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."));
219 // Forward declarations.
220 class LoopVectorizeHints;
221 class LoopVectorizationLegality;
222 class LoopVectorizationCostModel;
223 class LoopVectorizationRequirements;
225 /// \brief This modifies LoopAccessReport to initialize message with
226 /// loop-vectorizer-specific part.
227 class VectorizationReport : public LoopAccessReport {
229 VectorizationReport(Instruction *I = nullptr)
230 : LoopAccessReport("loop not vectorized: ", I) {}
232 /// \brief This allows promotion of the loop-access analysis report into the
233 /// loop-vectorizer report. It modifies the message to add the
234 /// loop-vectorizer-specific part of the message.
235 explicit VectorizationReport(const LoopAccessReport &R)
236 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
240 /// A helper function for converting Scalar types to vector types.
241 /// If the incoming type is void, we return void. If the VF is 1, we return
243 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
244 if (Scalar->isVoidTy() || VF == 1)
246 return VectorType::get(Scalar, VF);
249 /// InnerLoopVectorizer vectorizes loops which contain only one basic
250 /// block to a specified vectorization factor (VF).
251 /// This class performs the widening of scalars into vectors, or multiple
252 /// scalars. This class also implements the following features:
253 /// * It inserts an epilogue loop for handling loops that don't have iteration
254 /// counts that are known to be a multiple of the vectorization factor.
255 /// * It handles the code generation for reduction variables.
256 /// * Scalarization (implementation using scalars) of un-vectorizable
258 /// InnerLoopVectorizer does not perform any vectorization-legality
259 /// checks, and relies on the caller to check for the different legality
260 /// aspects. The InnerLoopVectorizer relies on the
261 /// LoopVectorizationLegality class to provide information about the induction
262 /// and reduction variables that were found to a given vectorization factor.
263 class InnerLoopVectorizer {
265 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
266 DominatorTree *DT, const TargetLibraryInfo *TLI,
267 const TargetTransformInfo *TTI, unsigned VecWidth,
268 unsigned UnrollFactor)
269 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
270 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
271 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
272 Legal(nullptr), AddedSafetyChecks(false) {}
274 // Perform the actual loop widening (vectorization).
275 void vectorize(LoopVectorizationLegality *L) {
277 // Create a new empty loop. Unlink the old loop and connect the new one.
279 // Widen each instruction in the old loop to a new one in the new loop.
280 // Use the Legality module to find the induction and reduction variables.
282 // Register the new loop and update the analysis passes.
286 // Return true if any runtime check is added.
287 bool IsSafetyChecksAdded() {
288 return AddedSafetyChecks;
291 virtual ~InnerLoopVectorizer() {}
294 /// A small list of PHINodes.
295 typedef SmallVector<PHINode*, 4> PhiVector;
296 /// When we unroll loops we have multiple vector values for each scalar.
297 /// This data structure holds the unrolled and vectorized values that
298 /// originated from one scalar instruction.
299 typedef SmallVector<Value*, 2> VectorParts;
301 // When we if-convert we need to create edge masks. We have to cache values
302 // so that we don't end up with exponential recursion/IR.
303 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
304 VectorParts> EdgeMaskCache;
306 /// \brief Add checks for strides that were assumed to be 1.
308 /// Returns the last check instruction and the first check instruction in the
309 /// pair as (first, last).
310 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
312 /// Create an empty loop, based on the loop ranges of the old loop.
313 void createEmptyLoop();
314 /// Copy and widen the instructions from the old loop.
315 virtual void vectorizeLoop();
317 /// \brief The Loop exit block may have single value PHI nodes where the
318 /// incoming value is 'Undef'. While vectorizing we only handled real values
319 /// that were defined inside the loop. Here we fix the 'undef case'.
323 /// A helper function that computes the predicate of the block BB, assuming
324 /// that the header block of the loop is set to True. It returns the *entry*
325 /// mask for the block BB.
326 VectorParts createBlockInMask(BasicBlock *BB);
327 /// A helper function that computes the predicate of the edge between SRC
329 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
331 /// A helper function to vectorize a single BB within the innermost loop.
332 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
334 /// Vectorize a single PHINode in a block. This method handles the induction
335 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
336 /// arbitrary length vectors.
337 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
338 unsigned UF, unsigned VF, PhiVector *PV);
340 /// Insert the new loop to the loop hierarchy and pass manager
341 /// and update the analysis passes.
342 void updateAnalysis();
344 /// This instruction is un-vectorizable. Implement it as a sequence
345 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
346 /// scalarized instruction behind an if block predicated on the control
347 /// dependence of the instruction.
348 virtual void scalarizeInstruction(Instruction *Instr,
349 bool IfPredicateStore=false);
351 /// Vectorize Load and Store instructions,
352 virtual void vectorizeMemoryInstruction(Instruction *Instr);
354 /// Create a broadcast instruction. This method generates a broadcast
355 /// instruction (shuffle) for loop invariant values and for the induction
356 /// value. If this is the induction variable then we extend it to N, N+1, ...
357 /// this is needed because each iteration in the loop corresponds to a SIMD
359 virtual Value *getBroadcastInstrs(Value *V);
361 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
362 /// to each vector element of Val. The sequence starts at StartIndex.
363 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
365 /// When we go over instructions in the basic block we rely on previous
366 /// values within the current basic block or on loop invariant values.
367 /// When we widen (vectorize) values we place them in the map. If the values
368 /// are not within the map, they have to be loop invariant, so we simply
369 /// broadcast them into a vector.
370 VectorParts &getVectorValue(Value *V);
372 /// Try to vectorize the interleaved access group that \p Instr belongs to.
373 void vectorizeInterleaveGroup(Instruction *Instr);
375 /// Generate a shuffle sequence that will reverse the vector Vec.
376 virtual Value *reverseVector(Value *Vec);
378 /// This is a helper class that holds the vectorizer state. It maps scalar
379 /// instructions to vector instructions. When the code is 'unrolled' then
380 /// then a single scalar value is mapped to multiple vector parts. The parts
381 /// are stored in the VectorPart type.
383 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
385 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
387 /// \return True if 'Key' is saved in the Value Map.
388 bool has(Value *Key) const { return MapStorage.count(Key); }
390 /// Initializes a new entry in the map. Sets all of the vector parts to the
391 /// save value in 'Val'.
392 /// \return A reference to a vector with splat values.
393 VectorParts &splat(Value *Key, Value *Val) {
394 VectorParts &Entry = MapStorage[Key];
395 Entry.assign(UF, Val);
399 ///\return A reference to the value that is stored at 'Key'.
400 VectorParts &get(Value *Key) {
401 VectorParts &Entry = MapStorage[Key];
404 assert(Entry.size() == UF);
409 /// The unroll factor. Each entry in the map stores this number of vector
413 /// Map storage. We use std::map and not DenseMap because insertions to a
414 /// dense map invalidates its iterators.
415 std::map<Value *, VectorParts> MapStorage;
418 /// The original loop.
420 /// Scev analysis to use.
428 /// Target Library Info.
429 const TargetLibraryInfo *TLI;
430 /// Target Transform Info.
431 const TargetTransformInfo *TTI;
433 /// The vectorization SIMD factor to use. Each vector will have this many
438 /// The vectorization unroll factor to use. Each scalar is vectorized to this
439 /// many different vector instructions.
442 /// The builder that we use
445 // --- Vectorization state ---
447 /// The vector-loop preheader.
448 BasicBlock *LoopVectorPreHeader;
449 /// The scalar-loop preheader.
450 BasicBlock *LoopScalarPreHeader;
451 /// Middle Block between the vector and the scalar.
452 BasicBlock *LoopMiddleBlock;
453 ///The ExitBlock of the scalar loop.
454 BasicBlock *LoopExitBlock;
455 ///The vector loop body.
456 SmallVector<BasicBlock *, 4> LoopVectorBody;
457 ///The scalar loop body.
458 BasicBlock *LoopScalarBody;
459 /// A list of all bypass blocks. The first block is the entry of the loop.
460 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
462 /// The new Induction variable which was added to the new block.
464 /// The induction variable of the old basic block.
465 PHINode *OldInduction;
466 /// Holds the extended (to the widest induction type) start index.
468 /// Maps scalars to widened vectors.
470 EdgeMaskCache MaskCache;
472 LoopVectorizationLegality *Legal;
474 // Record whether runtime check is added.
475 bool AddedSafetyChecks;
478 class InnerLoopUnroller : public InnerLoopVectorizer {
480 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
481 DominatorTree *DT, const TargetLibraryInfo *TLI,
482 const TargetTransformInfo *TTI, unsigned UnrollFactor)
483 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
486 void scalarizeInstruction(Instruction *Instr,
487 bool IfPredicateStore = false) override;
488 void vectorizeMemoryInstruction(Instruction *Instr) override;
489 Value *getBroadcastInstrs(Value *V) override;
490 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
491 Value *reverseVector(Value *Vec) override;
494 /// \brief Look for a meaningful debug location on the instruction or it's
496 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
501 if (I->getDebugLoc() != Empty)
504 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
505 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
506 if (OpInst->getDebugLoc() != Empty)
513 /// \brief Set the debug location in the builder using the debug location in the
515 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
516 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
517 B.SetCurrentDebugLocation(Inst->getDebugLoc());
519 B.SetCurrentDebugLocation(DebugLoc());
523 /// \return string containing a file name and a line # for the given loop.
524 static std::string getDebugLocString(const Loop *L) {
527 raw_string_ostream OS(Result);
528 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
529 LoopDbgLoc.print(OS);
531 // Just print the module name.
532 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
539 /// \brief Propagate known metadata from one instruction to another.
540 static void propagateMetadata(Instruction *To, const Instruction *From) {
541 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
542 From->getAllMetadataOtherThanDebugLoc(Metadata);
544 for (auto M : Metadata) {
545 unsigned Kind = M.first;
547 // These are safe to transfer (this is safe for TBAA, even when we
548 // if-convert, because should that metadata have had a control dependency
549 // on the condition, and thus actually aliased with some other
550 // non-speculated memory access when the condition was false, this would be
551 // caught by the runtime overlap checks).
552 if (Kind != LLVMContext::MD_tbaa &&
553 Kind != LLVMContext::MD_alias_scope &&
554 Kind != LLVMContext::MD_noalias &&
555 Kind != LLVMContext::MD_fpmath &&
556 Kind != LLVMContext::MD_nontemporal)
559 To->setMetadata(Kind, M.second);
563 /// \brief Propagate known metadata from one instruction to a vector of others.
564 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
566 if (Instruction *I = dyn_cast<Instruction>(V))
567 propagateMetadata(I, From);
570 /// \brief The group of interleaved loads/stores sharing the same stride and
571 /// close to each other.
573 /// Each member in this group has an index starting from 0, and the largest
574 /// index should be less than interleaved factor, which is equal to the absolute
575 /// value of the access's stride.
577 /// E.g. An interleaved load group of factor 4:
578 /// for (unsigned i = 0; i < 1024; i+=4) {
579 /// a = A[i]; // Member of index 0
580 /// b = A[i+1]; // Member of index 1
581 /// d = A[i+3]; // Member of index 3
585 /// An interleaved store group of factor 4:
586 /// for (unsigned i = 0; i < 1024; i+=4) {
588 /// A[i] = a; // Member of index 0
589 /// A[i+1] = b; // Member of index 1
590 /// A[i+2] = c; // Member of index 2
591 /// A[i+3] = d; // Member of index 3
594 /// Note: the interleaved load group could have gaps (missing members), but
595 /// the interleaved store group doesn't allow gaps.
596 class InterleaveGroup {
598 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
599 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
600 assert(Align && "The alignment should be non-zero");
602 Factor = std::abs(Stride);
603 assert(Factor > 1 && "Invalid interleave factor");
605 Reverse = Stride < 0;
609 bool isReverse() const { return Reverse; }
610 unsigned getFactor() const { return Factor; }
611 unsigned getAlignment() const { return Align; }
612 unsigned getNumMembers() const { return Members.size(); }
614 /// \brief Try to insert a new member \p Instr with index \p Index and
615 /// alignment \p NewAlign. The index is related to the leader and it could be
616 /// negative if it is the new leader.
618 /// \returns false if the instruction doesn't belong to the group.
619 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
620 assert(NewAlign && "The new member's alignment should be non-zero");
622 int Key = Index + SmallestKey;
624 // Skip if there is already a member with the same index.
625 if (Members.count(Key))
628 if (Key > LargestKey) {
629 // The largest index is always less than the interleave factor.
630 if (Index >= static_cast<int>(Factor))
634 } else if (Key < SmallestKey) {
635 // The largest index is always less than the interleave factor.
636 if (LargestKey - Key >= static_cast<int>(Factor))
642 // It's always safe to select the minimum alignment.
643 Align = std::min(Align, NewAlign);
644 Members[Key] = Instr;
648 /// \brief Get the member with the given index \p Index
650 /// \returns nullptr if contains no such member.
651 Instruction *getMember(unsigned Index) const {
652 int Key = SmallestKey + Index;
653 if (!Members.count(Key))
656 return Members.find(Key)->second;
659 /// \brief Get the index for the given member. Unlike the key in the member
660 /// map, the index starts from 0.
661 unsigned getIndex(Instruction *Instr) const {
662 for (auto I : Members)
663 if (I.second == Instr)
664 return I.first - SmallestKey;
666 llvm_unreachable("InterleaveGroup contains no such member");
669 Instruction *getInsertPos() const { return InsertPos; }
670 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
673 unsigned Factor; // Interleave Factor.
676 DenseMap<int, Instruction *> Members;
680 // To avoid breaking dependences, vectorized instructions of an interleave
681 // group should be inserted at either the first load or the last store in
684 // E.g. %even = load i32 // Insert Position
685 // %add = add i32 %even // Use of %even
689 // %odd = add i32 // Def of %odd
690 // store i32 %odd // Insert Position
691 Instruction *InsertPos;
694 /// \brief Drive the analysis of interleaved memory accesses in the loop.
696 /// Use this class to analyze interleaved accesses only when we can vectorize
697 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
698 /// on interleaved accesses is unsafe.
700 /// The analysis collects interleave groups and records the relationships
701 /// between the member and the group in a map.
702 class InterleavedAccessInfo {
704 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
705 : SE(SE), TheLoop(L), DT(DT) {}
707 ~InterleavedAccessInfo() {
708 SmallSet<InterleaveGroup *, 4> DelSet;
709 // Avoid releasing a pointer twice.
710 for (auto &I : InterleaveGroupMap)
711 DelSet.insert(I.second);
712 for (auto *Ptr : DelSet)
716 /// \brief Analyze the interleaved accesses and collect them in interleave
717 /// groups. Substitute symbolic strides using \p Strides.
718 void analyzeInterleaving(const ValueToValueMap &Strides);
720 /// \brief Check if \p Instr belongs to any interleave group.
721 bool isInterleaved(Instruction *Instr) const {
722 return InterleaveGroupMap.count(Instr);
725 /// \brief Get the interleave group that \p Instr belongs to.
727 /// \returns nullptr if doesn't have such group.
728 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
729 if (InterleaveGroupMap.count(Instr))
730 return InterleaveGroupMap.find(Instr)->second;
739 /// Holds the relationships between the members and the interleave group.
740 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
742 /// \brief The descriptor for a strided memory access.
743 struct StrideDescriptor {
744 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
746 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
748 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
750 int Stride; // The access's stride. It is negative for a reverse access.
751 const SCEV *Scev; // The scalar expression of this access
752 unsigned Size; // The size of the memory object.
753 unsigned Align; // The alignment of this access.
756 /// \brief Create a new interleave group with the given instruction \p Instr,
757 /// stride \p Stride and alignment \p Align.
759 /// \returns the newly created interleave group.
760 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
762 assert(!InterleaveGroupMap.count(Instr) &&
763 "Already in an interleaved access group");
764 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
765 return InterleaveGroupMap[Instr];
768 /// \brief Release the group and remove all the relationships.
769 void releaseGroup(InterleaveGroup *Group) {
770 for (unsigned i = 0; i < Group->getFactor(); i++)
771 if (Instruction *Member = Group->getMember(i))
772 InterleaveGroupMap.erase(Member);
777 /// \brief Collect all the accesses with a constant stride in program order.
778 void collectConstStridedAccesses(
779 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
780 const ValueToValueMap &Strides);
783 /// Utility class for getting and setting loop vectorizer hints in the form
784 /// of loop metadata.
785 /// This class keeps a number of loop annotations locally (as member variables)
786 /// and can, upon request, write them back as metadata on the loop. It will
787 /// initially scan the loop for existing metadata, and will update the local
788 /// values based on information in the loop.
789 /// We cannot write all values to metadata, as the mere presence of some info,
790 /// for example 'force', means a decision has been made. So, we need to be
791 /// careful NOT to add them if the user hasn't specifically asked so.
792 class LoopVectorizeHints {
799 /// Hint - associates name and validation with the hint value.
802 unsigned Value; // This may have to change for non-numeric values.
805 Hint(const char * Name, unsigned Value, HintKind Kind)
806 : Name(Name), Value(Value), Kind(Kind) { }
808 bool validate(unsigned Val) {
811 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
813 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
821 /// Vectorization width.
823 /// Vectorization interleave factor.
825 /// Vectorization forced
828 /// Return the loop metadata prefix.
829 static StringRef Prefix() { return "llvm.loop."; }
833 FK_Undefined = -1, ///< Not selected.
834 FK_Disabled = 0, ///< Forcing disabled.
835 FK_Enabled = 1, ///< Forcing enabled.
838 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
839 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
841 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
842 Force("vectorize.enable", FK_Undefined, HK_FORCE),
844 // Populate values with existing loop metadata.
845 getHintsFromMetadata();
847 // force-vector-interleave overrides DisableInterleaving.
848 if (VectorizerParams::isInterleaveForced())
849 Interleave.Value = VectorizerParams::VectorizationInterleave;
851 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
852 << "LV: Interleaving disabled by the pass manager\n");
855 /// Mark the loop L as already vectorized by setting the width to 1.
856 void setAlreadyVectorized() {
857 Width.Value = Interleave.Value = 1;
858 Hint Hints[] = {Width, Interleave};
859 writeHintsToMetadata(Hints);
862 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
863 if (getForce() == LoopVectorizeHints::FK_Disabled) {
864 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
865 emitOptimizationRemarkAnalysis(F->getContext(),
866 vectorizeAnalysisPassName(), *F,
867 L->getStartLoc(), emitRemark());
871 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
872 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
873 emitOptimizationRemarkAnalysis(F->getContext(),
874 vectorizeAnalysisPassName(), *F,
875 L->getStartLoc(), emitRemark());
879 if (getWidth() == 1 && getInterleave() == 1) {
880 // FIXME: Add a separate metadata to indicate when the loop has already
881 // been vectorized instead of setting width and count to 1.
882 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
883 // FIXME: Add interleave.disable metadata. This will allow
884 // vectorize.disable to be used without disabling the pass and errors
885 // to differentiate between disabled vectorization and a width of 1.
886 emitOptimizationRemarkAnalysis(
887 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
888 "loop not vectorized: vectorization and interleaving are explicitly "
889 "disabled, or vectorize width and interleave count are both set to "
897 /// Dumps all the hint information.
898 std::string emitRemark() const {
899 VectorizationReport R;
900 if (Force.Value == LoopVectorizeHints::FK_Disabled)
901 R << "vectorization is explicitly disabled";
903 R << "use -Rpass-analysis=loop-vectorize for more info";
904 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
906 if (Width.Value != 0)
907 R << ", Vector Width=" << Width.Value;
908 if (Interleave.Value != 0)
909 R << ", Interleave Count=" << Interleave.Value;
917 unsigned getWidth() const { return Width.Value; }
918 unsigned getInterleave() const { return Interleave.Value; }
919 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
920 const char *vectorizeAnalysisPassName() const {
921 // If hints are provided that don't disable vectorization use the
922 // AlwaysPrint pass name to force the frontend to print the diagnostic.
925 if (getForce() == LoopVectorizeHints::FK_Disabled)
927 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
929 return DiagnosticInfo::AlwaysPrint;
933 /// Find hints specified in the loop metadata and update local values.
934 void getHintsFromMetadata() {
935 MDNode *LoopID = TheLoop->getLoopID();
939 // First operand should refer to the loop id itself.
940 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
941 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
943 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
944 const MDString *S = nullptr;
945 SmallVector<Metadata *, 4> Args;
947 // The expected hint is either a MDString or a MDNode with the first
948 // operand a MDString.
949 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
950 if (!MD || MD->getNumOperands() == 0)
952 S = dyn_cast<MDString>(MD->getOperand(0));
953 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
954 Args.push_back(MD->getOperand(i));
956 S = dyn_cast<MDString>(LoopID->getOperand(i));
957 assert(Args.size() == 0 && "too many arguments for MDString");
963 // Check if the hint starts with the loop metadata prefix.
964 StringRef Name = S->getString();
965 if (Args.size() == 1)
966 setHint(Name, Args[0]);
970 /// Checks string hint with one operand and set value if valid.
971 void setHint(StringRef Name, Metadata *Arg) {
972 if (!Name.startswith(Prefix()))
974 Name = Name.substr(Prefix().size(), StringRef::npos);
976 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
978 unsigned Val = C->getZExtValue();
980 Hint *Hints[] = {&Width, &Interleave, &Force};
981 for (auto H : Hints) {
982 if (Name == H->Name) {
983 if (H->validate(Val))
986 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
992 /// Create a new hint from name / value pair.
993 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
994 LLVMContext &Context = TheLoop->getHeader()->getContext();
995 Metadata *MDs[] = {MDString::get(Context, Name),
996 ConstantAsMetadata::get(
997 ConstantInt::get(Type::getInt32Ty(Context), V))};
998 return MDNode::get(Context, MDs);
1001 /// Matches metadata with hint name.
1002 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1003 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1007 for (auto H : HintTypes)
1008 if (Name->getString().endswith(H.Name))
1013 /// Sets current hints into loop metadata, keeping other values intact.
1014 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1015 if (HintTypes.size() == 0)
1018 // Reserve the first element to LoopID (see below).
1019 SmallVector<Metadata *, 4> MDs(1);
1020 // If the loop already has metadata, then ignore the existing operands.
1021 MDNode *LoopID = TheLoop->getLoopID();
1023 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1024 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1025 // If node in update list, ignore old value.
1026 if (!matchesHintMetadataName(Node, HintTypes))
1027 MDs.push_back(Node);
1031 // Now, add the missing hints.
1032 for (auto H : HintTypes)
1033 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1035 // Replace current metadata node with new one.
1036 LLVMContext &Context = TheLoop->getHeader()->getContext();
1037 MDNode *NewLoopID = MDNode::get(Context, MDs);
1038 // Set operand 0 to refer to the loop id itself.
1039 NewLoopID->replaceOperandWith(0, NewLoopID);
1041 TheLoop->setLoopID(NewLoopID);
1044 /// The loop these hints belong to.
1045 const Loop *TheLoop;
1048 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1049 const LoopVectorizeHints &Hints,
1050 const LoopAccessReport &Message) {
1051 const char *Name = Hints.vectorizeAnalysisPassName();
1052 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1055 static void emitMissedWarning(Function *F, Loop *L,
1056 const LoopVectorizeHints &LH) {
1057 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1060 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1061 if (LH.getWidth() != 1)
1062 emitLoopVectorizeWarning(
1063 F->getContext(), *F, L->getStartLoc(),
1064 "failed explicitly specified loop vectorization");
1065 else if (LH.getInterleave() != 1)
1066 emitLoopInterleaveWarning(
1067 F->getContext(), *F, L->getStartLoc(),
1068 "failed explicitly specified loop interleaving");
1072 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1073 /// to what vectorization factor.
1074 /// This class does not look at the profitability of vectorization, only the
1075 /// legality. This class has two main kinds of checks:
1076 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1077 /// will change the order of memory accesses in a way that will change the
1078 /// correctness of the program.
1079 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1080 /// checks for a number of different conditions, such as the availability of a
1081 /// single induction variable, that all types are supported and vectorize-able,
1082 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1083 /// This class is also used by InnerLoopVectorizer for identifying
1084 /// induction variable and the different reduction variables.
1085 class LoopVectorizationLegality {
1087 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1088 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1089 Function *F, const TargetTransformInfo *TTI,
1090 LoopAccessAnalysis *LAA,
1091 LoopVectorizationRequirements *R,
1092 const LoopVectorizeHints *H)
1093 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1094 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1095 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1096 Requirements(R), Hints(H) {}
1098 /// ReductionList contains the reduction descriptors for all
1099 /// of the reductions that were found in the loop.
1100 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1102 /// InductionList saves induction variables and maps them to the
1103 /// induction descriptor.
1104 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1106 /// Returns true if it is legal to vectorize this loop.
1107 /// This does not mean that it is profitable to vectorize this
1108 /// loop, only that it is legal to do so.
1109 bool canVectorize();
1111 /// Returns the Induction variable.
1112 PHINode *getInduction() { return Induction; }
1114 /// Returns the reduction variables found in the loop.
1115 ReductionList *getReductionVars() { return &Reductions; }
1117 /// Returns the induction variables found in the loop.
1118 InductionList *getInductionVars() { return &Inductions; }
1120 /// Returns the widest induction type.
1121 Type *getWidestInductionType() { return WidestIndTy; }
1123 /// Returns True if V is an induction variable in this loop.
1124 bool isInductionVariable(const Value *V);
1126 /// Return true if the block BB needs to be predicated in order for the loop
1127 /// to be vectorized.
1128 bool blockNeedsPredication(BasicBlock *BB);
1130 /// Check if this pointer is consecutive when vectorizing. This happens
1131 /// when the last index of the GEP is the induction variable, or that the
1132 /// pointer itself is an induction variable.
1133 /// This check allows us to vectorize A[idx] into a wide load/store.
1135 /// 0 - Stride is unknown or non-consecutive.
1136 /// 1 - Address is consecutive.
1137 /// -1 - Address is consecutive, and decreasing.
1138 int isConsecutivePtr(Value *Ptr);
1140 /// Returns true if the value V is uniform within the loop.
1141 bool isUniform(Value *V);
1143 /// Returns true if this instruction will remain scalar after vectorization.
1144 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1146 /// Returns the information that we collected about runtime memory check.
1147 const RuntimePointerChecking *getRuntimePointerChecking() const {
1148 return LAI->getRuntimePointerChecking();
1151 const LoopAccessInfo *getLAI() const {
1155 /// \brief Check if \p Instr belongs to any interleaved access group.
1156 bool isAccessInterleaved(Instruction *Instr) {
1157 return InterleaveInfo.isInterleaved(Instr);
1160 /// \brief Get the interleaved access group that \p Instr belongs to.
1161 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1162 return InterleaveInfo.getInterleaveGroup(Instr);
1165 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1167 bool hasStride(Value *V) { return StrideSet.count(V); }
1168 bool mustCheckStrides() { return !StrideSet.empty(); }
1169 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1170 return StrideSet.begin();
1172 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1174 /// Returns true if the target machine supports masked store operation
1175 /// for the given \p DataType and kind of access to \p Ptr.
1176 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1177 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1179 /// Returns true if the target machine supports masked load operation
1180 /// for the given \p DataType and kind of access to \p Ptr.
1181 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1182 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1184 /// Returns true if vector representation of the instruction \p I
1186 bool isMaskRequired(const Instruction* I) {
1187 return (MaskedOp.count(I) != 0);
1189 unsigned getNumStores() const {
1190 return LAI->getNumStores();
1192 unsigned getNumLoads() const {
1193 return LAI->getNumLoads();
1195 unsigned getNumPredStores() const {
1196 return NumPredStores;
1199 /// Check if a single basic block loop is vectorizable.
1200 /// At this point we know that this is a loop with a constant trip count
1201 /// and we only need to check individual instructions.
1202 bool canVectorizeInstrs();
1204 /// When we vectorize loops we may change the order in which
1205 /// we read and write from memory. This method checks if it is
1206 /// legal to vectorize the code, considering only memory constrains.
1207 /// Returns true if the loop is vectorizable
1208 bool canVectorizeMemory();
1210 /// Return true if we can vectorize this loop using the IF-conversion
1212 bool canVectorizeWithIfConvert();
1214 /// Collect the variables that need to stay uniform after vectorization.
1215 void collectLoopUniforms();
1217 /// Return true if all of the instructions in the block can be speculatively
1218 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1219 /// and we know that we can read from them without segfault.
1220 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1222 /// \brief Collect memory access with loop invariant strides.
1224 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1226 void collectStridedAccess(Value *LoadOrStoreInst);
1228 /// Report an analysis message to assist the user in diagnosing loops that are
1229 /// not vectorized. These are handled as LoopAccessReport rather than
1230 /// VectorizationReport because the << operator of VectorizationReport returns
1231 /// LoopAccessReport.
1232 void emitAnalysis(const LoopAccessReport &Message) const {
1233 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1236 unsigned NumPredStores;
1238 /// The loop that we evaluate.
1241 ScalarEvolution *SE;
1242 /// Target Library Info.
1243 TargetLibraryInfo *TLI;
1245 Function *TheFunction;
1246 /// Target Transform Info
1247 const TargetTransformInfo *TTI;
1250 // LoopAccess analysis.
1251 LoopAccessAnalysis *LAA;
1252 // And the loop-accesses info corresponding to this loop. This pointer is
1253 // null until canVectorizeMemory sets it up.
1254 const LoopAccessInfo *LAI;
1256 /// The interleave access information contains groups of interleaved accesses
1257 /// with the same stride and close to each other.
1258 InterleavedAccessInfo InterleaveInfo;
1260 // --- vectorization state --- //
1262 /// Holds the integer induction variable. This is the counter of the
1265 /// Holds the reduction variables.
1266 ReductionList Reductions;
1267 /// Holds all of the induction variables that we found in the loop.
1268 /// Notice that inductions don't need to start at zero and that induction
1269 /// variables can be pointers.
1270 InductionList Inductions;
1271 /// Holds the widest induction type encountered.
1274 /// Allowed outside users. This holds the reduction
1275 /// vars which can be accessed from outside the loop.
1276 SmallPtrSet<Value*, 4> AllowedExit;
1277 /// This set holds the variables which are known to be uniform after
1279 SmallPtrSet<Instruction*, 4> Uniforms;
1281 /// Can we assume the absence of NaNs.
1282 bool HasFunNoNaNAttr;
1284 /// Vectorization requirements that will go through late-evaluation.
1285 LoopVectorizationRequirements *Requirements;
1287 /// Used to emit an analysis of any legality issues.
1288 const LoopVectorizeHints *Hints;
1290 ValueToValueMap Strides;
1291 SmallPtrSet<Value *, 8> StrideSet;
1293 /// While vectorizing these instructions we have to generate a
1294 /// call to the appropriate masked intrinsic
1295 SmallPtrSet<const Instruction*, 8> MaskedOp;
1298 /// LoopVectorizationCostModel - estimates the expected speedups due to
1300 /// In many cases vectorization is not profitable. This can happen because of
1301 /// a number of reasons. In this class we mainly attempt to predict the
1302 /// expected speedup/slowdowns due to the supported instruction set. We use the
1303 /// TargetTransformInfo to query the different backends for the cost of
1304 /// different operations.
1305 class LoopVectorizationCostModel {
1307 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1308 LoopVectorizationLegality *Legal,
1309 const TargetTransformInfo &TTI,
1310 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1311 const Function *F, const LoopVectorizeHints *Hints)
1312 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1313 TheFunction(F), Hints(Hints) {
1314 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
1317 /// Information about vectorization costs
1318 struct VectorizationFactor {
1319 unsigned Width; // Vector width with best cost
1320 unsigned Cost; // Cost of the loop with that width
1322 /// \return The most profitable vectorization factor and the cost of that VF.
1323 /// This method checks every power of two up to VF. If UserVF is not ZERO
1324 /// then this vectorization factor will be selected if vectorization is
1326 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1328 /// \return The size (in bits) of the widest type in the code that
1329 /// needs to be vectorized. We ignore values that remain scalar such as
1330 /// 64 bit loop indices.
1331 unsigned getWidestType();
1333 /// \return The desired interleave count.
1334 /// If interleave count has been specified by metadata it will be returned.
1335 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1336 /// are the selected vectorization factor and the cost of the selected VF.
1337 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1340 /// \return The most profitable unroll factor.
1341 /// This method finds the best unroll-factor based on register pressure and
1342 /// other parameters. VF and LoopCost are the selected vectorization factor
1343 /// and the cost of the selected VF.
1344 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1347 /// \brief A struct that represents some properties of the register usage
1349 struct RegisterUsage {
1350 /// Holds the number of loop invariant values that are used in the loop.
1351 unsigned LoopInvariantRegs;
1352 /// Holds the maximum number of concurrent live intervals in the loop.
1353 unsigned MaxLocalUsers;
1354 /// Holds the number of instructions in the loop.
1355 unsigned NumInstructions;
1358 /// \return information about the register usage of the loop.
1359 RegisterUsage calculateRegisterUsage();
1362 /// Returns the expected execution cost. The unit of the cost does
1363 /// not matter because we use the 'cost' units to compare different
1364 /// vector widths. The cost that is returned is *not* normalized by
1365 /// the factor width.
1366 unsigned expectedCost(unsigned VF);
1368 /// Returns the execution time cost of an instruction for a given vector
1369 /// width. Vector width of one means scalar.
1370 unsigned getInstructionCost(Instruction *I, unsigned VF);
1372 /// Returns whether the instruction is a load or store and will be a emitted
1373 /// as a vector operation.
1374 bool isConsecutiveLoadOrStore(Instruction *I);
1376 /// Report an analysis message to assist the user in diagnosing loops that are
1377 /// not vectorized. These are handled as LoopAccessReport rather than
1378 /// VectorizationReport because the << operator of VectorizationReport returns
1379 /// LoopAccessReport.
1380 void emitAnalysis(const LoopAccessReport &Message) const {
1381 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1384 /// Values used only by @llvm.assume calls.
1385 SmallPtrSet<const Value *, 32> EphValues;
1387 /// The loop that we evaluate.
1390 ScalarEvolution *SE;
1391 /// Loop Info analysis.
1393 /// Vectorization legality.
1394 LoopVectorizationLegality *Legal;
1395 /// Vector target information.
1396 const TargetTransformInfo &TTI;
1397 /// Target Library Info.
1398 const TargetLibraryInfo *TLI;
1399 const Function *TheFunction;
1400 // Loop Vectorize Hint.
1401 const LoopVectorizeHints *Hints;
1404 /// \brief This holds vectorization requirements that must be verified late in
1405 /// the process. The requirements are set by legalize and costmodel. Once
1406 /// vectorization has been determined to be possible and profitable the
1407 /// requirements can be verified by looking for metadata or compiler options.
1408 /// For example, some loops require FP commutativity which is only allowed if
1409 /// vectorization is explicitly specified or if the fast-math compiler option
1410 /// has been provided.
1411 /// Late evaluation of these requirements allows helpful diagnostics to be
1412 /// composed that tells the user what need to be done to vectorize the loop. For
1413 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1414 /// evaluation should be used only when diagnostics can generated that can be
1415 /// followed by a non-expert user.
1416 class LoopVectorizationRequirements {
1418 LoopVectorizationRequirements()
1419 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1421 void addUnsafeAlgebraInst(Instruction *I) {
1422 // First unsafe algebra instruction.
1423 if (!UnsafeAlgebraInst)
1424 UnsafeAlgebraInst = I;
1427 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1429 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1430 const char *Name = Hints.vectorizeAnalysisPassName();
1431 bool Failed = false;
1432 if (UnsafeAlgebraInst &&
1433 Hints.getForce() == LoopVectorizeHints::FK_Undefined &&
1434 Hints.getWidth() == 0) {
1435 emitOptimizationRemarkAnalysisFPCommute(
1436 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1437 VectorizationReport() << "vectorization requires changes in the "
1438 "order of operations, however IEEE 754 "
1439 "floating-point operations are not "
1444 if (NumRuntimePointerChecks >
1445 VectorizerParams::RuntimeMemoryCheckThreshold) {
1446 emitOptimizationRemarkAnalysisAliasing(
1447 F->getContext(), Name, *F, L->getStartLoc(),
1448 VectorizationReport()
1449 << "cannot prove pointers refer to independent arrays in memory. "
1450 "The loop requires "
1451 << NumRuntimePointerChecks
1452 << " runtime independence checks to vectorize the loop, but that "
1453 "would exceed the limit of "
1454 << VectorizerParams::RuntimeMemoryCheckThreshold << " checks");
1455 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1463 unsigned NumRuntimePointerChecks;
1464 Instruction *UnsafeAlgebraInst;
1467 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1469 return V.push_back(&L);
1471 for (Loop *InnerL : L)
1472 addInnerLoop(*InnerL, V);
1475 /// The LoopVectorize Pass.
1476 struct LoopVectorize : public FunctionPass {
1477 /// Pass identification, replacement for typeid
1480 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1482 DisableUnrolling(NoUnrolling),
1483 AlwaysVectorize(AlwaysVectorize) {
1484 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1487 ScalarEvolution *SE;
1489 TargetTransformInfo *TTI;
1491 BlockFrequencyInfo *BFI;
1492 TargetLibraryInfo *TLI;
1494 AssumptionCache *AC;
1495 LoopAccessAnalysis *LAA;
1496 bool DisableUnrolling;
1497 bool AlwaysVectorize;
1499 BlockFrequency ColdEntryFreq;
1501 bool runOnFunction(Function &F) override {
1502 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1503 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1504 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1505 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1506 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1507 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1508 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1509 AA = &getAnalysis<AliasAnalysis>();
1510 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1511 LAA = &getAnalysis<LoopAccessAnalysis>();
1513 // Compute some weights outside of the loop over the loops. Compute this
1514 // using a BranchProbability to re-use its scaling math.
1515 const BranchProbability ColdProb(1, 5); // 20%
1516 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1519 // 1. the target claims to have no vector registers, and
1520 // 2. interleaving won't help ILP.
1522 // The second condition is necessary because, even if the target has no
1523 // vector registers, loop vectorization may still enable scalar
1525 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1528 // Build up a worklist of inner-loops to vectorize. This is necessary as
1529 // the act of vectorizing or partially unrolling a loop creates new loops
1530 // and can invalidate iterators across the loops.
1531 SmallVector<Loop *, 8> Worklist;
1534 addInnerLoop(*L, Worklist);
1536 LoopsAnalyzed += Worklist.size();
1538 // Now walk the identified inner loops.
1539 bool Changed = false;
1540 while (!Worklist.empty())
1541 Changed |= processLoop(Worklist.pop_back_val());
1543 // Process each loop nest in the function.
1547 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1548 SmallVector<Metadata *, 4> MDs;
1549 // Reserve first location for self reference to the LoopID metadata node.
1550 MDs.push_back(nullptr);
1551 bool IsUnrollMetadata = false;
1552 MDNode *LoopID = L->getLoopID();
1554 // First find existing loop unrolling disable metadata.
1555 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1556 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1558 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1560 S && S->getString().startswith("llvm.loop.unroll.disable");
1562 MDs.push_back(LoopID->getOperand(i));
1566 if (!IsUnrollMetadata) {
1567 // Add runtime unroll disable metadata.
1568 LLVMContext &Context = L->getHeader()->getContext();
1569 SmallVector<Metadata *, 1> DisableOperands;
1570 DisableOperands.push_back(
1571 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1572 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1573 MDs.push_back(DisableNode);
1574 MDNode *NewLoopID = MDNode::get(Context, MDs);
1575 // Set operand 0 to refer to the loop id itself.
1576 NewLoopID->replaceOperandWith(0, NewLoopID);
1577 L->setLoopID(NewLoopID);
1581 bool processLoop(Loop *L) {
1582 assert(L->empty() && "Only process inner loops.");
1585 const std::string DebugLocStr = getDebugLocString(L);
1588 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1589 << L->getHeader()->getParent()->getName() << "\" from "
1590 << DebugLocStr << "\n");
1592 LoopVectorizeHints Hints(L, DisableUnrolling);
1594 DEBUG(dbgs() << "LV: Loop hints:"
1596 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1598 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1600 : "?")) << " width=" << Hints.getWidth()
1601 << " unroll=" << Hints.getInterleave() << "\n");
1603 // Function containing loop
1604 Function *F = L->getHeader()->getParent();
1606 // Looking at the diagnostic output is the only way to determine if a loop
1607 // was vectorized (other than looking at the IR or machine code), so it
1608 // is important to generate an optimization remark for each loop. Most of
1609 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1610 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1611 // less verbose reporting vectorized loops and unvectorized loops that may
1612 // benefit from vectorization, respectively.
1614 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1615 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1619 // Check the loop for a trip count threshold:
1620 // do not vectorize loops with a tiny trip count.
1621 const unsigned TC = SE->getSmallConstantTripCount(L);
1622 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1623 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1624 << "This loop is not worth vectorizing.");
1625 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1626 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1628 DEBUG(dbgs() << "\n");
1629 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1630 << "vectorization is not beneficial "
1631 "and is not explicitly forced");
1636 // Check if it is legal to vectorize the loop.
1637 LoopVectorizationRequirements Requirements;
1638 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1639 &Requirements, &Hints);
1640 if (!LVL.canVectorize()) {
1641 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1642 emitMissedWarning(F, L, Hints);
1646 // Use the cost model.
1647 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1649 // Check the function attributes to find out if this function should be
1650 // optimized for size.
1651 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1654 // Compute the weighted frequency of this loop being executed and see if it
1655 // is less than 20% of the function entry baseline frequency. Note that we
1656 // always have a canonical loop here because we think we *can* vectorize.
1657 // FIXME: This is hidden behind a flag due to pervasive problems with
1658 // exactly what block frequency models.
1659 if (LoopVectorizeWithBlockFrequency) {
1660 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1661 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1662 LoopEntryFreq < ColdEntryFreq)
1666 // Check the function attributes to see if implicit floats are allowed.
1667 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1668 // an integer loop and the vector instructions selected are purely integer
1669 // vector instructions?
1670 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1671 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1672 "attribute is used.\n");
1675 VectorizationReport()
1676 << "loop not vectorized due to NoImplicitFloat attribute");
1677 emitMissedWarning(F, L, Hints);
1681 // Select the optimal vectorization factor.
1682 const LoopVectorizationCostModel::VectorizationFactor VF =
1683 CM.selectVectorizationFactor(OptForSize);
1685 // Select the interleave count.
1686 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1688 // Get user interleave count.
1689 unsigned UserIC = Hints.getInterleave();
1691 // Identify the diagnostic messages that should be produced.
1692 std::string VecDiagMsg, IntDiagMsg;
1693 bool VectorizeLoop = true, InterleaveLoop = true;
1695 if (Requirements.doesNotMeet(F, L, Hints)) {
1696 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1698 emitMissedWarning(F, L, Hints);
1702 if (VF.Width == 1) {
1703 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1705 "the cost-model indicates that vectorization is not beneficial";
1706 VectorizeLoop = false;
1709 if (IC == 1 && UserIC <= 1) {
1710 // Tell the user interleaving is not beneficial.
1711 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1713 "the cost-model indicates that interleaving is not beneficial";
1714 InterleaveLoop = false;
1717 " and is explicitly disabled or interleave count is set to 1";
1718 } else if (IC > 1 && UserIC == 1) {
1719 // Tell the user interleaving is beneficial, but it explicitly disabled.
1721 << "LV: Interleaving is beneficial but is explicitly disabled.");
1722 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1723 "but is explicitly disabled or interleave count is set to 1";
1724 InterleaveLoop = false;
1727 // Override IC if user provided an interleave count.
1728 IC = UserIC > 0 ? UserIC : IC;
1730 // Emit diagnostic messages, if any.
1731 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1732 if (!VectorizeLoop && !InterleaveLoop) {
1733 // Do not vectorize or interleaving the loop.
1734 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1735 L->getStartLoc(), VecDiagMsg);
1736 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1737 L->getStartLoc(), IntDiagMsg);
1739 } else if (!VectorizeLoop && InterleaveLoop) {
1740 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1741 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1742 L->getStartLoc(), VecDiagMsg);
1743 } else if (VectorizeLoop && !InterleaveLoop) {
1744 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1745 << DebugLocStr << '\n');
1746 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1747 L->getStartLoc(), IntDiagMsg);
1748 } else if (VectorizeLoop && InterleaveLoop) {
1749 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1750 << DebugLocStr << '\n');
1751 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1754 if (!VectorizeLoop) {
1755 assert(IC > 1 && "interleave count should not be 1 or 0");
1756 // If we decided that it is not legal to vectorize the loop then
1758 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1759 Unroller.vectorize(&LVL);
1761 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1762 Twine("interleaved loop (interleaved count: ") +
1765 // If we decided that it is *legal* to vectorize the loop then do it.
1766 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1770 // Add metadata to disable runtime unrolling scalar loop when there's no
1771 // runtime check about strides and memory. Because at this situation,
1772 // scalar loop is rarely used not worthy to be unrolled.
1773 if (!LB.IsSafetyChecksAdded())
1774 AddRuntimeUnrollDisableMetaData(L);
1776 // Report the vectorization decision.
1777 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1778 Twine("vectorized loop (vectorization width: ") +
1779 Twine(VF.Width) + ", interleaved count: " +
1783 // Mark the loop as already vectorized to avoid vectorizing again.
1784 Hints.setAlreadyVectorized();
1786 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1790 void getAnalysisUsage(AnalysisUsage &AU) const override {
1791 AU.addRequired<AssumptionCacheTracker>();
1792 AU.addRequiredID(LoopSimplifyID);
1793 AU.addRequiredID(LCSSAID);
1794 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1795 AU.addRequired<DominatorTreeWrapperPass>();
1796 AU.addRequired<LoopInfoWrapperPass>();
1797 AU.addRequired<ScalarEvolutionWrapperPass>();
1798 AU.addRequired<TargetTransformInfoWrapperPass>();
1799 AU.addRequired<AliasAnalysis>();
1800 AU.addRequired<LoopAccessAnalysis>();
1801 AU.addPreserved<LoopInfoWrapperPass>();
1802 AU.addPreserved<DominatorTreeWrapperPass>();
1803 AU.addPreserved<AliasAnalysis>();
1808 } // end anonymous namespace
1810 //===----------------------------------------------------------------------===//
1811 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1812 // LoopVectorizationCostModel.
1813 //===----------------------------------------------------------------------===//
1815 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1816 // We need to place the broadcast of invariant variables outside the loop.
1817 Instruction *Instr = dyn_cast<Instruction>(V);
1819 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1820 Instr->getParent()) != LoopVectorBody.end());
1821 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1823 // Place the code for broadcasting invariant variables in the new preheader.
1824 IRBuilder<>::InsertPointGuard Guard(Builder);
1826 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1828 // Broadcast the scalar into all locations in the vector.
1829 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1834 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1836 assert(Val->getType()->isVectorTy() && "Must be a vector");
1837 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1838 "Elem must be an integer");
1839 assert(Step->getType() == Val->getType()->getScalarType() &&
1840 "Step has wrong type");
1841 // Create the types.
1842 Type *ITy = Val->getType()->getScalarType();
1843 VectorType *Ty = cast<VectorType>(Val->getType());
1844 int VLen = Ty->getNumElements();
1845 SmallVector<Constant*, 8> Indices;
1847 // Create a vector of consecutive numbers from zero to VF.
1848 for (int i = 0; i < VLen; ++i)
1849 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1851 // Add the consecutive indices to the vector value.
1852 Constant *Cv = ConstantVector::get(Indices);
1853 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1854 Step = Builder.CreateVectorSplat(VLen, Step);
1855 assert(Step->getType() == Val->getType() && "Invalid step vec");
1856 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1857 // which can be found from the original scalar operations.
1858 Step = Builder.CreateMul(Cv, Step);
1859 return Builder.CreateAdd(Val, Step, "induction");
1862 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1863 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1864 // Make sure that the pointer does not point to structs.
1865 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1868 // If this value is a pointer induction variable we know it is consecutive.
1869 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1870 if (Phi && Inductions.count(Phi)) {
1871 InductionDescriptor II = Inductions[Phi];
1872 return II.getConsecutiveDirection();
1875 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1879 unsigned NumOperands = Gep->getNumOperands();
1880 Value *GpPtr = Gep->getPointerOperand();
1881 // If this GEP value is a consecutive pointer induction variable and all of
1882 // the indices are constant then we know it is consecutive. We can
1883 Phi = dyn_cast<PHINode>(GpPtr);
1884 if (Phi && Inductions.count(Phi)) {
1886 // Make sure that the pointer does not point to structs.
1887 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1888 if (GepPtrType->getElementType()->isAggregateType())
1891 // Make sure that all of the index operands are loop invariant.
1892 for (unsigned i = 1; i < NumOperands; ++i)
1893 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1896 InductionDescriptor II = Inductions[Phi];
1897 return II.getConsecutiveDirection();
1900 unsigned InductionOperand = getGEPInductionOperand(Gep);
1902 // Check that all of the gep indices are uniform except for our induction
1904 for (unsigned i = 0; i != NumOperands; ++i)
1905 if (i != InductionOperand &&
1906 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1909 // We can emit wide load/stores only if the last non-zero index is the
1910 // induction variable.
1911 const SCEV *Last = nullptr;
1912 if (!Strides.count(Gep))
1913 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1915 // Because of the multiplication by a stride we can have a s/zext cast.
1916 // We are going to replace this stride by 1 so the cast is safe to ignore.
1918 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1919 // %0 = trunc i64 %indvars.iv to i32
1920 // %mul = mul i32 %0, %Stride1
1921 // %idxprom = zext i32 %mul to i64 << Safe cast.
1922 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1924 Last = replaceSymbolicStrideSCEV(SE, Strides,
1925 Gep->getOperand(InductionOperand), Gep);
1926 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1928 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1932 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1933 const SCEV *Step = AR->getStepRecurrence(*SE);
1935 // The memory is consecutive because the last index is consecutive
1936 // and all other indices are loop invariant.
1939 if (Step->isAllOnesValue())
1946 bool LoopVectorizationLegality::isUniform(Value *V) {
1947 return LAI->isUniform(V);
1950 InnerLoopVectorizer::VectorParts&
1951 InnerLoopVectorizer::getVectorValue(Value *V) {
1952 assert(V != Induction && "The new induction variable should not be used.");
1953 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1955 // If we have a stride that is replaced by one, do it here.
1956 if (Legal->hasStride(V))
1957 V = ConstantInt::get(V->getType(), 1);
1959 // If we have this scalar in the map, return it.
1960 if (WidenMap.has(V))
1961 return WidenMap.get(V);
1963 // If this scalar is unknown, assume that it is a constant or that it is
1964 // loop invariant. Broadcast V and save the value for future uses.
1965 Value *B = getBroadcastInstrs(V);
1966 return WidenMap.splat(V, B);
1969 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1970 assert(Vec->getType()->isVectorTy() && "Invalid type");
1971 SmallVector<Constant*, 8> ShuffleMask;
1972 for (unsigned i = 0; i < VF; ++i)
1973 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1975 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1976 ConstantVector::get(ShuffleMask),
1980 // Get a mask to interleave \p NumVec vectors into a wide vector.
1981 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
1982 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
1983 // <0, 4, 1, 5, 2, 6, 3, 7>
1984 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
1986 SmallVector<Constant *, 16> Mask;
1987 for (unsigned i = 0; i < VF; i++)
1988 for (unsigned j = 0; j < NumVec; j++)
1989 Mask.push_back(Builder.getInt32(j * VF + i));
1991 return ConstantVector::get(Mask);
1994 // Get the strided mask starting from index \p Start.
1995 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
1996 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
1997 unsigned Stride, unsigned VF) {
1998 SmallVector<Constant *, 16> Mask;
1999 for (unsigned i = 0; i < VF; i++)
2000 Mask.push_back(Builder.getInt32(Start + i * Stride));
2002 return ConstantVector::get(Mask);
2005 // Get a mask of two parts: The first part consists of sequential integers
2006 // starting from 0, The second part consists of UNDEFs.
2007 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2008 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2009 unsigned NumUndef) {
2010 SmallVector<Constant *, 16> Mask;
2011 for (unsigned i = 0; i < NumInt; i++)
2012 Mask.push_back(Builder.getInt32(i));
2014 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2015 for (unsigned i = 0; i < NumUndef; i++)
2016 Mask.push_back(Undef);
2018 return ConstantVector::get(Mask);
2021 // Concatenate two vectors with the same element type. The 2nd vector should
2022 // not have more elements than the 1st vector. If the 2nd vector has less
2023 // elements, extend it with UNDEFs.
2024 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2026 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2027 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2028 assert(VecTy1 && VecTy2 &&
2029 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2030 "Expect two vectors with the same element type");
2032 unsigned NumElts1 = VecTy1->getNumElements();
2033 unsigned NumElts2 = VecTy2->getNumElements();
2034 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2036 if (NumElts1 > NumElts2) {
2037 // Extend with UNDEFs.
2039 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2040 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2043 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2044 return Builder.CreateShuffleVector(V1, V2, Mask);
2047 // Concatenate vectors in the given list. All vectors have the same type.
2048 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2049 ArrayRef<Value *> InputList) {
2050 unsigned NumVec = InputList.size();
2051 assert(NumVec > 1 && "Should be at least two vectors");
2053 SmallVector<Value *, 8> ResList;
2054 ResList.append(InputList.begin(), InputList.end());
2056 SmallVector<Value *, 8> TmpList;
2057 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2058 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2059 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2060 "Only the last vector may have a different type");
2062 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2065 // Push the last vector if the total number of vectors is odd.
2066 if (NumVec % 2 != 0)
2067 TmpList.push_back(ResList[NumVec - 1]);
2070 NumVec = ResList.size();
2071 } while (NumVec > 1);
2076 // Try to vectorize the interleave group that \p Instr belongs to.
2078 // E.g. Translate following interleaved load group (factor = 3):
2079 // for (i = 0; i < N; i+=3) {
2080 // R = Pic[i]; // Member of index 0
2081 // G = Pic[i+1]; // Member of index 1
2082 // B = Pic[i+2]; // Member of index 2
2083 // ... // do something to R, G, B
2086 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2087 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2088 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2089 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2091 // Or translate following interleaved store group (factor = 3):
2092 // for (i = 0; i < N; i+=3) {
2093 // ... do something to R, G, B
2094 // Pic[i] = R; // Member of index 0
2095 // Pic[i+1] = G; // Member of index 1
2096 // Pic[i+2] = B; // Member of index 2
2099 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2100 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2101 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2102 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2103 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2104 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2105 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2106 assert(Group && "Fail to get an interleaved access group.");
2108 // Skip if current instruction is not the insert position.
2109 if (Instr != Group->getInsertPos())
2112 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2113 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2114 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2116 // Prepare for the vector type of the interleaved load/store.
2117 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2118 unsigned InterleaveFactor = Group->getFactor();
2119 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2120 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2122 // Prepare for the new pointers.
2123 setDebugLocFromInst(Builder, Ptr);
2124 VectorParts &PtrParts = getVectorValue(Ptr);
2125 SmallVector<Value *, 2> NewPtrs;
2126 unsigned Index = Group->getIndex(Instr);
2127 for (unsigned Part = 0; Part < UF; Part++) {
2128 // Extract the pointer for current instruction from the pointer vector. A
2129 // reverse access uses the pointer in the last lane.
2130 Value *NewPtr = Builder.CreateExtractElement(
2132 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2134 // Notice current instruction could be any index. Need to adjust the address
2135 // to the member of index 0.
2137 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2138 // b = A[i]; // Member of index 0
2139 // Current pointer is pointed to A[i+1], adjust it to A[i].
2141 // E.g. A[i+1] = a; // Member of index 1
2142 // A[i] = b; // Member of index 0
2143 // A[i+2] = c; // Member of index 2 (Current instruction)
2144 // Current pointer is pointed to A[i+2], adjust it to A[i].
2145 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2147 // Cast to the vector pointer type.
2148 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2151 setDebugLocFromInst(Builder, Instr);
2152 Value *UndefVec = UndefValue::get(VecTy);
2154 // Vectorize the interleaved load group.
2156 for (unsigned Part = 0; Part < UF; Part++) {
2157 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2158 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2160 for (unsigned i = 0; i < InterleaveFactor; i++) {
2161 Instruction *Member = Group->getMember(i);
2163 // Skip the gaps in the group.
2167 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2168 Value *StridedVec = Builder.CreateShuffleVector(
2169 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2171 // If this member has different type, cast the result type.
2172 if (Member->getType() != ScalarTy) {
2173 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2174 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2177 VectorParts &Entry = WidenMap.get(Member);
2179 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2182 propagateMetadata(NewLoadInstr, Instr);
2187 // The sub vector type for current instruction.
2188 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2190 // Vectorize the interleaved store group.
2191 for (unsigned Part = 0; Part < UF; Part++) {
2192 // Collect the stored vector from each member.
2193 SmallVector<Value *, 4> StoredVecs;
2194 for (unsigned i = 0; i < InterleaveFactor; i++) {
2195 // Interleaved store group doesn't allow a gap, so each index has a member
2196 Instruction *Member = Group->getMember(i);
2197 assert(Member && "Fail to get a member from an interleaved store group");
2200 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2201 if (Group->isReverse())
2202 StoredVec = reverseVector(StoredVec);
2204 // If this member has different type, cast it to an unified type.
2205 if (StoredVec->getType() != SubVT)
2206 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2208 StoredVecs.push_back(StoredVec);
2211 // Concatenate all vectors into a wide vector.
2212 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2214 // Interleave the elements in the wide vector.
2215 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2216 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2219 Instruction *NewStoreInstr =
2220 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2221 propagateMetadata(NewStoreInstr, Instr);
2225 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2226 // Attempt to issue a wide load.
2227 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2228 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2230 assert((LI || SI) && "Invalid Load/Store instruction");
2232 // Try to vectorize the interleave group if this access is interleaved.
2233 if (Legal->isAccessInterleaved(Instr))
2234 return vectorizeInterleaveGroup(Instr);
2236 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2237 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2238 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2239 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2240 // An alignment of 0 means target abi alignment. We need to use the scalar's
2241 // target abi alignment in such a case.
2242 const DataLayout &DL = Instr->getModule()->getDataLayout();
2244 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2245 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2246 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2247 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2249 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2250 !Legal->isMaskRequired(SI))
2251 return scalarizeInstruction(Instr, true);
2253 if (ScalarAllocatedSize != VectorElementSize)
2254 return scalarizeInstruction(Instr);
2256 // If the pointer is loop invariant or if it is non-consecutive,
2257 // scalarize the load.
2258 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2259 bool Reverse = ConsecutiveStride < 0;
2260 bool UniformLoad = LI && Legal->isUniform(Ptr);
2261 if (!ConsecutiveStride || UniformLoad)
2262 return scalarizeInstruction(Instr);
2264 Constant *Zero = Builder.getInt32(0);
2265 VectorParts &Entry = WidenMap.get(Instr);
2267 // Handle consecutive loads/stores.
2268 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2269 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2270 setDebugLocFromInst(Builder, Gep);
2271 Value *PtrOperand = Gep->getPointerOperand();
2272 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2273 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2275 // Create the new GEP with the new induction variable.
2276 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2277 Gep2->setOperand(0, FirstBasePtr);
2278 Gep2->setName("gep.indvar.base");
2279 Ptr = Builder.Insert(Gep2);
2281 setDebugLocFromInst(Builder, Gep);
2282 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2283 OrigLoop) && "Base ptr must be invariant");
2285 // The last index does not have to be the induction. It can be
2286 // consecutive and be a function of the index. For example A[I+1];
2287 unsigned NumOperands = Gep->getNumOperands();
2288 unsigned InductionOperand = getGEPInductionOperand(Gep);
2289 // Create the new GEP with the new induction variable.
2290 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2292 for (unsigned i = 0; i < NumOperands; ++i) {
2293 Value *GepOperand = Gep->getOperand(i);
2294 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2296 // Update last index or loop invariant instruction anchored in loop.
2297 if (i == InductionOperand ||
2298 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2299 assert((i == InductionOperand ||
2300 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2301 "Must be last index or loop invariant");
2303 VectorParts &GEPParts = getVectorValue(GepOperand);
2304 Value *Index = GEPParts[0];
2305 Index = Builder.CreateExtractElement(Index, Zero);
2306 Gep2->setOperand(i, Index);
2307 Gep2->setName("gep.indvar.idx");
2310 Ptr = Builder.Insert(Gep2);
2312 // Use the induction element ptr.
2313 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2314 setDebugLocFromInst(Builder, Ptr);
2315 VectorParts &PtrVal = getVectorValue(Ptr);
2316 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2319 VectorParts Mask = createBlockInMask(Instr->getParent());
2322 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2323 "We do not allow storing to uniform addresses");
2324 setDebugLocFromInst(Builder, SI);
2325 // We don't want to update the value in the map as it might be used in
2326 // another expression. So don't use a reference type for "StoredVal".
2327 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2329 for (unsigned Part = 0; Part < UF; ++Part) {
2330 // Calculate the pointer for the specific unroll-part.
2332 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2335 // If we store to reverse consecutive memory locations, then we need
2336 // to reverse the order of elements in the stored value.
2337 StoredVal[Part] = reverseVector(StoredVal[Part]);
2338 // If the address is consecutive but reversed, then the
2339 // wide store needs to start at the last vector element.
2340 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2341 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2342 Mask[Part] = reverseVector(Mask[Part]);
2345 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2346 DataTy->getPointerTo(AddressSpace));
2349 if (Legal->isMaskRequired(SI))
2350 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2353 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2354 propagateMetadata(NewSI, SI);
2360 assert(LI && "Must have a load instruction");
2361 setDebugLocFromInst(Builder, LI);
2362 for (unsigned Part = 0; Part < UF; ++Part) {
2363 // Calculate the pointer for the specific unroll-part.
2365 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2368 // If the address is consecutive but reversed, then the
2369 // wide load needs to start at the last vector element.
2370 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2371 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2372 Mask[Part] = reverseVector(Mask[Part]);
2376 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2377 DataTy->getPointerTo(AddressSpace));
2378 if (Legal->isMaskRequired(LI))
2379 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2380 UndefValue::get(DataTy),
2381 "wide.masked.load");
2383 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2384 propagateMetadata(NewLI, LI);
2385 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2389 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2390 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2391 // Holds vector parameters or scalars, in case of uniform vals.
2392 SmallVector<VectorParts, 4> Params;
2394 setDebugLocFromInst(Builder, Instr);
2396 // Find all of the vectorized parameters.
2397 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2398 Value *SrcOp = Instr->getOperand(op);
2400 // If we are accessing the old induction variable, use the new one.
2401 if (SrcOp == OldInduction) {
2402 Params.push_back(getVectorValue(SrcOp));
2406 // Try using previously calculated values.
2407 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2409 // If the src is an instruction that appeared earlier in the basic block,
2410 // then it should already be vectorized.
2411 if (SrcInst && OrigLoop->contains(SrcInst)) {
2412 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2413 // The parameter is a vector value from earlier.
2414 Params.push_back(WidenMap.get(SrcInst));
2416 // The parameter is a scalar from outside the loop. Maybe even a constant.
2417 VectorParts Scalars;
2418 Scalars.append(UF, SrcOp);
2419 Params.push_back(Scalars);
2423 assert(Params.size() == Instr->getNumOperands() &&
2424 "Invalid number of operands");
2426 // Does this instruction return a value ?
2427 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2429 Value *UndefVec = IsVoidRetTy ? nullptr :
2430 UndefValue::get(VectorType::get(Instr->getType(), VF));
2431 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2432 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2434 Instruction *InsertPt = Builder.GetInsertPoint();
2435 BasicBlock *IfBlock = Builder.GetInsertBlock();
2436 BasicBlock *CondBlock = nullptr;
2439 Loop *VectorLp = nullptr;
2440 if (IfPredicateStore) {
2441 assert(Instr->getParent()->getSinglePredecessor() &&
2442 "Only support single predecessor blocks");
2443 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2444 Instr->getParent());
2445 VectorLp = LI->getLoopFor(IfBlock);
2446 assert(VectorLp && "Must have a loop for this block");
2449 // For each vector unroll 'part':
2450 for (unsigned Part = 0; Part < UF; ++Part) {
2451 // For each scalar that we create:
2452 for (unsigned Width = 0; Width < VF; ++Width) {
2455 Value *Cmp = nullptr;
2456 if (IfPredicateStore) {
2457 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2458 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2459 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2460 LoopVectorBody.push_back(CondBlock);
2461 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2462 // Update Builder with newly created basic block.
2463 Builder.SetInsertPoint(InsertPt);
2466 Instruction *Cloned = Instr->clone();
2468 Cloned->setName(Instr->getName() + ".cloned");
2469 // Replace the operands of the cloned instructions with extracted scalars.
2470 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2471 Value *Op = Params[op][Part];
2472 // Param is a vector. Need to extract the right lane.
2473 if (Op->getType()->isVectorTy())
2474 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2475 Cloned->setOperand(op, Op);
2478 // Place the cloned scalar in the new loop.
2479 Builder.Insert(Cloned);
2481 // If the original scalar returns a value we need to place it in a vector
2482 // so that future users will be able to use it.
2484 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2485 Builder.getInt32(Width));
2487 if (IfPredicateStore) {
2488 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2489 LoopVectorBody.push_back(NewIfBlock);
2490 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2491 Builder.SetInsertPoint(InsertPt);
2492 ReplaceInstWithInst(IfBlock->getTerminator(),
2493 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2494 IfBlock = NewIfBlock;
2500 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2504 if (Instruction *I = dyn_cast<Instruction>(V))
2505 return I->getParent() == Loc->getParent() ? I : nullptr;
2509 std::pair<Instruction *, Instruction *>
2510 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2511 Instruction *tnullptr = nullptr;
2512 if (!Legal->mustCheckStrides())
2513 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2515 IRBuilder<> ChkBuilder(Loc);
2518 Value *Check = nullptr;
2519 Instruction *FirstInst = nullptr;
2520 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2521 SE = Legal->strides_end();
2523 Value *Ptr = stripIntegerCast(*SI);
2524 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2526 // Store the first instruction we create.
2527 FirstInst = getFirstInst(FirstInst, C, Loc);
2529 Check = ChkBuilder.CreateOr(Check, C);
2534 // We have to do this trickery because the IRBuilder might fold the check to a
2535 // constant expression in which case there is no Instruction anchored in a
2537 LLVMContext &Ctx = Loc->getContext();
2538 Instruction *TheCheck =
2539 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2540 ChkBuilder.Insert(TheCheck, "stride.not.one");
2541 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2543 return std::make_pair(FirstInst, TheCheck);
2546 void InnerLoopVectorizer::createEmptyLoop() {
2548 In this function we generate a new loop. The new loop will contain
2549 the vectorized instructions while the old loop will continue to run the
2552 [ ] <-- loop iteration number check.
2555 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2558 || [ ] <-- vector pre header.
2562 || [ ]_| <-- vector loop.
2565 | >[ ] <--- middle-block.
2568 -|- >[ ] <--- new preheader.
2572 | [ ]_| <-- old scalar loop to handle remainder.
2575 >[ ] <-- exit block.
2579 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2580 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2581 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2582 assert(VectorPH && "Invalid loop structure");
2583 assert(ExitBlock && "Must have an exit block");
2585 // Some loops have a single integer induction variable, while other loops
2586 // don't. One example is c++ iterators that often have multiple pointer
2587 // induction variables. In the code below we also support a case where we
2588 // don't have a single induction variable.
2589 OldInduction = Legal->getInduction();
2590 Type *IdxTy = Legal->getWidestInductionType();
2592 // Find the loop boundaries.
2593 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2594 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2596 // The exit count might have the type of i64 while the phi is i32. This can
2597 // happen if we have an induction variable that is sign extended before the
2598 // compare. The only way that we get a backedge taken count is that the
2599 // induction variable was signed and as such will not overflow. In such a case
2600 // truncation is legal.
2601 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2602 IdxTy->getPrimitiveSizeInBits())
2603 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2605 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2606 // Get the total trip count from the count by adding 1.
2607 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2608 SE->getConstant(BackedgeTakeCount->getType(), 1));
2610 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2612 // Expand the trip count and place the new instructions in the preheader.
2613 // Notice that the pre-header does not change, only the loop body.
2614 SCEVExpander Exp(*SE, DL, "induction");
2616 // The loop minimum iterations check below is to ensure the loop has enough
2617 // trip count so the generated vector loop will likely be executed and the
2618 // preparation and rounding-off costs will likely be worthy.
2620 // The minimum iteration check also covers case where the backedge-taken
2621 // count is uint##_max. Adding one to it will cause overflow and an
2622 // incorrect loop trip count being generated in the vector body. In this
2623 // case we also want to directly jump to the scalar remainder loop.
2624 Value *ExitCountValue = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2625 VectorPH->getTerminator());
2626 if (ExitCountValue->getType()->isPointerTy())
2627 ExitCountValue = CastInst::CreatePointerCast(ExitCountValue, IdxTy,
2628 "exitcount.ptrcnt.to.int",
2629 VectorPH->getTerminator());
2631 Instruction *CheckMinIters =
2632 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULT, ExitCountValue,
2633 ConstantInt::get(ExitCountValue->getType(), VF * UF),
2634 "min.iters.check", VectorPH->getTerminator());
2636 // The loop index does not have to start at Zero. Find the original start
2637 // value from the induction PHI node. If we don't have an induction variable
2638 // then we know that it starts at zero.
2639 Builder.SetInsertPoint(VectorPH->getTerminator());
2640 Value *StartIdx = ExtendedIdx =
2642 ? Builder.CreateZExt(OldInduction->getIncomingValueForBlock(VectorPH),
2644 : ConstantInt::get(IdxTy, 0);
2646 // Count holds the overall loop count (N).
2647 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2648 VectorPH->getTerminator());
2650 LoopBypassBlocks.push_back(VectorPH);
2652 // Split the single block loop into the two loop structure described above.
2653 BasicBlock *VecBody =
2654 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2655 BasicBlock *MiddleBlock =
2656 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2657 BasicBlock *ScalarPH =
2658 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2660 // Create and register the new vector loop.
2661 Loop* Lp = new Loop();
2662 Loop *ParentLoop = OrigLoop->getParentLoop();
2664 // Insert the new loop into the loop nest and register the new basic blocks
2665 // before calling any utilities such as SCEV that require valid LoopInfo.
2667 ParentLoop->addChildLoop(Lp);
2668 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2669 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2671 LI->addTopLevelLoop(Lp);
2673 Lp->addBasicBlockToLoop(VecBody, *LI);
2675 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2677 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2679 // Generate the induction variable.
2680 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2681 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2682 // The loop step is equal to the vectorization factor (num of SIMD elements)
2683 // times the unroll factor (num of SIMD instructions).
2684 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2686 // Generate code to check that the loop's trip count is not less than the
2687 // minimum loop iteration number threshold.
2688 BasicBlock *NewVectorPH =
2689 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "min.iters.checked");
2691 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2692 ReplaceInstWithInst(VectorPH->getTerminator(),
2693 BranchInst::Create(ScalarPH, NewVectorPH, CheckMinIters));
2694 VectorPH = NewVectorPH;
2696 // This is the IR builder that we use to add all of the logic for bypassing
2697 // the new vector loop.
2698 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2699 setDebugLocFromInst(BypassBuilder,
2700 getDebugLocFromInstOrOperands(OldInduction));
2702 // We may need to extend the index in case there is a type mismatch.
2703 // We know that the count starts at zero and does not overflow.
2704 if (Count->getType() != IdxTy) {
2705 // The exit count can be of pointer type. Convert it to the correct
2707 if (ExitCount->getType()->isPointerTy())
2708 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2710 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2713 // Add the start index to the loop count to get the new end index.
2714 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2716 // Now we need to generate the expression for N - (N % VF), which is
2717 // the part that the vectorized body will execute.
2718 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2719 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2720 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2721 "end.idx.rnd.down");
2723 // Now, compare the new count to zero. If it is zero skip the vector loop and
2724 // jump to the scalar loop.
2726 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2728 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2730 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2731 LoopBypassBlocks.push_back(VectorPH);
2732 ReplaceInstWithInst(VectorPH->getTerminator(),
2733 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2734 VectorPH = NewVectorPH;
2736 // Generate the code to check that the strides we assumed to be one are really
2737 // one. We want the new basic block to start at the first instruction in a
2738 // sequence of instructions that form a check.
2739 Instruction *StrideCheck;
2740 Instruction *FirstCheckInst;
2741 std::tie(FirstCheckInst, StrideCheck) =
2742 addStrideCheck(VectorPH->getTerminator());
2744 AddedSafetyChecks = true;
2745 // Create a new block containing the stride check.
2746 VectorPH->setName("vector.stridecheck");
2748 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2750 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2751 LoopBypassBlocks.push_back(VectorPH);
2753 // Replace the branch into the memory check block with a conditional branch
2754 // for the "few elements case".
2755 ReplaceInstWithInst(
2756 VectorPH->getTerminator(),
2757 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2759 VectorPH = NewVectorPH;
2762 // Generate the code that checks in runtime if arrays overlap. We put the
2763 // checks into a separate block to make the more common case of few elements
2765 Instruction *MemRuntimeCheck;
2766 std::tie(FirstCheckInst, MemRuntimeCheck) =
2767 Legal->getLAI()->addRuntimeChecks(VectorPH->getTerminator());
2768 if (MemRuntimeCheck) {
2769 AddedSafetyChecks = true;
2770 // Create a new block containing the memory check.
2771 VectorPH->setName("vector.memcheck");
2773 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2775 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2776 LoopBypassBlocks.push_back(VectorPH);
2778 // Replace the branch into the memory check block with a conditional branch
2779 // for the "few elements case".
2780 ReplaceInstWithInst(
2781 VectorPH->getTerminator(),
2782 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2784 VectorPH = NewVectorPH;
2787 // We are going to resume the execution of the scalar loop.
2788 // Go over all of the induction variables that we found and fix the
2789 // PHIs that are left in the scalar version of the loop.
2790 // The starting values of PHI nodes depend on the counter of the last
2791 // iteration in the vectorized loop.
2792 // If we come from a bypass edge then we need to start from the original
2795 // This variable saves the new starting index for the scalar loop.
2796 PHINode *ResumeIndex = nullptr;
2797 LoopVectorizationLegality::InductionList::iterator I, E;
2798 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2799 // Set builder to point to last bypass block.
2800 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2801 for (I = List->begin(), E = List->end(); I != E; ++I) {
2802 PHINode *OrigPhi = I->first;
2803 InductionDescriptor II = I->second;
2805 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2806 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2807 MiddleBlock->getTerminator());
2808 // We might have extended the type of the induction variable but we need a
2809 // truncated version for the scalar loop.
2810 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2811 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2812 MiddleBlock->getTerminator()) : nullptr;
2814 // Create phi nodes to merge from the backedge-taken check block.
2815 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2816 ScalarPH->getTerminator());
2817 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2819 PHINode *BCTruncResumeVal = nullptr;
2820 if (OrigPhi == OldInduction) {
2822 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2823 ScalarPH->getTerminator());
2824 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2827 Value *EndValue = nullptr;
2828 switch (II.getKind()) {
2829 case InductionDescriptor::IK_NoInduction:
2830 llvm_unreachable("Unknown induction");
2831 case InductionDescriptor::IK_IntInduction: {
2832 // Handle the integer induction counter.
2833 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2835 // We have the canonical induction variable.
2836 if (OrigPhi == OldInduction) {
2837 // Create a truncated version of the resume value for the scalar loop,
2838 // we might have promoted the type to a larger width.
2840 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2841 // The new PHI merges the original incoming value, in case of a bypass,
2842 // or the value at the end of the vectorized loop.
2843 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2844 TruncResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2845 TruncResumeVal->addIncoming(EndValue, VecBody);
2847 BCTruncResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[0]);
2849 // We know what the end value is.
2850 EndValue = IdxEndRoundDown;
2851 // We also know which PHI node holds it.
2852 ResumeIndex = ResumeVal;
2856 // Not the canonical induction variable - add the vector loop count to the
2858 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2859 II.getStartValue()->getType(),
2861 EndValue = II.transform(BypassBuilder, CRD);
2862 EndValue->setName("ind.end");
2865 case InductionDescriptor::IK_PtrInduction: {
2866 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2867 II.getStepValue()->getType(),
2869 EndValue = II.transform(BypassBuilder, CRD);
2870 EndValue->setName("ptr.ind.end");
2875 // The new PHI merges the original incoming value, in case of a bypass,
2876 // or the value at the end of the vectorized loop.
2877 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2878 if (OrigPhi == OldInduction)
2879 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2881 ResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2883 ResumeVal->addIncoming(EndValue, VecBody);
2885 // Fix the scalar body counter (PHI node).
2886 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2888 // The old induction's phi node in the scalar body needs the truncated
2890 if (OrigPhi == OldInduction) {
2891 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2892 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2894 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[0]);
2895 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2899 // If we are generating a new induction variable then we also need to
2900 // generate the code that calculates the exit value. This value is not
2901 // simply the end of the counter because we may skip the vectorized body
2902 // in case of a runtime check.
2904 assert(!ResumeIndex && "Unexpected resume value found");
2905 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2906 MiddleBlock->getTerminator());
2907 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2908 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2909 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2912 // Make sure that we found the index where scalar loop needs to continue.
2913 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2914 "Invalid resume Index");
2916 // Add a check in the middle block to see if we have completed
2917 // all of the iterations in the first vector loop.
2918 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2919 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2920 ResumeIndex, "cmp.n",
2921 MiddleBlock->getTerminator());
2922 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2923 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2925 // Create i+1 and fill the PHINode.
2926 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2927 Induction->addIncoming(StartIdx, VectorPH);
2928 Induction->addIncoming(NextIdx, VecBody);
2929 // Create the compare.
2930 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2931 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2933 // Now we have two terminators. Remove the old one from the block.
2934 VecBody->getTerminator()->eraseFromParent();
2936 // Get ready to start creating new instructions into the vectorized body.
2937 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2940 LoopVectorPreHeader = VectorPH;
2941 LoopScalarPreHeader = ScalarPH;
2942 LoopMiddleBlock = MiddleBlock;
2943 LoopExitBlock = ExitBlock;
2944 LoopVectorBody.push_back(VecBody);
2945 LoopScalarBody = OldBasicBlock;
2947 LoopVectorizeHints Hints(Lp, true);
2948 Hints.setAlreadyVectorized();
2952 struct CSEDenseMapInfo {
2953 static bool canHandle(Instruction *I) {
2954 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2955 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2957 static inline Instruction *getEmptyKey() {
2958 return DenseMapInfo<Instruction *>::getEmptyKey();
2960 static inline Instruction *getTombstoneKey() {
2961 return DenseMapInfo<Instruction *>::getTombstoneKey();
2963 static unsigned getHashValue(Instruction *I) {
2964 assert(canHandle(I) && "Unknown instruction!");
2965 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2966 I->value_op_end()));
2968 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2969 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2970 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2972 return LHS->isIdenticalTo(RHS);
2977 /// \brief Check whether this block is a predicated block.
2978 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2979 /// = ...; " blocks. We start with one vectorized basic block. For every
2980 /// conditional block we split this vectorized block. Therefore, every second
2981 /// block will be a predicated one.
2982 static bool isPredicatedBlock(unsigned BlockNum) {
2983 return BlockNum % 2;
2986 ///\brief Perform cse of induction variable instructions.
2987 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2988 // Perform simple cse.
2989 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2990 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2991 BasicBlock *BB = BBs[i];
2992 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2993 Instruction *In = I++;
2995 if (!CSEDenseMapInfo::canHandle(In))
2998 // Check if we can replace this instruction with any of the
2999 // visited instructions.
3000 if (Instruction *V = CSEMap.lookup(In)) {
3001 In->replaceAllUsesWith(V);
3002 In->eraseFromParent();
3005 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3006 // ...;" blocks for predicated stores. Every second block is a predicated
3008 if (isPredicatedBlock(i))
3016 /// \brief Adds a 'fast' flag to floating point operations.
3017 static Value *addFastMathFlag(Value *V) {
3018 if (isa<FPMathOperator>(V)){
3019 FastMathFlags Flags;
3020 Flags.setUnsafeAlgebra();
3021 cast<Instruction>(V)->setFastMathFlags(Flags);
3026 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3027 /// the result needs to be inserted and/or extracted from vectors.
3028 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3029 const TargetTransformInfo &TTI) {
3033 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3036 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3038 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3040 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3046 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3047 // Return the cost of the instruction, including scalarization overhead if it's
3048 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3049 // i.e. either vector version isn't available, or is too expensive.
3050 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3051 const TargetTransformInfo &TTI,
3052 const TargetLibraryInfo *TLI,
3053 bool &NeedToScalarize) {
3054 Function *F = CI->getCalledFunction();
3055 StringRef FnName = CI->getCalledFunction()->getName();
3056 Type *ScalarRetTy = CI->getType();
3057 SmallVector<Type *, 4> Tys, ScalarTys;
3058 for (auto &ArgOp : CI->arg_operands())
3059 ScalarTys.push_back(ArgOp->getType());
3061 // Estimate cost of scalarized vector call. The source operands are assumed
3062 // to be vectors, so we need to extract individual elements from there,
3063 // execute VF scalar calls, and then gather the result into the vector return
3065 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3067 return ScalarCallCost;
3069 // Compute corresponding vector type for return value and arguments.
3070 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3071 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3072 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3074 // Compute costs of unpacking argument values for the scalar calls and
3075 // packing the return values to a vector.
3076 unsigned ScalarizationCost =
3077 getScalarizationOverhead(RetTy, true, false, TTI);
3078 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3079 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3081 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3083 // If we can't emit a vector call for this function, then the currently found
3084 // cost is the cost we need to return.
3085 NeedToScalarize = true;
3086 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3089 // If the corresponding vector cost is cheaper, return its cost.
3090 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3091 if (VectorCallCost < Cost) {
3092 NeedToScalarize = false;
3093 return VectorCallCost;
3098 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3099 // factor VF. Return the cost of the instruction, including scalarization
3100 // overhead if it's needed.
3101 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3102 const TargetTransformInfo &TTI,
3103 const TargetLibraryInfo *TLI) {
3104 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3105 assert(ID && "Expected intrinsic call!");
3107 Type *RetTy = ToVectorTy(CI->getType(), VF);
3108 SmallVector<Type *, 4> Tys;
3109 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3110 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3112 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3115 void InnerLoopVectorizer::vectorizeLoop() {
3116 //===------------------------------------------------===//
3118 // Notice: any optimization or new instruction that go
3119 // into the code below should be also be implemented in
3122 //===------------------------------------------------===//
3123 Constant *Zero = Builder.getInt32(0);
3125 // In order to support reduction variables we need to be able to vectorize
3126 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3127 // stages. First, we create a new vector PHI node with no incoming edges.
3128 // We use this value when we vectorize all of the instructions that use the
3129 // PHI. Next, after all of the instructions in the block are complete we
3130 // add the new incoming edges to the PHI. At this point all of the
3131 // instructions in the basic block are vectorized, so we can use them to
3132 // construct the PHI.
3133 PhiVector RdxPHIsToFix;
3135 // Scan the loop in a topological order to ensure that defs are vectorized
3137 LoopBlocksDFS DFS(OrigLoop);
3140 // Vectorize all of the blocks in the original loop.
3141 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3142 be = DFS.endRPO(); bb != be; ++bb)
3143 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3145 // At this point every instruction in the original loop is widened to
3146 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3147 // that we vectorized. The PHI nodes are currently empty because we did
3148 // not want to introduce cycles. Notice that the remaining PHI nodes
3149 // that we need to fix are reduction variables.
3151 // Create the 'reduced' values for each of the induction vars.
3152 // The reduced values are the vector values that we scalarize and combine
3153 // after the loop is finished.
3154 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3156 PHINode *RdxPhi = *it;
3157 assert(RdxPhi && "Unable to recover vectorized PHI");
3159 // Find the reduction variable descriptor.
3160 assert(Legal->getReductionVars()->count(RdxPhi) &&
3161 "Unable to find the reduction variable");
3162 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3164 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3165 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3166 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3167 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3168 RdxDesc.getMinMaxRecurrenceKind();
3169 setDebugLocFromInst(Builder, ReductionStartValue);
3171 // We need to generate a reduction vector from the incoming scalar.
3172 // To do so, we need to generate the 'identity' vector and override
3173 // one of the elements with the incoming scalar reduction. We need
3174 // to do it in the vector-loop preheader.
3175 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3177 // This is the vector-clone of the value that leaves the loop.
3178 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3179 Type *VecTy = VectorExit[0]->getType();
3181 // Find the reduction identity variable. Zero for addition, or, xor,
3182 // one for multiplication, -1 for And.
3185 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3186 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3187 // MinMax reduction have the start value as their identify.
3189 VectorStart = Identity = ReductionStartValue;
3191 VectorStart = Identity =
3192 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3195 // Handle other reduction kinds:
3196 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3197 RK, VecTy->getScalarType());
3200 // This vector is the Identity vector where the first element is the
3201 // incoming scalar reduction.
3202 VectorStart = ReductionStartValue;
3204 Identity = ConstantVector::getSplat(VF, Iden);
3206 // This vector is the Identity vector where the first element is the
3207 // incoming scalar reduction.
3209 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3213 // Fix the vector-loop phi.
3215 // Reductions do not have to start at zero. They can start with
3216 // any loop invariant values.
3217 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3218 BasicBlock *Latch = OrigLoop->getLoopLatch();
3219 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3220 VectorParts &Val = getVectorValue(LoopVal);
3221 for (unsigned part = 0; part < UF; ++part) {
3222 // Make sure to add the reduction stat value only to the
3223 // first unroll part.
3224 Value *StartVal = (part == 0) ? VectorStart : Identity;
3225 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3226 LoopVectorPreHeader);
3227 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3228 LoopVectorBody.back());
3231 // Before each round, move the insertion point right between
3232 // the PHIs and the values we are going to write.
3233 // This allows us to write both PHINodes and the extractelement
3235 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3237 VectorParts RdxParts;
3238 setDebugLocFromInst(Builder, LoopExitInst);
3239 for (unsigned part = 0; part < UF; ++part) {
3240 // This PHINode contains the vectorized reduction variable, or
3241 // the initial value vector, if we bypass the vector loop.
3242 VectorParts &RdxExitVal = getVectorValue(LoopExitInst);
3243 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3244 Value *StartVal = (part == 0) ? VectorStart : Identity;
3245 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3246 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3247 NewPhi->addIncoming(RdxExitVal[part],
3248 LoopVectorBody.back());
3249 RdxParts.push_back(NewPhi);
3252 // Reduce all of the unrolled parts into a single vector.
3253 Value *ReducedPartRdx = RdxParts[0];
3254 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3255 setDebugLocFromInst(Builder, ReducedPartRdx);
3256 for (unsigned part = 1; part < UF; ++part) {
3257 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3258 // Floating point operations had to be 'fast' to enable the reduction.
3259 ReducedPartRdx = addFastMathFlag(
3260 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3261 ReducedPartRdx, "bin.rdx"));
3263 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3264 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3268 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3269 // and vector ops, reducing the set of values being computed by half each
3271 assert(isPowerOf2_32(VF) &&
3272 "Reduction emission only supported for pow2 vectors!");
3273 Value *TmpVec = ReducedPartRdx;
3274 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3275 for (unsigned i = VF; i != 1; i >>= 1) {
3276 // Move the upper half of the vector to the lower half.
3277 for (unsigned j = 0; j != i/2; ++j)
3278 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3280 // Fill the rest of the mask with undef.
3281 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3282 UndefValue::get(Builder.getInt32Ty()));
3285 Builder.CreateShuffleVector(TmpVec,
3286 UndefValue::get(TmpVec->getType()),
3287 ConstantVector::get(ShuffleMask),
3290 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3291 // Floating point operations had to be 'fast' to enable the reduction.
3292 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3293 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3295 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3299 // The result is in the first element of the vector.
3300 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3301 Builder.getInt32(0));
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 Type *PhiTy = P->getType();
3490 if (P == OldInduction) {
3491 // Handle the canonical induction variable. We might have had to
3493 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3495 // Handle other induction variables that are now based on the
3497 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3499 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3500 Broadcasted = II.transform(Builder, NormalizedIdx);
3501 Broadcasted->setName("offset.idx");
3503 Broadcasted = getBroadcastInstrs(Broadcasted);
3504 // After broadcasting the induction variable we need to make the vector
3505 // consecutive by adding 0, 1, 2, etc.
3506 for (unsigned part = 0; part < UF; ++part)
3507 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3510 case InductionDescriptor::IK_PtrInduction:
3511 // Handle the pointer induction variable case.
3512 assert(P->getType()->isPointerTy() && "Unexpected type.");
3513 // This is the normalized GEP that starts counting at zero.
3514 Value *NormalizedIdx =
3515 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3517 Builder.CreateSExtOrTrunc(NormalizedIdx, II.getStepValue()->getType());
3518 // This is the vector of results. Notice that we don't generate
3519 // vector geps because scalar geps result in better code.
3520 for (unsigned part = 0; part < UF; ++part) {
3522 int EltIndex = part;
3523 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3524 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3525 Value *SclrGep = II.transform(Builder, GlobalIdx);
3526 SclrGep->setName("next.gep");
3527 Entry[part] = SclrGep;
3531 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3532 for (unsigned int i = 0; i < VF; ++i) {
3533 int EltIndex = i + part * VF;
3534 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3535 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3536 Value *SclrGep = II.transform(Builder, GlobalIdx);
3537 SclrGep->setName("next.gep");
3538 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3539 Builder.getInt32(i),
3542 Entry[part] = VecVal;
3548 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3549 // For each instruction in the old loop.
3550 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3551 VectorParts &Entry = WidenMap.get(it);
3552 switch (it->getOpcode()) {
3553 case Instruction::Br:
3554 // Nothing to do for PHIs and BR, since we already took care of the
3555 // loop control flow instructions.
3557 case Instruction::PHI: {
3558 // Vectorize PHINodes.
3559 widenPHIInstruction(it, Entry, UF, VF, PV);
3563 case Instruction::Add:
3564 case Instruction::FAdd:
3565 case Instruction::Sub:
3566 case Instruction::FSub:
3567 case Instruction::Mul:
3568 case Instruction::FMul:
3569 case Instruction::UDiv:
3570 case Instruction::SDiv:
3571 case Instruction::FDiv:
3572 case Instruction::URem:
3573 case Instruction::SRem:
3574 case Instruction::FRem:
3575 case Instruction::Shl:
3576 case Instruction::LShr:
3577 case Instruction::AShr:
3578 case Instruction::And:
3579 case Instruction::Or:
3580 case Instruction::Xor: {
3581 // Just widen binops.
3582 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3583 setDebugLocFromInst(Builder, BinOp);
3584 VectorParts &A = getVectorValue(it->getOperand(0));
3585 VectorParts &B = getVectorValue(it->getOperand(1));
3587 // Use this vector value for all users of the original instruction.
3588 for (unsigned Part = 0; Part < UF; ++Part) {
3589 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3591 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3592 VecOp->copyIRFlags(BinOp);
3597 propagateMetadata(Entry, it);
3600 case Instruction::Select: {
3602 // If the selector is loop invariant we can create a select
3603 // instruction with a scalar condition. Otherwise, use vector-select.
3604 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3606 setDebugLocFromInst(Builder, it);
3608 // The condition can be loop invariant but still defined inside the
3609 // loop. This means that we can't just use the original 'cond' value.
3610 // We have to take the 'vectorized' value and pick the first lane.
3611 // Instcombine will make this a no-op.
3612 VectorParts &Cond = getVectorValue(it->getOperand(0));
3613 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3614 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3616 Value *ScalarCond = (VF == 1) ? Cond[0] :
3617 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3619 for (unsigned Part = 0; Part < UF; ++Part) {
3620 Entry[Part] = Builder.CreateSelect(
3621 InvariantCond ? ScalarCond : Cond[Part],
3626 propagateMetadata(Entry, it);
3630 case Instruction::ICmp:
3631 case Instruction::FCmp: {
3632 // Widen compares. Generate vector compares.
3633 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3634 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3635 setDebugLocFromInst(Builder, it);
3636 VectorParts &A = getVectorValue(it->getOperand(0));
3637 VectorParts &B = getVectorValue(it->getOperand(1));
3638 for (unsigned Part = 0; Part < UF; ++Part) {
3641 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3643 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3647 propagateMetadata(Entry, it);
3651 case Instruction::Store:
3652 case Instruction::Load:
3653 vectorizeMemoryInstruction(it);
3655 case Instruction::ZExt:
3656 case Instruction::SExt:
3657 case Instruction::FPToUI:
3658 case Instruction::FPToSI:
3659 case Instruction::FPExt:
3660 case Instruction::PtrToInt:
3661 case Instruction::IntToPtr:
3662 case Instruction::SIToFP:
3663 case Instruction::UIToFP:
3664 case Instruction::Trunc:
3665 case Instruction::FPTrunc:
3666 case Instruction::BitCast: {
3667 CastInst *CI = dyn_cast<CastInst>(it);
3668 setDebugLocFromInst(Builder, it);
3669 /// Optimize the special case where the source is the induction
3670 /// variable. Notice that we can only optimize the 'trunc' case
3671 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3672 /// c. other casts depend on pointer size.
3673 if (CI->getOperand(0) == OldInduction &&
3674 it->getOpcode() == Instruction::Trunc) {
3675 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3677 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3678 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3680 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3681 for (unsigned Part = 0; Part < UF; ++Part)
3682 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3683 propagateMetadata(Entry, it);
3686 /// Vectorize casts.
3687 Type *DestTy = (VF == 1) ? CI->getType() :
3688 VectorType::get(CI->getType(), VF);
3690 VectorParts &A = getVectorValue(it->getOperand(0));
3691 for (unsigned Part = 0; Part < UF; ++Part)
3692 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3693 propagateMetadata(Entry, it);
3697 case Instruction::Call: {
3698 // Ignore dbg intrinsics.
3699 if (isa<DbgInfoIntrinsic>(it))
3701 setDebugLocFromInst(Builder, it);
3703 Module *M = BB->getParent()->getParent();
3704 CallInst *CI = cast<CallInst>(it);
3706 StringRef FnName = CI->getCalledFunction()->getName();
3707 Function *F = CI->getCalledFunction();
3708 Type *RetTy = ToVectorTy(CI->getType(), VF);
3709 SmallVector<Type *, 4> Tys;
3710 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3711 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3713 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3715 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3716 ID == Intrinsic::lifetime_start)) {
3717 scalarizeInstruction(it);
3720 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3721 // version of the instruction.
3722 // Is it beneficial to perform intrinsic call compared to lib call?
3723 bool NeedToScalarize;
3724 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3725 bool UseVectorIntrinsic =
3726 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3727 if (!UseVectorIntrinsic && NeedToScalarize) {
3728 scalarizeInstruction(it);
3732 for (unsigned Part = 0; Part < UF; ++Part) {
3733 SmallVector<Value *, 4> Args;
3734 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3735 Value *Arg = CI->getArgOperand(i);
3736 // Some intrinsics have a scalar argument - don't replace it with a
3738 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3739 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3740 Arg = VectorArg[Part];
3742 Args.push_back(Arg);
3746 if (UseVectorIntrinsic) {
3747 // Use vector version of the intrinsic.
3748 Type *TysForDecl[] = {CI->getType()};
3750 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3751 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3753 // Use vector version of the library call.
3754 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3755 assert(!VFnName.empty() && "Vector function name is empty.");
3756 VectorF = M->getFunction(VFnName);
3758 // Generate a declaration
3759 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3761 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3762 VectorF->copyAttributesFrom(F);
3765 assert(VectorF && "Can't create vector function.");
3766 Entry[Part] = Builder.CreateCall(VectorF, Args);
3769 propagateMetadata(Entry, it);
3774 // All other instructions are unsupported. Scalarize them.
3775 scalarizeInstruction(it);
3778 }// end of for_each instr.
3781 void InnerLoopVectorizer::updateAnalysis() {
3782 // Forget the original basic block.
3783 SE->forgetLoop(OrigLoop);
3785 // Update the dominator tree information.
3786 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3787 "Entry does not dominate exit.");
3789 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3790 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3791 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3793 // Due to if predication of stores we might create a sequence of "if(pred)
3794 // a[i] = ...; " blocks.
3795 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3797 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3798 else if (isPredicatedBlock(i)) {
3799 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3801 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3805 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3806 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3807 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3808 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3810 DEBUG(DT->verifyDomTree());
3813 /// \brief Check whether it is safe to if-convert this phi node.
3815 /// Phi nodes with constant expressions that can trap are not safe to if
3817 static bool canIfConvertPHINodes(BasicBlock *BB) {
3818 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3819 PHINode *Phi = dyn_cast<PHINode>(I);
3822 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3823 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3830 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3831 if (!EnableIfConversion) {
3832 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3836 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3838 // A list of pointers that we can safely read and write to.
3839 SmallPtrSet<Value *, 8> SafePointes;
3841 // Collect safe addresses.
3842 for (Loop::block_iterator BI = TheLoop->block_begin(),
3843 BE = TheLoop->block_end(); BI != BE; ++BI) {
3844 BasicBlock *BB = *BI;
3846 if (blockNeedsPredication(BB))
3849 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3850 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3851 SafePointes.insert(LI->getPointerOperand());
3852 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3853 SafePointes.insert(SI->getPointerOperand());
3857 // Collect the blocks that need predication.
3858 BasicBlock *Header = TheLoop->getHeader();
3859 for (Loop::block_iterator BI = TheLoop->block_begin(),
3860 BE = TheLoop->block_end(); BI != BE; ++BI) {
3861 BasicBlock *BB = *BI;
3863 // We don't support switch statements inside loops.
3864 if (!isa<BranchInst>(BB->getTerminator())) {
3865 emitAnalysis(VectorizationReport(BB->getTerminator())
3866 << "loop contains a switch statement");
3870 // We must be able to predicate all blocks that need to be predicated.
3871 if (blockNeedsPredication(BB)) {
3872 if (!blockCanBePredicated(BB, SafePointes)) {
3873 emitAnalysis(VectorizationReport(BB->getTerminator())
3874 << "control flow cannot be substituted for a select");
3877 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3878 emitAnalysis(VectorizationReport(BB->getTerminator())
3879 << "control flow cannot be substituted for a select");
3884 // We can if-convert this loop.
3888 bool LoopVectorizationLegality::canVectorize() {
3889 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3890 // be canonicalized.
3891 if (!TheLoop->getLoopPreheader()) {
3893 VectorizationReport() <<
3894 "loop control flow is not understood by vectorizer");
3898 // We can only vectorize innermost loops.
3899 if (!TheLoop->empty()) {
3900 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3904 // We must have a single backedge.
3905 if (TheLoop->getNumBackEdges() != 1) {
3907 VectorizationReport() <<
3908 "loop control flow is not understood by vectorizer");
3912 // We must have a single exiting block.
3913 if (!TheLoop->getExitingBlock()) {
3915 VectorizationReport() <<
3916 "loop control flow is not understood by vectorizer");
3920 // We only handle bottom-tested loops, i.e. loop in which the condition is
3921 // checked at the end of each iteration. With that we can assume that all
3922 // instructions in the loop are executed the same number of times.
3923 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3925 VectorizationReport() <<
3926 "loop control flow is not understood by vectorizer");
3930 // We need to have a loop header.
3931 DEBUG(dbgs() << "LV: Found a loop: " <<
3932 TheLoop->getHeader()->getName() << '\n');
3934 // Check if we can if-convert non-single-bb loops.
3935 unsigned NumBlocks = TheLoop->getNumBlocks();
3936 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3937 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3941 // ScalarEvolution needs to be able to find the exit count.
3942 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3943 if (ExitCount == SE->getCouldNotCompute()) {
3944 emitAnalysis(VectorizationReport() <<
3945 "could not determine number of loop iterations");
3946 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3950 // Check if we can vectorize the instructions and CFG in this loop.
3951 if (!canVectorizeInstrs()) {
3952 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3956 // Go over each instruction and look at memory deps.
3957 if (!canVectorizeMemory()) {
3958 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3962 // Collect all of the variables that remain uniform after vectorization.
3963 collectLoopUniforms();
3965 DEBUG(dbgs() << "LV: We can vectorize this loop"
3966 << (LAI->getRuntimePointerChecking()->Need
3967 ? " (with a runtime bound check)"
3971 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
3973 // If an override option has been passed in for interleaved accesses, use it.
3974 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
3975 UseInterleaved = EnableInterleavedMemAccesses;
3977 // Analyze interleaved memory accesses.
3979 InterleaveInfo.analyzeInterleaving(Strides);
3981 // Okay! We can vectorize. At this point we don't have any other mem analysis
3982 // which may limit our maximum vectorization factor, so just return true with
3987 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3988 if (Ty->isPointerTy())
3989 return DL.getIntPtrType(Ty);
3991 // It is possible that char's or short's overflow when we ask for the loop's
3992 // trip count, work around this by changing the type size.
3993 if (Ty->getScalarSizeInBits() < 32)
3994 return Type::getInt32Ty(Ty->getContext());
3999 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4000 Ty0 = convertPointerToIntegerType(DL, Ty0);
4001 Ty1 = convertPointerToIntegerType(DL, Ty1);
4002 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4007 /// \brief Check that the instruction has outside loop users and is not an
4008 /// identified reduction variable.
4009 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4010 SmallPtrSetImpl<Value *> &Reductions) {
4011 // Reduction instructions are allowed to have exit users. All other
4012 // instructions must not have external users.
4013 if (!Reductions.count(Inst))
4014 //Check that all of the users of the loop are inside the BB.
4015 for (User *U : Inst->users()) {
4016 Instruction *UI = cast<Instruction>(U);
4017 // This user may be a reduction exit value.
4018 if (!TheLoop->contains(UI)) {
4019 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4026 bool LoopVectorizationLegality::canVectorizeInstrs() {
4027 BasicBlock *Header = TheLoop->getHeader();
4029 // Look for the attribute signaling the absence of NaNs.
4030 Function &F = *Header->getParent();
4031 const DataLayout &DL = F.getParent()->getDataLayout();
4032 if (F.hasFnAttribute("no-nans-fp-math"))
4034 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4036 // For each block in the loop.
4037 for (Loop::block_iterator bb = TheLoop->block_begin(),
4038 be = TheLoop->block_end(); bb != be; ++bb) {
4040 // Scan the instructions in the block and look for hazards.
4041 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4044 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4045 Type *PhiTy = Phi->getType();
4046 // Check that this PHI type is allowed.
4047 if (!PhiTy->isIntegerTy() &&
4048 !PhiTy->isFloatingPointTy() &&
4049 !PhiTy->isPointerTy()) {
4050 emitAnalysis(VectorizationReport(it)
4051 << "loop control flow is not understood by vectorizer");
4052 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4056 // If this PHINode is not in the header block, then we know that we
4057 // can convert it to select during if-conversion. No need to check if
4058 // the PHIs in this block are induction or reduction variables.
4059 if (*bb != Header) {
4060 // Check that this instruction has no outside users or is an
4061 // identified reduction value with an outside user.
4062 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4064 emitAnalysis(VectorizationReport(it) <<
4065 "value could not be identified as "
4066 "an induction or reduction variable");
4070 // We only allow if-converted PHIs with exactly two incoming values.
4071 if (Phi->getNumIncomingValues() != 2) {
4072 emitAnalysis(VectorizationReport(it)
4073 << "control flow not understood by vectorizer");
4074 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4078 InductionDescriptor ID;
4079 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4080 Inductions[Phi] = ID;
4081 // Get the widest type.
4083 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4085 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4087 // Int inductions are special because we only allow one IV.
4088 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4089 ID.getStepValue()->isOne()) {
4090 // Use the phi node with the widest type as induction. Use the last
4091 // one if there are multiple (no good reason for doing this other
4092 // than it is expedient).
4093 if (!Induction || PhiTy == WidestIndTy)
4097 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4099 // Until we explicitly handle the case of an induction variable with
4100 // an outside loop user we have to give up vectorizing this loop.
4101 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4102 emitAnalysis(VectorizationReport(it) <<
4103 "use of induction value outside of the "
4104 "loop is not handled by vectorizer");
4111 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4113 if (Reductions[Phi].hasUnsafeAlgebra())
4114 Requirements->addUnsafeAlgebraInst(
4115 Reductions[Phi].getUnsafeAlgebraInst());
4116 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4120 emitAnalysis(VectorizationReport(it) <<
4121 "value that could not be identified as "
4122 "reduction is used outside the loop");
4123 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4125 }// end of PHI handling
4127 // We handle calls that:
4128 // * Are debug info intrinsics.
4129 // * Have a mapping to an IR intrinsic.
4130 // * Have a vector version available.
4131 CallInst *CI = dyn_cast<CallInst>(it);
4132 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4133 !(CI->getCalledFunction() && TLI &&
4134 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4135 emitAnalysis(VectorizationReport(it) <<
4136 "call instruction cannot be vectorized");
4137 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4141 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4142 // second argument is the same (i.e. loop invariant)
4144 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4145 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4146 emitAnalysis(VectorizationReport(it)
4147 << "intrinsic instruction cannot be vectorized");
4148 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4153 // Check that the instruction return type is vectorizable.
4154 // Also, we can't vectorize extractelement instructions.
4155 if ((!VectorType::isValidElementType(it->getType()) &&
4156 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4157 emitAnalysis(VectorizationReport(it)
4158 << "instruction return type cannot be vectorized");
4159 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4163 // Check that the stored type is vectorizable.
4164 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4165 Type *T = ST->getValueOperand()->getType();
4166 if (!VectorType::isValidElementType(T)) {
4167 emitAnalysis(VectorizationReport(ST) <<
4168 "store instruction cannot be vectorized");
4171 if (EnableMemAccessVersioning)
4172 collectStridedAccess(ST);
4175 if (EnableMemAccessVersioning)
4176 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4177 collectStridedAccess(LI);
4179 // Reduction instructions are allowed to have exit users.
4180 // All other instructions must not have external users.
4181 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4182 emitAnalysis(VectorizationReport(it) <<
4183 "value cannot be used outside the loop");
4192 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4193 if (Inductions.empty()) {
4194 emitAnalysis(VectorizationReport()
4195 << "loop induction variable could not be identified");
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 // Ignore ephemeral values.
4656 if (EphValues.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.
4664 if (PHINode *PN = dyn_cast<PHINode>(it))
4665 if (!Legal->getReductionVars()->count(PN))
4668 // Examine the stored values.
4669 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4670 T = ST->getValueOperand()->getType();
4672 // Ignore loaded pointer types and stored pointer types that are not
4673 // consecutive. However, we do want to take consecutive stores/loads of
4674 // pointer vectors into account.
4675 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4678 MaxWidth = std::max(MaxWidth,
4679 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4686 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4688 unsigned LoopCost) {
4690 // -- The interleave heuristics --
4691 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4692 // There are many micro-architectural considerations that we can't predict
4693 // at this level. For example, frontend pressure (on decode or fetch) due to
4694 // code size, or the number and capabilities of the execution ports.
4696 // We use the following heuristics to select the interleave count:
4697 // 1. If the code has reductions, then we interleave to break the cross
4698 // iteration dependency.
4699 // 2. If the loop is really small, then we interleave to reduce the loop
4701 // 3. We don't interleave if we think that we will spill registers to memory
4702 // due to the increased register pressure.
4704 // When we optimize for size, we don't interleave.
4708 // We used the distance for the interleave count.
4709 if (Legal->getMaxSafeDepDistBytes() != -1U)
4712 // Do not interleave loops with a relatively small trip count.
4713 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4714 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4717 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4718 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4722 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4723 TargetNumRegisters = ForceTargetNumScalarRegs;
4725 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4726 TargetNumRegisters = ForceTargetNumVectorRegs;
4729 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4730 // We divide by these constants so assume that we have at least one
4731 // instruction that uses at least one register.
4732 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4733 R.NumInstructions = std::max(R.NumInstructions, 1U);
4735 // We calculate the interleave count using the following formula.
4736 // Subtract the number of loop invariants from the number of available
4737 // registers. These registers are used by all of the interleaved instances.
4738 // Next, divide the remaining registers by the number of registers that is
4739 // required by the loop, in order to estimate how many parallel instances
4740 // fit without causing spills. All of this is rounded down if necessary to be
4741 // a power of two. We want power of two interleave count to simplify any
4742 // addressing operations or alignment considerations.
4743 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4746 // Don't count the induction variable as interleaved.
4747 if (EnableIndVarRegisterHeur)
4748 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4749 std::max(1U, (R.MaxLocalUsers - 1)));
4751 // Clamp the interleave ranges to reasonable counts.
4752 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4754 // Check if the user has overridden the max.
4756 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4757 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4759 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4760 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4763 // If we did not calculate the cost for VF (because the user selected the VF)
4764 // then we calculate the cost of VF here.
4766 LoopCost = expectedCost(VF);
4768 // Clamp the calculated IC to be between the 1 and the max interleave count
4769 // that the target allows.
4770 if (IC > MaxInterleaveCount)
4771 IC = MaxInterleaveCount;
4775 // Interleave if we vectorized this loop and there is a reduction that could
4776 // benefit from interleaving.
4777 if (VF > 1 && Legal->getReductionVars()->size()) {
4778 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4782 // Note that if we've already vectorized the loop we will have done the
4783 // runtime check and so interleaving won't require further checks.
4784 bool InterleavingRequiresRuntimePointerCheck =
4785 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4787 // We want to interleave small loops in order to reduce the loop overhead and
4788 // potentially expose ILP opportunities.
4789 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4790 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4791 // We assume that the cost overhead is 1 and we use the cost model
4792 // to estimate the cost of the loop and interleave until the cost of the
4793 // loop overhead is about 5% of the cost of the loop.
4795 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4797 // Interleave until store/load ports (estimated by max interleave count) are
4799 unsigned NumStores = Legal->getNumStores();
4800 unsigned NumLoads = Legal->getNumLoads();
4801 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4802 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4804 // If we have a scalar reduction (vector reductions are already dealt with
4805 // by this point), we can increase the critical path length if the loop
4806 // we're interleaving is inside another loop. Limit, by default to 2, so the
4807 // critical path only gets increased by one reduction operation.
4808 if (Legal->getReductionVars()->size() &&
4809 TheLoop->getLoopDepth() > 1) {
4810 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4811 SmallIC = std::min(SmallIC, F);
4812 StoresIC = std::min(StoresIC, F);
4813 LoadsIC = std::min(LoadsIC, F);
4816 if (EnableLoadStoreRuntimeInterleave &&
4817 std::max(StoresIC, LoadsIC) > SmallIC) {
4818 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4819 return std::max(StoresIC, LoadsIC);
4822 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4826 // Interleave if this is a large loop (small loops are already dealt with by
4828 // point) that could benefit from interleaving.
4829 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4830 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4831 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4835 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4839 LoopVectorizationCostModel::RegisterUsage
4840 LoopVectorizationCostModel::calculateRegisterUsage() {
4841 // This function calculates the register usage by measuring the highest number
4842 // of values that are alive at a single location. Obviously, this is a very
4843 // rough estimation. We scan the loop in a topological order in order and
4844 // assign a number to each instruction. We use RPO to ensure that defs are
4845 // met before their users. We assume that each instruction that has in-loop
4846 // users starts an interval. We record every time that an in-loop value is
4847 // used, so we have a list of the first and last occurrences of each
4848 // instruction. Next, we transpose this data structure into a multi map that
4849 // holds the list of intervals that *end* at a specific location. This multi
4850 // map allows us to perform a linear search. We scan the instructions linearly
4851 // and record each time that a new interval starts, by placing it in a set.
4852 // If we find this value in the multi-map then we remove it from the set.
4853 // The max register usage is the maximum size of the set.
4854 // We also search for instructions that are defined outside the loop, but are
4855 // used inside the loop. We need this number separately from the max-interval
4856 // usage number because when we unroll, loop-invariant values do not take
4858 LoopBlocksDFS DFS(TheLoop);
4862 R.NumInstructions = 0;
4864 // Each 'key' in the map opens a new interval. The values
4865 // of the map are the index of the 'last seen' usage of the
4866 // instruction that is the key.
4867 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4868 // Maps instruction to its index.
4869 DenseMap<unsigned, Instruction*> IdxToInstr;
4870 // Marks the end of each interval.
4871 IntervalMap EndPoint;
4872 // Saves the list of instruction indices that are used in the loop.
4873 SmallSet<Instruction*, 8> Ends;
4874 // Saves the list of values that are used in the loop but are
4875 // defined outside the loop, such as arguments and constants.
4876 SmallPtrSet<Value*, 8> LoopInvariants;
4879 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4880 be = DFS.endRPO(); bb != be; ++bb) {
4881 R.NumInstructions += (*bb)->size();
4882 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4884 Instruction *I = it;
4885 IdxToInstr[Index++] = I;
4887 // Save the end location of each USE.
4888 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4889 Value *U = I->getOperand(i);
4890 Instruction *Instr = dyn_cast<Instruction>(U);
4892 // Ignore non-instruction values such as arguments, constants, etc.
4893 if (!Instr) continue;
4895 // If this instruction is outside the loop then record it and continue.
4896 if (!TheLoop->contains(Instr)) {
4897 LoopInvariants.insert(Instr);
4901 // Overwrite previous end points.
4902 EndPoint[Instr] = Index;
4908 // Saves the list of intervals that end with the index in 'key'.
4909 typedef SmallVector<Instruction*, 2> InstrList;
4910 DenseMap<unsigned, InstrList> TransposeEnds;
4912 // Transpose the EndPoints to a list of values that end at each index.
4913 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4915 TransposeEnds[it->second].push_back(it->first);
4917 SmallSet<Instruction*, 8> OpenIntervals;
4918 unsigned MaxUsage = 0;
4921 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4922 for (unsigned int i = 0; i < Index; ++i) {
4923 Instruction *I = IdxToInstr[i];
4924 // Ignore instructions that are never used within the loop.
4925 if (!Ends.count(I)) continue;
4927 // Ignore ephemeral values.
4928 if (EphValues.count(I))
4931 // Remove all of the instructions that end at this location.
4932 InstrList &List = TransposeEnds[i];
4933 for (unsigned int j=0, e = List.size(); j < e; ++j)
4934 OpenIntervals.erase(List[j]);
4936 // Count the number of live interals.
4937 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4939 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4940 OpenIntervals.size() << '\n');
4942 // Add the current instruction to the list of open intervals.
4943 OpenIntervals.insert(I);
4946 unsigned Invariant = LoopInvariants.size();
4947 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4948 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4949 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4951 R.LoopInvariantRegs = Invariant;
4952 R.MaxLocalUsers = MaxUsage;
4956 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4960 for (Loop::block_iterator bb = TheLoop->block_begin(),
4961 be = TheLoop->block_end(); bb != be; ++bb) {
4962 unsigned BlockCost = 0;
4963 BasicBlock *BB = *bb;
4965 // For each instruction in the old loop.
4966 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4967 // Skip dbg intrinsics.
4968 if (isa<DbgInfoIntrinsic>(it))
4971 // Ignore ephemeral values.
4972 if (EphValues.count(it))
4975 unsigned C = getInstructionCost(it, VF);
4977 // Check if we should override the cost.
4978 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4979 C = ForceTargetInstructionCost;
4982 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4983 VF << " For instruction: " << *it << '\n');
4986 // We assume that if-converted blocks have a 50% chance of being executed.
4987 // When the code is scalar then some of the blocks are avoided due to CF.
4988 // When the code is vectorized we execute all code paths.
4989 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4998 /// \brief Check whether the address computation for a non-consecutive memory
4999 /// access looks like an unlikely candidate for being merged into the indexing
5002 /// We look for a GEP which has one index that is an induction variable and all
5003 /// other indices are loop invariant. If the stride of this access is also
5004 /// within a small bound we decide that this address computation can likely be
5005 /// merged into the addressing mode.
5006 /// In all other cases, we identify the address computation as complex.
5007 static bool isLikelyComplexAddressComputation(Value *Ptr,
5008 LoopVectorizationLegality *Legal,
5009 ScalarEvolution *SE,
5010 const Loop *TheLoop) {
5011 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5015 // We are looking for a gep with all loop invariant indices except for one
5016 // which should be an induction variable.
5017 unsigned NumOperands = Gep->getNumOperands();
5018 for (unsigned i = 1; i < NumOperands; ++i) {
5019 Value *Opd = Gep->getOperand(i);
5020 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5021 !Legal->isInductionVariable(Opd))
5025 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5026 // can likely be merged into the address computation.
5027 unsigned MaxMergeDistance = 64;
5029 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5033 // Check the step is constant.
5034 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5035 // Calculate the pointer stride and check if it is consecutive.
5036 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5040 const APInt &APStepVal = C->getValue()->getValue();
5042 // Huge step value - give up.
5043 if (APStepVal.getBitWidth() > 64)
5046 int64_t StepVal = APStepVal.getSExtValue();
5048 return StepVal > MaxMergeDistance;
5051 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5052 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5058 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5059 // If we know that this instruction will remain uniform, check the cost of
5060 // the scalar version.
5061 if (Legal->isUniformAfterVectorization(I))
5064 Type *RetTy = I->getType();
5065 Type *VectorTy = ToVectorTy(RetTy, VF);
5067 // TODO: We need to estimate the cost of intrinsic calls.
5068 switch (I->getOpcode()) {
5069 case Instruction::GetElementPtr:
5070 // We mark this instruction as zero-cost because the cost of GEPs in
5071 // vectorized code depends on whether the corresponding memory instruction
5072 // is scalarized or not. Therefore, we handle GEPs with the memory
5073 // instruction cost.
5075 case Instruction::Br: {
5076 return TTI.getCFInstrCost(I->getOpcode());
5078 case Instruction::PHI:
5079 //TODO: IF-converted IFs become selects.
5081 case Instruction::Add:
5082 case Instruction::FAdd:
5083 case Instruction::Sub:
5084 case Instruction::FSub:
5085 case Instruction::Mul:
5086 case Instruction::FMul:
5087 case Instruction::UDiv:
5088 case Instruction::SDiv:
5089 case Instruction::FDiv:
5090 case Instruction::URem:
5091 case Instruction::SRem:
5092 case Instruction::FRem:
5093 case Instruction::Shl:
5094 case Instruction::LShr:
5095 case Instruction::AShr:
5096 case Instruction::And:
5097 case Instruction::Or:
5098 case Instruction::Xor: {
5099 // Since we will replace the stride by 1 the multiplication should go away.
5100 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5102 // Certain instructions can be cheaper to vectorize if they have a constant
5103 // second vector operand. One example of this are shifts on x86.
5104 TargetTransformInfo::OperandValueKind Op1VK =
5105 TargetTransformInfo::OK_AnyValue;
5106 TargetTransformInfo::OperandValueKind Op2VK =
5107 TargetTransformInfo::OK_AnyValue;
5108 TargetTransformInfo::OperandValueProperties Op1VP =
5109 TargetTransformInfo::OP_None;
5110 TargetTransformInfo::OperandValueProperties Op2VP =
5111 TargetTransformInfo::OP_None;
5112 Value *Op2 = I->getOperand(1);
5114 // Check for a splat of a constant or for a non uniform vector of constants.
5115 if (isa<ConstantInt>(Op2)) {
5116 ConstantInt *CInt = cast<ConstantInt>(Op2);
5117 if (CInt && CInt->getValue().isPowerOf2())
5118 Op2VP = TargetTransformInfo::OP_PowerOf2;
5119 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5120 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5121 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5122 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5124 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5125 if (CInt && CInt->getValue().isPowerOf2())
5126 Op2VP = TargetTransformInfo::OP_PowerOf2;
5127 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5131 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5134 case Instruction::Select: {
5135 SelectInst *SI = cast<SelectInst>(I);
5136 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5137 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5138 Type *CondTy = SI->getCondition()->getType();
5140 CondTy = VectorType::get(CondTy, VF);
5142 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5144 case Instruction::ICmp:
5145 case Instruction::FCmp: {
5146 Type *ValTy = I->getOperand(0)->getType();
5147 VectorTy = ToVectorTy(ValTy, VF);
5148 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5150 case Instruction::Store:
5151 case Instruction::Load: {
5152 StoreInst *SI = dyn_cast<StoreInst>(I);
5153 LoadInst *LI = dyn_cast<LoadInst>(I);
5154 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5156 VectorTy = ToVectorTy(ValTy, VF);
5158 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5159 unsigned AS = SI ? SI->getPointerAddressSpace() :
5160 LI->getPointerAddressSpace();
5161 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5162 // We add the cost of address computation here instead of with the gep
5163 // instruction because only here we know whether the operation is
5166 return TTI.getAddressComputationCost(VectorTy) +
5167 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5169 // For an interleaved access, calculate the total cost of the whole
5170 // interleave group.
5171 if (Legal->isAccessInterleaved(I)) {
5172 auto Group = Legal->getInterleavedAccessGroup(I);
5173 assert(Group && "Fail to get an interleaved access group.");
5175 // Only calculate the cost once at the insert position.
5176 if (Group->getInsertPos() != I)
5179 unsigned InterleaveFactor = Group->getFactor();
5181 VectorType::get(VectorTy->getVectorElementType(),
5182 VectorTy->getVectorNumElements() * InterleaveFactor);
5184 // Holds the indices of existing members in an interleaved load group.
5185 // An interleaved store group doesn't need this as it dones't allow gaps.
5186 SmallVector<unsigned, 4> Indices;
5188 for (unsigned i = 0; i < InterleaveFactor; i++)
5189 if (Group->getMember(i))
5190 Indices.push_back(i);
5193 // Calculate the cost of the whole interleaved group.
5194 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5195 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5196 Group->getAlignment(), AS);
5198 if (Group->isReverse())
5200 Group->getNumMembers() *
5201 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5203 // FIXME: The interleaved load group with a huge gap could be even more
5204 // expensive than scalar operations. Then we could ignore such group and
5205 // use scalar operations instead.
5209 // Scalarized loads/stores.
5210 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5211 bool Reverse = ConsecutiveStride < 0;
5212 const DataLayout &DL = I->getModule()->getDataLayout();
5213 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5214 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5215 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5216 bool IsComplexComputation =
5217 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5219 // The cost of extracting from the value vector and pointer vector.
5220 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5221 for (unsigned i = 0; i < VF; ++i) {
5222 // The cost of extracting the pointer operand.
5223 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5224 // In case of STORE, the cost of ExtractElement from the vector.
5225 // In case of LOAD, the cost of InsertElement into the returned
5227 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5228 Instruction::InsertElement,
5232 // The cost of the scalar loads/stores.
5233 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5234 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5239 // Wide load/stores.
5240 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5241 if (Legal->isMaskRequired(I))
5242 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5245 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5248 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5252 case Instruction::ZExt:
5253 case Instruction::SExt:
5254 case Instruction::FPToUI:
5255 case Instruction::FPToSI:
5256 case Instruction::FPExt:
5257 case Instruction::PtrToInt:
5258 case Instruction::IntToPtr:
5259 case Instruction::SIToFP:
5260 case Instruction::UIToFP:
5261 case Instruction::Trunc:
5262 case Instruction::FPTrunc:
5263 case Instruction::BitCast: {
5264 // We optimize the truncation of induction variable.
5265 // The cost of these is the same as the scalar operation.
5266 if (I->getOpcode() == Instruction::Trunc &&
5267 Legal->isInductionVariable(I->getOperand(0)))
5268 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5269 I->getOperand(0)->getType());
5271 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5272 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5274 case Instruction::Call: {
5275 bool NeedToScalarize;
5276 CallInst *CI = cast<CallInst>(I);
5277 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5278 if (getIntrinsicIDForCall(CI, TLI))
5279 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5283 // We are scalarizing the instruction. Return the cost of the scalar
5284 // instruction, plus the cost of insert and extract into vector
5285 // elements, times the vector width.
5288 if (!RetTy->isVoidTy() && VF != 1) {
5289 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5291 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5294 // The cost of inserting the results plus extracting each one of the
5296 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5299 // The cost of executing VF copies of the scalar instruction. This opcode
5300 // is unknown. Assume that it is the same as 'mul'.
5301 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5307 char LoopVectorize::ID = 0;
5308 static const char lv_name[] = "Loop Vectorization";
5309 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5310 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5311 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5312 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5313 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5314 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5315 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5316 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5317 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5318 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5319 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5320 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5323 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5324 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5328 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5329 // Check for a store.
5330 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5331 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5333 // Check for a load.
5334 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5335 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5341 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5342 bool IfPredicateStore) {
5343 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5344 // Holds vector parameters or scalars, in case of uniform vals.
5345 SmallVector<VectorParts, 4> Params;
5347 setDebugLocFromInst(Builder, Instr);
5349 // Find all of the vectorized parameters.
5350 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5351 Value *SrcOp = Instr->getOperand(op);
5353 // If we are accessing the old induction variable, use the new one.
5354 if (SrcOp == OldInduction) {
5355 Params.push_back(getVectorValue(SrcOp));
5359 // Try using previously calculated values.
5360 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5362 // If the src is an instruction that appeared earlier in the basic block
5363 // then it should already be vectorized.
5364 if (SrcInst && OrigLoop->contains(SrcInst)) {
5365 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5366 // The parameter is a vector value from earlier.
5367 Params.push_back(WidenMap.get(SrcInst));
5369 // The parameter is a scalar from outside the loop. Maybe even a constant.
5370 VectorParts Scalars;
5371 Scalars.append(UF, SrcOp);
5372 Params.push_back(Scalars);
5376 assert(Params.size() == Instr->getNumOperands() &&
5377 "Invalid number of operands");
5379 // Does this instruction return a value ?
5380 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5382 Value *UndefVec = IsVoidRetTy ? nullptr :
5383 UndefValue::get(Instr->getType());
5384 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5385 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5387 Instruction *InsertPt = Builder.GetInsertPoint();
5388 BasicBlock *IfBlock = Builder.GetInsertBlock();
5389 BasicBlock *CondBlock = nullptr;
5392 Loop *VectorLp = nullptr;
5393 if (IfPredicateStore) {
5394 assert(Instr->getParent()->getSinglePredecessor() &&
5395 "Only support single predecessor blocks");
5396 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5397 Instr->getParent());
5398 VectorLp = LI->getLoopFor(IfBlock);
5399 assert(VectorLp && "Must have a loop for this block");
5402 // For each vector unroll 'part':
5403 for (unsigned Part = 0; Part < UF; ++Part) {
5404 // For each scalar that we create:
5406 // Start an "if (pred) a[i] = ..." block.
5407 Value *Cmp = nullptr;
5408 if (IfPredicateStore) {
5409 if (Cond[Part]->getType()->isVectorTy())
5411 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5412 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5413 ConstantInt::get(Cond[Part]->getType(), 1));
5414 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5415 LoopVectorBody.push_back(CondBlock);
5416 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5417 // Update Builder with newly created basic block.
5418 Builder.SetInsertPoint(InsertPt);
5421 Instruction *Cloned = Instr->clone();
5423 Cloned->setName(Instr->getName() + ".cloned");
5424 // Replace the operands of the cloned instructions with extracted scalars.
5425 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5426 Value *Op = Params[op][Part];
5427 Cloned->setOperand(op, Op);
5430 // Place the cloned scalar in the new loop.
5431 Builder.Insert(Cloned);
5433 // If the original scalar returns a value we need to place it in a vector
5434 // so that future users will be able to use it.
5436 VecResults[Part] = Cloned;
5439 if (IfPredicateStore) {
5440 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5441 LoopVectorBody.push_back(NewIfBlock);
5442 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5443 Builder.SetInsertPoint(InsertPt);
5444 ReplaceInstWithInst(IfBlock->getTerminator(),
5445 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5446 IfBlock = NewIfBlock;
5451 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5452 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5453 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5455 return scalarizeInstruction(Instr, IfPredicateStore);
5458 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5462 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5466 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5467 // When unrolling and the VF is 1, we only need to add a simple scalar.
5468 Type *ITy = Val->getType();
5469 assert(!ITy->isVectorTy() && "Val must be a scalar");
5470 Constant *C = ConstantInt::get(ITy, StartIdx);
5471 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");