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 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
112 cl::desc("Sets the vectorization interleave 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 interleave count.
161 static const unsigned MaxInterleaveFactor = 16;
163 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
164 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's max interleave factor for "
168 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
169 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
170 cl::desc("A flag that overrides the target's max interleave 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 /// This class keeps a number of loop annotations locally (as member variables)
979 /// and can, upon request, write them back as metadata on the loop. It will
980 /// initially scan the loop for existing metadata, and will update the local
981 /// values based on information in the loop.
982 /// We cannot write all values to metadata, as the mere presence of some info,
983 /// for example 'force', means a decision has been made. So, we need to be
984 /// careful NOT to add them if the user hasn't specifically asked so.
985 class LoopVectorizeHints {
992 /// Hint - associates name and validation with the hint value.
995 unsigned Value; // This may have to change for non-numeric values.
998 Hint(const char * Name, unsigned Value, HintKind Kind)
999 : Name(Name), Value(Value), Kind(Kind) { }
1001 bool validate(unsigned Val) {
1004 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1006 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1014 /// Vectorization width.
1016 /// Vectorization interleave factor.
1018 /// Vectorization forced
1020 /// Array to help iterating through all hints.
1021 Hint *Hints[3]; // avoiding initialisation due to MSVC2012
1023 /// Return the loop metadata prefix.
1024 static StringRef Prefix() { return "llvm.loop."; }
1028 FK_Undefined = -1, ///< Not selected.
1029 FK_Disabled = 0, ///< Forcing disabled.
1030 FK_Enabled = 1, ///< Forcing enabled.
1033 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1034 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1035 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1036 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1038 // FIXME: Move this up initialisation when MSVC requirement is 2013+
1040 Hints[1] = &Interleave;
1043 // Populate values with existing loop metadata.
1044 getHintsFromMetadata();
1046 // force-vector-interleave overrides DisableInterleaving.
1047 if (VectorizationInterleave.getNumOccurrences() > 0)
1048 Interleave.Value = VectorizationInterleave;
1050 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1051 << "LV: Interleaving disabled by the pass manager\n");
1054 /// Mark the loop L as already vectorized by setting the width to 1.
1055 void setAlreadyVectorized() {
1056 Width.Value = Interleave.Value = 1;
1057 // FIXME: Change all lines below for this when we can use MSVC 2013+
1058 //writeHintsToMetadata({ Width, Unroll });
1059 std::vector<Hint> hints;
1061 hints.emplace_back(Width);
1062 hints.emplace_back(Interleave);
1063 writeHintsToMetadata(std::move(hints));
1066 /// Dumps all the hint information.
1067 std::string emitRemark() const {
1069 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1070 R << "vectorization is explicitly disabled";
1072 R << "use -Rpass-analysis=loop-vectorize for more info";
1073 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1074 R << " (Force=true";
1075 if (Width.Value != 0)
1076 R << ", Vector Width=" << Width.Value;
1077 if (Interleave.Value != 0)
1078 R << ", Interleave Count=" << Interleave.Value;
1086 unsigned getWidth() const { return Width.Value; }
1087 unsigned getInterleave() const { return Interleave.Value; }
1088 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1091 /// Find hints specified in the loop metadata and update local values.
1092 void getHintsFromMetadata() {
1093 MDNode *LoopID = TheLoop->getLoopID();
1097 // First operand should refer to the loop id itself.
1098 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1099 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1101 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1102 const MDString *S = nullptr;
1103 SmallVector<Value*, 4> Args;
1105 // The expected hint is either a MDString or a MDNode with the first
1106 // operand a MDString.
1107 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1108 if (!MD || MD->getNumOperands() == 0)
1110 S = dyn_cast<MDString>(MD->getOperand(0));
1111 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1112 Args.push_back(MD->getOperand(i));
1114 S = dyn_cast<MDString>(LoopID->getOperand(i));
1115 assert(Args.size() == 0 && "too many arguments for MDString");
1121 // Check if the hint starts with the loop metadata prefix.
1122 StringRef Name = S->getString();
1123 if (Args.size() == 1)
1124 setHint(Name, Args[0]);
1128 /// Checks string hint with one operand and set value if valid.
1129 void setHint(StringRef Name, Value *Arg) {
1130 if (!Name.startswith(Prefix()))
1132 Name = Name.substr(Prefix().size(), StringRef::npos);
1134 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1136 unsigned Val = C->getZExtValue();
1138 for (auto H : Hints) {
1139 if (Name == H->Name) {
1140 if (H->validate(Val))
1143 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1149 /// Create a new hint from name / value pair.
1150 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1151 LLVMContext &Context = TheLoop->getHeader()->getContext();
1152 SmallVector<Value*, 2> Vals;
1153 Vals.push_back(MDString::get(Context, Name));
1154 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
1155 return MDNode::get(Context, Vals);
1158 /// Matches metadata with hint name.
1159 bool matchesHintMetadataName(MDNode *Node, std::vector<Hint> &HintTypes) {
1160 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1164 for (auto H : HintTypes)
1165 if (Name->getName().endswith(H.Name))
1170 /// Sets current hints into loop metadata, keeping other values intact.
1171 void writeHintsToMetadata(std::vector<Hint> HintTypes) {
1172 if (HintTypes.size() == 0)
1175 // Reserve the first element to LoopID (see below).
1176 SmallVector<Value*, 4> Vals(1);
1177 // If the loop already has metadata, then ignore the existing operands.
1178 MDNode *LoopID = TheLoop->getLoopID();
1180 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1181 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1182 // If node in update list, ignore old value.
1183 if (!matchesHintMetadataName(Node, HintTypes))
1184 Vals.push_back(Node);
1188 // Now, add the missing hints.
1189 for (auto H : HintTypes)
1191 createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1193 // Replace current metadata node with new one.
1194 LLVMContext &Context = TheLoop->getHeader()->getContext();
1195 MDNode *NewLoopID = MDNode::get(Context, Vals);
1196 // Set operand 0 to refer to the loop id itself.
1197 NewLoopID->replaceOperandWith(0, NewLoopID);
1199 TheLoop->setLoopID(NewLoopID);
1201 LoopID->replaceAllUsesWith(NewLoopID);
1205 /// The loop these hints belong to.
1206 const Loop *TheLoop;
1209 static void emitMissedWarning(Function *F, Loop *L,
1210 const LoopVectorizeHints &LH) {
1211 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1212 L->getStartLoc(), LH.emitRemark());
1214 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1215 if (LH.getWidth() != 1)
1216 emitLoopVectorizeWarning(
1217 F->getContext(), *F, L->getStartLoc(),
1218 "failed explicitly specified loop vectorization");
1219 else if (LH.getInterleave() != 1)
1220 emitLoopInterleaveWarning(
1221 F->getContext(), *F, L->getStartLoc(),
1222 "failed explicitly specified loop interleaving");
1226 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1228 return V.push_back(&L);
1230 for (Loop *InnerL : L)
1231 addInnerLoop(*InnerL, V);
1234 /// The LoopVectorize Pass.
1235 struct LoopVectorize : public FunctionPass {
1236 /// Pass identification, replacement for typeid
1239 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1241 DisableUnrolling(NoUnrolling),
1242 AlwaysVectorize(AlwaysVectorize) {
1243 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1246 ScalarEvolution *SE;
1247 const DataLayout *DL;
1249 TargetTransformInfo *TTI;
1251 BlockFrequencyInfo *BFI;
1252 TargetLibraryInfo *TLI;
1254 bool DisableUnrolling;
1255 bool AlwaysVectorize;
1257 BlockFrequency ColdEntryFreq;
1259 bool runOnFunction(Function &F) override {
1260 SE = &getAnalysis<ScalarEvolution>();
1261 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1262 DL = DLP ? &DLP->getDataLayout() : nullptr;
1263 LI = &getAnalysis<LoopInfo>();
1264 TTI = &getAnalysis<TargetTransformInfo>();
1265 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1266 BFI = &getAnalysis<BlockFrequencyInfo>();
1267 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1268 AA = &getAnalysis<AliasAnalysis>();
1270 // Compute some weights outside of the loop over the loops. Compute this
1271 // using a BranchProbability to re-use its scaling math.
1272 const BranchProbability ColdProb(1, 5); // 20%
1273 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1275 // If the target claims to have no vector registers don't attempt
1277 if (!TTI->getNumberOfRegisters(true))
1281 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1282 << ": Missing data layout\n");
1286 // Build up a worklist of inner-loops to vectorize. This is necessary as
1287 // the act of vectorizing or partially unrolling a loop creates new loops
1288 // and can invalidate iterators across the loops.
1289 SmallVector<Loop *, 8> Worklist;
1292 addInnerLoop(*L, Worklist);
1294 LoopsAnalyzed += Worklist.size();
1296 // Now walk the identified inner loops.
1297 bool Changed = false;
1298 while (!Worklist.empty())
1299 Changed |= processLoop(Worklist.pop_back_val());
1301 // Process each loop nest in the function.
1305 bool processLoop(Loop *L) {
1306 assert(L->empty() && "Only process inner loops.");
1309 const std::string DebugLocStr = getDebugLocString(L);
1312 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1313 << L->getHeader()->getParent()->getName() << "\" from "
1314 << DebugLocStr << "\n");
1316 LoopVectorizeHints Hints(L, DisableUnrolling);
1318 DEBUG(dbgs() << "LV: Loop hints:"
1320 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1322 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1324 : "?")) << " width=" << Hints.getWidth()
1325 << " unroll=" << Hints.getInterleave() << "\n");
1327 // Function containing loop
1328 Function *F = L->getHeader()->getParent();
1330 // Looking at the diagnostic output is the only way to determine if a loop
1331 // was vectorized (other than looking at the IR or machine code), so it
1332 // is important to generate an optimization remark for each loop. Most of
1333 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1334 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1335 // less verbose reporting vectorized loops and unvectorized loops that may
1336 // benefit from vectorization, respectively.
1338 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1339 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1340 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1341 L->getStartLoc(), Hints.emitRemark());
1345 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1346 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1347 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1348 L->getStartLoc(), Hints.emitRemark());
1352 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1353 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1354 emitOptimizationRemarkAnalysis(
1355 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1356 "loop not vectorized: vector width and interleave count are "
1357 "explicitly set to 1");
1361 // Check the loop for a trip count threshold:
1362 // do not vectorize loops with a tiny trip count.
1363 BasicBlock *Latch = L->getLoopLatch();
1364 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1365 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1366 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1367 << "This loop is not worth vectorizing.");
1368 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1369 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1371 DEBUG(dbgs() << "\n");
1372 emitOptimizationRemarkAnalysis(
1373 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1374 "vectorization is not beneficial and is not explicitly forced");
1379 // Check if it is legal to vectorize the loop.
1380 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1381 if (!LVL.canVectorize()) {
1382 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1383 emitMissedWarning(F, L, Hints);
1387 // Use the cost model.
1388 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, F, &Hints);
1390 // Check the function attributes to find out if this function should be
1391 // optimized for size.
1392 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1393 F->hasFnAttribute(Attribute::OptimizeForSize);
1395 // Compute the weighted frequency of this loop being executed and see if it
1396 // is less than 20% of the function entry baseline frequency. Note that we
1397 // always have a canonical loop here because we think we *can* vectoriez.
1398 // FIXME: This is hidden behind a flag due to pervasive problems with
1399 // exactly what block frequency models.
1400 if (LoopVectorizeWithBlockFrequency) {
1401 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1402 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1403 LoopEntryFreq < ColdEntryFreq)
1407 // Check the function attributes to see if implicit floats are allowed.a
1408 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1409 // an integer loop and the vector instructions selected are purely integer
1410 // vector instructions?
1411 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1412 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1413 "attribute is used.\n");
1414 emitOptimizationRemarkAnalysis(
1415 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1416 "loop not vectorized due to NoImplicitFloat attribute");
1417 emitMissedWarning(F, L, Hints);
1421 // Select the optimal vectorization factor.
1422 const LoopVectorizationCostModel::VectorizationFactor VF =
1423 CM.selectVectorizationFactor(OptForSize);
1425 // Select the unroll factor.
1427 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1429 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1430 << DebugLocStr << '\n');
1431 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1433 if (VF.Width == 1) {
1434 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1437 emitOptimizationRemarkAnalysis(
1438 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1439 "not beneficial to vectorize and user disabled interleaving");
1442 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1444 // Report the unrolling decision.
1445 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1446 Twine("unrolled with interleaving factor " +
1448 " (vectorization not beneficial)"));
1450 // We decided not to vectorize, but we may want to unroll.
1452 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1453 Unroller.vectorize(&LVL);
1455 // If we decided that it is *legal* to vectorize the loop then do it.
1456 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1460 // Report the vectorization decision.
1461 emitOptimizationRemark(
1462 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1463 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1464 ", unrolling interleave factor: " + Twine(UF) + ")");
1467 // Mark the loop as already vectorized to avoid vectorizing again.
1468 Hints.setAlreadyVectorized();
1470 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1474 void getAnalysisUsage(AnalysisUsage &AU) const override {
1475 AU.addRequiredID(LoopSimplifyID);
1476 AU.addRequiredID(LCSSAID);
1477 AU.addRequired<BlockFrequencyInfo>();
1478 AU.addRequired<DominatorTreeWrapperPass>();
1479 AU.addRequired<LoopInfo>();
1480 AU.addRequired<ScalarEvolution>();
1481 AU.addRequired<TargetTransformInfo>();
1482 AU.addRequired<AliasAnalysis>();
1483 AU.addPreserved<LoopInfo>();
1484 AU.addPreserved<DominatorTreeWrapperPass>();
1485 AU.addPreserved<AliasAnalysis>();
1490 } // end anonymous namespace
1492 //===----------------------------------------------------------------------===//
1493 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1494 // LoopVectorizationCostModel.
1495 //===----------------------------------------------------------------------===//
1497 static Value *stripIntegerCast(Value *V) {
1498 if (CastInst *CI = dyn_cast<CastInst>(V))
1499 if (CI->getOperand(0)->getType()->isIntegerTy())
1500 return CI->getOperand(0);
1504 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1506 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1508 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1509 ValueToValueMap &PtrToStride,
1510 Value *Ptr, Value *OrigPtr = nullptr) {
1512 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1514 // If there is an entry in the map return the SCEV of the pointer with the
1515 // symbolic stride replaced by one.
1516 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1517 if (SI != PtrToStride.end()) {
1518 Value *StrideVal = SI->second;
1521 StrideVal = stripIntegerCast(StrideVal);
1523 // Replace symbolic stride by one.
1524 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1525 ValueToValueMap RewriteMap;
1526 RewriteMap[StrideVal] = One;
1529 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1530 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1535 // Otherwise, just return the SCEV of the original pointer.
1536 return SE->getSCEV(Ptr);
1539 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1540 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1541 unsigned ASId, ValueToValueMap &Strides) {
1542 // Get the stride replaced scev.
1543 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1544 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1545 assert(AR && "Invalid addrec expression");
1546 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1547 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1548 Pointers.push_back(Ptr);
1549 Starts.push_back(AR->getStart());
1550 Ends.push_back(ScEnd);
1551 IsWritePtr.push_back(WritePtr);
1552 DependencySetId.push_back(DepSetId);
1553 AliasSetId.push_back(ASId);
1556 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1557 // We need to place the broadcast of invariant variables outside the loop.
1558 Instruction *Instr = dyn_cast<Instruction>(V);
1560 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1561 Instr->getParent()) != LoopVectorBody.end());
1562 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1564 // Place the code for broadcasting invariant variables in the new preheader.
1565 IRBuilder<>::InsertPointGuard Guard(Builder);
1567 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1569 // Broadcast the scalar into all locations in the vector.
1570 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1575 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1577 assert(Val->getType()->isVectorTy() && "Must be a vector");
1578 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1579 "Elem must be an integer");
1580 // Create the types.
1581 Type *ITy = Val->getType()->getScalarType();
1582 VectorType *Ty = cast<VectorType>(Val->getType());
1583 int VLen = Ty->getNumElements();
1584 SmallVector<Constant*, 8> Indices;
1586 // Create a vector of consecutive numbers from zero to VF.
1587 for (int i = 0; i < VLen; ++i) {
1588 int64_t Idx = Negate ? (-i) : i;
1589 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1592 // Add the consecutive indices to the vector value.
1593 Constant *Cv = ConstantVector::get(Indices);
1594 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1595 return Builder.CreateAdd(Val, Cv, "induction");
1598 /// \brief Find the operand of the GEP that should be checked for consecutive
1599 /// stores. This ignores trailing indices that have no effect on the final
1601 static unsigned getGEPInductionOperand(const DataLayout *DL,
1602 const GetElementPtrInst *Gep) {
1603 unsigned LastOperand = Gep->getNumOperands() - 1;
1604 unsigned GEPAllocSize = DL->getTypeAllocSize(
1605 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1607 // Walk backwards and try to peel off zeros.
1608 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1609 // Find the type we're currently indexing into.
1610 gep_type_iterator GEPTI = gep_type_begin(Gep);
1611 std::advance(GEPTI, LastOperand - 1);
1613 // If it's a type with the same allocation size as the result of the GEP we
1614 // can peel off the zero index.
1615 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1623 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1624 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1625 // Make sure that the pointer does not point to structs.
1626 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1629 // If this value is a pointer induction variable we know it is consecutive.
1630 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1631 if (Phi && Inductions.count(Phi)) {
1632 InductionInfo II = Inductions[Phi];
1633 if (IK_PtrInduction == II.IK)
1635 else if (IK_ReversePtrInduction == II.IK)
1639 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1643 unsigned NumOperands = Gep->getNumOperands();
1644 Value *GpPtr = Gep->getPointerOperand();
1645 // If this GEP value is a consecutive pointer induction variable and all of
1646 // the indices are constant then we know it is consecutive. We can
1647 Phi = dyn_cast<PHINode>(GpPtr);
1648 if (Phi && Inductions.count(Phi)) {
1650 // Make sure that the pointer does not point to structs.
1651 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1652 if (GepPtrType->getElementType()->isAggregateType())
1655 // Make sure that all of the index operands are loop invariant.
1656 for (unsigned i = 1; i < NumOperands; ++i)
1657 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1660 InductionInfo II = Inductions[Phi];
1661 if (IK_PtrInduction == II.IK)
1663 else if (IK_ReversePtrInduction == II.IK)
1667 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1669 // Check that all of the gep indices are uniform except for our induction
1671 for (unsigned i = 0; i != NumOperands; ++i)
1672 if (i != InductionOperand &&
1673 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1676 // We can emit wide load/stores only if the last non-zero index is the
1677 // induction variable.
1678 const SCEV *Last = nullptr;
1679 if (!Strides.count(Gep))
1680 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1682 // Because of the multiplication by a stride we can have a s/zext cast.
1683 // We are going to replace this stride by 1 so the cast is safe to ignore.
1685 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1686 // %0 = trunc i64 %indvars.iv to i32
1687 // %mul = mul i32 %0, %Stride1
1688 // %idxprom = zext i32 %mul to i64 << Safe cast.
1689 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1691 Last = replaceSymbolicStrideSCEV(SE, Strides,
1692 Gep->getOperand(InductionOperand), Gep);
1693 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1695 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1699 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1700 const SCEV *Step = AR->getStepRecurrence(*SE);
1702 // The memory is consecutive because the last index is consecutive
1703 // and all other indices are loop invariant.
1706 if (Step->isAllOnesValue())
1713 bool LoopVectorizationLegality::isUniform(Value *V) {
1714 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1717 InnerLoopVectorizer::VectorParts&
1718 InnerLoopVectorizer::getVectorValue(Value *V) {
1719 assert(V != Induction && "The new induction variable should not be used.");
1720 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1722 // If we have a stride that is replaced by one, do it here.
1723 if (Legal->hasStride(V))
1724 V = ConstantInt::get(V->getType(), 1);
1726 // If we have this scalar in the map, return it.
1727 if (WidenMap.has(V))
1728 return WidenMap.get(V);
1730 // If this scalar is unknown, assume that it is a constant or that it is
1731 // loop invariant. Broadcast V and save the value for future uses.
1732 Value *B = getBroadcastInstrs(V);
1733 return WidenMap.splat(V, B);
1736 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1737 assert(Vec->getType()->isVectorTy() && "Invalid type");
1738 SmallVector<Constant*, 8> ShuffleMask;
1739 for (unsigned i = 0; i < VF; ++i)
1740 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1742 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1743 ConstantVector::get(ShuffleMask),
1747 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1748 // Attempt to issue a wide load.
1749 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1750 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1752 assert((LI || SI) && "Invalid Load/Store instruction");
1754 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1755 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1756 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1757 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1758 // An alignment of 0 means target abi alignment. We need to use the scalar's
1759 // target abi alignment in such a case.
1761 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1762 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1763 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1764 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1766 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1767 return scalarizeInstruction(Instr, true);
1769 if (ScalarAllocatedSize != VectorElementSize)
1770 return scalarizeInstruction(Instr);
1772 // If the pointer is loop invariant or if it is non-consecutive,
1773 // scalarize the load.
1774 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1775 bool Reverse = ConsecutiveStride < 0;
1776 bool UniformLoad = LI && Legal->isUniform(Ptr);
1777 if (!ConsecutiveStride || UniformLoad)
1778 return scalarizeInstruction(Instr);
1780 Constant *Zero = Builder.getInt32(0);
1781 VectorParts &Entry = WidenMap.get(Instr);
1783 // Handle consecutive loads/stores.
1784 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1785 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1786 setDebugLocFromInst(Builder, Gep);
1787 Value *PtrOperand = Gep->getPointerOperand();
1788 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1789 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1791 // Create the new GEP with the new induction variable.
1792 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1793 Gep2->setOperand(0, FirstBasePtr);
1794 Gep2->setName("gep.indvar.base");
1795 Ptr = Builder.Insert(Gep2);
1797 setDebugLocFromInst(Builder, Gep);
1798 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1799 OrigLoop) && "Base ptr must be invariant");
1801 // The last index does not have to be the induction. It can be
1802 // consecutive and be a function of the index. For example A[I+1];
1803 unsigned NumOperands = Gep->getNumOperands();
1804 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1805 // Create the new GEP with the new induction variable.
1806 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1808 for (unsigned i = 0; i < NumOperands; ++i) {
1809 Value *GepOperand = Gep->getOperand(i);
1810 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1812 // Update last index or loop invariant instruction anchored in loop.
1813 if (i == InductionOperand ||
1814 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1815 assert((i == InductionOperand ||
1816 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1817 "Must be last index or loop invariant");
1819 VectorParts &GEPParts = getVectorValue(GepOperand);
1820 Value *Index = GEPParts[0];
1821 Index = Builder.CreateExtractElement(Index, Zero);
1822 Gep2->setOperand(i, Index);
1823 Gep2->setName("gep.indvar.idx");
1826 Ptr = Builder.Insert(Gep2);
1828 // Use the induction element ptr.
1829 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1830 setDebugLocFromInst(Builder, Ptr);
1831 VectorParts &PtrVal = getVectorValue(Ptr);
1832 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1837 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1838 "We do not allow storing to uniform addresses");
1839 setDebugLocFromInst(Builder, SI);
1840 // We don't want to update the value in the map as it might be used in
1841 // another expression. So don't use a reference type for "StoredVal".
1842 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1844 for (unsigned Part = 0; Part < UF; ++Part) {
1845 // Calculate the pointer for the specific unroll-part.
1846 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1849 // If we store to reverse consecutive memory locations then we need
1850 // to reverse the order of elements in the stored value.
1851 StoredVal[Part] = reverseVector(StoredVal[Part]);
1852 // If the address is consecutive but reversed, then the
1853 // wide store needs to start at the last vector element.
1854 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1855 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1858 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1859 DataTy->getPointerTo(AddressSpace));
1861 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1862 propagateMetadata(NewSI, SI);
1868 assert(LI && "Must have a load instruction");
1869 setDebugLocFromInst(Builder, LI);
1870 for (unsigned Part = 0; Part < UF; ++Part) {
1871 // Calculate the pointer for the specific unroll-part.
1872 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1875 // If the address is consecutive but reversed, then the
1876 // wide store needs to start at the last vector element.
1877 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1878 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1881 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1882 DataTy->getPointerTo(AddressSpace));
1883 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1884 propagateMetadata(NewLI, LI);
1885 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1889 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1890 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1891 // Holds vector parameters or scalars, in case of uniform vals.
1892 SmallVector<VectorParts, 4> Params;
1894 setDebugLocFromInst(Builder, Instr);
1896 // Find all of the vectorized parameters.
1897 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1898 Value *SrcOp = Instr->getOperand(op);
1900 // If we are accessing the old induction variable, use the new one.
1901 if (SrcOp == OldInduction) {
1902 Params.push_back(getVectorValue(SrcOp));
1906 // Try using previously calculated values.
1907 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1909 // If the src is an instruction that appeared earlier in the basic block
1910 // then it should already be vectorized.
1911 if (SrcInst && OrigLoop->contains(SrcInst)) {
1912 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1913 // The parameter is a vector value from earlier.
1914 Params.push_back(WidenMap.get(SrcInst));
1916 // The parameter is a scalar from outside the loop. Maybe even a constant.
1917 VectorParts Scalars;
1918 Scalars.append(UF, SrcOp);
1919 Params.push_back(Scalars);
1923 assert(Params.size() == Instr->getNumOperands() &&
1924 "Invalid number of operands");
1926 // Does this instruction return a value ?
1927 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1929 Value *UndefVec = IsVoidRetTy ? nullptr :
1930 UndefValue::get(VectorType::get(Instr->getType(), VF));
1931 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1932 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1934 Instruction *InsertPt = Builder.GetInsertPoint();
1935 BasicBlock *IfBlock = Builder.GetInsertBlock();
1936 BasicBlock *CondBlock = nullptr;
1939 Loop *VectorLp = nullptr;
1940 if (IfPredicateStore) {
1941 assert(Instr->getParent()->getSinglePredecessor() &&
1942 "Only support single predecessor blocks");
1943 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1944 Instr->getParent());
1945 VectorLp = LI->getLoopFor(IfBlock);
1946 assert(VectorLp && "Must have a loop for this block");
1949 // For each vector unroll 'part':
1950 for (unsigned Part = 0; Part < UF; ++Part) {
1951 // For each scalar that we create:
1952 for (unsigned Width = 0; Width < VF; ++Width) {
1955 Value *Cmp = nullptr;
1956 if (IfPredicateStore) {
1957 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1958 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1959 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1960 LoopVectorBody.push_back(CondBlock);
1961 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1962 // Update Builder with newly created basic block.
1963 Builder.SetInsertPoint(InsertPt);
1966 Instruction *Cloned = Instr->clone();
1968 Cloned->setName(Instr->getName() + ".cloned");
1969 // Replace the operands of the cloned instructions with extracted scalars.
1970 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1971 Value *Op = Params[op][Part];
1972 // Param is a vector. Need to extract the right lane.
1973 if (Op->getType()->isVectorTy())
1974 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1975 Cloned->setOperand(op, Op);
1978 // Place the cloned scalar in the new loop.
1979 Builder.Insert(Cloned);
1981 // If the original scalar returns a value we need to place it in a vector
1982 // so that future users will be able to use it.
1984 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1985 Builder.getInt32(Width));
1987 if (IfPredicateStore) {
1988 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1989 LoopVectorBody.push_back(NewIfBlock);
1990 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1991 Builder.SetInsertPoint(InsertPt);
1992 Instruction *OldBr = IfBlock->getTerminator();
1993 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1994 OldBr->eraseFromParent();
1995 IfBlock = NewIfBlock;
2001 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2005 if (Instruction *I = dyn_cast<Instruction>(V))
2006 return I->getParent() == Loc->getParent() ? I : nullptr;
2010 std::pair<Instruction *, Instruction *>
2011 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2012 Instruction *tnullptr = nullptr;
2013 if (!Legal->mustCheckStrides())
2014 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2016 IRBuilder<> ChkBuilder(Loc);
2019 Value *Check = nullptr;
2020 Instruction *FirstInst = nullptr;
2021 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2022 SE = Legal->strides_end();
2024 Value *Ptr = stripIntegerCast(*SI);
2025 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2027 // Store the first instruction we create.
2028 FirstInst = getFirstInst(FirstInst, C, Loc);
2030 Check = ChkBuilder.CreateOr(Check, C);
2035 // We have to do this trickery because the IRBuilder might fold the check to a
2036 // constant expression in which case there is no Instruction anchored in a
2038 LLVMContext &Ctx = Loc->getContext();
2039 Instruction *TheCheck =
2040 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2041 ChkBuilder.Insert(TheCheck, "stride.not.one");
2042 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2044 return std::make_pair(FirstInst, TheCheck);
2047 std::pair<Instruction *, Instruction *>
2048 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2049 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2050 Legal->getRuntimePointerCheck();
2052 Instruction *tnullptr = nullptr;
2053 if (!PtrRtCheck->Need)
2054 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2056 unsigned NumPointers = PtrRtCheck->Pointers.size();
2057 SmallVector<TrackingVH<Value> , 2> Starts;
2058 SmallVector<TrackingVH<Value> , 2> Ends;
2060 LLVMContext &Ctx = Loc->getContext();
2061 SCEVExpander Exp(*SE, "induction");
2062 Instruction *FirstInst = nullptr;
2064 for (unsigned i = 0; i < NumPointers; ++i) {
2065 Value *Ptr = PtrRtCheck->Pointers[i];
2066 const SCEV *Sc = SE->getSCEV(Ptr);
2068 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2069 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2071 Starts.push_back(Ptr);
2072 Ends.push_back(Ptr);
2074 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2075 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2077 // Use this type for pointer arithmetic.
2078 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2080 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2081 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2082 Starts.push_back(Start);
2083 Ends.push_back(End);
2087 IRBuilder<> ChkBuilder(Loc);
2088 // Our instructions might fold to a constant.
2089 Value *MemoryRuntimeCheck = nullptr;
2090 for (unsigned i = 0; i < NumPointers; ++i) {
2091 for (unsigned j = i+1; j < NumPointers; ++j) {
2092 // No need to check if two readonly pointers intersect.
2093 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2096 // Only need to check pointers between two different dependency sets.
2097 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2099 // Only need to check pointers in the same alias set.
2100 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2103 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2104 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2106 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2107 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2108 "Trying to bounds check pointers with different address spaces");
2110 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2111 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2113 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2114 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2115 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2116 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2118 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2119 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2120 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2121 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2122 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2123 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2124 if (MemoryRuntimeCheck) {
2125 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2127 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2129 MemoryRuntimeCheck = IsConflict;
2133 // We have to do this trickery because the IRBuilder might fold the check to a
2134 // constant expression in which case there is no Instruction anchored in a
2136 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2137 ConstantInt::getTrue(Ctx));
2138 ChkBuilder.Insert(Check, "memcheck.conflict");
2139 FirstInst = getFirstInst(FirstInst, Check, Loc);
2140 return std::make_pair(FirstInst, Check);
2143 void InnerLoopVectorizer::createEmptyLoop() {
2145 In this function we generate a new loop. The new loop will contain
2146 the vectorized instructions while the old loop will continue to run the
2149 [ ] <-- Back-edge taken count overflow check.
2152 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2155 || [ ] <-- vector pre header.
2159 || [ ]_| <-- vector loop.
2162 | >[ ] <--- middle-block.
2165 -|- >[ ] <--- new preheader.
2169 | [ ]_| <-- old scalar loop to handle remainder.
2172 >[ ] <-- exit block.
2176 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2177 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2178 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2179 assert(BypassBlock && "Invalid loop structure");
2180 assert(ExitBlock && "Must have an exit block");
2182 // Some loops have a single integer induction variable, while other loops
2183 // don't. One example is c++ iterators that often have multiple pointer
2184 // induction variables. In the code below we also support a case where we
2185 // don't have a single induction variable.
2186 OldInduction = Legal->getInduction();
2187 Type *IdxTy = Legal->getWidestInductionType();
2189 // Find the loop boundaries.
2190 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2191 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2193 // The exit count might have the type of i64 while the phi is i32. This can
2194 // happen if we have an induction variable that is sign extended before the
2195 // compare. The only way that we get a backedge taken count is that the
2196 // induction variable was signed and as such will not overflow. In such a case
2197 // truncation is legal.
2198 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2199 IdxTy->getPrimitiveSizeInBits())
2200 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2202 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2203 // Get the total trip count from the count by adding 1.
2204 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2205 SE->getConstant(BackedgeTakeCount->getType(), 1));
2207 // Expand the trip count and place the new instructions in the preheader.
2208 // Notice that the pre-header does not change, only the loop body.
2209 SCEVExpander Exp(*SE, "induction");
2211 // We need to test whether the backedge-taken count is uint##_max. Adding one
2212 // to it will cause overflow and an incorrect loop trip count in the vector
2213 // body. In case of overflow we want to directly jump to the scalar remainder
2215 Value *BackedgeCount =
2216 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2217 BypassBlock->getTerminator());
2218 if (BackedgeCount->getType()->isPointerTy())
2219 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2220 "backedge.ptrcnt.to.int",
2221 BypassBlock->getTerminator());
2222 Instruction *CheckBCOverflow =
2223 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2224 Constant::getAllOnesValue(BackedgeCount->getType()),
2225 "backedge.overflow", BypassBlock->getTerminator());
2227 // The loop index does not have to start at Zero. Find the original start
2228 // value from the induction PHI node. If we don't have an induction variable
2229 // then we know that it starts at zero.
2230 Builder.SetInsertPoint(BypassBlock->getTerminator());
2231 Value *StartIdx = ExtendedIdx = OldInduction ?
2232 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2234 ConstantInt::get(IdxTy, 0);
2236 // We need an instruction to anchor the overflow check on. StartIdx needs to
2237 // be defined before the overflow check branch. Because the scalar preheader
2238 // is going to merge the start index and so the overflow branch block needs to
2239 // contain a definition of the start index.
2240 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2241 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2242 BypassBlock->getTerminator());
2244 // Count holds the overall loop count (N).
2245 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2246 BypassBlock->getTerminator());
2248 LoopBypassBlocks.push_back(BypassBlock);
2250 // Split the single block loop into the two loop structure described above.
2251 BasicBlock *VectorPH =
2252 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2253 BasicBlock *VecBody =
2254 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2255 BasicBlock *MiddleBlock =
2256 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2257 BasicBlock *ScalarPH =
2258 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2260 // Create and register the new vector loop.
2261 Loop* Lp = new Loop();
2262 Loop *ParentLoop = OrigLoop->getParentLoop();
2264 // Insert the new loop into the loop nest and register the new basic blocks
2265 // before calling any utilities such as SCEV that require valid LoopInfo.
2267 ParentLoop->addChildLoop(Lp);
2268 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2269 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2270 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2272 LI->addTopLevelLoop(Lp);
2274 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2276 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2278 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2280 // Generate the induction variable.
2281 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2282 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2283 // The loop step is equal to the vectorization factor (num of SIMD elements)
2284 // times the unroll factor (num of SIMD instructions).
2285 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2287 // This is the IR builder that we use to add all of the logic for bypassing
2288 // the new vector loop.
2289 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2290 setDebugLocFromInst(BypassBuilder,
2291 getDebugLocFromInstOrOperands(OldInduction));
2293 // We may need to extend the index in case there is a type mismatch.
2294 // We know that the count starts at zero and does not overflow.
2295 if (Count->getType() != IdxTy) {
2296 // The exit count can be of pointer type. Convert it to the correct
2298 if (ExitCount->getType()->isPointerTy())
2299 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2301 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2304 // Add the start index to the loop count to get the new end index.
2305 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2307 // Now we need to generate the expression for N - (N % VF), which is
2308 // the part that the vectorized body will execute.
2309 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2310 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2311 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2312 "end.idx.rnd.down");
2314 // Now, compare the new count to zero. If it is zero skip the vector loop and
2315 // jump to the scalar loop.
2317 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2319 BasicBlock *LastBypassBlock = BypassBlock;
2321 // Generate code to check that the loops trip count that we computed by adding
2322 // one to the backedge-taken count will not overflow.
2324 auto PastOverflowCheck =
2325 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2326 BasicBlock *CheckBlock =
2327 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2329 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2330 LoopBypassBlocks.push_back(CheckBlock);
2331 Instruction *OldTerm = LastBypassBlock->getTerminator();
2332 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2333 OldTerm->eraseFromParent();
2334 LastBypassBlock = CheckBlock;
2337 // Generate the code to check that the strides we assumed to be one are really
2338 // one. We want the new basic block to start at the first instruction in a
2339 // sequence of instructions that form a check.
2340 Instruction *StrideCheck;
2341 Instruction *FirstCheckInst;
2342 std::tie(FirstCheckInst, StrideCheck) =
2343 addStrideCheck(LastBypassBlock->getTerminator());
2345 // Create a new block containing the stride check.
2346 BasicBlock *CheckBlock =
2347 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2349 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2350 LoopBypassBlocks.push_back(CheckBlock);
2352 // Replace the branch into the memory check block with a conditional branch
2353 // for the "few elements case".
2354 Instruction *OldTerm = LastBypassBlock->getTerminator();
2355 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2356 OldTerm->eraseFromParent();
2359 LastBypassBlock = CheckBlock;
2362 // Generate the code that checks in runtime if arrays overlap. We put the
2363 // checks into a separate block to make the more common case of few elements
2365 Instruction *MemRuntimeCheck;
2366 std::tie(FirstCheckInst, MemRuntimeCheck) =
2367 addRuntimeCheck(LastBypassBlock->getTerminator());
2368 if (MemRuntimeCheck) {
2369 // Create a new block containing the memory check.
2370 BasicBlock *CheckBlock =
2371 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2373 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2374 LoopBypassBlocks.push_back(CheckBlock);
2376 // Replace the branch into the memory check block with a conditional branch
2377 // for the "few elements case".
2378 Instruction *OldTerm = LastBypassBlock->getTerminator();
2379 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2380 OldTerm->eraseFromParent();
2382 Cmp = MemRuntimeCheck;
2383 LastBypassBlock = CheckBlock;
2386 LastBypassBlock->getTerminator()->eraseFromParent();
2387 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2390 // We are going to resume the execution of the scalar loop.
2391 // Go over all of the induction variables that we found and fix the
2392 // PHIs that are left in the scalar version of the loop.
2393 // The starting values of PHI nodes depend on the counter of the last
2394 // iteration in the vectorized loop.
2395 // If we come from a bypass edge then we need to start from the original
2398 // This variable saves the new starting index for the scalar loop.
2399 PHINode *ResumeIndex = nullptr;
2400 LoopVectorizationLegality::InductionList::iterator I, E;
2401 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2402 // Set builder to point to last bypass block.
2403 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2404 for (I = List->begin(), E = List->end(); I != E; ++I) {
2405 PHINode *OrigPhi = I->first;
2406 LoopVectorizationLegality::InductionInfo II = I->second;
2408 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2409 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2410 MiddleBlock->getTerminator());
2411 // We might have extended the type of the induction variable but we need a
2412 // truncated version for the scalar loop.
2413 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2414 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2415 MiddleBlock->getTerminator()) : nullptr;
2417 // Create phi nodes to merge from the backedge-taken check block.
2418 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2419 ScalarPH->getTerminator());
2420 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2422 PHINode *BCTruncResumeVal = nullptr;
2423 if (OrigPhi == OldInduction) {
2425 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2426 ScalarPH->getTerminator());
2427 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2430 Value *EndValue = nullptr;
2432 case LoopVectorizationLegality::IK_NoInduction:
2433 llvm_unreachable("Unknown induction");
2434 case LoopVectorizationLegality::IK_IntInduction: {
2435 // Handle the integer induction counter.
2436 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2438 // We have the canonical induction variable.
2439 if (OrigPhi == OldInduction) {
2440 // Create a truncated version of the resume value for the scalar loop,
2441 // we might have promoted the type to a larger width.
2443 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2444 // The new PHI merges the original incoming value, in case of a bypass,
2445 // or the value at the end of the vectorized loop.
2446 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2447 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2448 TruncResumeVal->addIncoming(EndValue, VecBody);
2450 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2452 // We know what the end value is.
2453 EndValue = IdxEndRoundDown;
2454 // We also know which PHI node holds it.
2455 ResumeIndex = ResumeVal;
2459 // Not the canonical induction variable - add the vector loop count to the
2461 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2462 II.StartValue->getType(),
2464 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2467 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2468 // Convert the CountRoundDown variable to the PHI size.
2469 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2470 II.StartValue->getType(),
2472 // Handle reverse integer induction counter.
2473 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2476 case LoopVectorizationLegality::IK_PtrInduction: {
2477 // For pointer induction variables, calculate the offset using
2479 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2483 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2484 // The value at the end of the loop for the reverse pointer is calculated
2485 // by creating a GEP with a negative index starting from the start value.
2486 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2487 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2489 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2495 // The new PHI merges the original incoming value, in case of a bypass,
2496 // or the value at the end of the vectorized loop.
2497 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2498 if (OrigPhi == OldInduction)
2499 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2501 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2503 ResumeVal->addIncoming(EndValue, VecBody);
2505 // Fix the scalar body counter (PHI node).
2506 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2508 // The old induction's phi node in the scalar body needs the truncated
2510 if (OrigPhi == OldInduction) {
2511 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2512 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2514 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2515 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2519 // If we are generating a new induction variable then we also need to
2520 // generate the code that calculates the exit value. This value is not
2521 // simply the end of the counter because we may skip the vectorized body
2522 // in case of a runtime check.
2524 assert(!ResumeIndex && "Unexpected resume value found");
2525 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2526 MiddleBlock->getTerminator());
2527 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2528 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2529 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2532 // Make sure that we found the index where scalar loop needs to continue.
2533 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2534 "Invalid resume Index");
2536 // Add a check in the middle block to see if we have completed
2537 // all of the iterations in the first vector loop.
2538 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2539 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2540 ResumeIndex, "cmp.n",
2541 MiddleBlock->getTerminator());
2543 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2544 // Remove the old terminator.
2545 MiddleBlock->getTerminator()->eraseFromParent();
2547 // Create i+1 and fill the PHINode.
2548 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2549 Induction->addIncoming(StartIdx, VectorPH);
2550 Induction->addIncoming(NextIdx, VecBody);
2551 // Create the compare.
2552 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2553 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2555 // Now we have two terminators. Remove the old one from the block.
2556 VecBody->getTerminator()->eraseFromParent();
2558 // Get ready to start creating new instructions into the vectorized body.
2559 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2562 LoopVectorPreHeader = VectorPH;
2563 LoopScalarPreHeader = ScalarPH;
2564 LoopMiddleBlock = MiddleBlock;
2565 LoopExitBlock = ExitBlock;
2566 LoopVectorBody.push_back(VecBody);
2567 LoopScalarBody = OldBasicBlock;
2569 LoopVectorizeHints Hints(Lp, true);
2570 Hints.setAlreadyVectorized();
2573 /// This function returns the identity element (or neutral element) for
2574 /// the operation K.
2576 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2581 // Adding, Xoring, Oring zero to a number does not change it.
2582 return ConstantInt::get(Tp, 0);
2583 case RK_IntegerMult:
2584 // Multiplying a number by 1 does not change it.
2585 return ConstantInt::get(Tp, 1);
2587 // AND-ing a number with an all-1 value does not change it.
2588 return ConstantInt::get(Tp, -1, true);
2590 // Multiplying a number by 1 does not change it.
2591 return ConstantFP::get(Tp, 1.0L);
2593 // Adding zero to a number does not change it.
2594 return ConstantFP::get(Tp, 0.0L);
2596 llvm_unreachable("Unknown reduction kind");
2600 /// This function translates the reduction kind to an LLVM binary operator.
2602 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2604 case LoopVectorizationLegality::RK_IntegerAdd:
2605 return Instruction::Add;
2606 case LoopVectorizationLegality::RK_IntegerMult:
2607 return Instruction::Mul;
2608 case LoopVectorizationLegality::RK_IntegerOr:
2609 return Instruction::Or;
2610 case LoopVectorizationLegality::RK_IntegerAnd:
2611 return Instruction::And;
2612 case LoopVectorizationLegality::RK_IntegerXor:
2613 return Instruction::Xor;
2614 case LoopVectorizationLegality::RK_FloatMult:
2615 return Instruction::FMul;
2616 case LoopVectorizationLegality::RK_FloatAdd:
2617 return Instruction::FAdd;
2618 case LoopVectorizationLegality::RK_IntegerMinMax:
2619 return Instruction::ICmp;
2620 case LoopVectorizationLegality::RK_FloatMinMax:
2621 return Instruction::FCmp;
2623 llvm_unreachable("Unknown reduction operation");
2627 Value *createMinMaxOp(IRBuilder<> &Builder,
2628 LoopVectorizationLegality::MinMaxReductionKind RK,
2631 CmpInst::Predicate P = CmpInst::ICMP_NE;
2634 llvm_unreachable("Unknown min/max reduction kind");
2635 case LoopVectorizationLegality::MRK_UIntMin:
2636 P = CmpInst::ICMP_ULT;
2638 case LoopVectorizationLegality::MRK_UIntMax:
2639 P = CmpInst::ICMP_UGT;
2641 case LoopVectorizationLegality::MRK_SIntMin:
2642 P = CmpInst::ICMP_SLT;
2644 case LoopVectorizationLegality::MRK_SIntMax:
2645 P = CmpInst::ICMP_SGT;
2647 case LoopVectorizationLegality::MRK_FloatMin:
2648 P = CmpInst::FCMP_OLT;
2650 case LoopVectorizationLegality::MRK_FloatMax:
2651 P = CmpInst::FCMP_OGT;
2656 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2657 RK == LoopVectorizationLegality::MRK_FloatMax)
2658 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2660 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2662 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2667 struct CSEDenseMapInfo {
2668 static bool canHandle(Instruction *I) {
2669 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2670 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2672 static inline Instruction *getEmptyKey() {
2673 return DenseMapInfo<Instruction *>::getEmptyKey();
2675 static inline Instruction *getTombstoneKey() {
2676 return DenseMapInfo<Instruction *>::getTombstoneKey();
2678 static unsigned getHashValue(Instruction *I) {
2679 assert(canHandle(I) && "Unknown instruction!");
2680 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2681 I->value_op_end()));
2683 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2684 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2685 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2687 return LHS->isIdenticalTo(RHS);
2692 /// \brief Check whether this block is a predicated block.
2693 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2694 /// = ...; " blocks. We start with one vectorized basic block. For every
2695 /// conditional block we split this vectorized block. Therefore, every second
2696 /// block will be a predicated one.
2697 static bool isPredicatedBlock(unsigned BlockNum) {
2698 return BlockNum % 2;
2701 ///\brief Perform cse of induction variable instructions.
2702 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2703 // Perform simple cse.
2704 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2705 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2706 BasicBlock *BB = BBs[i];
2707 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2708 Instruction *In = I++;
2710 if (!CSEDenseMapInfo::canHandle(In))
2713 // Check if we can replace this instruction with any of the
2714 // visited instructions.
2715 if (Instruction *V = CSEMap.lookup(In)) {
2716 In->replaceAllUsesWith(V);
2717 In->eraseFromParent();
2720 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2721 // ...;" blocks for predicated stores. Every second block is a predicated
2723 if (isPredicatedBlock(i))
2731 /// \brief Adds a 'fast' flag to floating point operations.
2732 static Value *addFastMathFlag(Value *V) {
2733 if (isa<FPMathOperator>(V)){
2734 FastMathFlags Flags;
2735 Flags.setUnsafeAlgebra();
2736 cast<Instruction>(V)->setFastMathFlags(Flags);
2741 void InnerLoopVectorizer::vectorizeLoop() {
2742 //===------------------------------------------------===//
2744 // Notice: any optimization or new instruction that go
2745 // into the code below should be also be implemented in
2748 //===------------------------------------------------===//
2749 Constant *Zero = Builder.getInt32(0);
2751 // In order to support reduction variables we need to be able to vectorize
2752 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2753 // stages. First, we create a new vector PHI node with no incoming edges.
2754 // We use this value when we vectorize all of the instructions that use the
2755 // PHI. Next, after all of the instructions in the block are complete we
2756 // add the new incoming edges to the PHI. At this point all of the
2757 // instructions in the basic block are vectorized, so we can use them to
2758 // construct the PHI.
2759 PhiVector RdxPHIsToFix;
2761 // Scan the loop in a topological order to ensure that defs are vectorized
2763 LoopBlocksDFS DFS(OrigLoop);
2766 // Vectorize all of the blocks in the original loop.
2767 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2768 be = DFS.endRPO(); bb != be; ++bb)
2769 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2771 // At this point every instruction in the original loop is widened to
2772 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2773 // that we vectorized. The PHI nodes are currently empty because we did
2774 // not want to introduce cycles. Notice that the remaining PHI nodes
2775 // that we need to fix are reduction variables.
2777 // Create the 'reduced' values for each of the induction vars.
2778 // The reduced values are the vector values that we scalarize and combine
2779 // after the loop is finished.
2780 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2782 PHINode *RdxPhi = *it;
2783 assert(RdxPhi && "Unable to recover vectorized PHI");
2785 // Find the reduction variable descriptor.
2786 assert(Legal->getReductionVars()->count(RdxPhi) &&
2787 "Unable to find the reduction variable");
2788 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2789 (*Legal->getReductionVars())[RdxPhi];
2791 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2793 // We need to generate a reduction vector from the incoming scalar.
2794 // To do so, we need to generate the 'identity' vector and override
2795 // one of the elements with the incoming scalar reduction. We need
2796 // to do it in the vector-loop preheader.
2797 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2799 // This is the vector-clone of the value that leaves the loop.
2800 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2801 Type *VecTy = VectorExit[0]->getType();
2803 // Find the reduction identity variable. Zero for addition, or, xor,
2804 // one for multiplication, -1 for And.
2807 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2808 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2809 // MinMax reduction have the start value as their identify.
2811 VectorStart = Identity = RdxDesc.StartValue;
2813 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2818 // Handle other reduction kinds:
2820 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2821 VecTy->getScalarType());
2824 // This vector is the Identity vector where the first element is the
2825 // incoming scalar reduction.
2826 VectorStart = RdxDesc.StartValue;
2828 Identity = ConstantVector::getSplat(VF, Iden);
2830 // This vector is the Identity vector where the first element is the
2831 // incoming scalar reduction.
2832 VectorStart = Builder.CreateInsertElement(Identity,
2833 RdxDesc.StartValue, Zero);
2837 // Fix the vector-loop phi.
2838 // We created the induction variable so we know that the
2839 // preheader is the first entry.
2840 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2842 // Reductions do not have to start at zero. They can start with
2843 // any loop invariant values.
2844 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2845 BasicBlock *Latch = OrigLoop->getLoopLatch();
2846 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2847 VectorParts &Val = getVectorValue(LoopVal);
2848 for (unsigned part = 0; part < UF; ++part) {
2849 // Make sure to add the reduction stat value only to the
2850 // first unroll part.
2851 Value *StartVal = (part == 0) ? VectorStart : Identity;
2852 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2853 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2854 LoopVectorBody.back());
2857 // Before each round, move the insertion point right between
2858 // the PHIs and the values we are going to write.
2859 // This allows us to write both PHINodes and the extractelement
2861 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2863 VectorParts RdxParts;
2864 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2865 for (unsigned part = 0; part < UF; ++part) {
2866 // This PHINode contains the vectorized reduction variable, or
2867 // the initial value vector, if we bypass the vector loop.
2868 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2869 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2870 Value *StartVal = (part == 0) ? VectorStart : Identity;
2871 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2872 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2873 NewPhi->addIncoming(RdxExitVal[part],
2874 LoopVectorBody.back());
2875 RdxParts.push_back(NewPhi);
2878 // Reduce all of the unrolled parts into a single vector.
2879 Value *ReducedPartRdx = RdxParts[0];
2880 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2881 setDebugLocFromInst(Builder, ReducedPartRdx);
2882 for (unsigned part = 1; part < UF; ++part) {
2883 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2884 // Floating point operations had to be 'fast' to enable the reduction.
2885 ReducedPartRdx = addFastMathFlag(
2886 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2887 ReducedPartRdx, "bin.rdx"));
2889 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2890 ReducedPartRdx, RdxParts[part]);
2894 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2895 // and vector ops, reducing the set of values being computed by half each
2897 assert(isPowerOf2_32(VF) &&
2898 "Reduction emission only supported for pow2 vectors!");
2899 Value *TmpVec = ReducedPartRdx;
2900 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2901 for (unsigned i = VF; i != 1; i >>= 1) {
2902 // Move the upper half of the vector to the lower half.
2903 for (unsigned j = 0; j != i/2; ++j)
2904 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2906 // Fill the rest of the mask with undef.
2907 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2908 UndefValue::get(Builder.getInt32Ty()));
2911 Builder.CreateShuffleVector(TmpVec,
2912 UndefValue::get(TmpVec->getType()),
2913 ConstantVector::get(ShuffleMask),
2916 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2917 // Floating point operations had to be 'fast' to enable the reduction.
2918 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2919 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2921 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2924 // The result is in the first element of the vector.
2925 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2926 Builder.getInt32(0));
2929 // Create a phi node that merges control-flow from the backedge-taken check
2930 // block and the middle block.
2931 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2932 LoopScalarPreHeader->getTerminator());
2933 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2934 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2936 // Now, we need to fix the users of the reduction variable
2937 // inside and outside of the scalar remainder loop.
2938 // We know that the loop is in LCSSA form. We need to update the
2939 // PHI nodes in the exit blocks.
2940 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2941 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2942 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2943 if (!LCSSAPhi) break;
2945 // All PHINodes need to have a single entry edge, or two if
2946 // we already fixed them.
2947 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2949 // We found our reduction value exit-PHI. Update it with the
2950 // incoming bypass edge.
2951 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2952 // Add an edge coming from the bypass.
2953 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2956 }// end of the LCSSA phi scan.
2958 // Fix the scalar loop reduction variable with the incoming reduction sum
2959 // from the vector body and from the backedge value.
2960 int IncomingEdgeBlockIdx =
2961 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2962 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2963 // Pick the other block.
2964 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2965 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2966 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2967 }// end of for each redux variable.
2971 // Remove redundant induction instructions.
2972 cse(LoopVectorBody);
2975 void InnerLoopVectorizer::fixLCSSAPHIs() {
2976 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2977 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2978 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2979 if (!LCSSAPhi) break;
2980 if (LCSSAPhi->getNumIncomingValues() == 1)
2981 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2986 InnerLoopVectorizer::VectorParts
2987 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2988 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2991 // Look for cached value.
2992 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2993 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2994 if (ECEntryIt != MaskCache.end())
2995 return ECEntryIt->second;
2997 VectorParts SrcMask = createBlockInMask(Src);
2999 // The terminator has to be a branch inst!
3000 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3001 assert(BI && "Unexpected terminator found");
3003 if (BI->isConditional()) {
3004 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3006 if (BI->getSuccessor(0) != Dst)
3007 for (unsigned part = 0; part < UF; ++part)
3008 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3010 for (unsigned part = 0; part < UF; ++part)
3011 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3013 MaskCache[Edge] = EdgeMask;
3017 MaskCache[Edge] = SrcMask;
3021 InnerLoopVectorizer::VectorParts
3022 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3023 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3025 // Loop incoming mask is all-one.
3026 if (OrigLoop->getHeader() == BB) {
3027 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3028 return getVectorValue(C);
3031 // This is the block mask. We OR all incoming edges, and with zero.
3032 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3033 VectorParts BlockMask = getVectorValue(Zero);
3036 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3037 VectorParts EM = createEdgeMask(*it, BB);
3038 for (unsigned part = 0; part < UF; ++part)
3039 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3045 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3046 InnerLoopVectorizer::VectorParts &Entry,
3047 unsigned UF, unsigned VF, PhiVector *PV) {
3048 PHINode* P = cast<PHINode>(PN);
3049 // Handle reduction variables:
3050 if (Legal->getReductionVars()->count(P)) {
3051 for (unsigned part = 0; part < UF; ++part) {
3052 // This is phase one of vectorizing PHIs.
3053 Type *VecTy = (VF == 1) ? PN->getType() :
3054 VectorType::get(PN->getType(), VF);
3055 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3056 LoopVectorBody.back()-> getFirstInsertionPt());
3062 setDebugLocFromInst(Builder, P);
3063 // Check for PHI nodes that are lowered to vector selects.
3064 if (P->getParent() != OrigLoop->getHeader()) {
3065 // We know that all PHIs in non-header blocks are converted into
3066 // selects, so we don't have to worry about the insertion order and we
3067 // can just use the builder.
3068 // At this point we generate the predication tree. There may be
3069 // duplications since this is a simple recursive scan, but future
3070 // optimizations will clean it up.
3072 unsigned NumIncoming = P->getNumIncomingValues();
3074 // Generate a sequence of selects of the form:
3075 // SELECT(Mask3, In3,
3076 // SELECT(Mask2, In2,
3078 for (unsigned In = 0; In < NumIncoming; In++) {
3079 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3081 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3083 for (unsigned part = 0; part < UF; ++part) {
3084 // We might have single edge PHIs (blocks) - use an identity
3085 // 'select' for the first PHI operand.
3087 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3090 // Select between the current value and the previous incoming edge
3091 // based on the incoming mask.
3092 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3093 Entry[part], "predphi");
3099 // This PHINode must be an induction variable.
3100 // Make sure that we know about it.
3101 assert(Legal->getInductionVars()->count(P) &&
3102 "Not an induction variable");
3104 LoopVectorizationLegality::InductionInfo II =
3105 Legal->getInductionVars()->lookup(P);
3108 case LoopVectorizationLegality::IK_NoInduction:
3109 llvm_unreachable("Unknown induction");
3110 case LoopVectorizationLegality::IK_IntInduction: {
3111 assert(P->getType() == II.StartValue->getType() && "Types must match");
3112 Type *PhiTy = P->getType();
3114 if (P == OldInduction) {
3115 // Handle the canonical induction variable. We might have had to
3117 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3119 // Handle other induction variables that are now based on the
3121 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3123 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3124 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3127 Broadcasted = getBroadcastInstrs(Broadcasted);
3128 // After broadcasting the induction variable we need to make the vector
3129 // consecutive by adding 0, 1, 2, etc.
3130 for (unsigned part = 0; part < UF; ++part)
3131 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3134 case LoopVectorizationLegality::IK_ReverseIntInduction:
3135 case LoopVectorizationLegality::IK_PtrInduction:
3136 case LoopVectorizationLegality::IK_ReversePtrInduction:
3137 // Handle reverse integer and pointer inductions.
3138 Value *StartIdx = ExtendedIdx;
3139 // This is the normalized GEP that starts counting at zero.
3140 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3143 // Handle the reverse integer induction variable case.
3144 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3145 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3146 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3148 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3151 // This is a new value so do not hoist it out.
3152 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3153 // After broadcasting the induction variable we need to make the
3154 // vector consecutive by adding ... -3, -2, -1, 0.
3155 for (unsigned part = 0; part < UF; ++part)
3156 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3161 // Handle the pointer induction variable case.
3162 assert(P->getType()->isPointerTy() && "Unexpected type.");
3164 // Is this a reverse induction ptr or a consecutive induction ptr.
3165 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3168 // This is the vector of results. Notice that we don't generate
3169 // vector geps because scalar geps result in better code.
3170 for (unsigned part = 0; part < UF; ++part) {
3172 int EltIndex = (part) * (Reverse ? -1 : 1);
3173 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3176 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3178 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3180 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3182 Entry[part] = SclrGep;
3186 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3187 for (unsigned int i = 0; i < VF; ++i) {
3188 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3189 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3192 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3194 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3196 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3198 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3199 Builder.getInt32(i),
3202 Entry[part] = VecVal;
3208 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3209 // For each instruction in the old loop.
3210 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3211 VectorParts &Entry = WidenMap.get(it);
3212 switch (it->getOpcode()) {
3213 case Instruction::Br:
3214 // Nothing to do for PHIs and BR, since we already took care of the
3215 // loop control flow instructions.
3217 case Instruction::PHI:{
3218 // Vectorize PHINodes.
3219 widenPHIInstruction(it, Entry, UF, VF, PV);
3223 case Instruction::Add:
3224 case Instruction::FAdd:
3225 case Instruction::Sub:
3226 case Instruction::FSub:
3227 case Instruction::Mul:
3228 case Instruction::FMul:
3229 case Instruction::UDiv:
3230 case Instruction::SDiv:
3231 case Instruction::FDiv:
3232 case Instruction::URem:
3233 case Instruction::SRem:
3234 case Instruction::FRem:
3235 case Instruction::Shl:
3236 case Instruction::LShr:
3237 case Instruction::AShr:
3238 case Instruction::And:
3239 case Instruction::Or:
3240 case Instruction::Xor: {
3241 // Just widen binops.
3242 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3243 setDebugLocFromInst(Builder, BinOp);
3244 VectorParts &A = getVectorValue(it->getOperand(0));
3245 VectorParts &B = getVectorValue(it->getOperand(1));
3247 // Use this vector value for all users of the original instruction.
3248 for (unsigned Part = 0; Part < UF; ++Part) {
3249 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3251 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3252 VecOp->copyIRFlags(BinOp);
3257 propagateMetadata(Entry, it);
3260 case Instruction::Select: {
3262 // If the selector is loop invariant we can create a select
3263 // instruction with a scalar condition. Otherwise, use vector-select.
3264 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3266 setDebugLocFromInst(Builder, it);
3268 // The condition can be loop invariant but still defined inside the
3269 // loop. This means that we can't just use the original 'cond' value.
3270 // We have to take the 'vectorized' value and pick the first lane.
3271 // Instcombine will make this a no-op.
3272 VectorParts &Cond = getVectorValue(it->getOperand(0));
3273 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3274 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3276 Value *ScalarCond = (VF == 1) ? Cond[0] :
3277 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3279 for (unsigned Part = 0; Part < UF; ++Part) {
3280 Entry[Part] = Builder.CreateSelect(
3281 InvariantCond ? ScalarCond : Cond[Part],
3286 propagateMetadata(Entry, it);
3290 case Instruction::ICmp:
3291 case Instruction::FCmp: {
3292 // Widen compares. Generate vector compares.
3293 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3294 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3295 setDebugLocFromInst(Builder, it);
3296 VectorParts &A = getVectorValue(it->getOperand(0));
3297 VectorParts &B = getVectorValue(it->getOperand(1));
3298 for (unsigned Part = 0; Part < UF; ++Part) {
3301 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3303 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3307 propagateMetadata(Entry, it);
3311 case Instruction::Store:
3312 case Instruction::Load:
3313 vectorizeMemoryInstruction(it);
3315 case Instruction::ZExt:
3316 case Instruction::SExt:
3317 case Instruction::FPToUI:
3318 case Instruction::FPToSI:
3319 case Instruction::FPExt:
3320 case Instruction::PtrToInt:
3321 case Instruction::IntToPtr:
3322 case Instruction::SIToFP:
3323 case Instruction::UIToFP:
3324 case Instruction::Trunc:
3325 case Instruction::FPTrunc:
3326 case Instruction::BitCast: {
3327 CastInst *CI = dyn_cast<CastInst>(it);
3328 setDebugLocFromInst(Builder, it);
3329 /// Optimize the special case where the source is the induction
3330 /// variable. Notice that we can only optimize the 'trunc' case
3331 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3332 /// c. other casts depend on pointer size.
3333 if (CI->getOperand(0) == OldInduction &&
3334 it->getOpcode() == Instruction::Trunc) {
3335 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3337 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3338 for (unsigned Part = 0; Part < UF; ++Part)
3339 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3340 propagateMetadata(Entry, it);
3343 /// Vectorize casts.
3344 Type *DestTy = (VF == 1) ? CI->getType() :
3345 VectorType::get(CI->getType(), VF);
3347 VectorParts &A = getVectorValue(it->getOperand(0));
3348 for (unsigned Part = 0; Part < UF; ++Part)
3349 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3350 propagateMetadata(Entry, it);
3354 case Instruction::Call: {
3355 // Ignore dbg intrinsics.
3356 if (isa<DbgInfoIntrinsic>(it))
3358 setDebugLocFromInst(Builder, it);
3360 Module *M = BB->getParent()->getParent();
3361 CallInst *CI = cast<CallInst>(it);
3362 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3363 assert(ID && "Not an intrinsic call!");
3365 case Intrinsic::lifetime_end:
3366 case Intrinsic::lifetime_start:
3367 scalarizeInstruction(it);
3370 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3371 for (unsigned Part = 0; Part < UF; ++Part) {
3372 SmallVector<Value *, 4> Args;
3373 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3374 if (HasScalarOpd && i == 1) {
3375 Args.push_back(CI->getArgOperand(i));
3378 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3379 Args.push_back(Arg[Part]);
3381 Type *Tys[] = {CI->getType()};
3383 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3385 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3386 Entry[Part] = Builder.CreateCall(F, Args);
3389 propagateMetadata(Entry, it);
3396 // All other instructions are unsupported. Scalarize them.
3397 scalarizeInstruction(it);
3400 }// end of for_each instr.
3403 void InnerLoopVectorizer::updateAnalysis() {
3404 // Forget the original basic block.
3405 SE->forgetLoop(OrigLoop);
3407 // Update the dominator tree information.
3408 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3409 "Entry does not dominate exit.");
3411 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3412 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3413 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3415 // Due to if predication of stores we might create a sequence of "if(pred)
3416 // a[i] = ...; " blocks.
3417 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3419 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3420 else if (isPredicatedBlock(i)) {
3421 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3423 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3427 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3428 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3429 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3430 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3432 DEBUG(DT->verifyDomTree());
3435 /// \brief Check whether it is safe to if-convert this phi node.
3437 /// Phi nodes with constant expressions that can trap are not safe to if
3439 static bool canIfConvertPHINodes(BasicBlock *BB) {
3440 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3441 PHINode *Phi = dyn_cast<PHINode>(I);
3444 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3445 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3452 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3453 if (!EnableIfConversion) {
3454 emitAnalysis(Report() << "if-conversion is disabled");
3458 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3460 // A list of pointers that we can safely read and write to.
3461 SmallPtrSet<Value *, 8> SafePointes;
3463 // Collect safe addresses.
3464 for (Loop::block_iterator BI = TheLoop->block_begin(),
3465 BE = TheLoop->block_end(); BI != BE; ++BI) {
3466 BasicBlock *BB = *BI;
3468 if (blockNeedsPredication(BB))
3471 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3472 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3473 SafePointes.insert(LI->getPointerOperand());
3474 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3475 SafePointes.insert(SI->getPointerOperand());
3479 // Collect the blocks that need predication.
3480 BasicBlock *Header = TheLoop->getHeader();
3481 for (Loop::block_iterator BI = TheLoop->block_begin(),
3482 BE = TheLoop->block_end(); BI != BE; ++BI) {
3483 BasicBlock *BB = *BI;
3485 // We don't support switch statements inside loops.
3486 if (!isa<BranchInst>(BB->getTerminator())) {
3487 emitAnalysis(Report(BB->getTerminator())
3488 << "loop contains a switch statement");
3492 // We must be able to predicate all blocks that need to be predicated.
3493 if (blockNeedsPredication(BB)) {
3494 if (!blockCanBePredicated(BB, SafePointes)) {
3495 emitAnalysis(Report(BB->getTerminator())
3496 << "control flow cannot be substituted for a select");
3499 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3500 emitAnalysis(Report(BB->getTerminator())
3501 << "control flow cannot be substituted for a select");
3506 // We can if-convert this loop.
3510 bool LoopVectorizationLegality::canVectorize() {
3511 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3512 // be canonicalized.
3513 if (!TheLoop->getLoopPreheader()) {
3515 Report() << "loop control flow is not understood by vectorizer");
3519 // We can only vectorize innermost loops.
3520 if (TheLoop->getSubLoopsVector().size()) {
3521 emitAnalysis(Report() << "loop is not the innermost loop");
3525 // We must have a single backedge.
3526 if (TheLoop->getNumBackEdges() != 1) {
3528 Report() << "loop control flow is not understood by vectorizer");
3532 // We must have a single exiting block.
3533 if (!TheLoop->getExitingBlock()) {
3535 Report() << "loop control flow is not understood by vectorizer");
3539 // We need to have a loop header.
3540 DEBUG(dbgs() << "LV: Found a loop: " <<
3541 TheLoop->getHeader()->getName() << '\n');
3543 // Check if we can if-convert non-single-bb loops.
3544 unsigned NumBlocks = TheLoop->getNumBlocks();
3545 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3546 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3550 // ScalarEvolution needs to be able to find the exit count.
3551 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3552 if (ExitCount == SE->getCouldNotCompute()) {
3553 emitAnalysis(Report() << "could not determine number of loop iterations");
3554 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3558 // Check if we can vectorize the instructions and CFG in this loop.
3559 if (!canVectorizeInstrs()) {
3560 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3564 // Go over each instruction and look at memory deps.
3565 if (!canVectorizeMemory()) {
3566 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3570 // Collect all of the variables that remain uniform after vectorization.
3571 collectLoopUniforms();
3573 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3574 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3577 // Okay! We can vectorize. At this point we don't have any other mem analysis
3578 // which may limit our maximum vectorization factor, so just return true with
3583 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3584 if (Ty->isPointerTy())
3585 return DL.getIntPtrType(Ty);
3587 // It is possible that char's or short's overflow when we ask for the loop's
3588 // trip count, work around this by changing the type size.
3589 if (Ty->getScalarSizeInBits() < 32)
3590 return Type::getInt32Ty(Ty->getContext());
3595 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3596 Ty0 = convertPointerToIntegerType(DL, Ty0);
3597 Ty1 = convertPointerToIntegerType(DL, Ty1);
3598 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3603 /// \brief Check that the instruction has outside loop users and is not an
3604 /// identified reduction variable.
3605 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3606 SmallPtrSetImpl<Value *> &Reductions) {
3607 // Reduction instructions are allowed to have exit users. All other
3608 // instructions must not have external users.
3609 if (!Reductions.count(Inst))
3610 //Check that all of the users of the loop are inside the BB.
3611 for (User *U : Inst->users()) {
3612 Instruction *UI = cast<Instruction>(U);
3613 // This user may be a reduction exit value.
3614 if (!TheLoop->contains(UI)) {
3615 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3622 bool LoopVectorizationLegality::canVectorizeInstrs() {
3623 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3624 BasicBlock *Header = TheLoop->getHeader();
3626 // Look for the attribute signaling the absence of NaNs.
3627 Function &F = *Header->getParent();
3628 if (F.hasFnAttribute("no-nans-fp-math"))
3629 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3630 AttributeSet::FunctionIndex,
3631 "no-nans-fp-math").getValueAsString() == "true";
3633 // For each block in the loop.
3634 for (Loop::block_iterator bb = TheLoop->block_begin(),
3635 be = TheLoop->block_end(); bb != be; ++bb) {
3637 // Scan the instructions in the block and look for hazards.
3638 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3641 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3642 Type *PhiTy = Phi->getType();
3643 // Check that this PHI type is allowed.
3644 if (!PhiTy->isIntegerTy() &&
3645 !PhiTy->isFloatingPointTy() &&
3646 !PhiTy->isPointerTy()) {
3647 emitAnalysis(Report(it)
3648 << "loop control flow is not understood by vectorizer");
3649 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3653 // If this PHINode is not in the header block, then we know that we
3654 // can convert it to select during if-conversion. No need to check if
3655 // the PHIs in this block are induction or reduction variables.
3656 if (*bb != Header) {
3657 // Check that this instruction has no outside users or is an
3658 // identified reduction value with an outside user.
3659 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3661 emitAnalysis(Report(it) << "value could not be identified as "
3662 "an induction or reduction variable");
3666 // We only allow if-converted PHIs with more than two incoming values.
3667 if (Phi->getNumIncomingValues() != 2) {
3668 emitAnalysis(Report(it)
3669 << "control flow not understood by vectorizer");
3670 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3674 // This is the value coming from the preheader.
3675 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3676 // Check if this is an induction variable.
3677 InductionKind IK = isInductionVariable(Phi);
3679 if (IK_NoInduction != IK) {
3680 // Get the widest type.
3682 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3684 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3686 // Int inductions are special because we only allow one IV.
3687 if (IK == IK_IntInduction) {
3688 // Use the phi node with the widest type as induction. Use the last
3689 // one if there are multiple (no good reason for doing this other
3690 // than it is expedient).
3691 if (!Induction || PhiTy == WidestIndTy)
3695 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3696 Inductions[Phi] = InductionInfo(StartValue, IK);
3698 // Until we explicitly handle the case of an induction variable with
3699 // an outside loop user we have to give up vectorizing this loop.
3700 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3701 emitAnalysis(Report(it) << "use of induction value outside of the "
3702 "loop is not handled by vectorizer");
3709 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3710 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3713 if (AddReductionVar(Phi, RK_IntegerMult)) {
3714 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3717 if (AddReductionVar(Phi, RK_IntegerOr)) {
3718 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3721 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3722 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3725 if (AddReductionVar(Phi, RK_IntegerXor)) {
3726 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3729 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3730 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3733 if (AddReductionVar(Phi, RK_FloatMult)) {
3734 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3737 if (AddReductionVar(Phi, RK_FloatAdd)) {
3738 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3741 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3742 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3747 emitAnalysis(Report(it) << "value that could not be identified as "
3748 "reduction is used outside the loop");
3749 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3751 }// end of PHI handling
3753 // We still don't handle functions. However, we can ignore dbg intrinsic
3754 // calls and we do handle certain intrinsic and libm functions.
3755 CallInst *CI = dyn_cast<CallInst>(it);
3756 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3757 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3758 DEBUG(dbgs() << "LV: Found a call site.\n");
3762 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3763 // second argument is the same (i.e. loop invariant)
3765 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3766 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3767 emitAnalysis(Report(it)
3768 << "intrinsic instruction cannot be vectorized");
3769 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3774 // Check that the instruction return type is vectorizable.
3775 // Also, we can't vectorize extractelement instructions.
3776 if ((!VectorType::isValidElementType(it->getType()) &&
3777 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3778 emitAnalysis(Report(it)
3779 << "instruction return type cannot be vectorized");
3780 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3784 // Check that the stored type is vectorizable.
3785 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3786 Type *T = ST->getValueOperand()->getType();
3787 if (!VectorType::isValidElementType(T)) {
3788 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3791 if (EnableMemAccessVersioning)
3792 collectStridedAcccess(ST);
3795 if (EnableMemAccessVersioning)
3796 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3797 collectStridedAcccess(LI);
3799 // Reduction instructions are allowed to have exit users.
3800 // All other instructions must not have external users.
3801 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3802 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3811 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3812 if (Inductions.empty()) {
3813 emitAnalysis(Report()
3814 << "loop induction variable could not be identified");
3822 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3823 /// return the induction operand of the gep pointer.
3824 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3825 const DataLayout *DL, Loop *Lp) {
3826 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3830 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3832 // Check that all of the gep indices are uniform except for our induction
3834 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3835 if (i != InductionOperand &&
3836 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3838 return GEP->getOperand(InductionOperand);
3841 ///\brief Look for a cast use of the passed value.
3842 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3843 Value *UniqueCast = nullptr;
3844 for (User *U : Ptr->users()) {
3845 CastInst *CI = dyn_cast<CastInst>(U);
3846 if (CI && CI->getType() == Ty) {
3856 ///\brief Get the stride of a pointer access in a loop.
3857 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3858 /// pointer to the Value, or null otherwise.
3859 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3860 const DataLayout *DL, Loop *Lp) {
3861 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3862 if (!PtrTy || PtrTy->isAggregateType())
3865 // Try to remove a gep instruction to make the pointer (actually index at this
3866 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3867 // pointer, otherwise, we are analyzing the index.
3868 Value *OrigPtr = Ptr;
3870 // The size of the pointer access.
3871 int64_t PtrAccessSize = 1;
3873 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3874 const SCEV *V = SE->getSCEV(Ptr);
3878 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3879 V = C->getOperand();
3881 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3885 V = S->getStepRecurrence(*SE);
3889 // Strip off the size of access multiplication if we are still analyzing the
3891 if (OrigPtr == Ptr) {
3892 DL->getTypeAllocSize(PtrTy->getElementType());
3893 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3894 if (M->getOperand(0)->getSCEVType() != scConstant)
3897 const APInt &APStepVal =
3898 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3900 // Huge step value - give up.
3901 if (APStepVal.getBitWidth() > 64)
3904 int64_t StepVal = APStepVal.getSExtValue();
3905 if (PtrAccessSize != StepVal)
3907 V = M->getOperand(1);
3912 Type *StripedOffRecurrenceCast = nullptr;
3913 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3914 StripedOffRecurrenceCast = C->getType();
3915 V = C->getOperand();
3918 // Look for the loop invariant symbolic value.
3919 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3923 Value *Stride = U->getValue();
3924 if (!Lp->isLoopInvariant(Stride))
3927 // If we have stripped off the recurrence cast we have to make sure that we
3928 // return the value that is used in this loop so that we can replace it later.
3929 if (StripedOffRecurrenceCast)
3930 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3935 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3936 Value *Ptr = nullptr;
3937 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3938 Ptr = LI->getPointerOperand();
3939 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3940 Ptr = SI->getPointerOperand();
3944 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3948 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3949 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3950 Strides[Ptr] = Stride;
3951 StrideSet.insert(Stride);
3954 void LoopVectorizationLegality::collectLoopUniforms() {
3955 // We now know that the loop is vectorizable!
3956 // Collect variables that will remain uniform after vectorization.
3957 std::vector<Value*> Worklist;
3958 BasicBlock *Latch = TheLoop->getLoopLatch();
3960 // Start with the conditional branch and walk up the block.
3961 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3963 // Also add all consecutive pointer values; these values will be uniform
3964 // after vectorization (and subsequent cleanup) and, until revectorization is
3965 // supported, all dependencies must also be uniform.
3966 for (Loop::block_iterator B = TheLoop->block_begin(),
3967 BE = TheLoop->block_end(); B != BE; ++B)
3968 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3970 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3971 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3973 while (Worklist.size()) {
3974 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3975 Worklist.pop_back();
3977 // Look at instructions inside this loop.
3978 // Stop when reaching PHI nodes.
3979 // TODO: we need to follow values all over the loop, not only in this block.
3980 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3983 // This is a known uniform.
3986 // Insert all operands.
3987 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3992 /// \brief Analyses memory accesses in a loop.
3994 /// Checks whether run time pointer checks are needed and builds sets for data
3995 /// dependence checking.
3996 class AccessAnalysis {
3998 /// \brief Read or write access location.
3999 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4000 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4002 /// \brief Set of potential dependent memory accesses.
4003 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4005 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4006 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4008 /// \brief Register a load and whether it is only read from.
4009 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4010 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4011 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4012 Accesses.insert(MemAccessInfo(Ptr, false));
4014 ReadOnlyPtr.insert(Ptr);
4017 /// \brief Register a store.
4018 void addStore(AliasAnalysis::Location &Loc) {
4019 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4020 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4021 Accesses.insert(MemAccessInfo(Ptr, true));
4024 /// \brief Check whether we can check the pointers at runtime for
4025 /// non-intersection.
4026 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4027 unsigned &NumComparisons, ScalarEvolution *SE,
4028 Loop *TheLoop, ValueToValueMap &Strides,
4029 bool ShouldCheckStride = false);
4031 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4032 /// and builds sets of dependent accesses.
4033 void buildDependenceSets() {
4034 processMemAccesses();
4037 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4039 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4040 void resetDepChecks() { CheckDeps.clear(); }
4042 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4045 typedef SetVector<MemAccessInfo> PtrAccessSet;
4047 /// \brief Go over all memory access and check whether runtime pointer checks
4048 /// are needed /// and build sets of dependency check candidates.
4049 void processMemAccesses();
4051 /// Set of all accesses.
4052 PtrAccessSet Accesses;
4054 /// Set of accesses that need a further dependence check.
4055 MemAccessInfoSet CheckDeps;
4057 /// Set of pointers that are read only.
4058 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4060 const DataLayout *DL;
4062 /// An alias set tracker to partition the access set by underlying object and
4063 //intrinsic property (such as TBAA metadata).
4064 AliasSetTracker AST;
4066 /// Sets of potentially dependent accesses - members of one set share an
4067 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4068 /// dependence check.
4069 DepCandidates &DepCands;
4071 bool IsRTCheckNeeded;
4074 } // end anonymous namespace
4076 /// \brief Check whether a pointer can participate in a runtime bounds check.
4077 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4079 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4080 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4084 return AR->isAffine();
4087 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4088 /// the address space.
4089 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4090 const Loop *Lp, ValueToValueMap &StridesMap);
4092 bool AccessAnalysis::canCheckPtrAtRT(
4093 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4094 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4095 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4096 // Find pointers with computable bounds. We are going to use this information
4097 // to place a runtime bound check.
4098 bool CanDoRT = true;
4100 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4103 // We assign a consecutive id to access from different alias sets.
4104 // Accesses between different groups doesn't need to be checked.
4106 for (auto &AS : AST) {
4107 unsigned NumReadPtrChecks = 0;
4108 unsigned NumWritePtrChecks = 0;
4110 // We assign consecutive id to access from different dependence sets.
4111 // Accesses within the same set don't need a runtime check.
4112 unsigned RunningDepId = 1;
4113 DenseMap<Value *, unsigned> DepSetId;
4116 Value *Ptr = A.getValue();
4117 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4118 MemAccessInfo Access(Ptr, IsWrite);
4121 ++NumWritePtrChecks;
4125 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4126 // When we run after a failing dependency check we have to make sure we
4127 // don't have wrapping pointers.
4128 (!ShouldCheckStride ||
4129 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4130 // The id of the dependence set.
4133 if (IsDepCheckNeeded) {
4134 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4135 unsigned &LeaderId = DepSetId[Leader];
4137 LeaderId = RunningDepId++;
4140 // Each access has its own dependence set.
4141 DepId = RunningDepId++;
4143 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4145 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4151 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4152 NumComparisons += 0; // Only one dependence set.
4154 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4155 NumWritePtrChecks - 1));
4161 // If the pointers that we would use for the bounds comparison have different
4162 // address spaces, assume the values aren't directly comparable, so we can't
4163 // use them for the runtime check. We also have to assume they could
4164 // overlap. In the future there should be metadata for whether address spaces
4166 unsigned NumPointers = RtCheck.Pointers.size();
4167 for (unsigned i = 0; i < NumPointers; ++i) {
4168 for (unsigned j = i + 1; j < NumPointers; ++j) {
4169 // Only need to check pointers between two different dependency sets.
4170 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4172 // Only need to check pointers in the same alias set.
4173 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4176 Value *PtrI = RtCheck.Pointers[i];
4177 Value *PtrJ = RtCheck.Pointers[j];
4179 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4180 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4182 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4183 " different address spaces\n");
4192 void AccessAnalysis::processMemAccesses() {
4193 // We process the set twice: first we process read-write pointers, last we
4194 // process read-only pointers. This allows us to skip dependence tests for
4195 // read-only pointers.
4197 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4198 DEBUG(dbgs() << " AST: "; AST.dump());
4199 DEBUG(dbgs() << "LV: Accesses:\n");
4201 for (auto A : Accesses)
4202 dbgs() << "\t" << *A.getPointer() << " (" <<
4203 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4204 "read-only" : "read")) << ")\n";
4207 // The AliasSetTracker has nicely partitioned our pointers by metadata
4208 // compatibility and potential for underlying-object overlap. As a result, we
4209 // only need to check for potential pointer dependencies within each alias
4211 for (auto &AS : AST) {
4212 // Note that both the alias-set tracker and the alias sets themselves used
4213 // linked lists internally and so the iteration order here is deterministic
4214 // (matching the original instruction order within each set).
4216 bool SetHasWrite = false;
4218 // Map of pointers to last access encountered.
4219 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4220 UnderlyingObjToAccessMap ObjToLastAccess;
4222 // Set of access to check after all writes have been processed.
4223 PtrAccessSet DeferredAccesses;
4225 // Iterate over each alias set twice, once to process read/write pointers,
4226 // and then to process read-only pointers.
4227 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4228 bool UseDeferred = SetIteration > 0;
4229 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4232 Value *Ptr = A.getValue();
4233 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4235 // If we're using the deferred access set, then it contains only reads.
4236 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4237 if (UseDeferred && !IsReadOnlyPtr)
4239 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4241 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4242 S.count(MemAccessInfo(Ptr, false))) &&
4243 "Alias-set pointer not in the access set?");
4245 MemAccessInfo Access(Ptr, IsWrite);
4246 DepCands.insert(Access);
4248 // Memorize read-only pointers for later processing and skip them in the
4249 // first round (they need to be checked after we have seen all write
4250 // pointers). Note: we also mark pointer that are not consecutive as
4251 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4252 // the second check for "!IsWrite".
4253 if (!UseDeferred && IsReadOnlyPtr) {
4254 DeferredAccesses.insert(Access);
4258 // If this is a write - check other reads and writes for conflicts. If
4259 // this is a read only check other writes for conflicts (but only if
4260 // there is no other write to the ptr - this is an optimization to
4261 // catch "a[i] = a[i] + " without having to do a dependence check).
4262 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4263 CheckDeps.insert(Access);
4264 IsRTCheckNeeded = true;
4270 // Create sets of pointers connected by a shared alias set and
4271 // underlying object.
4272 typedef SmallVector<Value*, 16> ValueVector;
4273 ValueVector TempObjects;
4274 GetUnderlyingObjects(Ptr, TempObjects, DL);
4275 for (Value *UnderlyingObj : TempObjects) {
4276 UnderlyingObjToAccessMap::iterator Prev =
4277 ObjToLastAccess.find(UnderlyingObj);
4278 if (Prev != ObjToLastAccess.end())
4279 DepCands.unionSets(Access, Prev->second);
4281 ObjToLastAccess[UnderlyingObj] = Access;
4289 /// \brief Checks memory dependences among accesses to the same underlying
4290 /// object to determine whether there vectorization is legal or not (and at
4291 /// which vectorization factor).
4293 /// This class works under the assumption that we already checked that memory
4294 /// locations with different underlying pointers are "must-not alias".
4295 /// We use the ScalarEvolution framework to symbolically evalutate access
4296 /// functions pairs. Since we currently don't restructure the loop we can rely
4297 /// on the program order of memory accesses to determine their safety.
4298 /// At the moment we will only deem accesses as safe for:
4299 /// * A negative constant distance assuming program order.
4301 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4302 /// a[i] = tmp; y = a[i];
4304 /// The latter case is safe because later checks guarantuee that there can't
4305 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4306 /// the same variable: a header phi can only be an induction or a reduction, a
4307 /// reduction can't have a memory sink, an induction can't have a memory
4308 /// source). This is important and must not be violated (or we have to
4309 /// resort to checking for cycles through memory).
4311 /// * A positive constant distance assuming program order that is bigger
4312 /// than the biggest memory access.
4314 /// tmp = a[i] OR b[i] = x
4315 /// a[i+2] = tmp y = b[i+2];
4317 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4319 /// * Zero distances and all accesses have the same size.
4321 class MemoryDepChecker {
4323 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4324 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4326 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4327 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4328 ShouldRetryWithRuntimeCheck(false) {}
4330 /// \brief Register the location (instructions are given increasing numbers)
4331 /// of a write access.
4332 void addAccess(StoreInst *SI) {
4333 Value *Ptr = SI->getPointerOperand();
4334 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4335 InstMap.push_back(SI);
4339 /// \brief Register the location (instructions are given increasing numbers)
4340 /// of a write access.
4341 void addAccess(LoadInst *LI) {
4342 Value *Ptr = LI->getPointerOperand();
4343 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4344 InstMap.push_back(LI);
4348 /// \brief Check whether the dependencies between the accesses are safe.
4350 /// Only checks sets with elements in \p CheckDeps.
4351 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4352 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4354 /// \brief The maximum number of bytes of a vector register we can vectorize
4355 /// the accesses safely with.
4356 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4358 /// \brief In same cases when the dependency check fails we can still
4359 /// vectorize the loop with a dynamic array access check.
4360 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4363 ScalarEvolution *SE;
4364 const DataLayout *DL;
4365 const Loop *InnermostLoop;
4367 /// \brief Maps access locations (ptr, read/write) to program order.
4368 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4370 /// \brief Memory access instructions in program order.
4371 SmallVector<Instruction *, 16> InstMap;
4373 /// \brief The program order index to be used for the next instruction.
4376 // We can access this many bytes in parallel safely.
4377 unsigned MaxSafeDepDistBytes;
4379 /// \brief If we see a non-constant dependence distance we can still try to
4380 /// vectorize this loop with runtime checks.
4381 bool ShouldRetryWithRuntimeCheck;
4383 /// \brief Check whether there is a plausible dependence between the two
4386 /// Access \p A must happen before \p B in program order. The two indices
4387 /// identify the index into the program order map.
4389 /// This function checks whether there is a plausible dependence (or the
4390 /// absence of such can't be proved) between the two accesses. If there is a
4391 /// plausible dependence but the dependence distance is bigger than one
4392 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4393 /// distance is smaller than any other distance encountered so far).
4394 /// Otherwise, this function returns true signaling a possible dependence.
4395 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4396 const MemAccessInfo &B, unsigned BIdx,
4397 ValueToValueMap &Strides);
4399 /// \brief Check whether the data dependence could prevent store-load
4401 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4404 } // end anonymous namespace
4406 static bool isInBoundsGep(Value *Ptr) {
4407 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4408 return GEP->isInBounds();
4412 /// \brief Check whether the access through \p Ptr has a constant stride.
4413 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4414 const Loop *Lp, ValueToValueMap &StridesMap) {
4415 const Type *Ty = Ptr->getType();
4416 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4418 // Make sure that the pointer does not point to aggregate types.
4419 const PointerType *PtrTy = cast<PointerType>(Ty);
4420 if (PtrTy->getElementType()->isAggregateType()) {
4421 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4426 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4428 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4430 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4431 << *Ptr << " SCEV: " << *PtrScev << "\n");
4435 // The accesss function must stride over the innermost loop.
4436 if (Lp != AR->getLoop()) {
4437 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4438 *Ptr << " SCEV: " << *PtrScev << "\n");
4441 // The address calculation must not wrap. Otherwise, a dependence could be
4443 // An inbounds getelementptr that is a AddRec with a unit stride
4444 // cannot wrap per definition. The unit stride requirement is checked later.
4445 // An getelementptr without an inbounds attribute and unit stride would have
4446 // to access the pointer value "0" which is undefined behavior in address
4447 // space 0, therefore we can also vectorize this case.
4448 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4449 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4450 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4451 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4452 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4453 << *Ptr << " SCEV: " << *PtrScev << "\n");
4457 // Check the step is constant.
4458 const SCEV *Step = AR->getStepRecurrence(*SE);
4460 // Calculate the pointer stride and check if it is consecutive.
4461 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4463 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4464 " SCEV: " << *PtrScev << "\n");
4468 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4469 const APInt &APStepVal = C->getValue()->getValue();
4471 // Huge step value - give up.
4472 if (APStepVal.getBitWidth() > 64)
4475 int64_t StepVal = APStepVal.getSExtValue();
4478 int64_t Stride = StepVal / Size;
4479 int64_t Rem = StepVal % Size;
4483 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4484 // know we can't "wrap around the address space". In case of address space
4485 // zero we know that this won't happen without triggering undefined behavior.
4486 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4487 Stride != 1 && Stride != -1)
4493 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4494 unsigned TypeByteSize) {
4495 // If loads occur at a distance that is not a multiple of a feasible vector
4496 // factor store-load forwarding does not take place.
4497 // Positive dependences might cause troubles because vectorizing them might
4498 // prevent store-load forwarding making vectorized code run a lot slower.
4499 // a[i] = a[i-3] ^ a[i-8];
4500 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4501 // hence on your typical architecture store-load forwarding does not take
4502 // place. Vectorizing in such cases does not make sense.
4503 // Store-load forwarding distance.
4504 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4505 // Maximum vector factor.
4506 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4507 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4508 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4510 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4512 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4513 MaxVFWithoutSLForwardIssues = (vf >>=1);
4518 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4519 DEBUG(dbgs() << "LV: Distance " << Distance <<
4520 " that could cause a store-load forwarding conflict\n");
4524 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4525 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4526 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4530 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4531 const MemAccessInfo &B, unsigned BIdx,
4532 ValueToValueMap &Strides) {
4533 assert (AIdx < BIdx && "Must pass arguments in program order");
4535 Value *APtr = A.getPointer();
4536 Value *BPtr = B.getPointer();
4537 bool AIsWrite = A.getInt();
4538 bool BIsWrite = B.getInt();
4540 // Two reads are independent.
4541 if (!AIsWrite && !BIsWrite)
4544 // We cannot check pointers in different address spaces.
4545 if (APtr->getType()->getPointerAddressSpace() !=
4546 BPtr->getType()->getPointerAddressSpace())
4549 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4550 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4552 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4553 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4555 const SCEV *Src = AScev;
4556 const SCEV *Sink = BScev;
4558 // If the induction step is negative we have to invert source and sink of the
4560 if (StrideAPtr < 0) {
4563 std::swap(APtr, BPtr);
4564 std::swap(Src, Sink);
4565 std::swap(AIsWrite, BIsWrite);
4566 std::swap(AIdx, BIdx);
4567 std::swap(StrideAPtr, StrideBPtr);
4570 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4572 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4573 << "(Induction step: " << StrideAPtr << ")\n");
4574 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4575 << *InstMap[BIdx] << ": " << *Dist << "\n");
4577 // Need consecutive accesses. We don't want to vectorize
4578 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4579 // the address space.
4580 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4581 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4585 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4587 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4588 ShouldRetryWithRuntimeCheck = true;
4592 Type *ATy = APtr->getType()->getPointerElementType();
4593 Type *BTy = BPtr->getType()->getPointerElementType();
4594 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4596 // Negative distances are not plausible dependencies.
4597 const APInt &Val = C->getValue()->getValue();
4598 if (Val.isNegative()) {
4599 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4600 if (IsTrueDataDependence &&
4601 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4605 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4609 // Write to the same location with the same size.
4610 // Could be improved to assert type sizes are the same (i32 == float, etc).
4614 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4618 assert(Val.isStrictlyPositive() && "Expect a positive value");
4620 // Positive distance bigger than max vectorization factor.
4623 "LV: ReadWrite-Write positive dependency with different types\n");
4627 unsigned Distance = (unsigned) Val.getZExtValue();
4629 // Bail out early if passed-in parameters make vectorization not feasible.
4630 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4631 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4633 // The distance must be bigger than the size needed for a vectorized version
4634 // of the operation and the size of the vectorized operation must not be
4635 // bigger than the currrent maximum size.
4636 if (Distance < 2*TypeByteSize ||
4637 2*TypeByteSize > MaxSafeDepDistBytes ||
4638 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4639 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4640 << Val.getSExtValue() << '\n');
4644 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4645 Distance : MaxSafeDepDistBytes;
4647 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4648 if (IsTrueDataDependence &&
4649 couldPreventStoreLoadForward(Distance, TypeByteSize))
4652 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4653 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4658 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4659 MemAccessInfoSet &CheckDeps,
4660 ValueToValueMap &Strides) {
4662 MaxSafeDepDistBytes = -1U;
4663 while (!CheckDeps.empty()) {
4664 MemAccessInfo CurAccess = *CheckDeps.begin();
4666 // Get the relevant memory access set.
4667 EquivalenceClasses<MemAccessInfo>::iterator I =
4668 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4670 // Check accesses within this set.
4671 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4672 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4674 // Check every access pair.
4676 CheckDeps.erase(*AI);
4677 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4679 // Check every accessing instruction pair in program order.
4680 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4681 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4682 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4683 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4684 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4686 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4697 bool LoopVectorizationLegality::canVectorizeMemory() {
4699 typedef SmallVector<Value*, 16> ValueVector;
4700 typedef SmallPtrSet<Value*, 16> ValueSet;
4702 // Holds the Load and Store *instructions*.
4706 // Holds all the different accesses in the loop.
4707 unsigned NumReads = 0;
4708 unsigned NumReadWrites = 0;
4710 PtrRtCheck.Pointers.clear();
4711 PtrRtCheck.Need = false;
4713 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4714 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4717 for (Loop::block_iterator bb = TheLoop->block_begin(),
4718 be = TheLoop->block_end(); bb != be; ++bb) {
4720 // Scan the BB and collect legal loads and stores.
4721 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4724 // If this is a load, save it. If this instruction can read from memory
4725 // but is not a load, then we quit. Notice that we don't handle function
4726 // calls that read or write.
4727 if (it->mayReadFromMemory()) {
4728 // Many math library functions read the rounding mode. We will only
4729 // vectorize a loop if it contains known function calls that don't set
4730 // the flag. Therefore, it is safe to ignore this read from memory.
4731 CallInst *Call = dyn_cast<CallInst>(it);
4732 if (Call && getIntrinsicIDForCall(Call, TLI))
4735 LoadInst *Ld = dyn_cast<LoadInst>(it);
4736 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4737 emitAnalysis(Report(Ld)
4738 << "read with atomic ordering or volatile read");
4739 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4743 Loads.push_back(Ld);
4744 DepChecker.addAccess(Ld);
4748 // Save 'store' instructions. Abort if other instructions write to memory.
4749 if (it->mayWriteToMemory()) {
4750 StoreInst *St = dyn_cast<StoreInst>(it);
4752 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4755 if (!St->isSimple() && !IsAnnotatedParallel) {
4756 emitAnalysis(Report(St)
4757 << "write with atomic ordering or volatile write");
4758 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4762 Stores.push_back(St);
4763 DepChecker.addAccess(St);
4768 // Now we have two lists that hold the loads and the stores.
4769 // Next, we find the pointers that they use.
4771 // Check if we see any stores. If there are no stores, then we don't
4772 // care if the pointers are *restrict*.
4773 if (!Stores.size()) {
4774 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4778 AccessAnalysis::DepCandidates DependentAccesses;
4779 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4781 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4782 // multiple times on the same object. If the ptr is accessed twice, once
4783 // for read and once for write, it will only appear once (on the write
4784 // list). This is okay, since we are going to check for conflicts between
4785 // writes and between reads and writes, but not between reads and reads.
4788 ValueVector::iterator I, IE;
4789 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4790 StoreInst *ST = cast<StoreInst>(*I);
4791 Value* Ptr = ST->getPointerOperand();
4793 if (isUniform(Ptr)) {
4796 << "write to a loop invariant address could not be vectorized");
4797 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4801 // If we did *not* see this pointer before, insert it to the read-write
4802 // list. At this phase it is only a 'write' list.
4803 if (Seen.insert(Ptr)) {
4806 AliasAnalysis::Location Loc = AA->getLocation(ST);
4807 // The TBAA metadata could have a control dependency on the predication
4808 // condition, so we cannot rely on it when determining whether or not we
4809 // need runtime pointer checks.
4810 if (blockNeedsPredication(ST->getParent()))
4811 Loc.AATags.TBAA = nullptr;
4813 Accesses.addStore(Loc);
4817 if (IsAnnotatedParallel) {
4819 << "LV: A loop annotated parallel, ignore memory dependency "
4824 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4825 LoadInst *LD = cast<LoadInst>(*I);
4826 Value* Ptr = LD->getPointerOperand();
4827 // If we did *not* see this pointer before, insert it to the
4828 // read list. If we *did* see it before, then it is already in
4829 // the read-write list. This allows us to vectorize expressions
4830 // such as A[i] += x; Because the address of A[i] is a read-write
4831 // pointer. This only works if the index of A[i] is consecutive.
4832 // If the address of i is unknown (for example A[B[i]]) then we may
4833 // read a few words, modify, and write a few words, and some of the
4834 // words may be written to the same address.
4835 bool IsReadOnlyPtr = false;
4836 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4838 IsReadOnlyPtr = true;
4841 AliasAnalysis::Location Loc = AA->getLocation(LD);
4842 // The TBAA metadata could have a control dependency on the predication
4843 // condition, so we cannot rely on it when determining whether or not we
4844 // need runtime pointer checks.
4845 if (blockNeedsPredication(LD->getParent()))
4846 Loc.AATags.TBAA = nullptr;
4848 Accesses.addLoad(Loc, IsReadOnlyPtr);
4851 // If we write (or read-write) to a single destination and there are no
4852 // other reads in this loop then is it safe to vectorize.
4853 if (NumReadWrites == 1 && NumReads == 0) {
4854 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4858 // Build dependence sets and check whether we need a runtime pointer bounds
4860 Accesses.buildDependenceSets();
4861 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4863 // Find pointers with computable bounds. We are going to use this information
4864 // to place a runtime bound check.
4865 unsigned NumComparisons = 0;
4866 bool CanDoRT = false;
4868 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4871 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4872 " pointer comparisons.\n");
4874 // If we only have one set of dependences to check pointers among we don't
4875 // need a runtime check.
4876 if (NumComparisons == 0 && NeedRTCheck)
4877 NeedRTCheck = false;
4879 // Check that we did not collect too many pointers or found an unsizeable
4881 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4887 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4890 if (NeedRTCheck && !CanDoRT) {
4891 emitAnalysis(Report() << "cannot identify array bounds");
4892 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4893 "the array bounds.\n");
4898 PtrRtCheck.Need = NeedRTCheck;
4900 bool CanVecMem = true;
4901 if (Accesses.isDependencyCheckNeeded()) {
4902 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4903 CanVecMem = DepChecker.areDepsSafe(
4904 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4905 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4907 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4908 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4911 // Clear the dependency checks. We assume they are not needed.
4912 Accesses.resetDepChecks();
4915 PtrRtCheck.Need = true;
4917 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4918 TheLoop, Strides, true);
4919 // Check that we did not collect too many pointers or found an unsizeable
4921 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4922 if (!CanDoRT && NumComparisons > 0)
4923 emitAnalysis(Report()
4924 << "cannot check memory dependencies at runtime");
4926 emitAnalysis(Report()
4927 << NumComparisons << " exceeds limit of "
4928 << RuntimeMemoryCheckThreshold
4929 << " dependent memory operations checked at runtime");
4930 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4940 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4942 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4943 " need a runtime memory check.\n");
4948 static bool hasMultipleUsesOf(Instruction *I,
4949 SmallPtrSetImpl<Instruction *> &Insts) {
4950 unsigned NumUses = 0;
4951 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4952 if (Insts.count(dyn_cast<Instruction>(*Use)))
4961 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4962 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4963 if (!Set.count(dyn_cast<Instruction>(*Use)))
4968 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4969 ReductionKind Kind) {
4970 if (Phi->getNumIncomingValues() != 2)
4973 // Reduction variables are only found in the loop header block.
4974 if (Phi->getParent() != TheLoop->getHeader())
4977 // Obtain the reduction start value from the value that comes from the loop
4979 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4981 // ExitInstruction is the single value which is used outside the loop.
4982 // We only allow for a single reduction value to be used outside the loop.
4983 // This includes users of the reduction, variables (which form a cycle
4984 // which ends in the phi node).
4985 Instruction *ExitInstruction = nullptr;
4986 // Indicates that we found a reduction operation in our scan.
4987 bool FoundReduxOp = false;
4989 // We start with the PHI node and scan for all of the users of this
4990 // instruction. All users must be instructions that can be used as reduction
4991 // variables (such as ADD). We must have a single out-of-block user. The cycle
4992 // must include the original PHI.
4993 bool FoundStartPHI = false;
4995 // To recognize min/max patterns formed by a icmp select sequence, we store
4996 // the number of instruction we saw from the recognized min/max pattern,
4997 // to make sure we only see exactly the two instructions.
4998 unsigned NumCmpSelectPatternInst = 0;
4999 ReductionInstDesc ReduxDesc(false, nullptr);
5001 SmallPtrSet<Instruction *, 8> VisitedInsts;
5002 SmallVector<Instruction *, 8> Worklist;
5003 Worklist.push_back(Phi);
5004 VisitedInsts.insert(Phi);
5006 // A value in the reduction can be used:
5007 // - By the reduction:
5008 // - Reduction operation:
5009 // - One use of reduction value (safe).
5010 // - Multiple use of reduction value (not safe).
5012 // - All uses of the PHI must be the reduction (safe).
5013 // - Otherwise, not safe.
5014 // - By one instruction outside of the loop (safe).
5015 // - By further instructions outside of the loop (not safe).
5016 // - By an instruction that is not part of the reduction (not safe).
5018 // * An instruction type other than PHI or the reduction operation.
5019 // * A PHI in the header other than the initial PHI.
5020 while (!Worklist.empty()) {
5021 Instruction *Cur = Worklist.back();
5022 Worklist.pop_back();
5025 // If the instruction has no users then this is a broken chain and can't be
5026 // a reduction variable.
5027 if (Cur->use_empty())
5030 bool IsAPhi = isa<PHINode>(Cur);
5032 // A header PHI use other than the original PHI.
5033 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5036 // Reductions of instructions such as Div, and Sub is only possible if the
5037 // LHS is the reduction variable.
5038 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5039 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5040 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5043 // Any reduction instruction must be of one of the allowed kinds.
5044 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5045 if (!ReduxDesc.IsReduction)
5048 // A reduction operation must only have one use of the reduction value.
5049 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5050 hasMultipleUsesOf(Cur, VisitedInsts))
5053 // All inputs to a PHI node must be a reduction value.
5054 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5057 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5058 isa<SelectInst>(Cur)))
5059 ++NumCmpSelectPatternInst;
5060 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5061 isa<SelectInst>(Cur)))
5062 ++NumCmpSelectPatternInst;
5064 // Check whether we found a reduction operator.
5065 FoundReduxOp |= !IsAPhi;
5067 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5068 // onto the stack. This way we are going to have seen all inputs to PHI
5069 // nodes once we get to them.
5070 SmallVector<Instruction *, 8> NonPHIs;
5071 SmallVector<Instruction *, 8> PHIs;
5072 for (User *U : Cur->users()) {
5073 Instruction *UI = cast<Instruction>(U);
5075 // Check if we found the exit user.
5076 BasicBlock *Parent = UI->getParent();
5077 if (!TheLoop->contains(Parent)) {
5078 // Exit if you find multiple outside users or if the header phi node is
5079 // being used. In this case the user uses the value of the previous
5080 // iteration, in which case we would loose "VF-1" iterations of the
5081 // reduction operation if we vectorize.
5082 if (ExitInstruction != nullptr || Cur == Phi)
5085 // The instruction used by an outside user must be the last instruction
5086 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5087 // operations on the value.
5088 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5091 ExitInstruction = Cur;
5095 // Process instructions only once (termination). Each reduction cycle
5096 // value must only be used once, except by phi nodes and min/max
5097 // reductions which are represented as a cmp followed by a select.
5098 ReductionInstDesc IgnoredVal(false, nullptr);
5099 if (VisitedInsts.insert(UI)) {
5100 if (isa<PHINode>(UI))
5103 NonPHIs.push_back(UI);
5104 } else if (!isa<PHINode>(UI) &&
5105 ((!isa<FCmpInst>(UI) &&
5106 !isa<ICmpInst>(UI) &&
5107 !isa<SelectInst>(UI)) ||
5108 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5111 // Remember that we completed the cycle.
5113 FoundStartPHI = true;
5115 Worklist.append(PHIs.begin(), PHIs.end());
5116 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5119 // This means we have seen one but not the other instruction of the
5120 // pattern or more than just a select and cmp.
5121 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5122 NumCmpSelectPatternInst != 2)
5125 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5128 // We found a reduction var if we have reached the original phi node and we
5129 // only have a single instruction with out-of-loop users.
5131 // This instruction is allowed to have out-of-loop users.
5132 AllowedExit.insert(ExitInstruction);
5134 // Save the description of this reduction variable.
5135 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5136 ReduxDesc.MinMaxKind);
5137 Reductions[Phi] = RD;
5138 // We've ended the cycle. This is a reduction variable if we have an
5139 // outside user and it has a binary op.
5144 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5145 /// pattern corresponding to a min(X, Y) or max(X, Y).
5146 LoopVectorizationLegality::ReductionInstDesc
5147 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5148 ReductionInstDesc &Prev) {
5150 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5151 "Expect a select instruction");
5152 Instruction *Cmp = nullptr;
5153 SelectInst *Select = nullptr;
5155 // We must handle the select(cmp()) as a single instruction. Advance to the
5157 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5158 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5159 return ReductionInstDesc(false, I);
5160 return ReductionInstDesc(Select, Prev.MinMaxKind);
5163 // Only handle single use cases for now.
5164 if (!(Select = dyn_cast<SelectInst>(I)))
5165 return ReductionInstDesc(false, I);
5166 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5167 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5168 return ReductionInstDesc(false, I);
5169 if (!Cmp->hasOneUse())
5170 return ReductionInstDesc(false, I);
5175 // Look for a min/max pattern.
5176 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5177 return ReductionInstDesc(Select, MRK_UIntMin);
5178 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5179 return ReductionInstDesc(Select, MRK_UIntMax);
5180 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5181 return ReductionInstDesc(Select, MRK_SIntMax);
5182 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5183 return ReductionInstDesc(Select, MRK_SIntMin);
5184 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5185 return ReductionInstDesc(Select, MRK_FloatMin);
5186 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5187 return ReductionInstDesc(Select, MRK_FloatMax);
5188 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5189 return ReductionInstDesc(Select, MRK_FloatMin);
5190 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5191 return ReductionInstDesc(Select, MRK_FloatMax);
5193 return ReductionInstDesc(false, I);
5196 LoopVectorizationLegality::ReductionInstDesc
5197 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5199 ReductionInstDesc &Prev) {
5200 bool FP = I->getType()->isFloatingPointTy();
5201 bool FastMath = FP && I->hasUnsafeAlgebra();
5202 switch (I->getOpcode()) {
5204 return ReductionInstDesc(false, I);
5205 case Instruction::PHI:
5206 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5207 Kind != RK_FloatMinMax))
5208 return ReductionInstDesc(false, I);
5209 return ReductionInstDesc(I, Prev.MinMaxKind);
5210 case Instruction::Sub:
5211 case Instruction::Add:
5212 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5213 case Instruction::Mul:
5214 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5215 case Instruction::And:
5216 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5217 case Instruction::Or:
5218 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5219 case Instruction::Xor:
5220 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5221 case Instruction::FMul:
5222 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5223 case Instruction::FSub:
5224 case Instruction::FAdd:
5225 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5226 case Instruction::FCmp:
5227 case Instruction::ICmp:
5228 case Instruction::Select:
5229 if (Kind != RK_IntegerMinMax &&
5230 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5231 return ReductionInstDesc(false, I);
5232 return isMinMaxSelectCmpPattern(I, Prev);
5236 LoopVectorizationLegality::InductionKind
5237 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5238 Type *PhiTy = Phi->getType();
5239 // We only handle integer and pointer inductions variables.
5240 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5241 return IK_NoInduction;
5243 // Check that the PHI is consecutive.
5244 const SCEV *PhiScev = SE->getSCEV(Phi);
5245 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5247 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5248 return IK_NoInduction;
5250 const SCEV *Step = AR->getStepRecurrence(*SE);
5252 // Integer inductions need to have a stride of one.
5253 if (PhiTy->isIntegerTy()) {
5255 return IK_IntInduction;
5256 if (Step->isAllOnesValue())
5257 return IK_ReverseIntInduction;
5258 return IK_NoInduction;
5261 // Calculate the pointer stride and check if it is consecutive.
5262 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5264 return IK_NoInduction;
5266 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5267 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5268 if (C->getValue()->equalsInt(Size))
5269 return IK_PtrInduction;
5270 else if (C->getValue()->equalsInt(0 - Size))
5271 return IK_ReversePtrInduction;
5273 return IK_NoInduction;
5276 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5277 Value *In0 = const_cast<Value*>(V);
5278 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5282 return Inductions.count(PN);
5285 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5286 assert(TheLoop->contains(BB) && "Unknown block used");
5288 // Blocks that do not dominate the latch need predication.
5289 BasicBlock* Latch = TheLoop->getLoopLatch();
5290 return !DT->dominates(BB, Latch);
5293 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5294 SmallPtrSetImpl<Value *> &SafePtrs) {
5295 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5296 // We might be able to hoist the load.
5297 if (it->mayReadFromMemory()) {
5298 LoadInst *LI = dyn_cast<LoadInst>(it);
5299 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5303 // We don't predicate stores at the moment.
5304 if (it->mayWriteToMemory()) {
5305 StoreInst *SI = dyn_cast<StoreInst>(it);
5306 // We only support predication of stores in basic blocks with one
5308 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5309 !SafePtrs.count(SI->getPointerOperand()) ||
5310 !SI->getParent()->getSinglePredecessor())
5316 // Check that we don't have a constant expression that can trap as operand.
5317 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5319 if (Constant *C = dyn_cast<Constant>(*OI))
5324 // The instructions below can trap.
5325 switch (it->getOpcode()) {
5327 case Instruction::UDiv:
5328 case Instruction::SDiv:
5329 case Instruction::URem:
5330 case Instruction::SRem:
5338 LoopVectorizationCostModel::VectorizationFactor
5339 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5340 // Width 1 means no vectorize
5341 VectorizationFactor Factor = { 1U, 0U };
5342 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5343 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5344 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5348 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5349 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5350 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5354 // Find the trip count.
5355 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5356 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5358 unsigned WidestType = getWidestType();
5359 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5360 unsigned MaxSafeDepDist = -1U;
5361 if (Legal->getMaxSafeDepDistBytes() != -1U)
5362 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5363 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5364 WidestRegister : MaxSafeDepDist);
5365 unsigned MaxVectorSize = WidestRegister / WidestType;
5366 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5367 DEBUG(dbgs() << "LV: The Widest register is: "
5368 << WidestRegister << " bits.\n");
5370 if (MaxVectorSize == 0) {
5371 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5375 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5376 " into one vector!");
5378 unsigned VF = MaxVectorSize;
5380 // If we optimize the program for size, avoid creating the tail loop.
5382 // If we are unable to calculate the trip count then don't try to vectorize.
5384 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5385 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5389 // Find the maximum SIMD width that can fit within the trip count.
5390 VF = TC % MaxVectorSize;
5395 // If the trip count that we found modulo the vectorization factor is not
5396 // zero then we require a tail.
5398 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");
5399 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5404 int UserVF = Hints->getWidth();
5406 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5407 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5409 Factor.Width = UserVF;
5413 float Cost = expectedCost(1);
5415 const float ScalarCost = Cost;
5418 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5420 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5421 // Ignore scalar width, because the user explicitly wants vectorization.
5422 if (ForceVectorization && VF > 1) {
5424 Cost = expectedCost(Width) / (float)Width;
5427 for (unsigned i=2; i <= VF; i*=2) {
5428 // Notice that the vector loop needs to be executed less times, so
5429 // we need to divide the cost of the vector loops by the width of
5430 // the vector elements.
5431 float VectorCost = expectedCost(i) / (float)i;
5432 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5433 (int)VectorCost << ".\n");
5434 if (VectorCost < Cost) {
5440 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5441 << "LV: Vectorization seems to be not beneficial, "
5442 << "but was forced by a user.\n");
5443 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5444 Factor.Width = Width;
5445 Factor.Cost = Width * Cost;
5449 unsigned LoopVectorizationCostModel::getWidestType() {
5450 unsigned MaxWidth = 8;
5453 for (Loop::block_iterator bb = TheLoop->block_begin(),
5454 be = TheLoop->block_end(); bb != be; ++bb) {
5455 BasicBlock *BB = *bb;
5457 // For each instruction in the loop.
5458 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5459 Type *T = it->getType();
5461 // Only examine Loads, Stores and PHINodes.
5462 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5465 // Examine PHI nodes that are reduction variables.
5466 if (PHINode *PN = dyn_cast<PHINode>(it))
5467 if (!Legal->getReductionVars()->count(PN))
5470 // Examine the stored values.
5471 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5472 T = ST->getValueOperand()->getType();
5474 // Ignore loaded pointer types and stored pointer types that are not
5475 // consecutive. However, we do want to take consecutive stores/loads of
5476 // pointer vectors into account.
5477 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5480 MaxWidth = std::max(MaxWidth,
5481 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5489 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5491 unsigned LoopCost) {
5493 // -- The unroll heuristics --
5494 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5495 // There are many micro-architectural considerations that we can't predict
5496 // at this level. For example, frontend pressure (on decode or fetch) due to
5497 // code size, or the number and capabilities of the execution ports.
5499 // We use the following heuristics to select the unroll factor:
5500 // 1. If the code has reductions, then we unroll in order to break the cross
5501 // iteration dependency.
5502 // 2. If the loop is really small, then we unroll in order to reduce the loop
5504 // 3. We don't unroll if we think that we will spill registers to memory due
5505 // to the increased register pressure.
5507 // Use the user preference, unless 'auto' is selected.
5508 int UserUF = Hints->getInterleave();
5512 // When we optimize for size, we don't unroll.
5516 // We used the distance for the unroll factor.
5517 if (Legal->getMaxSafeDepDistBytes() != -1U)
5520 // Do not unroll loops with a relatively small trip count.
5521 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5522 TheLoop->getLoopLatch());
5523 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5526 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5527 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5531 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5532 TargetNumRegisters = ForceTargetNumScalarRegs;
5534 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5535 TargetNumRegisters = ForceTargetNumVectorRegs;
5538 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5539 // We divide by these constants so assume that we have at least one
5540 // instruction that uses at least one register.
5541 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5542 R.NumInstructions = std::max(R.NumInstructions, 1U);
5544 // We calculate the unroll factor using the following formula.
5545 // Subtract the number of loop invariants from the number of available
5546 // registers. These registers are used by all of the unrolled instances.
5547 // Next, divide the remaining registers by the number of registers that is
5548 // required by the loop, in order to estimate how many parallel instances
5549 // fit without causing spills. All of this is rounded down if necessary to be
5550 // a power of two. We want power of two unroll factors to simplify any
5551 // addressing operations or alignment considerations.
5552 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5555 // Don't count the induction variable as unrolled.
5556 if (EnableIndVarRegisterHeur)
5557 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5558 std::max(1U, (R.MaxLocalUsers - 1)));
5560 // Clamp the unroll factor ranges to reasonable factors.
5561 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5563 // Check if the user has overridden the unroll max.
5565 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5566 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5568 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5569 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5572 // If we did not calculate the cost for VF (because the user selected the VF)
5573 // then we calculate the cost of VF here.
5575 LoopCost = expectedCost(VF);
5577 // Clamp the calculated UF to be between the 1 and the max unroll factor
5578 // that the target allows.
5579 if (UF > MaxInterleaveSize)
5580 UF = MaxInterleaveSize;
5584 // Unroll if we vectorized this loop and there is a reduction that could
5585 // benefit from unrolling.
5586 if (VF > 1 && Legal->getReductionVars()->size()) {
5587 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5591 // Note that if we've already vectorized the loop we will have done the
5592 // runtime check and so unrolling won't require further checks.
5593 bool UnrollingRequiresRuntimePointerCheck =
5594 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5596 // We want to unroll small loops in order to reduce the loop overhead and
5597 // potentially expose ILP opportunities.
5598 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5599 if (!UnrollingRequiresRuntimePointerCheck &&
5600 LoopCost < SmallLoopCost) {
5601 // We assume that the cost overhead is 1 and we use the cost model
5602 // to estimate the cost of the loop and unroll until the cost of the
5603 // loop overhead is about 5% of the cost of the loop.
5604 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5606 // Unroll until store/load ports (estimated by max unroll factor) are
5608 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5609 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5611 // If we have a scalar reduction (vector reductions are already dealt with
5612 // by this point), we can increase the critical path length if the loop
5613 // we're unrolling is inside another loop. Limit, by default to 2, so the
5614 // critical path only gets increased by one reduction operation.
5615 if (Legal->getReductionVars()->size() &&
5616 TheLoop->getLoopDepth() > 1) {
5617 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5618 SmallUF = std::min(SmallUF, F);
5619 StoresUF = std::min(StoresUF, F);
5620 LoadsUF = std::min(LoadsUF, F);
5623 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5624 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5625 return std::max(StoresUF, LoadsUF);
5628 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5632 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5636 LoopVectorizationCostModel::RegisterUsage
5637 LoopVectorizationCostModel::calculateRegisterUsage() {
5638 // This function calculates the register usage by measuring the highest number
5639 // of values that are alive at a single location. Obviously, this is a very
5640 // rough estimation. We scan the loop in a topological order in order and
5641 // assign a number to each instruction. We use RPO to ensure that defs are
5642 // met before their users. We assume that each instruction that has in-loop
5643 // users starts an interval. We record every time that an in-loop value is
5644 // used, so we have a list of the first and last occurrences of each
5645 // instruction. Next, we transpose this data structure into a multi map that
5646 // holds the list of intervals that *end* at a specific location. This multi
5647 // map allows us to perform a linear search. We scan the instructions linearly
5648 // and record each time that a new interval starts, by placing it in a set.
5649 // If we find this value in the multi-map then we remove it from the set.
5650 // The max register usage is the maximum size of the set.
5651 // We also search for instructions that are defined outside the loop, but are
5652 // used inside the loop. We need this number separately from the max-interval
5653 // usage number because when we unroll, loop-invariant values do not take
5655 LoopBlocksDFS DFS(TheLoop);
5659 R.NumInstructions = 0;
5661 // Each 'key' in the map opens a new interval. The values
5662 // of the map are the index of the 'last seen' usage of the
5663 // instruction that is the key.
5664 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5665 // Maps instruction to its index.
5666 DenseMap<unsigned, Instruction*> IdxToInstr;
5667 // Marks the end of each interval.
5668 IntervalMap EndPoint;
5669 // Saves the list of instruction indices that are used in the loop.
5670 SmallSet<Instruction*, 8> Ends;
5671 // Saves the list of values that are used in the loop but are
5672 // defined outside the loop, such as arguments and constants.
5673 SmallPtrSet<Value*, 8> LoopInvariants;
5676 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5677 be = DFS.endRPO(); bb != be; ++bb) {
5678 R.NumInstructions += (*bb)->size();
5679 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5681 Instruction *I = it;
5682 IdxToInstr[Index++] = I;
5684 // Save the end location of each USE.
5685 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5686 Value *U = I->getOperand(i);
5687 Instruction *Instr = dyn_cast<Instruction>(U);
5689 // Ignore non-instruction values such as arguments, constants, etc.
5690 if (!Instr) continue;
5692 // If this instruction is outside the loop then record it and continue.
5693 if (!TheLoop->contains(Instr)) {
5694 LoopInvariants.insert(Instr);
5698 // Overwrite previous end points.
5699 EndPoint[Instr] = Index;
5705 // Saves the list of intervals that end with the index in 'key'.
5706 typedef SmallVector<Instruction*, 2> InstrList;
5707 DenseMap<unsigned, InstrList> TransposeEnds;
5709 // Transpose the EndPoints to a list of values that end at each index.
5710 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5712 TransposeEnds[it->second].push_back(it->first);
5714 SmallSet<Instruction*, 8> OpenIntervals;
5715 unsigned MaxUsage = 0;
5718 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5719 for (unsigned int i = 0; i < Index; ++i) {
5720 Instruction *I = IdxToInstr[i];
5721 // Ignore instructions that are never used within the loop.
5722 if (!Ends.count(I)) continue;
5724 // Remove all of the instructions that end at this location.
5725 InstrList &List = TransposeEnds[i];
5726 for (unsigned int j=0, e = List.size(); j < e; ++j)
5727 OpenIntervals.erase(List[j]);
5729 // Count the number of live interals.
5730 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5732 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5733 OpenIntervals.size() << '\n');
5735 // Add the current instruction to the list of open intervals.
5736 OpenIntervals.insert(I);
5739 unsigned Invariant = LoopInvariants.size();
5740 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5741 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5742 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5744 R.LoopInvariantRegs = Invariant;
5745 R.MaxLocalUsers = MaxUsage;
5749 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5753 for (Loop::block_iterator bb = TheLoop->block_begin(),
5754 be = TheLoop->block_end(); bb != be; ++bb) {
5755 unsigned BlockCost = 0;
5756 BasicBlock *BB = *bb;
5758 // For each instruction in the old loop.
5759 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5760 // Skip dbg intrinsics.
5761 if (isa<DbgInfoIntrinsic>(it))
5764 unsigned C = getInstructionCost(it, VF);
5766 // Check if we should override the cost.
5767 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5768 C = ForceTargetInstructionCost;
5771 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5772 VF << " For instruction: " << *it << '\n');
5775 // We assume that if-converted blocks have a 50% chance of being executed.
5776 // When the code is scalar then some of the blocks are avoided due to CF.
5777 // When the code is vectorized we execute all code paths.
5778 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5787 /// \brief Check whether the address computation for a non-consecutive memory
5788 /// access looks like an unlikely candidate for being merged into the indexing
5791 /// We look for a GEP which has one index that is an induction variable and all
5792 /// other indices are loop invariant. If the stride of this access is also
5793 /// within a small bound we decide that this address computation can likely be
5794 /// merged into the addressing mode.
5795 /// In all other cases, we identify the address computation as complex.
5796 static bool isLikelyComplexAddressComputation(Value *Ptr,
5797 LoopVectorizationLegality *Legal,
5798 ScalarEvolution *SE,
5799 const Loop *TheLoop) {
5800 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5804 // We are looking for a gep with all loop invariant indices except for one
5805 // which should be an induction variable.
5806 unsigned NumOperands = Gep->getNumOperands();
5807 for (unsigned i = 1; i < NumOperands; ++i) {
5808 Value *Opd = Gep->getOperand(i);
5809 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5810 !Legal->isInductionVariable(Opd))
5814 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5815 // can likely be merged into the address computation.
5816 unsigned MaxMergeDistance = 64;
5818 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5822 // Check the step is constant.
5823 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5824 // Calculate the pointer stride and check if it is consecutive.
5825 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5829 const APInt &APStepVal = C->getValue()->getValue();
5831 // Huge step value - give up.
5832 if (APStepVal.getBitWidth() > 64)
5835 int64_t StepVal = APStepVal.getSExtValue();
5837 return StepVal > MaxMergeDistance;
5840 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5841 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5847 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5848 // If we know that this instruction will remain uniform, check the cost of
5849 // the scalar version.
5850 if (Legal->isUniformAfterVectorization(I))
5853 Type *RetTy = I->getType();
5854 Type *VectorTy = ToVectorTy(RetTy, VF);
5856 // TODO: We need to estimate the cost of intrinsic calls.
5857 switch (I->getOpcode()) {
5858 case Instruction::GetElementPtr:
5859 // We mark this instruction as zero-cost because the cost of GEPs in
5860 // vectorized code depends on whether the corresponding memory instruction
5861 // is scalarized or not. Therefore, we handle GEPs with the memory
5862 // instruction cost.
5864 case Instruction::Br: {
5865 return TTI.getCFInstrCost(I->getOpcode());
5867 case Instruction::PHI:
5868 //TODO: IF-converted IFs become selects.
5870 case Instruction::Add:
5871 case Instruction::FAdd:
5872 case Instruction::Sub:
5873 case Instruction::FSub:
5874 case Instruction::Mul:
5875 case Instruction::FMul:
5876 case Instruction::UDiv:
5877 case Instruction::SDiv:
5878 case Instruction::FDiv:
5879 case Instruction::URem:
5880 case Instruction::SRem:
5881 case Instruction::FRem:
5882 case Instruction::Shl:
5883 case Instruction::LShr:
5884 case Instruction::AShr:
5885 case Instruction::And:
5886 case Instruction::Or:
5887 case Instruction::Xor: {
5888 // Since we will replace the stride by 1 the multiplication should go away.
5889 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5891 // Certain instructions can be cheaper to vectorize if they have a constant
5892 // second vector operand. One example of this are shifts on x86.
5893 TargetTransformInfo::OperandValueKind Op1VK =
5894 TargetTransformInfo::OK_AnyValue;
5895 TargetTransformInfo::OperandValueKind Op2VK =
5896 TargetTransformInfo::OK_AnyValue;
5897 TargetTransformInfo::OperandValueProperties Op1VP =
5898 TargetTransformInfo::OP_None;
5899 TargetTransformInfo::OperandValueProperties Op2VP =
5900 TargetTransformInfo::OP_None;
5901 Value *Op2 = I->getOperand(1);
5903 // Check for a splat of a constant or for a non uniform vector of constants.
5904 if (isa<ConstantInt>(Op2)) {
5905 ConstantInt *CInt = cast<ConstantInt>(Op2);
5906 if (CInt && CInt->getValue().isPowerOf2())
5907 Op2VP = TargetTransformInfo::OP_PowerOf2;
5908 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5909 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5910 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5911 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5913 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5914 if (CInt && CInt->getValue().isPowerOf2())
5915 Op2VP = TargetTransformInfo::OP_PowerOf2;
5916 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5920 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5923 case Instruction::Select: {
5924 SelectInst *SI = cast<SelectInst>(I);
5925 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5926 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5927 Type *CondTy = SI->getCondition()->getType();
5929 CondTy = VectorType::get(CondTy, VF);
5931 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5933 case Instruction::ICmp:
5934 case Instruction::FCmp: {
5935 Type *ValTy = I->getOperand(0)->getType();
5936 VectorTy = ToVectorTy(ValTy, VF);
5937 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5939 case Instruction::Store:
5940 case Instruction::Load: {
5941 StoreInst *SI = dyn_cast<StoreInst>(I);
5942 LoadInst *LI = dyn_cast<LoadInst>(I);
5943 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5945 VectorTy = ToVectorTy(ValTy, VF);
5947 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5948 unsigned AS = SI ? SI->getPointerAddressSpace() :
5949 LI->getPointerAddressSpace();
5950 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5951 // We add the cost of address computation here instead of with the gep
5952 // instruction because only here we know whether the operation is
5955 return TTI.getAddressComputationCost(VectorTy) +
5956 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5958 // Scalarized loads/stores.
5959 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5960 bool Reverse = ConsecutiveStride < 0;
5961 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5962 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5963 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5964 bool IsComplexComputation =
5965 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5967 // The cost of extracting from the value vector and pointer vector.
5968 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5969 for (unsigned i = 0; i < VF; ++i) {
5970 // The cost of extracting the pointer operand.
5971 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5972 // In case of STORE, the cost of ExtractElement from the vector.
5973 // In case of LOAD, the cost of InsertElement into the returned
5975 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5976 Instruction::InsertElement,
5980 // The cost of the scalar loads/stores.
5981 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5982 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5987 // Wide load/stores.
5988 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5989 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5992 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5996 case Instruction::ZExt:
5997 case Instruction::SExt:
5998 case Instruction::FPToUI:
5999 case Instruction::FPToSI:
6000 case Instruction::FPExt:
6001 case Instruction::PtrToInt:
6002 case Instruction::IntToPtr:
6003 case Instruction::SIToFP:
6004 case Instruction::UIToFP:
6005 case Instruction::Trunc:
6006 case Instruction::FPTrunc:
6007 case Instruction::BitCast: {
6008 // We optimize the truncation of induction variable.
6009 // The cost of these is the same as the scalar operation.
6010 if (I->getOpcode() == Instruction::Trunc &&
6011 Legal->isInductionVariable(I->getOperand(0)))
6012 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6013 I->getOperand(0)->getType());
6015 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6016 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6018 case Instruction::Call: {
6019 CallInst *CI = cast<CallInst>(I);
6020 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6021 assert(ID && "Not an intrinsic call!");
6022 Type *RetTy = ToVectorTy(CI->getType(), VF);
6023 SmallVector<Type*, 4> Tys;
6024 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6025 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6026 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6029 // We are scalarizing the instruction. Return the cost of the scalar
6030 // instruction, plus the cost of insert and extract into vector
6031 // elements, times the vector width.
6034 if (!RetTy->isVoidTy() && VF != 1) {
6035 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6037 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6040 // The cost of inserting the results plus extracting each one of the
6042 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6045 // The cost of executing VF copies of the scalar instruction. This opcode
6046 // is unknown. Assume that it is the same as 'mul'.
6047 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6053 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6054 if (Scalar->isVoidTy() || VF == 1)
6056 return VectorType::get(Scalar, VF);
6059 char LoopVectorize::ID = 0;
6060 static const char lv_name[] = "Loop Vectorization";
6061 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6062 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6063 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6064 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6065 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6066 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6067 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6068 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6069 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6070 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6073 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6074 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6078 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6079 // Check for a store.
6080 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6081 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6083 // Check for a load.
6084 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6085 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6091 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6092 bool IfPredicateStore) {
6093 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6094 // Holds vector parameters or scalars, in case of uniform vals.
6095 SmallVector<VectorParts, 4> Params;
6097 setDebugLocFromInst(Builder, Instr);
6099 // Find all of the vectorized parameters.
6100 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6101 Value *SrcOp = Instr->getOperand(op);
6103 // If we are accessing the old induction variable, use the new one.
6104 if (SrcOp == OldInduction) {
6105 Params.push_back(getVectorValue(SrcOp));
6109 // Try using previously calculated values.
6110 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6112 // If the src is an instruction that appeared earlier in the basic block
6113 // then it should already be vectorized.
6114 if (SrcInst && OrigLoop->contains(SrcInst)) {
6115 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6116 // The parameter is a vector value from earlier.
6117 Params.push_back(WidenMap.get(SrcInst));
6119 // The parameter is a scalar from outside the loop. Maybe even a constant.
6120 VectorParts Scalars;
6121 Scalars.append(UF, SrcOp);
6122 Params.push_back(Scalars);
6126 assert(Params.size() == Instr->getNumOperands() &&
6127 "Invalid number of operands");
6129 // Does this instruction return a value ?
6130 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6132 Value *UndefVec = IsVoidRetTy ? nullptr :
6133 UndefValue::get(Instr->getType());
6134 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6135 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6137 Instruction *InsertPt = Builder.GetInsertPoint();
6138 BasicBlock *IfBlock = Builder.GetInsertBlock();
6139 BasicBlock *CondBlock = nullptr;
6142 Loop *VectorLp = nullptr;
6143 if (IfPredicateStore) {
6144 assert(Instr->getParent()->getSinglePredecessor() &&
6145 "Only support single predecessor blocks");
6146 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6147 Instr->getParent());
6148 VectorLp = LI->getLoopFor(IfBlock);
6149 assert(VectorLp && "Must have a loop for this block");
6152 // For each vector unroll 'part':
6153 for (unsigned Part = 0; Part < UF; ++Part) {
6154 // For each scalar that we create:
6156 // Start an "if (pred) a[i] = ..." block.
6157 Value *Cmp = nullptr;
6158 if (IfPredicateStore) {
6159 if (Cond[Part]->getType()->isVectorTy())
6161 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6162 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6163 ConstantInt::get(Cond[Part]->getType(), 1));
6164 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6165 LoopVectorBody.push_back(CondBlock);
6166 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6167 // Update Builder with newly created basic block.
6168 Builder.SetInsertPoint(InsertPt);
6171 Instruction *Cloned = Instr->clone();
6173 Cloned->setName(Instr->getName() + ".cloned");
6174 // Replace the operands of the cloned instructions with extracted scalars.
6175 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6176 Value *Op = Params[op][Part];
6177 Cloned->setOperand(op, Op);
6180 // Place the cloned scalar in the new loop.
6181 Builder.Insert(Cloned);
6183 // If the original scalar returns a value we need to place it in a vector
6184 // so that future users will be able to use it.
6186 VecResults[Part] = Cloned;
6189 if (IfPredicateStore) {
6190 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6191 LoopVectorBody.push_back(NewIfBlock);
6192 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6193 Builder.SetInsertPoint(InsertPt);
6194 Instruction *OldBr = IfBlock->getTerminator();
6195 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6196 OldBr->eraseFromParent();
6197 IfBlock = NewIfBlock;
6202 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6203 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6204 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6206 return scalarizeInstruction(Instr, IfPredicateStore);
6209 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6213 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6217 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6219 // When unrolling and the VF is 1, we only need to add a simple scalar.
6220 Type *ITy = Val->getType();
6221 assert(!ITy->isVectorTy() && "Val must be a scalar");
6222 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6223 return Builder.CreateAdd(Val, C, "induction");