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/EquivalenceClasses.h"
52 #include "llvm/ADT/Hashing.h"
53 #include "llvm/ADT/MapVector.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/AliasAnalysis.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/LoopAccessAnalysis.h"
66 #include "llvm/Analysis/LoopInfo.h"
67 #include "llvm/Analysis/LoopIterator.h"
68 #include "llvm/Analysis/LoopPass.h"
69 #include "llvm/Analysis/ScalarEvolution.h"
70 #include "llvm/Analysis/ScalarEvolutionExpander.h"
71 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
72 #include "llvm/Analysis/TargetTransformInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/IR/Constants.h"
75 #include "llvm/IR/DataLayout.h"
76 #include "llvm/IR/DebugInfo.h"
77 #include "llvm/IR/DerivedTypes.h"
78 #include "llvm/IR/DiagnosticInfo.h"
79 #include "llvm/IR/Dominators.h"
80 #include "llvm/IR/Function.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/Instructions.h"
83 #include "llvm/IR/IntrinsicInst.h"
84 #include "llvm/IR/LLVMContext.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/IR/Type.h"
88 #include "llvm/IR/Value.h"
89 #include "llvm/IR/ValueHandle.h"
90 #include "llvm/IR/Verifier.h"
91 #include "llvm/Pass.h"
92 #include "llvm/Support/BranchProbability.h"
93 #include "llvm/Support/CommandLine.h"
94 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/raw_ostream.h"
96 #include "llvm/Transforms/Scalar.h"
97 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
98 #include "llvm/Transforms/Utils/Local.h"
99 #include "llvm/Analysis/VectorUtils.h"
100 #include "llvm/Transforms/Utils/LoopUtils.h"
105 using namespace llvm;
106 using namespace llvm::PatternMatch;
108 #define LV_NAME "loop-vectorize"
109 #define DEBUG_TYPE LV_NAME
111 STATISTIC(LoopsVectorized, "Number of loops vectorized");
112 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 static cl::opt<bool> EnableInterleavedMemAccesses(
142 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
143 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
145 /// Maximum factor for an interleaved memory access.
146 static cl::opt<unsigned> MaxInterleaveGroupFactor(
147 "max-interleave-group-factor", cl::Hidden,
148 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
151 /// We don't interleave loops with a known constant trip count below this
153 static const unsigned TinyTripCountInterleaveThreshold = 128;
155 static cl::opt<unsigned> ForceTargetNumScalarRegs(
156 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of scalar registers."));
159 static cl::opt<unsigned> ForceTargetNumVectorRegs(
160 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's number of vector registers."));
163 /// Maximum vectorization interleave count.
164 static const unsigned MaxInterleaveFactor = 16;
166 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
167 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max interleave factor for "
171 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
172 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
174 "vectorized loops."));
176 static cl::opt<unsigned> ForceTargetInstructionCost(
177 "force-target-instruction-cost", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's expected cost for "
179 "an instruction to a single constant value. Mostly "
180 "useful for getting consistent testing."));
182 static cl::opt<unsigned> SmallLoopCost(
183 "small-loop-cost", cl::init(20), cl::Hidden,
185 "The cost of a loop that is considered 'small' by the interleaver."));
187 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
188 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
189 cl::desc("Enable the use of the block frequency analysis to access PGO "
190 "heuristics minimizing code growth in cold regions and being more "
191 "aggressive in hot regions."));
193 // Runtime interleave loops for load/store throughput.
194 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
195 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
197 "Enable runtime interleaving until load/store ports are saturated"));
199 /// The number of stores in a loop that are allowed to need predication.
200 static cl::opt<unsigned> NumberOfStoresToPredicate(
201 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
202 cl::desc("Max number of stores to be predicated behind an if."));
204 static cl::opt<bool> EnableIndVarRegisterHeur(
205 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
206 cl::desc("Count the induction variable only once when interleaving"));
208 static cl::opt<bool> EnableCondStoresVectorization(
209 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
210 cl::desc("Enable if predication of stores during vectorization."));
212 static cl::opt<unsigned> MaxNestedScalarReductionIC(
213 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
214 cl::desc("The maximum interleave count to use when interleaving a scalar "
215 "reduction in a nested loop."));
217 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
218 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
219 cl::desc("The maximum allowed number of runtime memory checks with a "
220 "vectorize(enable) pragma."));
224 // Forward declarations.
225 class LoopVectorizeHints;
226 class LoopVectorizationLegality;
227 class LoopVectorizationCostModel;
228 class LoopVectorizationRequirements;
230 /// \brief This modifies LoopAccessReport to initialize message with
231 /// loop-vectorizer-specific part.
232 class VectorizationReport : public LoopAccessReport {
234 VectorizationReport(Instruction *I = nullptr)
235 : LoopAccessReport("loop not vectorized: ", I) {}
237 /// \brief This allows promotion of the loop-access analysis report into the
238 /// loop-vectorizer report. It modifies the message to add the
239 /// loop-vectorizer-specific part of the message.
240 explicit VectorizationReport(const LoopAccessReport &R)
241 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
245 /// A helper function for converting Scalar types to vector types.
246 /// If the incoming type is void, we return void. If the VF is 1, we return
248 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
249 if (Scalar->isVoidTy() || VF == 1)
251 return VectorType::get(Scalar, VF);
254 /// InnerLoopVectorizer vectorizes loops which contain only one basic
255 /// block to a specified vectorization factor (VF).
256 /// This class performs the widening of scalars into vectors, or multiple
257 /// scalars. This class also implements the following features:
258 /// * It inserts an epilogue loop for handling loops that don't have iteration
259 /// counts that are known to be a multiple of the vectorization factor.
260 /// * It handles the code generation for reduction variables.
261 /// * Scalarization (implementation using scalars) of un-vectorizable
263 /// InnerLoopVectorizer does not perform any vectorization-legality
264 /// checks, and relies on the caller to check for the different legality
265 /// aspects. The InnerLoopVectorizer relies on the
266 /// LoopVectorizationLegality class to provide information about the induction
267 /// and reduction variables that were found to a given vectorization factor.
268 class InnerLoopVectorizer {
270 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
271 DominatorTree *DT, const TargetLibraryInfo *TLI,
272 const TargetTransformInfo *TTI, unsigned VecWidth,
273 unsigned UnrollFactor)
274 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
275 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
276 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
277 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr),
278 AddedSafetyChecks(false) {}
280 // Perform the actual loop widening (vectorization).
281 void vectorize(LoopVectorizationLegality *L) {
283 // Create a new empty loop. Unlink the old loop and connect the new one.
285 // Widen each instruction in the old loop to a new one in the new loop.
286 // Use the Legality module to find the induction and reduction variables.
288 // Register the new loop and update the analysis passes.
292 // Return true if any runtime check is added.
293 bool IsSafetyChecksAdded() {
294 return AddedSafetyChecks;
297 virtual ~InnerLoopVectorizer() {}
300 /// A small list of PHINodes.
301 typedef SmallVector<PHINode*, 4> PhiVector;
302 /// When we unroll loops we have multiple vector values for each scalar.
303 /// This data structure holds the unrolled and vectorized values that
304 /// originated from one scalar instruction.
305 typedef SmallVector<Value*, 2> VectorParts;
307 // When we if-convert we need to create edge masks. We have to cache values
308 // so that we don't end up with exponential recursion/IR.
309 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
310 VectorParts> EdgeMaskCache;
312 /// \brief Add checks for strides that were assumed to be 1.
314 /// Returns the last check instruction and the first check instruction in the
315 /// pair as (first, last).
316 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
318 /// Create an empty loop, based on the loop ranges of the old loop.
319 void createEmptyLoop();
320 /// Create a new induction variable inside L.
321 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
322 Value *Step, Instruction *DL);
323 /// Copy and widen the instructions from the old loop.
324 virtual void vectorizeLoop();
326 /// \brief The Loop exit block may have single value PHI nodes where the
327 /// incoming value is 'Undef'. While vectorizing we only handled real values
328 /// that were defined inside the loop. Here we fix the 'undef case'.
332 /// A helper function that computes the predicate of the block BB, assuming
333 /// that the header block of the loop is set to True. It returns the *entry*
334 /// mask for the block BB.
335 VectorParts createBlockInMask(BasicBlock *BB);
336 /// A helper function that computes the predicate of the edge between SRC
338 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
340 /// A helper function to vectorize a single BB within the innermost loop.
341 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
343 /// Vectorize a single PHINode in a block. This method handles the induction
344 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
345 /// arbitrary length vectors.
346 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
347 unsigned UF, unsigned VF, PhiVector *PV);
349 /// Insert the new loop to the loop hierarchy and pass manager
350 /// and update the analysis passes.
351 void updateAnalysis();
353 /// This instruction is un-vectorizable. Implement it as a sequence
354 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
355 /// scalarized instruction behind an if block predicated on the control
356 /// dependence of the instruction.
357 virtual void scalarizeInstruction(Instruction *Instr,
358 bool IfPredicateStore=false);
360 /// Vectorize Load and Store instructions,
361 virtual void vectorizeMemoryInstruction(Instruction *Instr);
363 /// Create a broadcast instruction. This method generates a broadcast
364 /// instruction (shuffle) for loop invariant values and for the induction
365 /// value. If this is the induction variable then we extend it to N, N+1, ...
366 /// this is needed because each iteration in the loop corresponds to a SIMD
368 virtual Value *getBroadcastInstrs(Value *V);
370 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
371 /// to each vector element of Val. The sequence starts at StartIndex.
372 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
374 /// When we go over instructions in the basic block we rely on previous
375 /// values within the current basic block or on loop invariant values.
376 /// When we widen (vectorize) values we place them in the map. If the values
377 /// are not within the map, they have to be loop invariant, so we simply
378 /// broadcast them into a vector.
379 VectorParts &getVectorValue(Value *V);
381 /// Try to vectorize the interleaved access group that \p Instr belongs to.
382 void vectorizeInterleaveGroup(Instruction *Instr);
384 /// Generate a shuffle sequence that will reverse the vector Vec.
385 virtual Value *reverseVector(Value *Vec);
387 /// Returns (and creates if needed) the original loop trip count.
388 Value *getOrCreateTripCount(Loop *NewLoop);
390 /// Returns (and creates if needed) the trip count of the widened loop.
391 Value *getOrCreateVectorTripCount(Loop *NewLoop);
393 /// Emit a bypass check to see if the trip count would overflow, or we
394 /// wouldn't have enough iterations to execute one vector loop.
395 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
396 /// Emit a bypass check to see if the vector trip count is nonzero.
397 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
398 /// Emit bypass checks to check if strides we've assumed to be one really are.
399 void emitStrideChecks(Loop *L, BasicBlock *Bypass);
400 /// Emit bypass checks to check any memory assumptions we may have made.
401 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
403 /// This is a helper class that holds the vectorizer state. It maps scalar
404 /// instructions to vector instructions. When the code is 'unrolled' then
405 /// then a single scalar value is mapped to multiple vector parts. The parts
406 /// are stored in the VectorPart type.
408 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
410 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
412 /// \return True if 'Key' is saved in the Value Map.
413 bool has(Value *Key) const { return MapStorage.count(Key); }
415 /// Initializes a new entry in the map. Sets all of the vector parts to the
416 /// save value in 'Val'.
417 /// \return A reference to a vector with splat values.
418 VectorParts &splat(Value *Key, Value *Val) {
419 VectorParts &Entry = MapStorage[Key];
420 Entry.assign(UF, Val);
424 ///\return A reference to the value that is stored at 'Key'.
425 VectorParts &get(Value *Key) {
426 VectorParts &Entry = MapStorage[Key];
429 assert(Entry.size() == UF);
434 /// The unroll factor. Each entry in the map stores this number of vector
438 /// Map storage. We use std::map and not DenseMap because insertions to a
439 /// dense map invalidates its iterators.
440 std::map<Value *, VectorParts> MapStorage;
443 /// The original loop.
445 /// Scev analysis to use.
453 /// Target Library Info.
454 const TargetLibraryInfo *TLI;
455 /// Target Transform Info.
456 const TargetTransformInfo *TTI;
458 /// The vectorization SIMD factor to use. Each vector will have this many
463 /// The vectorization unroll factor to use. Each scalar is vectorized to this
464 /// many different vector instructions.
467 /// The builder that we use
470 // --- Vectorization state ---
472 /// The vector-loop preheader.
473 BasicBlock *LoopVectorPreHeader;
474 /// The scalar-loop preheader.
475 BasicBlock *LoopScalarPreHeader;
476 /// Middle Block between the vector and the scalar.
477 BasicBlock *LoopMiddleBlock;
478 ///The ExitBlock of the scalar loop.
479 BasicBlock *LoopExitBlock;
480 ///The vector loop body.
481 SmallVector<BasicBlock *, 4> LoopVectorBody;
482 ///The scalar loop body.
483 BasicBlock *LoopScalarBody;
484 /// A list of all bypass blocks. The first block is the entry of the loop.
485 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
487 /// The new Induction variable which was added to the new block.
489 /// The induction variable of the old basic block.
490 PHINode *OldInduction;
491 /// Maps scalars to widened vectors.
493 EdgeMaskCache MaskCache;
494 /// Trip count of the original loop.
496 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
497 Value *VectorTripCount;
499 LoopVectorizationLegality *Legal;
501 // Record whether runtime check is added.
502 bool AddedSafetyChecks;
505 class InnerLoopUnroller : public InnerLoopVectorizer {
507 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
508 DominatorTree *DT, const TargetLibraryInfo *TLI,
509 const TargetTransformInfo *TTI, unsigned UnrollFactor)
510 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
513 void scalarizeInstruction(Instruction *Instr,
514 bool IfPredicateStore = false) override;
515 void vectorizeMemoryInstruction(Instruction *Instr) override;
516 Value *getBroadcastInstrs(Value *V) override;
517 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
518 Value *reverseVector(Value *Vec) override;
521 /// \brief Look for a meaningful debug location on the instruction or it's
523 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
528 if (I->getDebugLoc() != Empty)
531 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
532 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
533 if (OpInst->getDebugLoc() != Empty)
540 /// \brief Set the debug location in the builder using the debug location in the
542 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
543 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
544 B.SetCurrentDebugLocation(Inst->getDebugLoc());
546 B.SetCurrentDebugLocation(DebugLoc());
550 /// \return string containing a file name and a line # for the given loop.
551 static std::string getDebugLocString(const Loop *L) {
554 raw_string_ostream OS(Result);
555 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
556 LoopDbgLoc.print(OS);
558 // Just print the module name.
559 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
566 /// \brief Propagate known metadata from one instruction to another.
567 static void propagateMetadata(Instruction *To, const Instruction *From) {
568 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
569 From->getAllMetadataOtherThanDebugLoc(Metadata);
571 for (auto M : Metadata) {
572 unsigned Kind = M.first;
574 // These are safe to transfer (this is safe for TBAA, even when we
575 // if-convert, because should that metadata have had a control dependency
576 // on the condition, and thus actually aliased with some other
577 // non-speculated memory access when the condition was false, this would be
578 // caught by the runtime overlap checks).
579 if (Kind != LLVMContext::MD_tbaa &&
580 Kind != LLVMContext::MD_alias_scope &&
581 Kind != LLVMContext::MD_noalias &&
582 Kind != LLVMContext::MD_fpmath &&
583 Kind != LLVMContext::MD_nontemporal)
586 To->setMetadata(Kind, M.second);
590 /// \brief Propagate known metadata from one instruction to a vector of others.
591 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
593 if (Instruction *I = dyn_cast<Instruction>(V))
594 propagateMetadata(I, From);
597 /// \brief The group of interleaved loads/stores sharing the same stride and
598 /// close to each other.
600 /// Each member in this group has an index starting from 0, and the largest
601 /// index should be less than interleaved factor, which is equal to the absolute
602 /// value of the access's stride.
604 /// E.g. An interleaved load group of factor 4:
605 /// for (unsigned i = 0; i < 1024; i+=4) {
606 /// a = A[i]; // Member of index 0
607 /// b = A[i+1]; // Member of index 1
608 /// d = A[i+3]; // Member of index 3
612 /// An interleaved store group of factor 4:
613 /// for (unsigned i = 0; i < 1024; i+=4) {
615 /// A[i] = a; // Member of index 0
616 /// A[i+1] = b; // Member of index 1
617 /// A[i+2] = c; // Member of index 2
618 /// A[i+3] = d; // Member of index 3
621 /// Note: the interleaved load group could have gaps (missing members), but
622 /// the interleaved store group doesn't allow gaps.
623 class InterleaveGroup {
625 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
626 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
627 assert(Align && "The alignment should be non-zero");
629 Factor = std::abs(Stride);
630 assert(Factor > 1 && "Invalid interleave factor");
632 Reverse = Stride < 0;
636 bool isReverse() const { return Reverse; }
637 unsigned getFactor() const { return Factor; }
638 unsigned getAlignment() const { return Align; }
639 unsigned getNumMembers() const { return Members.size(); }
641 /// \brief Try to insert a new member \p Instr with index \p Index and
642 /// alignment \p NewAlign. The index is related to the leader and it could be
643 /// negative if it is the new leader.
645 /// \returns false if the instruction doesn't belong to the group.
646 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
647 assert(NewAlign && "The new member's alignment should be non-zero");
649 int Key = Index + SmallestKey;
651 // Skip if there is already a member with the same index.
652 if (Members.count(Key))
655 if (Key > LargestKey) {
656 // The largest index is always less than the interleave factor.
657 if (Index >= static_cast<int>(Factor))
661 } else if (Key < SmallestKey) {
662 // The largest index is always less than the interleave factor.
663 if (LargestKey - Key >= static_cast<int>(Factor))
669 // It's always safe to select the minimum alignment.
670 Align = std::min(Align, NewAlign);
671 Members[Key] = Instr;
675 /// \brief Get the member with the given index \p Index
677 /// \returns nullptr if contains no such member.
678 Instruction *getMember(unsigned Index) const {
679 int Key = SmallestKey + Index;
680 if (!Members.count(Key))
683 return Members.find(Key)->second;
686 /// \brief Get the index for the given member. Unlike the key in the member
687 /// map, the index starts from 0.
688 unsigned getIndex(Instruction *Instr) const {
689 for (auto I : Members)
690 if (I.second == Instr)
691 return I.first - SmallestKey;
693 llvm_unreachable("InterleaveGroup contains no such member");
696 Instruction *getInsertPos() const { return InsertPos; }
697 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
700 unsigned Factor; // Interleave Factor.
703 DenseMap<int, Instruction *> Members;
707 // To avoid breaking dependences, vectorized instructions of an interleave
708 // group should be inserted at either the first load or the last store in
711 // E.g. %even = load i32 // Insert Position
712 // %add = add i32 %even // Use of %even
716 // %odd = add i32 // Def of %odd
717 // store i32 %odd // Insert Position
718 Instruction *InsertPos;
721 /// \brief Drive the analysis of interleaved memory accesses in the loop.
723 /// Use this class to analyze interleaved accesses only when we can vectorize
724 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
725 /// on interleaved accesses is unsafe.
727 /// The analysis collects interleave groups and records the relationships
728 /// between the member and the group in a map.
729 class InterleavedAccessInfo {
731 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
732 : SE(SE), TheLoop(L), DT(DT) {}
734 ~InterleavedAccessInfo() {
735 SmallSet<InterleaveGroup *, 4> DelSet;
736 // Avoid releasing a pointer twice.
737 for (auto &I : InterleaveGroupMap)
738 DelSet.insert(I.second);
739 for (auto *Ptr : DelSet)
743 /// \brief Analyze the interleaved accesses and collect them in interleave
744 /// groups. Substitute symbolic strides using \p Strides.
745 void analyzeInterleaving(const ValueToValueMap &Strides);
747 /// \brief Check if \p Instr belongs to any interleave group.
748 bool isInterleaved(Instruction *Instr) const {
749 return InterleaveGroupMap.count(Instr);
752 /// \brief Get the interleave group that \p Instr belongs to.
754 /// \returns nullptr if doesn't have such group.
755 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
756 if (InterleaveGroupMap.count(Instr))
757 return InterleaveGroupMap.find(Instr)->second;
766 /// Holds the relationships between the members and the interleave group.
767 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
769 /// \brief The descriptor for a strided memory access.
770 struct StrideDescriptor {
771 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
773 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
775 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
777 int Stride; // The access's stride. It is negative for a reverse access.
778 const SCEV *Scev; // The scalar expression of this access
779 unsigned Size; // The size of the memory object.
780 unsigned Align; // The alignment of this access.
783 /// \brief Create a new interleave group with the given instruction \p Instr,
784 /// stride \p Stride and alignment \p Align.
786 /// \returns the newly created interleave group.
787 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
789 assert(!InterleaveGroupMap.count(Instr) &&
790 "Already in an interleaved access group");
791 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
792 return InterleaveGroupMap[Instr];
795 /// \brief Release the group and remove all the relationships.
796 void releaseGroup(InterleaveGroup *Group) {
797 for (unsigned i = 0; i < Group->getFactor(); i++)
798 if (Instruction *Member = Group->getMember(i))
799 InterleaveGroupMap.erase(Member);
804 /// \brief Collect all the accesses with a constant stride in program order.
805 void collectConstStridedAccesses(
806 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
807 const ValueToValueMap &Strides);
810 /// Utility class for getting and setting loop vectorizer hints in the form
811 /// of loop metadata.
812 /// This class keeps a number of loop annotations locally (as member variables)
813 /// and can, upon request, write them back as metadata on the loop. It will
814 /// initially scan the loop for existing metadata, and will update the local
815 /// values based on information in the loop.
816 /// We cannot write all values to metadata, as the mere presence of some info,
817 /// for example 'force', means a decision has been made. So, we need to be
818 /// careful NOT to add them if the user hasn't specifically asked so.
819 class LoopVectorizeHints {
826 /// Hint - associates name and validation with the hint value.
829 unsigned Value; // This may have to change for non-numeric values.
832 Hint(const char * Name, unsigned Value, HintKind Kind)
833 : Name(Name), Value(Value), Kind(Kind) { }
835 bool validate(unsigned Val) {
838 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
840 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
848 /// Vectorization width.
850 /// Vectorization interleave factor.
852 /// Vectorization forced
855 /// Return the loop metadata prefix.
856 static StringRef Prefix() { return "llvm.loop."; }
860 FK_Undefined = -1, ///< Not selected.
861 FK_Disabled = 0, ///< Forcing disabled.
862 FK_Enabled = 1, ///< Forcing enabled.
865 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
866 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
868 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
869 Force("vectorize.enable", FK_Undefined, HK_FORCE),
871 // Populate values with existing loop metadata.
872 getHintsFromMetadata();
874 // force-vector-interleave overrides DisableInterleaving.
875 if (VectorizerParams::isInterleaveForced())
876 Interleave.Value = VectorizerParams::VectorizationInterleave;
878 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
879 << "LV: Interleaving disabled by the pass manager\n");
882 /// Mark the loop L as already vectorized by setting the width to 1.
883 void setAlreadyVectorized() {
884 Width.Value = Interleave.Value = 1;
885 Hint Hints[] = {Width, Interleave};
886 writeHintsToMetadata(Hints);
889 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
890 if (getForce() == LoopVectorizeHints::FK_Disabled) {
891 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
892 emitOptimizationRemarkAnalysis(F->getContext(),
893 vectorizeAnalysisPassName(), *F,
894 L->getStartLoc(), emitRemark());
898 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
899 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
900 emitOptimizationRemarkAnalysis(F->getContext(),
901 vectorizeAnalysisPassName(), *F,
902 L->getStartLoc(), emitRemark());
906 if (getWidth() == 1 && getInterleave() == 1) {
907 // FIXME: Add a separate metadata to indicate when the loop has already
908 // been vectorized instead of setting width and count to 1.
909 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
910 // FIXME: Add interleave.disable metadata. This will allow
911 // vectorize.disable to be used without disabling the pass and errors
912 // to differentiate between disabled vectorization and a width of 1.
913 emitOptimizationRemarkAnalysis(
914 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
915 "loop not vectorized: vectorization and interleaving are explicitly "
916 "disabled, or vectorize width and interleave count are both set to "
924 /// Dumps all the hint information.
925 std::string emitRemark() const {
926 VectorizationReport R;
927 if (Force.Value == LoopVectorizeHints::FK_Disabled)
928 R << "vectorization is explicitly disabled";
930 R << "use -Rpass-analysis=loop-vectorize for more info";
931 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
933 if (Width.Value != 0)
934 R << ", Vector Width=" << Width.Value;
935 if (Interleave.Value != 0)
936 R << ", Interleave Count=" << Interleave.Value;
944 unsigned getWidth() const { return Width.Value; }
945 unsigned getInterleave() const { return Interleave.Value; }
946 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
947 const char *vectorizeAnalysisPassName() const {
948 // If hints are provided that don't disable vectorization use the
949 // AlwaysPrint pass name to force the frontend to print the diagnostic.
952 if (getForce() == LoopVectorizeHints::FK_Disabled)
954 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
956 return DiagnosticInfo::AlwaysPrint;
959 bool allowReordering() const {
960 // When enabling loop hints are provided we allow the vectorizer to change
961 // the order of operations that is given by the scalar loop. This is not
962 // enabled by default because can be unsafe or inefficient. For example,
963 // reordering floating-point operations will change the way round-off
964 // error accumulates in the loop.
965 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
969 /// Find hints specified in the loop metadata and update local values.
970 void getHintsFromMetadata() {
971 MDNode *LoopID = TheLoop->getLoopID();
975 // First operand should refer to the loop id itself.
976 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
977 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
979 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
980 const MDString *S = nullptr;
981 SmallVector<Metadata *, 4> Args;
983 // The expected hint is either a MDString or a MDNode with the first
984 // operand a MDString.
985 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
986 if (!MD || MD->getNumOperands() == 0)
988 S = dyn_cast<MDString>(MD->getOperand(0));
989 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
990 Args.push_back(MD->getOperand(i));
992 S = dyn_cast<MDString>(LoopID->getOperand(i));
993 assert(Args.size() == 0 && "too many arguments for MDString");
999 // Check if the hint starts with the loop metadata prefix.
1000 StringRef Name = S->getString();
1001 if (Args.size() == 1)
1002 setHint(Name, Args[0]);
1006 /// Checks string hint with one operand and set value if valid.
1007 void setHint(StringRef Name, Metadata *Arg) {
1008 if (!Name.startswith(Prefix()))
1010 Name = Name.substr(Prefix().size(), StringRef::npos);
1012 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1014 unsigned Val = C->getZExtValue();
1016 Hint *Hints[] = {&Width, &Interleave, &Force};
1017 for (auto H : Hints) {
1018 if (Name == H->Name) {
1019 if (H->validate(Val))
1022 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1028 /// Create a new hint from name / value pair.
1029 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1030 LLVMContext &Context = TheLoop->getHeader()->getContext();
1031 Metadata *MDs[] = {MDString::get(Context, Name),
1032 ConstantAsMetadata::get(
1033 ConstantInt::get(Type::getInt32Ty(Context), V))};
1034 return MDNode::get(Context, MDs);
1037 /// Matches metadata with hint name.
1038 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1039 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1043 for (auto H : HintTypes)
1044 if (Name->getString().endswith(H.Name))
1049 /// Sets current hints into loop metadata, keeping other values intact.
1050 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1051 if (HintTypes.size() == 0)
1054 // Reserve the first element to LoopID (see below).
1055 SmallVector<Metadata *, 4> MDs(1);
1056 // If the loop already has metadata, then ignore the existing operands.
1057 MDNode *LoopID = TheLoop->getLoopID();
1059 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1060 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1061 // If node in update list, ignore old value.
1062 if (!matchesHintMetadataName(Node, HintTypes))
1063 MDs.push_back(Node);
1067 // Now, add the missing hints.
1068 for (auto H : HintTypes)
1069 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1071 // Replace current metadata node with new one.
1072 LLVMContext &Context = TheLoop->getHeader()->getContext();
1073 MDNode *NewLoopID = MDNode::get(Context, MDs);
1074 // Set operand 0 to refer to the loop id itself.
1075 NewLoopID->replaceOperandWith(0, NewLoopID);
1077 TheLoop->setLoopID(NewLoopID);
1080 /// The loop these hints belong to.
1081 const Loop *TheLoop;
1084 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1085 const LoopVectorizeHints &Hints,
1086 const LoopAccessReport &Message) {
1087 const char *Name = Hints.vectorizeAnalysisPassName();
1088 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1091 static void emitMissedWarning(Function *F, Loop *L,
1092 const LoopVectorizeHints &LH) {
1093 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1096 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1097 if (LH.getWidth() != 1)
1098 emitLoopVectorizeWarning(
1099 F->getContext(), *F, L->getStartLoc(),
1100 "failed explicitly specified loop vectorization");
1101 else if (LH.getInterleave() != 1)
1102 emitLoopInterleaveWarning(
1103 F->getContext(), *F, L->getStartLoc(),
1104 "failed explicitly specified loop interleaving");
1108 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1109 /// to what vectorization factor.
1110 /// This class does not look at the profitability of vectorization, only the
1111 /// legality. This class has two main kinds of checks:
1112 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1113 /// will change the order of memory accesses in a way that will change the
1114 /// correctness of the program.
1115 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1116 /// checks for a number of different conditions, such as the availability of a
1117 /// single induction variable, that all types are supported and vectorize-able,
1118 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1119 /// This class is also used by InnerLoopVectorizer for identifying
1120 /// induction variable and the different reduction variables.
1121 class LoopVectorizationLegality {
1123 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1124 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1125 Function *F, const TargetTransformInfo *TTI,
1126 LoopAccessAnalysis *LAA,
1127 LoopVectorizationRequirements *R,
1128 const LoopVectorizeHints *H)
1129 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1130 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1131 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1132 Requirements(R), Hints(H) {}
1134 /// ReductionList contains the reduction descriptors for all
1135 /// of the reductions that were found in the loop.
1136 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1138 /// InductionList saves induction variables and maps them to the
1139 /// induction descriptor.
1140 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1142 /// Returns true if it is legal to vectorize this loop.
1143 /// This does not mean that it is profitable to vectorize this
1144 /// loop, only that it is legal to do so.
1145 bool canVectorize();
1147 /// Returns the Induction variable.
1148 PHINode *getInduction() { return Induction; }
1150 /// Returns the reduction variables found in the loop.
1151 ReductionList *getReductionVars() { return &Reductions; }
1153 /// Returns the induction variables found in the loop.
1154 InductionList *getInductionVars() { return &Inductions; }
1156 /// Returns the widest induction type.
1157 Type *getWidestInductionType() { return WidestIndTy; }
1159 /// Returns True if V is an induction variable in this loop.
1160 bool isInductionVariable(const Value *V);
1162 /// Return true if the block BB needs to be predicated in order for the loop
1163 /// to be vectorized.
1164 bool blockNeedsPredication(BasicBlock *BB);
1166 /// Check if this pointer is consecutive when vectorizing. This happens
1167 /// when the last index of the GEP is the induction variable, or that the
1168 /// pointer itself is an induction variable.
1169 /// This check allows us to vectorize A[idx] into a wide load/store.
1171 /// 0 - Stride is unknown or non-consecutive.
1172 /// 1 - Address is consecutive.
1173 /// -1 - Address is consecutive, and decreasing.
1174 int isConsecutivePtr(Value *Ptr);
1176 /// Returns true if the value V is uniform within the loop.
1177 bool isUniform(Value *V);
1179 /// Returns true if this instruction will remain scalar after vectorization.
1180 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1182 /// Returns the information that we collected about runtime memory check.
1183 const RuntimePointerChecking *getRuntimePointerChecking() const {
1184 return LAI->getRuntimePointerChecking();
1187 const LoopAccessInfo *getLAI() const {
1191 /// \brief Check if \p Instr belongs to any interleaved access group.
1192 bool isAccessInterleaved(Instruction *Instr) {
1193 return InterleaveInfo.isInterleaved(Instr);
1196 /// \brief Get the interleaved access group that \p Instr belongs to.
1197 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1198 return InterleaveInfo.getInterleaveGroup(Instr);
1201 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1203 bool hasStride(Value *V) { return StrideSet.count(V); }
1204 bool mustCheckStrides() { return !StrideSet.empty(); }
1205 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1206 return StrideSet.begin();
1208 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1210 /// Returns true if the target machine supports masked store operation
1211 /// for the given \p DataType and kind of access to \p Ptr.
1212 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1213 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1215 /// Returns true if the target machine supports masked load operation
1216 /// for the given \p DataType and kind of access to \p Ptr.
1217 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1218 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1220 /// Returns true if vector representation of the instruction \p I
1222 bool isMaskRequired(const Instruction* I) {
1223 return (MaskedOp.count(I) != 0);
1225 unsigned getNumStores() const {
1226 return LAI->getNumStores();
1228 unsigned getNumLoads() const {
1229 return LAI->getNumLoads();
1231 unsigned getNumPredStores() const {
1232 return NumPredStores;
1235 /// Check if a single basic block loop is vectorizable.
1236 /// At this point we know that this is a loop with a constant trip count
1237 /// and we only need to check individual instructions.
1238 bool canVectorizeInstrs();
1240 /// When we vectorize loops we may change the order in which
1241 /// we read and write from memory. This method checks if it is
1242 /// legal to vectorize the code, considering only memory constrains.
1243 /// Returns true if the loop is vectorizable
1244 bool canVectorizeMemory();
1246 /// Return true if we can vectorize this loop using the IF-conversion
1248 bool canVectorizeWithIfConvert();
1250 /// Collect the variables that need to stay uniform after vectorization.
1251 void collectLoopUniforms();
1253 /// Return true if all of the instructions in the block can be speculatively
1254 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1255 /// and we know that we can read from them without segfault.
1256 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1258 /// \brief Collect memory access with loop invariant strides.
1260 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1262 void collectStridedAccess(Value *LoadOrStoreInst);
1264 /// Report an analysis message to assist the user in diagnosing loops that are
1265 /// not vectorized. These are handled as LoopAccessReport rather than
1266 /// VectorizationReport because the << operator of VectorizationReport returns
1267 /// LoopAccessReport.
1268 void emitAnalysis(const LoopAccessReport &Message) const {
1269 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1272 unsigned NumPredStores;
1274 /// The loop that we evaluate.
1277 ScalarEvolution *SE;
1278 /// Target Library Info.
1279 TargetLibraryInfo *TLI;
1281 Function *TheFunction;
1282 /// Target Transform Info
1283 const TargetTransformInfo *TTI;
1286 // LoopAccess analysis.
1287 LoopAccessAnalysis *LAA;
1288 // And the loop-accesses info corresponding to this loop. This pointer is
1289 // null until canVectorizeMemory sets it up.
1290 const LoopAccessInfo *LAI;
1292 /// The interleave access information contains groups of interleaved accesses
1293 /// with the same stride and close to each other.
1294 InterleavedAccessInfo InterleaveInfo;
1296 // --- vectorization state --- //
1298 /// Holds the integer induction variable. This is the counter of the
1301 /// Holds the reduction variables.
1302 ReductionList Reductions;
1303 /// Holds all of the induction variables that we found in the loop.
1304 /// Notice that inductions don't need to start at zero and that induction
1305 /// variables can be pointers.
1306 InductionList Inductions;
1307 /// Holds the widest induction type encountered.
1310 /// Allowed outside users. This holds the reduction
1311 /// vars which can be accessed from outside the loop.
1312 SmallPtrSet<Value*, 4> AllowedExit;
1313 /// This set holds the variables which are known to be uniform after
1315 SmallPtrSet<Instruction*, 4> Uniforms;
1317 /// Can we assume the absence of NaNs.
1318 bool HasFunNoNaNAttr;
1320 /// Vectorization requirements that will go through late-evaluation.
1321 LoopVectorizationRequirements *Requirements;
1323 /// Used to emit an analysis of any legality issues.
1324 const LoopVectorizeHints *Hints;
1326 ValueToValueMap Strides;
1327 SmallPtrSet<Value *, 8> StrideSet;
1329 /// While vectorizing these instructions we have to generate a
1330 /// call to the appropriate masked intrinsic
1331 SmallPtrSet<const Instruction*, 8> MaskedOp;
1334 /// LoopVectorizationCostModel - estimates the expected speedups due to
1336 /// In many cases vectorization is not profitable. This can happen because of
1337 /// a number of reasons. In this class we mainly attempt to predict the
1338 /// expected speedup/slowdowns due to the supported instruction set. We use the
1339 /// TargetTransformInfo to query the different backends for the cost of
1340 /// different operations.
1341 class LoopVectorizationCostModel {
1343 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1344 LoopVectorizationLegality *Legal,
1345 const TargetTransformInfo &TTI,
1346 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1347 const Function *F, const LoopVectorizeHints *Hints,
1348 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1349 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1350 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1352 /// Information about vectorization costs
1353 struct VectorizationFactor {
1354 unsigned Width; // Vector width with best cost
1355 unsigned Cost; // Cost of the loop with that width
1357 /// \return The most profitable vectorization factor and the cost of that VF.
1358 /// This method checks every power of two up to VF. If UserVF is not ZERO
1359 /// then this vectorization factor will be selected if vectorization is
1361 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1363 /// \return The size (in bits) of the widest type in the code that
1364 /// needs to be vectorized. We ignore values that remain scalar such as
1365 /// 64 bit loop indices.
1366 unsigned getWidestType();
1368 /// \return The desired interleave count.
1369 /// If interleave count has been specified by metadata it will be returned.
1370 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1371 /// are the selected vectorization factor and the cost of the selected VF.
1372 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1375 /// \return The most profitable unroll factor.
1376 /// This method finds the best unroll-factor based on register pressure and
1377 /// other parameters. VF and LoopCost are the selected vectorization factor
1378 /// and the cost of the selected VF.
1379 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1382 /// \brief A struct that represents some properties of the register usage
1384 struct RegisterUsage {
1385 /// Holds the number of loop invariant values that are used in the loop.
1386 unsigned LoopInvariantRegs;
1387 /// Holds the maximum number of concurrent live intervals in the loop.
1388 unsigned MaxLocalUsers;
1389 /// Holds the number of instructions in the loop.
1390 unsigned NumInstructions;
1393 /// \return information about the register usage of the loop.
1394 RegisterUsage calculateRegisterUsage();
1397 /// Returns the expected execution cost. The unit of the cost does
1398 /// not matter because we use the 'cost' units to compare different
1399 /// vector widths. The cost that is returned is *not* normalized by
1400 /// the factor width.
1401 unsigned expectedCost(unsigned VF);
1403 /// Returns the execution time cost of an instruction for a given vector
1404 /// width. Vector width of one means scalar.
1405 unsigned getInstructionCost(Instruction *I, unsigned VF);
1407 /// Returns whether the instruction is a load or store and will be a emitted
1408 /// as a vector operation.
1409 bool isConsecutiveLoadOrStore(Instruction *I);
1411 /// Report an analysis message to assist the user in diagnosing loops that are
1412 /// not vectorized. These are handled as LoopAccessReport rather than
1413 /// VectorizationReport because the << operator of VectorizationReport returns
1414 /// LoopAccessReport.
1415 void emitAnalysis(const LoopAccessReport &Message) const {
1416 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1419 /// The loop that we evaluate.
1422 ScalarEvolution *SE;
1423 /// Loop Info analysis.
1425 /// Vectorization legality.
1426 LoopVectorizationLegality *Legal;
1427 /// Vector target information.
1428 const TargetTransformInfo &TTI;
1429 /// Target Library Info.
1430 const TargetLibraryInfo *TLI;
1431 const Function *TheFunction;
1432 // Loop Vectorize Hint.
1433 const LoopVectorizeHints *Hints;
1434 // Values to ignore in the cost model.
1435 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1438 /// \brief This holds vectorization requirements that must be verified late in
1439 /// the process. The requirements are set by legalize and costmodel. Once
1440 /// vectorization has been determined to be possible and profitable the
1441 /// requirements can be verified by looking for metadata or compiler options.
1442 /// For example, some loops require FP commutativity which is only allowed if
1443 /// vectorization is explicitly specified or if the fast-math compiler option
1444 /// has been provided.
1445 /// Late evaluation of these requirements allows helpful diagnostics to be
1446 /// composed that tells the user what need to be done to vectorize the loop. For
1447 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1448 /// evaluation should be used only when diagnostics can generated that can be
1449 /// followed by a non-expert user.
1450 class LoopVectorizationRequirements {
1452 LoopVectorizationRequirements()
1453 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1455 void addUnsafeAlgebraInst(Instruction *I) {
1456 // First unsafe algebra instruction.
1457 if (!UnsafeAlgebraInst)
1458 UnsafeAlgebraInst = I;
1461 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1463 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1464 const char *Name = Hints.vectorizeAnalysisPassName();
1465 bool Failed = false;
1466 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1467 emitOptimizationRemarkAnalysisFPCommute(
1468 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1469 VectorizationReport() << "cannot prove it is safe to reorder "
1470 "floating-point operations");
1474 // Test if runtime memcheck thresholds are exceeded.
1475 bool PragmaThresholdReached =
1476 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1477 bool ThresholdReached =
1478 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1479 if ((ThresholdReached && !Hints.allowReordering()) ||
1480 PragmaThresholdReached) {
1481 emitOptimizationRemarkAnalysisAliasing(
1482 F->getContext(), Name, *F, L->getStartLoc(),
1483 VectorizationReport()
1484 << "cannot prove it is safe to reorder memory operations");
1485 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1493 unsigned NumRuntimePointerChecks;
1494 Instruction *UnsafeAlgebraInst;
1497 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1499 return V.push_back(&L);
1501 for (Loop *InnerL : L)
1502 addInnerLoop(*InnerL, V);
1505 /// The LoopVectorize Pass.
1506 struct LoopVectorize : public FunctionPass {
1507 /// Pass identification, replacement for typeid
1510 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1512 DisableUnrolling(NoUnrolling),
1513 AlwaysVectorize(AlwaysVectorize) {
1514 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1517 ScalarEvolution *SE;
1519 TargetTransformInfo *TTI;
1521 BlockFrequencyInfo *BFI;
1522 TargetLibraryInfo *TLI;
1524 AssumptionCache *AC;
1525 LoopAccessAnalysis *LAA;
1526 bool DisableUnrolling;
1527 bool AlwaysVectorize;
1529 BlockFrequency ColdEntryFreq;
1531 bool runOnFunction(Function &F) override {
1532 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1533 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1534 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1535 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1536 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1537 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1538 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1539 AA = &getAnalysis<AliasAnalysis>();
1540 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1541 LAA = &getAnalysis<LoopAccessAnalysis>();
1543 // Compute some weights outside of the loop over the loops. Compute this
1544 // using a BranchProbability to re-use its scaling math.
1545 const BranchProbability ColdProb(1, 5); // 20%
1546 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1549 // 1. the target claims to have no vector registers, and
1550 // 2. interleaving won't help ILP.
1552 // The second condition is necessary because, even if the target has no
1553 // vector registers, loop vectorization may still enable scalar
1555 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1558 // Build up a worklist of inner-loops to vectorize. This is necessary as
1559 // the act of vectorizing or partially unrolling a loop creates new loops
1560 // and can invalidate iterators across the loops.
1561 SmallVector<Loop *, 8> Worklist;
1564 addInnerLoop(*L, Worklist);
1566 LoopsAnalyzed += Worklist.size();
1568 // Now walk the identified inner loops.
1569 bool Changed = false;
1570 while (!Worklist.empty())
1571 Changed |= processLoop(Worklist.pop_back_val());
1573 // Process each loop nest in the function.
1577 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1578 SmallVector<Metadata *, 4> MDs;
1579 // Reserve first location for self reference to the LoopID metadata node.
1580 MDs.push_back(nullptr);
1581 bool IsUnrollMetadata = false;
1582 MDNode *LoopID = L->getLoopID();
1584 // First find existing loop unrolling disable metadata.
1585 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1586 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1588 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1590 S && S->getString().startswith("llvm.loop.unroll.disable");
1592 MDs.push_back(LoopID->getOperand(i));
1596 if (!IsUnrollMetadata) {
1597 // Add runtime unroll disable metadata.
1598 LLVMContext &Context = L->getHeader()->getContext();
1599 SmallVector<Metadata *, 1> DisableOperands;
1600 DisableOperands.push_back(
1601 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1602 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1603 MDs.push_back(DisableNode);
1604 MDNode *NewLoopID = MDNode::get(Context, MDs);
1605 // Set operand 0 to refer to the loop id itself.
1606 NewLoopID->replaceOperandWith(0, NewLoopID);
1607 L->setLoopID(NewLoopID);
1611 bool processLoop(Loop *L) {
1612 assert(L->empty() && "Only process inner loops.");
1615 const std::string DebugLocStr = getDebugLocString(L);
1618 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1619 << L->getHeader()->getParent()->getName() << "\" from "
1620 << DebugLocStr << "\n");
1622 LoopVectorizeHints Hints(L, DisableUnrolling);
1624 DEBUG(dbgs() << "LV: Loop hints:"
1626 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1628 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1630 : "?")) << " width=" << Hints.getWidth()
1631 << " unroll=" << Hints.getInterleave() << "\n");
1633 // Function containing loop
1634 Function *F = L->getHeader()->getParent();
1636 // Looking at the diagnostic output is the only way to determine if a loop
1637 // was vectorized (other than looking at the IR or machine code), so it
1638 // is important to generate an optimization remark for each loop. Most of
1639 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1640 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1641 // less verbose reporting vectorized loops and unvectorized loops that may
1642 // benefit from vectorization, respectively.
1644 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1645 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1649 // Check the loop for a trip count threshold:
1650 // do not vectorize loops with a tiny trip count.
1651 const unsigned TC = SE->getSmallConstantTripCount(L);
1652 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1653 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1654 << "This loop is not worth vectorizing.");
1655 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1656 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1658 DEBUG(dbgs() << "\n");
1659 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1660 << "vectorization is not beneficial "
1661 "and is not explicitly forced");
1666 // Check if it is legal to vectorize the loop.
1667 LoopVectorizationRequirements Requirements;
1668 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1669 &Requirements, &Hints);
1670 if (!LVL.canVectorize()) {
1671 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1672 emitMissedWarning(F, L, Hints);
1676 // Collect values we want to ignore in the cost model. This includes
1677 // type-promoting instructions we identified during reduction detection.
1678 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1679 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1680 for (auto &Reduction : *LVL.getReductionVars()) {
1681 RecurrenceDescriptor &RedDes = Reduction.second;
1682 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1683 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1686 // Use the cost model.
1687 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints,
1690 // Check the function attributes to find out if this function should be
1691 // optimized for size.
1692 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1695 // Compute the weighted frequency of this loop being executed and see if it
1696 // is less than 20% of the function entry baseline frequency. Note that we
1697 // always have a canonical loop here because we think we *can* vectorize.
1698 // FIXME: This is hidden behind a flag due to pervasive problems with
1699 // exactly what block frequency models.
1700 if (LoopVectorizeWithBlockFrequency) {
1701 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1702 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1703 LoopEntryFreq < ColdEntryFreq)
1707 // Check the function attributes to see if implicit floats are allowed.
1708 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1709 // an integer loop and the vector instructions selected are purely integer
1710 // vector instructions?
1711 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1712 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1713 "attribute is used.\n");
1716 VectorizationReport()
1717 << "loop not vectorized due to NoImplicitFloat attribute");
1718 emitMissedWarning(F, L, Hints);
1722 // Select the optimal vectorization factor.
1723 const LoopVectorizationCostModel::VectorizationFactor VF =
1724 CM.selectVectorizationFactor(OptForSize);
1726 // Select the interleave count.
1727 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1729 // Get user interleave count.
1730 unsigned UserIC = Hints.getInterleave();
1732 // Identify the diagnostic messages that should be produced.
1733 std::string VecDiagMsg, IntDiagMsg;
1734 bool VectorizeLoop = true, InterleaveLoop = true;
1736 if (Requirements.doesNotMeet(F, L, Hints)) {
1737 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1739 emitMissedWarning(F, L, Hints);
1743 if (VF.Width == 1) {
1744 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1746 "the cost-model indicates that vectorization is not beneficial";
1747 VectorizeLoop = false;
1750 if (IC == 1 && UserIC <= 1) {
1751 // Tell the user interleaving is not beneficial.
1752 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1754 "the cost-model indicates that interleaving is not beneficial";
1755 InterleaveLoop = false;
1758 " and is explicitly disabled or interleave count is set to 1";
1759 } else if (IC > 1 && UserIC == 1) {
1760 // Tell the user interleaving is beneficial, but it explicitly disabled.
1762 << "LV: Interleaving is beneficial but is explicitly disabled.");
1763 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1764 "but is explicitly disabled or interleave count is set to 1";
1765 InterleaveLoop = false;
1768 // Override IC if user provided an interleave count.
1769 IC = UserIC > 0 ? UserIC : IC;
1771 // Emit diagnostic messages, if any.
1772 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1773 if (!VectorizeLoop && !InterleaveLoop) {
1774 // Do not vectorize or interleaving the loop.
1775 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1776 L->getStartLoc(), VecDiagMsg);
1777 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1778 L->getStartLoc(), IntDiagMsg);
1780 } else if (!VectorizeLoop && InterleaveLoop) {
1781 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1782 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1783 L->getStartLoc(), VecDiagMsg);
1784 } else if (VectorizeLoop && !InterleaveLoop) {
1785 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1786 << DebugLocStr << '\n');
1787 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1788 L->getStartLoc(), IntDiagMsg);
1789 } else if (VectorizeLoop && InterleaveLoop) {
1790 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1791 << DebugLocStr << '\n');
1792 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1795 if (!VectorizeLoop) {
1796 assert(IC > 1 && "interleave count should not be 1 or 0");
1797 // If we decided that it is not legal to vectorize the loop then
1799 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1800 Unroller.vectorize(&LVL);
1802 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1803 Twine("interleaved loop (interleaved count: ") +
1806 // If we decided that it is *legal* to vectorize the loop then do it.
1807 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1811 // Add metadata to disable runtime unrolling scalar loop when there's no
1812 // runtime check about strides and memory. Because at this situation,
1813 // scalar loop is rarely used not worthy to be unrolled.
1814 if (!LB.IsSafetyChecksAdded())
1815 AddRuntimeUnrollDisableMetaData(L);
1817 // Report the vectorization decision.
1818 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1819 Twine("vectorized loop (vectorization width: ") +
1820 Twine(VF.Width) + ", interleaved count: " +
1824 // Mark the loop as already vectorized to avoid vectorizing again.
1825 Hints.setAlreadyVectorized();
1827 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1831 void getAnalysisUsage(AnalysisUsage &AU) const override {
1832 AU.addRequired<AssumptionCacheTracker>();
1833 AU.addRequiredID(LoopSimplifyID);
1834 AU.addRequiredID(LCSSAID);
1835 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1836 AU.addRequired<DominatorTreeWrapperPass>();
1837 AU.addRequired<LoopInfoWrapperPass>();
1838 AU.addRequired<ScalarEvolutionWrapperPass>();
1839 AU.addRequired<TargetTransformInfoWrapperPass>();
1840 AU.addRequired<AliasAnalysis>();
1841 AU.addRequired<LoopAccessAnalysis>();
1842 AU.addPreserved<LoopInfoWrapperPass>();
1843 AU.addPreserved<DominatorTreeWrapperPass>();
1844 AU.addPreserved<AliasAnalysis>();
1849 } // end anonymous namespace
1851 //===----------------------------------------------------------------------===//
1852 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1853 // LoopVectorizationCostModel.
1854 //===----------------------------------------------------------------------===//
1856 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1857 // We need to place the broadcast of invariant variables outside the loop.
1858 Instruction *Instr = dyn_cast<Instruction>(V);
1860 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1861 Instr->getParent()) != LoopVectorBody.end());
1862 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1864 // Place the code for broadcasting invariant variables in the new preheader.
1865 IRBuilder<>::InsertPointGuard Guard(Builder);
1867 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1869 // Broadcast the scalar into all locations in the vector.
1870 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1875 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1877 assert(Val->getType()->isVectorTy() && "Must be a vector");
1878 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1879 "Elem must be an integer");
1880 assert(Step->getType() == Val->getType()->getScalarType() &&
1881 "Step has wrong type");
1882 // Create the types.
1883 Type *ITy = Val->getType()->getScalarType();
1884 VectorType *Ty = cast<VectorType>(Val->getType());
1885 int VLen = Ty->getNumElements();
1886 SmallVector<Constant*, 8> Indices;
1888 // Create a vector of consecutive numbers from zero to VF.
1889 for (int i = 0; i < VLen; ++i)
1890 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1892 // Add the consecutive indices to the vector value.
1893 Constant *Cv = ConstantVector::get(Indices);
1894 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1895 Step = Builder.CreateVectorSplat(VLen, Step);
1896 assert(Step->getType() == Val->getType() && "Invalid step vec");
1897 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1898 // which can be found from the original scalar operations.
1899 Step = Builder.CreateMul(Cv, Step);
1900 return Builder.CreateAdd(Val, Step, "induction");
1903 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1904 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1905 // Make sure that the pointer does not point to structs.
1906 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1909 // If this value is a pointer induction variable we know it is consecutive.
1910 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1911 if (Phi && Inductions.count(Phi)) {
1912 InductionDescriptor II = Inductions[Phi];
1913 return II.getConsecutiveDirection();
1916 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1920 unsigned NumOperands = Gep->getNumOperands();
1921 Value *GpPtr = Gep->getPointerOperand();
1922 // If this GEP value is a consecutive pointer induction variable and all of
1923 // the indices are constant then we know it is consecutive. We can
1924 Phi = dyn_cast<PHINode>(GpPtr);
1925 if (Phi && Inductions.count(Phi)) {
1927 // Make sure that the pointer does not point to structs.
1928 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1929 if (GepPtrType->getElementType()->isAggregateType())
1932 // Make sure that all of the index operands are loop invariant.
1933 for (unsigned i = 1; i < NumOperands; ++i)
1934 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1937 InductionDescriptor II = Inductions[Phi];
1938 return II.getConsecutiveDirection();
1941 unsigned InductionOperand = getGEPInductionOperand(Gep);
1943 // Check that all of the gep indices are uniform except for our induction
1945 for (unsigned i = 0; i != NumOperands; ++i)
1946 if (i != InductionOperand &&
1947 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1950 // We can emit wide load/stores only if the last non-zero index is the
1951 // induction variable.
1952 const SCEV *Last = nullptr;
1953 if (!Strides.count(Gep))
1954 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1956 // Because of the multiplication by a stride we can have a s/zext cast.
1957 // We are going to replace this stride by 1 so the cast is safe to ignore.
1959 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1960 // %0 = trunc i64 %indvars.iv to i32
1961 // %mul = mul i32 %0, %Stride1
1962 // %idxprom = zext i32 %mul to i64 << Safe cast.
1963 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1965 Last = replaceSymbolicStrideSCEV(SE, Strides,
1966 Gep->getOperand(InductionOperand), Gep);
1967 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1969 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1973 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1974 const SCEV *Step = AR->getStepRecurrence(*SE);
1976 // The memory is consecutive because the last index is consecutive
1977 // and all other indices are loop invariant.
1980 if (Step->isAllOnesValue())
1987 bool LoopVectorizationLegality::isUniform(Value *V) {
1988 return LAI->isUniform(V);
1991 InnerLoopVectorizer::VectorParts&
1992 InnerLoopVectorizer::getVectorValue(Value *V) {
1993 assert(V != Induction && "The new induction variable should not be used.");
1994 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1996 // If we have a stride that is replaced by one, do it here.
1997 if (Legal->hasStride(V))
1998 V = ConstantInt::get(V->getType(), 1);
2000 // If we have this scalar in the map, return it.
2001 if (WidenMap.has(V))
2002 return WidenMap.get(V);
2004 // If this scalar is unknown, assume that it is a constant or that it is
2005 // loop invariant. Broadcast V and save the value for future uses.
2006 Value *B = getBroadcastInstrs(V);
2007 return WidenMap.splat(V, B);
2010 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2011 assert(Vec->getType()->isVectorTy() && "Invalid type");
2012 SmallVector<Constant*, 8> ShuffleMask;
2013 for (unsigned i = 0; i < VF; ++i)
2014 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2016 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2017 ConstantVector::get(ShuffleMask),
2021 // Get a mask to interleave \p NumVec vectors into a wide vector.
2022 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2023 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2024 // <0, 4, 1, 5, 2, 6, 3, 7>
2025 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2027 SmallVector<Constant *, 16> Mask;
2028 for (unsigned i = 0; i < VF; i++)
2029 for (unsigned j = 0; j < NumVec; j++)
2030 Mask.push_back(Builder.getInt32(j * VF + i));
2032 return ConstantVector::get(Mask);
2035 // Get the strided mask starting from index \p Start.
2036 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2037 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2038 unsigned Stride, unsigned VF) {
2039 SmallVector<Constant *, 16> Mask;
2040 for (unsigned i = 0; i < VF; i++)
2041 Mask.push_back(Builder.getInt32(Start + i * Stride));
2043 return ConstantVector::get(Mask);
2046 // Get a mask of two parts: The first part consists of sequential integers
2047 // starting from 0, The second part consists of UNDEFs.
2048 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2049 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2050 unsigned NumUndef) {
2051 SmallVector<Constant *, 16> Mask;
2052 for (unsigned i = 0; i < NumInt; i++)
2053 Mask.push_back(Builder.getInt32(i));
2055 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2056 for (unsigned i = 0; i < NumUndef; i++)
2057 Mask.push_back(Undef);
2059 return ConstantVector::get(Mask);
2062 // Concatenate two vectors with the same element type. The 2nd vector should
2063 // not have more elements than the 1st vector. If the 2nd vector has less
2064 // elements, extend it with UNDEFs.
2065 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2067 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2068 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2069 assert(VecTy1 && VecTy2 &&
2070 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2071 "Expect two vectors with the same element type");
2073 unsigned NumElts1 = VecTy1->getNumElements();
2074 unsigned NumElts2 = VecTy2->getNumElements();
2075 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2077 if (NumElts1 > NumElts2) {
2078 // Extend with UNDEFs.
2080 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2081 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2084 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2085 return Builder.CreateShuffleVector(V1, V2, Mask);
2088 // Concatenate vectors in the given list. All vectors have the same type.
2089 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2090 ArrayRef<Value *> InputList) {
2091 unsigned NumVec = InputList.size();
2092 assert(NumVec > 1 && "Should be at least two vectors");
2094 SmallVector<Value *, 8> ResList;
2095 ResList.append(InputList.begin(), InputList.end());
2097 SmallVector<Value *, 8> TmpList;
2098 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2099 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2100 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2101 "Only the last vector may have a different type");
2103 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2106 // Push the last vector if the total number of vectors is odd.
2107 if (NumVec % 2 != 0)
2108 TmpList.push_back(ResList[NumVec - 1]);
2111 NumVec = ResList.size();
2112 } while (NumVec > 1);
2117 // Try to vectorize the interleave group that \p Instr belongs to.
2119 // E.g. Translate following interleaved load group (factor = 3):
2120 // for (i = 0; i < N; i+=3) {
2121 // R = Pic[i]; // Member of index 0
2122 // G = Pic[i+1]; // Member of index 1
2123 // B = Pic[i+2]; // Member of index 2
2124 // ... // do something to R, G, B
2127 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2128 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2129 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2130 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2132 // Or translate following interleaved store group (factor = 3):
2133 // for (i = 0; i < N; i+=3) {
2134 // ... do something to R, G, B
2135 // Pic[i] = R; // Member of index 0
2136 // Pic[i+1] = G; // Member of index 1
2137 // Pic[i+2] = B; // Member of index 2
2140 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2141 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2142 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2143 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2144 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2145 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2146 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2147 assert(Group && "Fail to get an interleaved access group.");
2149 // Skip if current instruction is not the insert position.
2150 if (Instr != Group->getInsertPos())
2153 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2154 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2155 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2157 // Prepare for the vector type of the interleaved load/store.
2158 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2159 unsigned InterleaveFactor = Group->getFactor();
2160 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2161 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2163 // Prepare for the new pointers.
2164 setDebugLocFromInst(Builder, Ptr);
2165 VectorParts &PtrParts = getVectorValue(Ptr);
2166 SmallVector<Value *, 2> NewPtrs;
2167 unsigned Index = Group->getIndex(Instr);
2168 for (unsigned Part = 0; Part < UF; Part++) {
2169 // Extract the pointer for current instruction from the pointer vector. A
2170 // reverse access uses the pointer in the last lane.
2171 Value *NewPtr = Builder.CreateExtractElement(
2173 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2175 // Notice current instruction could be any index. Need to adjust the address
2176 // to the member of index 0.
2178 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2179 // b = A[i]; // Member of index 0
2180 // Current pointer is pointed to A[i+1], adjust it to A[i].
2182 // E.g. A[i+1] = a; // Member of index 1
2183 // A[i] = b; // Member of index 0
2184 // A[i+2] = c; // Member of index 2 (Current instruction)
2185 // Current pointer is pointed to A[i+2], adjust it to A[i].
2186 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2188 // Cast to the vector pointer type.
2189 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2192 setDebugLocFromInst(Builder, Instr);
2193 Value *UndefVec = UndefValue::get(VecTy);
2195 // Vectorize the interleaved load group.
2197 for (unsigned Part = 0; Part < UF; Part++) {
2198 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2199 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2201 for (unsigned i = 0; i < InterleaveFactor; i++) {
2202 Instruction *Member = Group->getMember(i);
2204 // Skip the gaps in the group.
2208 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2209 Value *StridedVec = Builder.CreateShuffleVector(
2210 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2212 // If this member has different type, cast the result type.
2213 if (Member->getType() != ScalarTy) {
2214 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2215 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2218 VectorParts &Entry = WidenMap.get(Member);
2220 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2223 propagateMetadata(NewLoadInstr, Instr);
2228 // The sub vector type for current instruction.
2229 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2231 // Vectorize the interleaved store group.
2232 for (unsigned Part = 0; Part < UF; Part++) {
2233 // Collect the stored vector from each member.
2234 SmallVector<Value *, 4> StoredVecs;
2235 for (unsigned i = 0; i < InterleaveFactor; i++) {
2236 // Interleaved store group doesn't allow a gap, so each index has a member
2237 Instruction *Member = Group->getMember(i);
2238 assert(Member && "Fail to get a member from an interleaved store group");
2241 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2242 if (Group->isReverse())
2243 StoredVec = reverseVector(StoredVec);
2245 // If this member has different type, cast it to an unified type.
2246 if (StoredVec->getType() != SubVT)
2247 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2249 StoredVecs.push_back(StoredVec);
2252 // Concatenate all vectors into a wide vector.
2253 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2255 // Interleave the elements in the wide vector.
2256 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2257 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2260 Instruction *NewStoreInstr =
2261 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2262 propagateMetadata(NewStoreInstr, Instr);
2266 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2267 // Attempt to issue a wide load.
2268 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2269 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2271 assert((LI || SI) && "Invalid Load/Store instruction");
2273 // Try to vectorize the interleave group if this access is interleaved.
2274 if (Legal->isAccessInterleaved(Instr))
2275 return vectorizeInterleaveGroup(Instr);
2277 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2278 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2279 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2280 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2281 // An alignment of 0 means target abi alignment. We need to use the scalar's
2282 // target abi alignment in such a case.
2283 const DataLayout &DL = Instr->getModule()->getDataLayout();
2285 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2286 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2287 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2288 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2290 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2291 !Legal->isMaskRequired(SI))
2292 return scalarizeInstruction(Instr, true);
2294 if (ScalarAllocatedSize != VectorElementSize)
2295 return scalarizeInstruction(Instr);
2297 // If the pointer is loop invariant or if it is non-consecutive,
2298 // scalarize the load.
2299 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2300 bool Reverse = ConsecutiveStride < 0;
2301 bool UniformLoad = LI && Legal->isUniform(Ptr);
2302 if (!ConsecutiveStride || UniformLoad)
2303 return scalarizeInstruction(Instr);
2305 Constant *Zero = Builder.getInt32(0);
2306 VectorParts &Entry = WidenMap.get(Instr);
2308 // Handle consecutive loads/stores.
2309 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2310 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2311 setDebugLocFromInst(Builder, Gep);
2312 Value *PtrOperand = Gep->getPointerOperand();
2313 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2314 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2316 // Create the new GEP with the new induction variable.
2317 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2318 Gep2->setOperand(0, FirstBasePtr);
2319 Gep2->setName("gep.indvar.base");
2320 Ptr = Builder.Insert(Gep2);
2322 setDebugLocFromInst(Builder, Gep);
2323 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2324 OrigLoop) && "Base ptr must be invariant");
2326 // The last index does not have to be the induction. It can be
2327 // consecutive and be a function of the index. For example A[I+1];
2328 unsigned NumOperands = Gep->getNumOperands();
2329 unsigned InductionOperand = getGEPInductionOperand(Gep);
2330 // Create the new GEP with the new induction variable.
2331 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2333 for (unsigned i = 0; i < NumOperands; ++i) {
2334 Value *GepOperand = Gep->getOperand(i);
2335 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2337 // Update last index or loop invariant instruction anchored in loop.
2338 if (i == InductionOperand ||
2339 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2340 assert((i == InductionOperand ||
2341 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2342 "Must be last index or loop invariant");
2344 VectorParts &GEPParts = getVectorValue(GepOperand);
2345 Value *Index = GEPParts[0];
2346 Index = Builder.CreateExtractElement(Index, Zero);
2347 Gep2->setOperand(i, Index);
2348 Gep2->setName("gep.indvar.idx");
2351 Ptr = Builder.Insert(Gep2);
2353 // Use the induction element ptr.
2354 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2355 setDebugLocFromInst(Builder, Ptr);
2356 VectorParts &PtrVal = getVectorValue(Ptr);
2357 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2360 VectorParts Mask = createBlockInMask(Instr->getParent());
2363 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2364 "We do not allow storing to uniform addresses");
2365 setDebugLocFromInst(Builder, SI);
2366 // We don't want to update the value in the map as it might be used in
2367 // another expression. So don't use a reference type for "StoredVal".
2368 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2370 for (unsigned Part = 0; Part < UF; ++Part) {
2371 // Calculate the pointer for the specific unroll-part.
2373 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2376 // If we store to reverse consecutive memory locations, then we need
2377 // to reverse the order of elements in the stored value.
2378 StoredVal[Part] = reverseVector(StoredVal[Part]);
2379 // If the address is consecutive but reversed, then the
2380 // wide store needs to start at the last vector element.
2381 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2382 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2383 Mask[Part] = reverseVector(Mask[Part]);
2386 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2387 DataTy->getPointerTo(AddressSpace));
2390 if (Legal->isMaskRequired(SI))
2391 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2394 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2395 propagateMetadata(NewSI, SI);
2401 assert(LI && "Must have a load instruction");
2402 setDebugLocFromInst(Builder, LI);
2403 for (unsigned Part = 0; Part < UF; ++Part) {
2404 // Calculate the pointer for the specific unroll-part.
2406 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2409 // If the address is consecutive but reversed, then the
2410 // wide load needs to start at the last vector element.
2411 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2412 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2413 Mask[Part] = reverseVector(Mask[Part]);
2417 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2418 DataTy->getPointerTo(AddressSpace));
2419 if (Legal->isMaskRequired(LI))
2420 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2421 UndefValue::get(DataTy),
2422 "wide.masked.load");
2424 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2425 propagateMetadata(NewLI, LI);
2426 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2430 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2431 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2432 // Holds vector parameters or scalars, in case of uniform vals.
2433 SmallVector<VectorParts, 4> Params;
2435 setDebugLocFromInst(Builder, Instr);
2437 // Find all of the vectorized parameters.
2438 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2439 Value *SrcOp = Instr->getOperand(op);
2441 // If we are accessing the old induction variable, use the new one.
2442 if (SrcOp == OldInduction) {
2443 Params.push_back(getVectorValue(SrcOp));
2447 // Try using previously calculated values.
2448 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2450 // If the src is an instruction that appeared earlier in the basic block,
2451 // then it should already be vectorized.
2452 if (SrcInst && OrigLoop->contains(SrcInst)) {
2453 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2454 // The parameter is a vector value from earlier.
2455 Params.push_back(WidenMap.get(SrcInst));
2457 // The parameter is a scalar from outside the loop. Maybe even a constant.
2458 VectorParts Scalars;
2459 Scalars.append(UF, SrcOp);
2460 Params.push_back(Scalars);
2464 assert(Params.size() == Instr->getNumOperands() &&
2465 "Invalid number of operands");
2467 // Does this instruction return a value ?
2468 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2470 Value *UndefVec = IsVoidRetTy ? nullptr :
2471 UndefValue::get(VectorType::get(Instr->getType(), VF));
2472 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2473 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2475 Instruction *InsertPt = Builder.GetInsertPoint();
2476 BasicBlock *IfBlock = Builder.GetInsertBlock();
2477 BasicBlock *CondBlock = nullptr;
2480 Loop *VectorLp = nullptr;
2481 if (IfPredicateStore) {
2482 assert(Instr->getParent()->getSinglePredecessor() &&
2483 "Only support single predecessor blocks");
2484 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2485 Instr->getParent());
2486 VectorLp = LI->getLoopFor(IfBlock);
2487 assert(VectorLp && "Must have a loop for this block");
2490 // For each vector unroll 'part':
2491 for (unsigned Part = 0; Part < UF; ++Part) {
2492 // For each scalar that we create:
2493 for (unsigned Width = 0; Width < VF; ++Width) {
2496 Value *Cmp = nullptr;
2497 if (IfPredicateStore) {
2498 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2499 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2500 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2501 LoopVectorBody.push_back(CondBlock);
2502 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2503 // Update Builder with newly created basic block.
2504 Builder.SetInsertPoint(InsertPt);
2507 Instruction *Cloned = Instr->clone();
2509 Cloned->setName(Instr->getName() + ".cloned");
2510 // Replace the operands of the cloned instructions with extracted scalars.
2511 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2512 Value *Op = Params[op][Part];
2513 // Param is a vector. Need to extract the right lane.
2514 if (Op->getType()->isVectorTy())
2515 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2516 Cloned->setOperand(op, Op);
2519 // Place the cloned scalar in the new loop.
2520 Builder.Insert(Cloned);
2522 // If the original scalar returns a value we need to place it in a vector
2523 // so that future users will be able to use it.
2525 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2526 Builder.getInt32(Width));
2528 if (IfPredicateStore) {
2529 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2530 LoopVectorBody.push_back(NewIfBlock);
2531 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2532 Builder.SetInsertPoint(InsertPt);
2533 ReplaceInstWithInst(IfBlock->getTerminator(),
2534 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2535 IfBlock = NewIfBlock;
2541 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2545 if (Instruction *I = dyn_cast<Instruction>(V))
2546 return I->getParent() == Loc->getParent() ? I : nullptr;
2550 std::pair<Instruction *, Instruction *>
2551 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2552 Instruction *tnullptr = nullptr;
2553 if (!Legal->mustCheckStrides())
2554 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2556 IRBuilder<> ChkBuilder(Loc);
2559 Value *Check = nullptr;
2560 Instruction *FirstInst = nullptr;
2561 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2562 SE = Legal->strides_end();
2564 Value *Ptr = stripIntegerCast(*SI);
2565 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2567 // Store the first instruction we create.
2568 FirstInst = getFirstInst(FirstInst, C, Loc);
2570 Check = ChkBuilder.CreateOr(Check, C);
2575 // We have to do this trickery because the IRBuilder might fold the check to a
2576 // constant expression in which case there is no Instruction anchored in a
2578 LLVMContext &Ctx = Loc->getContext();
2579 Instruction *TheCheck =
2580 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2581 ChkBuilder.Insert(TheCheck, "stride.not.one");
2582 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2584 return std::make_pair(FirstInst, TheCheck);
2587 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L,
2592 BasicBlock *Header = L->getHeader();
2593 BasicBlock *Latch = L->getLoopLatch();
2594 // As we're just creating this loop, it's possible no latch exists
2595 // yet. If so, use the header as this will be a single block loop.
2599 IRBuilder<> Builder(Header->getFirstInsertionPt());
2600 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2601 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2603 Builder.SetInsertPoint(Latch->getTerminator());
2605 // Create i+1 and fill the PHINode.
2606 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2607 Induction->addIncoming(Start, L->getLoopPreheader());
2608 Induction->addIncoming(Next, Latch);
2609 // Create the compare.
2610 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2611 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2613 // Now we have two terminators. Remove the old one from the block.
2614 Latch->getTerminator()->eraseFromParent();
2619 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2623 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2624 // Find the loop boundaries.
2625 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2626 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2628 Type *IdxTy = Legal->getWidestInductionType();
2630 // The exit count might have the type of i64 while the phi is i32. This can
2631 // happen if we have an induction variable that is sign extended before the
2632 // compare. The only way that we get a backedge taken count is that the
2633 // induction variable was signed and as such will not overflow. In such a case
2634 // truncation is legal.
2635 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2636 IdxTy->getPrimitiveSizeInBits())
2637 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2639 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2640 // Get the total trip count from the count by adding 1.
2641 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2642 SE->getConstant(BackedgeTakeCount->getType(), 1));
2644 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2646 // Expand the trip count and place the new instructions in the preheader.
2647 // Notice that the pre-header does not change, only the loop body.
2648 SCEVExpander Exp(*SE, DL, "induction");
2650 // Count holds the overall loop count (N).
2651 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2652 L->getLoopPreheader()->getTerminator());
2654 if (TripCount->getType()->isPointerTy())
2656 CastInst::CreatePointerCast(TripCount, IdxTy,
2657 "exitcount.ptrcnt.to.int",
2658 L->getLoopPreheader()->getTerminator());
2663 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2664 if (VectorTripCount)
2665 return VectorTripCount;
2667 Value *TC = getOrCreateTripCount(L);
2668 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2670 // Now we need to generate the expression for N - (N % VF), which is
2671 // the part that the vectorized body will execute.
2672 // The loop step is equal to the vectorization factor (num of SIMD elements)
2673 // times the unroll factor (num of SIMD instructions).
2674 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2675 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2676 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2678 return VectorTripCount;
2681 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
2682 BasicBlock *Bypass) {
2683 Value *Count = getOrCreateTripCount(L);
2684 BasicBlock *BB = L->getLoopPreheader();
2685 IRBuilder<> Builder(BB->getTerminator());
2687 // Generate code to check that the loop's trip count that we computed by
2688 // adding one to the backedge-taken count will not overflow.
2689 Value *CheckMinIters =
2690 Builder.CreateICmpULT(Count,
2691 ConstantInt::get(Count->getType(), VF * UF),
2694 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2695 "min.iters.checked");
2696 if (L->getParentLoop())
2697 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2698 ReplaceInstWithInst(BB->getTerminator(),
2699 BranchInst::Create(Bypass, NewBB, CheckMinIters));
2700 LoopBypassBlocks.push_back(BB);
2703 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
2704 BasicBlock *Bypass) {
2705 Value *TC = getOrCreateVectorTripCount(L);
2706 BasicBlock *BB = L->getLoopPreheader();
2707 IRBuilder<> Builder(BB->getTerminator());
2709 // Now, compare the new count to zero. If it is zero skip the vector loop and
2710 // jump to the scalar loop.
2711 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
2714 // Generate code to check that the loop's trip count that we computed by
2715 // adding one to the backedge-taken count will not overflow.
2716 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2718 if (L->getParentLoop())
2719 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2720 ReplaceInstWithInst(BB->getTerminator(),
2721 BranchInst::Create(Bypass, NewBB, Cmp));
2722 LoopBypassBlocks.push_back(BB);
2725 void InnerLoopVectorizer::emitStrideChecks(Loop *L,
2726 BasicBlock *Bypass) {
2727 BasicBlock *BB = L->getLoopPreheader();
2729 // Generate the code to check that the strides we assumed to be one are really
2730 // one. We want the new basic block to start at the first instruction in a
2731 // sequence of instructions that form a check.
2732 Instruction *StrideCheck;
2733 Instruction *FirstCheckInst;
2734 std::tie(FirstCheckInst, StrideCheck) = addStrideCheck(BB->getTerminator());
2738 // Create a new block containing the stride check.
2739 BB->setName("vector.stridecheck");
2740 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2741 if (L->getParentLoop())
2742 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2743 ReplaceInstWithInst(BB->getTerminator(),
2744 BranchInst::Create(Bypass, NewBB, StrideCheck));
2745 LoopBypassBlocks.push_back(BB);
2746 AddedSafetyChecks = true;
2749 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L,
2750 BasicBlock *Bypass) {
2751 BasicBlock *BB = L->getLoopPreheader();
2753 // Generate the code that checks in runtime if arrays overlap. We put the
2754 // checks into a separate block to make the more common case of few elements
2756 Instruction *FirstCheckInst;
2757 Instruction *MemRuntimeCheck;
2758 std::tie(FirstCheckInst, MemRuntimeCheck) =
2759 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
2760 if (!MemRuntimeCheck)
2763 // Create a new block containing the memory check.
2764 BB->setName("vector.memcheck");
2765 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2766 if (L->getParentLoop())
2767 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2768 ReplaceInstWithInst(BB->getTerminator(),
2769 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
2770 LoopBypassBlocks.push_back(BB);
2771 AddedSafetyChecks = true;
2775 void InnerLoopVectorizer::createEmptyLoop() {
2777 In this function we generate a new loop. The new loop will contain
2778 the vectorized instructions while the old loop will continue to run the
2781 [ ] <-- loop iteration number check.
2784 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2787 || [ ] <-- vector pre header.
2791 | [ ]_| <-- vector loop.
2794 | -[ ] <--- middle-block.
2797 -|- >[ ] <--- new preheader.
2801 | [ ]_| <-- old scalar loop to handle remainder.
2804 >[ ] <-- exit block.
2808 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2809 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2810 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2811 assert(VectorPH && "Invalid loop structure");
2812 assert(ExitBlock && "Must have an exit block");
2814 // Some loops have a single integer induction variable, while other loops
2815 // don't. One example is c++ iterators that often have multiple pointer
2816 // induction variables. In the code below we also support a case where we
2817 // don't have a single induction variable.
2819 // We try to obtain an induction variable from the original loop as hard
2820 // as possible. However if we don't find one that:
2822 // - counts from zero, stepping by one
2823 // - is the size of the widest induction variable type
2824 // then we create a new one.
2825 OldInduction = Legal->getInduction();
2826 Type *IdxTy = Legal->getWidestInductionType();
2828 // Split the single block loop into the two loop structure described above.
2829 BasicBlock *VecBody =
2830 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2831 BasicBlock *MiddleBlock =
2832 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2833 BasicBlock *ScalarPH =
2834 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2836 // Create and register the new vector loop.
2837 Loop* Lp = new Loop();
2838 Loop *ParentLoop = OrigLoop->getParentLoop();
2840 // Insert the new loop into the loop nest and register the new basic blocks
2841 // before calling any utilities such as SCEV that require valid LoopInfo.
2843 ParentLoop->addChildLoop(Lp);
2844 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2845 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2847 LI->addTopLevelLoop(Lp);
2849 Lp->addBasicBlockToLoop(VecBody, *LI);
2851 // Find the loop boundaries.
2852 Value *Count = getOrCreateTripCount(Lp);
2854 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2856 // We need to test whether the backedge-taken count is uint##_max. Adding one
2857 // to it will cause overflow and an incorrect loop trip count in the vector
2858 // body. In case of overflow we want to directly jump to the scalar remainder
2860 emitMinimumIterationCountCheck(Lp, ScalarPH);
2861 // Now, compare the new count to zero. If it is zero skip the vector loop and
2862 // jump to the scalar loop.
2863 emitVectorLoopEnteredCheck(Lp, ScalarPH);
2864 // Generate the code to check that the strides we assumed to be one are really
2865 // one. We want the new basic block to start at the first instruction in a
2866 // sequence of instructions that form a check.
2867 emitStrideChecks(Lp, ScalarPH);
2868 // Generate the code that checks in runtime if arrays overlap. We put the
2869 // checks into a separate block to make the more common case of few elements
2871 emitMemRuntimeChecks(Lp, ScalarPH);
2873 // Generate the induction variable.
2874 // The loop step is equal to the vectorization factor (num of SIMD elements)
2875 // times the unroll factor (num of SIMD instructions).
2876 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
2877 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2879 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
2880 getDebugLocFromInstOrOperands(OldInduction));
2882 // We are going to resume the execution of the scalar loop.
2883 // Go over all of the induction variables that we found and fix the
2884 // PHIs that are left in the scalar version of the loop.
2885 // The starting values of PHI nodes depend on the counter of the last
2886 // iteration in the vectorized loop.
2887 // If we come from a bypass edge then we need to start from the original
2890 // This variable saves the new starting index for the scalar loop. It is used
2891 // to test if there are any tail iterations left once the vector loop has
2893 LoopVectorizationLegality::InductionList::iterator I, E;
2894 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2895 for (I = List->begin(), E = List->end(); I != E; ++I) {
2896 PHINode *OrigPhi = I->first;
2897 InductionDescriptor II = I->second;
2899 // Create phi nodes to merge from the backedge-taken check block.
2900 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2902 ScalarPH->getTerminator());
2904 if (OrigPhi == OldInduction) {
2905 // We know what the end value is.
2906 EndValue = CountRoundDown;
2908 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
2909 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
2910 II.getStepValue()->getType(),
2912 EndValue = II.transform(B, CRD);
2913 EndValue->setName("ind.end");
2916 // The new PHI merges the original incoming value, in case of a bypass,
2917 // or the value at the end of the vectorized loop.
2918 BCResumeVal->addIncoming(EndValue, MiddleBlock);
2920 // Fix the scalar body counter (PHI node).
2921 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2923 // The old induction's phi node in the scalar body needs the truncated
2925 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2926 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2927 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2930 // Add a check in the middle block to see if we have completed
2931 // all of the iterations in the first vector loop.
2932 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2933 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
2934 CountRoundDown, "cmp.n",
2935 MiddleBlock->getTerminator());
2936 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2937 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2939 // Get ready to start creating new instructions into the vectorized body.
2940 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2943 LoopVectorPreHeader = Lp->getLoopPreheader();
2944 LoopScalarPreHeader = ScalarPH;
2945 LoopMiddleBlock = MiddleBlock;
2946 LoopExitBlock = ExitBlock;
2947 LoopVectorBody.push_back(VecBody);
2948 LoopScalarBody = OldBasicBlock;
2950 LoopVectorizeHints Hints(Lp, true);
2951 Hints.setAlreadyVectorized();
2955 struct CSEDenseMapInfo {
2956 static bool canHandle(Instruction *I) {
2957 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2958 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2960 static inline Instruction *getEmptyKey() {
2961 return DenseMapInfo<Instruction *>::getEmptyKey();
2963 static inline Instruction *getTombstoneKey() {
2964 return DenseMapInfo<Instruction *>::getTombstoneKey();
2966 static unsigned getHashValue(Instruction *I) {
2967 assert(canHandle(I) && "Unknown instruction!");
2968 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2969 I->value_op_end()));
2971 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2972 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2973 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2975 return LHS->isIdenticalTo(RHS);
2980 /// \brief Check whether this block is a predicated block.
2981 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2982 /// = ...; " blocks. We start with one vectorized basic block. For every
2983 /// conditional block we split this vectorized block. Therefore, every second
2984 /// block will be a predicated one.
2985 static bool isPredicatedBlock(unsigned BlockNum) {
2986 return BlockNum % 2;
2989 ///\brief Perform cse of induction variable instructions.
2990 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2991 // Perform simple cse.
2992 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2993 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2994 BasicBlock *BB = BBs[i];
2995 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2996 Instruction *In = I++;
2998 if (!CSEDenseMapInfo::canHandle(In))
3001 // Check if we can replace this instruction with any of the
3002 // visited instructions.
3003 if (Instruction *V = CSEMap.lookup(In)) {
3004 In->replaceAllUsesWith(V);
3005 In->eraseFromParent();
3008 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3009 // ...;" blocks for predicated stores. Every second block is a predicated
3011 if (isPredicatedBlock(i))
3019 /// \brief Adds a 'fast' flag to floating point operations.
3020 static Value *addFastMathFlag(Value *V) {
3021 if (isa<FPMathOperator>(V)){
3022 FastMathFlags Flags;
3023 Flags.setUnsafeAlgebra();
3024 cast<Instruction>(V)->setFastMathFlags(Flags);
3029 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3030 /// the result needs to be inserted and/or extracted from vectors.
3031 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3032 const TargetTransformInfo &TTI) {
3036 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3039 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3041 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3043 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3049 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3050 // Return the cost of the instruction, including scalarization overhead if it's
3051 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3052 // i.e. either vector version isn't available, or is too expensive.
3053 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3054 const TargetTransformInfo &TTI,
3055 const TargetLibraryInfo *TLI,
3056 bool &NeedToScalarize) {
3057 Function *F = CI->getCalledFunction();
3058 StringRef FnName = CI->getCalledFunction()->getName();
3059 Type *ScalarRetTy = CI->getType();
3060 SmallVector<Type *, 4> Tys, ScalarTys;
3061 for (auto &ArgOp : CI->arg_operands())
3062 ScalarTys.push_back(ArgOp->getType());
3064 // Estimate cost of scalarized vector call. The source operands are assumed
3065 // to be vectors, so we need to extract individual elements from there,
3066 // execute VF scalar calls, and then gather the result into the vector return
3068 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3070 return ScalarCallCost;
3072 // Compute corresponding vector type for return value and arguments.
3073 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3074 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3075 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3077 // Compute costs of unpacking argument values for the scalar calls and
3078 // packing the return values to a vector.
3079 unsigned ScalarizationCost =
3080 getScalarizationOverhead(RetTy, true, false, TTI);
3081 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3082 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3084 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3086 // If we can't emit a vector call for this function, then the currently found
3087 // cost is the cost we need to return.
3088 NeedToScalarize = true;
3089 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3092 // If the corresponding vector cost is cheaper, return its cost.
3093 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3094 if (VectorCallCost < Cost) {
3095 NeedToScalarize = false;
3096 return VectorCallCost;
3101 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3102 // factor VF. Return the cost of the instruction, including scalarization
3103 // overhead if it's needed.
3104 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3105 const TargetTransformInfo &TTI,
3106 const TargetLibraryInfo *TLI) {
3107 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3108 assert(ID && "Expected intrinsic call!");
3110 Type *RetTy = ToVectorTy(CI->getType(), VF);
3111 SmallVector<Type *, 4> Tys;
3112 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3113 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3115 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3118 void InnerLoopVectorizer::vectorizeLoop() {
3119 //===------------------------------------------------===//
3121 // Notice: any optimization or new instruction that go
3122 // into the code below should be also be implemented in
3125 //===------------------------------------------------===//
3126 Constant *Zero = Builder.getInt32(0);
3128 // In order to support reduction variables we need to be able to vectorize
3129 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3130 // stages. First, we create a new vector PHI node with no incoming edges.
3131 // We use this value when we vectorize all of the instructions that use the
3132 // PHI. Next, after all of the instructions in the block are complete we
3133 // add the new incoming edges to the PHI. At this point all of the
3134 // instructions in the basic block are vectorized, so we can use them to
3135 // construct the PHI.
3136 PhiVector RdxPHIsToFix;
3138 // Scan the loop in a topological order to ensure that defs are vectorized
3140 LoopBlocksDFS DFS(OrigLoop);
3143 // Vectorize all of the blocks in the original loop.
3144 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3145 be = DFS.endRPO(); bb != be; ++bb)
3146 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3148 // At this point every instruction in the original loop is widened to
3149 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3150 // that we vectorized. The PHI nodes are currently empty because we did
3151 // not want to introduce cycles. Notice that the remaining PHI nodes
3152 // that we need to fix are reduction variables.
3154 // Create the 'reduced' values for each of the induction vars.
3155 // The reduced values are the vector values that we scalarize and combine
3156 // after the loop is finished.
3157 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3159 PHINode *RdxPhi = *it;
3160 assert(RdxPhi && "Unable to recover vectorized PHI");
3162 // Find the reduction variable descriptor.
3163 assert(Legal->getReductionVars()->count(RdxPhi) &&
3164 "Unable to find the reduction variable");
3165 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3167 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3168 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3169 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3170 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3171 RdxDesc.getMinMaxRecurrenceKind();
3172 setDebugLocFromInst(Builder, ReductionStartValue);
3174 // We need to generate a reduction vector from the incoming scalar.
3175 // To do so, we need to generate the 'identity' vector and override
3176 // one of the elements with the incoming scalar reduction. We need
3177 // to do it in the vector-loop preheader.
3178 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3180 // This is the vector-clone of the value that leaves the loop.
3181 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3182 Type *VecTy = VectorExit[0]->getType();
3184 // Find the reduction identity variable. Zero for addition, or, xor,
3185 // one for multiplication, -1 for And.
3188 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3189 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3190 // MinMax reduction have the start value as their identify.
3192 VectorStart = Identity = ReductionStartValue;
3194 VectorStart = Identity =
3195 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3198 // Handle other reduction kinds:
3199 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3200 RK, VecTy->getScalarType());
3203 // This vector is the Identity vector where the first element is the
3204 // incoming scalar reduction.
3205 VectorStart = ReductionStartValue;
3207 Identity = ConstantVector::getSplat(VF, Iden);
3209 // This vector is the Identity vector where the first element is the
3210 // incoming scalar reduction.
3212 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3216 // Fix the vector-loop phi.
3218 // Reductions do not have to start at zero. They can start with
3219 // any loop invariant values.
3220 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3221 BasicBlock *Latch = OrigLoop->getLoopLatch();
3222 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3223 VectorParts &Val = getVectorValue(LoopVal);
3224 for (unsigned part = 0; part < UF; ++part) {
3225 // Make sure to add the reduction stat value only to the
3226 // first unroll part.
3227 Value *StartVal = (part == 0) ? VectorStart : Identity;
3228 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3229 LoopVectorPreHeader);
3230 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3231 LoopVectorBody.back());
3234 // Before each round, move the insertion point right between
3235 // the PHIs and the values we are going to write.
3236 // This allows us to write both PHINodes and the extractelement
3238 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3240 VectorParts RdxParts = getVectorValue(LoopExitInst);
3241 setDebugLocFromInst(Builder, LoopExitInst);
3243 // If the vector reduction can be performed in a smaller type, we truncate
3244 // then extend the loop exit value to enable InstCombine to evaluate the
3245 // entire expression in the smaller type.
3246 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3247 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3248 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3249 for (unsigned part = 0; part < UF; ++part) {
3250 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3251 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3252 : Builder.CreateZExt(Trunc, VecTy);
3253 for (Value::user_iterator UI = RdxParts[part]->user_begin();
3254 UI != RdxParts[part]->user_end();)
3256 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
3257 RdxParts[part] = Extnd;
3262 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3263 for (unsigned part = 0; part < UF; ++part)
3264 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3267 // Reduce all of the unrolled parts into a single vector.
3268 Value *ReducedPartRdx = RdxParts[0];
3269 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3270 setDebugLocFromInst(Builder, ReducedPartRdx);
3271 for (unsigned part = 1; part < UF; ++part) {
3272 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3273 // Floating point operations had to be 'fast' to enable the reduction.
3274 ReducedPartRdx = addFastMathFlag(
3275 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3276 ReducedPartRdx, "bin.rdx"));
3278 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3279 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3283 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3284 // and vector ops, reducing the set of values being computed by half each
3286 assert(isPowerOf2_32(VF) &&
3287 "Reduction emission only supported for pow2 vectors!");
3288 Value *TmpVec = ReducedPartRdx;
3289 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3290 for (unsigned i = VF; i != 1; i >>= 1) {
3291 // Move the upper half of the vector to the lower half.
3292 for (unsigned j = 0; j != i/2; ++j)
3293 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3295 // Fill the rest of the mask with undef.
3296 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3297 UndefValue::get(Builder.getInt32Ty()));
3300 Builder.CreateShuffleVector(TmpVec,
3301 UndefValue::get(TmpVec->getType()),
3302 ConstantVector::get(ShuffleMask),
3305 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3306 // Floating point operations had to be 'fast' to enable the reduction.
3307 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3308 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3310 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3314 // The result is in the first element of the vector.
3315 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3316 Builder.getInt32(0));
3318 // If the reduction can be performed in a smaller type, we need to extend
3319 // the reduction to the wider type before we branch to the original loop.
3320 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3323 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3324 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3327 // Create a phi node that merges control-flow from the backedge-taken check
3328 // block and the middle block.
3329 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3330 LoopScalarPreHeader->getTerminator());
3331 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
3332 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
3333 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3335 // Now, we need to fix the users of the reduction variable
3336 // inside and outside of the scalar remainder loop.
3337 // We know that the loop is in LCSSA form. We need to update the
3338 // PHI nodes in the exit blocks.
3339 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3340 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3341 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3342 if (!LCSSAPhi) break;
3344 // All PHINodes need to have a single entry edge, or two if
3345 // we already fixed them.
3346 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3348 // We found our reduction value exit-PHI. Update it with the
3349 // incoming bypass edge.
3350 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3351 // Add an edge coming from the bypass.
3352 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3355 }// end of the LCSSA phi scan.
3357 // Fix the scalar loop reduction variable with the incoming reduction sum
3358 // from the vector body and from the backedge value.
3359 int IncomingEdgeBlockIdx =
3360 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3361 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3362 // Pick the other block.
3363 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3364 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3365 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3366 }// end of for each redux variable.
3370 // Remove redundant induction instructions.
3371 cse(LoopVectorBody);
3374 void InnerLoopVectorizer::fixLCSSAPHIs() {
3375 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3376 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3377 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3378 if (!LCSSAPhi) break;
3379 if (LCSSAPhi->getNumIncomingValues() == 1)
3380 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3385 InnerLoopVectorizer::VectorParts
3386 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3387 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3390 // Look for cached value.
3391 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3392 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3393 if (ECEntryIt != MaskCache.end())
3394 return ECEntryIt->second;
3396 VectorParts SrcMask = createBlockInMask(Src);
3398 // The terminator has to be a branch inst!
3399 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3400 assert(BI && "Unexpected terminator found");
3402 if (BI->isConditional()) {
3403 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3405 if (BI->getSuccessor(0) != Dst)
3406 for (unsigned part = 0; part < UF; ++part)
3407 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3409 for (unsigned part = 0; part < UF; ++part)
3410 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3412 MaskCache[Edge] = EdgeMask;
3416 MaskCache[Edge] = SrcMask;
3420 InnerLoopVectorizer::VectorParts
3421 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3422 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3424 // Loop incoming mask is all-one.
3425 if (OrigLoop->getHeader() == BB) {
3426 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3427 return getVectorValue(C);
3430 // This is the block mask. We OR all incoming edges, and with zero.
3431 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3432 VectorParts BlockMask = getVectorValue(Zero);
3435 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3436 VectorParts EM = createEdgeMask(*it, BB);
3437 for (unsigned part = 0; part < UF; ++part)
3438 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3444 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3445 InnerLoopVectorizer::VectorParts &Entry,
3446 unsigned UF, unsigned VF, PhiVector *PV) {
3447 PHINode* P = cast<PHINode>(PN);
3448 // Handle reduction variables:
3449 if (Legal->getReductionVars()->count(P)) {
3450 for (unsigned part = 0; part < UF; ++part) {
3451 // This is phase one of vectorizing PHIs.
3452 Type *VecTy = (VF == 1) ? PN->getType() :
3453 VectorType::get(PN->getType(), VF);
3454 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3455 LoopVectorBody.back()-> getFirstInsertionPt());
3461 setDebugLocFromInst(Builder, P);
3462 // Check for PHI nodes that are lowered to vector selects.
3463 if (P->getParent() != OrigLoop->getHeader()) {
3464 // We know that all PHIs in non-header blocks are converted into
3465 // selects, so we don't have to worry about the insertion order and we
3466 // can just use the builder.
3467 // At this point we generate the predication tree. There may be
3468 // duplications since this is a simple recursive scan, but future
3469 // optimizations will clean it up.
3471 unsigned NumIncoming = P->getNumIncomingValues();
3473 // Generate a sequence of selects of the form:
3474 // SELECT(Mask3, In3,
3475 // SELECT(Mask2, In2,
3477 for (unsigned In = 0; In < NumIncoming; In++) {
3478 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3480 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3482 for (unsigned part = 0; part < UF; ++part) {
3483 // We might have single edge PHIs (blocks) - use an identity
3484 // 'select' for the first PHI operand.
3486 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3489 // Select between the current value and the previous incoming edge
3490 // based on the incoming mask.
3491 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3492 Entry[part], "predphi");
3498 // This PHINode must be an induction variable.
3499 // Make sure that we know about it.
3500 assert(Legal->getInductionVars()->count(P) &&
3501 "Not an induction variable");
3503 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3505 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3506 // which can be found from the original scalar operations.
3507 switch (II.getKind()) {
3508 case InductionDescriptor::IK_NoInduction:
3509 llvm_unreachable("Unknown induction");
3510 case InductionDescriptor::IK_IntInduction: {
3511 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3512 // Handle other induction variables that are now based on the
3514 Value *V = Induction;
3515 if (P != OldInduction) {
3516 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3517 V = II.transform(Builder, V);
3518 V->setName("offset.idx");
3520 Value *Broadcasted = getBroadcastInstrs(V);
3521 // After broadcasting the induction variable we need to make the vector
3522 // consecutive by adding 0, 1, 2, etc.
3523 for (unsigned part = 0; part < UF; ++part)
3524 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3527 case InductionDescriptor::IK_PtrInduction:
3528 // Handle the pointer induction variable case.
3529 assert(P->getType()->isPointerTy() && "Unexpected type.");
3530 // This is the normalized GEP that starts counting at zero.
3531 Value *PtrInd = Induction;
3532 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3533 // This is the vector of results. Notice that we don't generate
3534 // vector geps because scalar geps result in better code.
3535 for (unsigned part = 0; part < UF; ++part) {
3537 int EltIndex = part;
3538 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3539 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3540 Value *SclrGep = II.transform(Builder, GlobalIdx);
3541 SclrGep->setName("next.gep");
3542 Entry[part] = SclrGep;
3546 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3547 for (unsigned int i = 0; i < VF; ++i) {
3548 int EltIndex = i + part * VF;
3549 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3550 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3551 Value *SclrGep = II.transform(Builder, GlobalIdx);
3552 SclrGep->setName("next.gep");
3553 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3554 Builder.getInt32(i),
3557 Entry[part] = VecVal;
3563 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3564 // For each instruction in the old loop.
3565 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3566 VectorParts &Entry = WidenMap.get(it);
3567 switch (it->getOpcode()) {
3568 case Instruction::Br:
3569 // Nothing to do for PHIs and BR, since we already took care of the
3570 // loop control flow instructions.
3572 case Instruction::PHI: {
3573 // Vectorize PHINodes.
3574 widenPHIInstruction(it, Entry, UF, VF, PV);
3578 case Instruction::Add:
3579 case Instruction::FAdd:
3580 case Instruction::Sub:
3581 case Instruction::FSub:
3582 case Instruction::Mul:
3583 case Instruction::FMul:
3584 case Instruction::UDiv:
3585 case Instruction::SDiv:
3586 case Instruction::FDiv:
3587 case Instruction::URem:
3588 case Instruction::SRem:
3589 case Instruction::FRem:
3590 case Instruction::Shl:
3591 case Instruction::LShr:
3592 case Instruction::AShr:
3593 case Instruction::And:
3594 case Instruction::Or:
3595 case Instruction::Xor: {
3596 // Just widen binops.
3597 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3598 setDebugLocFromInst(Builder, BinOp);
3599 VectorParts &A = getVectorValue(it->getOperand(0));
3600 VectorParts &B = getVectorValue(it->getOperand(1));
3602 // Use this vector value for all users of the original instruction.
3603 for (unsigned Part = 0; Part < UF; ++Part) {
3604 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3606 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3607 VecOp->copyIRFlags(BinOp);
3612 propagateMetadata(Entry, it);
3615 case Instruction::Select: {
3617 // If the selector is loop invariant we can create a select
3618 // instruction with a scalar condition. Otherwise, use vector-select.
3619 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3621 setDebugLocFromInst(Builder, it);
3623 // The condition can be loop invariant but still defined inside the
3624 // loop. This means that we can't just use the original 'cond' value.
3625 // We have to take the 'vectorized' value and pick the first lane.
3626 // Instcombine will make this a no-op.
3627 VectorParts &Cond = getVectorValue(it->getOperand(0));
3628 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3629 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3631 Value *ScalarCond = (VF == 1) ? Cond[0] :
3632 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3634 for (unsigned Part = 0; Part < UF; ++Part) {
3635 Entry[Part] = Builder.CreateSelect(
3636 InvariantCond ? ScalarCond : Cond[Part],
3641 propagateMetadata(Entry, it);
3645 case Instruction::ICmp:
3646 case Instruction::FCmp: {
3647 // Widen compares. Generate vector compares.
3648 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3649 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3650 setDebugLocFromInst(Builder, it);
3651 VectorParts &A = getVectorValue(it->getOperand(0));
3652 VectorParts &B = getVectorValue(it->getOperand(1));
3653 for (unsigned Part = 0; Part < UF; ++Part) {
3656 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3658 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3662 propagateMetadata(Entry, it);
3666 case Instruction::Store:
3667 case Instruction::Load:
3668 vectorizeMemoryInstruction(it);
3670 case Instruction::ZExt:
3671 case Instruction::SExt:
3672 case Instruction::FPToUI:
3673 case Instruction::FPToSI:
3674 case Instruction::FPExt:
3675 case Instruction::PtrToInt:
3676 case Instruction::IntToPtr:
3677 case Instruction::SIToFP:
3678 case Instruction::UIToFP:
3679 case Instruction::Trunc:
3680 case Instruction::FPTrunc:
3681 case Instruction::BitCast: {
3682 CastInst *CI = dyn_cast<CastInst>(it);
3683 setDebugLocFromInst(Builder, it);
3684 /// Optimize the special case where the source is the induction
3685 /// variable. Notice that we can only optimize the 'trunc' case
3686 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3687 /// c. other casts depend on pointer size.
3688 if (CI->getOperand(0) == OldInduction &&
3689 it->getOpcode() == Instruction::Trunc) {
3690 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3692 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3693 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3695 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3696 for (unsigned Part = 0; Part < UF; ++Part)
3697 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3698 propagateMetadata(Entry, it);
3701 /// Vectorize casts.
3702 Type *DestTy = (VF == 1) ? CI->getType() :
3703 VectorType::get(CI->getType(), VF);
3705 VectorParts &A = getVectorValue(it->getOperand(0));
3706 for (unsigned Part = 0; Part < UF; ++Part)
3707 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3708 propagateMetadata(Entry, it);
3712 case Instruction::Call: {
3713 // Ignore dbg intrinsics.
3714 if (isa<DbgInfoIntrinsic>(it))
3716 setDebugLocFromInst(Builder, it);
3718 Module *M = BB->getParent()->getParent();
3719 CallInst *CI = cast<CallInst>(it);
3721 StringRef FnName = CI->getCalledFunction()->getName();
3722 Function *F = CI->getCalledFunction();
3723 Type *RetTy = ToVectorTy(CI->getType(), VF);
3724 SmallVector<Type *, 4> Tys;
3725 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3726 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3728 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3730 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3731 ID == Intrinsic::lifetime_start)) {
3732 scalarizeInstruction(it);
3735 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3736 // version of the instruction.
3737 // Is it beneficial to perform intrinsic call compared to lib call?
3738 bool NeedToScalarize;
3739 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3740 bool UseVectorIntrinsic =
3741 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3742 if (!UseVectorIntrinsic && NeedToScalarize) {
3743 scalarizeInstruction(it);
3747 for (unsigned Part = 0; Part < UF; ++Part) {
3748 SmallVector<Value *, 4> Args;
3749 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3750 Value *Arg = CI->getArgOperand(i);
3751 // Some intrinsics have a scalar argument - don't replace it with a
3753 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3754 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3755 Arg = VectorArg[Part];
3757 Args.push_back(Arg);
3761 if (UseVectorIntrinsic) {
3762 // Use vector version of the intrinsic.
3763 Type *TysForDecl[] = {CI->getType()};
3765 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3766 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3768 // Use vector version of the library call.
3769 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3770 assert(!VFnName.empty() && "Vector function name is empty.");
3771 VectorF = M->getFunction(VFnName);
3773 // Generate a declaration
3774 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3776 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3777 VectorF->copyAttributesFrom(F);
3780 assert(VectorF && "Can't create vector function.");
3781 Entry[Part] = Builder.CreateCall(VectorF, Args);
3784 propagateMetadata(Entry, it);
3789 // All other instructions are unsupported. Scalarize them.
3790 scalarizeInstruction(it);
3793 }// end of for_each instr.
3796 void InnerLoopVectorizer::updateAnalysis() {
3797 // Forget the original basic block.
3798 SE->forgetLoop(OrigLoop);
3800 // Update the dominator tree information.
3801 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3802 "Entry does not dominate exit.");
3804 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3805 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3806 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3808 // Due to if predication of stores we might create a sequence of "if(pred)
3809 // a[i] = ...; " blocks.
3810 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3812 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3813 else if (isPredicatedBlock(i)) {
3814 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3816 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3820 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody.back());
3821 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3822 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3823 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3825 DEBUG(DT->verifyDomTree());
3828 /// \brief Check whether it is safe to if-convert this phi node.
3830 /// Phi nodes with constant expressions that can trap are not safe to if
3832 static bool canIfConvertPHINodes(BasicBlock *BB) {
3833 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3834 PHINode *Phi = dyn_cast<PHINode>(I);
3837 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3838 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3845 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3846 if (!EnableIfConversion) {
3847 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3851 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3853 // A list of pointers that we can safely read and write to.
3854 SmallPtrSet<Value *, 8> SafePointes;
3856 // Collect safe addresses.
3857 for (Loop::block_iterator BI = TheLoop->block_begin(),
3858 BE = TheLoop->block_end(); BI != BE; ++BI) {
3859 BasicBlock *BB = *BI;
3861 if (blockNeedsPredication(BB))
3864 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3865 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3866 SafePointes.insert(LI->getPointerOperand());
3867 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3868 SafePointes.insert(SI->getPointerOperand());
3872 // Collect the blocks that need predication.
3873 BasicBlock *Header = TheLoop->getHeader();
3874 for (Loop::block_iterator BI = TheLoop->block_begin(),
3875 BE = TheLoop->block_end(); BI != BE; ++BI) {
3876 BasicBlock *BB = *BI;
3878 // We don't support switch statements inside loops.
3879 if (!isa<BranchInst>(BB->getTerminator())) {
3880 emitAnalysis(VectorizationReport(BB->getTerminator())
3881 << "loop contains a switch statement");
3885 // We must be able to predicate all blocks that need to be predicated.
3886 if (blockNeedsPredication(BB)) {
3887 if (!blockCanBePredicated(BB, SafePointes)) {
3888 emitAnalysis(VectorizationReport(BB->getTerminator())
3889 << "control flow cannot be substituted for a select");
3892 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3893 emitAnalysis(VectorizationReport(BB->getTerminator())
3894 << "control flow cannot be substituted for a select");
3899 // We can if-convert this loop.
3903 bool LoopVectorizationLegality::canVectorize() {
3904 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3905 // be canonicalized.
3906 if (!TheLoop->getLoopPreheader()) {
3908 VectorizationReport() <<
3909 "loop control flow is not understood by vectorizer");
3913 // We can only vectorize innermost loops.
3914 if (!TheLoop->empty()) {
3915 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3919 // We must have a single backedge.
3920 if (TheLoop->getNumBackEdges() != 1) {
3922 VectorizationReport() <<
3923 "loop control flow is not understood by vectorizer");
3927 // We must have a single exiting block.
3928 if (!TheLoop->getExitingBlock()) {
3930 VectorizationReport() <<
3931 "loop control flow is not understood by vectorizer");
3935 // We only handle bottom-tested loops, i.e. loop in which the condition is
3936 // checked at the end of each iteration. With that we can assume that all
3937 // instructions in the loop are executed the same number of times.
3938 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3940 VectorizationReport() <<
3941 "loop control flow is not understood by vectorizer");
3945 // We need to have a loop header.
3946 DEBUG(dbgs() << "LV: Found a loop: " <<
3947 TheLoop->getHeader()->getName() << '\n');
3949 // Check if we can if-convert non-single-bb loops.
3950 unsigned NumBlocks = TheLoop->getNumBlocks();
3951 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3952 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3956 // ScalarEvolution needs to be able to find the exit count.
3957 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3958 if (ExitCount == SE->getCouldNotCompute()) {
3959 emitAnalysis(VectorizationReport() <<
3960 "could not determine number of loop iterations");
3961 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3965 // Check if we can vectorize the instructions and CFG in this loop.
3966 if (!canVectorizeInstrs()) {
3967 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3971 // Go over each instruction and look at memory deps.
3972 if (!canVectorizeMemory()) {
3973 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3977 // Collect all of the variables that remain uniform after vectorization.
3978 collectLoopUniforms();
3980 DEBUG(dbgs() << "LV: We can vectorize this loop"
3981 << (LAI->getRuntimePointerChecking()->Need
3982 ? " (with a runtime bound check)"
3986 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
3988 // If an override option has been passed in for interleaved accesses, use it.
3989 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
3990 UseInterleaved = EnableInterleavedMemAccesses;
3992 // Analyze interleaved memory accesses.
3994 InterleaveInfo.analyzeInterleaving(Strides);
3996 // Okay! We can vectorize. At this point we don't have any other mem analysis
3997 // which may limit our maximum vectorization factor, so just return true with
4002 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4003 if (Ty->isPointerTy())
4004 return DL.getIntPtrType(Ty);
4006 // It is possible that char's or short's overflow when we ask for the loop's
4007 // trip count, work around this by changing the type size.
4008 if (Ty->getScalarSizeInBits() < 32)
4009 return Type::getInt32Ty(Ty->getContext());
4014 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4015 Ty0 = convertPointerToIntegerType(DL, Ty0);
4016 Ty1 = convertPointerToIntegerType(DL, Ty1);
4017 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4022 /// \brief Check that the instruction has outside loop users and is not an
4023 /// identified reduction variable.
4024 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4025 SmallPtrSetImpl<Value *> &Reductions) {
4026 // Reduction instructions are allowed to have exit users. All other
4027 // instructions must not have external users.
4028 if (!Reductions.count(Inst))
4029 //Check that all of the users of the loop are inside the BB.
4030 for (User *U : Inst->users()) {
4031 Instruction *UI = cast<Instruction>(U);
4032 // This user may be a reduction exit value.
4033 if (!TheLoop->contains(UI)) {
4034 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4041 bool LoopVectorizationLegality::canVectorizeInstrs() {
4042 BasicBlock *Header = TheLoop->getHeader();
4044 // Look for the attribute signaling the absence of NaNs.
4045 Function &F = *Header->getParent();
4046 const DataLayout &DL = F.getParent()->getDataLayout();
4047 if (F.hasFnAttribute("no-nans-fp-math"))
4049 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4051 // For each block in the loop.
4052 for (Loop::block_iterator bb = TheLoop->block_begin(),
4053 be = TheLoop->block_end(); bb != be; ++bb) {
4055 // Scan the instructions in the block and look for hazards.
4056 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4059 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4060 Type *PhiTy = Phi->getType();
4061 // Check that this PHI type is allowed.
4062 if (!PhiTy->isIntegerTy() &&
4063 !PhiTy->isFloatingPointTy() &&
4064 !PhiTy->isPointerTy()) {
4065 emitAnalysis(VectorizationReport(it)
4066 << "loop control flow is not understood by vectorizer");
4067 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4071 // If this PHINode is not in the header block, then we know that we
4072 // can convert it to select during if-conversion. No need to check if
4073 // the PHIs in this block are induction or reduction variables.
4074 if (*bb != Header) {
4075 // Check that this instruction has no outside users or is an
4076 // identified reduction value with an outside user.
4077 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4079 emitAnalysis(VectorizationReport(it) <<
4080 "value could not be identified as "
4081 "an induction or reduction variable");
4085 // We only allow if-converted PHIs with exactly two incoming values.
4086 if (Phi->getNumIncomingValues() != 2) {
4087 emitAnalysis(VectorizationReport(it)
4088 << "control flow not understood by vectorizer");
4089 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4093 InductionDescriptor ID;
4094 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4095 Inductions[Phi] = ID;
4096 // Get the widest type.
4098 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4100 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4102 // Int inductions are special because we only allow one IV.
4103 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4104 ID.getStepValue()->isOne() &&
4105 isa<Constant>(ID.getStartValue()) &&
4106 cast<Constant>(ID.getStartValue())->isNullValue()) {
4107 // Use the phi node with the widest type as induction. Use the last
4108 // one if there are multiple (no good reason for doing this other
4109 // than it is expedient). We've checked that it begins at zero and
4110 // steps by one, so this is a canonical induction variable.
4111 if (!Induction || PhiTy == WidestIndTy)
4115 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4117 // Until we explicitly handle the case of an induction variable with
4118 // an outside loop user we have to give up vectorizing this loop.
4119 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4120 emitAnalysis(VectorizationReport(it) <<
4121 "use of induction value outside of the "
4122 "loop is not handled by vectorizer");
4129 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4131 if (Reductions[Phi].hasUnsafeAlgebra())
4132 Requirements->addUnsafeAlgebraInst(
4133 Reductions[Phi].getUnsafeAlgebraInst());
4134 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4138 emitAnalysis(VectorizationReport(it) <<
4139 "value that could not be identified as "
4140 "reduction is used outside the loop");
4141 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4143 }// end of PHI handling
4145 // We handle calls that:
4146 // * Are debug info intrinsics.
4147 // * Have a mapping to an IR intrinsic.
4148 // * Have a vector version available.
4149 CallInst *CI = dyn_cast<CallInst>(it);
4150 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4151 !(CI->getCalledFunction() && TLI &&
4152 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4153 emitAnalysis(VectorizationReport(it) <<
4154 "call instruction cannot be vectorized");
4155 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4159 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4160 // second argument is the same (i.e. loop invariant)
4162 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4163 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4164 emitAnalysis(VectorizationReport(it)
4165 << "intrinsic instruction cannot be vectorized");
4166 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4171 // Check that the instruction return type is vectorizable.
4172 // Also, we can't vectorize extractelement instructions.
4173 if ((!VectorType::isValidElementType(it->getType()) &&
4174 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4175 emitAnalysis(VectorizationReport(it)
4176 << "instruction return type cannot be vectorized");
4177 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4181 // Check that the stored type is vectorizable.
4182 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4183 Type *T = ST->getValueOperand()->getType();
4184 if (!VectorType::isValidElementType(T)) {
4185 emitAnalysis(VectorizationReport(ST) <<
4186 "store instruction cannot be vectorized");
4189 if (EnableMemAccessVersioning)
4190 collectStridedAccess(ST);
4193 if (EnableMemAccessVersioning)
4194 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4195 collectStridedAccess(LI);
4197 // Reduction instructions are allowed to have exit users.
4198 // All other instructions must not have external users.
4199 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4200 emitAnalysis(VectorizationReport(it) <<
4201 "value cannot be used outside the loop");
4210 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4211 if (Inductions.empty()) {
4212 emitAnalysis(VectorizationReport()
4213 << "loop induction variable could not be identified");
4218 // Now we know the widest induction type, check if our found induction
4219 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4220 // will create another.
4221 if (Induction && WidestIndTy != Induction->getType())
4222 Induction = nullptr;
4227 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4228 Value *Ptr = nullptr;
4229 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4230 Ptr = LI->getPointerOperand();
4231 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4232 Ptr = SI->getPointerOperand();
4236 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4240 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4241 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4242 Strides[Ptr] = Stride;
4243 StrideSet.insert(Stride);
4246 void LoopVectorizationLegality::collectLoopUniforms() {
4247 // We now know that the loop is vectorizable!
4248 // Collect variables that will remain uniform after vectorization.
4249 std::vector<Value*> Worklist;
4250 BasicBlock *Latch = TheLoop->getLoopLatch();
4252 // Start with the conditional branch and walk up the block.
4253 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4255 // Also add all consecutive pointer values; these values will be uniform
4256 // after vectorization (and subsequent cleanup) and, until revectorization is
4257 // supported, all dependencies must also be uniform.
4258 for (Loop::block_iterator B = TheLoop->block_begin(),
4259 BE = TheLoop->block_end(); B != BE; ++B)
4260 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4262 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4263 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4265 while (!Worklist.empty()) {
4266 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4267 Worklist.pop_back();
4269 // Look at instructions inside this loop.
4270 // Stop when reaching PHI nodes.
4271 // TODO: we need to follow values all over the loop, not only in this block.
4272 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4275 // This is a known uniform.
4278 // Insert all operands.
4279 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4283 bool LoopVectorizationLegality::canVectorizeMemory() {
4284 LAI = &LAA->getInfo(TheLoop, Strides);
4285 auto &OptionalReport = LAI->getReport();
4287 emitAnalysis(VectorizationReport(*OptionalReport));
4288 if (!LAI->canVectorizeMemory())
4291 if (LAI->hasStoreToLoopInvariantAddress()) {
4293 VectorizationReport()
4294 << "write to a loop invariant address could not be vectorized");
4295 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4299 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4304 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4305 Value *In0 = const_cast<Value*>(V);
4306 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4310 return Inductions.count(PN);
4313 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4314 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4317 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4318 SmallPtrSetImpl<Value *> &SafePtrs) {
4320 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4321 // Check that we don't have a constant expression that can trap as operand.
4322 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4324 if (Constant *C = dyn_cast<Constant>(*OI))
4328 // We might be able to hoist the load.
4329 if (it->mayReadFromMemory()) {
4330 LoadInst *LI = dyn_cast<LoadInst>(it);
4333 if (!SafePtrs.count(LI->getPointerOperand())) {
4334 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4335 MaskedOp.insert(LI);
4342 // We don't predicate stores at the moment.
4343 if (it->mayWriteToMemory()) {
4344 StoreInst *SI = dyn_cast<StoreInst>(it);
4345 // We only support predication of stores in basic blocks with one
4350 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4351 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4353 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4354 !isSinglePredecessor) {
4355 // Build a masked store if it is legal for the target, otherwise scalarize
4357 bool isLegalMaskedOp =
4358 isLegalMaskedStore(SI->getValueOperand()->getType(),
4359 SI->getPointerOperand());
4360 if (isLegalMaskedOp) {
4362 MaskedOp.insert(SI);
4371 // The instructions below can trap.
4372 switch (it->getOpcode()) {
4374 case Instruction::UDiv:
4375 case Instruction::SDiv:
4376 case Instruction::URem:
4377 case Instruction::SRem:
4385 void InterleavedAccessInfo::collectConstStridedAccesses(
4386 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4387 const ValueToValueMap &Strides) {
4388 // Holds load/store instructions in program order.
4389 SmallVector<Instruction *, 16> AccessList;
4391 for (auto *BB : TheLoop->getBlocks()) {
4392 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4394 for (auto &I : *BB) {
4395 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4397 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4401 AccessList.push_back(&I);
4405 if (AccessList.empty())
4408 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4409 for (auto I : AccessList) {
4410 LoadInst *LI = dyn_cast<LoadInst>(I);
4411 StoreInst *SI = dyn_cast<StoreInst>(I);
4413 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4414 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4416 // The factor of the corresponding interleave group.
4417 unsigned Factor = std::abs(Stride);
4419 // Ignore the access if the factor is too small or too large.
4420 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4423 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4424 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4425 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4427 // An alignment of 0 means target ABI alignment.
4428 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4430 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4432 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4436 // Analyze interleaved accesses and collect them into interleave groups.
4438 // Notice that the vectorization on interleaved groups will change instruction
4439 // orders and may break dependences. But the memory dependence check guarantees
4440 // that there is no overlap between two pointers of different strides, element
4441 // sizes or underlying bases.
4443 // For pointers sharing the same stride, element size and underlying base, no
4444 // need to worry about Read-After-Write dependences and Write-After-Read
4447 // E.g. The RAW dependence: A[i] = a;
4449 // This won't exist as it is a store-load forwarding conflict, which has
4450 // already been checked and forbidden in the dependence check.
4452 // E.g. The WAR dependence: a = A[i]; // (1)
4454 // The store group of (2) is always inserted at or below (2), and the load group
4455 // of (1) is always inserted at or above (1). The dependence is safe.
4456 void InterleavedAccessInfo::analyzeInterleaving(
4457 const ValueToValueMap &Strides) {
4458 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4460 // Holds all the stride accesses.
4461 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4462 collectConstStridedAccesses(StrideAccesses, Strides);
4464 if (StrideAccesses.empty())
4467 // Holds all interleaved store groups temporarily.
4468 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4470 // Search the load-load/write-write pair B-A in bottom-up order and try to
4471 // insert B into the interleave group of A according to 3 rules:
4472 // 1. A and B have the same stride.
4473 // 2. A and B have the same memory object size.
4474 // 3. B belongs to the group according to the distance.
4476 // The bottom-up order can avoid breaking the Write-After-Write dependences
4477 // between two pointers of the same base.
4478 // E.g. A[i] = a; (1)
4481 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4482 // above (1), which guarantees that (1) is always above (2).
4483 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4485 Instruction *A = I->first;
4486 StrideDescriptor DesA = I->second;
4488 InterleaveGroup *Group = getInterleaveGroup(A);
4490 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4491 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4494 if (A->mayWriteToMemory())
4495 StoreGroups.insert(Group);
4497 for (auto II = std::next(I); II != E; ++II) {
4498 Instruction *B = II->first;
4499 StrideDescriptor DesB = II->second;
4501 // Ignore if B is already in a group or B is a different memory operation.
4502 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4505 // Check the rule 1 and 2.
4506 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4509 // Calculate the distance and prepare for the rule 3.
4510 const SCEVConstant *DistToA =
4511 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4515 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4517 // Skip if the distance is not multiple of size as they are not in the
4519 if (DistanceToA % static_cast<int>(DesA.Size))
4522 // The index of B is the index of A plus the related index to A.
4524 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4526 // Try to insert B into the group.
4527 if (Group->insertMember(B, IndexB, DesB.Align)) {
4528 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4529 << " into the interleave group with" << *A << '\n');
4530 InterleaveGroupMap[B] = Group;
4532 // Set the first load in program order as the insert position.
4533 if (B->mayReadFromMemory())
4534 Group->setInsertPos(B);
4536 } // Iteration on instruction B
4537 } // Iteration on instruction A
4539 // Remove interleaved store groups with gaps.
4540 for (InterleaveGroup *Group : StoreGroups)
4541 if (Group->getNumMembers() != Group->getFactor())
4542 releaseGroup(Group);
4545 LoopVectorizationCostModel::VectorizationFactor
4546 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4547 // Width 1 means no vectorize
4548 VectorizationFactor Factor = { 1U, 0U };
4549 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4550 emitAnalysis(VectorizationReport() <<
4551 "runtime pointer checks needed. Enable vectorization of this "
4552 "loop with '#pragma clang loop vectorize(enable)' when "
4553 "compiling with -Os/-Oz");
4555 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4559 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4560 emitAnalysis(VectorizationReport() <<
4561 "store that is conditionally executed prevents vectorization");
4562 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4566 // Find the trip count.
4567 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4568 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4570 unsigned WidestType = getWidestType();
4571 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4572 unsigned MaxSafeDepDist = -1U;
4573 if (Legal->getMaxSafeDepDistBytes() != -1U)
4574 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4575 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4576 WidestRegister : MaxSafeDepDist);
4577 unsigned MaxVectorSize = WidestRegister / WidestType;
4578 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4579 DEBUG(dbgs() << "LV: The Widest register is: "
4580 << WidestRegister << " bits.\n");
4582 if (MaxVectorSize == 0) {
4583 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4587 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4588 " into one vector!");
4590 unsigned VF = MaxVectorSize;
4592 // If we optimize the program for size, avoid creating the tail loop.
4594 // If we are unable to calculate the trip count then don't try to vectorize.
4597 (VectorizationReport() <<
4598 "unable to calculate the loop count due to complex control flow");
4599 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4603 // Find the maximum SIMD width that can fit within the trip count.
4604 VF = TC % MaxVectorSize;
4609 // If the trip count that we found modulo the vectorization factor is not
4610 // zero then we require a tail.
4611 emitAnalysis(VectorizationReport() <<
4612 "cannot optimize for size and vectorize at the "
4613 "same time. Enable vectorization of this loop "
4614 "with '#pragma clang loop vectorize(enable)' "
4615 "when compiling with -Os/-Oz");
4616 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4621 int UserVF = Hints->getWidth();
4623 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4624 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4626 Factor.Width = UserVF;
4630 float Cost = expectedCost(1);
4632 const float ScalarCost = Cost;
4635 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4637 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4638 // Ignore scalar width, because the user explicitly wants vectorization.
4639 if (ForceVectorization && VF > 1) {
4641 Cost = expectedCost(Width) / (float)Width;
4644 for (unsigned i=2; i <= VF; i*=2) {
4645 // Notice that the vector loop needs to be executed less times, so
4646 // we need to divide the cost of the vector loops by the width of
4647 // the vector elements.
4648 float VectorCost = expectedCost(i) / (float)i;
4649 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4650 (int)VectorCost << ".\n");
4651 if (VectorCost < Cost) {
4657 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4658 << "LV: Vectorization seems to be not beneficial, "
4659 << "but was forced by a user.\n");
4660 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4661 Factor.Width = Width;
4662 Factor.Cost = Width * Cost;
4666 unsigned LoopVectorizationCostModel::getWidestType() {
4667 unsigned MaxWidth = 8;
4668 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4671 for (Loop::block_iterator bb = TheLoop->block_begin(),
4672 be = TheLoop->block_end(); bb != be; ++bb) {
4673 BasicBlock *BB = *bb;
4675 // For each instruction in the loop.
4676 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4677 Type *T = it->getType();
4679 // Skip ignored values.
4680 if (ValuesToIgnore.count(it))
4683 // Only examine Loads, Stores and PHINodes.
4684 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4687 // Examine PHI nodes that are reduction variables. Update the type to
4688 // account for the recurrence type.
4689 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4690 if (!Legal->getReductionVars()->count(PN))
4692 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4693 T = RdxDesc.getRecurrenceType();
4696 // Examine the stored values.
4697 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4698 T = ST->getValueOperand()->getType();
4700 // Ignore loaded pointer types and stored pointer types that are not
4701 // consecutive. However, we do want to take consecutive stores/loads of
4702 // pointer vectors into account.
4703 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4706 MaxWidth = std::max(MaxWidth,
4707 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4714 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4716 unsigned LoopCost) {
4718 // -- The interleave heuristics --
4719 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4720 // There are many micro-architectural considerations that we can't predict
4721 // at this level. For example, frontend pressure (on decode or fetch) due to
4722 // code size, or the number and capabilities of the execution ports.
4724 // We use the following heuristics to select the interleave count:
4725 // 1. If the code has reductions, then we interleave to break the cross
4726 // iteration dependency.
4727 // 2. If the loop is really small, then we interleave to reduce the loop
4729 // 3. We don't interleave if we think that we will spill registers to memory
4730 // due to the increased register pressure.
4732 // When we optimize for size, we don't interleave.
4736 // We used the distance for the interleave count.
4737 if (Legal->getMaxSafeDepDistBytes() != -1U)
4740 // Do not interleave loops with a relatively small trip count.
4741 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4742 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4745 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4746 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4750 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4751 TargetNumRegisters = ForceTargetNumScalarRegs;
4753 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4754 TargetNumRegisters = ForceTargetNumVectorRegs;
4757 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4758 // We divide by these constants so assume that we have at least one
4759 // instruction that uses at least one register.
4760 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4761 R.NumInstructions = std::max(R.NumInstructions, 1U);
4763 // We calculate the interleave count using the following formula.
4764 // Subtract the number of loop invariants from the number of available
4765 // registers. These registers are used by all of the interleaved instances.
4766 // Next, divide the remaining registers by the number of registers that is
4767 // required by the loop, in order to estimate how many parallel instances
4768 // fit without causing spills. All of this is rounded down if necessary to be
4769 // a power of two. We want power of two interleave count to simplify any
4770 // addressing operations or alignment considerations.
4771 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4774 // Don't count the induction variable as interleaved.
4775 if (EnableIndVarRegisterHeur)
4776 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4777 std::max(1U, (R.MaxLocalUsers - 1)));
4779 // Clamp the interleave ranges to reasonable counts.
4780 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4782 // Check if the user has overridden the max.
4784 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4785 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4787 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4788 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4791 // If we did not calculate the cost for VF (because the user selected the VF)
4792 // then we calculate the cost of VF here.
4794 LoopCost = expectedCost(VF);
4796 // Clamp the calculated IC to be between the 1 and the max interleave count
4797 // that the target allows.
4798 if (IC > MaxInterleaveCount)
4799 IC = MaxInterleaveCount;
4803 // Interleave if we vectorized this loop and there is a reduction that could
4804 // benefit from interleaving.
4805 if (VF > 1 && Legal->getReductionVars()->size()) {
4806 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4810 // Note that if we've already vectorized the loop we will have done the
4811 // runtime check and so interleaving won't require further checks.
4812 bool InterleavingRequiresRuntimePointerCheck =
4813 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4815 // We want to interleave small loops in order to reduce the loop overhead and
4816 // potentially expose ILP opportunities.
4817 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4818 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4819 // We assume that the cost overhead is 1 and we use the cost model
4820 // to estimate the cost of the loop and interleave until the cost of the
4821 // loop overhead is about 5% of the cost of the loop.
4823 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4825 // Interleave until store/load ports (estimated by max interleave count) are
4827 unsigned NumStores = Legal->getNumStores();
4828 unsigned NumLoads = Legal->getNumLoads();
4829 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4830 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4832 // If we have a scalar reduction (vector reductions are already dealt with
4833 // by this point), we can increase the critical path length if the loop
4834 // we're interleaving is inside another loop. Limit, by default to 2, so the
4835 // critical path only gets increased by one reduction operation.
4836 if (Legal->getReductionVars()->size() &&
4837 TheLoop->getLoopDepth() > 1) {
4838 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4839 SmallIC = std::min(SmallIC, F);
4840 StoresIC = std::min(StoresIC, F);
4841 LoadsIC = std::min(LoadsIC, F);
4844 if (EnableLoadStoreRuntimeInterleave &&
4845 std::max(StoresIC, LoadsIC) > SmallIC) {
4846 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4847 return std::max(StoresIC, LoadsIC);
4850 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4854 // Interleave if this is a large loop (small loops are already dealt with by
4856 // point) that could benefit from interleaving.
4857 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4858 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4859 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4863 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4867 LoopVectorizationCostModel::RegisterUsage
4868 LoopVectorizationCostModel::calculateRegisterUsage() {
4869 // This function calculates the register usage by measuring the highest number
4870 // of values that are alive at a single location. Obviously, this is a very
4871 // rough estimation. We scan the loop in a topological order in order and
4872 // assign a number to each instruction. We use RPO to ensure that defs are
4873 // met before their users. We assume that each instruction that has in-loop
4874 // users starts an interval. We record every time that an in-loop value is
4875 // used, so we have a list of the first and last occurrences of each
4876 // instruction. Next, we transpose this data structure into a multi map that
4877 // holds the list of intervals that *end* at a specific location. This multi
4878 // map allows us to perform a linear search. We scan the instructions linearly
4879 // and record each time that a new interval starts, by placing it in a set.
4880 // If we find this value in the multi-map then we remove it from the set.
4881 // The max register usage is the maximum size of the set.
4882 // We also search for instructions that are defined outside the loop, but are
4883 // used inside the loop. We need this number separately from the max-interval
4884 // usage number because when we unroll, loop-invariant values do not take
4886 LoopBlocksDFS DFS(TheLoop);
4890 R.NumInstructions = 0;
4892 // Each 'key' in the map opens a new interval. The values
4893 // of the map are the index of the 'last seen' usage of the
4894 // instruction that is the key.
4895 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4896 // Maps instruction to its index.
4897 DenseMap<unsigned, Instruction*> IdxToInstr;
4898 // Marks the end of each interval.
4899 IntervalMap EndPoint;
4900 // Saves the list of instruction indices that are used in the loop.
4901 SmallSet<Instruction*, 8> Ends;
4902 // Saves the list of values that are used in the loop but are
4903 // defined outside the loop, such as arguments and constants.
4904 SmallPtrSet<Value*, 8> LoopInvariants;
4907 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4908 be = DFS.endRPO(); bb != be; ++bb) {
4909 R.NumInstructions += (*bb)->size();
4910 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4912 Instruction *I = it;
4913 IdxToInstr[Index++] = I;
4915 // Save the end location of each USE.
4916 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4917 Value *U = I->getOperand(i);
4918 Instruction *Instr = dyn_cast<Instruction>(U);
4920 // Ignore non-instruction values such as arguments, constants, etc.
4921 if (!Instr) continue;
4923 // If this instruction is outside the loop then record it and continue.
4924 if (!TheLoop->contains(Instr)) {
4925 LoopInvariants.insert(Instr);
4929 // Overwrite previous end points.
4930 EndPoint[Instr] = Index;
4936 // Saves the list of intervals that end with the index in 'key'.
4937 typedef SmallVector<Instruction*, 2> InstrList;
4938 DenseMap<unsigned, InstrList> TransposeEnds;
4940 // Transpose the EndPoints to a list of values that end at each index.
4941 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4943 TransposeEnds[it->second].push_back(it->first);
4945 SmallSet<Instruction*, 8> OpenIntervals;
4946 unsigned MaxUsage = 0;
4949 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4950 for (unsigned int i = 0; i < Index; ++i) {
4951 Instruction *I = IdxToInstr[i];
4952 // Ignore instructions that are never used within the loop.
4953 if (!Ends.count(I)) continue;
4955 // Skip ignored values.
4956 if (ValuesToIgnore.count(I))
4959 // Remove all of the instructions that end at this location.
4960 InstrList &List = TransposeEnds[i];
4961 for (unsigned int j=0, e = List.size(); j < e; ++j)
4962 OpenIntervals.erase(List[j]);
4964 // Count the number of live interals.
4965 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4967 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4968 OpenIntervals.size() << '\n');
4970 // Add the current instruction to the list of open intervals.
4971 OpenIntervals.insert(I);
4974 unsigned Invariant = LoopInvariants.size();
4975 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4976 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4977 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4979 R.LoopInvariantRegs = Invariant;
4980 R.MaxLocalUsers = MaxUsage;
4984 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4988 for (Loop::block_iterator bb = TheLoop->block_begin(),
4989 be = TheLoop->block_end(); bb != be; ++bb) {
4990 unsigned BlockCost = 0;
4991 BasicBlock *BB = *bb;
4993 // For each instruction in the old loop.
4994 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4995 // Skip dbg intrinsics.
4996 if (isa<DbgInfoIntrinsic>(it))
4999 // Skip ignored values.
5000 if (ValuesToIgnore.count(it))
5003 unsigned C = getInstructionCost(it, VF);
5005 // Check if we should override the cost.
5006 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5007 C = ForceTargetInstructionCost;
5010 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5011 VF << " For instruction: " << *it << '\n');
5014 // We assume that if-converted blocks have a 50% chance of being executed.
5015 // When the code is scalar then some of the blocks are avoided due to CF.
5016 // When the code is vectorized we execute all code paths.
5017 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5026 /// \brief Check whether the address computation for a non-consecutive memory
5027 /// access looks like an unlikely candidate for being merged into the indexing
5030 /// We look for a GEP which has one index that is an induction variable and all
5031 /// other indices are loop invariant. If the stride of this access is also
5032 /// within a small bound we decide that this address computation can likely be
5033 /// merged into the addressing mode.
5034 /// In all other cases, we identify the address computation as complex.
5035 static bool isLikelyComplexAddressComputation(Value *Ptr,
5036 LoopVectorizationLegality *Legal,
5037 ScalarEvolution *SE,
5038 const Loop *TheLoop) {
5039 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5043 // We are looking for a gep with all loop invariant indices except for one
5044 // which should be an induction variable.
5045 unsigned NumOperands = Gep->getNumOperands();
5046 for (unsigned i = 1; i < NumOperands; ++i) {
5047 Value *Opd = Gep->getOperand(i);
5048 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5049 !Legal->isInductionVariable(Opd))
5053 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5054 // can likely be merged into the address computation.
5055 unsigned MaxMergeDistance = 64;
5057 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5061 // Check the step is constant.
5062 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5063 // Calculate the pointer stride and check if it is consecutive.
5064 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5068 const APInt &APStepVal = C->getValue()->getValue();
5070 // Huge step value - give up.
5071 if (APStepVal.getBitWidth() > 64)
5074 int64_t StepVal = APStepVal.getSExtValue();
5076 return StepVal > MaxMergeDistance;
5079 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5080 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5086 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5087 // If we know that this instruction will remain uniform, check the cost of
5088 // the scalar version.
5089 if (Legal->isUniformAfterVectorization(I))
5092 Type *RetTy = I->getType();
5093 Type *VectorTy = ToVectorTy(RetTy, VF);
5095 // TODO: We need to estimate the cost of intrinsic calls.
5096 switch (I->getOpcode()) {
5097 case Instruction::GetElementPtr:
5098 // We mark this instruction as zero-cost because the cost of GEPs in
5099 // vectorized code depends on whether the corresponding memory instruction
5100 // is scalarized or not. Therefore, we handle GEPs with the memory
5101 // instruction cost.
5103 case Instruction::Br: {
5104 return TTI.getCFInstrCost(I->getOpcode());
5106 case Instruction::PHI:
5107 //TODO: IF-converted IFs become selects.
5109 case Instruction::Add:
5110 case Instruction::FAdd:
5111 case Instruction::Sub:
5112 case Instruction::FSub:
5113 case Instruction::Mul:
5114 case Instruction::FMul:
5115 case Instruction::UDiv:
5116 case Instruction::SDiv:
5117 case Instruction::FDiv:
5118 case Instruction::URem:
5119 case Instruction::SRem:
5120 case Instruction::FRem:
5121 case Instruction::Shl:
5122 case Instruction::LShr:
5123 case Instruction::AShr:
5124 case Instruction::And:
5125 case Instruction::Or:
5126 case Instruction::Xor: {
5127 // Since we will replace the stride by 1 the multiplication should go away.
5128 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5130 // Certain instructions can be cheaper to vectorize if they have a constant
5131 // second vector operand. One example of this are shifts on x86.
5132 TargetTransformInfo::OperandValueKind Op1VK =
5133 TargetTransformInfo::OK_AnyValue;
5134 TargetTransformInfo::OperandValueKind Op2VK =
5135 TargetTransformInfo::OK_AnyValue;
5136 TargetTransformInfo::OperandValueProperties Op1VP =
5137 TargetTransformInfo::OP_None;
5138 TargetTransformInfo::OperandValueProperties Op2VP =
5139 TargetTransformInfo::OP_None;
5140 Value *Op2 = I->getOperand(1);
5142 // Check for a splat of a constant or for a non uniform vector of constants.
5143 if (isa<ConstantInt>(Op2)) {
5144 ConstantInt *CInt = cast<ConstantInt>(Op2);
5145 if (CInt && CInt->getValue().isPowerOf2())
5146 Op2VP = TargetTransformInfo::OP_PowerOf2;
5147 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5148 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5149 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5150 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5152 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5153 if (CInt && CInt->getValue().isPowerOf2())
5154 Op2VP = TargetTransformInfo::OP_PowerOf2;
5155 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5159 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5162 case Instruction::Select: {
5163 SelectInst *SI = cast<SelectInst>(I);
5164 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5165 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5166 Type *CondTy = SI->getCondition()->getType();
5168 CondTy = VectorType::get(CondTy, VF);
5170 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5172 case Instruction::ICmp:
5173 case Instruction::FCmp: {
5174 Type *ValTy = I->getOperand(0)->getType();
5175 VectorTy = ToVectorTy(ValTy, VF);
5176 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5178 case Instruction::Store:
5179 case Instruction::Load: {
5180 StoreInst *SI = dyn_cast<StoreInst>(I);
5181 LoadInst *LI = dyn_cast<LoadInst>(I);
5182 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5184 VectorTy = ToVectorTy(ValTy, VF);
5186 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5187 unsigned AS = SI ? SI->getPointerAddressSpace() :
5188 LI->getPointerAddressSpace();
5189 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5190 // We add the cost of address computation here instead of with the gep
5191 // instruction because only here we know whether the operation is
5194 return TTI.getAddressComputationCost(VectorTy) +
5195 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5197 // For an interleaved access, calculate the total cost of the whole
5198 // interleave group.
5199 if (Legal->isAccessInterleaved(I)) {
5200 auto Group = Legal->getInterleavedAccessGroup(I);
5201 assert(Group && "Fail to get an interleaved access group.");
5203 // Only calculate the cost once at the insert position.
5204 if (Group->getInsertPos() != I)
5207 unsigned InterleaveFactor = Group->getFactor();
5209 VectorType::get(VectorTy->getVectorElementType(),
5210 VectorTy->getVectorNumElements() * InterleaveFactor);
5212 // Holds the indices of existing members in an interleaved load group.
5213 // An interleaved store group doesn't need this as it dones't allow gaps.
5214 SmallVector<unsigned, 4> Indices;
5216 for (unsigned i = 0; i < InterleaveFactor; i++)
5217 if (Group->getMember(i))
5218 Indices.push_back(i);
5221 // Calculate the cost of the whole interleaved group.
5222 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5223 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5224 Group->getAlignment(), AS);
5226 if (Group->isReverse())
5228 Group->getNumMembers() *
5229 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5231 // FIXME: The interleaved load group with a huge gap could be even more
5232 // expensive than scalar operations. Then we could ignore such group and
5233 // use scalar operations instead.
5237 // Scalarized loads/stores.
5238 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5239 bool Reverse = ConsecutiveStride < 0;
5240 const DataLayout &DL = I->getModule()->getDataLayout();
5241 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5242 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5243 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5244 bool IsComplexComputation =
5245 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5247 // The cost of extracting from the value vector and pointer vector.
5248 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5249 for (unsigned i = 0; i < VF; ++i) {
5250 // The cost of extracting the pointer operand.
5251 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5252 // In case of STORE, the cost of ExtractElement from the vector.
5253 // In case of LOAD, the cost of InsertElement into the returned
5255 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5256 Instruction::InsertElement,
5260 // The cost of the scalar loads/stores.
5261 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5262 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5267 // Wide load/stores.
5268 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5269 if (Legal->isMaskRequired(I))
5270 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5273 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5276 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5280 case Instruction::ZExt:
5281 case Instruction::SExt:
5282 case Instruction::FPToUI:
5283 case Instruction::FPToSI:
5284 case Instruction::FPExt:
5285 case Instruction::PtrToInt:
5286 case Instruction::IntToPtr:
5287 case Instruction::SIToFP:
5288 case Instruction::UIToFP:
5289 case Instruction::Trunc:
5290 case Instruction::FPTrunc:
5291 case Instruction::BitCast: {
5292 // We optimize the truncation of induction variable.
5293 // The cost of these is the same as the scalar operation.
5294 if (I->getOpcode() == Instruction::Trunc &&
5295 Legal->isInductionVariable(I->getOperand(0)))
5296 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5297 I->getOperand(0)->getType());
5299 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5300 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5302 case Instruction::Call: {
5303 bool NeedToScalarize;
5304 CallInst *CI = cast<CallInst>(I);
5305 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5306 if (getIntrinsicIDForCall(CI, TLI))
5307 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5311 // We are scalarizing the instruction. Return the cost of the scalar
5312 // instruction, plus the cost of insert and extract into vector
5313 // elements, times the vector width.
5316 if (!RetTy->isVoidTy() && VF != 1) {
5317 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5319 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5322 // The cost of inserting the results plus extracting each one of the
5324 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5327 // The cost of executing VF copies of the scalar instruction. This opcode
5328 // is unknown. Assume that it is the same as 'mul'.
5329 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5335 char LoopVectorize::ID = 0;
5336 static const char lv_name[] = "Loop Vectorization";
5337 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5338 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5339 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5340 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5341 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5342 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5343 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5344 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5345 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5346 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5347 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5348 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5351 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5352 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5356 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5357 // Check for a store.
5358 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5359 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5361 // Check for a load.
5362 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5363 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5369 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5370 bool IfPredicateStore) {
5371 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5372 // Holds vector parameters or scalars, in case of uniform vals.
5373 SmallVector<VectorParts, 4> Params;
5375 setDebugLocFromInst(Builder, Instr);
5377 // Find all of the vectorized parameters.
5378 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5379 Value *SrcOp = Instr->getOperand(op);
5381 // If we are accessing the old induction variable, use the new one.
5382 if (SrcOp == OldInduction) {
5383 Params.push_back(getVectorValue(SrcOp));
5387 // Try using previously calculated values.
5388 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5390 // If the src is an instruction that appeared earlier in the basic block
5391 // then it should already be vectorized.
5392 if (SrcInst && OrigLoop->contains(SrcInst)) {
5393 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5394 // The parameter is a vector value from earlier.
5395 Params.push_back(WidenMap.get(SrcInst));
5397 // The parameter is a scalar from outside the loop. Maybe even a constant.
5398 VectorParts Scalars;
5399 Scalars.append(UF, SrcOp);
5400 Params.push_back(Scalars);
5404 assert(Params.size() == Instr->getNumOperands() &&
5405 "Invalid number of operands");
5407 // Does this instruction return a value ?
5408 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5410 Value *UndefVec = IsVoidRetTy ? nullptr :
5411 UndefValue::get(Instr->getType());
5412 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5413 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5415 Instruction *InsertPt = Builder.GetInsertPoint();
5416 BasicBlock *IfBlock = Builder.GetInsertBlock();
5417 BasicBlock *CondBlock = nullptr;
5420 Loop *VectorLp = nullptr;
5421 if (IfPredicateStore) {
5422 assert(Instr->getParent()->getSinglePredecessor() &&
5423 "Only support single predecessor blocks");
5424 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5425 Instr->getParent());
5426 VectorLp = LI->getLoopFor(IfBlock);
5427 assert(VectorLp && "Must have a loop for this block");
5430 // For each vector unroll 'part':
5431 for (unsigned Part = 0; Part < UF; ++Part) {
5432 // For each scalar that we create:
5434 // Start an "if (pred) a[i] = ..." block.
5435 Value *Cmp = nullptr;
5436 if (IfPredicateStore) {
5437 if (Cond[Part]->getType()->isVectorTy())
5439 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5440 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5441 ConstantInt::get(Cond[Part]->getType(), 1));
5442 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5443 LoopVectorBody.push_back(CondBlock);
5444 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5445 // Update Builder with newly created basic block.
5446 Builder.SetInsertPoint(InsertPt);
5449 Instruction *Cloned = Instr->clone();
5451 Cloned->setName(Instr->getName() + ".cloned");
5452 // Replace the operands of the cloned instructions with extracted scalars.
5453 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5454 Value *Op = Params[op][Part];
5455 Cloned->setOperand(op, Op);
5458 // Place the cloned scalar in the new loop.
5459 Builder.Insert(Cloned);
5461 // If the original scalar returns a value we need to place it in a vector
5462 // so that future users will be able to use it.
5464 VecResults[Part] = Cloned;
5467 if (IfPredicateStore) {
5468 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5469 LoopVectorBody.push_back(NewIfBlock);
5470 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5471 Builder.SetInsertPoint(InsertPt);
5472 ReplaceInstWithInst(IfBlock->getTerminator(),
5473 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5474 IfBlock = NewIfBlock;
5479 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5480 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5481 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5483 return scalarizeInstruction(Instr, IfPredicateStore);
5486 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5490 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5494 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5495 // When unrolling and the VF is 1, we only need to add a simple scalar.
5496 Type *ITy = Val->getType();
5497 assert(!ITy->isVectorTy() && "Val must be a scalar");
5498 Constant *C = ConstantInt::get(ITy, StartIdx);
5499 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");