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 Legal(nullptr), AddedSafetyChecks(false) {}
279 // Perform the actual loop widening (vectorization).
280 void vectorize(LoopVectorizationLegality *L) {
282 // Create a new empty loop. Unlink the old loop and connect the new one.
284 // Widen each instruction in the old loop to a new one in the new loop.
285 // Use the Legality module to find the induction and reduction variables.
287 // Register the new loop and update the analysis passes.
291 // Return true if any runtime check is added.
292 bool IsSafetyChecksAdded() {
293 return AddedSafetyChecks;
296 virtual ~InnerLoopVectorizer() {}
299 /// A small list of PHINodes.
300 typedef SmallVector<PHINode*, 4> PhiVector;
301 /// When we unroll loops we have multiple vector values for each scalar.
302 /// This data structure holds the unrolled and vectorized values that
303 /// originated from one scalar instruction.
304 typedef SmallVector<Value*, 2> VectorParts;
306 // When we if-convert we need to create edge masks. We have to cache values
307 // so that we don't end up with exponential recursion/IR.
308 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
309 VectorParts> EdgeMaskCache;
311 /// \brief Add checks for strides that were assumed to be 1.
313 /// Returns the last check instruction and the first check instruction in the
314 /// pair as (first, last).
315 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
317 /// Create an empty loop, based on the loop ranges of the old loop.
318 void createEmptyLoop();
319 /// Create a new induction variable inside L.
320 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
321 Value *Step, Instruction *DL);
322 /// Copy and widen the instructions from the old loop.
323 virtual void vectorizeLoop();
325 /// \brief The Loop exit block may have single value PHI nodes where the
326 /// incoming value is 'Undef'. While vectorizing we only handled real values
327 /// that were defined inside the loop. Here we fix the 'undef case'.
331 /// A helper function that computes the predicate of the block BB, assuming
332 /// that the header block of the loop is set to True. It returns the *entry*
333 /// mask for the block BB.
334 VectorParts createBlockInMask(BasicBlock *BB);
335 /// A helper function that computes the predicate of the edge between SRC
337 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
339 /// A helper function to vectorize a single BB within the innermost loop.
340 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
342 /// Vectorize a single PHINode in a block. This method handles the induction
343 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
344 /// arbitrary length vectors.
345 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
346 unsigned UF, unsigned VF, PhiVector *PV);
348 /// Insert the new loop to the loop hierarchy and pass manager
349 /// and update the analysis passes.
350 void updateAnalysis();
352 /// This instruction is un-vectorizable. Implement it as a sequence
353 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
354 /// scalarized instruction behind an if block predicated on the control
355 /// dependence of the instruction.
356 virtual void scalarizeInstruction(Instruction *Instr,
357 bool IfPredicateStore=false);
359 /// Vectorize Load and Store instructions,
360 virtual void vectorizeMemoryInstruction(Instruction *Instr);
362 /// Create a broadcast instruction. This method generates a broadcast
363 /// instruction (shuffle) for loop invariant values and for the induction
364 /// value. If this is the induction variable then we extend it to N, N+1, ...
365 /// this is needed because each iteration in the loop corresponds to a SIMD
367 virtual Value *getBroadcastInstrs(Value *V);
369 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
370 /// to each vector element of Val. The sequence starts at StartIndex.
371 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
373 /// When we go over instructions in the basic block we rely on previous
374 /// values within the current basic block or on loop invariant values.
375 /// When we widen (vectorize) values we place them in the map. If the values
376 /// are not within the map, they have to be loop invariant, so we simply
377 /// broadcast them into a vector.
378 VectorParts &getVectorValue(Value *V);
380 /// Try to vectorize the interleaved access group that \p Instr belongs to.
381 void vectorizeInterleaveGroup(Instruction *Instr);
383 /// Generate a shuffle sequence that will reverse the vector Vec.
384 virtual Value *reverseVector(Value *Vec);
386 /// This is a helper class that holds the vectorizer state. It maps scalar
387 /// instructions to vector instructions. When the code is 'unrolled' then
388 /// then a single scalar value is mapped to multiple vector parts. The parts
389 /// are stored in the VectorPart type.
391 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
393 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
395 /// \return True if 'Key' is saved in the Value Map.
396 bool has(Value *Key) const { return MapStorage.count(Key); }
398 /// Initializes a new entry in the map. Sets all of the vector parts to the
399 /// save value in 'Val'.
400 /// \return A reference to a vector with splat values.
401 VectorParts &splat(Value *Key, Value *Val) {
402 VectorParts &Entry = MapStorage[Key];
403 Entry.assign(UF, Val);
407 ///\return A reference to the value that is stored at 'Key'.
408 VectorParts &get(Value *Key) {
409 VectorParts &Entry = MapStorage[Key];
412 assert(Entry.size() == UF);
417 /// The unroll factor. Each entry in the map stores this number of vector
421 /// Map storage. We use std::map and not DenseMap because insertions to a
422 /// dense map invalidates its iterators.
423 std::map<Value *, VectorParts> MapStorage;
426 /// The original loop.
428 /// Scev analysis to use.
436 /// Target Library Info.
437 const TargetLibraryInfo *TLI;
438 /// Target Transform Info.
439 const TargetTransformInfo *TTI;
441 /// The vectorization SIMD factor to use. Each vector will have this many
446 /// The vectorization unroll factor to use. Each scalar is vectorized to this
447 /// many different vector instructions.
450 /// The builder that we use
453 // --- Vectorization state ---
455 /// The vector-loop preheader.
456 BasicBlock *LoopVectorPreHeader;
457 /// The scalar-loop preheader.
458 BasicBlock *LoopScalarPreHeader;
459 /// Middle Block between the vector and the scalar.
460 BasicBlock *LoopMiddleBlock;
461 ///The ExitBlock of the scalar loop.
462 BasicBlock *LoopExitBlock;
463 ///The vector loop body.
464 SmallVector<BasicBlock *, 4> LoopVectorBody;
465 ///The scalar loop body.
466 BasicBlock *LoopScalarBody;
467 /// A list of all bypass blocks. The first block is the entry of the loop.
468 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
470 /// The new Induction variable which was added to the new block.
472 /// The induction variable of the old basic block.
473 PHINode *OldInduction;
474 /// Maps scalars to widened vectors.
476 EdgeMaskCache MaskCache;
478 LoopVectorizationLegality *Legal;
480 // Record whether runtime check is added.
481 bool AddedSafetyChecks;
484 class InnerLoopUnroller : public InnerLoopVectorizer {
486 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
487 DominatorTree *DT, const TargetLibraryInfo *TLI,
488 const TargetTransformInfo *TTI, unsigned UnrollFactor)
489 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
492 void scalarizeInstruction(Instruction *Instr,
493 bool IfPredicateStore = false) override;
494 void vectorizeMemoryInstruction(Instruction *Instr) override;
495 Value *getBroadcastInstrs(Value *V) override;
496 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
497 Value *reverseVector(Value *Vec) override;
500 /// \brief Look for a meaningful debug location on the instruction or it's
502 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
507 if (I->getDebugLoc() != Empty)
510 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
511 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
512 if (OpInst->getDebugLoc() != Empty)
519 /// \brief Set the debug location in the builder using the debug location in the
521 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
522 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
523 B.SetCurrentDebugLocation(Inst->getDebugLoc());
525 B.SetCurrentDebugLocation(DebugLoc());
529 /// \return string containing a file name and a line # for the given loop.
530 static std::string getDebugLocString(const Loop *L) {
533 raw_string_ostream OS(Result);
534 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
535 LoopDbgLoc.print(OS);
537 // Just print the module name.
538 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
545 /// \brief Propagate known metadata from one instruction to another.
546 static void propagateMetadata(Instruction *To, const Instruction *From) {
547 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
548 From->getAllMetadataOtherThanDebugLoc(Metadata);
550 for (auto M : Metadata) {
551 unsigned Kind = M.first;
553 // These are safe to transfer (this is safe for TBAA, even when we
554 // if-convert, because should that metadata have had a control dependency
555 // on the condition, and thus actually aliased with some other
556 // non-speculated memory access when the condition was false, this would be
557 // caught by the runtime overlap checks).
558 if (Kind != LLVMContext::MD_tbaa &&
559 Kind != LLVMContext::MD_alias_scope &&
560 Kind != LLVMContext::MD_noalias &&
561 Kind != LLVMContext::MD_fpmath &&
562 Kind != LLVMContext::MD_nontemporal)
565 To->setMetadata(Kind, M.second);
569 /// \brief Propagate known metadata from one instruction to a vector of others.
570 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
572 if (Instruction *I = dyn_cast<Instruction>(V))
573 propagateMetadata(I, From);
576 /// \brief The group of interleaved loads/stores sharing the same stride and
577 /// close to each other.
579 /// Each member in this group has an index starting from 0, and the largest
580 /// index should be less than interleaved factor, which is equal to the absolute
581 /// value of the access's stride.
583 /// E.g. An interleaved load group of factor 4:
584 /// for (unsigned i = 0; i < 1024; i+=4) {
585 /// a = A[i]; // Member of index 0
586 /// b = A[i+1]; // Member of index 1
587 /// d = A[i+3]; // Member of index 3
591 /// An interleaved store group of factor 4:
592 /// for (unsigned i = 0; i < 1024; i+=4) {
594 /// A[i] = a; // Member of index 0
595 /// A[i+1] = b; // Member of index 1
596 /// A[i+2] = c; // Member of index 2
597 /// A[i+3] = d; // Member of index 3
600 /// Note: the interleaved load group could have gaps (missing members), but
601 /// the interleaved store group doesn't allow gaps.
602 class InterleaveGroup {
604 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
605 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
606 assert(Align && "The alignment should be non-zero");
608 Factor = std::abs(Stride);
609 assert(Factor > 1 && "Invalid interleave factor");
611 Reverse = Stride < 0;
615 bool isReverse() const { return Reverse; }
616 unsigned getFactor() const { return Factor; }
617 unsigned getAlignment() const { return Align; }
618 unsigned getNumMembers() const { return Members.size(); }
620 /// \brief Try to insert a new member \p Instr with index \p Index and
621 /// alignment \p NewAlign. The index is related to the leader and it could be
622 /// negative if it is the new leader.
624 /// \returns false if the instruction doesn't belong to the group.
625 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
626 assert(NewAlign && "The new member's alignment should be non-zero");
628 int Key = Index + SmallestKey;
630 // Skip if there is already a member with the same index.
631 if (Members.count(Key))
634 if (Key > LargestKey) {
635 // The largest index is always less than the interleave factor.
636 if (Index >= static_cast<int>(Factor))
640 } else if (Key < SmallestKey) {
641 // The largest index is always less than the interleave factor.
642 if (LargestKey - Key >= static_cast<int>(Factor))
648 // It's always safe to select the minimum alignment.
649 Align = std::min(Align, NewAlign);
650 Members[Key] = Instr;
654 /// \brief Get the member with the given index \p Index
656 /// \returns nullptr if contains no such member.
657 Instruction *getMember(unsigned Index) const {
658 int Key = SmallestKey + Index;
659 if (!Members.count(Key))
662 return Members.find(Key)->second;
665 /// \brief Get the index for the given member. Unlike the key in the member
666 /// map, the index starts from 0.
667 unsigned getIndex(Instruction *Instr) const {
668 for (auto I : Members)
669 if (I.second == Instr)
670 return I.first - SmallestKey;
672 llvm_unreachable("InterleaveGroup contains no such member");
675 Instruction *getInsertPos() const { return InsertPos; }
676 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
679 unsigned Factor; // Interleave Factor.
682 DenseMap<int, Instruction *> Members;
686 // To avoid breaking dependences, vectorized instructions of an interleave
687 // group should be inserted at either the first load or the last store in
690 // E.g. %even = load i32 // Insert Position
691 // %add = add i32 %even // Use of %even
695 // %odd = add i32 // Def of %odd
696 // store i32 %odd // Insert Position
697 Instruction *InsertPos;
700 /// \brief Drive the analysis of interleaved memory accesses in the loop.
702 /// Use this class to analyze interleaved accesses only when we can vectorize
703 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
704 /// on interleaved accesses is unsafe.
706 /// The analysis collects interleave groups and records the relationships
707 /// between the member and the group in a map.
708 class InterleavedAccessInfo {
710 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
711 : SE(SE), TheLoop(L), DT(DT) {}
713 ~InterleavedAccessInfo() {
714 SmallSet<InterleaveGroup *, 4> DelSet;
715 // Avoid releasing a pointer twice.
716 for (auto &I : InterleaveGroupMap)
717 DelSet.insert(I.second);
718 for (auto *Ptr : DelSet)
722 /// \brief Analyze the interleaved accesses and collect them in interleave
723 /// groups. Substitute symbolic strides using \p Strides.
724 void analyzeInterleaving(const ValueToValueMap &Strides);
726 /// \brief Check if \p Instr belongs to any interleave group.
727 bool isInterleaved(Instruction *Instr) const {
728 return InterleaveGroupMap.count(Instr);
731 /// \brief Get the interleave group that \p Instr belongs to.
733 /// \returns nullptr if doesn't have such group.
734 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
735 if (InterleaveGroupMap.count(Instr))
736 return InterleaveGroupMap.find(Instr)->second;
745 /// Holds the relationships between the members and the interleave group.
746 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
748 /// \brief The descriptor for a strided memory access.
749 struct StrideDescriptor {
750 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
752 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
754 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
756 int Stride; // The access's stride. It is negative for a reverse access.
757 const SCEV *Scev; // The scalar expression of this access
758 unsigned Size; // The size of the memory object.
759 unsigned Align; // The alignment of this access.
762 /// \brief Create a new interleave group with the given instruction \p Instr,
763 /// stride \p Stride and alignment \p Align.
765 /// \returns the newly created interleave group.
766 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
768 assert(!InterleaveGroupMap.count(Instr) &&
769 "Already in an interleaved access group");
770 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
771 return InterleaveGroupMap[Instr];
774 /// \brief Release the group and remove all the relationships.
775 void releaseGroup(InterleaveGroup *Group) {
776 for (unsigned i = 0; i < Group->getFactor(); i++)
777 if (Instruction *Member = Group->getMember(i))
778 InterleaveGroupMap.erase(Member);
783 /// \brief Collect all the accesses with a constant stride in program order.
784 void collectConstStridedAccesses(
785 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
786 const ValueToValueMap &Strides);
789 /// Utility class for getting and setting loop vectorizer hints in the form
790 /// of loop metadata.
791 /// This class keeps a number of loop annotations locally (as member variables)
792 /// and can, upon request, write them back as metadata on the loop. It will
793 /// initially scan the loop for existing metadata, and will update the local
794 /// values based on information in the loop.
795 /// We cannot write all values to metadata, as the mere presence of some info,
796 /// for example 'force', means a decision has been made. So, we need to be
797 /// careful NOT to add them if the user hasn't specifically asked so.
798 class LoopVectorizeHints {
805 /// Hint - associates name and validation with the hint value.
808 unsigned Value; // This may have to change for non-numeric values.
811 Hint(const char * Name, unsigned Value, HintKind Kind)
812 : Name(Name), Value(Value), Kind(Kind) { }
814 bool validate(unsigned Val) {
817 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
819 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
827 /// Vectorization width.
829 /// Vectorization interleave factor.
831 /// Vectorization forced
834 /// Return the loop metadata prefix.
835 static StringRef Prefix() { return "llvm.loop."; }
839 FK_Undefined = -1, ///< Not selected.
840 FK_Disabled = 0, ///< Forcing disabled.
841 FK_Enabled = 1, ///< Forcing enabled.
844 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
845 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
847 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
848 Force("vectorize.enable", FK_Undefined, HK_FORCE),
850 // Populate values with existing loop metadata.
851 getHintsFromMetadata();
853 // force-vector-interleave overrides DisableInterleaving.
854 if (VectorizerParams::isInterleaveForced())
855 Interleave.Value = VectorizerParams::VectorizationInterleave;
857 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
858 << "LV: Interleaving disabled by the pass manager\n");
861 /// Mark the loop L as already vectorized by setting the width to 1.
862 void setAlreadyVectorized() {
863 Width.Value = Interleave.Value = 1;
864 Hint Hints[] = {Width, Interleave};
865 writeHintsToMetadata(Hints);
868 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
869 if (getForce() == LoopVectorizeHints::FK_Disabled) {
870 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
871 emitOptimizationRemarkAnalysis(F->getContext(),
872 vectorizeAnalysisPassName(), *F,
873 L->getStartLoc(), emitRemark());
877 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
878 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
879 emitOptimizationRemarkAnalysis(F->getContext(),
880 vectorizeAnalysisPassName(), *F,
881 L->getStartLoc(), emitRemark());
885 if (getWidth() == 1 && getInterleave() == 1) {
886 // FIXME: Add a separate metadata to indicate when the loop has already
887 // been vectorized instead of setting width and count to 1.
888 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
889 // FIXME: Add interleave.disable metadata. This will allow
890 // vectorize.disable to be used without disabling the pass and errors
891 // to differentiate between disabled vectorization and a width of 1.
892 emitOptimizationRemarkAnalysis(
893 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
894 "loop not vectorized: vectorization and interleaving are explicitly "
895 "disabled, or vectorize width and interleave count are both set to "
903 /// Dumps all the hint information.
904 std::string emitRemark() const {
905 VectorizationReport R;
906 if (Force.Value == LoopVectorizeHints::FK_Disabled)
907 R << "vectorization is explicitly disabled";
909 R << "use -Rpass-analysis=loop-vectorize for more info";
910 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
912 if (Width.Value != 0)
913 R << ", Vector Width=" << Width.Value;
914 if (Interleave.Value != 0)
915 R << ", Interleave Count=" << Interleave.Value;
923 unsigned getWidth() const { return Width.Value; }
924 unsigned getInterleave() const { return Interleave.Value; }
925 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
926 const char *vectorizeAnalysisPassName() const {
927 // If hints are provided that don't disable vectorization use the
928 // AlwaysPrint pass name to force the frontend to print the diagnostic.
931 if (getForce() == LoopVectorizeHints::FK_Disabled)
933 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
935 return DiagnosticInfo::AlwaysPrint;
938 bool allowReordering() const {
939 // When enabling loop hints are provided we allow the vectorizer to change
940 // the order of operations that is given by the scalar loop. This is not
941 // enabled by default because can be unsafe or inefficient. For example,
942 // reordering floating-point operations will change the way round-off
943 // error accumulates in the loop.
944 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
948 /// Find hints specified in the loop metadata and update local values.
949 void getHintsFromMetadata() {
950 MDNode *LoopID = TheLoop->getLoopID();
954 // First operand should refer to the loop id itself.
955 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
956 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
958 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
959 const MDString *S = nullptr;
960 SmallVector<Metadata *, 4> Args;
962 // The expected hint is either a MDString or a MDNode with the first
963 // operand a MDString.
964 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
965 if (!MD || MD->getNumOperands() == 0)
967 S = dyn_cast<MDString>(MD->getOperand(0));
968 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
969 Args.push_back(MD->getOperand(i));
971 S = dyn_cast<MDString>(LoopID->getOperand(i));
972 assert(Args.size() == 0 && "too many arguments for MDString");
978 // Check if the hint starts with the loop metadata prefix.
979 StringRef Name = S->getString();
980 if (Args.size() == 1)
981 setHint(Name, Args[0]);
985 /// Checks string hint with one operand and set value if valid.
986 void setHint(StringRef Name, Metadata *Arg) {
987 if (!Name.startswith(Prefix()))
989 Name = Name.substr(Prefix().size(), StringRef::npos);
991 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
993 unsigned Val = C->getZExtValue();
995 Hint *Hints[] = {&Width, &Interleave, &Force};
996 for (auto H : Hints) {
997 if (Name == H->Name) {
998 if (H->validate(Val))
1001 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1007 /// Create a new hint from name / value pair.
1008 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1009 LLVMContext &Context = TheLoop->getHeader()->getContext();
1010 Metadata *MDs[] = {MDString::get(Context, Name),
1011 ConstantAsMetadata::get(
1012 ConstantInt::get(Type::getInt32Ty(Context), V))};
1013 return MDNode::get(Context, MDs);
1016 /// Matches metadata with hint name.
1017 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1018 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1022 for (auto H : HintTypes)
1023 if (Name->getString().endswith(H.Name))
1028 /// Sets current hints into loop metadata, keeping other values intact.
1029 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1030 if (HintTypes.size() == 0)
1033 // Reserve the first element to LoopID (see below).
1034 SmallVector<Metadata *, 4> MDs(1);
1035 // If the loop already has metadata, then ignore the existing operands.
1036 MDNode *LoopID = TheLoop->getLoopID();
1038 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1039 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1040 // If node in update list, ignore old value.
1041 if (!matchesHintMetadataName(Node, HintTypes))
1042 MDs.push_back(Node);
1046 // Now, add the missing hints.
1047 for (auto H : HintTypes)
1048 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1050 // Replace current metadata node with new one.
1051 LLVMContext &Context = TheLoop->getHeader()->getContext();
1052 MDNode *NewLoopID = MDNode::get(Context, MDs);
1053 // Set operand 0 to refer to the loop id itself.
1054 NewLoopID->replaceOperandWith(0, NewLoopID);
1056 TheLoop->setLoopID(NewLoopID);
1059 /// The loop these hints belong to.
1060 const Loop *TheLoop;
1063 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1064 const LoopVectorizeHints &Hints,
1065 const LoopAccessReport &Message) {
1066 const char *Name = Hints.vectorizeAnalysisPassName();
1067 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1070 static void emitMissedWarning(Function *F, Loop *L,
1071 const LoopVectorizeHints &LH) {
1072 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1075 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1076 if (LH.getWidth() != 1)
1077 emitLoopVectorizeWarning(
1078 F->getContext(), *F, L->getStartLoc(),
1079 "failed explicitly specified loop vectorization");
1080 else if (LH.getInterleave() != 1)
1081 emitLoopInterleaveWarning(
1082 F->getContext(), *F, L->getStartLoc(),
1083 "failed explicitly specified loop interleaving");
1087 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1088 /// to what vectorization factor.
1089 /// This class does not look at the profitability of vectorization, only the
1090 /// legality. This class has two main kinds of checks:
1091 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1092 /// will change the order of memory accesses in a way that will change the
1093 /// correctness of the program.
1094 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1095 /// checks for a number of different conditions, such as the availability of a
1096 /// single induction variable, that all types are supported and vectorize-able,
1097 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1098 /// This class is also used by InnerLoopVectorizer for identifying
1099 /// induction variable and the different reduction variables.
1100 class LoopVectorizationLegality {
1102 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1103 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1104 Function *F, const TargetTransformInfo *TTI,
1105 LoopAccessAnalysis *LAA,
1106 LoopVectorizationRequirements *R,
1107 const LoopVectorizeHints *H)
1108 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1109 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1110 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1111 Requirements(R), Hints(H) {}
1113 /// ReductionList contains the reduction descriptors for all
1114 /// of the reductions that were found in the loop.
1115 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1117 /// InductionList saves induction variables and maps them to the
1118 /// induction descriptor.
1119 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1121 /// Returns true if it is legal to vectorize this loop.
1122 /// This does not mean that it is profitable to vectorize this
1123 /// loop, only that it is legal to do so.
1124 bool canVectorize();
1126 /// Returns the Induction variable.
1127 PHINode *getInduction() { return Induction; }
1129 /// Returns the reduction variables found in the loop.
1130 ReductionList *getReductionVars() { return &Reductions; }
1132 /// Returns the induction variables found in the loop.
1133 InductionList *getInductionVars() { return &Inductions; }
1135 /// Returns the widest induction type.
1136 Type *getWidestInductionType() { return WidestIndTy; }
1138 /// Returns True if V is an induction variable in this loop.
1139 bool isInductionVariable(const Value *V);
1141 /// Return true if the block BB needs to be predicated in order for the loop
1142 /// to be vectorized.
1143 bool blockNeedsPredication(BasicBlock *BB);
1145 /// Check if this pointer is consecutive when vectorizing. This happens
1146 /// when the last index of the GEP is the induction variable, or that the
1147 /// pointer itself is an induction variable.
1148 /// This check allows us to vectorize A[idx] into a wide load/store.
1150 /// 0 - Stride is unknown or non-consecutive.
1151 /// 1 - Address is consecutive.
1152 /// -1 - Address is consecutive, and decreasing.
1153 int isConsecutivePtr(Value *Ptr);
1155 /// Returns true if the value V is uniform within the loop.
1156 bool isUniform(Value *V);
1158 /// Returns true if this instruction will remain scalar after vectorization.
1159 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1161 /// Returns the information that we collected about runtime memory check.
1162 const RuntimePointerChecking *getRuntimePointerChecking() const {
1163 return LAI->getRuntimePointerChecking();
1166 const LoopAccessInfo *getLAI() const {
1170 /// \brief Check if \p Instr belongs to any interleaved access group.
1171 bool isAccessInterleaved(Instruction *Instr) {
1172 return InterleaveInfo.isInterleaved(Instr);
1175 /// \brief Get the interleaved access group that \p Instr belongs to.
1176 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1177 return InterleaveInfo.getInterleaveGroup(Instr);
1180 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1182 bool hasStride(Value *V) { return StrideSet.count(V); }
1183 bool mustCheckStrides() { return !StrideSet.empty(); }
1184 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1185 return StrideSet.begin();
1187 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1189 /// Returns true if the target machine supports masked store operation
1190 /// for the given \p DataType and kind of access to \p Ptr.
1191 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1192 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1194 /// Returns true if the target machine supports masked load operation
1195 /// for the given \p DataType and kind of access to \p Ptr.
1196 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1197 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1199 /// Returns true if vector representation of the instruction \p I
1201 bool isMaskRequired(const Instruction* I) {
1202 return (MaskedOp.count(I) != 0);
1204 unsigned getNumStores() const {
1205 return LAI->getNumStores();
1207 unsigned getNumLoads() const {
1208 return LAI->getNumLoads();
1210 unsigned getNumPredStores() const {
1211 return NumPredStores;
1214 /// Check if a single basic block loop is vectorizable.
1215 /// At this point we know that this is a loop with a constant trip count
1216 /// and we only need to check individual instructions.
1217 bool canVectorizeInstrs();
1219 /// When we vectorize loops we may change the order in which
1220 /// we read and write from memory. This method checks if it is
1221 /// legal to vectorize the code, considering only memory constrains.
1222 /// Returns true if the loop is vectorizable
1223 bool canVectorizeMemory();
1225 /// Return true if we can vectorize this loop using the IF-conversion
1227 bool canVectorizeWithIfConvert();
1229 /// Collect the variables that need to stay uniform after vectorization.
1230 void collectLoopUniforms();
1232 /// Return true if all of the instructions in the block can be speculatively
1233 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1234 /// and we know that we can read from them without segfault.
1235 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1237 /// \brief Collect memory access with loop invariant strides.
1239 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1241 void collectStridedAccess(Value *LoadOrStoreInst);
1243 /// Report an analysis message to assist the user in diagnosing loops that are
1244 /// not vectorized. These are handled as LoopAccessReport rather than
1245 /// VectorizationReport because the << operator of VectorizationReport returns
1246 /// LoopAccessReport.
1247 void emitAnalysis(const LoopAccessReport &Message) const {
1248 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1251 unsigned NumPredStores;
1253 /// The loop that we evaluate.
1256 ScalarEvolution *SE;
1257 /// Target Library Info.
1258 TargetLibraryInfo *TLI;
1260 Function *TheFunction;
1261 /// Target Transform Info
1262 const TargetTransformInfo *TTI;
1265 // LoopAccess analysis.
1266 LoopAccessAnalysis *LAA;
1267 // And the loop-accesses info corresponding to this loop. This pointer is
1268 // null until canVectorizeMemory sets it up.
1269 const LoopAccessInfo *LAI;
1271 /// The interleave access information contains groups of interleaved accesses
1272 /// with the same stride and close to each other.
1273 InterleavedAccessInfo InterleaveInfo;
1275 // --- vectorization state --- //
1277 /// Holds the integer induction variable. This is the counter of the
1280 /// Holds the reduction variables.
1281 ReductionList Reductions;
1282 /// Holds all of the induction variables that we found in the loop.
1283 /// Notice that inductions don't need to start at zero and that induction
1284 /// variables can be pointers.
1285 InductionList Inductions;
1286 /// Holds the widest induction type encountered.
1289 /// Allowed outside users. This holds the reduction
1290 /// vars which can be accessed from outside the loop.
1291 SmallPtrSet<Value*, 4> AllowedExit;
1292 /// This set holds the variables which are known to be uniform after
1294 SmallPtrSet<Instruction*, 4> Uniforms;
1296 /// Can we assume the absence of NaNs.
1297 bool HasFunNoNaNAttr;
1299 /// Vectorization requirements that will go through late-evaluation.
1300 LoopVectorizationRequirements *Requirements;
1302 /// Used to emit an analysis of any legality issues.
1303 const LoopVectorizeHints *Hints;
1305 ValueToValueMap Strides;
1306 SmallPtrSet<Value *, 8> StrideSet;
1308 /// While vectorizing these instructions we have to generate a
1309 /// call to the appropriate masked intrinsic
1310 SmallPtrSet<const Instruction*, 8> MaskedOp;
1313 /// LoopVectorizationCostModel - estimates the expected speedups due to
1315 /// In many cases vectorization is not profitable. This can happen because of
1316 /// a number of reasons. In this class we mainly attempt to predict the
1317 /// expected speedup/slowdowns due to the supported instruction set. We use the
1318 /// TargetTransformInfo to query the different backends for the cost of
1319 /// different operations.
1320 class LoopVectorizationCostModel {
1322 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1323 LoopVectorizationLegality *Legal,
1324 const TargetTransformInfo &TTI,
1325 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1326 const Function *F, const LoopVectorizeHints *Hints,
1327 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1328 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1329 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1331 /// Information about vectorization costs
1332 struct VectorizationFactor {
1333 unsigned Width; // Vector width with best cost
1334 unsigned Cost; // Cost of the loop with that width
1336 /// \return The most profitable vectorization factor and the cost of that VF.
1337 /// This method checks every power of two up to VF. If UserVF is not ZERO
1338 /// then this vectorization factor will be selected if vectorization is
1340 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1342 /// \return The size (in bits) of the widest type in the code that
1343 /// needs to be vectorized. We ignore values that remain scalar such as
1344 /// 64 bit loop indices.
1345 unsigned getWidestType();
1347 /// \return The desired interleave count.
1348 /// If interleave count has been specified by metadata it will be returned.
1349 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1350 /// are the selected vectorization factor and the cost of the selected VF.
1351 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1354 /// \return The most profitable unroll factor.
1355 /// This method finds the best unroll-factor based on register pressure and
1356 /// other parameters. VF and LoopCost are the selected vectorization factor
1357 /// and the cost of the selected VF.
1358 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1361 /// \brief A struct that represents some properties of the register usage
1363 struct RegisterUsage {
1364 /// Holds the number of loop invariant values that are used in the loop.
1365 unsigned LoopInvariantRegs;
1366 /// Holds the maximum number of concurrent live intervals in the loop.
1367 unsigned MaxLocalUsers;
1368 /// Holds the number of instructions in the loop.
1369 unsigned NumInstructions;
1372 /// \return information about the register usage of the loop.
1373 RegisterUsage calculateRegisterUsage();
1376 /// Returns the expected execution cost. The unit of the cost does
1377 /// not matter because we use the 'cost' units to compare different
1378 /// vector widths. The cost that is returned is *not* normalized by
1379 /// the factor width.
1380 unsigned expectedCost(unsigned VF);
1382 /// Returns the execution time cost of an instruction for a given vector
1383 /// width. Vector width of one means scalar.
1384 unsigned getInstructionCost(Instruction *I, unsigned VF);
1386 /// Returns whether the instruction is a load or store and will be a emitted
1387 /// as a vector operation.
1388 bool isConsecutiveLoadOrStore(Instruction *I);
1390 /// Report an analysis message to assist the user in diagnosing loops that are
1391 /// not vectorized. These are handled as LoopAccessReport rather than
1392 /// VectorizationReport because the << operator of VectorizationReport returns
1393 /// LoopAccessReport.
1394 void emitAnalysis(const LoopAccessReport &Message) const {
1395 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1398 /// The loop that we evaluate.
1401 ScalarEvolution *SE;
1402 /// Loop Info analysis.
1404 /// Vectorization legality.
1405 LoopVectorizationLegality *Legal;
1406 /// Vector target information.
1407 const TargetTransformInfo &TTI;
1408 /// Target Library Info.
1409 const TargetLibraryInfo *TLI;
1410 const Function *TheFunction;
1411 // Loop Vectorize Hint.
1412 const LoopVectorizeHints *Hints;
1413 // Values to ignore in the cost model.
1414 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1417 /// \brief This holds vectorization requirements that must be verified late in
1418 /// the process. The requirements are set by legalize and costmodel. Once
1419 /// vectorization has been determined to be possible and profitable the
1420 /// requirements can be verified by looking for metadata or compiler options.
1421 /// For example, some loops require FP commutativity which is only allowed if
1422 /// vectorization is explicitly specified or if the fast-math compiler option
1423 /// has been provided.
1424 /// Late evaluation of these requirements allows helpful diagnostics to be
1425 /// composed that tells the user what need to be done to vectorize the loop. For
1426 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1427 /// evaluation should be used only when diagnostics can generated that can be
1428 /// followed by a non-expert user.
1429 class LoopVectorizationRequirements {
1431 LoopVectorizationRequirements()
1432 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1434 void addUnsafeAlgebraInst(Instruction *I) {
1435 // First unsafe algebra instruction.
1436 if (!UnsafeAlgebraInst)
1437 UnsafeAlgebraInst = I;
1440 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1442 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1443 const char *Name = Hints.vectorizeAnalysisPassName();
1444 bool Failed = false;
1445 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1446 emitOptimizationRemarkAnalysisFPCommute(
1447 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1448 VectorizationReport() << "cannot prove it is safe to reorder "
1449 "floating-point operations");
1453 // Test if runtime memcheck thresholds are exceeded.
1454 bool PragmaThresholdReached =
1455 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1456 bool ThresholdReached =
1457 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1458 if ((ThresholdReached && !Hints.allowReordering()) ||
1459 PragmaThresholdReached) {
1460 emitOptimizationRemarkAnalysisAliasing(
1461 F->getContext(), Name, *F, L->getStartLoc(),
1462 VectorizationReport()
1463 << "cannot prove it is safe to reorder memory operations");
1464 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1472 unsigned NumRuntimePointerChecks;
1473 Instruction *UnsafeAlgebraInst;
1476 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1478 return V.push_back(&L);
1480 for (Loop *InnerL : L)
1481 addInnerLoop(*InnerL, V);
1484 /// The LoopVectorize Pass.
1485 struct LoopVectorize : public FunctionPass {
1486 /// Pass identification, replacement for typeid
1489 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1491 DisableUnrolling(NoUnrolling),
1492 AlwaysVectorize(AlwaysVectorize) {
1493 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1496 ScalarEvolution *SE;
1498 TargetTransformInfo *TTI;
1500 BlockFrequencyInfo *BFI;
1501 TargetLibraryInfo *TLI;
1503 AssumptionCache *AC;
1504 LoopAccessAnalysis *LAA;
1505 bool DisableUnrolling;
1506 bool AlwaysVectorize;
1508 BlockFrequency ColdEntryFreq;
1510 bool runOnFunction(Function &F) override {
1511 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1512 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1513 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1514 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1515 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1516 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1517 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1518 AA = &getAnalysis<AliasAnalysis>();
1519 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1520 LAA = &getAnalysis<LoopAccessAnalysis>();
1522 // Compute some weights outside of the loop over the loops. Compute this
1523 // using a BranchProbability to re-use its scaling math.
1524 const BranchProbability ColdProb(1, 5); // 20%
1525 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1528 // 1. the target claims to have no vector registers, and
1529 // 2. interleaving won't help ILP.
1531 // The second condition is necessary because, even if the target has no
1532 // vector registers, loop vectorization may still enable scalar
1534 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1537 // Build up a worklist of inner-loops to vectorize. This is necessary as
1538 // the act of vectorizing or partially unrolling a loop creates new loops
1539 // and can invalidate iterators across the loops.
1540 SmallVector<Loop *, 8> Worklist;
1543 addInnerLoop(*L, Worklist);
1545 LoopsAnalyzed += Worklist.size();
1547 // Now walk the identified inner loops.
1548 bool Changed = false;
1549 while (!Worklist.empty())
1550 Changed |= processLoop(Worklist.pop_back_val());
1552 // Process each loop nest in the function.
1556 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1557 SmallVector<Metadata *, 4> MDs;
1558 // Reserve first location for self reference to the LoopID metadata node.
1559 MDs.push_back(nullptr);
1560 bool IsUnrollMetadata = false;
1561 MDNode *LoopID = L->getLoopID();
1563 // First find existing loop unrolling disable metadata.
1564 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1565 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1567 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1569 S && S->getString().startswith("llvm.loop.unroll.disable");
1571 MDs.push_back(LoopID->getOperand(i));
1575 if (!IsUnrollMetadata) {
1576 // Add runtime unroll disable metadata.
1577 LLVMContext &Context = L->getHeader()->getContext();
1578 SmallVector<Metadata *, 1> DisableOperands;
1579 DisableOperands.push_back(
1580 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1581 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1582 MDs.push_back(DisableNode);
1583 MDNode *NewLoopID = MDNode::get(Context, MDs);
1584 // Set operand 0 to refer to the loop id itself.
1585 NewLoopID->replaceOperandWith(0, NewLoopID);
1586 L->setLoopID(NewLoopID);
1590 bool processLoop(Loop *L) {
1591 assert(L->empty() && "Only process inner loops.");
1594 const std::string DebugLocStr = getDebugLocString(L);
1597 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1598 << L->getHeader()->getParent()->getName() << "\" from "
1599 << DebugLocStr << "\n");
1601 LoopVectorizeHints Hints(L, DisableUnrolling);
1603 DEBUG(dbgs() << "LV: Loop hints:"
1605 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1607 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1609 : "?")) << " width=" << Hints.getWidth()
1610 << " unroll=" << Hints.getInterleave() << "\n");
1612 // Function containing loop
1613 Function *F = L->getHeader()->getParent();
1615 // Looking at the diagnostic output is the only way to determine if a loop
1616 // was vectorized (other than looking at the IR or machine code), so it
1617 // is important to generate an optimization remark for each loop. Most of
1618 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1619 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1620 // less verbose reporting vectorized loops and unvectorized loops that may
1621 // benefit from vectorization, respectively.
1623 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1624 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1628 // Check the loop for a trip count threshold:
1629 // do not vectorize loops with a tiny trip count.
1630 const unsigned TC = SE->getSmallConstantTripCount(L);
1631 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1632 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1633 << "This loop is not worth vectorizing.");
1634 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1635 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1637 DEBUG(dbgs() << "\n");
1638 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1639 << "vectorization is not beneficial "
1640 "and is not explicitly forced");
1645 // Check if it is legal to vectorize the loop.
1646 LoopVectorizationRequirements Requirements;
1647 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1648 &Requirements, &Hints);
1649 if (!LVL.canVectorize()) {
1650 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1651 emitMissedWarning(F, L, Hints);
1655 // Collect values we want to ignore in the cost model. This includes
1656 // type-promoting instructions we identified during reduction detection.
1657 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1658 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1659 for (auto &Reduction : *LVL.getReductionVars()) {
1660 RecurrenceDescriptor &RedDes = Reduction.second;
1661 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1662 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1665 // Use the cost model.
1666 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints,
1669 // Check the function attributes to find out if this function should be
1670 // optimized for size.
1671 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1674 // Compute the weighted frequency of this loop being executed and see if it
1675 // is less than 20% of the function entry baseline frequency. Note that we
1676 // always have a canonical loop here because we think we *can* vectorize.
1677 // FIXME: This is hidden behind a flag due to pervasive problems with
1678 // exactly what block frequency models.
1679 if (LoopVectorizeWithBlockFrequency) {
1680 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1681 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1682 LoopEntryFreq < ColdEntryFreq)
1686 // Check the function attributes to see if implicit floats are allowed.
1687 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1688 // an integer loop and the vector instructions selected are purely integer
1689 // vector instructions?
1690 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1691 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1692 "attribute is used.\n");
1695 VectorizationReport()
1696 << "loop not vectorized due to NoImplicitFloat attribute");
1697 emitMissedWarning(F, L, Hints);
1701 // Select the optimal vectorization factor.
1702 const LoopVectorizationCostModel::VectorizationFactor VF =
1703 CM.selectVectorizationFactor(OptForSize);
1705 // Select the interleave count.
1706 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1708 // Get user interleave count.
1709 unsigned UserIC = Hints.getInterleave();
1711 // Identify the diagnostic messages that should be produced.
1712 std::string VecDiagMsg, IntDiagMsg;
1713 bool VectorizeLoop = true, InterleaveLoop = true;
1715 if (Requirements.doesNotMeet(F, L, Hints)) {
1716 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1718 emitMissedWarning(F, L, Hints);
1722 if (VF.Width == 1) {
1723 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1725 "the cost-model indicates that vectorization is not beneficial";
1726 VectorizeLoop = false;
1729 if (IC == 1 && UserIC <= 1) {
1730 // Tell the user interleaving is not beneficial.
1731 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1733 "the cost-model indicates that interleaving is not beneficial";
1734 InterleaveLoop = false;
1737 " and is explicitly disabled or interleave count is set to 1";
1738 } else if (IC > 1 && UserIC == 1) {
1739 // Tell the user interleaving is beneficial, but it explicitly disabled.
1741 << "LV: Interleaving is beneficial but is explicitly disabled.");
1742 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1743 "but is explicitly disabled or interleave count is set to 1";
1744 InterleaveLoop = false;
1747 // Override IC if user provided an interleave count.
1748 IC = UserIC > 0 ? UserIC : IC;
1750 // Emit diagnostic messages, if any.
1751 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1752 if (!VectorizeLoop && !InterleaveLoop) {
1753 // Do not vectorize or interleaving the loop.
1754 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1755 L->getStartLoc(), VecDiagMsg);
1756 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1757 L->getStartLoc(), IntDiagMsg);
1759 } else if (!VectorizeLoop && InterleaveLoop) {
1760 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1761 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1762 L->getStartLoc(), VecDiagMsg);
1763 } else if (VectorizeLoop && !InterleaveLoop) {
1764 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1765 << DebugLocStr << '\n');
1766 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1767 L->getStartLoc(), IntDiagMsg);
1768 } else if (VectorizeLoop && InterleaveLoop) {
1769 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1770 << DebugLocStr << '\n');
1771 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1774 if (!VectorizeLoop) {
1775 assert(IC > 1 && "interleave count should not be 1 or 0");
1776 // If we decided that it is not legal to vectorize the loop then
1778 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1779 Unroller.vectorize(&LVL);
1781 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1782 Twine("interleaved loop (interleaved count: ") +
1785 // If we decided that it is *legal* to vectorize the loop then do it.
1786 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1790 // Add metadata to disable runtime unrolling scalar loop when there's no
1791 // runtime check about strides and memory. Because at this situation,
1792 // scalar loop is rarely used not worthy to be unrolled.
1793 if (!LB.IsSafetyChecksAdded())
1794 AddRuntimeUnrollDisableMetaData(L);
1796 // Report the vectorization decision.
1797 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1798 Twine("vectorized loop (vectorization width: ") +
1799 Twine(VF.Width) + ", interleaved count: " +
1803 // Mark the loop as already vectorized to avoid vectorizing again.
1804 Hints.setAlreadyVectorized();
1806 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1810 void getAnalysisUsage(AnalysisUsage &AU) const override {
1811 AU.addRequired<AssumptionCacheTracker>();
1812 AU.addRequiredID(LoopSimplifyID);
1813 AU.addRequiredID(LCSSAID);
1814 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1815 AU.addRequired<DominatorTreeWrapperPass>();
1816 AU.addRequired<LoopInfoWrapperPass>();
1817 AU.addRequired<ScalarEvolutionWrapperPass>();
1818 AU.addRequired<TargetTransformInfoWrapperPass>();
1819 AU.addRequired<AliasAnalysis>();
1820 AU.addRequired<LoopAccessAnalysis>();
1821 AU.addPreserved<LoopInfoWrapperPass>();
1822 AU.addPreserved<DominatorTreeWrapperPass>();
1823 AU.addPreserved<AliasAnalysis>();
1828 } // end anonymous namespace
1830 //===----------------------------------------------------------------------===//
1831 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1832 // LoopVectorizationCostModel.
1833 //===----------------------------------------------------------------------===//
1835 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1836 // We need to place the broadcast of invariant variables outside the loop.
1837 Instruction *Instr = dyn_cast<Instruction>(V);
1839 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1840 Instr->getParent()) != LoopVectorBody.end());
1841 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1843 // Place the code for broadcasting invariant variables in the new preheader.
1844 IRBuilder<>::InsertPointGuard Guard(Builder);
1846 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1848 // Broadcast the scalar into all locations in the vector.
1849 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1854 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1856 assert(Val->getType()->isVectorTy() && "Must be a vector");
1857 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1858 "Elem must be an integer");
1859 assert(Step->getType() == Val->getType()->getScalarType() &&
1860 "Step has wrong type");
1861 // Create the types.
1862 Type *ITy = Val->getType()->getScalarType();
1863 VectorType *Ty = cast<VectorType>(Val->getType());
1864 int VLen = Ty->getNumElements();
1865 SmallVector<Constant*, 8> Indices;
1867 // Create a vector of consecutive numbers from zero to VF.
1868 for (int i = 0; i < VLen; ++i)
1869 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1871 // Add the consecutive indices to the vector value.
1872 Constant *Cv = ConstantVector::get(Indices);
1873 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1874 Step = Builder.CreateVectorSplat(VLen, Step);
1875 assert(Step->getType() == Val->getType() && "Invalid step vec");
1876 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1877 // which can be found from the original scalar operations.
1878 Step = Builder.CreateMul(Cv, Step);
1879 return Builder.CreateAdd(Val, Step, "induction");
1882 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1883 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1884 // Make sure that the pointer does not point to structs.
1885 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1888 // If this value is a pointer induction variable we know it is consecutive.
1889 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1890 if (Phi && Inductions.count(Phi)) {
1891 InductionDescriptor II = Inductions[Phi];
1892 return II.getConsecutiveDirection();
1895 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1899 unsigned NumOperands = Gep->getNumOperands();
1900 Value *GpPtr = Gep->getPointerOperand();
1901 // If this GEP value is a consecutive pointer induction variable and all of
1902 // the indices are constant then we know it is consecutive. We can
1903 Phi = dyn_cast<PHINode>(GpPtr);
1904 if (Phi && Inductions.count(Phi)) {
1906 // Make sure that the pointer does not point to structs.
1907 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1908 if (GepPtrType->getElementType()->isAggregateType())
1911 // Make sure that all of the index operands are loop invariant.
1912 for (unsigned i = 1; i < NumOperands; ++i)
1913 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1916 InductionDescriptor II = Inductions[Phi];
1917 return II.getConsecutiveDirection();
1920 unsigned InductionOperand = getGEPInductionOperand(Gep);
1922 // Check that all of the gep indices are uniform except for our induction
1924 for (unsigned i = 0; i != NumOperands; ++i)
1925 if (i != InductionOperand &&
1926 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1929 // We can emit wide load/stores only if the last non-zero index is the
1930 // induction variable.
1931 const SCEV *Last = nullptr;
1932 if (!Strides.count(Gep))
1933 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1935 // Because of the multiplication by a stride we can have a s/zext cast.
1936 // We are going to replace this stride by 1 so the cast is safe to ignore.
1938 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1939 // %0 = trunc i64 %indvars.iv to i32
1940 // %mul = mul i32 %0, %Stride1
1941 // %idxprom = zext i32 %mul to i64 << Safe cast.
1942 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1944 Last = replaceSymbolicStrideSCEV(SE, Strides,
1945 Gep->getOperand(InductionOperand), Gep);
1946 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1948 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1952 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1953 const SCEV *Step = AR->getStepRecurrence(*SE);
1955 // The memory is consecutive because the last index is consecutive
1956 // and all other indices are loop invariant.
1959 if (Step->isAllOnesValue())
1966 bool LoopVectorizationLegality::isUniform(Value *V) {
1967 return LAI->isUniform(V);
1970 InnerLoopVectorizer::VectorParts&
1971 InnerLoopVectorizer::getVectorValue(Value *V) {
1972 assert(V != Induction && "The new induction variable should not be used.");
1973 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1975 // If we have a stride that is replaced by one, do it here.
1976 if (Legal->hasStride(V))
1977 V = ConstantInt::get(V->getType(), 1);
1979 // If we have this scalar in the map, return it.
1980 if (WidenMap.has(V))
1981 return WidenMap.get(V);
1983 // If this scalar is unknown, assume that it is a constant or that it is
1984 // loop invariant. Broadcast V and save the value for future uses.
1985 Value *B = getBroadcastInstrs(V);
1986 return WidenMap.splat(V, B);
1989 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1990 assert(Vec->getType()->isVectorTy() && "Invalid type");
1991 SmallVector<Constant*, 8> ShuffleMask;
1992 for (unsigned i = 0; i < VF; ++i)
1993 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1995 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1996 ConstantVector::get(ShuffleMask),
2000 // Get a mask to interleave \p NumVec vectors into a wide vector.
2001 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2002 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2003 // <0, 4, 1, 5, 2, 6, 3, 7>
2004 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2006 SmallVector<Constant *, 16> Mask;
2007 for (unsigned i = 0; i < VF; i++)
2008 for (unsigned j = 0; j < NumVec; j++)
2009 Mask.push_back(Builder.getInt32(j * VF + i));
2011 return ConstantVector::get(Mask);
2014 // Get the strided mask starting from index \p Start.
2015 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2016 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2017 unsigned Stride, unsigned VF) {
2018 SmallVector<Constant *, 16> Mask;
2019 for (unsigned i = 0; i < VF; i++)
2020 Mask.push_back(Builder.getInt32(Start + i * Stride));
2022 return ConstantVector::get(Mask);
2025 // Get a mask of two parts: The first part consists of sequential integers
2026 // starting from 0, The second part consists of UNDEFs.
2027 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2028 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2029 unsigned NumUndef) {
2030 SmallVector<Constant *, 16> Mask;
2031 for (unsigned i = 0; i < NumInt; i++)
2032 Mask.push_back(Builder.getInt32(i));
2034 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2035 for (unsigned i = 0; i < NumUndef; i++)
2036 Mask.push_back(Undef);
2038 return ConstantVector::get(Mask);
2041 // Concatenate two vectors with the same element type. The 2nd vector should
2042 // not have more elements than the 1st vector. If the 2nd vector has less
2043 // elements, extend it with UNDEFs.
2044 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2046 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2047 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2048 assert(VecTy1 && VecTy2 &&
2049 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2050 "Expect two vectors with the same element type");
2052 unsigned NumElts1 = VecTy1->getNumElements();
2053 unsigned NumElts2 = VecTy2->getNumElements();
2054 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2056 if (NumElts1 > NumElts2) {
2057 // Extend with UNDEFs.
2059 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2060 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2063 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2064 return Builder.CreateShuffleVector(V1, V2, Mask);
2067 // Concatenate vectors in the given list. All vectors have the same type.
2068 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2069 ArrayRef<Value *> InputList) {
2070 unsigned NumVec = InputList.size();
2071 assert(NumVec > 1 && "Should be at least two vectors");
2073 SmallVector<Value *, 8> ResList;
2074 ResList.append(InputList.begin(), InputList.end());
2076 SmallVector<Value *, 8> TmpList;
2077 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2078 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2079 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2080 "Only the last vector may have a different type");
2082 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2085 // Push the last vector if the total number of vectors is odd.
2086 if (NumVec % 2 != 0)
2087 TmpList.push_back(ResList[NumVec - 1]);
2090 NumVec = ResList.size();
2091 } while (NumVec > 1);
2096 // Try to vectorize the interleave group that \p Instr belongs to.
2098 // E.g. Translate following interleaved load group (factor = 3):
2099 // for (i = 0; i < N; i+=3) {
2100 // R = Pic[i]; // Member of index 0
2101 // G = Pic[i+1]; // Member of index 1
2102 // B = Pic[i+2]; // Member of index 2
2103 // ... // do something to R, G, B
2106 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2107 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2108 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2109 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2111 // Or translate following interleaved store group (factor = 3):
2112 // for (i = 0; i < N; i+=3) {
2113 // ... do something to R, G, B
2114 // Pic[i] = R; // Member of index 0
2115 // Pic[i+1] = G; // Member of index 1
2116 // Pic[i+2] = B; // Member of index 2
2119 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2120 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2121 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2122 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2123 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2124 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2125 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2126 assert(Group && "Fail to get an interleaved access group.");
2128 // Skip if current instruction is not the insert position.
2129 if (Instr != Group->getInsertPos())
2132 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2133 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2134 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2136 // Prepare for the vector type of the interleaved load/store.
2137 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2138 unsigned InterleaveFactor = Group->getFactor();
2139 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2140 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2142 // Prepare for the new pointers.
2143 setDebugLocFromInst(Builder, Ptr);
2144 VectorParts &PtrParts = getVectorValue(Ptr);
2145 SmallVector<Value *, 2> NewPtrs;
2146 unsigned Index = Group->getIndex(Instr);
2147 for (unsigned Part = 0; Part < UF; Part++) {
2148 // Extract the pointer for current instruction from the pointer vector. A
2149 // reverse access uses the pointer in the last lane.
2150 Value *NewPtr = Builder.CreateExtractElement(
2152 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2154 // Notice current instruction could be any index. Need to adjust the address
2155 // to the member of index 0.
2157 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2158 // b = A[i]; // Member of index 0
2159 // Current pointer is pointed to A[i+1], adjust it to A[i].
2161 // E.g. A[i+1] = a; // Member of index 1
2162 // A[i] = b; // Member of index 0
2163 // A[i+2] = c; // Member of index 2 (Current instruction)
2164 // Current pointer is pointed to A[i+2], adjust it to A[i].
2165 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2167 // Cast to the vector pointer type.
2168 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2171 setDebugLocFromInst(Builder, Instr);
2172 Value *UndefVec = UndefValue::get(VecTy);
2174 // Vectorize the interleaved load group.
2176 for (unsigned Part = 0; Part < UF; Part++) {
2177 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2178 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2180 for (unsigned i = 0; i < InterleaveFactor; i++) {
2181 Instruction *Member = Group->getMember(i);
2183 // Skip the gaps in the group.
2187 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2188 Value *StridedVec = Builder.CreateShuffleVector(
2189 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2191 // If this member has different type, cast the result type.
2192 if (Member->getType() != ScalarTy) {
2193 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2194 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2197 VectorParts &Entry = WidenMap.get(Member);
2199 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2202 propagateMetadata(NewLoadInstr, Instr);
2207 // The sub vector type for current instruction.
2208 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2210 // Vectorize the interleaved store group.
2211 for (unsigned Part = 0; Part < UF; Part++) {
2212 // Collect the stored vector from each member.
2213 SmallVector<Value *, 4> StoredVecs;
2214 for (unsigned i = 0; i < InterleaveFactor; i++) {
2215 // Interleaved store group doesn't allow a gap, so each index has a member
2216 Instruction *Member = Group->getMember(i);
2217 assert(Member && "Fail to get a member from an interleaved store group");
2220 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2221 if (Group->isReverse())
2222 StoredVec = reverseVector(StoredVec);
2224 // If this member has different type, cast it to an unified type.
2225 if (StoredVec->getType() != SubVT)
2226 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2228 StoredVecs.push_back(StoredVec);
2231 // Concatenate all vectors into a wide vector.
2232 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2234 // Interleave the elements in the wide vector.
2235 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2236 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2239 Instruction *NewStoreInstr =
2240 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2241 propagateMetadata(NewStoreInstr, Instr);
2245 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2246 // Attempt to issue a wide load.
2247 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2248 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2250 assert((LI || SI) && "Invalid Load/Store instruction");
2252 // Try to vectorize the interleave group if this access is interleaved.
2253 if (Legal->isAccessInterleaved(Instr))
2254 return vectorizeInterleaveGroup(Instr);
2256 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2257 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2258 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2259 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2260 // An alignment of 0 means target abi alignment. We need to use the scalar's
2261 // target abi alignment in such a case.
2262 const DataLayout &DL = Instr->getModule()->getDataLayout();
2264 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2265 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2266 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2267 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2269 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2270 !Legal->isMaskRequired(SI))
2271 return scalarizeInstruction(Instr, true);
2273 if (ScalarAllocatedSize != VectorElementSize)
2274 return scalarizeInstruction(Instr);
2276 // If the pointer is loop invariant or if it is non-consecutive,
2277 // scalarize the load.
2278 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2279 bool Reverse = ConsecutiveStride < 0;
2280 bool UniformLoad = LI && Legal->isUniform(Ptr);
2281 if (!ConsecutiveStride || UniformLoad)
2282 return scalarizeInstruction(Instr);
2284 Constant *Zero = Builder.getInt32(0);
2285 VectorParts &Entry = WidenMap.get(Instr);
2287 // Handle consecutive loads/stores.
2288 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2289 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2290 setDebugLocFromInst(Builder, Gep);
2291 Value *PtrOperand = Gep->getPointerOperand();
2292 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2293 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2295 // Create the new GEP with the new induction variable.
2296 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2297 Gep2->setOperand(0, FirstBasePtr);
2298 Gep2->setName("gep.indvar.base");
2299 Ptr = Builder.Insert(Gep2);
2301 setDebugLocFromInst(Builder, Gep);
2302 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2303 OrigLoop) && "Base ptr must be invariant");
2305 // The last index does not have to be the induction. It can be
2306 // consecutive and be a function of the index. For example A[I+1];
2307 unsigned NumOperands = Gep->getNumOperands();
2308 unsigned InductionOperand = getGEPInductionOperand(Gep);
2309 // Create the new GEP with the new induction variable.
2310 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2312 for (unsigned i = 0; i < NumOperands; ++i) {
2313 Value *GepOperand = Gep->getOperand(i);
2314 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2316 // Update last index or loop invariant instruction anchored in loop.
2317 if (i == InductionOperand ||
2318 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2319 assert((i == InductionOperand ||
2320 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2321 "Must be last index or loop invariant");
2323 VectorParts &GEPParts = getVectorValue(GepOperand);
2324 Value *Index = GEPParts[0];
2325 Index = Builder.CreateExtractElement(Index, Zero);
2326 Gep2->setOperand(i, Index);
2327 Gep2->setName("gep.indvar.idx");
2330 Ptr = Builder.Insert(Gep2);
2332 // Use the induction element ptr.
2333 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2334 setDebugLocFromInst(Builder, Ptr);
2335 VectorParts &PtrVal = getVectorValue(Ptr);
2336 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2339 VectorParts Mask = createBlockInMask(Instr->getParent());
2342 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2343 "We do not allow storing to uniform addresses");
2344 setDebugLocFromInst(Builder, SI);
2345 // We don't want to update the value in the map as it might be used in
2346 // another expression. So don't use a reference type for "StoredVal".
2347 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2349 for (unsigned Part = 0; Part < UF; ++Part) {
2350 // Calculate the pointer for the specific unroll-part.
2352 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2355 // If we store to reverse consecutive memory locations, then we need
2356 // to reverse the order of elements in the stored value.
2357 StoredVal[Part] = reverseVector(StoredVal[Part]);
2358 // If the address is consecutive but reversed, then the
2359 // wide store needs to start at the last vector element.
2360 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2361 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2362 Mask[Part] = reverseVector(Mask[Part]);
2365 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2366 DataTy->getPointerTo(AddressSpace));
2369 if (Legal->isMaskRequired(SI))
2370 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2373 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2374 propagateMetadata(NewSI, SI);
2380 assert(LI && "Must have a load instruction");
2381 setDebugLocFromInst(Builder, LI);
2382 for (unsigned Part = 0; Part < UF; ++Part) {
2383 // Calculate the pointer for the specific unroll-part.
2385 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2388 // If the address is consecutive but reversed, then the
2389 // wide load needs to start at the last vector element.
2390 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2391 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2392 Mask[Part] = reverseVector(Mask[Part]);
2396 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2397 DataTy->getPointerTo(AddressSpace));
2398 if (Legal->isMaskRequired(LI))
2399 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2400 UndefValue::get(DataTy),
2401 "wide.masked.load");
2403 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2404 propagateMetadata(NewLI, LI);
2405 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2409 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2410 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2411 // Holds vector parameters or scalars, in case of uniform vals.
2412 SmallVector<VectorParts, 4> Params;
2414 setDebugLocFromInst(Builder, Instr);
2416 // Find all of the vectorized parameters.
2417 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2418 Value *SrcOp = Instr->getOperand(op);
2420 // If we are accessing the old induction variable, use the new one.
2421 if (SrcOp == OldInduction) {
2422 Params.push_back(getVectorValue(SrcOp));
2426 // Try using previously calculated values.
2427 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2429 // If the src is an instruction that appeared earlier in the basic block,
2430 // then it should already be vectorized.
2431 if (SrcInst && OrigLoop->contains(SrcInst)) {
2432 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2433 // The parameter is a vector value from earlier.
2434 Params.push_back(WidenMap.get(SrcInst));
2436 // The parameter is a scalar from outside the loop. Maybe even a constant.
2437 VectorParts Scalars;
2438 Scalars.append(UF, SrcOp);
2439 Params.push_back(Scalars);
2443 assert(Params.size() == Instr->getNumOperands() &&
2444 "Invalid number of operands");
2446 // Does this instruction return a value ?
2447 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2449 Value *UndefVec = IsVoidRetTy ? nullptr :
2450 UndefValue::get(VectorType::get(Instr->getType(), VF));
2451 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2452 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2454 Instruction *InsertPt = Builder.GetInsertPoint();
2455 BasicBlock *IfBlock = Builder.GetInsertBlock();
2456 BasicBlock *CondBlock = nullptr;
2459 Loop *VectorLp = nullptr;
2460 if (IfPredicateStore) {
2461 assert(Instr->getParent()->getSinglePredecessor() &&
2462 "Only support single predecessor blocks");
2463 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2464 Instr->getParent());
2465 VectorLp = LI->getLoopFor(IfBlock);
2466 assert(VectorLp && "Must have a loop for this block");
2469 // For each vector unroll 'part':
2470 for (unsigned Part = 0; Part < UF; ++Part) {
2471 // For each scalar that we create:
2472 for (unsigned Width = 0; Width < VF; ++Width) {
2475 Value *Cmp = nullptr;
2476 if (IfPredicateStore) {
2477 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2478 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2479 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2480 LoopVectorBody.push_back(CondBlock);
2481 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2482 // Update Builder with newly created basic block.
2483 Builder.SetInsertPoint(InsertPt);
2486 Instruction *Cloned = Instr->clone();
2488 Cloned->setName(Instr->getName() + ".cloned");
2489 // Replace the operands of the cloned instructions with extracted scalars.
2490 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2491 Value *Op = Params[op][Part];
2492 // Param is a vector. Need to extract the right lane.
2493 if (Op->getType()->isVectorTy())
2494 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2495 Cloned->setOperand(op, Op);
2498 // Place the cloned scalar in the new loop.
2499 Builder.Insert(Cloned);
2501 // If the original scalar returns a value we need to place it in a vector
2502 // so that future users will be able to use it.
2504 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2505 Builder.getInt32(Width));
2507 if (IfPredicateStore) {
2508 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2509 LoopVectorBody.push_back(NewIfBlock);
2510 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2511 Builder.SetInsertPoint(InsertPt);
2512 ReplaceInstWithInst(IfBlock->getTerminator(),
2513 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2514 IfBlock = NewIfBlock;
2520 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2524 if (Instruction *I = dyn_cast<Instruction>(V))
2525 return I->getParent() == Loc->getParent() ? I : nullptr;
2529 std::pair<Instruction *, Instruction *>
2530 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2531 Instruction *tnullptr = nullptr;
2532 if (!Legal->mustCheckStrides())
2533 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2535 IRBuilder<> ChkBuilder(Loc);
2538 Value *Check = nullptr;
2539 Instruction *FirstInst = nullptr;
2540 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2541 SE = Legal->strides_end();
2543 Value *Ptr = stripIntegerCast(*SI);
2544 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2546 // Store the first instruction we create.
2547 FirstInst = getFirstInst(FirstInst, C, Loc);
2549 Check = ChkBuilder.CreateOr(Check, C);
2554 // We have to do this trickery because the IRBuilder might fold the check to a
2555 // constant expression in which case there is no Instruction anchored in a
2557 LLVMContext &Ctx = Loc->getContext();
2558 Instruction *TheCheck =
2559 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2560 ChkBuilder.Insert(TheCheck, "stride.not.one");
2561 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2563 return std::make_pair(FirstInst, TheCheck);
2566 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L,
2571 BasicBlock *Header = L->getHeader();
2572 BasicBlock *Latch = L->getLoopLatch();
2573 // As we're just creating this loop, it's possible no latch exists
2574 // yet. If so, use the header as this will be a single block loop.
2578 IRBuilder<> Builder(Header->getFirstInsertionPt());
2579 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2580 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2582 Builder.SetInsertPoint(Latch->getTerminator());
2584 // Create i+1 and fill the PHINode.
2585 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2586 Induction->addIncoming(Start, L->getLoopPreheader());
2587 Induction->addIncoming(Next, Latch);
2588 // Create the compare.
2589 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2590 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2592 // Now we have two terminators. Remove the old one from the block.
2593 Latch->getTerminator()->eraseFromParent();
2598 void InnerLoopVectorizer::createEmptyLoop() {
2600 In this function we generate a new loop. The new loop will contain
2601 the vectorized instructions while the old loop will continue to run the
2604 [ ] <-- loop iteration number check.
2607 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2610 || [ ] <-- vector pre header.
2614 || [ ]_| <-- vector loop.
2617 | >[ ] <--- middle-block.
2620 -|- >[ ] <--- new preheader.
2624 | [ ]_| <-- old scalar loop to handle remainder.
2627 >[ ] <-- exit block.
2631 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2632 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2633 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2634 assert(VectorPH && "Invalid loop structure");
2635 assert(ExitBlock && "Must have an exit block");
2637 // Some loops have a single integer induction variable, while other loops
2638 // don't. One example is c++ iterators that often have multiple pointer
2639 // induction variables. In the code below we also support a case where we
2640 // don't have a single induction variable.
2642 // We try to obtain an induction variable from the original loop as hard
2643 // as possible. However if we don't find one that:
2645 // - counts from zero, stepping by one
2646 // - is the size of the widest induction variable type
2647 // then we create a new one.
2648 OldInduction = Legal->getInduction();
2649 Type *IdxTy = Legal->getWidestInductionType();
2651 // Find the loop boundaries.
2652 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2653 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2655 // The exit count might have the type of i64 while the phi is i32. This can
2656 // happen if we have an induction variable that is sign extended before the
2657 // compare. The only way that we get a backedge taken count is that the
2658 // induction variable was signed and as such will not overflow. In such a case
2659 // truncation is legal.
2660 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2661 IdxTy->getPrimitiveSizeInBits())
2662 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2664 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2665 // Get the total trip count from the count by adding 1.
2666 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2667 SE->getConstant(BackedgeTakeCount->getType(), 1));
2669 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2671 // Expand the trip count and place the new instructions in the preheader.
2672 // Notice that the pre-header does not change, only the loop body.
2673 SCEVExpander Exp(*SE, DL, "induction");
2675 // The loop minimum iterations check below is to ensure the loop has enough
2676 // trip count so the generated vector loop will likely be executed and the
2677 // preparation and rounding-off costs will likely be worthy.
2679 // The minimum iteration check also covers case where the backedge-taken
2680 // count is uint##_max. Adding one to it will cause overflow and an
2681 // incorrect loop trip count being generated in the vector body. In this
2682 // case we also want to directly jump to the scalar remainder loop.
2683 Value *ExitCountValue = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2684 VectorPH->getTerminator());
2685 if (ExitCountValue->getType()->isPointerTy())
2686 ExitCountValue = CastInst::CreatePointerCast(ExitCountValue, IdxTy,
2687 "exitcount.ptrcnt.to.int",
2688 VectorPH->getTerminator());
2690 Instruction *CheckMinIters =
2691 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULT, ExitCountValue,
2692 ConstantInt::get(ExitCountValue->getType(), VF * UF),
2693 "min.iters.check", VectorPH->getTerminator());
2695 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2697 LoopBypassBlocks.push_back(VectorPH);
2699 // Split the single block loop into the two loop structure described above.
2700 BasicBlock *VecBody =
2701 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2702 BasicBlock *MiddleBlock =
2703 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2704 BasicBlock *ScalarPH =
2705 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2707 // Create and register the new vector loop.
2708 Loop* Lp = new Loop();
2709 Loop *ParentLoop = OrigLoop->getParentLoop();
2711 // Insert the new loop into the loop nest and register the new basic blocks
2712 // before calling any utilities such as SCEV that require valid LoopInfo.
2714 ParentLoop->addChildLoop(Lp);
2715 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2716 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2718 LI->addTopLevelLoop(Lp);
2720 Lp->addBasicBlockToLoop(VecBody, *LI);
2722 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2724 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2726 // Generate code to check that the loop's trip count is not less than the
2727 // minimum loop iteration number threshold.
2728 BasicBlock *NewVectorPH =
2729 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "min.iters.checked");
2731 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2732 ReplaceInstWithInst(VectorPH->getTerminator(),
2733 BranchInst::Create(ScalarPH, NewVectorPH, CheckMinIters));
2734 VectorPH = NewVectorPH;
2736 // This is the IR builder that we use to add all of the logic for bypassing
2737 // the new vector loop.
2738 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2739 setDebugLocFromInst(BypassBuilder,
2740 getDebugLocFromInstOrOperands(OldInduction));
2742 // Add the start index to the loop count to get the new end index.
2743 Value *IdxEnd = BypassBuilder.CreateAdd(ExitCountValue, StartIdx, "end.idx");
2745 // Now we need to generate the expression for N - (N % VF), which is
2746 // the part that the vectorized body will execute.
2747 // The loop step is equal to the vectorization factor (num of SIMD elements)
2748 // times the unroll factor (num of SIMD instructions).
2749 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2750 Value *R = BypassBuilder.CreateURem(ExitCountValue, Step, "n.mod.vf");
2751 Value *CountRoundDown = BypassBuilder.CreateSub(ExitCountValue, R, "n.vec");
2752 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2753 "end.idx.rnd.down");
2755 // Generate the induction variable.
2757 createInductionVariable(Lp, StartIdx, IdxEndRoundDown, Step,
2758 getDebugLocFromInstOrOperands(OldInduction));
2760 // Now, compare the new count to zero. If it is zero skip the vector loop and
2761 // jump to the scalar loop.
2763 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2765 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2767 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2768 LoopBypassBlocks.push_back(VectorPH);
2769 ReplaceInstWithInst(VectorPH->getTerminator(),
2770 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2771 VectorPH = NewVectorPH;
2773 // Generate the code to check that the strides we assumed to be one are really
2774 // one. We want the new basic block to start at the first instruction in a
2775 // sequence of instructions that form a check.
2776 Instruction *StrideCheck;
2777 Instruction *FirstCheckInst;
2778 std::tie(FirstCheckInst, StrideCheck) =
2779 addStrideCheck(VectorPH->getTerminator());
2781 AddedSafetyChecks = true;
2782 // Create a new block containing the stride check.
2783 VectorPH->setName("vector.stridecheck");
2785 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2787 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2788 LoopBypassBlocks.push_back(VectorPH);
2790 // Replace the branch into the memory check block with a conditional branch
2791 // for the "few elements case".
2792 ReplaceInstWithInst(
2793 VectorPH->getTerminator(),
2794 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2796 VectorPH = NewVectorPH;
2799 // Generate the code that checks in runtime if arrays overlap. We put the
2800 // checks into a separate block to make the more common case of few elements
2802 Instruction *MemRuntimeCheck;
2803 std::tie(FirstCheckInst, MemRuntimeCheck) =
2804 Legal->getLAI()->addRuntimeChecks(VectorPH->getTerminator());
2805 if (MemRuntimeCheck) {
2806 AddedSafetyChecks = true;
2807 // Create a new block containing the memory check.
2808 VectorPH->setName("vector.memcheck");
2810 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2812 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2813 LoopBypassBlocks.push_back(VectorPH);
2815 // Replace the branch into the memory check block with a conditional branch
2816 // for the "few elements case".
2817 ReplaceInstWithInst(
2818 VectorPH->getTerminator(),
2819 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2821 VectorPH = NewVectorPH;
2824 // We are going to resume the execution of the scalar loop.
2825 // Go over all of the induction variables that we found and fix the
2826 // PHIs that are left in the scalar version of the loop.
2827 // The starting values of PHI nodes depend on the counter of the last
2828 // iteration in the vectorized loop.
2829 // If we come from a bypass edge then we need to start from the original
2832 // This variable saves the new starting index for the scalar loop. It is used
2833 // to test if there are any tail iterations left once the vector loop has
2835 PHINode *ResumeIndex = nullptr;
2836 LoopVectorizationLegality::InductionList::iterator I, E;
2837 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2838 // Set builder to point to last bypass block.
2839 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2840 for (I = List->begin(), E = List->end(); I != E; ++I) {
2841 PHINode *OrigPhi = I->first;
2842 InductionDescriptor II = I->second;
2844 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
2845 MiddleBlock->getTerminator());
2846 // Create phi nodes to merge from the backedge-taken check block.
2847 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2849 ScalarPH->getTerminator());
2850 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2853 if (OrigPhi == OldInduction) {
2854 // We know what the end value is.
2855 EndValue = IdxEndRoundDown;
2856 // We also know which PHI node holds it.
2857 ResumeIndex = ResumeVal;
2859 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2860 II.getStepValue()->getType(),
2862 EndValue = II.transform(BypassBuilder, CRD);
2863 EndValue->setName("ind.end");
2866 // The new PHI merges the original incoming value, in case of a bypass,
2867 // or the value at the end of the vectorized loop.
2868 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2869 ResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2870 ResumeVal->addIncoming(EndValue, VecBody);
2872 // Fix the scalar body counter (PHI node).
2873 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2875 // The old induction's phi node in the scalar body needs the truncated
2877 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[0]);
2878 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2881 // If we are generating a new induction variable then we also need to
2882 // generate the code that calculates the exit value. This value is not
2883 // simply the end of the counter because we may skip the vectorized body
2884 // in case of a runtime check.
2886 assert(!ResumeIndex && "Unexpected resume value found");
2887 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2888 MiddleBlock->getTerminator());
2889 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2890 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2891 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2894 // Make sure that we found the index where scalar loop needs to continue.
2895 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2896 "Invalid resume Index");
2898 // Add a check in the middle block to see if we have completed
2899 // all of the iterations in the first vector loop.
2900 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2901 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2902 ResumeIndex, "cmp.n",
2903 MiddleBlock->getTerminator());
2904 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2905 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2907 // Get ready to start creating new instructions into the vectorized body.
2908 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2911 LoopVectorPreHeader = VectorPH;
2912 LoopScalarPreHeader = ScalarPH;
2913 LoopMiddleBlock = MiddleBlock;
2914 LoopExitBlock = ExitBlock;
2915 LoopVectorBody.push_back(VecBody);
2916 LoopScalarBody = OldBasicBlock;
2918 LoopVectorizeHints Hints(Lp, true);
2919 Hints.setAlreadyVectorized();
2923 struct CSEDenseMapInfo {
2924 static bool canHandle(Instruction *I) {
2925 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2926 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2928 static inline Instruction *getEmptyKey() {
2929 return DenseMapInfo<Instruction *>::getEmptyKey();
2931 static inline Instruction *getTombstoneKey() {
2932 return DenseMapInfo<Instruction *>::getTombstoneKey();
2934 static unsigned getHashValue(Instruction *I) {
2935 assert(canHandle(I) && "Unknown instruction!");
2936 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2937 I->value_op_end()));
2939 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2940 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2941 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2943 return LHS->isIdenticalTo(RHS);
2948 /// \brief Check whether this block is a predicated block.
2949 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2950 /// = ...; " blocks. We start with one vectorized basic block. For every
2951 /// conditional block we split this vectorized block. Therefore, every second
2952 /// block will be a predicated one.
2953 static bool isPredicatedBlock(unsigned BlockNum) {
2954 return BlockNum % 2;
2957 ///\brief Perform cse of induction variable instructions.
2958 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2959 // Perform simple cse.
2960 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2961 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2962 BasicBlock *BB = BBs[i];
2963 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2964 Instruction *In = I++;
2966 if (!CSEDenseMapInfo::canHandle(In))
2969 // Check if we can replace this instruction with any of the
2970 // visited instructions.
2971 if (Instruction *V = CSEMap.lookup(In)) {
2972 In->replaceAllUsesWith(V);
2973 In->eraseFromParent();
2976 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2977 // ...;" blocks for predicated stores. Every second block is a predicated
2979 if (isPredicatedBlock(i))
2987 /// \brief Adds a 'fast' flag to floating point operations.
2988 static Value *addFastMathFlag(Value *V) {
2989 if (isa<FPMathOperator>(V)){
2990 FastMathFlags Flags;
2991 Flags.setUnsafeAlgebra();
2992 cast<Instruction>(V)->setFastMathFlags(Flags);
2997 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
2998 /// the result needs to be inserted and/or extracted from vectors.
2999 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3000 const TargetTransformInfo &TTI) {
3004 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3007 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3009 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3011 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3017 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3018 // Return the cost of the instruction, including scalarization overhead if it's
3019 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3020 // i.e. either vector version isn't available, or is too expensive.
3021 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3022 const TargetTransformInfo &TTI,
3023 const TargetLibraryInfo *TLI,
3024 bool &NeedToScalarize) {
3025 Function *F = CI->getCalledFunction();
3026 StringRef FnName = CI->getCalledFunction()->getName();
3027 Type *ScalarRetTy = CI->getType();
3028 SmallVector<Type *, 4> Tys, ScalarTys;
3029 for (auto &ArgOp : CI->arg_operands())
3030 ScalarTys.push_back(ArgOp->getType());
3032 // Estimate cost of scalarized vector call. The source operands are assumed
3033 // to be vectors, so we need to extract individual elements from there,
3034 // execute VF scalar calls, and then gather the result into the vector return
3036 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3038 return ScalarCallCost;
3040 // Compute corresponding vector type for return value and arguments.
3041 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3042 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3043 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3045 // Compute costs of unpacking argument values for the scalar calls and
3046 // packing the return values to a vector.
3047 unsigned ScalarizationCost =
3048 getScalarizationOverhead(RetTy, true, false, TTI);
3049 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3050 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3052 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3054 // If we can't emit a vector call for this function, then the currently found
3055 // cost is the cost we need to return.
3056 NeedToScalarize = true;
3057 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3060 // If the corresponding vector cost is cheaper, return its cost.
3061 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3062 if (VectorCallCost < Cost) {
3063 NeedToScalarize = false;
3064 return VectorCallCost;
3069 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3070 // factor VF. Return the cost of the instruction, including scalarization
3071 // overhead if it's needed.
3072 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3073 const TargetTransformInfo &TTI,
3074 const TargetLibraryInfo *TLI) {
3075 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3076 assert(ID && "Expected intrinsic call!");
3078 Type *RetTy = ToVectorTy(CI->getType(), VF);
3079 SmallVector<Type *, 4> Tys;
3080 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3081 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3083 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3086 void InnerLoopVectorizer::vectorizeLoop() {
3087 //===------------------------------------------------===//
3089 // Notice: any optimization or new instruction that go
3090 // into the code below should be also be implemented in
3093 //===------------------------------------------------===//
3094 Constant *Zero = Builder.getInt32(0);
3096 // In order to support reduction variables we need to be able to vectorize
3097 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3098 // stages. First, we create a new vector PHI node with no incoming edges.
3099 // We use this value when we vectorize all of the instructions that use the
3100 // PHI. Next, after all of the instructions in the block are complete we
3101 // add the new incoming edges to the PHI. At this point all of the
3102 // instructions in the basic block are vectorized, so we can use them to
3103 // construct the PHI.
3104 PhiVector RdxPHIsToFix;
3106 // Scan the loop in a topological order to ensure that defs are vectorized
3108 LoopBlocksDFS DFS(OrigLoop);
3111 // Vectorize all of the blocks in the original loop.
3112 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3113 be = DFS.endRPO(); bb != be; ++bb)
3114 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3116 // At this point every instruction in the original loop is widened to
3117 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3118 // that we vectorized. The PHI nodes are currently empty because we did
3119 // not want to introduce cycles. Notice that the remaining PHI nodes
3120 // that we need to fix are reduction variables.
3122 // Create the 'reduced' values for each of the induction vars.
3123 // The reduced values are the vector values that we scalarize and combine
3124 // after the loop is finished.
3125 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3127 PHINode *RdxPhi = *it;
3128 assert(RdxPhi && "Unable to recover vectorized PHI");
3130 // Find the reduction variable descriptor.
3131 assert(Legal->getReductionVars()->count(RdxPhi) &&
3132 "Unable to find the reduction variable");
3133 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3135 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3136 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3137 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3138 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3139 RdxDesc.getMinMaxRecurrenceKind();
3140 setDebugLocFromInst(Builder, ReductionStartValue);
3142 // We need to generate a reduction vector from the incoming scalar.
3143 // To do so, we need to generate the 'identity' vector and override
3144 // one of the elements with the incoming scalar reduction. We need
3145 // to do it in the vector-loop preheader.
3146 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3148 // This is the vector-clone of the value that leaves the loop.
3149 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3150 Type *VecTy = VectorExit[0]->getType();
3152 // Find the reduction identity variable. Zero for addition, or, xor,
3153 // one for multiplication, -1 for And.
3156 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3157 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3158 // MinMax reduction have the start value as their identify.
3160 VectorStart = Identity = ReductionStartValue;
3162 VectorStart = Identity =
3163 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3166 // Handle other reduction kinds:
3167 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3168 RK, VecTy->getScalarType());
3171 // This vector is the Identity vector where the first element is the
3172 // incoming scalar reduction.
3173 VectorStart = ReductionStartValue;
3175 Identity = ConstantVector::getSplat(VF, Iden);
3177 // This vector is the Identity vector where the first element is the
3178 // incoming scalar reduction.
3180 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3184 // Fix the vector-loop phi.
3186 // Reductions do not have to start at zero. They can start with
3187 // any loop invariant values.
3188 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3189 BasicBlock *Latch = OrigLoop->getLoopLatch();
3190 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3191 VectorParts &Val = getVectorValue(LoopVal);
3192 for (unsigned part = 0; part < UF; ++part) {
3193 // Make sure to add the reduction stat value only to the
3194 // first unroll part.
3195 Value *StartVal = (part == 0) ? VectorStart : Identity;
3196 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3197 LoopVectorPreHeader);
3198 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3199 LoopVectorBody.back());
3202 // Before each round, move the insertion point right between
3203 // the PHIs and the values we are going to write.
3204 // This allows us to write both PHINodes and the extractelement
3206 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3208 VectorParts RdxParts, &RdxExitVal = getVectorValue(LoopExitInst);
3209 setDebugLocFromInst(Builder, LoopExitInst);
3210 for (unsigned part = 0; part < UF; ++part) {
3211 // This PHINode contains the vectorized reduction variable, or
3212 // the initial value vector, if we bypass the vector loop.
3213 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3214 Value *StartVal = (part == 0) ? VectorStart : Identity;
3215 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3216 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3217 NewPhi->addIncoming(RdxExitVal[part],
3218 LoopVectorBody.back());
3219 RdxParts.push_back(NewPhi);
3222 // If the vector reduction can be performed in a smaller type, we truncate
3223 // then extend the loop exit value to enable InstCombine to evaluate the
3224 // entire expression in the smaller type.
3225 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3226 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3227 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3228 for (unsigned part = 0; part < UF; ++part) {
3229 Value *Trunc = Builder.CreateTrunc(RdxExitVal[part], RdxVecTy);
3230 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3231 : Builder.CreateZExt(Trunc, VecTy);
3232 for (Value::user_iterator UI = RdxExitVal[part]->user_begin();
3233 UI != RdxExitVal[part]->user_end();)
3235 (*UI++)->replaceUsesOfWith(RdxExitVal[part], Extnd);
3239 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3240 for (unsigned part = 0; part < UF; ++part)
3241 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3244 // Reduce all of the unrolled parts into a single vector.
3245 Value *ReducedPartRdx = RdxParts[0];
3246 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3247 setDebugLocFromInst(Builder, ReducedPartRdx);
3248 for (unsigned part = 1; part < UF; ++part) {
3249 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3250 // Floating point operations had to be 'fast' to enable the reduction.
3251 ReducedPartRdx = addFastMathFlag(
3252 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3253 ReducedPartRdx, "bin.rdx"));
3255 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3256 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3260 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3261 // and vector ops, reducing the set of values being computed by half each
3263 assert(isPowerOf2_32(VF) &&
3264 "Reduction emission only supported for pow2 vectors!");
3265 Value *TmpVec = ReducedPartRdx;
3266 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3267 for (unsigned i = VF; i != 1; i >>= 1) {
3268 // Move the upper half of the vector to the lower half.
3269 for (unsigned j = 0; j != i/2; ++j)
3270 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3272 // Fill the rest of the mask with undef.
3273 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3274 UndefValue::get(Builder.getInt32Ty()));
3277 Builder.CreateShuffleVector(TmpVec,
3278 UndefValue::get(TmpVec->getType()),
3279 ConstantVector::get(ShuffleMask),
3282 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3283 // Floating point operations had to be 'fast' to enable the reduction.
3284 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3285 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3287 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3291 // The result is in the first element of the vector.
3292 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3293 Builder.getInt32(0));
3295 // If the reduction can be performed in a smaller type, we need to extend
3296 // the reduction to the wider type before we branch to the original loop.
3297 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3300 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3301 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3304 // Create a phi node that merges control-flow from the backedge-taken check
3305 // block and the middle block.
3306 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3307 LoopScalarPreHeader->getTerminator());
3308 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3309 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3311 // Now, we need to fix the users of the reduction variable
3312 // inside and outside of the scalar remainder loop.
3313 // We know that the loop is in LCSSA form. We need to update the
3314 // PHI nodes in the exit blocks.
3315 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3316 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3317 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3318 if (!LCSSAPhi) break;
3320 // All PHINodes need to have a single entry edge, or two if
3321 // we already fixed them.
3322 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3324 // We found our reduction value exit-PHI. Update it with the
3325 // incoming bypass edge.
3326 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3327 // Add an edge coming from the bypass.
3328 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3331 }// end of the LCSSA phi scan.
3333 // Fix the scalar loop reduction variable with the incoming reduction sum
3334 // from the vector body and from the backedge value.
3335 int IncomingEdgeBlockIdx =
3336 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3337 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3338 // Pick the other block.
3339 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3340 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3341 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3342 }// end of for each redux variable.
3346 // Remove redundant induction instructions.
3347 cse(LoopVectorBody);
3350 void InnerLoopVectorizer::fixLCSSAPHIs() {
3351 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3352 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3353 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3354 if (!LCSSAPhi) break;
3355 if (LCSSAPhi->getNumIncomingValues() == 1)
3356 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3361 InnerLoopVectorizer::VectorParts
3362 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3363 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3366 // Look for cached value.
3367 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3368 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3369 if (ECEntryIt != MaskCache.end())
3370 return ECEntryIt->second;
3372 VectorParts SrcMask = createBlockInMask(Src);
3374 // The terminator has to be a branch inst!
3375 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3376 assert(BI && "Unexpected terminator found");
3378 if (BI->isConditional()) {
3379 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3381 if (BI->getSuccessor(0) != Dst)
3382 for (unsigned part = 0; part < UF; ++part)
3383 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3385 for (unsigned part = 0; part < UF; ++part)
3386 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3388 MaskCache[Edge] = EdgeMask;
3392 MaskCache[Edge] = SrcMask;
3396 InnerLoopVectorizer::VectorParts
3397 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3398 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3400 // Loop incoming mask is all-one.
3401 if (OrigLoop->getHeader() == BB) {
3402 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3403 return getVectorValue(C);
3406 // This is the block mask. We OR all incoming edges, and with zero.
3407 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3408 VectorParts BlockMask = getVectorValue(Zero);
3411 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3412 VectorParts EM = createEdgeMask(*it, BB);
3413 for (unsigned part = 0; part < UF; ++part)
3414 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3420 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3421 InnerLoopVectorizer::VectorParts &Entry,
3422 unsigned UF, unsigned VF, PhiVector *PV) {
3423 PHINode* P = cast<PHINode>(PN);
3424 // Handle reduction variables:
3425 if (Legal->getReductionVars()->count(P)) {
3426 for (unsigned part = 0; part < UF; ++part) {
3427 // This is phase one of vectorizing PHIs.
3428 Type *VecTy = (VF == 1) ? PN->getType() :
3429 VectorType::get(PN->getType(), VF);
3430 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3431 LoopVectorBody.back()-> getFirstInsertionPt());
3437 setDebugLocFromInst(Builder, P);
3438 // Check for PHI nodes that are lowered to vector selects.
3439 if (P->getParent() != OrigLoop->getHeader()) {
3440 // We know that all PHIs in non-header blocks are converted into
3441 // selects, so we don't have to worry about the insertion order and we
3442 // can just use the builder.
3443 // At this point we generate the predication tree. There may be
3444 // duplications since this is a simple recursive scan, but future
3445 // optimizations will clean it up.
3447 unsigned NumIncoming = P->getNumIncomingValues();
3449 // Generate a sequence of selects of the form:
3450 // SELECT(Mask3, In3,
3451 // SELECT(Mask2, In2,
3453 for (unsigned In = 0; In < NumIncoming; In++) {
3454 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3456 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3458 for (unsigned part = 0; part < UF; ++part) {
3459 // We might have single edge PHIs (blocks) - use an identity
3460 // 'select' for the first PHI operand.
3462 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3465 // Select between the current value and the previous incoming edge
3466 // based on the incoming mask.
3467 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3468 Entry[part], "predphi");
3474 // This PHINode must be an induction variable.
3475 // Make sure that we know about it.
3476 assert(Legal->getInductionVars()->count(P) &&
3477 "Not an induction variable");
3479 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3481 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3482 // which can be found from the original scalar operations.
3483 switch (II.getKind()) {
3484 case InductionDescriptor::IK_NoInduction:
3485 llvm_unreachable("Unknown induction");
3486 case InductionDescriptor::IK_IntInduction: {
3487 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3488 // Handle other induction variables that are now based on the
3490 Value *V = Induction;
3491 if (P != OldInduction) {
3492 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3493 V = II.transform(Builder, V);
3494 V->setName("offset.idx");
3496 Value *Broadcasted = getBroadcastInstrs(V);
3497 // After broadcasting the induction variable we need to make the vector
3498 // consecutive by adding 0, 1, 2, etc.
3499 for (unsigned part = 0; part < UF; ++part)
3500 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3503 case InductionDescriptor::IK_PtrInduction:
3504 // Handle the pointer induction variable case.
3505 assert(P->getType()->isPointerTy() && "Unexpected type.");
3506 // This is the normalized GEP that starts counting at zero.
3507 Value *PtrInd = Induction;
3508 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3509 // This is the vector of results. Notice that we don't generate
3510 // vector geps because scalar geps result in better code.
3511 for (unsigned part = 0; part < UF; ++part) {
3513 int EltIndex = part;
3514 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3515 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3516 Value *SclrGep = II.transform(Builder, GlobalIdx);
3517 SclrGep->setName("next.gep");
3518 Entry[part] = SclrGep;
3522 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3523 for (unsigned int i = 0; i < VF; ++i) {
3524 int EltIndex = i + part * VF;
3525 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3526 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3527 Value *SclrGep = II.transform(Builder, GlobalIdx);
3528 SclrGep->setName("next.gep");
3529 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3530 Builder.getInt32(i),
3533 Entry[part] = VecVal;
3539 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3540 // For each instruction in the old loop.
3541 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3542 VectorParts &Entry = WidenMap.get(it);
3543 switch (it->getOpcode()) {
3544 case Instruction::Br:
3545 // Nothing to do for PHIs and BR, since we already took care of the
3546 // loop control flow instructions.
3548 case Instruction::PHI: {
3549 // Vectorize PHINodes.
3550 widenPHIInstruction(it, Entry, UF, VF, PV);
3554 case Instruction::Add:
3555 case Instruction::FAdd:
3556 case Instruction::Sub:
3557 case Instruction::FSub:
3558 case Instruction::Mul:
3559 case Instruction::FMul:
3560 case Instruction::UDiv:
3561 case Instruction::SDiv:
3562 case Instruction::FDiv:
3563 case Instruction::URem:
3564 case Instruction::SRem:
3565 case Instruction::FRem:
3566 case Instruction::Shl:
3567 case Instruction::LShr:
3568 case Instruction::AShr:
3569 case Instruction::And:
3570 case Instruction::Or:
3571 case Instruction::Xor: {
3572 // Just widen binops.
3573 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3574 setDebugLocFromInst(Builder, BinOp);
3575 VectorParts &A = getVectorValue(it->getOperand(0));
3576 VectorParts &B = getVectorValue(it->getOperand(1));
3578 // Use this vector value for all users of the original instruction.
3579 for (unsigned Part = 0; Part < UF; ++Part) {
3580 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3582 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3583 VecOp->copyIRFlags(BinOp);
3588 propagateMetadata(Entry, it);
3591 case Instruction::Select: {
3593 // If the selector is loop invariant we can create a select
3594 // instruction with a scalar condition. Otherwise, use vector-select.
3595 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3597 setDebugLocFromInst(Builder, it);
3599 // The condition can be loop invariant but still defined inside the
3600 // loop. This means that we can't just use the original 'cond' value.
3601 // We have to take the 'vectorized' value and pick the first lane.
3602 // Instcombine will make this a no-op.
3603 VectorParts &Cond = getVectorValue(it->getOperand(0));
3604 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3605 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3607 Value *ScalarCond = (VF == 1) ? Cond[0] :
3608 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3610 for (unsigned Part = 0; Part < UF; ++Part) {
3611 Entry[Part] = Builder.CreateSelect(
3612 InvariantCond ? ScalarCond : Cond[Part],
3617 propagateMetadata(Entry, it);
3621 case Instruction::ICmp:
3622 case Instruction::FCmp: {
3623 // Widen compares. Generate vector compares.
3624 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3625 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3626 setDebugLocFromInst(Builder, it);
3627 VectorParts &A = getVectorValue(it->getOperand(0));
3628 VectorParts &B = getVectorValue(it->getOperand(1));
3629 for (unsigned Part = 0; Part < UF; ++Part) {
3632 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3634 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3638 propagateMetadata(Entry, it);
3642 case Instruction::Store:
3643 case Instruction::Load:
3644 vectorizeMemoryInstruction(it);
3646 case Instruction::ZExt:
3647 case Instruction::SExt:
3648 case Instruction::FPToUI:
3649 case Instruction::FPToSI:
3650 case Instruction::FPExt:
3651 case Instruction::PtrToInt:
3652 case Instruction::IntToPtr:
3653 case Instruction::SIToFP:
3654 case Instruction::UIToFP:
3655 case Instruction::Trunc:
3656 case Instruction::FPTrunc:
3657 case Instruction::BitCast: {
3658 CastInst *CI = dyn_cast<CastInst>(it);
3659 setDebugLocFromInst(Builder, it);
3660 /// Optimize the special case where the source is the induction
3661 /// variable. Notice that we can only optimize the 'trunc' case
3662 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3663 /// c. other casts depend on pointer size.
3664 if (CI->getOperand(0) == OldInduction &&
3665 it->getOpcode() == Instruction::Trunc) {
3666 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3668 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3669 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3671 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3672 for (unsigned Part = 0; Part < UF; ++Part)
3673 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3674 propagateMetadata(Entry, it);
3677 /// Vectorize casts.
3678 Type *DestTy = (VF == 1) ? CI->getType() :
3679 VectorType::get(CI->getType(), VF);
3681 VectorParts &A = getVectorValue(it->getOperand(0));
3682 for (unsigned Part = 0; Part < UF; ++Part)
3683 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3684 propagateMetadata(Entry, it);
3688 case Instruction::Call: {
3689 // Ignore dbg intrinsics.
3690 if (isa<DbgInfoIntrinsic>(it))
3692 setDebugLocFromInst(Builder, it);
3694 Module *M = BB->getParent()->getParent();
3695 CallInst *CI = cast<CallInst>(it);
3697 StringRef FnName = CI->getCalledFunction()->getName();
3698 Function *F = CI->getCalledFunction();
3699 Type *RetTy = ToVectorTy(CI->getType(), VF);
3700 SmallVector<Type *, 4> Tys;
3701 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3702 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3704 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3706 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3707 ID == Intrinsic::lifetime_start)) {
3708 scalarizeInstruction(it);
3711 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3712 // version of the instruction.
3713 // Is it beneficial to perform intrinsic call compared to lib call?
3714 bool NeedToScalarize;
3715 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3716 bool UseVectorIntrinsic =
3717 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3718 if (!UseVectorIntrinsic && NeedToScalarize) {
3719 scalarizeInstruction(it);
3723 for (unsigned Part = 0; Part < UF; ++Part) {
3724 SmallVector<Value *, 4> Args;
3725 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3726 Value *Arg = CI->getArgOperand(i);
3727 // Some intrinsics have a scalar argument - don't replace it with a
3729 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3730 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3731 Arg = VectorArg[Part];
3733 Args.push_back(Arg);
3737 if (UseVectorIntrinsic) {
3738 // Use vector version of the intrinsic.
3739 Type *TysForDecl[] = {CI->getType()};
3741 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3742 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3744 // Use vector version of the library call.
3745 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3746 assert(!VFnName.empty() && "Vector function name is empty.");
3747 VectorF = M->getFunction(VFnName);
3749 // Generate a declaration
3750 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3752 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3753 VectorF->copyAttributesFrom(F);
3756 assert(VectorF && "Can't create vector function.");
3757 Entry[Part] = Builder.CreateCall(VectorF, Args);
3760 propagateMetadata(Entry, it);
3765 // All other instructions are unsupported. Scalarize them.
3766 scalarizeInstruction(it);
3769 }// end of for_each instr.
3772 void InnerLoopVectorizer::updateAnalysis() {
3773 // Forget the original basic block.
3774 SE->forgetLoop(OrigLoop);
3776 // Update the dominator tree information.
3777 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3778 "Entry does not dominate exit.");
3780 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3781 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3782 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3784 // Due to if predication of stores we might create a sequence of "if(pred)
3785 // a[i] = ...; " blocks.
3786 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3788 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3789 else if (isPredicatedBlock(i)) {
3790 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3792 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3796 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3797 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3798 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3799 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3801 DEBUG(DT->verifyDomTree());
3804 /// \brief Check whether it is safe to if-convert this phi node.
3806 /// Phi nodes with constant expressions that can trap are not safe to if
3808 static bool canIfConvertPHINodes(BasicBlock *BB) {
3809 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3810 PHINode *Phi = dyn_cast<PHINode>(I);
3813 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3814 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3821 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3822 if (!EnableIfConversion) {
3823 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3827 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3829 // A list of pointers that we can safely read and write to.
3830 SmallPtrSet<Value *, 8> SafePointes;
3832 // Collect safe addresses.
3833 for (Loop::block_iterator BI = TheLoop->block_begin(),
3834 BE = TheLoop->block_end(); BI != BE; ++BI) {
3835 BasicBlock *BB = *BI;
3837 if (blockNeedsPredication(BB))
3840 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3841 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3842 SafePointes.insert(LI->getPointerOperand());
3843 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3844 SafePointes.insert(SI->getPointerOperand());
3848 // Collect the blocks that need predication.
3849 BasicBlock *Header = TheLoop->getHeader();
3850 for (Loop::block_iterator BI = TheLoop->block_begin(),
3851 BE = TheLoop->block_end(); BI != BE; ++BI) {
3852 BasicBlock *BB = *BI;
3854 // We don't support switch statements inside loops.
3855 if (!isa<BranchInst>(BB->getTerminator())) {
3856 emitAnalysis(VectorizationReport(BB->getTerminator())
3857 << "loop contains a switch statement");
3861 // We must be able to predicate all blocks that need to be predicated.
3862 if (blockNeedsPredication(BB)) {
3863 if (!blockCanBePredicated(BB, SafePointes)) {
3864 emitAnalysis(VectorizationReport(BB->getTerminator())
3865 << "control flow cannot be substituted for a select");
3868 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3869 emitAnalysis(VectorizationReport(BB->getTerminator())
3870 << "control flow cannot be substituted for a select");
3875 // We can if-convert this loop.
3879 bool LoopVectorizationLegality::canVectorize() {
3880 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3881 // be canonicalized.
3882 if (!TheLoop->getLoopPreheader()) {
3884 VectorizationReport() <<
3885 "loop control flow is not understood by vectorizer");
3889 // We can only vectorize innermost loops.
3890 if (!TheLoop->empty()) {
3891 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3895 // We must have a single backedge.
3896 if (TheLoop->getNumBackEdges() != 1) {
3898 VectorizationReport() <<
3899 "loop control flow is not understood by vectorizer");
3903 // We must have a single exiting block.
3904 if (!TheLoop->getExitingBlock()) {
3906 VectorizationReport() <<
3907 "loop control flow is not understood by vectorizer");
3911 // We only handle bottom-tested loops, i.e. loop in which the condition is
3912 // checked at the end of each iteration. With that we can assume that all
3913 // instructions in the loop are executed the same number of times.
3914 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3916 VectorizationReport() <<
3917 "loop control flow is not understood by vectorizer");
3921 // We need to have a loop header.
3922 DEBUG(dbgs() << "LV: Found a loop: " <<
3923 TheLoop->getHeader()->getName() << '\n');
3925 // Check if we can if-convert non-single-bb loops.
3926 unsigned NumBlocks = TheLoop->getNumBlocks();
3927 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3928 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3932 // ScalarEvolution needs to be able to find the exit count.
3933 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3934 if (ExitCount == SE->getCouldNotCompute()) {
3935 emitAnalysis(VectorizationReport() <<
3936 "could not determine number of loop iterations");
3937 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3941 // Check if we can vectorize the instructions and CFG in this loop.
3942 if (!canVectorizeInstrs()) {
3943 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3947 // Go over each instruction and look at memory deps.
3948 if (!canVectorizeMemory()) {
3949 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3953 // Collect all of the variables that remain uniform after vectorization.
3954 collectLoopUniforms();
3956 DEBUG(dbgs() << "LV: We can vectorize this loop"
3957 << (LAI->getRuntimePointerChecking()->Need
3958 ? " (with a runtime bound check)"
3962 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
3964 // If an override option has been passed in for interleaved accesses, use it.
3965 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
3966 UseInterleaved = EnableInterleavedMemAccesses;
3968 // Analyze interleaved memory accesses.
3970 InterleaveInfo.analyzeInterleaving(Strides);
3972 // Okay! We can vectorize. At this point we don't have any other mem analysis
3973 // which may limit our maximum vectorization factor, so just return true with
3978 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3979 if (Ty->isPointerTy())
3980 return DL.getIntPtrType(Ty);
3982 // It is possible that char's or short's overflow when we ask for the loop's
3983 // trip count, work around this by changing the type size.
3984 if (Ty->getScalarSizeInBits() < 32)
3985 return Type::getInt32Ty(Ty->getContext());
3990 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3991 Ty0 = convertPointerToIntegerType(DL, Ty0);
3992 Ty1 = convertPointerToIntegerType(DL, Ty1);
3993 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3998 /// \brief Check that the instruction has outside loop users and is not an
3999 /// identified reduction variable.
4000 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4001 SmallPtrSetImpl<Value *> &Reductions) {
4002 // Reduction instructions are allowed to have exit users. All other
4003 // instructions must not have external users.
4004 if (!Reductions.count(Inst))
4005 //Check that all of the users of the loop are inside the BB.
4006 for (User *U : Inst->users()) {
4007 Instruction *UI = cast<Instruction>(U);
4008 // This user may be a reduction exit value.
4009 if (!TheLoop->contains(UI)) {
4010 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4017 bool LoopVectorizationLegality::canVectorizeInstrs() {
4018 BasicBlock *Header = TheLoop->getHeader();
4020 // Look for the attribute signaling the absence of NaNs.
4021 Function &F = *Header->getParent();
4022 const DataLayout &DL = F.getParent()->getDataLayout();
4023 if (F.hasFnAttribute("no-nans-fp-math"))
4025 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4027 // For each block in the loop.
4028 for (Loop::block_iterator bb = TheLoop->block_begin(),
4029 be = TheLoop->block_end(); bb != be; ++bb) {
4031 // Scan the instructions in the block and look for hazards.
4032 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4035 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4036 Type *PhiTy = Phi->getType();
4037 // Check that this PHI type is allowed.
4038 if (!PhiTy->isIntegerTy() &&
4039 !PhiTy->isFloatingPointTy() &&
4040 !PhiTy->isPointerTy()) {
4041 emitAnalysis(VectorizationReport(it)
4042 << "loop control flow is not understood by vectorizer");
4043 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4047 // If this PHINode is not in the header block, then we know that we
4048 // can convert it to select during if-conversion. No need to check if
4049 // the PHIs in this block are induction or reduction variables.
4050 if (*bb != Header) {
4051 // Check that this instruction has no outside users or is an
4052 // identified reduction value with an outside user.
4053 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4055 emitAnalysis(VectorizationReport(it) <<
4056 "value could not be identified as "
4057 "an induction or reduction variable");
4061 // We only allow if-converted PHIs with exactly two incoming values.
4062 if (Phi->getNumIncomingValues() != 2) {
4063 emitAnalysis(VectorizationReport(it)
4064 << "control flow not understood by vectorizer");
4065 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4069 InductionDescriptor ID;
4070 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4071 Inductions[Phi] = ID;
4072 // Get the widest type.
4074 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4076 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4078 // Int inductions are special because we only allow one IV.
4079 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4080 ID.getStepValue()->isOne() &&
4081 isa<Constant>(ID.getStartValue()) &&
4082 cast<Constant>(ID.getStartValue())->isNullValue()) {
4083 // Use the phi node with the widest type as induction. Use the last
4084 // one if there are multiple (no good reason for doing this other
4085 // than it is expedient). We've checked that it begins at zero and
4086 // steps by one, so this is a canonical induction variable.
4087 if (!Induction || PhiTy == WidestIndTy)
4091 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4093 // Until we explicitly handle the case of an induction variable with
4094 // an outside loop user we have to give up vectorizing this loop.
4095 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4096 emitAnalysis(VectorizationReport(it) <<
4097 "use of induction value outside of the "
4098 "loop is not handled by vectorizer");
4105 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4107 if (Reductions[Phi].hasUnsafeAlgebra())
4108 Requirements->addUnsafeAlgebraInst(
4109 Reductions[Phi].getUnsafeAlgebraInst());
4110 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4114 emitAnalysis(VectorizationReport(it) <<
4115 "value that could not be identified as "
4116 "reduction is used outside the loop");
4117 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4119 }// end of PHI handling
4121 // We handle calls that:
4122 // * Are debug info intrinsics.
4123 // * Have a mapping to an IR intrinsic.
4124 // * Have a vector version available.
4125 CallInst *CI = dyn_cast<CallInst>(it);
4126 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4127 !(CI->getCalledFunction() && TLI &&
4128 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4129 emitAnalysis(VectorizationReport(it) <<
4130 "call instruction cannot be vectorized");
4131 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4135 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4136 // second argument is the same (i.e. loop invariant)
4138 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4139 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4140 emitAnalysis(VectorizationReport(it)
4141 << "intrinsic instruction cannot be vectorized");
4142 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4147 // Check that the instruction return type is vectorizable.
4148 // Also, we can't vectorize extractelement instructions.
4149 if ((!VectorType::isValidElementType(it->getType()) &&
4150 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4151 emitAnalysis(VectorizationReport(it)
4152 << "instruction return type cannot be vectorized");
4153 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4157 // Check that the stored type is vectorizable.
4158 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4159 Type *T = ST->getValueOperand()->getType();
4160 if (!VectorType::isValidElementType(T)) {
4161 emitAnalysis(VectorizationReport(ST) <<
4162 "store instruction cannot be vectorized");
4165 if (EnableMemAccessVersioning)
4166 collectStridedAccess(ST);
4169 if (EnableMemAccessVersioning)
4170 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4171 collectStridedAccess(LI);
4173 // Reduction instructions are allowed to have exit users.
4174 // All other instructions must not have external users.
4175 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4176 emitAnalysis(VectorizationReport(it) <<
4177 "value cannot be used outside the loop");
4186 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4187 if (Inductions.empty()) {
4188 emitAnalysis(VectorizationReport()
4189 << "loop induction variable could not be identified");
4194 // Now we know the widest induction type, check if our found induction
4195 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4196 // will create another.
4197 if (Induction && WidestIndTy != Induction->getType())
4198 Induction = nullptr;
4203 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4204 Value *Ptr = nullptr;
4205 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4206 Ptr = LI->getPointerOperand();
4207 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4208 Ptr = SI->getPointerOperand();
4212 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4216 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4217 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4218 Strides[Ptr] = Stride;
4219 StrideSet.insert(Stride);
4222 void LoopVectorizationLegality::collectLoopUniforms() {
4223 // We now know that the loop is vectorizable!
4224 // Collect variables that will remain uniform after vectorization.
4225 std::vector<Value*> Worklist;
4226 BasicBlock *Latch = TheLoop->getLoopLatch();
4228 // Start with the conditional branch and walk up the block.
4229 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4231 // Also add all consecutive pointer values; these values will be uniform
4232 // after vectorization (and subsequent cleanup) and, until revectorization is
4233 // supported, all dependencies must also be uniform.
4234 for (Loop::block_iterator B = TheLoop->block_begin(),
4235 BE = TheLoop->block_end(); B != BE; ++B)
4236 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4238 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4239 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4241 while (!Worklist.empty()) {
4242 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4243 Worklist.pop_back();
4245 // Look at instructions inside this loop.
4246 // Stop when reaching PHI nodes.
4247 // TODO: we need to follow values all over the loop, not only in this block.
4248 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4251 // This is a known uniform.
4254 // Insert all operands.
4255 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4259 bool LoopVectorizationLegality::canVectorizeMemory() {
4260 LAI = &LAA->getInfo(TheLoop, Strides);
4261 auto &OptionalReport = LAI->getReport();
4263 emitAnalysis(VectorizationReport(*OptionalReport));
4264 if (!LAI->canVectorizeMemory())
4267 if (LAI->hasStoreToLoopInvariantAddress()) {
4269 VectorizationReport()
4270 << "write to a loop invariant address could not be vectorized");
4271 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4275 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4280 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4281 Value *In0 = const_cast<Value*>(V);
4282 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4286 return Inductions.count(PN);
4289 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4290 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4293 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4294 SmallPtrSetImpl<Value *> &SafePtrs) {
4296 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4297 // Check that we don't have a constant expression that can trap as operand.
4298 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4300 if (Constant *C = dyn_cast<Constant>(*OI))
4304 // We might be able to hoist the load.
4305 if (it->mayReadFromMemory()) {
4306 LoadInst *LI = dyn_cast<LoadInst>(it);
4309 if (!SafePtrs.count(LI->getPointerOperand())) {
4310 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4311 MaskedOp.insert(LI);
4318 // We don't predicate stores at the moment.
4319 if (it->mayWriteToMemory()) {
4320 StoreInst *SI = dyn_cast<StoreInst>(it);
4321 // We only support predication of stores in basic blocks with one
4326 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4327 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4329 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4330 !isSinglePredecessor) {
4331 // Build a masked store if it is legal for the target, otherwise scalarize
4333 bool isLegalMaskedOp =
4334 isLegalMaskedStore(SI->getValueOperand()->getType(),
4335 SI->getPointerOperand());
4336 if (isLegalMaskedOp) {
4338 MaskedOp.insert(SI);
4347 // The instructions below can trap.
4348 switch (it->getOpcode()) {
4350 case Instruction::UDiv:
4351 case Instruction::SDiv:
4352 case Instruction::URem:
4353 case Instruction::SRem:
4361 void InterleavedAccessInfo::collectConstStridedAccesses(
4362 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4363 const ValueToValueMap &Strides) {
4364 // Holds load/store instructions in program order.
4365 SmallVector<Instruction *, 16> AccessList;
4367 for (auto *BB : TheLoop->getBlocks()) {
4368 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4370 for (auto &I : *BB) {
4371 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4373 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4377 AccessList.push_back(&I);
4381 if (AccessList.empty())
4384 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4385 for (auto I : AccessList) {
4386 LoadInst *LI = dyn_cast<LoadInst>(I);
4387 StoreInst *SI = dyn_cast<StoreInst>(I);
4389 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4390 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4392 // The factor of the corresponding interleave group.
4393 unsigned Factor = std::abs(Stride);
4395 // Ignore the access if the factor is too small or too large.
4396 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4399 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4400 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4401 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4403 // An alignment of 0 means target ABI alignment.
4404 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4406 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4408 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4412 // Analyze interleaved accesses and collect them into interleave groups.
4414 // Notice that the vectorization on interleaved groups will change instruction
4415 // orders and may break dependences. But the memory dependence check guarantees
4416 // that there is no overlap between two pointers of different strides, element
4417 // sizes or underlying bases.
4419 // For pointers sharing the same stride, element size and underlying base, no
4420 // need to worry about Read-After-Write dependences and Write-After-Read
4423 // E.g. The RAW dependence: A[i] = a;
4425 // This won't exist as it is a store-load forwarding conflict, which has
4426 // already been checked and forbidden in the dependence check.
4428 // E.g. The WAR dependence: a = A[i]; // (1)
4430 // The store group of (2) is always inserted at or below (2), and the load group
4431 // of (1) is always inserted at or above (1). The dependence is safe.
4432 void InterleavedAccessInfo::analyzeInterleaving(
4433 const ValueToValueMap &Strides) {
4434 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4436 // Holds all the stride accesses.
4437 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4438 collectConstStridedAccesses(StrideAccesses, Strides);
4440 if (StrideAccesses.empty())
4443 // Holds all interleaved store groups temporarily.
4444 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4446 // Search the load-load/write-write pair B-A in bottom-up order and try to
4447 // insert B into the interleave group of A according to 3 rules:
4448 // 1. A and B have the same stride.
4449 // 2. A and B have the same memory object size.
4450 // 3. B belongs to the group according to the distance.
4452 // The bottom-up order can avoid breaking the Write-After-Write dependences
4453 // between two pointers of the same base.
4454 // E.g. A[i] = a; (1)
4457 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4458 // above (1), which guarantees that (1) is always above (2).
4459 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4461 Instruction *A = I->first;
4462 StrideDescriptor DesA = I->second;
4464 InterleaveGroup *Group = getInterleaveGroup(A);
4466 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4467 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4470 if (A->mayWriteToMemory())
4471 StoreGroups.insert(Group);
4473 for (auto II = std::next(I); II != E; ++II) {
4474 Instruction *B = II->first;
4475 StrideDescriptor DesB = II->second;
4477 // Ignore if B is already in a group or B is a different memory operation.
4478 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4481 // Check the rule 1 and 2.
4482 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4485 // Calculate the distance and prepare for the rule 3.
4486 const SCEVConstant *DistToA =
4487 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4491 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4493 // Skip if the distance is not multiple of size as they are not in the
4495 if (DistanceToA % static_cast<int>(DesA.Size))
4498 // The index of B is the index of A plus the related index to A.
4500 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4502 // Try to insert B into the group.
4503 if (Group->insertMember(B, IndexB, DesB.Align)) {
4504 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4505 << " into the interleave group with" << *A << '\n');
4506 InterleaveGroupMap[B] = Group;
4508 // Set the first load in program order as the insert position.
4509 if (B->mayReadFromMemory())
4510 Group->setInsertPos(B);
4512 } // Iteration on instruction B
4513 } // Iteration on instruction A
4515 // Remove interleaved store groups with gaps.
4516 for (InterleaveGroup *Group : StoreGroups)
4517 if (Group->getNumMembers() != Group->getFactor())
4518 releaseGroup(Group);
4521 LoopVectorizationCostModel::VectorizationFactor
4522 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4523 // Width 1 means no vectorize
4524 VectorizationFactor Factor = { 1U, 0U };
4525 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4526 emitAnalysis(VectorizationReport() <<
4527 "runtime pointer checks needed. Enable vectorization of this "
4528 "loop with '#pragma clang loop vectorize(enable)' when "
4529 "compiling with -Os/-Oz");
4531 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4535 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4536 emitAnalysis(VectorizationReport() <<
4537 "store that is conditionally executed prevents vectorization");
4538 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4542 // Find the trip count.
4543 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4544 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4546 unsigned WidestType = getWidestType();
4547 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4548 unsigned MaxSafeDepDist = -1U;
4549 if (Legal->getMaxSafeDepDistBytes() != -1U)
4550 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4551 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4552 WidestRegister : MaxSafeDepDist);
4553 unsigned MaxVectorSize = WidestRegister / WidestType;
4554 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4555 DEBUG(dbgs() << "LV: The Widest register is: "
4556 << WidestRegister << " bits.\n");
4558 if (MaxVectorSize == 0) {
4559 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4563 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4564 " into one vector!");
4566 unsigned VF = MaxVectorSize;
4568 // If we optimize the program for size, avoid creating the tail loop.
4570 // If we are unable to calculate the trip count then don't try to vectorize.
4573 (VectorizationReport() <<
4574 "unable to calculate the loop count due to complex control flow");
4575 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4579 // Find the maximum SIMD width that can fit within the trip count.
4580 VF = TC % MaxVectorSize;
4585 // If the trip count that we found modulo the vectorization factor is not
4586 // zero then we require a tail.
4587 emitAnalysis(VectorizationReport() <<
4588 "cannot optimize for size and vectorize at the "
4589 "same time. Enable vectorization of this loop "
4590 "with '#pragma clang loop vectorize(enable)' "
4591 "when compiling with -Os/-Oz");
4592 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4597 int UserVF = Hints->getWidth();
4599 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4600 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4602 Factor.Width = UserVF;
4606 float Cost = expectedCost(1);
4608 const float ScalarCost = Cost;
4611 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4613 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4614 // Ignore scalar width, because the user explicitly wants vectorization.
4615 if (ForceVectorization && VF > 1) {
4617 Cost = expectedCost(Width) / (float)Width;
4620 for (unsigned i=2; i <= VF; i*=2) {
4621 // Notice that the vector loop needs to be executed less times, so
4622 // we need to divide the cost of the vector loops by the width of
4623 // the vector elements.
4624 float VectorCost = expectedCost(i) / (float)i;
4625 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4626 (int)VectorCost << ".\n");
4627 if (VectorCost < Cost) {
4633 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4634 << "LV: Vectorization seems to be not beneficial, "
4635 << "but was forced by a user.\n");
4636 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4637 Factor.Width = Width;
4638 Factor.Cost = Width * Cost;
4642 unsigned LoopVectorizationCostModel::getWidestType() {
4643 unsigned MaxWidth = 8;
4644 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4647 for (Loop::block_iterator bb = TheLoop->block_begin(),
4648 be = TheLoop->block_end(); bb != be; ++bb) {
4649 BasicBlock *BB = *bb;
4651 // For each instruction in the loop.
4652 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4653 Type *T = it->getType();
4655 // Skip ignored values.
4656 if (ValuesToIgnore.count(it))
4659 // Only examine Loads, Stores and PHINodes.
4660 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4663 // Examine PHI nodes that are reduction variables. Update the type to
4664 // account for the recurrence type.
4665 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4666 if (!Legal->getReductionVars()->count(PN))
4668 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4669 T = RdxDesc.getRecurrenceType();
4672 // Examine the stored values.
4673 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4674 T = ST->getValueOperand()->getType();
4676 // Ignore loaded pointer types and stored pointer types that are not
4677 // consecutive. However, we do want to take consecutive stores/loads of
4678 // pointer vectors into account.
4679 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4682 MaxWidth = std::max(MaxWidth,
4683 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4690 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4692 unsigned LoopCost) {
4694 // -- The interleave heuristics --
4695 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4696 // There are many micro-architectural considerations that we can't predict
4697 // at this level. For example, frontend pressure (on decode or fetch) due to
4698 // code size, or the number and capabilities of the execution ports.
4700 // We use the following heuristics to select the interleave count:
4701 // 1. If the code has reductions, then we interleave to break the cross
4702 // iteration dependency.
4703 // 2. If the loop is really small, then we interleave to reduce the loop
4705 // 3. We don't interleave if we think that we will spill registers to memory
4706 // due to the increased register pressure.
4708 // When we optimize for size, we don't interleave.
4712 // We used the distance for the interleave count.
4713 if (Legal->getMaxSafeDepDistBytes() != -1U)
4716 // Do not interleave loops with a relatively small trip count.
4717 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4718 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4721 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4722 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4726 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4727 TargetNumRegisters = ForceTargetNumScalarRegs;
4729 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4730 TargetNumRegisters = ForceTargetNumVectorRegs;
4733 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4734 // We divide by these constants so assume that we have at least one
4735 // instruction that uses at least one register.
4736 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4737 R.NumInstructions = std::max(R.NumInstructions, 1U);
4739 // We calculate the interleave count using the following formula.
4740 // Subtract the number of loop invariants from the number of available
4741 // registers. These registers are used by all of the interleaved instances.
4742 // Next, divide the remaining registers by the number of registers that is
4743 // required by the loop, in order to estimate how many parallel instances
4744 // fit without causing spills. All of this is rounded down if necessary to be
4745 // a power of two. We want power of two interleave count to simplify any
4746 // addressing operations or alignment considerations.
4747 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4750 // Don't count the induction variable as interleaved.
4751 if (EnableIndVarRegisterHeur)
4752 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4753 std::max(1U, (R.MaxLocalUsers - 1)));
4755 // Clamp the interleave ranges to reasonable counts.
4756 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4758 // Check if the user has overridden the max.
4760 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4761 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4763 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4764 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4767 // If we did not calculate the cost for VF (because the user selected the VF)
4768 // then we calculate the cost of VF here.
4770 LoopCost = expectedCost(VF);
4772 // Clamp the calculated IC to be between the 1 and the max interleave count
4773 // that the target allows.
4774 if (IC > MaxInterleaveCount)
4775 IC = MaxInterleaveCount;
4779 // Interleave if we vectorized this loop and there is a reduction that could
4780 // benefit from interleaving.
4781 if (VF > 1 && Legal->getReductionVars()->size()) {
4782 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4786 // Note that if we've already vectorized the loop we will have done the
4787 // runtime check and so interleaving won't require further checks.
4788 bool InterleavingRequiresRuntimePointerCheck =
4789 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4791 // We want to interleave small loops in order to reduce the loop overhead and
4792 // potentially expose ILP opportunities.
4793 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4794 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4795 // We assume that the cost overhead is 1 and we use the cost model
4796 // to estimate the cost of the loop and interleave until the cost of the
4797 // loop overhead is about 5% of the cost of the loop.
4799 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4801 // Interleave until store/load ports (estimated by max interleave count) are
4803 unsigned NumStores = Legal->getNumStores();
4804 unsigned NumLoads = Legal->getNumLoads();
4805 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4806 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4808 // If we have a scalar reduction (vector reductions are already dealt with
4809 // by this point), we can increase the critical path length if the loop
4810 // we're interleaving is inside another loop. Limit, by default to 2, so the
4811 // critical path only gets increased by one reduction operation.
4812 if (Legal->getReductionVars()->size() &&
4813 TheLoop->getLoopDepth() > 1) {
4814 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4815 SmallIC = std::min(SmallIC, F);
4816 StoresIC = std::min(StoresIC, F);
4817 LoadsIC = std::min(LoadsIC, F);
4820 if (EnableLoadStoreRuntimeInterleave &&
4821 std::max(StoresIC, LoadsIC) > SmallIC) {
4822 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4823 return std::max(StoresIC, LoadsIC);
4826 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4830 // Interleave if this is a large loop (small loops are already dealt with by
4832 // point) that could benefit from interleaving.
4833 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4834 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4835 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4839 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4843 LoopVectorizationCostModel::RegisterUsage
4844 LoopVectorizationCostModel::calculateRegisterUsage() {
4845 // This function calculates the register usage by measuring the highest number
4846 // of values that are alive at a single location. Obviously, this is a very
4847 // rough estimation. We scan the loop in a topological order in order and
4848 // assign a number to each instruction. We use RPO to ensure that defs are
4849 // met before their users. We assume that each instruction that has in-loop
4850 // users starts an interval. We record every time that an in-loop value is
4851 // used, so we have a list of the first and last occurrences of each
4852 // instruction. Next, we transpose this data structure into a multi map that
4853 // holds the list of intervals that *end* at a specific location. This multi
4854 // map allows us to perform a linear search. We scan the instructions linearly
4855 // and record each time that a new interval starts, by placing it in a set.
4856 // If we find this value in the multi-map then we remove it from the set.
4857 // The max register usage is the maximum size of the set.
4858 // We also search for instructions that are defined outside the loop, but are
4859 // used inside the loop. We need this number separately from the max-interval
4860 // usage number because when we unroll, loop-invariant values do not take
4862 LoopBlocksDFS DFS(TheLoop);
4866 R.NumInstructions = 0;
4868 // Each 'key' in the map opens a new interval. The values
4869 // of the map are the index of the 'last seen' usage of the
4870 // instruction that is the key.
4871 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4872 // Maps instruction to its index.
4873 DenseMap<unsigned, Instruction*> IdxToInstr;
4874 // Marks the end of each interval.
4875 IntervalMap EndPoint;
4876 // Saves the list of instruction indices that are used in the loop.
4877 SmallSet<Instruction*, 8> Ends;
4878 // Saves the list of values that are used in the loop but are
4879 // defined outside the loop, such as arguments and constants.
4880 SmallPtrSet<Value*, 8> LoopInvariants;
4883 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4884 be = DFS.endRPO(); bb != be; ++bb) {
4885 R.NumInstructions += (*bb)->size();
4886 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4888 Instruction *I = it;
4889 IdxToInstr[Index++] = I;
4891 // Save the end location of each USE.
4892 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4893 Value *U = I->getOperand(i);
4894 Instruction *Instr = dyn_cast<Instruction>(U);
4896 // Ignore non-instruction values such as arguments, constants, etc.
4897 if (!Instr) continue;
4899 // If this instruction is outside the loop then record it and continue.
4900 if (!TheLoop->contains(Instr)) {
4901 LoopInvariants.insert(Instr);
4905 // Overwrite previous end points.
4906 EndPoint[Instr] = Index;
4912 // Saves the list of intervals that end with the index in 'key'.
4913 typedef SmallVector<Instruction*, 2> InstrList;
4914 DenseMap<unsigned, InstrList> TransposeEnds;
4916 // Transpose the EndPoints to a list of values that end at each index.
4917 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4919 TransposeEnds[it->second].push_back(it->first);
4921 SmallSet<Instruction*, 8> OpenIntervals;
4922 unsigned MaxUsage = 0;
4925 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4926 for (unsigned int i = 0; i < Index; ++i) {
4927 Instruction *I = IdxToInstr[i];
4928 // Ignore instructions that are never used within the loop.
4929 if (!Ends.count(I)) continue;
4931 // Skip ignored values.
4932 if (ValuesToIgnore.count(I))
4935 // Remove all of the instructions that end at this location.
4936 InstrList &List = TransposeEnds[i];
4937 for (unsigned int j=0, e = List.size(); j < e; ++j)
4938 OpenIntervals.erase(List[j]);
4940 // Count the number of live interals.
4941 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4943 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4944 OpenIntervals.size() << '\n');
4946 // Add the current instruction to the list of open intervals.
4947 OpenIntervals.insert(I);
4950 unsigned Invariant = LoopInvariants.size();
4951 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4952 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4953 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4955 R.LoopInvariantRegs = Invariant;
4956 R.MaxLocalUsers = MaxUsage;
4960 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4964 for (Loop::block_iterator bb = TheLoop->block_begin(),
4965 be = TheLoop->block_end(); bb != be; ++bb) {
4966 unsigned BlockCost = 0;
4967 BasicBlock *BB = *bb;
4969 // For each instruction in the old loop.
4970 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4971 // Skip dbg intrinsics.
4972 if (isa<DbgInfoIntrinsic>(it))
4975 // Skip ignored values.
4976 if (ValuesToIgnore.count(it))
4979 unsigned C = getInstructionCost(it, VF);
4981 // Check if we should override the cost.
4982 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4983 C = ForceTargetInstructionCost;
4986 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4987 VF << " For instruction: " << *it << '\n');
4990 // We assume that if-converted blocks have a 50% chance of being executed.
4991 // When the code is scalar then some of the blocks are avoided due to CF.
4992 // When the code is vectorized we execute all code paths.
4993 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5002 /// \brief Check whether the address computation for a non-consecutive memory
5003 /// access looks like an unlikely candidate for being merged into the indexing
5006 /// We look for a GEP which has one index that is an induction variable and all
5007 /// other indices are loop invariant. If the stride of this access is also
5008 /// within a small bound we decide that this address computation can likely be
5009 /// merged into the addressing mode.
5010 /// In all other cases, we identify the address computation as complex.
5011 static bool isLikelyComplexAddressComputation(Value *Ptr,
5012 LoopVectorizationLegality *Legal,
5013 ScalarEvolution *SE,
5014 const Loop *TheLoop) {
5015 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5019 // We are looking for a gep with all loop invariant indices except for one
5020 // which should be an induction variable.
5021 unsigned NumOperands = Gep->getNumOperands();
5022 for (unsigned i = 1; i < NumOperands; ++i) {
5023 Value *Opd = Gep->getOperand(i);
5024 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5025 !Legal->isInductionVariable(Opd))
5029 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5030 // can likely be merged into the address computation.
5031 unsigned MaxMergeDistance = 64;
5033 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5037 // Check the step is constant.
5038 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5039 // Calculate the pointer stride and check if it is consecutive.
5040 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5044 const APInt &APStepVal = C->getValue()->getValue();
5046 // Huge step value - give up.
5047 if (APStepVal.getBitWidth() > 64)
5050 int64_t StepVal = APStepVal.getSExtValue();
5052 return StepVal > MaxMergeDistance;
5055 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5056 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5062 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5063 // If we know that this instruction will remain uniform, check the cost of
5064 // the scalar version.
5065 if (Legal->isUniformAfterVectorization(I))
5068 Type *RetTy = I->getType();
5069 Type *VectorTy = ToVectorTy(RetTy, VF);
5071 // TODO: We need to estimate the cost of intrinsic calls.
5072 switch (I->getOpcode()) {
5073 case Instruction::GetElementPtr:
5074 // We mark this instruction as zero-cost because the cost of GEPs in
5075 // vectorized code depends on whether the corresponding memory instruction
5076 // is scalarized or not. Therefore, we handle GEPs with the memory
5077 // instruction cost.
5079 case Instruction::Br: {
5080 return TTI.getCFInstrCost(I->getOpcode());
5082 case Instruction::PHI:
5083 //TODO: IF-converted IFs become selects.
5085 case Instruction::Add:
5086 case Instruction::FAdd:
5087 case Instruction::Sub:
5088 case Instruction::FSub:
5089 case Instruction::Mul:
5090 case Instruction::FMul:
5091 case Instruction::UDiv:
5092 case Instruction::SDiv:
5093 case Instruction::FDiv:
5094 case Instruction::URem:
5095 case Instruction::SRem:
5096 case Instruction::FRem:
5097 case Instruction::Shl:
5098 case Instruction::LShr:
5099 case Instruction::AShr:
5100 case Instruction::And:
5101 case Instruction::Or:
5102 case Instruction::Xor: {
5103 // Since we will replace the stride by 1 the multiplication should go away.
5104 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5106 // Certain instructions can be cheaper to vectorize if they have a constant
5107 // second vector operand. One example of this are shifts on x86.
5108 TargetTransformInfo::OperandValueKind Op1VK =
5109 TargetTransformInfo::OK_AnyValue;
5110 TargetTransformInfo::OperandValueKind Op2VK =
5111 TargetTransformInfo::OK_AnyValue;
5112 TargetTransformInfo::OperandValueProperties Op1VP =
5113 TargetTransformInfo::OP_None;
5114 TargetTransformInfo::OperandValueProperties Op2VP =
5115 TargetTransformInfo::OP_None;
5116 Value *Op2 = I->getOperand(1);
5118 // Check for a splat of a constant or for a non uniform vector of constants.
5119 if (isa<ConstantInt>(Op2)) {
5120 ConstantInt *CInt = cast<ConstantInt>(Op2);
5121 if (CInt && CInt->getValue().isPowerOf2())
5122 Op2VP = TargetTransformInfo::OP_PowerOf2;
5123 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5124 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5125 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5126 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5128 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5129 if (CInt && CInt->getValue().isPowerOf2())
5130 Op2VP = TargetTransformInfo::OP_PowerOf2;
5131 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5135 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5138 case Instruction::Select: {
5139 SelectInst *SI = cast<SelectInst>(I);
5140 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5141 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5142 Type *CondTy = SI->getCondition()->getType();
5144 CondTy = VectorType::get(CondTy, VF);
5146 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5148 case Instruction::ICmp:
5149 case Instruction::FCmp: {
5150 Type *ValTy = I->getOperand(0)->getType();
5151 VectorTy = ToVectorTy(ValTy, VF);
5152 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5154 case Instruction::Store:
5155 case Instruction::Load: {
5156 StoreInst *SI = dyn_cast<StoreInst>(I);
5157 LoadInst *LI = dyn_cast<LoadInst>(I);
5158 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5160 VectorTy = ToVectorTy(ValTy, VF);
5162 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5163 unsigned AS = SI ? SI->getPointerAddressSpace() :
5164 LI->getPointerAddressSpace();
5165 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5166 // We add the cost of address computation here instead of with the gep
5167 // instruction because only here we know whether the operation is
5170 return TTI.getAddressComputationCost(VectorTy) +
5171 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5173 // For an interleaved access, calculate the total cost of the whole
5174 // interleave group.
5175 if (Legal->isAccessInterleaved(I)) {
5176 auto Group = Legal->getInterleavedAccessGroup(I);
5177 assert(Group && "Fail to get an interleaved access group.");
5179 // Only calculate the cost once at the insert position.
5180 if (Group->getInsertPos() != I)
5183 unsigned InterleaveFactor = Group->getFactor();
5185 VectorType::get(VectorTy->getVectorElementType(),
5186 VectorTy->getVectorNumElements() * InterleaveFactor);
5188 // Holds the indices of existing members in an interleaved load group.
5189 // An interleaved store group doesn't need this as it dones't allow gaps.
5190 SmallVector<unsigned, 4> Indices;
5192 for (unsigned i = 0; i < InterleaveFactor; i++)
5193 if (Group->getMember(i))
5194 Indices.push_back(i);
5197 // Calculate the cost of the whole interleaved group.
5198 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5199 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5200 Group->getAlignment(), AS);
5202 if (Group->isReverse())
5204 Group->getNumMembers() *
5205 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5207 // FIXME: The interleaved load group with a huge gap could be even more
5208 // expensive than scalar operations. Then we could ignore such group and
5209 // use scalar operations instead.
5213 // Scalarized loads/stores.
5214 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5215 bool Reverse = ConsecutiveStride < 0;
5216 const DataLayout &DL = I->getModule()->getDataLayout();
5217 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5218 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5219 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5220 bool IsComplexComputation =
5221 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5223 // The cost of extracting from the value vector and pointer vector.
5224 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5225 for (unsigned i = 0; i < VF; ++i) {
5226 // The cost of extracting the pointer operand.
5227 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5228 // In case of STORE, the cost of ExtractElement from the vector.
5229 // In case of LOAD, the cost of InsertElement into the returned
5231 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5232 Instruction::InsertElement,
5236 // The cost of the scalar loads/stores.
5237 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5238 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5243 // Wide load/stores.
5244 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5245 if (Legal->isMaskRequired(I))
5246 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5249 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5252 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5256 case Instruction::ZExt:
5257 case Instruction::SExt:
5258 case Instruction::FPToUI:
5259 case Instruction::FPToSI:
5260 case Instruction::FPExt:
5261 case Instruction::PtrToInt:
5262 case Instruction::IntToPtr:
5263 case Instruction::SIToFP:
5264 case Instruction::UIToFP:
5265 case Instruction::Trunc:
5266 case Instruction::FPTrunc:
5267 case Instruction::BitCast: {
5268 // We optimize the truncation of induction variable.
5269 // The cost of these is the same as the scalar operation.
5270 if (I->getOpcode() == Instruction::Trunc &&
5271 Legal->isInductionVariable(I->getOperand(0)))
5272 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5273 I->getOperand(0)->getType());
5275 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5276 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5278 case Instruction::Call: {
5279 bool NeedToScalarize;
5280 CallInst *CI = cast<CallInst>(I);
5281 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5282 if (getIntrinsicIDForCall(CI, TLI))
5283 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5287 // We are scalarizing the instruction. Return the cost of the scalar
5288 // instruction, plus the cost of insert and extract into vector
5289 // elements, times the vector width.
5292 if (!RetTy->isVoidTy() && VF != 1) {
5293 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5295 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5298 // The cost of inserting the results plus extracting each one of the
5300 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5303 // The cost of executing VF copies of the scalar instruction. This opcode
5304 // is unknown. Assume that it is the same as 'mul'.
5305 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5311 char LoopVectorize::ID = 0;
5312 static const char lv_name[] = "Loop Vectorization";
5313 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5314 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5315 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5316 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5317 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5318 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5319 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5320 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5321 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5322 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5323 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5324 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5327 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5328 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5332 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5333 // Check for a store.
5334 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5335 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5337 // Check for a load.
5338 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5339 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5345 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5346 bool IfPredicateStore) {
5347 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5348 // Holds vector parameters or scalars, in case of uniform vals.
5349 SmallVector<VectorParts, 4> Params;
5351 setDebugLocFromInst(Builder, Instr);
5353 // Find all of the vectorized parameters.
5354 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5355 Value *SrcOp = Instr->getOperand(op);
5357 // If we are accessing the old induction variable, use the new one.
5358 if (SrcOp == OldInduction) {
5359 Params.push_back(getVectorValue(SrcOp));
5363 // Try using previously calculated values.
5364 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5366 // If the src is an instruction that appeared earlier in the basic block
5367 // then it should already be vectorized.
5368 if (SrcInst && OrigLoop->contains(SrcInst)) {
5369 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5370 // The parameter is a vector value from earlier.
5371 Params.push_back(WidenMap.get(SrcInst));
5373 // The parameter is a scalar from outside the loop. Maybe even a constant.
5374 VectorParts Scalars;
5375 Scalars.append(UF, SrcOp);
5376 Params.push_back(Scalars);
5380 assert(Params.size() == Instr->getNumOperands() &&
5381 "Invalid number of operands");
5383 // Does this instruction return a value ?
5384 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5386 Value *UndefVec = IsVoidRetTy ? nullptr :
5387 UndefValue::get(Instr->getType());
5388 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5389 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5391 Instruction *InsertPt = Builder.GetInsertPoint();
5392 BasicBlock *IfBlock = Builder.GetInsertBlock();
5393 BasicBlock *CondBlock = nullptr;
5396 Loop *VectorLp = nullptr;
5397 if (IfPredicateStore) {
5398 assert(Instr->getParent()->getSinglePredecessor() &&
5399 "Only support single predecessor blocks");
5400 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5401 Instr->getParent());
5402 VectorLp = LI->getLoopFor(IfBlock);
5403 assert(VectorLp && "Must have a loop for this block");
5406 // For each vector unroll 'part':
5407 for (unsigned Part = 0; Part < UF; ++Part) {
5408 // For each scalar that we create:
5410 // Start an "if (pred) a[i] = ..." block.
5411 Value *Cmp = nullptr;
5412 if (IfPredicateStore) {
5413 if (Cond[Part]->getType()->isVectorTy())
5415 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5416 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5417 ConstantInt::get(Cond[Part]->getType(), 1));
5418 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5419 LoopVectorBody.push_back(CondBlock);
5420 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5421 // Update Builder with newly created basic block.
5422 Builder.SetInsertPoint(InsertPt);
5425 Instruction *Cloned = Instr->clone();
5427 Cloned->setName(Instr->getName() + ".cloned");
5428 // Replace the operands of the cloned instructions with extracted scalars.
5429 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5430 Value *Op = Params[op][Part];
5431 Cloned->setOperand(op, Op);
5434 // Place the cloned scalar in the new loop.
5435 Builder.Insert(Cloned);
5437 // If the original scalar returns a value we need to place it in a vector
5438 // so that future users will be able to use it.
5440 VecResults[Part] = Cloned;
5443 if (IfPredicateStore) {
5444 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5445 LoopVectorBody.push_back(NewIfBlock);
5446 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5447 Builder.SetInsertPoint(InsertPt);
5448 ReplaceInstWithInst(IfBlock->getTerminator(),
5449 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5450 IfBlock = NewIfBlock;
5455 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5456 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5457 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5459 return scalarizeInstruction(Instr, IfPredicateStore);
5462 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5466 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5470 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5471 // When unrolling and the VF is 1, we only need to add a simple scalar.
5472 Type *ITy = Val->getType();
5473 assert(!ITy->isVectorTy() && "Val must be a scalar");
5474 Constant *C = ConstantInt::get(ITy, StartIdx);
5475 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");