1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/EquivalenceClasses.h"
52 #include "llvm/ADT/Hashing.h"
53 #include "llvm/ADT/MapVector.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/AliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/LoopAccessAnalysis.h"
66 #include "llvm/Analysis/LoopInfo.h"
67 #include "llvm/Analysis/LoopIterator.h"
68 #include "llvm/Analysis/LoopPass.h"
69 #include "llvm/Analysis/ScalarEvolution.h"
70 #include "llvm/Analysis/ScalarEvolutionExpander.h"
71 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
72 #include "llvm/Analysis/TargetTransformInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/IR/Constants.h"
75 #include "llvm/IR/DataLayout.h"
76 #include "llvm/IR/DebugInfo.h"
77 #include "llvm/IR/DerivedTypes.h"
78 #include "llvm/IR/DiagnosticInfo.h"
79 #include "llvm/IR/Dominators.h"
80 #include "llvm/IR/Function.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/Instructions.h"
83 #include "llvm/IR/IntrinsicInst.h"
84 #include "llvm/IR/LLVMContext.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/IR/Type.h"
88 #include "llvm/IR/Value.h"
89 #include "llvm/IR/ValueHandle.h"
90 #include "llvm/IR/Verifier.h"
91 #include "llvm/Pass.h"
92 #include "llvm/Support/BranchProbability.h"
93 #include "llvm/Support/CommandLine.h"
94 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/raw_ostream.h"
96 #include "llvm/Transforms/Scalar.h"
97 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
98 #include "llvm/Transforms/Utils/Local.h"
99 #include "llvm/Analysis/VectorUtils.h"
100 #include "llvm/Transforms/Utils/LoopUtils.h"
105 using namespace llvm;
106 using namespace llvm::PatternMatch;
108 #define LV_NAME "loop-vectorize"
109 #define DEBUG_TYPE LV_NAME
111 STATISTIC(LoopsVectorized, "Number of loops vectorized");
112 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 static cl::opt<bool> EnableInterleavedMemAccesses(
142 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
143 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
145 /// Maximum factor for an interleaved memory access.
146 static cl::opt<unsigned> MaxInterleaveGroupFactor(
147 "max-interleave-group-factor", cl::Hidden,
148 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
151 /// We don't interleave loops with a known constant trip count below this
153 static const unsigned TinyTripCountInterleaveThreshold = 128;
155 static cl::opt<unsigned> ForceTargetNumScalarRegs(
156 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of scalar registers."));
159 static cl::opt<unsigned> ForceTargetNumVectorRegs(
160 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's number of vector registers."));
163 /// Maximum vectorization interleave count.
164 static const unsigned MaxInterleaveFactor = 16;
166 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
167 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max interleave factor for "
171 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
172 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
174 "vectorized loops."));
176 static cl::opt<unsigned> ForceTargetInstructionCost(
177 "force-target-instruction-cost", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's expected cost for "
179 "an instruction to a single constant value. Mostly "
180 "useful for getting consistent testing."));
182 static cl::opt<unsigned> SmallLoopCost(
183 "small-loop-cost", cl::init(20), cl::Hidden,
185 "The cost of a loop that is considered 'small' by the interleaver."));
187 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
188 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
189 cl::desc("Enable the use of the block frequency analysis to access PGO "
190 "heuristics minimizing code growth in cold regions and being more "
191 "aggressive in hot regions."));
193 // Runtime interleave loops for load/store throughput.
194 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
195 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
197 "Enable runtime interleaving until load/store ports are saturated"));
199 /// The number of stores in a loop that are allowed to need predication.
200 static cl::opt<unsigned> NumberOfStoresToPredicate(
201 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
202 cl::desc("Max number of stores to be predicated behind an if."));
204 static cl::opt<bool> EnableIndVarRegisterHeur(
205 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
206 cl::desc("Count the induction variable only once when interleaving"));
208 static cl::opt<bool> EnableCondStoresVectorization(
209 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
210 cl::desc("Enable if predication of stores during vectorization."));
212 static cl::opt<unsigned> MaxNestedScalarReductionIC(
213 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
214 cl::desc("The maximum interleave count to use when interleaving a scalar "
215 "reduction in a nested loop."));
217 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
218 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
219 cl::desc("The maximum allowed number of runtime memory checks with a "
220 "vectorize(enable) pragma."));
224 // Forward declarations.
225 class LoopVectorizeHints;
226 class LoopVectorizationLegality;
227 class LoopVectorizationCostModel;
228 class LoopVectorizationRequirements;
230 /// \brief This modifies LoopAccessReport to initialize message with
231 /// loop-vectorizer-specific part.
232 class VectorizationReport : public LoopAccessReport {
234 VectorizationReport(Instruction *I = nullptr)
235 : LoopAccessReport("loop not vectorized: ", I) {}
237 /// \brief This allows promotion of the loop-access analysis report into the
238 /// loop-vectorizer report. It modifies the message to add the
239 /// loop-vectorizer-specific part of the message.
240 explicit VectorizationReport(const LoopAccessReport &R)
241 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
245 /// A helper function for converting Scalar types to vector types.
246 /// If the incoming type is void, we return void. If the VF is 1, we return
248 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
249 if (Scalar->isVoidTy() || VF == 1)
251 return VectorType::get(Scalar, VF);
254 /// InnerLoopVectorizer vectorizes loops which contain only one basic
255 /// block to a specified vectorization factor (VF).
256 /// This class performs the widening of scalars into vectors, or multiple
257 /// scalars. This class also implements the following features:
258 /// * It inserts an epilogue loop for handling loops that don't have iteration
259 /// counts that are known to be a multiple of the vectorization factor.
260 /// * It handles the code generation for reduction variables.
261 /// * Scalarization (implementation using scalars) of un-vectorizable
263 /// InnerLoopVectorizer does not perform any vectorization-legality
264 /// checks, and relies on the caller to check for the different legality
265 /// aspects. The InnerLoopVectorizer relies on the
266 /// LoopVectorizationLegality class to provide information about the induction
267 /// and reduction variables that were found to a given vectorization factor.
268 class InnerLoopVectorizer {
270 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
271 DominatorTree *DT, const TargetLibraryInfo *TLI,
272 const TargetTransformInfo *TTI, unsigned VecWidth,
273 unsigned UnrollFactor)
274 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
275 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
276 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
277 Legal(nullptr), AddedSafetyChecks(false) {}
279 // Perform the actual loop widening (vectorization).
280 void vectorize(LoopVectorizationLegality *L) {
282 // Create a new empty loop. Unlink the old loop and connect the new one.
284 // Widen each instruction in the old loop to a new one in the new loop.
285 // Use the Legality module to find the induction and reduction variables.
287 // Register the new loop and update the analysis passes.
291 // Return true if any runtime check is added.
292 bool IsSafetyChecksAdded() {
293 return AddedSafetyChecks;
296 virtual ~InnerLoopVectorizer() {}
299 /// A small list of PHINodes.
300 typedef SmallVector<PHINode*, 4> PhiVector;
301 /// When we unroll loops we have multiple vector values for each scalar.
302 /// This data structure holds the unrolled and vectorized values that
303 /// originated from one scalar instruction.
304 typedef SmallVector<Value*, 2> VectorParts;
306 // When we if-convert we need to create edge masks. We have to cache values
307 // so that we don't end up with exponential recursion/IR.
308 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
309 VectorParts> EdgeMaskCache;
311 /// \brief Add checks for strides that were assumed to be 1.
313 /// Returns the last check instruction and the first check instruction in the
314 /// pair as (first, last).
315 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
317 /// Create an empty loop, based on the loop ranges of the old loop.
318 void createEmptyLoop();
319 /// Copy and widen the instructions from the old loop.
320 virtual void vectorizeLoop();
322 /// \brief The Loop exit block may have single value PHI nodes where the
323 /// incoming value is 'Undef'. While vectorizing we only handled real values
324 /// that were defined inside the loop. Here we fix the 'undef case'.
328 /// A helper function that computes the predicate of the block BB, assuming
329 /// that the header block of the loop is set to True. It returns the *entry*
330 /// mask for the block BB.
331 VectorParts createBlockInMask(BasicBlock *BB);
332 /// A helper function that computes the predicate of the edge between SRC
334 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
336 /// A helper function to vectorize a single BB within the innermost loop.
337 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
339 /// Vectorize a single PHINode in a block. This method handles the induction
340 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
341 /// arbitrary length vectors.
342 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
343 unsigned UF, unsigned VF, PhiVector *PV);
345 /// Insert the new loop to the loop hierarchy and pass manager
346 /// and update the analysis passes.
347 void updateAnalysis();
349 /// This instruction is un-vectorizable. Implement it as a sequence
350 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
351 /// scalarized instruction behind an if block predicated on the control
352 /// dependence of the instruction.
353 virtual void scalarizeInstruction(Instruction *Instr,
354 bool IfPredicateStore=false);
356 /// Vectorize Load and Store instructions,
357 virtual void vectorizeMemoryInstruction(Instruction *Instr);
359 /// Create a broadcast instruction. This method generates a broadcast
360 /// instruction (shuffle) for loop invariant values and for the induction
361 /// value. If this is the induction variable then we extend it to N, N+1, ...
362 /// this is needed because each iteration in the loop corresponds to a SIMD
364 virtual Value *getBroadcastInstrs(Value *V);
366 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
367 /// to each vector element of Val. The sequence starts at StartIndex.
368 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
370 /// When we go over instructions in the basic block we rely on previous
371 /// values within the current basic block or on loop invariant values.
372 /// When we widen (vectorize) values we place them in the map. If the values
373 /// are not within the map, they have to be loop invariant, so we simply
374 /// broadcast them into a vector.
375 VectorParts &getVectorValue(Value *V);
377 /// Try to vectorize the interleaved access group that \p Instr belongs to.
378 void vectorizeInterleaveGroup(Instruction *Instr);
380 /// Generate a shuffle sequence that will reverse the vector Vec.
381 virtual Value *reverseVector(Value *Vec);
383 /// This is a helper class that holds the vectorizer state. It maps scalar
384 /// instructions to vector instructions. When the code is 'unrolled' then
385 /// then a single scalar value is mapped to multiple vector parts. The parts
386 /// are stored in the VectorPart type.
388 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
390 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
392 /// \return True if 'Key' is saved in the Value Map.
393 bool has(Value *Key) const { return MapStorage.count(Key); }
395 /// Initializes a new entry in the map. Sets all of the vector parts to the
396 /// save value in 'Val'.
397 /// \return A reference to a vector with splat values.
398 VectorParts &splat(Value *Key, Value *Val) {
399 VectorParts &Entry = MapStorage[Key];
400 Entry.assign(UF, Val);
404 ///\return A reference to the value that is stored at 'Key'.
405 VectorParts &get(Value *Key) {
406 VectorParts &Entry = MapStorage[Key];
409 assert(Entry.size() == UF);
414 /// The unroll factor. Each entry in the map stores this number of vector
418 /// Map storage. We use std::map and not DenseMap because insertions to a
419 /// dense map invalidates its iterators.
420 std::map<Value *, VectorParts> MapStorage;
423 /// The original loop.
425 /// Scev analysis to use.
433 /// Target Library Info.
434 const TargetLibraryInfo *TLI;
435 /// Target Transform Info.
436 const TargetTransformInfo *TTI;
438 /// The vectorization SIMD factor to use. Each vector will have this many
443 /// The vectorization unroll factor to use. Each scalar is vectorized to this
444 /// many different vector instructions.
447 /// The builder that we use
450 // --- Vectorization state ---
452 /// The vector-loop preheader.
453 BasicBlock *LoopVectorPreHeader;
454 /// The scalar-loop preheader.
455 BasicBlock *LoopScalarPreHeader;
456 /// Middle Block between the vector and the scalar.
457 BasicBlock *LoopMiddleBlock;
458 ///The ExitBlock of the scalar loop.
459 BasicBlock *LoopExitBlock;
460 ///The vector loop body.
461 SmallVector<BasicBlock *, 4> LoopVectorBody;
462 ///The scalar loop body.
463 BasicBlock *LoopScalarBody;
464 /// A list of all bypass blocks. The first block is the entry of the loop.
465 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
467 /// The new Induction variable which was added to the new block.
469 /// The induction variable of the old basic block.
470 PHINode *OldInduction;
471 /// Maps scalars to widened vectors.
473 EdgeMaskCache MaskCache;
475 LoopVectorizationLegality *Legal;
477 // Record whether runtime check is added.
478 bool AddedSafetyChecks;
481 class InnerLoopUnroller : public InnerLoopVectorizer {
483 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
484 DominatorTree *DT, const TargetLibraryInfo *TLI,
485 const TargetTransformInfo *TTI, unsigned UnrollFactor)
486 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
489 void scalarizeInstruction(Instruction *Instr,
490 bool IfPredicateStore = false) override;
491 void vectorizeMemoryInstruction(Instruction *Instr) override;
492 Value *getBroadcastInstrs(Value *V) override;
493 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
494 Value *reverseVector(Value *Vec) override;
497 /// \brief Look for a meaningful debug location on the instruction or it's
499 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
504 if (I->getDebugLoc() != Empty)
507 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
508 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
509 if (OpInst->getDebugLoc() != Empty)
516 /// \brief Set the debug location in the builder using the debug location in the
518 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
519 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
520 B.SetCurrentDebugLocation(Inst->getDebugLoc());
522 B.SetCurrentDebugLocation(DebugLoc());
526 /// \return string containing a file name and a line # for the given loop.
527 static std::string getDebugLocString(const Loop *L) {
530 raw_string_ostream OS(Result);
531 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
532 LoopDbgLoc.print(OS);
534 // Just print the module name.
535 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
542 /// \brief Propagate known metadata from one instruction to another.
543 static void propagateMetadata(Instruction *To, const Instruction *From) {
544 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
545 From->getAllMetadataOtherThanDebugLoc(Metadata);
547 for (auto M : Metadata) {
548 unsigned Kind = M.first;
550 // These are safe to transfer (this is safe for TBAA, even when we
551 // if-convert, because should that metadata have had a control dependency
552 // on the condition, and thus actually aliased with some other
553 // non-speculated memory access when the condition was false, this would be
554 // caught by the runtime overlap checks).
555 if (Kind != LLVMContext::MD_tbaa &&
556 Kind != LLVMContext::MD_alias_scope &&
557 Kind != LLVMContext::MD_noalias &&
558 Kind != LLVMContext::MD_fpmath &&
559 Kind != LLVMContext::MD_nontemporal)
562 To->setMetadata(Kind, M.second);
566 /// \brief Propagate known metadata from one instruction to a vector of others.
567 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
569 if (Instruction *I = dyn_cast<Instruction>(V))
570 propagateMetadata(I, From);
573 /// \brief The group of interleaved loads/stores sharing the same stride and
574 /// close to each other.
576 /// Each member in this group has an index starting from 0, and the largest
577 /// index should be less than interleaved factor, which is equal to the absolute
578 /// value of the access's stride.
580 /// E.g. An interleaved load group of factor 4:
581 /// for (unsigned i = 0; i < 1024; i+=4) {
582 /// a = A[i]; // Member of index 0
583 /// b = A[i+1]; // Member of index 1
584 /// d = A[i+3]; // Member of index 3
588 /// An interleaved store group of factor 4:
589 /// for (unsigned i = 0; i < 1024; i+=4) {
591 /// A[i] = a; // Member of index 0
592 /// A[i+1] = b; // Member of index 1
593 /// A[i+2] = c; // Member of index 2
594 /// A[i+3] = d; // Member of index 3
597 /// Note: the interleaved load group could have gaps (missing members), but
598 /// the interleaved store group doesn't allow gaps.
599 class InterleaveGroup {
601 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
602 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
603 assert(Align && "The alignment should be non-zero");
605 Factor = std::abs(Stride);
606 assert(Factor > 1 && "Invalid interleave factor");
608 Reverse = Stride < 0;
612 bool isReverse() const { return Reverse; }
613 unsigned getFactor() const { return Factor; }
614 unsigned getAlignment() const { return Align; }
615 unsigned getNumMembers() const { return Members.size(); }
617 /// \brief Try to insert a new member \p Instr with index \p Index and
618 /// alignment \p NewAlign. The index is related to the leader and it could be
619 /// negative if it is the new leader.
621 /// \returns false if the instruction doesn't belong to the group.
622 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
623 assert(NewAlign && "The new member's alignment should be non-zero");
625 int Key = Index + SmallestKey;
627 // Skip if there is already a member with the same index.
628 if (Members.count(Key))
631 if (Key > LargestKey) {
632 // The largest index is always less than the interleave factor.
633 if (Index >= static_cast<int>(Factor))
637 } else if (Key < SmallestKey) {
638 // The largest index is always less than the interleave factor.
639 if (LargestKey - Key >= static_cast<int>(Factor))
645 // It's always safe to select the minimum alignment.
646 Align = std::min(Align, NewAlign);
647 Members[Key] = Instr;
651 /// \brief Get the member with the given index \p Index
653 /// \returns nullptr if contains no such member.
654 Instruction *getMember(unsigned Index) const {
655 int Key = SmallestKey + Index;
656 if (!Members.count(Key))
659 return Members.find(Key)->second;
662 /// \brief Get the index for the given member. Unlike the key in the member
663 /// map, the index starts from 0.
664 unsigned getIndex(Instruction *Instr) const {
665 for (auto I : Members)
666 if (I.second == Instr)
667 return I.first - SmallestKey;
669 llvm_unreachable("InterleaveGroup contains no such member");
672 Instruction *getInsertPos() const { return InsertPos; }
673 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
676 unsigned Factor; // Interleave Factor.
679 DenseMap<int, Instruction *> Members;
683 // To avoid breaking dependences, vectorized instructions of an interleave
684 // group should be inserted at either the first load or the last store in
687 // E.g. %even = load i32 // Insert Position
688 // %add = add i32 %even // Use of %even
692 // %odd = add i32 // Def of %odd
693 // store i32 %odd // Insert Position
694 Instruction *InsertPos;
697 /// \brief Drive the analysis of interleaved memory accesses in the loop.
699 /// Use this class to analyze interleaved accesses only when we can vectorize
700 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
701 /// on interleaved accesses is unsafe.
703 /// The analysis collects interleave groups and records the relationships
704 /// between the member and the group in a map.
705 class InterleavedAccessInfo {
707 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
708 : SE(SE), TheLoop(L), DT(DT) {}
710 ~InterleavedAccessInfo() {
711 SmallSet<InterleaveGroup *, 4> DelSet;
712 // Avoid releasing a pointer twice.
713 for (auto &I : InterleaveGroupMap)
714 DelSet.insert(I.second);
715 for (auto *Ptr : DelSet)
719 /// \brief Analyze the interleaved accesses and collect them in interleave
720 /// groups. Substitute symbolic strides using \p Strides.
721 void analyzeInterleaving(const ValueToValueMap &Strides);
723 /// \brief Check if \p Instr belongs to any interleave group.
724 bool isInterleaved(Instruction *Instr) const {
725 return InterleaveGroupMap.count(Instr);
728 /// \brief Get the interleave group that \p Instr belongs to.
730 /// \returns nullptr if doesn't have such group.
731 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
732 if (InterleaveGroupMap.count(Instr))
733 return InterleaveGroupMap.find(Instr)->second;
742 /// Holds the relationships between the members and the interleave group.
743 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
745 /// \brief The descriptor for a strided memory access.
746 struct StrideDescriptor {
747 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
749 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
751 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
753 int Stride; // The access's stride. It is negative for a reverse access.
754 const SCEV *Scev; // The scalar expression of this access
755 unsigned Size; // The size of the memory object.
756 unsigned Align; // The alignment of this access.
759 /// \brief Create a new interleave group with the given instruction \p Instr,
760 /// stride \p Stride and alignment \p Align.
762 /// \returns the newly created interleave group.
763 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
765 assert(!InterleaveGroupMap.count(Instr) &&
766 "Already in an interleaved access group");
767 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
768 return InterleaveGroupMap[Instr];
771 /// \brief Release the group and remove all the relationships.
772 void releaseGroup(InterleaveGroup *Group) {
773 for (unsigned i = 0; i < Group->getFactor(); i++)
774 if (Instruction *Member = Group->getMember(i))
775 InterleaveGroupMap.erase(Member);
780 /// \brief Collect all the accesses with a constant stride in program order.
781 void collectConstStridedAccesses(
782 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
783 const ValueToValueMap &Strides);
786 /// Utility class for getting and setting loop vectorizer hints in the form
787 /// of loop metadata.
788 /// This class keeps a number of loop annotations locally (as member variables)
789 /// and can, upon request, write them back as metadata on the loop. It will
790 /// initially scan the loop for existing metadata, and will update the local
791 /// values based on information in the loop.
792 /// We cannot write all values to metadata, as the mere presence of some info,
793 /// for example 'force', means a decision has been made. So, we need to be
794 /// careful NOT to add them if the user hasn't specifically asked so.
795 class LoopVectorizeHints {
802 /// Hint - associates name and validation with the hint value.
805 unsigned Value; // This may have to change for non-numeric values.
808 Hint(const char * Name, unsigned Value, HintKind Kind)
809 : Name(Name), Value(Value), Kind(Kind) { }
811 bool validate(unsigned Val) {
814 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
816 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
824 /// Vectorization width.
826 /// Vectorization interleave factor.
828 /// Vectorization forced
831 /// Return the loop metadata prefix.
832 static StringRef Prefix() { return "llvm.loop."; }
836 FK_Undefined = -1, ///< Not selected.
837 FK_Disabled = 0, ///< Forcing disabled.
838 FK_Enabled = 1, ///< Forcing enabled.
841 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
842 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
844 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
845 Force("vectorize.enable", FK_Undefined, HK_FORCE),
847 // Populate values with existing loop metadata.
848 getHintsFromMetadata();
850 // force-vector-interleave overrides DisableInterleaving.
851 if (VectorizerParams::isInterleaveForced())
852 Interleave.Value = VectorizerParams::VectorizationInterleave;
854 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
855 << "LV: Interleaving disabled by the pass manager\n");
858 /// Mark the loop L as already vectorized by setting the width to 1.
859 void setAlreadyVectorized() {
860 Width.Value = Interleave.Value = 1;
861 Hint Hints[] = {Width, Interleave};
862 writeHintsToMetadata(Hints);
865 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
866 if (getForce() == LoopVectorizeHints::FK_Disabled) {
867 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
868 emitOptimizationRemarkAnalysis(F->getContext(),
869 vectorizeAnalysisPassName(), *F,
870 L->getStartLoc(), emitRemark());
874 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
875 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
876 emitOptimizationRemarkAnalysis(F->getContext(),
877 vectorizeAnalysisPassName(), *F,
878 L->getStartLoc(), emitRemark());
882 if (getWidth() == 1 && getInterleave() == 1) {
883 // FIXME: Add a separate metadata to indicate when the loop has already
884 // been vectorized instead of setting width and count to 1.
885 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
886 // FIXME: Add interleave.disable metadata. This will allow
887 // vectorize.disable to be used without disabling the pass and errors
888 // to differentiate between disabled vectorization and a width of 1.
889 emitOptimizationRemarkAnalysis(
890 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
891 "loop not vectorized: vectorization and interleaving are explicitly "
892 "disabled, or vectorize width and interleave count are both set to "
900 /// Dumps all the hint information.
901 std::string emitRemark() const {
902 VectorizationReport R;
903 if (Force.Value == LoopVectorizeHints::FK_Disabled)
904 R << "vectorization is explicitly disabled";
906 R << "use -Rpass-analysis=loop-vectorize for more info";
907 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
909 if (Width.Value != 0)
910 R << ", Vector Width=" << Width.Value;
911 if (Interleave.Value != 0)
912 R << ", Interleave Count=" << Interleave.Value;
920 unsigned getWidth() const { return Width.Value; }
921 unsigned getInterleave() const { return Interleave.Value; }
922 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
923 const char *vectorizeAnalysisPassName() const {
924 // If hints are provided that don't disable vectorization use the
925 // AlwaysPrint pass name to force the frontend to print the diagnostic.
928 if (getForce() == LoopVectorizeHints::FK_Disabled)
930 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
932 return DiagnosticInfo::AlwaysPrint;
935 bool allowReordering() const {
936 // When enabling loop hints are provided we allow the vectorizer to change
937 // the order of operations that is given by the scalar loop. This is not
938 // enabled by default because can be unsafe or inefficient. For example,
939 // reordering floating-point operations will change the way round-off
940 // error accumulates in the loop.
941 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
945 /// Find hints specified in the loop metadata and update local values.
946 void getHintsFromMetadata() {
947 MDNode *LoopID = TheLoop->getLoopID();
951 // First operand should refer to the loop id itself.
952 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
953 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
955 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
956 const MDString *S = nullptr;
957 SmallVector<Metadata *, 4> Args;
959 // The expected hint is either a MDString or a MDNode with the first
960 // operand a MDString.
961 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
962 if (!MD || MD->getNumOperands() == 0)
964 S = dyn_cast<MDString>(MD->getOperand(0));
965 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
966 Args.push_back(MD->getOperand(i));
968 S = dyn_cast<MDString>(LoopID->getOperand(i));
969 assert(Args.size() == 0 && "too many arguments for MDString");
975 // Check if the hint starts with the loop metadata prefix.
976 StringRef Name = S->getString();
977 if (Args.size() == 1)
978 setHint(Name, Args[0]);
982 /// Checks string hint with one operand and set value if valid.
983 void setHint(StringRef Name, Metadata *Arg) {
984 if (!Name.startswith(Prefix()))
986 Name = Name.substr(Prefix().size(), StringRef::npos);
988 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
990 unsigned Val = C->getZExtValue();
992 Hint *Hints[] = {&Width, &Interleave, &Force};
993 for (auto H : Hints) {
994 if (Name == H->Name) {
995 if (H->validate(Val))
998 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1004 /// Create a new hint from name / value pair.
1005 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1006 LLVMContext &Context = TheLoop->getHeader()->getContext();
1007 Metadata *MDs[] = {MDString::get(Context, Name),
1008 ConstantAsMetadata::get(
1009 ConstantInt::get(Type::getInt32Ty(Context), V))};
1010 return MDNode::get(Context, MDs);
1013 /// Matches metadata with hint name.
1014 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1015 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1019 for (auto H : HintTypes)
1020 if (Name->getString().endswith(H.Name))
1025 /// Sets current hints into loop metadata, keeping other values intact.
1026 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1027 if (HintTypes.size() == 0)
1030 // Reserve the first element to LoopID (see below).
1031 SmallVector<Metadata *, 4> MDs(1);
1032 // If the loop already has metadata, then ignore the existing operands.
1033 MDNode *LoopID = TheLoop->getLoopID();
1035 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1036 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1037 // If node in update list, ignore old value.
1038 if (!matchesHintMetadataName(Node, HintTypes))
1039 MDs.push_back(Node);
1043 // Now, add the missing hints.
1044 for (auto H : HintTypes)
1045 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1047 // Replace current metadata node with new one.
1048 LLVMContext &Context = TheLoop->getHeader()->getContext();
1049 MDNode *NewLoopID = MDNode::get(Context, MDs);
1050 // Set operand 0 to refer to the loop id itself.
1051 NewLoopID->replaceOperandWith(0, NewLoopID);
1053 TheLoop->setLoopID(NewLoopID);
1056 /// The loop these hints belong to.
1057 const Loop *TheLoop;
1060 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1061 const LoopVectorizeHints &Hints,
1062 const LoopAccessReport &Message) {
1063 const char *Name = Hints.vectorizeAnalysisPassName();
1064 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1067 static void emitMissedWarning(Function *F, Loop *L,
1068 const LoopVectorizeHints &LH) {
1069 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1072 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1073 if (LH.getWidth() != 1)
1074 emitLoopVectorizeWarning(
1075 F->getContext(), *F, L->getStartLoc(),
1076 "failed explicitly specified loop vectorization");
1077 else if (LH.getInterleave() != 1)
1078 emitLoopInterleaveWarning(
1079 F->getContext(), *F, L->getStartLoc(),
1080 "failed explicitly specified loop interleaving");
1084 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1085 /// to what vectorization factor.
1086 /// This class does not look at the profitability of vectorization, only the
1087 /// legality. This class has two main kinds of checks:
1088 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1089 /// will change the order of memory accesses in a way that will change the
1090 /// correctness of the program.
1091 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1092 /// checks for a number of different conditions, such as the availability of a
1093 /// single induction variable, that all types are supported and vectorize-able,
1094 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1095 /// This class is also used by InnerLoopVectorizer for identifying
1096 /// induction variable and the different reduction variables.
1097 class LoopVectorizationLegality {
1099 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1100 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1101 Function *F, const TargetTransformInfo *TTI,
1102 LoopAccessAnalysis *LAA,
1103 LoopVectorizationRequirements *R,
1104 const LoopVectorizeHints *H)
1105 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1106 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1107 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1108 Requirements(R), Hints(H) {}
1110 /// ReductionList contains the reduction descriptors for all
1111 /// of the reductions that were found in the loop.
1112 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1114 /// InductionList saves induction variables and maps them to the
1115 /// induction descriptor.
1116 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1118 /// Returns true if it is legal to vectorize this loop.
1119 /// This does not mean that it is profitable to vectorize this
1120 /// loop, only that it is legal to do so.
1121 bool canVectorize();
1123 /// Returns the Induction variable.
1124 PHINode *getInduction() { return Induction; }
1126 /// Returns the reduction variables found in the loop.
1127 ReductionList *getReductionVars() { return &Reductions; }
1129 /// Returns the induction variables found in the loop.
1130 InductionList *getInductionVars() { return &Inductions; }
1132 /// Returns the widest induction type.
1133 Type *getWidestInductionType() { return WidestIndTy; }
1135 /// Returns True if V is an induction variable in this loop.
1136 bool isInductionVariable(const Value *V);
1138 /// Return true if the block BB needs to be predicated in order for the loop
1139 /// to be vectorized.
1140 bool blockNeedsPredication(BasicBlock *BB);
1142 /// Check if this pointer is consecutive when vectorizing. This happens
1143 /// when the last index of the GEP is the induction variable, or that the
1144 /// pointer itself is an induction variable.
1145 /// This check allows us to vectorize A[idx] into a wide load/store.
1147 /// 0 - Stride is unknown or non-consecutive.
1148 /// 1 - Address is consecutive.
1149 /// -1 - Address is consecutive, and decreasing.
1150 int isConsecutivePtr(Value *Ptr);
1152 /// Returns true if the value V is uniform within the loop.
1153 bool isUniform(Value *V);
1155 /// Returns true if this instruction will remain scalar after vectorization.
1156 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1158 /// Returns the information that we collected about runtime memory check.
1159 const RuntimePointerChecking *getRuntimePointerChecking() const {
1160 return LAI->getRuntimePointerChecking();
1163 const LoopAccessInfo *getLAI() const {
1167 /// \brief Check if \p Instr belongs to any interleaved access group.
1168 bool isAccessInterleaved(Instruction *Instr) {
1169 return InterleaveInfo.isInterleaved(Instr);
1172 /// \brief Get the interleaved access group that \p Instr belongs to.
1173 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1174 return InterleaveInfo.getInterleaveGroup(Instr);
1177 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1179 bool hasStride(Value *V) { return StrideSet.count(V); }
1180 bool mustCheckStrides() { return !StrideSet.empty(); }
1181 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1182 return StrideSet.begin();
1184 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1186 /// Returns true if the target machine supports masked store operation
1187 /// for the given \p DataType and kind of access to \p Ptr.
1188 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1189 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1191 /// Returns true if the target machine supports masked load operation
1192 /// for the given \p DataType and kind of access to \p Ptr.
1193 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1194 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1196 /// Returns true if vector representation of the instruction \p I
1198 bool isMaskRequired(const Instruction* I) {
1199 return (MaskedOp.count(I) != 0);
1201 unsigned getNumStores() const {
1202 return LAI->getNumStores();
1204 unsigned getNumLoads() const {
1205 return LAI->getNumLoads();
1207 unsigned getNumPredStores() const {
1208 return NumPredStores;
1211 /// Check if a single basic block loop is vectorizable.
1212 /// At this point we know that this is a loop with a constant trip count
1213 /// and we only need to check individual instructions.
1214 bool canVectorizeInstrs();
1216 /// When we vectorize loops we may change the order in which
1217 /// we read and write from memory. This method checks if it is
1218 /// legal to vectorize the code, considering only memory constrains.
1219 /// Returns true if the loop is vectorizable
1220 bool canVectorizeMemory();
1222 /// Return true if we can vectorize this loop using the IF-conversion
1224 bool canVectorizeWithIfConvert();
1226 /// Collect the variables that need to stay uniform after vectorization.
1227 void collectLoopUniforms();
1229 /// Return true if all of the instructions in the block can be speculatively
1230 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1231 /// and we know that we can read from them without segfault.
1232 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1234 /// \brief Collect memory access with loop invariant strides.
1236 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1238 void collectStridedAccess(Value *LoadOrStoreInst);
1240 /// Report an analysis message to assist the user in diagnosing loops that are
1241 /// not vectorized. These are handled as LoopAccessReport rather than
1242 /// VectorizationReport because the << operator of VectorizationReport returns
1243 /// LoopAccessReport.
1244 void emitAnalysis(const LoopAccessReport &Message) const {
1245 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1248 unsigned NumPredStores;
1250 /// The loop that we evaluate.
1253 ScalarEvolution *SE;
1254 /// Target Library Info.
1255 TargetLibraryInfo *TLI;
1257 Function *TheFunction;
1258 /// Target Transform Info
1259 const TargetTransformInfo *TTI;
1262 // LoopAccess analysis.
1263 LoopAccessAnalysis *LAA;
1264 // And the loop-accesses info corresponding to this loop. This pointer is
1265 // null until canVectorizeMemory sets it up.
1266 const LoopAccessInfo *LAI;
1268 /// The interleave access information contains groups of interleaved accesses
1269 /// with the same stride and close to each other.
1270 InterleavedAccessInfo InterleaveInfo;
1272 // --- vectorization state --- //
1274 /// Holds the integer induction variable. This is the counter of the
1277 /// Holds the reduction variables.
1278 ReductionList Reductions;
1279 /// Holds all of the induction variables that we found in the loop.
1280 /// Notice that inductions don't need to start at zero and that induction
1281 /// variables can be pointers.
1282 InductionList Inductions;
1283 /// Holds the widest induction type encountered.
1286 /// Allowed outside users. This holds the reduction
1287 /// vars which can be accessed from outside the loop.
1288 SmallPtrSet<Value*, 4> AllowedExit;
1289 /// This set holds the variables which are known to be uniform after
1291 SmallPtrSet<Instruction*, 4> Uniforms;
1293 /// Can we assume the absence of NaNs.
1294 bool HasFunNoNaNAttr;
1296 /// Vectorization requirements that will go through late-evaluation.
1297 LoopVectorizationRequirements *Requirements;
1299 /// Used to emit an analysis of any legality issues.
1300 const LoopVectorizeHints *Hints;
1302 ValueToValueMap Strides;
1303 SmallPtrSet<Value *, 8> StrideSet;
1305 /// While vectorizing these instructions we have to generate a
1306 /// call to the appropriate masked intrinsic
1307 SmallPtrSet<const Instruction*, 8> MaskedOp;
1310 /// LoopVectorizationCostModel - estimates the expected speedups due to
1312 /// In many cases vectorization is not profitable. This can happen because of
1313 /// a number of reasons. In this class we mainly attempt to predict the
1314 /// expected speedup/slowdowns due to the supported instruction set. We use the
1315 /// TargetTransformInfo to query the different backends for the cost of
1316 /// different operations.
1317 class LoopVectorizationCostModel {
1319 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1320 LoopVectorizationLegality *Legal,
1321 const TargetTransformInfo &TTI,
1322 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1323 const Function *F, const LoopVectorizeHints *Hints,
1324 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1325 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1326 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1328 /// Information about vectorization costs
1329 struct VectorizationFactor {
1330 unsigned Width; // Vector width with best cost
1331 unsigned Cost; // Cost of the loop with that width
1333 /// \return The most profitable vectorization factor and the cost of that VF.
1334 /// This method checks every power of two up to VF. If UserVF is not ZERO
1335 /// then this vectorization factor will be selected if vectorization is
1337 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1339 /// \return The size (in bits) of the widest type in the code that
1340 /// needs to be vectorized. We ignore values that remain scalar such as
1341 /// 64 bit loop indices.
1342 unsigned getWidestType();
1344 /// \return The desired interleave count.
1345 /// If interleave count has been specified by metadata it will be returned.
1346 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1347 /// are the selected vectorization factor and the cost of the selected VF.
1348 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1351 /// \return The most profitable unroll factor.
1352 /// This method finds the best unroll-factor based on register pressure and
1353 /// other parameters. VF and LoopCost are the selected vectorization factor
1354 /// and the cost of the selected VF.
1355 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1358 /// \brief A struct that represents some properties of the register usage
1360 struct RegisterUsage {
1361 /// Holds the number of loop invariant values that are used in the loop.
1362 unsigned LoopInvariantRegs;
1363 /// Holds the maximum number of concurrent live intervals in the loop.
1364 unsigned MaxLocalUsers;
1365 /// Holds the number of instructions in the loop.
1366 unsigned NumInstructions;
1369 /// \return information about the register usage of the loop.
1370 RegisterUsage calculateRegisterUsage();
1373 /// Returns the expected execution cost. The unit of the cost does
1374 /// not matter because we use the 'cost' units to compare different
1375 /// vector widths. The cost that is returned is *not* normalized by
1376 /// the factor width.
1377 unsigned expectedCost(unsigned VF);
1379 /// Returns the execution time cost of an instruction for a given vector
1380 /// width. Vector width of one means scalar.
1381 unsigned getInstructionCost(Instruction *I, unsigned VF);
1383 /// Returns whether the instruction is a load or store and will be a emitted
1384 /// as a vector operation.
1385 bool isConsecutiveLoadOrStore(Instruction *I);
1387 /// Report an analysis message to assist the user in diagnosing loops that are
1388 /// not vectorized. These are handled as LoopAccessReport rather than
1389 /// VectorizationReport because the << operator of VectorizationReport returns
1390 /// LoopAccessReport.
1391 void emitAnalysis(const LoopAccessReport &Message) const {
1392 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1395 /// The loop that we evaluate.
1398 ScalarEvolution *SE;
1399 /// Loop Info analysis.
1401 /// Vectorization legality.
1402 LoopVectorizationLegality *Legal;
1403 /// Vector target information.
1404 const TargetTransformInfo &TTI;
1405 /// Target Library Info.
1406 const TargetLibraryInfo *TLI;
1407 const Function *TheFunction;
1408 // Loop Vectorize Hint.
1409 const LoopVectorizeHints *Hints;
1410 // Values to ignore in the cost model.
1411 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1414 /// \brief This holds vectorization requirements that must be verified late in
1415 /// the process. The requirements are set by legalize and costmodel. Once
1416 /// vectorization has been determined to be possible and profitable the
1417 /// requirements can be verified by looking for metadata or compiler options.
1418 /// For example, some loops require FP commutativity which is only allowed if
1419 /// vectorization is explicitly specified or if the fast-math compiler option
1420 /// has been provided.
1421 /// Late evaluation of these requirements allows helpful diagnostics to be
1422 /// composed that tells the user what need to be done to vectorize the loop. For
1423 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1424 /// evaluation should be used only when diagnostics can generated that can be
1425 /// followed by a non-expert user.
1426 class LoopVectorizationRequirements {
1428 LoopVectorizationRequirements()
1429 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1431 void addUnsafeAlgebraInst(Instruction *I) {
1432 // First unsafe algebra instruction.
1433 if (!UnsafeAlgebraInst)
1434 UnsafeAlgebraInst = I;
1437 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1439 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1440 const char *Name = Hints.vectorizeAnalysisPassName();
1441 bool Failed = false;
1442 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1443 emitOptimizationRemarkAnalysisFPCommute(
1444 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1445 VectorizationReport() << "cannot prove it is safe to reorder "
1446 "floating-point operations");
1450 // Test if runtime memcheck thresholds are exceeded.
1451 bool PragmaThresholdReached =
1452 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1453 bool ThresholdReached =
1454 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1455 if ((ThresholdReached && !Hints.allowReordering()) ||
1456 PragmaThresholdReached) {
1457 emitOptimizationRemarkAnalysisAliasing(
1458 F->getContext(), Name, *F, L->getStartLoc(),
1459 VectorizationReport()
1460 << "cannot prove it is safe to reorder memory operations");
1461 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1469 unsigned NumRuntimePointerChecks;
1470 Instruction *UnsafeAlgebraInst;
1473 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1475 return V.push_back(&L);
1477 for (Loop *InnerL : L)
1478 addInnerLoop(*InnerL, V);
1481 /// The LoopVectorize Pass.
1482 struct LoopVectorize : public FunctionPass {
1483 /// Pass identification, replacement for typeid
1486 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1488 DisableUnrolling(NoUnrolling),
1489 AlwaysVectorize(AlwaysVectorize) {
1490 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1493 ScalarEvolution *SE;
1495 TargetTransformInfo *TTI;
1497 BlockFrequencyInfo *BFI;
1498 TargetLibraryInfo *TLI;
1500 AssumptionCache *AC;
1501 LoopAccessAnalysis *LAA;
1502 bool DisableUnrolling;
1503 bool AlwaysVectorize;
1505 BlockFrequency ColdEntryFreq;
1507 bool runOnFunction(Function &F) override {
1508 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1509 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1510 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1511 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1512 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1513 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1514 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1515 AA = &getAnalysis<AliasAnalysis>();
1516 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1517 LAA = &getAnalysis<LoopAccessAnalysis>();
1519 // Compute some weights outside of the loop over the loops. Compute this
1520 // using a BranchProbability to re-use its scaling math.
1521 const BranchProbability ColdProb(1, 5); // 20%
1522 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1525 // 1. the target claims to have no vector registers, and
1526 // 2. interleaving won't help ILP.
1528 // The second condition is necessary because, even if the target has no
1529 // vector registers, loop vectorization may still enable scalar
1531 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1534 // Build up a worklist of inner-loops to vectorize. This is necessary as
1535 // the act of vectorizing or partially unrolling a loop creates new loops
1536 // and can invalidate iterators across the loops.
1537 SmallVector<Loop *, 8> Worklist;
1540 addInnerLoop(*L, Worklist);
1542 LoopsAnalyzed += Worklist.size();
1544 // Now walk the identified inner loops.
1545 bool Changed = false;
1546 while (!Worklist.empty())
1547 Changed |= processLoop(Worklist.pop_back_val());
1549 // Process each loop nest in the function.
1553 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1554 SmallVector<Metadata *, 4> MDs;
1555 // Reserve first location for self reference to the LoopID metadata node.
1556 MDs.push_back(nullptr);
1557 bool IsUnrollMetadata = false;
1558 MDNode *LoopID = L->getLoopID();
1560 // First find existing loop unrolling disable metadata.
1561 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1562 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1564 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1566 S && S->getString().startswith("llvm.loop.unroll.disable");
1568 MDs.push_back(LoopID->getOperand(i));
1572 if (!IsUnrollMetadata) {
1573 // Add runtime unroll disable metadata.
1574 LLVMContext &Context = L->getHeader()->getContext();
1575 SmallVector<Metadata *, 1> DisableOperands;
1576 DisableOperands.push_back(
1577 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1578 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1579 MDs.push_back(DisableNode);
1580 MDNode *NewLoopID = MDNode::get(Context, MDs);
1581 // Set operand 0 to refer to the loop id itself.
1582 NewLoopID->replaceOperandWith(0, NewLoopID);
1583 L->setLoopID(NewLoopID);
1587 bool processLoop(Loop *L) {
1588 assert(L->empty() && "Only process inner loops.");
1591 const std::string DebugLocStr = getDebugLocString(L);
1594 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1595 << L->getHeader()->getParent()->getName() << "\" from "
1596 << DebugLocStr << "\n");
1598 LoopVectorizeHints Hints(L, DisableUnrolling);
1600 DEBUG(dbgs() << "LV: Loop hints:"
1602 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1604 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1606 : "?")) << " width=" << Hints.getWidth()
1607 << " unroll=" << Hints.getInterleave() << "\n");
1609 // Function containing loop
1610 Function *F = L->getHeader()->getParent();
1612 // Looking at the diagnostic output is the only way to determine if a loop
1613 // was vectorized (other than looking at the IR or machine code), so it
1614 // is important to generate an optimization remark for each loop. Most of
1615 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1616 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1617 // less verbose reporting vectorized loops and unvectorized loops that may
1618 // benefit from vectorization, respectively.
1620 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1621 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1625 // Check the loop for a trip count threshold:
1626 // do not vectorize loops with a tiny trip count.
1627 const unsigned TC = SE->getSmallConstantTripCount(L);
1628 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1629 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1630 << "This loop is not worth vectorizing.");
1631 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1632 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1634 DEBUG(dbgs() << "\n");
1635 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1636 << "vectorization is not beneficial "
1637 "and is not explicitly forced");
1642 // Check if it is legal to vectorize the loop.
1643 LoopVectorizationRequirements Requirements;
1644 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1645 &Requirements, &Hints);
1646 if (!LVL.canVectorize()) {
1647 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1648 emitMissedWarning(F, L, Hints);
1652 // Collect values we want to ignore in the cost model. This includes
1653 // type-promoting instructions we identified during reduction detection.
1654 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1655 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1656 for (auto &Reduction : *LVL.getReductionVars()) {
1657 RecurrenceDescriptor &RedDes = Reduction.second;
1658 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1659 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1662 // Use the cost model.
1663 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints,
1666 // Check the function attributes to find out if this function should be
1667 // optimized for size.
1668 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1671 // Compute the weighted frequency of this loop being executed and see if it
1672 // is less than 20% of the function entry baseline frequency. Note that we
1673 // always have a canonical loop here because we think we *can* vectorize.
1674 // FIXME: This is hidden behind a flag due to pervasive problems with
1675 // exactly what block frequency models.
1676 if (LoopVectorizeWithBlockFrequency) {
1677 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1678 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1679 LoopEntryFreq < ColdEntryFreq)
1683 // Check the function attributes to see if implicit floats are allowed.
1684 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1685 // an integer loop and the vector instructions selected are purely integer
1686 // vector instructions?
1687 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1688 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1689 "attribute is used.\n");
1692 VectorizationReport()
1693 << "loop not vectorized due to NoImplicitFloat attribute");
1694 emitMissedWarning(F, L, Hints);
1698 // Select the optimal vectorization factor.
1699 const LoopVectorizationCostModel::VectorizationFactor VF =
1700 CM.selectVectorizationFactor(OptForSize);
1702 // Select the interleave count.
1703 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1705 // Get user interleave count.
1706 unsigned UserIC = Hints.getInterleave();
1708 // Identify the diagnostic messages that should be produced.
1709 std::string VecDiagMsg, IntDiagMsg;
1710 bool VectorizeLoop = true, InterleaveLoop = true;
1712 if (Requirements.doesNotMeet(F, L, Hints)) {
1713 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1715 emitMissedWarning(F, L, Hints);
1719 if (VF.Width == 1) {
1720 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1722 "the cost-model indicates that vectorization is not beneficial";
1723 VectorizeLoop = false;
1726 if (IC == 1 && UserIC <= 1) {
1727 // Tell the user interleaving is not beneficial.
1728 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1730 "the cost-model indicates that interleaving is not beneficial";
1731 InterleaveLoop = false;
1734 " and is explicitly disabled or interleave count is set to 1";
1735 } else if (IC > 1 && UserIC == 1) {
1736 // Tell the user interleaving is beneficial, but it explicitly disabled.
1738 << "LV: Interleaving is beneficial but is explicitly disabled.");
1739 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1740 "but is explicitly disabled or interleave count is set to 1";
1741 InterleaveLoop = false;
1744 // Override IC if user provided an interleave count.
1745 IC = UserIC > 0 ? UserIC : IC;
1747 // Emit diagnostic messages, if any.
1748 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1749 if (!VectorizeLoop && !InterleaveLoop) {
1750 // Do not vectorize or interleaving the loop.
1751 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1752 L->getStartLoc(), VecDiagMsg);
1753 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1754 L->getStartLoc(), IntDiagMsg);
1756 } else if (!VectorizeLoop && InterleaveLoop) {
1757 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1758 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1759 L->getStartLoc(), VecDiagMsg);
1760 } else if (VectorizeLoop && !InterleaveLoop) {
1761 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1762 << DebugLocStr << '\n');
1763 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1764 L->getStartLoc(), IntDiagMsg);
1765 } else if (VectorizeLoop && InterleaveLoop) {
1766 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1767 << DebugLocStr << '\n');
1768 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1771 if (!VectorizeLoop) {
1772 assert(IC > 1 && "interleave count should not be 1 or 0");
1773 // If we decided that it is not legal to vectorize the loop then
1775 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1776 Unroller.vectorize(&LVL);
1778 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1779 Twine("interleaved loop (interleaved count: ") +
1782 // If we decided that it is *legal* to vectorize the loop then do it.
1783 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1787 // Add metadata to disable runtime unrolling scalar loop when there's no
1788 // runtime check about strides and memory. Because at this situation,
1789 // scalar loop is rarely used not worthy to be unrolled.
1790 if (!LB.IsSafetyChecksAdded())
1791 AddRuntimeUnrollDisableMetaData(L);
1793 // Report the vectorization decision.
1794 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1795 Twine("vectorized loop (vectorization width: ") +
1796 Twine(VF.Width) + ", interleaved count: " +
1800 // Mark the loop as already vectorized to avoid vectorizing again.
1801 Hints.setAlreadyVectorized();
1803 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1807 void getAnalysisUsage(AnalysisUsage &AU) const override {
1808 AU.addRequired<AssumptionCacheTracker>();
1809 AU.addRequiredID(LoopSimplifyID);
1810 AU.addRequiredID(LCSSAID);
1811 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1812 AU.addRequired<DominatorTreeWrapperPass>();
1813 AU.addRequired<LoopInfoWrapperPass>();
1814 AU.addRequired<ScalarEvolutionWrapperPass>();
1815 AU.addRequired<TargetTransformInfoWrapperPass>();
1816 AU.addRequired<AliasAnalysis>();
1817 AU.addRequired<LoopAccessAnalysis>();
1818 AU.addPreserved<LoopInfoWrapperPass>();
1819 AU.addPreserved<DominatorTreeWrapperPass>();
1820 AU.addPreserved<AliasAnalysis>();
1825 } // end anonymous namespace
1827 //===----------------------------------------------------------------------===//
1828 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1829 // LoopVectorizationCostModel.
1830 //===----------------------------------------------------------------------===//
1832 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1833 // We need to place the broadcast of invariant variables outside the loop.
1834 Instruction *Instr = dyn_cast<Instruction>(V);
1836 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1837 Instr->getParent()) != LoopVectorBody.end());
1838 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1840 // Place the code for broadcasting invariant variables in the new preheader.
1841 IRBuilder<>::InsertPointGuard Guard(Builder);
1843 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1845 // Broadcast the scalar into all locations in the vector.
1846 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1851 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1853 assert(Val->getType()->isVectorTy() && "Must be a vector");
1854 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1855 "Elem must be an integer");
1856 assert(Step->getType() == Val->getType()->getScalarType() &&
1857 "Step has wrong type");
1858 // Create the types.
1859 Type *ITy = Val->getType()->getScalarType();
1860 VectorType *Ty = cast<VectorType>(Val->getType());
1861 int VLen = Ty->getNumElements();
1862 SmallVector<Constant*, 8> Indices;
1864 // Create a vector of consecutive numbers from zero to VF.
1865 for (int i = 0; i < VLen; ++i)
1866 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1868 // Add the consecutive indices to the vector value.
1869 Constant *Cv = ConstantVector::get(Indices);
1870 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1871 Step = Builder.CreateVectorSplat(VLen, Step);
1872 assert(Step->getType() == Val->getType() && "Invalid step vec");
1873 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1874 // which can be found from the original scalar operations.
1875 Step = Builder.CreateMul(Cv, Step);
1876 return Builder.CreateAdd(Val, Step, "induction");
1879 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1880 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1881 // Make sure that the pointer does not point to structs.
1882 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1885 // If this value is a pointer induction variable we know it is consecutive.
1886 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1887 if (Phi && Inductions.count(Phi)) {
1888 InductionDescriptor II = Inductions[Phi];
1889 return II.getConsecutiveDirection();
1892 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1896 unsigned NumOperands = Gep->getNumOperands();
1897 Value *GpPtr = Gep->getPointerOperand();
1898 // If this GEP value is a consecutive pointer induction variable and all of
1899 // the indices are constant then we know it is consecutive. We can
1900 Phi = dyn_cast<PHINode>(GpPtr);
1901 if (Phi && Inductions.count(Phi)) {
1903 // Make sure that the pointer does not point to structs.
1904 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1905 if (GepPtrType->getElementType()->isAggregateType())
1908 // Make sure that all of the index operands are loop invariant.
1909 for (unsigned i = 1; i < NumOperands; ++i)
1910 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1913 InductionDescriptor II = Inductions[Phi];
1914 return II.getConsecutiveDirection();
1917 unsigned InductionOperand = getGEPInductionOperand(Gep);
1919 // Check that all of the gep indices are uniform except for our induction
1921 for (unsigned i = 0; i != NumOperands; ++i)
1922 if (i != InductionOperand &&
1923 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1926 // We can emit wide load/stores only if the last non-zero index is the
1927 // induction variable.
1928 const SCEV *Last = nullptr;
1929 if (!Strides.count(Gep))
1930 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1932 // Because of the multiplication by a stride we can have a s/zext cast.
1933 // We are going to replace this stride by 1 so the cast is safe to ignore.
1935 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1936 // %0 = trunc i64 %indvars.iv to i32
1937 // %mul = mul i32 %0, %Stride1
1938 // %idxprom = zext i32 %mul to i64 << Safe cast.
1939 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1941 Last = replaceSymbolicStrideSCEV(SE, Strides,
1942 Gep->getOperand(InductionOperand), Gep);
1943 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1945 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1949 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1950 const SCEV *Step = AR->getStepRecurrence(*SE);
1952 // The memory is consecutive because the last index is consecutive
1953 // and all other indices are loop invariant.
1956 if (Step->isAllOnesValue())
1963 bool LoopVectorizationLegality::isUniform(Value *V) {
1964 return LAI->isUniform(V);
1967 InnerLoopVectorizer::VectorParts&
1968 InnerLoopVectorizer::getVectorValue(Value *V) {
1969 assert(V != Induction && "The new induction variable should not be used.");
1970 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1972 // If we have a stride that is replaced by one, do it here.
1973 if (Legal->hasStride(V))
1974 V = ConstantInt::get(V->getType(), 1);
1976 // If we have this scalar in the map, return it.
1977 if (WidenMap.has(V))
1978 return WidenMap.get(V);
1980 // If this scalar is unknown, assume that it is a constant or that it is
1981 // loop invariant. Broadcast V and save the value for future uses.
1982 Value *B = getBroadcastInstrs(V);
1983 return WidenMap.splat(V, B);
1986 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1987 assert(Vec->getType()->isVectorTy() && "Invalid type");
1988 SmallVector<Constant*, 8> ShuffleMask;
1989 for (unsigned i = 0; i < VF; ++i)
1990 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1992 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1993 ConstantVector::get(ShuffleMask),
1997 // Get a mask to interleave \p NumVec vectors into a wide vector.
1998 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
1999 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2000 // <0, 4, 1, 5, 2, 6, 3, 7>
2001 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2003 SmallVector<Constant *, 16> Mask;
2004 for (unsigned i = 0; i < VF; i++)
2005 for (unsigned j = 0; j < NumVec; j++)
2006 Mask.push_back(Builder.getInt32(j * VF + i));
2008 return ConstantVector::get(Mask);
2011 // Get the strided mask starting from index \p Start.
2012 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2013 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2014 unsigned Stride, unsigned VF) {
2015 SmallVector<Constant *, 16> Mask;
2016 for (unsigned i = 0; i < VF; i++)
2017 Mask.push_back(Builder.getInt32(Start + i * Stride));
2019 return ConstantVector::get(Mask);
2022 // Get a mask of two parts: The first part consists of sequential integers
2023 // starting from 0, The second part consists of UNDEFs.
2024 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2025 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2026 unsigned NumUndef) {
2027 SmallVector<Constant *, 16> Mask;
2028 for (unsigned i = 0; i < NumInt; i++)
2029 Mask.push_back(Builder.getInt32(i));
2031 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2032 for (unsigned i = 0; i < NumUndef; i++)
2033 Mask.push_back(Undef);
2035 return ConstantVector::get(Mask);
2038 // Concatenate two vectors with the same element type. The 2nd vector should
2039 // not have more elements than the 1st vector. If the 2nd vector has less
2040 // elements, extend it with UNDEFs.
2041 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2043 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2044 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2045 assert(VecTy1 && VecTy2 &&
2046 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2047 "Expect two vectors with the same element type");
2049 unsigned NumElts1 = VecTy1->getNumElements();
2050 unsigned NumElts2 = VecTy2->getNumElements();
2051 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2053 if (NumElts1 > NumElts2) {
2054 // Extend with UNDEFs.
2056 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2057 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2060 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2061 return Builder.CreateShuffleVector(V1, V2, Mask);
2064 // Concatenate vectors in the given list. All vectors have the same type.
2065 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2066 ArrayRef<Value *> InputList) {
2067 unsigned NumVec = InputList.size();
2068 assert(NumVec > 1 && "Should be at least two vectors");
2070 SmallVector<Value *, 8> ResList;
2071 ResList.append(InputList.begin(), InputList.end());
2073 SmallVector<Value *, 8> TmpList;
2074 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2075 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2076 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2077 "Only the last vector may have a different type");
2079 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2082 // Push the last vector if the total number of vectors is odd.
2083 if (NumVec % 2 != 0)
2084 TmpList.push_back(ResList[NumVec - 1]);
2087 NumVec = ResList.size();
2088 } while (NumVec > 1);
2093 // Try to vectorize the interleave group that \p Instr belongs to.
2095 // E.g. Translate following interleaved load group (factor = 3):
2096 // for (i = 0; i < N; i+=3) {
2097 // R = Pic[i]; // Member of index 0
2098 // G = Pic[i+1]; // Member of index 1
2099 // B = Pic[i+2]; // Member of index 2
2100 // ... // do something to R, G, B
2103 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2104 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2105 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2106 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2108 // Or translate following interleaved store group (factor = 3):
2109 // for (i = 0; i < N; i+=3) {
2110 // ... do something to R, G, B
2111 // Pic[i] = R; // Member of index 0
2112 // Pic[i+1] = G; // Member of index 1
2113 // Pic[i+2] = B; // Member of index 2
2116 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2117 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2118 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2119 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2120 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2121 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2122 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2123 assert(Group && "Fail to get an interleaved access group.");
2125 // Skip if current instruction is not the insert position.
2126 if (Instr != Group->getInsertPos())
2129 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2130 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2131 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2133 // Prepare for the vector type of the interleaved load/store.
2134 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2135 unsigned InterleaveFactor = Group->getFactor();
2136 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2137 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2139 // Prepare for the new pointers.
2140 setDebugLocFromInst(Builder, Ptr);
2141 VectorParts &PtrParts = getVectorValue(Ptr);
2142 SmallVector<Value *, 2> NewPtrs;
2143 unsigned Index = Group->getIndex(Instr);
2144 for (unsigned Part = 0; Part < UF; Part++) {
2145 // Extract the pointer for current instruction from the pointer vector. A
2146 // reverse access uses the pointer in the last lane.
2147 Value *NewPtr = Builder.CreateExtractElement(
2149 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2151 // Notice current instruction could be any index. Need to adjust the address
2152 // to the member of index 0.
2154 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2155 // b = A[i]; // Member of index 0
2156 // Current pointer is pointed to A[i+1], adjust it to A[i].
2158 // E.g. A[i+1] = a; // Member of index 1
2159 // A[i] = b; // Member of index 0
2160 // A[i+2] = c; // Member of index 2 (Current instruction)
2161 // Current pointer is pointed to A[i+2], adjust it to A[i].
2162 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2164 // Cast to the vector pointer type.
2165 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2168 setDebugLocFromInst(Builder, Instr);
2169 Value *UndefVec = UndefValue::get(VecTy);
2171 // Vectorize the interleaved load group.
2173 for (unsigned Part = 0; Part < UF; Part++) {
2174 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2175 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2177 for (unsigned i = 0; i < InterleaveFactor; i++) {
2178 Instruction *Member = Group->getMember(i);
2180 // Skip the gaps in the group.
2184 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2185 Value *StridedVec = Builder.CreateShuffleVector(
2186 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2188 // If this member has different type, cast the result type.
2189 if (Member->getType() != ScalarTy) {
2190 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2191 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2194 VectorParts &Entry = WidenMap.get(Member);
2196 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2199 propagateMetadata(NewLoadInstr, Instr);
2204 // The sub vector type for current instruction.
2205 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2207 // Vectorize the interleaved store group.
2208 for (unsigned Part = 0; Part < UF; Part++) {
2209 // Collect the stored vector from each member.
2210 SmallVector<Value *, 4> StoredVecs;
2211 for (unsigned i = 0; i < InterleaveFactor; i++) {
2212 // Interleaved store group doesn't allow a gap, so each index has a member
2213 Instruction *Member = Group->getMember(i);
2214 assert(Member && "Fail to get a member from an interleaved store group");
2217 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2218 if (Group->isReverse())
2219 StoredVec = reverseVector(StoredVec);
2221 // If this member has different type, cast it to an unified type.
2222 if (StoredVec->getType() != SubVT)
2223 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2225 StoredVecs.push_back(StoredVec);
2228 // Concatenate all vectors into a wide vector.
2229 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2231 // Interleave the elements in the wide vector.
2232 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2233 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2236 Instruction *NewStoreInstr =
2237 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2238 propagateMetadata(NewStoreInstr, Instr);
2242 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2243 // Attempt to issue a wide load.
2244 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2245 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2247 assert((LI || SI) && "Invalid Load/Store instruction");
2249 // Try to vectorize the interleave group if this access is interleaved.
2250 if (Legal->isAccessInterleaved(Instr))
2251 return vectorizeInterleaveGroup(Instr);
2253 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2254 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2255 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2256 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2257 // An alignment of 0 means target abi alignment. We need to use the scalar's
2258 // target abi alignment in such a case.
2259 const DataLayout &DL = Instr->getModule()->getDataLayout();
2261 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2262 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2263 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2264 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2266 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2267 !Legal->isMaskRequired(SI))
2268 return scalarizeInstruction(Instr, true);
2270 if (ScalarAllocatedSize != VectorElementSize)
2271 return scalarizeInstruction(Instr);
2273 // If the pointer is loop invariant or if it is non-consecutive,
2274 // scalarize the load.
2275 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2276 bool Reverse = ConsecutiveStride < 0;
2277 bool UniformLoad = LI && Legal->isUniform(Ptr);
2278 if (!ConsecutiveStride || UniformLoad)
2279 return scalarizeInstruction(Instr);
2281 Constant *Zero = Builder.getInt32(0);
2282 VectorParts &Entry = WidenMap.get(Instr);
2284 // Handle consecutive loads/stores.
2285 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2286 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2287 setDebugLocFromInst(Builder, Gep);
2288 Value *PtrOperand = Gep->getPointerOperand();
2289 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2290 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2292 // Create the new GEP with the new induction variable.
2293 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2294 Gep2->setOperand(0, FirstBasePtr);
2295 Gep2->setName("gep.indvar.base");
2296 Ptr = Builder.Insert(Gep2);
2298 setDebugLocFromInst(Builder, Gep);
2299 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2300 OrigLoop) && "Base ptr must be invariant");
2302 // The last index does not have to be the induction. It can be
2303 // consecutive and be a function of the index. For example A[I+1];
2304 unsigned NumOperands = Gep->getNumOperands();
2305 unsigned InductionOperand = getGEPInductionOperand(Gep);
2306 // Create the new GEP with the new induction variable.
2307 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2309 for (unsigned i = 0; i < NumOperands; ++i) {
2310 Value *GepOperand = Gep->getOperand(i);
2311 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2313 // Update last index or loop invariant instruction anchored in loop.
2314 if (i == InductionOperand ||
2315 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2316 assert((i == InductionOperand ||
2317 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2318 "Must be last index or loop invariant");
2320 VectorParts &GEPParts = getVectorValue(GepOperand);
2321 Value *Index = GEPParts[0];
2322 Index = Builder.CreateExtractElement(Index, Zero);
2323 Gep2->setOperand(i, Index);
2324 Gep2->setName("gep.indvar.idx");
2327 Ptr = Builder.Insert(Gep2);
2329 // Use the induction element ptr.
2330 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2331 setDebugLocFromInst(Builder, Ptr);
2332 VectorParts &PtrVal = getVectorValue(Ptr);
2333 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2336 VectorParts Mask = createBlockInMask(Instr->getParent());
2339 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2340 "We do not allow storing to uniform addresses");
2341 setDebugLocFromInst(Builder, SI);
2342 // We don't want to update the value in the map as it might be used in
2343 // another expression. So don't use a reference type for "StoredVal".
2344 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2346 for (unsigned Part = 0; Part < UF; ++Part) {
2347 // Calculate the pointer for the specific unroll-part.
2349 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2352 // If we store to reverse consecutive memory locations, then we need
2353 // to reverse the order of elements in the stored value.
2354 StoredVal[Part] = reverseVector(StoredVal[Part]);
2355 // If the address is consecutive but reversed, then the
2356 // wide store needs to start at the last vector element.
2357 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2358 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2359 Mask[Part] = reverseVector(Mask[Part]);
2362 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2363 DataTy->getPointerTo(AddressSpace));
2366 if (Legal->isMaskRequired(SI))
2367 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2370 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2371 propagateMetadata(NewSI, SI);
2377 assert(LI && "Must have a load instruction");
2378 setDebugLocFromInst(Builder, LI);
2379 for (unsigned Part = 0; Part < UF; ++Part) {
2380 // Calculate the pointer for the specific unroll-part.
2382 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2385 // If the address is consecutive but reversed, then the
2386 // wide load needs to start at the last vector element.
2387 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2388 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2389 Mask[Part] = reverseVector(Mask[Part]);
2393 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2394 DataTy->getPointerTo(AddressSpace));
2395 if (Legal->isMaskRequired(LI))
2396 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2397 UndefValue::get(DataTy),
2398 "wide.masked.load");
2400 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2401 propagateMetadata(NewLI, LI);
2402 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2406 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2407 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2408 // Holds vector parameters or scalars, in case of uniform vals.
2409 SmallVector<VectorParts, 4> Params;
2411 setDebugLocFromInst(Builder, Instr);
2413 // Find all of the vectorized parameters.
2414 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2415 Value *SrcOp = Instr->getOperand(op);
2417 // If we are accessing the old induction variable, use the new one.
2418 if (SrcOp == OldInduction) {
2419 Params.push_back(getVectorValue(SrcOp));
2423 // Try using previously calculated values.
2424 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2426 // If the src is an instruction that appeared earlier in the basic block,
2427 // then it should already be vectorized.
2428 if (SrcInst && OrigLoop->contains(SrcInst)) {
2429 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2430 // The parameter is a vector value from earlier.
2431 Params.push_back(WidenMap.get(SrcInst));
2433 // The parameter is a scalar from outside the loop. Maybe even a constant.
2434 VectorParts Scalars;
2435 Scalars.append(UF, SrcOp);
2436 Params.push_back(Scalars);
2440 assert(Params.size() == Instr->getNumOperands() &&
2441 "Invalid number of operands");
2443 // Does this instruction return a value ?
2444 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2446 Value *UndefVec = IsVoidRetTy ? nullptr :
2447 UndefValue::get(VectorType::get(Instr->getType(), VF));
2448 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2449 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2451 Instruction *InsertPt = Builder.GetInsertPoint();
2452 BasicBlock *IfBlock = Builder.GetInsertBlock();
2453 BasicBlock *CondBlock = nullptr;
2456 Loop *VectorLp = nullptr;
2457 if (IfPredicateStore) {
2458 assert(Instr->getParent()->getSinglePredecessor() &&
2459 "Only support single predecessor blocks");
2460 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2461 Instr->getParent());
2462 VectorLp = LI->getLoopFor(IfBlock);
2463 assert(VectorLp && "Must have a loop for this block");
2466 // For each vector unroll 'part':
2467 for (unsigned Part = 0; Part < UF; ++Part) {
2468 // For each scalar that we create:
2469 for (unsigned Width = 0; Width < VF; ++Width) {
2472 Value *Cmp = nullptr;
2473 if (IfPredicateStore) {
2474 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2475 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2476 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2477 LoopVectorBody.push_back(CondBlock);
2478 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2479 // Update Builder with newly created basic block.
2480 Builder.SetInsertPoint(InsertPt);
2483 Instruction *Cloned = Instr->clone();
2485 Cloned->setName(Instr->getName() + ".cloned");
2486 // Replace the operands of the cloned instructions with extracted scalars.
2487 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2488 Value *Op = Params[op][Part];
2489 // Param is a vector. Need to extract the right lane.
2490 if (Op->getType()->isVectorTy())
2491 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2492 Cloned->setOperand(op, Op);
2495 // Place the cloned scalar in the new loop.
2496 Builder.Insert(Cloned);
2498 // If the original scalar returns a value we need to place it in a vector
2499 // so that future users will be able to use it.
2501 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2502 Builder.getInt32(Width));
2504 if (IfPredicateStore) {
2505 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2506 LoopVectorBody.push_back(NewIfBlock);
2507 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2508 Builder.SetInsertPoint(InsertPt);
2509 ReplaceInstWithInst(IfBlock->getTerminator(),
2510 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2511 IfBlock = NewIfBlock;
2517 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2521 if (Instruction *I = dyn_cast<Instruction>(V))
2522 return I->getParent() == Loc->getParent() ? I : nullptr;
2526 std::pair<Instruction *, Instruction *>
2527 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2528 Instruction *tnullptr = nullptr;
2529 if (!Legal->mustCheckStrides())
2530 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2532 IRBuilder<> ChkBuilder(Loc);
2535 Value *Check = nullptr;
2536 Instruction *FirstInst = nullptr;
2537 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2538 SE = Legal->strides_end();
2540 Value *Ptr = stripIntegerCast(*SI);
2541 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2543 // Store the first instruction we create.
2544 FirstInst = getFirstInst(FirstInst, C, Loc);
2546 Check = ChkBuilder.CreateOr(Check, C);
2551 // We have to do this trickery because the IRBuilder might fold the check to a
2552 // constant expression in which case there is no Instruction anchored in a
2554 LLVMContext &Ctx = Loc->getContext();
2555 Instruction *TheCheck =
2556 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2557 ChkBuilder.Insert(TheCheck, "stride.not.one");
2558 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2560 return std::make_pair(FirstInst, TheCheck);
2563 void InnerLoopVectorizer::createEmptyLoop() {
2565 In this function we generate a new loop. The new loop will contain
2566 the vectorized instructions while the old loop will continue to run the
2569 [ ] <-- loop iteration number check.
2572 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2575 || [ ] <-- vector pre header.
2579 || [ ]_| <-- vector loop.
2582 | >[ ] <--- middle-block.
2585 -|- >[ ] <--- new preheader.
2589 | [ ]_| <-- old scalar loop to handle remainder.
2592 >[ ] <-- exit block.
2596 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2597 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2598 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2599 assert(VectorPH && "Invalid loop structure");
2600 assert(ExitBlock && "Must have an exit block");
2602 // Some loops have a single integer induction variable, while other loops
2603 // don't. One example is c++ iterators that often have multiple pointer
2604 // induction variables. In the code below we also support a case where we
2605 // don't have a single induction variable.
2607 // We try to obtain an induction variable from the original loop as hard
2608 // as possible. However if we don't find one that:
2610 // - counts from zero, stepping by one
2611 // - is the size of the widest induction variable type
2612 // then we create a new one.
2613 OldInduction = Legal->getInduction();
2614 Type *IdxTy = Legal->getWidestInductionType();
2616 // Find the loop boundaries.
2617 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2618 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2620 // The exit count might have the type of i64 while the phi is i32. This can
2621 // happen if we have an induction variable that is sign extended before the
2622 // compare. The only way that we get a backedge taken count is that the
2623 // induction variable was signed and as such will not overflow. In such a case
2624 // truncation is legal.
2625 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2626 IdxTy->getPrimitiveSizeInBits())
2627 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2629 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2630 // Get the total trip count from the count by adding 1.
2631 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2632 SE->getConstant(BackedgeTakeCount->getType(), 1));
2634 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2636 // Expand the trip count and place the new instructions in the preheader.
2637 // Notice that the pre-header does not change, only the loop body.
2638 SCEVExpander Exp(*SE, DL, "induction");
2640 // The loop minimum iterations check below is to ensure the loop has enough
2641 // trip count so the generated vector loop will likely be executed and the
2642 // preparation and rounding-off costs will likely be worthy.
2644 // The minimum iteration check also covers case where the backedge-taken
2645 // count is uint##_max. Adding one to it will cause overflow and an
2646 // incorrect loop trip count being generated in the vector body. In this
2647 // case we also want to directly jump to the scalar remainder loop.
2648 Value *ExitCountValue = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2649 VectorPH->getTerminator());
2650 if (ExitCountValue->getType()->isPointerTy())
2651 ExitCountValue = CastInst::CreatePointerCast(ExitCountValue, IdxTy,
2652 "exitcount.ptrcnt.to.int",
2653 VectorPH->getTerminator());
2655 Instruction *CheckMinIters =
2656 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULT, ExitCountValue,
2657 ConstantInt::get(ExitCountValue->getType(), VF * UF),
2658 "min.iters.check", VectorPH->getTerminator());
2660 Builder.SetInsertPoint(VectorPH->getTerminator());
2661 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2663 LoopBypassBlocks.push_back(VectorPH);
2665 // Split the single block loop into the two loop structure described above.
2666 BasicBlock *VecBody =
2667 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2668 BasicBlock *MiddleBlock =
2669 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2670 BasicBlock *ScalarPH =
2671 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2673 // Create and register the new vector loop.
2674 Loop* Lp = new Loop();
2675 Loop *ParentLoop = OrigLoop->getParentLoop();
2677 // Insert the new loop into the loop nest and register the new basic blocks
2678 // before calling any utilities such as SCEV that require valid LoopInfo.
2680 ParentLoop->addChildLoop(Lp);
2681 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2682 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2684 LI->addTopLevelLoop(Lp);
2686 Lp->addBasicBlockToLoop(VecBody, *LI);
2688 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2690 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2692 // Generate the induction variable.
2693 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2694 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2695 // The loop step is equal to the vectorization factor (num of SIMD elements)
2696 // times the unroll factor (num of SIMD instructions).
2697 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2699 // Generate code to check that the loop's trip count is not less than the
2700 // minimum loop iteration number threshold.
2701 BasicBlock *NewVectorPH =
2702 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "min.iters.checked");
2704 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2705 ReplaceInstWithInst(VectorPH->getTerminator(),
2706 BranchInst::Create(ScalarPH, NewVectorPH, CheckMinIters));
2707 VectorPH = NewVectorPH;
2709 // This is the IR builder that we use to add all of the logic for bypassing
2710 // the new vector loop.
2711 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2712 setDebugLocFromInst(BypassBuilder,
2713 getDebugLocFromInstOrOperands(OldInduction));
2715 // Add the start index to the loop count to get the new end index.
2716 Value *IdxEnd = BypassBuilder.CreateAdd(ExitCountValue, StartIdx, "end.idx");
2718 // Now we need to generate the expression for N - (N % VF), which is
2719 // the part that the vectorized body will execute.
2720 Value *R = BypassBuilder.CreateURem(ExitCountValue, Step, "n.mod.vf");
2721 Value *CountRoundDown = BypassBuilder.CreateSub(ExitCountValue, R, "n.vec");
2722 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2723 "end.idx.rnd.down");
2725 // Now, compare the new count to zero. If it is zero skip the vector loop and
2726 // jump to the scalar loop.
2728 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2730 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2732 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2733 LoopBypassBlocks.push_back(VectorPH);
2734 ReplaceInstWithInst(VectorPH->getTerminator(),
2735 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2736 VectorPH = NewVectorPH;
2738 // Generate the code to check that the strides we assumed to be one are really
2739 // one. We want the new basic block to start at the first instruction in a
2740 // sequence of instructions that form a check.
2741 Instruction *StrideCheck;
2742 Instruction *FirstCheckInst;
2743 std::tie(FirstCheckInst, StrideCheck) =
2744 addStrideCheck(VectorPH->getTerminator());
2746 AddedSafetyChecks = true;
2747 // Create a new block containing the stride check.
2748 VectorPH->setName("vector.stridecheck");
2750 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2752 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2753 LoopBypassBlocks.push_back(VectorPH);
2755 // Replace the branch into the memory check block with a conditional branch
2756 // for the "few elements case".
2757 ReplaceInstWithInst(
2758 VectorPH->getTerminator(),
2759 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2761 VectorPH = NewVectorPH;
2764 // Generate the code that checks in runtime if arrays overlap. We put the
2765 // checks into a separate block to make the more common case of few elements
2767 Instruction *MemRuntimeCheck;
2768 std::tie(FirstCheckInst, MemRuntimeCheck) =
2769 Legal->getLAI()->addRuntimeChecks(VectorPH->getTerminator());
2770 if (MemRuntimeCheck) {
2771 AddedSafetyChecks = true;
2772 // Create a new block containing the memory check.
2773 VectorPH->setName("vector.memcheck");
2775 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2777 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2778 LoopBypassBlocks.push_back(VectorPH);
2780 // Replace the branch into the memory check block with a conditional branch
2781 // for the "few elements case".
2782 ReplaceInstWithInst(
2783 VectorPH->getTerminator(),
2784 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2786 VectorPH = NewVectorPH;
2789 // We are going to resume the execution of the scalar loop.
2790 // Go over all of the induction variables that we found and fix the
2791 // PHIs that are left in the scalar version of the loop.
2792 // The starting values of PHI nodes depend on the counter of the last
2793 // iteration in the vectorized loop.
2794 // If we come from a bypass edge then we need to start from the original
2797 // This variable saves the new starting index for the scalar loop. It is used
2798 // to test if there are any tail iterations left once the vector loop has
2800 PHINode *ResumeIndex = nullptr;
2801 LoopVectorizationLegality::InductionList::iterator I, E;
2802 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2803 // Set builder to point to last bypass block.
2804 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2805 for (I = List->begin(), E = List->end(); I != E; ++I) {
2806 PHINode *OrigPhi = I->first;
2807 InductionDescriptor II = I->second;
2809 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
2810 MiddleBlock->getTerminator());
2811 // Create phi nodes to merge from the backedge-taken check block.
2812 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2814 ScalarPH->getTerminator());
2815 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2818 if (OrigPhi == OldInduction) {
2819 // We know what the end value is.
2820 EndValue = IdxEndRoundDown;
2821 // We also know which PHI node holds it.
2822 ResumeIndex = ResumeVal;
2824 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2825 II.getStepValue()->getType(),
2827 EndValue = II.transform(BypassBuilder, CRD);
2828 EndValue->setName("ind.end");
2831 // The new PHI merges the original incoming value, in case of a bypass,
2832 // or the value at the end of the vectorized loop.
2833 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2834 ResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2835 ResumeVal->addIncoming(EndValue, VecBody);
2837 // Fix the scalar body counter (PHI node).
2838 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2840 // The old induction's phi node in the scalar body needs the truncated
2842 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[0]);
2843 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2846 // If we are generating a new induction variable then we also need to
2847 // generate the code that calculates the exit value. This value is not
2848 // simply the end of the counter because we may skip the vectorized body
2849 // in case of a runtime check.
2851 assert(!ResumeIndex && "Unexpected resume value found");
2852 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2853 MiddleBlock->getTerminator());
2854 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2855 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2856 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2859 // Make sure that we found the index where scalar loop needs to continue.
2860 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2861 "Invalid resume Index");
2863 // Add a check in the middle block to see if we have completed
2864 // all of the iterations in the first vector loop.
2865 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2866 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2867 ResumeIndex, "cmp.n",
2868 MiddleBlock->getTerminator());
2869 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2870 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2872 // Create i+1 and fill the PHINode.
2873 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2874 Induction->addIncoming(StartIdx, VectorPH);
2875 Induction->addIncoming(NextIdx, VecBody);
2876 // Create the compare.
2877 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2878 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2880 // Now we have two terminators. Remove the old one from the block.
2881 VecBody->getTerminator()->eraseFromParent();
2883 // Get ready to start creating new instructions into the vectorized body.
2884 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2887 LoopVectorPreHeader = VectorPH;
2888 LoopScalarPreHeader = ScalarPH;
2889 LoopMiddleBlock = MiddleBlock;
2890 LoopExitBlock = ExitBlock;
2891 LoopVectorBody.push_back(VecBody);
2892 LoopScalarBody = OldBasicBlock;
2894 LoopVectorizeHints Hints(Lp, true);
2895 Hints.setAlreadyVectorized();
2899 struct CSEDenseMapInfo {
2900 static bool canHandle(Instruction *I) {
2901 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2902 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2904 static inline Instruction *getEmptyKey() {
2905 return DenseMapInfo<Instruction *>::getEmptyKey();
2907 static inline Instruction *getTombstoneKey() {
2908 return DenseMapInfo<Instruction *>::getTombstoneKey();
2910 static unsigned getHashValue(Instruction *I) {
2911 assert(canHandle(I) && "Unknown instruction!");
2912 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2913 I->value_op_end()));
2915 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2916 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2917 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2919 return LHS->isIdenticalTo(RHS);
2924 /// \brief Check whether this block is a predicated block.
2925 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2926 /// = ...; " blocks. We start with one vectorized basic block. For every
2927 /// conditional block we split this vectorized block. Therefore, every second
2928 /// block will be a predicated one.
2929 static bool isPredicatedBlock(unsigned BlockNum) {
2930 return BlockNum % 2;
2933 ///\brief Perform cse of induction variable instructions.
2934 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2935 // Perform simple cse.
2936 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2937 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2938 BasicBlock *BB = BBs[i];
2939 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2940 Instruction *In = I++;
2942 if (!CSEDenseMapInfo::canHandle(In))
2945 // Check if we can replace this instruction with any of the
2946 // visited instructions.
2947 if (Instruction *V = CSEMap.lookup(In)) {
2948 In->replaceAllUsesWith(V);
2949 In->eraseFromParent();
2952 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2953 // ...;" blocks for predicated stores. Every second block is a predicated
2955 if (isPredicatedBlock(i))
2963 /// \brief Adds a 'fast' flag to floating point operations.
2964 static Value *addFastMathFlag(Value *V) {
2965 if (isa<FPMathOperator>(V)){
2966 FastMathFlags Flags;
2967 Flags.setUnsafeAlgebra();
2968 cast<Instruction>(V)->setFastMathFlags(Flags);
2973 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
2974 /// the result needs to be inserted and/or extracted from vectors.
2975 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
2976 const TargetTransformInfo &TTI) {
2980 assert(Ty->isVectorTy() && "Can only scalarize vectors");
2983 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
2985 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
2987 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
2993 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
2994 // Return the cost of the instruction, including scalarization overhead if it's
2995 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
2996 // i.e. either vector version isn't available, or is too expensive.
2997 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
2998 const TargetTransformInfo &TTI,
2999 const TargetLibraryInfo *TLI,
3000 bool &NeedToScalarize) {
3001 Function *F = CI->getCalledFunction();
3002 StringRef FnName = CI->getCalledFunction()->getName();
3003 Type *ScalarRetTy = CI->getType();
3004 SmallVector<Type *, 4> Tys, ScalarTys;
3005 for (auto &ArgOp : CI->arg_operands())
3006 ScalarTys.push_back(ArgOp->getType());
3008 // Estimate cost of scalarized vector call. The source operands are assumed
3009 // to be vectors, so we need to extract individual elements from there,
3010 // execute VF scalar calls, and then gather the result into the vector return
3012 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3014 return ScalarCallCost;
3016 // Compute corresponding vector type for return value and arguments.
3017 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3018 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3019 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3021 // Compute costs of unpacking argument values for the scalar calls and
3022 // packing the return values to a vector.
3023 unsigned ScalarizationCost =
3024 getScalarizationOverhead(RetTy, true, false, TTI);
3025 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3026 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3028 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3030 // If we can't emit a vector call for this function, then the currently found
3031 // cost is the cost we need to return.
3032 NeedToScalarize = true;
3033 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3036 // If the corresponding vector cost is cheaper, return its cost.
3037 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3038 if (VectorCallCost < Cost) {
3039 NeedToScalarize = false;
3040 return VectorCallCost;
3045 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3046 // factor VF. Return the cost of the instruction, including scalarization
3047 // overhead if it's needed.
3048 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3049 const TargetTransformInfo &TTI,
3050 const TargetLibraryInfo *TLI) {
3051 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3052 assert(ID && "Expected intrinsic call!");
3054 Type *RetTy = ToVectorTy(CI->getType(), VF);
3055 SmallVector<Type *, 4> Tys;
3056 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3057 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3059 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3062 void InnerLoopVectorizer::vectorizeLoop() {
3063 //===------------------------------------------------===//
3065 // Notice: any optimization or new instruction that go
3066 // into the code below should be also be implemented in
3069 //===------------------------------------------------===//
3070 Constant *Zero = Builder.getInt32(0);
3072 // In order to support reduction variables we need to be able to vectorize
3073 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3074 // stages. First, we create a new vector PHI node with no incoming edges.
3075 // We use this value when we vectorize all of the instructions that use the
3076 // PHI. Next, after all of the instructions in the block are complete we
3077 // add the new incoming edges to the PHI. At this point all of the
3078 // instructions in the basic block are vectorized, so we can use them to
3079 // construct the PHI.
3080 PhiVector RdxPHIsToFix;
3082 // Scan the loop in a topological order to ensure that defs are vectorized
3084 LoopBlocksDFS DFS(OrigLoop);
3087 // Vectorize all of the blocks in the original loop.
3088 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3089 be = DFS.endRPO(); bb != be; ++bb)
3090 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3092 // At this point every instruction in the original loop is widened to
3093 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3094 // that we vectorized. The PHI nodes are currently empty because we did
3095 // not want to introduce cycles. Notice that the remaining PHI nodes
3096 // that we need to fix are reduction variables.
3098 // Create the 'reduced' values for each of the induction vars.
3099 // The reduced values are the vector values that we scalarize and combine
3100 // after the loop is finished.
3101 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3103 PHINode *RdxPhi = *it;
3104 assert(RdxPhi && "Unable to recover vectorized PHI");
3106 // Find the reduction variable descriptor.
3107 assert(Legal->getReductionVars()->count(RdxPhi) &&
3108 "Unable to find the reduction variable");
3109 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3111 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3112 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3113 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3114 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3115 RdxDesc.getMinMaxRecurrenceKind();
3116 setDebugLocFromInst(Builder, ReductionStartValue);
3118 // We need to generate a reduction vector from the incoming scalar.
3119 // To do so, we need to generate the 'identity' vector and override
3120 // one of the elements with the incoming scalar reduction. We need
3121 // to do it in the vector-loop preheader.
3122 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3124 // This is the vector-clone of the value that leaves the loop.
3125 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3126 Type *VecTy = VectorExit[0]->getType();
3128 // Find the reduction identity variable. Zero for addition, or, xor,
3129 // one for multiplication, -1 for And.
3132 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3133 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3134 // MinMax reduction have the start value as their identify.
3136 VectorStart = Identity = ReductionStartValue;
3138 VectorStart = Identity =
3139 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3142 // Handle other reduction kinds:
3143 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3144 RK, VecTy->getScalarType());
3147 // This vector is the Identity vector where the first element is the
3148 // incoming scalar reduction.
3149 VectorStart = ReductionStartValue;
3151 Identity = ConstantVector::getSplat(VF, Iden);
3153 // This vector is the Identity vector where the first element is the
3154 // incoming scalar reduction.
3156 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3160 // Fix the vector-loop phi.
3162 // Reductions do not have to start at zero. They can start with
3163 // any loop invariant values.
3164 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3165 BasicBlock *Latch = OrigLoop->getLoopLatch();
3166 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3167 VectorParts &Val = getVectorValue(LoopVal);
3168 for (unsigned part = 0; part < UF; ++part) {
3169 // Make sure to add the reduction stat value only to the
3170 // first unroll part.
3171 Value *StartVal = (part == 0) ? VectorStart : Identity;
3172 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3173 LoopVectorPreHeader);
3174 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3175 LoopVectorBody.back());
3178 // Before each round, move the insertion point right between
3179 // the PHIs and the values we are going to write.
3180 // This allows us to write both PHINodes and the extractelement
3182 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3184 VectorParts RdxParts, &RdxExitVal = getVectorValue(LoopExitInst);
3185 setDebugLocFromInst(Builder, LoopExitInst);
3186 for (unsigned part = 0; part < UF; ++part) {
3187 // This PHINode contains the vectorized reduction variable, or
3188 // the initial value vector, if we bypass the vector loop.
3189 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3190 Value *StartVal = (part == 0) ? VectorStart : Identity;
3191 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3192 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3193 NewPhi->addIncoming(RdxExitVal[part],
3194 LoopVectorBody.back());
3195 RdxParts.push_back(NewPhi);
3198 // If the vector reduction can be performed in a smaller type, we truncate
3199 // then extend the loop exit value to enable InstCombine to evaluate the
3200 // entire expression in the smaller type.
3201 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3202 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3203 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3204 for (unsigned part = 0; part < UF; ++part) {
3205 Value *Trunc = Builder.CreateTrunc(RdxExitVal[part], RdxVecTy);
3206 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3207 : Builder.CreateZExt(Trunc, VecTy);
3208 for (Value::user_iterator UI = RdxExitVal[part]->user_begin();
3209 UI != RdxExitVal[part]->user_end();)
3211 (*UI++)->replaceUsesOfWith(RdxExitVal[part], Extnd);
3215 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3216 for (unsigned part = 0; part < UF; ++part)
3217 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3220 // Reduce all of the unrolled parts into a single vector.
3221 Value *ReducedPartRdx = RdxParts[0];
3222 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3223 setDebugLocFromInst(Builder, ReducedPartRdx);
3224 for (unsigned part = 1; part < UF; ++part) {
3225 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3226 // Floating point operations had to be 'fast' to enable the reduction.
3227 ReducedPartRdx = addFastMathFlag(
3228 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3229 ReducedPartRdx, "bin.rdx"));
3231 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3232 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3236 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3237 // and vector ops, reducing the set of values being computed by half each
3239 assert(isPowerOf2_32(VF) &&
3240 "Reduction emission only supported for pow2 vectors!");
3241 Value *TmpVec = ReducedPartRdx;
3242 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3243 for (unsigned i = VF; i != 1; i >>= 1) {
3244 // Move the upper half of the vector to the lower half.
3245 for (unsigned j = 0; j != i/2; ++j)
3246 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3248 // Fill the rest of the mask with undef.
3249 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3250 UndefValue::get(Builder.getInt32Ty()));
3253 Builder.CreateShuffleVector(TmpVec,
3254 UndefValue::get(TmpVec->getType()),
3255 ConstantVector::get(ShuffleMask),
3258 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3259 // Floating point operations had to be 'fast' to enable the reduction.
3260 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3261 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3263 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3267 // The result is in the first element of the vector.
3268 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3269 Builder.getInt32(0));
3271 // If the reduction can be performed in a smaller type, we need to extend
3272 // the reduction to the wider type before we branch to the original loop.
3273 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3276 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3277 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3280 // Create a phi node that merges control-flow from the backedge-taken check
3281 // block and the middle block.
3282 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3283 LoopScalarPreHeader->getTerminator());
3284 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3285 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3287 // Now, we need to fix the users of the reduction variable
3288 // inside and outside of the scalar remainder loop.
3289 // We know that the loop is in LCSSA form. We need to update the
3290 // PHI nodes in the exit blocks.
3291 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3292 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3293 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3294 if (!LCSSAPhi) break;
3296 // All PHINodes need to have a single entry edge, or two if
3297 // we already fixed them.
3298 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3300 // We found our reduction value exit-PHI. Update it with the
3301 // incoming bypass edge.
3302 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3303 // Add an edge coming from the bypass.
3304 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3307 }// end of the LCSSA phi scan.
3309 // Fix the scalar loop reduction variable with the incoming reduction sum
3310 // from the vector body and from the backedge value.
3311 int IncomingEdgeBlockIdx =
3312 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3313 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3314 // Pick the other block.
3315 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3316 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3317 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3318 }// end of for each redux variable.
3322 // Remove redundant induction instructions.
3323 cse(LoopVectorBody);
3326 void InnerLoopVectorizer::fixLCSSAPHIs() {
3327 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3328 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3329 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3330 if (!LCSSAPhi) break;
3331 if (LCSSAPhi->getNumIncomingValues() == 1)
3332 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3337 InnerLoopVectorizer::VectorParts
3338 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3339 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3342 // Look for cached value.
3343 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3344 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3345 if (ECEntryIt != MaskCache.end())
3346 return ECEntryIt->second;
3348 VectorParts SrcMask = createBlockInMask(Src);
3350 // The terminator has to be a branch inst!
3351 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3352 assert(BI && "Unexpected terminator found");
3354 if (BI->isConditional()) {
3355 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3357 if (BI->getSuccessor(0) != Dst)
3358 for (unsigned part = 0; part < UF; ++part)
3359 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3361 for (unsigned part = 0; part < UF; ++part)
3362 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3364 MaskCache[Edge] = EdgeMask;
3368 MaskCache[Edge] = SrcMask;
3372 InnerLoopVectorizer::VectorParts
3373 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3374 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3376 // Loop incoming mask is all-one.
3377 if (OrigLoop->getHeader() == BB) {
3378 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3379 return getVectorValue(C);
3382 // This is the block mask. We OR all incoming edges, and with zero.
3383 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3384 VectorParts BlockMask = getVectorValue(Zero);
3387 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3388 VectorParts EM = createEdgeMask(*it, BB);
3389 for (unsigned part = 0; part < UF; ++part)
3390 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3396 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3397 InnerLoopVectorizer::VectorParts &Entry,
3398 unsigned UF, unsigned VF, PhiVector *PV) {
3399 PHINode* P = cast<PHINode>(PN);
3400 // Handle reduction variables:
3401 if (Legal->getReductionVars()->count(P)) {
3402 for (unsigned part = 0; part < UF; ++part) {
3403 // This is phase one of vectorizing PHIs.
3404 Type *VecTy = (VF == 1) ? PN->getType() :
3405 VectorType::get(PN->getType(), VF);
3406 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3407 LoopVectorBody.back()-> getFirstInsertionPt());
3413 setDebugLocFromInst(Builder, P);
3414 // Check for PHI nodes that are lowered to vector selects.
3415 if (P->getParent() != OrigLoop->getHeader()) {
3416 // We know that all PHIs in non-header blocks are converted into
3417 // selects, so we don't have to worry about the insertion order and we
3418 // can just use the builder.
3419 // At this point we generate the predication tree. There may be
3420 // duplications since this is a simple recursive scan, but future
3421 // optimizations will clean it up.
3423 unsigned NumIncoming = P->getNumIncomingValues();
3425 // Generate a sequence of selects of the form:
3426 // SELECT(Mask3, In3,
3427 // SELECT(Mask2, In2,
3429 for (unsigned In = 0; In < NumIncoming; In++) {
3430 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3432 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3434 for (unsigned part = 0; part < UF; ++part) {
3435 // We might have single edge PHIs (blocks) - use an identity
3436 // 'select' for the first PHI operand.
3438 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3441 // Select between the current value and the previous incoming edge
3442 // based on the incoming mask.
3443 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3444 Entry[part], "predphi");
3450 // This PHINode must be an induction variable.
3451 // Make sure that we know about it.
3452 assert(Legal->getInductionVars()->count(P) &&
3453 "Not an induction variable");
3455 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3457 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3458 // which can be found from the original scalar operations.
3459 switch (II.getKind()) {
3460 case InductionDescriptor::IK_NoInduction:
3461 llvm_unreachable("Unknown induction");
3462 case InductionDescriptor::IK_IntInduction: {
3463 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3464 // Handle other induction variables that are now based on the
3466 Value *V = Induction;
3467 if (P != OldInduction) {
3468 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3469 V = II.transform(Builder, V);
3470 V->setName("offset.idx");
3472 Value *Broadcasted = getBroadcastInstrs(V);
3473 // After broadcasting the induction variable we need to make the vector
3474 // consecutive by adding 0, 1, 2, etc.
3475 for (unsigned part = 0; part < UF; ++part)
3476 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3479 case InductionDescriptor::IK_PtrInduction:
3480 // Handle the pointer induction variable case.
3481 assert(P->getType()->isPointerTy() && "Unexpected type.");
3482 // This is the normalized GEP that starts counting at zero.
3483 Value *PtrInd = Induction;
3484 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3485 // This is the vector of results. Notice that we don't generate
3486 // vector geps because scalar geps result in better code.
3487 for (unsigned part = 0; part < UF; ++part) {
3489 int EltIndex = part;
3490 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3491 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3492 Value *SclrGep = II.transform(Builder, GlobalIdx);
3493 SclrGep->setName("next.gep");
3494 Entry[part] = SclrGep;
3498 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3499 for (unsigned int i = 0; i < VF; ++i) {
3500 int EltIndex = i + part * VF;
3501 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3502 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3503 Value *SclrGep = II.transform(Builder, GlobalIdx);
3504 SclrGep->setName("next.gep");
3505 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3506 Builder.getInt32(i),
3509 Entry[part] = VecVal;
3515 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3516 // For each instruction in the old loop.
3517 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3518 VectorParts &Entry = WidenMap.get(it);
3519 switch (it->getOpcode()) {
3520 case Instruction::Br:
3521 // Nothing to do for PHIs and BR, since we already took care of the
3522 // loop control flow instructions.
3524 case Instruction::PHI: {
3525 // Vectorize PHINodes.
3526 widenPHIInstruction(it, Entry, UF, VF, PV);
3530 case Instruction::Add:
3531 case Instruction::FAdd:
3532 case Instruction::Sub:
3533 case Instruction::FSub:
3534 case Instruction::Mul:
3535 case Instruction::FMul:
3536 case Instruction::UDiv:
3537 case Instruction::SDiv:
3538 case Instruction::FDiv:
3539 case Instruction::URem:
3540 case Instruction::SRem:
3541 case Instruction::FRem:
3542 case Instruction::Shl:
3543 case Instruction::LShr:
3544 case Instruction::AShr:
3545 case Instruction::And:
3546 case Instruction::Or:
3547 case Instruction::Xor: {
3548 // Just widen binops.
3549 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3550 setDebugLocFromInst(Builder, BinOp);
3551 VectorParts &A = getVectorValue(it->getOperand(0));
3552 VectorParts &B = getVectorValue(it->getOperand(1));
3554 // Use this vector value for all users of the original instruction.
3555 for (unsigned Part = 0; Part < UF; ++Part) {
3556 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3558 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3559 VecOp->copyIRFlags(BinOp);
3564 propagateMetadata(Entry, it);
3567 case Instruction::Select: {
3569 // If the selector is loop invariant we can create a select
3570 // instruction with a scalar condition. Otherwise, use vector-select.
3571 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3573 setDebugLocFromInst(Builder, it);
3575 // The condition can be loop invariant but still defined inside the
3576 // loop. This means that we can't just use the original 'cond' value.
3577 // We have to take the 'vectorized' value and pick the first lane.
3578 // Instcombine will make this a no-op.
3579 VectorParts &Cond = getVectorValue(it->getOperand(0));
3580 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3581 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3583 Value *ScalarCond = (VF == 1) ? Cond[0] :
3584 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3586 for (unsigned Part = 0; Part < UF; ++Part) {
3587 Entry[Part] = Builder.CreateSelect(
3588 InvariantCond ? ScalarCond : Cond[Part],
3593 propagateMetadata(Entry, it);
3597 case Instruction::ICmp:
3598 case Instruction::FCmp: {
3599 // Widen compares. Generate vector compares.
3600 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3601 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3602 setDebugLocFromInst(Builder, it);
3603 VectorParts &A = getVectorValue(it->getOperand(0));
3604 VectorParts &B = getVectorValue(it->getOperand(1));
3605 for (unsigned Part = 0; Part < UF; ++Part) {
3608 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3610 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3614 propagateMetadata(Entry, it);
3618 case Instruction::Store:
3619 case Instruction::Load:
3620 vectorizeMemoryInstruction(it);
3622 case Instruction::ZExt:
3623 case Instruction::SExt:
3624 case Instruction::FPToUI:
3625 case Instruction::FPToSI:
3626 case Instruction::FPExt:
3627 case Instruction::PtrToInt:
3628 case Instruction::IntToPtr:
3629 case Instruction::SIToFP:
3630 case Instruction::UIToFP:
3631 case Instruction::Trunc:
3632 case Instruction::FPTrunc:
3633 case Instruction::BitCast: {
3634 CastInst *CI = dyn_cast<CastInst>(it);
3635 setDebugLocFromInst(Builder, it);
3636 /// Optimize the special case where the source is the induction
3637 /// variable. Notice that we can only optimize the 'trunc' case
3638 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3639 /// c. other casts depend on pointer size.
3640 if (CI->getOperand(0) == OldInduction &&
3641 it->getOpcode() == Instruction::Trunc) {
3642 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3644 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3645 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3647 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3648 for (unsigned Part = 0; Part < UF; ++Part)
3649 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3650 propagateMetadata(Entry, it);
3653 /// Vectorize casts.
3654 Type *DestTy = (VF == 1) ? CI->getType() :
3655 VectorType::get(CI->getType(), VF);
3657 VectorParts &A = getVectorValue(it->getOperand(0));
3658 for (unsigned Part = 0; Part < UF; ++Part)
3659 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3660 propagateMetadata(Entry, it);
3664 case Instruction::Call: {
3665 // Ignore dbg intrinsics.
3666 if (isa<DbgInfoIntrinsic>(it))
3668 setDebugLocFromInst(Builder, it);
3670 Module *M = BB->getParent()->getParent();
3671 CallInst *CI = cast<CallInst>(it);
3673 StringRef FnName = CI->getCalledFunction()->getName();
3674 Function *F = CI->getCalledFunction();
3675 Type *RetTy = ToVectorTy(CI->getType(), VF);
3676 SmallVector<Type *, 4> Tys;
3677 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3678 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3680 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3682 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3683 ID == Intrinsic::lifetime_start)) {
3684 scalarizeInstruction(it);
3687 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3688 // version of the instruction.
3689 // Is it beneficial to perform intrinsic call compared to lib call?
3690 bool NeedToScalarize;
3691 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3692 bool UseVectorIntrinsic =
3693 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3694 if (!UseVectorIntrinsic && NeedToScalarize) {
3695 scalarizeInstruction(it);
3699 for (unsigned Part = 0; Part < UF; ++Part) {
3700 SmallVector<Value *, 4> Args;
3701 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3702 Value *Arg = CI->getArgOperand(i);
3703 // Some intrinsics have a scalar argument - don't replace it with a
3705 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3706 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3707 Arg = VectorArg[Part];
3709 Args.push_back(Arg);
3713 if (UseVectorIntrinsic) {
3714 // Use vector version of the intrinsic.
3715 Type *TysForDecl[] = {CI->getType()};
3717 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3718 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3720 // Use vector version of the library call.
3721 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3722 assert(!VFnName.empty() && "Vector function name is empty.");
3723 VectorF = M->getFunction(VFnName);
3725 // Generate a declaration
3726 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3728 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3729 VectorF->copyAttributesFrom(F);
3732 assert(VectorF && "Can't create vector function.");
3733 Entry[Part] = Builder.CreateCall(VectorF, Args);
3736 propagateMetadata(Entry, it);
3741 // All other instructions are unsupported. Scalarize them.
3742 scalarizeInstruction(it);
3745 }// end of for_each instr.
3748 void InnerLoopVectorizer::updateAnalysis() {
3749 // Forget the original basic block.
3750 SE->forgetLoop(OrigLoop);
3752 // Update the dominator tree information.
3753 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3754 "Entry does not dominate exit.");
3756 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3757 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3758 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3760 // Due to if predication of stores we might create a sequence of "if(pred)
3761 // a[i] = ...; " blocks.
3762 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3764 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3765 else if (isPredicatedBlock(i)) {
3766 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3768 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3772 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3773 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3774 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3775 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3777 DEBUG(DT->verifyDomTree());
3780 /// \brief Check whether it is safe to if-convert this phi node.
3782 /// Phi nodes with constant expressions that can trap are not safe to if
3784 static bool canIfConvertPHINodes(BasicBlock *BB) {
3785 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3786 PHINode *Phi = dyn_cast<PHINode>(I);
3789 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3790 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3797 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3798 if (!EnableIfConversion) {
3799 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3803 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3805 // A list of pointers that we can safely read and write to.
3806 SmallPtrSet<Value *, 8> SafePointes;
3808 // Collect safe addresses.
3809 for (Loop::block_iterator BI = TheLoop->block_begin(),
3810 BE = TheLoop->block_end(); BI != BE; ++BI) {
3811 BasicBlock *BB = *BI;
3813 if (blockNeedsPredication(BB))
3816 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3817 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3818 SafePointes.insert(LI->getPointerOperand());
3819 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3820 SafePointes.insert(SI->getPointerOperand());
3824 // Collect the blocks that need predication.
3825 BasicBlock *Header = TheLoop->getHeader();
3826 for (Loop::block_iterator BI = TheLoop->block_begin(),
3827 BE = TheLoop->block_end(); BI != BE; ++BI) {
3828 BasicBlock *BB = *BI;
3830 // We don't support switch statements inside loops.
3831 if (!isa<BranchInst>(BB->getTerminator())) {
3832 emitAnalysis(VectorizationReport(BB->getTerminator())
3833 << "loop contains a switch statement");
3837 // We must be able to predicate all blocks that need to be predicated.
3838 if (blockNeedsPredication(BB)) {
3839 if (!blockCanBePredicated(BB, SafePointes)) {
3840 emitAnalysis(VectorizationReport(BB->getTerminator())
3841 << "control flow cannot be substituted for a select");
3844 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3845 emitAnalysis(VectorizationReport(BB->getTerminator())
3846 << "control flow cannot be substituted for a select");
3851 // We can if-convert this loop.
3855 bool LoopVectorizationLegality::canVectorize() {
3856 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3857 // be canonicalized.
3858 if (!TheLoop->getLoopPreheader()) {
3860 VectorizationReport() <<
3861 "loop control flow is not understood by vectorizer");
3865 // We can only vectorize innermost loops.
3866 if (!TheLoop->empty()) {
3867 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3871 // We must have a single backedge.
3872 if (TheLoop->getNumBackEdges() != 1) {
3874 VectorizationReport() <<
3875 "loop control flow is not understood by vectorizer");
3879 // We must have a single exiting block.
3880 if (!TheLoop->getExitingBlock()) {
3882 VectorizationReport() <<
3883 "loop control flow is not understood by vectorizer");
3887 // We only handle bottom-tested loops, i.e. loop in which the condition is
3888 // checked at the end of each iteration. With that we can assume that all
3889 // instructions in the loop are executed the same number of times.
3890 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3892 VectorizationReport() <<
3893 "loop control flow is not understood by vectorizer");
3897 // We need to have a loop header.
3898 DEBUG(dbgs() << "LV: Found a loop: " <<
3899 TheLoop->getHeader()->getName() << '\n');
3901 // Check if we can if-convert non-single-bb loops.
3902 unsigned NumBlocks = TheLoop->getNumBlocks();
3903 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3904 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3908 // ScalarEvolution needs to be able to find the exit count.
3909 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3910 if (ExitCount == SE->getCouldNotCompute()) {
3911 emitAnalysis(VectorizationReport() <<
3912 "could not determine number of loop iterations");
3913 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3917 // Check if we can vectorize the instructions and CFG in this loop.
3918 if (!canVectorizeInstrs()) {
3919 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3923 // Go over each instruction and look at memory deps.
3924 if (!canVectorizeMemory()) {
3925 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3929 // Collect all of the variables that remain uniform after vectorization.
3930 collectLoopUniforms();
3932 DEBUG(dbgs() << "LV: We can vectorize this loop"
3933 << (LAI->getRuntimePointerChecking()->Need
3934 ? " (with a runtime bound check)"
3938 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
3940 // If an override option has been passed in for interleaved accesses, use it.
3941 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
3942 UseInterleaved = EnableInterleavedMemAccesses;
3944 // Analyze interleaved memory accesses.
3946 InterleaveInfo.analyzeInterleaving(Strides);
3948 // Okay! We can vectorize. At this point we don't have any other mem analysis
3949 // which may limit our maximum vectorization factor, so just return true with
3954 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3955 if (Ty->isPointerTy())
3956 return DL.getIntPtrType(Ty);
3958 // It is possible that char's or short's overflow when we ask for the loop's
3959 // trip count, work around this by changing the type size.
3960 if (Ty->getScalarSizeInBits() < 32)
3961 return Type::getInt32Ty(Ty->getContext());
3966 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3967 Ty0 = convertPointerToIntegerType(DL, Ty0);
3968 Ty1 = convertPointerToIntegerType(DL, Ty1);
3969 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3974 /// \brief Check that the instruction has outside loop users and is not an
3975 /// identified reduction variable.
3976 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3977 SmallPtrSetImpl<Value *> &Reductions) {
3978 // Reduction instructions are allowed to have exit users. All other
3979 // instructions must not have external users.
3980 if (!Reductions.count(Inst))
3981 //Check that all of the users of the loop are inside the BB.
3982 for (User *U : Inst->users()) {
3983 Instruction *UI = cast<Instruction>(U);
3984 // This user may be a reduction exit value.
3985 if (!TheLoop->contains(UI)) {
3986 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3993 bool LoopVectorizationLegality::canVectorizeInstrs() {
3994 BasicBlock *Header = TheLoop->getHeader();
3996 // Look for the attribute signaling the absence of NaNs.
3997 Function &F = *Header->getParent();
3998 const DataLayout &DL = F.getParent()->getDataLayout();
3999 if (F.hasFnAttribute("no-nans-fp-math"))
4001 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4003 // For each block in the loop.
4004 for (Loop::block_iterator bb = TheLoop->block_begin(),
4005 be = TheLoop->block_end(); bb != be; ++bb) {
4007 // Scan the instructions in the block and look for hazards.
4008 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4011 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4012 Type *PhiTy = Phi->getType();
4013 // Check that this PHI type is allowed.
4014 if (!PhiTy->isIntegerTy() &&
4015 !PhiTy->isFloatingPointTy() &&
4016 !PhiTy->isPointerTy()) {
4017 emitAnalysis(VectorizationReport(it)
4018 << "loop control flow is not understood by vectorizer");
4019 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4023 // If this PHINode is not in the header block, then we know that we
4024 // can convert it to select during if-conversion. No need to check if
4025 // the PHIs in this block are induction or reduction variables.
4026 if (*bb != Header) {
4027 // Check that this instruction has no outside users or is an
4028 // identified reduction value with an outside user.
4029 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4031 emitAnalysis(VectorizationReport(it) <<
4032 "value could not be identified as "
4033 "an induction or reduction variable");
4037 // We only allow if-converted PHIs with exactly two incoming values.
4038 if (Phi->getNumIncomingValues() != 2) {
4039 emitAnalysis(VectorizationReport(it)
4040 << "control flow not understood by vectorizer");
4041 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4045 InductionDescriptor ID;
4046 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4047 Inductions[Phi] = ID;
4048 // Get the widest type.
4050 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4052 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4054 // Int inductions are special because we only allow one IV.
4055 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4056 ID.getStepValue()->isOne() &&
4057 isa<Constant>(ID.getStartValue()) &&
4058 cast<Constant>(ID.getStartValue())->isNullValue()) {
4059 // Use the phi node with the widest type as induction. Use the last
4060 // one if there are multiple (no good reason for doing this other
4061 // than it is expedient). We've checked that it begins at zero and
4062 // steps by one, so this is a canonical induction variable.
4063 if (!Induction || PhiTy == WidestIndTy)
4067 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4069 // Until we explicitly handle the case of an induction variable with
4070 // an outside loop user we have to give up vectorizing this loop.
4071 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4072 emitAnalysis(VectorizationReport(it) <<
4073 "use of induction value outside of the "
4074 "loop is not handled by vectorizer");
4081 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4083 if (Reductions[Phi].hasUnsafeAlgebra())
4084 Requirements->addUnsafeAlgebraInst(
4085 Reductions[Phi].getUnsafeAlgebraInst());
4086 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4090 emitAnalysis(VectorizationReport(it) <<
4091 "value that could not be identified as "
4092 "reduction is used outside the loop");
4093 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4095 }// end of PHI handling
4097 // We handle calls that:
4098 // * Are debug info intrinsics.
4099 // * Have a mapping to an IR intrinsic.
4100 // * Have a vector version available.
4101 CallInst *CI = dyn_cast<CallInst>(it);
4102 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4103 !(CI->getCalledFunction() && TLI &&
4104 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4105 emitAnalysis(VectorizationReport(it) <<
4106 "call instruction cannot be vectorized");
4107 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4111 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4112 // second argument is the same (i.e. loop invariant)
4114 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4115 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4116 emitAnalysis(VectorizationReport(it)
4117 << "intrinsic instruction cannot be vectorized");
4118 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4123 // Check that the instruction return type is vectorizable.
4124 // Also, we can't vectorize extractelement instructions.
4125 if ((!VectorType::isValidElementType(it->getType()) &&
4126 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4127 emitAnalysis(VectorizationReport(it)
4128 << "instruction return type cannot be vectorized");
4129 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4133 // Check that the stored type is vectorizable.
4134 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4135 Type *T = ST->getValueOperand()->getType();
4136 if (!VectorType::isValidElementType(T)) {
4137 emitAnalysis(VectorizationReport(ST) <<
4138 "store instruction cannot be vectorized");
4141 if (EnableMemAccessVersioning)
4142 collectStridedAccess(ST);
4145 if (EnableMemAccessVersioning)
4146 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4147 collectStridedAccess(LI);
4149 // Reduction instructions are allowed to have exit users.
4150 // All other instructions must not have external users.
4151 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4152 emitAnalysis(VectorizationReport(it) <<
4153 "value cannot be used outside the loop");
4162 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4163 if (Inductions.empty()) {
4164 emitAnalysis(VectorizationReport()
4165 << "loop induction variable could not be identified");
4170 // Now we know the widest induction type, check if our found induction
4171 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4172 // will create another.
4173 if (Induction && WidestIndTy != Induction->getType())
4174 Induction = nullptr;
4179 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4180 Value *Ptr = nullptr;
4181 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4182 Ptr = LI->getPointerOperand();
4183 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4184 Ptr = SI->getPointerOperand();
4188 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4192 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4193 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4194 Strides[Ptr] = Stride;
4195 StrideSet.insert(Stride);
4198 void LoopVectorizationLegality::collectLoopUniforms() {
4199 // We now know that the loop is vectorizable!
4200 // Collect variables that will remain uniform after vectorization.
4201 std::vector<Value*> Worklist;
4202 BasicBlock *Latch = TheLoop->getLoopLatch();
4204 // Start with the conditional branch and walk up the block.
4205 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4207 // Also add all consecutive pointer values; these values will be uniform
4208 // after vectorization (and subsequent cleanup) and, until revectorization is
4209 // supported, all dependencies must also be uniform.
4210 for (Loop::block_iterator B = TheLoop->block_begin(),
4211 BE = TheLoop->block_end(); B != BE; ++B)
4212 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4214 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4215 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4217 while (!Worklist.empty()) {
4218 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4219 Worklist.pop_back();
4221 // Look at instructions inside this loop.
4222 // Stop when reaching PHI nodes.
4223 // TODO: we need to follow values all over the loop, not only in this block.
4224 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4227 // This is a known uniform.
4230 // Insert all operands.
4231 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4235 bool LoopVectorizationLegality::canVectorizeMemory() {
4236 LAI = &LAA->getInfo(TheLoop, Strides);
4237 auto &OptionalReport = LAI->getReport();
4239 emitAnalysis(VectorizationReport(*OptionalReport));
4240 if (!LAI->canVectorizeMemory())
4243 if (LAI->hasStoreToLoopInvariantAddress()) {
4245 VectorizationReport()
4246 << "write to a loop invariant address could not be vectorized");
4247 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4251 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4256 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4257 Value *In0 = const_cast<Value*>(V);
4258 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4262 return Inductions.count(PN);
4265 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4266 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4269 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4270 SmallPtrSetImpl<Value *> &SafePtrs) {
4272 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4273 // Check that we don't have a constant expression that can trap as operand.
4274 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4276 if (Constant *C = dyn_cast<Constant>(*OI))
4280 // We might be able to hoist the load.
4281 if (it->mayReadFromMemory()) {
4282 LoadInst *LI = dyn_cast<LoadInst>(it);
4285 if (!SafePtrs.count(LI->getPointerOperand())) {
4286 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4287 MaskedOp.insert(LI);
4294 // We don't predicate stores at the moment.
4295 if (it->mayWriteToMemory()) {
4296 StoreInst *SI = dyn_cast<StoreInst>(it);
4297 // We only support predication of stores in basic blocks with one
4302 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4303 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4305 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4306 !isSinglePredecessor) {
4307 // Build a masked store if it is legal for the target, otherwise scalarize
4309 bool isLegalMaskedOp =
4310 isLegalMaskedStore(SI->getValueOperand()->getType(),
4311 SI->getPointerOperand());
4312 if (isLegalMaskedOp) {
4314 MaskedOp.insert(SI);
4323 // The instructions below can trap.
4324 switch (it->getOpcode()) {
4326 case Instruction::UDiv:
4327 case Instruction::SDiv:
4328 case Instruction::URem:
4329 case Instruction::SRem:
4337 void InterleavedAccessInfo::collectConstStridedAccesses(
4338 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4339 const ValueToValueMap &Strides) {
4340 // Holds load/store instructions in program order.
4341 SmallVector<Instruction *, 16> AccessList;
4343 for (auto *BB : TheLoop->getBlocks()) {
4344 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4346 for (auto &I : *BB) {
4347 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4349 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4353 AccessList.push_back(&I);
4357 if (AccessList.empty())
4360 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4361 for (auto I : AccessList) {
4362 LoadInst *LI = dyn_cast<LoadInst>(I);
4363 StoreInst *SI = dyn_cast<StoreInst>(I);
4365 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4366 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4368 // The factor of the corresponding interleave group.
4369 unsigned Factor = std::abs(Stride);
4371 // Ignore the access if the factor is too small or too large.
4372 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4375 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4376 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4377 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4379 // An alignment of 0 means target ABI alignment.
4380 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4382 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4384 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4388 // Analyze interleaved accesses and collect them into interleave groups.
4390 // Notice that the vectorization on interleaved groups will change instruction
4391 // orders and may break dependences. But the memory dependence check guarantees
4392 // that there is no overlap between two pointers of different strides, element
4393 // sizes or underlying bases.
4395 // For pointers sharing the same stride, element size and underlying base, no
4396 // need to worry about Read-After-Write dependences and Write-After-Read
4399 // E.g. The RAW dependence: A[i] = a;
4401 // This won't exist as it is a store-load forwarding conflict, which has
4402 // already been checked and forbidden in the dependence check.
4404 // E.g. The WAR dependence: a = A[i]; // (1)
4406 // The store group of (2) is always inserted at or below (2), and the load group
4407 // of (1) is always inserted at or above (1). The dependence is safe.
4408 void InterleavedAccessInfo::analyzeInterleaving(
4409 const ValueToValueMap &Strides) {
4410 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4412 // Holds all the stride accesses.
4413 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4414 collectConstStridedAccesses(StrideAccesses, Strides);
4416 if (StrideAccesses.empty())
4419 // Holds all interleaved store groups temporarily.
4420 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4422 // Search the load-load/write-write pair B-A in bottom-up order and try to
4423 // insert B into the interleave group of A according to 3 rules:
4424 // 1. A and B have the same stride.
4425 // 2. A and B have the same memory object size.
4426 // 3. B belongs to the group according to the distance.
4428 // The bottom-up order can avoid breaking the Write-After-Write dependences
4429 // between two pointers of the same base.
4430 // E.g. A[i] = a; (1)
4433 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4434 // above (1), which guarantees that (1) is always above (2).
4435 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4437 Instruction *A = I->first;
4438 StrideDescriptor DesA = I->second;
4440 InterleaveGroup *Group = getInterleaveGroup(A);
4442 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4443 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4446 if (A->mayWriteToMemory())
4447 StoreGroups.insert(Group);
4449 for (auto II = std::next(I); II != E; ++II) {
4450 Instruction *B = II->first;
4451 StrideDescriptor DesB = II->second;
4453 // Ignore if B is already in a group or B is a different memory operation.
4454 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4457 // Check the rule 1 and 2.
4458 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4461 // Calculate the distance and prepare for the rule 3.
4462 const SCEVConstant *DistToA =
4463 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4467 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4469 // Skip if the distance is not multiple of size as they are not in the
4471 if (DistanceToA % static_cast<int>(DesA.Size))
4474 // The index of B is the index of A plus the related index to A.
4476 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4478 // Try to insert B into the group.
4479 if (Group->insertMember(B, IndexB, DesB.Align)) {
4480 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4481 << " into the interleave group with" << *A << '\n');
4482 InterleaveGroupMap[B] = Group;
4484 // Set the first load in program order as the insert position.
4485 if (B->mayReadFromMemory())
4486 Group->setInsertPos(B);
4488 } // Iteration on instruction B
4489 } // Iteration on instruction A
4491 // Remove interleaved store groups with gaps.
4492 for (InterleaveGroup *Group : StoreGroups)
4493 if (Group->getNumMembers() != Group->getFactor())
4494 releaseGroup(Group);
4497 LoopVectorizationCostModel::VectorizationFactor
4498 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4499 // Width 1 means no vectorize
4500 VectorizationFactor Factor = { 1U, 0U };
4501 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4502 emitAnalysis(VectorizationReport() <<
4503 "runtime pointer checks needed. Enable vectorization of this "
4504 "loop with '#pragma clang loop vectorize(enable)' when "
4505 "compiling with -Os/-Oz");
4507 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4511 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4512 emitAnalysis(VectorizationReport() <<
4513 "store that is conditionally executed prevents vectorization");
4514 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4518 // Find the trip count.
4519 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4520 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4522 unsigned WidestType = getWidestType();
4523 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4524 unsigned MaxSafeDepDist = -1U;
4525 if (Legal->getMaxSafeDepDistBytes() != -1U)
4526 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4527 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4528 WidestRegister : MaxSafeDepDist);
4529 unsigned MaxVectorSize = WidestRegister / WidestType;
4530 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4531 DEBUG(dbgs() << "LV: The Widest register is: "
4532 << WidestRegister << " bits.\n");
4534 if (MaxVectorSize == 0) {
4535 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4539 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4540 " into one vector!");
4542 unsigned VF = MaxVectorSize;
4544 // If we optimize the program for size, avoid creating the tail loop.
4546 // If we are unable to calculate the trip count then don't try to vectorize.
4549 (VectorizationReport() <<
4550 "unable to calculate the loop count due to complex control flow");
4551 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4555 // Find the maximum SIMD width that can fit within the trip count.
4556 VF = TC % MaxVectorSize;
4561 // If the trip count that we found modulo the vectorization factor is not
4562 // zero then we require a tail.
4563 emitAnalysis(VectorizationReport() <<
4564 "cannot optimize for size and vectorize at the "
4565 "same time. Enable vectorization of this loop "
4566 "with '#pragma clang loop vectorize(enable)' "
4567 "when compiling with -Os/-Oz");
4568 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4573 int UserVF = Hints->getWidth();
4575 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4576 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4578 Factor.Width = UserVF;
4582 float Cost = expectedCost(1);
4584 const float ScalarCost = Cost;
4587 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4589 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4590 // Ignore scalar width, because the user explicitly wants vectorization.
4591 if (ForceVectorization && VF > 1) {
4593 Cost = expectedCost(Width) / (float)Width;
4596 for (unsigned i=2; i <= VF; i*=2) {
4597 // Notice that the vector loop needs to be executed less times, so
4598 // we need to divide the cost of the vector loops by the width of
4599 // the vector elements.
4600 float VectorCost = expectedCost(i) / (float)i;
4601 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4602 (int)VectorCost << ".\n");
4603 if (VectorCost < Cost) {
4609 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4610 << "LV: Vectorization seems to be not beneficial, "
4611 << "but was forced by a user.\n");
4612 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4613 Factor.Width = Width;
4614 Factor.Cost = Width * Cost;
4618 unsigned LoopVectorizationCostModel::getWidestType() {
4619 unsigned MaxWidth = 8;
4620 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4623 for (Loop::block_iterator bb = TheLoop->block_begin(),
4624 be = TheLoop->block_end(); bb != be; ++bb) {
4625 BasicBlock *BB = *bb;
4627 // For each instruction in the loop.
4628 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4629 Type *T = it->getType();
4631 // Skip ignored values.
4632 if (ValuesToIgnore.count(it))
4635 // Only examine Loads, Stores and PHINodes.
4636 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4639 // Examine PHI nodes that are reduction variables. Update the type to
4640 // account for the recurrence type.
4641 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4642 if (!Legal->getReductionVars()->count(PN))
4644 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4645 T = RdxDesc.getRecurrenceType();
4648 // Examine the stored values.
4649 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4650 T = ST->getValueOperand()->getType();
4652 // Ignore loaded pointer types and stored pointer types that are not
4653 // consecutive. However, we do want to take consecutive stores/loads of
4654 // pointer vectors into account.
4655 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4658 MaxWidth = std::max(MaxWidth,
4659 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4666 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4668 unsigned LoopCost) {
4670 // -- The interleave heuristics --
4671 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4672 // There are many micro-architectural considerations that we can't predict
4673 // at this level. For example, frontend pressure (on decode or fetch) due to
4674 // code size, or the number and capabilities of the execution ports.
4676 // We use the following heuristics to select the interleave count:
4677 // 1. If the code has reductions, then we interleave to break the cross
4678 // iteration dependency.
4679 // 2. If the loop is really small, then we interleave to reduce the loop
4681 // 3. We don't interleave if we think that we will spill registers to memory
4682 // due to the increased register pressure.
4684 // When we optimize for size, we don't interleave.
4688 // We used the distance for the interleave count.
4689 if (Legal->getMaxSafeDepDistBytes() != -1U)
4692 // Do not interleave loops with a relatively small trip count.
4693 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4694 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4697 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4698 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4702 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4703 TargetNumRegisters = ForceTargetNumScalarRegs;
4705 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4706 TargetNumRegisters = ForceTargetNumVectorRegs;
4709 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4710 // We divide by these constants so assume that we have at least one
4711 // instruction that uses at least one register.
4712 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4713 R.NumInstructions = std::max(R.NumInstructions, 1U);
4715 // We calculate the interleave count using the following formula.
4716 // Subtract the number of loop invariants from the number of available
4717 // registers. These registers are used by all of the interleaved instances.
4718 // Next, divide the remaining registers by the number of registers that is
4719 // required by the loop, in order to estimate how many parallel instances
4720 // fit without causing spills. All of this is rounded down if necessary to be
4721 // a power of two. We want power of two interleave count to simplify any
4722 // addressing operations or alignment considerations.
4723 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4726 // Don't count the induction variable as interleaved.
4727 if (EnableIndVarRegisterHeur)
4728 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4729 std::max(1U, (R.MaxLocalUsers - 1)));
4731 // Clamp the interleave ranges to reasonable counts.
4732 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4734 // Check if the user has overridden the max.
4736 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4737 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4739 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4740 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4743 // If we did not calculate the cost for VF (because the user selected the VF)
4744 // then we calculate the cost of VF here.
4746 LoopCost = expectedCost(VF);
4748 // Clamp the calculated IC to be between the 1 and the max interleave count
4749 // that the target allows.
4750 if (IC > MaxInterleaveCount)
4751 IC = MaxInterleaveCount;
4755 // Interleave if we vectorized this loop and there is a reduction that could
4756 // benefit from interleaving.
4757 if (VF > 1 && Legal->getReductionVars()->size()) {
4758 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4762 // Note that if we've already vectorized the loop we will have done the
4763 // runtime check and so interleaving won't require further checks.
4764 bool InterleavingRequiresRuntimePointerCheck =
4765 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4767 // We want to interleave small loops in order to reduce the loop overhead and
4768 // potentially expose ILP opportunities.
4769 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4770 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4771 // We assume that the cost overhead is 1 and we use the cost model
4772 // to estimate the cost of the loop and interleave until the cost of the
4773 // loop overhead is about 5% of the cost of the loop.
4775 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4777 // Interleave until store/load ports (estimated by max interleave count) are
4779 unsigned NumStores = Legal->getNumStores();
4780 unsigned NumLoads = Legal->getNumLoads();
4781 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4782 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4784 // If we have a scalar reduction (vector reductions are already dealt with
4785 // by this point), we can increase the critical path length if the loop
4786 // we're interleaving is inside another loop. Limit, by default to 2, so the
4787 // critical path only gets increased by one reduction operation.
4788 if (Legal->getReductionVars()->size() &&
4789 TheLoop->getLoopDepth() > 1) {
4790 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4791 SmallIC = std::min(SmallIC, F);
4792 StoresIC = std::min(StoresIC, F);
4793 LoadsIC = std::min(LoadsIC, F);
4796 if (EnableLoadStoreRuntimeInterleave &&
4797 std::max(StoresIC, LoadsIC) > SmallIC) {
4798 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4799 return std::max(StoresIC, LoadsIC);
4802 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4806 // Interleave if this is a large loop (small loops are already dealt with by
4808 // point) that could benefit from interleaving.
4809 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4810 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4811 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4815 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4819 LoopVectorizationCostModel::RegisterUsage
4820 LoopVectorizationCostModel::calculateRegisterUsage() {
4821 // This function calculates the register usage by measuring the highest number
4822 // of values that are alive at a single location. Obviously, this is a very
4823 // rough estimation. We scan the loop in a topological order in order and
4824 // assign a number to each instruction. We use RPO to ensure that defs are
4825 // met before their users. We assume that each instruction that has in-loop
4826 // users starts an interval. We record every time that an in-loop value is
4827 // used, so we have a list of the first and last occurrences of each
4828 // instruction. Next, we transpose this data structure into a multi map that
4829 // holds the list of intervals that *end* at a specific location. This multi
4830 // map allows us to perform a linear search. We scan the instructions linearly
4831 // and record each time that a new interval starts, by placing it in a set.
4832 // If we find this value in the multi-map then we remove it from the set.
4833 // The max register usage is the maximum size of the set.
4834 // We also search for instructions that are defined outside the loop, but are
4835 // used inside the loop. We need this number separately from the max-interval
4836 // usage number because when we unroll, loop-invariant values do not take
4838 LoopBlocksDFS DFS(TheLoop);
4842 R.NumInstructions = 0;
4844 // Each 'key' in the map opens a new interval. The values
4845 // of the map are the index of the 'last seen' usage of the
4846 // instruction that is the key.
4847 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4848 // Maps instruction to its index.
4849 DenseMap<unsigned, Instruction*> IdxToInstr;
4850 // Marks the end of each interval.
4851 IntervalMap EndPoint;
4852 // Saves the list of instruction indices that are used in the loop.
4853 SmallSet<Instruction*, 8> Ends;
4854 // Saves the list of values that are used in the loop but are
4855 // defined outside the loop, such as arguments and constants.
4856 SmallPtrSet<Value*, 8> LoopInvariants;
4859 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4860 be = DFS.endRPO(); bb != be; ++bb) {
4861 R.NumInstructions += (*bb)->size();
4862 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4864 Instruction *I = it;
4865 IdxToInstr[Index++] = I;
4867 // Save the end location of each USE.
4868 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4869 Value *U = I->getOperand(i);
4870 Instruction *Instr = dyn_cast<Instruction>(U);
4872 // Ignore non-instruction values such as arguments, constants, etc.
4873 if (!Instr) continue;
4875 // If this instruction is outside the loop then record it and continue.
4876 if (!TheLoop->contains(Instr)) {
4877 LoopInvariants.insert(Instr);
4881 // Overwrite previous end points.
4882 EndPoint[Instr] = Index;
4888 // Saves the list of intervals that end with the index in 'key'.
4889 typedef SmallVector<Instruction*, 2> InstrList;
4890 DenseMap<unsigned, InstrList> TransposeEnds;
4892 // Transpose the EndPoints to a list of values that end at each index.
4893 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4895 TransposeEnds[it->second].push_back(it->first);
4897 SmallSet<Instruction*, 8> OpenIntervals;
4898 unsigned MaxUsage = 0;
4901 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4902 for (unsigned int i = 0; i < Index; ++i) {
4903 Instruction *I = IdxToInstr[i];
4904 // Ignore instructions that are never used within the loop.
4905 if (!Ends.count(I)) continue;
4907 // Skip ignored values.
4908 if (ValuesToIgnore.count(I))
4911 // Remove all of the instructions that end at this location.
4912 InstrList &List = TransposeEnds[i];
4913 for (unsigned int j=0, e = List.size(); j < e; ++j)
4914 OpenIntervals.erase(List[j]);
4916 // Count the number of live interals.
4917 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4919 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4920 OpenIntervals.size() << '\n');
4922 // Add the current instruction to the list of open intervals.
4923 OpenIntervals.insert(I);
4926 unsigned Invariant = LoopInvariants.size();
4927 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4928 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4929 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4931 R.LoopInvariantRegs = Invariant;
4932 R.MaxLocalUsers = MaxUsage;
4936 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4940 for (Loop::block_iterator bb = TheLoop->block_begin(),
4941 be = TheLoop->block_end(); bb != be; ++bb) {
4942 unsigned BlockCost = 0;
4943 BasicBlock *BB = *bb;
4945 // For each instruction in the old loop.
4946 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4947 // Skip dbg intrinsics.
4948 if (isa<DbgInfoIntrinsic>(it))
4951 // Skip ignored values.
4952 if (ValuesToIgnore.count(it))
4955 unsigned C = getInstructionCost(it, VF);
4957 // Check if we should override the cost.
4958 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4959 C = ForceTargetInstructionCost;
4962 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4963 VF << " For instruction: " << *it << '\n');
4966 // We assume that if-converted blocks have a 50% chance of being executed.
4967 // When the code is scalar then some of the blocks are avoided due to CF.
4968 // When the code is vectorized we execute all code paths.
4969 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4978 /// \brief Check whether the address computation for a non-consecutive memory
4979 /// access looks like an unlikely candidate for being merged into the indexing
4982 /// We look for a GEP which has one index that is an induction variable and all
4983 /// other indices are loop invariant. If the stride of this access is also
4984 /// within a small bound we decide that this address computation can likely be
4985 /// merged into the addressing mode.
4986 /// In all other cases, we identify the address computation as complex.
4987 static bool isLikelyComplexAddressComputation(Value *Ptr,
4988 LoopVectorizationLegality *Legal,
4989 ScalarEvolution *SE,
4990 const Loop *TheLoop) {
4991 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4995 // We are looking for a gep with all loop invariant indices except for one
4996 // which should be an induction variable.
4997 unsigned NumOperands = Gep->getNumOperands();
4998 for (unsigned i = 1; i < NumOperands; ++i) {
4999 Value *Opd = Gep->getOperand(i);
5000 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5001 !Legal->isInductionVariable(Opd))
5005 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5006 // can likely be merged into the address computation.
5007 unsigned MaxMergeDistance = 64;
5009 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5013 // Check the step is constant.
5014 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5015 // Calculate the pointer stride and check if it is consecutive.
5016 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5020 const APInt &APStepVal = C->getValue()->getValue();
5022 // Huge step value - give up.
5023 if (APStepVal.getBitWidth() > 64)
5026 int64_t StepVal = APStepVal.getSExtValue();
5028 return StepVal > MaxMergeDistance;
5031 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5032 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5038 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5039 // If we know that this instruction will remain uniform, check the cost of
5040 // the scalar version.
5041 if (Legal->isUniformAfterVectorization(I))
5044 Type *RetTy = I->getType();
5045 Type *VectorTy = ToVectorTy(RetTy, VF);
5047 // TODO: We need to estimate the cost of intrinsic calls.
5048 switch (I->getOpcode()) {
5049 case Instruction::GetElementPtr:
5050 // We mark this instruction as zero-cost because the cost of GEPs in
5051 // vectorized code depends on whether the corresponding memory instruction
5052 // is scalarized or not. Therefore, we handle GEPs with the memory
5053 // instruction cost.
5055 case Instruction::Br: {
5056 return TTI.getCFInstrCost(I->getOpcode());
5058 case Instruction::PHI:
5059 //TODO: IF-converted IFs become selects.
5061 case Instruction::Add:
5062 case Instruction::FAdd:
5063 case Instruction::Sub:
5064 case Instruction::FSub:
5065 case Instruction::Mul:
5066 case Instruction::FMul:
5067 case Instruction::UDiv:
5068 case Instruction::SDiv:
5069 case Instruction::FDiv:
5070 case Instruction::URem:
5071 case Instruction::SRem:
5072 case Instruction::FRem:
5073 case Instruction::Shl:
5074 case Instruction::LShr:
5075 case Instruction::AShr:
5076 case Instruction::And:
5077 case Instruction::Or:
5078 case Instruction::Xor: {
5079 // Since we will replace the stride by 1 the multiplication should go away.
5080 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5082 // Certain instructions can be cheaper to vectorize if they have a constant
5083 // second vector operand. One example of this are shifts on x86.
5084 TargetTransformInfo::OperandValueKind Op1VK =
5085 TargetTransformInfo::OK_AnyValue;
5086 TargetTransformInfo::OperandValueKind Op2VK =
5087 TargetTransformInfo::OK_AnyValue;
5088 TargetTransformInfo::OperandValueProperties Op1VP =
5089 TargetTransformInfo::OP_None;
5090 TargetTransformInfo::OperandValueProperties Op2VP =
5091 TargetTransformInfo::OP_None;
5092 Value *Op2 = I->getOperand(1);
5094 // Check for a splat of a constant or for a non uniform vector of constants.
5095 if (isa<ConstantInt>(Op2)) {
5096 ConstantInt *CInt = cast<ConstantInt>(Op2);
5097 if (CInt && CInt->getValue().isPowerOf2())
5098 Op2VP = TargetTransformInfo::OP_PowerOf2;
5099 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5100 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5101 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5102 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5104 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5105 if (CInt && CInt->getValue().isPowerOf2())
5106 Op2VP = TargetTransformInfo::OP_PowerOf2;
5107 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5111 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5114 case Instruction::Select: {
5115 SelectInst *SI = cast<SelectInst>(I);
5116 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5117 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5118 Type *CondTy = SI->getCondition()->getType();
5120 CondTy = VectorType::get(CondTy, VF);
5122 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5124 case Instruction::ICmp:
5125 case Instruction::FCmp: {
5126 Type *ValTy = I->getOperand(0)->getType();
5127 VectorTy = ToVectorTy(ValTy, VF);
5128 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5130 case Instruction::Store:
5131 case Instruction::Load: {
5132 StoreInst *SI = dyn_cast<StoreInst>(I);
5133 LoadInst *LI = dyn_cast<LoadInst>(I);
5134 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5136 VectorTy = ToVectorTy(ValTy, VF);
5138 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5139 unsigned AS = SI ? SI->getPointerAddressSpace() :
5140 LI->getPointerAddressSpace();
5141 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5142 // We add the cost of address computation here instead of with the gep
5143 // instruction because only here we know whether the operation is
5146 return TTI.getAddressComputationCost(VectorTy) +
5147 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5149 // For an interleaved access, calculate the total cost of the whole
5150 // interleave group.
5151 if (Legal->isAccessInterleaved(I)) {
5152 auto Group = Legal->getInterleavedAccessGroup(I);
5153 assert(Group && "Fail to get an interleaved access group.");
5155 // Only calculate the cost once at the insert position.
5156 if (Group->getInsertPos() != I)
5159 unsigned InterleaveFactor = Group->getFactor();
5161 VectorType::get(VectorTy->getVectorElementType(),
5162 VectorTy->getVectorNumElements() * InterleaveFactor);
5164 // Holds the indices of existing members in an interleaved load group.
5165 // An interleaved store group doesn't need this as it dones't allow gaps.
5166 SmallVector<unsigned, 4> Indices;
5168 for (unsigned i = 0; i < InterleaveFactor; i++)
5169 if (Group->getMember(i))
5170 Indices.push_back(i);
5173 // Calculate the cost of the whole interleaved group.
5174 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5175 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5176 Group->getAlignment(), AS);
5178 if (Group->isReverse())
5180 Group->getNumMembers() *
5181 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5183 // FIXME: The interleaved load group with a huge gap could be even more
5184 // expensive than scalar operations. Then we could ignore such group and
5185 // use scalar operations instead.
5189 // Scalarized loads/stores.
5190 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5191 bool Reverse = ConsecutiveStride < 0;
5192 const DataLayout &DL = I->getModule()->getDataLayout();
5193 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5194 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5195 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5196 bool IsComplexComputation =
5197 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5199 // The cost of extracting from the value vector and pointer vector.
5200 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5201 for (unsigned i = 0; i < VF; ++i) {
5202 // The cost of extracting the pointer operand.
5203 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5204 // In case of STORE, the cost of ExtractElement from the vector.
5205 // In case of LOAD, the cost of InsertElement into the returned
5207 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5208 Instruction::InsertElement,
5212 // The cost of the scalar loads/stores.
5213 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5214 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5219 // Wide load/stores.
5220 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5221 if (Legal->isMaskRequired(I))
5222 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5225 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5228 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5232 case Instruction::ZExt:
5233 case Instruction::SExt:
5234 case Instruction::FPToUI:
5235 case Instruction::FPToSI:
5236 case Instruction::FPExt:
5237 case Instruction::PtrToInt:
5238 case Instruction::IntToPtr:
5239 case Instruction::SIToFP:
5240 case Instruction::UIToFP:
5241 case Instruction::Trunc:
5242 case Instruction::FPTrunc:
5243 case Instruction::BitCast: {
5244 // We optimize the truncation of induction variable.
5245 // The cost of these is the same as the scalar operation.
5246 if (I->getOpcode() == Instruction::Trunc &&
5247 Legal->isInductionVariable(I->getOperand(0)))
5248 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5249 I->getOperand(0)->getType());
5251 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5252 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5254 case Instruction::Call: {
5255 bool NeedToScalarize;
5256 CallInst *CI = cast<CallInst>(I);
5257 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5258 if (getIntrinsicIDForCall(CI, TLI))
5259 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5263 // We are scalarizing the instruction. Return the cost of the scalar
5264 // instruction, plus the cost of insert and extract into vector
5265 // elements, times the vector width.
5268 if (!RetTy->isVoidTy() && VF != 1) {
5269 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5271 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5274 // The cost of inserting the results plus extracting each one of the
5276 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5279 // The cost of executing VF copies of the scalar instruction. This opcode
5280 // is unknown. Assume that it is the same as 'mul'.
5281 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5287 char LoopVectorize::ID = 0;
5288 static const char lv_name[] = "Loop Vectorization";
5289 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5290 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5291 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5292 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5293 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5294 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5295 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5296 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5297 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5298 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5299 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5300 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5303 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5304 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5308 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5309 // Check for a store.
5310 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5311 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5313 // Check for a load.
5314 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5315 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5321 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5322 bool IfPredicateStore) {
5323 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5324 // Holds vector parameters or scalars, in case of uniform vals.
5325 SmallVector<VectorParts, 4> Params;
5327 setDebugLocFromInst(Builder, Instr);
5329 // Find all of the vectorized parameters.
5330 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5331 Value *SrcOp = Instr->getOperand(op);
5333 // If we are accessing the old induction variable, use the new one.
5334 if (SrcOp == OldInduction) {
5335 Params.push_back(getVectorValue(SrcOp));
5339 // Try using previously calculated values.
5340 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5342 // If the src is an instruction that appeared earlier in the basic block
5343 // then it should already be vectorized.
5344 if (SrcInst && OrigLoop->contains(SrcInst)) {
5345 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5346 // The parameter is a vector value from earlier.
5347 Params.push_back(WidenMap.get(SrcInst));
5349 // The parameter is a scalar from outside the loop. Maybe even a constant.
5350 VectorParts Scalars;
5351 Scalars.append(UF, SrcOp);
5352 Params.push_back(Scalars);
5356 assert(Params.size() == Instr->getNumOperands() &&
5357 "Invalid number of operands");
5359 // Does this instruction return a value ?
5360 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5362 Value *UndefVec = IsVoidRetTy ? nullptr :
5363 UndefValue::get(Instr->getType());
5364 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5365 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5367 Instruction *InsertPt = Builder.GetInsertPoint();
5368 BasicBlock *IfBlock = Builder.GetInsertBlock();
5369 BasicBlock *CondBlock = nullptr;
5372 Loop *VectorLp = nullptr;
5373 if (IfPredicateStore) {
5374 assert(Instr->getParent()->getSinglePredecessor() &&
5375 "Only support single predecessor blocks");
5376 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5377 Instr->getParent());
5378 VectorLp = LI->getLoopFor(IfBlock);
5379 assert(VectorLp && "Must have a loop for this block");
5382 // For each vector unroll 'part':
5383 for (unsigned Part = 0; Part < UF; ++Part) {
5384 // For each scalar that we create:
5386 // Start an "if (pred) a[i] = ..." block.
5387 Value *Cmp = nullptr;
5388 if (IfPredicateStore) {
5389 if (Cond[Part]->getType()->isVectorTy())
5391 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5392 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5393 ConstantInt::get(Cond[Part]->getType(), 1));
5394 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5395 LoopVectorBody.push_back(CondBlock);
5396 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5397 // Update Builder with newly created basic block.
5398 Builder.SetInsertPoint(InsertPt);
5401 Instruction *Cloned = Instr->clone();
5403 Cloned->setName(Instr->getName() + ".cloned");
5404 // Replace the operands of the cloned instructions with extracted scalars.
5405 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5406 Value *Op = Params[op][Part];
5407 Cloned->setOperand(op, Op);
5410 // Place the cloned scalar in the new loop.
5411 Builder.Insert(Cloned);
5413 // If the original scalar returns a value we need to place it in a vector
5414 // so that future users will be able to use it.
5416 VecResults[Part] = Cloned;
5419 if (IfPredicateStore) {
5420 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5421 LoopVectorBody.push_back(NewIfBlock);
5422 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5423 Builder.SetInsertPoint(InsertPt);
5424 ReplaceInstWithInst(IfBlock->getTerminator(),
5425 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5426 IfBlock = NewIfBlock;
5431 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5432 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5433 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5435 return scalarizeInstruction(Instr, IfPredicateStore);
5438 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5442 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5446 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5447 // When unrolling and the VF is 1, we only need to add a simple scalar.
5448 Type *ITy = Val->getType();
5449 assert(!ITy->isVectorTy() && "Val must be a scalar");
5450 Constant *C = ConstantInt::get(ITy, StartIdx);
5451 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");