1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/EquivalenceClasses.h"
52 #include "llvm/ADT/Hashing.h"
53 #include "llvm/ADT/MapVector.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/AliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/LoopAccessAnalysis.h"
66 #include "llvm/Analysis/LoopInfo.h"
67 #include "llvm/Analysis/LoopIterator.h"
68 #include "llvm/Analysis/LoopPass.h"
69 #include "llvm/Analysis/ScalarEvolution.h"
70 #include "llvm/Analysis/ScalarEvolutionExpander.h"
71 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
72 #include "llvm/Analysis/TargetTransformInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/IR/Constants.h"
75 #include "llvm/IR/DataLayout.h"
76 #include "llvm/IR/DebugInfo.h"
77 #include "llvm/IR/DerivedTypes.h"
78 #include "llvm/IR/DiagnosticInfo.h"
79 #include "llvm/IR/Dominators.h"
80 #include "llvm/IR/Function.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/Instructions.h"
83 #include "llvm/IR/IntrinsicInst.h"
84 #include "llvm/IR/LLVMContext.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/IR/Type.h"
88 #include "llvm/IR/Value.h"
89 #include "llvm/IR/ValueHandle.h"
90 #include "llvm/IR/Verifier.h"
91 #include "llvm/Pass.h"
92 #include "llvm/Support/BranchProbability.h"
93 #include "llvm/Support/CommandLine.h"
94 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/raw_ostream.h"
96 #include "llvm/Transforms/Scalar.h"
97 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
98 #include "llvm/Transforms/Utils/Local.h"
99 #include "llvm/Analysis/VectorUtils.h"
100 #include "llvm/Transforms/Utils/LoopUtils.h"
105 using namespace llvm;
106 using namespace llvm::PatternMatch;
108 #define LV_NAME "loop-vectorize"
109 #define DEBUG_TYPE LV_NAME
111 STATISTIC(LoopsVectorized, "Number of loops vectorized");
112 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 static cl::opt<bool> EnableInterleavedMemAccesses(
142 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
143 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
145 /// Maximum factor for an interleaved memory access.
146 static cl::opt<unsigned> MaxInterleaveGroupFactor(
147 "max-interleave-group-factor", cl::Hidden,
148 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
151 /// We don't interleave loops with a known constant trip count below this
153 static const unsigned TinyTripCountInterleaveThreshold = 128;
155 static cl::opt<unsigned> ForceTargetNumScalarRegs(
156 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of scalar registers."));
159 static cl::opt<unsigned> ForceTargetNumVectorRegs(
160 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's number of vector registers."));
163 /// Maximum vectorization interleave count.
164 static const unsigned MaxInterleaveFactor = 16;
166 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
167 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max interleave factor for "
171 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
172 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
174 "vectorized loops."));
176 static cl::opt<unsigned> ForceTargetInstructionCost(
177 "force-target-instruction-cost", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's expected cost for "
179 "an instruction to a single constant value. Mostly "
180 "useful for getting consistent testing."));
182 static cl::opt<unsigned> SmallLoopCost(
183 "small-loop-cost", cl::init(20), cl::Hidden,
185 "The cost of a loop that is considered 'small' by the interleaver."));
187 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
188 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
189 cl::desc("Enable the use of the block frequency analysis to access PGO "
190 "heuristics minimizing code growth in cold regions and being more "
191 "aggressive in hot regions."));
193 // Runtime interleave loops for load/store throughput.
194 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
195 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
197 "Enable runtime interleaving until load/store ports are saturated"));
199 /// The number of stores in a loop that are allowed to need predication.
200 static cl::opt<unsigned> NumberOfStoresToPredicate(
201 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
202 cl::desc("Max number of stores to be predicated behind an if."));
204 static cl::opt<bool> EnableIndVarRegisterHeur(
205 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
206 cl::desc("Count the induction variable only once when interleaving"));
208 static cl::opt<bool> EnableCondStoresVectorization(
209 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
210 cl::desc("Enable if predication of stores during vectorization."));
212 static cl::opt<unsigned> MaxNestedScalarReductionIC(
213 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
214 cl::desc("The maximum interleave count to use when interleaving a scalar "
215 "reduction in a nested loop."));
217 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
218 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
219 cl::desc("The maximum allowed number of runtime memory checks with a "
220 "vectorize(enable) pragma."));
224 // Forward declarations.
225 class LoopVectorizeHints;
226 class LoopVectorizationLegality;
227 class LoopVectorizationCostModel;
228 class LoopVectorizationRequirements;
230 /// \brief This modifies LoopAccessReport to initialize message with
231 /// loop-vectorizer-specific part.
232 class VectorizationReport : public LoopAccessReport {
234 VectorizationReport(Instruction *I = nullptr)
235 : LoopAccessReport("loop not vectorized: ", I) {}
237 /// \brief This allows promotion of the loop-access analysis report into the
238 /// loop-vectorizer report. It modifies the message to add the
239 /// loop-vectorizer-specific part of the message.
240 explicit VectorizationReport(const LoopAccessReport &R)
241 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
245 /// A helper function for converting Scalar types to vector types.
246 /// If the incoming type is void, we return void. If the VF is 1, we return
248 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
249 if (Scalar->isVoidTy() || VF == 1)
251 return VectorType::get(Scalar, VF);
254 /// InnerLoopVectorizer vectorizes loops which contain only one basic
255 /// block to a specified vectorization factor (VF).
256 /// This class performs the widening of scalars into vectors, or multiple
257 /// scalars. This class also implements the following features:
258 /// * It inserts an epilogue loop for handling loops that don't have iteration
259 /// counts that are known to be a multiple of the vectorization factor.
260 /// * It handles the code generation for reduction variables.
261 /// * Scalarization (implementation using scalars) of un-vectorizable
263 /// InnerLoopVectorizer does not perform any vectorization-legality
264 /// checks, and relies on the caller to check for the different legality
265 /// aspects. The InnerLoopVectorizer relies on the
266 /// LoopVectorizationLegality class to provide information about the induction
267 /// and reduction variables that were found to a given vectorization factor.
268 class InnerLoopVectorizer {
270 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
271 DominatorTree *DT, const TargetLibraryInfo *TLI,
272 const TargetTransformInfo *TTI, unsigned VecWidth,
273 unsigned UnrollFactor)
274 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
275 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
276 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
277 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr),
278 AddedSafetyChecks(false) {}
280 // Perform the actual loop widening (vectorization).
281 void vectorize(LoopVectorizationLegality *L) {
283 // Create a new empty loop. Unlink the old loop and connect the new one.
285 // Widen each instruction in the old loop to a new one in the new loop.
286 // Use the Legality module to find the induction and reduction variables.
288 // Register the new loop and update the analysis passes.
292 // Return true if any runtime check is added.
293 bool IsSafetyChecksAdded() {
294 return AddedSafetyChecks;
297 virtual ~InnerLoopVectorizer() {}
300 /// A small list of PHINodes.
301 typedef SmallVector<PHINode*, 4> PhiVector;
302 /// When we unroll loops we have multiple vector values for each scalar.
303 /// This data structure holds the unrolled and vectorized values that
304 /// originated from one scalar instruction.
305 typedef SmallVector<Value*, 2> VectorParts;
307 // When we if-convert we need to create edge masks. We have to cache values
308 // so that we don't end up with exponential recursion/IR.
309 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
310 VectorParts> EdgeMaskCache;
312 /// \brief Add checks for strides that were assumed to be 1.
314 /// Returns the last check instruction and the first check instruction in the
315 /// pair as (first, last).
316 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
318 /// Create an empty loop, based on the loop ranges of the old loop.
319 void createEmptyLoop();
320 /// Create a new induction variable inside L.
321 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
322 Value *Step, Instruction *DL);
323 /// Copy and widen the instructions from the old loop.
324 virtual void vectorizeLoop();
326 /// \brief The Loop exit block may have single value PHI nodes where the
327 /// incoming value is 'Undef'. While vectorizing we only handled real values
328 /// that were defined inside the loop. Here we fix the 'undef case'.
332 /// A helper function that computes the predicate of the block BB, assuming
333 /// that the header block of the loop is set to True. It returns the *entry*
334 /// mask for the block BB.
335 VectorParts createBlockInMask(BasicBlock *BB);
336 /// A helper function that computes the predicate of the edge between SRC
338 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
340 /// A helper function to vectorize a single BB within the innermost loop.
341 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
343 /// Vectorize a single PHINode in a block. This method handles the induction
344 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
345 /// arbitrary length vectors.
346 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
347 unsigned UF, unsigned VF, PhiVector *PV);
349 /// Insert the new loop to the loop hierarchy and pass manager
350 /// and update the analysis passes.
351 void updateAnalysis();
353 /// This instruction is un-vectorizable. Implement it as a sequence
354 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
355 /// scalarized instruction behind an if block predicated on the control
356 /// dependence of the instruction.
357 virtual void scalarizeInstruction(Instruction *Instr,
358 bool IfPredicateStore=false);
360 /// Vectorize Load and Store instructions,
361 virtual void vectorizeMemoryInstruction(Instruction *Instr);
363 /// Create a broadcast instruction. This method generates a broadcast
364 /// instruction (shuffle) for loop invariant values and for the induction
365 /// value. If this is the induction variable then we extend it to N, N+1, ...
366 /// this is needed because each iteration in the loop corresponds to a SIMD
368 virtual Value *getBroadcastInstrs(Value *V);
370 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
371 /// to each vector element of Val. The sequence starts at StartIndex.
372 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
374 /// When we go over instructions in the basic block we rely on previous
375 /// values within the current basic block or on loop invariant values.
376 /// When we widen (vectorize) values we place them in the map. If the values
377 /// are not within the map, they have to be loop invariant, so we simply
378 /// broadcast them into a vector.
379 VectorParts &getVectorValue(Value *V);
381 /// Try to vectorize the interleaved access group that \p Instr belongs to.
382 void vectorizeInterleaveGroup(Instruction *Instr);
384 /// Generate a shuffle sequence that will reverse the vector Vec.
385 virtual Value *reverseVector(Value *Vec);
387 /// Returns (and creates if needed) the original loop trip count.
388 Value *getOrCreateTripCount(Loop *NewLoop);
390 /// Returns (and creates if needed) the trip count of the widened loop.
391 Value *getOrCreateVectorTripCount(Loop *NewLoop);
393 /// This is a helper class that holds the vectorizer state. It maps scalar
394 /// instructions to vector instructions. When the code is 'unrolled' then
395 /// then a single scalar value is mapped to multiple vector parts. The parts
396 /// are stored in the VectorPart type.
398 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
400 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
402 /// \return True if 'Key' is saved in the Value Map.
403 bool has(Value *Key) const { return MapStorage.count(Key); }
405 /// Initializes a new entry in the map. Sets all of the vector parts to the
406 /// save value in 'Val'.
407 /// \return A reference to a vector with splat values.
408 VectorParts &splat(Value *Key, Value *Val) {
409 VectorParts &Entry = MapStorage[Key];
410 Entry.assign(UF, Val);
414 ///\return A reference to the value that is stored at 'Key'.
415 VectorParts &get(Value *Key) {
416 VectorParts &Entry = MapStorage[Key];
419 assert(Entry.size() == UF);
424 /// The unroll factor. Each entry in the map stores this number of vector
428 /// Map storage. We use std::map and not DenseMap because insertions to a
429 /// dense map invalidates its iterators.
430 std::map<Value *, VectorParts> MapStorage;
433 /// The original loop.
435 /// Scev analysis to use.
443 /// Target Library Info.
444 const TargetLibraryInfo *TLI;
445 /// Target Transform Info.
446 const TargetTransformInfo *TTI;
448 /// The vectorization SIMD factor to use. Each vector will have this many
453 /// The vectorization unroll factor to use. Each scalar is vectorized to this
454 /// many different vector instructions.
457 /// The builder that we use
460 // --- Vectorization state ---
462 /// The vector-loop preheader.
463 BasicBlock *LoopVectorPreHeader;
464 /// The scalar-loop preheader.
465 BasicBlock *LoopScalarPreHeader;
466 /// Middle Block between the vector and the scalar.
467 BasicBlock *LoopMiddleBlock;
468 ///The ExitBlock of the scalar loop.
469 BasicBlock *LoopExitBlock;
470 ///The vector loop body.
471 SmallVector<BasicBlock *, 4> LoopVectorBody;
472 ///The scalar loop body.
473 BasicBlock *LoopScalarBody;
474 /// A list of all bypass blocks. The first block is the entry of the loop.
475 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
477 /// The new Induction variable which was added to the new block.
479 /// The induction variable of the old basic block.
480 PHINode *OldInduction;
481 /// Maps scalars to widened vectors.
483 EdgeMaskCache MaskCache;
484 /// Trip count of the original loop.
486 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
487 Value *VectorTripCount;
489 LoopVectorizationLegality *Legal;
491 // Record whether runtime check is added.
492 bool AddedSafetyChecks;
495 class InnerLoopUnroller : public InnerLoopVectorizer {
497 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
498 DominatorTree *DT, const TargetLibraryInfo *TLI,
499 const TargetTransformInfo *TTI, unsigned UnrollFactor)
500 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
503 void scalarizeInstruction(Instruction *Instr,
504 bool IfPredicateStore = false) override;
505 void vectorizeMemoryInstruction(Instruction *Instr) override;
506 Value *getBroadcastInstrs(Value *V) override;
507 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
508 Value *reverseVector(Value *Vec) override;
511 /// \brief Look for a meaningful debug location on the instruction or it's
513 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
518 if (I->getDebugLoc() != Empty)
521 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
522 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
523 if (OpInst->getDebugLoc() != Empty)
530 /// \brief Set the debug location in the builder using the debug location in the
532 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
533 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
534 B.SetCurrentDebugLocation(Inst->getDebugLoc());
536 B.SetCurrentDebugLocation(DebugLoc());
540 /// \return string containing a file name and a line # for the given loop.
541 static std::string getDebugLocString(const Loop *L) {
544 raw_string_ostream OS(Result);
545 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
546 LoopDbgLoc.print(OS);
548 // Just print the module name.
549 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
556 /// \brief Propagate known metadata from one instruction to another.
557 static void propagateMetadata(Instruction *To, const Instruction *From) {
558 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
559 From->getAllMetadataOtherThanDebugLoc(Metadata);
561 for (auto M : Metadata) {
562 unsigned Kind = M.first;
564 // These are safe to transfer (this is safe for TBAA, even when we
565 // if-convert, because should that metadata have had a control dependency
566 // on the condition, and thus actually aliased with some other
567 // non-speculated memory access when the condition was false, this would be
568 // caught by the runtime overlap checks).
569 if (Kind != LLVMContext::MD_tbaa &&
570 Kind != LLVMContext::MD_alias_scope &&
571 Kind != LLVMContext::MD_noalias &&
572 Kind != LLVMContext::MD_fpmath &&
573 Kind != LLVMContext::MD_nontemporal)
576 To->setMetadata(Kind, M.second);
580 /// \brief Propagate known metadata from one instruction to a vector of others.
581 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
583 if (Instruction *I = dyn_cast<Instruction>(V))
584 propagateMetadata(I, From);
587 /// \brief The group of interleaved loads/stores sharing the same stride and
588 /// close to each other.
590 /// Each member in this group has an index starting from 0, and the largest
591 /// index should be less than interleaved factor, which is equal to the absolute
592 /// value of the access's stride.
594 /// E.g. An interleaved load group of factor 4:
595 /// for (unsigned i = 0; i < 1024; i+=4) {
596 /// a = A[i]; // Member of index 0
597 /// b = A[i+1]; // Member of index 1
598 /// d = A[i+3]; // Member of index 3
602 /// An interleaved store group of factor 4:
603 /// for (unsigned i = 0; i < 1024; i+=4) {
605 /// A[i] = a; // Member of index 0
606 /// A[i+1] = b; // Member of index 1
607 /// A[i+2] = c; // Member of index 2
608 /// A[i+3] = d; // Member of index 3
611 /// Note: the interleaved load group could have gaps (missing members), but
612 /// the interleaved store group doesn't allow gaps.
613 class InterleaveGroup {
615 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
616 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
617 assert(Align && "The alignment should be non-zero");
619 Factor = std::abs(Stride);
620 assert(Factor > 1 && "Invalid interleave factor");
622 Reverse = Stride < 0;
626 bool isReverse() const { return Reverse; }
627 unsigned getFactor() const { return Factor; }
628 unsigned getAlignment() const { return Align; }
629 unsigned getNumMembers() const { return Members.size(); }
631 /// \brief Try to insert a new member \p Instr with index \p Index and
632 /// alignment \p NewAlign. The index is related to the leader and it could be
633 /// negative if it is the new leader.
635 /// \returns false if the instruction doesn't belong to the group.
636 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
637 assert(NewAlign && "The new member's alignment should be non-zero");
639 int Key = Index + SmallestKey;
641 // Skip if there is already a member with the same index.
642 if (Members.count(Key))
645 if (Key > LargestKey) {
646 // The largest index is always less than the interleave factor.
647 if (Index >= static_cast<int>(Factor))
651 } else if (Key < SmallestKey) {
652 // The largest index is always less than the interleave factor.
653 if (LargestKey - Key >= static_cast<int>(Factor))
659 // It's always safe to select the minimum alignment.
660 Align = std::min(Align, NewAlign);
661 Members[Key] = Instr;
665 /// \brief Get the member with the given index \p Index
667 /// \returns nullptr if contains no such member.
668 Instruction *getMember(unsigned Index) const {
669 int Key = SmallestKey + Index;
670 if (!Members.count(Key))
673 return Members.find(Key)->second;
676 /// \brief Get the index for the given member. Unlike the key in the member
677 /// map, the index starts from 0.
678 unsigned getIndex(Instruction *Instr) const {
679 for (auto I : Members)
680 if (I.second == Instr)
681 return I.first - SmallestKey;
683 llvm_unreachable("InterleaveGroup contains no such member");
686 Instruction *getInsertPos() const { return InsertPos; }
687 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
690 unsigned Factor; // Interleave Factor.
693 DenseMap<int, Instruction *> Members;
697 // To avoid breaking dependences, vectorized instructions of an interleave
698 // group should be inserted at either the first load or the last store in
701 // E.g. %even = load i32 // Insert Position
702 // %add = add i32 %even // Use of %even
706 // %odd = add i32 // Def of %odd
707 // store i32 %odd // Insert Position
708 Instruction *InsertPos;
711 /// \brief Drive the analysis of interleaved memory accesses in the loop.
713 /// Use this class to analyze interleaved accesses only when we can vectorize
714 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
715 /// on interleaved accesses is unsafe.
717 /// The analysis collects interleave groups and records the relationships
718 /// between the member and the group in a map.
719 class InterleavedAccessInfo {
721 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
722 : SE(SE), TheLoop(L), DT(DT) {}
724 ~InterleavedAccessInfo() {
725 SmallSet<InterleaveGroup *, 4> DelSet;
726 // Avoid releasing a pointer twice.
727 for (auto &I : InterleaveGroupMap)
728 DelSet.insert(I.second);
729 for (auto *Ptr : DelSet)
733 /// \brief Analyze the interleaved accesses and collect them in interleave
734 /// groups. Substitute symbolic strides using \p Strides.
735 void analyzeInterleaving(const ValueToValueMap &Strides);
737 /// \brief Check if \p Instr belongs to any interleave group.
738 bool isInterleaved(Instruction *Instr) const {
739 return InterleaveGroupMap.count(Instr);
742 /// \brief Get the interleave group that \p Instr belongs to.
744 /// \returns nullptr if doesn't have such group.
745 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
746 if (InterleaveGroupMap.count(Instr))
747 return InterleaveGroupMap.find(Instr)->second;
756 /// Holds the relationships between the members and the interleave group.
757 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
759 /// \brief The descriptor for a strided memory access.
760 struct StrideDescriptor {
761 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
763 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
765 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
767 int Stride; // The access's stride. It is negative for a reverse access.
768 const SCEV *Scev; // The scalar expression of this access
769 unsigned Size; // The size of the memory object.
770 unsigned Align; // The alignment of this access.
773 /// \brief Create a new interleave group with the given instruction \p Instr,
774 /// stride \p Stride and alignment \p Align.
776 /// \returns the newly created interleave group.
777 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
779 assert(!InterleaveGroupMap.count(Instr) &&
780 "Already in an interleaved access group");
781 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
782 return InterleaveGroupMap[Instr];
785 /// \brief Release the group and remove all the relationships.
786 void releaseGroup(InterleaveGroup *Group) {
787 for (unsigned i = 0; i < Group->getFactor(); i++)
788 if (Instruction *Member = Group->getMember(i))
789 InterleaveGroupMap.erase(Member);
794 /// \brief Collect all the accesses with a constant stride in program order.
795 void collectConstStridedAccesses(
796 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
797 const ValueToValueMap &Strides);
800 /// Utility class for getting and setting loop vectorizer hints in the form
801 /// of loop metadata.
802 /// This class keeps a number of loop annotations locally (as member variables)
803 /// and can, upon request, write them back as metadata on the loop. It will
804 /// initially scan the loop for existing metadata, and will update the local
805 /// values based on information in the loop.
806 /// We cannot write all values to metadata, as the mere presence of some info,
807 /// for example 'force', means a decision has been made. So, we need to be
808 /// careful NOT to add them if the user hasn't specifically asked so.
809 class LoopVectorizeHints {
816 /// Hint - associates name and validation with the hint value.
819 unsigned Value; // This may have to change for non-numeric values.
822 Hint(const char * Name, unsigned Value, HintKind Kind)
823 : Name(Name), Value(Value), Kind(Kind) { }
825 bool validate(unsigned Val) {
828 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
830 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
838 /// Vectorization width.
840 /// Vectorization interleave factor.
842 /// Vectorization forced
845 /// Return the loop metadata prefix.
846 static StringRef Prefix() { return "llvm.loop."; }
850 FK_Undefined = -1, ///< Not selected.
851 FK_Disabled = 0, ///< Forcing disabled.
852 FK_Enabled = 1, ///< Forcing enabled.
855 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
856 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
858 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
859 Force("vectorize.enable", FK_Undefined, HK_FORCE),
861 // Populate values with existing loop metadata.
862 getHintsFromMetadata();
864 // force-vector-interleave overrides DisableInterleaving.
865 if (VectorizerParams::isInterleaveForced())
866 Interleave.Value = VectorizerParams::VectorizationInterleave;
868 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
869 << "LV: Interleaving disabled by the pass manager\n");
872 /// Mark the loop L as already vectorized by setting the width to 1.
873 void setAlreadyVectorized() {
874 Width.Value = Interleave.Value = 1;
875 Hint Hints[] = {Width, Interleave};
876 writeHintsToMetadata(Hints);
879 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
880 if (getForce() == LoopVectorizeHints::FK_Disabled) {
881 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
882 emitOptimizationRemarkAnalysis(F->getContext(),
883 vectorizeAnalysisPassName(), *F,
884 L->getStartLoc(), emitRemark());
888 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
889 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
890 emitOptimizationRemarkAnalysis(F->getContext(),
891 vectorizeAnalysisPassName(), *F,
892 L->getStartLoc(), emitRemark());
896 if (getWidth() == 1 && getInterleave() == 1) {
897 // FIXME: Add a separate metadata to indicate when the loop has already
898 // been vectorized instead of setting width and count to 1.
899 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
900 // FIXME: Add interleave.disable metadata. This will allow
901 // vectorize.disable to be used without disabling the pass and errors
902 // to differentiate between disabled vectorization and a width of 1.
903 emitOptimizationRemarkAnalysis(
904 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
905 "loop not vectorized: vectorization and interleaving are explicitly "
906 "disabled, or vectorize width and interleave count are both set to "
914 /// Dumps all the hint information.
915 std::string emitRemark() const {
916 VectorizationReport R;
917 if (Force.Value == LoopVectorizeHints::FK_Disabled)
918 R << "vectorization is explicitly disabled";
920 R << "use -Rpass-analysis=loop-vectorize for more info";
921 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
923 if (Width.Value != 0)
924 R << ", Vector Width=" << Width.Value;
925 if (Interleave.Value != 0)
926 R << ", Interleave Count=" << Interleave.Value;
934 unsigned getWidth() const { return Width.Value; }
935 unsigned getInterleave() const { return Interleave.Value; }
936 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
937 const char *vectorizeAnalysisPassName() const {
938 // If hints are provided that don't disable vectorization use the
939 // AlwaysPrint pass name to force the frontend to print the diagnostic.
942 if (getForce() == LoopVectorizeHints::FK_Disabled)
944 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
946 return DiagnosticInfo::AlwaysPrint;
949 bool allowReordering() const {
950 // When enabling loop hints are provided we allow the vectorizer to change
951 // the order of operations that is given by the scalar loop. This is not
952 // enabled by default because can be unsafe or inefficient. For example,
953 // reordering floating-point operations will change the way round-off
954 // error accumulates in the loop.
955 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
959 /// Find hints specified in the loop metadata and update local values.
960 void getHintsFromMetadata() {
961 MDNode *LoopID = TheLoop->getLoopID();
965 // First operand should refer to the loop id itself.
966 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
967 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
969 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
970 const MDString *S = nullptr;
971 SmallVector<Metadata *, 4> Args;
973 // The expected hint is either a MDString or a MDNode with the first
974 // operand a MDString.
975 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
976 if (!MD || MD->getNumOperands() == 0)
978 S = dyn_cast<MDString>(MD->getOperand(0));
979 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
980 Args.push_back(MD->getOperand(i));
982 S = dyn_cast<MDString>(LoopID->getOperand(i));
983 assert(Args.size() == 0 && "too many arguments for MDString");
989 // Check if the hint starts with the loop metadata prefix.
990 StringRef Name = S->getString();
991 if (Args.size() == 1)
992 setHint(Name, Args[0]);
996 /// Checks string hint with one operand and set value if valid.
997 void setHint(StringRef Name, Metadata *Arg) {
998 if (!Name.startswith(Prefix()))
1000 Name = Name.substr(Prefix().size(), StringRef::npos);
1002 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1004 unsigned Val = C->getZExtValue();
1006 Hint *Hints[] = {&Width, &Interleave, &Force};
1007 for (auto H : Hints) {
1008 if (Name == H->Name) {
1009 if (H->validate(Val))
1012 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1018 /// Create a new hint from name / value pair.
1019 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1020 LLVMContext &Context = TheLoop->getHeader()->getContext();
1021 Metadata *MDs[] = {MDString::get(Context, Name),
1022 ConstantAsMetadata::get(
1023 ConstantInt::get(Type::getInt32Ty(Context), V))};
1024 return MDNode::get(Context, MDs);
1027 /// Matches metadata with hint name.
1028 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1029 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1033 for (auto H : HintTypes)
1034 if (Name->getString().endswith(H.Name))
1039 /// Sets current hints into loop metadata, keeping other values intact.
1040 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1041 if (HintTypes.size() == 0)
1044 // Reserve the first element to LoopID (see below).
1045 SmallVector<Metadata *, 4> MDs(1);
1046 // If the loop already has metadata, then ignore the existing operands.
1047 MDNode *LoopID = TheLoop->getLoopID();
1049 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1050 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1051 // If node in update list, ignore old value.
1052 if (!matchesHintMetadataName(Node, HintTypes))
1053 MDs.push_back(Node);
1057 // Now, add the missing hints.
1058 for (auto H : HintTypes)
1059 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1061 // Replace current metadata node with new one.
1062 LLVMContext &Context = TheLoop->getHeader()->getContext();
1063 MDNode *NewLoopID = MDNode::get(Context, MDs);
1064 // Set operand 0 to refer to the loop id itself.
1065 NewLoopID->replaceOperandWith(0, NewLoopID);
1067 TheLoop->setLoopID(NewLoopID);
1070 /// The loop these hints belong to.
1071 const Loop *TheLoop;
1074 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1075 const LoopVectorizeHints &Hints,
1076 const LoopAccessReport &Message) {
1077 const char *Name = Hints.vectorizeAnalysisPassName();
1078 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1081 static void emitMissedWarning(Function *F, Loop *L,
1082 const LoopVectorizeHints &LH) {
1083 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1086 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1087 if (LH.getWidth() != 1)
1088 emitLoopVectorizeWarning(
1089 F->getContext(), *F, L->getStartLoc(),
1090 "failed explicitly specified loop vectorization");
1091 else if (LH.getInterleave() != 1)
1092 emitLoopInterleaveWarning(
1093 F->getContext(), *F, L->getStartLoc(),
1094 "failed explicitly specified loop interleaving");
1098 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1099 /// to what vectorization factor.
1100 /// This class does not look at the profitability of vectorization, only the
1101 /// legality. This class has two main kinds of checks:
1102 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1103 /// will change the order of memory accesses in a way that will change the
1104 /// correctness of the program.
1105 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1106 /// checks for a number of different conditions, such as the availability of a
1107 /// single induction variable, that all types are supported and vectorize-able,
1108 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1109 /// This class is also used by InnerLoopVectorizer for identifying
1110 /// induction variable and the different reduction variables.
1111 class LoopVectorizationLegality {
1113 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1114 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1115 Function *F, const TargetTransformInfo *TTI,
1116 LoopAccessAnalysis *LAA,
1117 LoopVectorizationRequirements *R,
1118 const LoopVectorizeHints *H)
1119 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1120 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1121 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1122 Requirements(R), Hints(H) {}
1124 /// ReductionList contains the reduction descriptors for all
1125 /// of the reductions that were found in the loop.
1126 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1128 /// InductionList saves induction variables and maps them to the
1129 /// induction descriptor.
1130 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1132 /// Returns true if it is legal to vectorize this loop.
1133 /// This does not mean that it is profitable to vectorize this
1134 /// loop, only that it is legal to do so.
1135 bool canVectorize();
1137 /// Returns the Induction variable.
1138 PHINode *getInduction() { return Induction; }
1140 /// Returns the reduction variables found in the loop.
1141 ReductionList *getReductionVars() { return &Reductions; }
1143 /// Returns the induction variables found in the loop.
1144 InductionList *getInductionVars() { return &Inductions; }
1146 /// Returns the widest induction type.
1147 Type *getWidestInductionType() { return WidestIndTy; }
1149 /// Returns True if V is an induction variable in this loop.
1150 bool isInductionVariable(const Value *V);
1152 /// Return true if the block BB needs to be predicated in order for the loop
1153 /// to be vectorized.
1154 bool blockNeedsPredication(BasicBlock *BB);
1156 /// Check if this pointer is consecutive when vectorizing. This happens
1157 /// when the last index of the GEP is the induction variable, or that the
1158 /// pointer itself is an induction variable.
1159 /// This check allows us to vectorize A[idx] into a wide load/store.
1161 /// 0 - Stride is unknown or non-consecutive.
1162 /// 1 - Address is consecutive.
1163 /// -1 - Address is consecutive, and decreasing.
1164 int isConsecutivePtr(Value *Ptr);
1166 /// Returns true if the value V is uniform within the loop.
1167 bool isUniform(Value *V);
1169 /// Returns true if this instruction will remain scalar after vectorization.
1170 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1172 /// Returns the information that we collected about runtime memory check.
1173 const RuntimePointerChecking *getRuntimePointerChecking() const {
1174 return LAI->getRuntimePointerChecking();
1177 const LoopAccessInfo *getLAI() const {
1181 /// \brief Check if \p Instr belongs to any interleaved access group.
1182 bool isAccessInterleaved(Instruction *Instr) {
1183 return InterleaveInfo.isInterleaved(Instr);
1186 /// \brief Get the interleaved access group that \p Instr belongs to.
1187 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1188 return InterleaveInfo.getInterleaveGroup(Instr);
1191 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1193 bool hasStride(Value *V) { return StrideSet.count(V); }
1194 bool mustCheckStrides() { return !StrideSet.empty(); }
1195 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1196 return StrideSet.begin();
1198 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1200 /// Returns true if the target machine supports masked store operation
1201 /// for the given \p DataType and kind of access to \p Ptr.
1202 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1203 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1205 /// Returns true if the target machine supports masked load operation
1206 /// for the given \p DataType and kind of access to \p Ptr.
1207 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1208 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1210 /// Returns true if vector representation of the instruction \p I
1212 bool isMaskRequired(const Instruction* I) {
1213 return (MaskedOp.count(I) != 0);
1215 unsigned getNumStores() const {
1216 return LAI->getNumStores();
1218 unsigned getNumLoads() const {
1219 return LAI->getNumLoads();
1221 unsigned getNumPredStores() const {
1222 return NumPredStores;
1225 /// Check if a single basic block loop is vectorizable.
1226 /// At this point we know that this is a loop with a constant trip count
1227 /// and we only need to check individual instructions.
1228 bool canVectorizeInstrs();
1230 /// When we vectorize loops we may change the order in which
1231 /// we read and write from memory. This method checks if it is
1232 /// legal to vectorize the code, considering only memory constrains.
1233 /// Returns true if the loop is vectorizable
1234 bool canVectorizeMemory();
1236 /// Return true if we can vectorize this loop using the IF-conversion
1238 bool canVectorizeWithIfConvert();
1240 /// Collect the variables that need to stay uniform after vectorization.
1241 void collectLoopUniforms();
1243 /// Return true if all of the instructions in the block can be speculatively
1244 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1245 /// and we know that we can read from them without segfault.
1246 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1248 /// \brief Collect memory access with loop invariant strides.
1250 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1252 void collectStridedAccess(Value *LoadOrStoreInst);
1254 /// Report an analysis message to assist the user in diagnosing loops that are
1255 /// not vectorized. These are handled as LoopAccessReport rather than
1256 /// VectorizationReport because the << operator of VectorizationReport returns
1257 /// LoopAccessReport.
1258 void emitAnalysis(const LoopAccessReport &Message) const {
1259 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1262 unsigned NumPredStores;
1264 /// The loop that we evaluate.
1267 ScalarEvolution *SE;
1268 /// Target Library Info.
1269 TargetLibraryInfo *TLI;
1271 Function *TheFunction;
1272 /// Target Transform Info
1273 const TargetTransformInfo *TTI;
1276 // LoopAccess analysis.
1277 LoopAccessAnalysis *LAA;
1278 // And the loop-accesses info corresponding to this loop. This pointer is
1279 // null until canVectorizeMemory sets it up.
1280 const LoopAccessInfo *LAI;
1282 /// The interleave access information contains groups of interleaved accesses
1283 /// with the same stride and close to each other.
1284 InterleavedAccessInfo InterleaveInfo;
1286 // --- vectorization state --- //
1288 /// Holds the integer induction variable. This is the counter of the
1291 /// Holds the reduction variables.
1292 ReductionList Reductions;
1293 /// Holds all of the induction variables that we found in the loop.
1294 /// Notice that inductions don't need to start at zero and that induction
1295 /// variables can be pointers.
1296 InductionList Inductions;
1297 /// Holds the widest induction type encountered.
1300 /// Allowed outside users. This holds the reduction
1301 /// vars which can be accessed from outside the loop.
1302 SmallPtrSet<Value*, 4> AllowedExit;
1303 /// This set holds the variables which are known to be uniform after
1305 SmallPtrSet<Instruction*, 4> Uniforms;
1307 /// Can we assume the absence of NaNs.
1308 bool HasFunNoNaNAttr;
1310 /// Vectorization requirements that will go through late-evaluation.
1311 LoopVectorizationRequirements *Requirements;
1313 /// Used to emit an analysis of any legality issues.
1314 const LoopVectorizeHints *Hints;
1316 ValueToValueMap Strides;
1317 SmallPtrSet<Value *, 8> StrideSet;
1319 /// While vectorizing these instructions we have to generate a
1320 /// call to the appropriate masked intrinsic
1321 SmallPtrSet<const Instruction*, 8> MaskedOp;
1324 /// LoopVectorizationCostModel - estimates the expected speedups due to
1326 /// In many cases vectorization is not profitable. This can happen because of
1327 /// a number of reasons. In this class we mainly attempt to predict the
1328 /// expected speedup/slowdowns due to the supported instruction set. We use the
1329 /// TargetTransformInfo to query the different backends for the cost of
1330 /// different operations.
1331 class LoopVectorizationCostModel {
1333 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1334 LoopVectorizationLegality *Legal,
1335 const TargetTransformInfo &TTI,
1336 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1337 const Function *F, const LoopVectorizeHints *Hints,
1338 SmallPtrSetImpl<const Value *> &ValuesToIgnore)
1339 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1340 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1342 /// Information about vectorization costs
1343 struct VectorizationFactor {
1344 unsigned Width; // Vector width with best cost
1345 unsigned Cost; // Cost of the loop with that width
1347 /// \return The most profitable vectorization factor and the cost of that VF.
1348 /// This method checks every power of two up to VF. If UserVF is not ZERO
1349 /// then this vectorization factor will be selected if vectorization is
1351 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1353 /// \return The size (in bits) of the widest type in the code that
1354 /// needs to be vectorized. We ignore values that remain scalar such as
1355 /// 64 bit loop indices.
1356 unsigned getWidestType();
1358 /// \return The desired interleave count.
1359 /// If interleave count has been specified by metadata it will be returned.
1360 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1361 /// are the selected vectorization factor and the cost of the selected VF.
1362 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1365 /// \return The most profitable unroll factor.
1366 /// This method finds the best unroll-factor based on register pressure and
1367 /// other parameters. VF and LoopCost are the selected vectorization factor
1368 /// and the cost of the selected VF.
1369 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1372 /// \brief A struct that represents some properties of the register usage
1374 struct RegisterUsage {
1375 /// Holds the number of loop invariant values that are used in the loop.
1376 unsigned LoopInvariantRegs;
1377 /// Holds the maximum number of concurrent live intervals in the loop.
1378 unsigned MaxLocalUsers;
1379 /// Holds the number of instructions in the loop.
1380 unsigned NumInstructions;
1383 /// \return information about the register usage of the loop.
1384 RegisterUsage calculateRegisterUsage();
1387 /// Returns the expected execution cost. The unit of the cost does
1388 /// not matter because we use the 'cost' units to compare different
1389 /// vector widths. The cost that is returned is *not* normalized by
1390 /// the factor width.
1391 unsigned expectedCost(unsigned VF);
1393 /// Returns the execution time cost of an instruction for a given vector
1394 /// width. Vector width of one means scalar.
1395 unsigned getInstructionCost(Instruction *I, unsigned VF);
1397 /// Returns whether the instruction is a load or store and will be a emitted
1398 /// as a vector operation.
1399 bool isConsecutiveLoadOrStore(Instruction *I);
1401 /// Report an analysis message to assist the user in diagnosing loops that are
1402 /// not vectorized. These are handled as LoopAccessReport rather than
1403 /// VectorizationReport because the << operator of VectorizationReport returns
1404 /// LoopAccessReport.
1405 void emitAnalysis(const LoopAccessReport &Message) const {
1406 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1409 /// The loop that we evaluate.
1412 ScalarEvolution *SE;
1413 /// Loop Info analysis.
1415 /// Vectorization legality.
1416 LoopVectorizationLegality *Legal;
1417 /// Vector target information.
1418 const TargetTransformInfo &TTI;
1419 /// Target Library Info.
1420 const TargetLibraryInfo *TLI;
1421 const Function *TheFunction;
1422 // Loop Vectorize Hint.
1423 const LoopVectorizeHints *Hints;
1424 // Values to ignore in the cost model.
1425 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1428 /// \brief This holds vectorization requirements that must be verified late in
1429 /// the process. The requirements are set by legalize and costmodel. Once
1430 /// vectorization has been determined to be possible and profitable the
1431 /// requirements can be verified by looking for metadata or compiler options.
1432 /// For example, some loops require FP commutativity which is only allowed if
1433 /// vectorization is explicitly specified or if the fast-math compiler option
1434 /// has been provided.
1435 /// Late evaluation of these requirements allows helpful diagnostics to be
1436 /// composed that tells the user what need to be done to vectorize the loop. For
1437 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1438 /// evaluation should be used only when diagnostics can generated that can be
1439 /// followed by a non-expert user.
1440 class LoopVectorizationRequirements {
1442 LoopVectorizationRequirements()
1443 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1445 void addUnsafeAlgebraInst(Instruction *I) {
1446 // First unsafe algebra instruction.
1447 if (!UnsafeAlgebraInst)
1448 UnsafeAlgebraInst = I;
1451 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1453 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1454 const char *Name = Hints.vectorizeAnalysisPassName();
1455 bool Failed = false;
1456 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1457 emitOptimizationRemarkAnalysisFPCommute(
1458 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1459 VectorizationReport() << "cannot prove it is safe to reorder "
1460 "floating-point operations");
1464 // Test if runtime memcheck thresholds are exceeded.
1465 bool PragmaThresholdReached =
1466 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1467 bool ThresholdReached =
1468 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1469 if ((ThresholdReached && !Hints.allowReordering()) ||
1470 PragmaThresholdReached) {
1471 emitOptimizationRemarkAnalysisAliasing(
1472 F->getContext(), Name, *F, L->getStartLoc(),
1473 VectorizationReport()
1474 << "cannot prove it is safe to reorder memory operations");
1475 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1483 unsigned NumRuntimePointerChecks;
1484 Instruction *UnsafeAlgebraInst;
1487 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1489 return V.push_back(&L);
1491 for (Loop *InnerL : L)
1492 addInnerLoop(*InnerL, V);
1495 /// The LoopVectorize Pass.
1496 struct LoopVectorize : public FunctionPass {
1497 /// Pass identification, replacement for typeid
1500 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1502 DisableUnrolling(NoUnrolling),
1503 AlwaysVectorize(AlwaysVectorize) {
1504 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1507 ScalarEvolution *SE;
1509 TargetTransformInfo *TTI;
1511 BlockFrequencyInfo *BFI;
1512 TargetLibraryInfo *TLI;
1514 AssumptionCache *AC;
1515 LoopAccessAnalysis *LAA;
1516 bool DisableUnrolling;
1517 bool AlwaysVectorize;
1519 BlockFrequency ColdEntryFreq;
1521 bool runOnFunction(Function &F) override {
1522 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1523 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1524 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1525 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1526 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1527 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1528 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1529 AA = &getAnalysis<AliasAnalysis>();
1530 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1531 LAA = &getAnalysis<LoopAccessAnalysis>();
1533 // Compute some weights outside of the loop over the loops. Compute this
1534 // using a BranchProbability to re-use its scaling math.
1535 const BranchProbability ColdProb(1, 5); // 20%
1536 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1539 // 1. the target claims to have no vector registers, and
1540 // 2. interleaving won't help ILP.
1542 // The second condition is necessary because, even if the target has no
1543 // vector registers, loop vectorization may still enable scalar
1545 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1548 // Build up a worklist of inner-loops to vectorize. This is necessary as
1549 // the act of vectorizing or partially unrolling a loop creates new loops
1550 // and can invalidate iterators across the loops.
1551 SmallVector<Loop *, 8> Worklist;
1554 addInnerLoop(*L, Worklist);
1556 LoopsAnalyzed += Worklist.size();
1558 // Now walk the identified inner loops.
1559 bool Changed = false;
1560 while (!Worklist.empty())
1561 Changed |= processLoop(Worklist.pop_back_val());
1563 // Process each loop nest in the function.
1567 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1568 SmallVector<Metadata *, 4> MDs;
1569 // Reserve first location for self reference to the LoopID metadata node.
1570 MDs.push_back(nullptr);
1571 bool IsUnrollMetadata = false;
1572 MDNode *LoopID = L->getLoopID();
1574 // First find existing loop unrolling disable metadata.
1575 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1576 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1578 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1580 S && S->getString().startswith("llvm.loop.unroll.disable");
1582 MDs.push_back(LoopID->getOperand(i));
1586 if (!IsUnrollMetadata) {
1587 // Add runtime unroll disable metadata.
1588 LLVMContext &Context = L->getHeader()->getContext();
1589 SmallVector<Metadata *, 1> DisableOperands;
1590 DisableOperands.push_back(
1591 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1592 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1593 MDs.push_back(DisableNode);
1594 MDNode *NewLoopID = MDNode::get(Context, MDs);
1595 // Set operand 0 to refer to the loop id itself.
1596 NewLoopID->replaceOperandWith(0, NewLoopID);
1597 L->setLoopID(NewLoopID);
1601 bool processLoop(Loop *L) {
1602 assert(L->empty() && "Only process inner loops.");
1605 const std::string DebugLocStr = getDebugLocString(L);
1608 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1609 << L->getHeader()->getParent()->getName() << "\" from "
1610 << DebugLocStr << "\n");
1612 LoopVectorizeHints Hints(L, DisableUnrolling);
1614 DEBUG(dbgs() << "LV: Loop hints:"
1616 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1618 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1620 : "?")) << " width=" << Hints.getWidth()
1621 << " unroll=" << Hints.getInterleave() << "\n");
1623 // Function containing loop
1624 Function *F = L->getHeader()->getParent();
1626 // Looking at the diagnostic output is the only way to determine if a loop
1627 // was vectorized (other than looking at the IR or machine code), so it
1628 // is important to generate an optimization remark for each loop. Most of
1629 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1630 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1631 // less verbose reporting vectorized loops and unvectorized loops that may
1632 // benefit from vectorization, respectively.
1634 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1635 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1639 // Check the loop for a trip count threshold:
1640 // do not vectorize loops with a tiny trip count.
1641 const unsigned TC = SE->getSmallConstantTripCount(L);
1642 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1643 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1644 << "This loop is not worth vectorizing.");
1645 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1646 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1648 DEBUG(dbgs() << "\n");
1649 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1650 << "vectorization is not beneficial "
1651 "and is not explicitly forced");
1656 // Check if it is legal to vectorize the loop.
1657 LoopVectorizationRequirements Requirements;
1658 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1659 &Requirements, &Hints);
1660 if (!LVL.canVectorize()) {
1661 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1662 emitMissedWarning(F, L, Hints);
1666 // Collect values we want to ignore in the cost model. This includes
1667 // type-promoting instructions we identified during reduction detection.
1668 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1669 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1670 for (auto &Reduction : *LVL.getReductionVars()) {
1671 RecurrenceDescriptor &RedDes = Reduction.second;
1672 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1673 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1676 // Use the cost model.
1677 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints,
1680 // Check the function attributes to find out if this function should be
1681 // optimized for size.
1682 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1685 // Compute the weighted frequency of this loop being executed and see if it
1686 // is less than 20% of the function entry baseline frequency. Note that we
1687 // always have a canonical loop here because we think we *can* vectorize.
1688 // FIXME: This is hidden behind a flag due to pervasive problems with
1689 // exactly what block frequency models.
1690 if (LoopVectorizeWithBlockFrequency) {
1691 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1692 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1693 LoopEntryFreq < ColdEntryFreq)
1697 // Check the function attributes to see if implicit floats are allowed.
1698 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1699 // an integer loop and the vector instructions selected are purely integer
1700 // vector instructions?
1701 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1702 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1703 "attribute is used.\n");
1706 VectorizationReport()
1707 << "loop not vectorized due to NoImplicitFloat attribute");
1708 emitMissedWarning(F, L, Hints);
1712 // Select the optimal vectorization factor.
1713 const LoopVectorizationCostModel::VectorizationFactor VF =
1714 CM.selectVectorizationFactor(OptForSize);
1716 // Select the interleave count.
1717 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1719 // Get user interleave count.
1720 unsigned UserIC = Hints.getInterleave();
1722 // Identify the diagnostic messages that should be produced.
1723 std::string VecDiagMsg, IntDiagMsg;
1724 bool VectorizeLoop = true, InterleaveLoop = true;
1726 if (Requirements.doesNotMeet(F, L, Hints)) {
1727 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1729 emitMissedWarning(F, L, Hints);
1733 if (VF.Width == 1) {
1734 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1736 "the cost-model indicates that vectorization is not beneficial";
1737 VectorizeLoop = false;
1740 if (IC == 1 && UserIC <= 1) {
1741 // Tell the user interleaving is not beneficial.
1742 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1744 "the cost-model indicates that interleaving is not beneficial";
1745 InterleaveLoop = false;
1748 " and is explicitly disabled or interleave count is set to 1";
1749 } else if (IC > 1 && UserIC == 1) {
1750 // Tell the user interleaving is beneficial, but it explicitly disabled.
1752 << "LV: Interleaving is beneficial but is explicitly disabled.");
1753 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1754 "but is explicitly disabled or interleave count is set to 1";
1755 InterleaveLoop = false;
1758 // Override IC if user provided an interleave count.
1759 IC = UserIC > 0 ? UserIC : IC;
1761 // Emit diagnostic messages, if any.
1762 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1763 if (!VectorizeLoop && !InterleaveLoop) {
1764 // Do not vectorize or interleaving the loop.
1765 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1766 L->getStartLoc(), VecDiagMsg);
1767 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1768 L->getStartLoc(), IntDiagMsg);
1770 } else if (!VectorizeLoop && InterleaveLoop) {
1771 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1772 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1773 L->getStartLoc(), VecDiagMsg);
1774 } else if (VectorizeLoop && !InterleaveLoop) {
1775 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1776 << DebugLocStr << '\n');
1777 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1778 L->getStartLoc(), IntDiagMsg);
1779 } else if (VectorizeLoop && InterleaveLoop) {
1780 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1781 << DebugLocStr << '\n');
1782 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1785 if (!VectorizeLoop) {
1786 assert(IC > 1 && "interleave count should not be 1 or 0");
1787 // If we decided that it is not legal to vectorize the loop then
1789 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1790 Unroller.vectorize(&LVL);
1792 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1793 Twine("interleaved loop (interleaved count: ") +
1796 // If we decided that it is *legal* to vectorize the loop then do it.
1797 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1801 // Add metadata to disable runtime unrolling scalar loop when there's no
1802 // runtime check about strides and memory. Because at this situation,
1803 // scalar loop is rarely used not worthy to be unrolled.
1804 if (!LB.IsSafetyChecksAdded())
1805 AddRuntimeUnrollDisableMetaData(L);
1807 // Report the vectorization decision.
1808 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1809 Twine("vectorized loop (vectorization width: ") +
1810 Twine(VF.Width) + ", interleaved count: " +
1814 // Mark the loop as already vectorized to avoid vectorizing again.
1815 Hints.setAlreadyVectorized();
1817 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1821 void getAnalysisUsage(AnalysisUsage &AU) const override {
1822 AU.addRequired<AssumptionCacheTracker>();
1823 AU.addRequiredID(LoopSimplifyID);
1824 AU.addRequiredID(LCSSAID);
1825 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1826 AU.addRequired<DominatorTreeWrapperPass>();
1827 AU.addRequired<LoopInfoWrapperPass>();
1828 AU.addRequired<ScalarEvolutionWrapperPass>();
1829 AU.addRequired<TargetTransformInfoWrapperPass>();
1830 AU.addRequired<AliasAnalysis>();
1831 AU.addRequired<LoopAccessAnalysis>();
1832 AU.addPreserved<LoopInfoWrapperPass>();
1833 AU.addPreserved<DominatorTreeWrapperPass>();
1834 AU.addPreserved<AliasAnalysis>();
1839 } // end anonymous namespace
1841 //===----------------------------------------------------------------------===//
1842 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1843 // LoopVectorizationCostModel.
1844 //===----------------------------------------------------------------------===//
1846 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1847 // We need to place the broadcast of invariant variables outside the loop.
1848 Instruction *Instr = dyn_cast<Instruction>(V);
1850 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1851 Instr->getParent()) != LoopVectorBody.end());
1852 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1854 // Place the code for broadcasting invariant variables in the new preheader.
1855 IRBuilder<>::InsertPointGuard Guard(Builder);
1857 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1859 // Broadcast the scalar into all locations in the vector.
1860 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1865 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1867 assert(Val->getType()->isVectorTy() && "Must be a vector");
1868 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1869 "Elem must be an integer");
1870 assert(Step->getType() == Val->getType()->getScalarType() &&
1871 "Step has wrong type");
1872 // Create the types.
1873 Type *ITy = Val->getType()->getScalarType();
1874 VectorType *Ty = cast<VectorType>(Val->getType());
1875 int VLen = Ty->getNumElements();
1876 SmallVector<Constant*, 8> Indices;
1878 // Create a vector of consecutive numbers from zero to VF.
1879 for (int i = 0; i < VLen; ++i)
1880 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1882 // Add the consecutive indices to the vector value.
1883 Constant *Cv = ConstantVector::get(Indices);
1884 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1885 Step = Builder.CreateVectorSplat(VLen, Step);
1886 assert(Step->getType() == Val->getType() && "Invalid step vec");
1887 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1888 // which can be found from the original scalar operations.
1889 Step = Builder.CreateMul(Cv, Step);
1890 return Builder.CreateAdd(Val, Step, "induction");
1893 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1894 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1895 // Make sure that the pointer does not point to structs.
1896 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1899 // If this value is a pointer induction variable we know it is consecutive.
1900 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1901 if (Phi && Inductions.count(Phi)) {
1902 InductionDescriptor II = Inductions[Phi];
1903 return II.getConsecutiveDirection();
1906 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1910 unsigned NumOperands = Gep->getNumOperands();
1911 Value *GpPtr = Gep->getPointerOperand();
1912 // If this GEP value is a consecutive pointer induction variable and all of
1913 // the indices are constant then we know it is consecutive. We can
1914 Phi = dyn_cast<PHINode>(GpPtr);
1915 if (Phi && Inductions.count(Phi)) {
1917 // Make sure that the pointer does not point to structs.
1918 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1919 if (GepPtrType->getElementType()->isAggregateType())
1922 // Make sure that all of the index operands are loop invariant.
1923 for (unsigned i = 1; i < NumOperands; ++i)
1924 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1927 InductionDescriptor II = Inductions[Phi];
1928 return II.getConsecutiveDirection();
1931 unsigned InductionOperand = getGEPInductionOperand(Gep);
1933 // Check that all of the gep indices are uniform except for our induction
1935 for (unsigned i = 0; i != NumOperands; ++i)
1936 if (i != InductionOperand &&
1937 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1940 // We can emit wide load/stores only if the last non-zero index is the
1941 // induction variable.
1942 const SCEV *Last = nullptr;
1943 if (!Strides.count(Gep))
1944 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1946 // Because of the multiplication by a stride we can have a s/zext cast.
1947 // We are going to replace this stride by 1 so the cast is safe to ignore.
1949 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1950 // %0 = trunc i64 %indvars.iv to i32
1951 // %mul = mul i32 %0, %Stride1
1952 // %idxprom = zext i32 %mul to i64 << Safe cast.
1953 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1955 Last = replaceSymbolicStrideSCEV(SE, Strides,
1956 Gep->getOperand(InductionOperand), Gep);
1957 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1959 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1963 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1964 const SCEV *Step = AR->getStepRecurrence(*SE);
1966 // The memory is consecutive because the last index is consecutive
1967 // and all other indices are loop invariant.
1970 if (Step->isAllOnesValue())
1977 bool LoopVectorizationLegality::isUniform(Value *V) {
1978 return LAI->isUniform(V);
1981 InnerLoopVectorizer::VectorParts&
1982 InnerLoopVectorizer::getVectorValue(Value *V) {
1983 assert(V != Induction && "The new induction variable should not be used.");
1984 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1986 // If we have a stride that is replaced by one, do it here.
1987 if (Legal->hasStride(V))
1988 V = ConstantInt::get(V->getType(), 1);
1990 // If we have this scalar in the map, return it.
1991 if (WidenMap.has(V))
1992 return WidenMap.get(V);
1994 // If this scalar is unknown, assume that it is a constant or that it is
1995 // loop invariant. Broadcast V and save the value for future uses.
1996 Value *B = getBroadcastInstrs(V);
1997 return WidenMap.splat(V, B);
2000 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2001 assert(Vec->getType()->isVectorTy() && "Invalid type");
2002 SmallVector<Constant*, 8> ShuffleMask;
2003 for (unsigned i = 0; i < VF; ++i)
2004 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2006 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2007 ConstantVector::get(ShuffleMask),
2011 // Get a mask to interleave \p NumVec vectors into a wide vector.
2012 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2013 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2014 // <0, 4, 1, 5, 2, 6, 3, 7>
2015 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2017 SmallVector<Constant *, 16> Mask;
2018 for (unsigned i = 0; i < VF; i++)
2019 for (unsigned j = 0; j < NumVec; j++)
2020 Mask.push_back(Builder.getInt32(j * VF + i));
2022 return ConstantVector::get(Mask);
2025 // Get the strided mask starting from index \p Start.
2026 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2027 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2028 unsigned Stride, unsigned VF) {
2029 SmallVector<Constant *, 16> Mask;
2030 for (unsigned i = 0; i < VF; i++)
2031 Mask.push_back(Builder.getInt32(Start + i * Stride));
2033 return ConstantVector::get(Mask);
2036 // Get a mask of two parts: The first part consists of sequential integers
2037 // starting from 0, The second part consists of UNDEFs.
2038 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2039 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2040 unsigned NumUndef) {
2041 SmallVector<Constant *, 16> Mask;
2042 for (unsigned i = 0; i < NumInt; i++)
2043 Mask.push_back(Builder.getInt32(i));
2045 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2046 for (unsigned i = 0; i < NumUndef; i++)
2047 Mask.push_back(Undef);
2049 return ConstantVector::get(Mask);
2052 // Concatenate two vectors with the same element type. The 2nd vector should
2053 // not have more elements than the 1st vector. If the 2nd vector has less
2054 // elements, extend it with UNDEFs.
2055 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2057 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2058 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2059 assert(VecTy1 && VecTy2 &&
2060 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2061 "Expect two vectors with the same element type");
2063 unsigned NumElts1 = VecTy1->getNumElements();
2064 unsigned NumElts2 = VecTy2->getNumElements();
2065 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2067 if (NumElts1 > NumElts2) {
2068 // Extend with UNDEFs.
2070 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2071 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2074 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2075 return Builder.CreateShuffleVector(V1, V2, Mask);
2078 // Concatenate vectors in the given list. All vectors have the same type.
2079 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2080 ArrayRef<Value *> InputList) {
2081 unsigned NumVec = InputList.size();
2082 assert(NumVec > 1 && "Should be at least two vectors");
2084 SmallVector<Value *, 8> ResList;
2085 ResList.append(InputList.begin(), InputList.end());
2087 SmallVector<Value *, 8> TmpList;
2088 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2089 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2090 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2091 "Only the last vector may have a different type");
2093 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2096 // Push the last vector if the total number of vectors is odd.
2097 if (NumVec % 2 != 0)
2098 TmpList.push_back(ResList[NumVec - 1]);
2101 NumVec = ResList.size();
2102 } while (NumVec > 1);
2107 // Try to vectorize the interleave group that \p Instr belongs to.
2109 // E.g. Translate following interleaved load group (factor = 3):
2110 // for (i = 0; i < N; i+=3) {
2111 // R = Pic[i]; // Member of index 0
2112 // G = Pic[i+1]; // Member of index 1
2113 // B = Pic[i+2]; // Member of index 2
2114 // ... // do something to R, G, B
2117 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2118 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2119 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2120 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2122 // Or translate following interleaved store group (factor = 3):
2123 // for (i = 0; i < N; i+=3) {
2124 // ... do something to R, G, B
2125 // Pic[i] = R; // Member of index 0
2126 // Pic[i+1] = G; // Member of index 1
2127 // Pic[i+2] = B; // Member of index 2
2130 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2131 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2132 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2133 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2134 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2135 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2136 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2137 assert(Group && "Fail to get an interleaved access group.");
2139 // Skip if current instruction is not the insert position.
2140 if (Instr != Group->getInsertPos())
2143 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2144 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2145 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2147 // Prepare for the vector type of the interleaved load/store.
2148 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2149 unsigned InterleaveFactor = Group->getFactor();
2150 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2151 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2153 // Prepare for the new pointers.
2154 setDebugLocFromInst(Builder, Ptr);
2155 VectorParts &PtrParts = getVectorValue(Ptr);
2156 SmallVector<Value *, 2> NewPtrs;
2157 unsigned Index = Group->getIndex(Instr);
2158 for (unsigned Part = 0; Part < UF; Part++) {
2159 // Extract the pointer for current instruction from the pointer vector. A
2160 // reverse access uses the pointer in the last lane.
2161 Value *NewPtr = Builder.CreateExtractElement(
2163 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2165 // Notice current instruction could be any index. Need to adjust the address
2166 // to the member of index 0.
2168 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2169 // b = A[i]; // Member of index 0
2170 // Current pointer is pointed to A[i+1], adjust it to A[i].
2172 // E.g. A[i+1] = a; // Member of index 1
2173 // A[i] = b; // Member of index 0
2174 // A[i+2] = c; // Member of index 2 (Current instruction)
2175 // Current pointer is pointed to A[i+2], adjust it to A[i].
2176 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2178 // Cast to the vector pointer type.
2179 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2182 setDebugLocFromInst(Builder, Instr);
2183 Value *UndefVec = UndefValue::get(VecTy);
2185 // Vectorize the interleaved load group.
2187 for (unsigned Part = 0; Part < UF; Part++) {
2188 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2189 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2191 for (unsigned i = 0; i < InterleaveFactor; i++) {
2192 Instruction *Member = Group->getMember(i);
2194 // Skip the gaps in the group.
2198 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2199 Value *StridedVec = Builder.CreateShuffleVector(
2200 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2202 // If this member has different type, cast the result type.
2203 if (Member->getType() != ScalarTy) {
2204 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2205 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2208 VectorParts &Entry = WidenMap.get(Member);
2210 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2213 propagateMetadata(NewLoadInstr, Instr);
2218 // The sub vector type for current instruction.
2219 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2221 // Vectorize the interleaved store group.
2222 for (unsigned Part = 0; Part < UF; Part++) {
2223 // Collect the stored vector from each member.
2224 SmallVector<Value *, 4> StoredVecs;
2225 for (unsigned i = 0; i < InterleaveFactor; i++) {
2226 // Interleaved store group doesn't allow a gap, so each index has a member
2227 Instruction *Member = Group->getMember(i);
2228 assert(Member && "Fail to get a member from an interleaved store group");
2231 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2232 if (Group->isReverse())
2233 StoredVec = reverseVector(StoredVec);
2235 // If this member has different type, cast it to an unified type.
2236 if (StoredVec->getType() != SubVT)
2237 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2239 StoredVecs.push_back(StoredVec);
2242 // Concatenate all vectors into a wide vector.
2243 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2245 // Interleave the elements in the wide vector.
2246 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2247 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2250 Instruction *NewStoreInstr =
2251 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2252 propagateMetadata(NewStoreInstr, Instr);
2256 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2257 // Attempt to issue a wide load.
2258 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2259 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2261 assert((LI || SI) && "Invalid Load/Store instruction");
2263 // Try to vectorize the interleave group if this access is interleaved.
2264 if (Legal->isAccessInterleaved(Instr))
2265 return vectorizeInterleaveGroup(Instr);
2267 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2268 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2269 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2270 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2271 // An alignment of 0 means target abi alignment. We need to use the scalar's
2272 // target abi alignment in such a case.
2273 const DataLayout &DL = Instr->getModule()->getDataLayout();
2275 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2276 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2277 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2278 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2280 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2281 !Legal->isMaskRequired(SI))
2282 return scalarizeInstruction(Instr, true);
2284 if (ScalarAllocatedSize != VectorElementSize)
2285 return scalarizeInstruction(Instr);
2287 // If the pointer is loop invariant or if it is non-consecutive,
2288 // scalarize the load.
2289 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2290 bool Reverse = ConsecutiveStride < 0;
2291 bool UniformLoad = LI && Legal->isUniform(Ptr);
2292 if (!ConsecutiveStride || UniformLoad)
2293 return scalarizeInstruction(Instr);
2295 Constant *Zero = Builder.getInt32(0);
2296 VectorParts &Entry = WidenMap.get(Instr);
2298 // Handle consecutive loads/stores.
2299 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2300 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2301 setDebugLocFromInst(Builder, Gep);
2302 Value *PtrOperand = Gep->getPointerOperand();
2303 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2304 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2306 // Create the new GEP with the new induction variable.
2307 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2308 Gep2->setOperand(0, FirstBasePtr);
2309 Gep2->setName("gep.indvar.base");
2310 Ptr = Builder.Insert(Gep2);
2312 setDebugLocFromInst(Builder, Gep);
2313 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2314 OrigLoop) && "Base ptr must be invariant");
2316 // The last index does not have to be the induction. It can be
2317 // consecutive and be a function of the index. For example A[I+1];
2318 unsigned NumOperands = Gep->getNumOperands();
2319 unsigned InductionOperand = getGEPInductionOperand(Gep);
2320 // Create the new GEP with the new induction variable.
2321 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2323 for (unsigned i = 0; i < NumOperands; ++i) {
2324 Value *GepOperand = Gep->getOperand(i);
2325 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2327 // Update last index or loop invariant instruction anchored in loop.
2328 if (i == InductionOperand ||
2329 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2330 assert((i == InductionOperand ||
2331 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2332 "Must be last index or loop invariant");
2334 VectorParts &GEPParts = getVectorValue(GepOperand);
2335 Value *Index = GEPParts[0];
2336 Index = Builder.CreateExtractElement(Index, Zero);
2337 Gep2->setOperand(i, Index);
2338 Gep2->setName("gep.indvar.idx");
2341 Ptr = Builder.Insert(Gep2);
2343 // Use the induction element ptr.
2344 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2345 setDebugLocFromInst(Builder, Ptr);
2346 VectorParts &PtrVal = getVectorValue(Ptr);
2347 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2350 VectorParts Mask = createBlockInMask(Instr->getParent());
2353 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2354 "We do not allow storing to uniform addresses");
2355 setDebugLocFromInst(Builder, SI);
2356 // We don't want to update the value in the map as it might be used in
2357 // another expression. So don't use a reference type for "StoredVal".
2358 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2360 for (unsigned Part = 0; Part < UF; ++Part) {
2361 // Calculate the pointer for the specific unroll-part.
2363 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2366 // If we store to reverse consecutive memory locations, then we need
2367 // to reverse the order of elements in the stored value.
2368 StoredVal[Part] = reverseVector(StoredVal[Part]);
2369 // If the address is consecutive but reversed, then the
2370 // wide store needs to start at the last vector element.
2371 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2372 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2373 Mask[Part] = reverseVector(Mask[Part]);
2376 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2377 DataTy->getPointerTo(AddressSpace));
2380 if (Legal->isMaskRequired(SI))
2381 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2384 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2385 propagateMetadata(NewSI, SI);
2391 assert(LI && "Must have a load instruction");
2392 setDebugLocFromInst(Builder, LI);
2393 for (unsigned Part = 0; Part < UF; ++Part) {
2394 // Calculate the pointer for the specific unroll-part.
2396 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2399 // If the address is consecutive but reversed, then the
2400 // wide load needs to start at the last vector element.
2401 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2402 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2403 Mask[Part] = reverseVector(Mask[Part]);
2407 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2408 DataTy->getPointerTo(AddressSpace));
2409 if (Legal->isMaskRequired(LI))
2410 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2411 UndefValue::get(DataTy),
2412 "wide.masked.load");
2414 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2415 propagateMetadata(NewLI, LI);
2416 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2420 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2421 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2422 // Holds vector parameters or scalars, in case of uniform vals.
2423 SmallVector<VectorParts, 4> Params;
2425 setDebugLocFromInst(Builder, Instr);
2427 // Find all of the vectorized parameters.
2428 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2429 Value *SrcOp = Instr->getOperand(op);
2431 // If we are accessing the old induction variable, use the new one.
2432 if (SrcOp == OldInduction) {
2433 Params.push_back(getVectorValue(SrcOp));
2437 // Try using previously calculated values.
2438 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2440 // If the src is an instruction that appeared earlier in the basic block,
2441 // then it should already be vectorized.
2442 if (SrcInst && OrigLoop->contains(SrcInst)) {
2443 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2444 // The parameter is a vector value from earlier.
2445 Params.push_back(WidenMap.get(SrcInst));
2447 // The parameter is a scalar from outside the loop. Maybe even a constant.
2448 VectorParts Scalars;
2449 Scalars.append(UF, SrcOp);
2450 Params.push_back(Scalars);
2454 assert(Params.size() == Instr->getNumOperands() &&
2455 "Invalid number of operands");
2457 // Does this instruction return a value ?
2458 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2460 Value *UndefVec = IsVoidRetTy ? nullptr :
2461 UndefValue::get(VectorType::get(Instr->getType(), VF));
2462 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2463 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2465 Instruction *InsertPt = Builder.GetInsertPoint();
2466 BasicBlock *IfBlock = Builder.GetInsertBlock();
2467 BasicBlock *CondBlock = nullptr;
2470 Loop *VectorLp = nullptr;
2471 if (IfPredicateStore) {
2472 assert(Instr->getParent()->getSinglePredecessor() &&
2473 "Only support single predecessor blocks");
2474 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2475 Instr->getParent());
2476 VectorLp = LI->getLoopFor(IfBlock);
2477 assert(VectorLp && "Must have a loop for this block");
2480 // For each vector unroll 'part':
2481 for (unsigned Part = 0; Part < UF; ++Part) {
2482 // For each scalar that we create:
2483 for (unsigned Width = 0; Width < VF; ++Width) {
2486 Value *Cmp = nullptr;
2487 if (IfPredicateStore) {
2488 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2489 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2490 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2491 LoopVectorBody.push_back(CondBlock);
2492 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2493 // Update Builder with newly created basic block.
2494 Builder.SetInsertPoint(InsertPt);
2497 Instruction *Cloned = Instr->clone();
2499 Cloned->setName(Instr->getName() + ".cloned");
2500 // Replace the operands of the cloned instructions with extracted scalars.
2501 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2502 Value *Op = Params[op][Part];
2503 // Param is a vector. Need to extract the right lane.
2504 if (Op->getType()->isVectorTy())
2505 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2506 Cloned->setOperand(op, Op);
2509 // Place the cloned scalar in the new loop.
2510 Builder.Insert(Cloned);
2512 // If the original scalar returns a value we need to place it in a vector
2513 // so that future users will be able to use it.
2515 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2516 Builder.getInt32(Width));
2518 if (IfPredicateStore) {
2519 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2520 LoopVectorBody.push_back(NewIfBlock);
2521 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2522 Builder.SetInsertPoint(InsertPt);
2523 ReplaceInstWithInst(IfBlock->getTerminator(),
2524 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2525 IfBlock = NewIfBlock;
2531 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2535 if (Instruction *I = dyn_cast<Instruction>(V))
2536 return I->getParent() == Loc->getParent() ? I : nullptr;
2540 std::pair<Instruction *, Instruction *>
2541 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2542 Instruction *tnullptr = nullptr;
2543 if (!Legal->mustCheckStrides())
2544 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2546 IRBuilder<> ChkBuilder(Loc);
2549 Value *Check = nullptr;
2550 Instruction *FirstInst = nullptr;
2551 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2552 SE = Legal->strides_end();
2554 Value *Ptr = stripIntegerCast(*SI);
2555 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2557 // Store the first instruction we create.
2558 FirstInst = getFirstInst(FirstInst, C, Loc);
2560 Check = ChkBuilder.CreateOr(Check, C);
2565 // We have to do this trickery because the IRBuilder might fold the check to a
2566 // constant expression in which case there is no Instruction anchored in a
2568 LLVMContext &Ctx = Loc->getContext();
2569 Instruction *TheCheck =
2570 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2571 ChkBuilder.Insert(TheCheck, "stride.not.one");
2572 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2574 return std::make_pair(FirstInst, TheCheck);
2577 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L,
2582 BasicBlock *Header = L->getHeader();
2583 BasicBlock *Latch = L->getLoopLatch();
2584 // As we're just creating this loop, it's possible no latch exists
2585 // yet. If so, use the header as this will be a single block loop.
2589 IRBuilder<> Builder(Header->getFirstInsertionPt());
2590 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2591 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2593 Builder.SetInsertPoint(Latch->getTerminator());
2595 // Create i+1 and fill the PHINode.
2596 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2597 Induction->addIncoming(Start, L->getLoopPreheader());
2598 Induction->addIncoming(Next, Latch);
2599 // Create the compare.
2600 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2601 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2603 // Now we have two terminators. Remove the old one from the block.
2604 Latch->getTerminator()->eraseFromParent();
2609 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2613 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2614 // Find the loop boundaries.
2615 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2616 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2618 Type *IdxTy = Legal->getWidestInductionType();
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 = L->getHeader()->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 // Count holds the overall loop count (N).
2641 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2642 L->getLoopPreheader()->getTerminator());
2644 if (TripCount->getType()->isPointerTy())
2646 CastInst::CreatePointerCast(TripCount, IdxTy,
2647 "exitcount.ptrcnt.to.int",
2648 L->getLoopPreheader()->getTerminator());
2653 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2654 if (VectorTripCount)
2655 return VectorTripCount;
2657 Value *TC = getOrCreateTripCount(L);
2658 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2660 // Now we need to generate the expression for N - (N % VF), which is
2661 // the part that the vectorized body will execute.
2662 // The loop step is equal to the vectorization factor (num of SIMD elements)
2663 // times the unroll factor (num of SIMD instructions).
2664 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2665 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2666 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2668 return VectorTripCount;
2671 void InnerLoopVectorizer::createEmptyLoop() {
2673 In this function we generate a new loop. The new loop will contain
2674 the vectorized instructions while the old loop will continue to run the
2677 [ ] <-- loop iteration number check.
2680 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2683 || [ ] <-- vector pre header.
2687 || [ ]_| <-- vector loop.
2690 | >[ ] <--- middle-block.
2693 -|- >[ ] <--- new preheader.
2697 | [ ]_| <-- old scalar loop to handle remainder.
2700 >[ ] <-- exit block.
2704 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2705 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2706 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2707 assert(VectorPH && "Invalid loop structure");
2708 assert(ExitBlock && "Must have an exit block");
2710 // Some loops have a single integer induction variable, while other loops
2711 // don't. One example is c++ iterators that often have multiple pointer
2712 // induction variables. In the code below we also support a case where we
2713 // don't have a single induction variable.
2715 // We try to obtain an induction variable from the original loop as hard
2716 // as possible. However if we don't find one that:
2718 // - counts from zero, stepping by one
2719 // - is the size of the widest induction variable type
2720 // then we create a new one.
2721 OldInduction = Legal->getInduction();
2722 Type *IdxTy = Legal->getWidestInductionType();
2724 // Split the single block loop into the two loop structure described above.
2725 BasicBlock *VecBody =
2726 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2727 BasicBlock *MiddleBlock =
2728 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2729 BasicBlock *ScalarPH =
2730 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2732 // Create and register the new vector loop.
2733 Loop* Lp = new Loop();
2734 Loop *ParentLoop = OrigLoop->getParentLoop();
2736 // Insert the new loop into the loop nest and register the new basic blocks
2737 // before calling any utilities such as SCEV that require valid LoopInfo.
2739 ParentLoop->addChildLoop(Lp);
2740 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2741 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2743 LI->addTopLevelLoop(Lp);
2745 Lp->addBasicBlockToLoop(VecBody, *LI);
2747 // Find the loop boundaries.
2748 Value *Count = getOrCreateTripCount(Lp);
2750 // The loop minimum iterations check below is to ensure the loop has enough
2751 // trip count so the generated vector loop will likely be executed and the
2752 // preparation and rounding-off costs will likely be worthy.
2754 // The minimum iteration check also covers case where the backedge-taken
2755 // count is uint##_max. Adding one to it will cause overflow and an
2756 // incorrect loop trip count being generated in the vector body. In this
2757 // case we also want to directly jump to the scalar remainder loop.
2758 Instruction *CheckMinIters =
2759 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULT, Count,
2760 ConstantInt::get(Count->getType(), VF * UF),
2761 "min.iters.check", VectorPH->getTerminator());
2763 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2765 LoopBypassBlocks.push_back(VectorPH);
2767 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2769 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2771 // Generate code to check that the loop's trip count is not less than the
2772 // minimum loop iteration number threshold.
2773 BasicBlock *NewVectorPH =
2774 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "min.iters.checked");
2776 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2777 ReplaceInstWithInst(VectorPH->getTerminator(),
2778 BranchInst::Create(ScalarPH, NewVectorPH, CheckMinIters));
2779 VectorPH = NewVectorPH;
2781 // This is the IR builder that we use to add all of the logic for bypassing
2782 // the new vector loop.
2783 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2784 setDebugLocFromInst(BypassBuilder,
2785 getDebugLocFromInstOrOperands(OldInduction));
2787 // Add the start index to the loop count to get the new end index.
2788 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
2790 // Generate the induction variable.
2791 // The loop step is equal to the vectorization factor (num of SIMD elements)
2792 // times the unroll factor (num of SIMD instructions).
2793 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2795 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
2796 getDebugLocFromInstOrOperands(OldInduction));
2798 // Now, compare the new count to zero. If it is zero skip the vector loop and
2799 // jump to the scalar loop.
2801 BypassBuilder.CreateICmpEQ(CountRoundDown, StartIdx, "cmp.zero");
2803 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2805 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2806 LoopBypassBlocks.push_back(VectorPH);
2807 ReplaceInstWithInst(VectorPH->getTerminator(),
2808 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2809 VectorPH = NewVectorPH;
2811 // Generate the code to check that the strides we assumed to be one are really
2812 // one. We want the new basic block to start at the first instruction in a
2813 // sequence of instructions that form a check.
2814 Instruction *StrideCheck;
2815 Instruction *FirstCheckInst;
2816 std::tie(FirstCheckInst, StrideCheck) =
2817 addStrideCheck(VectorPH->getTerminator());
2819 AddedSafetyChecks = true;
2820 // Create a new block containing the stride check.
2821 VectorPH->setName("vector.stridecheck");
2823 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2825 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2826 LoopBypassBlocks.push_back(VectorPH);
2828 // Replace the branch into the memory check block with a conditional branch
2829 // for the "few elements case".
2830 ReplaceInstWithInst(
2831 VectorPH->getTerminator(),
2832 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2834 VectorPH = NewVectorPH;
2837 // Generate the code that checks in runtime if arrays overlap. We put the
2838 // checks into a separate block to make the more common case of few elements
2840 Instruction *MemRuntimeCheck;
2841 std::tie(FirstCheckInst, MemRuntimeCheck) =
2842 Legal->getLAI()->addRuntimeChecks(VectorPH->getTerminator());
2843 if (MemRuntimeCheck) {
2844 AddedSafetyChecks = true;
2845 // Create a new block containing the memory check.
2846 VectorPH->setName("vector.memcheck");
2848 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2850 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2851 LoopBypassBlocks.push_back(VectorPH);
2853 // Replace the branch into the memory check block with a conditional branch
2854 // for the "few elements case".
2855 ReplaceInstWithInst(
2856 VectorPH->getTerminator(),
2857 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2859 VectorPH = NewVectorPH;
2862 // We are going to resume the execution of the scalar loop.
2863 // Go over all of the induction variables that we found and fix the
2864 // PHIs that are left in the scalar version of the loop.
2865 // The starting values of PHI nodes depend on the counter of the last
2866 // iteration in the vectorized loop.
2867 // If we come from a bypass edge then we need to start from the original
2870 // This variable saves the new starting index for the scalar loop. It is used
2871 // to test if there are any tail iterations left once the vector loop has
2873 PHINode *ResumeIndex = nullptr;
2874 LoopVectorizationLegality::InductionList::iterator I, E;
2875 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2876 // Set builder to point to last bypass block.
2877 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2878 for (I = List->begin(), E = List->end(); I != E; ++I) {
2879 PHINode *OrigPhi = I->first;
2880 InductionDescriptor II = I->second;
2882 PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
2883 MiddleBlock->getTerminator());
2884 // Create phi nodes to merge from the backedge-taken check block.
2885 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2887 ScalarPH->getTerminator());
2888 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2891 if (OrigPhi == OldInduction) {
2892 // We know what the end value is.
2893 EndValue = CountRoundDown;
2894 // We also know which PHI node holds it.
2895 ResumeIndex = ResumeVal;
2897 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2898 II.getStepValue()->getType(),
2900 EndValue = II.transform(BypassBuilder, CRD);
2901 EndValue->setName("ind.end");
2904 // The new PHI merges the original incoming value, in case of a bypass,
2905 // or the value at the end of the vectorized loop.
2906 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2907 ResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2908 ResumeVal->addIncoming(EndValue, VecBody);
2910 // Fix the scalar body counter (PHI node).
2911 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2913 // The old induction's phi node in the scalar body needs the truncated
2915 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[0]);
2916 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2919 // If we are generating a new induction variable then we also need to
2920 // generate the code that calculates the exit value. This value is not
2921 // simply the end of the counter because we may skip the vectorized body
2922 // in case of a runtime check.
2924 assert(!ResumeIndex && "Unexpected resume value found");
2925 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2926 MiddleBlock->getTerminator());
2927 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2928 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2929 ResumeIndex->addIncoming(CountRoundDown, VecBody);
2932 // Make sure that we found the index where scalar loop needs to continue.
2933 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2934 "Invalid resume Index");
2936 // Add a check in the middle block to see if we have completed
2937 // all of the iterations in the first vector loop.
2938 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2939 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
2940 ResumeIndex, "cmp.n",
2941 MiddleBlock->getTerminator());
2942 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2943 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2945 // Get ready to start creating new instructions into the vectorized body.
2946 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2949 LoopVectorPreHeader = VectorPH;
2950 LoopScalarPreHeader = ScalarPH;
2951 LoopMiddleBlock = MiddleBlock;
2952 LoopExitBlock = ExitBlock;
2953 LoopVectorBody.push_back(VecBody);
2954 LoopScalarBody = OldBasicBlock;
2956 LoopVectorizeHints Hints(Lp, true);
2957 Hints.setAlreadyVectorized();
2961 struct CSEDenseMapInfo {
2962 static bool canHandle(Instruction *I) {
2963 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2964 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2966 static inline Instruction *getEmptyKey() {
2967 return DenseMapInfo<Instruction *>::getEmptyKey();
2969 static inline Instruction *getTombstoneKey() {
2970 return DenseMapInfo<Instruction *>::getTombstoneKey();
2972 static unsigned getHashValue(Instruction *I) {
2973 assert(canHandle(I) && "Unknown instruction!");
2974 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2975 I->value_op_end()));
2977 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2978 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2979 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2981 return LHS->isIdenticalTo(RHS);
2986 /// \brief Check whether this block is a predicated block.
2987 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2988 /// = ...; " blocks. We start with one vectorized basic block. For every
2989 /// conditional block we split this vectorized block. Therefore, every second
2990 /// block will be a predicated one.
2991 static bool isPredicatedBlock(unsigned BlockNum) {
2992 return BlockNum % 2;
2995 ///\brief Perform cse of induction variable instructions.
2996 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2997 // Perform simple cse.
2998 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2999 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3000 BasicBlock *BB = BBs[i];
3001 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3002 Instruction *In = I++;
3004 if (!CSEDenseMapInfo::canHandle(In))
3007 // Check if we can replace this instruction with any of the
3008 // visited instructions.
3009 if (Instruction *V = CSEMap.lookup(In)) {
3010 In->replaceAllUsesWith(V);
3011 In->eraseFromParent();
3014 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3015 // ...;" blocks for predicated stores. Every second block is a predicated
3017 if (isPredicatedBlock(i))
3025 /// \brief Adds a 'fast' flag to floating point operations.
3026 static Value *addFastMathFlag(Value *V) {
3027 if (isa<FPMathOperator>(V)){
3028 FastMathFlags Flags;
3029 Flags.setUnsafeAlgebra();
3030 cast<Instruction>(V)->setFastMathFlags(Flags);
3035 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3036 /// the result needs to be inserted and/or extracted from vectors.
3037 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3038 const TargetTransformInfo &TTI) {
3042 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3045 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3047 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3049 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3055 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3056 // Return the cost of the instruction, including scalarization overhead if it's
3057 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3058 // i.e. either vector version isn't available, or is too expensive.
3059 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3060 const TargetTransformInfo &TTI,
3061 const TargetLibraryInfo *TLI,
3062 bool &NeedToScalarize) {
3063 Function *F = CI->getCalledFunction();
3064 StringRef FnName = CI->getCalledFunction()->getName();
3065 Type *ScalarRetTy = CI->getType();
3066 SmallVector<Type *, 4> Tys, ScalarTys;
3067 for (auto &ArgOp : CI->arg_operands())
3068 ScalarTys.push_back(ArgOp->getType());
3070 // Estimate cost of scalarized vector call. The source operands are assumed
3071 // to be vectors, so we need to extract individual elements from there,
3072 // execute VF scalar calls, and then gather the result into the vector return
3074 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3076 return ScalarCallCost;
3078 // Compute corresponding vector type for return value and arguments.
3079 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3080 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3081 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3083 // Compute costs of unpacking argument values for the scalar calls and
3084 // packing the return values to a vector.
3085 unsigned ScalarizationCost =
3086 getScalarizationOverhead(RetTy, true, false, TTI);
3087 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3088 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3090 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3092 // If we can't emit a vector call for this function, then the currently found
3093 // cost is the cost we need to return.
3094 NeedToScalarize = true;
3095 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3098 // If the corresponding vector cost is cheaper, return its cost.
3099 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3100 if (VectorCallCost < Cost) {
3101 NeedToScalarize = false;
3102 return VectorCallCost;
3107 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3108 // factor VF. Return the cost of the instruction, including scalarization
3109 // overhead if it's needed.
3110 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3111 const TargetTransformInfo &TTI,
3112 const TargetLibraryInfo *TLI) {
3113 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3114 assert(ID && "Expected intrinsic call!");
3116 Type *RetTy = ToVectorTy(CI->getType(), VF);
3117 SmallVector<Type *, 4> Tys;
3118 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3119 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3121 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3124 void InnerLoopVectorizer::vectorizeLoop() {
3125 //===------------------------------------------------===//
3127 // Notice: any optimization or new instruction that go
3128 // into the code below should be also be implemented in
3131 //===------------------------------------------------===//
3132 Constant *Zero = Builder.getInt32(0);
3134 // In order to support reduction variables we need to be able to vectorize
3135 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3136 // stages. First, we create a new vector PHI node with no incoming edges.
3137 // We use this value when we vectorize all of the instructions that use the
3138 // PHI. Next, after all of the instructions in the block are complete we
3139 // add the new incoming edges to the PHI. At this point all of the
3140 // instructions in the basic block are vectorized, so we can use them to
3141 // construct the PHI.
3142 PhiVector RdxPHIsToFix;
3144 // Scan the loop in a topological order to ensure that defs are vectorized
3146 LoopBlocksDFS DFS(OrigLoop);
3149 // Vectorize all of the blocks in the original loop.
3150 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3151 be = DFS.endRPO(); bb != be; ++bb)
3152 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3154 // At this point every instruction in the original loop is widened to
3155 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3156 // that we vectorized. The PHI nodes are currently empty because we did
3157 // not want to introduce cycles. Notice that the remaining PHI nodes
3158 // that we need to fix are reduction variables.
3160 // Create the 'reduced' values for each of the induction vars.
3161 // The reduced values are the vector values that we scalarize and combine
3162 // after the loop is finished.
3163 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3165 PHINode *RdxPhi = *it;
3166 assert(RdxPhi && "Unable to recover vectorized PHI");
3168 // Find the reduction variable descriptor.
3169 assert(Legal->getReductionVars()->count(RdxPhi) &&
3170 "Unable to find the reduction variable");
3171 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3173 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3174 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3175 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3176 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3177 RdxDesc.getMinMaxRecurrenceKind();
3178 setDebugLocFromInst(Builder, ReductionStartValue);
3180 // We need to generate a reduction vector from the incoming scalar.
3181 // To do so, we need to generate the 'identity' vector and override
3182 // one of the elements with the incoming scalar reduction. We need
3183 // to do it in the vector-loop preheader.
3184 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3186 // This is the vector-clone of the value that leaves the loop.
3187 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3188 Type *VecTy = VectorExit[0]->getType();
3190 // Find the reduction identity variable. Zero for addition, or, xor,
3191 // one for multiplication, -1 for And.
3194 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3195 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3196 // MinMax reduction have the start value as their identify.
3198 VectorStart = Identity = ReductionStartValue;
3200 VectorStart = Identity =
3201 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3204 // Handle other reduction kinds:
3205 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3206 RK, VecTy->getScalarType());
3209 // This vector is the Identity vector where the first element is the
3210 // incoming scalar reduction.
3211 VectorStart = ReductionStartValue;
3213 Identity = ConstantVector::getSplat(VF, Iden);
3215 // This vector is the Identity vector where the first element is the
3216 // incoming scalar reduction.
3218 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3222 // Fix the vector-loop phi.
3224 // Reductions do not have to start at zero. They can start with
3225 // any loop invariant values.
3226 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3227 BasicBlock *Latch = OrigLoop->getLoopLatch();
3228 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3229 VectorParts &Val = getVectorValue(LoopVal);
3230 for (unsigned part = 0; part < UF; ++part) {
3231 // Make sure to add the reduction stat value only to the
3232 // first unroll part.
3233 Value *StartVal = (part == 0) ? VectorStart : Identity;
3234 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3235 LoopVectorPreHeader);
3236 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3237 LoopVectorBody.back());
3240 // Before each round, move the insertion point right between
3241 // the PHIs and the values we are going to write.
3242 // This allows us to write both PHINodes and the extractelement
3244 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3246 VectorParts RdxParts, &RdxExitVal = getVectorValue(LoopExitInst);
3247 setDebugLocFromInst(Builder, LoopExitInst);
3248 for (unsigned part = 0; part < UF; ++part) {
3249 // This PHINode contains the vectorized reduction variable, or
3250 // the initial value vector, if we bypass the vector loop.
3251 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3252 Value *StartVal = (part == 0) ? VectorStart : Identity;
3253 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3254 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3255 NewPhi->addIncoming(RdxExitVal[part],
3256 LoopVectorBody.back());
3257 RdxParts.push_back(NewPhi);
3260 // If the vector reduction can be performed in a smaller type, we truncate
3261 // then extend the loop exit value to enable InstCombine to evaluate the
3262 // entire expression in the smaller type.
3263 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3264 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3265 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3266 for (unsigned part = 0; part < UF; ++part) {
3267 Value *Trunc = Builder.CreateTrunc(RdxExitVal[part], RdxVecTy);
3268 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3269 : Builder.CreateZExt(Trunc, VecTy);
3270 for (Value::user_iterator UI = RdxExitVal[part]->user_begin();
3271 UI != RdxExitVal[part]->user_end();)
3273 (*UI++)->replaceUsesOfWith(RdxExitVal[part], Extnd);
3277 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3278 for (unsigned part = 0; part < UF; ++part)
3279 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3282 // Reduce all of the unrolled parts into a single vector.
3283 Value *ReducedPartRdx = RdxParts[0];
3284 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3285 setDebugLocFromInst(Builder, ReducedPartRdx);
3286 for (unsigned part = 1; part < UF; ++part) {
3287 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3288 // Floating point operations had to be 'fast' to enable the reduction.
3289 ReducedPartRdx = addFastMathFlag(
3290 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3291 ReducedPartRdx, "bin.rdx"));
3293 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3294 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3298 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3299 // and vector ops, reducing the set of values being computed by half each
3301 assert(isPowerOf2_32(VF) &&
3302 "Reduction emission only supported for pow2 vectors!");
3303 Value *TmpVec = ReducedPartRdx;
3304 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3305 for (unsigned i = VF; i != 1; i >>= 1) {
3306 // Move the upper half of the vector to the lower half.
3307 for (unsigned j = 0; j != i/2; ++j)
3308 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3310 // Fill the rest of the mask with undef.
3311 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3312 UndefValue::get(Builder.getInt32Ty()));
3315 Builder.CreateShuffleVector(TmpVec,
3316 UndefValue::get(TmpVec->getType()),
3317 ConstantVector::get(ShuffleMask),
3320 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3321 // Floating point operations had to be 'fast' to enable the reduction.
3322 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3323 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3325 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3329 // The result is in the first element of the vector.
3330 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3331 Builder.getInt32(0));
3333 // If the reduction can be performed in a smaller type, we need to extend
3334 // the reduction to the wider type before we branch to the original loop.
3335 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3338 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3339 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3342 // Create a phi node that merges control-flow from the backedge-taken check
3343 // block and the middle block.
3344 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3345 LoopScalarPreHeader->getTerminator());
3346 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3347 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3349 // Now, we need to fix the users of the reduction variable
3350 // inside and outside of the scalar remainder loop.
3351 // We know that the loop is in LCSSA form. We need to update the
3352 // PHI nodes in the exit blocks.
3353 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3354 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3355 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3356 if (!LCSSAPhi) break;
3358 // All PHINodes need to have a single entry edge, or two if
3359 // we already fixed them.
3360 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3362 // We found our reduction value exit-PHI. Update it with the
3363 // incoming bypass edge.
3364 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3365 // Add an edge coming from the bypass.
3366 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3369 }// end of the LCSSA phi scan.
3371 // Fix the scalar loop reduction variable with the incoming reduction sum
3372 // from the vector body and from the backedge value.
3373 int IncomingEdgeBlockIdx =
3374 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3375 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3376 // Pick the other block.
3377 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3378 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3379 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3380 }// end of for each redux variable.
3384 // Remove redundant induction instructions.
3385 cse(LoopVectorBody);
3388 void InnerLoopVectorizer::fixLCSSAPHIs() {
3389 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3390 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3391 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3392 if (!LCSSAPhi) break;
3393 if (LCSSAPhi->getNumIncomingValues() == 1)
3394 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3399 InnerLoopVectorizer::VectorParts
3400 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3401 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3404 // Look for cached value.
3405 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3406 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3407 if (ECEntryIt != MaskCache.end())
3408 return ECEntryIt->second;
3410 VectorParts SrcMask = createBlockInMask(Src);
3412 // The terminator has to be a branch inst!
3413 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3414 assert(BI && "Unexpected terminator found");
3416 if (BI->isConditional()) {
3417 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3419 if (BI->getSuccessor(0) != Dst)
3420 for (unsigned part = 0; part < UF; ++part)
3421 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3423 for (unsigned part = 0; part < UF; ++part)
3424 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3426 MaskCache[Edge] = EdgeMask;
3430 MaskCache[Edge] = SrcMask;
3434 InnerLoopVectorizer::VectorParts
3435 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3436 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3438 // Loop incoming mask is all-one.
3439 if (OrigLoop->getHeader() == BB) {
3440 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3441 return getVectorValue(C);
3444 // This is the block mask. We OR all incoming edges, and with zero.
3445 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3446 VectorParts BlockMask = getVectorValue(Zero);
3449 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3450 VectorParts EM = createEdgeMask(*it, BB);
3451 for (unsigned part = 0; part < UF; ++part)
3452 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3458 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3459 InnerLoopVectorizer::VectorParts &Entry,
3460 unsigned UF, unsigned VF, PhiVector *PV) {
3461 PHINode* P = cast<PHINode>(PN);
3462 // Handle reduction variables:
3463 if (Legal->getReductionVars()->count(P)) {
3464 for (unsigned part = 0; part < UF; ++part) {
3465 // This is phase one of vectorizing PHIs.
3466 Type *VecTy = (VF == 1) ? PN->getType() :
3467 VectorType::get(PN->getType(), VF);
3468 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3469 LoopVectorBody.back()-> getFirstInsertionPt());
3475 setDebugLocFromInst(Builder, P);
3476 // Check for PHI nodes that are lowered to vector selects.
3477 if (P->getParent() != OrigLoop->getHeader()) {
3478 // We know that all PHIs in non-header blocks are converted into
3479 // selects, so we don't have to worry about the insertion order and we
3480 // can just use the builder.
3481 // At this point we generate the predication tree. There may be
3482 // duplications since this is a simple recursive scan, but future
3483 // optimizations will clean it up.
3485 unsigned NumIncoming = P->getNumIncomingValues();
3487 // Generate a sequence of selects of the form:
3488 // SELECT(Mask3, In3,
3489 // SELECT(Mask2, In2,
3491 for (unsigned In = 0; In < NumIncoming; In++) {
3492 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3494 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3496 for (unsigned part = 0; part < UF; ++part) {
3497 // We might have single edge PHIs (blocks) - use an identity
3498 // 'select' for the first PHI operand.
3500 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3503 // Select between the current value and the previous incoming edge
3504 // based on the incoming mask.
3505 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3506 Entry[part], "predphi");
3512 // This PHINode must be an induction variable.
3513 // Make sure that we know about it.
3514 assert(Legal->getInductionVars()->count(P) &&
3515 "Not an induction variable");
3517 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3519 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3520 // which can be found from the original scalar operations.
3521 switch (II.getKind()) {
3522 case InductionDescriptor::IK_NoInduction:
3523 llvm_unreachable("Unknown induction");
3524 case InductionDescriptor::IK_IntInduction: {
3525 assert(P->getType() == II.getStartValue()->getType() && "Types must match");
3526 // Handle other induction variables that are now based on the
3528 Value *V = Induction;
3529 if (P != OldInduction) {
3530 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3531 V = II.transform(Builder, V);
3532 V->setName("offset.idx");
3534 Value *Broadcasted = getBroadcastInstrs(V);
3535 // After broadcasting the induction variable we need to make the vector
3536 // consecutive by adding 0, 1, 2, etc.
3537 for (unsigned part = 0; part < UF; ++part)
3538 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3541 case InductionDescriptor::IK_PtrInduction:
3542 // Handle the pointer induction variable case.
3543 assert(P->getType()->isPointerTy() && "Unexpected type.");
3544 // This is the normalized GEP that starts counting at zero.
3545 Value *PtrInd = Induction;
3546 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3547 // This is the vector of results. Notice that we don't generate
3548 // vector geps because scalar geps result in better code.
3549 for (unsigned part = 0; part < UF; ++part) {
3551 int EltIndex = part;
3552 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3553 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3554 Value *SclrGep = II.transform(Builder, GlobalIdx);
3555 SclrGep->setName("next.gep");
3556 Entry[part] = SclrGep;
3560 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3561 for (unsigned int i = 0; i < VF; ++i) {
3562 int EltIndex = i + part * VF;
3563 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3564 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3565 Value *SclrGep = II.transform(Builder, GlobalIdx);
3566 SclrGep->setName("next.gep");
3567 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3568 Builder.getInt32(i),
3571 Entry[part] = VecVal;
3577 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3578 // For each instruction in the old loop.
3579 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3580 VectorParts &Entry = WidenMap.get(it);
3581 switch (it->getOpcode()) {
3582 case Instruction::Br:
3583 // Nothing to do for PHIs and BR, since we already took care of the
3584 // loop control flow instructions.
3586 case Instruction::PHI: {
3587 // Vectorize PHINodes.
3588 widenPHIInstruction(it, Entry, UF, VF, PV);
3592 case Instruction::Add:
3593 case Instruction::FAdd:
3594 case Instruction::Sub:
3595 case Instruction::FSub:
3596 case Instruction::Mul:
3597 case Instruction::FMul:
3598 case Instruction::UDiv:
3599 case Instruction::SDiv:
3600 case Instruction::FDiv:
3601 case Instruction::URem:
3602 case Instruction::SRem:
3603 case Instruction::FRem:
3604 case Instruction::Shl:
3605 case Instruction::LShr:
3606 case Instruction::AShr:
3607 case Instruction::And:
3608 case Instruction::Or:
3609 case Instruction::Xor: {
3610 // Just widen binops.
3611 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3612 setDebugLocFromInst(Builder, BinOp);
3613 VectorParts &A = getVectorValue(it->getOperand(0));
3614 VectorParts &B = getVectorValue(it->getOperand(1));
3616 // Use this vector value for all users of the original instruction.
3617 for (unsigned Part = 0; Part < UF; ++Part) {
3618 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3620 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3621 VecOp->copyIRFlags(BinOp);
3626 propagateMetadata(Entry, it);
3629 case Instruction::Select: {
3631 // If the selector is loop invariant we can create a select
3632 // instruction with a scalar condition. Otherwise, use vector-select.
3633 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3635 setDebugLocFromInst(Builder, it);
3637 // The condition can be loop invariant but still defined inside the
3638 // loop. This means that we can't just use the original 'cond' value.
3639 // We have to take the 'vectorized' value and pick the first lane.
3640 // Instcombine will make this a no-op.
3641 VectorParts &Cond = getVectorValue(it->getOperand(0));
3642 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3643 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3645 Value *ScalarCond = (VF == 1) ? Cond[0] :
3646 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3648 for (unsigned Part = 0; Part < UF; ++Part) {
3649 Entry[Part] = Builder.CreateSelect(
3650 InvariantCond ? ScalarCond : Cond[Part],
3655 propagateMetadata(Entry, it);
3659 case Instruction::ICmp:
3660 case Instruction::FCmp: {
3661 // Widen compares. Generate vector compares.
3662 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3663 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3664 setDebugLocFromInst(Builder, it);
3665 VectorParts &A = getVectorValue(it->getOperand(0));
3666 VectorParts &B = getVectorValue(it->getOperand(1));
3667 for (unsigned Part = 0; Part < UF; ++Part) {
3670 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3672 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3676 propagateMetadata(Entry, it);
3680 case Instruction::Store:
3681 case Instruction::Load:
3682 vectorizeMemoryInstruction(it);
3684 case Instruction::ZExt:
3685 case Instruction::SExt:
3686 case Instruction::FPToUI:
3687 case Instruction::FPToSI:
3688 case Instruction::FPExt:
3689 case Instruction::PtrToInt:
3690 case Instruction::IntToPtr:
3691 case Instruction::SIToFP:
3692 case Instruction::UIToFP:
3693 case Instruction::Trunc:
3694 case Instruction::FPTrunc:
3695 case Instruction::BitCast: {
3696 CastInst *CI = dyn_cast<CastInst>(it);
3697 setDebugLocFromInst(Builder, it);
3698 /// Optimize the special case where the source is the induction
3699 /// variable. Notice that we can only optimize the 'trunc' case
3700 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3701 /// c. other casts depend on pointer size.
3702 if (CI->getOperand(0) == OldInduction &&
3703 it->getOpcode() == Instruction::Trunc) {
3704 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3706 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3707 InductionDescriptor II = Legal->getInductionVars()->lookup(OldInduction);
3709 ConstantInt::getSigned(CI->getType(), II.getStepValue()->getSExtValue());
3710 for (unsigned Part = 0; Part < UF; ++Part)
3711 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3712 propagateMetadata(Entry, it);
3715 /// Vectorize casts.
3716 Type *DestTy = (VF == 1) ? CI->getType() :
3717 VectorType::get(CI->getType(), VF);
3719 VectorParts &A = getVectorValue(it->getOperand(0));
3720 for (unsigned Part = 0; Part < UF; ++Part)
3721 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3722 propagateMetadata(Entry, it);
3726 case Instruction::Call: {
3727 // Ignore dbg intrinsics.
3728 if (isa<DbgInfoIntrinsic>(it))
3730 setDebugLocFromInst(Builder, it);
3732 Module *M = BB->getParent()->getParent();
3733 CallInst *CI = cast<CallInst>(it);
3735 StringRef FnName = CI->getCalledFunction()->getName();
3736 Function *F = CI->getCalledFunction();
3737 Type *RetTy = ToVectorTy(CI->getType(), VF);
3738 SmallVector<Type *, 4> Tys;
3739 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3740 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3742 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3744 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3745 ID == Intrinsic::lifetime_start)) {
3746 scalarizeInstruction(it);
3749 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3750 // version of the instruction.
3751 // Is it beneficial to perform intrinsic call compared to lib call?
3752 bool NeedToScalarize;
3753 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3754 bool UseVectorIntrinsic =
3755 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3756 if (!UseVectorIntrinsic && NeedToScalarize) {
3757 scalarizeInstruction(it);
3761 for (unsigned Part = 0; Part < UF; ++Part) {
3762 SmallVector<Value *, 4> Args;
3763 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3764 Value *Arg = CI->getArgOperand(i);
3765 // Some intrinsics have a scalar argument - don't replace it with a
3767 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3768 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3769 Arg = VectorArg[Part];
3771 Args.push_back(Arg);
3775 if (UseVectorIntrinsic) {
3776 // Use vector version of the intrinsic.
3777 Type *TysForDecl[] = {CI->getType()};
3779 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3780 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3782 // Use vector version of the library call.
3783 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3784 assert(!VFnName.empty() && "Vector function name is empty.");
3785 VectorF = M->getFunction(VFnName);
3787 // Generate a declaration
3788 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3790 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3791 VectorF->copyAttributesFrom(F);
3794 assert(VectorF && "Can't create vector function.");
3795 Entry[Part] = Builder.CreateCall(VectorF, Args);
3798 propagateMetadata(Entry, it);
3803 // All other instructions are unsupported. Scalarize them.
3804 scalarizeInstruction(it);
3807 }// end of for_each instr.
3810 void InnerLoopVectorizer::updateAnalysis() {
3811 // Forget the original basic block.
3812 SE->forgetLoop(OrigLoop);
3814 // Update the dominator tree information.
3815 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3816 "Entry does not dominate exit.");
3818 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3819 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3820 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3822 // Due to if predication of stores we might create a sequence of "if(pred)
3823 // a[i] = ...; " blocks.
3824 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3826 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3827 else if (isPredicatedBlock(i)) {
3828 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3830 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3834 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3835 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3836 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3837 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3839 DEBUG(DT->verifyDomTree());
3842 /// \brief Check whether it is safe to if-convert this phi node.
3844 /// Phi nodes with constant expressions that can trap are not safe to if
3846 static bool canIfConvertPHINodes(BasicBlock *BB) {
3847 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3848 PHINode *Phi = dyn_cast<PHINode>(I);
3851 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3852 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3859 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3860 if (!EnableIfConversion) {
3861 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3865 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3867 // A list of pointers that we can safely read and write to.
3868 SmallPtrSet<Value *, 8> SafePointes;
3870 // Collect safe addresses.
3871 for (Loop::block_iterator BI = TheLoop->block_begin(),
3872 BE = TheLoop->block_end(); BI != BE; ++BI) {
3873 BasicBlock *BB = *BI;
3875 if (blockNeedsPredication(BB))
3878 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3879 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3880 SafePointes.insert(LI->getPointerOperand());
3881 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3882 SafePointes.insert(SI->getPointerOperand());
3886 // Collect the blocks that need predication.
3887 BasicBlock *Header = TheLoop->getHeader();
3888 for (Loop::block_iterator BI = TheLoop->block_begin(),
3889 BE = TheLoop->block_end(); BI != BE; ++BI) {
3890 BasicBlock *BB = *BI;
3892 // We don't support switch statements inside loops.
3893 if (!isa<BranchInst>(BB->getTerminator())) {
3894 emitAnalysis(VectorizationReport(BB->getTerminator())
3895 << "loop contains a switch statement");
3899 // We must be able to predicate all blocks that need to be predicated.
3900 if (blockNeedsPredication(BB)) {
3901 if (!blockCanBePredicated(BB, SafePointes)) {
3902 emitAnalysis(VectorizationReport(BB->getTerminator())
3903 << "control flow cannot be substituted for a select");
3906 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3907 emitAnalysis(VectorizationReport(BB->getTerminator())
3908 << "control flow cannot be substituted for a select");
3913 // We can if-convert this loop.
3917 bool LoopVectorizationLegality::canVectorize() {
3918 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3919 // be canonicalized.
3920 if (!TheLoop->getLoopPreheader()) {
3922 VectorizationReport() <<
3923 "loop control flow is not understood by vectorizer");
3927 // We can only vectorize innermost loops.
3928 if (!TheLoop->empty()) {
3929 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3933 // We must have a single backedge.
3934 if (TheLoop->getNumBackEdges() != 1) {
3936 VectorizationReport() <<
3937 "loop control flow is not understood by vectorizer");
3941 // We must have a single exiting block.
3942 if (!TheLoop->getExitingBlock()) {
3944 VectorizationReport() <<
3945 "loop control flow is not understood by vectorizer");
3949 // We only handle bottom-tested loops, i.e. loop in which the condition is
3950 // checked at the end of each iteration. With that we can assume that all
3951 // instructions in the loop are executed the same number of times.
3952 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3954 VectorizationReport() <<
3955 "loop control flow is not understood by vectorizer");
3959 // We need to have a loop header.
3960 DEBUG(dbgs() << "LV: Found a loop: " <<
3961 TheLoop->getHeader()->getName() << '\n');
3963 // Check if we can if-convert non-single-bb loops.
3964 unsigned NumBlocks = TheLoop->getNumBlocks();
3965 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3966 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3970 // ScalarEvolution needs to be able to find the exit count.
3971 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3972 if (ExitCount == SE->getCouldNotCompute()) {
3973 emitAnalysis(VectorizationReport() <<
3974 "could not determine number of loop iterations");
3975 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3979 // Check if we can vectorize the instructions and CFG in this loop.
3980 if (!canVectorizeInstrs()) {
3981 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3985 // Go over each instruction and look at memory deps.
3986 if (!canVectorizeMemory()) {
3987 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3991 // Collect all of the variables that remain uniform after vectorization.
3992 collectLoopUniforms();
3994 DEBUG(dbgs() << "LV: We can vectorize this loop"
3995 << (LAI->getRuntimePointerChecking()->Need
3996 ? " (with a runtime bound check)"
4000 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4002 // If an override option has been passed in for interleaved accesses, use it.
4003 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4004 UseInterleaved = EnableInterleavedMemAccesses;
4006 // Analyze interleaved memory accesses.
4008 InterleaveInfo.analyzeInterleaving(Strides);
4010 // Okay! We can vectorize. At this point we don't have any other mem analysis
4011 // which may limit our maximum vectorization factor, so just return true with
4016 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4017 if (Ty->isPointerTy())
4018 return DL.getIntPtrType(Ty);
4020 // It is possible that char's or short's overflow when we ask for the loop's
4021 // trip count, work around this by changing the type size.
4022 if (Ty->getScalarSizeInBits() < 32)
4023 return Type::getInt32Ty(Ty->getContext());
4028 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4029 Ty0 = convertPointerToIntegerType(DL, Ty0);
4030 Ty1 = convertPointerToIntegerType(DL, Ty1);
4031 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4036 /// \brief Check that the instruction has outside loop users and is not an
4037 /// identified reduction variable.
4038 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4039 SmallPtrSetImpl<Value *> &Reductions) {
4040 // Reduction instructions are allowed to have exit users. All other
4041 // instructions must not have external users.
4042 if (!Reductions.count(Inst))
4043 //Check that all of the users of the loop are inside the BB.
4044 for (User *U : Inst->users()) {
4045 Instruction *UI = cast<Instruction>(U);
4046 // This user may be a reduction exit value.
4047 if (!TheLoop->contains(UI)) {
4048 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4055 bool LoopVectorizationLegality::canVectorizeInstrs() {
4056 BasicBlock *Header = TheLoop->getHeader();
4058 // Look for the attribute signaling the absence of NaNs.
4059 Function &F = *Header->getParent();
4060 const DataLayout &DL = F.getParent()->getDataLayout();
4061 if (F.hasFnAttribute("no-nans-fp-math"))
4063 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4065 // For each block in the loop.
4066 for (Loop::block_iterator bb = TheLoop->block_begin(),
4067 be = TheLoop->block_end(); bb != be; ++bb) {
4069 // Scan the instructions in the block and look for hazards.
4070 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4073 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4074 Type *PhiTy = Phi->getType();
4075 // Check that this PHI type is allowed.
4076 if (!PhiTy->isIntegerTy() &&
4077 !PhiTy->isFloatingPointTy() &&
4078 !PhiTy->isPointerTy()) {
4079 emitAnalysis(VectorizationReport(it)
4080 << "loop control flow is not understood by vectorizer");
4081 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4085 // If this PHINode is not in the header block, then we know that we
4086 // can convert it to select during if-conversion. No need to check if
4087 // the PHIs in this block are induction or reduction variables.
4088 if (*bb != Header) {
4089 // Check that this instruction has no outside users or is an
4090 // identified reduction value with an outside user.
4091 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4093 emitAnalysis(VectorizationReport(it) <<
4094 "value could not be identified as "
4095 "an induction or reduction variable");
4099 // We only allow if-converted PHIs with exactly two incoming values.
4100 if (Phi->getNumIncomingValues() != 2) {
4101 emitAnalysis(VectorizationReport(it)
4102 << "control flow not understood by vectorizer");
4103 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4107 InductionDescriptor ID;
4108 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4109 Inductions[Phi] = ID;
4110 // Get the widest type.
4112 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4114 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4116 // Int inductions are special because we only allow one IV.
4117 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4118 ID.getStepValue()->isOne() &&
4119 isa<Constant>(ID.getStartValue()) &&
4120 cast<Constant>(ID.getStartValue())->isNullValue()) {
4121 // Use the phi node with the widest type as induction. Use the last
4122 // one if there are multiple (no good reason for doing this other
4123 // than it is expedient). We've checked that it begins at zero and
4124 // steps by one, so this is a canonical induction variable.
4125 if (!Induction || PhiTy == WidestIndTy)
4129 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4131 // Until we explicitly handle the case of an induction variable with
4132 // an outside loop user we have to give up vectorizing this loop.
4133 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4134 emitAnalysis(VectorizationReport(it) <<
4135 "use of induction value outside of the "
4136 "loop is not handled by vectorizer");
4143 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4145 if (Reductions[Phi].hasUnsafeAlgebra())
4146 Requirements->addUnsafeAlgebraInst(
4147 Reductions[Phi].getUnsafeAlgebraInst());
4148 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4152 emitAnalysis(VectorizationReport(it) <<
4153 "value that could not be identified as "
4154 "reduction is used outside the loop");
4155 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4157 }// end of PHI handling
4159 // We handle calls that:
4160 // * Are debug info intrinsics.
4161 // * Have a mapping to an IR intrinsic.
4162 // * Have a vector version available.
4163 CallInst *CI = dyn_cast<CallInst>(it);
4164 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4165 !(CI->getCalledFunction() && TLI &&
4166 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4167 emitAnalysis(VectorizationReport(it) <<
4168 "call instruction cannot be vectorized");
4169 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4173 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4174 // second argument is the same (i.e. loop invariant)
4176 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4177 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4178 emitAnalysis(VectorizationReport(it)
4179 << "intrinsic instruction cannot be vectorized");
4180 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4185 // Check that the instruction return type is vectorizable.
4186 // Also, we can't vectorize extractelement instructions.
4187 if ((!VectorType::isValidElementType(it->getType()) &&
4188 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4189 emitAnalysis(VectorizationReport(it)
4190 << "instruction return type cannot be vectorized");
4191 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4195 // Check that the stored type is vectorizable.
4196 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4197 Type *T = ST->getValueOperand()->getType();
4198 if (!VectorType::isValidElementType(T)) {
4199 emitAnalysis(VectorizationReport(ST) <<
4200 "store instruction cannot be vectorized");
4203 if (EnableMemAccessVersioning)
4204 collectStridedAccess(ST);
4207 if (EnableMemAccessVersioning)
4208 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4209 collectStridedAccess(LI);
4211 // Reduction instructions are allowed to have exit users.
4212 // All other instructions must not have external users.
4213 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4214 emitAnalysis(VectorizationReport(it) <<
4215 "value cannot be used outside the loop");
4224 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4225 if (Inductions.empty()) {
4226 emitAnalysis(VectorizationReport()
4227 << "loop induction variable could not be identified");
4232 // Now we know the widest induction type, check if our found induction
4233 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4234 // will create another.
4235 if (Induction && WidestIndTy != Induction->getType())
4236 Induction = nullptr;
4241 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4242 Value *Ptr = nullptr;
4243 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4244 Ptr = LI->getPointerOperand();
4245 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4246 Ptr = SI->getPointerOperand();
4250 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4254 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4255 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4256 Strides[Ptr] = Stride;
4257 StrideSet.insert(Stride);
4260 void LoopVectorizationLegality::collectLoopUniforms() {
4261 // We now know that the loop is vectorizable!
4262 // Collect variables that will remain uniform after vectorization.
4263 std::vector<Value*> Worklist;
4264 BasicBlock *Latch = TheLoop->getLoopLatch();
4266 // Start with the conditional branch and walk up the block.
4267 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4269 // Also add all consecutive pointer values; these values will be uniform
4270 // after vectorization (and subsequent cleanup) and, until revectorization is
4271 // supported, all dependencies must also be uniform.
4272 for (Loop::block_iterator B = TheLoop->block_begin(),
4273 BE = TheLoop->block_end(); B != BE; ++B)
4274 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4276 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4277 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4279 while (!Worklist.empty()) {
4280 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4281 Worklist.pop_back();
4283 // Look at instructions inside this loop.
4284 // Stop when reaching PHI nodes.
4285 // TODO: we need to follow values all over the loop, not only in this block.
4286 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4289 // This is a known uniform.
4292 // Insert all operands.
4293 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4297 bool LoopVectorizationLegality::canVectorizeMemory() {
4298 LAI = &LAA->getInfo(TheLoop, Strides);
4299 auto &OptionalReport = LAI->getReport();
4301 emitAnalysis(VectorizationReport(*OptionalReport));
4302 if (!LAI->canVectorizeMemory())
4305 if (LAI->hasStoreToLoopInvariantAddress()) {
4307 VectorizationReport()
4308 << "write to a loop invariant address could not be vectorized");
4309 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4313 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4318 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4319 Value *In0 = const_cast<Value*>(V);
4320 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4324 return Inductions.count(PN);
4327 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4328 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4331 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4332 SmallPtrSetImpl<Value *> &SafePtrs) {
4334 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4335 // Check that we don't have a constant expression that can trap as operand.
4336 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4338 if (Constant *C = dyn_cast<Constant>(*OI))
4342 // We might be able to hoist the load.
4343 if (it->mayReadFromMemory()) {
4344 LoadInst *LI = dyn_cast<LoadInst>(it);
4347 if (!SafePtrs.count(LI->getPointerOperand())) {
4348 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4349 MaskedOp.insert(LI);
4356 // We don't predicate stores at the moment.
4357 if (it->mayWriteToMemory()) {
4358 StoreInst *SI = dyn_cast<StoreInst>(it);
4359 // We only support predication of stores in basic blocks with one
4364 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4365 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4367 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4368 !isSinglePredecessor) {
4369 // Build a masked store if it is legal for the target, otherwise scalarize
4371 bool isLegalMaskedOp =
4372 isLegalMaskedStore(SI->getValueOperand()->getType(),
4373 SI->getPointerOperand());
4374 if (isLegalMaskedOp) {
4376 MaskedOp.insert(SI);
4385 // The instructions below can trap.
4386 switch (it->getOpcode()) {
4388 case Instruction::UDiv:
4389 case Instruction::SDiv:
4390 case Instruction::URem:
4391 case Instruction::SRem:
4399 void InterleavedAccessInfo::collectConstStridedAccesses(
4400 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4401 const ValueToValueMap &Strides) {
4402 // Holds load/store instructions in program order.
4403 SmallVector<Instruction *, 16> AccessList;
4405 for (auto *BB : TheLoop->getBlocks()) {
4406 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4408 for (auto &I : *BB) {
4409 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4411 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4415 AccessList.push_back(&I);
4419 if (AccessList.empty())
4422 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4423 for (auto I : AccessList) {
4424 LoadInst *LI = dyn_cast<LoadInst>(I);
4425 StoreInst *SI = dyn_cast<StoreInst>(I);
4427 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4428 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4430 // The factor of the corresponding interleave group.
4431 unsigned Factor = std::abs(Stride);
4433 // Ignore the access if the factor is too small or too large.
4434 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4437 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4438 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4439 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4441 // An alignment of 0 means target ABI alignment.
4442 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4444 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4446 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4450 // Analyze interleaved accesses and collect them into interleave groups.
4452 // Notice that the vectorization on interleaved groups will change instruction
4453 // orders and may break dependences. But the memory dependence check guarantees
4454 // that there is no overlap between two pointers of different strides, element
4455 // sizes or underlying bases.
4457 // For pointers sharing the same stride, element size and underlying base, no
4458 // need to worry about Read-After-Write dependences and Write-After-Read
4461 // E.g. The RAW dependence: A[i] = a;
4463 // This won't exist as it is a store-load forwarding conflict, which has
4464 // already been checked and forbidden in the dependence check.
4466 // E.g. The WAR dependence: a = A[i]; // (1)
4468 // The store group of (2) is always inserted at or below (2), and the load group
4469 // of (1) is always inserted at or above (1). The dependence is safe.
4470 void InterleavedAccessInfo::analyzeInterleaving(
4471 const ValueToValueMap &Strides) {
4472 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4474 // Holds all the stride accesses.
4475 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4476 collectConstStridedAccesses(StrideAccesses, Strides);
4478 if (StrideAccesses.empty())
4481 // Holds all interleaved store groups temporarily.
4482 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4484 // Search the load-load/write-write pair B-A in bottom-up order and try to
4485 // insert B into the interleave group of A according to 3 rules:
4486 // 1. A and B have the same stride.
4487 // 2. A and B have the same memory object size.
4488 // 3. B belongs to the group according to the distance.
4490 // The bottom-up order can avoid breaking the Write-After-Write dependences
4491 // between two pointers of the same base.
4492 // E.g. A[i] = a; (1)
4495 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4496 // above (1), which guarantees that (1) is always above (2).
4497 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4499 Instruction *A = I->first;
4500 StrideDescriptor DesA = I->second;
4502 InterleaveGroup *Group = getInterleaveGroup(A);
4504 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4505 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4508 if (A->mayWriteToMemory())
4509 StoreGroups.insert(Group);
4511 for (auto II = std::next(I); II != E; ++II) {
4512 Instruction *B = II->first;
4513 StrideDescriptor DesB = II->second;
4515 // Ignore if B is already in a group or B is a different memory operation.
4516 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4519 // Check the rule 1 and 2.
4520 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4523 // Calculate the distance and prepare for the rule 3.
4524 const SCEVConstant *DistToA =
4525 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4529 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4531 // Skip if the distance is not multiple of size as they are not in the
4533 if (DistanceToA % static_cast<int>(DesA.Size))
4536 // The index of B is the index of A plus the related index to A.
4538 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4540 // Try to insert B into the group.
4541 if (Group->insertMember(B, IndexB, DesB.Align)) {
4542 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4543 << " into the interleave group with" << *A << '\n');
4544 InterleaveGroupMap[B] = Group;
4546 // Set the first load in program order as the insert position.
4547 if (B->mayReadFromMemory())
4548 Group->setInsertPos(B);
4550 } // Iteration on instruction B
4551 } // Iteration on instruction A
4553 // Remove interleaved store groups with gaps.
4554 for (InterleaveGroup *Group : StoreGroups)
4555 if (Group->getNumMembers() != Group->getFactor())
4556 releaseGroup(Group);
4559 LoopVectorizationCostModel::VectorizationFactor
4560 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4561 // Width 1 means no vectorize
4562 VectorizationFactor Factor = { 1U, 0U };
4563 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4564 emitAnalysis(VectorizationReport() <<
4565 "runtime pointer checks needed. Enable vectorization of this "
4566 "loop with '#pragma clang loop vectorize(enable)' when "
4567 "compiling with -Os/-Oz");
4569 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4573 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4574 emitAnalysis(VectorizationReport() <<
4575 "store that is conditionally executed prevents vectorization");
4576 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4580 // Find the trip count.
4581 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4582 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4584 unsigned WidestType = getWidestType();
4585 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4586 unsigned MaxSafeDepDist = -1U;
4587 if (Legal->getMaxSafeDepDistBytes() != -1U)
4588 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4589 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4590 WidestRegister : MaxSafeDepDist);
4591 unsigned MaxVectorSize = WidestRegister / WidestType;
4592 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4593 DEBUG(dbgs() << "LV: The Widest register is: "
4594 << WidestRegister << " bits.\n");
4596 if (MaxVectorSize == 0) {
4597 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4601 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4602 " into one vector!");
4604 unsigned VF = MaxVectorSize;
4606 // If we optimize the program for size, avoid creating the tail loop.
4608 // If we are unable to calculate the trip count then don't try to vectorize.
4611 (VectorizationReport() <<
4612 "unable to calculate the loop count due to complex control flow");
4613 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4617 // Find the maximum SIMD width that can fit within the trip count.
4618 VF = TC % MaxVectorSize;
4623 // If the trip count that we found modulo the vectorization factor is not
4624 // zero then we require a tail.
4625 emitAnalysis(VectorizationReport() <<
4626 "cannot optimize for size and vectorize at the "
4627 "same time. Enable vectorization of this loop "
4628 "with '#pragma clang loop vectorize(enable)' "
4629 "when compiling with -Os/-Oz");
4630 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4635 int UserVF = Hints->getWidth();
4637 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4638 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4640 Factor.Width = UserVF;
4644 float Cost = expectedCost(1);
4646 const float ScalarCost = Cost;
4649 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4651 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4652 // Ignore scalar width, because the user explicitly wants vectorization.
4653 if (ForceVectorization && VF > 1) {
4655 Cost = expectedCost(Width) / (float)Width;
4658 for (unsigned i=2; i <= VF; i*=2) {
4659 // Notice that the vector loop needs to be executed less times, so
4660 // we need to divide the cost of the vector loops by the width of
4661 // the vector elements.
4662 float VectorCost = expectedCost(i) / (float)i;
4663 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4664 (int)VectorCost << ".\n");
4665 if (VectorCost < Cost) {
4671 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4672 << "LV: Vectorization seems to be not beneficial, "
4673 << "but was forced by a user.\n");
4674 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4675 Factor.Width = Width;
4676 Factor.Cost = Width * Cost;
4680 unsigned LoopVectorizationCostModel::getWidestType() {
4681 unsigned MaxWidth = 8;
4682 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4685 for (Loop::block_iterator bb = TheLoop->block_begin(),
4686 be = TheLoop->block_end(); bb != be; ++bb) {
4687 BasicBlock *BB = *bb;
4689 // For each instruction in the loop.
4690 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4691 Type *T = it->getType();
4693 // Skip ignored values.
4694 if (ValuesToIgnore.count(it))
4697 // Only examine Loads, Stores and PHINodes.
4698 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4701 // Examine PHI nodes that are reduction variables. Update the type to
4702 // account for the recurrence type.
4703 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4704 if (!Legal->getReductionVars()->count(PN))
4706 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4707 T = RdxDesc.getRecurrenceType();
4710 // Examine the stored values.
4711 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4712 T = ST->getValueOperand()->getType();
4714 // Ignore loaded pointer types and stored pointer types that are not
4715 // consecutive. However, we do want to take consecutive stores/loads of
4716 // pointer vectors into account.
4717 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4720 MaxWidth = std::max(MaxWidth,
4721 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4728 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4730 unsigned LoopCost) {
4732 // -- The interleave heuristics --
4733 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4734 // There are many micro-architectural considerations that we can't predict
4735 // at this level. For example, frontend pressure (on decode or fetch) due to
4736 // code size, or the number and capabilities of the execution ports.
4738 // We use the following heuristics to select the interleave count:
4739 // 1. If the code has reductions, then we interleave to break the cross
4740 // iteration dependency.
4741 // 2. If the loop is really small, then we interleave to reduce the loop
4743 // 3. We don't interleave if we think that we will spill registers to memory
4744 // due to the increased register pressure.
4746 // When we optimize for size, we don't interleave.
4750 // We used the distance for the interleave count.
4751 if (Legal->getMaxSafeDepDistBytes() != -1U)
4754 // Do not interleave loops with a relatively small trip count.
4755 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4756 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4759 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4760 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4764 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4765 TargetNumRegisters = ForceTargetNumScalarRegs;
4767 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4768 TargetNumRegisters = ForceTargetNumVectorRegs;
4771 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4772 // We divide by these constants so assume that we have at least one
4773 // instruction that uses at least one register.
4774 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4775 R.NumInstructions = std::max(R.NumInstructions, 1U);
4777 // We calculate the interleave count using the following formula.
4778 // Subtract the number of loop invariants from the number of available
4779 // registers. These registers are used by all of the interleaved instances.
4780 // Next, divide the remaining registers by the number of registers that is
4781 // required by the loop, in order to estimate how many parallel instances
4782 // fit without causing spills. All of this is rounded down if necessary to be
4783 // a power of two. We want power of two interleave count to simplify any
4784 // addressing operations or alignment considerations.
4785 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4788 // Don't count the induction variable as interleaved.
4789 if (EnableIndVarRegisterHeur)
4790 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4791 std::max(1U, (R.MaxLocalUsers - 1)));
4793 // Clamp the interleave ranges to reasonable counts.
4794 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4796 // Check if the user has overridden the max.
4798 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4799 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4801 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4802 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4805 // If we did not calculate the cost for VF (because the user selected the VF)
4806 // then we calculate the cost of VF here.
4808 LoopCost = expectedCost(VF);
4810 // Clamp the calculated IC to be between the 1 and the max interleave count
4811 // that the target allows.
4812 if (IC > MaxInterleaveCount)
4813 IC = MaxInterleaveCount;
4817 // Interleave if we vectorized this loop and there is a reduction that could
4818 // benefit from interleaving.
4819 if (VF > 1 && Legal->getReductionVars()->size()) {
4820 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4824 // Note that if we've already vectorized the loop we will have done the
4825 // runtime check and so interleaving won't require further checks.
4826 bool InterleavingRequiresRuntimePointerCheck =
4827 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4829 // We want to interleave small loops in order to reduce the loop overhead and
4830 // potentially expose ILP opportunities.
4831 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4832 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4833 // We assume that the cost overhead is 1 and we use the cost model
4834 // to estimate the cost of the loop and interleave until the cost of the
4835 // loop overhead is about 5% of the cost of the loop.
4837 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4839 // Interleave until store/load ports (estimated by max interleave count) are
4841 unsigned NumStores = Legal->getNumStores();
4842 unsigned NumLoads = Legal->getNumLoads();
4843 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4844 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4846 // If we have a scalar reduction (vector reductions are already dealt with
4847 // by this point), we can increase the critical path length if the loop
4848 // we're interleaving is inside another loop. Limit, by default to 2, so the
4849 // critical path only gets increased by one reduction operation.
4850 if (Legal->getReductionVars()->size() &&
4851 TheLoop->getLoopDepth() > 1) {
4852 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4853 SmallIC = std::min(SmallIC, F);
4854 StoresIC = std::min(StoresIC, F);
4855 LoadsIC = std::min(LoadsIC, F);
4858 if (EnableLoadStoreRuntimeInterleave &&
4859 std::max(StoresIC, LoadsIC) > SmallIC) {
4860 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4861 return std::max(StoresIC, LoadsIC);
4864 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4868 // Interleave if this is a large loop (small loops are already dealt with by
4870 // point) that could benefit from interleaving.
4871 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4872 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4873 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4877 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4881 LoopVectorizationCostModel::RegisterUsage
4882 LoopVectorizationCostModel::calculateRegisterUsage() {
4883 // This function calculates the register usage by measuring the highest number
4884 // of values that are alive at a single location. Obviously, this is a very
4885 // rough estimation. We scan the loop in a topological order in order and
4886 // assign a number to each instruction. We use RPO to ensure that defs are
4887 // met before their users. We assume that each instruction that has in-loop
4888 // users starts an interval. We record every time that an in-loop value is
4889 // used, so we have a list of the first and last occurrences of each
4890 // instruction. Next, we transpose this data structure into a multi map that
4891 // holds the list of intervals that *end* at a specific location. This multi
4892 // map allows us to perform a linear search. We scan the instructions linearly
4893 // and record each time that a new interval starts, by placing it in a set.
4894 // If we find this value in the multi-map then we remove it from the set.
4895 // The max register usage is the maximum size of the set.
4896 // We also search for instructions that are defined outside the loop, but are
4897 // used inside the loop. We need this number separately from the max-interval
4898 // usage number because when we unroll, loop-invariant values do not take
4900 LoopBlocksDFS DFS(TheLoop);
4904 R.NumInstructions = 0;
4906 // Each 'key' in the map opens a new interval. The values
4907 // of the map are the index of the 'last seen' usage of the
4908 // instruction that is the key.
4909 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4910 // Maps instruction to its index.
4911 DenseMap<unsigned, Instruction*> IdxToInstr;
4912 // Marks the end of each interval.
4913 IntervalMap EndPoint;
4914 // Saves the list of instruction indices that are used in the loop.
4915 SmallSet<Instruction*, 8> Ends;
4916 // Saves the list of values that are used in the loop but are
4917 // defined outside the loop, such as arguments and constants.
4918 SmallPtrSet<Value*, 8> LoopInvariants;
4921 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4922 be = DFS.endRPO(); bb != be; ++bb) {
4923 R.NumInstructions += (*bb)->size();
4924 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4926 Instruction *I = it;
4927 IdxToInstr[Index++] = I;
4929 // Save the end location of each USE.
4930 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4931 Value *U = I->getOperand(i);
4932 Instruction *Instr = dyn_cast<Instruction>(U);
4934 // Ignore non-instruction values such as arguments, constants, etc.
4935 if (!Instr) continue;
4937 // If this instruction is outside the loop then record it and continue.
4938 if (!TheLoop->contains(Instr)) {
4939 LoopInvariants.insert(Instr);
4943 // Overwrite previous end points.
4944 EndPoint[Instr] = Index;
4950 // Saves the list of intervals that end with the index in 'key'.
4951 typedef SmallVector<Instruction*, 2> InstrList;
4952 DenseMap<unsigned, InstrList> TransposeEnds;
4954 // Transpose the EndPoints to a list of values that end at each index.
4955 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4957 TransposeEnds[it->second].push_back(it->first);
4959 SmallSet<Instruction*, 8> OpenIntervals;
4960 unsigned MaxUsage = 0;
4963 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4964 for (unsigned int i = 0; i < Index; ++i) {
4965 Instruction *I = IdxToInstr[i];
4966 // Ignore instructions that are never used within the loop.
4967 if (!Ends.count(I)) continue;
4969 // Skip ignored values.
4970 if (ValuesToIgnore.count(I))
4973 // Remove all of the instructions that end at this location.
4974 InstrList &List = TransposeEnds[i];
4975 for (unsigned int j=0, e = List.size(); j < e; ++j)
4976 OpenIntervals.erase(List[j]);
4978 // Count the number of live interals.
4979 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4981 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4982 OpenIntervals.size() << '\n');
4984 // Add the current instruction to the list of open intervals.
4985 OpenIntervals.insert(I);
4988 unsigned Invariant = LoopInvariants.size();
4989 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4990 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4991 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4993 R.LoopInvariantRegs = Invariant;
4994 R.MaxLocalUsers = MaxUsage;
4998 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5002 for (Loop::block_iterator bb = TheLoop->block_begin(),
5003 be = TheLoop->block_end(); bb != be; ++bb) {
5004 unsigned BlockCost = 0;
5005 BasicBlock *BB = *bb;
5007 // For each instruction in the old loop.
5008 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5009 // Skip dbg intrinsics.
5010 if (isa<DbgInfoIntrinsic>(it))
5013 // Skip ignored values.
5014 if (ValuesToIgnore.count(it))
5017 unsigned C = getInstructionCost(it, VF);
5019 // Check if we should override the cost.
5020 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5021 C = ForceTargetInstructionCost;
5024 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5025 VF << " For instruction: " << *it << '\n');
5028 // We assume that if-converted blocks have a 50% chance of being executed.
5029 // When the code is scalar then some of the blocks are avoided due to CF.
5030 // When the code is vectorized we execute all code paths.
5031 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5040 /// \brief Check whether the address computation for a non-consecutive memory
5041 /// access looks like an unlikely candidate for being merged into the indexing
5044 /// We look for a GEP which has one index that is an induction variable and all
5045 /// other indices are loop invariant. If the stride of this access is also
5046 /// within a small bound we decide that this address computation can likely be
5047 /// merged into the addressing mode.
5048 /// In all other cases, we identify the address computation as complex.
5049 static bool isLikelyComplexAddressComputation(Value *Ptr,
5050 LoopVectorizationLegality *Legal,
5051 ScalarEvolution *SE,
5052 const Loop *TheLoop) {
5053 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5057 // We are looking for a gep with all loop invariant indices except for one
5058 // which should be an induction variable.
5059 unsigned NumOperands = Gep->getNumOperands();
5060 for (unsigned i = 1; i < NumOperands; ++i) {
5061 Value *Opd = Gep->getOperand(i);
5062 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5063 !Legal->isInductionVariable(Opd))
5067 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5068 // can likely be merged into the address computation.
5069 unsigned MaxMergeDistance = 64;
5071 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5075 // Check the step is constant.
5076 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5077 // Calculate the pointer stride and check if it is consecutive.
5078 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5082 const APInt &APStepVal = C->getValue()->getValue();
5084 // Huge step value - give up.
5085 if (APStepVal.getBitWidth() > 64)
5088 int64_t StepVal = APStepVal.getSExtValue();
5090 return StepVal > MaxMergeDistance;
5093 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5094 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5100 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5101 // If we know that this instruction will remain uniform, check the cost of
5102 // the scalar version.
5103 if (Legal->isUniformAfterVectorization(I))
5106 Type *RetTy = I->getType();
5107 Type *VectorTy = ToVectorTy(RetTy, VF);
5109 // TODO: We need to estimate the cost of intrinsic calls.
5110 switch (I->getOpcode()) {
5111 case Instruction::GetElementPtr:
5112 // We mark this instruction as zero-cost because the cost of GEPs in
5113 // vectorized code depends on whether the corresponding memory instruction
5114 // is scalarized or not. Therefore, we handle GEPs with the memory
5115 // instruction cost.
5117 case Instruction::Br: {
5118 return TTI.getCFInstrCost(I->getOpcode());
5120 case Instruction::PHI:
5121 //TODO: IF-converted IFs become selects.
5123 case Instruction::Add:
5124 case Instruction::FAdd:
5125 case Instruction::Sub:
5126 case Instruction::FSub:
5127 case Instruction::Mul:
5128 case Instruction::FMul:
5129 case Instruction::UDiv:
5130 case Instruction::SDiv:
5131 case Instruction::FDiv:
5132 case Instruction::URem:
5133 case Instruction::SRem:
5134 case Instruction::FRem:
5135 case Instruction::Shl:
5136 case Instruction::LShr:
5137 case Instruction::AShr:
5138 case Instruction::And:
5139 case Instruction::Or:
5140 case Instruction::Xor: {
5141 // Since we will replace the stride by 1 the multiplication should go away.
5142 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5144 // Certain instructions can be cheaper to vectorize if they have a constant
5145 // second vector operand. One example of this are shifts on x86.
5146 TargetTransformInfo::OperandValueKind Op1VK =
5147 TargetTransformInfo::OK_AnyValue;
5148 TargetTransformInfo::OperandValueKind Op2VK =
5149 TargetTransformInfo::OK_AnyValue;
5150 TargetTransformInfo::OperandValueProperties Op1VP =
5151 TargetTransformInfo::OP_None;
5152 TargetTransformInfo::OperandValueProperties Op2VP =
5153 TargetTransformInfo::OP_None;
5154 Value *Op2 = I->getOperand(1);
5156 // Check for a splat of a constant or for a non uniform vector of constants.
5157 if (isa<ConstantInt>(Op2)) {
5158 ConstantInt *CInt = cast<ConstantInt>(Op2);
5159 if (CInt && CInt->getValue().isPowerOf2())
5160 Op2VP = TargetTransformInfo::OP_PowerOf2;
5161 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5162 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5163 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5164 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5166 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5167 if (CInt && CInt->getValue().isPowerOf2())
5168 Op2VP = TargetTransformInfo::OP_PowerOf2;
5169 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5173 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5176 case Instruction::Select: {
5177 SelectInst *SI = cast<SelectInst>(I);
5178 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5179 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5180 Type *CondTy = SI->getCondition()->getType();
5182 CondTy = VectorType::get(CondTy, VF);
5184 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5186 case Instruction::ICmp:
5187 case Instruction::FCmp: {
5188 Type *ValTy = I->getOperand(0)->getType();
5189 VectorTy = ToVectorTy(ValTy, VF);
5190 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5192 case Instruction::Store:
5193 case Instruction::Load: {
5194 StoreInst *SI = dyn_cast<StoreInst>(I);
5195 LoadInst *LI = dyn_cast<LoadInst>(I);
5196 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5198 VectorTy = ToVectorTy(ValTy, VF);
5200 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5201 unsigned AS = SI ? SI->getPointerAddressSpace() :
5202 LI->getPointerAddressSpace();
5203 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5204 // We add the cost of address computation here instead of with the gep
5205 // instruction because only here we know whether the operation is
5208 return TTI.getAddressComputationCost(VectorTy) +
5209 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5211 // For an interleaved access, calculate the total cost of the whole
5212 // interleave group.
5213 if (Legal->isAccessInterleaved(I)) {
5214 auto Group = Legal->getInterleavedAccessGroup(I);
5215 assert(Group && "Fail to get an interleaved access group.");
5217 // Only calculate the cost once at the insert position.
5218 if (Group->getInsertPos() != I)
5221 unsigned InterleaveFactor = Group->getFactor();
5223 VectorType::get(VectorTy->getVectorElementType(),
5224 VectorTy->getVectorNumElements() * InterleaveFactor);
5226 // Holds the indices of existing members in an interleaved load group.
5227 // An interleaved store group doesn't need this as it dones't allow gaps.
5228 SmallVector<unsigned, 4> Indices;
5230 for (unsigned i = 0; i < InterleaveFactor; i++)
5231 if (Group->getMember(i))
5232 Indices.push_back(i);
5235 // Calculate the cost of the whole interleaved group.
5236 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5237 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5238 Group->getAlignment(), AS);
5240 if (Group->isReverse())
5242 Group->getNumMembers() *
5243 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5245 // FIXME: The interleaved load group with a huge gap could be even more
5246 // expensive than scalar operations. Then we could ignore such group and
5247 // use scalar operations instead.
5251 // Scalarized loads/stores.
5252 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5253 bool Reverse = ConsecutiveStride < 0;
5254 const DataLayout &DL = I->getModule()->getDataLayout();
5255 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5256 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5257 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5258 bool IsComplexComputation =
5259 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5261 // The cost of extracting from the value vector and pointer vector.
5262 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5263 for (unsigned i = 0; i < VF; ++i) {
5264 // The cost of extracting the pointer operand.
5265 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5266 // In case of STORE, the cost of ExtractElement from the vector.
5267 // In case of LOAD, the cost of InsertElement into the returned
5269 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5270 Instruction::InsertElement,
5274 // The cost of the scalar loads/stores.
5275 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5276 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5281 // Wide load/stores.
5282 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5283 if (Legal->isMaskRequired(I))
5284 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5287 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5290 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5294 case Instruction::ZExt:
5295 case Instruction::SExt:
5296 case Instruction::FPToUI:
5297 case Instruction::FPToSI:
5298 case Instruction::FPExt:
5299 case Instruction::PtrToInt:
5300 case Instruction::IntToPtr:
5301 case Instruction::SIToFP:
5302 case Instruction::UIToFP:
5303 case Instruction::Trunc:
5304 case Instruction::FPTrunc:
5305 case Instruction::BitCast: {
5306 // We optimize the truncation of induction variable.
5307 // The cost of these is the same as the scalar operation.
5308 if (I->getOpcode() == Instruction::Trunc &&
5309 Legal->isInductionVariable(I->getOperand(0)))
5310 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5311 I->getOperand(0)->getType());
5313 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5314 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5316 case Instruction::Call: {
5317 bool NeedToScalarize;
5318 CallInst *CI = cast<CallInst>(I);
5319 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5320 if (getIntrinsicIDForCall(CI, TLI))
5321 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5325 // We are scalarizing the instruction. Return the cost of the scalar
5326 // instruction, plus the cost of insert and extract into vector
5327 // elements, times the vector width.
5330 if (!RetTy->isVoidTy() && VF != 1) {
5331 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5333 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5336 // The cost of inserting the results plus extracting each one of the
5338 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5341 // The cost of executing VF copies of the scalar instruction. This opcode
5342 // is unknown. Assume that it is the same as 'mul'.
5343 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5349 char LoopVectorize::ID = 0;
5350 static const char lv_name[] = "Loop Vectorization";
5351 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5352 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5353 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5354 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5355 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5356 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5357 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5358 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5359 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5360 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5361 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5362 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5365 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5366 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5370 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5371 // Check for a store.
5372 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5373 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5375 // Check for a load.
5376 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5377 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5383 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5384 bool IfPredicateStore) {
5385 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5386 // Holds vector parameters or scalars, in case of uniform vals.
5387 SmallVector<VectorParts, 4> Params;
5389 setDebugLocFromInst(Builder, Instr);
5391 // Find all of the vectorized parameters.
5392 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5393 Value *SrcOp = Instr->getOperand(op);
5395 // If we are accessing the old induction variable, use the new one.
5396 if (SrcOp == OldInduction) {
5397 Params.push_back(getVectorValue(SrcOp));
5401 // Try using previously calculated values.
5402 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5404 // If the src is an instruction that appeared earlier in the basic block
5405 // then it should already be vectorized.
5406 if (SrcInst && OrigLoop->contains(SrcInst)) {
5407 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5408 // The parameter is a vector value from earlier.
5409 Params.push_back(WidenMap.get(SrcInst));
5411 // The parameter is a scalar from outside the loop. Maybe even a constant.
5412 VectorParts Scalars;
5413 Scalars.append(UF, SrcOp);
5414 Params.push_back(Scalars);
5418 assert(Params.size() == Instr->getNumOperands() &&
5419 "Invalid number of operands");
5421 // Does this instruction return a value ?
5422 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5424 Value *UndefVec = IsVoidRetTy ? nullptr :
5425 UndefValue::get(Instr->getType());
5426 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5427 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5429 Instruction *InsertPt = Builder.GetInsertPoint();
5430 BasicBlock *IfBlock = Builder.GetInsertBlock();
5431 BasicBlock *CondBlock = nullptr;
5434 Loop *VectorLp = nullptr;
5435 if (IfPredicateStore) {
5436 assert(Instr->getParent()->getSinglePredecessor() &&
5437 "Only support single predecessor blocks");
5438 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5439 Instr->getParent());
5440 VectorLp = LI->getLoopFor(IfBlock);
5441 assert(VectorLp && "Must have a loop for this block");
5444 // For each vector unroll 'part':
5445 for (unsigned Part = 0; Part < UF; ++Part) {
5446 // For each scalar that we create:
5448 // Start an "if (pred) a[i] = ..." block.
5449 Value *Cmp = nullptr;
5450 if (IfPredicateStore) {
5451 if (Cond[Part]->getType()->isVectorTy())
5453 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5454 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5455 ConstantInt::get(Cond[Part]->getType(), 1));
5456 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5457 LoopVectorBody.push_back(CondBlock);
5458 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5459 // Update Builder with newly created basic block.
5460 Builder.SetInsertPoint(InsertPt);
5463 Instruction *Cloned = Instr->clone();
5465 Cloned->setName(Instr->getName() + ".cloned");
5466 // Replace the operands of the cloned instructions with extracted scalars.
5467 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5468 Value *Op = Params[op][Part];
5469 Cloned->setOperand(op, Op);
5472 // Place the cloned scalar in the new loop.
5473 Builder.Insert(Cloned);
5475 // If the original scalar returns a value we need to place it in a vector
5476 // so that future users will be able to use it.
5478 VecResults[Part] = Cloned;
5481 if (IfPredicateStore) {
5482 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5483 LoopVectorBody.push_back(NewIfBlock);
5484 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5485 Builder.SetInsertPoint(InsertPt);
5486 ReplaceInstWithInst(IfBlock->getTerminator(),
5487 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5488 IfBlock = NewIfBlock;
5493 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5494 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5495 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5497 return scalarizeInstruction(Instr, IfPredicateStore);
5500 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5504 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5508 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5509 // When unrolling and the VF is 1, we only need to add a simple scalar.
5510 Type *ITy = Val->getType();
5511 assert(!ITy->isVectorTy() && "Val must be a scalar");
5512 Constant *C = ConstantInt::get(ITy, StartIdx);
5513 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");