IfPredicateStore = false) override;
478 void vectorizeMemoryInstruction(Instruction *Instr) override;
479 Value *getBroadcastInstrs(Value *V) override;
480 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
481 Value *reverseVector(Value *Vec) override;
484 /// \brief Look for a meaningful debug location on the instruction or it's
486 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
491 if (I->getDebugLoc() != Empty)
494 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
495 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
496 if (OpInst->getDebugLoc() != Empty)
503 /// \brief Set the debug location in the builder using the debug location in the
505 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
506 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
507 B.SetCurrentDebugLocation(Inst->getDebugLoc());
509 B.SetCurrentDebugLocation(DebugLoc());
513 /// \return string containing a file name and a line # for the given loop.
514 static std::string getDebugLocString(const Loop *L) {
517 raw_string_ostream OS(Result);
518 const DebugLoc LoopDbgLoc = L->getStartLoc();
519 if (!LoopDbgLoc.isUnknown())
520 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
522 // Just print the module name.
523 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
530 /// \brief Propagate known metadata from one instruction to another.
531 static void propagateMetadata(Instruction *To, const Instruction *From) {
532 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
533 From->getAllMetadataOtherThanDebugLoc(Metadata);
535 for (auto M : Metadata) {
536 unsigned Kind = M.first;
538 // These are safe to transfer (this is safe for TBAA, even when we
539 // if-convert, because should that metadata have had a control dependency
540 // on the condition, and thus actually aliased with some other
541 // non-speculated memory access when the condition was false, this would be
542 // caught by the runtime overlap checks).
543 if (Kind != LLVMContext::MD_tbaa &&
544 Kind != LLVMContext::MD_alias_scope &&
545 Kind != LLVMContext::MD_noalias &&
546 Kind != LLVMContext::MD_fpmath)
549 To->setMetadata(Kind, M.second);
553 /// \brief Propagate known metadata from one instruction to a vector of others.
554 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
556 if (Instruction *I = dyn_cast<Instruction>(V))
557 propagateMetadata(I, From);
560 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
561 /// to what vectorization factor.
562 /// This class does not look at the profitability of vectorization, only the
563 /// legality. This class has two main kinds of checks:
564 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
565 /// will change the order of memory accesses in a way that will change the
566 /// correctness of the program.
567 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
568 /// checks for a number of different conditions, such as the availability of a
569 /// single induction variable, that all types are supported and vectorize-able,
570 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
571 /// This class is also used by InnerLoopVectorizer for identifying
572 /// induction variable and the different reduction variables.
573 class LoopVectorizationLegality {
577 unsigned NumPredStores;
579 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
580 DominatorTree *DT, TargetLibraryInfo *TLI,
581 AliasAnalysis *AA, Function *F)
582 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
583 DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
584 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
587 /// This enum represents the kinds of reductions that we support.
589 RK_NoReduction, ///< Not a reduction.
590 RK_IntegerAdd, ///< Sum of integers.
591 RK_IntegerMult, ///< Product of integers.
592 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
593 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
594 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
595 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
596 RK_FloatAdd, ///< Sum of floats.
597 RK_FloatMult, ///< Product of floats.
598 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
601 /// This enum represents the kinds of inductions that we support.
603 IK_NoInduction, ///< Not an induction variable.
604 IK_IntInduction, ///< Integer induction variable. Step = 1.
605 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
606 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
607 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
610 // This enum represents the kind of minmax reduction.
611 enum MinMaxReductionKind {
621 /// This struct holds information about reduction variables.
622 struct ReductionDescriptor {
623 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
624 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
626 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
627 MinMaxReductionKind MK)
628 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
630 // The starting value of the reduction.
631 // It does not have to be zero!
632 TrackingVH<Value> StartValue;
633 // The instruction who's value is used outside the loop.
634 Instruction *LoopExitInstr;
635 // The kind of the reduction.
637 // If this a min/max reduction the kind of reduction.
638 MinMaxReductionKind MinMaxKind;
641 /// This POD struct holds information about a potential reduction operation.
642 struct ReductionInstDesc {
643 ReductionInstDesc(bool IsRedux, Instruction *I) :
644 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
646 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
647 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
649 // Is this instruction a reduction candidate.
651 // The last instruction in a min/max pattern (select of the select(icmp())
652 // pattern), or the current reduction instruction otherwise.
653 Instruction *PatternLastInst;
654 // If this is a min/max pattern the comparison predicate.
655 MinMaxReductionKind MinMaxKind;
658 /// This struct holds information about the memory runtime legality
659 /// check that a group of pointers do not overlap.
660 struct RuntimePointerCheck {
661 RuntimePointerCheck() : Need(false) {}
663 /// Reset the state of the pointer runtime information.
670 DependencySetId.clear();
674 /// Insert a pointer and calculate the start and end SCEVs.
675 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
676 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
678 /// This flag indicates if we need to add the runtime check.
680 /// Holds the pointers that we need to check.
681 SmallVector<TrackingVH<Value>, 2> Pointers;
682 /// Holds the pointer value at the beginning of the loop.
683 SmallVector<const SCEV*, 2> Starts;
684 /// Holds the pointer value at the end of the loop.
685 SmallVector<const SCEV*, 2> Ends;
686 /// Holds the information if this pointer is used for writing to memory.
687 SmallVector<bool, 2> IsWritePtr;
688 /// Holds the id of the set of pointers that could be dependent because of a
689 /// shared underlying object.
690 SmallVector<unsigned, 2> DependencySetId;
691 /// Holds the id of the disjoint alias set to which this pointer belongs.
692 SmallVector<unsigned, 2> AliasSetId;
695 /// A struct for saving information about induction variables.
696 struct InductionInfo {
697 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
698 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
700 TrackingVH<Value> StartValue;
705 /// ReductionList contains the reduction descriptors for all
706 /// of the reductions that were found in the loop.
707 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
709 /// InductionList saves induction variables and maps them to the
710 /// induction descriptor.
711 typedef MapVector<PHINode*, InductionInfo> InductionList;
713 /// Returns true if it is legal to vectorize this loop.
714 /// This does not mean that it is profitable to vectorize this
715 /// loop, only that it is legal to do so.
718 /// Returns the Induction variable.
719 PHINode *getInduction() { return Induction; }
721 /// Returns the reduction variables found in the loop.
722 ReductionList *getReductionVars() { return &Reductions; }
724 /// Returns the induction variables found in the loop.
725 InductionList *getInductionVars() { return &Inductions; }
727 /// Returns the widest induction type.
728 Type *getWidestInductionType() { return WidestIndTy; }
730 /// Returns True if V is an induction variable in this loop.
731 bool isInductionVariable(const Value *V);
733 /// Return true if the block BB needs to be predicated in order for the loop
734 /// to be vectorized.
735 bool blockNeedsPredication(BasicBlock *BB);
737 /// Check if this pointer is consecutive when vectorizing. This happens
738 /// when the last index of the GEP is the induction variable, or that the
739 /// pointer itself is an induction variable.
740 /// This check allows us to vectorize A[idx] into a wide load/store.
742 /// 0 - Stride is unknown or non-consecutive.
743 /// 1 - Address is consecutive.
744 /// -1 - Address is consecutive, and decreasing.
745 int isConsecutivePtr(Value *Ptr);
747 /// Returns true if the value V is uniform within the loop.
748 bool isUniform(Value *V);
750 /// Returns true if this instruction will remain scalar after vectorization.
751 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
753 /// Returns the information that we collected about runtime memory check.
754 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
756 /// This function returns the identity element (or neutral element) for
758 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
760 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
762 bool hasStride(Value *V) { return StrideSet.count(V); }
763 bool mustCheckStrides() { return !StrideSet.empty(); }
764 SmallPtrSet<Value *, 8>::iterator strides_begin() {
765 return StrideSet.begin();
767 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
770 /// Check if a single basic block loop is vectorizable.
771 /// At this point we know that this is a loop with a constant trip count
772 /// and we only need to check individual instructions.
773 bool canVectorizeInstrs();
775 /// When we vectorize loops we may change the order in which
776 /// we read and write from memory. This method checks if it is
777 /// legal to vectorize the code, considering only memory constrains.
778 /// Returns true if the loop is vectorizable
779 bool canVectorizeMemory();
781 /// Return true if we can vectorize this loop using the IF-conversion
783 bool canVectorizeWithIfConvert();
785 /// Collect the variables that need to stay uniform after vectorization.
786 void collectLoopUniforms();
788 /// Return true if all of the instructions in the block can be speculatively
789 /// executed. \p SafePtrs is a list of addresses that are known to be legal
790 /// and we know that we can read from them without segfault.
791 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
793 /// Returns True, if 'Phi' is the kind of reduction variable for type
794 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
795 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
796 /// Returns a struct describing if the instruction 'I' can be a reduction
797 /// variable of type 'Kind'. If the reduction is a min/max pattern of
798 /// select(icmp()) this function advances the instruction pointer 'I' from the
799 /// compare instruction to the select instruction and stores this pointer in
800 /// 'PatternLastInst' member of the returned struct.
801 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
802 ReductionInstDesc &Desc);
803 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
804 /// pattern corresponding to a min(X, Y) or max(X, Y).
805 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
806 ReductionInstDesc &Prev);
807 /// Returns the induction kind of Phi. This function may return NoInduction
808 /// if the PHI is not an induction variable.
809 InductionKind isInductionVariable(PHINode *Phi);
811 /// \brief Collect memory access with loop invariant strides.
813 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
815 void collectStridedAcccess(Value *LoadOrStoreInst);
817 /// Report an analysis message to assist the user in diagnosing loops that are
819 void emitAnalysis(Report &Message) {
820 DebugLoc DL = TheLoop->getStartLoc();
821 if (Instruction *I = Message.getInstr())
822 DL = I->getDebugLoc();
823 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
824 *TheFunction, DL, Message.str());
827 /// The loop that we evaluate.
831 /// DataLayout analysis.
832 const DataLayout *DL;
835 /// Target Library Info.
836 TargetLibraryInfo *TLI;
840 Function *TheFunction;
842 // --- vectorization state --- //
844 /// Holds the integer induction variable. This is the counter of the
847 /// Holds the reduction variables.
848 ReductionList Reductions;
849 /// Holds all of the induction variables that we found in the loop.
850 /// Notice that inductions don't need to start at zero and that induction
851 /// variables can be pointers.
852 InductionList Inductions;
853 /// Holds the widest induction type encountered.
856 /// Allowed outside users. This holds the reduction
857 /// vars which can be accessed from outside the loop.
858 SmallPtrSet<Value*, 4> AllowedExit;
859 /// This set holds the variables which are known to be uniform after
861 SmallPtrSet<Instruction*, 4> Uniforms;
862 /// We need to check that all of the pointers in this list are disjoint
864 RuntimePointerCheck PtrRtCheck;
865 /// Can we assume the absence of NaNs.
866 bool HasFunNoNaNAttr;
868 unsigned MaxSafeDepDistBytes;
870 ValueToValueMap Strides;
871 SmallPtrSet<Value *, 8> StrideSet;
874 /// LoopVectorizationCostModel - estimates the expected speedups due to
876 /// In many cases vectorization is not profitable. This can happen because of
877 /// a number of reasons. In this class we mainly attempt to predict the
878 /// expected speedup/slowdowns due to the supported instruction set. We use the
879 /// TargetTransformInfo to query the different backends for the cost of
880 /// different operations.
881 class LoopVectorizationCostModel {
883 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
884 LoopVectorizationLegality *Legal,
885 const TargetTransformInfo &TTI,
886 const DataLayout *DL, const TargetLibraryInfo *TLI,
887 const Function *F, const LoopVectorizeHints *Hints)
888 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI), TheFunction(F), Hints(Hints) {}
890 /// Information about vectorization costs
891 struct VectorizationFactor {
892 unsigned Width; // Vector width with best cost
893 unsigned Cost; // Cost of the loop with that width
895 /// \return The most profitable vectorization factor and the cost of that VF.
896 /// This method checks every power of two up to VF. If UserVF is not ZERO
897 /// then this vectorization factor will be selected if vectorization is
899 VectorizationFactor selectVectorizationFactor(bool OptForSize);
901 /// \return The size (in bits) of the widest type in the code that
902 /// needs to be vectorized. We ignore values that remain scalar such as
903 /// 64 bit loop indices.
904 unsigned getWidestType();
906 /// \return The most profitable unroll factor.
907 /// If UserUF is non-zero then this method finds the best unroll-factor
908 /// based on register pressure and other parameters.
909 /// VF and LoopCost are the selected vectorization factor and the cost of the
911 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
913 /// \brief A struct that represents some properties of the register usage
915 struct RegisterUsage {
916 /// Holds the number of loop invariant values that are used in the loop.
917 unsigned LoopInvariantRegs;
918 /// Holds the maximum number of concurrent live intervals in the loop.
919 unsigned MaxLocalUsers;
920 /// Holds the number of instructions in the loop.
921 unsigned NumInstructions;
924 /// \return information about the register usage of the loop.
925 RegisterUsage calculateRegisterUsage();
928 /// Returns the expected execution cost. The unit of the cost does
929 /// not matter because we use the 'cost' units to compare different
930 /// vector widths. The cost that is returned is *not* normalized by
931 /// the factor width.
932 unsigned expectedCost(unsigned VF);
934 /// Returns the execution time cost of an instruction for a given vector
935 /// width. Vector width of one means scalar.
936 unsigned getInstructionCost(Instruction *I, unsigned VF);
938 /// A helper function for converting Scalar types to vector types.
939 /// If the incoming type is void, we return void. If the VF is 1, we return
941 static Type* ToVectorTy(Type *Scalar, unsigned VF);
943 /// Returns whether the instruction is a load or store and will be a emitted
944 /// as a vector operation.
945 bool isConsecutiveLoadOrStore(Instruction *I);
947 /// Report an analysis message to assist the user in diagnosing loops that are
949 void emitAnalysis(Report &Message) {
950 DebugLoc DL = TheLoop->getStartLoc();
951 if (Instruction *I = Message.getInstr())
952 DL = I->getDebugLoc();
953 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
954 *TheFunction, DL, Message.str());
957 /// The loop that we evaluate.
961 /// Loop Info analysis.
963 /// Vectorization legality.
964 LoopVectorizationLegality *Legal;
965 /// Vector target information.
966 const TargetTransformInfo &TTI;
967 /// Target data layout information.
968 const DataLayout *DL;
969 /// Target Library Info.
970 const TargetLibraryInfo *TLI;
971 const Function *TheFunction;
972 // Loop Vectorize Hint.
973 const LoopVectorizeHints *Hints;
976 /// Utility class for getting and setting loop vectorizer hints in the form
977 /// of loop metadata.
978 class LoopVectorizeHints {
981 FK_Undefined = -1, ///< Not selected.
982 FK_Disabled = 0, ///< Forcing disabled.
983 FK_Enabled = 1, ///< Forcing enabled.
986 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
987 : Width(VectorizationFactor),
988 Unroll(DisableUnrolling),
990 LoopID(L->getLoopID()) {
992 // force-vector-unroll overrides DisableUnrolling.
993 if (VectorizationUnroll.getNumOccurrences() > 0)
994 Unroll = VectorizationUnroll;
996 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
997 << "LV: Unrolling disabled by the pass manager\n");
1000 /// Return the loop metadata prefix.
1001 static StringRef Prefix() { return "llvm.loop."; }
1003 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
1004 SmallVector<Value*, 2> Vals;
1005 Vals.push_back(MDString::get(Context, Name));
1006 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
1007 return MDNode::get(Context, Vals);
1010 /// Mark the loop L as already vectorized by setting the width to 1.
1011 void setAlreadyVectorized(Loop *L) {
1012 LLVMContext &Context = L->getHeader()->getContext();
1016 // Create a new loop id with one more operand for the already_vectorized
1017 // hint. If the loop already has a loop id then copy the existing operands.
1018 SmallVector<Value*, 4> Vals(1);
1020 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
1021 Vals.push_back(LoopID->getOperand(i));
1024 createHint(Context, Twine(Prefix(), "vectorize.width").str(), Width));
1026 createHint(Context, Twine(Prefix(), "interleave.count").str(), 1));
1028 MDNode *NewLoopID = MDNode::get(Context, Vals);
1029 // Set operand 0 to refer to the loop id itself.
1030 NewLoopID->replaceOperandWith(0, NewLoopID);
1032 L->setLoopID(NewLoopID);
1034 LoopID->replaceAllUsesWith(NewLoopID);
1039 std::string emitRemark() const {
1041 if (Force == LoopVectorizeHints::FK_Disabled)
1042 R << "vectorization is explicitly disabled";
1044 R << "use -Rpass-analysis=loop-vectorize for more info";
1045 if (Force == LoopVectorizeHints::FK_Enabled) {
1046 R << " (Force=true";
1048 R << ", Vector Width=" << Width;
1050 R << ", Interleave Count=" << Unroll;
1058 unsigned getWidth() const { return Width; }
1059 unsigned getUnroll() const { return Unroll; }
1060 enum ForceKind getForce() const { return Force; }
1061 MDNode *getLoopID() const { return LoopID; }
1064 /// Find hints specified in the loop metadata.
1065 void getHints(const Loop *L) {
1069 // First operand should refer to the loop id itself.
1070 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1071 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1073 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1074 const MDString *S = nullptr;
1075 SmallVector<Value*, 4> Args;
1077 // The expected hint is either a MDString or a MDNode with the first
1078 // operand a MDString.
1079 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1080 if (!MD || MD->getNumOperands() == 0)
1082 S = dyn_cast<MDString>(MD->getOperand(0));
1083 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1084 Args.push_back(MD->getOperand(i));
1086 S = dyn_cast<MDString>(LoopID->getOperand(i));
1087 assert(Args.size() == 0 && "too many arguments for MDString");
1093 // Check if the hint starts with the loop metadata prefix.
1094 StringRef Hint = S->getString();
1095 if (!Hint.startswith(Prefix()))
1097 // Remove the prefix.
1098 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1100 if (Args.size() == 1)
1101 getHint(Hint, Args[0]);
1105 // Check string hint with one operand.
1106 void getHint(StringRef Hint, Value *Arg) {
1107 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1109 unsigned Val = C->getZExtValue();
1111 if (Hint == "vectorize.width") {
1112 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1115 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1116 } else if (Hint == "vectorize.enable") {
1117 if (C->getBitWidth() == 1)
1118 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1119 : LoopVectorizeHints::FK_Disabled;
1121 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1122 } else if (Hint == "interleave.count") {
1123 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1126 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1128 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1132 /// Vectorization width.
1134 /// Vectorization unroll factor.
1136 /// Vectorization forced
1137 enum ForceKind Force;
1142 static void emitMissedWarning(Function *F, Loop *L,
1143 const LoopVectorizeHints &LH) {
1144 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1145 L->getStartLoc(), LH.emitRemark());
1147 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1148 if (LH.getWidth() != 1)
1149 emitLoopVectorizeWarning(
1150 F->getContext(), *F, L->getStartLoc(),
1151 "failed explicitly specified loop vectorization");
1152 else if (LH.getUnroll() != 1)
1153 emitLoopInterleaveWarning(
1154 F->getContext(), *F, L->getStartLoc(),
1155 "failed explicitly specified loop interleaving");
1159 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1161 return V.push_back(&L);
1163 for (Loop *InnerL : L)
1164 addInnerLoop(*InnerL, V);
1167 /// The LoopVectorize Pass.
1168 struct LoopVectorize : public FunctionPass {
1169 /// Pass identification, replacement for typeid
1172 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1174 DisableUnrolling(NoUnrolling),
1175 AlwaysVectorize(AlwaysVectorize) {
1176 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1179 ScalarEvolution *SE;
1180 const DataLayout *DL;
1182 TargetTransformInfo *TTI;
1184 BlockFrequencyInfo *BFI;
1185 TargetLibraryInfo *TLI;
1187 bool DisableUnrolling;
1188 bool AlwaysVectorize;
1190 BlockFrequency ColdEntryFreq;
1192 bool runOnFunction(Function &F) override {
1193 SE = &getAnalysis<ScalarEvolution>();
1194 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1195 DL = DLP ? &DLP->getDataLayout() : nullptr;
1196 LI = &getAnalysis<LoopInfo>();
1197 TTI = &getAnalysis<TargetTransformInfo>();
1198 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1199 BFI = &getAnalysis<BlockFrequencyInfo>();
1200 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1201 AA = &getAnalysis<AliasAnalysis>();
1203 // Compute some weights outside of the loop over the loops. Compute this
1204 // using a BranchProbability to re-use its scaling math.
1205 const BranchProbability ColdProb(1, 5); // 20%
1206 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1208 // If the target claims to have no vector registers don't attempt
1210 if (!TTI->getNumberOfRegisters(true))
1214 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1215 << ": Missing data layout\n");
1219 // Build up a worklist of inner-loops to vectorize. This is necessary as
1220 // the act of vectorizing or partially unrolling a loop creates new loops
1221 // and can invalidate iterators across the loops.
1222 SmallVector<Loop *, 8> Worklist;
1225 addInnerLoop(*L, Worklist);
1227 LoopsAnalyzed += Worklist.size();
1229 // Now walk the identified inner loops.
1230 bool Changed = false;
1231 while (!Worklist.empty())
1232 Changed |= processLoop(Worklist.pop_back_val());
1234 // Process each loop nest in the function.
1238 bool processLoop(Loop *L) {
1239 assert(L->empty() && "Only process inner loops.");
1242 const std::string DebugLocStr = getDebugLocString(L);
1245 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1246 << L->getHeader()->getParent()->getName() << "\" from "
1247 << DebugLocStr << "\n");
1249 LoopVectorizeHints Hints(L, DisableUnrolling);
1251 DEBUG(dbgs() << "LV: Loop hints:"
1253 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1255 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1257 : "?")) << " width=" << Hints.getWidth()
1258 << " unroll=" << Hints.getUnroll() << "\n");
1260 // Function containing loop
1261 Function *F = L->getHeader()->getParent();
1263 // Looking at the diagnostic output is the only way to determine if a loop
1264 // was vectorized (other than looking at the IR or machine code), so it
1265 // is important to generate an optimization remark for each loop. Most of
1266 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1267 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1268 // less verbose reporting vectorized loops and unvectorized loops that may
1269 // benefit from vectorization, respectively.
1271 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1272 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1273 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1274 L->getStartLoc(), Hints.emitRemark());
1278 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1279 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1280 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1281 L->getStartLoc(), Hints.emitRemark());
1285 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1286 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1287 emitOptimizationRemarkAnalysis(
1288 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1289 "loop not vectorized: vector width and interleave count are "
1290 "explicitly set to 1");
1294 // Check the loop for a trip count threshold:
1295 // do not vectorize loops with a tiny trip count.
1296 BasicBlock *Latch = L->getLoopLatch();
1297 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1298 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1299 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1300 << "This loop is not worth vectorizing.");
1301 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1302 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1304 DEBUG(dbgs() << "\n");
1305 emitOptimizationRemarkAnalysis(
1306 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1307 "vectorization is not beneficial and is not explicitly forced");
1312 // Check if it is legal to vectorize the loop.
1313 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1314 if (!LVL.canVectorize()) {
1315 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1316 emitMissedWarning(F, L, Hints);
1320 // Use the cost model.
1321 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, F, &Hints);
1323 // Check the function attributes to find out if this function should be
1324 // optimized for size.
1325 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1326 F->hasFnAttribute(Attribute::OptimizeForSize);
1328 // Compute the weighted frequency of this loop being executed and see if it
1329 // is less than 20% of the function entry baseline frequency. Note that we
1330 // always have a canonical loop here because we think we *can* vectoriez.
1331 // FIXME: This is hidden behind a flag due to pervasive problems with
1332 // exactly what block frequency models.
1333 if (LoopVectorizeWithBlockFrequency) {
1334 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1335 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1336 LoopEntryFreq < ColdEntryFreq)
1340 // Check the function attributes to see if implicit floats are allowed.a
1341 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1342 // an integer loop and the vector instructions selected are purely integer
1343 // vector instructions?
1344 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1345 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1346 "attribute is used.\n");
1347 emitOptimizationRemarkAnalysis(
1348 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1349 "loop not vectorized due to NoImplicitFloat attribute");
1350 emitMissedWarning(F, L, Hints);
1354 // Select the optimal vectorization factor.
1355 const LoopVectorizationCostModel::VectorizationFactor VF =
1356 CM.selectVectorizationFactor(OptForSize);
1358 // Select the unroll factor.
1360 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1362 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1363 << DebugLocStr << '\n');
1364 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1366 if (VF.Width == 1) {
1367 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1370 emitOptimizationRemarkAnalysis(
1371 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1372 "not beneficial to vectorize and user disabled interleaving");
1375 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1377 // Report the unrolling decision.
1378 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1379 Twine("unrolled with interleaving factor " +
1381 " (vectorization not beneficial)"));
1383 // We decided not to vectorize, but we may want to unroll.
1385 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1386 Unroller.vectorize(&LVL);
1388 // If we decided that it is *legal* to vectorize the loop then do it.
1389 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1393 // Report the vectorization decision.
1394 emitOptimizationRemark(
1395 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1396 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1397 ", unrolling interleave factor: " + Twine(UF) + ")");
1400 // Mark the loop as already vectorized to avoid vectorizing again.
1401 Hints.setAlreadyVectorized(L);
1403 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1407 void getAnalysisUsage(AnalysisUsage &AU) const override {
1408 AU.addRequiredID(LoopSimplifyID);
1409 AU.addRequiredID(LCSSAID);
1410 AU.addRequired<BlockFrequencyInfo>();
1411 AU.addRequired<DominatorTreeWrapperPass>();
1412 AU.addRequired<LoopInfo>();
1413 AU.addRequired<ScalarEvolution>();
1414 AU.addRequired<TargetTransformInfo>();
1415 AU.addRequired<AliasAnalysis>();
1416 AU.addPreserved<LoopInfo>();
1417 AU.addPreserved<DominatorTreeWrapperPass>();
1418 AU.addPreserved<AliasAnalysis>();
1423 } // end anonymous namespace
1425 //===----------------------------------------------------------------------===//
1426 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1427 // LoopVectorizationCostModel.
1428 //===----------------------------------------------------------------------===//
1430 static Value *stripIntegerCast(Value *V) {
1431 if (CastInst *CI = dyn_cast<CastInst>(V))
1432 if (CI->getOperand(0)->getType()->isIntegerTy())
1433 return CI->getOperand(0);
1437 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1439 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1441 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1442 ValueToValueMap &PtrToStride,
1443 Value *Ptr, Value *OrigPtr = nullptr) {
1445 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1447 // If there is an entry in the map return the SCEV of the pointer with the
1448 // symbolic stride replaced by one.
1449 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1450 if (SI != PtrToStride.end()) {
1451 Value *StrideVal = SI->second;
1454 StrideVal = stripIntegerCast(StrideVal);
1456 // Replace symbolic stride by one.
1457 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1458 ValueToValueMap RewriteMap;
1459 RewriteMap[StrideVal] = One;
1462 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1463 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1468 // Otherwise, just return the SCEV of the original pointer.
1469 return SE->getSCEV(Ptr);
1472 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1473 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1474 unsigned ASId, ValueToValueMap &Strides) {
1475 // Get the stride replaced scev.
1476 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1477 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1478 assert(AR && "Invalid addrec expression");
1479 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1480 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1481 Pointers.push_back(Ptr);
1482 Starts.push_back(AR->getStart());
1483 Ends.push_back(ScEnd);
1484 IsWritePtr.push_back(WritePtr);
1485 DependencySetId.push_back(DepSetId);
1486 AliasSetId.push_back(ASId);
1489 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1490 // We need to place the broadcast of invariant variables outside the loop.
1491 Instruction *Instr = dyn_cast<Instruction>(V);
1493 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1494 Instr->getParent()) != LoopVectorBody.end());
1495 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1497 // Place the code for broadcasting invariant variables in the new preheader.
1498 IRBuilder<>::InsertPointGuard Guard(Builder);
1500 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1502 // Broadcast the scalar into all locations in the vector.
1503 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1508 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1510 assert(Val->getType()->isVectorTy() && "Must be a vector");
1511 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1512 "Elem must be an integer");
1513 // Create the types.
1514 Type *ITy = Val->getType()->getScalarType();
1515 VectorType *Ty = cast<VectorType>(Val->getType());
1516 int VLen = Ty->getNumElements();
1517 SmallVector<Constant*, 8> Indices;
1519 // Create a vector of consecutive numbers from zero to VF.
1520 for (int i = 0; i < VLen; ++i) {
1521 int64_t Idx = Negate ? (-i) : i;
1522 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1525 // Add the consecutive indices to the vector value.
1526 Constant *Cv = ConstantVector::get(Indices);
1527 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1528 return Builder.CreateAdd(Val, Cv, "induction");
1531 /// \brief Find the operand of the GEP that should be checked for consecutive
1532 /// stores. This ignores trailing indices that have no effect on the final
1534 static unsigned getGEPInductionOperand(const DataLayout *DL,
1535 const GetElementPtrInst *Gep) {
1536 unsigned LastOperand = Gep->getNumOperands() - 1;
1537 unsigned GEPAllocSize = DL->getTypeAllocSize(
1538 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1540 // Walk backwards and try to peel off zeros.
1541 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1542 // Find the type we're currently indexing into.
1543 gep_type_iterator GEPTI = gep_type_begin(Gep);
1544 std::advance(GEPTI, LastOperand - 1);
1546 // If it's a type with the same allocation size as the result of the GEP we
1547 // can peel off the zero index.
1548 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1556 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1557 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1558 // Make sure that the pointer does not point to structs.
1559 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1562 // If this value is a pointer induction variable we know it is consecutive.
1563 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1564 if (Phi && Inductions.count(Phi)) {
1565 InductionInfo II = Inductions[Phi];
1566 if (IK_PtrInduction == II.IK)
1568 else if (IK_ReversePtrInduction == II.IK)
1572 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1576 unsigned NumOperands = Gep->getNumOperands();
1577 Value *GpPtr = Gep->getPointerOperand();
1578 // If this GEP value is a consecutive pointer induction variable and all of
1579 // the indices are constant then we know it is consecutive. We can
1580 Phi = dyn_cast<PHINode>(GpPtr);
1581 if (Phi && Inductions.count(Phi)) {
1583 // Make sure that the pointer does not point to structs.
1584 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1585 if (GepPtrType->getElementType()->isAggregateType())
1588 // Make sure that all of the index operands are loop invariant.
1589 for (unsigned i = 1; i < NumOperands; ++i)
1590 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1593 InductionInfo II = Inductions[Phi];
1594 if (IK_PtrInduction == II.IK)
1596 else if (IK_ReversePtrInduction == II.IK)
1600 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1602 // Check that all of the gep indices are uniform except for our induction
1604 for (unsigned i = 0; i != NumOperands; ++i)
1605 if (i != InductionOperand &&
1606 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1609 // We can emit wide load/stores only if the last non-zero index is the
1610 // induction variable.
1611 const SCEV *Last = nullptr;
1612 if (!Strides.count(Gep))
1613 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1615 // Because of the multiplication by a stride we can have a s/zext cast.
1616 // We are going to replace this stride by 1 so the cast is safe to ignore.
1618 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1619 // %0 = trunc i64 %indvars.iv to i32
1620 // %mul = mul i32 %0, %Stride1
1621 // %idxprom = zext i32 %mul to i64 << Safe cast.
1622 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1624 Last = replaceSymbolicStrideSCEV(SE, Strides,
1625 Gep->getOperand(InductionOperand), Gep);
1626 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1628 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1632 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1633 const SCEV *Step = AR->getStepRecurrence(*SE);
1635 // The memory is consecutive because the last index is consecutive
1636 // and all other indices are loop invariant.
1639 if (Step->isAllOnesValue())
1646 bool LoopVectorizationLegality::isUniform(Value *V) {
1647 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1650 InnerLoopVectorizer::VectorParts&
1651 InnerLoopVectorizer::getVectorValue(Value *V) {
1652 assert(V != Induction && "The new induction variable should not be used.");
1653 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1655 // If we have a stride that is replaced by one, do it here.
1656 if (Legal->hasStride(V))
1657 V = ConstantInt::get(V->getType(), 1);
1659 // If we have this scalar in the map, return it.
1660 if (WidenMap.has(V))
1661 return WidenMap.get(V);
1663 // If this scalar is unknown, assume that it is a constant or that it is
1664 // loop invariant. Broadcast V and save the value for future uses.
1665 Value *B = getBroadcastInstrs(V);
1666 return WidenMap.splat(V, B);
1669 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1670 assert(Vec->getType()->isVectorTy() && "Invalid type");
1671 SmallVector<Constant*, 8> ShuffleMask;
1672 for (unsigned i = 0; i < VF; ++i)
1673 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1675 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1676 ConstantVector::get(ShuffleMask),
1680 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1681 // Attempt to issue a wide load.
1682 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1683 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1685 assert((LI || SI) && "Invalid Load/Store instruction");
1687 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1688 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1689 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1690 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1691 // An alignment of 0 means target abi alignment. We need to use the scalar's
1692 // target abi alignment in such a case.
1694 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1695 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1696 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1697 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1699 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1700 return scalarizeInstruction(Instr, true);
1702 if (ScalarAllocatedSize != VectorElementSize)
1703 return scalarizeInstruction(Instr);
1705 // If the pointer is loop invariant or if it is non-consecutive,
1706 // scalarize the load.
1707 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1708 bool Reverse = ConsecutiveStride < 0;
1709 bool UniformLoad = LI && Legal->isUniform(Ptr);
1710 if (!ConsecutiveStride || UniformLoad)
1711 return scalarizeInstruction(Instr);
1713 Constant *Zero = Builder.getInt32(0);
1714 VectorParts &Entry = WidenMap.get(Instr);
1716 // Handle consecutive loads/stores.
1717 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1718 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1719 setDebugLocFromInst(Builder, Gep);
1720 Value *PtrOperand = Gep->getPointerOperand();
1721 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1722 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1724 // Create the new GEP with the new induction variable.
1725 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1726 Gep2->setOperand(0, FirstBasePtr);
1727 Gep2->setName("gep.indvar.base");
1728 Ptr = Builder.Insert(Gep2);
1730 setDebugLocFromInst(Builder, Gep);
1731 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1732 OrigLoop) && "Base ptr must be invariant");
1734 // The last index does not have to be the induction. It can be
1735 // consecutive and be a function of the index. For example A[I+1];
1736 unsigned NumOperands = Gep->getNumOperands();
1737 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1738 // Create the new GEP with the new induction variable.
1739 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1741 for (unsigned i = 0; i < NumOperands; ++i) {
1742 Value *GepOperand = Gep->getOperand(i);
1743 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1745 // Update last index or loop invariant instruction anchored in loop.
1746 if (i == InductionOperand ||
1747 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1748 assert((i == InductionOperand ||
1749 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1750 "Must be last index or loop invariant");
1752 VectorParts &GEPParts = getVectorValue(GepOperand);
1753 Value *Index = GEPParts[0];
1754 Index = Builder.CreateExtractElement(Index, Zero);
1755 Gep2->setOperand(i, Index);
1756 Gep2->setName("gep.indvar.idx");
1759 Ptr = Builder.Insert(Gep2);
1761 // Use the induction element ptr.
1762 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1763 setDebugLocFromInst(Builder, Ptr);
1764 VectorParts &PtrVal = getVectorValue(Ptr);
1765 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1770 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1771 "We do not allow storing to uniform addresses");
1772 setDebugLocFromInst(Builder, SI);
1773 // We don't want to update the value in the map as it might be used in
1774 // another expression. So don't use a reference type for "StoredVal".
1775 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1777 for (unsigned Part = 0; Part < UF; ++Part) {
1778 // Calculate the pointer for the specific unroll-part.
1779 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1782 // If we store to reverse consecutive memory locations then we need
1783 // to reverse the order of elements in the stored value.
1784 StoredVal[Part] = reverseVector(StoredVal[Part]);
1785 // If the address is consecutive but reversed, then the
1786 // wide store needs to start at the last vector element.
1787 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1788 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1791 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1792 DataTy->getPointerTo(AddressSpace));
1794 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1795 propagateMetadata(NewSI, SI);
1801 assert(LI && "Must have a load instruction");
1802 setDebugLocFromInst(Builder, LI);
1803 for (unsigned Part = 0; Part < UF; ++Part) {
1804 // Calculate the pointer for the specific unroll-part.
1805 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1808 // If the address is consecutive but reversed, then the
1809 // wide store needs to start at the last vector element.
1810 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1811 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1814 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1815 DataTy->getPointerTo(AddressSpace));
1816 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1817 propagateMetadata(NewLI, LI);
1818 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1822 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1823 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1824 // Holds vector parameters or scalars, in case of uniform vals.
1825 SmallVector<VectorParts, 4> Params;
1827 setDebugLocFromInst(Builder, Instr);
1829 // Find all of the vectorized parameters.
1830 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1831 Value *SrcOp = Instr->getOperand(op);
1833 // If we are accessing the old induction variable, use the new one.
1834 if (SrcOp == OldInduction) {
1835 Params.push_back(getVectorValue(SrcOp));
1839 // Try using previously calculated values.
1840 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1842 // If the src is an instruction that appeared earlier in the basic block
1843 // then it should already be vectorized.
1844 if (SrcInst && OrigLoop->contains(SrcInst)) {
1845 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1846 // The parameter is a vector value from earlier.
1847 Params.push_back(WidenMap.get(SrcInst));
1849 // The parameter is a scalar from outside the loop. Maybe even a constant.
1850 VectorParts Scalars;
1851 Scalars.append(UF, SrcOp);
1852 Params.push_back(Scalars);
1856 assert(Params.size() == Instr->getNumOperands() &&
1857 "Invalid number of operands");
1859 // Does this instruction return a value ?
1860 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1862 Value *UndefVec = IsVoidRetTy ? nullptr :
1863 UndefValue::get(VectorType::get(Instr->getType(), VF));
1864 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1865 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1867 Instruction *InsertPt = Builder.GetInsertPoint();
1868 BasicBlock *IfBlock = Builder.GetInsertBlock();
1869 BasicBlock *CondBlock = nullptr;
1872 Loop *VectorLp = nullptr;
1873 if (IfPredicateStore) {
1874 assert(Instr->getParent()->getSinglePredecessor() &&
1875 "Only support single predecessor blocks");
1876 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1877 Instr->getParent());
1878 VectorLp = LI->getLoopFor(IfBlock);
1879 assert(VectorLp && "Must have a loop for this block");
1882 // For each vector unroll 'part':
1883 for (unsigned Part = 0; Part < UF; ++Part) {
1884 // For each scalar that we create:
1885 for (unsigned Width = 0; Width < VF; ++Width) {
1888 Value *Cmp = nullptr;
1889 if (IfPredicateStore) {
1890 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1891 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1892 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1893 LoopVectorBody.push_back(CondBlock);
1894 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1895 // Update Builder with newly created basic block.
1896 Builder.SetInsertPoint(InsertPt);
1899 Instruction *Cloned = Instr->clone();
1901 Cloned->setName(Instr->getName() + ".cloned");
1902 // Replace the operands of the cloned instructions with extracted scalars.
1903 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1904 Value *Op = Params[op][Part];
1905 // Param is a vector. Need to extract the right lane.
1906 if (Op->getType()->isVectorTy())
1907 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1908 Cloned->setOperand(op, Op);
1911 // Place the cloned scalar in the new loop.
1912 Builder.Insert(Cloned);
1914 // If the original scalar returns a value we need to place it in a vector
1915 // so that future users will be able to use it.
1917 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1918 Builder.getInt32(Width));
1920 if (IfPredicateStore) {
1921 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1922 LoopVectorBody.push_back(NewIfBlock);
1923 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1924 Builder.SetInsertPoint(InsertPt);
1925 Instruction *OldBr = IfBlock->getTerminator();
1926 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1927 OldBr->eraseFromParent();
1928 IfBlock = NewIfBlock;
1934 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1938 if (Instruction *I = dyn_cast<Instruction>(V))
1939 return I->getParent() == Loc->getParent() ? I : nullptr;
1943 std::pair<Instruction *, Instruction *>
1944 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1945 Instruction *tnullptr = nullptr;
1946 if (!Legal->mustCheckStrides())
1947 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1949 IRBuilder<> ChkBuilder(Loc);
1952 Value *Check = nullptr;
1953 Instruction *FirstInst = nullptr;
1954 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1955 SE = Legal->strides_end();
1957 Value *Ptr = stripIntegerCast(*SI);
1958 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1960 // Store the first instruction we create.
1961 FirstInst = getFirstInst(FirstInst, C, Loc);
1963 Check = ChkBuilder.CreateOr(Check, C);
1968 // We have to do this trickery because the IRBuilder might fold the check to a
1969 // constant expression in which case there is no Instruction anchored in a
1971 LLVMContext &Ctx = Loc->getContext();
1972 Instruction *TheCheck =
1973 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1974 ChkBuilder.Insert(TheCheck, "stride.not.one");
1975 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1977 return std::make_pair(FirstInst, TheCheck);
1980 std::pair<Instruction *, Instruction *>
1981 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1982 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1983 Legal->getRuntimePointerCheck();
1985 Instruction *tnullptr = nullptr;
1986 if (!PtrRtCheck->Need)
1987 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1989 unsigned NumPointers = PtrRtCheck->Pointers.size();
1990 SmallVector<TrackingVH<Value> , 2> Starts;
1991 SmallVector<TrackingVH<Value> , 2> Ends;
1993 LLVMContext &Ctx = Loc->getContext();
1994 SCEVExpander Exp(*SE, "induction");
1995 Instruction *FirstInst = nullptr;
1997 for (unsigned i = 0; i < NumPointers; ++i) {
1998 Value *Ptr = PtrRtCheck->Pointers[i];
1999 const SCEV *Sc = SE->getSCEV(Ptr);
2001 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2002 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2004 Starts.push_back(Ptr);
2005 Ends.push_back(Ptr);
2007 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2008 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2010 // Use this type for pointer arithmetic.
2011 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2013 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2014 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2015 Starts.push_back(Start);
2016 Ends.push_back(End);
2020 IRBuilder<> ChkBuilder(Loc);
2021 // Our instructions might fold to a constant.
2022 Value *MemoryRuntimeCheck = nullptr;
2023 for (unsigned i = 0; i < NumPointers; ++i) {
2024 for (unsigned j = i+1; j < NumPointers; ++j) {
2025 // No need to check if two readonly pointers intersect.
2026 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2029 // Only need to check pointers between two different dependency sets.
2030 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2032 // Only need to check pointers in the same alias set.
2033 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2036 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2037 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2039 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2040 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2041 "Trying to bounds check pointers with different address spaces");
2043 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2044 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2046 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2047 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2048 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2049 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2051 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2052 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2053 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2054 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2055 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2056 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2057 if (MemoryRuntimeCheck) {
2058 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2060 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2062 MemoryRuntimeCheck = IsConflict;
2066 // We have to do this trickery because the IRBuilder might fold the check to a
2067 // constant expression in which case there is no Instruction anchored in a
2069 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2070 ConstantInt::getTrue(Ctx));
2071 ChkBuilder.Insert(Check, "memcheck.conflict");
2072 FirstInst = getFirstInst(FirstInst, Check, Loc);
2073 return std::make_pair(FirstInst, Check);
2076 void InnerLoopVectorizer::createEmptyLoop() {
2078 In this function we generate a new loop. The new loop will contain
2079 the vectorized instructions while the old loop will continue to run the
2082 [ ] <-- Back-edge taken count overflow check.
2085 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2088 || [ ] <-- vector pre header.
2092 || [ ]_| <-- vector loop.
2095 | >[ ] <--- middle-block.
2098 -|- >[ ] <--- new preheader.
2102 | [ ]_| <-- old scalar loop to handle remainder.
2105 >[ ] <-- exit block.
2109 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2110 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2111 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2112 assert(BypassBlock && "Invalid loop structure");
2113 assert(ExitBlock && "Must have an exit block");
2115 // Some loops have a single integer induction variable, while other loops
2116 // don't. One example is c++ iterators that often have multiple pointer
2117 // induction variables. In the code below we also support a case where we
2118 // don't have a single induction variable.
2119 OldInduction = Legal->getInduction();
2120 Type *IdxTy = Legal->getWidestInductionType();
2122 // Find the loop boundaries.
2123 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2124 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2126 // The exit count might have the type of i64 while the phi is i32. This can
2127 // happen if we have an induction variable that is sign extended before the
2128 // compare. The only way that we get a backedge taken count is that the
2129 // induction variable was signed and as such will not overflow. In such a case
2130 // truncation is legal.
2131 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2132 IdxTy->getPrimitiveSizeInBits())
2133 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2135 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2136 // Get the total trip count from the count by adding 1.
2137 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2138 SE->getConstant(BackedgeTakeCount->getType(), 1));
2140 // Expand the trip count and place the new instructions in the preheader.
2141 // Notice that the pre-header does not change, only the loop body.
2142 SCEVExpander Exp(*SE, "induction");
2144 // We need to test whether the backedge-taken count is uint##_max. Adding one
2145 // to it will cause overflow and an incorrect loop trip count in the vector
2146 // body. In case of overflow we want to directly jump to the scalar remainder
2148 Value *BackedgeCount =
2149 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2150 BypassBlock->getTerminator());
2151 if (BackedgeCount->getType()->isPointerTy())
2152 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2153 "backedge.ptrcnt.to.int",
2154 BypassBlock->getTerminator());
2155 Instruction *CheckBCOverflow =
2156 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2157 Constant::getAllOnesValue(BackedgeCount->getType()),
2158 "backedge.overflow", BypassBlock->getTerminator());
2160 // The loop index does not have to start at Zero. Find the original start
2161 // value from the induction PHI node. If we don't have an induction variable
2162 // then we know that it starts at zero.
2163 Builder.SetInsertPoint(BypassBlock->getTerminator());
2164 Value *StartIdx = ExtendedIdx = OldInduction ?
2165 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2167 ConstantInt::get(IdxTy, 0);
2169 // We need an instruction to anchor the overflow check on. StartIdx needs to
2170 // be defined before the overflow check branch. Because the scalar preheader
2171 // is going to merge the start index and so the overflow branch block needs to
2172 // contain a definition of the start index.
2173 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2174 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2175 BypassBlock->getTerminator());
2177 // Count holds the overall loop count (N).
2178 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2179 BypassBlock->getTerminator());
2181 LoopBypassBlocks.push_back(BypassBlock);
2183 // Split the single block loop into the two loop structure described above.
2184 BasicBlock *VectorPH =
2185 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2186 BasicBlock *VecBody =
2187 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2188 BasicBlock *MiddleBlock =
2189 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2190 BasicBlock *ScalarPH =
2191 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2193 // Create and register the new vector loop.
2194 Loop* Lp = new Loop();
2195 Loop *ParentLoop = OrigLoop->getParentLoop();
2197 // Insert the new loop into the loop nest and register the new basic blocks
2198 // before calling any utilities such as SCEV that require valid LoopInfo.
2200 ParentLoop->addChildLoop(Lp);
2201 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2202 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2203 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2205 LI->addTopLevelLoop(Lp);
2207 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2209 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2211 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2213 // Generate the induction variable.
2214 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2215 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2216 // The loop step is equal to the vectorization factor (num of SIMD elements)
2217 // times the unroll factor (num of SIMD instructions).
2218 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2220 // This is the IR builder that we use to add all of the logic for bypassing
2221 // the new vector loop.
2222 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2223 setDebugLocFromInst(BypassBuilder,
2224 getDebugLocFromInstOrOperands(OldInduction));
2226 // We may need to extend the index in case there is a type mismatch.
2227 // We know that the count starts at zero and does not overflow.
2228 if (Count->getType() != IdxTy) {
2229 // The exit count can be of pointer type. Convert it to the correct
2231 if (ExitCount->getType()->isPointerTy())
2232 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2234 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2237 // Add the start index to the loop count to get the new end index.
2238 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2240 // Now we need to generate the expression for N - (N % VF), which is
2241 // the part that the vectorized body will execute.
2242 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2243 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2244 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2245 "end.idx.rnd.down");
2247 // Now, compare the new count to zero. If it is zero skip the vector loop and
2248 // jump to the scalar loop.
2250 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2252 BasicBlock *LastBypassBlock = BypassBlock;
2254 // Generate code to check that the loops trip count that we computed by adding
2255 // one to the backedge-taken count will not overflow.
2257 auto PastOverflowCheck =
2258 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2259 BasicBlock *CheckBlock =
2260 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2262 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2263 LoopBypassBlocks.push_back(CheckBlock);
2264 Instruction *OldTerm = LastBypassBlock->getTerminator();
2265 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2266 OldTerm->eraseFromParent();
2267 LastBypassBlock = CheckBlock;
2270 // Generate the code to check that the strides we assumed to be one are really
2271 // one. We want the new basic block to start at the first instruction in a
2272 // sequence of instructions that form a check.
2273 Instruction *StrideCheck;
2274 Instruction *FirstCheckInst;
2275 std::tie(FirstCheckInst, StrideCheck) =
2276 addStrideCheck(LastBypassBlock->getTerminator());
2278 // Create a new block containing the stride check.
2279 BasicBlock *CheckBlock =
2280 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2282 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2283 LoopBypassBlocks.push_back(CheckBlock);
2285 // Replace the branch into the memory check block with a conditional branch
2286 // for the "few elements case".
2287 Instruction *OldTerm = LastBypassBlock->getTerminator();
2288 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2289 OldTerm->eraseFromParent();
2292 LastBypassBlock = CheckBlock;
2295 // Generate the code that checks in runtime if arrays overlap. We put the
2296 // checks into a separate block to make the more common case of few elements
2298 Instruction *MemRuntimeCheck;
2299 std::tie(FirstCheckInst, MemRuntimeCheck) =
2300 addRuntimeCheck(LastBypassBlock->getTerminator());
2301 if (MemRuntimeCheck) {
2302 // Create a new block containing the memory check.
2303 BasicBlock *CheckBlock =
2304 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2306 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2307 LoopBypassBlocks.push_back(CheckBlock);
2309 // Replace the branch into the memory check block with a conditional branch
2310 // for the "few elements case".
2311 Instruction *OldTerm = LastBypassBlock->getTerminator();
2312 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2313 OldTerm->eraseFromParent();
2315 Cmp = MemRuntimeCheck;
2316 LastBypassBlock = CheckBlock;
2319 LastBypassBlock->getTerminator()->eraseFromParent();
2320 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2323 // We are going to resume the execution of the scalar loop.
2324 // Go over all of the induction variables that we found and fix the
2325 // PHIs that are left in the scalar version of the loop.
2326 // The starting values of PHI nodes depend on the counter of the last
2327 // iteration in the vectorized loop.
2328 // If we come from a bypass edge then we need to start from the original
2331 // This variable saves the new starting index for the scalar loop.
2332 PHINode *ResumeIndex = nullptr;
2333 LoopVectorizationLegality::InductionList::iterator I, E;
2334 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2335 // Set builder to point to last bypass block.
2336 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2337 for (I = List->begin(), E = List->end(); I != E; ++I) {
2338 PHINode *OrigPhi = I->first;
2339 LoopVectorizationLegality::InductionInfo II = I->second;
2341 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2342 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2343 MiddleBlock->getTerminator());
2344 // We might have extended the type of the induction variable but we need a
2345 // truncated version for the scalar loop.
2346 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2347 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2348 MiddleBlock->getTerminator()) : nullptr;
2350 // Create phi nodes to merge from the backedge-taken check block.
2351 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2352 ScalarPH->getTerminator());
2353 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2355 PHINode *BCTruncResumeVal = nullptr;
2356 if (OrigPhi == OldInduction) {
2358 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2359 ScalarPH->getTerminator());
2360 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2363 Value *EndValue = nullptr;
2365 case LoopVectorizationLegality::IK_NoInduction:
2366 llvm_unreachable("Unknown induction");
2367 case LoopVectorizationLegality::IK_IntInduction: {
2368 // Handle the integer induction counter.
2369 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2371 // We have the canonical induction variable.
2372 if (OrigPhi == OldInduction) {
2373 // Create a truncated version of the resume value for the scalar loop,
2374 // we might have promoted the type to a larger width.
2376 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2377 // The new PHI merges the original incoming value, in case of a bypass,
2378 // or the value at the end of the vectorized loop.
2379 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2380 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2381 TruncResumeVal->addIncoming(EndValue, VecBody);
2383 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2385 // We know what the end value is.
2386 EndValue = IdxEndRoundDown;
2387 // We also know which PHI node holds it.
2388 ResumeIndex = ResumeVal;
2392 // Not the canonical induction variable - add the vector loop count to the
2394 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2395 II.StartValue->getType(),
2397 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2400 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2401 // Convert the CountRoundDown variable to the PHI size.
2402 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2403 II.StartValue->getType(),
2405 // Handle reverse integer induction counter.
2406 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2409 case LoopVectorizationLegality::IK_PtrInduction: {
2410 // For pointer induction variables, calculate the offset using
2412 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2416 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2417 // The value at the end of the loop for the reverse pointer is calculated
2418 // by creating a GEP with a negative index starting from the start value.
2419 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2420 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2422 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2428 // The new PHI merges the original incoming value, in case of a bypass,
2429 // or the value at the end of the vectorized loop.
2430 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2431 if (OrigPhi == OldInduction)
2432 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2434 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2436 ResumeVal->addIncoming(EndValue, VecBody);
2438 // Fix the scalar body counter (PHI node).
2439 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2441 // The old induction's phi node in the scalar body needs the truncated
2443 if (OrigPhi == OldInduction) {
2444 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2445 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2447 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2448 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2452 // If we are generating a new induction variable then we also need to
2453 // generate the code that calculates the exit value. This value is not
2454 // simply the end of the counter because we may skip the vectorized body
2455 // in case of a runtime check.
2457 assert(!ResumeIndex && "Unexpected resume value found");
2458 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2459 MiddleBlock->getTerminator());
2460 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2461 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2462 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2465 // Make sure that we found the index where scalar loop needs to continue.
2466 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2467 "Invalid resume Index");
2469 // Add a check in the middle block to see if we have completed
2470 // all of the iterations in the first vector loop.
2471 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2472 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2473 ResumeIndex, "cmp.n",
2474 MiddleBlock->getTerminator());
2476 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2477 // Remove the old terminator.
2478 MiddleBlock->getTerminator()->eraseFromParent();
2480 // Create i+1 and fill the PHINode.
2481 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2482 Induction->addIncoming(StartIdx, VectorPH);
2483 Induction->addIncoming(NextIdx, VecBody);
2484 // Create the compare.
2485 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2486 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2488 // Now we have two terminators. Remove the old one from the block.
2489 VecBody->getTerminator()->eraseFromParent();
2491 // Get ready to start creating new instructions into the vectorized body.
2492 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2495 LoopVectorPreHeader = VectorPH;
2496 LoopScalarPreHeader = ScalarPH;
2497 LoopMiddleBlock = MiddleBlock;
2498 LoopExitBlock = ExitBlock;
2499 LoopVectorBody.push_back(VecBody);
2500 LoopScalarBody = OldBasicBlock;
2502 LoopVectorizeHints Hints(Lp, true);
2503 Hints.setAlreadyVectorized(Lp);
2506 /// This function returns the identity element (or neutral element) for
2507 /// the operation K.
2509 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2514 // Adding, Xoring, Oring zero to a number does not change it.
2515 return ConstantInt::get(Tp, 0);
2516 case RK_IntegerMult:
2517 // Multiplying a number by 1 does not change it.
2518 return ConstantInt::get(Tp, 1);
2520 // AND-ing a number with an all-1 value does not change it.
2521 return ConstantInt::get(Tp, -1, true);
2523 // Multiplying a number by 1 does not change it.
2524 return ConstantFP::get(Tp, 1.0L);
2526 // Adding zero to a number does not change it.
2527 return ConstantFP::get(Tp, 0.0L);
2529 llvm_unreachable("Unknown reduction kind");
2533 /// This function translates the reduction kind to an LLVM binary operator.
2535 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2537 case LoopVectorizationLegality::RK_IntegerAdd:
2538 return Instruction::Add;
2539 case LoopVectorizationLegality::RK_IntegerMult:
2540 return Instruction::Mul;
2541 case LoopVectorizationLegality::RK_IntegerOr:
2542 return Instruction::Or;
2543 case LoopVectorizationLegality::RK_IntegerAnd:
2544 return Instruction::And;
2545 case LoopVectorizationLegality::RK_IntegerXor:
2546 return Instruction::Xor;
2547 case LoopVectorizationLegality::RK_FloatMult:
2548 return Instruction::FMul;
2549 case LoopVectorizationLegality::RK_FloatAdd:
2550 return Instruction::FAdd;
2551 case LoopVectorizationLegality::RK_IntegerMinMax:
2552 return Instruction::ICmp;
2553 case LoopVectorizationLegality::RK_FloatMinMax:
2554 return Instruction::FCmp;
2556 llvm_unreachable("Unknown reduction operation");
2560 Value *createMinMaxOp(IRBuilder<> &Builder,
2561 LoopVectorizationLegality::MinMaxReductionKind RK,
2564 CmpInst::Predicate P = CmpInst::ICMP_NE;
2567 llvm_unreachable("Unknown min/max reduction kind");
2568 case LoopVectorizationLegality::MRK_UIntMin:
2569 P = CmpInst::ICMP_ULT;
2571 case LoopVectorizationLegality::MRK_UIntMax:
2572 P = CmpInst::ICMP_UGT;
2574 case LoopVectorizationLegality::MRK_SIntMin:
2575 P = CmpInst::ICMP_SLT;
2577 case LoopVectorizationLegality::MRK_SIntMax:
2578 P = CmpInst::ICMP_SGT;
2580 case LoopVectorizationLegality::MRK_FloatMin:
2581 P = CmpInst::FCMP_OLT;
2583 case LoopVectorizationLegality::MRK_FloatMax:
2584 P = CmpInst::FCMP_OGT;
2589 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2590 RK == LoopVectorizationLegality::MRK_FloatMax)
2591 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2593 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2595 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2600 struct CSEDenseMapInfo {
2601 static bool canHandle(Instruction *I) {
2602 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2603 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2605 static inline Instruction *getEmptyKey() {
2606 return DenseMapInfo<Instruction *>::getEmptyKey();
2608 static inline Instruction *getTombstoneKey() {
2609 return DenseMapInfo<Instruction *>::getTombstoneKey();
2611 static unsigned getHashValue(Instruction *I) {
2612 assert(canHandle(I) && "Unknown instruction!");
2613 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2614 I->value_op_end()));
2616 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2617 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2618 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2620 return LHS->isIdenticalTo(RHS);
2625 /// \brief Check whether this block is a predicated block.
2626 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2627 /// = ...; " blocks. We start with one vectorized basic block. For every
2628 /// conditional block we split this vectorized block. Therefore, every second
2629 /// block will be a predicated one.
2630 static bool isPredicatedBlock(unsigned BlockNum) {
2631 return BlockNum % 2;
2634 ///\brief Perform cse of induction variable instructions.
2635 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2636 // Perform simple cse.
2637 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2638 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2639 BasicBlock *BB = BBs[i];
2640 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2641 Instruction *In = I++;
2643 if (!CSEDenseMapInfo::canHandle(In))
2646 // Check if we can replace this instruction with any of the
2647 // visited instructions.
2648 if (Instruction *V = CSEMap.lookup(In)) {
2649 In->replaceAllUsesWith(V);
2650 In->eraseFromParent();
2653 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2654 // ...;" blocks for predicated stores. Every second block is a predicated
2656 if (isPredicatedBlock(i))
2664 /// \brief Adds a 'fast' flag to floating point operations.
2665 static Value *addFastMathFlag(Value *V) {
2666 if (isa<FPMathOperator>(V)){
2667 FastMathFlags Flags;
2668 Flags.setUnsafeAlgebra();
2669 cast<Instruction>(V)->setFastMathFlags(Flags);
2674 void InnerLoopVectorizer::vectorizeLoop() {
2675 //===------------------------------------------------===//
2677 // Notice: any optimization or new instruction that go
2678 // into the code below should be also be implemented in
2681 //===------------------------------------------------===//
2682 Constant *Zero = Builder.getInt32(0);
2684 // In order to support reduction variables we need to be able to vectorize
2685 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2686 // stages. First, we create a new vector PHI node with no incoming edges.
2687 // We use this value when we vectorize all of the instructions that use the
2688 // PHI. Next, after all of the instructions in the block are complete we
2689 // add the new incoming edges to the PHI. At this point all of the
2690 // instructions in the basic block are vectorized, so we can use them to
2691 // construct the PHI.
2692 PhiVector RdxPHIsToFix;
2694 // Scan the loop in a topological order to ensure that defs are vectorized
2696 LoopBlocksDFS DFS(OrigLoop);
2699 // Vectorize all of the blocks in the original loop.
2700 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2701 be = DFS.endRPO(); bb != be; ++bb)
2702 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2704 // At this point every instruction in the original loop is widened to
2705 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2706 // that we vectorized. The PHI nodes are currently empty because we did
2707 // not want to introduce cycles. Notice that the remaining PHI nodes
2708 // that we need to fix are reduction variables.
2710 // Create the 'reduced' values for each of the induction vars.
2711 // The reduced values are the vector values that we scalarize and combine
2712 // after the loop is finished.
2713 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2715 PHINode *RdxPhi = *it;
2716 assert(RdxPhi && "Unable to recover vectorized PHI");
2718 // Find the reduction variable descriptor.
2719 assert(Legal->getReductionVars()->count(RdxPhi) &&
2720 "Unable to find the reduction variable");
2721 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2722 (*Legal->getReductionVars())[RdxPhi];
2724 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2726 // We need to generate a reduction vector from the incoming scalar.
2727 // To do so, we need to generate the 'identity' vector and override
2728 // one of the elements with the incoming scalar reduction. We need
2729 // to do it in the vector-loop preheader.
2730 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2732 // This is the vector-clone of the value that leaves the loop.
2733 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2734 Type *VecTy = VectorExit[0]->getType();
2736 // Find the reduction identity variable. Zero for addition, or, xor,
2737 // one for multiplication, -1 for And.
2740 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2741 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2742 // MinMax reduction have the start value as their identify.
2744 VectorStart = Identity = RdxDesc.StartValue;
2746 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2751 // Handle other reduction kinds:
2753 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2754 VecTy->getScalarType());
2757 // This vector is the Identity vector where the first element is the
2758 // incoming scalar reduction.
2759 VectorStart = RdxDesc.StartValue;
2761 Identity = ConstantVector::getSplat(VF, Iden);
2763 // This vector is the Identity vector where the first element is the
2764 // incoming scalar reduction.
2765 VectorStart = Builder.CreateInsertElement(Identity,
2766 RdxDesc.StartValue, Zero);
2770 // Fix the vector-loop phi.
2771 // We created the induction variable so we know that the
2772 // preheader is the first entry.
2773 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2775 // Reductions do not have to start at zero. They can start with
2776 // any loop invariant values.
2777 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2778 BasicBlock *Latch = OrigLoop->getLoopLatch();
2779 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2780 VectorParts &Val = getVectorValue(LoopVal);
2781 for (unsigned part = 0; part < UF; ++part) {
2782 // Make sure to add the reduction stat value only to the
2783 // first unroll part.
2784 Value *StartVal = (part == 0) ? VectorStart : Identity;
2785 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2786 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2787 LoopVectorBody.back());
2790 // Before each round, move the insertion point right between
2791 // the PHIs and the values we are going to write.
2792 // This allows us to write both PHINodes and the extractelement
2794 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2796 VectorParts RdxParts;
2797 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2798 for (unsigned part = 0; part < UF; ++part) {
2799 // This PHINode contains the vectorized reduction variable, or
2800 // the initial value vector, if we bypass the vector loop.
2801 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2802 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2803 Value *StartVal = (part == 0) ? VectorStart : Identity;
2804 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2805 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2806 NewPhi->addIncoming(RdxExitVal[part],
2807 LoopVectorBody.back());
2808 RdxParts.push_back(NewPhi);
2811 // Reduce all of the unrolled parts into a single vector.
2812 Value *ReducedPartRdx = RdxParts[0];
2813 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2814 setDebugLocFromInst(Builder, ReducedPartRdx);
2815 for (unsigned part = 1; part < UF; ++part) {
2816 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2817 // Floating point operations had to be 'fast' to enable the reduction.
2818 ReducedPartRdx = addFastMathFlag(
2819 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2820 ReducedPartRdx, "bin.rdx"));
2822 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2823 ReducedPartRdx, RdxParts[part]);
2827 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2828 // and vector ops, reducing the set of values being computed by half each
2830 assert(isPowerOf2_32(VF) &&
2831 "Reduction emission only supported for pow2 vectors!");
2832 Value *TmpVec = ReducedPartRdx;
2833 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2834 for (unsigned i = VF; i != 1; i >>= 1) {
2835 // Move the upper half of the vector to the lower half.
2836 for (unsigned j = 0; j != i/2; ++j)
2837 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2839 // Fill the rest of the mask with undef.
2840 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2841 UndefValue::get(Builder.getInt32Ty()));
2844 Builder.CreateShuffleVector(TmpVec,
2845 UndefValue::get(TmpVec->getType()),
2846 ConstantVector::get(ShuffleMask),
2849 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2850 // Floating point operations had to be 'fast' to enable the reduction.
2851 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2852 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2854 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2857 // The result is in the first element of the vector.
2858 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2859 Builder.getInt32(0));
2862 // Create a phi node that merges control-flow from the backedge-taken check
2863 // block and the middle block.
2864 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2865 LoopScalarPreHeader->getTerminator());
2866 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2867 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2869 // Now, we need to fix the users of the reduction variable
2870 // inside and outside of the scalar remainder loop.
2871 // We know that the loop is in LCSSA form. We need to update the
2872 // PHI nodes in the exit blocks.
2873 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2874 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2875 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2876 if (!LCSSAPhi) break;
2878 // All PHINodes need to have a single entry edge, or two if
2879 // we already fixed them.
2880 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2882 // We found our reduction value exit-PHI. Update it with the
2883 // incoming bypass edge.
2884 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2885 // Add an edge coming from the bypass.
2886 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2889 }// end of the LCSSA phi scan.
2891 // Fix the scalar loop reduction variable with the incoming reduction sum
2892 // from the vector body and from the backedge value.
2893 int IncomingEdgeBlockIdx =
2894 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2895 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2896 // Pick the other block.
2897 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2898 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2899 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2900 }// end of for each redux variable.
2904 // Remove redundant induction instructions.
2905 cse(LoopVectorBody);
2908 void InnerLoopVectorizer::fixLCSSAPHIs() {
2909 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2910 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2911 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2912 if (!LCSSAPhi) break;
2913 if (LCSSAPhi->getNumIncomingValues() == 1)
2914 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2919 InnerLoopVectorizer::VectorParts
2920 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2921 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2924 // Look for cached value.
2925 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2926 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2927 if (ECEntryIt != MaskCache.end())
2928 return ECEntryIt->second;
2930 VectorParts SrcMask = createBlockInMask(Src);
2932 // The terminator has to be a branch inst!
2933 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2934 assert(BI && "Unexpected terminator found");
2936 if (BI->isConditional()) {
2937 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2939 if (BI->getSuccessor(0) != Dst)
2940 for (unsigned part = 0; part < UF; ++part)
2941 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2943 for (unsigned part = 0; part < UF; ++part)
2944 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2946 MaskCache[Edge] = EdgeMask;
2950 MaskCache[Edge] = SrcMask;
2954 InnerLoopVectorizer::VectorParts
2955 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2956 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2958 // Loop incoming mask is all-one.
2959 if (OrigLoop->getHeader() == BB) {
2960 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2961 return getVectorValue(C);
2964 // This is the block mask. We OR all incoming edges, and with zero.
2965 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2966 VectorParts BlockMask = getVectorValue(Zero);
2969 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2970 VectorParts EM = createEdgeMask(*it, BB);
2971 for (unsigned part = 0; part < UF; ++part)
2972 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2978 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2979 InnerLoopVectorizer::VectorParts &Entry,
2980 unsigned UF, unsigned VF, PhiVector *PV) {
2981 PHINode* P = cast<PHINode>(PN);
2982 // Handle reduction variables:
2983 if (Legal->getReductionVars()->count(P)) {
2984 for (unsigned part = 0; part < UF; ++part) {
2985 // This is phase one of vectorizing PHIs.
2986 Type *VecTy = (VF == 1) ? PN->getType() :
2987 VectorType::get(PN->getType(), VF);
2988 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2989 LoopVectorBody.back()-> getFirstInsertionPt());
2995 setDebugLocFromInst(Builder, P);
2996 // Check for PHI nodes that are lowered to vector selects.
2997 if (P->getParent() != OrigLoop->getHeader()) {
2998 // We know that all PHIs in non-header blocks are converted into
2999 // selects, so we don't have to worry about the insertion order and we
3000 // can just use the builder.
3001 // At this point we generate the predication tree. There may be
3002 // duplications since this is a simple recursive scan, but future
3003 // optimizations will clean it up.
3005 unsigned NumIncoming = P->getNumIncomingValues();
3007 // Generate a sequence of selects of the form:
3008 // SELECT(Mask3, In3,
3009 // SELECT(Mask2, In2,
3011 for (unsigned In = 0; In < NumIncoming; In++) {
3012 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3014 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3016 for (unsigned part = 0; part < UF; ++part) {
3017 // We might have single edge PHIs (blocks) - use an identity
3018 // 'select' for the first PHI operand.
3020 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3023 // Select between the current value and the previous incoming edge
3024 // based on the incoming mask.
3025 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3026 Entry[part], "predphi");
3032 // This PHINode must be an induction variable.
3033 // Make sure that we know about it.
3034 assert(Legal->getInductionVars()->count(P) &&
3035 "Not an induction variable");
3037 LoopVectorizationLegality::InductionInfo II =
3038 Legal->getInductionVars()->lookup(P);
3041 case LoopVectorizationLegality::IK_NoInduction:
3042 llvm_unreachable("Unknown induction");
3043 case LoopVectorizationLegality::IK_IntInduction: {
3044 assert(P->getType() == II.StartValue->getType() && "Types must match");
3045 Type *PhiTy = P->getType();
3047 if (P == OldInduction) {
3048 // Handle the canonical induction variable. We might have had to
3050 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3052 // Handle other induction variables that are now based on the
3054 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3056 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3057 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3060 Broadcasted = getBroadcastInstrs(Broadcasted);
3061 // After broadcasting the induction variable we need to make the vector
3062 // consecutive by adding 0, 1, 2, etc.
3063 for (unsigned part = 0; part < UF; ++part)
3064 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3067 case LoopVectorizationLegality::IK_ReverseIntInduction:
3068 case LoopVectorizationLegality::IK_PtrInduction:
3069 case LoopVectorizationLegality::IK_ReversePtrInduction:
3070 // Handle reverse integer and pointer inductions.
3071 Value *StartIdx = ExtendedIdx;
3072 // This is the normalized GEP that starts counting at zero.
3073 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3076 // Handle the reverse integer induction variable case.
3077 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3078 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3079 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3081 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3084 // This is a new value so do not hoist it out.
3085 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3086 // After broadcasting the induction variable we need to make the
3087 // vector consecutive by adding ... -3, -2, -1, 0.
3088 for (unsigned part = 0; part < UF; ++part)
3089 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3094 // Handle the pointer induction variable case.
3095 assert(P->getType()->isPointerTy() && "Unexpected type.");
3097 // Is this a reverse induction ptr or a consecutive induction ptr.
3098 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3101 // This is the vector of results. Notice that we don't generate
3102 // vector geps because scalar geps result in better code.
3103 for (unsigned part = 0; part < UF; ++part) {
3105 int EltIndex = (part) * (Reverse ? -1 : 1);
3106 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3109 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3111 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3113 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3115 Entry[part] = SclrGep;
3119 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3120 for (unsigned int i = 0; i < VF; ++i) {
3121 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3122 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3125 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3127 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3129 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3131 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3132 Builder.getInt32(i),
3135 Entry[part] = VecVal;
3141 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3142 // For each instruction in the old loop.
3143 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3144 VectorParts &Entry = WidenMap.get(it);
3145 switch (it->getOpcode()) {
3146 case Instruction::Br:
3147 // Nothing to do for PHIs and BR, since we already took care of the
3148 // loop control flow instructions.
3150 case Instruction::PHI:{
3151 // Vectorize PHINodes.
3152 widenPHIInstruction(it, Entry, UF, VF, PV);
3156 case Instruction::Add:
3157 case Instruction::FAdd:
3158 case Instruction::Sub:
3159 case Instruction::FSub:
3160 case Instruction::Mul:
3161 case Instruction::FMul:
3162 case Instruction::UDiv:
3163 case Instruction::SDiv:
3164 case Instruction::FDiv:
3165 case Instruction::URem:
3166 case Instruction::SRem:
3167 case Instruction::FRem:
3168 case Instruction::Shl:
3169 case Instruction::LShr:
3170 case Instruction::AShr:
3171 case Instruction::And:
3172 case Instruction::Or:
3173 case Instruction::Xor: {
3174 // Just widen binops.
3175 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3176 setDebugLocFromInst(Builder, BinOp);
3177 VectorParts &A = getVectorValue(it->getOperand(0));
3178 VectorParts &B = getVectorValue(it->getOperand(1));
3180 // Use this vector value for all users of the original instruction.
3181 for (unsigned Part = 0; Part < UF; ++Part) {
3182 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3184 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3185 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3186 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3187 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3188 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3190 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3191 VecOp->setIsExact(BinOp->isExact());
3193 // Copy the fast-math flags.
3194 if (VecOp && isa<FPMathOperator>(V))
3195 VecOp->setFastMathFlags(it->getFastMathFlags());
3200 propagateMetadata(Entry, it);
3203 case Instruction::Select: {
3205 // If the selector is loop invariant we can create a select
3206 // instruction with a scalar condition. Otherwise, use vector-select.
3207 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3209 setDebugLocFromInst(Builder, it);
3211 // The condition can be loop invariant but still defined inside the
3212 // loop. This means that we can't just use the original 'cond' value.
3213 // We have to take the 'vectorized' value and pick the first lane.
3214 // Instcombine will make this a no-op.
3215 VectorParts &Cond = getVectorValue(it->getOperand(0));
3216 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3217 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3219 Value *ScalarCond = (VF == 1) ? Cond[0] :
3220 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3222 for (unsigned Part = 0; Part < UF; ++Part) {
3223 Entry[Part] = Builder.CreateSelect(
3224 InvariantCond ? ScalarCond : Cond[Part],
3229 propagateMetadata(Entry, it);
3233 case Instruction::ICmp:
3234 case Instruction::FCmp: {
3235 // Widen compares. Generate vector compares.
3236 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3237 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3238 setDebugLocFromInst(Builder, it);
3239 VectorParts &A = getVectorValue(it->getOperand(0));
3240 VectorParts &B = getVectorValue(it->getOperand(1));
3241 for (unsigned Part = 0; Part < UF; ++Part) {
3244 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3246 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3250 propagateMetadata(Entry, it);
3254 case Instruction::Store:
3255 case Instruction::Load:
3256 vectorizeMemoryInstruction(it);
3258 case Instruction::ZExt:
3259 case Instruction::SExt:
3260 case Instruction::FPToUI:
3261 case Instruction::FPToSI:
3262 case Instruction::FPExt:
3263 case Instruction::PtrToInt:
3264 case Instruction::IntToPtr:
3265 case Instruction::SIToFP:
3266 case Instruction::UIToFP:
3267 case Instruction::Trunc:
3268 case Instruction::FPTrunc:
3269 case Instruction::BitCast: {
3270 CastInst *CI = dyn_cast<CastInst>(it);
3271 setDebugLocFromInst(Builder, it);
3272 /// Optimize the special case where the source is the induction
3273 /// variable. Notice that we can only optimize the 'trunc' case
3274 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3275 /// c. other casts depend on pointer size.
3276 if (CI->getOperand(0) == OldInduction &&
3277 it->getOpcode() == Instruction::Trunc) {
3278 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3280 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3281 for (unsigned Part = 0; Part < UF; ++Part)
3282 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3283 propagateMetadata(Entry, it);
3286 /// Vectorize casts.
3287 Type *DestTy = (VF == 1) ? CI->getType() :
3288 VectorType::get(CI->getType(), VF);
3290 VectorParts &A = getVectorValue(it->getOperand(0));
3291 for (unsigned Part = 0; Part < UF; ++Part)
3292 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3293 propagateMetadata(Entry, it);
3297 case Instruction::Call: {
3298 // Ignore dbg intrinsics.
3299 if (isa<DbgInfoIntrinsic>(it))
3301 setDebugLocFromInst(Builder, it);
3303 Module *M = BB->getParent()->getParent();
3304 CallInst *CI = cast<CallInst>(it);
3305 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3306 assert(ID && "Not an intrinsic call!");
3308 case Intrinsic::lifetime_end:
3309 case Intrinsic::lifetime_start:
3310 scalarizeInstruction(it);
3313 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3314 for (unsigned Part = 0; Part < UF; ++Part) {
3315 SmallVector<Value *, 4> Args;
3316 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3317 if (HasScalarOpd && i == 1) {
3318 Args.push_back(CI->getArgOperand(i));
3321 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3322 Args.push_back(Arg[Part]);
3324 Type *Tys[] = {CI->getType()};
3326 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3328 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3329 Entry[Part] = Builder.CreateCall(F, Args);
3332 propagateMetadata(Entry, it);
3339 // All other instructions are unsupported. Scalarize them.
3340 scalarizeInstruction(it);
3343 }// end of for_each instr.
3346 void InnerLoopVectorizer::updateAnalysis() {
3347 // Forget the original basic block.
3348 SE->forgetLoop(OrigLoop);
3350 // Update the dominator tree information.
3351 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3352 "Entry does not dominate exit.");
3354 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3355 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3356 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3358 // Due to if predication of stores we might create a sequence of "if(pred)
3359 // a[i] = ...; " blocks.
3360 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3362 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3363 else if (isPredicatedBlock(i)) {
3364 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3366 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3370 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3371 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3372 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3373 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3375 DEBUG(DT->verifyDomTree());
3378 /// \brief Check whether it is safe to if-convert this phi node.
3380 /// Phi nodes with constant expressions that can trap are not safe to if
3382 static bool canIfConvertPHINodes(BasicBlock *BB) {
3383 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3384 PHINode *Phi = dyn_cast<PHINode>(I);
3387 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3388 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3395 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3396 if (!EnableIfConversion) {
3397 emitAnalysis(Report() << "if-conversion is disabled");
3401 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3403 // A list of pointers that we can safely read and write to.
3404 SmallPtrSet<Value *, 8> SafePointes;
3406 // Collect safe addresses.
3407 for (Loop::block_iterator BI = TheLoop->block_begin(),
3408 BE = TheLoop->block_end(); BI != BE; ++BI) {
3409 BasicBlock *BB = *BI;
3411 if (blockNeedsPredication(BB))
3414 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3415 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3416 SafePointes.insert(LI->getPointerOperand());
3417 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3418 SafePointes.insert(SI->getPointerOperand());
3422 // Collect the blocks that need predication.
3423 BasicBlock *Header = TheLoop->getHeader();
3424 for (Loop::block_iterator BI = TheLoop->block_begin(),
3425 BE = TheLoop->block_end(); BI != BE; ++BI) {
3426 BasicBlock *BB = *BI;
3428 // We don't support switch statements inside loops.
3429 if (!isa<BranchInst>(BB->getTerminator())) {
3430 emitAnalysis(Report(BB->getTerminator())
3431 << "loop contains a switch statement");
3435 // We must be able to predicate all blocks that need to be predicated.
3436 if (blockNeedsPredication(BB)) {
3437 if (!blockCanBePredicated(BB, SafePointes)) {
3438 emitAnalysis(Report(BB->getTerminator())
3439 << "control flow cannot be substituted for a select");
3442 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3443 emitAnalysis(Report(BB->getTerminator())
3444 << "control flow cannot be substituted for a select");
3449 // We can if-convert this loop.
3453 bool LoopVectorizationLegality::canVectorize() {
3454 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3455 // be canonicalized.
3456 if (!TheLoop->getLoopPreheader()) {
3458 Report() << "loop control flow is not understood by vectorizer");
3462 // We can only vectorize innermost loops.
3463 if (TheLoop->getSubLoopsVector().size()) {
3464 emitAnalysis(Report() << "loop is not the innermost loop");
3468 // We must have a single backedge.
3469 if (TheLoop->getNumBackEdges() != 1) {
3471 Report() << "loop control flow is not understood by vectorizer");
3475 // We must have a single exiting block.
3476 if (!TheLoop->getExitingBlock()) {
3478 Report() << "loop control flow is not understood by vectorizer");
3482 // We need to have a loop header.
3483 DEBUG(dbgs() << "LV: Found a loop: " <<
3484 TheLoop->getHeader()->getName() << '\n');
3486 // Check if we can if-convert non-single-bb loops.
3487 unsigned NumBlocks = TheLoop->getNumBlocks();
3488 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3489 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3493 // ScalarEvolution needs to be able to find the exit count.
3494 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3495 if (ExitCount == SE->getCouldNotCompute()) {
3496 emitAnalysis(Report() << "could not determine number of loop iterations");
3497 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3501 // Check if we can vectorize the instructions and CFG in this loop.
3502 if (!canVectorizeInstrs()) {
3503 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3507 // Go over each instruction and look at memory deps.
3508 if (!canVectorizeMemory()) {
3509 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3513 // Collect all of the variables that remain uniform after vectorization.
3514 collectLoopUniforms();
3516 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3517 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3520 // Okay! We can vectorize. At this point we don't have any other mem analysis
3521 // which may limit our maximum vectorization factor, so just return true with
3526 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3527 if (Ty->isPointerTy())
3528 return DL.getIntPtrType(Ty);
3530 // It is possible that char's or short's overflow when we ask for the loop's
3531 // trip count, work around this by changing the type size.
3532 if (Ty->getScalarSizeInBits() < 32)
3533 return Type::getInt32Ty(Ty->getContext());
3538 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3539 Ty0 = convertPointerToIntegerType(DL, Ty0);
3540 Ty1 = convertPointerToIntegerType(DL, Ty1);
3541 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3546 /// \brief Check that the instruction has outside loop users and is not an
3547 /// identified reduction variable.
3548 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3549 SmallPtrSetImpl<Value *> &Reductions) {
3550 // Reduction instructions are allowed to have exit users. All other
3551 // instructions must not have external users.
3552 if (!Reductions.count(Inst))
3553 //Check that all of the users of the loop are inside the BB.
3554 for (User *U : Inst->users()) {
3555 Instruction *UI = cast<Instruction>(U);
3556 // This user may be a reduction exit value.
3557 if (!TheLoop->contains(UI)) {
3558 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3565 bool LoopVectorizationLegality::canVectorizeInstrs() {
3566 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3567 BasicBlock *Header = TheLoop->getHeader();
3569 // Look for the attribute signaling the absence of NaNs.
3570 Function &F = *Header->getParent();
3571 if (F.hasFnAttribute("no-nans-fp-math"))
3572 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3573 AttributeSet::FunctionIndex,
3574 "no-nans-fp-math").getValueAsString() == "true";
3576 // For each block in the loop.
3577 for (Loop::block_iterator bb = TheLoop->block_begin(),
3578 be = TheLoop->block_end(); bb != be; ++bb) {
3580 // Scan the instructions in the block and look for hazards.
3581 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3584 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3585 Type *PhiTy = Phi->getType();
3586 // Check that this PHI type is allowed.
3587 if (!PhiTy->isIntegerTy() &&
3588 !PhiTy->isFloatingPointTy() &&
3589 !PhiTy->isPointerTy()) {
3590 emitAnalysis(Report(it)
3591 << "loop control flow is not understood by vectorizer");
3592 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3596 // If this PHINode is not in the header block, then we know that we
3597 // can convert it to select during if-conversion. No need to check if
3598 // the PHIs in this block are induction or reduction variables.
3599 if (*bb != Header) {
3600 // Check that this instruction has no outside users or is an
3601 // identified reduction value with an outside user.
3602 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3604 emitAnalysis(Report(it) << "value could not be identified as "
3605 "an induction or reduction variable");
3609 // We only allow if-converted PHIs with more than two incoming values.
3610 if (Phi->getNumIncomingValues() != 2) {
3611 emitAnalysis(Report(it)
3612 << "control flow not understood by vectorizer");
3613 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3617 // This is the value coming from the preheader.
3618 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3619 // Check if this is an induction variable.
3620 InductionKind IK = isInductionVariable(Phi);
3622 if (IK_NoInduction != IK) {
3623 // Get the widest type.
3625 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3627 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3629 // Int inductions are special because we only allow one IV.
3630 if (IK == IK_IntInduction) {
3631 // Use the phi node with the widest type as induction. Use the last
3632 // one if there are multiple (no good reason for doing this other
3633 // than it is expedient).
3634 if (!Induction || PhiTy == WidestIndTy)
3638 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3639 Inductions[Phi] = InductionInfo(StartValue, IK);
3641 // Until we explicitly handle the case of an induction variable with
3642 // an outside loop user we have to give up vectorizing this loop.
3643 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3644 emitAnalysis(Report(it) << "use of induction value outside of the "
3645 "loop is not handled by vectorizer");
3652 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3653 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3656 if (AddReductionVar(Phi, RK_IntegerMult)) {
3657 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3660 if (AddReductionVar(Phi, RK_IntegerOr)) {
3661 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3664 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3665 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3668 if (AddReductionVar(Phi, RK_IntegerXor)) {
3669 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3672 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3673 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3676 if (AddReductionVar(Phi, RK_FloatMult)) {
3677 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3680 if (AddReductionVar(Phi, RK_FloatAdd)) {
3681 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3684 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3685 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3690 emitAnalysis(Report(it) << "value that could not be identified as "
3691 "reduction is used outside the loop");
3692 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3694 }// end of PHI handling
3696 // We still don't handle functions. However, we can ignore dbg intrinsic
3697 // calls and we do handle certain intrinsic and libm functions.
3698 CallInst *CI = dyn_cast<CallInst>(it);
3699 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3700 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3701 DEBUG(dbgs() << "LV: Found a call site.\n");
3705 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3706 // second argument is the same (i.e. loop invariant)
3708 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3709 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3710 emitAnalysis(Report(it)
3711 << "intrinsic instruction cannot be vectorized");
3712 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3717 // Check that the instruction return type is vectorizable.
3718 // Also, we can't vectorize extractelement instructions.
3719 if ((!VectorType::isValidElementType(it->getType()) &&
3720 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3721 emitAnalysis(Report(it)
3722 << "instruction return type cannot be vectorized");
3723 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3727 // Check that the stored type is vectorizable.
3728 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3729 Type *T = ST->getValueOperand()->getType();
3730 if (!VectorType::isValidElementType(T)) {
3731 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3734 if (EnableMemAccessVersioning)
3735 collectStridedAcccess(ST);
3738 if (EnableMemAccessVersioning)
3739 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3740 collectStridedAcccess(LI);
3742 // Reduction instructions are allowed to have exit users.
3743 // All other instructions must not have external users.
3744 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3745 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3754 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3755 if (Inductions.empty()) {
3756 emitAnalysis(Report()
3757 << "loop induction variable could not be identified");
3765 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3766 /// return the induction operand of the gep pointer.
3767 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3768 const DataLayout *DL, Loop *Lp) {
3769 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3773 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3775 // Check that all of the gep indices are uniform except for our induction
3777 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3778 if (i != InductionOperand &&
3779 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3781 return GEP->getOperand(InductionOperand);
3784 ///\brief Look for a cast use of the passed value.
3785 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3786 Value *UniqueCast = nullptr;
3787 for (User *U : Ptr->users()) {
3788 CastInst *CI = dyn_cast<CastInst>(U);
3789 if (CI && CI->getType() == Ty) {
3799 ///\brief Get the stride of a pointer access in a loop.
3800 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3801 /// pointer to the Value, or null otherwise.
3802 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3803 const DataLayout *DL, Loop *Lp) {
3804 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3805 if (!PtrTy || PtrTy->isAggregateType())
3808 // Try to remove a gep instruction to make the pointer (actually index at this
3809 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3810 // pointer, otherwise, we are analyzing the index.
3811 Value *OrigPtr = Ptr;
3813 // The size of the pointer access.
3814 int64_t PtrAccessSize = 1;
3816 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3817 const SCEV *V = SE->getSCEV(Ptr);
3821 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3822 V = C->getOperand();
3824 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3828 V = S->getStepRecurrence(*SE);
3832 // Strip off the size of access multiplication if we are still analyzing the
3834 if (OrigPtr == Ptr) {
3835 DL->getTypeAllocSize(PtrTy->getElementType());
3836 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3837 if (M->getOperand(0)->getSCEVType() != scConstant)
3840 const APInt &APStepVal =
3841 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3843 // Huge step value - give up.
3844 if (APStepVal.getBitWidth() > 64)
3847 int64_t StepVal = APStepVal.getSExtValue();
3848 if (PtrAccessSize != StepVal)
3850 V = M->getOperand(1);
3855 Type *StripedOffRecurrenceCast = nullptr;
3856 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3857 StripedOffRecurrenceCast = C->getType();
3858 V = C->getOperand();
3861 // Look for the loop invariant symbolic value.
3862 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3866 Value *Stride = U->getValue();
3867 if (!Lp->isLoopInvariant(Stride))
3870 // If we have stripped off the recurrence cast we have to make sure that we
3871 // return the value that is used in this loop so that we can replace it later.
3872 if (StripedOffRecurrenceCast)
3873 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3878 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3879 Value *Ptr = nullptr;
3880 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3881 Ptr = LI->getPointerOperand();
3882 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3883 Ptr = SI->getPointerOperand();
3887 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3891 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3892 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3893 Strides[Ptr] = Stride;
3894 StrideSet.insert(Stride);
3897 void LoopVectorizationLegality::collectLoopUniforms() {
3898 // We now know that the loop is vectorizable!
3899 // Collect variables that will remain uniform after vectorization.
3900 std::vector<Value*> Worklist;
3901 BasicBlock *Latch = TheLoop->getLoopLatch();
3903 // Start with the conditional branch and walk up the block.
3904 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3906 // Also add all consecutive pointer values; these values will be uniform
3907 // after vectorization (and subsequent cleanup) and, until revectorization is
3908 // supported, all dependencies must also be uniform.
3909 for (Loop::block_iterator B = TheLoop->block_begin(),
3910 BE = TheLoop->block_end(); B != BE; ++B)
3911 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3913 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3914 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3916 while (Worklist.size()) {
3917 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3918 Worklist.pop_back();
3920 // Look at instructions inside this loop.
3921 // Stop when reaching PHI nodes.
3922 // TODO: we need to follow values all over the loop, not only in this block.
3923 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3926 // This is a known uniform.
3929 // Insert all operands.
3930 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3935 /// \brief Analyses memory accesses in a loop.
3937 /// Checks whether run time pointer checks are needed and builds sets for data
3938 /// dependence checking.
3939 class AccessAnalysis {
3941 /// \brief Read or write access location.
3942 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3943 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3945 /// \brief Set of potential dependent memory accesses.
3946 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3948 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
3949 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
3951 /// \brief Register a load and whether it is only read from.
3952 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
3953 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3954 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3955 Accesses.insert(MemAccessInfo(Ptr, false));
3957 ReadOnlyPtr.insert(Ptr);
3960 /// \brief Register a store.
3961 void addStore(AliasAnalysis::Location &Loc) {
3962 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3963 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3964 Accesses.insert(MemAccessInfo(Ptr, true));
3967 /// \brief Check whether we can check the pointers at runtime for
3968 /// non-intersection.
3969 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3970 unsigned &NumComparisons, ScalarEvolution *SE,
3971 Loop *TheLoop, ValueToValueMap &Strides,
3972 bool ShouldCheckStride = false);
3974 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3975 /// and builds sets of dependent accesses.
3976 void buildDependenceSets() {
3977 processMemAccesses();
3980 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3982 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3983 void resetDepChecks() { CheckDeps.clear(); }
3985 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3988 typedef SetVector<MemAccessInfo> PtrAccessSet;
3990 /// \brief Go over all memory access and check whether runtime pointer checks
3991 /// are needed /// and build sets of dependency check candidates.
3992 void processMemAccesses();
3994 /// Set of all accesses.
3995 PtrAccessSet Accesses;
3997 /// Set of accesses that need a further dependence check.
3998 MemAccessInfoSet CheckDeps;
4000 /// Set of pointers that are read only.
4001 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4003 const DataLayout *DL;
4005 /// An alias set tracker to partition the access set by underlying object and
4006 //intrinsic property (such as TBAA metadata).
4007 AliasSetTracker AST;
4009 /// Sets of potentially dependent accesses - members of one set share an
4010 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4011 /// dependence check.
4012 DepCandidates &DepCands;
4014 bool IsRTCheckNeeded;
4017 } // end anonymous namespace
4019 /// \brief Check whether a pointer can participate in a runtime bounds check.
4020 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4022 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4023 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4027 return AR->isAffine();
4030 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4031 /// the address space.
4032 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4033 const Loop *Lp, ValueToValueMap &StridesMap);
4035 bool AccessAnalysis::canCheckPtrAtRT(
4036 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4037 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4038 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4039 // Find pointers with computable bounds. We are going to use this information
4040 // to place a runtime bound check.
4041 bool CanDoRT = true;
4043 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4046 // We assign a consecutive id to access from different alias sets.
4047 // Accesses between different groups doesn't need to be checked.
4049 for (auto &AS : AST) {
4050 unsigned NumReadPtrChecks = 0;
4051 unsigned NumWritePtrChecks = 0;
4053 // We assign consecutive id to access from different dependence sets.
4054 // Accesses within the same set don't need a runtime check.
4055 unsigned RunningDepId = 1;
4056 DenseMap<Value *, unsigned> DepSetId;
4059 Value *Ptr = A.getValue();
4060 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4061 MemAccessInfo Access(Ptr, IsWrite);
4064 ++NumWritePtrChecks;
4068 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4069 // When we run after a failing dependency check we have to make sure we
4070 // don't have wrapping pointers.
4071 (!ShouldCheckStride ||
4072 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4073 // The id of the dependence set.
4076 if (IsDepCheckNeeded) {
4077 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4078 unsigned &LeaderId = DepSetId[Leader];
4080 LeaderId = RunningDepId++;
4083 // Each access has its own dependence set.
4084 DepId = RunningDepId++;
4086 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4088 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4094 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4095 NumComparisons += 0; // Only one dependence set.
4097 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4098 NumWritePtrChecks - 1));
4104 // If the pointers that we would use for the bounds comparison have different
4105 // address spaces, assume the values aren't directly comparable, so we can't
4106 // use them for the runtime check. We also have to assume they could
4107 // overlap. In the future there should be metadata for whether address spaces
4109 unsigned NumPointers = RtCheck.Pointers.size();
4110 for (unsigned i = 0; i < NumPointers; ++i) {
4111 for (unsigned j = i + 1; j < NumPointers; ++j) {
4112 // Only need to check pointers between two different dependency sets.
4113 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4115 // Only need to check pointers in the same alias set.
4116 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4119 Value *PtrI = RtCheck.Pointers[i];
4120 Value *PtrJ = RtCheck.Pointers[j];
4122 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4123 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4125 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4126 " different address spaces\n");
4135 void AccessAnalysis::processMemAccesses() {
4136 // We process the set twice: first we process read-write pointers, last we
4137 // process read-only pointers. This allows us to skip dependence tests for
4138 // read-only pointers.
4140 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4141 DEBUG(dbgs() << " AST: "; AST.dump());
4142 DEBUG(dbgs() << "LV: Accesses:\n");
4144 for (auto A : Accesses)
4145 dbgs() << "\t" << *A.getPointer() << " (" <<
4146 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4147 "read-only" : "read")) << ")\n";
4150 // The AliasSetTracker has nicely partitioned our pointers by metadata
4151 // compatibility and potential for underlying-object overlap. As a result, we
4152 // only need to check for potential pointer dependencies within each alias
4154 for (auto &AS : AST) {
4155 // Note that both the alias-set tracker and the alias sets themselves used
4156 // linked lists internally and so the iteration order here is deterministic
4157 // (matching the original instruction order within each set).
4159 bool SetHasWrite = false;
4161 // Map of pointers to last access encountered.
4162 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4163 UnderlyingObjToAccessMap ObjToLastAccess;
4165 // Set of access to check after all writes have been processed.
4166 PtrAccessSet DeferredAccesses;
4168 // Iterate over each alias set twice, once to process read/write pointers,
4169 // and then to process read-only pointers.
4170 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4171 bool UseDeferred = SetIteration > 0;
4172 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4175 Value *Ptr = A.getValue();
4176 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4178 // If we're using the deferred access set, then it contains only reads.
4179 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4180 if (UseDeferred && !IsReadOnlyPtr)
4182 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4184 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4185 S.count(MemAccessInfo(Ptr, false))) &&
4186 "Alias-set pointer not in the access set?");
4188 MemAccessInfo Access(Ptr, IsWrite);
4189 DepCands.insert(Access);
4191 // Memorize read-only pointers for later processing and skip them in the
4192 // first round (they need to be checked after we have seen all write
4193 // pointers). Note: we also mark pointer that are not consecutive as
4194 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4195 // the second check for "!IsWrite".
4196 if (!UseDeferred && IsReadOnlyPtr) {
4197 DeferredAccesses.insert(Access);
4201 // If this is a write - check other reads and writes for conflicts. If
4202 // this is a read only check other writes for conflicts (but only if
4203 // there is no other write to the ptr - this is an optimization to
4204 // catch "a[i] = a[i] + " without having to do a dependence check).
4205 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4206 CheckDeps.insert(Access);
4207 IsRTCheckNeeded = true;
4213 // Create sets of pointers connected by a shared alias set and
4214 // underlying object.
4215 typedef SmallVector<Value*, 16> ValueVector;
4216 ValueVector TempObjects;
4217 GetUnderlyingObjects(Ptr, TempObjects, DL);
4218 for (Value *UnderlyingObj : TempObjects) {
4219 UnderlyingObjToAccessMap::iterator Prev =
4220 ObjToLastAccess.find(UnderlyingObj);
4221 if (Prev != ObjToLastAccess.end())
4222 DepCands.unionSets(Access, Prev->second);
4224 ObjToLastAccess[UnderlyingObj] = Access;
4232 /// \brief Checks memory dependences among accesses to the same underlying
4233 /// object to determine whether there vectorization is legal or not (and at
4234 /// which vectorization factor).
4236 /// This class works under the assumption that we already checked that memory
4237 /// locations with different underlying pointers are "must-not alias".
4238 /// We use the ScalarEvolution framework to symbolically evalutate access
4239 /// functions pairs. Since we currently don't restructure the loop we can rely
4240 /// on the program order of memory accesses to determine their safety.
4241 /// At the moment we will only deem accesses as safe for:
4242 /// * A negative constant distance assuming program order.
4244 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4245 /// a[i] = tmp; y = a[i];
4247 /// The latter case is safe because later checks guarantuee that there can't
4248 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4249 /// the same variable: a header phi can only be an induction or a reduction, a
4250 /// reduction can't have a memory sink, an induction can't have a memory
4251 /// source). This is important and must not be violated (or we have to
4252 /// resort to checking for cycles through memory).
4254 /// * A positive constant distance assuming program order that is bigger
4255 /// than the biggest memory access.
4257 /// tmp = a[i] OR b[i] = x
4258 /// a[i+2] = tmp y = b[i+2];
4260 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4262 /// * Zero distances and all accesses have the same size.
4264 class MemoryDepChecker {
4266 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4267 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4269 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4270 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4271 ShouldRetryWithRuntimeCheck(false) {}
4273 /// \brief Register the location (instructions are given increasing numbers)
4274 /// of a write access.
4275 void addAccess(StoreInst *SI) {
4276 Value *Ptr = SI->getPointerOperand();
4277 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4278 InstMap.push_back(SI);
4282 /// \brief Register the location (instructions are given increasing numbers)
4283 /// of a write access.
4284 void addAccess(LoadInst *LI) {
4285 Value *Ptr = LI->getPointerOperand();
4286 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4287 InstMap.push_back(LI);
4291 /// \brief Check whether the dependencies between the accesses are safe.
4293 /// Only checks sets with elements in \p CheckDeps.
4294 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4295 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4297 /// \brief The maximum number of bytes of a vector register we can vectorize
4298 /// the accesses safely with.
4299 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4301 /// \brief In same cases when the dependency check fails we can still
4302 /// vectorize the loop with a dynamic array access check.
4303 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4306 ScalarEvolution *SE;
4307 const DataLayout *DL;
4308 const Loop *InnermostLoop;
4310 /// \brief Maps access locations (ptr, read/write) to program order.
4311 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4313 /// \brief Memory access instructions in program order.
4314 SmallVector<Instruction *, 16> InstMap;
4316 /// \brief The program order index to be used for the next instruction.
4319 // We can access this many bytes in parallel safely.
4320 unsigned MaxSafeDepDistBytes;
4322 /// \brief If we see a non-constant dependence distance we can still try to
4323 /// vectorize this loop with runtime checks.
4324 bool ShouldRetryWithRuntimeCheck;
4326 /// \brief Check whether there is a plausible dependence between the two
4329 /// Access \p A must happen before \p B in program order. The two indices
4330 /// identify the index into the program order map.
4332 /// This function checks whether there is a plausible dependence (or the
4333 /// absence of such can't be proved) between the two accesses. If there is a
4334 /// plausible dependence but the dependence distance is bigger than one
4335 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4336 /// distance is smaller than any other distance encountered so far).
4337 /// Otherwise, this function returns true signaling a possible dependence.
4338 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4339 const MemAccessInfo &B, unsigned BIdx,
4340 ValueToValueMap &Strides);
4342 /// \brief Check whether the data dependence could prevent store-load
4344 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4347 } // end anonymous namespace
4349 static bool isInBoundsGep(Value *Ptr) {
4350 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4351 return GEP->isInBounds();
4355 /// \brief Check whether the access through \p Ptr has a constant stride.
4356 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4357 const Loop *Lp, ValueToValueMap &StridesMap) {
4358 const Type *Ty = Ptr->getType();
4359 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4361 // Make sure that the pointer does not point to aggregate types.
4362 const PointerType *PtrTy = cast<PointerType>(Ty);
4363 if (PtrTy->getElementType()->isAggregateType()) {
4364 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4369 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4371 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4373 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4374 << *Ptr << " SCEV: " << *PtrScev << "\n");
4378 // The accesss function must stride over the innermost loop.
4379 if (Lp != AR->getLoop()) {
4380 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4381 *Ptr << " SCEV: " << *PtrScev << "\n");
4384 // The address calculation must not wrap. Otherwise, a dependence could be
4386 // An inbounds getelementptr that is a AddRec with a unit stride
4387 // cannot wrap per definition. The unit stride requirement is checked later.
4388 // An getelementptr without an inbounds attribute and unit stride would have
4389 // to access the pointer value "0" which is undefined behavior in address
4390 // space 0, therefore we can also vectorize this case.
4391 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4392 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4393 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4394 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4395 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4396 << *Ptr << " SCEV: " << *PtrScev << "\n");
4400 // Check the step is constant.
4401 const SCEV *Step = AR->getStepRecurrence(*SE);
4403 // Calculate the pointer stride and check if it is consecutive.
4404 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4406 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4407 " SCEV: " << *PtrScev << "\n");
4411 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4412 const APInt &APStepVal = C->getValue()->getValue();
4414 // Huge step value - give up.
4415 if (APStepVal.getBitWidth() > 64)
4418 int64_t StepVal = APStepVal.getSExtValue();
4421 int64_t Stride = StepVal / Size;
4422 int64_t Rem = StepVal % Size;
4426 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4427 // know we can't "wrap around the address space". In case of address space
4428 // zero we know that this won't happen without triggering undefined behavior.
4429 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4430 Stride != 1 && Stride != -1)
4436 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4437 unsigned TypeByteSize) {
4438 // If loads occur at a distance that is not a multiple of a feasible vector
4439 // factor store-load forwarding does not take place.
4440 // Positive dependences might cause troubles because vectorizing them might
4441 // prevent store-load forwarding making vectorized code run a lot slower.
4442 // a[i] = a[i-3] ^ a[i-8];
4443 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4444 // hence on your typical architecture store-load forwarding does not take
4445 // place. Vectorizing in such cases does not make sense.
4446 // Store-load forwarding distance.
4447 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4448 // Maximum vector factor.
4449 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4450 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4451 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4453 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4455 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4456 MaxVFWithoutSLForwardIssues = (vf >>=1);
4461 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4462 DEBUG(dbgs() << "LV: Distance " << Distance <<
4463 " that could cause a store-load forwarding conflict\n");
4467 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4468 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4469 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4473 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4474 const MemAccessInfo &B, unsigned BIdx,
4475 ValueToValueMap &Strides) {
4476 assert (AIdx < BIdx && "Must pass arguments in program order");
4478 Value *APtr = A.getPointer();
4479 Value *BPtr = B.getPointer();
4480 bool AIsWrite = A.getInt();
4481 bool BIsWrite = B.getInt();
4483 // Two reads are independent.
4484 if (!AIsWrite && !BIsWrite)
4487 // We cannot check pointers in different address spaces.
4488 if (APtr->getType()->getPointerAddressSpace() !=
4489 BPtr->getType()->getPointerAddressSpace())
4492 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4493 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4495 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4496 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4498 const SCEV *Src = AScev;
4499 const SCEV *Sink = BScev;
4501 // If the induction step is negative we have to invert source and sink of the
4503 if (StrideAPtr < 0) {
4506 std::swap(APtr, BPtr);
4507 std::swap(Src, Sink);
4508 std::swap(AIsWrite, BIsWrite);
4509 std::swap(AIdx, BIdx);
4510 std::swap(StrideAPtr, StrideBPtr);
4513 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4515 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4516 << "(Induction step: " << StrideAPtr << ")\n");
4517 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4518 << *InstMap[BIdx] << ": " << *Dist << "\n");
4520 // Need consecutive accesses. We don't want to vectorize
4521 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4522 // the address space.
4523 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4524 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4528 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4530 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4531 ShouldRetryWithRuntimeCheck = true;
4535 Type *ATy = APtr->getType()->getPointerElementType();
4536 Type *BTy = BPtr->getType()->getPointerElementType();
4537 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4539 // Negative distances are not plausible dependencies.
4540 const APInt &Val = C->getValue()->getValue();
4541 if (Val.isNegative()) {
4542 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4543 if (IsTrueDataDependence &&
4544 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4548 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4552 // Write to the same location with the same size.
4553 // Could be improved to assert type sizes are the same (i32 == float, etc).
4557 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4561 assert(Val.isStrictlyPositive() && "Expect a positive value");
4563 // Positive distance bigger than max vectorization factor.
4566 "LV: ReadWrite-Write positive dependency with different types\n");
4570 unsigned Distance = (unsigned) Val.getZExtValue();
4572 // Bail out early if passed-in parameters make vectorization not feasible.
4573 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4574 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4576 // The distance must be bigger than the size needed for a vectorized version
4577 // of the operation and the size of the vectorized operation must not be
4578 // bigger than the currrent maximum size.
4579 if (Distance < 2*TypeByteSize ||
4580 2*TypeByteSize > MaxSafeDepDistBytes ||
4581 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4582 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4583 << Val.getSExtValue() << '\n');
4587 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4588 Distance : MaxSafeDepDistBytes;
4590 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4591 if (IsTrueDataDependence &&
4592 couldPreventStoreLoadForward(Distance, TypeByteSize))
4595 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4596 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4601 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4602 MemAccessInfoSet &CheckDeps,
4603 ValueToValueMap &Strides) {
4605 MaxSafeDepDistBytes = -1U;
4606 while (!CheckDeps.empty()) {
4607 MemAccessInfo CurAccess = *CheckDeps.begin();
4609 // Get the relevant memory access set.
4610 EquivalenceClasses<MemAccessInfo>::iterator I =
4611 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4613 // Check accesses within this set.
4614 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4615 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4617 // Check every access pair.
4619 CheckDeps.erase(*AI);
4620 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4622 // Check every accessing instruction pair in program order.
4623 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4624 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4625 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4626 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4627 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4629 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4640 bool LoopVectorizationLegality::canVectorizeMemory() {
4642 typedef SmallVector<Value*, 16> ValueVector;
4643 typedef SmallPtrSet<Value*, 16> ValueSet;
4645 // Holds the Load and Store *instructions*.
4649 // Holds all the different accesses in the loop.
4650 unsigned NumReads = 0;
4651 unsigned NumReadWrites = 0;
4653 PtrRtCheck.Pointers.clear();
4654 PtrRtCheck.Need = false;
4656 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4657 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4660 for (Loop::block_iterator bb = TheLoop->block_begin(),
4661 be = TheLoop->block_end(); bb != be; ++bb) {
4663 // Scan the BB and collect legal loads and stores.
4664 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4667 // If this is a load, save it. If this instruction can read from memory
4668 // but is not a load, then we quit. Notice that we don't handle function
4669 // calls that read or write.
4670 if (it->mayReadFromMemory()) {
4671 // Many math library functions read the rounding mode. We will only
4672 // vectorize a loop if it contains known function calls that don't set
4673 // the flag. Therefore, it is safe to ignore this read from memory.
4674 CallInst *Call = dyn_cast<CallInst>(it);
4675 if (Call && getIntrinsicIDForCall(Call, TLI))
4678 LoadInst *Ld = dyn_cast<LoadInst>(it);
4679 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4680 emitAnalysis(Report(Ld)
4681 << "read with atomic ordering or volatile read");
4682 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4686 Loads.push_back(Ld);
4687 DepChecker.addAccess(Ld);
4691 // Save 'store' instructions. Abort if other instructions write to memory.
4692 if (it->mayWriteToMemory()) {
4693 StoreInst *St = dyn_cast<StoreInst>(it);
4695 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4698 if (!St->isSimple() && !IsAnnotatedParallel) {
4699 emitAnalysis(Report(St)
4700 << "write with atomic ordering or volatile write");
4701 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4705 Stores.push_back(St);
4706 DepChecker.addAccess(St);
4711 // Now we have two lists that hold the loads and the stores.
4712 // Next, we find the pointers that they use.
4714 // Check if we see any stores. If there are no stores, then we don't
4715 // care if the pointers are *restrict*.
4716 if (!Stores.size()) {
4717 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4721 AccessAnalysis::DepCandidates DependentAccesses;
4722 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4724 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4725 // multiple times on the same object. If the ptr is accessed twice, once
4726 // for read and once for write, it will only appear once (on the write
4727 // list). This is okay, since we are going to check for conflicts between
4728 // writes and between reads and writes, but not between reads and reads.
4731 ValueVector::iterator I, IE;
4732 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4733 StoreInst *ST = cast<StoreInst>(*I);
4734 Value* Ptr = ST->getPointerOperand();
4736 if (isUniform(Ptr)) {
4739 << "write to a loop invariant address could not be vectorized");
4740 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4744 // If we did *not* see this pointer before, insert it to the read-write
4745 // list. At this phase it is only a 'write' list.
4746 if (Seen.insert(Ptr)) {
4749 AliasAnalysis::Location Loc = AA->getLocation(ST);
4750 // The TBAA metadata could have a control dependency on the predication
4751 // condition, so we cannot rely on it when determining whether or not we
4752 // need runtime pointer checks.
4753 if (blockNeedsPredication(ST->getParent()))
4754 Loc.AATags.TBAA = nullptr;
4756 Accesses.addStore(Loc);
4760 if (IsAnnotatedParallel) {
4762 << "LV: A loop annotated parallel, ignore memory dependency "
4767 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4768 LoadInst *LD = cast<LoadInst>(*I);
4769 Value* Ptr = LD->getPointerOperand();
4770 // If we did *not* see this pointer before, insert it to the
4771 // read list. If we *did* see it before, then it is already in
4772 // the read-write list. This allows us to vectorize expressions
4773 // such as A[i] += x; Because the address of A[i] is a read-write
4774 // pointer. This only works if the index of A[i] is consecutive.
4775 // If the address of i is unknown (for example A[B[i]]) then we may
4776 // read a few words, modify, and write a few words, and some of the
4777 // words may be written to the same address.
4778 bool IsReadOnlyPtr = false;
4779 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4781 IsReadOnlyPtr = true;
4784 AliasAnalysis::Location Loc = AA->getLocation(LD);
4785 // The TBAA metadata could have a control dependency on the predication
4786 // condition, so we cannot rely on it when determining whether or not we
4787 // need runtime pointer checks.
4788 if (blockNeedsPredication(LD->getParent()))
4789 Loc.AATags.TBAA = nullptr;
4791 Accesses.addLoad(Loc, IsReadOnlyPtr);
4794 // If we write (or read-write) to a single destination and there are no
4795 // other reads in this loop then is it safe to vectorize.
4796 if (NumReadWrites == 1 && NumReads == 0) {
4797 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4801 // Build dependence sets and check whether we need a runtime pointer bounds
4803 Accesses.buildDependenceSets();
4804 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4806 // Find pointers with computable bounds. We are going to use this information
4807 // to place a runtime bound check.
4808 unsigned NumComparisons = 0;
4809 bool CanDoRT = false;
4811 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4814 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4815 " pointer comparisons.\n");
4817 // If we only have one set of dependences to check pointers among we don't
4818 // need a runtime check.
4819 if (NumComparisons == 0 && NeedRTCheck)
4820 NeedRTCheck = false;
4822 // Check that we did not collect too many pointers or found an unsizeable
4824 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4830 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4833 if (NeedRTCheck && !CanDoRT) {
4834 emitAnalysis(Report() << "cannot identify array bounds");
4835 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4836 "the array bounds.\n");
4841 PtrRtCheck.Need = NeedRTCheck;
4843 bool CanVecMem = true;
4844 if (Accesses.isDependencyCheckNeeded()) {
4845 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4846 CanVecMem = DepChecker.areDepsSafe(
4847 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4848 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4850 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4851 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4854 // Clear the dependency checks. We assume they are not needed.
4855 Accesses.resetDepChecks();
4858 PtrRtCheck.Need = true;
4860 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4861 TheLoop, Strides, true);
4862 // Check that we did not collect too many pointers or found an unsizeable
4864 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4865 if (!CanDoRT && NumComparisons > 0)
4866 emitAnalysis(Report()
4867 << "cannot check memory dependencies at runtime");
4869 emitAnalysis(Report()
4870 << NumComparisons << " exceeds limit of "
4871 << RuntimeMemoryCheckThreshold
4872 << " dependent memory operations checked at runtime");
4873 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4883 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4885 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4886 " need a runtime memory check.\n");
4891 static bool hasMultipleUsesOf(Instruction *I,
4892 SmallPtrSetImpl<Instruction *> &Insts) {
4893 unsigned NumUses = 0;
4894 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4895 if (Insts.count(dyn_cast<Instruction>(*Use)))
4904 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4905 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4906 if (!Set.count(dyn_cast<Instruction>(*Use)))
4911 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4912 ReductionKind Kind) {
4913 if (Phi->getNumIncomingValues() != 2)
4916 // Reduction variables are only found in the loop header block.
4917 if (Phi->getParent() != TheLoop->getHeader())
4920 // Obtain the reduction start value from the value that comes from the loop
4922 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4924 // ExitInstruction is the single value which is used outside the loop.
4925 // We only allow for a single reduction value to be used outside the loop.
4926 // This includes users of the reduction, variables (which form a cycle
4927 // which ends in the phi node).
4928 Instruction *ExitInstruction = nullptr;
4929 // Indicates that we found a reduction operation in our scan.
4930 bool FoundReduxOp = false;
4932 // We start with the PHI node and scan for all of the users of this
4933 // instruction. All users must be instructions that can be used as reduction
4934 // variables (such as ADD). We must have a single out-of-block user. The cycle
4935 // must include the original PHI.
4936 bool FoundStartPHI = false;
4938 // To recognize min/max patterns formed by a icmp select sequence, we store
4939 // the number of instruction we saw from the recognized min/max pattern,
4940 // to make sure we only see exactly the two instructions.
4941 unsigned NumCmpSelectPatternInst = 0;
4942 ReductionInstDesc ReduxDesc(false, nullptr);
4944 SmallPtrSet<Instruction *, 8> VisitedInsts;
4945 SmallVector<Instruction *, 8> Worklist;
4946 Worklist.push_back(Phi);
4947 VisitedInsts.insert(Phi);
4949 // A value in the reduction can be used:
4950 // - By the reduction:
4951 // - Reduction operation:
4952 // - One use of reduction value (safe).
4953 // - Multiple use of reduction value (not safe).
4955 // - All uses of the PHI must be the reduction (safe).
4956 // - Otherwise, not safe.
4957 // - By one instruction outside of the loop (safe).
4958 // - By further instructions outside of the loop (not safe).
4959 // - By an instruction that is not part of the reduction (not safe).
4961 // * An instruction type other than PHI or the reduction operation.
4962 // * A PHI in the header other than the initial PHI.
4963 while (!Worklist.empty()) {
4964 Instruction *Cur = Worklist.back();
4965 Worklist.pop_back();
4968 // If the instruction has no users then this is a broken chain and can't be
4969 // a reduction variable.
4970 if (Cur->use_empty())
4973 bool IsAPhi = isa<PHINode>(Cur);
4975 // A header PHI use other than the original PHI.
4976 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4979 // Reductions of instructions such as Div, and Sub is only possible if the
4980 // LHS is the reduction variable.
4981 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4982 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4983 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4986 // Any reduction instruction must be of one of the allowed kinds.
4987 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4988 if (!ReduxDesc.IsReduction)
4991 // A reduction operation must only have one use of the reduction value.
4992 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4993 hasMultipleUsesOf(Cur, VisitedInsts))
4996 // All inputs to a PHI node must be a reduction value.
4997 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5000 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5001 isa<SelectInst>(Cur)))
5002 ++NumCmpSelectPatternInst;
5003 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5004 isa<SelectInst>(Cur)))
5005 ++NumCmpSelectPatternInst;
5007 // Check whether we found a reduction operator.
5008 FoundReduxOp |= !IsAPhi;
5010 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5011 // onto the stack. This way we are going to have seen all inputs to PHI
5012 // nodes once we get to them.
5013 SmallVector<Instruction *, 8> NonPHIs;
5014 SmallVector<Instruction *, 8> PHIs;
5015 for (User *U : Cur->users()) {
5016 Instruction *UI = cast<Instruction>(U);
5018 // Check if we found the exit user.
5019 BasicBlock *Parent = UI->getParent();
5020 if (!TheLoop->contains(Parent)) {
5021 // Exit if you find multiple outside users or if the header phi node is
5022 // being used. In this case the user uses the value of the previous
5023 // iteration, in which case we would loose "VF-1" iterations of the
5024 // reduction operation if we vectorize.
5025 if (ExitInstruction != nullptr || Cur == Phi)
5028 // The instruction used by an outside user must be the last instruction
5029 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5030 // operations on the value.
5031 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5034 ExitInstruction = Cur;
5038 // Process instructions only once (termination). Each reduction cycle
5039 // value must only be used once, except by phi nodes and min/max
5040 // reductions which are represented as a cmp followed by a select.
5041 ReductionInstDesc IgnoredVal(false, nullptr);
5042 if (VisitedInsts.insert(UI)) {
5043 if (isa<PHINode>(UI))
5046 NonPHIs.push_back(UI);
5047 } else if (!isa<PHINode>(UI) &&
5048 ((!isa<FCmpInst>(UI) &&
5049 !isa<ICmpInst>(UI) &&
5050 !isa<SelectInst>(UI)) ||
5051 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5054 // Remember that we completed the cycle.
5056 FoundStartPHI = true;
5058 Worklist.append(PHIs.begin(), PHIs.end());
5059 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5062 // This means we have seen one but not the other instruction of the
5063 // pattern or more than just a select and cmp.
5064 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5065 NumCmpSelectPatternInst != 2)
5068 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5071 // We found a reduction var if we have reached the original phi node and we
5072 // only have a single instruction with out-of-loop users.
5074 // This instruction is allowed to have out-of-loop users.
5075 AllowedExit.insert(ExitInstruction);
5077 // Save the description of this reduction variable.
5078 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5079 ReduxDesc.MinMaxKind);
5080 Reductions[Phi] = RD;
5081 // We've ended the cycle. This is a reduction variable if we have an
5082 // outside user and it has a binary op.
5087 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5088 /// pattern corresponding to a min(X, Y) or max(X, Y).
5089 LoopVectorizationLegality::ReductionInstDesc
5090 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5091 ReductionInstDesc &Prev) {
5093 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5094 "Expect a select instruction");
5095 Instruction *Cmp = nullptr;
5096 SelectInst *Select = nullptr;
5098 // We must handle the select(cmp()) as a single instruction. Advance to the
5100 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5101 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5102 return ReductionInstDesc(false, I);
5103 return ReductionInstDesc(Select, Prev.MinMaxKind);
5106 // Only handle single use cases for now.
5107 if (!(Select = dyn_cast<SelectInst>(I)))
5108 return ReductionInstDesc(false, I);
5109 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5110 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5111 return ReductionInstDesc(false, I);
5112 if (!Cmp->hasOneUse())
5113 return ReductionInstDesc(false, I);
5118 // Look for a min/max pattern.
5119 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5120 return ReductionInstDesc(Select, MRK_UIntMin);
5121 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5122 return ReductionInstDesc(Select, MRK_UIntMax);
5123 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5124 return ReductionInstDesc(Select, MRK_SIntMax);
5125 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5126 return ReductionInstDesc(Select, MRK_SIntMin);
5127 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5128 return ReductionInstDesc(Select, MRK_FloatMin);
5129 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5130 return ReductionInstDesc(Select, MRK_FloatMax);
5131 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5132 return ReductionInstDesc(Select, MRK_FloatMin);
5133 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5134 return ReductionInstDesc(Select, MRK_FloatMax);
5136 return ReductionInstDesc(false, I);
5139 LoopVectorizationLegality::ReductionInstDesc
5140 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5142 ReductionInstDesc &Prev) {
5143 bool FP = I->getType()->isFloatingPointTy();
5144 bool FastMath = FP && I->hasUnsafeAlgebra();
5145 switch (I->getOpcode()) {
5147 return ReductionInstDesc(false, I);
5148 case Instruction::PHI:
5149 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5150 Kind != RK_FloatMinMax))
5151 return ReductionInstDesc(false, I);
5152 return ReductionInstDesc(I, Prev.MinMaxKind);
5153 case Instruction::Sub:
5154 case Instruction::Add:
5155 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5156 case Instruction::Mul:
5157 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5158 case Instruction::And:
5159 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5160 case Instruction::Or:
5161 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5162 case Instruction::Xor:
5163 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5164 case Instruction::FMul:
5165 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5166 case Instruction::FSub:
5167 case Instruction::FAdd:
5168 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5169 case Instruction::FCmp:
5170 case Instruction::ICmp:
5171 case Instruction::Select:
5172 if (Kind != RK_IntegerMinMax &&
5173 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5174 return ReductionInstDesc(false, I);
5175 return isMinMaxSelectCmpPattern(I, Prev);
5179 LoopVectorizationLegality::InductionKind
5180 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5181 Type *PhiTy = Phi->getType();
5182 // We only handle integer and pointer inductions variables.
5183 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5184 return IK_NoInduction;
5186 // Check that the PHI is consecutive.
5187 const SCEV *PhiScev = SE->getSCEV(Phi);
5188 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5190 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5191 return IK_NoInduction;
5193 const SCEV *Step = AR->getStepRecurrence(*SE);
5195 // Integer inductions need to have a stride of one.
5196 if (PhiTy->isIntegerTy()) {
5198 return IK_IntInduction;
5199 if (Step->isAllOnesValue())
5200 return IK_ReverseIntInduction;
5201 return IK_NoInduction;
5204 // Calculate the pointer stride and check if it is consecutive.
5205 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5207 return IK_NoInduction;
5209 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5210 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5211 if (C->getValue()->equalsInt(Size))
5212 return IK_PtrInduction;
5213 else if (C->getValue()->equalsInt(0 - Size))
5214 return IK_ReversePtrInduction;
5216 return IK_NoInduction;
5219 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5220 Value *In0 = const_cast<Value*>(V);
5221 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5225 return Inductions.count(PN);
5228 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5229 assert(TheLoop->contains(BB) && "Unknown block used");
5231 // Blocks that do not dominate the latch need predication.
5232 BasicBlock* Latch = TheLoop->getLoopLatch();
5233 return !DT->dominates(BB, Latch);
5236 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5237 SmallPtrSetImpl<Value *> &SafePtrs) {
5238 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5239 // We might be able to hoist the load.
5240 if (it->mayReadFromMemory()) {
5241 LoadInst *LI = dyn_cast<LoadInst>(it);
5242 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5246 // We don't predicate stores at the moment.
5247 if (it->mayWriteToMemory()) {
5248 StoreInst *SI = dyn_cast<StoreInst>(it);
5249 // We only support predication of stores in basic blocks with one
5251 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5252 !SafePtrs.count(SI->getPointerOperand()) ||
5253 !SI->getParent()->getSinglePredecessor())
5259 // Check that we don't have a constant expression that can trap as operand.
5260 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5262 if (Constant *C = dyn_cast<Constant>(*OI))
5267 // The instructions below can trap.
5268 switch (it->getOpcode()) {
5270 case Instruction::UDiv:
5271 case Instruction::SDiv:
5272 case Instruction::URem:
5273 case Instruction::SRem:
5281 LoopVectorizationCostModel::VectorizationFactor
5282 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5283 // Width 1 means no vectorize
5284 VectorizationFactor Factor = { 1U, 0U };
5285 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5286 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5287 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5291 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5292 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5293 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5297 // Find the trip count.
5298 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5299 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5301 unsigned WidestType = getWidestType();
5302 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5303 unsigned MaxSafeDepDist = -1U;
5304 if (Legal->getMaxSafeDepDistBytes() != -1U)
5305 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5306 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5307 WidestRegister : MaxSafeDepDist);
5308 unsigned MaxVectorSize = WidestRegister / WidestType;
5309 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5310 DEBUG(dbgs() << "LV: The Widest register is: "
5311 << WidestRegister << " bits.\n");
5313 if (MaxVectorSize == 0) {
5314 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5318 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5319 " into one vector!");
5321 unsigned VF = MaxVectorSize;
5323 // If we optimize the program for size, avoid creating the tail loop.
5325 // If we are unable to calculate the trip count then don't try to vectorize.
5327 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5328 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5332 // Find the maximum SIMD width that can fit within the trip count.
5333 VF = TC % MaxVectorSize;
5338 // If the trip count that we found modulo the vectorization factor is not
5339 // zero then we require a tail.
5341 emitAnalysis(Report() << "cannot optimize for size and vectorize at the same time. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5342 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5347 int UserVF = Hints->getWidth();
5349 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5350 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5352 Factor.Width = UserVF;
5356 float Cost = expectedCost(1);
5358 const float ScalarCost = Cost;
5361 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5363 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5364 // Ignore scalar width, because the user explicitly wants vectorization.
5365 if (ForceVectorization && VF > 1) {
5367 Cost = expectedCost(Width) / (float)Width;
5370 for (unsigned i=2; i <= VF; i*=2) {
5371 // Notice that the vector loop needs to be executed less times, so
5372 // we need to divide the cost of the vector loops by the width of
5373 // the vector elements.
5374 float VectorCost = expectedCost(i) / (float)i;
5375 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5376 (int)VectorCost << ".\n");
5377 if (VectorCost < Cost) {
5383 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5384 << "LV: Vectorization seems to be not beneficial, "
5385 << "but was forced by a user.\n");
5386 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5387 Factor.Width = Width;
5388 Factor.Cost = Width * Cost;
5392 unsigned LoopVectorizationCostModel::getWidestType() {
5393 unsigned MaxWidth = 8;
5396 for (Loop::block_iterator bb = TheLoop->block_begin(),
5397 be = TheLoop->block_end(); bb != be; ++bb) {
5398 BasicBlock *BB = *bb;
5400 // For each instruction in the loop.
5401 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5402 Type *T = it->getType();
5404 // Only examine Loads, Stores and PHINodes.
5405 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5408 // Examine PHI nodes that are reduction variables.
5409 if (PHINode *PN = dyn_cast<PHINode>(it))
5410 if (!Legal->getReductionVars()->count(PN))
5413 // Examine the stored values.
5414 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5415 T = ST->getValueOperand()->getType();
5417 // Ignore loaded pointer types and stored pointer types that are not
5418 // consecutive. However, we do want to take consecutive stores/loads of
5419 // pointer vectors into account.
5420 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5423 MaxWidth = std::max(MaxWidth,
5424 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5432 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5434 unsigned LoopCost) {
5436 // -- The unroll heuristics --
5437 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5438 // There are many micro-architectural considerations that we can't predict
5439 // at this level. For example frontend pressure (on decode or fetch) due to
5440 // code size, or the number and capabilities of the execution ports.
5442 // We use the following heuristics to select the unroll factor:
5443 // 1. If the code has reductions the we unroll in order to break the cross
5444 // iteration dependency.
5445 // 2. If the loop is really small then we unroll in order to reduce the loop
5447 // 3. We don't unroll if we think that we will spill registers to memory due
5448 // to the increased register pressure.
5450 // Use the user preference, unless 'auto' is selected.
5451 int UserUF = Hints->getUnroll();
5455 // When we optimize for size we don't unroll.
5459 // We used the distance for the unroll factor.
5460 if (Legal->getMaxSafeDepDistBytes() != -1U)
5463 // Do not unroll loops with a relatively small trip count.
5464 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5465 TheLoop->getLoopLatch());
5466 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5469 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5470 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5474 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5475 TargetNumRegisters = ForceTargetNumScalarRegs;
5477 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5478 TargetNumRegisters = ForceTargetNumVectorRegs;
5481 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5482 // We divide by these constants so assume that we have at least one
5483 // instruction that uses at least one register.
5484 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5485 R.NumInstructions = std::max(R.NumInstructions, 1U);
5487 // We calculate the unroll factor using the following formula.
5488 // Subtract the number of loop invariants from the number of available
5489 // registers. These registers are used by all of the unrolled instances.
5490 // Next, divide the remaining registers by the number of registers that is
5491 // required by the loop, in order to estimate how many parallel instances
5492 // fit without causing spills. All of this is rounded down if necessary to be
5493 // a power of two. We want power of two unroll factors to simplify any
5494 // addressing operations or alignment considerations.
5495 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5498 // Don't count the induction variable as unrolled.
5499 if (EnableIndVarRegisterHeur)
5500 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5501 std::max(1U, (R.MaxLocalUsers - 1)));
5503 // Clamp the unroll factor ranges to reasonable factors.
5504 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5506 // Check if the user has overridden the unroll max.
5508 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5509 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5511 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5512 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5515 // If we did not calculate the cost for VF (because the user selected the VF)
5516 // then we calculate the cost of VF here.
5518 LoopCost = expectedCost(VF);
5520 // Clamp the calculated UF to be between the 1 and the max unroll factor
5521 // that the target allows.
5522 if (UF > MaxUnrollSize)
5527 // Unroll if we vectorized this loop and there is a reduction that could
5528 // benefit from unrolling.
5529 if (VF > 1 && Legal->getReductionVars()->size()) {
5530 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5534 // Note that if we've already vectorized the loop we will have done the
5535 // runtime check and so unrolling won't require further checks.
5536 bool UnrollingRequiresRuntimePointerCheck =
5537 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5539 // We want to unroll small loops in order to reduce the loop overhead and
5540 // potentially expose ILP opportunities.
5541 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5542 if (!UnrollingRequiresRuntimePointerCheck &&
5543 LoopCost < SmallLoopCost) {
5544 // We assume that the cost overhead is 1 and we use the cost model
5545 // to estimate the cost of the loop and unroll until the cost of the
5546 // loop overhead is about 5% of the cost of the loop.
5547 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5549 // Unroll until store/load ports (estimated by max unroll factor) are
5551 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5552 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5554 // If we have a scalar reduction (vector reductions are already dealt with
5555 // by this point), we can increase the critical path length if the loop
5556 // we're unrolling is inside another loop. Limit, by default to 2, so the
5557 // critical path only gets increased by one reduction operation.
5558 if (Legal->getReductionVars()->size() &&
5559 TheLoop->getLoopDepth() > 1) {
5560 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5561 SmallUF = std::min(SmallUF, F);
5562 StoresUF = std::min(StoresUF, F);
5563 LoadsUF = std::min(LoadsUF, F);
5566 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5567 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5568 return std::max(StoresUF, LoadsUF);
5571 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5575 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5579 LoopVectorizationCostModel::RegisterUsage
5580 LoopVectorizationCostModel::calculateRegisterUsage() {
5581 // This function calculates the register usage by measuring the highest number
5582 // of values that are alive at a single location. Obviously, this is a very
5583 // rough estimation. We scan the loop in a topological order in order and
5584 // assign a number to each instruction. We use RPO to ensure that defs are
5585 // met before their users. We assume that each instruction that has in-loop
5586 // users starts an interval. We record every time that an in-loop value is
5587 // used, so we have a list of the first and last occurrences of each
5588 // instruction. Next, we transpose this data structure into a multi map that
5589 // holds the list of intervals that *end* at a specific location. This multi
5590 // map allows us to perform a linear search. We scan the instructions linearly
5591 // and record each time that a new interval starts, by placing it in a set.
5592 // If we find this value in the multi-map then we remove it from the set.
5593 // The max register usage is the maximum size of the set.
5594 // We also search for instructions that are defined outside the loop, but are
5595 // used inside the loop. We need this number separately from the max-interval
5596 // usage number because when we unroll, loop-invariant values do not take
5598 LoopBlocksDFS DFS(TheLoop);
5602 R.NumInstructions = 0;
5604 // Each 'key' in the map opens a new interval. The values
5605 // of the map are the index of the 'last seen' usage of the
5606 // instruction that is the key.
5607 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5608 // Maps instruction to its index.
5609 DenseMap<unsigned, Instruction*> IdxToInstr;
5610 // Marks the end of each interval.
5611 IntervalMap EndPoint;
5612 // Saves the list of instruction indices that are used in the loop.
5613 SmallSet<Instruction*, 8> Ends;
5614 // Saves the list of values that are used in the loop but are
5615 // defined outside the loop, such as arguments and constants.
5616 SmallPtrSet<Value*, 8> LoopInvariants;
5619 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5620 be = DFS.endRPO(); bb != be; ++bb) {
5621 R.NumInstructions += (*bb)->size();
5622 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5624 Instruction *I = it;
5625 IdxToInstr[Index++] = I;
5627 // Save the end location of each USE.
5628 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5629 Value *U = I->getOperand(i);
5630 Instruction *Instr = dyn_cast<Instruction>(U);
5632 // Ignore non-instruction values such as arguments, constants, etc.
5633 if (!Instr) continue;
5635 // If this instruction is outside the loop then record it and continue.
5636 if (!TheLoop->contains(Instr)) {
5637 LoopInvariants.insert(Instr);
5641 // Overwrite previous end points.
5642 EndPoint[Instr] = Index;
5648 // Saves the list of intervals that end with the index in 'key'.
5649 typedef SmallVector<Instruction*, 2> InstrList;
5650 DenseMap<unsigned, InstrList> TransposeEnds;
5652 // Transpose the EndPoints to a list of values that end at each index.
5653 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5655 TransposeEnds[it->second].push_back(it->first);
5657 SmallSet<Instruction*, 8> OpenIntervals;
5658 unsigned MaxUsage = 0;
5661 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5662 for (unsigned int i = 0; i < Index; ++i) {
5663 Instruction *I = IdxToInstr[i];
5664 // Ignore instructions that are never used within the loop.
5665 if (!Ends.count(I)) continue;
5667 // Remove all of the instructions that end at this location.
5668 InstrList &List = TransposeEnds[i];
5669 for (unsigned int j=0, e = List.size(); j < e; ++j)
5670 OpenIntervals.erase(List[j]);
5672 // Count the number of live interals.
5673 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5675 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5676 OpenIntervals.size() << '\n');
5678 // Add the current instruction to the list of open intervals.
5679 OpenIntervals.insert(I);
5682 unsigned Invariant = LoopInvariants.size();
5683 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5684 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5685 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5687 R.LoopInvariantRegs = Invariant;
5688 R.MaxLocalUsers = MaxUsage;
5692 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5696 for (Loop::block_iterator bb = TheLoop->block_begin(),
5697 be = TheLoop->block_end(); bb != be; ++bb) {
5698 unsigned BlockCost = 0;
5699 BasicBlock *BB = *bb;
5701 // For each instruction in the old loop.
5702 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5703 // Skip dbg intrinsics.
5704 if (isa<DbgInfoIntrinsic>(it))
5707 unsigned C = getInstructionCost(it, VF);
5709 // Check if we should override the cost.
5710 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5711 C = ForceTargetInstructionCost;
5714 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5715 VF << " For instruction: " << *it << '\n');
5718 // We assume that if-converted blocks have a 50% chance of being executed.
5719 // When the code is scalar then some of the blocks are avoided due to CF.
5720 // When the code is vectorized we execute all code paths.
5721 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5730 /// \brief Check whether the address computation for a non-consecutive memory
5731 /// access looks like an unlikely candidate for being merged into the indexing
5734 /// We look for a GEP which has one index that is an induction variable and all
5735 /// other indices are loop invariant. If the stride of this access is also
5736 /// within a small bound we decide that this address computation can likely be
5737 /// merged into the addressing mode.
5738 /// In all other cases, we identify the address computation as complex.
5739 static bool isLikelyComplexAddressComputation(Value *Ptr,
5740 LoopVectorizationLegality *Legal,
5741 ScalarEvolution *SE,
5742 const Loop *TheLoop) {
5743 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5747 // We are looking for a gep with all loop invariant indices except for one
5748 // which should be an induction variable.
5749 unsigned NumOperands = Gep->getNumOperands();
5750 for (unsigned i = 1; i < NumOperands; ++i) {
5751 Value *Opd = Gep->getOperand(i);
5752 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5753 !Legal->isInductionVariable(Opd))
5757 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5758 // can likely be merged into the address computation.
5759 unsigned MaxMergeDistance = 64;
5761 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5765 // Check the step is constant.
5766 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5767 // Calculate the pointer stride and check if it is consecutive.
5768 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5772 const APInt &APStepVal = C->getValue()->getValue();
5774 // Huge step value - give up.
5775 if (APStepVal.getBitWidth() > 64)
5778 int64_t StepVal = APStepVal.getSExtValue();
5780 return StepVal > MaxMergeDistance;
5783 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5784 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5790 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5791 // If we know that this instruction will remain uniform, check the cost of
5792 // the scalar version.
5793 if (Legal->isUniformAfterVectorization(I))
5796 Type *RetTy = I->getType();
5797 Type *VectorTy = ToVectorTy(RetTy, VF);
5799 // TODO: We need to estimate the cost of intrinsic calls.
5800 switch (I->getOpcode()) {
5801 case Instruction::GetElementPtr:
5802 // We mark this instruction as zero-cost because the cost of GEPs in
5803 // vectorized code depends on whether the corresponding memory instruction
5804 // is scalarized or not. Therefore, we handle GEPs with the memory
5805 // instruction cost.
5807 case Instruction::Br: {
5808 return TTI.getCFInstrCost(I->getOpcode());
5810 case Instruction::PHI:
5811 //TODO: IF-converted IFs become selects.
5813 case Instruction::Add:
5814 case Instruction::FAdd:
5815 case Instruction::Sub:
5816 case Instruction::FSub:
5817 case Instruction::Mul:
5818 case Instruction::FMul:
5819 case Instruction::UDiv:
5820 case Instruction::SDiv:
5821 case Instruction::FDiv:
5822 case Instruction::URem:
5823 case Instruction::SRem:
5824 case Instruction::FRem:
5825 case Instruction::Shl:
5826 case Instruction::LShr:
5827 case Instruction::AShr:
5828 case Instruction::And:
5829 case Instruction::Or:
5830 case Instruction::Xor: {
5831 // Since we will replace the stride by 1 the multiplication should go away.
5832 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5834 // Certain instructions can be cheaper to vectorize if they have a constant
5835 // second vector operand. One example of this are shifts on x86.
5836 TargetTransformInfo::OperandValueKind Op1VK =
5837 TargetTransformInfo::OK_AnyValue;
5838 TargetTransformInfo::OperandValueKind Op2VK =
5839 TargetTransformInfo::OK_AnyValue;
5840 Value *Op2 = I->getOperand(1);
5842 // Check for a splat of a constant or for a non uniform vector of constants.
5843 if (isa<ConstantInt>(Op2))
5844 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5845 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5846 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5847 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5848 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5851 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5853 case Instruction::Select: {
5854 SelectInst *SI = cast<SelectInst>(I);
5855 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5856 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5857 Type *CondTy = SI->getCondition()->getType();
5859 CondTy = VectorType::get(CondTy, VF);
5861 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5863 case Instruction::ICmp:
5864 case Instruction::FCmp: {
5865 Type *ValTy = I->getOperand(0)->getType();
5866 VectorTy = ToVectorTy(ValTy, VF);
5867 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5869 case Instruction::Store:
5870 case Instruction::Load: {
5871 StoreInst *SI = dyn_cast<StoreInst>(I);
5872 LoadInst *LI = dyn_cast<LoadInst>(I);
5873 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5875 VectorTy = ToVectorTy(ValTy, VF);
5877 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5878 unsigned AS = SI ? SI->getPointerAddressSpace() :
5879 LI->getPointerAddressSpace();
5880 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5881 // We add the cost of address computation here instead of with the gep
5882 // instruction because only here we know whether the operation is
5885 return TTI.getAddressComputationCost(VectorTy) +
5886 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5888 // Scalarized loads/stores.
5889 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5890 bool Reverse = ConsecutiveStride < 0;
5891 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5892 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5893 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5894 bool IsComplexComputation =
5895 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5897 // The cost of extracting from the value vector and pointer vector.
5898 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5899 for (unsigned i = 0; i < VF; ++i) {
5900 // The cost of extracting the pointer operand.
5901 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5902 // In case of STORE, the cost of ExtractElement from the vector.
5903 // In case of LOAD, the cost of InsertElement into the returned
5905 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5906 Instruction::InsertElement,
5910 // The cost of the scalar loads/stores.
5911 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5912 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5917 // Wide load/stores.
5918 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5919 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5922 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5926 case Instruction::ZExt:
5927 case Instruction::SExt:
5928 case Instruction::FPToUI:
5929 case Instruction::FPToSI:
5930 case Instruction::FPExt:
5931 case Instruction::PtrToInt:
5932 case Instruction::IntToPtr:
5933 case Instruction::SIToFP:
5934 case Instruction::UIToFP:
5935 case Instruction::Trunc:
5936 case Instruction::FPTrunc:
5937 case Instruction::BitCast: {
5938 // We optimize the truncation of induction variable.
5939 // The cost of these is the same as the scalar operation.
5940 if (I->getOpcode() == Instruction::Trunc &&
5941 Legal->isInductionVariable(I->getOperand(0)))
5942 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5943 I->getOperand(0)->getType());
5945 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5946 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5948 case Instruction::Call: {
5949 CallInst *CI = cast<CallInst>(I);
5950 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5951 assert(ID && "Not an intrinsic call!");
5952 Type *RetTy = ToVectorTy(CI->getType(), VF);
5953 SmallVector<Type*, 4> Tys;
5954 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5955 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5956 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5959 // We are scalarizing the instruction. Return the cost of the scalar
5960 // instruction, plus the cost of insert and extract into vector
5961 // elements, times the vector width.
5964 if (!RetTy->isVoidTy() && VF != 1) {
5965 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5967 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5970 // The cost of inserting the results plus extracting each one of the
5972 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5975 // The cost of executing VF copies of the scalar instruction. This opcode
5976 // is unknown. Assume that it is the same as 'mul'.
5977 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5983 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5984 if (Scalar->isVoidTy() || VF == 1)
5986 return VectorType::get(Scalar, VF);
5989 char LoopVectorize::ID = 0;
5990 static const char lv_name[] = "Loop Vectorization";
5991 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5992 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5993 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5994 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5995 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5996 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5997 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5998 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5999 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6000 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6003 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6004 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6008 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6009 // Check for a store.
6010 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6011 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6013 // Check for a load.
6014 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6015 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6021 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6022 bool IfPredicateStore) {
6023 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6024 // Holds vector parameters or scalars, in case of uniform vals.
6025 SmallVector<VectorParts, 4> Params;
6027 setDebugLocFromInst(Builder, Instr);
6029 // Find all of the vectorized parameters.
6030 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6031 Value *SrcOp = Instr->getOperand(op);
6033 // If we are accessing the old induction variable, use the new one.
6034 if (SrcOp == OldInduction) {
6035 Params.push_back(getVectorValue(SrcOp));
6039 // Try using previously calculated values.
6040 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6042 // If the src is an instruction that appeared earlier in the basic block
6043 // then it should already be vectorized.
6044 if (SrcInst && OrigLoop->contains(SrcInst)) {
6045 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6046 // The parameter is a vector value from earlier.
6047 Params.push_back(WidenMap.get(SrcInst));
6049 // The parameter is a scalar from outside the loop. Maybe even a constant.
6050 VectorParts Scalars;
6051 Scalars.append(UF, SrcOp);
6052 Params.push_back(Scalars);
6056 assert(Params.size() == Instr->getNumOperands() &&
6057 "Invalid number of operands");
6059 // Does this instruction return a value ?
6060 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6062 Value *UndefVec = IsVoidRetTy ? nullptr :
6063 UndefValue::get(Instr->getType());
6064 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6065 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6067 Instruction *InsertPt = Builder.GetInsertPoint();
6068 BasicBlock *IfBlock = Builder.GetInsertBlock();
6069 BasicBlock *CondBlock = nullptr;
6072 Loop *VectorLp = nullptr;
6073 if (IfPredicateStore) {
6074 assert(Instr->getParent()->getSinglePredecessor() &&
6075 "Only support single predecessor blocks");
6076 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6077 Instr->getParent());
6078 VectorLp = LI->getLoopFor(IfBlock);
6079 assert(VectorLp && "Must have a loop for this block");
6082 // For each vector unroll 'part':
6083 for (unsigned Part = 0; Part < UF; ++Part) {
6084 // For each scalar that we create:
6086 // Start an "if (pred) a[i] = ..." block.
6087 Value *Cmp = nullptr;
6088 if (IfPredicateStore) {
6089 if (Cond[Part]->getType()->isVectorTy())
6091 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6092 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6093 ConstantInt::get(Cond[Part]->getType(), 1));
6094 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6095 LoopVectorBody.push_back(CondBlock);
6096 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6097 // Update Builder with newly created basic block.
6098 Builder.SetInsertPoint(InsertPt);
6101 Instruction *Cloned = Instr->clone();
6103 Cloned->setName(Instr->getName() + ".cloned");
6104 // Replace the operands of the cloned instructions with extracted scalars.
6105 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6106 Value *Op = Params[op][Part];
6107 Cloned->setOperand(op, Op);
6110 // Place the cloned scalar in the new loop.
6111 Builder.Insert(Cloned);
6113 // If the original scalar returns a value we need to place it in a vector
6114 // so that future users will be able to use it.
6116 VecResults[Part] = Cloned;
6119 if (IfPredicateStore) {
6120 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6121 LoopVectorBody.push_back(NewIfBlock);
6122 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6123 Builder.SetInsertPoint(InsertPt);
6124 Instruction *OldBr = IfBlock->getTerminator();
6125 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6126 OldBr->eraseFromParent();
6127 IfBlock = NewIfBlock;
6132 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6133 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6134 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6136 return scalarizeInstruction(Instr, IfPredicateStore);
6139 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6143 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6147 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6149 // When unrolling and the VF is 1, we only need to add a simple scalar.
6150 Type *ITy = Val->getType();
6151 assert(!ITy->isVectorTy() && "Val must be a scalar");
6152 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6153 return Builder.CreateAdd(Val, C, "induction");