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 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/StringExtras.h"
59 #include "llvm/Analysis/AliasAnalysis.h"
60 #include "llvm/Analysis/BasicAliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/DemandedBits.h"
66 #include "llvm/Analysis/GlobalsModRef.h"
67 #include "llvm/Analysis/LoopAccessAnalysis.h"
68 #include "llvm/Analysis/LoopInfo.h"
69 #include "llvm/Analysis/LoopIterator.h"
70 #include "llvm/Analysis/LoopPass.h"
71 #include "llvm/Analysis/ScalarEvolution.h"
72 #include "llvm/Analysis/ScalarEvolutionExpander.h"
73 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
74 #include "llvm/Analysis/TargetTransformInfo.h"
75 #include "llvm/Analysis/ValueTracking.h"
76 #include "llvm/IR/Constants.h"
77 #include "llvm/IR/DataLayout.h"
78 #include "llvm/IR/DebugInfo.h"
79 #include "llvm/IR/DerivedTypes.h"
80 #include "llvm/IR/DiagnosticInfo.h"
81 #include "llvm/IR/Dominators.h"
82 #include "llvm/IR/Function.h"
83 #include "llvm/IR/IRBuilder.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/LLVMContext.h"
87 #include "llvm/IR/Module.h"
88 #include "llvm/IR/PatternMatch.h"
89 #include "llvm/IR/Type.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/IR/Verifier.h"
93 #include "llvm/Pass.h"
94 #include "llvm/Support/BranchProbability.h"
95 #include "llvm/Support/CommandLine.h"
96 #include "llvm/Support/Debug.h"
97 #include "llvm/Support/raw_ostream.h"
98 #include "llvm/Transforms/Scalar.h"
99 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
100 #include "llvm/Transforms/Utils/Local.h"
101 #include "llvm/Analysis/VectorUtils.h"
102 #include "llvm/Transforms/Utils/LoopUtils.h"
104 #include <functional>
108 using namespace llvm;
109 using namespace llvm::PatternMatch;
111 #define LV_NAME "loop-vectorize"
112 #define DEBUG_TYPE LV_NAME
114 STATISTIC(LoopsVectorized, "Number of loops vectorized");
115 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
129 /// This enables versioning on the strides of symbolically striding memory
130 /// accesses in code like the following.
131 /// for (i = 0; i < N; ++i)
132 /// A[i * Stride1] += B[i * Stride2] ...
134 /// Will be roughly translated to
135 /// if (Stride1 == 1 && Stride2 == 1) {
136 /// for (i = 0; i < N; i+=4)
140 static cl::opt<bool> EnableMemAccessVersioning(
141 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
142 cl::desc("Enable symblic stride memory access versioning"));
144 static cl::opt<bool> EnableInterleavedMemAccesses(
145 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
146 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
148 /// Maximum factor for an interleaved memory access.
149 static cl::opt<unsigned> MaxInterleaveGroupFactor(
150 "max-interleave-group-factor", cl::Hidden,
151 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
154 /// We don't interleave loops with a known constant trip count below this
156 static const unsigned TinyTripCountInterleaveThreshold = 128;
158 static cl::opt<unsigned> ForceTargetNumScalarRegs(
159 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of scalar registers."));
162 static cl::opt<unsigned> ForceTargetNumVectorRegs(
163 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
164 cl::desc("A flag that overrides the target's number of vector registers."));
166 /// Maximum vectorization interleave count.
167 static const unsigned MaxInterleaveFactor = 16;
169 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
170 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
171 cl::desc("A flag that overrides the target's max interleave factor for "
174 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
175 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
176 cl::desc("A flag that overrides the target's max interleave factor for "
177 "vectorized loops."));
179 static cl::opt<unsigned> ForceTargetInstructionCost(
180 "force-target-instruction-cost", cl::init(0), cl::Hidden,
181 cl::desc("A flag that overrides the target's expected cost for "
182 "an instruction to a single constant value. Mostly "
183 "useful for getting consistent testing."));
185 static cl::opt<unsigned> SmallLoopCost(
186 "small-loop-cost", cl::init(20), cl::Hidden,
188 "The cost of a loop that is considered 'small' by the interleaver."));
190 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
191 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
192 cl::desc("Enable the use of the block frequency analysis to access PGO "
193 "heuristics minimizing code growth in cold regions and being more "
194 "aggressive in hot regions."));
196 // Runtime interleave loops for load/store throughput.
197 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
198 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
200 "Enable runtime interleaving until load/store ports are saturated"));
202 /// The number of stores in a loop that are allowed to need predication.
203 static cl::opt<unsigned> NumberOfStoresToPredicate(
204 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
205 cl::desc("Max number of stores to be predicated behind an if."));
207 static cl::opt<bool> EnableIndVarRegisterHeur(
208 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
209 cl::desc("Count the induction variable only once when interleaving"));
211 static cl::opt<bool> EnableCondStoresVectorization(
212 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
213 cl::desc("Enable if predication of stores during vectorization."));
215 static cl::opt<unsigned> MaxNestedScalarReductionIC(
216 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
217 cl::desc("The maximum interleave count to use when interleaving a scalar "
218 "reduction in a nested loop."));
220 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
221 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
222 cl::desc("The maximum allowed number of runtime memory checks with a "
223 "vectorize(enable) pragma."));
227 // Forward declarations.
228 class LoopVectorizeHints;
229 class LoopVectorizationLegality;
230 class LoopVectorizationCostModel;
231 class LoopVectorizationRequirements;
233 /// \brief This modifies LoopAccessReport to initialize message with
234 /// loop-vectorizer-specific part.
235 class VectorizationReport : public LoopAccessReport {
237 VectorizationReport(Instruction *I = nullptr)
238 : LoopAccessReport("loop not vectorized: ", I) {}
240 /// \brief This allows promotion of the loop-access analysis report into the
241 /// loop-vectorizer report. It modifies the message to add the
242 /// loop-vectorizer-specific part of the message.
243 explicit VectorizationReport(const LoopAccessReport &R)
244 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
248 /// A helper function for converting Scalar types to vector types.
249 /// If the incoming type is void, we return void. If the VF is 1, we return
251 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
252 if (Scalar->isVoidTy() || VF == 1)
254 return VectorType::get(Scalar, VF);
257 /// InnerLoopVectorizer vectorizes loops which contain only one basic
258 /// block to a specified vectorization factor (VF).
259 /// This class performs the widening of scalars into vectors, or multiple
260 /// scalars. This class also implements the following features:
261 /// * It inserts an epilogue loop for handling loops that don't have iteration
262 /// counts that are known to be a multiple of the vectorization factor.
263 /// * It handles the code generation for reduction variables.
264 /// * Scalarization (implementation using scalars) of un-vectorizable
266 /// InnerLoopVectorizer does not perform any vectorization-legality
267 /// checks, and relies on the caller to check for the different legality
268 /// aspects. The InnerLoopVectorizer relies on the
269 /// LoopVectorizationLegality class to provide information about the induction
270 /// and reduction variables that were found to a given vectorization factor.
271 class InnerLoopVectorizer {
273 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
274 DominatorTree *DT, const TargetLibraryInfo *TLI,
275 const TargetTransformInfo *TTI, unsigned VecWidth,
276 unsigned UnrollFactor)
277 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
278 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
279 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
280 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr),
281 AddedSafetyChecks(false) {}
283 // Perform the actual loop widening (vectorization).
284 // MinimumBitWidths maps scalar integer values to the smallest bitwidth they
285 // can be validly truncated to. The cost model has assumed this truncation
286 // will happen when vectorizing.
287 void vectorize(LoopVectorizationLegality *L,
288 DenseMap<Instruction*,uint64_t> MinimumBitWidths) {
289 MinBWs = MinimumBitWidths;
291 // Create a new empty loop. Unlink the old loop and connect the new one.
293 // Widen each instruction in the old loop to a new one in the new loop.
294 // Use the Legality module to find the induction and reduction variables.
298 // Return true if any runtime check is added.
299 bool IsSafetyChecksAdded() {
300 return AddedSafetyChecks;
303 virtual ~InnerLoopVectorizer() {}
306 /// A small list of PHINodes.
307 typedef SmallVector<PHINode*, 4> PhiVector;
308 /// When we unroll loops we have multiple vector values for each scalar.
309 /// This data structure holds the unrolled and vectorized values that
310 /// originated from one scalar instruction.
311 typedef SmallVector<Value*, 2> VectorParts;
313 // When we if-convert we need to create edge masks. We have to cache values
314 // so that we don't end up with exponential recursion/IR.
315 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
316 VectorParts> EdgeMaskCache;
318 /// \brief Add checks for strides that were assumed to be 1.
320 /// Returns the last check instruction and the first check instruction in the
321 /// pair as (first, last).
322 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
324 /// Create an empty loop, based on the loop ranges of the old loop.
325 void createEmptyLoop();
326 /// Create a new induction variable inside L.
327 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
328 Value *Step, Instruction *DL);
329 /// Copy and widen the instructions from the old loop.
330 virtual void vectorizeLoop();
332 /// \brief The Loop exit block may have single value PHI nodes where the
333 /// incoming value is 'Undef'. While vectorizing we only handled real values
334 /// that were defined inside the loop. Here we fix the 'undef case'.
338 /// Shrinks vector element sizes based on information in "MinBWs".
339 void truncateToMinimalBitwidths();
341 /// A helper function that computes the predicate of the block BB, assuming
342 /// that the header block of the loop is set to True. It returns the *entry*
343 /// mask for the block BB.
344 VectorParts createBlockInMask(BasicBlock *BB);
345 /// A helper function that computes the predicate of the edge between SRC
347 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
349 /// A helper function to vectorize a single BB within the innermost loop.
350 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
352 /// Vectorize a single PHINode in a block. This method handles the induction
353 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
354 /// arbitrary length vectors.
355 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
356 unsigned UF, unsigned VF, PhiVector *PV);
358 /// Insert the new loop to the loop hierarchy and pass manager
359 /// and update the analysis passes.
360 void updateAnalysis();
362 /// This instruction is un-vectorizable. Implement it as a sequence
363 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
364 /// scalarized instruction behind an if block predicated on the control
365 /// dependence of the instruction.
366 virtual void scalarizeInstruction(Instruction *Instr,
367 bool IfPredicateStore=false);
369 /// Vectorize Load and Store instructions,
370 virtual void vectorizeMemoryInstruction(Instruction *Instr);
372 /// Create a broadcast instruction. This method generates a broadcast
373 /// instruction (shuffle) for loop invariant values and for the induction
374 /// value. If this is the induction variable then we extend it to N, N+1, ...
375 /// this is needed because each iteration in the loop corresponds to a SIMD
377 virtual Value *getBroadcastInstrs(Value *V);
379 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
380 /// to each vector element of Val. The sequence starts at StartIndex.
381 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
383 /// When we go over instructions in the basic block we rely on previous
384 /// values within the current basic block or on loop invariant values.
385 /// When we widen (vectorize) values we place them in the map. If the values
386 /// are not within the map, they have to be loop invariant, so we simply
387 /// broadcast them into a vector.
388 VectorParts &getVectorValue(Value *V);
390 /// Try to vectorize the interleaved access group that \p Instr belongs to.
391 void vectorizeInterleaveGroup(Instruction *Instr);
393 /// Generate a shuffle sequence that will reverse the vector Vec.
394 virtual Value *reverseVector(Value *Vec);
396 /// Returns (and creates if needed) the original loop trip count.
397 Value *getOrCreateTripCount(Loop *NewLoop);
399 /// Returns (and creates if needed) the trip count of the widened loop.
400 Value *getOrCreateVectorTripCount(Loop *NewLoop);
402 /// Emit a bypass check to see if the trip count would overflow, or we
403 /// wouldn't have enough iterations to execute one vector loop.
404 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
405 /// Emit a bypass check to see if the vector trip count is nonzero.
406 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
407 /// Emit bypass checks to check if strides we've assumed to be one really are.
408 void emitStrideChecks(Loop *L, BasicBlock *Bypass);
409 /// Emit bypass checks to check any memory assumptions we may have made.
410 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
412 /// This is a helper class that holds the vectorizer state. It maps scalar
413 /// instructions to vector instructions. When the code is 'unrolled' then
414 /// then a single scalar value is mapped to multiple vector parts. The parts
415 /// are stored in the VectorPart type.
417 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
419 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
421 /// \return True if 'Key' is saved in the Value Map.
422 bool has(Value *Key) const { return MapStorage.count(Key); }
424 /// Initializes a new entry in the map. Sets all of the vector parts to the
425 /// save value in 'Val'.
426 /// \return A reference to a vector with splat values.
427 VectorParts &splat(Value *Key, Value *Val) {
428 VectorParts &Entry = MapStorage[Key];
429 Entry.assign(UF, Val);
433 ///\return A reference to the value that is stored at 'Key'.
434 VectorParts &get(Value *Key) {
435 VectorParts &Entry = MapStorage[Key];
438 assert(Entry.size() == UF);
443 /// The unroll factor. Each entry in the map stores this number of vector
447 /// Map storage. We use std::map and not DenseMap because insertions to a
448 /// dense map invalidates its iterators.
449 std::map<Value *, VectorParts> MapStorage;
452 /// The original loop.
454 /// Scev analysis to use.
462 /// Target Library Info.
463 const TargetLibraryInfo *TLI;
464 /// Target Transform Info.
465 const TargetTransformInfo *TTI;
467 /// The vectorization SIMD factor to use. Each vector will have this many
472 /// The vectorization unroll factor to use. Each scalar is vectorized to this
473 /// many different vector instructions.
476 /// The builder that we use
479 // --- Vectorization state ---
481 /// The vector-loop preheader.
482 BasicBlock *LoopVectorPreHeader;
483 /// The scalar-loop preheader.
484 BasicBlock *LoopScalarPreHeader;
485 /// Middle Block between the vector and the scalar.
486 BasicBlock *LoopMiddleBlock;
487 ///The ExitBlock of the scalar loop.
488 BasicBlock *LoopExitBlock;
489 ///The vector loop body.
490 SmallVector<BasicBlock *, 4> LoopVectorBody;
491 ///The scalar loop body.
492 BasicBlock *LoopScalarBody;
493 /// A list of all bypass blocks. The first block is the entry of the loop.
494 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
496 /// The new Induction variable which was added to the new block.
498 /// The induction variable of the old basic block.
499 PHINode *OldInduction;
500 /// Maps scalars to widened vectors.
502 /// Store instructions that should be predicated, as a pair
503 /// <StoreInst, Predicate>
504 SmallVector<std::pair<StoreInst*,Value*>, 4> PredicatedStores;
505 EdgeMaskCache MaskCache;
506 /// Trip count of the original loop.
508 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
509 Value *VectorTripCount;
511 /// Map of scalar integer values to the smallest bitwidth they can be legally
512 /// represented as. The vector equivalents of these values should be truncated
514 DenseMap<Instruction*,uint64_t> MinBWs;
515 LoopVectorizationLegality *Legal;
517 // Record whether runtime check is added.
518 bool AddedSafetyChecks;
521 class InnerLoopUnroller : public InnerLoopVectorizer {
523 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
524 DominatorTree *DT, const TargetLibraryInfo *TLI,
525 const TargetTransformInfo *TTI, unsigned UnrollFactor)
526 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
529 void scalarizeInstruction(Instruction *Instr,
530 bool IfPredicateStore = false) override;
531 void vectorizeMemoryInstruction(Instruction *Instr) override;
532 Value *getBroadcastInstrs(Value *V) override;
533 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
534 Value *reverseVector(Value *Vec) override;
537 /// \brief Look for a meaningful debug location on the instruction or it's
539 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
544 if (I->getDebugLoc() != Empty)
547 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
548 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
549 if (OpInst->getDebugLoc() != Empty)
556 /// \brief Set the debug location in the builder using the debug location in the
558 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
559 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
560 B.SetCurrentDebugLocation(Inst->getDebugLoc());
562 B.SetCurrentDebugLocation(DebugLoc());
566 /// \return string containing a file name and a line # for the given loop.
567 static std::string getDebugLocString(const Loop *L) {
570 raw_string_ostream OS(Result);
571 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
572 LoopDbgLoc.print(OS);
574 // Just print the module name.
575 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
582 /// \brief Propagate known metadata from one instruction to another.
583 static void propagateMetadata(Instruction *To, const Instruction *From) {
584 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
585 From->getAllMetadataOtherThanDebugLoc(Metadata);
587 for (auto M : Metadata) {
588 unsigned Kind = M.first;
590 // These are safe to transfer (this is safe for TBAA, even when we
591 // if-convert, because should that metadata have had a control dependency
592 // on the condition, and thus actually aliased with some other
593 // non-speculated memory access when the condition was false, this would be
594 // caught by the runtime overlap checks).
595 if (Kind != LLVMContext::MD_tbaa &&
596 Kind != LLVMContext::MD_alias_scope &&
597 Kind != LLVMContext::MD_noalias &&
598 Kind != LLVMContext::MD_fpmath &&
599 Kind != LLVMContext::MD_nontemporal)
602 To->setMetadata(Kind, M.second);
606 /// \brief Propagate known metadata from one instruction to a vector of others.
607 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
609 if (Instruction *I = dyn_cast<Instruction>(V))
610 propagateMetadata(I, From);
613 /// \brief The group of interleaved loads/stores sharing the same stride and
614 /// close to each other.
616 /// Each member in this group has an index starting from 0, and the largest
617 /// index should be less than interleaved factor, which is equal to the absolute
618 /// value of the access's stride.
620 /// E.g. An interleaved load group of factor 4:
621 /// for (unsigned i = 0; i < 1024; i+=4) {
622 /// a = A[i]; // Member of index 0
623 /// b = A[i+1]; // Member of index 1
624 /// d = A[i+3]; // Member of index 3
628 /// An interleaved store group of factor 4:
629 /// for (unsigned i = 0; i < 1024; i+=4) {
631 /// A[i] = a; // Member of index 0
632 /// A[i+1] = b; // Member of index 1
633 /// A[i+2] = c; // Member of index 2
634 /// A[i+3] = d; // Member of index 3
637 /// Note: the interleaved load group could have gaps (missing members), but
638 /// the interleaved store group doesn't allow gaps.
639 class InterleaveGroup {
641 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
642 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
643 assert(Align && "The alignment should be non-zero");
645 Factor = std::abs(Stride);
646 assert(Factor > 1 && "Invalid interleave factor");
648 Reverse = Stride < 0;
652 bool isReverse() const { return Reverse; }
653 unsigned getFactor() const { return Factor; }
654 unsigned getAlignment() const { return Align; }
655 unsigned getNumMembers() const { return Members.size(); }
657 /// \brief Try to insert a new member \p Instr with index \p Index and
658 /// alignment \p NewAlign. The index is related to the leader and it could be
659 /// negative if it is the new leader.
661 /// \returns false if the instruction doesn't belong to the group.
662 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
663 assert(NewAlign && "The new member's alignment should be non-zero");
665 int Key = Index + SmallestKey;
667 // Skip if there is already a member with the same index.
668 if (Members.count(Key))
671 if (Key > LargestKey) {
672 // The largest index is always less than the interleave factor.
673 if (Index >= static_cast<int>(Factor))
677 } else if (Key < SmallestKey) {
678 // The largest index is always less than the interleave factor.
679 if (LargestKey - Key >= static_cast<int>(Factor))
685 // It's always safe to select the minimum alignment.
686 Align = std::min(Align, NewAlign);
687 Members[Key] = Instr;
691 /// \brief Get the member with the given index \p Index
693 /// \returns nullptr if contains no such member.
694 Instruction *getMember(unsigned Index) const {
695 int Key = SmallestKey + Index;
696 if (!Members.count(Key))
699 return Members.find(Key)->second;
702 /// \brief Get the index for the given member. Unlike the key in the member
703 /// map, the index starts from 0.
704 unsigned getIndex(Instruction *Instr) const {
705 for (auto I : Members)
706 if (I.second == Instr)
707 return I.first - SmallestKey;
709 llvm_unreachable("InterleaveGroup contains no such member");
712 Instruction *getInsertPos() const { return InsertPos; }
713 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
716 unsigned Factor; // Interleave Factor.
719 DenseMap<int, Instruction *> Members;
723 // To avoid breaking dependences, vectorized instructions of an interleave
724 // group should be inserted at either the first load or the last store in
727 // E.g. %even = load i32 // Insert Position
728 // %add = add i32 %even // Use of %even
732 // %odd = add i32 // Def of %odd
733 // store i32 %odd // Insert Position
734 Instruction *InsertPos;
737 /// \brief Drive the analysis of interleaved memory accesses in the loop.
739 /// Use this class to analyze interleaved accesses only when we can vectorize
740 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
741 /// on interleaved accesses is unsafe.
743 /// The analysis collects interleave groups and records the relationships
744 /// between the member and the group in a map.
745 class InterleavedAccessInfo {
747 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
748 : SE(SE), TheLoop(L), DT(DT) {}
750 ~InterleavedAccessInfo() {
751 SmallSet<InterleaveGroup *, 4> DelSet;
752 // Avoid releasing a pointer twice.
753 for (auto &I : InterleaveGroupMap)
754 DelSet.insert(I.second);
755 for (auto *Ptr : DelSet)
759 /// \brief Analyze the interleaved accesses and collect them in interleave
760 /// groups. Substitute symbolic strides using \p Strides.
761 void analyzeInterleaving(const ValueToValueMap &Strides);
763 /// \brief Check if \p Instr belongs to any interleave group.
764 bool isInterleaved(Instruction *Instr) const {
765 return InterleaveGroupMap.count(Instr);
768 /// \brief Get the interleave group that \p Instr belongs to.
770 /// \returns nullptr if doesn't have such group.
771 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
772 if (InterleaveGroupMap.count(Instr))
773 return InterleaveGroupMap.find(Instr)->second;
782 /// Holds the relationships between the members and the interleave group.
783 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
785 /// \brief The descriptor for a strided memory access.
786 struct StrideDescriptor {
787 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
789 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
791 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
793 int Stride; // The access's stride. It is negative for a reverse access.
794 const SCEV *Scev; // The scalar expression of this access
795 unsigned Size; // The size of the memory object.
796 unsigned Align; // The alignment of this access.
799 /// \brief Create a new interleave group with the given instruction \p Instr,
800 /// stride \p Stride and alignment \p Align.
802 /// \returns the newly created interleave group.
803 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
805 assert(!InterleaveGroupMap.count(Instr) &&
806 "Already in an interleaved access group");
807 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
808 return InterleaveGroupMap[Instr];
811 /// \brief Release the group and remove all the relationships.
812 void releaseGroup(InterleaveGroup *Group) {
813 for (unsigned i = 0; i < Group->getFactor(); i++)
814 if (Instruction *Member = Group->getMember(i))
815 InterleaveGroupMap.erase(Member);
820 /// \brief Collect all the accesses with a constant stride in program order.
821 void collectConstStridedAccesses(
822 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
823 const ValueToValueMap &Strides);
826 /// Utility class for getting and setting loop vectorizer hints in the form
827 /// of loop metadata.
828 /// This class keeps a number of loop annotations locally (as member variables)
829 /// and can, upon request, write them back as metadata on the loop. It will
830 /// initially scan the loop for existing metadata, and will update the local
831 /// values based on information in the loop.
832 /// We cannot write all values to metadata, as the mere presence of some info,
833 /// for example 'force', means a decision has been made. So, we need to be
834 /// careful NOT to add them if the user hasn't specifically asked so.
835 class LoopVectorizeHints {
842 /// Hint - associates name and validation with the hint value.
845 unsigned Value; // This may have to change for non-numeric values.
848 Hint(const char * Name, unsigned Value, HintKind Kind)
849 : Name(Name), Value(Value), Kind(Kind) { }
851 bool validate(unsigned Val) {
854 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
856 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
864 /// Vectorization width.
866 /// Vectorization interleave factor.
868 /// Vectorization forced
871 /// Return the loop metadata prefix.
872 static StringRef Prefix() { return "llvm.loop."; }
876 FK_Undefined = -1, ///< Not selected.
877 FK_Disabled = 0, ///< Forcing disabled.
878 FK_Enabled = 1, ///< Forcing enabled.
881 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
882 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
884 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
885 Force("vectorize.enable", FK_Undefined, HK_FORCE),
887 // Populate values with existing loop metadata.
888 getHintsFromMetadata();
890 // force-vector-interleave overrides DisableInterleaving.
891 if (VectorizerParams::isInterleaveForced())
892 Interleave.Value = VectorizerParams::VectorizationInterleave;
894 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
895 << "LV: Interleaving disabled by the pass manager\n");
898 /// Mark the loop L as already vectorized by setting the width to 1.
899 void setAlreadyVectorized() {
900 Width.Value = Interleave.Value = 1;
901 Hint Hints[] = {Width, Interleave};
902 writeHintsToMetadata(Hints);
905 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
906 if (getForce() == LoopVectorizeHints::FK_Disabled) {
907 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
908 emitOptimizationRemarkAnalysis(F->getContext(),
909 vectorizeAnalysisPassName(), *F,
910 L->getStartLoc(), emitRemark());
914 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
915 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
916 emitOptimizationRemarkAnalysis(F->getContext(),
917 vectorizeAnalysisPassName(), *F,
918 L->getStartLoc(), emitRemark());
922 if (getWidth() == 1 && getInterleave() == 1) {
923 // FIXME: Add a separate metadata to indicate when the loop has already
924 // been vectorized instead of setting width and count to 1.
925 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
926 // FIXME: Add interleave.disable metadata. This will allow
927 // vectorize.disable to be used without disabling the pass and errors
928 // to differentiate between disabled vectorization and a width of 1.
929 emitOptimizationRemarkAnalysis(
930 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
931 "loop not vectorized: vectorization and interleaving are explicitly "
932 "disabled, or vectorize width and interleave count are both set to "
940 /// Dumps all the hint information.
941 std::string emitRemark() const {
942 VectorizationReport R;
943 if (Force.Value == LoopVectorizeHints::FK_Disabled)
944 R << "vectorization is explicitly disabled";
946 R << "use -Rpass-analysis=loop-vectorize for more info";
947 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
949 if (Width.Value != 0)
950 R << ", Vector Width=" << Width.Value;
951 if (Interleave.Value != 0)
952 R << ", Interleave Count=" << Interleave.Value;
960 unsigned getWidth() const { return Width.Value; }
961 unsigned getInterleave() const { return Interleave.Value; }
962 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
963 const char *vectorizeAnalysisPassName() const {
964 // If hints are provided that don't disable vectorization use the
965 // AlwaysPrint pass name to force the frontend to print the diagnostic.
968 if (getForce() == LoopVectorizeHints::FK_Disabled)
970 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
972 return DiagnosticInfo::AlwaysPrint;
975 bool allowReordering() const {
976 // When enabling loop hints are provided we allow the vectorizer to change
977 // the order of operations that is given by the scalar loop. This is not
978 // enabled by default because can be unsafe or inefficient. For example,
979 // reordering floating-point operations will change the way round-off
980 // error accumulates in the loop.
981 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
985 /// Find hints specified in the loop metadata and update local values.
986 void getHintsFromMetadata() {
987 MDNode *LoopID = TheLoop->getLoopID();
991 // First operand should refer to the loop id itself.
992 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
993 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
995 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
996 const MDString *S = nullptr;
997 SmallVector<Metadata *, 4> Args;
999 // The expected hint is either a MDString or a MDNode with the first
1000 // operand a MDString.
1001 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1002 if (!MD || MD->getNumOperands() == 0)
1004 S = dyn_cast<MDString>(MD->getOperand(0));
1005 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1006 Args.push_back(MD->getOperand(i));
1008 S = dyn_cast<MDString>(LoopID->getOperand(i));
1009 assert(Args.size() == 0 && "too many arguments for MDString");
1015 // Check if the hint starts with the loop metadata prefix.
1016 StringRef Name = S->getString();
1017 if (Args.size() == 1)
1018 setHint(Name, Args[0]);
1022 /// Checks string hint with one operand and set value if valid.
1023 void setHint(StringRef Name, Metadata *Arg) {
1024 if (!Name.startswith(Prefix()))
1026 Name = Name.substr(Prefix().size(), StringRef::npos);
1028 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1030 unsigned Val = C->getZExtValue();
1032 Hint *Hints[] = {&Width, &Interleave, &Force};
1033 for (auto H : Hints) {
1034 if (Name == H->Name) {
1035 if (H->validate(Val))
1038 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1044 /// Create a new hint from name / value pair.
1045 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1046 LLVMContext &Context = TheLoop->getHeader()->getContext();
1047 Metadata *MDs[] = {MDString::get(Context, Name),
1048 ConstantAsMetadata::get(
1049 ConstantInt::get(Type::getInt32Ty(Context), V))};
1050 return MDNode::get(Context, MDs);
1053 /// Matches metadata with hint name.
1054 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1055 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1059 for (auto H : HintTypes)
1060 if (Name->getString().endswith(H.Name))
1065 /// Sets current hints into loop metadata, keeping other values intact.
1066 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1067 if (HintTypes.size() == 0)
1070 // Reserve the first element to LoopID (see below).
1071 SmallVector<Metadata *, 4> MDs(1);
1072 // If the loop already has metadata, then ignore the existing operands.
1073 MDNode *LoopID = TheLoop->getLoopID();
1075 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1076 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1077 // If node in update list, ignore old value.
1078 if (!matchesHintMetadataName(Node, HintTypes))
1079 MDs.push_back(Node);
1083 // Now, add the missing hints.
1084 for (auto H : HintTypes)
1085 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1087 // Replace current metadata node with new one.
1088 LLVMContext &Context = TheLoop->getHeader()->getContext();
1089 MDNode *NewLoopID = MDNode::get(Context, MDs);
1090 // Set operand 0 to refer to the loop id itself.
1091 NewLoopID->replaceOperandWith(0, NewLoopID);
1093 TheLoop->setLoopID(NewLoopID);
1096 /// The loop these hints belong to.
1097 const Loop *TheLoop;
1100 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1101 const LoopVectorizeHints &Hints,
1102 const LoopAccessReport &Message) {
1103 const char *Name = Hints.vectorizeAnalysisPassName();
1104 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1107 static void emitMissedWarning(Function *F, Loop *L,
1108 const LoopVectorizeHints &LH) {
1109 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1112 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1113 if (LH.getWidth() != 1)
1114 emitLoopVectorizeWarning(
1115 F->getContext(), *F, L->getStartLoc(),
1116 "failed explicitly specified loop vectorization");
1117 else if (LH.getInterleave() != 1)
1118 emitLoopInterleaveWarning(
1119 F->getContext(), *F, L->getStartLoc(),
1120 "failed explicitly specified loop interleaving");
1124 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1125 /// to what vectorization factor.
1126 /// This class does not look at the profitability of vectorization, only the
1127 /// legality. This class has two main kinds of checks:
1128 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1129 /// will change the order of memory accesses in a way that will change the
1130 /// correctness of the program.
1131 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1132 /// checks for a number of different conditions, such as the availability of a
1133 /// single induction variable, that all types are supported and vectorize-able,
1134 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1135 /// This class is also used by InnerLoopVectorizer for identifying
1136 /// induction variable and the different reduction variables.
1137 class LoopVectorizationLegality {
1139 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1140 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1141 Function *F, const TargetTransformInfo *TTI,
1142 LoopAccessAnalysis *LAA,
1143 LoopVectorizationRequirements *R,
1144 const LoopVectorizeHints *H)
1145 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1146 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1147 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1148 Requirements(R), Hints(H) {}
1150 /// ReductionList contains the reduction descriptors for all
1151 /// of the reductions that were found in the loop.
1152 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1154 /// InductionList saves induction variables and maps them to the
1155 /// induction descriptor.
1156 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1158 /// Returns true if it is legal to vectorize this loop.
1159 /// This does not mean that it is profitable to vectorize this
1160 /// loop, only that it is legal to do so.
1161 bool canVectorize();
1163 /// Returns the Induction variable.
1164 PHINode *getInduction() { return Induction; }
1166 /// Returns the reduction variables found in the loop.
1167 ReductionList *getReductionVars() { return &Reductions; }
1169 /// Returns the induction variables found in the loop.
1170 InductionList *getInductionVars() { return &Inductions; }
1172 /// Returns the widest induction type.
1173 Type *getWidestInductionType() { return WidestIndTy; }
1175 /// Returns True if V is an induction variable in this loop.
1176 bool isInductionVariable(const Value *V);
1178 /// Return true if the block BB needs to be predicated in order for the loop
1179 /// to be vectorized.
1180 bool blockNeedsPredication(BasicBlock *BB);
1182 /// Check if this pointer is consecutive when vectorizing. This happens
1183 /// when the last index of the GEP is the induction variable, or that the
1184 /// pointer itself is an induction variable.
1185 /// This check allows us to vectorize A[idx] into a wide load/store.
1187 /// 0 - Stride is unknown or non-consecutive.
1188 /// 1 - Address is consecutive.
1189 /// -1 - Address is consecutive, and decreasing.
1190 int isConsecutivePtr(Value *Ptr);
1192 /// Returns true if the value V is uniform within the loop.
1193 bool isUniform(Value *V);
1195 /// Returns true if this instruction will remain scalar after vectorization.
1196 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1198 /// Returns the information that we collected about runtime memory check.
1199 const RuntimePointerChecking *getRuntimePointerChecking() const {
1200 return LAI->getRuntimePointerChecking();
1203 const LoopAccessInfo *getLAI() const {
1207 /// \brief Check if \p Instr belongs to any interleaved access group.
1208 bool isAccessInterleaved(Instruction *Instr) {
1209 return InterleaveInfo.isInterleaved(Instr);
1212 /// \brief Get the interleaved access group that \p Instr belongs to.
1213 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1214 return InterleaveInfo.getInterleaveGroup(Instr);
1217 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1219 bool hasStride(Value *V) { return StrideSet.count(V); }
1220 bool mustCheckStrides() { return !StrideSet.empty(); }
1221 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1222 return StrideSet.begin();
1224 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1226 /// Returns true if the target machine supports masked store operation
1227 /// for the given \p DataType and kind of access to \p Ptr.
1228 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1229 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1231 /// Returns true if the target machine supports masked load operation
1232 /// for the given \p DataType and kind of access to \p Ptr.
1233 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1234 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1236 /// Returns true if vector representation of the instruction \p I
1238 bool isMaskRequired(const Instruction* I) {
1239 return (MaskedOp.count(I) != 0);
1241 unsigned getNumStores() const {
1242 return LAI->getNumStores();
1244 unsigned getNumLoads() const {
1245 return LAI->getNumLoads();
1247 unsigned getNumPredStores() const {
1248 return NumPredStores;
1251 /// Check if a single basic block loop is vectorizable.
1252 /// At this point we know that this is a loop with a constant trip count
1253 /// and we only need to check individual instructions.
1254 bool canVectorizeInstrs();
1256 /// When we vectorize loops we may change the order in which
1257 /// we read and write from memory. This method checks if it is
1258 /// legal to vectorize the code, considering only memory constrains.
1259 /// Returns true if the loop is vectorizable
1260 bool canVectorizeMemory();
1262 /// Return true if we can vectorize this loop using the IF-conversion
1264 bool canVectorizeWithIfConvert();
1266 /// Collect the variables that need to stay uniform after vectorization.
1267 void collectLoopUniforms();
1269 /// Return true if all of the instructions in the block can be speculatively
1270 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1271 /// and we know that we can read from them without segfault.
1272 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1274 /// \brief Collect memory access with loop invariant strides.
1276 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1278 void collectStridedAccess(Value *LoadOrStoreInst);
1280 /// Report an analysis message to assist the user in diagnosing loops that are
1281 /// not vectorized. These are handled as LoopAccessReport rather than
1282 /// VectorizationReport because the << operator of VectorizationReport returns
1283 /// LoopAccessReport.
1284 void emitAnalysis(const LoopAccessReport &Message) const {
1285 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1288 unsigned NumPredStores;
1290 /// The loop that we evaluate.
1293 ScalarEvolution *SE;
1294 /// Target Library Info.
1295 TargetLibraryInfo *TLI;
1297 Function *TheFunction;
1298 /// Target Transform Info
1299 const TargetTransformInfo *TTI;
1302 // LoopAccess analysis.
1303 LoopAccessAnalysis *LAA;
1304 // And the loop-accesses info corresponding to this loop. This pointer is
1305 // null until canVectorizeMemory sets it up.
1306 const LoopAccessInfo *LAI;
1308 /// The interleave access information contains groups of interleaved accesses
1309 /// with the same stride and close to each other.
1310 InterleavedAccessInfo InterleaveInfo;
1312 // --- vectorization state --- //
1314 /// Holds the integer induction variable. This is the counter of the
1317 /// Holds the reduction variables.
1318 ReductionList Reductions;
1319 /// Holds all of the induction variables that we found in the loop.
1320 /// Notice that inductions don't need to start at zero and that induction
1321 /// variables can be pointers.
1322 InductionList Inductions;
1323 /// Holds the widest induction type encountered.
1326 /// Allowed outside users. This holds the reduction
1327 /// vars which can be accessed from outside the loop.
1328 SmallPtrSet<Value*, 4> AllowedExit;
1329 /// This set holds the variables which are known to be uniform after
1331 SmallPtrSet<Instruction*, 4> Uniforms;
1333 /// Can we assume the absence of NaNs.
1334 bool HasFunNoNaNAttr;
1336 /// Vectorization requirements that will go through late-evaluation.
1337 LoopVectorizationRequirements *Requirements;
1339 /// Used to emit an analysis of any legality issues.
1340 const LoopVectorizeHints *Hints;
1342 ValueToValueMap Strides;
1343 SmallPtrSet<Value *, 8> StrideSet;
1345 /// While vectorizing these instructions we have to generate a
1346 /// call to the appropriate masked intrinsic
1347 SmallPtrSet<const Instruction*, 8> MaskedOp;
1350 /// LoopVectorizationCostModel - estimates the expected speedups due to
1352 /// In many cases vectorization is not profitable. This can happen because of
1353 /// a number of reasons. In this class we mainly attempt to predict the
1354 /// expected speedup/slowdowns due to the supported instruction set. We use the
1355 /// TargetTransformInfo to query the different backends for the cost of
1356 /// different operations.
1357 class LoopVectorizationCostModel {
1359 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1360 LoopVectorizationLegality *Legal,
1361 const TargetTransformInfo &TTI,
1362 const TargetLibraryInfo *TLI, DemandedBits *DB,
1363 AssumptionCache *AC,
1364 const Function *F, const LoopVectorizeHints *Hints,
1365 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1366 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1367 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1369 /// Information about vectorization costs
1370 struct VectorizationFactor {
1371 unsigned Width; // Vector width with best cost
1372 unsigned Cost; // Cost of the loop with that width
1374 /// \return The most profitable vectorization factor and the cost of that VF.
1375 /// This method checks every power of two up to VF. If UserVF is not ZERO
1376 /// then this vectorization factor will be selected if vectorization is
1378 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1380 /// \return The size (in bits) of the widest type in the code that
1381 /// needs to be vectorized. We ignore values that remain scalar such as
1382 /// 64 bit loop indices.
1383 unsigned getWidestType();
1385 /// \return The desired interleave count.
1386 /// If interleave count has been specified by metadata it will be returned.
1387 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1388 /// are the selected vectorization factor and the cost of the selected VF.
1389 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1392 /// \return The most profitable unroll factor.
1393 /// This method finds the best unroll-factor based on register pressure and
1394 /// other parameters. VF and LoopCost are the selected vectorization factor
1395 /// and the cost of the selected VF.
1396 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1399 /// \brief A struct that represents some properties of the register usage
1401 struct RegisterUsage {
1402 /// Holds the number of loop invariant values that are used in the loop.
1403 unsigned LoopInvariantRegs;
1404 /// Holds the maximum number of concurrent live intervals in the loop.
1405 unsigned MaxLocalUsers;
1406 /// Holds the number of instructions in the loop.
1407 unsigned NumInstructions;
1410 /// \return information about the register usage of the loop.
1411 RegisterUsage calculateRegisterUsage();
1414 /// Returns the expected execution cost. The unit of the cost does
1415 /// not matter because we use the 'cost' units to compare different
1416 /// vector widths. The cost that is returned is *not* normalized by
1417 /// the factor width.
1418 unsigned expectedCost(unsigned VF);
1420 /// Returns the execution time cost of an instruction for a given vector
1421 /// width. Vector width of one means scalar.
1422 unsigned getInstructionCost(Instruction *I, unsigned VF);
1424 /// Returns whether the instruction is a load or store and will be a emitted
1425 /// as a vector operation.
1426 bool isConsecutiveLoadOrStore(Instruction *I);
1428 /// Report an analysis message to assist the user in diagnosing loops that are
1429 /// not vectorized. These are handled as LoopAccessReport rather than
1430 /// VectorizationReport because the << operator of VectorizationReport returns
1431 /// LoopAccessReport.
1432 void emitAnalysis(const LoopAccessReport &Message) const {
1433 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1437 /// Map of scalar integer values to the smallest bitwidth they can be legally
1438 /// represented as. The vector equivalents of these values should be truncated
1440 DenseMap<Instruction*,uint64_t> MinBWs;
1442 /// The loop that we evaluate.
1445 ScalarEvolution *SE;
1446 /// Loop Info analysis.
1448 /// Vectorization legality.
1449 LoopVectorizationLegality *Legal;
1450 /// Vector target information.
1451 const TargetTransformInfo &TTI;
1452 /// Target Library Info.
1453 const TargetLibraryInfo *TLI;
1454 /// Demanded bits analysis
1456 const Function *TheFunction;
1457 // Loop Vectorize Hint.
1458 const LoopVectorizeHints *Hints;
1459 // Values to ignore in the cost model.
1460 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1463 /// \brief This holds vectorization requirements that must be verified late in
1464 /// the process. The requirements are set by legalize and costmodel. Once
1465 /// vectorization has been determined to be possible and profitable the
1466 /// requirements can be verified by looking for metadata or compiler options.
1467 /// For example, some loops require FP commutativity which is only allowed if
1468 /// vectorization is explicitly specified or if the fast-math compiler option
1469 /// has been provided.
1470 /// Late evaluation of these requirements allows helpful diagnostics to be
1471 /// composed that tells the user what need to be done to vectorize the loop. For
1472 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1473 /// evaluation should be used only when diagnostics can generated that can be
1474 /// followed by a non-expert user.
1475 class LoopVectorizationRequirements {
1477 LoopVectorizationRequirements()
1478 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1480 void addUnsafeAlgebraInst(Instruction *I) {
1481 // First unsafe algebra instruction.
1482 if (!UnsafeAlgebraInst)
1483 UnsafeAlgebraInst = I;
1486 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1488 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1489 const char *Name = Hints.vectorizeAnalysisPassName();
1490 bool Failed = false;
1491 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1492 emitOptimizationRemarkAnalysisFPCommute(
1493 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1494 VectorizationReport() << "cannot prove it is safe to reorder "
1495 "floating-point operations");
1499 // Test if runtime memcheck thresholds are exceeded.
1500 bool PragmaThresholdReached =
1501 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1502 bool ThresholdReached =
1503 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1504 if ((ThresholdReached && !Hints.allowReordering()) ||
1505 PragmaThresholdReached) {
1506 emitOptimizationRemarkAnalysisAliasing(
1507 F->getContext(), Name, *F, L->getStartLoc(),
1508 VectorizationReport()
1509 << "cannot prove it is safe to reorder memory operations");
1510 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1518 unsigned NumRuntimePointerChecks;
1519 Instruction *UnsafeAlgebraInst;
1522 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1524 return V.push_back(&L);
1526 for (Loop *InnerL : L)
1527 addInnerLoop(*InnerL, V);
1530 /// The LoopVectorize Pass.
1531 struct LoopVectorize : public FunctionPass {
1532 /// Pass identification, replacement for typeid
1535 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1537 DisableUnrolling(NoUnrolling),
1538 AlwaysVectorize(AlwaysVectorize) {
1539 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1542 ScalarEvolution *SE;
1544 TargetTransformInfo *TTI;
1546 BlockFrequencyInfo *BFI;
1547 TargetLibraryInfo *TLI;
1550 AssumptionCache *AC;
1551 LoopAccessAnalysis *LAA;
1552 bool DisableUnrolling;
1553 bool AlwaysVectorize;
1555 BlockFrequency ColdEntryFreq;
1557 bool runOnFunction(Function &F) override {
1558 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1559 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1560 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1561 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1562 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1563 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1564 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1565 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1566 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1567 LAA = &getAnalysis<LoopAccessAnalysis>();
1568 DB = &getAnalysis<DemandedBits>();
1570 // Compute some weights outside of the loop over the loops. Compute this
1571 // using a BranchProbability to re-use its scaling math.
1572 const BranchProbability ColdProb(1, 5); // 20%
1573 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1576 // 1. the target claims to have no vector registers, and
1577 // 2. interleaving won't help ILP.
1579 // The second condition is necessary because, even if the target has no
1580 // vector registers, loop vectorization may still enable scalar
1582 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1585 // Build up a worklist of inner-loops to vectorize. This is necessary as
1586 // the act of vectorizing or partially unrolling a loop creates new loops
1587 // and can invalidate iterators across the loops.
1588 SmallVector<Loop *, 8> Worklist;
1591 addInnerLoop(*L, Worklist);
1593 LoopsAnalyzed += Worklist.size();
1595 // Now walk the identified inner loops.
1596 bool Changed = false;
1597 while (!Worklist.empty())
1598 Changed |= processLoop(Worklist.pop_back_val());
1600 // Process each loop nest in the function.
1604 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1605 SmallVector<Metadata *, 4> MDs;
1606 // Reserve first location for self reference to the LoopID metadata node.
1607 MDs.push_back(nullptr);
1608 bool IsUnrollMetadata = false;
1609 MDNode *LoopID = L->getLoopID();
1611 // First find existing loop unrolling disable metadata.
1612 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1613 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1615 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1617 S && S->getString().startswith("llvm.loop.unroll.disable");
1619 MDs.push_back(LoopID->getOperand(i));
1623 if (!IsUnrollMetadata) {
1624 // Add runtime unroll disable metadata.
1625 LLVMContext &Context = L->getHeader()->getContext();
1626 SmallVector<Metadata *, 1> DisableOperands;
1627 DisableOperands.push_back(
1628 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1629 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1630 MDs.push_back(DisableNode);
1631 MDNode *NewLoopID = MDNode::get(Context, MDs);
1632 // Set operand 0 to refer to the loop id itself.
1633 NewLoopID->replaceOperandWith(0, NewLoopID);
1634 L->setLoopID(NewLoopID);
1638 bool processLoop(Loop *L) {
1639 assert(L->empty() && "Only process inner loops.");
1642 const std::string DebugLocStr = getDebugLocString(L);
1645 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1646 << L->getHeader()->getParent()->getName() << "\" from "
1647 << DebugLocStr << "\n");
1649 LoopVectorizeHints Hints(L, DisableUnrolling);
1651 DEBUG(dbgs() << "LV: Loop hints:"
1653 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1655 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1657 : "?")) << " width=" << Hints.getWidth()
1658 << " unroll=" << Hints.getInterleave() << "\n");
1660 // Function containing loop
1661 Function *F = L->getHeader()->getParent();
1663 // Looking at the diagnostic output is the only way to determine if a loop
1664 // was vectorized (other than looking at the IR or machine code), so it
1665 // is important to generate an optimization remark for each loop. Most of
1666 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1667 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1668 // less verbose reporting vectorized loops and unvectorized loops that may
1669 // benefit from vectorization, respectively.
1671 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1672 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1676 // Check the loop for a trip count threshold:
1677 // do not vectorize loops with a tiny trip count.
1678 const unsigned TC = SE->getSmallConstantTripCount(L);
1679 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1680 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1681 << "This loop is not worth vectorizing.");
1682 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1683 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1685 DEBUG(dbgs() << "\n");
1686 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1687 << "vectorization is not beneficial "
1688 "and is not explicitly forced");
1693 // Check if it is legal to vectorize the loop.
1694 LoopVectorizationRequirements Requirements;
1695 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1696 &Requirements, &Hints);
1697 if (!LVL.canVectorize()) {
1698 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1699 emitMissedWarning(F, L, Hints);
1703 // Collect values we want to ignore in the cost model. This includes
1704 // type-promoting instructions we identified during reduction detection.
1705 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1706 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1707 for (auto &Reduction : *LVL.getReductionVars()) {
1708 RecurrenceDescriptor &RedDes = Reduction.second;
1709 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1710 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1713 // Use the cost model.
1714 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, DB, AC, F, &Hints,
1717 // Check the function attributes to find out if this function should be
1718 // optimized for size.
1719 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1722 // Compute the weighted frequency of this loop being executed and see if it
1723 // is less than 20% of the function entry baseline frequency. Note that we
1724 // always have a canonical loop here because we think we *can* vectorize.
1725 // FIXME: This is hidden behind a flag due to pervasive problems with
1726 // exactly what block frequency models.
1727 if (LoopVectorizeWithBlockFrequency) {
1728 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1729 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1730 LoopEntryFreq < ColdEntryFreq)
1734 // Check the function attributes to see if implicit floats are allowed.
1735 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1736 // an integer loop and the vector instructions selected are purely integer
1737 // vector instructions?
1738 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1739 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1740 "attribute is used.\n");
1743 VectorizationReport()
1744 << "loop not vectorized due to NoImplicitFloat attribute");
1745 emitMissedWarning(F, L, Hints);
1749 // Select the optimal vectorization factor.
1750 const LoopVectorizationCostModel::VectorizationFactor VF =
1751 CM.selectVectorizationFactor(OptForSize);
1753 // Select the interleave count.
1754 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1756 // Get user interleave count.
1757 unsigned UserIC = Hints.getInterleave();
1759 // Identify the diagnostic messages that should be produced.
1760 std::string VecDiagMsg, IntDiagMsg;
1761 bool VectorizeLoop = true, InterleaveLoop = true;
1763 if (Requirements.doesNotMeet(F, L, Hints)) {
1764 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1766 emitMissedWarning(F, L, Hints);
1770 if (VF.Width == 1) {
1771 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1773 "the cost-model indicates that vectorization is not beneficial";
1774 VectorizeLoop = false;
1777 if (IC == 1 && UserIC <= 1) {
1778 // Tell the user interleaving is not beneficial.
1779 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1781 "the cost-model indicates that interleaving is not beneficial";
1782 InterleaveLoop = false;
1785 " and is explicitly disabled or interleave count is set to 1";
1786 } else if (IC > 1 && UserIC == 1) {
1787 // Tell the user interleaving is beneficial, but it explicitly disabled.
1789 << "LV: Interleaving is beneficial but is explicitly disabled.");
1790 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1791 "but is explicitly disabled or interleave count is set to 1";
1792 InterleaveLoop = false;
1795 // Override IC if user provided an interleave count.
1796 IC = UserIC > 0 ? UserIC : IC;
1798 // Emit diagnostic messages, if any.
1799 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1800 if (!VectorizeLoop && !InterleaveLoop) {
1801 // Do not vectorize or interleaving the loop.
1802 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1803 L->getStartLoc(), VecDiagMsg);
1804 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1805 L->getStartLoc(), IntDiagMsg);
1807 } else if (!VectorizeLoop && InterleaveLoop) {
1808 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1809 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1810 L->getStartLoc(), VecDiagMsg);
1811 } else if (VectorizeLoop && !InterleaveLoop) {
1812 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1813 << DebugLocStr << '\n');
1814 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1815 L->getStartLoc(), IntDiagMsg);
1816 } else if (VectorizeLoop && InterleaveLoop) {
1817 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1818 << DebugLocStr << '\n');
1819 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1822 if (!VectorizeLoop) {
1823 assert(IC > 1 && "interleave count should not be 1 or 0");
1824 // If we decided that it is not legal to vectorize the loop then
1826 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1827 Unroller.vectorize(&LVL, CM.MinBWs);
1829 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1830 Twine("interleaved loop (interleaved count: ") +
1833 // If we decided that it is *legal* to vectorize the loop then do it.
1834 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1835 LB.vectorize(&LVL, CM.MinBWs);
1838 // Add metadata to disable runtime unrolling scalar loop when there's no
1839 // runtime check about strides and memory. Because at this situation,
1840 // scalar loop is rarely used not worthy to be unrolled.
1841 if (!LB.IsSafetyChecksAdded())
1842 AddRuntimeUnrollDisableMetaData(L);
1844 // Report the vectorization decision.
1845 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1846 Twine("vectorized loop (vectorization width: ") +
1847 Twine(VF.Width) + ", interleaved count: " +
1851 // Mark the loop as already vectorized to avoid vectorizing again.
1852 Hints.setAlreadyVectorized();
1854 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1858 void getAnalysisUsage(AnalysisUsage &AU) const override {
1859 AU.addRequired<AssumptionCacheTracker>();
1860 AU.addRequiredID(LoopSimplifyID);
1861 AU.addRequiredID(LCSSAID);
1862 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1863 AU.addRequired<DominatorTreeWrapperPass>();
1864 AU.addRequired<LoopInfoWrapperPass>();
1865 AU.addRequired<ScalarEvolutionWrapperPass>();
1866 AU.addRequired<TargetTransformInfoWrapperPass>();
1867 AU.addRequired<AAResultsWrapperPass>();
1868 AU.addRequired<LoopAccessAnalysis>();
1869 AU.addRequired<DemandedBits>();
1870 AU.addPreserved<LoopInfoWrapperPass>();
1871 AU.addPreserved<DominatorTreeWrapperPass>();
1872 AU.addPreserved<BasicAAWrapperPass>();
1873 AU.addPreserved<AAResultsWrapperPass>();
1874 AU.addPreserved<GlobalsAAWrapperPass>();
1879 } // end anonymous namespace
1881 //===----------------------------------------------------------------------===//
1882 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1883 // LoopVectorizationCostModel.
1884 //===----------------------------------------------------------------------===//
1886 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1887 // We need to place the broadcast of invariant variables outside the loop.
1888 Instruction *Instr = dyn_cast<Instruction>(V);
1890 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1891 Instr->getParent()) != LoopVectorBody.end());
1892 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1894 // Place the code for broadcasting invariant variables in the new preheader.
1895 IRBuilder<>::InsertPointGuard Guard(Builder);
1897 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1899 // Broadcast the scalar into all locations in the vector.
1900 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1905 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1907 assert(Val->getType()->isVectorTy() && "Must be a vector");
1908 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1909 "Elem must be an integer");
1910 assert(Step->getType() == Val->getType()->getScalarType() &&
1911 "Step has wrong type");
1912 // Create the types.
1913 Type *ITy = Val->getType()->getScalarType();
1914 VectorType *Ty = cast<VectorType>(Val->getType());
1915 int VLen = Ty->getNumElements();
1916 SmallVector<Constant*, 8> Indices;
1918 // Create a vector of consecutive numbers from zero to VF.
1919 for (int i = 0; i < VLen; ++i)
1920 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1922 // Add the consecutive indices to the vector value.
1923 Constant *Cv = ConstantVector::get(Indices);
1924 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1925 Step = Builder.CreateVectorSplat(VLen, Step);
1926 assert(Step->getType() == Val->getType() && "Invalid step vec");
1927 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1928 // which can be found from the original scalar operations.
1929 Step = Builder.CreateMul(Cv, Step);
1930 return Builder.CreateAdd(Val, Step, "induction");
1933 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1934 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1935 // Make sure that the pointer does not point to structs.
1936 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1939 // If this value is a pointer induction variable we know it is consecutive.
1940 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1941 if (Phi && Inductions.count(Phi)) {
1942 InductionDescriptor II = Inductions[Phi];
1943 return II.getConsecutiveDirection();
1946 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1950 unsigned NumOperands = Gep->getNumOperands();
1951 Value *GpPtr = Gep->getPointerOperand();
1952 // If this GEP value is a consecutive pointer induction variable and all of
1953 // the indices are constant then we know it is consecutive. We can
1954 Phi = dyn_cast<PHINode>(GpPtr);
1955 if (Phi && Inductions.count(Phi)) {
1957 // Make sure that the pointer does not point to structs.
1958 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1959 if (GepPtrType->getElementType()->isAggregateType())
1962 // Make sure that all of the index operands are loop invariant.
1963 for (unsigned i = 1; i < NumOperands; ++i)
1964 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1967 InductionDescriptor II = Inductions[Phi];
1968 return II.getConsecutiveDirection();
1971 unsigned InductionOperand = getGEPInductionOperand(Gep);
1973 // Check that all of the gep indices are uniform except for our induction
1975 for (unsigned i = 0; i != NumOperands; ++i)
1976 if (i != InductionOperand &&
1977 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1980 // We can emit wide load/stores only if the last non-zero index is the
1981 // induction variable.
1982 const SCEV *Last = nullptr;
1983 if (!Strides.count(Gep))
1984 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1986 // Because of the multiplication by a stride we can have a s/zext cast.
1987 // We are going to replace this stride by 1 so the cast is safe to ignore.
1989 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1990 // %0 = trunc i64 %indvars.iv to i32
1991 // %mul = mul i32 %0, %Stride1
1992 // %idxprom = zext i32 %mul to i64 << Safe cast.
1993 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1995 Last = replaceSymbolicStrideSCEV(SE, Strides,
1996 Gep->getOperand(InductionOperand), Gep);
1997 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1999 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2003 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2004 const SCEV *Step = AR->getStepRecurrence(*SE);
2006 // The memory is consecutive because the last index is consecutive
2007 // and all other indices are loop invariant.
2010 if (Step->isAllOnesValue())
2017 bool LoopVectorizationLegality::isUniform(Value *V) {
2018 return LAI->isUniform(V);
2021 InnerLoopVectorizer::VectorParts&
2022 InnerLoopVectorizer::getVectorValue(Value *V) {
2023 assert(V != Induction && "The new induction variable should not be used.");
2024 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2026 // If we have a stride that is replaced by one, do it here.
2027 if (Legal->hasStride(V))
2028 V = ConstantInt::get(V->getType(), 1);
2030 // If we have this scalar in the map, return it.
2031 if (WidenMap.has(V))
2032 return WidenMap.get(V);
2034 // If this scalar is unknown, assume that it is a constant or that it is
2035 // loop invariant. Broadcast V and save the value for future uses.
2036 Value *B = getBroadcastInstrs(V);
2038 return WidenMap.splat(V, B);
2041 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2042 assert(Vec->getType()->isVectorTy() && "Invalid type");
2043 SmallVector<Constant*, 8> ShuffleMask;
2044 for (unsigned i = 0; i < VF; ++i)
2045 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2047 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2048 ConstantVector::get(ShuffleMask),
2052 // Get a mask to interleave \p NumVec vectors into a wide vector.
2053 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2054 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2055 // <0, 4, 1, 5, 2, 6, 3, 7>
2056 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2058 SmallVector<Constant *, 16> Mask;
2059 for (unsigned i = 0; i < VF; i++)
2060 for (unsigned j = 0; j < NumVec; j++)
2061 Mask.push_back(Builder.getInt32(j * VF + i));
2063 return ConstantVector::get(Mask);
2066 // Get the strided mask starting from index \p Start.
2067 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2068 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2069 unsigned Stride, unsigned VF) {
2070 SmallVector<Constant *, 16> Mask;
2071 for (unsigned i = 0; i < VF; i++)
2072 Mask.push_back(Builder.getInt32(Start + i * Stride));
2074 return ConstantVector::get(Mask);
2077 // Get a mask of two parts: The first part consists of sequential integers
2078 // starting from 0, The second part consists of UNDEFs.
2079 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2080 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2081 unsigned NumUndef) {
2082 SmallVector<Constant *, 16> Mask;
2083 for (unsigned i = 0; i < NumInt; i++)
2084 Mask.push_back(Builder.getInt32(i));
2086 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2087 for (unsigned i = 0; i < NumUndef; i++)
2088 Mask.push_back(Undef);
2090 return ConstantVector::get(Mask);
2093 // Concatenate two vectors with the same element type. The 2nd vector should
2094 // not have more elements than the 1st vector. If the 2nd vector has less
2095 // elements, extend it with UNDEFs.
2096 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2098 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2099 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2100 assert(VecTy1 && VecTy2 &&
2101 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2102 "Expect two vectors with the same element type");
2104 unsigned NumElts1 = VecTy1->getNumElements();
2105 unsigned NumElts2 = VecTy2->getNumElements();
2106 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2108 if (NumElts1 > NumElts2) {
2109 // Extend with UNDEFs.
2111 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2112 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2115 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2116 return Builder.CreateShuffleVector(V1, V2, Mask);
2119 // Concatenate vectors in the given list. All vectors have the same type.
2120 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2121 ArrayRef<Value *> InputList) {
2122 unsigned NumVec = InputList.size();
2123 assert(NumVec > 1 && "Should be at least two vectors");
2125 SmallVector<Value *, 8> ResList;
2126 ResList.append(InputList.begin(), InputList.end());
2128 SmallVector<Value *, 8> TmpList;
2129 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2130 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2131 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2132 "Only the last vector may have a different type");
2134 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2137 // Push the last vector if the total number of vectors is odd.
2138 if (NumVec % 2 != 0)
2139 TmpList.push_back(ResList[NumVec - 1]);
2142 NumVec = ResList.size();
2143 } while (NumVec > 1);
2148 // Try to vectorize the interleave group that \p Instr belongs to.
2150 // E.g. Translate following interleaved load group (factor = 3):
2151 // for (i = 0; i < N; i+=3) {
2152 // R = Pic[i]; // Member of index 0
2153 // G = Pic[i+1]; // Member of index 1
2154 // B = Pic[i+2]; // Member of index 2
2155 // ... // do something to R, G, B
2158 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2159 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2160 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2161 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2163 // Or translate following interleaved store group (factor = 3):
2164 // for (i = 0; i < N; i+=3) {
2165 // ... do something to R, G, B
2166 // Pic[i] = R; // Member of index 0
2167 // Pic[i+1] = G; // Member of index 1
2168 // Pic[i+2] = B; // Member of index 2
2171 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2172 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2173 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2174 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2175 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2176 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2177 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2178 assert(Group && "Fail to get an interleaved access group.");
2180 // Skip if current instruction is not the insert position.
2181 if (Instr != Group->getInsertPos())
2184 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2185 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2186 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2188 // Prepare for the vector type of the interleaved load/store.
2189 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2190 unsigned InterleaveFactor = Group->getFactor();
2191 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2192 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2194 // Prepare for the new pointers.
2195 setDebugLocFromInst(Builder, Ptr);
2196 VectorParts &PtrParts = getVectorValue(Ptr);
2197 SmallVector<Value *, 2> NewPtrs;
2198 unsigned Index = Group->getIndex(Instr);
2199 for (unsigned Part = 0; Part < UF; Part++) {
2200 // Extract the pointer for current instruction from the pointer vector. A
2201 // reverse access uses the pointer in the last lane.
2202 Value *NewPtr = Builder.CreateExtractElement(
2204 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2206 // Notice current instruction could be any index. Need to adjust the address
2207 // to the member of index 0.
2209 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2210 // b = A[i]; // Member of index 0
2211 // Current pointer is pointed to A[i+1], adjust it to A[i].
2213 // E.g. A[i+1] = a; // Member of index 1
2214 // A[i] = b; // Member of index 0
2215 // A[i+2] = c; // Member of index 2 (Current instruction)
2216 // Current pointer is pointed to A[i+2], adjust it to A[i].
2217 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2219 // Cast to the vector pointer type.
2220 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2223 setDebugLocFromInst(Builder, Instr);
2224 Value *UndefVec = UndefValue::get(VecTy);
2226 // Vectorize the interleaved load group.
2228 for (unsigned Part = 0; Part < UF; Part++) {
2229 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2230 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2232 for (unsigned i = 0; i < InterleaveFactor; i++) {
2233 Instruction *Member = Group->getMember(i);
2235 // Skip the gaps in the group.
2239 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2240 Value *StridedVec = Builder.CreateShuffleVector(
2241 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2243 // If this member has different type, cast the result type.
2244 if (Member->getType() != ScalarTy) {
2245 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2246 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2249 VectorParts &Entry = WidenMap.get(Member);
2251 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2254 propagateMetadata(NewLoadInstr, Instr);
2259 // The sub vector type for current instruction.
2260 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2262 // Vectorize the interleaved store group.
2263 for (unsigned Part = 0; Part < UF; Part++) {
2264 // Collect the stored vector from each member.
2265 SmallVector<Value *, 4> StoredVecs;
2266 for (unsigned i = 0; i < InterleaveFactor; i++) {
2267 // Interleaved store group doesn't allow a gap, so each index has a member
2268 Instruction *Member = Group->getMember(i);
2269 assert(Member && "Fail to get a member from an interleaved store group");
2272 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2273 if (Group->isReverse())
2274 StoredVec = reverseVector(StoredVec);
2276 // If this member has different type, cast it to an unified type.
2277 if (StoredVec->getType() != SubVT)
2278 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2280 StoredVecs.push_back(StoredVec);
2283 // Concatenate all vectors into a wide vector.
2284 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2286 // Interleave the elements in the wide vector.
2287 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2288 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2291 Instruction *NewStoreInstr =
2292 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2293 propagateMetadata(NewStoreInstr, Instr);
2297 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2298 // Attempt to issue a wide load.
2299 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2300 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2302 assert((LI || SI) && "Invalid Load/Store instruction");
2304 // Try to vectorize the interleave group if this access is interleaved.
2305 if (Legal->isAccessInterleaved(Instr))
2306 return vectorizeInterleaveGroup(Instr);
2308 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2309 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2310 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2311 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2312 // An alignment of 0 means target abi alignment. We need to use the scalar's
2313 // target abi alignment in such a case.
2314 const DataLayout &DL = Instr->getModule()->getDataLayout();
2316 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2317 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2318 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2319 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2321 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2322 !Legal->isMaskRequired(SI))
2323 return scalarizeInstruction(Instr, true);
2325 if (ScalarAllocatedSize != VectorElementSize)
2326 return scalarizeInstruction(Instr);
2328 // If the pointer is loop invariant or if it is non-consecutive,
2329 // scalarize the load.
2330 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2331 bool Reverse = ConsecutiveStride < 0;
2332 bool UniformLoad = LI && Legal->isUniform(Ptr);
2333 if (!ConsecutiveStride || UniformLoad)
2334 return scalarizeInstruction(Instr);
2336 Constant *Zero = Builder.getInt32(0);
2337 VectorParts &Entry = WidenMap.get(Instr);
2339 // Handle consecutive loads/stores.
2340 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2341 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2342 setDebugLocFromInst(Builder, Gep);
2343 Value *PtrOperand = Gep->getPointerOperand();
2344 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2345 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2347 // Create the new GEP with the new induction variable.
2348 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2349 Gep2->setOperand(0, FirstBasePtr);
2350 Gep2->setName("gep.indvar.base");
2351 Ptr = Builder.Insert(Gep2);
2353 setDebugLocFromInst(Builder, Gep);
2354 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2355 OrigLoop) && "Base ptr must be invariant");
2357 // The last index does not have to be the induction. It can be
2358 // consecutive and be a function of the index. For example A[I+1];
2359 unsigned NumOperands = Gep->getNumOperands();
2360 unsigned InductionOperand = getGEPInductionOperand(Gep);
2361 // Create the new GEP with the new induction variable.
2362 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2364 for (unsigned i = 0; i < NumOperands; ++i) {
2365 Value *GepOperand = Gep->getOperand(i);
2366 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2368 // Update last index or loop invariant instruction anchored in loop.
2369 if (i == InductionOperand ||
2370 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2371 assert((i == InductionOperand ||
2372 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2373 "Must be last index or loop invariant");
2375 VectorParts &GEPParts = getVectorValue(GepOperand);
2376 Value *Index = GEPParts[0];
2377 Index = Builder.CreateExtractElement(Index, Zero);
2378 Gep2->setOperand(i, Index);
2379 Gep2->setName("gep.indvar.idx");
2382 Ptr = Builder.Insert(Gep2);
2384 // Use the induction element ptr.
2385 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2386 setDebugLocFromInst(Builder, Ptr);
2387 VectorParts &PtrVal = getVectorValue(Ptr);
2388 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2391 VectorParts Mask = createBlockInMask(Instr->getParent());
2394 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2395 "We do not allow storing to uniform addresses");
2396 setDebugLocFromInst(Builder, SI);
2397 // We don't want to update the value in the map as it might be used in
2398 // another expression. So don't use a reference type for "StoredVal".
2399 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2401 for (unsigned Part = 0; Part < UF; ++Part) {
2402 // Calculate the pointer for the specific unroll-part.
2404 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2407 // If we store to reverse consecutive memory locations, then we need
2408 // to reverse the order of elements in the stored value.
2409 StoredVal[Part] = reverseVector(StoredVal[Part]);
2410 // If the address is consecutive but reversed, then the
2411 // wide store needs to start at the last vector element.
2412 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2413 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2414 Mask[Part] = reverseVector(Mask[Part]);
2417 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2418 DataTy->getPointerTo(AddressSpace));
2421 if (Legal->isMaskRequired(SI))
2422 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2425 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2426 propagateMetadata(NewSI, SI);
2432 assert(LI && "Must have a load instruction");
2433 setDebugLocFromInst(Builder, LI);
2434 for (unsigned Part = 0; Part < UF; ++Part) {
2435 // Calculate the pointer for the specific unroll-part.
2437 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2440 // If the address is consecutive but reversed, then the
2441 // wide load needs to start at the last vector element.
2442 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2443 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2444 Mask[Part] = reverseVector(Mask[Part]);
2448 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2449 DataTy->getPointerTo(AddressSpace));
2450 if (Legal->isMaskRequired(LI))
2451 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2452 UndefValue::get(DataTy),
2453 "wide.masked.load");
2455 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2456 propagateMetadata(NewLI, LI);
2457 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2461 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2462 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2463 // Holds vector parameters or scalars, in case of uniform vals.
2464 SmallVector<VectorParts, 4> Params;
2466 setDebugLocFromInst(Builder, Instr);
2468 // Find all of the vectorized parameters.
2469 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2470 Value *SrcOp = Instr->getOperand(op);
2472 // If we are accessing the old induction variable, use the new one.
2473 if (SrcOp == OldInduction) {
2474 Params.push_back(getVectorValue(SrcOp));
2478 // Try using previously calculated values.
2479 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2481 // If the src is an instruction that appeared earlier in the basic block,
2482 // then it should already be vectorized.
2483 if (SrcInst && OrigLoop->contains(SrcInst)) {
2484 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2485 // The parameter is a vector value from earlier.
2486 Params.push_back(WidenMap.get(SrcInst));
2488 // The parameter is a scalar from outside the loop. Maybe even a constant.
2489 VectorParts Scalars;
2490 Scalars.append(UF, SrcOp);
2491 Params.push_back(Scalars);
2495 assert(Params.size() == Instr->getNumOperands() &&
2496 "Invalid number of operands");
2498 // Does this instruction return a value ?
2499 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2501 Value *UndefVec = IsVoidRetTy ? nullptr :
2502 UndefValue::get(VectorType::get(Instr->getType(), VF));
2503 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2504 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2507 if (IfPredicateStore) {
2508 assert(Instr->getParent()->getSinglePredecessor() &&
2509 "Only support single predecessor blocks");
2510 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2511 Instr->getParent());
2514 // For each vector unroll 'part':
2515 for (unsigned Part = 0; Part < UF; ++Part) {
2516 // For each scalar that we create:
2517 for (unsigned Width = 0; Width < VF; ++Width) {
2520 Value *Cmp = nullptr;
2521 if (IfPredicateStore) {
2522 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2523 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2526 Instruction *Cloned = Instr->clone();
2528 Cloned->setName(Instr->getName() + ".cloned");
2529 // Replace the operands of the cloned instructions with extracted scalars.
2530 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2531 Value *Op = Params[op][Part];
2532 // Param is a vector. Need to extract the right lane.
2533 if (Op->getType()->isVectorTy())
2534 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2535 Cloned->setOperand(op, Op);
2538 // Place the cloned scalar in the new loop.
2539 Builder.Insert(Cloned);
2541 // If the original scalar returns a value we need to place it in a vector
2542 // so that future users will be able to use it.
2544 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2545 Builder.getInt32(Width));
2547 if (IfPredicateStore)
2548 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
2554 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2558 if (Instruction *I = dyn_cast<Instruction>(V))
2559 return I->getParent() == Loc->getParent() ? I : nullptr;
2563 std::pair<Instruction *, Instruction *>
2564 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2565 Instruction *tnullptr = nullptr;
2566 if (!Legal->mustCheckStrides())
2567 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2569 IRBuilder<> ChkBuilder(Loc);
2572 Value *Check = nullptr;
2573 Instruction *FirstInst = nullptr;
2574 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2575 SE = Legal->strides_end();
2577 Value *Ptr = stripIntegerCast(*SI);
2578 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2580 // Store the first instruction we create.
2581 FirstInst = getFirstInst(FirstInst, C, Loc);
2583 Check = ChkBuilder.CreateOr(Check, C);
2588 // We have to do this trickery because the IRBuilder might fold the check to a
2589 // constant expression in which case there is no Instruction anchored in a
2591 LLVMContext &Ctx = Loc->getContext();
2592 Instruction *TheCheck =
2593 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2594 ChkBuilder.Insert(TheCheck, "stride.not.one");
2595 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2597 return std::make_pair(FirstInst, TheCheck);
2600 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L,
2605 BasicBlock *Header = L->getHeader();
2606 BasicBlock *Latch = L->getLoopLatch();
2607 // As we're just creating this loop, it's possible no latch exists
2608 // yet. If so, use the header as this will be a single block loop.
2612 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
2613 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2614 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2616 Builder.SetInsertPoint(Latch->getTerminator());
2618 // Create i+1 and fill the PHINode.
2619 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2620 Induction->addIncoming(Start, L->getLoopPreheader());
2621 Induction->addIncoming(Next, Latch);
2622 // Create the compare.
2623 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2624 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2626 // Now we have two terminators. Remove the old one from the block.
2627 Latch->getTerminator()->eraseFromParent();
2632 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2636 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2637 // Find the loop boundaries.
2638 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(OrigLoop);
2639 assert(BackedgeTakenCount != SE->getCouldNotCompute() && "Invalid loop count");
2641 Type *IdxTy = Legal->getWidestInductionType();
2643 // The exit count might have the type of i64 while the phi is i32. This can
2644 // happen if we have an induction variable that is sign extended before the
2645 // compare. The only way that we get a backedge taken count is that the
2646 // induction variable was signed and as such will not overflow. In such a case
2647 // truncation is legal.
2648 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
2649 IdxTy->getPrimitiveSizeInBits())
2650 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
2651 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
2653 // Get the total trip count from the count by adding 1.
2654 const SCEV *ExitCount = SE->getAddExpr(
2655 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
2657 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2659 // Expand the trip count and place the new instructions in the preheader.
2660 // Notice that the pre-header does not change, only the loop body.
2661 SCEVExpander Exp(*SE, DL, "induction");
2663 // Count holds the overall loop count (N).
2664 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2665 L->getLoopPreheader()->getTerminator());
2667 if (TripCount->getType()->isPointerTy())
2669 CastInst::CreatePointerCast(TripCount, IdxTy,
2670 "exitcount.ptrcnt.to.int",
2671 L->getLoopPreheader()->getTerminator());
2676 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2677 if (VectorTripCount)
2678 return VectorTripCount;
2680 Value *TC = getOrCreateTripCount(L);
2681 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2683 // Now we need to generate the expression for N - (N % VF), which is
2684 // the part that the vectorized body will execute.
2685 // The loop step is equal to the vectorization factor (num of SIMD elements)
2686 // times the unroll factor (num of SIMD instructions).
2687 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2688 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2689 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2691 return VectorTripCount;
2694 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
2695 BasicBlock *Bypass) {
2696 Value *Count = getOrCreateTripCount(L);
2697 BasicBlock *BB = L->getLoopPreheader();
2698 IRBuilder<> Builder(BB->getTerminator());
2700 // Generate code to check that the loop's trip count that we computed by
2701 // adding one to the backedge-taken count will not overflow.
2702 Value *CheckMinIters =
2703 Builder.CreateICmpULT(Count,
2704 ConstantInt::get(Count->getType(), VF * UF),
2707 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2708 "min.iters.checked");
2709 if (L->getParentLoop())
2710 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2711 ReplaceInstWithInst(BB->getTerminator(),
2712 BranchInst::Create(Bypass, NewBB, CheckMinIters));
2713 LoopBypassBlocks.push_back(BB);
2716 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
2717 BasicBlock *Bypass) {
2718 Value *TC = getOrCreateVectorTripCount(L);
2719 BasicBlock *BB = L->getLoopPreheader();
2720 IRBuilder<> Builder(BB->getTerminator());
2722 // Now, compare the new count to zero. If it is zero skip the vector loop and
2723 // jump to the scalar loop.
2724 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
2727 // Generate code to check that the loop's trip count that we computed by
2728 // adding one to the backedge-taken count will not overflow.
2729 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2731 if (L->getParentLoop())
2732 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2733 ReplaceInstWithInst(BB->getTerminator(),
2734 BranchInst::Create(Bypass, NewBB, Cmp));
2735 LoopBypassBlocks.push_back(BB);
2738 void InnerLoopVectorizer::emitStrideChecks(Loop *L,
2739 BasicBlock *Bypass) {
2740 BasicBlock *BB = L->getLoopPreheader();
2742 // Generate the code to check that the strides we assumed to be one are really
2743 // one. We want the new basic block to start at the first instruction in a
2744 // sequence of instructions that form a check.
2745 Instruction *StrideCheck;
2746 Instruction *FirstCheckInst;
2747 std::tie(FirstCheckInst, StrideCheck) = addStrideCheck(BB->getTerminator());
2751 // Create a new block containing the stride check.
2752 BB->setName("vector.stridecheck");
2753 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2754 if (L->getParentLoop())
2755 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2756 ReplaceInstWithInst(BB->getTerminator(),
2757 BranchInst::Create(Bypass, NewBB, StrideCheck));
2758 LoopBypassBlocks.push_back(BB);
2759 AddedSafetyChecks = true;
2762 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L,
2763 BasicBlock *Bypass) {
2764 BasicBlock *BB = L->getLoopPreheader();
2766 // Generate the code that checks in runtime if arrays overlap. We put the
2767 // checks into a separate block to make the more common case of few elements
2769 Instruction *FirstCheckInst;
2770 Instruction *MemRuntimeCheck;
2771 std::tie(FirstCheckInst, MemRuntimeCheck) =
2772 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
2773 if (!MemRuntimeCheck)
2776 // Create a new block containing the memory check.
2777 BB->setName("vector.memcheck");
2778 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2779 if (L->getParentLoop())
2780 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2781 ReplaceInstWithInst(BB->getTerminator(),
2782 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
2783 LoopBypassBlocks.push_back(BB);
2784 AddedSafetyChecks = true;
2788 void InnerLoopVectorizer::createEmptyLoop() {
2790 In this function we generate a new loop. The new loop will contain
2791 the vectorized instructions while the old loop will continue to run the
2794 [ ] <-- loop iteration number check.
2797 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2800 || [ ] <-- vector pre header.
2804 | [ ]_| <-- vector loop.
2807 | -[ ] <--- middle-block.
2810 -|- >[ ] <--- new preheader.
2814 | [ ]_| <-- old scalar loop to handle remainder.
2817 >[ ] <-- exit block.
2821 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2822 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2823 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2824 assert(VectorPH && "Invalid loop structure");
2825 assert(ExitBlock && "Must have an exit block");
2827 // Some loops have a single integer induction variable, while other loops
2828 // don't. One example is c++ iterators that often have multiple pointer
2829 // induction variables. In the code below we also support a case where we
2830 // don't have a single induction variable.
2832 // We try to obtain an induction variable from the original loop as hard
2833 // as possible. However if we don't find one that:
2835 // - counts from zero, stepping by one
2836 // - is the size of the widest induction variable type
2837 // then we create a new one.
2838 OldInduction = Legal->getInduction();
2839 Type *IdxTy = Legal->getWidestInductionType();
2841 // Split the single block loop into the two loop structure described above.
2842 BasicBlock *VecBody =
2843 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2844 BasicBlock *MiddleBlock =
2845 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2846 BasicBlock *ScalarPH =
2847 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2849 // Create and register the new vector loop.
2850 Loop* Lp = new Loop();
2851 Loop *ParentLoop = OrigLoop->getParentLoop();
2853 // Insert the new loop into the loop nest and register the new basic blocks
2854 // before calling any utilities such as SCEV that require valid LoopInfo.
2856 ParentLoop->addChildLoop(Lp);
2857 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2858 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2860 LI->addTopLevelLoop(Lp);
2862 Lp->addBasicBlockToLoop(VecBody, *LI);
2864 // Find the loop boundaries.
2865 Value *Count = getOrCreateTripCount(Lp);
2867 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2869 // We need to test whether the backedge-taken count is uint##_max. Adding one
2870 // to it will cause overflow and an incorrect loop trip count in the vector
2871 // body. In case of overflow we want to directly jump to the scalar remainder
2873 emitMinimumIterationCountCheck(Lp, ScalarPH);
2874 // Now, compare the new count to zero. If it is zero skip the vector loop and
2875 // jump to the scalar loop.
2876 emitVectorLoopEnteredCheck(Lp, ScalarPH);
2877 // Generate the code to check that the strides we assumed to be one are really
2878 // one. We want the new basic block to start at the first instruction in a
2879 // sequence of instructions that form a check.
2880 emitStrideChecks(Lp, ScalarPH);
2881 // Generate the code that checks in runtime if arrays overlap. We put the
2882 // checks into a separate block to make the more common case of few elements
2884 emitMemRuntimeChecks(Lp, ScalarPH);
2886 // Generate the induction variable.
2887 // The loop step is equal to the vectorization factor (num of SIMD elements)
2888 // times the unroll factor (num of SIMD instructions).
2889 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
2890 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2892 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
2893 getDebugLocFromInstOrOperands(OldInduction));
2895 // We are going to resume the execution of the scalar loop.
2896 // Go over all of the induction variables that we found and fix the
2897 // PHIs that are left in the scalar version of the loop.
2898 // The starting values of PHI nodes depend on the counter of the last
2899 // iteration in the vectorized loop.
2900 // If we come from a bypass edge then we need to start from the original
2903 // This variable saves the new starting index for the scalar loop. It is used
2904 // to test if there are any tail iterations left once the vector loop has
2906 LoopVectorizationLegality::InductionList::iterator I, E;
2907 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2908 for (I = List->begin(), E = List->end(); I != E; ++I) {
2909 PHINode *OrigPhi = I->first;
2910 InductionDescriptor II = I->second;
2912 // Create phi nodes to merge from the backedge-taken check block.
2913 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2915 ScalarPH->getTerminator());
2917 if (OrigPhi == OldInduction) {
2918 // We know what the end value is.
2919 EndValue = CountRoundDown;
2921 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
2922 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
2923 II.getStepValue()->getType(),
2925 EndValue = II.transform(B, CRD);
2926 EndValue->setName("ind.end");
2929 // The new PHI merges the original incoming value, in case of a bypass,
2930 // or the value at the end of the vectorized loop.
2931 BCResumeVal->addIncoming(EndValue, MiddleBlock);
2933 // Fix the scalar body counter (PHI node).
2934 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2936 // The old induction's phi node in the scalar body needs the truncated
2938 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2939 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2940 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2943 // Add a check in the middle block to see if we have completed
2944 // all of the iterations in the first vector loop.
2945 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2946 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
2947 CountRoundDown, "cmp.n",
2948 MiddleBlock->getTerminator());
2949 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2950 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2952 // Get ready to start creating new instructions into the vectorized body.
2953 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
2956 LoopVectorPreHeader = Lp->getLoopPreheader();
2957 LoopScalarPreHeader = ScalarPH;
2958 LoopMiddleBlock = MiddleBlock;
2959 LoopExitBlock = ExitBlock;
2960 LoopVectorBody.push_back(VecBody);
2961 LoopScalarBody = OldBasicBlock;
2963 LoopVectorizeHints Hints(Lp, true);
2964 Hints.setAlreadyVectorized();
2968 struct CSEDenseMapInfo {
2969 static bool canHandle(Instruction *I) {
2970 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2971 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2973 static inline Instruction *getEmptyKey() {
2974 return DenseMapInfo<Instruction *>::getEmptyKey();
2976 static inline Instruction *getTombstoneKey() {
2977 return DenseMapInfo<Instruction *>::getTombstoneKey();
2979 static unsigned getHashValue(Instruction *I) {
2980 assert(canHandle(I) && "Unknown instruction!");
2981 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2982 I->value_op_end()));
2984 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2985 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2986 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2988 return LHS->isIdenticalTo(RHS);
2993 /// \brief Check whether this block is a predicated block.
2994 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2995 /// = ...; " blocks. We start with one vectorized basic block. For every
2996 /// conditional block we split this vectorized block. Therefore, every second
2997 /// block will be a predicated one.
2998 static bool isPredicatedBlock(unsigned BlockNum) {
2999 return BlockNum % 2;
3002 ///\brief Perform cse of induction variable instructions.
3003 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
3004 // Perform simple cse.
3005 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3006 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3007 BasicBlock *BB = BBs[i];
3008 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3009 Instruction *In = &*I++;
3011 if (!CSEDenseMapInfo::canHandle(In))
3014 // Check if we can replace this instruction with any of the
3015 // visited instructions.
3016 if (Instruction *V = CSEMap.lookup(In)) {
3017 In->replaceAllUsesWith(V);
3018 In->eraseFromParent();
3021 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3022 // ...;" blocks for predicated stores. Every second block is a predicated
3024 if (isPredicatedBlock(i))
3032 /// \brief Adds a 'fast' flag to floating point operations.
3033 static Value *addFastMathFlag(Value *V) {
3034 if (isa<FPMathOperator>(V)){
3035 FastMathFlags Flags;
3036 Flags.setUnsafeAlgebra();
3037 cast<Instruction>(V)->setFastMathFlags(Flags);
3042 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3043 /// the result needs to be inserted and/or extracted from vectors.
3044 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3045 const TargetTransformInfo &TTI) {
3049 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3052 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3054 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3056 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3062 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3063 // Return the cost of the instruction, including scalarization overhead if it's
3064 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3065 // i.e. either vector version isn't available, or is too expensive.
3066 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3067 const TargetTransformInfo &TTI,
3068 const TargetLibraryInfo *TLI,
3069 bool &NeedToScalarize) {
3070 Function *F = CI->getCalledFunction();
3071 StringRef FnName = CI->getCalledFunction()->getName();
3072 Type *ScalarRetTy = CI->getType();
3073 SmallVector<Type *, 4> Tys, ScalarTys;
3074 for (auto &ArgOp : CI->arg_operands())
3075 ScalarTys.push_back(ArgOp->getType());
3077 // Estimate cost of scalarized vector call. The source operands are assumed
3078 // to be vectors, so we need to extract individual elements from there,
3079 // execute VF scalar calls, and then gather the result into the vector return
3081 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3083 return ScalarCallCost;
3085 // Compute corresponding vector type for return value and arguments.
3086 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3087 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3088 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3090 // Compute costs of unpacking argument values for the scalar calls and
3091 // packing the return values to a vector.
3092 unsigned ScalarizationCost =
3093 getScalarizationOverhead(RetTy, true, false, TTI);
3094 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3095 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3097 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3099 // If we can't emit a vector call for this function, then the currently found
3100 // cost is the cost we need to return.
3101 NeedToScalarize = true;
3102 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3105 // If the corresponding vector cost is cheaper, return its cost.
3106 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3107 if (VectorCallCost < Cost) {
3108 NeedToScalarize = false;
3109 return VectorCallCost;
3114 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3115 // factor VF. Return the cost of the instruction, including scalarization
3116 // overhead if it's needed.
3117 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3118 const TargetTransformInfo &TTI,
3119 const TargetLibraryInfo *TLI) {
3120 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3121 assert(ID && "Expected intrinsic call!");
3123 Type *RetTy = ToVectorTy(CI->getType(), VF);
3124 SmallVector<Type *, 4> Tys;
3125 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3126 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3128 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3131 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3132 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3133 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3134 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3136 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3137 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3138 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3139 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3142 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3143 // For every instruction `I` in MinBWs, truncate the operands, create a
3144 // truncated version of `I` and reextend its result. InstCombine runs
3145 // later and will remove any ext/trunc pairs.
3147 for (auto &KV : MinBWs) {
3148 VectorParts &Parts = WidenMap.get(KV.first);
3149 for (Value *&I : Parts) {
3152 Type *OriginalTy = I->getType();
3153 Type *ScalarTruncatedTy = IntegerType::get(OriginalTy->getContext(),
3155 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3156 OriginalTy->getVectorNumElements());
3157 if (TruncatedTy == OriginalTy)
3160 IRBuilder<> B(cast<Instruction>(I));
3161 auto ShrinkOperand = [&](Value *V) -> Value* {
3162 if (auto *ZI = dyn_cast<ZExtInst>(V))
3163 if (ZI->getSrcTy() == TruncatedTy)
3164 return ZI->getOperand(0);
3165 return B.CreateZExtOrTrunc(V, TruncatedTy);
3168 // The actual instruction modification depends on the instruction type,
3170 Value *NewI = nullptr;
3171 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3172 NewI = B.CreateBinOp(BO->getOpcode(),
3173 ShrinkOperand(BO->getOperand(0)),
3174 ShrinkOperand(BO->getOperand(1)));
3175 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3176 } else if (ICmpInst *CI = dyn_cast<ICmpInst>(I)) {
3177 NewI = B.CreateICmp(CI->getPredicate(),
3178 ShrinkOperand(CI->getOperand(0)),
3179 ShrinkOperand(CI->getOperand(1)));
3180 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
3181 NewI = B.CreateSelect(SI->getCondition(),
3182 ShrinkOperand(SI->getTrueValue()),
3183 ShrinkOperand(SI->getFalseValue()));
3184 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3185 switch (CI->getOpcode()) {
3186 default: llvm_unreachable("Unhandled cast!");
3187 case Instruction::Trunc:
3188 NewI = ShrinkOperand(CI->getOperand(0));
3190 case Instruction::SExt:
3191 NewI = B.CreateSExtOrTrunc(CI->getOperand(0),
3192 smallestIntegerVectorType(OriginalTy,
3195 case Instruction::ZExt:
3196 NewI = B.CreateZExtOrTrunc(CI->getOperand(0),
3197 smallestIntegerVectorType(OriginalTy,
3201 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
3202 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3204 B.CreateZExtOrTrunc(SI->getOperand(0),
3205 VectorType::get(ScalarTruncatedTy, Elements0));
3206 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3208 B.CreateZExtOrTrunc(SI->getOperand(1),
3209 VectorType::get(ScalarTruncatedTy, Elements1));
3211 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3212 } else if (isa<LoadInst>(I)) {
3213 // Don't do anything with the operands, just extend the result.
3216 llvm_unreachable("Unhandled instruction type!");
3219 // Lastly, extend the result.
3220 NewI->takeName(cast<Instruction>(I));
3221 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3222 I->replaceAllUsesWith(Res);
3223 cast<Instruction>(I)->eraseFromParent();
3228 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3229 for (auto &KV : MinBWs) {
3230 VectorParts &Parts = WidenMap.get(KV.first);
3231 for (Value *&I : Parts) {
3232 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3233 if (Inst && Inst->use_empty()) {
3234 Value *NewI = Inst->getOperand(0);
3235 Inst->eraseFromParent();
3242 void InnerLoopVectorizer::vectorizeLoop() {
3243 //===------------------------------------------------===//
3245 // Notice: any optimization or new instruction that go
3246 // into the code below should be also be implemented in
3249 //===------------------------------------------------===//
3250 Constant *Zero = Builder.getInt32(0);
3252 // In order to support reduction variables we need to be able to vectorize
3253 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3254 // stages. First, we create a new vector PHI node with no incoming edges.
3255 // We use this value when we vectorize all of the instructions that use the
3256 // PHI. Next, after all of the instructions in the block are complete we
3257 // add the new incoming edges to the PHI. At this point all of the
3258 // instructions in the basic block are vectorized, so we can use them to
3259 // construct the PHI.
3260 PhiVector RdxPHIsToFix;
3262 // Scan the loop in a topological order to ensure that defs are vectorized
3264 LoopBlocksDFS DFS(OrigLoop);
3267 // Vectorize all of the blocks in the original loop.
3268 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3269 be = DFS.endRPO(); bb != be; ++bb)
3270 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3272 // Insert truncates and extends for any truncated instructions as hints to
3275 truncateToMinimalBitwidths();
3277 // At this point every instruction in the original loop is widened to
3278 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3279 // that we vectorized. The PHI nodes are currently empty because we did
3280 // not want to introduce cycles. Notice that the remaining PHI nodes
3281 // that we need to fix are reduction variables.
3283 // Create the 'reduced' values for each of the induction vars.
3284 // The reduced values are the vector values that we scalarize and combine
3285 // after the loop is finished.
3286 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3288 PHINode *RdxPhi = *it;
3289 assert(RdxPhi && "Unable to recover vectorized PHI");
3291 // Find the reduction variable descriptor.
3292 assert(Legal->getReductionVars()->count(RdxPhi) &&
3293 "Unable to find the reduction variable");
3294 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3296 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3297 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3298 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3299 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3300 RdxDesc.getMinMaxRecurrenceKind();
3301 setDebugLocFromInst(Builder, ReductionStartValue);
3303 // We need to generate a reduction vector from the incoming scalar.
3304 // To do so, we need to generate the 'identity' vector and override
3305 // one of the elements with the incoming scalar reduction. We need
3306 // to do it in the vector-loop preheader.
3307 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3309 // This is the vector-clone of the value that leaves the loop.
3310 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3311 Type *VecTy = VectorExit[0]->getType();
3313 // Find the reduction identity variable. Zero for addition, or, xor,
3314 // one for multiplication, -1 for And.
3317 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3318 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3319 // MinMax reduction have the start value as their identify.
3321 VectorStart = Identity = ReductionStartValue;
3323 VectorStart = Identity =
3324 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3327 // Handle other reduction kinds:
3328 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3329 RK, VecTy->getScalarType());
3332 // This vector is the Identity vector where the first element is the
3333 // incoming scalar reduction.
3334 VectorStart = ReductionStartValue;
3336 Identity = ConstantVector::getSplat(VF, Iden);
3338 // This vector is the Identity vector where the first element is the
3339 // incoming scalar reduction.
3341 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3345 // Fix the vector-loop phi.
3347 // Reductions do not have to start at zero. They can start with
3348 // any loop invariant values.
3349 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3350 BasicBlock *Latch = OrigLoop->getLoopLatch();
3351 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3352 VectorParts &Val = getVectorValue(LoopVal);
3353 for (unsigned part = 0; part < UF; ++part) {
3354 // Make sure to add the reduction stat value only to the
3355 // first unroll part.
3356 Value *StartVal = (part == 0) ? VectorStart : Identity;
3357 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3358 LoopVectorPreHeader);
3359 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3360 LoopVectorBody.back());
3363 // Before each round, move the insertion point right between
3364 // the PHIs and the values we are going to write.
3365 // This allows us to write both PHINodes and the extractelement
3367 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3369 VectorParts RdxParts = getVectorValue(LoopExitInst);
3370 setDebugLocFromInst(Builder, LoopExitInst);
3372 // If the vector reduction can be performed in a smaller type, we truncate
3373 // then extend the loop exit value to enable InstCombine to evaluate the
3374 // entire expression in the smaller type.
3375 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3376 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3377 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3378 for (unsigned part = 0; part < UF; ++part) {
3379 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3380 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3381 : Builder.CreateZExt(Trunc, VecTy);
3382 for (Value::user_iterator UI = RdxParts[part]->user_begin();
3383 UI != RdxParts[part]->user_end();)
3385 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
3386 RdxParts[part] = Extnd;
3391 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3392 for (unsigned part = 0; part < UF; ++part)
3393 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3396 // Reduce all of the unrolled parts into a single vector.
3397 Value *ReducedPartRdx = RdxParts[0];
3398 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3399 setDebugLocFromInst(Builder, ReducedPartRdx);
3400 for (unsigned part = 1; part < UF; ++part) {
3401 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3402 // Floating point operations had to be 'fast' to enable the reduction.
3403 ReducedPartRdx = addFastMathFlag(
3404 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3405 ReducedPartRdx, "bin.rdx"));
3407 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3408 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3412 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3413 // and vector ops, reducing the set of values being computed by half each
3415 assert(isPowerOf2_32(VF) &&
3416 "Reduction emission only supported for pow2 vectors!");
3417 Value *TmpVec = ReducedPartRdx;
3418 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3419 for (unsigned i = VF; i != 1; i >>= 1) {
3420 // Move the upper half of the vector to the lower half.
3421 for (unsigned j = 0; j != i/2; ++j)
3422 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3424 // Fill the rest of the mask with undef.
3425 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3426 UndefValue::get(Builder.getInt32Ty()));
3429 Builder.CreateShuffleVector(TmpVec,
3430 UndefValue::get(TmpVec->getType()),
3431 ConstantVector::get(ShuffleMask),
3434 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3435 // Floating point operations had to be 'fast' to enable the reduction.
3436 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3437 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3439 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3443 // The result is in the first element of the vector.
3444 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3445 Builder.getInt32(0));
3447 // If the reduction can be performed in a smaller type, we need to extend
3448 // the reduction to the wider type before we branch to the original loop.
3449 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3452 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3453 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3456 // Create a phi node that merges control-flow from the backedge-taken check
3457 // block and the middle block.
3458 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3459 LoopScalarPreHeader->getTerminator());
3460 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
3461 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
3462 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3464 // Now, we need to fix the users of the reduction variable
3465 // inside and outside of the scalar remainder loop.
3466 // We know that the loop is in LCSSA form. We need to update the
3467 // PHI nodes in the exit blocks.
3468 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3469 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3470 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3471 if (!LCSSAPhi) break;
3473 // All PHINodes need to have a single entry edge, or two if
3474 // we already fixed them.
3475 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3477 // We found our reduction value exit-PHI. Update it with the
3478 // incoming bypass edge.
3479 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3480 // Add an edge coming from the bypass.
3481 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3484 }// end of the LCSSA phi scan.
3486 // Fix the scalar loop reduction variable with the incoming reduction sum
3487 // from the vector body and from the backedge value.
3488 int IncomingEdgeBlockIdx =
3489 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3490 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3491 // Pick the other block.
3492 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3493 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3494 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3495 }// end of for each redux variable.
3499 // Make sure DomTree is updated.
3502 // Predicate any stores.
3503 for (auto KV : PredicatedStores) {
3504 BasicBlock::iterator I(KV.first);
3505 auto *BB = SplitBlock(I->getParent(), &*std::next(I), DT, LI);
3506 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
3507 /*BranchWeights=*/nullptr, DT);
3509 I->getParent()->setName("pred.store.if");
3510 BB->setName("pred.store.continue");
3512 DEBUG(DT->verifyDomTree());
3513 // Remove redundant induction instructions.
3514 cse(LoopVectorBody);
3517 void InnerLoopVectorizer::fixLCSSAPHIs() {
3518 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3519 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3520 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3521 if (!LCSSAPhi) break;
3522 if (LCSSAPhi->getNumIncomingValues() == 1)
3523 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3528 InnerLoopVectorizer::VectorParts
3529 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3530 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3533 // Look for cached value.
3534 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3535 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3536 if (ECEntryIt != MaskCache.end())
3537 return ECEntryIt->second;
3539 VectorParts SrcMask = createBlockInMask(Src);
3541 // The terminator has to be a branch inst!
3542 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3543 assert(BI && "Unexpected terminator found");
3545 if (BI->isConditional()) {
3546 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3548 if (BI->getSuccessor(0) != Dst)
3549 for (unsigned part = 0; part < UF; ++part)
3550 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3552 for (unsigned part = 0; part < UF; ++part)
3553 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3555 MaskCache[Edge] = EdgeMask;
3559 MaskCache[Edge] = SrcMask;
3563 InnerLoopVectorizer::VectorParts
3564 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3565 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3567 // Loop incoming mask is all-one.
3568 if (OrigLoop->getHeader() == BB) {
3569 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3570 return getVectorValue(C);
3573 // This is the block mask. We OR all incoming edges, and with zero.
3574 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3575 VectorParts BlockMask = getVectorValue(Zero);
3578 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3579 VectorParts EM = createEdgeMask(*it, BB);
3580 for (unsigned part = 0; part < UF; ++part)
3581 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3587 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3588 InnerLoopVectorizer::VectorParts &Entry,
3589 unsigned UF, unsigned VF, PhiVector *PV) {
3590 PHINode* P = cast<PHINode>(PN);
3591 // Handle reduction variables:
3592 if (Legal->getReductionVars()->count(P)) {
3593 for (unsigned part = 0; part < UF; ++part) {
3594 // This is phase one of vectorizing PHIs.
3595 Type *VecTy = (VF == 1) ? PN->getType() :
3596 VectorType::get(PN->getType(), VF);
3597 Entry[part] = PHINode::Create(
3598 VecTy, 2, "vec.phi", &*LoopVectorBody.back()->getFirstInsertionPt());
3604 setDebugLocFromInst(Builder, P);
3605 // Check for PHI nodes that are lowered to vector selects.
3606 if (P->getParent() != OrigLoop->getHeader()) {
3607 // We know that all PHIs in non-header blocks are converted into
3608 // selects, so we don't have to worry about the insertion order and we
3609 // can just use the builder.
3610 // At this point we generate the predication tree. There may be
3611 // duplications since this is a simple recursive scan, but future
3612 // optimizations will clean it up.
3614 unsigned NumIncoming = P->getNumIncomingValues();
3616 // Generate a sequence of selects of the form:
3617 // SELECT(Mask3, In3,
3618 // SELECT(Mask2, In2,
3620 for (unsigned In = 0; In < NumIncoming; In++) {
3621 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3623 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3625 for (unsigned part = 0; part < UF; ++part) {
3626 // We might have single edge PHIs (blocks) - use an identity
3627 // 'select' for the first PHI operand.
3629 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3632 // Select between the current value and the previous incoming edge
3633 // based on the incoming mask.
3634 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3635 Entry[part], "predphi");
3641 // This PHINode must be an induction variable.
3642 // Make sure that we know about it.
3643 assert(Legal->getInductionVars()->count(P) &&
3644 "Not an induction variable");
3646 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3648 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3649 // which can be found from the original scalar operations.
3650 switch (II.getKind()) {
3651 case InductionDescriptor::IK_NoInduction:
3652 llvm_unreachable("Unknown induction");
3653 case InductionDescriptor::IK_IntInduction: {
3654 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3655 // Handle other induction variables that are now based on the
3657 Value *V = Induction;
3658 if (P != OldInduction) {
3659 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3660 V = II.transform(Builder, V);
3661 V->setName("offset.idx");
3663 Value *Broadcasted = getBroadcastInstrs(V);
3664 // After broadcasting the induction variable we need to make the vector
3665 // consecutive by adding 0, 1, 2, etc.
3666 for (unsigned part = 0; part < UF; ++part)
3667 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3670 case InductionDescriptor::IK_PtrInduction:
3671 // Handle the pointer induction variable case.
3672 assert(P->getType()->isPointerTy() && "Unexpected type.");
3673 // This is the normalized GEP that starts counting at zero.
3674 Value *PtrInd = Induction;
3675 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3676 // This is the vector of results. Notice that we don't generate
3677 // vector geps because scalar geps result in better code.
3678 for (unsigned part = 0; part < UF; ++part) {
3680 int EltIndex = part;
3681 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3682 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3683 Value *SclrGep = II.transform(Builder, GlobalIdx);
3684 SclrGep->setName("next.gep");
3685 Entry[part] = SclrGep;
3689 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3690 for (unsigned int i = 0; i < VF; ++i) {
3691 int EltIndex = i + part * VF;
3692 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3693 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3694 Value *SclrGep = II.transform(Builder, GlobalIdx);
3695 SclrGep->setName("next.gep");
3696 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3697 Builder.getInt32(i),
3700 Entry[part] = VecVal;
3706 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3707 // For each instruction in the old loop.
3708 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3709 VectorParts &Entry = WidenMap.get(&*it);
3711 switch (it->getOpcode()) {
3712 case Instruction::Br:
3713 // Nothing to do for PHIs and BR, since we already took care of the
3714 // loop control flow instructions.
3716 case Instruction::PHI: {
3717 // Vectorize PHINodes.
3718 widenPHIInstruction(&*it, Entry, UF, VF, PV);
3722 case Instruction::Add:
3723 case Instruction::FAdd:
3724 case Instruction::Sub:
3725 case Instruction::FSub:
3726 case Instruction::Mul:
3727 case Instruction::FMul:
3728 case Instruction::UDiv:
3729 case Instruction::SDiv:
3730 case Instruction::FDiv:
3731 case Instruction::URem:
3732 case Instruction::SRem:
3733 case Instruction::FRem:
3734 case Instruction::Shl:
3735 case Instruction::LShr:
3736 case Instruction::AShr:
3737 case Instruction::And:
3738 case Instruction::Or:
3739 case Instruction::Xor: {
3740 // Just widen binops.
3741 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3742 setDebugLocFromInst(Builder, BinOp);
3743 VectorParts &A = getVectorValue(it->getOperand(0));
3744 VectorParts &B = getVectorValue(it->getOperand(1));
3746 // Use this vector value for all users of the original instruction.
3747 for (unsigned Part = 0; Part < UF; ++Part) {
3748 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3750 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3751 VecOp->copyIRFlags(BinOp);
3756 propagateMetadata(Entry, &*it);
3759 case Instruction::Select: {
3761 // If the selector is loop invariant we can create a select
3762 // instruction with a scalar condition. Otherwise, use vector-select.
3763 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3765 setDebugLocFromInst(Builder, &*it);
3767 // The condition can be loop invariant but still defined inside the
3768 // loop. This means that we can't just use the original 'cond' value.
3769 // We have to take the 'vectorized' value and pick the first lane.
3770 // Instcombine will make this a no-op.
3771 VectorParts &Cond = getVectorValue(it->getOperand(0));
3772 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3773 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3775 Value *ScalarCond = (VF == 1) ? Cond[0] :
3776 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3778 for (unsigned Part = 0; Part < UF; ++Part) {
3779 Entry[Part] = Builder.CreateSelect(
3780 InvariantCond ? ScalarCond : Cond[Part],
3785 propagateMetadata(Entry, &*it);
3789 case Instruction::ICmp:
3790 case Instruction::FCmp: {
3791 // Widen compares. Generate vector compares.
3792 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3793 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3794 setDebugLocFromInst(Builder, &*it);
3795 VectorParts &A = getVectorValue(it->getOperand(0));
3796 VectorParts &B = getVectorValue(it->getOperand(1));
3797 for (unsigned Part = 0; Part < UF; ++Part) {
3800 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3801 cast<FCmpInst>(C)->copyFastMathFlags(&*it);
3803 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3808 propagateMetadata(Entry, &*it);
3812 case Instruction::Store:
3813 case Instruction::Load:
3814 vectorizeMemoryInstruction(&*it);
3816 case Instruction::ZExt:
3817 case Instruction::SExt:
3818 case Instruction::FPToUI:
3819 case Instruction::FPToSI:
3820 case Instruction::FPExt:
3821 case Instruction::PtrToInt:
3822 case Instruction::IntToPtr:
3823 case Instruction::SIToFP:
3824 case Instruction::UIToFP:
3825 case Instruction::Trunc:
3826 case Instruction::FPTrunc:
3827 case Instruction::BitCast: {
3828 CastInst *CI = dyn_cast<CastInst>(it);
3829 setDebugLocFromInst(Builder, &*it);
3830 /// Optimize the special case where the source is the induction
3831 /// variable. Notice that we can only optimize the 'trunc' case
3832 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3833 /// c. other casts depend on pointer size.
3834 if (CI->getOperand(0) == OldInduction &&
3835 it->getOpcode() == Instruction::Trunc) {
3836 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3838 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3839 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3841 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3842 for (unsigned Part = 0; Part < UF; ++Part)
3843 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3844 propagateMetadata(Entry, &*it);
3847 /// Vectorize casts.
3848 Type *DestTy = (VF == 1) ? CI->getType() :
3849 VectorType::get(CI->getType(), VF);
3851 VectorParts &A = getVectorValue(it->getOperand(0));
3852 for (unsigned Part = 0; Part < UF; ++Part)
3853 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3854 propagateMetadata(Entry, &*it);
3858 case Instruction::Call: {
3859 // Ignore dbg intrinsics.
3860 if (isa<DbgInfoIntrinsic>(it))
3862 setDebugLocFromInst(Builder, &*it);
3864 Module *M = BB->getParent()->getParent();
3865 CallInst *CI = cast<CallInst>(it);
3867 StringRef FnName = CI->getCalledFunction()->getName();
3868 Function *F = CI->getCalledFunction();
3869 Type *RetTy = ToVectorTy(CI->getType(), VF);
3870 SmallVector<Type *, 4> Tys;
3871 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3872 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3874 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3876 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3877 ID == Intrinsic::lifetime_start)) {
3878 scalarizeInstruction(&*it);
3881 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3882 // version of the instruction.
3883 // Is it beneficial to perform intrinsic call compared to lib call?
3884 bool NeedToScalarize;
3885 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3886 bool UseVectorIntrinsic =
3887 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3888 if (!UseVectorIntrinsic && NeedToScalarize) {
3889 scalarizeInstruction(&*it);
3893 for (unsigned Part = 0; Part < UF; ++Part) {
3894 SmallVector<Value *, 4> Args;
3895 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3896 Value *Arg = CI->getArgOperand(i);
3897 // Some intrinsics have a scalar argument - don't replace it with a
3899 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3900 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3901 Arg = VectorArg[Part];
3903 Args.push_back(Arg);
3907 if (UseVectorIntrinsic) {
3908 // Use vector version of the intrinsic.
3909 Type *TysForDecl[] = {CI->getType()};
3911 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3912 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3914 // Use vector version of the library call.
3915 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3916 assert(!VFnName.empty() && "Vector function name is empty.");
3917 VectorF = M->getFunction(VFnName);
3919 // Generate a declaration
3920 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3922 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3923 VectorF->copyAttributesFrom(F);
3926 assert(VectorF && "Can't create vector function.");
3927 Entry[Part] = Builder.CreateCall(VectorF, Args);
3930 propagateMetadata(Entry, &*it);
3935 // All other instructions are unsupported. Scalarize them.
3936 scalarizeInstruction(&*it);
3939 }// end of for_each instr.
3942 void InnerLoopVectorizer::updateAnalysis() {
3943 // Forget the original basic block.
3944 SE->forgetLoop(OrigLoop);
3946 // Update the dominator tree information.
3947 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3948 "Entry does not dominate exit.");
3950 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3951 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3952 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3954 // We don't predicate stores by this point, so the vector body should be a
3956 assert(LoopVectorBody.size() == 1 && "Expected single block loop!");
3957 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3959 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody.back());
3960 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3961 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3962 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3964 DEBUG(DT->verifyDomTree());
3967 /// \brief Check whether it is safe to if-convert this phi node.
3969 /// Phi nodes with constant expressions that can trap are not safe to if
3971 static bool canIfConvertPHINodes(BasicBlock *BB) {
3972 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3973 PHINode *Phi = dyn_cast<PHINode>(I);
3976 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3977 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3984 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3985 if (!EnableIfConversion) {
3986 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3990 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3992 // A list of pointers that we can safely read and write to.
3993 SmallPtrSet<Value *, 8> SafePointes;
3995 // Collect safe addresses.
3996 for (Loop::block_iterator BI = TheLoop->block_begin(),
3997 BE = TheLoop->block_end(); BI != BE; ++BI) {
3998 BasicBlock *BB = *BI;
4000 if (blockNeedsPredication(BB))
4003 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
4004 if (LoadInst *LI = dyn_cast<LoadInst>(I))
4005 SafePointes.insert(LI->getPointerOperand());
4006 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
4007 SafePointes.insert(SI->getPointerOperand());
4011 // Collect the blocks that need predication.
4012 BasicBlock *Header = TheLoop->getHeader();
4013 for (Loop::block_iterator BI = TheLoop->block_begin(),
4014 BE = TheLoop->block_end(); BI != BE; ++BI) {
4015 BasicBlock *BB = *BI;
4017 // We don't support switch statements inside loops.
4018 if (!isa<BranchInst>(BB->getTerminator())) {
4019 emitAnalysis(VectorizationReport(BB->getTerminator())
4020 << "loop contains a switch statement");
4024 // We must be able to predicate all blocks that need to be predicated.
4025 if (blockNeedsPredication(BB)) {
4026 if (!blockCanBePredicated(BB, SafePointes)) {
4027 emitAnalysis(VectorizationReport(BB->getTerminator())
4028 << "control flow cannot be substituted for a select");
4031 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
4032 emitAnalysis(VectorizationReport(BB->getTerminator())
4033 << "control flow cannot be substituted for a select");
4038 // We can if-convert this loop.
4042 bool LoopVectorizationLegality::canVectorize() {
4043 // We must have a loop in canonical form. Loops with indirectbr in them cannot
4044 // be canonicalized.
4045 if (!TheLoop->getLoopPreheader()) {
4047 VectorizationReport() <<
4048 "loop control flow is not understood by vectorizer");
4052 // We can only vectorize innermost loops.
4053 if (!TheLoop->empty()) {
4054 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
4058 // We must have a single backedge.
4059 if (TheLoop->getNumBackEdges() != 1) {
4061 VectorizationReport() <<
4062 "loop control flow is not understood by vectorizer");
4066 // We must have a single exiting block.
4067 if (!TheLoop->getExitingBlock()) {
4069 VectorizationReport() <<
4070 "loop control flow is not understood by vectorizer");
4074 // We only handle bottom-tested loops, i.e. loop in which the condition is
4075 // checked at the end of each iteration. With that we can assume that all
4076 // instructions in the loop are executed the same number of times.
4077 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
4079 VectorizationReport() <<
4080 "loop control flow is not understood by vectorizer");
4084 // We need to have a loop header.
4085 DEBUG(dbgs() << "LV: Found a loop: " <<
4086 TheLoop->getHeader()->getName() << '\n');
4088 // Check if we can if-convert non-single-bb loops.
4089 unsigned NumBlocks = TheLoop->getNumBlocks();
4090 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4091 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4095 // ScalarEvolution needs to be able to find the exit count.
4096 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
4097 if (ExitCount == SE->getCouldNotCompute()) {
4098 emitAnalysis(VectorizationReport() <<
4099 "could not determine number of loop iterations");
4100 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4104 // Check if we can vectorize the instructions and CFG in this loop.
4105 if (!canVectorizeInstrs()) {
4106 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4110 // Go over each instruction and look at memory deps.
4111 if (!canVectorizeMemory()) {
4112 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4116 // Collect all of the variables that remain uniform after vectorization.
4117 collectLoopUniforms();
4119 DEBUG(dbgs() << "LV: We can vectorize this loop"
4120 << (LAI->getRuntimePointerChecking()->Need
4121 ? " (with a runtime bound check)"
4125 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4127 // If an override option has been passed in for interleaved accesses, use it.
4128 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4129 UseInterleaved = EnableInterleavedMemAccesses;
4131 // Analyze interleaved memory accesses.
4133 InterleaveInfo.analyzeInterleaving(Strides);
4135 // Okay! We can vectorize. At this point we don't have any other mem analysis
4136 // which may limit our maximum vectorization factor, so just return true with
4141 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4142 if (Ty->isPointerTy())
4143 return DL.getIntPtrType(Ty);
4145 // It is possible that char's or short's overflow when we ask for the loop's
4146 // trip count, work around this by changing the type size.
4147 if (Ty->getScalarSizeInBits() < 32)
4148 return Type::getInt32Ty(Ty->getContext());
4153 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4154 Ty0 = convertPointerToIntegerType(DL, Ty0);
4155 Ty1 = convertPointerToIntegerType(DL, Ty1);
4156 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4161 /// \brief Check that the instruction has outside loop users and is not an
4162 /// identified reduction variable.
4163 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4164 SmallPtrSetImpl<Value *> &Reductions) {
4165 // Reduction instructions are allowed to have exit users. All other
4166 // instructions must not have external users.
4167 if (!Reductions.count(Inst))
4168 //Check that all of the users of the loop are inside the BB.
4169 for (User *U : Inst->users()) {
4170 Instruction *UI = cast<Instruction>(U);
4171 // This user may be a reduction exit value.
4172 if (!TheLoop->contains(UI)) {
4173 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4180 bool LoopVectorizationLegality::canVectorizeInstrs() {
4181 BasicBlock *Header = TheLoop->getHeader();
4183 // Look for the attribute signaling the absence of NaNs.
4184 Function &F = *Header->getParent();
4185 const DataLayout &DL = F.getParent()->getDataLayout();
4186 if (F.hasFnAttribute("no-nans-fp-math"))
4188 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4190 // For each block in the loop.
4191 for (Loop::block_iterator bb = TheLoop->block_begin(),
4192 be = TheLoop->block_end(); bb != be; ++bb) {
4194 // Scan the instructions in the block and look for hazards.
4195 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4198 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4199 Type *PhiTy = Phi->getType();
4200 // Check that this PHI type is allowed.
4201 if (!PhiTy->isIntegerTy() &&
4202 !PhiTy->isFloatingPointTy() &&
4203 !PhiTy->isPointerTy()) {
4204 emitAnalysis(VectorizationReport(&*it)
4205 << "loop control flow is not understood by vectorizer");
4206 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4210 // If this PHINode is not in the header block, then we know that we
4211 // can convert it to select during if-conversion. No need to check if
4212 // the PHIs in this block are induction or reduction variables.
4213 if (*bb != Header) {
4214 // Check that this instruction has no outside users or is an
4215 // identified reduction value with an outside user.
4216 if (!hasOutsideLoopUser(TheLoop, &*it, AllowedExit))
4218 emitAnalysis(VectorizationReport(&*it) <<
4219 "value could not be identified as "
4220 "an induction or reduction variable");
4224 // We only allow if-converted PHIs with exactly two incoming values.
4225 if (Phi->getNumIncomingValues() != 2) {
4226 emitAnalysis(VectorizationReport(&*it)
4227 << "control flow not understood by vectorizer");
4228 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4232 InductionDescriptor ID;
4233 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4234 Inductions[Phi] = ID;
4235 // Get the widest type.
4237 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4239 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4241 // Int inductions are special because we only allow one IV.
4242 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4243 ID.getStepValue()->isOne() &&
4244 isa<Constant>(ID.getStartValue()) &&
4245 cast<Constant>(ID.getStartValue())->isNullValue()) {
4246 // Use the phi node with the widest type as induction. Use the last
4247 // one if there are multiple (no good reason for doing this other
4248 // than it is expedient). We've checked that it begins at zero and
4249 // steps by one, so this is a canonical induction variable.
4250 if (!Induction || PhiTy == WidestIndTy)
4254 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4256 // Until we explicitly handle the case of an induction variable with
4257 // an outside loop user we have to give up vectorizing this loop.
4258 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4259 emitAnalysis(VectorizationReport(&*it) <<
4260 "use of induction value outside of the "
4261 "loop is not handled by vectorizer");
4268 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4270 if (Reductions[Phi].hasUnsafeAlgebra())
4271 Requirements->addUnsafeAlgebraInst(
4272 Reductions[Phi].getUnsafeAlgebraInst());
4273 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4277 emitAnalysis(VectorizationReport(&*it) <<
4278 "value that could not be identified as "
4279 "reduction is used outside the loop");
4280 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4282 }// end of PHI handling
4284 // We handle calls that:
4285 // * Are debug info intrinsics.
4286 // * Have a mapping to an IR intrinsic.
4287 // * Have a vector version available.
4288 CallInst *CI = dyn_cast<CallInst>(it);
4289 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4290 !(CI->getCalledFunction() && TLI &&
4291 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4292 emitAnalysis(VectorizationReport(&*it)
4293 << "call instruction cannot be vectorized");
4294 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4298 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4299 // second argument is the same (i.e. loop invariant)
4301 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4302 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4303 emitAnalysis(VectorizationReport(&*it)
4304 << "intrinsic instruction cannot be vectorized");
4305 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4310 // Check that the instruction return type is vectorizable.
4311 // Also, we can't vectorize extractelement instructions.
4312 if ((!VectorType::isValidElementType(it->getType()) &&
4313 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4314 emitAnalysis(VectorizationReport(&*it)
4315 << "instruction return type cannot be vectorized");
4316 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4320 // Check that the stored type is vectorizable.
4321 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4322 Type *T = ST->getValueOperand()->getType();
4323 if (!VectorType::isValidElementType(T)) {
4324 emitAnalysis(VectorizationReport(ST) <<
4325 "store instruction cannot be vectorized");
4328 if (EnableMemAccessVersioning)
4329 collectStridedAccess(ST);
4332 if (EnableMemAccessVersioning)
4333 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4334 collectStridedAccess(LI);
4336 // Reduction instructions are allowed to have exit users.
4337 // All other instructions must not have external users.
4338 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4339 emitAnalysis(VectorizationReport(&*it) <<
4340 "value cannot be used outside the loop");
4349 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4350 if (Inductions.empty()) {
4351 emitAnalysis(VectorizationReport()
4352 << "loop induction variable could not be identified");
4357 // Now we know the widest induction type, check if our found induction
4358 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4359 // will create another.
4360 if (Induction && WidestIndTy != Induction->getType())
4361 Induction = nullptr;
4366 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4367 Value *Ptr = nullptr;
4368 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4369 Ptr = LI->getPointerOperand();
4370 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4371 Ptr = SI->getPointerOperand();
4375 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4379 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4380 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4381 Strides[Ptr] = Stride;
4382 StrideSet.insert(Stride);
4385 void LoopVectorizationLegality::collectLoopUniforms() {
4386 // We now know that the loop is vectorizable!
4387 // Collect variables that will remain uniform after vectorization.
4388 std::vector<Value*> Worklist;
4389 BasicBlock *Latch = TheLoop->getLoopLatch();
4391 // Start with the conditional branch and walk up the block.
4392 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4394 // Also add all consecutive pointer values; these values will be uniform
4395 // after vectorization (and subsequent cleanup) and, until revectorization is
4396 // supported, all dependencies must also be uniform.
4397 for (Loop::block_iterator B = TheLoop->block_begin(),
4398 BE = TheLoop->block_end(); B != BE; ++B)
4399 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4401 if (I->getType()->isPointerTy() && isConsecutivePtr(&*I))
4402 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4404 while (!Worklist.empty()) {
4405 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4406 Worklist.pop_back();
4408 // Look at instructions inside this loop.
4409 // Stop when reaching PHI nodes.
4410 // TODO: we need to follow values all over the loop, not only in this block.
4411 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4414 // This is a known uniform.
4417 // Insert all operands.
4418 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4422 bool LoopVectorizationLegality::canVectorizeMemory() {
4423 LAI = &LAA->getInfo(TheLoop, Strides);
4424 auto &OptionalReport = LAI->getReport();
4426 emitAnalysis(VectorizationReport(*OptionalReport));
4427 if (!LAI->canVectorizeMemory())
4430 if (LAI->hasStoreToLoopInvariantAddress()) {
4432 VectorizationReport()
4433 << "write to a loop invariant address could not be vectorized");
4434 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4438 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4443 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4444 Value *In0 = const_cast<Value*>(V);
4445 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4449 return Inductions.count(PN);
4452 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4453 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4456 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4457 SmallPtrSetImpl<Value *> &SafePtrs) {
4459 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4460 // Check that we don't have a constant expression that can trap as operand.
4461 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4463 if (Constant *C = dyn_cast<Constant>(*OI))
4467 // We might be able to hoist the load.
4468 if (it->mayReadFromMemory()) {
4469 LoadInst *LI = dyn_cast<LoadInst>(it);
4472 if (!SafePtrs.count(LI->getPointerOperand())) {
4473 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4474 MaskedOp.insert(LI);
4481 // We don't predicate stores at the moment.
4482 if (it->mayWriteToMemory()) {
4483 StoreInst *SI = dyn_cast<StoreInst>(it);
4484 // We only support predication of stores in basic blocks with one
4489 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4490 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4492 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4493 !isSinglePredecessor) {
4494 // Build a masked store if it is legal for the target, otherwise scalarize
4496 bool isLegalMaskedOp =
4497 isLegalMaskedStore(SI->getValueOperand()->getType(),
4498 SI->getPointerOperand());
4499 if (isLegalMaskedOp) {
4501 MaskedOp.insert(SI);
4510 // The instructions below can trap.
4511 switch (it->getOpcode()) {
4513 case Instruction::UDiv:
4514 case Instruction::SDiv:
4515 case Instruction::URem:
4516 case Instruction::SRem:
4524 void InterleavedAccessInfo::collectConstStridedAccesses(
4525 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4526 const ValueToValueMap &Strides) {
4527 // Holds load/store instructions in program order.
4528 SmallVector<Instruction *, 16> AccessList;
4530 for (auto *BB : TheLoop->getBlocks()) {
4531 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4533 for (auto &I : *BB) {
4534 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4536 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4540 AccessList.push_back(&I);
4544 if (AccessList.empty())
4547 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4548 for (auto I : AccessList) {
4549 LoadInst *LI = dyn_cast<LoadInst>(I);
4550 StoreInst *SI = dyn_cast<StoreInst>(I);
4552 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4553 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4555 // The factor of the corresponding interleave group.
4556 unsigned Factor = std::abs(Stride);
4558 // Ignore the access if the factor is too small or too large.
4559 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4562 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4563 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4564 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4566 // An alignment of 0 means target ABI alignment.
4567 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4569 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4571 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4575 // Analyze interleaved accesses and collect them into interleave groups.
4577 // Notice that the vectorization on interleaved groups will change instruction
4578 // orders and may break dependences. But the memory dependence check guarantees
4579 // that there is no overlap between two pointers of different strides, element
4580 // sizes or underlying bases.
4582 // For pointers sharing the same stride, element size and underlying base, no
4583 // need to worry about Read-After-Write dependences and Write-After-Read
4586 // E.g. The RAW dependence: A[i] = a;
4588 // This won't exist as it is a store-load forwarding conflict, which has
4589 // already been checked and forbidden in the dependence check.
4591 // E.g. The WAR dependence: a = A[i]; // (1)
4593 // The store group of (2) is always inserted at or below (2), and the load group
4594 // of (1) is always inserted at or above (1). The dependence is safe.
4595 void InterleavedAccessInfo::analyzeInterleaving(
4596 const ValueToValueMap &Strides) {
4597 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4599 // Holds all the stride accesses.
4600 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4601 collectConstStridedAccesses(StrideAccesses, Strides);
4603 if (StrideAccesses.empty())
4606 // Holds all interleaved store groups temporarily.
4607 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4609 // Search the load-load/write-write pair B-A in bottom-up order and try to
4610 // insert B into the interleave group of A according to 3 rules:
4611 // 1. A and B have the same stride.
4612 // 2. A and B have the same memory object size.
4613 // 3. B belongs to the group according to the distance.
4615 // The bottom-up order can avoid breaking the Write-After-Write dependences
4616 // between two pointers of the same base.
4617 // E.g. A[i] = a; (1)
4620 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4621 // above (1), which guarantees that (1) is always above (2).
4622 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4624 Instruction *A = I->first;
4625 StrideDescriptor DesA = I->second;
4627 InterleaveGroup *Group = getInterleaveGroup(A);
4629 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4630 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4633 if (A->mayWriteToMemory())
4634 StoreGroups.insert(Group);
4636 for (auto II = std::next(I); II != E; ++II) {
4637 Instruction *B = II->first;
4638 StrideDescriptor DesB = II->second;
4640 // Ignore if B is already in a group or B is a different memory operation.
4641 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4644 // Check the rule 1 and 2.
4645 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4648 // Calculate the distance and prepare for the rule 3.
4649 const SCEVConstant *DistToA =
4650 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4654 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4656 // Skip if the distance is not multiple of size as they are not in the
4658 if (DistanceToA % static_cast<int>(DesA.Size))
4661 // The index of B is the index of A plus the related index to A.
4663 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4665 // Try to insert B into the group.
4666 if (Group->insertMember(B, IndexB, DesB.Align)) {
4667 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4668 << " into the interleave group with" << *A << '\n');
4669 InterleaveGroupMap[B] = Group;
4671 // Set the first load in program order as the insert position.
4672 if (B->mayReadFromMemory())
4673 Group->setInsertPos(B);
4675 } // Iteration on instruction B
4676 } // Iteration on instruction A
4678 // Remove interleaved store groups with gaps.
4679 for (InterleaveGroup *Group : StoreGroups)
4680 if (Group->getNumMembers() != Group->getFactor())
4681 releaseGroup(Group);
4684 LoopVectorizationCostModel::VectorizationFactor
4685 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4686 // Width 1 means no vectorize
4687 VectorizationFactor Factor = { 1U, 0U };
4688 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4689 emitAnalysis(VectorizationReport() <<
4690 "runtime pointer checks needed. Enable vectorization of this "
4691 "loop with '#pragma clang loop vectorize(enable)' when "
4692 "compiling with -Os/-Oz");
4694 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4698 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4699 emitAnalysis(VectorizationReport() <<
4700 "store that is conditionally executed prevents vectorization");
4701 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4705 // Find the trip count.
4706 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4707 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4709 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
4710 unsigned WidestType = getWidestType();
4711 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4712 unsigned MaxSafeDepDist = -1U;
4713 if (Legal->getMaxSafeDepDistBytes() != -1U)
4714 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4715 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4716 WidestRegister : MaxSafeDepDist);
4717 unsigned MaxVectorSize = WidestRegister / WidestType;
4718 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4719 DEBUG(dbgs() << "LV: The Widest register is: "
4720 << WidestRegister << " bits.\n");
4722 if (MaxVectorSize == 0) {
4723 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4727 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4728 " into one vector!");
4730 unsigned VF = MaxVectorSize;
4732 // If we optimize the program for size, avoid creating the tail loop.
4734 // If we are unable to calculate the trip count then don't try to vectorize.
4737 (VectorizationReport() <<
4738 "unable to calculate the loop count due to complex control flow");
4739 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4743 // Find the maximum SIMD width that can fit within the trip count.
4744 VF = TC % MaxVectorSize;
4749 // If the trip count that we found modulo the vectorization factor is not
4750 // zero then we require a tail.
4751 emitAnalysis(VectorizationReport() <<
4752 "cannot optimize for size and vectorize at the "
4753 "same time. Enable vectorization of this loop "
4754 "with '#pragma clang loop vectorize(enable)' "
4755 "when compiling with -Os/-Oz");
4756 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4761 int UserVF = Hints->getWidth();
4763 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4764 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4766 Factor.Width = UserVF;
4770 float Cost = expectedCost(1);
4772 const float ScalarCost = Cost;
4775 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4777 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4778 // Ignore scalar width, because the user explicitly wants vectorization.
4779 if (ForceVectorization && VF > 1) {
4781 Cost = expectedCost(Width) / (float)Width;
4784 for (unsigned i=2; i <= VF; i*=2) {
4785 // Notice that the vector loop needs to be executed less times, so
4786 // we need to divide the cost of the vector loops by the width of
4787 // the vector elements.
4788 float VectorCost = expectedCost(i) / (float)i;
4789 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4790 (int)VectorCost << ".\n");
4791 if (VectorCost < Cost) {
4797 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4798 << "LV: Vectorization seems to be not beneficial, "
4799 << "but was forced by a user.\n");
4800 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4801 Factor.Width = Width;
4802 Factor.Cost = Width * Cost;
4806 unsigned LoopVectorizationCostModel::getWidestType() {
4807 unsigned MaxWidth = 8;
4808 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4811 for (Loop::block_iterator bb = TheLoop->block_begin(),
4812 be = TheLoop->block_end(); bb != be; ++bb) {
4813 BasicBlock *BB = *bb;
4815 // For each instruction in the loop.
4816 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4817 Type *T = it->getType();
4819 // Skip ignored values.
4820 if (ValuesToIgnore.count(&*it))
4823 // Only examine Loads, Stores and PHINodes.
4824 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4827 // Examine PHI nodes that are reduction variables. Update the type to
4828 // account for the recurrence type.
4829 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4830 if (!Legal->getReductionVars()->count(PN))
4832 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4833 T = RdxDesc.getRecurrenceType();
4836 // Examine the stored values.
4837 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4838 T = ST->getValueOperand()->getType();
4840 // Ignore loaded pointer types and stored pointer types that are not
4841 // consecutive. However, we do want to take consecutive stores/loads of
4842 // pointer vectors into account.
4843 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&*it))
4846 MaxWidth = std::max(MaxWidth,
4847 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4854 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4856 unsigned LoopCost) {
4858 // -- The interleave heuristics --
4859 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4860 // There are many micro-architectural considerations that we can't predict
4861 // at this level. For example, frontend pressure (on decode or fetch) due to
4862 // code size, or the number and capabilities of the execution ports.
4864 // We use the following heuristics to select the interleave count:
4865 // 1. If the code has reductions, then we interleave to break the cross
4866 // iteration dependency.
4867 // 2. If the loop is really small, then we interleave to reduce the loop
4869 // 3. We don't interleave if we think that we will spill registers to memory
4870 // due to the increased register pressure.
4872 // When we optimize for size, we don't interleave.
4876 // We used the distance for the interleave count.
4877 if (Legal->getMaxSafeDepDistBytes() != -1U)
4880 // Do not interleave loops with a relatively small trip count.
4881 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4882 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4885 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4886 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4890 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4891 TargetNumRegisters = ForceTargetNumScalarRegs;
4893 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4894 TargetNumRegisters = ForceTargetNumVectorRegs;
4897 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4898 // We divide by these constants so assume that we have at least one
4899 // instruction that uses at least one register.
4900 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4901 R.NumInstructions = std::max(R.NumInstructions, 1U);
4903 // We calculate the interleave count using the following formula.
4904 // Subtract the number of loop invariants from the number of available
4905 // registers. These registers are used by all of the interleaved instances.
4906 // Next, divide the remaining registers by the number of registers that is
4907 // required by the loop, in order to estimate how many parallel instances
4908 // fit without causing spills. All of this is rounded down if necessary to be
4909 // a power of two. We want power of two interleave count to simplify any
4910 // addressing operations or alignment considerations.
4911 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4914 // Don't count the induction variable as interleaved.
4915 if (EnableIndVarRegisterHeur)
4916 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4917 std::max(1U, (R.MaxLocalUsers - 1)));
4919 // Clamp the interleave ranges to reasonable counts.
4920 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4922 // Check if the user has overridden the max.
4924 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4925 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4927 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4928 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4931 // If we did not calculate the cost for VF (because the user selected the VF)
4932 // then we calculate the cost of VF here.
4934 LoopCost = expectedCost(VF);
4936 // Clamp the calculated IC to be between the 1 and the max interleave count
4937 // that the target allows.
4938 if (IC > MaxInterleaveCount)
4939 IC = MaxInterleaveCount;
4943 // Interleave if we vectorized this loop and there is a reduction that could
4944 // benefit from interleaving.
4945 if (VF > 1 && Legal->getReductionVars()->size()) {
4946 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4950 // Note that if we've already vectorized the loop we will have done the
4951 // runtime check and so interleaving won't require further checks.
4952 bool InterleavingRequiresRuntimePointerCheck =
4953 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4955 // We want to interleave small loops in order to reduce the loop overhead and
4956 // potentially expose ILP opportunities.
4957 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4958 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4959 // We assume that the cost overhead is 1 and we use the cost model
4960 // to estimate the cost of the loop and interleave until the cost of the
4961 // loop overhead is about 5% of the cost of the loop.
4963 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4965 // Interleave until store/load ports (estimated by max interleave count) are
4967 unsigned NumStores = Legal->getNumStores();
4968 unsigned NumLoads = Legal->getNumLoads();
4969 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4970 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4972 // If we have a scalar reduction (vector reductions are already dealt with
4973 // by this point), we can increase the critical path length if the loop
4974 // we're interleaving is inside another loop. Limit, by default to 2, so the
4975 // critical path only gets increased by one reduction operation.
4976 if (Legal->getReductionVars()->size() &&
4977 TheLoop->getLoopDepth() > 1) {
4978 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4979 SmallIC = std::min(SmallIC, F);
4980 StoresIC = std::min(StoresIC, F);
4981 LoadsIC = std::min(LoadsIC, F);
4984 if (EnableLoadStoreRuntimeInterleave &&
4985 std::max(StoresIC, LoadsIC) > SmallIC) {
4986 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4987 return std::max(StoresIC, LoadsIC);
4990 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4994 // Interleave if this is a large loop (small loops are already dealt with by
4996 // point) that could benefit from interleaving.
4997 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4998 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4999 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
5003 DEBUG(dbgs() << "LV: Not Interleaving.\n");
5007 LoopVectorizationCostModel::RegisterUsage
5008 LoopVectorizationCostModel::calculateRegisterUsage() {
5009 // This function calculates the register usage by measuring the highest number
5010 // of values that are alive at a single location. Obviously, this is a very
5011 // rough estimation. We scan the loop in a topological order in order and
5012 // assign a number to each instruction. We use RPO to ensure that defs are
5013 // met before their users. We assume that each instruction that has in-loop
5014 // users starts an interval. We record every time that an in-loop value is
5015 // used, so we have a list of the first and last occurrences of each
5016 // instruction. Next, we transpose this data structure into a multi map that
5017 // holds the list of intervals that *end* at a specific location. This multi
5018 // map allows us to perform a linear search. We scan the instructions linearly
5019 // and record each time that a new interval starts, by placing it in a set.
5020 // If we find this value in the multi-map then we remove it from the set.
5021 // The max register usage is the maximum size of the set.
5022 // We also search for instructions that are defined outside the loop, but are
5023 // used inside the loop. We need this number separately from the max-interval
5024 // usage number because when we unroll, loop-invariant values do not take
5026 LoopBlocksDFS DFS(TheLoop);
5030 R.NumInstructions = 0;
5032 // Each 'key' in the map opens a new interval. The values
5033 // of the map are the index of the 'last seen' usage of the
5034 // instruction that is the key.
5035 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5036 // Maps instruction to its index.
5037 DenseMap<unsigned, Instruction*> IdxToInstr;
5038 // Marks the end of each interval.
5039 IntervalMap EndPoint;
5040 // Saves the list of instruction indices that are used in the loop.
5041 SmallSet<Instruction*, 8> Ends;
5042 // Saves the list of values that are used in the loop but are
5043 // defined outside the loop, such as arguments and constants.
5044 SmallPtrSet<Value*, 8> LoopInvariants;
5047 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5048 be = DFS.endRPO(); bb != be; ++bb) {
5049 R.NumInstructions += (*bb)->size();
5050 for (Instruction &I : **bb) {
5051 IdxToInstr[Index++] = &I;
5053 // Save the end location of each USE.
5054 for (unsigned i = 0; i < I.getNumOperands(); ++i) {
5055 Value *U = I.getOperand(i);
5056 Instruction *Instr = dyn_cast<Instruction>(U);
5058 // Ignore non-instruction values such as arguments, constants, etc.
5059 if (!Instr) continue;
5061 // If this instruction is outside the loop then record it and continue.
5062 if (!TheLoop->contains(Instr)) {
5063 LoopInvariants.insert(Instr);
5067 // Overwrite previous end points.
5068 EndPoint[Instr] = Index;
5074 // Saves the list of intervals that end with the index in 'key'.
5075 typedef SmallVector<Instruction*, 2> InstrList;
5076 DenseMap<unsigned, InstrList> TransposeEnds;
5078 // Transpose the EndPoints to a list of values that end at each index.
5079 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5081 TransposeEnds[it->second].push_back(it->first);
5083 SmallSet<Instruction*, 8> OpenIntervals;
5084 unsigned MaxUsage = 0;
5087 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5088 for (unsigned int i = 0; i < Index; ++i) {
5089 Instruction *I = IdxToInstr[i];
5090 // Ignore instructions that are never used within the loop.
5091 if (!Ends.count(I)) continue;
5093 // Skip ignored values.
5094 if (ValuesToIgnore.count(I))
5097 // Remove all of the instructions that end at this location.
5098 InstrList &List = TransposeEnds[i];
5099 for (unsigned int j=0, e = List.size(); j < e; ++j)
5100 OpenIntervals.erase(List[j]);
5102 // Count the number of live interals.
5103 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5105 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5106 OpenIntervals.size() << '\n');
5108 // Add the current instruction to the list of open intervals.
5109 OpenIntervals.insert(I);
5112 unsigned Invariant = LoopInvariants.size();
5113 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5114 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5115 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5117 R.LoopInvariantRegs = Invariant;
5118 R.MaxLocalUsers = MaxUsage;
5122 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5126 for (Loop::block_iterator bb = TheLoop->block_begin(),
5127 be = TheLoop->block_end(); bb != be; ++bb) {
5128 unsigned BlockCost = 0;
5129 BasicBlock *BB = *bb;
5131 // For each instruction in the old loop.
5132 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5133 // Skip dbg intrinsics.
5134 if (isa<DbgInfoIntrinsic>(it))
5137 // Skip ignored values.
5138 if (ValuesToIgnore.count(&*it))
5141 unsigned C = getInstructionCost(&*it, VF);
5143 // Check if we should override the cost.
5144 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5145 C = ForceTargetInstructionCost;
5148 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5149 VF << " For instruction: " << *it << '\n');
5152 // We assume that if-converted blocks have a 50% chance of being executed.
5153 // When the code is scalar then some of the blocks are avoided due to CF.
5154 // When the code is vectorized we execute all code paths.
5155 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5164 /// \brief Check whether the address computation for a non-consecutive memory
5165 /// access looks like an unlikely candidate for being merged into the indexing
5168 /// We look for a GEP which has one index that is an induction variable and all
5169 /// other indices are loop invariant. If the stride of this access is also
5170 /// within a small bound we decide that this address computation can likely be
5171 /// merged into the addressing mode.
5172 /// In all other cases, we identify the address computation as complex.
5173 static bool isLikelyComplexAddressComputation(Value *Ptr,
5174 LoopVectorizationLegality *Legal,
5175 ScalarEvolution *SE,
5176 const Loop *TheLoop) {
5177 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5181 // We are looking for a gep with all loop invariant indices except for one
5182 // which should be an induction variable.
5183 unsigned NumOperands = Gep->getNumOperands();
5184 for (unsigned i = 1; i < NumOperands; ++i) {
5185 Value *Opd = Gep->getOperand(i);
5186 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5187 !Legal->isInductionVariable(Opd))
5191 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5192 // can likely be merged into the address computation.
5193 unsigned MaxMergeDistance = 64;
5195 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5199 // Check the step is constant.
5200 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5201 // Calculate the pointer stride and check if it is consecutive.
5202 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5206 const APInt &APStepVal = C->getValue()->getValue();
5208 // Huge step value - give up.
5209 if (APStepVal.getBitWidth() > 64)
5212 int64_t StepVal = APStepVal.getSExtValue();
5214 return StepVal > MaxMergeDistance;
5217 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5218 return Legal->hasStride(I->getOperand(0)) ||
5219 Legal->hasStride(I->getOperand(1));
5223 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5224 // If we know that this instruction will remain uniform, check the cost of
5225 // the scalar version.
5226 if (Legal->isUniformAfterVectorization(I))
5229 Type *RetTy = I->getType();
5230 if (VF > 1 && MinBWs.count(I))
5231 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
5232 Type *VectorTy = ToVectorTy(RetTy, VF);
5234 // TODO: We need to estimate the cost of intrinsic calls.
5235 switch (I->getOpcode()) {
5236 case Instruction::GetElementPtr:
5237 // We mark this instruction as zero-cost because the cost of GEPs in
5238 // vectorized code depends on whether the corresponding memory instruction
5239 // is scalarized or not. Therefore, we handle GEPs with the memory
5240 // instruction cost.
5242 case Instruction::Br: {
5243 return TTI.getCFInstrCost(I->getOpcode());
5245 case Instruction::PHI:
5246 //TODO: IF-converted IFs become selects.
5248 case Instruction::Add:
5249 case Instruction::FAdd:
5250 case Instruction::Sub:
5251 case Instruction::FSub:
5252 case Instruction::Mul:
5253 case Instruction::FMul:
5254 case Instruction::UDiv:
5255 case Instruction::SDiv:
5256 case Instruction::FDiv:
5257 case Instruction::URem:
5258 case Instruction::SRem:
5259 case Instruction::FRem:
5260 case Instruction::Shl:
5261 case Instruction::LShr:
5262 case Instruction::AShr:
5263 case Instruction::And:
5264 case Instruction::Or:
5265 case Instruction::Xor: {
5266 // Since we will replace the stride by 1 the multiplication should go away.
5267 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5269 // Certain instructions can be cheaper to vectorize if they have a constant
5270 // second vector operand. One example of this are shifts on x86.
5271 TargetTransformInfo::OperandValueKind Op1VK =
5272 TargetTransformInfo::OK_AnyValue;
5273 TargetTransformInfo::OperandValueKind Op2VK =
5274 TargetTransformInfo::OK_AnyValue;
5275 TargetTransformInfo::OperandValueProperties Op1VP =
5276 TargetTransformInfo::OP_None;
5277 TargetTransformInfo::OperandValueProperties Op2VP =
5278 TargetTransformInfo::OP_None;
5279 Value *Op2 = I->getOperand(1);
5281 // Check for a splat of a constant or for a non uniform vector of constants.
5282 if (isa<ConstantInt>(Op2)) {
5283 ConstantInt *CInt = cast<ConstantInt>(Op2);
5284 if (CInt && CInt->getValue().isPowerOf2())
5285 Op2VP = TargetTransformInfo::OP_PowerOf2;
5286 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5287 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5288 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5289 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5291 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5292 if (CInt && CInt->getValue().isPowerOf2())
5293 Op2VP = TargetTransformInfo::OP_PowerOf2;
5294 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5298 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5301 case Instruction::Select: {
5302 SelectInst *SI = cast<SelectInst>(I);
5303 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5304 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5305 Type *CondTy = SI->getCondition()->getType();
5307 CondTy = VectorType::get(CondTy, VF);
5309 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5311 case Instruction::ICmp:
5312 case Instruction::FCmp: {
5313 Type *ValTy = I->getOperand(0)->getType();
5314 if (VF > 1 && MinBWs.count(dyn_cast<Instruction>(I->getOperand(0))))
5315 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[I]);
5316 VectorTy = ToVectorTy(ValTy, VF);
5317 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5319 case Instruction::Store:
5320 case Instruction::Load: {
5321 StoreInst *SI = dyn_cast<StoreInst>(I);
5322 LoadInst *LI = dyn_cast<LoadInst>(I);
5323 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5325 VectorTy = ToVectorTy(ValTy, VF);
5327 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5328 unsigned AS = SI ? SI->getPointerAddressSpace() :
5329 LI->getPointerAddressSpace();
5330 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5331 // We add the cost of address computation here instead of with the gep
5332 // instruction because only here we know whether the operation is
5335 return TTI.getAddressComputationCost(VectorTy) +
5336 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5338 // For an interleaved access, calculate the total cost of the whole
5339 // interleave group.
5340 if (Legal->isAccessInterleaved(I)) {
5341 auto Group = Legal->getInterleavedAccessGroup(I);
5342 assert(Group && "Fail to get an interleaved access group.");
5344 // Only calculate the cost once at the insert position.
5345 if (Group->getInsertPos() != I)
5348 unsigned InterleaveFactor = Group->getFactor();
5350 VectorType::get(VectorTy->getVectorElementType(),
5351 VectorTy->getVectorNumElements() * InterleaveFactor);
5353 // Holds the indices of existing members in an interleaved load group.
5354 // An interleaved store group doesn't need this as it dones't allow gaps.
5355 SmallVector<unsigned, 4> Indices;
5357 for (unsigned i = 0; i < InterleaveFactor; i++)
5358 if (Group->getMember(i))
5359 Indices.push_back(i);
5362 // Calculate the cost of the whole interleaved group.
5363 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5364 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5365 Group->getAlignment(), AS);
5367 if (Group->isReverse())
5369 Group->getNumMembers() *
5370 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5372 // FIXME: The interleaved load group with a huge gap could be even more
5373 // expensive than scalar operations. Then we could ignore such group and
5374 // use scalar operations instead.
5378 // Scalarized loads/stores.
5379 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5380 bool Reverse = ConsecutiveStride < 0;
5381 const DataLayout &DL = I->getModule()->getDataLayout();
5382 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5383 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5384 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5385 bool IsComplexComputation =
5386 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5388 // The cost of extracting from the value vector and pointer vector.
5389 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5390 for (unsigned i = 0; i < VF; ++i) {
5391 // The cost of extracting the pointer operand.
5392 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5393 // In case of STORE, the cost of ExtractElement from the vector.
5394 // In case of LOAD, the cost of InsertElement into the returned
5396 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5397 Instruction::InsertElement,
5401 // The cost of the scalar loads/stores.
5402 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5403 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5408 // Wide load/stores.
5409 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5410 if (Legal->isMaskRequired(I))
5411 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5414 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5417 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5421 case Instruction::ZExt:
5422 case Instruction::SExt:
5423 case Instruction::FPToUI:
5424 case Instruction::FPToSI:
5425 case Instruction::FPExt:
5426 case Instruction::PtrToInt:
5427 case Instruction::IntToPtr:
5428 case Instruction::SIToFP:
5429 case Instruction::UIToFP:
5430 case Instruction::Trunc:
5431 case Instruction::FPTrunc:
5432 case Instruction::BitCast: {
5433 // We optimize the truncation of induction variable.
5434 // The cost of these is the same as the scalar operation.
5435 if (I->getOpcode() == Instruction::Trunc &&
5436 Legal->isInductionVariable(I->getOperand(0)))
5437 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5438 I->getOperand(0)->getType());
5440 Type *SrcScalarTy = I->getOperand(0)->getType();
5441 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
5442 if (VF > 1 && MinBWs.count(I)) {
5443 // This cast is going to be shrunk. This may remove the cast or it might
5444 // turn it into slightly different cast. For example, if MinBW == 16,
5445 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
5447 // Calculate the modified src and dest types.
5448 Type *MinVecTy = VectorTy;
5449 if (I->getOpcode() == Instruction::Trunc) {
5450 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
5451 VectorTy = largestIntegerVectorType(ToVectorTy(I->getType(), VF),
5453 } else if (I->getOpcode() == Instruction::ZExt ||
5454 I->getOpcode() == Instruction::SExt) {
5455 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
5456 VectorTy = smallestIntegerVectorType(ToVectorTy(I->getType(), VF),
5461 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5463 case Instruction::Call: {
5464 bool NeedToScalarize;
5465 CallInst *CI = cast<CallInst>(I);
5466 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5467 if (getIntrinsicIDForCall(CI, TLI))
5468 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5472 // We are scalarizing the instruction. Return the cost of the scalar
5473 // instruction, plus the cost of insert and extract into vector
5474 // elements, times the vector width.
5477 if (!RetTy->isVoidTy() && VF != 1) {
5478 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5480 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5483 // The cost of inserting the results plus extracting each one of the
5485 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5488 // The cost of executing VF copies of the scalar instruction. This opcode
5489 // is unknown. Assume that it is the same as 'mul'.
5490 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5496 char LoopVectorize::ID = 0;
5497 static const char lv_name[] = "Loop Vectorization";
5498 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5499 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5500 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
5501 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
5502 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
5503 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5504 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5505 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5506 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5507 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5508 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5509 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5510 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5511 INITIALIZE_PASS_DEPENDENCY(DemandedBits)
5512 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5515 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5516 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5520 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5521 // Check for a store.
5522 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5523 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5525 // Check for a load.
5526 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5527 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5533 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5534 bool IfPredicateStore) {
5535 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5536 // Holds vector parameters or scalars, in case of uniform vals.
5537 SmallVector<VectorParts, 4> Params;
5539 setDebugLocFromInst(Builder, Instr);
5541 // Find all of the vectorized parameters.
5542 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5543 Value *SrcOp = Instr->getOperand(op);
5545 // If we are accessing the old induction variable, use the new one.
5546 if (SrcOp == OldInduction) {
5547 Params.push_back(getVectorValue(SrcOp));
5551 // Try using previously calculated values.
5552 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5554 // If the src is an instruction that appeared earlier in the basic block
5555 // then it should already be vectorized.
5556 if (SrcInst && OrigLoop->contains(SrcInst)) {
5557 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5558 // The parameter is a vector value from earlier.
5559 Params.push_back(WidenMap.get(SrcInst));
5561 // The parameter is a scalar from outside the loop. Maybe even a constant.
5562 VectorParts Scalars;
5563 Scalars.append(UF, SrcOp);
5564 Params.push_back(Scalars);
5568 assert(Params.size() == Instr->getNumOperands() &&
5569 "Invalid number of operands");
5571 // Does this instruction return a value ?
5572 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5574 Value *UndefVec = IsVoidRetTy ? nullptr :
5575 UndefValue::get(Instr->getType());
5576 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5577 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5580 if (IfPredicateStore) {
5581 assert(Instr->getParent()->getSinglePredecessor() &&
5582 "Only support single predecessor blocks");
5583 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5584 Instr->getParent());
5587 // For each vector unroll 'part':
5588 for (unsigned Part = 0; Part < UF; ++Part) {
5589 // For each scalar that we create:
5591 // Start an "if (pred) a[i] = ..." block.
5592 Value *Cmp = nullptr;
5593 if (IfPredicateStore) {
5594 if (Cond[Part]->getType()->isVectorTy())
5596 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5597 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5598 ConstantInt::get(Cond[Part]->getType(), 1));
5601 Instruction *Cloned = Instr->clone();
5603 Cloned->setName(Instr->getName() + ".cloned");
5604 // Replace the operands of the cloned instructions with extracted scalars.
5605 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5606 Value *Op = Params[op][Part];
5607 Cloned->setOperand(op, Op);
5610 // Place the cloned scalar in the new loop.
5611 Builder.Insert(Cloned);
5613 // If the original scalar returns a value we need to place it in a vector
5614 // so that future users will be able to use it.
5616 VecResults[Part] = Cloned;
5619 if (IfPredicateStore)
5620 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
5625 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5626 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5627 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5629 return scalarizeInstruction(Instr, IfPredicateStore);
5632 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5636 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5640 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5641 // When unrolling and the VF is 1, we only need to add a simple scalar.
5642 Type *ITy = Val->getType();
5643 assert(!ITy->isVectorTy() && "Val must be a scalar");
5644 Constant *C = ConstantInt::get(ITy, StartIdx);
5645 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");