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 static cl::opt<bool> MaximizeBandwidth(
130 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
131 cl::desc("Maximize bandwidth when selecting vectorization factor which "
132 "will be determined by the smallest type in loop."));
134 /// This enables versioning on the strides of symbolically striding memory
135 /// accesses in code like the following.
136 /// for (i = 0; i < N; ++i)
137 /// A[i * Stride1] += B[i * Stride2] ...
139 /// Will be roughly translated to
140 /// if (Stride1 == 1 && Stride2 == 1) {
141 /// for (i = 0; i < N; i+=4)
145 static cl::opt<bool> EnableMemAccessVersioning(
146 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
147 cl::desc("Enable symblic stride memory access versioning"));
149 static cl::opt<bool> EnableInterleavedMemAccesses(
150 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
151 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
153 /// Maximum factor for an interleaved memory access.
154 static cl::opt<unsigned> MaxInterleaveGroupFactor(
155 "max-interleave-group-factor", cl::Hidden,
156 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
159 /// We don't interleave loops with a known constant trip count below this
161 static const unsigned TinyTripCountInterleaveThreshold = 128;
163 static cl::opt<unsigned> ForceTargetNumScalarRegs(
164 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's number of scalar registers."));
167 static cl::opt<unsigned> ForceTargetNumVectorRegs(
168 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
169 cl::desc("A flag that overrides the target's number of vector registers."));
171 /// Maximum vectorization interleave count.
172 static const unsigned MaxInterleaveFactor = 16;
174 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
175 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
176 cl::desc("A flag that overrides the target's max interleave factor for "
179 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
180 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
181 cl::desc("A flag that overrides the target's max interleave factor for "
182 "vectorized loops."));
184 static cl::opt<unsigned> ForceTargetInstructionCost(
185 "force-target-instruction-cost", cl::init(0), cl::Hidden,
186 cl::desc("A flag that overrides the target's expected cost for "
187 "an instruction to a single constant value. Mostly "
188 "useful for getting consistent testing."));
190 static cl::opt<unsigned> SmallLoopCost(
191 "small-loop-cost", cl::init(20), cl::Hidden,
193 "The cost of a loop that is considered 'small' by the interleaver."));
195 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
196 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
197 cl::desc("Enable the use of the block frequency analysis to access PGO "
198 "heuristics minimizing code growth in cold regions and being more "
199 "aggressive in hot regions."));
201 // Runtime interleave loops for load/store throughput.
202 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
203 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
205 "Enable runtime interleaving until load/store ports are saturated"));
207 /// The number of stores in a loop that are allowed to need predication.
208 static cl::opt<unsigned> NumberOfStoresToPredicate(
209 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
210 cl::desc("Max number of stores to be predicated behind an if."));
212 static cl::opt<bool> EnableIndVarRegisterHeur(
213 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
214 cl::desc("Count the induction variable only once when interleaving"));
216 static cl::opt<bool> EnableCondStoresVectorization(
217 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
218 cl::desc("Enable if predication of stores during vectorization."));
220 static cl::opt<unsigned> MaxNestedScalarReductionIC(
221 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
222 cl::desc("The maximum interleave count to use when interleaving a scalar "
223 "reduction in a nested loop."));
225 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
226 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
227 cl::desc("The maximum allowed number of runtime memory checks with a "
228 "vectorize(enable) pragma."));
232 // Forward declarations.
233 class LoopVectorizeHints;
234 class LoopVectorizationLegality;
235 class LoopVectorizationCostModel;
236 class LoopVectorizationRequirements;
238 /// \brief This modifies LoopAccessReport to initialize message with
239 /// loop-vectorizer-specific part.
240 class VectorizationReport : public LoopAccessReport {
242 VectorizationReport(Instruction *I = nullptr)
243 : LoopAccessReport("loop not vectorized: ", I) {}
245 /// \brief This allows promotion of the loop-access analysis report into the
246 /// loop-vectorizer report. It modifies the message to add the
247 /// loop-vectorizer-specific part of the message.
248 explicit VectorizationReport(const LoopAccessReport &R)
249 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
253 /// A helper function for converting Scalar types to vector types.
254 /// If the incoming type is void, we return void. If the VF is 1, we return
256 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
257 if (Scalar->isVoidTy() || VF == 1)
259 return VectorType::get(Scalar, VF);
262 /// InnerLoopVectorizer vectorizes loops which contain only one basic
263 /// block to a specified vectorization factor (VF).
264 /// This class performs the widening of scalars into vectors, or multiple
265 /// scalars. This class also implements the following features:
266 /// * It inserts an epilogue loop for handling loops that don't have iteration
267 /// counts that are known to be a multiple of the vectorization factor.
268 /// * It handles the code generation for reduction variables.
269 /// * Scalarization (implementation using scalars) of un-vectorizable
271 /// InnerLoopVectorizer does not perform any vectorization-legality
272 /// checks, and relies on the caller to check for the different legality
273 /// aspects. The InnerLoopVectorizer relies on the
274 /// LoopVectorizationLegality class to provide information about the induction
275 /// and reduction variables that were found to a given vectorization factor.
276 class InnerLoopVectorizer {
278 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
279 DominatorTree *DT, const TargetLibraryInfo *TLI,
280 const TargetTransformInfo *TTI, unsigned VecWidth,
281 unsigned UnrollFactor)
282 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
283 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
284 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
285 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr),
286 AddedSafetyChecks(false) {}
288 // Perform the actual loop widening (vectorization).
289 // MinimumBitWidths maps scalar integer values to the smallest bitwidth they
290 // can be validly truncated to. The cost model has assumed this truncation
291 // will happen when vectorizing.
292 void vectorize(LoopVectorizationLegality *L,
293 DenseMap<Instruction*,uint64_t> MinimumBitWidths) {
294 MinBWs = MinimumBitWidths;
296 // Create a new empty loop. Unlink the old loop and connect the new one.
298 // Widen each instruction in the old loop to a new one in the new loop.
299 // Use the Legality module to find the induction and reduction variables.
303 // Return true if any runtime check is added.
304 bool IsSafetyChecksAdded() {
305 return AddedSafetyChecks;
308 virtual ~InnerLoopVectorizer() {}
311 /// A small list of PHINodes.
312 typedef SmallVector<PHINode*, 4> PhiVector;
313 /// When we unroll loops we have multiple vector values for each scalar.
314 /// This data structure holds the unrolled and vectorized values that
315 /// originated from one scalar instruction.
316 typedef SmallVector<Value*, 2> VectorParts;
318 // When we if-convert we need to create edge masks. We have to cache values
319 // so that we don't end up with exponential recursion/IR.
320 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
321 VectorParts> EdgeMaskCache;
323 /// \brief Add checks for strides that were assumed to be 1.
325 /// Returns the last check instruction and the first check instruction in the
326 /// pair as (first, last).
327 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
329 /// Create an empty loop, based on the loop ranges of the old loop.
330 void createEmptyLoop();
331 /// Create a new induction variable inside L.
332 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
333 Value *Step, Instruction *DL);
334 /// Copy and widen the instructions from the old loop.
335 virtual void vectorizeLoop();
337 /// \brief The Loop exit block may have single value PHI nodes where the
338 /// incoming value is 'Undef'. While vectorizing we only handled real values
339 /// that were defined inside the loop. Here we fix the 'undef case'.
343 /// Shrinks vector element sizes based on information in "MinBWs".
344 void truncateToMinimalBitwidths();
346 /// A helper function that computes the predicate of the block BB, assuming
347 /// that the header block of the loop is set to True. It returns the *entry*
348 /// mask for the block BB.
349 VectorParts createBlockInMask(BasicBlock *BB);
350 /// A helper function that computes the predicate of the edge between SRC
352 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
354 /// A helper function to vectorize a single BB within the innermost loop.
355 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
357 /// Vectorize a single PHINode in a block. This method handles the induction
358 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
359 /// arbitrary length vectors.
360 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
361 unsigned UF, unsigned VF, PhiVector *PV);
363 /// Insert the new loop to the loop hierarchy and pass manager
364 /// and update the analysis passes.
365 void updateAnalysis();
367 /// This instruction is un-vectorizable. Implement it as a sequence
368 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
369 /// scalarized instruction behind an if block predicated on the control
370 /// dependence of the instruction.
371 virtual void scalarizeInstruction(Instruction *Instr,
372 bool IfPredicateStore=false);
374 /// Vectorize Load and Store instructions,
375 virtual void vectorizeMemoryInstruction(Instruction *Instr);
377 /// Create a broadcast instruction. This method generates a broadcast
378 /// instruction (shuffle) for loop invariant values and for the induction
379 /// value. If this is the induction variable then we extend it to N, N+1, ...
380 /// this is needed because each iteration in the loop corresponds to a SIMD
382 virtual Value *getBroadcastInstrs(Value *V);
384 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
385 /// to each vector element of Val. The sequence starts at StartIndex.
386 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
388 /// When we go over instructions in the basic block we rely on previous
389 /// values within the current basic block or on loop invariant values.
390 /// When we widen (vectorize) values we place them in the map. If the values
391 /// are not within the map, they have to be loop invariant, so we simply
392 /// broadcast them into a vector.
393 VectorParts &getVectorValue(Value *V);
395 /// Try to vectorize the interleaved access group that \p Instr belongs to.
396 void vectorizeInterleaveGroup(Instruction *Instr);
398 /// Generate a shuffle sequence that will reverse the vector Vec.
399 virtual Value *reverseVector(Value *Vec);
401 /// Returns (and creates if needed) the original loop trip count.
402 Value *getOrCreateTripCount(Loop *NewLoop);
404 /// Returns (and creates if needed) the trip count of the widened loop.
405 Value *getOrCreateVectorTripCount(Loop *NewLoop);
407 /// Emit a bypass check to see if the trip count would overflow, or we
408 /// wouldn't have enough iterations to execute one vector loop.
409 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
410 /// Emit a bypass check to see if the vector trip count is nonzero.
411 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
412 /// Emit bypass checks to check if strides we've assumed to be one really are.
413 void emitStrideChecks(Loop *L, BasicBlock *Bypass);
414 /// Emit bypass checks to check any memory assumptions we may have made.
415 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
417 /// This is a helper class that holds the vectorizer state. It maps scalar
418 /// instructions to vector instructions. When the code is 'unrolled' then
419 /// then a single scalar value is mapped to multiple vector parts. The parts
420 /// are stored in the VectorPart type.
422 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
424 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
426 /// \return True if 'Key' is saved in the Value Map.
427 bool has(Value *Key) const { return MapStorage.count(Key); }
429 /// Initializes a new entry in the map. Sets all of the vector parts to the
430 /// save value in 'Val'.
431 /// \return A reference to a vector with splat values.
432 VectorParts &splat(Value *Key, Value *Val) {
433 VectorParts &Entry = MapStorage[Key];
434 Entry.assign(UF, Val);
438 ///\return A reference to the value that is stored at 'Key'.
439 VectorParts &get(Value *Key) {
440 VectorParts &Entry = MapStorage[Key];
443 assert(Entry.size() == UF);
448 /// The unroll factor. Each entry in the map stores this number of vector
452 /// Map storage. We use std::map and not DenseMap because insertions to a
453 /// dense map invalidates its iterators.
454 std::map<Value *, VectorParts> MapStorage;
457 /// The original loop.
459 /// Scev analysis to use.
467 /// Target Library Info.
468 const TargetLibraryInfo *TLI;
469 /// Target Transform Info.
470 const TargetTransformInfo *TTI;
472 /// The vectorization SIMD factor to use. Each vector will have this many
477 /// The vectorization unroll factor to use. Each scalar is vectorized to this
478 /// many different vector instructions.
481 /// The builder that we use
484 // --- Vectorization state ---
486 /// The vector-loop preheader.
487 BasicBlock *LoopVectorPreHeader;
488 /// The scalar-loop preheader.
489 BasicBlock *LoopScalarPreHeader;
490 /// Middle Block between the vector and the scalar.
491 BasicBlock *LoopMiddleBlock;
492 ///The ExitBlock of the scalar loop.
493 BasicBlock *LoopExitBlock;
494 ///The vector loop body.
495 SmallVector<BasicBlock *, 4> LoopVectorBody;
496 ///The scalar loop body.
497 BasicBlock *LoopScalarBody;
498 /// A list of all bypass blocks. The first block is the entry of the loop.
499 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
501 /// The new Induction variable which was added to the new block.
503 /// The induction variable of the old basic block.
504 PHINode *OldInduction;
505 /// Maps scalars to widened vectors.
507 /// Store instructions that should be predicated, as a pair
508 /// <StoreInst, Predicate>
509 SmallVector<std::pair<StoreInst*,Value*>, 4> PredicatedStores;
510 EdgeMaskCache MaskCache;
511 /// Trip count of the original loop.
513 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
514 Value *VectorTripCount;
516 /// Map of scalar integer values to the smallest bitwidth they can be legally
517 /// represented as. The vector equivalents of these values should be truncated
519 DenseMap<Instruction*,uint64_t> MinBWs;
520 LoopVectorizationLegality *Legal;
522 // Record whether runtime check is added.
523 bool AddedSafetyChecks;
526 class InnerLoopUnroller : public InnerLoopVectorizer {
528 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
529 DominatorTree *DT, const TargetLibraryInfo *TLI,
530 const TargetTransformInfo *TTI, unsigned UnrollFactor)
531 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
534 void scalarizeInstruction(Instruction *Instr,
535 bool IfPredicateStore = false) override;
536 void vectorizeMemoryInstruction(Instruction *Instr) override;
537 Value *getBroadcastInstrs(Value *V) override;
538 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
539 Value *reverseVector(Value *Vec) override;
542 /// \brief Look for a meaningful debug location on the instruction or it's
544 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
549 if (I->getDebugLoc() != Empty)
552 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
553 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
554 if (OpInst->getDebugLoc() != Empty)
561 /// \brief Set the debug location in the builder using the debug location in the
563 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
564 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
565 B.SetCurrentDebugLocation(Inst->getDebugLoc());
567 B.SetCurrentDebugLocation(DebugLoc());
571 /// \return string containing a file name and a line # for the given loop.
572 static std::string getDebugLocString(const Loop *L) {
575 raw_string_ostream OS(Result);
576 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
577 LoopDbgLoc.print(OS);
579 // Just print the module name.
580 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
587 /// \brief Propagate known metadata from one instruction to another.
588 static void propagateMetadata(Instruction *To, const Instruction *From) {
589 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
590 From->getAllMetadataOtherThanDebugLoc(Metadata);
592 for (auto M : Metadata) {
593 unsigned Kind = M.first;
595 // These are safe to transfer (this is safe for TBAA, even when we
596 // if-convert, because should that metadata have had a control dependency
597 // on the condition, and thus actually aliased with some other
598 // non-speculated memory access when the condition was false, this would be
599 // caught by the runtime overlap checks).
600 if (Kind != LLVMContext::MD_tbaa &&
601 Kind != LLVMContext::MD_alias_scope &&
602 Kind != LLVMContext::MD_noalias &&
603 Kind != LLVMContext::MD_fpmath &&
604 Kind != LLVMContext::MD_nontemporal)
607 To->setMetadata(Kind, M.second);
611 /// \brief Propagate known metadata from one instruction to a vector of others.
612 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
614 if (Instruction *I = dyn_cast<Instruction>(V))
615 propagateMetadata(I, From);
618 /// \brief The group of interleaved loads/stores sharing the same stride and
619 /// close to each other.
621 /// Each member in this group has an index starting from 0, and the largest
622 /// index should be less than interleaved factor, which is equal to the absolute
623 /// value of the access's stride.
625 /// E.g. An interleaved load group of factor 4:
626 /// for (unsigned i = 0; i < 1024; i+=4) {
627 /// a = A[i]; // Member of index 0
628 /// b = A[i+1]; // Member of index 1
629 /// d = A[i+3]; // Member of index 3
633 /// An interleaved store group of factor 4:
634 /// for (unsigned i = 0; i < 1024; i+=4) {
636 /// A[i] = a; // Member of index 0
637 /// A[i+1] = b; // Member of index 1
638 /// A[i+2] = c; // Member of index 2
639 /// A[i+3] = d; // Member of index 3
642 /// Note: the interleaved load group could have gaps (missing members), but
643 /// the interleaved store group doesn't allow gaps.
644 class InterleaveGroup {
646 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
647 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
648 assert(Align && "The alignment should be non-zero");
650 Factor = std::abs(Stride);
651 assert(Factor > 1 && "Invalid interleave factor");
653 Reverse = Stride < 0;
657 bool isReverse() const { return Reverse; }
658 unsigned getFactor() const { return Factor; }
659 unsigned getAlignment() const { return Align; }
660 unsigned getNumMembers() const { return Members.size(); }
662 /// \brief Try to insert a new member \p Instr with index \p Index and
663 /// alignment \p NewAlign. The index is related to the leader and it could be
664 /// negative if it is the new leader.
666 /// \returns false if the instruction doesn't belong to the group.
667 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
668 assert(NewAlign && "The new member's alignment should be non-zero");
670 int Key = Index + SmallestKey;
672 // Skip if there is already a member with the same index.
673 if (Members.count(Key))
676 if (Key > LargestKey) {
677 // The largest index is always less than the interleave factor.
678 if (Index >= static_cast<int>(Factor))
682 } else if (Key < SmallestKey) {
683 // The largest index is always less than the interleave factor.
684 if (LargestKey - Key >= static_cast<int>(Factor))
690 // It's always safe to select the minimum alignment.
691 Align = std::min(Align, NewAlign);
692 Members[Key] = Instr;
696 /// \brief Get the member with the given index \p Index
698 /// \returns nullptr if contains no such member.
699 Instruction *getMember(unsigned Index) const {
700 int Key = SmallestKey + Index;
701 if (!Members.count(Key))
704 return Members.find(Key)->second;
707 /// \brief Get the index for the given member. Unlike the key in the member
708 /// map, the index starts from 0.
709 unsigned getIndex(Instruction *Instr) const {
710 for (auto I : Members)
711 if (I.second == Instr)
712 return I.first - SmallestKey;
714 llvm_unreachable("InterleaveGroup contains no such member");
717 Instruction *getInsertPos() const { return InsertPos; }
718 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
721 unsigned Factor; // Interleave Factor.
724 DenseMap<int, Instruction *> Members;
728 // To avoid breaking dependences, vectorized instructions of an interleave
729 // group should be inserted at either the first load or the last store in
732 // E.g. %even = load i32 // Insert Position
733 // %add = add i32 %even // Use of %even
737 // %odd = add i32 // Def of %odd
738 // store i32 %odd // Insert Position
739 Instruction *InsertPos;
742 /// \brief Drive the analysis of interleaved memory accesses in the loop.
744 /// Use this class to analyze interleaved accesses only when we can vectorize
745 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
746 /// on interleaved accesses is unsafe.
748 /// The analysis collects interleave groups and records the relationships
749 /// between the member and the group in a map.
750 class InterleavedAccessInfo {
752 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
753 : SE(SE), TheLoop(L), DT(DT) {}
755 ~InterleavedAccessInfo() {
756 SmallSet<InterleaveGroup *, 4> DelSet;
757 // Avoid releasing a pointer twice.
758 for (auto &I : InterleaveGroupMap)
759 DelSet.insert(I.second);
760 for (auto *Ptr : DelSet)
764 /// \brief Analyze the interleaved accesses and collect them in interleave
765 /// groups. Substitute symbolic strides using \p Strides.
766 void analyzeInterleaving(const ValueToValueMap &Strides);
768 /// \brief Check if \p Instr belongs to any interleave group.
769 bool isInterleaved(Instruction *Instr) const {
770 return InterleaveGroupMap.count(Instr);
773 /// \brief Get the interleave group that \p Instr belongs to.
775 /// \returns nullptr if doesn't have such group.
776 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
777 if (InterleaveGroupMap.count(Instr))
778 return InterleaveGroupMap.find(Instr)->second;
787 /// Holds the relationships between the members and the interleave group.
788 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
790 /// \brief The descriptor for a strided memory access.
791 struct StrideDescriptor {
792 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
794 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
796 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
798 int Stride; // The access's stride. It is negative for a reverse access.
799 const SCEV *Scev; // The scalar expression of this access
800 unsigned Size; // The size of the memory object.
801 unsigned Align; // The alignment of this access.
804 /// \brief Create a new interleave group with the given instruction \p Instr,
805 /// stride \p Stride and alignment \p Align.
807 /// \returns the newly created interleave group.
808 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
810 assert(!InterleaveGroupMap.count(Instr) &&
811 "Already in an interleaved access group");
812 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
813 return InterleaveGroupMap[Instr];
816 /// \brief Release the group and remove all the relationships.
817 void releaseGroup(InterleaveGroup *Group) {
818 for (unsigned i = 0; i < Group->getFactor(); i++)
819 if (Instruction *Member = Group->getMember(i))
820 InterleaveGroupMap.erase(Member);
825 /// \brief Collect all the accesses with a constant stride in program order.
826 void collectConstStridedAccesses(
827 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
828 const ValueToValueMap &Strides);
831 /// Utility class for getting and setting loop vectorizer hints in the form
832 /// of loop metadata.
833 /// This class keeps a number of loop annotations locally (as member variables)
834 /// and can, upon request, write them back as metadata on the loop. It will
835 /// initially scan the loop for existing metadata, and will update the local
836 /// values based on information in the loop.
837 /// We cannot write all values to metadata, as the mere presence of some info,
838 /// for example 'force', means a decision has been made. So, we need to be
839 /// careful NOT to add them if the user hasn't specifically asked so.
840 class LoopVectorizeHints {
847 /// Hint - associates name and validation with the hint value.
850 unsigned Value; // This may have to change for non-numeric values.
853 Hint(const char * Name, unsigned Value, HintKind Kind)
854 : Name(Name), Value(Value), Kind(Kind) { }
856 bool validate(unsigned Val) {
859 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
861 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
869 /// Vectorization width.
871 /// Vectorization interleave factor.
873 /// Vectorization forced
876 /// Return the loop metadata prefix.
877 static StringRef Prefix() { return "llvm.loop."; }
881 FK_Undefined = -1, ///< Not selected.
882 FK_Disabled = 0, ///< Forcing disabled.
883 FK_Enabled = 1, ///< Forcing enabled.
886 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
887 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
889 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
890 Force("vectorize.enable", FK_Undefined, HK_FORCE),
892 // Populate values with existing loop metadata.
893 getHintsFromMetadata();
895 // force-vector-interleave overrides DisableInterleaving.
896 if (VectorizerParams::isInterleaveForced())
897 Interleave.Value = VectorizerParams::VectorizationInterleave;
899 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
900 << "LV: Interleaving disabled by the pass manager\n");
903 /// Mark the loop L as already vectorized by setting the width to 1.
904 void setAlreadyVectorized() {
905 Width.Value = Interleave.Value = 1;
906 Hint Hints[] = {Width, Interleave};
907 writeHintsToMetadata(Hints);
910 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
911 if (getForce() == LoopVectorizeHints::FK_Disabled) {
912 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
913 emitOptimizationRemarkAnalysis(F->getContext(),
914 vectorizeAnalysisPassName(), *F,
915 L->getStartLoc(), emitRemark());
919 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
920 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
921 emitOptimizationRemarkAnalysis(F->getContext(),
922 vectorizeAnalysisPassName(), *F,
923 L->getStartLoc(), emitRemark());
927 if (getWidth() == 1 && getInterleave() == 1) {
928 // FIXME: Add a separate metadata to indicate when the loop has already
929 // been vectorized instead of setting width and count to 1.
930 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
931 // FIXME: Add interleave.disable metadata. This will allow
932 // vectorize.disable to be used without disabling the pass and errors
933 // to differentiate between disabled vectorization and a width of 1.
934 emitOptimizationRemarkAnalysis(
935 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
936 "loop not vectorized: vectorization and interleaving are explicitly "
937 "disabled, or vectorize width and interleave count are both set to "
945 /// Dumps all the hint information.
946 std::string emitRemark() const {
947 VectorizationReport R;
948 if (Force.Value == LoopVectorizeHints::FK_Disabled)
949 R << "vectorization is explicitly disabled";
951 R << "use -Rpass-analysis=loop-vectorize for more info";
952 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
954 if (Width.Value != 0)
955 R << ", Vector Width=" << Width.Value;
956 if (Interleave.Value != 0)
957 R << ", Interleave Count=" << Interleave.Value;
965 unsigned getWidth() const { return Width.Value; }
966 unsigned getInterleave() const { return Interleave.Value; }
967 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
968 const char *vectorizeAnalysisPassName() const {
969 // If hints are provided that don't disable vectorization use the
970 // AlwaysPrint pass name to force the frontend to print the diagnostic.
973 if (getForce() == LoopVectorizeHints::FK_Disabled)
975 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
977 return DiagnosticInfo::AlwaysPrint;
980 bool allowReordering() const {
981 // When enabling loop hints are provided we allow the vectorizer to change
982 // the order of operations that is given by the scalar loop. This is not
983 // enabled by default because can be unsafe or inefficient. For example,
984 // reordering floating-point operations will change the way round-off
985 // error accumulates in the loop.
986 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
990 /// Find hints specified in the loop metadata and update local values.
991 void getHintsFromMetadata() {
992 MDNode *LoopID = TheLoop->getLoopID();
996 // First operand should refer to the loop id itself.
997 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
998 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1000 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1001 const MDString *S = nullptr;
1002 SmallVector<Metadata *, 4> Args;
1004 // The expected hint is either a MDString or a MDNode with the first
1005 // operand a MDString.
1006 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1007 if (!MD || MD->getNumOperands() == 0)
1009 S = dyn_cast<MDString>(MD->getOperand(0));
1010 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1011 Args.push_back(MD->getOperand(i));
1013 S = dyn_cast<MDString>(LoopID->getOperand(i));
1014 assert(Args.size() == 0 && "too many arguments for MDString");
1020 // Check if the hint starts with the loop metadata prefix.
1021 StringRef Name = S->getString();
1022 if (Args.size() == 1)
1023 setHint(Name, Args[0]);
1027 /// Checks string hint with one operand and set value if valid.
1028 void setHint(StringRef Name, Metadata *Arg) {
1029 if (!Name.startswith(Prefix()))
1031 Name = Name.substr(Prefix().size(), StringRef::npos);
1033 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1035 unsigned Val = C->getZExtValue();
1037 Hint *Hints[] = {&Width, &Interleave, &Force};
1038 for (auto H : Hints) {
1039 if (Name == H->Name) {
1040 if (H->validate(Val))
1043 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1049 /// Create a new hint from name / value pair.
1050 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1051 LLVMContext &Context = TheLoop->getHeader()->getContext();
1052 Metadata *MDs[] = {MDString::get(Context, Name),
1053 ConstantAsMetadata::get(
1054 ConstantInt::get(Type::getInt32Ty(Context), V))};
1055 return MDNode::get(Context, MDs);
1058 /// Matches metadata with hint name.
1059 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1060 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1064 for (auto H : HintTypes)
1065 if (Name->getString().endswith(H.Name))
1070 /// Sets current hints into loop metadata, keeping other values intact.
1071 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1072 if (HintTypes.size() == 0)
1075 // Reserve the first element to LoopID (see below).
1076 SmallVector<Metadata *, 4> MDs(1);
1077 // If the loop already has metadata, then ignore the existing operands.
1078 MDNode *LoopID = TheLoop->getLoopID();
1080 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1081 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1082 // If node in update list, ignore old value.
1083 if (!matchesHintMetadataName(Node, HintTypes))
1084 MDs.push_back(Node);
1088 // Now, add the missing hints.
1089 for (auto H : HintTypes)
1090 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1092 // Replace current metadata node with new one.
1093 LLVMContext &Context = TheLoop->getHeader()->getContext();
1094 MDNode *NewLoopID = MDNode::get(Context, MDs);
1095 // Set operand 0 to refer to the loop id itself.
1096 NewLoopID->replaceOperandWith(0, NewLoopID);
1098 TheLoop->setLoopID(NewLoopID);
1101 /// The loop these hints belong to.
1102 const Loop *TheLoop;
1105 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1106 const LoopVectorizeHints &Hints,
1107 const LoopAccessReport &Message) {
1108 const char *Name = Hints.vectorizeAnalysisPassName();
1109 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1112 static void emitMissedWarning(Function *F, Loop *L,
1113 const LoopVectorizeHints &LH) {
1114 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1117 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1118 if (LH.getWidth() != 1)
1119 emitLoopVectorizeWarning(
1120 F->getContext(), *F, L->getStartLoc(),
1121 "failed explicitly specified loop vectorization");
1122 else if (LH.getInterleave() != 1)
1123 emitLoopInterleaveWarning(
1124 F->getContext(), *F, L->getStartLoc(),
1125 "failed explicitly specified loop interleaving");
1129 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1130 /// to what vectorization factor.
1131 /// This class does not look at the profitability of vectorization, only the
1132 /// legality. This class has two main kinds of checks:
1133 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1134 /// will change the order of memory accesses in a way that will change the
1135 /// correctness of the program.
1136 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1137 /// checks for a number of different conditions, such as the availability of a
1138 /// single induction variable, that all types are supported and vectorize-able,
1139 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1140 /// This class is also used by InnerLoopVectorizer for identifying
1141 /// induction variable and the different reduction variables.
1142 class LoopVectorizationLegality {
1144 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1145 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1146 Function *F, const TargetTransformInfo *TTI,
1147 LoopAccessAnalysis *LAA,
1148 LoopVectorizationRequirements *R,
1149 const LoopVectorizeHints *H)
1150 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1151 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1152 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1153 Requirements(R), Hints(H) {}
1155 /// ReductionList contains the reduction descriptors for all
1156 /// of the reductions that were found in the loop.
1157 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1159 /// InductionList saves induction variables and maps them to the
1160 /// induction descriptor.
1161 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1163 /// Returns true if it is legal to vectorize this loop.
1164 /// This does not mean that it is profitable to vectorize this
1165 /// loop, only that it is legal to do so.
1166 bool canVectorize();
1168 /// Returns the Induction variable.
1169 PHINode *getInduction() { return Induction; }
1171 /// Returns the reduction variables found in the loop.
1172 ReductionList *getReductionVars() { return &Reductions; }
1174 /// Returns the induction variables found in the loop.
1175 InductionList *getInductionVars() { return &Inductions; }
1177 /// Returns the widest induction type.
1178 Type *getWidestInductionType() { return WidestIndTy; }
1180 /// Returns True if V is an induction variable in this loop.
1181 bool isInductionVariable(const Value *V);
1183 /// Return true if the block BB needs to be predicated in order for the loop
1184 /// to be vectorized.
1185 bool blockNeedsPredication(BasicBlock *BB);
1187 /// Check if this pointer is consecutive when vectorizing. This happens
1188 /// when the last index of the GEP is the induction variable, or that the
1189 /// pointer itself is an induction variable.
1190 /// This check allows us to vectorize A[idx] into a wide load/store.
1192 /// 0 - Stride is unknown or non-consecutive.
1193 /// 1 - Address is consecutive.
1194 /// -1 - Address is consecutive, and decreasing.
1195 int isConsecutivePtr(Value *Ptr);
1197 /// Returns true if the value V is uniform within the loop.
1198 bool isUniform(Value *V);
1200 /// Returns true if this instruction will remain scalar after vectorization.
1201 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1203 /// Returns the information that we collected about runtime memory check.
1204 const RuntimePointerChecking *getRuntimePointerChecking() const {
1205 return LAI->getRuntimePointerChecking();
1208 const LoopAccessInfo *getLAI() const {
1212 /// \brief Check if \p Instr belongs to any interleaved access group.
1213 bool isAccessInterleaved(Instruction *Instr) {
1214 return InterleaveInfo.isInterleaved(Instr);
1217 /// \brief Get the interleaved access group that \p Instr belongs to.
1218 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1219 return InterleaveInfo.getInterleaveGroup(Instr);
1222 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1224 bool hasStride(Value *V) { return StrideSet.count(V); }
1225 bool mustCheckStrides() { return !StrideSet.empty(); }
1226 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1227 return StrideSet.begin();
1229 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1231 /// Returns true if the target machine supports masked store operation
1232 /// for the given \p DataType and kind of access to \p Ptr.
1233 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1234 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1236 /// Returns true if the target machine supports masked load operation
1237 /// for the given \p DataType and kind of access to \p Ptr.
1238 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1239 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1241 /// Returns true if vector representation of the instruction \p I
1243 bool isMaskRequired(const Instruction* I) {
1244 return (MaskedOp.count(I) != 0);
1246 unsigned getNumStores() const {
1247 return LAI->getNumStores();
1249 unsigned getNumLoads() const {
1250 return LAI->getNumLoads();
1252 unsigned getNumPredStores() const {
1253 return NumPredStores;
1256 /// Check if a single basic block loop is vectorizable.
1257 /// At this point we know that this is a loop with a constant trip count
1258 /// and we only need to check individual instructions.
1259 bool canVectorizeInstrs();
1261 /// When we vectorize loops we may change the order in which
1262 /// we read and write from memory. This method checks if it is
1263 /// legal to vectorize the code, considering only memory constrains.
1264 /// Returns true if the loop is vectorizable
1265 bool canVectorizeMemory();
1267 /// Return true if we can vectorize this loop using the IF-conversion
1269 bool canVectorizeWithIfConvert();
1271 /// Collect the variables that need to stay uniform after vectorization.
1272 void collectLoopUniforms();
1274 /// Return true if all of the instructions in the block can be speculatively
1275 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1276 /// and we know that we can read from them without segfault.
1277 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1279 /// \brief Collect memory access with loop invariant strides.
1281 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1283 void collectStridedAccess(Value *LoadOrStoreInst);
1285 /// Report an analysis message to assist the user in diagnosing loops that are
1286 /// not vectorized. These are handled as LoopAccessReport rather than
1287 /// VectorizationReport because the << operator of VectorizationReport returns
1288 /// LoopAccessReport.
1289 void emitAnalysis(const LoopAccessReport &Message) const {
1290 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1293 unsigned NumPredStores;
1295 /// The loop that we evaluate.
1298 ScalarEvolution *SE;
1299 /// Target Library Info.
1300 TargetLibraryInfo *TLI;
1302 Function *TheFunction;
1303 /// Target Transform Info
1304 const TargetTransformInfo *TTI;
1307 // LoopAccess analysis.
1308 LoopAccessAnalysis *LAA;
1309 // And the loop-accesses info corresponding to this loop. This pointer is
1310 // null until canVectorizeMemory sets it up.
1311 const LoopAccessInfo *LAI;
1313 /// The interleave access information contains groups of interleaved accesses
1314 /// with the same stride and close to each other.
1315 InterleavedAccessInfo InterleaveInfo;
1317 // --- vectorization state --- //
1319 /// Holds the integer induction variable. This is the counter of the
1322 /// Holds the reduction variables.
1323 ReductionList Reductions;
1324 /// Holds all of the induction variables that we found in the loop.
1325 /// Notice that inductions don't need to start at zero and that induction
1326 /// variables can be pointers.
1327 InductionList Inductions;
1328 /// Holds the widest induction type encountered.
1331 /// Allowed outside users. This holds the reduction
1332 /// vars which can be accessed from outside the loop.
1333 SmallPtrSet<Value*, 4> AllowedExit;
1334 /// This set holds the variables which are known to be uniform after
1336 SmallPtrSet<Instruction*, 4> Uniforms;
1338 /// Can we assume the absence of NaNs.
1339 bool HasFunNoNaNAttr;
1341 /// Vectorization requirements that will go through late-evaluation.
1342 LoopVectorizationRequirements *Requirements;
1344 /// Used to emit an analysis of any legality issues.
1345 const LoopVectorizeHints *Hints;
1347 ValueToValueMap Strides;
1348 SmallPtrSet<Value *, 8> StrideSet;
1350 /// While vectorizing these instructions we have to generate a
1351 /// call to the appropriate masked intrinsic
1352 SmallPtrSet<const Instruction*, 8> MaskedOp;
1355 /// LoopVectorizationCostModel - estimates the expected speedups due to
1357 /// In many cases vectorization is not profitable. This can happen because of
1358 /// a number of reasons. In this class we mainly attempt to predict the
1359 /// expected speedup/slowdowns due to the supported instruction set. We use the
1360 /// TargetTransformInfo to query the different backends for the cost of
1361 /// different operations.
1362 class LoopVectorizationCostModel {
1364 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1365 LoopVectorizationLegality *Legal,
1366 const TargetTransformInfo &TTI,
1367 const TargetLibraryInfo *TLI, DemandedBits *DB,
1368 AssumptionCache *AC,
1369 const Function *F, const LoopVectorizeHints *Hints,
1370 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1371 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1372 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1374 /// Information about vectorization costs
1375 struct VectorizationFactor {
1376 unsigned Width; // Vector width with best cost
1377 unsigned Cost; // Cost of the loop with that width
1379 /// \return The most profitable vectorization factor and the cost of that VF.
1380 /// This method checks every power of two up to VF. If UserVF is not ZERO
1381 /// then this vectorization factor will be selected if vectorization is
1383 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1385 /// \return The size (in bits) of the smallest and widest types in the code
1386 /// that needs to be vectorized. We ignore values that remain scalar such as
1387 /// 64 bit loop indices.
1388 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1390 /// \return The desired interleave count.
1391 /// If interleave count has been specified by metadata it will be returned.
1392 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1393 /// are the selected vectorization factor and the cost of the selected VF.
1394 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1397 /// \return The most profitable unroll factor.
1398 /// This method finds the best unroll-factor based on register pressure and
1399 /// other parameters. VF and LoopCost are the selected vectorization factor
1400 /// and the cost of the selected VF.
1401 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1404 /// \brief A struct that represents some properties of the register usage
1406 struct RegisterUsage {
1407 /// Holds the number of loop invariant values that are used in the loop.
1408 unsigned LoopInvariantRegs;
1409 /// Holds the maximum number of concurrent live intervals in the loop.
1410 unsigned MaxLocalUsers;
1411 /// Holds the number of instructions in the loop.
1412 unsigned NumInstructions;
1415 /// \return Returns information about the register usages of the loop for the
1416 /// given vectorization factors.
1417 SmallVector<RegisterUsage, 8>
1418 calculateRegisterUsage(const SmallVector<unsigned, 8> &VFs);
1421 /// Returns the expected execution cost. The unit of the cost does
1422 /// not matter because we use the 'cost' units to compare different
1423 /// vector widths. The cost that is returned is *not* normalized by
1424 /// the factor width.
1425 unsigned expectedCost(unsigned VF);
1427 /// Returns the execution time cost of an instruction for a given vector
1428 /// width. Vector width of one means scalar.
1429 unsigned getInstructionCost(Instruction *I, unsigned VF);
1431 /// Returns whether the instruction is a load or store and will be a emitted
1432 /// as a vector operation.
1433 bool isConsecutiveLoadOrStore(Instruction *I);
1435 /// Report an analysis message to assist the user in diagnosing loops that are
1436 /// not vectorized. These are handled as LoopAccessReport rather than
1437 /// VectorizationReport because the << operator of VectorizationReport returns
1438 /// LoopAccessReport.
1439 void emitAnalysis(const LoopAccessReport &Message) const {
1440 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1444 /// Map of scalar integer values to the smallest bitwidth they can be legally
1445 /// represented as. The vector equivalents of these values should be truncated
1447 DenseMap<Instruction*,uint64_t> MinBWs;
1449 /// The loop that we evaluate.
1452 ScalarEvolution *SE;
1453 /// Loop Info analysis.
1455 /// Vectorization legality.
1456 LoopVectorizationLegality *Legal;
1457 /// Vector target information.
1458 const TargetTransformInfo &TTI;
1459 /// Target Library Info.
1460 const TargetLibraryInfo *TLI;
1461 /// Demanded bits analysis
1463 const Function *TheFunction;
1464 // Loop Vectorize Hint.
1465 const LoopVectorizeHints *Hints;
1466 // Values to ignore in the cost model.
1467 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1470 /// \brief This holds vectorization requirements that must be verified late in
1471 /// the process. The requirements are set by legalize and costmodel. Once
1472 /// vectorization has been determined to be possible and profitable the
1473 /// requirements can be verified by looking for metadata or compiler options.
1474 /// For example, some loops require FP commutativity which is only allowed if
1475 /// vectorization is explicitly specified or if the fast-math compiler option
1476 /// has been provided.
1477 /// Late evaluation of these requirements allows helpful diagnostics to be
1478 /// composed that tells the user what need to be done to vectorize the loop. For
1479 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1480 /// evaluation should be used only when diagnostics can generated that can be
1481 /// followed by a non-expert user.
1482 class LoopVectorizationRequirements {
1484 LoopVectorizationRequirements()
1485 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1487 void addUnsafeAlgebraInst(Instruction *I) {
1488 // First unsafe algebra instruction.
1489 if (!UnsafeAlgebraInst)
1490 UnsafeAlgebraInst = I;
1493 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1495 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1496 const char *Name = Hints.vectorizeAnalysisPassName();
1497 bool Failed = false;
1498 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1499 emitOptimizationRemarkAnalysisFPCommute(
1500 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1501 VectorizationReport() << "cannot prove it is safe to reorder "
1502 "floating-point operations");
1506 // Test if runtime memcheck thresholds are exceeded.
1507 bool PragmaThresholdReached =
1508 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1509 bool ThresholdReached =
1510 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1511 if ((ThresholdReached && !Hints.allowReordering()) ||
1512 PragmaThresholdReached) {
1513 emitOptimizationRemarkAnalysisAliasing(
1514 F->getContext(), Name, *F, L->getStartLoc(),
1515 VectorizationReport()
1516 << "cannot prove it is safe to reorder memory operations");
1517 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1525 unsigned NumRuntimePointerChecks;
1526 Instruction *UnsafeAlgebraInst;
1529 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1531 return V.push_back(&L);
1533 for (Loop *InnerL : L)
1534 addInnerLoop(*InnerL, V);
1537 /// The LoopVectorize Pass.
1538 struct LoopVectorize : public FunctionPass {
1539 /// Pass identification, replacement for typeid
1542 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1544 DisableUnrolling(NoUnrolling),
1545 AlwaysVectorize(AlwaysVectorize) {
1546 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1549 ScalarEvolution *SE;
1551 TargetTransformInfo *TTI;
1553 BlockFrequencyInfo *BFI;
1554 TargetLibraryInfo *TLI;
1557 AssumptionCache *AC;
1558 LoopAccessAnalysis *LAA;
1559 bool DisableUnrolling;
1560 bool AlwaysVectorize;
1562 BlockFrequency ColdEntryFreq;
1564 bool runOnFunction(Function &F) override {
1565 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1566 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1567 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1568 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1569 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1570 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1571 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1572 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1573 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1574 LAA = &getAnalysis<LoopAccessAnalysis>();
1575 DB = &getAnalysis<DemandedBits>();
1577 // Compute some weights outside of the loop over the loops. Compute this
1578 // using a BranchProbability to re-use its scaling math.
1579 const BranchProbability ColdProb(1, 5); // 20%
1580 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1583 // 1. the target claims to have no vector registers, and
1584 // 2. interleaving won't help ILP.
1586 // The second condition is necessary because, even if the target has no
1587 // vector registers, loop vectorization may still enable scalar
1589 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1592 // Build up a worklist of inner-loops to vectorize. This is necessary as
1593 // the act of vectorizing or partially unrolling a loop creates new loops
1594 // and can invalidate iterators across the loops.
1595 SmallVector<Loop *, 8> Worklist;
1598 addInnerLoop(*L, Worklist);
1600 LoopsAnalyzed += Worklist.size();
1602 // Now walk the identified inner loops.
1603 bool Changed = false;
1604 while (!Worklist.empty())
1605 Changed |= processLoop(Worklist.pop_back_val());
1607 // Process each loop nest in the function.
1611 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1612 SmallVector<Metadata *, 4> MDs;
1613 // Reserve first location for self reference to the LoopID metadata node.
1614 MDs.push_back(nullptr);
1615 bool IsUnrollMetadata = false;
1616 MDNode *LoopID = L->getLoopID();
1618 // First find existing loop unrolling disable metadata.
1619 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1620 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1622 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1624 S && S->getString().startswith("llvm.loop.unroll.disable");
1626 MDs.push_back(LoopID->getOperand(i));
1630 if (!IsUnrollMetadata) {
1631 // Add runtime unroll disable metadata.
1632 LLVMContext &Context = L->getHeader()->getContext();
1633 SmallVector<Metadata *, 1> DisableOperands;
1634 DisableOperands.push_back(
1635 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1636 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1637 MDs.push_back(DisableNode);
1638 MDNode *NewLoopID = MDNode::get(Context, MDs);
1639 // Set operand 0 to refer to the loop id itself.
1640 NewLoopID->replaceOperandWith(0, NewLoopID);
1641 L->setLoopID(NewLoopID);
1645 bool processLoop(Loop *L) {
1646 assert(L->empty() && "Only process inner loops.");
1649 const std::string DebugLocStr = getDebugLocString(L);
1652 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1653 << L->getHeader()->getParent()->getName() << "\" from "
1654 << DebugLocStr << "\n");
1656 LoopVectorizeHints Hints(L, DisableUnrolling);
1658 DEBUG(dbgs() << "LV: Loop hints:"
1660 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1662 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1664 : "?")) << " width=" << Hints.getWidth()
1665 << " unroll=" << Hints.getInterleave() << "\n");
1667 // Function containing loop
1668 Function *F = L->getHeader()->getParent();
1670 // Looking at the diagnostic output is the only way to determine if a loop
1671 // was vectorized (other than looking at the IR or machine code), so it
1672 // is important to generate an optimization remark for each loop. Most of
1673 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1674 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1675 // less verbose reporting vectorized loops and unvectorized loops that may
1676 // benefit from vectorization, respectively.
1678 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1679 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1683 // Check the loop for a trip count threshold:
1684 // do not vectorize loops with a tiny trip count.
1685 const unsigned TC = SE->getSmallConstantTripCount(L);
1686 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1687 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1688 << "This loop is not worth vectorizing.");
1689 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1690 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1692 DEBUG(dbgs() << "\n");
1693 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1694 << "vectorization is not beneficial "
1695 "and is not explicitly forced");
1700 // Check if it is legal to vectorize the loop.
1701 LoopVectorizationRequirements Requirements;
1702 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1703 &Requirements, &Hints);
1704 if (!LVL.canVectorize()) {
1705 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1706 emitMissedWarning(F, L, Hints);
1710 // Collect values we want to ignore in the cost model. This includes
1711 // type-promoting instructions we identified during reduction detection.
1712 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1713 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1714 for (auto &Reduction : *LVL.getReductionVars()) {
1715 RecurrenceDescriptor &RedDes = Reduction.second;
1716 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1717 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1720 // Use the cost model.
1721 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, DB, AC, F, &Hints,
1724 // Check the function attributes to find out if this function should be
1725 // optimized for size.
1726 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1729 // Compute the weighted frequency of this loop being executed and see if it
1730 // is less than 20% of the function entry baseline frequency. Note that we
1731 // always have a canonical loop here because we think we *can* vectorize.
1732 // FIXME: This is hidden behind a flag due to pervasive problems with
1733 // exactly what block frequency models.
1734 if (LoopVectorizeWithBlockFrequency) {
1735 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1736 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1737 LoopEntryFreq < ColdEntryFreq)
1741 // Check the function attributes to see if implicit floats are allowed.
1742 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1743 // an integer loop and the vector instructions selected are purely integer
1744 // vector instructions?
1745 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1746 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1747 "attribute is used.\n");
1750 VectorizationReport()
1751 << "loop not vectorized due to NoImplicitFloat attribute");
1752 emitMissedWarning(F, L, Hints);
1756 // Select the optimal vectorization factor.
1757 const LoopVectorizationCostModel::VectorizationFactor VF =
1758 CM.selectVectorizationFactor(OptForSize);
1760 // Select the interleave count.
1761 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1763 // Get user interleave count.
1764 unsigned UserIC = Hints.getInterleave();
1766 // Identify the diagnostic messages that should be produced.
1767 std::string VecDiagMsg, IntDiagMsg;
1768 bool VectorizeLoop = true, InterleaveLoop = true;
1770 if (Requirements.doesNotMeet(F, L, Hints)) {
1771 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1773 emitMissedWarning(F, L, Hints);
1777 if (VF.Width == 1) {
1778 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1780 "the cost-model indicates that vectorization is not beneficial";
1781 VectorizeLoop = false;
1784 if (IC == 1 && UserIC <= 1) {
1785 // Tell the user interleaving is not beneficial.
1786 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1788 "the cost-model indicates that interleaving is not beneficial";
1789 InterleaveLoop = false;
1792 " and is explicitly disabled or interleave count is set to 1";
1793 } else if (IC > 1 && UserIC == 1) {
1794 // Tell the user interleaving is beneficial, but it explicitly disabled.
1796 << "LV: Interleaving is beneficial but is explicitly disabled.");
1797 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1798 "but is explicitly disabled or interleave count is set to 1";
1799 InterleaveLoop = false;
1802 // Override IC if user provided an interleave count.
1803 IC = UserIC > 0 ? UserIC : IC;
1805 // Emit diagnostic messages, if any.
1806 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1807 if (!VectorizeLoop && !InterleaveLoop) {
1808 // Do not vectorize or interleaving the loop.
1809 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1810 L->getStartLoc(), VecDiagMsg);
1811 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1812 L->getStartLoc(), IntDiagMsg);
1814 } else if (!VectorizeLoop && InterleaveLoop) {
1815 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1816 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1817 L->getStartLoc(), VecDiagMsg);
1818 } else if (VectorizeLoop && !InterleaveLoop) {
1819 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1820 << DebugLocStr << '\n');
1821 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1822 L->getStartLoc(), IntDiagMsg);
1823 } else if (VectorizeLoop && InterleaveLoop) {
1824 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1825 << DebugLocStr << '\n');
1826 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1829 if (!VectorizeLoop) {
1830 assert(IC > 1 && "interleave count should not be 1 or 0");
1831 // If we decided that it is not legal to vectorize the loop then
1833 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1834 Unroller.vectorize(&LVL, CM.MinBWs);
1836 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1837 Twine("interleaved loop (interleaved count: ") +
1840 // If we decided that it is *legal* to vectorize the loop then do it.
1841 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1842 LB.vectorize(&LVL, CM.MinBWs);
1845 // Add metadata to disable runtime unrolling scalar loop when there's no
1846 // runtime check about strides and memory. Because at this situation,
1847 // scalar loop is rarely used not worthy to be unrolled.
1848 if (!LB.IsSafetyChecksAdded())
1849 AddRuntimeUnrollDisableMetaData(L);
1851 // Report the vectorization decision.
1852 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1853 Twine("vectorized loop (vectorization width: ") +
1854 Twine(VF.Width) + ", interleaved count: " +
1858 // Mark the loop as already vectorized to avoid vectorizing again.
1859 Hints.setAlreadyVectorized();
1861 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1865 void getAnalysisUsage(AnalysisUsage &AU) const override {
1866 AU.addRequired<AssumptionCacheTracker>();
1867 AU.addRequiredID(LoopSimplifyID);
1868 AU.addRequiredID(LCSSAID);
1869 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1870 AU.addRequired<DominatorTreeWrapperPass>();
1871 AU.addRequired<LoopInfoWrapperPass>();
1872 AU.addRequired<ScalarEvolutionWrapperPass>();
1873 AU.addRequired<TargetTransformInfoWrapperPass>();
1874 AU.addRequired<AAResultsWrapperPass>();
1875 AU.addRequired<LoopAccessAnalysis>();
1876 AU.addRequired<DemandedBits>();
1877 AU.addPreserved<LoopInfoWrapperPass>();
1878 AU.addPreserved<DominatorTreeWrapperPass>();
1879 AU.addPreserved<BasicAAWrapperPass>();
1880 AU.addPreserved<AAResultsWrapperPass>();
1881 AU.addPreserved<GlobalsAAWrapperPass>();
1886 } // end anonymous namespace
1888 //===----------------------------------------------------------------------===//
1889 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1890 // LoopVectorizationCostModel.
1891 //===----------------------------------------------------------------------===//
1893 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1894 // We need to place the broadcast of invariant variables outside the loop.
1895 Instruction *Instr = dyn_cast<Instruction>(V);
1897 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1898 Instr->getParent()) != LoopVectorBody.end());
1899 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1901 // Place the code for broadcasting invariant variables in the new preheader.
1902 IRBuilder<>::InsertPointGuard Guard(Builder);
1904 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1906 // Broadcast the scalar into all locations in the vector.
1907 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1912 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1914 assert(Val->getType()->isVectorTy() && "Must be a vector");
1915 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1916 "Elem must be an integer");
1917 assert(Step->getType() == Val->getType()->getScalarType() &&
1918 "Step has wrong type");
1919 // Create the types.
1920 Type *ITy = Val->getType()->getScalarType();
1921 VectorType *Ty = cast<VectorType>(Val->getType());
1922 int VLen = Ty->getNumElements();
1923 SmallVector<Constant*, 8> Indices;
1925 // Create a vector of consecutive numbers from zero to VF.
1926 for (int i = 0; i < VLen; ++i)
1927 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1929 // Add the consecutive indices to the vector value.
1930 Constant *Cv = ConstantVector::get(Indices);
1931 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1932 Step = Builder.CreateVectorSplat(VLen, Step);
1933 assert(Step->getType() == Val->getType() && "Invalid step vec");
1934 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1935 // which can be found from the original scalar operations.
1936 Step = Builder.CreateMul(Cv, Step);
1937 return Builder.CreateAdd(Val, Step, "induction");
1940 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1941 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1942 // Make sure that the pointer does not point to structs.
1943 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1946 // If this value is a pointer induction variable we know it is consecutive.
1947 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1948 if (Phi && Inductions.count(Phi)) {
1949 InductionDescriptor II = Inductions[Phi];
1950 return II.getConsecutiveDirection();
1953 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1957 unsigned NumOperands = Gep->getNumOperands();
1958 Value *GpPtr = Gep->getPointerOperand();
1959 // If this GEP value is a consecutive pointer induction variable and all of
1960 // the indices are constant then we know it is consecutive. We can
1961 Phi = dyn_cast<PHINode>(GpPtr);
1962 if (Phi && Inductions.count(Phi)) {
1964 // Make sure that the pointer does not point to structs.
1965 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1966 if (GepPtrType->getElementType()->isAggregateType())
1969 // Make sure that all of the index operands are loop invariant.
1970 for (unsigned i = 1; i < NumOperands; ++i)
1971 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1974 InductionDescriptor II = Inductions[Phi];
1975 return II.getConsecutiveDirection();
1978 unsigned InductionOperand = getGEPInductionOperand(Gep);
1980 // Check that all of the gep indices are uniform except for our induction
1982 for (unsigned i = 0; i != NumOperands; ++i)
1983 if (i != InductionOperand &&
1984 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1987 // We can emit wide load/stores only if the last non-zero index is the
1988 // induction variable.
1989 const SCEV *Last = nullptr;
1990 if (!Strides.count(Gep))
1991 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1993 // Because of the multiplication by a stride we can have a s/zext cast.
1994 // We are going to replace this stride by 1 so the cast is safe to ignore.
1996 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1997 // %0 = trunc i64 %indvars.iv to i32
1998 // %mul = mul i32 %0, %Stride1
1999 // %idxprom = zext i32 %mul to i64 << Safe cast.
2000 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
2002 Last = replaceSymbolicStrideSCEV(SE, Strides,
2003 Gep->getOperand(InductionOperand), Gep);
2004 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
2006 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2010 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2011 const SCEV *Step = AR->getStepRecurrence(*SE);
2013 // The memory is consecutive because the last index is consecutive
2014 // and all other indices are loop invariant.
2017 if (Step->isAllOnesValue())
2024 bool LoopVectorizationLegality::isUniform(Value *V) {
2025 return LAI->isUniform(V);
2028 InnerLoopVectorizer::VectorParts&
2029 InnerLoopVectorizer::getVectorValue(Value *V) {
2030 assert(V != Induction && "The new induction variable should not be used.");
2031 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2033 // If we have a stride that is replaced by one, do it here.
2034 if (Legal->hasStride(V))
2035 V = ConstantInt::get(V->getType(), 1);
2037 // If we have this scalar in the map, return it.
2038 if (WidenMap.has(V))
2039 return WidenMap.get(V);
2041 // If this scalar is unknown, assume that it is a constant or that it is
2042 // loop invariant. Broadcast V and save the value for future uses.
2043 Value *B = getBroadcastInstrs(V);
2045 return WidenMap.splat(V, B);
2048 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2049 assert(Vec->getType()->isVectorTy() && "Invalid type");
2050 SmallVector<Constant*, 8> ShuffleMask;
2051 for (unsigned i = 0; i < VF; ++i)
2052 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2054 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2055 ConstantVector::get(ShuffleMask),
2059 // Get a mask to interleave \p NumVec vectors into a wide vector.
2060 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2061 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2062 // <0, 4, 1, 5, 2, 6, 3, 7>
2063 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2065 SmallVector<Constant *, 16> Mask;
2066 for (unsigned i = 0; i < VF; i++)
2067 for (unsigned j = 0; j < NumVec; j++)
2068 Mask.push_back(Builder.getInt32(j * VF + i));
2070 return ConstantVector::get(Mask);
2073 // Get the strided mask starting from index \p Start.
2074 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2075 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2076 unsigned Stride, unsigned VF) {
2077 SmallVector<Constant *, 16> Mask;
2078 for (unsigned i = 0; i < VF; i++)
2079 Mask.push_back(Builder.getInt32(Start + i * Stride));
2081 return ConstantVector::get(Mask);
2084 // Get a mask of two parts: The first part consists of sequential integers
2085 // starting from 0, The second part consists of UNDEFs.
2086 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2087 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2088 unsigned NumUndef) {
2089 SmallVector<Constant *, 16> Mask;
2090 for (unsigned i = 0; i < NumInt; i++)
2091 Mask.push_back(Builder.getInt32(i));
2093 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2094 for (unsigned i = 0; i < NumUndef; i++)
2095 Mask.push_back(Undef);
2097 return ConstantVector::get(Mask);
2100 // Concatenate two vectors with the same element type. The 2nd vector should
2101 // not have more elements than the 1st vector. If the 2nd vector has less
2102 // elements, extend it with UNDEFs.
2103 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2105 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2106 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2107 assert(VecTy1 && VecTy2 &&
2108 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2109 "Expect two vectors with the same element type");
2111 unsigned NumElts1 = VecTy1->getNumElements();
2112 unsigned NumElts2 = VecTy2->getNumElements();
2113 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2115 if (NumElts1 > NumElts2) {
2116 // Extend with UNDEFs.
2118 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2119 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2122 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2123 return Builder.CreateShuffleVector(V1, V2, Mask);
2126 // Concatenate vectors in the given list. All vectors have the same type.
2127 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2128 ArrayRef<Value *> InputList) {
2129 unsigned NumVec = InputList.size();
2130 assert(NumVec > 1 && "Should be at least two vectors");
2132 SmallVector<Value *, 8> ResList;
2133 ResList.append(InputList.begin(), InputList.end());
2135 SmallVector<Value *, 8> TmpList;
2136 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2137 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2138 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2139 "Only the last vector may have a different type");
2141 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2144 // Push the last vector if the total number of vectors is odd.
2145 if (NumVec % 2 != 0)
2146 TmpList.push_back(ResList[NumVec - 1]);
2149 NumVec = ResList.size();
2150 } while (NumVec > 1);
2155 // Try to vectorize the interleave group that \p Instr belongs to.
2157 // E.g. Translate following interleaved load group (factor = 3):
2158 // for (i = 0; i < N; i+=3) {
2159 // R = Pic[i]; // Member of index 0
2160 // G = Pic[i+1]; // Member of index 1
2161 // B = Pic[i+2]; // Member of index 2
2162 // ... // do something to R, G, B
2165 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2166 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2167 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2168 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2170 // Or translate following interleaved store group (factor = 3):
2171 // for (i = 0; i < N; i+=3) {
2172 // ... do something to R, G, B
2173 // Pic[i] = R; // Member of index 0
2174 // Pic[i+1] = G; // Member of index 1
2175 // Pic[i+2] = B; // Member of index 2
2178 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2179 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2180 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2181 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2182 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2183 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2184 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2185 assert(Group && "Fail to get an interleaved access group.");
2187 // Skip if current instruction is not the insert position.
2188 if (Instr != Group->getInsertPos())
2191 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2192 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2193 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2195 // Prepare for the vector type of the interleaved load/store.
2196 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2197 unsigned InterleaveFactor = Group->getFactor();
2198 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2199 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2201 // Prepare for the new pointers.
2202 setDebugLocFromInst(Builder, Ptr);
2203 VectorParts &PtrParts = getVectorValue(Ptr);
2204 SmallVector<Value *, 2> NewPtrs;
2205 unsigned Index = Group->getIndex(Instr);
2206 for (unsigned Part = 0; Part < UF; Part++) {
2207 // Extract the pointer for current instruction from the pointer vector. A
2208 // reverse access uses the pointer in the last lane.
2209 Value *NewPtr = Builder.CreateExtractElement(
2211 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2213 // Notice current instruction could be any index. Need to adjust the address
2214 // to the member of index 0.
2216 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2217 // b = A[i]; // Member of index 0
2218 // Current pointer is pointed to A[i+1], adjust it to A[i].
2220 // E.g. A[i+1] = a; // Member of index 1
2221 // A[i] = b; // Member of index 0
2222 // A[i+2] = c; // Member of index 2 (Current instruction)
2223 // Current pointer is pointed to A[i+2], adjust it to A[i].
2224 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2226 // Cast to the vector pointer type.
2227 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2230 setDebugLocFromInst(Builder, Instr);
2231 Value *UndefVec = UndefValue::get(VecTy);
2233 // Vectorize the interleaved load group.
2235 for (unsigned Part = 0; Part < UF; Part++) {
2236 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2237 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2239 for (unsigned i = 0; i < InterleaveFactor; i++) {
2240 Instruction *Member = Group->getMember(i);
2242 // Skip the gaps in the group.
2246 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2247 Value *StridedVec = Builder.CreateShuffleVector(
2248 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2250 // If this member has different type, cast the result type.
2251 if (Member->getType() != ScalarTy) {
2252 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2253 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2256 VectorParts &Entry = WidenMap.get(Member);
2258 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2261 propagateMetadata(NewLoadInstr, Instr);
2266 // The sub vector type for current instruction.
2267 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2269 // Vectorize the interleaved store group.
2270 for (unsigned Part = 0; Part < UF; Part++) {
2271 // Collect the stored vector from each member.
2272 SmallVector<Value *, 4> StoredVecs;
2273 for (unsigned i = 0; i < InterleaveFactor; i++) {
2274 // Interleaved store group doesn't allow a gap, so each index has a member
2275 Instruction *Member = Group->getMember(i);
2276 assert(Member && "Fail to get a member from an interleaved store group");
2279 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2280 if (Group->isReverse())
2281 StoredVec = reverseVector(StoredVec);
2283 // If this member has different type, cast it to an unified type.
2284 if (StoredVec->getType() != SubVT)
2285 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2287 StoredVecs.push_back(StoredVec);
2290 // Concatenate all vectors into a wide vector.
2291 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2293 // Interleave the elements in the wide vector.
2294 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2295 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2298 Instruction *NewStoreInstr =
2299 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2300 propagateMetadata(NewStoreInstr, Instr);
2304 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2305 // Attempt to issue a wide load.
2306 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2307 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2309 assert((LI || SI) && "Invalid Load/Store instruction");
2311 // Try to vectorize the interleave group if this access is interleaved.
2312 if (Legal->isAccessInterleaved(Instr))
2313 return vectorizeInterleaveGroup(Instr);
2315 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2316 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2317 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2318 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2319 // An alignment of 0 means target abi alignment. We need to use the scalar's
2320 // target abi alignment in such a case.
2321 const DataLayout &DL = Instr->getModule()->getDataLayout();
2323 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2324 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2325 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2326 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2328 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2329 !Legal->isMaskRequired(SI))
2330 return scalarizeInstruction(Instr, true);
2332 if (ScalarAllocatedSize != VectorElementSize)
2333 return scalarizeInstruction(Instr);
2335 // If the pointer is loop invariant or if it is non-consecutive,
2336 // scalarize the load.
2337 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2338 bool Reverse = ConsecutiveStride < 0;
2339 bool UniformLoad = LI && Legal->isUniform(Ptr);
2340 if (!ConsecutiveStride || UniformLoad)
2341 return scalarizeInstruction(Instr);
2343 Constant *Zero = Builder.getInt32(0);
2344 VectorParts &Entry = WidenMap.get(Instr);
2346 // Handle consecutive loads/stores.
2347 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2348 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2349 setDebugLocFromInst(Builder, Gep);
2350 Value *PtrOperand = Gep->getPointerOperand();
2351 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2352 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2354 // Create the new GEP with the new induction variable.
2355 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2356 Gep2->setOperand(0, FirstBasePtr);
2357 Gep2->setName("gep.indvar.base");
2358 Ptr = Builder.Insert(Gep2);
2360 setDebugLocFromInst(Builder, Gep);
2361 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2362 OrigLoop) && "Base ptr must be invariant");
2364 // The last index does not have to be the induction. It can be
2365 // consecutive and be a function of the index. For example A[I+1];
2366 unsigned NumOperands = Gep->getNumOperands();
2367 unsigned InductionOperand = getGEPInductionOperand(Gep);
2368 // Create the new GEP with the new induction variable.
2369 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2371 for (unsigned i = 0; i < NumOperands; ++i) {
2372 Value *GepOperand = Gep->getOperand(i);
2373 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2375 // Update last index or loop invariant instruction anchored in loop.
2376 if (i == InductionOperand ||
2377 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2378 assert((i == InductionOperand ||
2379 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2380 "Must be last index or loop invariant");
2382 VectorParts &GEPParts = getVectorValue(GepOperand);
2383 Value *Index = GEPParts[0];
2384 Index = Builder.CreateExtractElement(Index, Zero);
2385 Gep2->setOperand(i, Index);
2386 Gep2->setName("gep.indvar.idx");
2389 Ptr = Builder.Insert(Gep2);
2391 // Use the induction element ptr.
2392 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2393 setDebugLocFromInst(Builder, Ptr);
2394 VectorParts &PtrVal = getVectorValue(Ptr);
2395 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2398 VectorParts Mask = createBlockInMask(Instr->getParent());
2401 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2402 "We do not allow storing to uniform addresses");
2403 setDebugLocFromInst(Builder, SI);
2404 // We don't want to update the value in the map as it might be used in
2405 // another expression. So don't use a reference type for "StoredVal".
2406 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2408 for (unsigned Part = 0; Part < UF; ++Part) {
2409 // Calculate the pointer for the specific unroll-part.
2411 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2414 // If we store to reverse consecutive memory locations, then we need
2415 // to reverse the order of elements in the stored value.
2416 StoredVal[Part] = reverseVector(StoredVal[Part]);
2417 // If the address is consecutive but reversed, then the
2418 // wide store needs to start at the last vector element.
2419 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2420 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2421 Mask[Part] = reverseVector(Mask[Part]);
2424 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2425 DataTy->getPointerTo(AddressSpace));
2428 if (Legal->isMaskRequired(SI))
2429 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2432 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2433 propagateMetadata(NewSI, SI);
2439 assert(LI && "Must have a load instruction");
2440 setDebugLocFromInst(Builder, LI);
2441 for (unsigned Part = 0; Part < UF; ++Part) {
2442 // Calculate the pointer for the specific unroll-part.
2444 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2447 // If the address is consecutive but reversed, then the
2448 // wide load needs to start at the last vector element.
2449 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2450 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2451 Mask[Part] = reverseVector(Mask[Part]);
2455 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2456 DataTy->getPointerTo(AddressSpace));
2457 if (Legal->isMaskRequired(LI))
2458 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2459 UndefValue::get(DataTy),
2460 "wide.masked.load");
2462 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2463 propagateMetadata(NewLI, LI);
2464 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2468 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2469 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2470 // Holds vector parameters or scalars, in case of uniform vals.
2471 SmallVector<VectorParts, 4> Params;
2473 setDebugLocFromInst(Builder, Instr);
2475 // Find all of the vectorized parameters.
2476 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2477 Value *SrcOp = Instr->getOperand(op);
2479 // If we are accessing the old induction variable, use the new one.
2480 if (SrcOp == OldInduction) {
2481 Params.push_back(getVectorValue(SrcOp));
2485 // Try using previously calculated values.
2486 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2488 // If the src is an instruction that appeared earlier in the basic block,
2489 // then it should already be vectorized.
2490 if (SrcInst && OrigLoop->contains(SrcInst)) {
2491 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2492 // The parameter is a vector value from earlier.
2493 Params.push_back(WidenMap.get(SrcInst));
2495 // The parameter is a scalar from outside the loop. Maybe even a constant.
2496 VectorParts Scalars;
2497 Scalars.append(UF, SrcOp);
2498 Params.push_back(Scalars);
2502 assert(Params.size() == Instr->getNumOperands() &&
2503 "Invalid number of operands");
2505 // Does this instruction return a value ?
2506 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2508 Value *UndefVec = IsVoidRetTy ? nullptr :
2509 UndefValue::get(VectorType::get(Instr->getType(), VF));
2510 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2511 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2514 if (IfPredicateStore) {
2515 assert(Instr->getParent()->getSinglePredecessor() &&
2516 "Only support single predecessor blocks");
2517 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2518 Instr->getParent());
2521 // For each vector unroll 'part':
2522 for (unsigned Part = 0; Part < UF; ++Part) {
2523 // For each scalar that we create:
2524 for (unsigned Width = 0; Width < VF; ++Width) {
2527 Value *Cmp = nullptr;
2528 if (IfPredicateStore) {
2529 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2530 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2533 Instruction *Cloned = Instr->clone();
2535 Cloned->setName(Instr->getName() + ".cloned");
2536 // Replace the operands of the cloned instructions with extracted scalars.
2537 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2538 Value *Op = Params[op][Part];
2539 // Param is a vector. Need to extract the right lane.
2540 if (Op->getType()->isVectorTy())
2541 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2542 Cloned->setOperand(op, Op);
2545 // Place the cloned scalar in the new loop.
2546 Builder.Insert(Cloned);
2548 // If the original scalar returns a value we need to place it in a vector
2549 // so that future users will be able to use it.
2551 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2552 Builder.getInt32(Width));
2554 if (IfPredicateStore)
2555 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
2561 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2565 if (Instruction *I = dyn_cast<Instruction>(V))
2566 return I->getParent() == Loc->getParent() ? I : nullptr;
2570 std::pair<Instruction *, Instruction *>
2571 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2572 Instruction *tnullptr = nullptr;
2573 if (!Legal->mustCheckStrides())
2574 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2576 IRBuilder<> ChkBuilder(Loc);
2579 Value *Check = nullptr;
2580 Instruction *FirstInst = nullptr;
2581 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2582 SE = Legal->strides_end();
2584 Value *Ptr = stripIntegerCast(*SI);
2585 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2587 // Store the first instruction we create.
2588 FirstInst = getFirstInst(FirstInst, C, Loc);
2590 Check = ChkBuilder.CreateOr(Check, C);
2595 // We have to do this trickery because the IRBuilder might fold the check to a
2596 // constant expression in which case there is no Instruction anchored in a
2598 LLVMContext &Ctx = Loc->getContext();
2599 Instruction *TheCheck =
2600 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2601 ChkBuilder.Insert(TheCheck, "stride.not.one");
2602 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2604 return std::make_pair(FirstInst, TheCheck);
2607 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L,
2612 BasicBlock *Header = L->getHeader();
2613 BasicBlock *Latch = L->getLoopLatch();
2614 // As we're just creating this loop, it's possible no latch exists
2615 // yet. If so, use the header as this will be a single block loop.
2619 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
2620 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2621 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2623 Builder.SetInsertPoint(Latch->getTerminator());
2625 // Create i+1 and fill the PHINode.
2626 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2627 Induction->addIncoming(Start, L->getLoopPreheader());
2628 Induction->addIncoming(Next, Latch);
2629 // Create the compare.
2630 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2631 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2633 // Now we have two terminators. Remove the old one from the block.
2634 Latch->getTerminator()->eraseFromParent();
2639 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2643 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2644 // Find the loop boundaries.
2645 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(OrigLoop);
2646 assert(BackedgeTakenCount != SE->getCouldNotCompute() && "Invalid loop count");
2648 Type *IdxTy = Legal->getWidestInductionType();
2650 // The exit count might have the type of i64 while the phi is i32. This can
2651 // happen if we have an induction variable that is sign extended before the
2652 // compare. The only way that we get a backedge taken count is that the
2653 // induction variable was signed and as such will not overflow. In such a case
2654 // truncation is legal.
2655 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
2656 IdxTy->getPrimitiveSizeInBits())
2657 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
2658 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
2660 // Get the total trip count from the count by adding 1.
2661 const SCEV *ExitCount = SE->getAddExpr(
2662 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
2664 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2666 // Expand the trip count and place the new instructions in the preheader.
2667 // Notice that the pre-header does not change, only the loop body.
2668 SCEVExpander Exp(*SE, DL, "induction");
2670 // Count holds the overall loop count (N).
2671 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2672 L->getLoopPreheader()->getTerminator());
2674 if (TripCount->getType()->isPointerTy())
2676 CastInst::CreatePointerCast(TripCount, IdxTy,
2677 "exitcount.ptrcnt.to.int",
2678 L->getLoopPreheader()->getTerminator());
2683 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2684 if (VectorTripCount)
2685 return VectorTripCount;
2687 Value *TC = getOrCreateTripCount(L);
2688 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2690 // Now we need to generate the expression for N - (N % VF), which is
2691 // the part that the vectorized body will execute.
2692 // The loop step is equal to the vectorization factor (num of SIMD elements)
2693 // times the unroll factor (num of SIMD instructions).
2694 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2695 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2696 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2698 return VectorTripCount;
2701 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
2702 BasicBlock *Bypass) {
2703 Value *Count = getOrCreateTripCount(L);
2704 BasicBlock *BB = L->getLoopPreheader();
2705 IRBuilder<> Builder(BB->getTerminator());
2707 // Generate code to check that the loop's trip count that we computed by
2708 // adding one to the backedge-taken count will not overflow.
2709 Value *CheckMinIters =
2710 Builder.CreateICmpULT(Count,
2711 ConstantInt::get(Count->getType(), VF * UF),
2714 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2715 "min.iters.checked");
2716 if (L->getParentLoop())
2717 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2718 ReplaceInstWithInst(BB->getTerminator(),
2719 BranchInst::Create(Bypass, NewBB, CheckMinIters));
2720 LoopBypassBlocks.push_back(BB);
2723 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
2724 BasicBlock *Bypass) {
2725 Value *TC = getOrCreateVectorTripCount(L);
2726 BasicBlock *BB = L->getLoopPreheader();
2727 IRBuilder<> Builder(BB->getTerminator());
2729 // Now, compare the new count to zero. If it is zero skip the vector loop and
2730 // jump to the scalar loop.
2731 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
2734 // Generate code to check that the loop's trip count that we computed by
2735 // adding one to the backedge-taken count will not overflow.
2736 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2738 if (L->getParentLoop())
2739 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2740 ReplaceInstWithInst(BB->getTerminator(),
2741 BranchInst::Create(Bypass, NewBB, Cmp));
2742 LoopBypassBlocks.push_back(BB);
2745 void InnerLoopVectorizer::emitStrideChecks(Loop *L,
2746 BasicBlock *Bypass) {
2747 BasicBlock *BB = L->getLoopPreheader();
2749 // Generate the code to check that the strides we assumed to be one are really
2750 // one. We want the new basic block to start at the first instruction in a
2751 // sequence of instructions that form a check.
2752 Instruction *StrideCheck;
2753 Instruction *FirstCheckInst;
2754 std::tie(FirstCheckInst, StrideCheck) = addStrideCheck(BB->getTerminator());
2758 // Create a new block containing the stride check.
2759 BB->setName("vector.stridecheck");
2760 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2761 if (L->getParentLoop())
2762 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2763 ReplaceInstWithInst(BB->getTerminator(),
2764 BranchInst::Create(Bypass, NewBB, StrideCheck));
2765 LoopBypassBlocks.push_back(BB);
2766 AddedSafetyChecks = true;
2769 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L,
2770 BasicBlock *Bypass) {
2771 BasicBlock *BB = L->getLoopPreheader();
2773 // Generate the code that checks in runtime if arrays overlap. We put the
2774 // checks into a separate block to make the more common case of few elements
2776 Instruction *FirstCheckInst;
2777 Instruction *MemRuntimeCheck;
2778 std::tie(FirstCheckInst, MemRuntimeCheck) =
2779 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
2780 if (!MemRuntimeCheck)
2783 // Create a new block containing the memory check.
2784 BB->setName("vector.memcheck");
2785 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2786 if (L->getParentLoop())
2787 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2788 ReplaceInstWithInst(BB->getTerminator(),
2789 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
2790 LoopBypassBlocks.push_back(BB);
2791 AddedSafetyChecks = true;
2795 void InnerLoopVectorizer::createEmptyLoop() {
2797 In this function we generate a new loop. The new loop will contain
2798 the vectorized instructions while the old loop will continue to run the
2801 [ ] <-- loop iteration number check.
2804 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2807 || [ ] <-- vector pre header.
2811 | [ ]_| <-- vector loop.
2814 | -[ ] <--- middle-block.
2817 -|- >[ ] <--- new preheader.
2821 | [ ]_| <-- old scalar loop to handle remainder.
2824 >[ ] <-- exit block.
2828 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2829 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2830 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2831 assert(VectorPH && "Invalid loop structure");
2832 assert(ExitBlock && "Must have an exit block");
2834 // Some loops have a single integer induction variable, while other loops
2835 // don't. One example is c++ iterators that often have multiple pointer
2836 // induction variables. In the code below we also support a case where we
2837 // don't have a single induction variable.
2839 // We try to obtain an induction variable from the original loop as hard
2840 // as possible. However if we don't find one that:
2842 // - counts from zero, stepping by one
2843 // - is the size of the widest induction variable type
2844 // then we create a new one.
2845 OldInduction = Legal->getInduction();
2846 Type *IdxTy = Legal->getWidestInductionType();
2848 // Split the single block loop into the two loop structure described above.
2849 BasicBlock *VecBody =
2850 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2851 BasicBlock *MiddleBlock =
2852 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2853 BasicBlock *ScalarPH =
2854 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2856 // Create and register the new vector loop.
2857 Loop* Lp = new Loop();
2858 Loop *ParentLoop = OrigLoop->getParentLoop();
2860 // Insert the new loop into the loop nest and register the new basic blocks
2861 // before calling any utilities such as SCEV that require valid LoopInfo.
2863 ParentLoop->addChildLoop(Lp);
2864 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2865 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2867 LI->addTopLevelLoop(Lp);
2869 Lp->addBasicBlockToLoop(VecBody, *LI);
2871 // Find the loop boundaries.
2872 Value *Count = getOrCreateTripCount(Lp);
2874 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2876 // We need to test whether the backedge-taken count is uint##_max. Adding one
2877 // to it will cause overflow and an incorrect loop trip count in the vector
2878 // body. In case of overflow we want to directly jump to the scalar remainder
2880 emitMinimumIterationCountCheck(Lp, ScalarPH);
2881 // Now, compare the new count to zero. If it is zero skip the vector loop and
2882 // jump to the scalar loop.
2883 emitVectorLoopEnteredCheck(Lp, ScalarPH);
2884 // Generate the code to check that the strides we assumed to be one are really
2885 // one. We want the new basic block to start at the first instruction in a
2886 // sequence of instructions that form a check.
2887 emitStrideChecks(Lp, ScalarPH);
2888 // Generate the code that checks in runtime if arrays overlap. We put the
2889 // checks into a separate block to make the more common case of few elements
2891 emitMemRuntimeChecks(Lp, ScalarPH);
2893 // Generate the induction variable.
2894 // The loop step is equal to the vectorization factor (num of SIMD elements)
2895 // times the unroll factor (num of SIMD instructions).
2896 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
2897 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2899 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
2900 getDebugLocFromInstOrOperands(OldInduction));
2902 // We are going to resume the execution of the scalar loop.
2903 // Go over all of the induction variables that we found and fix the
2904 // PHIs that are left in the scalar version of the loop.
2905 // The starting values of PHI nodes depend on the counter of the last
2906 // iteration in the vectorized loop.
2907 // If we come from a bypass edge then we need to start from the original
2910 // This variable saves the new starting index for the scalar loop. It is used
2911 // to test if there are any tail iterations left once the vector loop has
2913 LoopVectorizationLegality::InductionList::iterator I, E;
2914 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2915 for (I = List->begin(), E = List->end(); I != E; ++I) {
2916 PHINode *OrigPhi = I->first;
2917 InductionDescriptor II = I->second;
2919 // Create phi nodes to merge from the backedge-taken check block.
2920 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2922 ScalarPH->getTerminator());
2924 if (OrigPhi == OldInduction) {
2925 // We know what the end value is.
2926 EndValue = CountRoundDown;
2928 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
2929 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
2930 II.getStepValue()->getType(),
2932 EndValue = II.transform(B, CRD);
2933 EndValue->setName("ind.end");
2936 // The new PHI merges the original incoming value, in case of a bypass,
2937 // or the value at the end of the vectorized loop.
2938 BCResumeVal->addIncoming(EndValue, MiddleBlock);
2940 // Fix the scalar body counter (PHI node).
2941 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2943 // The old induction's phi node in the scalar body needs the truncated
2945 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2946 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2947 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2950 // Add a check in the middle block to see if we have completed
2951 // all of the iterations in the first vector loop.
2952 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2953 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
2954 CountRoundDown, "cmp.n",
2955 MiddleBlock->getTerminator());
2956 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2957 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2959 // Get ready to start creating new instructions into the vectorized body.
2960 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
2963 LoopVectorPreHeader = Lp->getLoopPreheader();
2964 LoopScalarPreHeader = ScalarPH;
2965 LoopMiddleBlock = MiddleBlock;
2966 LoopExitBlock = ExitBlock;
2967 LoopVectorBody.push_back(VecBody);
2968 LoopScalarBody = OldBasicBlock;
2970 LoopVectorizeHints Hints(Lp, true);
2971 Hints.setAlreadyVectorized();
2975 struct CSEDenseMapInfo {
2976 static bool canHandle(Instruction *I) {
2977 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2978 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2980 static inline Instruction *getEmptyKey() {
2981 return DenseMapInfo<Instruction *>::getEmptyKey();
2983 static inline Instruction *getTombstoneKey() {
2984 return DenseMapInfo<Instruction *>::getTombstoneKey();
2986 static unsigned getHashValue(Instruction *I) {
2987 assert(canHandle(I) && "Unknown instruction!");
2988 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2989 I->value_op_end()));
2991 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2992 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2993 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2995 return LHS->isIdenticalTo(RHS);
3000 /// \brief Check whether this block is a predicated block.
3001 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
3002 /// = ...; " blocks. We start with one vectorized basic block. For every
3003 /// conditional block we split this vectorized block. Therefore, every second
3004 /// block will be a predicated one.
3005 static bool isPredicatedBlock(unsigned BlockNum) {
3006 return BlockNum % 2;
3009 ///\brief Perform cse of induction variable instructions.
3010 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
3011 // Perform simple cse.
3012 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3013 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3014 BasicBlock *BB = BBs[i];
3015 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3016 Instruction *In = &*I++;
3018 if (!CSEDenseMapInfo::canHandle(In))
3021 // Check if we can replace this instruction with any of the
3022 // visited instructions.
3023 if (Instruction *V = CSEMap.lookup(In)) {
3024 In->replaceAllUsesWith(V);
3025 In->eraseFromParent();
3028 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3029 // ...;" blocks for predicated stores. Every second block is a predicated
3031 if (isPredicatedBlock(i))
3039 /// \brief Adds a 'fast' flag to floating point operations.
3040 static Value *addFastMathFlag(Value *V) {
3041 if (isa<FPMathOperator>(V)){
3042 FastMathFlags Flags;
3043 Flags.setUnsafeAlgebra();
3044 cast<Instruction>(V)->setFastMathFlags(Flags);
3049 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3050 /// the result needs to be inserted and/or extracted from vectors.
3051 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3052 const TargetTransformInfo &TTI) {
3056 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3059 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3061 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3063 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3069 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3070 // Return the cost of the instruction, including scalarization overhead if it's
3071 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3072 // i.e. either vector version isn't available, or is too expensive.
3073 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3074 const TargetTransformInfo &TTI,
3075 const TargetLibraryInfo *TLI,
3076 bool &NeedToScalarize) {
3077 Function *F = CI->getCalledFunction();
3078 StringRef FnName = CI->getCalledFunction()->getName();
3079 Type *ScalarRetTy = CI->getType();
3080 SmallVector<Type *, 4> Tys, ScalarTys;
3081 for (auto &ArgOp : CI->arg_operands())
3082 ScalarTys.push_back(ArgOp->getType());
3084 // Estimate cost of scalarized vector call. The source operands are assumed
3085 // to be vectors, so we need to extract individual elements from there,
3086 // execute VF scalar calls, and then gather the result into the vector return
3088 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3090 return ScalarCallCost;
3092 // Compute corresponding vector type for return value and arguments.
3093 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3094 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3095 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3097 // Compute costs of unpacking argument values for the scalar calls and
3098 // packing the return values to a vector.
3099 unsigned ScalarizationCost =
3100 getScalarizationOverhead(RetTy, true, false, TTI);
3101 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3102 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3104 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3106 // If we can't emit a vector call for this function, then the currently found
3107 // cost is the cost we need to return.
3108 NeedToScalarize = true;
3109 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3112 // If the corresponding vector cost is cheaper, return its cost.
3113 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3114 if (VectorCallCost < Cost) {
3115 NeedToScalarize = false;
3116 return VectorCallCost;
3121 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3122 // factor VF. Return the cost of the instruction, including scalarization
3123 // overhead if it's needed.
3124 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3125 const TargetTransformInfo &TTI,
3126 const TargetLibraryInfo *TLI) {
3127 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3128 assert(ID && "Expected intrinsic call!");
3130 Type *RetTy = ToVectorTy(CI->getType(), VF);
3131 SmallVector<Type *, 4> Tys;
3132 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3133 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3135 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3138 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3139 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3140 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3141 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3143 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3144 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3145 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3146 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3149 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3150 // For every instruction `I` in MinBWs, truncate the operands, create a
3151 // truncated version of `I` and reextend its result. InstCombine runs
3152 // later and will remove any ext/trunc pairs.
3154 for (auto &KV : MinBWs) {
3155 VectorParts &Parts = WidenMap.get(KV.first);
3156 for (Value *&I : Parts) {
3159 Type *OriginalTy = I->getType();
3160 Type *ScalarTruncatedTy = IntegerType::get(OriginalTy->getContext(),
3162 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3163 OriginalTy->getVectorNumElements());
3164 if (TruncatedTy == OriginalTy)
3167 IRBuilder<> B(cast<Instruction>(I));
3168 auto ShrinkOperand = [&](Value *V) -> Value* {
3169 if (auto *ZI = dyn_cast<ZExtInst>(V))
3170 if (ZI->getSrcTy() == TruncatedTy)
3171 return ZI->getOperand(0);
3172 return B.CreateZExtOrTrunc(V, TruncatedTy);
3175 // The actual instruction modification depends on the instruction type,
3177 Value *NewI = nullptr;
3178 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3179 NewI = B.CreateBinOp(BO->getOpcode(),
3180 ShrinkOperand(BO->getOperand(0)),
3181 ShrinkOperand(BO->getOperand(1)));
3182 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3183 } else if (ICmpInst *CI = dyn_cast<ICmpInst>(I)) {
3184 NewI = B.CreateICmp(CI->getPredicate(),
3185 ShrinkOperand(CI->getOperand(0)),
3186 ShrinkOperand(CI->getOperand(1)));
3187 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
3188 NewI = B.CreateSelect(SI->getCondition(),
3189 ShrinkOperand(SI->getTrueValue()),
3190 ShrinkOperand(SI->getFalseValue()));
3191 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3192 switch (CI->getOpcode()) {
3193 default: llvm_unreachable("Unhandled cast!");
3194 case Instruction::Trunc:
3195 NewI = ShrinkOperand(CI->getOperand(0));
3197 case Instruction::SExt:
3198 NewI = B.CreateSExtOrTrunc(CI->getOperand(0),
3199 smallestIntegerVectorType(OriginalTy,
3202 case Instruction::ZExt:
3203 NewI = B.CreateZExtOrTrunc(CI->getOperand(0),
3204 smallestIntegerVectorType(OriginalTy,
3208 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
3209 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3211 B.CreateZExtOrTrunc(SI->getOperand(0),
3212 VectorType::get(ScalarTruncatedTy, Elements0));
3213 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3215 B.CreateZExtOrTrunc(SI->getOperand(1),
3216 VectorType::get(ScalarTruncatedTy, Elements1));
3218 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3219 } else if (isa<LoadInst>(I)) {
3220 // Don't do anything with the operands, just extend the result.
3223 llvm_unreachable("Unhandled instruction type!");
3226 // Lastly, extend the result.
3227 NewI->takeName(cast<Instruction>(I));
3228 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3229 I->replaceAllUsesWith(Res);
3230 cast<Instruction>(I)->eraseFromParent();
3235 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3236 for (auto &KV : MinBWs) {
3237 VectorParts &Parts = WidenMap.get(KV.first);
3238 for (Value *&I : Parts) {
3239 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3240 if (Inst && Inst->use_empty()) {
3241 Value *NewI = Inst->getOperand(0);
3242 Inst->eraseFromParent();
3249 void InnerLoopVectorizer::vectorizeLoop() {
3250 //===------------------------------------------------===//
3252 // Notice: any optimization or new instruction that go
3253 // into the code below should be also be implemented in
3256 //===------------------------------------------------===//
3257 Constant *Zero = Builder.getInt32(0);
3259 // In order to support reduction variables we need to be able to vectorize
3260 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3261 // stages. First, we create a new vector PHI node with no incoming edges.
3262 // We use this value when we vectorize all of the instructions that use the
3263 // PHI. Next, after all of the instructions in the block are complete we
3264 // add the new incoming edges to the PHI. At this point all of the
3265 // instructions in the basic block are vectorized, so we can use them to
3266 // construct the PHI.
3267 PhiVector RdxPHIsToFix;
3269 // Scan the loop in a topological order to ensure that defs are vectorized
3271 LoopBlocksDFS DFS(OrigLoop);
3274 // Vectorize all of the blocks in the original loop.
3275 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3276 be = DFS.endRPO(); bb != be; ++bb)
3277 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3279 // Insert truncates and extends for any truncated instructions as hints to
3282 truncateToMinimalBitwidths();
3284 // At this point every instruction in the original loop is widened to
3285 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3286 // that we vectorized. The PHI nodes are currently empty because we did
3287 // not want to introduce cycles. Notice that the remaining PHI nodes
3288 // that we need to fix are reduction variables.
3290 // Create the 'reduced' values for each of the induction vars.
3291 // The reduced values are the vector values that we scalarize and combine
3292 // after the loop is finished.
3293 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3295 PHINode *RdxPhi = *it;
3296 assert(RdxPhi && "Unable to recover vectorized PHI");
3298 // Find the reduction variable descriptor.
3299 assert(Legal->getReductionVars()->count(RdxPhi) &&
3300 "Unable to find the reduction variable");
3301 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3303 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3304 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3305 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3306 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3307 RdxDesc.getMinMaxRecurrenceKind();
3308 setDebugLocFromInst(Builder, ReductionStartValue);
3310 // We need to generate a reduction vector from the incoming scalar.
3311 // To do so, we need to generate the 'identity' vector and override
3312 // one of the elements with the incoming scalar reduction. We need
3313 // to do it in the vector-loop preheader.
3314 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3316 // This is the vector-clone of the value that leaves the loop.
3317 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3318 Type *VecTy = VectorExit[0]->getType();
3320 // Find the reduction identity variable. Zero for addition, or, xor,
3321 // one for multiplication, -1 for And.
3324 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3325 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3326 // MinMax reduction have the start value as their identify.
3328 VectorStart = Identity = ReductionStartValue;
3330 VectorStart = Identity =
3331 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3334 // Handle other reduction kinds:
3335 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3336 RK, VecTy->getScalarType());
3339 // This vector is the Identity vector where the first element is the
3340 // incoming scalar reduction.
3341 VectorStart = ReductionStartValue;
3343 Identity = ConstantVector::getSplat(VF, Iden);
3345 // This vector is the Identity vector where the first element is the
3346 // incoming scalar reduction.
3348 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3352 // Fix the vector-loop phi.
3354 // Reductions do not have to start at zero. They can start with
3355 // any loop invariant values.
3356 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3357 BasicBlock *Latch = OrigLoop->getLoopLatch();
3358 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3359 VectorParts &Val = getVectorValue(LoopVal);
3360 for (unsigned part = 0; part < UF; ++part) {
3361 // Make sure to add the reduction stat value only to the
3362 // first unroll part.
3363 Value *StartVal = (part == 0) ? VectorStart : Identity;
3364 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3365 LoopVectorPreHeader);
3366 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3367 LoopVectorBody.back());
3370 // Before each round, move the insertion point right between
3371 // the PHIs and the values we are going to write.
3372 // This allows us to write both PHINodes and the extractelement
3374 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3376 VectorParts RdxParts = getVectorValue(LoopExitInst);
3377 setDebugLocFromInst(Builder, LoopExitInst);
3379 // If the vector reduction can be performed in a smaller type, we truncate
3380 // then extend the loop exit value to enable InstCombine to evaluate the
3381 // entire expression in the smaller type.
3382 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3383 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3384 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3385 for (unsigned part = 0; part < UF; ++part) {
3386 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3387 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3388 : Builder.CreateZExt(Trunc, VecTy);
3389 for (Value::user_iterator UI = RdxParts[part]->user_begin();
3390 UI != RdxParts[part]->user_end();)
3392 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
3393 RdxParts[part] = Extnd;
3398 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3399 for (unsigned part = 0; part < UF; ++part)
3400 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3403 // Reduce all of the unrolled parts into a single vector.
3404 Value *ReducedPartRdx = RdxParts[0];
3405 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3406 setDebugLocFromInst(Builder, ReducedPartRdx);
3407 for (unsigned part = 1; part < UF; ++part) {
3408 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3409 // Floating point operations had to be 'fast' to enable the reduction.
3410 ReducedPartRdx = addFastMathFlag(
3411 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3412 ReducedPartRdx, "bin.rdx"));
3414 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3415 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3419 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3420 // and vector ops, reducing the set of values being computed by half each
3422 assert(isPowerOf2_32(VF) &&
3423 "Reduction emission only supported for pow2 vectors!");
3424 Value *TmpVec = ReducedPartRdx;
3425 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3426 for (unsigned i = VF; i != 1; i >>= 1) {
3427 // Move the upper half of the vector to the lower half.
3428 for (unsigned j = 0; j != i/2; ++j)
3429 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3431 // Fill the rest of the mask with undef.
3432 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3433 UndefValue::get(Builder.getInt32Ty()));
3436 Builder.CreateShuffleVector(TmpVec,
3437 UndefValue::get(TmpVec->getType()),
3438 ConstantVector::get(ShuffleMask),
3441 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3442 // Floating point operations had to be 'fast' to enable the reduction.
3443 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3444 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3446 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3450 // The result is in the first element of the vector.
3451 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3452 Builder.getInt32(0));
3454 // If the reduction can be performed in a smaller type, we need to extend
3455 // the reduction to the wider type before we branch to the original loop.
3456 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3459 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3460 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3463 // Create a phi node that merges control-flow from the backedge-taken check
3464 // block and the middle block.
3465 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3466 LoopScalarPreHeader->getTerminator());
3467 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
3468 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
3469 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3471 // Now, we need to fix the users of the reduction variable
3472 // inside and outside of the scalar remainder loop.
3473 // We know that the loop is in LCSSA form. We need to update the
3474 // PHI nodes in the exit blocks.
3475 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3476 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3477 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3478 if (!LCSSAPhi) break;
3480 // All PHINodes need to have a single entry edge, or two if
3481 // we already fixed them.
3482 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3484 // We found our reduction value exit-PHI. Update it with the
3485 // incoming bypass edge.
3486 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3487 // Add an edge coming from the bypass.
3488 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3491 }// end of the LCSSA phi scan.
3493 // Fix the scalar loop reduction variable with the incoming reduction sum
3494 // from the vector body and from the backedge value.
3495 int IncomingEdgeBlockIdx =
3496 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3497 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3498 // Pick the other block.
3499 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3500 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3501 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3502 }// end of for each redux variable.
3506 // Make sure DomTree is updated.
3509 // Predicate any stores.
3510 for (auto KV : PredicatedStores) {
3511 BasicBlock::iterator I(KV.first);
3512 auto *BB = SplitBlock(I->getParent(), &*std::next(I), DT, LI);
3513 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
3514 /*BranchWeights=*/nullptr, DT);
3516 I->getParent()->setName("pred.store.if");
3517 BB->setName("pred.store.continue");
3519 DEBUG(DT->verifyDomTree());
3520 // Remove redundant induction instructions.
3521 cse(LoopVectorBody);
3524 void InnerLoopVectorizer::fixLCSSAPHIs() {
3525 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3526 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3527 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3528 if (!LCSSAPhi) break;
3529 if (LCSSAPhi->getNumIncomingValues() == 1)
3530 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3535 InnerLoopVectorizer::VectorParts
3536 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3537 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3540 // Look for cached value.
3541 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3542 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3543 if (ECEntryIt != MaskCache.end())
3544 return ECEntryIt->second;
3546 VectorParts SrcMask = createBlockInMask(Src);
3548 // The terminator has to be a branch inst!
3549 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3550 assert(BI && "Unexpected terminator found");
3552 if (BI->isConditional()) {
3553 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3555 if (BI->getSuccessor(0) != Dst)
3556 for (unsigned part = 0; part < UF; ++part)
3557 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3559 for (unsigned part = 0; part < UF; ++part)
3560 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3562 MaskCache[Edge] = EdgeMask;
3566 MaskCache[Edge] = SrcMask;
3570 InnerLoopVectorizer::VectorParts
3571 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3572 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3574 // Loop incoming mask is all-one.
3575 if (OrigLoop->getHeader() == BB) {
3576 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3577 return getVectorValue(C);
3580 // This is the block mask. We OR all incoming edges, and with zero.
3581 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3582 VectorParts BlockMask = getVectorValue(Zero);
3585 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3586 VectorParts EM = createEdgeMask(*it, BB);
3587 for (unsigned part = 0; part < UF; ++part)
3588 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3594 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3595 InnerLoopVectorizer::VectorParts &Entry,
3596 unsigned UF, unsigned VF, PhiVector *PV) {
3597 PHINode* P = cast<PHINode>(PN);
3598 // Handle reduction variables:
3599 if (Legal->getReductionVars()->count(P)) {
3600 for (unsigned part = 0; part < UF; ++part) {
3601 // This is phase one of vectorizing PHIs.
3602 Type *VecTy = (VF == 1) ? PN->getType() :
3603 VectorType::get(PN->getType(), VF);
3604 Entry[part] = PHINode::Create(
3605 VecTy, 2, "vec.phi", &*LoopVectorBody.back()->getFirstInsertionPt());
3611 setDebugLocFromInst(Builder, P);
3612 // Check for PHI nodes that are lowered to vector selects.
3613 if (P->getParent() != OrigLoop->getHeader()) {
3614 // We know that all PHIs in non-header blocks are converted into
3615 // selects, so we don't have to worry about the insertion order and we
3616 // can just use the builder.
3617 // At this point we generate the predication tree. There may be
3618 // duplications since this is a simple recursive scan, but future
3619 // optimizations will clean it up.
3621 unsigned NumIncoming = P->getNumIncomingValues();
3623 // Generate a sequence of selects of the form:
3624 // SELECT(Mask3, In3,
3625 // SELECT(Mask2, In2,
3627 for (unsigned In = 0; In < NumIncoming; In++) {
3628 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3630 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3632 for (unsigned part = 0; part < UF; ++part) {
3633 // We might have single edge PHIs (blocks) - use an identity
3634 // 'select' for the first PHI operand.
3636 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3639 // Select between the current value and the previous incoming edge
3640 // based on the incoming mask.
3641 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3642 Entry[part], "predphi");
3648 // This PHINode must be an induction variable.
3649 // Make sure that we know about it.
3650 assert(Legal->getInductionVars()->count(P) &&
3651 "Not an induction variable");
3653 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3655 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3656 // which can be found from the original scalar operations.
3657 switch (II.getKind()) {
3658 case InductionDescriptor::IK_NoInduction:
3659 llvm_unreachable("Unknown induction");
3660 case InductionDescriptor::IK_IntInduction: {
3661 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3662 // Handle other induction variables that are now based on the
3664 Value *V = Induction;
3665 if (P != OldInduction) {
3666 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3667 V = II.transform(Builder, V);
3668 V->setName("offset.idx");
3670 Value *Broadcasted = getBroadcastInstrs(V);
3671 // After broadcasting the induction variable we need to make the vector
3672 // consecutive by adding 0, 1, 2, etc.
3673 for (unsigned part = 0; part < UF; ++part)
3674 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3677 case InductionDescriptor::IK_PtrInduction:
3678 // Handle the pointer induction variable case.
3679 assert(P->getType()->isPointerTy() && "Unexpected type.");
3680 // This is the normalized GEP that starts counting at zero.
3681 Value *PtrInd = Induction;
3682 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3683 // This is the vector of results. Notice that we don't generate
3684 // vector geps because scalar geps result in better code.
3685 for (unsigned part = 0; part < UF; ++part) {
3687 int EltIndex = part;
3688 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3689 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3690 Value *SclrGep = II.transform(Builder, GlobalIdx);
3691 SclrGep->setName("next.gep");
3692 Entry[part] = SclrGep;
3696 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3697 for (unsigned int i = 0; i < VF; ++i) {
3698 int EltIndex = i + part * VF;
3699 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3700 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3701 Value *SclrGep = II.transform(Builder, GlobalIdx);
3702 SclrGep->setName("next.gep");
3703 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3704 Builder.getInt32(i),
3707 Entry[part] = VecVal;
3713 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3714 // For each instruction in the old loop.
3715 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3716 VectorParts &Entry = WidenMap.get(&*it);
3718 switch (it->getOpcode()) {
3719 case Instruction::Br:
3720 // Nothing to do for PHIs and BR, since we already took care of the
3721 // loop control flow instructions.
3723 case Instruction::PHI: {
3724 // Vectorize PHINodes.
3725 widenPHIInstruction(&*it, Entry, UF, VF, PV);
3729 case Instruction::Add:
3730 case Instruction::FAdd:
3731 case Instruction::Sub:
3732 case Instruction::FSub:
3733 case Instruction::Mul:
3734 case Instruction::FMul:
3735 case Instruction::UDiv:
3736 case Instruction::SDiv:
3737 case Instruction::FDiv:
3738 case Instruction::URem:
3739 case Instruction::SRem:
3740 case Instruction::FRem:
3741 case Instruction::Shl:
3742 case Instruction::LShr:
3743 case Instruction::AShr:
3744 case Instruction::And:
3745 case Instruction::Or:
3746 case Instruction::Xor: {
3747 // Just widen binops.
3748 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3749 setDebugLocFromInst(Builder, BinOp);
3750 VectorParts &A = getVectorValue(it->getOperand(0));
3751 VectorParts &B = getVectorValue(it->getOperand(1));
3753 // Use this vector value for all users of the original instruction.
3754 for (unsigned Part = 0; Part < UF; ++Part) {
3755 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3757 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3758 VecOp->copyIRFlags(BinOp);
3763 propagateMetadata(Entry, &*it);
3766 case Instruction::Select: {
3768 // If the selector is loop invariant we can create a select
3769 // instruction with a scalar condition. Otherwise, use vector-select.
3770 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3772 setDebugLocFromInst(Builder, &*it);
3774 // The condition can be loop invariant but still defined inside the
3775 // loop. This means that we can't just use the original 'cond' value.
3776 // We have to take the 'vectorized' value and pick the first lane.
3777 // Instcombine will make this a no-op.
3778 VectorParts &Cond = getVectorValue(it->getOperand(0));
3779 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3780 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3782 Value *ScalarCond = (VF == 1) ? Cond[0] :
3783 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3785 for (unsigned Part = 0; Part < UF; ++Part) {
3786 Entry[Part] = Builder.CreateSelect(
3787 InvariantCond ? ScalarCond : Cond[Part],
3792 propagateMetadata(Entry, &*it);
3796 case Instruction::ICmp:
3797 case Instruction::FCmp: {
3798 // Widen compares. Generate vector compares.
3799 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3800 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3801 setDebugLocFromInst(Builder, &*it);
3802 VectorParts &A = getVectorValue(it->getOperand(0));
3803 VectorParts &B = getVectorValue(it->getOperand(1));
3804 for (unsigned Part = 0; Part < UF; ++Part) {
3807 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3808 cast<FCmpInst>(C)->copyFastMathFlags(&*it);
3810 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3815 propagateMetadata(Entry, &*it);
3819 case Instruction::Store:
3820 case Instruction::Load:
3821 vectorizeMemoryInstruction(&*it);
3823 case Instruction::ZExt:
3824 case Instruction::SExt:
3825 case Instruction::FPToUI:
3826 case Instruction::FPToSI:
3827 case Instruction::FPExt:
3828 case Instruction::PtrToInt:
3829 case Instruction::IntToPtr:
3830 case Instruction::SIToFP:
3831 case Instruction::UIToFP:
3832 case Instruction::Trunc:
3833 case Instruction::FPTrunc:
3834 case Instruction::BitCast: {
3835 CastInst *CI = dyn_cast<CastInst>(it);
3836 setDebugLocFromInst(Builder, &*it);
3837 /// Optimize the special case where the source is the induction
3838 /// variable. Notice that we can only optimize the 'trunc' case
3839 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3840 /// c. other casts depend on pointer size.
3841 if (CI->getOperand(0) == OldInduction &&
3842 it->getOpcode() == Instruction::Trunc) {
3843 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3845 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3846 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3848 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3849 for (unsigned Part = 0; Part < UF; ++Part)
3850 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3851 propagateMetadata(Entry, &*it);
3854 /// Vectorize casts.
3855 Type *DestTy = (VF == 1) ? CI->getType() :
3856 VectorType::get(CI->getType(), VF);
3858 VectorParts &A = getVectorValue(it->getOperand(0));
3859 for (unsigned Part = 0; Part < UF; ++Part)
3860 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3861 propagateMetadata(Entry, &*it);
3865 case Instruction::Call: {
3866 // Ignore dbg intrinsics.
3867 if (isa<DbgInfoIntrinsic>(it))
3869 setDebugLocFromInst(Builder, &*it);
3871 Module *M = BB->getParent()->getParent();
3872 CallInst *CI = cast<CallInst>(it);
3874 StringRef FnName = CI->getCalledFunction()->getName();
3875 Function *F = CI->getCalledFunction();
3876 Type *RetTy = ToVectorTy(CI->getType(), VF);
3877 SmallVector<Type *, 4> Tys;
3878 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3879 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3881 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3883 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3884 ID == Intrinsic::lifetime_start)) {
3885 scalarizeInstruction(&*it);
3888 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3889 // version of the instruction.
3890 // Is it beneficial to perform intrinsic call compared to lib call?
3891 bool NeedToScalarize;
3892 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3893 bool UseVectorIntrinsic =
3894 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3895 if (!UseVectorIntrinsic && NeedToScalarize) {
3896 scalarizeInstruction(&*it);
3900 for (unsigned Part = 0; Part < UF; ++Part) {
3901 SmallVector<Value *, 4> Args;
3902 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3903 Value *Arg = CI->getArgOperand(i);
3904 // Some intrinsics have a scalar argument - don't replace it with a
3906 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3907 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3908 Arg = VectorArg[Part];
3910 Args.push_back(Arg);
3914 if (UseVectorIntrinsic) {
3915 // Use vector version of the intrinsic.
3916 Type *TysForDecl[] = {CI->getType()};
3918 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3919 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3921 // Use vector version of the library call.
3922 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3923 assert(!VFnName.empty() && "Vector function name is empty.");
3924 VectorF = M->getFunction(VFnName);
3926 // Generate a declaration
3927 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3929 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3930 VectorF->copyAttributesFrom(F);
3933 assert(VectorF && "Can't create vector function.");
3934 Entry[Part] = Builder.CreateCall(VectorF, Args);
3937 propagateMetadata(Entry, &*it);
3942 // All other instructions are unsupported. Scalarize them.
3943 scalarizeInstruction(&*it);
3946 }// end of for_each instr.
3949 void InnerLoopVectorizer::updateAnalysis() {
3950 // Forget the original basic block.
3951 SE->forgetLoop(OrigLoop);
3953 // Update the dominator tree information.
3954 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3955 "Entry does not dominate exit.");
3957 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3958 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3959 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3961 // We don't predicate stores by this point, so the vector body should be a
3963 assert(LoopVectorBody.size() == 1 && "Expected single block loop!");
3964 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3966 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody.back());
3967 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3968 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3969 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3971 DEBUG(DT->verifyDomTree());
3974 /// \brief Check whether it is safe to if-convert this phi node.
3976 /// Phi nodes with constant expressions that can trap are not safe to if
3978 static bool canIfConvertPHINodes(BasicBlock *BB) {
3979 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3980 PHINode *Phi = dyn_cast<PHINode>(I);
3983 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3984 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3991 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3992 if (!EnableIfConversion) {
3993 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3997 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3999 // A list of pointers that we can safely read and write to.
4000 SmallPtrSet<Value *, 8> SafePointes;
4002 // Collect safe addresses.
4003 for (Loop::block_iterator BI = TheLoop->block_begin(),
4004 BE = TheLoop->block_end(); BI != BE; ++BI) {
4005 BasicBlock *BB = *BI;
4007 if (blockNeedsPredication(BB))
4010 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
4011 if (LoadInst *LI = dyn_cast<LoadInst>(I))
4012 SafePointes.insert(LI->getPointerOperand());
4013 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
4014 SafePointes.insert(SI->getPointerOperand());
4018 // Collect the blocks that need predication.
4019 BasicBlock *Header = TheLoop->getHeader();
4020 for (Loop::block_iterator BI = TheLoop->block_begin(),
4021 BE = TheLoop->block_end(); BI != BE; ++BI) {
4022 BasicBlock *BB = *BI;
4024 // We don't support switch statements inside loops.
4025 if (!isa<BranchInst>(BB->getTerminator())) {
4026 emitAnalysis(VectorizationReport(BB->getTerminator())
4027 << "loop contains a switch statement");
4031 // We must be able to predicate all blocks that need to be predicated.
4032 if (blockNeedsPredication(BB)) {
4033 if (!blockCanBePredicated(BB, SafePointes)) {
4034 emitAnalysis(VectorizationReport(BB->getTerminator())
4035 << "control flow cannot be substituted for a select");
4038 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
4039 emitAnalysis(VectorizationReport(BB->getTerminator())
4040 << "control flow cannot be substituted for a select");
4045 // We can if-convert this loop.
4049 bool LoopVectorizationLegality::canVectorize() {
4050 // We must have a loop in canonical form. Loops with indirectbr in them cannot
4051 // be canonicalized.
4052 if (!TheLoop->getLoopPreheader()) {
4054 VectorizationReport() <<
4055 "loop control flow is not understood by vectorizer");
4059 // We can only vectorize innermost loops.
4060 if (!TheLoop->empty()) {
4061 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
4065 // We must have a single backedge.
4066 if (TheLoop->getNumBackEdges() != 1) {
4068 VectorizationReport() <<
4069 "loop control flow is not understood by vectorizer");
4073 // We must have a single exiting block.
4074 if (!TheLoop->getExitingBlock()) {
4076 VectorizationReport() <<
4077 "loop control flow is not understood by vectorizer");
4081 // We only handle bottom-tested loops, i.e. loop in which the condition is
4082 // checked at the end of each iteration. With that we can assume that all
4083 // instructions in the loop are executed the same number of times.
4084 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
4086 VectorizationReport() <<
4087 "loop control flow is not understood by vectorizer");
4091 // We need to have a loop header.
4092 DEBUG(dbgs() << "LV: Found a loop: " <<
4093 TheLoop->getHeader()->getName() << '\n');
4095 // Check if we can if-convert non-single-bb loops.
4096 unsigned NumBlocks = TheLoop->getNumBlocks();
4097 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4098 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4102 // ScalarEvolution needs to be able to find the exit count.
4103 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
4104 if (ExitCount == SE->getCouldNotCompute()) {
4105 emitAnalysis(VectorizationReport() <<
4106 "could not determine number of loop iterations");
4107 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4111 // Check if we can vectorize the instructions and CFG in this loop.
4112 if (!canVectorizeInstrs()) {
4113 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4117 // Go over each instruction and look at memory deps.
4118 if (!canVectorizeMemory()) {
4119 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4123 // Collect all of the variables that remain uniform after vectorization.
4124 collectLoopUniforms();
4126 DEBUG(dbgs() << "LV: We can vectorize this loop"
4127 << (LAI->getRuntimePointerChecking()->Need
4128 ? " (with a runtime bound check)"
4132 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4134 // If an override option has been passed in for interleaved accesses, use it.
4135 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4136 UseInterleaved = EnableInterleavedMemAccesses;
4138 // Analyze interleaved memory accesses.
4140 InterleaveInfo.analyzeInterleaving(Strides);
4142 // Okay! We can vectorize. At this point we don't have any other mem analysis
4143 // which may limit our maximum vectorization factor, so just return true with
4148 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4149 if (Ty->isPointerTy())
4150 return DL.getIntPtrType(Ty);
4152 // It is possible that char's or short's overflow when we ask for the loop's
4153 // trip count, work around this by changing the type size.
4154 if (Ty->getScalarSizeInBits() < 32)
4155 return Type::getInt32Ty(Ty->getContext());
4160 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4161 Ty0 = convertPointerToIntegerType(DL, Ty0);
4162 Ty1 = convertPointerToIntegerType(DL, Ty1);
4163 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4168 /// \brief Check that the instruction has outside loop users and is not an
4169 /// identified reduction variable.
4170 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4171 SmallPtrSetImpl<Value *> &Reductions) {
4172 // Reduction instructions are allowed to have exit users. All other
4173 // instructions must not have external users.
4174 if (!Reductions.count(Inst))
4175 //Check that all of the users of the loop are inside the BB.
4176 for (User *U : Inst->users()) {
4177 Instruction *UI = cast<Instruction>(U);
4178 // This user may be a reduction exit value.
4179 if (!TheLoop->contains(UI)) {
4180 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4187 bool LoopVectorizationLegality::canVectorizeInstrs() {
4188 BasicBlock *Header = TheLoop->getHeader();
4190 // Look for the attribute signaling the absence of NaNs.
4191 Function &F = *Header->getParent();
4192 const DataLayout &DL = F.getParent()->getDataLayout();
4193 if (F.hasFnAttribute("no-nans-fp-math"))
4195 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4197 // For each block in the loop.
4198 for (Loop::block_iterator bb = TheLoop->block_begin(),
4199 be = TheLoop->block_end(); bb != be; ++bb) {
4201 // Scan the instructions in the block and look for hazards.
4202 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4205 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4206 Type *PhiTy = Phi->getType();
4207 // Check that this PHI type is allowed.
4208 if (!PhiTy->isIntegerTy() &&
4209 !PhiTy->isFloatingPointTy() &&
4210 !PhiTy->isPointerTy()) {
4211 emitAnalysis(VectorizationReport(&*it)
4212 << "loop control flow is not understood by vectorizer");
4213 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4217 // If this PHINode is not in the header block, then we know that we
4218 // can convert it to select during if-conversion. No need to check if
4219 // the PHIs in this block are induction or reduction variables.
4220 if (*bb != Header) {
4221 // Check that this instruction has no outside users or is an
4222 // identified reduction value with an outside user.
4223 if (!hasOutsideLoopUser(TheLoop, &*it, AllowedExit))
4225 emitAnalysis(VectorizationReport(&*it) <<
4226 "value could not be identified as "
4227 "an induction or reduction variable");
4231 // We only allow if-converted PHIs with exactly two incoming values.
4232 if (Phi->getNumIncomingValues() != 2) {
4233 emitAnalysis(VectorizationReport(&*it)
4234 << "control flow not understood by vectorizer");
4235 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4239 InductionDescriptor ID;
4240 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4241 Inductions[Phi] = ID;
4242 // Get the widest type.
4244 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4246 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4248 // Int inductions are special because we only allow one IV.
4249 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4250 ID.getStepValue()->isOne() &&
4251 isa<Constant>(ID.getStartValue()) &&
4252 cast<Constant>(ID.getStartValue())->isNullValue()) {
4253 // Use the phi node with the widest type as induction. Use the last
4254 // one if there are multiple (no good reason for doing this other
4255 // than it is expedient). We've checked that it begins at zero and
4256 // steps by one, so this is a canonical induction variable.
4257 if (!Induction || PhiTy == WidestIndTy)
4261 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4263 // Until we explicitly handle the case of an induction variable with
4264 // an outside loop user we have to give up vectorizing this loop.
4265 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4266 emitAnalysis(VectorizationReport(&*it) <<
4267 "use of induction value outside of the "
4268 "loop is not handled by vectorizer");
4275 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4277 if (Reductions[Phi].hasUnsafeAlgebra())
4278 Requirements->addUnsafeAlgebraInst(
4279 Reductions[Phi].getUnsafeAlgebraInst());
4280 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4284 emitAnalysis(VectorizationReport(&*it) <<
4285 "value that could not be identified as "
4286 "reduction is used outside the loop");
4287 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4289 }// end of PHI handling
4291 // We handle calls that:
4292 // * Are debug info intrinsics.
4293 // * Have a mapping to an IR intrinsic.
4294 // * Have a vector version available.
4295 CallInst *CI = dyn_cast<CallInst>(it);
4296 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4297 !(CI->getCalledFunction() && TLI &&
4298 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4299 emitAnalysis(VectorizationReport(&*it)
4300 << "call instruction cannot be vectorized");
4301 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4305 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4306 // second argument is the same (i.e. loop invariant)
4308 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4309 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4310 emitAnalysis(VectorizationReport(&*it)
4311 << "intrinsic instruction cannot be vectorized");
4312 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4317 // Check that the instruction return type is vectorizable.
4318 // Also, we can't vectorize extractelement instructions.
4319 if ((!VectorType::isValidElementType(it->getType()) &&
4320 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4321 emitAnalysis(VectorizationReport(&*it)
4322 << "instruction return type cannot be vectorized");
4323 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4327 // Check that the stored type is vectorizable.
4328 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4329 Type *T = ST->getValueOperand()->getType();
4330 if (!VectorType::isValidElementType(T)) {
4331 emitAnalysis(VectorizationReport(ST) <<
4332 "store instruction cannot be vectorized");
4335 if (EnableMemAccessVersioning)
4336 collectStridedAccess(ST);
4339 if (EnableMemAccessVersioning)
4340 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4341 collectStridedAccess(LI);
4343 // Reduction instructions are allowed to have exit users.
4344 // All other instructions must not have external users.
4345 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4346 emitAnalysis(VectorizationReport(&*it) <<
4347 "value cannot be used outside the loop");
4356 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4357 if (Inductions.empty()) {
4358 emitAnalysis(VectorizationReport()
4359 << "loop induction variable could not be identified");
4364 // Now we know the widest induction type, check if our found induction
4365 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4366 // will create another.
4367 if (Induction && WidestIndTy != Induction->getType())
4368 Induction = nullptr;
4373 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4374 Value *Ptr = nullptr;
4375 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4376 Ptr = LI->getPointerOperand();
4377 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4378 Ptr = SI->getPointerOperand();
4382 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4386 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4387 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4388 Strides[Ptr] = Stride;
4389 StrideSet.insert(Stride);
4392 void LoopVectorizationLegality::collectLoopUniforms() {
4393 // We now know that the loop is vectorizable!
4394 // Collect variables that will remain uniform after vectorization.
4395 std::vector<Value*> Worklist;
4396 BasicBlock *Latch = TheLoop->getLoopLatch();
4398 // Start with the conditional branch and walk up the block.
4399 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4401 // Also add all consecutive pointer values; these values will be uniform
4402 // after vectorization (and subsequent cleanup) and, until revectorization is
4403 // supported, all dependencies must also be uniform.
4404 for (Loop::block_iterator B = TheLoop->block_begin(),
4405 BE = TheLoop->block_end(); B != BE; ++B)
4406 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4408 if (I->getType()->isPointerTy() && isConsecutivePtr(&*I))
4409 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4411 while (!Worklist.empty()) {
4412 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4413 Worklist.pop_back();
4415 // Look at instructions inside this loop.
4416 // Stop when reaching PHI nodes.
4417 // TODO: we need to follow values all over the loop, not only in this block.
4418 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4421 // This is a known uniform.
4424 // Insert all operands.
4425 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4429 bool LoopVectorizationLegality::canVectorizeMemory() {
4430 LAI = &LAA->getInfo(TheLoop, Strides);
4431 auto &OptionalReport = LAI->getReport();
4433 emitAnalysis(VectorizationReport(*OptionalReport));
4434 if (!LAI->canVectorizeMemory())
4437 if (LAI->hasStoreToLoopInvariantAddress()) {
4439 VectorizationReport()
4440 << "write to a loop invariant address could not be vectorized");
4441 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4445 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4450 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4451 Value *In0 = const_cast<Value*>(V);
4452 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4456 return Inductions.count(PN);
4459 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4460 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4463 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4464 SmallPtrSetImpl<Value *> &SafePtrs) {
4466 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4467 // Check that we don't have a constant expression that can trap as operand.
4468 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4470 if (Constant *C = dyn_cast<Constant>(*OI))
4474 // We might be able to hoist the load.
4475 if (it->mayReadFromMemory()) {
4476 LoadInst *LI = dyn_cast<LoadInst>(it);
4479 if (!SafePtrs.count(LI->getPointerOperand())) {
4480 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4481 MaskedOp.insert(LI);
4488 // We don't predicate stores at the moment.
4489 if (it->mayWriteToMemory()) {
4490 StoreInst *SI = dyn_cast<StoreInst>(it);
4491 // We only support predication of stores in basic blocks with one
4496 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4497 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4499 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4500 !isSinglePredecessor) {
4501 // Build a masked store if it is legal for the target, otherwise scalarize
4503 bool isLegalMaskedOp =
4504 isLegalMaskedStore(SI->getValueOperand()->getType(),
4505 SI->getPointerOperand());
4506 if (isLegalMaskedOp) {
4508 MaskedOp.insert(SI);
4517 // The instructions below can trap.
4518 switch (it->getOpcode()) {
4520 case Instruction::UDiv:
4521 case Instruction::SDiv:
4522 case Instruction::URem:
4523 case Instruction::SRem:
4531 void InterleavedAccessInfo::collectConstStridedAccesses(
4532 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4533 const ValueToValueMap &Strides) {
4534 // Holds load/store instructions in program order.
4535 SmallVector<Instruction *, 16> AccessList;
4537 for (auto *BB : TheLoop->getBlocks()) {
4538 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4540 for (auto &I : *BB) {
4541 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4543 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4547 AccessList.push_back(&I);
4551 if (AccessList.empty())
4554 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4555 for (auto I : AccessList) {
4556 LoadInst *LI = dyn_cast<LoadInst>(I);
4557 StoreInst *SI = dyn_cast<StoreInst>(I);
4559 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4560 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4562 // The factor of the corresponding interleave group.
4563 unsigned Factor = std::abs(Stride);
4565 // Ignore the access if the factor is too small or too large.
4566 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4569 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4570 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4571 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4573 // An alignment of 0 means target ABI alignment.
4574 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4576 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4578 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4582 // Analyze interleaved accesses and collect them into interleave groups.
4584 // Notice that the vectorization on interleaved groups will change instruction
4585 // orders and may break dependences. But the memory dependence check guarantees
4586 // that there is no overlap between two pointers of different strides, element
4587 // sizes or underlying bases.
4589 // For pointers sharing the same stride, element size and underlying base, no
4590 // need to worry about Read-After-Write dependences and Write-After-Read
4593 // E.g. The RAW dependence: A[i] = a;
4595 // This won't exist as it is a store-load forwarding conflict, which has
4596 // already been checked and forbidden in the dependence check.
4598 // E.g. The WAR dependence: a = A[i]; // (1)
4600 // The store group of (2) is always inserted at or below (2), and the load group
4601 // of (1) is always inserted at or above (1). The dependence is safe.
4602 void InterleavedAccessInfo::analyzeInterleaving(
4603 const ValueToValueMap &Strides) {
4604 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4606 // Holds all the stride accesses.
4607 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4608 collectConstStridedAccesses(StrideAccesses, Strides);
4610 if (StrideAccesses.empty())
4613 // Holds all interleaved store groups temporarily.
4614 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4616 // Search the load-load/write-write pair B-A in bottom-up order and try to
4617 // insert B into the interleave group of A according to 3 rules:
4618 // 1. A and B have the same stride.
4619 // 2. A and B have the same memory object size.
4620 // 3. B belongs to the group according to the distance.
4622 // The bottom-up order can avoid breaking the Write-After-Write dependences
4623 // between two pointers of the same base.
4624 // E.g. A[i] = a; (1)
4627 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4628 // above (1), which guarantees that (1) is always above (2).
4629 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4631 Instruction *A = I->first;
4632 StrideDescriptor DesA = I->second;
4634 InterleaveGroup *Group = getInterleaveGroup(A);
4636 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4637 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4640 if (A->mayWriteToMemory())
4641 StoreGroups.insert(Group);
4643 for (auto II = std::next(I); II != E; ++II) {
4644 Instruction *B = II->first;
4645 StrideDescriptor DesB = II->second;
4647 // Ignore if B is already in a group or B is a different memory operation.
4648 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4651 // Check the rule 1 and 2.
4652 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4655 // Calculate the distance and prepare for the rule 3.
4656 const SCEVConstant *DistToA =
4657 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4661 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4663 // Skip if the distance is not multiple of size as they are not in the
4665 if (DistanceToA % static_cast<int>(DesA.Size))
4668 // The index of B is the index of A plus the related index to A.
4670 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4672 // Try to insert B into the group.
4673 if (Group->insertMember(B, IndexB, DesB.Align)) {
4674 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4675 << " into the interleave group with" << *A << '\n');
4676 InterleaveGroupMap[B] = Group;
4678 // Set the first load in program order as the insert position.
4679 if (B->mayReadFromMemory())
4680 Group->setInsertPos(B);
4682 } // Iteration on instruction B
4683 } // Iteration on instruction A
4685 // Remove interleaved store groups with gaps.
4686 for (InterleaveGroup *Group : StoreGroups)
4687 if (Group->getNumMembers() != Group->getFactor())
4688 releaseGroup(Group);
4691 LoopVectorizationCostModel::VectorizationFactor
4692 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4693 // Width 1 means no vectorize
4694 VectorizationFactor Factor = { 1U, 0U };
4695 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4696 emitAnalysis(VectorizationReport() <<
4697 "runtime pointer checks needed. Enable vectorization of this "
4698 "loop with '#pragma clang loop vectorize(enable)' when "
4699 "compiling with -Os/-Oz");
4701 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4705 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4706 emitAnalysis(VectorizationReport() <<
4707 "store that is conditionally executed prevents vectorization");
4708 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4712 // Find the trip count.
4713 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4714 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4716 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
4717 unsigned SmallestType, WidestType;
4718 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
4719 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4720 unsigned MaxSafeDepDist = -1U;
4721 if (Legal->getMaxSafeDepDistBytes() != -1U)
4722 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4723 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4724 WidestRegister : MaxSafeDepDist);
4725 unsigned MaxVectorSize = WidestRegister / WidestType;
4727 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "
4728 << WidestType << " bits.\n");
4729 DEBUG(dbgs() << "LV: The Widest register is: "
4730 << WidestRegister << " bits.\n");
4732 if (MaxVectorSize == 0) {
4733 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4737 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4738 " into one vector!");
4740 unsigned VF = MaxVectorSize;
4741 if (MaximizeBandwidth && !OptForSize) {
4742 // Collect all viable vectorization factors.
4743 SmallVector<unsigned, 8> VFs;
4744 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
4745 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
4748 // For each VF calculate its register usage.
4749 auto RUs = calculateRegisterUsage(VFs);
4751 // Select the largest VF which doesn't require more registers than existing
4753 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
4754 for (int i = RUs.size() - 1; i >= 0; --i) {
4755 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
4762 // If we optimize the program for size, avoid creating the tail loop.
4764 // If we are unable to calculate the trip count then don't try to vectorize.
4767 (VectorizationReport() <<
4768 "unable to calculate the loop count due to complex control flow");
4769 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4773 // Find the maximum SIMD width that can fit within the trip count.
4774 VF = TC % MaxVectorSize;
4779 // If the trip count that we found modulo the vectorization factor is not
4780 // zero then we require a tail.
4781 emitAnalysis(VectorizationReport() <<
4782 "cannot optimize for size and vectorize at the "
4783 "same time. Enable vectorization of this loop "
4784 "with '#pragma clang loop vectorize(enable)' "
4785 "when compiling with -Os/-Oz");
4786 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4791 int UserVF = Hints->getWidth();
4793 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4794 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4796 Factor.Width = UserVF;
4800 float Cost = expectedCost(1);
4802 const float ScalarCost = Cost;
4805 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4807 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4808 // Ignore scalar width, because the user explicitly wants vectorization.
4809 if (ForceVectorization && VF > 1) {
4811 Cost = expectedCost(Width) / (float)Width;
4814 for (unsigned i=2; i <= VF; i*=2) {
4815 // Notice that the vector loop needs to be executed less times, so
4816 // we need to divide the cost of the vector loops by the width of
4817 // the vector elements.
4818 float VectorCost = expectedCost(i) / (float)i;
4819 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4820 (int)VectorCost << ".\n");
4821 if (VectorCost < Cost) {
4827 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4828 << "LV: Vectorization seems to be not beneficial, "
4829 << "but was forced by a user.\n");
4830 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4831 Factor.Width = Width;
4832 Factor.Cost = Width * Cost;
4836 std::pair<unsigned, unsigned>
4837 LoopVectorizationCostModel::getSmallestAndWidestTypes() {
4838 unsigned MinWidth = -1U;
4839 unsigned MaxWidth = 8;
4840 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4843 for (Loop::block_iterator bb = TheLoop->block_begin(),
4844 be = TheLoop->block_end(); bb != be; ++bb) {
4845 BasicBlock *BB = *bb;
4847 // For each instruction in the loop.
4848 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4849 Type *T = it->getType();
4851 // Skip ignored values.
4852 if (ValuesToIgnore.count(&*it))
4855 // Only examine Loads, Stores and PHINodes.
4856 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4859 // Examine PHI nodes that are reduction variables. Update the type to
4860 // account for the recurrence type.
4861 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4862 if (!Legal->getReductionVars()->count(PN))
4864 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4865 T = RdxDesc.getRecurrenceType();
4868 // Examine the stored values.
4869 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4870 T = ST->getValueOperand()->getType();
4872 // Ignore loaded pointer types and stored pointer types that are not
4873 // consecutive. However, we do want to take consecutive stores/loads of
4874 // pointer vectors into account.
4875 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&*it))
4878 MinWidth = std::min(MinWidth,
4879 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4880 MaxWidth = std::max(MaxWidth,
4881 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4885 return {MinWidth, MaxWidth};
4888 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4890 unsigned LoopCost) {
4892 // -- The interleave heuristics --
4893 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4894 // There are many micro-architectural considerations that we can't predict
4895 // at this level. For example, frontend pressure (on decode or fetch) due to
4896 // code size, or the number and capabilities of the execution ports.
4898 // We use the following heuristics to select the interleave count:
4899 // 1. If the code has reductions, then we interleave to break the cross
4900 // iteration dependency.
4901 // 2. If the loop is really small, then we interleave to reduce the loop
4903 // 3. We don't interleave if we think that we will spill registers to memory
4904 // due to the increased register pressure.
4906 // When we optimize for size, we don't interleave.
4910 // We used the distance for the interleave count.
4911 if (Legal->getMaxSafeDepDistBytes() != -1U)
4914 // Do not interleave loops with a relatively small trip count.
4915 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4916 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4919 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4920 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4924 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4925 TargetNumRegisters = ForceTargetNumScalarRegs;
4927 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4928 TargetNumRegisters = ForceTargetNumVectorRegs;
4931 RegisterUsage R = calculateRegisterUsage({VF})[0];
4932 // We divide by these constants so assume that we have at least one
4933 // instruction that uses at least one register.
4934 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4935 R.NumInstructions = std::max(R.NumInstructions, 1U);
4937 // We calculate the interleave count using the following formula.
4938 // Subtract the number of loop invariants from the number of available
4939 // registers. These registers are used by all of the interleaved instances.
4940 // Next, divide the remaining registers by the number of registers that is
4941 // required by the loop, in order to estimate how many parallel instances
4942 // fit without causing spills. All of this is rounded down if necessary to be
4943 // a power of two. We want power of two interleave count to simplify any
4944 // addressing operations or alignment considerations.
4945 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4948 // Don't count the induction variable as interleaved.
4949 if (EnableIndVarRegisterHeur)
4950 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4951 std::max(1U, (R.MaxLocalUsers - 1)));
4953 // Clamp the interleave ranges to reasonable counts.
4954 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4956 // Check if the user has overridden the max.
4958 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4959 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4961 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4962 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4965 // If we did not calculate the cost for VF (because the user selected the VF)
4966 // then we calculate the cost of VF here.
4968 LoopCost = expectedCost(VF);
4970 // Clamp the calculated IC to be between the 1 and the max interleave count
4971 // that the target allows.
4972 if (IC > MaxInterleaveCount)
4973 IC = MaxInterleaveCount;
4977 // Interleave if we vectorized this loop and there is a reduction that could
4978 // benefit from interleaving.
4979 if (VF > 1 && Legal->getReductionVars()->size()) {
4980 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4984 // Note that if we've already vectorized the loop we will have done the
4985 // runtime check and so interleaving won't require further checks.
4986 bool InterleavingRequiresRuntimePointerCheck =
4987 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4989 // We want to interleave small loops in order to reduce the loop overhead and
4990 // potentially expose ILP opportunities.
4991 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4992 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4993 // We assume that the cost overhead is 1 and we use the cost model
4994 // to estimate the cost of the loop and interleave until the cost of the
4995 // loop overhead is about 5% of the cost of the loop.
4997 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4999 // Interleave until store/load ports (estimated by max interleave count) are
5001 unsigned NumStores = Legal->getNumStores();
5002 unsigned NumLoads = Legal->getNumLoads();
5003 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
5004 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
5006 // If we have a scalar reduction (vector reductions are already dealt with
5007 // by this point), we can increase the critical path length if the loop
5008 // we're interleaving is inside another loop. Limit, by default to 2, so the
5009 // critical path only gets increased by one reduction operation.
5010 if (Legal->getReductionVars()->size() &&
5011 TheLoop->getLoopDepth() > 1) {
5012 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
5013 SmallIC = std::min(SmallIC, F);
5014 StoresIC = std::min(StoresIC, F);
5015 LoadsIC = std::min(LoadsIC, F);
5018 if (EnableLoadStoreRuntimeInterleave &&
5019 std::max(StoresIC, LoadsIC) > SmallIC) {
5020 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
5021 return std::max(StoresIC, LoadsIC);
5024 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
5028 // Interleave if this is a large loop (small loops are already dealt with by
5030 // point) that could benefit from interleaving.
5031 bool HasReductions = (Legal->getReductionVars()->size() > 0);
5032 if (TTI.enableAggressiveInterleaving(HasReductions)) {
5033 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
5037 DEBUG(dbgs() << "LV: Not Interleaving.\n");
5041 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
5042 LoopVectorizationCostModel::calculateRegisterUsage(
5043 const SmallVector<unsigned, 8> &VFs) {
5044 // This function calculates the register usage by measuring the highest number
5045 // of values that are alive at a single location. Obviously, this is a very
5046 // rough estimation. We scan the loop in a topological order in order and
5047 // assign a number to each instruction. We use RPO to ensure that defs are
5048 // met before their users. We assume that each instruction that has in-loop
5049 // users starts an interval. We record every time that an in-loop value is
5050 // used, so we have a list of the first and last occurrences of each
5051 // instruction. Next, we transpose this data structure into a multi map that
5052 // holds the list of intervals that *end* at a specific location. This multi
5053 // map allows us to perform a linear search. We scan the instructions linearly
5054 // and record each time that a new interval starts, by placing it in a set.
5055 // If we find this value in the multi-map then we remove it from the set.
5056 // The max register usage is the maximum size of the set.
5057 // We also search for instructions that are defined outside the loop, but are
5058 // used inside the loop. We need this number separately from the max-interval
5059 // usage number because when we unroll, loop-invariant values do not take
5061 LoopBlocksDFS DFS(TheLoop);
5065 RU.NumInstructions = 0;
5067 // Each 'key' in the map opens a new interval. The values
5068 // of the map are the index of the 'last seen' usage of the
5069 // instruction that is the key.
5070 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5071 // Maps instruction to its index.
5072 DenseMap<unsigned, Instruction*> IdxToInstr;
5073 // Marks the end of each interval.
5074 IntervalMap EndPoint;
5075 // Saves the list of instruction indices that are used in the loop.
5076 SmallSet<Instruction*, 8> Ends;
5077 // Saves the list of values that are used in the loop but are
5078 // defined outside the loop, such as arguments and constants.
5079 SmallPtrSet<Value*, 8> LoopInvariants;
5082 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5083 be = DFS.endRPO(); bb != be; ++bb) {
5084 RU.NumInstructions += (*bb)->size();
5085 for (Instruction &I : **bb) {
5086 IdxToInstr[Index++] = &I;
5088 // Save the end location of each USE.
5089 for (unsigned i = 0; i < I.getNumOperands(); ++i) {
5090 Value *U = I.getOperand(i);
5091 Instruction *Instr = dyn_cast<Instruction>(U);
5093 // Ignore non-instruction values such as arguments, constants, etc.
5094 if (!Instr) continue;
5096 // If this instruction is outside the loop then record it and continue.
5097 if (!TheLoop->contains(Instr)) {
5098 LoopInvariants.insert(Instr);
5102 // Overwrite previous end points.
5103 EndPoint[Instr] = Index;
5109 // Saves the list of intervals that end with the index in 'key'.
5110 typedef SmallVector<Instruction*, 2> InstrList;
5111 DenseMap<unsigned, InstrList> TransposeEnds;
5113 // Transpose the EndPoints to a list of values that end at each index.
5114 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5116 TransposeEnds[it->second].push_back(it->first);
5118 SmallSet<Instruction*, 8> OpenIntervals;
5120 // Get the size of the widest register.
5121 unsigned MaxSafeDepDist = -1U;
5122 if (Legal->getMaxSafeDepDistBytes() != -1U)
5123 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5124 unsigned WidestRegister =
5125 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
5126 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
5128 SmallVector<RegisterUsage, 8> RUs(VFs.size());
5129 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
5131 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5133 for (unsigned int i = 0; i < Index; ++i) {
5134 Instruction *I = IdxToInstr[i];
5135 // Ignore instructions that are never used within the loop.
5136 if (!Ends.count(I)) continue;
5138 // Skip ignored values.
5139 if (ValuesToIgnore.count(I))
5142 // Remove all of the instructions that end at this location.
5143 InstrList &List = TransposeEnds[i];
5144 for (unsigned int j = 0, e = List.size(); j < e; ++j)
5145 OpenIntervals.erase(List[j]);
5147 // For each VF find the maximum usage of registers.
5148 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
5149 // Count the number of live interals.
5150 unsigned RegUsage = 0;
5151 for (auto Inst : OpenIntervals) {
5153 DL.getTypeSizeInBits(Inst->getType()->getScalarType());
5154 RegUsage += std::max<unsigned>(1, VFs[j] * TypeSize / WidestRegister);
5156 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
5159 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "
5160 << OpenIntervals.size() << '\n');
5162 // Add the current instruction to the list of open intervals.
5163 OpenIntervals.insert(I);
5166 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
5167 unsigned Invariant = 0;
5168 for (auto Inst : LoopInvariants) {
5170 DL.getTypeSizeInBits(Inst->getType()->getScalarType());
5171 Invariant += std::max<unsigned>(1, VFs[i] * TypeSize / WidestRegister);
5174 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n');
5175 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n');
5176 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5177 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n');
5179 RU.LoopInvariantRegs = Invariant;
5180 RU.MaxLocalUsers = MaxUsages[i];
5187 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5191 for (Loop::block_iterator bb = TheLoop->block_begin(),
5192 be = TheLoop->block_end(); bb != be; ++bb) {
5193 unsigned BlockCost = 0;
5194 BasicBlock *BB = *bb;
5196 // For each instruction in the old loop.
5197 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5198 // Skip dbg intrinsics.
5199 if (isa<DbgInfoIntrinsic>(it))
5202 // Skip ignored values.
5203 if (ValuesToIgnore.count(&*it))
5206 unsigned C = getInstructionCost(&*it, VF);
5208 // Check if we should override the cost.
5209 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5210 C = ForceTargetInstructionCost;
5213 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5214 VF << " For instruction: " << *it << '\n');
5217 // We assume that if-converted blocks have a 50% chance of being executed.
5218 // When the code is scalar then some of the blocks are avoided due to CF.
5219 // When the code is vectorized we execute all code paths.
5220 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5229 /// \brief Check whether the address computation for a non-consecutive memory
5230 /// access looks like an unlikely candidate for being merged into the indexing
5233 /// We look for a GEP which has one index that is an induction variable and all
5234 /// other indices are loop invariant. If the stride of this access is also
5235 /// within a small bound we decide that this address computation can likely be
5236 /// merged into the addressing mode.
5237 /// In all other cases, we identify the address computation as complex.
5238 static bool isLikelyComplexAddressComputation(Value *Ptr,
5239 LoopVectorizationLegality *Legal,
5240 ScalarEvolution *SE,
5241 const Loop *TheLoop) {
5242 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5246 // We are looking for a gep with all loop invariant indices except for one
5247 // which should be an induction variable.
5248 unsigned NumOperands = Gep->getNumOperands();
5249 for (unsigned i = 1; i < NumOperands; ++i) {
5250 Value *Opd = Gep->getOperand(i);
5251 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5252 !Legal->isInductionVariable(Opd))
5256 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5257 // can likely be merged into the address computation.
5258 unsigned MaxMergeDistance = 64;
5260 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5264 // Check the step is constant.
5265 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5266 // Calculate the pointer stride and check if it is consecutive.
5267 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5271 const APInt &APStepVal = C->getValue()->getValue();
5273 // Huge step value - give up.
5274 if (APStepVal.getBitWidth() > 64)
5277 int64_t StepVal = APStepVal.getSExtValue();
5279 return StepVal > MaxMergeDistance;
5282 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5283 return Legal->hasStride(I->getOperand(0)) ||
5284 Legal->hasStride(I->getOperand(1));
5288 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5289 // If we know that this instruction will remain uniform, check the cost of
5290 // the scalar version.
5291 if (Legal->isUniformAfterVectorization(I))
5294 Type *RetTy = I->getType();
5295 if (VF > 1 && MinBWs.count(I))
5296 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
5297 Type *VectorTy = ToVectorTy(RetTy, VF);
5299 // TODO: We need to estimate the cost of intrinsic calls.
5300 switch (I->getOpcode()) {
5301 case Instruction::GetElementPtr:
5302 // We mark this instruction as zero-cost because the cost of GEPs in
5303 // vectorized code depends on whether the corresponding memory instruction
5304 // is scalarized or not. Therefore, we handle GEPs with the memory
5305 // instruction cost.
5307 case Instruction::Br: {
5308 return TTI.getCFInstrCost(I->getOpcode());
5310 case Instruction::PHI:
5311 //TODO: IF-converted IFs become selects.
5313 case Instruction::Add:
5314 case Instruction::FAdd:
5315 case Instruction::Sub:
5316 case Instruction::FSub:
5317 case Instruction::Mul:
5318 case Instruction::FMul:
5319 case Instruction::UDiv:
5320 case Instruction::SDiv:
5321 case Instruction::FDiv:
5322 case Instruction::URem:
5323 case Instruction::SRem:
5324 case Instruction::FRem:
5325 case Instruction::Shl:
5326 case Instruction::LShr:
5327 case Instruction::AShr:
5328 case Instruction::And:
5329 case Instruction::Or:
5330 case Instruction::Xor: {
5331 // Since we will replace the stride by 1 the multiplication should go away.
5332 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5334 // Certain instructions can be cheaper to vectorize if they have a constant
5335 // second vector operand. One example of this are shifts on x86.
5336 TargetTransformInfo::OperandValueKind Op1VK =
5337 TargetTransformInfo::OK_AnyValue;
5338 TargetTransformInfo::OperandValueKind Op2VK =
5339 TargetTransformInfo::OK_AnyValue;
5340 TargetTransformInfo::OperandValueProperties Op1VP =
5341 TargetTransformInfo::OP_None;
5342 TargetTransformInfo::OperandValueProperties Op2VP =
5343 TargetTransformInfo::OP_None;
5344 Value *Op2 = I->getOperand(1);
5346 // Check for a splat of a constant or for a non uniform vector of constants.
5347 if (isa<ConstantInt>(Op2)) {
5348 ConstantInt *CInt = cast<ConstantInt>(Op2);
5349 if (CInt && CInt->getValue().isPowerOf2())
5350 Op2VP = TargetTransformInfo::OP_PowerOf2;
5351 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5352 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5353 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5354 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5356 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5357 if (CInt && CInt->getValue().isPowerOf2())
5358 Op2VP = TargetTransformInfo::OP_PowerOf2;
5359 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5363 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5366 case Instruction::Select: {
5367 SelectInst *SI = cast<SelectInst>(I);
5368 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5369 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5370 Type *CondTy = SI->getCondition()->getType();
5372 CondTy = VectorType::get(CondTy, VF);
5374 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5376 case Instruction::ICmp:
5377 case Instruction::FCmp: {
5378 Type *ValTy = I->getOperand(0)->getType();
5379 if (VF > 1 && MinBWs.count(dyn_cast<Instruction>(I->getOperand(0))))
5380 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[I]);
5381 VectorTy = ToVectorTy(ValTy, VF);
5382 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5384 case Instruction::Store:
5385 case Instruction::Load: {
5386 StoreInst *SI = dyn_cast<StoreInst>(I);
5387 LoadInst *LI = dyn_cast<LoadInst>(I);
5388 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5390 VectorTy = ToVectorTy(ValTy, VF);
5392 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5393 unsigned AS = SI ? SI->getPointerAddressSpace() :
5394 LI->getPointerAddressSpace();
5395 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5396 // We add the cost of address computation here instead of with the gep
5397 // instruction because only here we know whether the operation is
5400 return TTI.getAddressComputationCost(VectorTy) +
5401 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5403 // For an interleaved access, calculate the total cost of the whole
5404 // interleave group.
5405 if (Legal->isAccessInterleaved(I)) {
5406 auto Group = Legal->getInterleavedAccessGroup(I);
5407 assert(Group && "Fail to get an interleaved access group.");
5409 // Only calculate the cost once at the insert position.
5410 if (Group->getInsertPos() != I)
5413 unsigned InterleaveFactor = Group->getFactor();
5415 VectorType::get(VectorTy->getVectorElementType(),
5416 VectorTy->getVectorNumElements() * InterleaveFactor);
5418 // Holds the indices of existing members in an interleaved load group.
5419 // An interleaved store group doesn't need this as it dones't allow gaps.
5420 SmallVector<unsigned, 4> Indices;
5422 for (unsigned i = 0; i < InterleaveFactor; i++)
5423 if (Group->getMember(i))
5424 Indices.push_back(i);
5427 // Calculate the cost of the whole interleaved group.
5428 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5429 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5430 Group->getAlignment(), AS);
5432 if (Group->isReverse())
5434 Group->getNumMembers() *
5435 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5437 // FIXME: The interleaved load group with a huge gap could be even more
5438 // expensive than scalar operations. Then we could ignore such group and
5439 // use scalar operations instead.
5443 // Scalarized loads/stores.
5444 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5445 bool Reverse = ConsecutiveStride < 0;
5446 const DataLayout &DL = I->getModule()->getDataLayout();
5447 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5448 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5449 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5450 bool IsComplexComputation =
5451 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5453 // The cost of extracting from the value vector and pointer vector.
5454 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5455 for (unsigned i = 0; i < VF; ++i) {
5456 // The cost of extracting the pointer operand.
5457 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5458 // In case of STORE, the cost of ExtractElement from the vector.
5459 // In case of LOAD, the cost of InsertElement into the returned
5461 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5462 Instruction::InsertElement,
5466 // The cost of the scalar loads/stores.
5467 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5468 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5473 // Wide load/stores.
5474 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5475 if (Legal->isMaskRequired(I))
5476 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5479 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5482 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5486 case Instruction::ZExt:
5487 case Instruction::SExt:
5488 case Instruction::FPToUI:
5489 case Instruction::FPToSI:
5490 case Instruction::FPExt:
5491 case Instruction::PtrToInt:
5492 case Instruction::IntToPtr:
5493 case Instruction::SIToFP:
5494 case Instruction::UIToFP:
5495 case Instruction::Trunc:
5496 case Instruction::FPTrunc:
5497 case Instruction::BitCast: {
5498 // We optimize the truncation of induction variable.
5499 // The cost of these is the same as the scalar operation.
5500 if (I->getOpcode() == Instruction::Trunc &&
5501 Legal->isInductionVariable(I->getOperand(0)))
5502 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5503 I->getOperand(0)->getType());
5505 Type *SrcScalarTy = I->getOperand(0)->getType();
5506 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
5507 if (VF > 1 && MinBWs.count(I)) {
5508 // This cast is going to be shrunk. This may remove the cast or it might
5509 // turn it into slightly different cast. For example, if MinBW == 16,
5510 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
5512 // Calculate the modified src and dest types.
5513 Type *MinVecTy = VectorTy;
5514 if (I->getOpcode() == Instruction::Trunc) {
5515 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
5516 VectorTy = largestIntegerVectorType(ToVectorTy(I->getType(), VF),
5518 } else if (I->getOpcode() == Instruction::ZExt ||
5519 I->getOpcode() == Instruction::SExt) {
5520 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
5521 VectorTy = smallestIntegerVectorType(ToVectorTy(I->getType(), VF),
5526 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5528 case Instruction::Call: {
5529 bool NeedToScalarize;
5530 CallInst *CI = cast<CallInst>(I);
5531 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5532 if (getIntrinsicIDForCall(CI, TLI))
5533 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5537 // We are scalarizing the instruction. Return the cost of the scalar
5538 // instruction, plus the cost of insert and extract into vector
5539 // elements, times the vector width.
5542 if (!RetTy->isVoidTy() && VF != 1) {
5543 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5545 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5548 // The cost of inserting the results plus extracting each one of the
5550 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5553 // The cost of executing VF copies of the scalar instruction. This opcode
5554 // is unknown. Assume that it is the same as 'mul'.
5555 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5561 char LoopVectorize::ID = 0;
5562 static const char lv_name[] = "Loop Vectorization";
5563 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5564 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5565 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
5566 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
5567 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
5568 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5569 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5570 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5571 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5572 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5573 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5574 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5575 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5576 INITIALIZE_PASS_DEPENDENCY(DemandedBits)
5577 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5580 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5581 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5585 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5586 // Check for a store.
5587 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5588 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5590 // Check for a load.
5591 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5592 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5598 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5599 bool IfPredicateStore) {
5600 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5601 // Holds vector parameters or scalars, in case of uniform vals.
5602 SmallVector<VectorParts, 4> Params;
5604 setDebugLocFromInst(Builder, Instr);
5606 // Find all of the vectorized parameters.
5607 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5608 Value *SrcOp = Instr->getOperand(op);
5610 // If we are accessing the old induction variable, use the new one.
5611 if (SrcOp == OldInduction) {
5612 Params.push_back(getVectorValue(SrcOp));
5616 // Try using previously calculated values.
5617 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5619 // If the src is an instruction that appeared earlier in the basic block
5620 // then it should already be vectorized.
5621 if (SrcInst && OrigLoop->contains(SrcInst)) {
5622 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5623 // The parameter is a vector value from earlier.
5624 Params.push_back(WidenMap.get(SrcInst));
5626 // The parameter is a scalar from outside the loop. Maybe even a constant.
5627 VectorParts Scalars;
5628 Scalars.append(UF, SrcOp);
5629 Params.push_back(Scalars);
5633 assert(Params.size() == Instr->getNumOperands() &&
5634 "Invalid number of operands");
5636 // Does this instruction return a value ?
5637 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5639 Value *UndefVec = IsVoidRetTy ? nullptr :
5640 UndefValue::get(Instr->getType());
5641 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5642 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5645 if (IfPredicateStore) {
5646 assert(Instr->getParent()->getSinglePredecessor() &&
5647 "Only support single predecessor blocks");
5648 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5649 Instr->getParent());
5652 // For each vector unroll 'part':
5653 for (unsigned Part = 0; Part < UF; ++Part) {
5654 // For each scalar that we create:
5656 // Start an "if (pred) a[i] = ..." block.
5657 Value *Cmp = nullptr;
5658 if (IfPredicateStore) {
5659 if (Cond[Part]->getType()->isVectorTy())
5661 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5662 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5663 ConstantInt::get(Cond[Part]->getType(), 1));
5666 Instruction *Cloned = Instr->clone();
5668 Cloned->setName(Instr->getName() + ".cloned");
5669 // Replace the operands of the cloned instructions with extracted scalars.
5670 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5671 Value *Op = Params[op][Part];
5672 Cloned->setOperand(op, Op);
5675 // Place the cloned scalar in the new loop.
5676 Builder.Insert(Cloned);
5678 // If the original scalar returns a value we need to place it in a vector
5679 // so that future users will be able to use it.
5681 VecResults[Part] = Cloned;
5684 if (IfPredicateStore)
5685 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
5690 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5691 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5692 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5694 return scalarizeInstruction(Instr, IfPredicateStore);
5697 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5701 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5705 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5706 // When unrolling and the VF is 1, we only need to add a simple scalar.
5707 Type *ITy = Val->getType();
5708 assert(!ITy->isVectorTy() && "Val must be a scalar");
5709 Constant *C = ConstantInt::get(ITy, StartIdx);
5710 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");