1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/EquivalenceClasses.h"
52 #include "llvm/ADT/Hashing.h"
53 #include "llvm/ADT/MapVector.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/AliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/LoopAccessAnalysis.h"
66 #include "llvm/Analysis/LoopInfo.h"
67 #include "llvm/Analysis/LoopIterator.h"
68 #include "llvm/Analysis/LoopPass.h"
69 #include "llvm/Analysis/ScalarEvolution.h"
70 #include "llvm/Analysis/ScalarEvolutionExpander.h"
71 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
72 #include "llvm/Analysis/TargetTransformInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/IR/Constants.h"
75 #include "llvm/IR/DataLayout.h"
76 #include "llvm/IR/DebugInfo.h"
77 #include "llvm/IR/DerivedTypes.h"
78 #include "llvm/IR/DiagnosticInfo.h"
79 #include "llvm/IR/Dominators.h"
80 #include "llvm/IR/Function.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/Instructions.h"
83 #include "llvm/IR/IntrinsicInst.h"
84 #include "llvm/IR/LLVMContext.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/IR/Type.h"
88 #include "llvm/IR/Value.h"
89 #include "llvm/IR/ValueHandle.h"
90 #include "llvm/IR/Verifier.h"
91 #include "llvm/Pass.h"
92 #include "llvm/Support/BranchProbability.h"
93 #include "llvm/Support/CommandLine.h"
94 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/raw_ostream.h"
96 #include "llvm/Transforms/Scalar.h"
97 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
98 #include "llvm/Transforms/Utils/Local.h"
99 #include "llvm/Analysis/VectorUtils.h"
100 #include "llvm/Transforms/Utils/LoopUtils.h"
105 using namespace llvm;
106 using namespace llvm::PatternMatch;
108 #define LV_NAME "loop-vectorize"
109 #define DEBUG_TYPE LV_NAME
111 STATISTIC(LoopsVectorized, "Number of loops vectorized");
112 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 static cl::opt<bool> EnableInterleavedMemAccesses(
142 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
143 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
145 /// Maximum factor for an interleaved memory access.
146 static cl::opt<unsigned> MaxInterleaveGroupFactor(
147 "max-interleave-group-factor", cl::Hidden,
148 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
151 /// We don't interleave loops with a known constant trip count below this
153 static const unsigned TinyTripCountInterleaveThreshold = 128;
155 static cl::opt<unsigned> ForceTargetNumScalarRegs(
156 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of scalar registers."));
159 static cl::opt<unsigned> ForceTargetNumVectorRegs(
160 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's number of vector registers."));
163 /// Maximum vectorization interleave count.
164 static const unsigned MaxInterleaveFactor = 16;
166 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
167 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max interleave factor for "
171 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
172 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
174 "vectorized loops."));
176 static cl::opt<unsigned> ForceTargetInstructionCost(
177 "force-target-instruction-cost", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's expected cost for "
179 "an instruction to a single constant value. Mostly "
180 "useful for getting consistent testing."));
182 static cl::opt<unsigned> SmallLoopCost(
183 "small-loop-cost", cl::init(20), cl::Hidden,
185 "The cost of a loop that is considered 'small' by the interleaver."));
187 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
188 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
189 cl::desc("Enable the use of the block frequency analysis to access PGO "
190 "heuristics minimizing code growth in cold regions and being more "
191 "aggressive in hot regions."));
193 // Runtime interleave loops for load/store throughput.
194 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
195 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
197 "Enable runtime interleaving until load/store ports are saturated"));
199 /// The number of stores in a loop that are allowed to need predication.
200 static cl::opt<unsigned> NumberOfStoresToPredicate(
201 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
202 cl::desc("Max number of stores to be predicated behind an if."));
204 static cl::opt<bool> EnableIndVarRegisterHeur(
205 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
206 cl::desc("Count the induction variable only once when interleaving"));
208 static cl::opt<bool> EnableCondStoresVectorization(
209 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
210 cl::desc("Enable if predication of stores during vectorization."));
212 static cl::opt<unsigned> MaxNestedScalarReductionIC(
213 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
214 cl::desc("The maximum interleave count to use when interleaving a scalar "
215 "reduction in a nested loop."));
219 // Forward declarations.
220 class LoopVectorizeHints;
221 class LoopVectorizationLegality;
222 class LoopVectorizationCostModel;
223 class LoopVectorizationRequirements;
225 /// \brief This modifies LoopAccessReport to initialize message with
226 /// loop-vectorizer-specific part.
227 class VectorizationReport : public LoopAccessReport {
229 VectorizationReport(Instruction *I = nullptr)
230 : LoopAccessReport("loop not vectorized: ", I) {}
232 /// \brief This allows promotion of the loop-access analysis report into the
233 /// loop-vectorizer report. It modifies the message to add the
234 /// loop-vectorizer-specific part of the message.
235 explicit VectorizationReport(const LoopAccessReport &R)
236 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
240 /// A helper function for converting Scalar types to vector types.
241 /// If the incoming type is void, we return void. If the VF is 1, we return
243 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
244 if (Scalar->isVoidTy() || VF == 1)
246 return VectorType::get(Scalar, VF);
249 /// InnerLoopVectorizer vectorizes loops which contain only one basic
250 /// block to a specified vectorization factor (VF).
251 /// This class performs the widening of scalars into vectors, or multiple
252 /// scalars. This class also implements the following features:
253 /// * It inserts an epilogue loop for handling loops that don't have iteration
254 /// counts that are known to be a multiple of the vectorization factor.
255 /// * It handles the code generation for reduction variables.
256 /// * Scalarization (implementation using scalars) of un-vectorizable
258 /// InnerLoopVectorizer does not perform any vectorization-legality
259 /// checks, and relies on the caller to check for the different legality
260 /// aspects. The InnerLoopVectorizer relies on the
261 /// LoopVectorizationLegality class to provide information about the induction
262 /// and reduction variables that were found to a given vectorization factor.
263 class InnerLoopVectorizer {
265 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
266 DominatorTree *DT, const TargetLibraryInfo *TLI,
267 const TargetTransformInfo *TTI, unsigned VecWidth,
268 unsigned UnrollFactor)
269 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
270 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
271 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
272 Legal(nullptr), AddedSafetyChecks(false) {}
274 // Perform the actual loop widening (vectorization).
275 void vectorize(LoopVectorizationLegality *L) {
277 // Create a new empty loop. Unlink the old loop and connect the new one.
279 // Widen each instruction in the old loop to a new one in the new loop.
280 // Use the Legality module to find the induction and reduction variables.
282 // Register the new loop and update the analysis passes.
286 // Return true if any runtime check is added.
287 bool IsSafetyChecksAdded() {
288 return AddedSafetyChecks;
291 virtual ~InnerLoopVectorizer() {}
294 /// A small list of PHINodes.
295 typedef SmallVector<PHINode*, 4> PhiVector;
296 /// When we unroll loops we have multiple vector values for each scalar.
297 /// This data structure holds the unrolled and vectorized values that
298 /// originated from one scalar instruction.
299 typedef SmallVector<Value*, 2> VectorParts;
301 // When we if-convert we need to create edge masks. We have to cache values
302 // so that we don't end up with exponential recursion/IR.
303 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
304 VectorParts> EdgeMaskCache;
306 /// \brief Add checks for strides that were assumed to be 1.
308 /// Returns the last check instruction and the first check instruction in the
309 /// pair as (first, last).
310 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
312 /// Create an empty loop, based on the loop ranges of the old loop.
313 void createEmptyLoop();
314 /// Copy and widen the instructions from the old loop.
315 virtual void vectorizeLoop();
317 /// \brief The Loop exit block may have single value PHI nodes where the
318 /// incoming value is 'Undef'. While vectorizing we only handled real values
319 /// that were defined inside the loop. Here we fix the 'undef case'.
323 /// A helper function that computes the predicate of the block BB, assuming
324 /// that the header block of the loop is set to True. It returns the *entry*
325 /// mask for the block BB.
326 VectorParts createBlockInMask(BasicBlock *BB);
327 /// A helper function that computes the predicate of the edge between SRC
329 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
331 /// A helper function to vectorize a single BB within the innermost loop.
332 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
334 /// Vectorize a single PHINode in a block. This method handles the induction
335 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
336 /// arbitrary length vectors.
337 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
338 unsigned UF, unsigned VF, PhiVector *PV);
340 /// Insert the new loop to the loop hierarchy and pass manager
341 /// and update the analysis passes.
342 void updateAnalysis();
344 /// This instruction is un-vectorizable. Implement it as a sequence
345 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
346 /// scalarized instruction behind an if block predicated on the control
347 /// dependence of the instruction.
348 virtual void scalarizeInstruction(Instruction *Instr,
349 bool IfPredicateStore=false);
351 /// Vectorize Load and Store instructions,
352 virtual void vectorizeMemoryInstruction(Instruction *Instr);
354 /// Create a broadcast instruction. This method generates a broadcast
355 /// instruction (shuffle) for loop invariant values and for the induction
356 /// value. If this is the induction variable then we extend it to N, N+1, ...
357 /// this is needed because each iteration in the loop corresponds to a SIMD
359 virtual Value *getBroadcastInstrs(Value *V);
361 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
362 /// to each vector element of Val. The sequence starts at StartIndex.
363 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
365 /// When we go over instructions in the basic block we rely on previous
366 /// values within the current basic block or on loop invariant values.
367 /// When we widen (vectorize) values we place them in the map. If the values
368 /// are not within the map, they have to be loop invariant, so we simply
369 /// broadcast them into a vector.
370 VectorParts &getVectorValue(Value *V);
372 /// Try to vectorize the interleaved access group that \p Instr belongs to.
373 void vectorizeInterleaveGroup(Instruction *Instr);
375 /// Generate a shuffle sequence that will reverse the vector Vec.
376 virtual Value *reverseVector(Value *Vec);
378 /// This is a helper class that holds the vectorizer state. It maps scalar
379 /// instructions to vector instructions. When the code is 'unrolled' then
380 /// then a single scalar value is mapped to multiple vector parts. The parts
381 /// are stored in the VectorPart type.
383 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
385 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
387 /// \return True if 'Key' is saved in the Value Map.
388 bool has(Value *Key) const { return MapStorage.count(Key); }
390 /// Initializes a new entry in the map. Sets all of the vector parts to the
391 /// save value in 'Val'.
392 /// \return A reference to a vector with splat values.
393 VectorParts &splat(Value *Key, Value *Val) {
394 VectorParts &Entry = MapStorage[Key];
395 Entry.assign(UF, Val);
399 ///\return A reference to the value that is stored at 'Key'.
400 VectorParts &get(Value *Key) {
401 VectorParts &Entry = MapStorage[Key];
404 assert(Entry.size() == UF);
409 /// The unroll factor. Each entry in the map stores this number of vector
413 /// Map storage. We use std::map and not DenseMap because insertions to a
414 /// dense map invalidates its iterators.
415 std::map<Value *, VectorParts> MapStorage;
418 /// The original loop.
420 /// Scev analysis to use.
428 /// Target Library Info.
429 const TargetLibraryInfo *TLI;
430 /// Target Transform Info.
431 const TargetTransformInfo *TTI;
433 /// The vectorization SIMD factor to use. Each vector will have this many
438 /// The vectorization unroll factor to use. Each scalar is vectorized to this
439 /// many different vector instructions.
442 /// The builder that we use
445 // --- Vectorization state ---
447 /// The vector-loop preheader.
448 BasicBlock *LoopVectorPreHeader;
449 /// The scalar-loop preheader.
450 BasicBlock *LoopScalarPreHeader;
451 /// Middle Block between the vector and the scalar.
452 BasicBlock *LoopMiddleBlock;
453 ///The ExitBlock of the scalar loop.
454 BasicBlock *LoopExitBlock;
455 ///The vector loop body.
456 SmallVector<BasicBlock *, 4> LoopVectorBody;
457 ///The scalar loop body.
458 BasicBlock *LoopScalarBody;
459 /// A list of all bypass blocks. The first block is the entry of the loop.
460 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
462 /// The new Induction variable which was added to the new block.
464 /// The induction variable of the old basic block.
465 PHINode *OldInduction;
466 /// Holds the extended (to the widest induction type) start index.
468 /// Maps scalars to widened vectors.
470 EdgeMaskCache MaskCache;
472 LoopVectorizationLegality *Legal;
474 // Record whether runtime check is added.
475 bool AddedSafetyChecks;
478 class InnerLoopUnroller : public InnerLoopVectorizer {
480 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
481 DominatorTree *DT, const TargetLibraryInfo *TLI,
482 const TargetTransformInfo *TTI, unsigned UnrollFactor)
483 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
486 void scalarizeInstruction(Instruction *Instr,
487 bool IfPredicateStore = false) override;
488 void vectorizeMemoryInstruction(Instruction *Instr) override;
489 Value *getBroadcastInstrs(Value *V) override;
490 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
491 Value *reverseVector(Value *Vec) override;
494 /// \brief Look for a meaningful debug location on the instruction or it's
496 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
501 if (I->getDebugLoc() != Empty)
504 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
505 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
506 if (OpInst->getDebugLoc() != Empty)
513 /// \brief Set the debug location in the builder using the debug location in the
515 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
516 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
517 B.SetCurrentDebugLocation(Inst->getDebugLoc());
519 B.SetCurrentDebugLocation(DebugLoc());
523 /// \return string containing a file name and a line # for the given loop.
524 static std::string getDebugLocString(const Loop *L) {
527 raw_string_ostream OS(Result);
528 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
529 LoopDbgLoc.print(OS);
531 // Just print the module name.
532 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
539 /// \brief Propagate known metadata from one instruction to another.
540 static void propagateMetadata(Instruction *To, const Instruction *From) {
541 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
542 From->getAllMetadataOtherThanDebugLoc(Metadata);
544 for (auto M : Metadata) {
545 unsigned Kind = M.first;
547 // These are safe to transfer (this is safe for TBAA, even when we
548 // if-convert, because should that metadata have had a control dependency
549 // on the condition, and thus actually aliased with some other
550 // non-speculated memory access when the condition was false, this would be
551 // caught by the runtime overlap checks).
552 if (Kind != LLVMContext::MD_tbaa &&
553 Kind != LLVMContext::MD_alias_scope &&
554 Kind != LLVMContext::MD_noalias &&
555 Kind != LLVMContext::MD_fpmath &&
556 Kind != LLVMContext::MD_nontemporal)
559 To->setMetadata(Kind, M.second);
563 /// \brief Propagate known metadata from one instruction to a vector of others.
564 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
566 if (Instruction *I = dyn_cast<Instruction>(V))
567 propagateMetadata(I, From);
570 /// \brief The group of interleaved loads/stores sharing the same stride and
571 /// close to each other.
573 /// Each member in this group has an index starting from 0, and the largest
574 /// index should be less than interleaved factor, which is equal to the absolute
575 /// value of the access's stride.
577 /// E.g. An interleaved load group of factor 4:
578 /// for (unsigned i = 0; i < 1024; i+=4) {
579 /// a = A[i]; // Member of index 0
580 /// b = A[i+1]; // Member of index 1
581 /// d = A[i+3]; // Member of index 3
585 /// An interleaved store group of factor 4:
586 /// for (unsigned i = 0; i < 1024; i+=4) {
588 /// A[i] = a; // Member of index 0
589 /// A[i+1] = b; // Member of index 1
590 /// A[i+2] = c; // Member of index 2
591 /// A[i+3] = d; // Member of index 3
594 /// Note: the interleaved load group could have gaps (missing members), but
595 /// the interleaved store group doesn't allow gaps.
596 class InterleaveGroup {
598 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
599 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
600 assert(Align && "The alignment should be non-zero");
602 Factor = std::abs(Stride);
603 assert(Factor > 1 && "Invalid interleave factor");
605 Reverse = Stride < 0;
609 bool isReverse() const { return Reverse; }
610 unsigned getFactor() const { return Factor; }
611 unsigned getAlignment() const { return Align; }
612 unsigned getNumMembers() const { return Members.size(); }
614 /// \brief Try to insert a new member \p Instr with index \p Index and
615 /// alignment \p NewAlign. The index is related to the leader and it could be
616 /// negative if it is the new leader.
618 /// \returns false if the instruction doesn't belong to the group.
619 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
620 assert(NewAlign && "The new member's alignment should be non-zero");
622 int Key = Index + SmallestKey;
624 // Skip if there is already a member with the same index.
625 if (Members.count(Key))
628 if (Key > LargestKey) {
629 // The largest index is always less than the interleave factor.
630 if (Index >= static_cast<int>(Factor))
634 } else if (Key < SmallestKey) {
635 // The largest index is always less than the interleave factor.
636 if (LargestKey - Key >= static_cast<int>(Factor))
642 // It's always safe to select the minimum alignment.
643 Align = std::min(Align, NewAlign);
644 Members[Key] = Instr;
648 /// \brief Get the member with the given index \p Index
650 /// \returns nullptr if contains no such member.
651 Instruction *getMember(unsigned Index) const {
652 int Key = SmallestKey + Index;
653 if (!Members.count(Key))
656 return Members.find(Key)->second;
659 /// \brief Get the index for the given member. Unlike the key in the member
660 /// map, the index starts from 0.
661 unsigned getIndex(Instruction *Instr) const {
662 for (auto I : Members)
663 if (I.second == Instr)
664 return I.first - SmallestKey;
666 llvm_unreachable("InterleaveGroup contains no such member");
669 Instruction *getInsertPos() const { return InsertPos; }
670 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
673 unsigned Factor; // Interleave Factor.
676 DenseMap<int, Instruction *> Members;
680 // To avoid breaking dependences, vectorized instructions of an interleave
681 // group should be inserted at either the first load or the last store in
684 // E.g. %even = load i32 // Insert Position
685 // %add = add i32 %even // Use of %even
689 // %odd = add i32 // Def of %odd
690 // store i32 %odd // Insert Position
691 Instruction *InsertPos;
694 /// \brief Drive the analysis of interleaved memory accesses in the loop.
696 /// Use this class to analyze interleaved accesses only when we can vectorize
697 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
698 /// on interleaved accesses is unsafe.
700 /// The analysis collects interleave groups and records the relationships
701 /// between the member and the group in a map.
702 class InterleavedAccessInfo {
704 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
705 : SE(SE), TheLoop(L), DT(DT) {}
707 ~InterleavedAccessInfo() {
708 SmallSet<InterleaveGroup *, 4> DelSet;
709 // Avoid releasing a pointer twice.
710 for (auto &I : InterleaveGroupMap)
711 DelSet.insert(I.second);
712 for (auto *Ptr : DelSet)
716 /// \brief Analyze the interleaved accesses and collect them in interleave
717 /// groups. Substitute symbolic strides using \p Strides.
718 void analyzeInterleaving(const ValueToValueMap &Strides);
720 /// \brief Check if \p Instr belongs to any interleave group.
721 bool isInterleaved(Instruction *Instr) const {
722 return InterleaveGroupMap.count(Instr);
725 /// \brief Get the interleave group that \p Instr belongs to.
727 /// \returns nullptr if doesn't have such group.
728 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
729 if (InterleaveGroupMap.count(Instr))
730 return InterleaveGroupMap.find(Instr)->second;
739 /// Holds the relationships between the members and the interleave group.
740 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
742 /// \brief The descriptor for a strided memory access.
743 struct StrideDescriptor {
744 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
746 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
748 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
750 int Stride; // The access's stride. It is negative for a reverse access.
751 const SCEV *Scev; // The scalar expression of this access
752 unsigned Size; // The size of the memory object.
753 unsigned Align; // The alignment of this access.
756 /// \brief Create a new interleave group with the given instruction \p Instr,
757 /// stride \p Stride and alignment \p Align.
759 /// \returns the newly created interleave group.
760 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
762 assert(!InterleaveGroupMap.count(Instr) &&
763 "Already in an interleaved access group");
764 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
765 return InterleaveGroupMap[Instr];
768 /// \brief Release the group and remove all the relationships.
769 void releaseGroup(InterleaveGroup *Group) {
770 for (unsigned i = 0; i < Group->getFactor(); i++)
771 if (Instruction *Member = Group->getMember(i))
772 InterleaveGroupMap.erase(Member);
777 /// \brief Collect all the accesses with a constant stride in program order.
778 void collectConstStridedAccesses(
779 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
780 const ValueToValueMap &Strides);
783 /// Utility class for getting and setting loop vectorizer hints in the form
784 /// of loop metadata.
785 /// This class keeps a number of loop annotations locally (as member variables)
786 /// and can, upon request, write them back as metadata on the loop. It will
787 /// initially scan the loop for existing metadata, and will update the local
788 /// values based on information in the loop.
789 /// We cannot write all values to metadata, as the mere presence of some info,
790 /// for example 'force', means a decision has been made. So, we need to be
791 /// careful NOT to add them if the user hasn't specifically asked so.
792 class LoopVectorizeHints {
799 /// Hint - associates name and validation with the hint value.
802 unsigned Value; // This may have to change for non-numeric values.
805 Hint(const char * Name, unsigned Value, HintKind Kind)
806 : Name(Name), Value(Value), Kind(Kind) { }
808 bool validate(unsigned Val) {
811 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
813 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
821 /// Vectorization width.
823 /// Vectorization interleave factor.
825 /// Vectorization forced
828 /// Return the loop metadata prefix.
829 static StringRef Prefix() { return "llvm.loop."; }
833 FK_Undefined = -1, ///< Not selected.
834 FK_Disabled = 0, ///< Forcing disabled.
835 FK_Enabled = 1, ///< Forcing enabled.
838 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
839 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
841 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
842 Force("vectorize.enable", FK_Undefined, HK_FORCE),
844 // Populate values with existing loop metadata.
845 getHintsFromMetadata();
847 // force-vector-interleave overrides DisableInterleaving.
848 if (VectorizerParams::isInterleaveForced())
849 Interleave.Value = VectorizerParams::VectorizationInterleave;
851 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
852 << "LV: Interleaving disabled by the pass manager\n");
855 /// Mark the loop L as already vectorized by setting the width to 1.
856 void setAlreadyVectorized() {
857 Width.Value = Interleave.Value = 1;
858 Hint Hints[] = {Width, Interleave};
859 writeHintsToMetadata(Hints);
862 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
863 if (getForce() == LoopVectorizeHints::FK_Disabled) {
864 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
865 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
866 L->getStartLoc(), emitRemark());
870 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
871 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
872 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
873 L->getStartLoc(), emitRemark());
877 if (getWidth() == 1 && getInterleave() == 1) {
878 // FIXME: Add a separate metadata to indicate when the loop has already
879 // been vectorized instead of setting width and count to 1.
880 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
881 // FIXME: Add interleave.disable metadata. This will allow
882 // vectorize.disable to be used without disabling the pass and errors
883 // to differentiate between disabled vectorization and a width of 1.
884 emitOptimizationRemarkAnalysis(
885 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
886 "loop not vectorized: vectorization and interleaving are explicitly "
887 "disabled, or vectorize width and interleave count are both set to "
895 /// Dumps all the hint information.
896 std::string emitRemark() const {
897 VectorizationReport R;
898 if (Force.Value == LoopVectorizeHints::FK_Disabled)
899 R << "vectorization is explicitly disabled";
901 R << "use -Rpass-analysis=loop-vectorize for more info";
902 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
904 if (Width.Value != 0)
905 R << ", Vector Width=" << Width.Value;
906 if (Interleave.Value != 0)
907 R << ", Interleave Count=" << Interleave.Value;
915 unsigned getWidth() const { return Width.Value; }
916 unsigned getInterleave() const { return Interleave.Value; }
917 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
918 bool isForced() const {
919 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1 ||
924 /// Find hints specified in the loop metadata and update local values.
925 void getHintsFromMetadata() {
926 MDNode *LoopID = TheLoop->getLoopID();
930 // First operand should refer to the loop id itself.
931 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
932 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
934 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
935 const MDString *S = nullptr;
936 SmallVector<Metadata *, 4> Args;
938 // The expected hint is either a MDString or a MDNode with the first
939 // operand a MDString.
940 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
941 if (!MD || MD->getNumOperands() == 0)
943 S = dyn_cast<MDString>(MD->getOperand(0));
944 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
945 Args.push_back(MD->getOperand(i));
947 S = dyn_cast<MDString>(LoopID->getOperand(i));
948 assert(Args.size() == 0 && "too many arguments for MDString");
954 // Check if the hint starts with the loop metadata prefix.
955 StringRef Name = S->getString();
956 if (Args.size() == 1)
957 setHint(Name, Args[0]);
961 /// Checks string hint with one operand and set value if valid.
962 void setHint(StringRef Name, Metadata *Arg) {
963 if (!Name.startswith(Prefix()))
965 Name = Name.substr(Prefix().size(), StringRef::npos);
967 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
969 unsigned Val = C->getZExtValue();
971 Hint *Hints[] = {&Width, &Interleave, &Force};
972 for (auto H : Hints) {
973 if (Name == H->Name) {
974 if (H->validate(Val))
977 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
983 /// Create a new hint from name / value pair.
984 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
985 LLVMContext &Context = TheLoop->getHeader()->getContext();
986 Metadata *MDs[] = {MDString::get(Context, Name),
987 ConstantAsMetadata::get(
988 ConstantInt::get(Type::getInt32Ty(Context), V))};
989 return MDNode::get(Context, MDs);
992 /// Matches metadata with hint name.
993 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
994 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
998 for (auto H : HintTypes)
999 if (Name->getString().endswith(H.Name))
1004 /// Sets current hints into loop metadata, keeping other values intact.
1005 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1006 if (HintTypes.size() == 0)
1009 // Reserve the first element to LoopID (see below).
1010 SmallVector<Metadata *, 4> MDs(1);
1011 // If the loop already has metadata, then ignore the existing operands.
1012 MDNode *LoopID = TheLoop->getLoopID();
1014 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1015 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1016 // If node in update list, ignore old value.
1017 if (!matchesHintMetadataName(Node, HintTypes))
1018 MDs.push_back(Node);
1022 // Now, add the missing hints.
1023 for (auto H : HintTypes)
1024 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1026 // Replace current metadata node with new one.
1027 LLVMContext &Context = TheLoop->getHeader()->getContext();
1028 MDNode *NewLoopID = MDNode::get(Context, MDs);
1029 // Set operand 0 to refer to the loop id itself.
1030 NewLoopID->replaceOperandWith(0, NewLoopID);
1032 TheLoop->setLoopID(NewLoopID);
1035 /// The loop these hints belong to.
1036 const Loop *TheLoop;
1039 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1040 const LoopVectorizeHints &Hints,
1041 const LoopAccessReport &Message) {
1042 // If a loop hint is provided the diagnostic is always produced.
1043 const char *Name = Hints.isForced() ? DiagnosticInfo::AlwaysPrint : LV_NAME;
1044 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1047 static void emitMissedWarning(Function *F, Loop *L,
1048 const LoopVectorizeHints &LH) {
1049 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1050 L->getStartLoc(), LH.emitRemark());
1052 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1053 if (LH.getWidth() != 1)
1054 emitLoopVectorizeWarning(
1055 F->getContext(), *F, L->getStartLoc(),
1056 "failed explicitly specified loop vectorization");
1057 else if (LH.getInterleave() != 1)
1058 emitLoopInterleaveWarning(
1059 F->getContext(), *F, L->getStartLoc(),
1060 "failed explicitly specified loop interleaving");
1064 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1065 /// to what vectorization factor.
1066 /// This class does not look at the profitability of vectorization, only the
1067 /// legality. This class has two main kinds of checks:
1068 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1069 /// will change the order of memory accesses in a way that will change the
1070 /// correctness of the program.
1071 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1072 /// checks for a number of different conditions, such as the availability of a
1073 /// single induction variable, that all types are supported and vectorize-able,
1074 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1075 /// This class is also used by InnerLoopVectorizer for identifying
1076 /// induction variable and the different reduction variables.
1077 class LoopVectorizationLegality {
1079 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1080 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1081 Function *F, const TargetTransformInfo *TTI,
1082 LoopAccessAnalysis *LAA,
1083 LoopVectorizationRequirements *R,
1084 const LoopVectorizeHints *H)
1085 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1086 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1087 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1088 Requirements(R), Hints(H) {}
1090 /// This enum represents the kinds of inductions that we support.
1091 enum InductionKind {
1092 IK_NoInduction, ///< Not an induction variable.
1093 IK_IntInduction, ///< Integer induction variable. Step = C.
1094 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
1097 /// A struct for saving information about induction variables.
1098 struct InductionInfo {
1099 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
1100 : StartValue(Start), IK(K), StepValue(Step) {
1101 assert(IK != IK_NoInduction && "Not an induction");
1102 assert(StartValue && "StartValue is null");
1103 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
1104 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
1105 "StartValue is not a pointer for pointer induction");
1106 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
1107 "StartValue is not an integer for integer induction");
1108 assert(StepValue->getType()->isIntegerTy() &&
1109 "StepValue is not an integer");
1112 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
1114 /// Get the consecutive direction. Returns:
1115 /// 0 - unknown or non-consecutive.
1116 /// 1 - consecutive and increasing.
1117 /// -1 - consecutive and decreasing.
1118 int getConsecutiveDirection() const {
1119 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
1120 return StepValue->getSExtValue();
1124 /// Compute the transformed value of Index at offset StartValue using step
1126 /// For integer induction, returns StartValue + Index * StepValue.
1127 /// For pointer induction, returns StartValue[Index * StepValue].
1128 /// FIXME: The newly created binary instructions should contain nsw/nuw
1129 /// flags, which can be found from the original scalar operations.
1130 Value *transform(IRBuilder<> &B, Value *Index) const {
1132 case IK_IntInduction:
1133 assert(Index->getType() == StartValue->getType() &&
1134 "Index type does not match StartValue type");
1135 if (StepValue->isMinusOne())
1136 return B.CreateSub(StartValue, Index);
1137 if (!StepValue->isOne())
1138 Index = B.CreateMul(Index, StepValue);
1139 return B.CreateAdd(StartValue, Index);
1141 case IK_PtrInduction:
1142 assert(Index->getType() == StepValue->getType() &&
1143 "Index type does not match StepValue type");
1144 if (StepValue->isMinusOne())
1145 Index = B.CreateNeg(Index);
1146 else if (!StepValue->isOne())
1147 Index = B.CreateMul(Index, StepValue);
1148 return B.CreateGEP(nullptr, StartValue, Index);
1150 case IK_NoInduction:
1153 llvm_unreachable("invalid enum");
1157 TrackingVH<Value> StartValue;
1161 ConstantInt *StepValue;
1164 /// ReductionList contains the reduction descriptors for all
1165 /// of the reductions that were found in the loop.
1166 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1168 /// InductionList saves induction variables and maps them to the
1169 /// induction descriptor.
1170 typedef MapVector<PHINode*, InductionInfo> InductionList;
1172 /// Returns true if it is legal to vectorize this loop.
1173 /// This does not mean that it is profitable to vectorize this
1174 /// loop, only that it is legal to do so.
1175 bool canVectorize();
1177 /// Returns the Induction variable.
1178 PHINode *getInduction() { return Induction; }
1180 /// Returns the reduction variables found in the loop.
1181 ReductionList *getReductionVars() { return &Reductions; }
1183 /// Returns the induction variables found in the loop.
1184 InductionList *getInductionVars() { return &Inductions; }
1186 /// Returns the widest induction type.
1187 Type *getWidestInductionType() { return WidestIndTy; }
1189 /// Returns True if V is an induction variable in this loop.
1190 bool isInductionVariable(const Value *V);
1192 /// Return true if the block BB needs to be predicated in order for the loop
1193 /// to be vectorized.
1194 bool blockNeedsPredication(BasicBlock *BB);
1196 /// Check if this pointer is consecutive when vectorizing. This happens
1197 /// when the last index of the GEP is the induction variable, or that the
1198 /// pointer itself is an induction variable.
1199 /// This check allows us to vectorize A[idx] into a wide load/store.
1201 /// 0 - Stride is unknown or non-consecutive.
1202 /// 1 - Address is consecutive.
1203 /// -1 - Address is consecutive, and decreasing.
1204 int isConsecutivePtr(Value *Ptr);
1206 /// Returns true if the value V is uniform within the loop.
1207 bool isUniform(Value *V);
1209 /// Returns true if this instruction will remain scalar after vectorization.
1210 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1212 /// Returns the information that we collected about runtime memory check.
1213 const RuntimePointerChecking *getRuntimePointerChecking() const {
1214 return LAI->getRuntimePointerChecking();
1217 const LoopAccessInfo *getLAI() const {
1221 /// \brief Check if \p Instr belongs to any interleaved access group.
1222 bool isAccessInterleaved(Instruction *Instr) {
1223 return InterleaveInfo.isInterleaved(Instr);
1226 /// \brief Get the interleaved access group that \p Instr belongs to.
1227 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1228 return InterleaveInfo.getInterleaveGroup(Instr);
1231 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1233 bool hasStride(Value *V) { return StrideSet.count(V); }
1234 bool mustCheckStrides() { return !StrideSet.empty(); }
1235 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1236 return StrideSet.begin();
1238 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1240 /// Returns true if the target machine supports masked store operation
1241 /// for the given \p DataType and kind of access to \p Ptr.
1242 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1243 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1245 /// Returns true if the target machine supports masked load operation
1246 /// for the given \p DataType and kind of access to \p Ptr.
1247 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1248 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1250 /// Returns true if vector representation of the instruction \p I
1252 bool isMaskRequired(const Instruction* I) {
1253 return (MaskedOp.count(I) != 0);
1255 unsigned getNumStores() const {
1256 return LAI->getNumStores();
1258 unsigned getNumLoads() const {
1259 return LAI->getNumLoads();
1261 unsigned getNumPredStores() const {
1262 return NumPredStores;
1265 /// Check if a single basic block loop is vectorizable.
1266 /// At this point we know that this is a loop with a constant trip count
1267 /// and we only need to check individual instructions.
1268 bool canVectorizeInstrs();
1270 /// When we vectorize loops we may change the order in which
1271 /// we read and write from memory. This method checks if it is
1272 /// legal to vectorize the code, considering only memory constrains.
1273 /// Returns true if the loop is vectorizable
1274 bool canVectorizeMemory();
1276 /// Return true if we can vectorize this loop using the IF-conversion
1278 bool canVectorizeWithIfConvert();
1280 /// Collect the variables that need to stay uniform after vectorization.
1281 void collectLoopUniforms();
1283 /// Return true if all of the instructions in the block can be speculatively
1284 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1285 /// and we know that we can read from them without segfault.
1286 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1288 /// Returns the induction kind of Phi and record the step. This function may
1289 /// return NoInduction if the PHI is not an induction variable.
1290 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
1292 /// \brief Collect memory access with loop invariant strides.
1294 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1296 void collectStridedAccess(Value *LoadOrStoreInst);
1298 /// Report an analysis message to assist the user in diagnosing loops that are
1299 /// not vectorized. These are handled as LoopAccessReport rather than
1300 /// VectorizationReport because the << operator of VectorizationReport returns
1301 /// LoopAccessReport.
1302 void emitAnalysis(const LoopAccessReport &Message) const {
1303 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1306 unsigned NumPredStores;
1308 /// The loop that we evaluate.
1311 ScalarEvolution *SE;
1312 /// Target Library Info.
1313 TargetLibraryInfo *TLI;
1315 Function *TheFunction;
1316 /// Target Transform Info
1317 const TargetTransformInfo *TTI;
1320 // LoopAccess analysis.
1321 LoopAccessAnalysis *LAA;
1322 // And the loop-accesses info corresponding to this loop. This pointer is
1323 // null until canVectorizeMemory sets it up.
1324 const LoopAccessInfo *LAI;
1326 /// The interleave access information contains groups of interleaved accesses
1327 /// with the same stride and close to each other.
1328 InterleavedAccessInfo InterleaveInfo;
1330 // --- vectorization state --- //
1332 /// Holds the integer induction variable. This is the counter of the
1335 /// Holds the reduction variables.
1336 ReductionList Reductions;
1337 /// Holds all of the induction variables that we found in the loop.
1338 /// Notice that inductions don't need to start at zero and that induction
1339 /// variables can be pointers.
1340 InductionList Inductions;
1341 /// Holds the widest induction type encountered.
1344 /// Allowed outside users. This holds the reduction
1345 /// vars which can be accessed from outside the loop.
1346 SmallPtrSet<Value*, 4> AllowedExit;
1347 /// This set holds the variables which are known to be uniform after
1349 SmallPtrSet<Instruction*, 4> Uniforms;
1351 /// Can we assume the absence of NaNs.
1352 bool HasFunNoNaNAttr;
1354 /// Vectorization requirements that will go through late-evaluation.
1355 LoopVectorizationRequirements *Requirements;
1357 /// Used to emit an analysis of any legality issues.
1358 const LoopVectorizeHints *Hints;
1360 ValueToValueMap Strides;
1361 SmallPtrSet<Value *, 8> StrideSet;
1363 /// While vectorizing these instructions we have to generate a
1364 /// call to the appropriate masked intrinsic
1365 SmallPtrSet<const Instruction*, 8> MaskedOp;
1368 /// LoopVectorizationCostModel - estimates the expected speedups due to
1370 /// In many cases vectorization is not profitable. This can happen because of
1371 /// a number of reasons. In this class we mainly attempt to predict the
1372 /// expected speedup/slowdowns due to the supported instruction set. We use the
1373 /// TargetTransformInfo to query the different backends for the cost of
1374 /// different operations.
1375 class LoopVectorizationCostModel {
1377 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1378 LoopVectorizationLegality *Legal,
1379 const TargetTransformInfo &TTI,
1380 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1381 const Function *F, const LoopVectorizeHints *Hints)
1382 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1383 TheFunction(F), Hints(Hints) {
1384 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
1387 /// Information about vectorization costs
1388 struct VectorizationFactor {
1389 unsigned Width; // Vector width with best cost
1390 unsigned Cost; // Cost of the loop with that width
1392 /// \return The most profitable vectorization factor and the cost of that VF.
1393 /// This method checks every power of two up to VF. If UserVF is not ZERO
1394 /// then this vectorization factor will be selected if vectorization is
1396 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1398 /// \return The size (in bits) of the widest type in the code that
1399 /// needs to be vectorized. We ignore values that remain scalar such as
1400 /// 64 bit loop indices.
1401 unsigned getWidestType();
1403 /// \return The desired interleave count.
1404 /// If interleave count has been specified by metadata it will be returned.
1405 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1406 /// are the selected vectorization factor and the cost of the selected VF.
1407 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1410 /// \return The most profitable unroll factor.
1411 /// This method finds the best unroll-factor based on register pressure and
1412 /// other parameters. VF and LoopCost are the selected vectorization factor
1413 /// and the cost of the selected VF.
1414 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1417 /// \brief A struct that represents some properties of the register usage
1419 struct RegisterUsage {
1420 /// Holds the number of loop invariant values that are used in the loop.
1421 unsigned LoopInvariantRegs;
1422 /// Holds the maximum number of concurrent live intervals in the loop.
1423 unsigned MaxLocalUsers;
1424 /// Holds the number of instructions in the loop.
1425 unsigned NumInstructions;
1428 /// \return information about the register usage of the loop.
1429 RegisterUsage calculateRegisterUsage();
1432 /// Returns the expected execution cost. The unit of the cost does
1433 /// not matter because we use the 'cost' units to compare different
1434 /// vector widths. The cost that is returned is *not* normalized by
1435 /// the factor width.
1436 unsigned expectedCost(unsigned VF);
1438 /// Returns the execution time cost of an instruction for a given vector
1439 /// width. Vector width of one means scalar.
1440 unsigned getInstructionCost(Instruction *I, unsigned VF);
1442 /// Returns whether the instruction is a load or store and will be a emitted
1443 /// as a vector operation.
1444 bool isConsecutiveLoadOrStore(Instruction *I);
1446 /// Report an analysis message to assist the user in diagnosing loops that are
1447 /// not vectorized. These are handled as LoopAccessReport rather than
1448 /// VectorizationReport because the << operator of VectorizationReport returns
1449 /// LoopAccessReport.
1450 void emitAnalysis(const LoopAccessReport &Message) const {
1451 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1454 /// Values used only by @llvm.assume calls.
1455 SmallPtrSet<const Value *, 32> EphValues;
1457 /// The loop that we evaluate.
1460 ScalarEvolution *SE;
1461 /// Loop Info analysis.
1463 /// Vectorization legality.
1464 LoopVectorizationLegality *Legal;
1465 /// Vector target information.
1466 const TargetTransformInfo &TTI;
1467 /// Target Library Info.
1468 const TargetLibraryInfo *TLI;
1469 const Function *TheFunction;
1470 // Loop Vectorize Hint.
1471 const LoopVectorizeHints *Hints;
1474 /// \brief This holds vectorization requirements that must be verified late in
1475 /// the process. The requirements are set by legalize and costmodel. Once
1476 /// vectorization has been determined to be possible and profitable the
1477 /// requirements can be verified by looking for metadata or compiler options.
1478 /// For example, some loops require FP commutativity which is only allowed if
1479 /// vectorization is explicitly specified or if the fast-math compiler option
1480 /// has been provided.
1481 /// Late evaluation of these requirements allows helpful diagnostics to be
1482 /// composed that tells the user what need to be done to vectorize the loop. For
1483 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1484 /// evaluation should be used only when diagnostics can generated that can be
1485 /// followed by a non-expert user.
1486 class LoopVectorizationRequirements {
1488 LoopVectorizationRequirements()
1489 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1491 void addUnsafeAlgebraInst(Instruction *I) {
1492 // First unsafe algebra instruction.
1493 if (!UnsafeAlgebraInst)
1494 UnsafeAlgebraInst = I;
1497 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1499 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1500 // If a loop hint is provided the diagnostic is always produced.
1501 const char *Name = Hints.isForced() ? DiagnosticInfo::AlwaysPrint : LV_NAME;
1502 bool Failed = false;
1503 if (UnsafeAlgebraInst &&
1504 Hints.getForce() == LoopVectorizeHints::FK_Undefined &&
1505 Hints.getWidth() == 0) {
1506 emitOptimizationRemarkAnalysisFPCommute(
1507 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1508 VectorizationReport() << "vectorization requires changes in the "
1509 "order of operations, however IEEE 754 "
1510 "floating-point operations are not "
1515 if (NumRuntimePointerChecks >
1516 VectorizerParams::RuntimeMemoryCheckThreshold) {
1517 emitOptimizationRemarkAnalysisAliasing(
1518 F->getContext(), Name, *F, L->getStartLoc(),
1519 VectorizationReport()
1520 << "cannot prove pointers refer to independent arrays in memory. "
1521 "The loop requires "
1522 << NumRuntimePointerChecks
1523 << " runtime independence checks to vectorize the loop, but that "
1524 "would exceed the limit of "
1525 << VectorizerParams::RuntimeMemoryCheckThreshold << " checks");
1526 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1534 unsigned NumRuntimePointerChecks;
1535 Instruction *UnsafeAlgebraInst;
1538 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1540 return V.push_back(&L);
1542 for (Loop *InnerL : L)
1543 addInnerLoop(*InnerL, V);
1546 /// The LoopVectorize Pass.
1547 struct LoopVectorize : public FunctionPass {
1548 /// Pass identification, replacement for typeid
1551 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1553 DisableUnrolling(NoUnrolling),
1554 AlwaysVectorize(AlwaysVectorize) {
1555 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1558 ScalarEvolution *SE;
1560 TargetTransformInfo *TTI;
1562 BlockFrequencyInfo *BFI;
1563 TargetLibraryInfo *TLI;
1565 AssumptionCache *AC;
1566 LoopAccessAnalysis *LAA;
1567 bool DisableUnrolling;
1568 bool AlwaysVectorize;
1570 BlockFrequency ColdEntryFreq;
1572 bool runOnFunction(Function &F) override {
1573 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1574 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1575 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1576 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1577 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1578 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1579 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1580 AA = &getAnalysis<AliasAnalysis>();
1581 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1582 LAA = &getAnalysis<LoopAccessAnalysis>();
1584 // Compute some weights outside of the loop over the loops. Compute this
1585 // using a BranchProbability to re-use its scaling math.
1586 const BranchProbability ColdProb(1, 5); // 20%
1587 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1590 // 1. the target claims to have no vector registers, and
1591 // 2. interleaving won't help ILP.
1593 // The second condition is necessary because, even if the target has no
1594 // vector registers, loop vectorization may still enable scalar
1596 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1599 // Build up a worklist of inner-loops to vectorize. This is necessary as
1600 // the act of vectorizing or partially unrolling a loop creates new loops
1601 // and can invalidate iterators across the loops.
1602 SmallVector<Loop *, 8> Worklist;
1605 addInnerLoop(*L, Worklist);
1607 LoopsAnalyzed += Worklist.size();
1609 // Now walk the identified inner loops.
1610 bool Changed = false;
1611 while (!Worklist.empty())
1612 Changed |= processLoop(Worklist.pop_back_val());
1614 // Process each loop nest in the function.
1618 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1619 SmallVector<Metadata *, 4> MDs;
1620 // Reserve first location for self reference to the LoopID metadata node.
1621 MDs.push_back(nullptr);
1622 bool IsUnrollMetadata = false;
1623 MDNode *LoopID = L->getLoopID();
1625 // First find existing loop unrolling disable metadata.
1626 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1627 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1629 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1631 S && S->getString().startswith("llvm.loop.unroll.disable");
1633 MDs.push_back(LoopID->getOperand(i));
1637 if (!IsUnrollMetadata) {
1638 // Add runtime unroll disable metadata.
1639 LLVMContext &Context = L->getHeader()->getContext();
1640 SmallVector<Metadata *, 1> DisableOperands;
1641 DisableOperands.push_back(
1642 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1643 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1644 MDs.push_back(DisableNode);
1645 MDNode *NewLoopID = MDNode::get(Context, MDs);
1646 // Set operand 0 to refer to the loop id itself.
1647 NewLoopID->replaceOperandWith(0, NewLoopID);
1648 L->setLoopID(NewLoopID);
1652 bool processLoop(Loop *L) {
1653 assert(L->empty() && "Only process inner loops.");
1656 const std::string DebugLocStr = getDebugLocString(L);
1659 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1660 << L->getHeader()->getParent()->getName() << "\" from "
1661 << DebugLocStr << "\n");
1663 LoopVectorizeHints Hints(L, DisableUnrolling);
1665 DEBUG(dbgs() << "LV: Loop hints:"
1667 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1669 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1671 : "?")) << " width=" << Hints.getWidth()
1672 << " unroll=" << Hints.getInterleave() << "\n");
1674 // Function containing loop
1675 Function *F = L->getHeader()->getParent();
1677 // Looking at the diagnostic output is the only way to determine if a loop
1678 // was vectorized (other than looking at the IR or machine code), so it
1679 // is important to generate an optimization remark for each loop. Most of
1680 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1681 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1682 // less verbose reporting vectorized loops and unvectorized loops that may
1683 // benefit from vectorization, respectively.
1685 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1686 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1690 // Check the loop for a trip count threshold:
1691 // do not vectorize loops with a tiny trip count.
1692 const unsigned TC = SE->getSmallConstantTripCount(L);
1693 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1694 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1695 << "This loop is not worth vectorizing.");
1696 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1697 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1699 DEBUG(dbgs() << "\n");
1700 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1701 << "vectorization is not beneficial "
1702 "and is not explicitly forced");
1707 // Check if it is legal to vectorize the loop.
1708 LoopVectorizationRequirements Requirements;
1709 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1710 &Requirements, &Hints);
1711 if (!LVL.canVectorize()) {
1712 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1713 emitMissedWarning(F, L, Hints);
1717 // Use the cost model.
1718 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1720 // Check the function attributes to find out if this function should be
1721 // optimized for size.
1722 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1725 // Compute the weighted frequency of this loop being executed and see if it
1726 // is less than 20% of the function entry baseline frequency. Note that we
1727 // always have a canonical loop here because we think we *can* vectorize.
1728 // FIXME: This is hidden behind a flag due to pervasive problems with
1729 // exactly what block frequency models.
1730 if (LoopVectorizeWithBlockFrequency) {
1731 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1732 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1733 LoopEntryFreq < ColdEntryFreq)
1737 // Check the function attributes to see if implicit floats are allowed.
1738 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1739 // an integer loop and the vector instructions selected are purely integer
1740 // vector instructions?
1741 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1742 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1743 "attribute is used.\n");
1746 VectorizationReport()
1747 << "loop not vectorized due to NoImplicitFloat attribute");
1748 emitMissedWarning(F, L, Hints);
1752 // Select the optimal vectorization factor.
1753 const LoopVectorizationCostModel::VectorizationFactor VF =
1754 CM.selectVectorizationFactor(OptForSize);
1756 // Select the interleave count.
1757 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1759 // Get user interleave count.
1760 unsigned UserIC = Hints.getInterleave();
1762 // Identify the diagnostic messages that should be produced.
1763 std::string VecDiagMsg, IntDiagMsg;
1764 bool VectorizeLoop = true, InterleaveLoop = true;
1766 if (Requirements.doesNotMeet(F, L, Hints)) {
1767 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1769 emitMissedWarning(F, L, Hints);
1773 if (VF.Width == 1) {
1774 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1776 "the cost-model indicates that vectorization is not beneficial";
1777 VectorizeLoop = false;
1780 if (IC == 1 && UserIC <= 1) {
1781 // Tell the user interleaving is not beneficial.
1782 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1784 "the cost-model indicates that interleaving is not beneficial";
1785 InterleaveLoop = false;
1788 " and is explicitly disabled or interleave count is set to 1";
1789 } else if (IC > 1 && UserIC == 1) {
1790 // Tell the user interleaving is beneficial, but it explicitly disabled.
1792 << "LV: Interleaving is beneficial but is explicitly disabled.");
1793 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1794 "but is explicitly disabled or interleave count is set to 1";
1795 InterleaveLoop = false;
1798 // Override IC if user provided an interleave count.
1799 IC = UserIC > 0 ? UserIC : IC;
1801 // Emit diagnostic messages, if any.
1802 if (!VectorizeLoop && !InterleaveLoop) {
1803 // Do not vectorize or interleaving the loop.
1804 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1805 L->getStartLoc(), VecDiagMsg);
1806 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1807 L->getStartLoc(), IntDiagMsg);
1809 } else if (!VectorizeLoop && InterleaveLoop) {
1810 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1811 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1812 L->getStartLoc(), VecDiagMsg);
1813 } else if (VectorizeLoop && !InterleaveLoop) {
1814 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1815 << DebugLocStr << '\n');
1816 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1817 L->getStartLoc(), IntDiagMsg);
1818 } else if (VectorizeLoop && InterleaveLoop) {
1819 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1820 << DebugLocStr << '\n');
1821 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1824 if (!VectorizeLoop) {
1825 assert(IC > 1 && "interleave count should not be 1 or 0");
1826 // If we decided that it is not legal to vectorize the loop then
1828 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1829 Unroller.vectorize(&LVL);
1831 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1832 Twine("interleaved loop (interleaved count: ") +
1835 // If we decided that it is *legal* to vectorize the loop then do it.
1836 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1840 // Add metadata to disable runtime unrolling scalar loop when there's no
1841 // runtime check about strides and memory. Because at this situation,
1842 // scalar loop is rarely used not worthy to be unrolled.
1843 if (!LB.IsSafetyChecksAdded())
1844 AddRuntimeUnrollDisableMetaData(L);
1846 // Report the vectorization decision.
1847 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1848 Twine("vectorized loop (vectorization width: ") +
1849 Twine(VF.Width) + ", interleaved count: " +
1853 // Mark the loop as already vectorized to avoid vectorizing again.
1854 Hints.setAlreadyVectorized();
1856 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1860 void getAnalysisUsage(AnalysisUsage &AU) const override {
1861 AU.addRequired<AssumptionCacheTracker>();
1862 AU.addRequiredID(LoopSimplifyID);
1863 AU.addRequiredID(LCSSAID);
1864 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1865 AU.addRequired<DominatorTreeWrapperPass>();
1866 AU.addRequired<LoopInfoWrapperPass>();
1867 AU.addRequired<ScalarEvolutionWrapperPass>();
1868 AU.addRequired<TargetTransformInfoWrapperPass>();
1869 AU.addRequired<AliasAnalysis>();
1870 AU.addRequired<LoopAccessAnalysis>();
1871 AU.addPreserved<LoopInfoWrapperPass>();
1872 AU.addPreserved<DominatorTreeWrapperPass>();
1873 AU.addPreserved<AliasAnalysis>();
1878 } // end anonymous namespace
1880 //===----------------------------------------------------------------------===//
1881 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1882 // LoopVectorizationCostModel.
1883 //===----------------------------------------------------------------------===//
1885 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1886 // We need to place the broadcast of invariant variables outside the loop.
1887 Instruction *Instr = dyn_cast<Instruction>(V);
1889 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1890 Instr->getParent()) != LoopVectorBody.end());
1891 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1893 // Place the code for broadcasting invariant variables in the new preheader.
1894 IRBuilder<>::InsertPointGuard Guard(Builder);
1896 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1898 // Broadcast the scalar into all locations in the vector.
1899 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1904 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1906 assert(Val->getType()->isVectorTy() && "Must be a vector");
1907 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1908 "Elem must be an integer");
1909 assert(Step->getType() == Val->getType()->getScalarType() &&
1910 "Step has wrong type");
1911 // Create the types.
1912 Type *ITy = Val->getType()->getScalarType();
1913 VectorType *Ty = cast<VectorType>(Val->getType());
1914 int VLen = Ty->getNumElements();
1915 SmallVector<Constant*, 8> Indices;
1917 // Create a vector of consecutive numbers from zero to VF.
1918 for (int i = 0; i < VLen; ++i)
1919 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1921 // Add the consecutive indices to the vector value.
1922 Constant *Cv = ConstantVector::get(Indices);
1923 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1924 Step = Builder.CreateVectorSplat(VLen, Step);
1925 assert(Step->getType() == Val->getType() && "Invalid step vec");
1926 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1927 // which can be found from the original scalar operations.
1928 Step = Builder.CreateMul(Cv, Step);
1929 return Builder.CreateAdd(Val, Step, "induction");
1932 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1933 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1934 // Make sure that the pointer does not point to structs.
1935 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1938 // If this value is a pointer induction variable we know it is consecutive.
1939 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1940 if (Phi && Inductions.count(Phi)) {
1941 InductionInfo II = Inductions[Phi];
1942 return II.getConsecutiveDirection();
1945 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1949 unsigned NumOperands = Gep->getNumOperands();
1950 Value *GpPtr = Gep->getPointerOperand();
1951 // If this GEP value is a consecutive pointer induction variable and all of
1952 // the indices are constant then we know it is consecutive. We can
1953 Phi = dyn_cast<PHINode>(GpPtr);
1954 if (Phi && Inductions.count(Phi)) {
1956 // Make sure that the pointer does not point to structs.
1957 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1958 if (GepPtrType->getElementType()->isAggregateType())
1961 // Make sure that all of the index operands are loop invariant.
1962 for (unsigned i = 1; i < NumOperands; ++i)
1963 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1966 InductionInfo II = Inductions[Phi];
1967 return II.getConsecutiveDirection();
1970 unsigned InductionOperand = getGEPInductionOperand(Gep);
1972 // Check that all of the gep indices are uniform except for our induction
1974 for (unsigned i = 0; i != NumOperands; ++i)
1975 if (i != InductionOperand &&
1976 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1979 // We can emit wide load/stores only if the last non-zero index is the
1980 // induction variable.
1981 const SCEV *Last = nullptr;
1982 if (!Strides.count(Gep))
1983 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1985 // Because of the multiplication by a stride we can have a s/zext cast.
1986 // We are going to replace this stride by 1 so the cast is safe to ignore.
1988 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1989 // %0 = trunc i64 %indvars.iv to i32
1990 // %mul = mul i32 %0, %Stride1
1991 // %idxprom = zext i32 %mul to i64 << Safe cast.
1992 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1994 Last = replaceSymbolicStrideSCEV(SE, Strides,
1995 Gep->getOperand(InductionOperand), Gep);
1996 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1998 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2002 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2003 const SCEV *Step = AR->getStepRecurrence(*SE);
2005 // The memory is consecutive because the last index is consecutive
2006 // and all other indices are loop invariant.
2009 if (Step->isAllOnesValue())
2016 bool LoopVectorizationLegality::isUniform(Value *V) {
2017 return LAI->isUniform(V);
2020 InnerLoopVectorizer::VectorParts&
2021 InnerLoopVectorizer::getVectorValue(Value *V) {
2022 assert(V != Induction && "The new induction variable should not be used.");
2023 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2025 // If we have a stride that is replaced by one, do it here.
2026 if (Legal->hasStride(V))
2027 V = ConstantInt::get(V->getType(), 1);
2029 // If we have this scalar in the map, return it.
2030 if (WidenMap.has(V))
2031 return WidenMap.get(V);
2033 // If this scalar is unknown, assume that it is a constant or that it is
2034 // loop invariant. Broadcast V and save the value for future uses.
2035 Value *B = getBroadcastInstrs(V);
2036 return WidenMap.splat(V, B);
2039 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2040 assert(Vec->getType()->isVectorTy() && "Invalid type");
2041 SmallVector<Constant*, 8> ShuffleMask;
2042 for (unsigned i = 0; i < VF; ++i)
2043 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2045 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2046 ConstantVector::get(ShuffleMask),
2050 // Get a mask to interleave \p NumVec vectors into a wide vector.
2051 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2052 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2053 // <0, 4, 1, 5, 2, 6, 3, 7>
2054 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2056 SmallVector<Constant *, 16> Mask;
2057 for (unsigned i = 0; i < VF; i++)
2058 for (unsigned j = 0; j < NumVec; j++)
2059 Mask.push_back(Builder.getInt32(j * VF + i));
2061 return ConstantVector::get(Mask);
2064 // Get the strided mask starting from index \p Start.
2065 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2066 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2067 unsigned Stride, unsigned VF) {
2068 SmallVector<Constant *, 16> Mask;
2069 for (unsigned i = 0; i < VF; i++)
2070 Mask.push_back(Builder.getInt32(Start + i * Stride));
2072 return ConstantVector::get(Mask);
2075 // Get a mask of two parts: The first part consists of sequential integers
2076 // starting from 0, The second part consists of UNDEFs.
2077 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2078 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2079 unsigned NumUndef) {
2080 SmallVector<Constant *, 16> Mask;
2081 for (unsigned i = 0; i < NumInt; i++)
2082 Mask.push_back(Builder.getInt32(i));
2084 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2085 for (unsigned i = 0; i < NumUndef; i++)
2086 Mask.push_back(Undef);
2088 return ConstantVector::get(Mask);
2091 // Concatenate two vectors with the same element type. The 2nd vector should
2092 // not have more elements than the 1st vector. If the 2nd vector has less
2093 // elements, extend it with UNDEFs.
2094 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2096 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2097 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2098 assert(VecTy1 && VecTy2 &&
2099 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2100 "Expect two vectors with the same element type");
2102 unsigned NumElts1 = VecTy1->getNumElements();
2103 unsigned NumElts2 = VecTy2->getNumElements();
2104 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2106 if (NumElts1 > NumElts2) {
2107 // Extend with UNDEFs.
2109 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2110 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2113 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2114 return Builder.CreateShuffleVector(V1, V2, Mask);
2117 // Concatenate vectors in the given list. All vectors have the same type.
2118 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2119 ArrayRef<Value *> InputList) {
2120 unsigned NumVec = InputList.size();
2121 assert(NumVec > 1 && "Should be at least two vectors");
2123 SmallVector<Value *, 8> ResList;
2124 ResList.append(InputList.begin(), InputList.end());
2126 SmallVector<Value *, 8> TmpList;
2127 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2128 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2129 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2130 "Only the last vector may have a different type");
2132 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2135 // Push the last vector if the total number of vectors is odd.
2136 if (NumVec % 2 != 0)
2137 TmpList.push_back(ResList[NumVec - 1]);
2140 NumVec = ResList.size();
2141 } while (NumVec > 1);
2146 // Try to vectorize the interleave group that \p Instr belongs to.
2148 // E.g. Translate following interleaved load group (factor = 3):
2149 // for (i = 0; i < N; i+=3) {
2150 // R = Pic[i]; // Member of index 0
2151 // G = Pic[i+1]; // Member of index 1
2152 // B = Pic[i+2]; // Member of index 2
2153 // ... // do something to R, G, B
2156 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2157 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2158 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2159 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2161 // Or translate following interleaved store group (factor = 3):
2162 // for (i = 0; i < N; i+=3) {
2163 // ... do something to R, G, B
2164 // Pic[i] = R; // Member of index 0
2165 // Pic[i+1] = G; // Member of index 1
2166 // Pic[i+2] = B; // Member of index 2
2169 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2170 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2171 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2172 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2173 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2174 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2175 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2176 assert(Group && "Fail to get an interleaved access group.");
2178 // Skip if current instruction is not the insert position.
2179 if (Instr != Group->getInsertPos())
2182 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2183 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2184 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2186 // Prepare for the vector type of the interleaved load/store.
2187 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2188 unsigned InterleaveFactor = Group->getFactor();
2189 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2190 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2192 // Prepare for the new pointers.
2193 setDebugLocFromInst(Builder, Ptr);
2194 VectorParts &PtrParts = getVectorValue(Ptr);
2195 SmallVector<Value *, 2> NewPtrs;
2196 unsigned Index = Group->getIndex(Instr);
2197 for (unsigned Part = 0; Part < UF; Part++) {
2198 // Extract the pointer for current instruction from the pointer vector. A
2199 // reverse access uses the pointer in the last lane.
2200 Value *NewPtr = Builder.CreateExtractElement(
2202 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2204 // Notice current instruction could be any index. Need to adjust the address
2205 // to the member of index 0.
2207 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2208 // b = A[i]; // Member of index 0
2209 // Current pointer is pointed to A[i+1], adjust it to A[i].
2211 // E.g. A[i+1] = a; // Member of index 1
2212 // A[i] = b; // Member of index 0
2213 // A[i+2] = c; // Member of index 2 (Current instruction)
2214 // Current pointer is pointed to A[i+2], adjust it to A[i].
2215 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2217 // Cast to the vector pointer type.
2218 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2221 setDebugLocFromInst(Builder, Instr);
2222 Value *UndefVec = UndefValue::get(VecTy);
2224 // Vectorize the interleaved load group.
2226 for (unsigned Part = 0; Part < UF; Part++) {
2227 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2228 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2230 for (unsigned i = 0; i < InterleaveFactor; i++) {
2231 Instruction *Member = Group->getMember(i);
2233 // Skip the gaps in the group.
2237 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2238 Value *StridedVec = Builder.CreateShuffleVector(
2239 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2241 // If this member has different type, cast the result type.
2242 if (Member->getType() != ScalarTy) {
2243 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2244 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2247 VectorParts &Entry = WidenMap.get(Member);
2249 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2252 propagateMetadata(NewLoadInstr, Instr);
2257 // The sub vector type for current instruction.
2258 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2260 // Vectorize the interleaved store group.
2261 for (unsigned Part = 0; Part < UF; Part++) {
2262 // Collect the stored vector from each member.
2263 SmallVector<Value *, 4> StoredVecs;
2264 for (unsigned i = 0; i < InterleaveFactor; i++) {
2265 // Interleaved store group doesn't allow a gap, so each index has a member
2266 Instruction *Member = Group->getMember(i);
2267 assert(Member && "Fail to get a member from an interleaved store group");
2270 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2271 if (Group->isReverse())
2272 StoredVec = reverseVector(StoredVec);
2274 // If this member has different type, cast it to an unified type.
2275 if (StoredVec->getType() != SubVT)
2276 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2278 StoredVecs.push_back(StoredVec);
2281 // Concatenate all vectors into a wide vector.
2282 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2284 // Interleave the elements in the wide vector.
2285 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2286 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2289 Instruction *NewStoreInstr =
2290 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2291 propagateMetadata(NewStoreInstr, Instr);
2295 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2296 // Attempt to issue a wide load.
2297 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2298 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2300 assert((LI || SI) && "Invalid Load/Store instruction");
2302 // Try to vectorize the interleave group if this access is interleaved.
2303 if (Legal->isAccessInterleaved(Instr))
2304 return vectorizeInterleaveGroup(Instr);
2306 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2307 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2308 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2309 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2310 // An alignment of 0 means target abi alignment. We need to use the scalar's
2311 // target abi alignment in such a case.
2312 const DataLayout &DL = Instr->getModule()->getDataLayout();
2314 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2315 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2316 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2317 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2319 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2320 !Legal->isMaskRequired(SI))
2321 return scalarizeInstruction(Instr, true);
2323 if (ScalarAllocatedSize != VectorElementSize)
2324 return scalarizeInstruction(Instr);
2326 // If the pointer is loop invariant or if it is non-consecutive,
2327 // scalarize the load.
2328 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2329 bool Reverse = ConsecutiveStride < 0;
2330 bool UniformLoad = LI && Legal->isUniform(Ptr);
2331 if (!ConsecutiveStride || UniformLoad)
2332 return scalarizeInstruction(Instr);
2334 Constant *Zero = Builder.getInt32(0);
2335 VectorParts &Entry = WidenMap.get(Instr);
2337 // Handle consecutive loads/stores.
2338 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2339 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2340 setDebugLocFromInst(Builder, Gep);
2341 Value *PtrOperand = Gep->getPointerOperand();
2342 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2343 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2345 // Create the new GEP with the new induction variable.
2346 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2347 Gep2->setOperand(0, FirstBasePtr);
2348 Gep2->setName("gep.indvar.base");
2349 Ptr = Builder.Insert(Gep2);
2351 setDebugLocFromInst(Builder, Gep);
2352 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2353 OrigLoop) && "Base ptr must be invariant");
2355 // The last index does not have to be the induction. It can be
2356 // consecutive and be a function of the index. For example A[I+1];
2357 unsigned NumOperands = Gep->getNumOperands();
2358 unsigned InductionOperand = getGEPInductionOperand(Gep);
2359 // Create the new GEP with the new induction variable.
2360 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2362 for (unsigned i = 0; i < NumOperands; ++i) {
2363 Value *GepOperand = Gep->getOperand(i);
2364 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2366 // Update last index or loop invariant instruction anchored in loop.
2367 if (i == InductionOperand ||
2368 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2369 assert((i == InductionOperand ||
2370 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2371 "Must be last index or loop invariant");
2373 VectorParts &GEPParts = getVectorValue(GepOperand);
2374 Value *Index = GEPParts[0];
2375 Index = Builder.CreateExtractElement(Index, Zero);
2376 Gep2->setOperand(i, Index);
2377 Gep2->setName("gep.indvar.idx");
2380 Ptr = Builder.Insert(Gep2);
2382 // Use the induction element ptr.
2383 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2384 setDebugLocFromInst(Builder, Ptr);
2385 VectorParts &PtrVal = getVectorValue(Ptr);
2386 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2389 VectorParts Mask = createBlockInMask(Instr->getParent());
2392 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2393 "We do not allow storing to uniform addresses");
2394 setDebugLocFromInst(Builder, SI);
2395 // We don't want to update the value in the map as it might be used in
2396 // another expression. So don't use a reference type for "StoredVal".
2397 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2399 for (unsigned Part = 0; Part < UF; ++Part) {
2400 // Calculate the pointer for the specific unroll-part.
2402 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2405 // If we store to reverse consecutive memory locations, then we need
2406 // to reverse the order of elements in the stored value.
2407 StoredVal[Part] = reverseVector(StoredVal[Part]);
2408 // If the address is consecutive but reversed, then the
2409 // wide store needs to start at the last vector element.
2410 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2411 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2412 Mask[Part] = reverseVector(Mask[Part]);
2415 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2416 DataTy->getPointerTo(AddressSpace));
2419 if (Legal->isMaskRequired(SI))
2420 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2423 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2424 propagateMetadata(NewSI, SI);
2430 assert(LI && "Must have a load instruction");
2431 setDebugLocFromInst(Builder, LI);
2432 for (unsigned Part = 0; Part < UF; ++Part) {
2433 // Calculate the pointer for the specific unroll-part.
2435 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2438 // If the address is consecutive but reversed, then the
2439 // wide load needs to start at the last vector element.
2440 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2441 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2442 Mask[Part] = reverseVector(Mask[Part]);
2446 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2447 DataTy->getPointerTo(AddressSpace));
2448 if (Legal->isMaskRequired(LI))
2449 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2450 UndefValue::get(DataTy),
2451 "wide.masked.load");
2453 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2454 propagateMetadata(NewLI, LI);
2455 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2459 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2460 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2461 // Holds vector parameters or scalars, in case of uniform vals.
2462 SmallVector<VectorParts, 4> Params;
2464 setDebugLocFromInst(Builder, Instr);
2466 // Find all of the vectorized parameters.
2467 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2468 Value *SrcOp = Instr->getOperand(op);
2470 // If we are accessing the old induction variable, use the new one.
2471 if (SrcOp == OldInduction) {
2472 Params.push_back(getVectorValue(SrcOp));
2476 // Try using previously calculated values.
2477 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2479 // If the src is an instruction that appeared earlier in the basic block,
2480 // then it should already be vectorized.
2481 if (SrcInst && OrigLoop->contains(SrcInst)) {
2482 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2483 // The parameter is a vector value from earlier.
2484 Params.push_back(WidenMap.get(SrcInst));
2486 // The parameter is a scalar from outside the loop. Maybe even a constant.
2487 VectorParts Scalars;
2488 Scalars.append(UF, SrcOp);
2489 Params.push_back(Scalars);
2493 assert(Params.size() == Instr->getNumOperands() &&
2494 "Invalid number of operands");
2496 // Does this instruction return a value ?
2497 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2499 Value *UndefVec = IsVoidRetTy ? nullptr :
2500 UndefValue::get(VectorType::get(Instr->getType(), VF));
2501 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2502 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2504 Instruction *InsertPt = Builder.GetInsertPoint();
2505 BasicBlock *IfBlock = Builder.GetInsertBlock();
2506 BasicBlock *CondBlock = nullptr;
2509 Loop *VectorLp = nullptr;
2510 if (IfPredicateStore) {
2511 assert(Instr->getParent()->getSinglePredecessor() &&
2512 "Only support single predecessor blocks");
2513 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2514 Instr->getParent());
2515 VectorLp = LI->getLoopFor(IfBlock);
2516 assert(VectorLp && "Must have a loop for this block");
2519 // For each vector unroll 'part':
2520 for (unsigned Part = 0; Part < UF; ++Part) {
2521 // For each scalar that we create:
2522 for (unsigned Width = 0; Width < VF; ++Width) {
2525 Value *Cmp = nullptr;
2526 if (IfPredicateStore) {
2527 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2528 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2529 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2530 LoopVectorBody.push_back(CondBlock);
2531 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2532 // Update Builder with newly created basic block.
2533 Builder.SetInsertPoint(InsertPt);
2536 Instruction *Cloned = Instr->clone();
2538 Cloned->setName(Instr->getName() + ".cloned");
2539 // Replace the operands of the cloned instructions with extracted scalars.
2540 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2541 Value *Op = Params[op][Part];
2542 // Param is a vector. Need to extract the right lane.
2543 if (Op->getType()->isVectorTy())
2544 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2545 Cloned->setOperand(op, Op);
2548 // Place the cloned scalar in the new loop.
2549 Builder.Insert(Cloned);
2551 // If the original scalar returns a value we need to place it in a vector
2552 // so that future users will be able to use it.
2554 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2555 Builder.getInt32(Width));
2557 if (IfPredicateStore) {
2558 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2559 LoopVectorBody.push_back(NewIfBlock);
2560 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2561 Builder.SetInsertPoint(InsertPt);
2562 ReplaceInstWithInst(IfBlock->getTerminator(),
2563 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2564 IfBlock = NewIfBlock;
2570 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2574 if (Instruction *I = dyn_cast<Instruction>(V))
2575 return I->getParent() == Loc->getParent() ? I : nullptr;
2579 std::pair<Instruction *, Instruction *>
2580 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2581 Instruction *tnullptr = nullptr;
2582 if (!Legal->mustCheckStrides())
2583 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2585 IRBuilder<> ChkBuilder(Loc);
2588 Value *Check = nullptr;
2589 Instruction *FirstInst = nullptr;
2590 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2591 SE = Legal->strides_end();
2593 Value *Ptr = stripIntegerCast(*SI);
2594 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2596 // Store the first instruction we create.
2597 FirstInst = getFirstInst(FirstInst, C, Loc);
2599 Check = ChkBuilder.CreateOr(Check, C);
2604 // We have to do this trickery because the IRBuilder might fold the check to a
2605 // constant expression in which case there is no Instruction anchored in a
2607 LLVMContext &Ctx = Loc->getContext();
2608 Instruction *TheCheck =
2609 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2610 ChkBuilder.Insert(TheCheck, "stride.not.one");
2611 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2613 return std::make_pair(FirstInst, TheCheck);
2616 void InnerLoopVectorizer::createEmptyLoop() {
2618 In this function we generate a new loop. The new loop will contain
2619 the vectorized instructions while the old loop will continue to run the
2622 [ ] <-- loop iteration number check.
2625 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2628 || [ ] <-- vector pre header.
2632 || [ ]_| <-- vector loop.
2635 | >[ ] <--- middle-block.
2638 -|- >[ ] <--- new preheader.
2642 | [ ]_| <-- old scalar loop to handle remainder.
2645 >[ ] <-- exit block.
2649 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2650 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2651 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2652 assert(VectorPH && "Invalid loop structure");
2653 assert(ExitBlock && "Must have an exit block");
2655 // Some loops have a single integer induction variable, while other loops
2656 // don't. One example is c++ iterators that often have multiple pointer
2657 // induction variables. In the code below we also support a case where we
2658 // don't have a single induction variable.
2659 OldInduction = Legal->getInduction();
2660 Type *IdxTy = Legal->getWidestInductionType();
2662 // Find the loop boundaries.
2663 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2664 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2666 // The exit count might have the type of i64 while the phi is i32. This can
2667 // happen if we have an induction variable that is sign extended before the
2668 // compare. The only way that we get a backedge taken count is that the
2669 // induction variable was signed and as such will not overflow. In such a case
2670 // truncation is legal.
2671 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2672 IdxTy->getPrimitiveSizeInBits())
2673 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2675 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2676 // Get the total trip count from the count by adding 1.
2677 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2678 SE->getConstant(BackedgeTakeCount->getType(), 1));
2680 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2682 // Expand the trip count and place the new instructions in the preheader.
2683 // Notice that the pre-header does not change, only the loop body.
2684 SCEVExpander Exp(*SE, DL, "induction");
2686 // The loop minimum iterations check below is to ensure the loop has enough
2687 // trip count so the generated vector loop will likely be executed and the
2688 // preparation and rounding-off costs will likely be worthy.
2690 // The minimum iteration check also covers case where the backedge-taken
2691 // count is uint##_max. Adding one to it will cause overflow and an
2692 // incorrect loop trip count being generated in the vector body. In this
2693 // case we also want to directly jump to the scalar remainder loop.
2694 Value *ExitCountValue = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2695 VectorPH->getTerminator());
2696 if (ExitCountValue->getType()->isPointerTy())
2697 ExitCountValue = CastInst::CreatePointerCast(ExitCountValue, IdxTy,
2698 "exitcount.ptrcnt.to.int",
2699 VectorPH->getTerminator());
2701 Instruction *CheckMinIters =
2702 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULT, ExitCountValue,
2703 ConstantInt::get(ExitCountValue->getType(), VF * UF),
2704 "min.iters.check", VectorPH->getTerminator());
2706 // The loop index does not have to start at Zero. Find the original start
2707 // value from the induction PHI node. If we don't have an induction variable
2708 // then we know that it starts at zero.
2709 Builder.SetInsertPoint(VectorPH->getTerminator());
2710 Value *StartIdx = ExtendedIdx =
2712 ? Builder.CreateZExt(OldInduction->getIncomingValueForBlock(VectorPH),
2714 : ConstantInt::get(IdxTy, 0);
2716 // Count holds the overall loop count (N).
2717 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2718 VectorPH->getTerminator());
2720 LoopBypassBlocks.push_back(VectorPH);
2722 // Split the single block loop into the two loop structure described above.
2723 BasicBlock *VecBody =
2724 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2725 BasicBlock *MiddleBlock =
2726 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2727 BasicBlock *ScalarPH =
2728 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2730 // Create and register the new vector loop.
2731 Loop* Lp = new Loop();
2732 Loop *ParentLoop = OrigLoop->getParentLoop();
2734 // Insert the new loop into the loop nest and register the new basic blocks
2735 // before calling any utilities such as SCEV that require valid LoopInfo.
2737 ParentLoop->addChildLoop(Lp);
2738 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2739 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2741 LI->addTopLevelLoop(Lp);
2743 Lp->addBasicBlockToLoop(VecBody, *LI);
2745 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2747 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2749 // Generate the induction variable.
2750 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2751 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2752 // The loop step is equal to the vectorization factor (num of SIMD elements)
2753 // times the unroll factor (num of SIMD instructions).
2754 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2756 // Generate code to check that the loop's trip count is not less than the
2757 // minimum loop iteration number threshold.
2758 BasicBlock *NewVectorPH =
2759 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "min.iters.checked");
2761 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2762 ReplaceInstWithInst(VectorPH->getTerminator(),
2763 BranchInst::Create(ScalarPH, NewVectorPH, CheckMinIters));
2764 VectorPH = NewVectorPH;
2766 // This is the IR builder that we use to add all of the logic for bypassing
2767 // the new vector loop.
2768 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2769 setDebugLocFromInst(BypassBuilder,
2770 getDebugLocFromInstOrOperands(OldInduction));
2772 // We may need to extend the index in case there is a type mismatch.
2773 // We know that the count starts at zero and does not overflow.
2774 if (Count->getType() != IdxTy) {
2775 // The exit count can be of pointer type. Convert it to the correct
2777 if (ExitCount->getType()->isPointerTy())
2778 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2780 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2783 // Add the start index to the loop count to get the new end index.
2784 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2786 // Now we need to generate the expression for N - (N % VF), which is
2787 // the part that the vectorized body will execute.
2788 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2789 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2790 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2791 "end.idx.rnd.down");
2793 // Now, compare the new count to zero. If it is zero skip the vector loop and
2794 // jump to the scalar loop.
2796 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2798 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2800 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2801 LoopBypassBlocks.push_back(VectorPH);
2802 ReplaceInstWithInst(VectorPH->getTerminator(),
2803 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2804 VectorPH = NewVectorPH;
2806 // Generate the code to check that the strides we assumed to be one are really
2807 // one. We want the new basic block to start at the first instruction in a
2808 // sequence of instructions that form a check.
2809 Instruction *StrideCheck;
2810 Instruction *FirstCheckInst;
2811 std::tie(FirstCheckInst, StrideCheck) =
2812 addStrideCheck(VectorPH->getTerminator());
2814 AddedSafetyChecks = true;
2815 // Create a new block containing the stride check.
2816 VectorPH->setName("vector.stridecheck");
2818 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2820 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2821 LoopBypassBlocks.push_back(VectorPH);
2823 // Replace the branch into the memory check block with a conditional branch
2824 // for the "few elements case".
2825 ReplaceInstWithInst(
2826 VectorPH->getTerminator(),
2827 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2829 VectorPH = NewVectorPH;
2832 // Generate the code that checks in runtime if arrays overlap. We put the
2833 // checks into a separate block to make the more common case of few elements
2835 Instruction *MemRuntimeCheck;
2836 std::tie(FirstCheckInst, MemRuntimeCheck) =
2837 Legal->getLAI()->addRuntimeChecks(VectorPH->getTerminator());
2838 if (MemRuntimeCheck) {
2839 AddedSafetyChecks = true;
2840 // Create a new block containing the memory check.
2841 VectorPH->setName("vector.memcheck");
2843 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2845 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2846 LoopBypassBlocks.push_back(VectorPH);
2848 // Replace the branch into the memory check block with a conditional branch
2849 // for the "few elements case".
2850 ReplaceInstWithInst(
2851 VectorPH->getTerminator(),
2852 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2854 VectorPH = NewVectorPH;
2857 // We are going to resume the execution of the scalar loop.
2858 // Go over all of the induction variables that we found and fix the
2859 // PHIs that are left in the scalar version of the loop.
2860 // The starting values of PHI nodes depend on the counter of the last
2861 // iteration in the vectorized loop.
2862 // If we come from a bypass edge then we need to start from the original
2865 // This variable saves the new starting index for the scalar loop.
2866 PHINode *ResumeIndex = nullptr;
2867 LoopVectorizationLegality::InductionList::iterator I, E;
2868 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2869 // Set builder to point to last bypass block.
2870 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2871 for (I = List->begin(), E = List->end(); I != E; ++I) {
2872 PHINode *OrigPhi = I->first;
2873 LoopVectorizationLegality::InductionInfo II = I->second;
2875 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2876 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2877 MiddleBlock->getTerminator());
2878 // We might have extended the type of the induction variable but we need a
2879 // truncated version for the scalar loop.
2880 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2881 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2882 MiddleBlock->getTerminator()) : nullptr;
2884 // Create phi nodes to merge from the backedge-taken check block.
2885 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2886 ScalarPH->getTerminator());
2887 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2889 PHINode *BCTruncResumeVal = nullptr;
2890 if (OrigPhi == OldInduction) {
2892 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2893 ScalarPH->getTerminator());
2894 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2897 Value *EndValue = nullptr;
2899 case LoopVectorizationLegality::IK_NoInduction:
2900 llvm_unreachable("Unknown induction");
2901 case LoopVectorizationLegality::IK_IntInduction: {
2902 // Handle the integer induction counter.
2903 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2905 // We have the canonical induction variable.
2906 if (OrigPhi == OldInduction) {
2907 // Create a truncated version of the resume value for the scalar loop,
2908 // we might have promoted the type to a larger width.
2910 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2911 // The new PHI merges the original incoming value, in case of a bypass,
2912 // or the value at the end of the vectorized loop.
2913 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2914 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2915 TruncResumeVal->addIncoming(EndValue, VecBody);
2917 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2919 // We know what the end value is.
2920 EndValue = IdxEndRoundDown;
2921 // We also know which PHI node holds it.
2922 ResumeIndex = ResumeVal;
2926 // Not the canonical induction variable - add the vector loop count to the
2928 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2929 II.StartValue->getType(),
2931 EndValue = II.transform(BypassBuilder, CRD);
2932 EndValue->setName("ind.end");
2935 case LoopVectorizationLegality::IK_PtrInduction: {
2936 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2937 II.StepValue->getType(),
2939 EndValue = II.transform(BypassBuilder, CRD);
2940 EndValue->setName("ptr.ind.end");
2945 // The new PHI merges the original incoming value, in case of a bypass,
2946 // or the value at the end of the vectorized loop.
2947 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2948 if (OrigPhi == OldInduction)
2949 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2951 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2953 ResumeVal->addIncoming(EndValue, VecBody);
2955 // Fix the scalar body counter (PHI node).
2956 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2958 // The old induction's phi node in the scalar body needs the truncated
2960 if (OrigPhi == OldInduction) {
2961 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2962 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2964 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2965 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2969 // If we are generating a new induction variable then we also need to
2970 // generate the code that calculates the exit value. This value is not
2971 // simply the end of the counter because we may skip the vectorized body
2972 // in case of a runtime check.
2974 assert(!ResumeIndex && "Unexpected resume value found");
2975 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2976 MiddleBlock->getTerminator());
2977 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2978 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2979 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2982 // Make sure that we found the index where scalar loop needs to continue.
2983 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2984 "Invalid resume Index");
2986 // Add a check in the middle block to see if we have completed
2987 // all of the iterations in the first vector loop.
2988 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2989 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2990 ResumeIndex, "cmp.n",
2991 MiddleBlock->getTerminator());
2992 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2993 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2995 // Create i+1 and fill the PHINode.
2996 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2997 Induction->addIncoming(StartIdx, VectorPH);
2998 Induction->addIncoming(NextIdx, VecBody);
2999 // Create the compare.
3000 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
3001 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
3003 // Now we have two terminators. Remove the old one from the block.
3004 VecBody->getTerminator()->eraseFromParent();
3006 // Get ready to start creating new instructions into the vectorized body.
3007 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
3010 LoopVectorPreHeader = VectorPH;
3011 LoopScalarPreHeader = ScalarPH;
3012 LoopMiddleBlock = MiddleBlock;
3013 LoopExitBlock = ExitBlock;
3014 LoopVectorBody.push_back(VecBody);
3015 LoopScalarBody = OldBasicBlock;
3017 LoopVectorizeHints Hints(Lp, true);
3018 Hints.setAlreadyVectorized();
3022 struct CSEDenseMapInfo {
3023 static bool canHandle(Instruction *I) {
3024 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3025 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3027 static inline Instruction *getEmptyKey() {
3028 return DenseMapInfo<Instruction *>::getEmptyKey();
3030 static inline Instruction *getTombstoneKey() {
3031 return DenseMapInfo<Instruction *>::getTombstoneKey();
3033 static unsigned getHashValue(Instruction *I) {
3034 assert(canHandle(I) && "Unknown instruction!");
3035 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3036 I->value_op_end()));
3038 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3039 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3040 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3042 return LHS->isIdenticalTo(RHS);
3047 /// \brief Check whether this block is a predicated block.
3048 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
3049 /// = ...; " blocks. We start with one vectorized basic block. For every
3050 /// conditional block we split this vectorized block. Therefore, every second
3051 /// block will be a predicated one.
3052 static bool isPredicatedBlock(unsigned BlockNum) {
3053 return BlockNum % 2;
3056 ///\brief Perform cse of induction variable instructions.
3057 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
3058 // Perform simple cse.
3059 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3060 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3061 BasicBlock *BB = BBs[i];
3062 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3063 Instruction *In = I++;
3065 if (!CSEDenseMapInfo::canHandle(In))
3068 // Check if we can replace this instruction with any of the
3069 // visited instructions.
3070 if (Instruction *V = CSEMap.lookup(In)) {
3071 In->replaceAllUsesWith(V);
3072 In->eraseFromParent();
3075 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3076 // ...;" blocks for predicated stores. Every second block is a predicated
3078 if (isPredicatedBlock(i))
3086 /// \brief Adds a 'fast' flag to floating point operations.
3087 static Value *addFastMathFlag(Value *V) {
3088 if (isa<FPMathOperator>(V)){
3089 FastMathFlags Flags;
3090 Flags.setUnsafeAlgebra();
3091 cast<Instruction>(V)->setFastMathFlags(Flags);
3096 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3097 /// the result needs to be inserted and/or extracted from vectors.
3098 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3099 const TargetTransformInfo &TTI) {
3103 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3106 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3108 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3110 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3116 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3117 // Return the cost of the instruction, including scalarization overhead if it's
3118 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3119 // i.e. either vector version isn't available, or is too expensive.
3120 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3121 const TargetTransformInfo &TTI,
3122 const TargetLibraryInfo *TLI,
3123 bool &NeedToScalarize) {
3124 Function *F = CI->getCalledFunction();
3125 StringRef FnName = CI->getCalledFunction()->getName();
3126 Type *ScalarRetTy = CI->getType();
3127 SmallVector<Type *, 4> Tys, ScalarTys;
3128 for (auto &ArgOp : CI->arg_operands())
3129 ScalarTys.push_back(ArgOp->getType());
3131 // Estimate cost of scalarized vector call. The source operands are assumed
3132 // to be vectors, so we need to extract individual elements from there,
3133 // execute VF scalar calls, and then gather the result into the vector return
3135 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3137 return ScalarCallCost;
3139 // Compute corresponding vector type for return value and arguments.
3140 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3141 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3142 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3144 // Compute costs of unpacking argument values for the scalar calls and
3145 // packing the return values to a vector.
3146 unsigned ScalarizationCost =
3147 getScalarizationOverhead(RetTy, true, false, TTI);
3148 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3149 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3151 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3153 // If we can't emit a vector call for this function, then the currently found
3154 // cost is the cost we need to return.
3155 NeedToScalarize = true;
3156 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3159 // If the corresponding vector cost is cheaper, return its cost.
3160 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3161 if (VectorCallCost < Cost) {
3162 NeedToScalarize = false;
3163 return VectorCallCost;
3168 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3169 // factor VF. Return the cost of the instruction, including scalarization
3170 // overhead if it's needed.
3171 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3172 const TargetTransformInfo &TTI,
3173 const TargetLibraryInfo *TLI) {
3174 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3175 assert(ID && "Expected intrinsic call!");
3177 Type *RetTy = ToVectorTy(CI->getType(), VF);
3178 SmallVector<Type *, 4> Tys;
3179 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3180 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3182 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3185 void InnerLoopVectorizer::vectorizeLoop() {
3186 //===------------------------------------------------===//
3188 // Notice: any optimization or new instruction that go
3189 // into the code below should be also be implemented in
3192 //===------------------------------------------------===//
3193 Constant *Zero = Builder.getInt32(0);
3195 // In order to support reduction variables we need to be able to vectorize
3196 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3197 // stages. First, we create a new vector PHI node with no incoming edges.
3198 // We use this value when we vectorize all of the instructions that use the
3199 // PHI. Next, after all of the instructions in the block are complete we
3200 // add the new incoming edges to the PHI. At this point all of the
3201 // instructions in the basic block are vectorized, so we can use them to
3202 // construct the PHI.
3203 PhiVector RdxPHIsToFix;
3205 // Scan the loop in a topological order to ensure that defs are vectorized
3207 LoopBlocksDFS DFS(OrigLoop);
3210 // Vectorize all of the blocks in the original loop.
3211 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3212 be = DFS.endRPO(); bb != be; ++bb)
3213 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3215 // At this point every instruction in the original loop is widened to
3216 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3217 // that we vectorized. The PHI nodes are currently empty because we did
3218 // not want to introduce cycles. Notice that the remaining PHI nodes
3219 // that we need to fix are reduction variables.
3221 // Create the 'reduced' values for each of the induction vars.
3222 // The reduced values are the vector values that we scalarize and combine
3223 // after the loop is finished.
3224 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3226 PHINode *RdxPhi = *it;
3227 assert(RdxPhi && "Unable to recover vectorized PHI");
3229 // Find the reduction variable descriptor.
3230 assert(Legal->getReductionVars()->count(RdxPhi) &&
3231 "Unable to find the reduction variable");
3232 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3234 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3235 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3236 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3237 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3238 RdxDesc.getMinMaxRecurrenceKind();
3239 setDebugLocFromInst(Builder, ReductionStartValue);
3241 // We need to generate a reduction vector from the incoming scalar.
3242 // To do so, we need to generate the 'identity' vector and override
3243 // one of the elements with the incoming scalar reduction. We need
3244 // to do it in the vector-loop preheader.
3245 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3247 // This is the vector-clone of the value that leaves the loop.
3248 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3249 Type *VecTy = VectorExit[0]->getType();
3251 // Find the reduction identity variable. Zero for addition, or, xor,
3252 // one for multiplication, -1 for And.
3255 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3256 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3257 // MinMax reduction have the start value as their identify.
3259 VectorStart = Identity = ReductionStartValue;
3261 VectorStart = Identity =
3262 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3265 // Handle other reduction kinds:
3266 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3267 RK, VecTy->getScalarType());
3270 // This vector is the Identity vector where the first element is the
3271 // incoming scalar reduction.
3272 VectorStart = ReductionStartValue;
3274 Identity = ConstantVector::getSplat(VF, Iden);
3276 // This vector is the Identity vector where the first element is the
3277 // incoming scalar reduction.
3279 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3283 // Fix the vector-loop phi.
3285 // Reductions do not have to start at zero. They can start with
3286 // any loop invariant values.
3287 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3288 BasicBlock *Latch = OrigLoop->getLoopLatch();
3289 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3290 VectorParts &Val = getVectorValue(LoopVal);
3291 for (unsigned part = 0; part < UF; ++part) {
3292 // Make sure to add the reduction stat value only to the
3293 // first unroll part.
3294 Value *StartVal = (part == 0) ? VectorStart : Identity;
3295 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3296 LoopVectorPreHeader);
3297 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3298 LoopVectorBody.back());
3301 // Before each round, move the insertion point right between
3302 // the PHIs and the values we are going to write.
3303 // This allows us to write both PHINodes and the extractelement
3305 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3307 VectorParts RdxParts;
3308 setDebugLocFromInst(Builder, LoopExitInst);
3309 for (unsigned part = 0; part < UF; ++part) {
3310 // This PHINode contains the vectorized reduction variable, or
3311 // the initial value vector, if we bypass the vector loop.
3312 VectorParts &RdxExitVal = getVectorValue(LoopExitInst);
3313 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3314 Value *StartVal = (part == 0) ? VectorStart : Identity;
3315 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3316 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3317 NewPhi->addIncoming(RdxExitVal[part],
3318 LoopVectorBody.back());
3319 RdxParts.push_back(NewPhi);
3322 // Reduce all of the unrolled parts into a single vector.
3323 Value *ReducedPartRdx = RdxParts[0];
3324 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3325 setDebugLocFromInst(Builder, ReducedPartRdx);
3326 for (unsigned part = 1; part < UF; ++part) {
3327 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3328 // Floating point operations had to be 'fast' to enable the reduction.
3329 ReducedPartRdx = addFastMathFlag(
3330 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3331 ReducedPartRdx, "bin.rdx"));
3333 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3334 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3338 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3339 // and vector ops, reducing the set of values being computed by half each
3341 assert(isPowerOf2_32(VF) &&
3342 "Reduction emission only supported for pow2 vectors!");
3343 Value *TmpVec = ReducedPartRdx;
3344 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3345 for (unsigned i = VF; i != 1; i >>= 1) {
3346 // Move the upper half of the vector to the lower half.
3347 for (unsigned j = 0; j != i/2; ++j)
3348 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3350 // Fill the rest of the mask with undef.
3351 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3352 UndefValue::get(Builder.getInt32Ty()));
3355 Builder.CreateShuffleVector(TmpVec,
3356 UndefValue::get(TmpVec->getType()),
3357 ConstantVector::get(ShuffleMask),
3360 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3361 // Floating point operations had to be 'fast' to enable the reduction.
3362 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3363 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3365 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3369 // The result is in the first element of the vector.
3370 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3371 Builder.getInt32(0));
3374 // Create a phi node that merges control-flow from the backedge-taken check
3375 // block and the middle block.
3376 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3377 LoopScalarPreHeader->getTerminator());
3378 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3379 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3381 // Now, we need to fix the users of the reduction variable
3382 // inside and outside of the scalar remainder loop.
3383 // We know that the loop is in LCSSA form. We need to update the
3384 // PHI nodes in the exit blocks.
3385 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3386 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3387 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3388 if (!LCSSAPhi) break;
3390 // All PHINodes need to have a single entry edge, or two if
3391 // we already fixed them.
3392 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3394 // We found our reduction value exit-PHI. Update it with the
3395 // incoming bypass edge.
3396 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3397 // Add an edge coming from the bypass.
3398 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3401 }// end of the LCSSA phi scan.
3403 // Fix the scalar loop reduction variable with the incoming reduction sum
3404 // from the vector body and from the backedge value.
3405 int IncomingEdgeBlockIdx =
3406 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3407 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3408 // Pick the other block.
3409 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3410 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3411 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3412 }// end of for each redux variable.
3416 // Remove redundant induction instructions.
3417 cse(LoopVectorBody);
3420 void InnerLoopVectorizer::fixLCSSAPHIs() {
3421 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3422 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3423 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3424 if (!LCSSAPhi) break;
3425 if (LCSSAPhi->getNumIncomingValues() == 1)
3426 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3431 InnerLoopVectorizer::VectorParts
3432 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3433 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3436 // Look for cached value.
3437 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3438 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3439 if (ECEntryIt != MaskCache.end())
3440 return ECEntryIt->second;
3442 VectorParts SrcMask = createBlockInMask(Src);
3444 // The terminator has to be a branch inst!
3445 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3446 assert(BI && "Unexpected terminator found");
3448 if (BI->isConditional()) {
3449 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3451 if (BI->getSuccessor(0) != Dst)
3452 for (unsigned part = 0; part < UF; ++part)
3453 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3455 for (unsigned part = 0; part < UF; ++part)
3456 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3458 MaskCache[Edge] = EdgeMask;
3462 MaskCache[Edge] = SrcMask;
3466 InnerLoopVectorizer::VectorParts
3467 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3468 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3470 // Loop incoming mask is all-one.
3471 if (OrigLoop->getHeader() == BB) {
3472 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3473 return getVectorValue(C);
3476 // This is the block mask. We OR all incoming edges, and with zero.
3477 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3478 VectorParts BlockMask = getVectorValue(Zero);
3481 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3482 VectorParts EM = createEdgeMask(*it, BB);
3483 for (unsigned part = 0; part < UF; ++part)
3484 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3490 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3491 InnerLoopVectorizer::VectorParts &Entry,
3492 unsigned UF, unsigned VF, PhiVector *PV) {
3493 PHINode* P = cast<PHINode>(PN);
3494 // Handle reduction variables:
3495 if (Legal->getReductionVars()->count(P)) {
3496 for (unsigned part = 0; part < UF; ++part) {
3497 // This is phase one of vectorizing PHIs.
3498 Type *VecTy = (VF == 1) ? PN->getType() :
3499 VectorType::get(PN->getType(), VF);
3500 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3501 LoopVectorBody.back()-> getFirstInsertionPt());
3507 setDebugLocFromInst(Builder, P);
3508 // Check for PHI nodes that are lowered to vector selects.
3509 if (P->getParent() != OrigLoop->getHeader()) {
3510 // We know that all PHIs in non-header blocks are converted into
3511 // selects, so we don't have to worry about the insertion order and we
3512 // can just use the builder.
3513 // At this point we generate the predication tree. There may be
3514 // duplications since this is a simple recursive scan, but future
3515 // optimizations will clean it up.
3517 unsigned NumIncoming = P->getNumIncomingValues();
3519 // Generate a sequence of selects of the form:
3520 // SELECT(Mask3, In3,
3521 // SELECT(Mask2, In2,
3523 for (unsigned In = 0; In < NumIncoming; In++) {
3524 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3526 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3528 for (unsigned part = 0; part < UF; ++part) {
3529 // We might have single edge PHIs (blocks) - use an identity
3530 // 'select' for the first PHI operand.
3532 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3535 // Select between the current value and the previous incoming edge
3536 // based on the incoming mask.
3537 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3538 Entry[part], "predphi");
3544 // This PHINode must be an induction variable.
3545 // Make sure that we know about it.
3546 assert(Legal->getInductionVars()->count(P) &&
3547 "Not an induction variable");
3549 LoopVectorizationLegality::InductionInfo II =
3550 Legal->getInductionVars()->lookup(P);
3552 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3553 // which can be found from the original scalar operations.
3555 case LoopVectorizationLegality::IK_NoInduction:
3556 llvm_unreachable("Unknown induction");
3557 case LoopVectorizationLegality::IK_IntInduction: {
3558 assert(P->getType() == II.StartValue->getType() && "Types must match");
3559 Type *PhiTy = P->getType();
3561 if (P == OldInduction) {
3562 // Handle the canonical induction variable. We might have had to
3564 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3566 // Handle other induction variables that are now based on the
3568 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3570 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3571 Broadcasted = II.transform(Builder, NormalizedIdx);
3572 Broadcasted->setName("offset.idx");
3574 Broadcasted = getBroadcastInstrs(Broadcasted);
3575 // After broadcasting the induction variable we need to make the vector
3576 // consecutive by adding 0, 1, 2, etc.
3577 for (unsigned part = 0; part < UF; ++part)
3578 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3581 case LoopVectorizationLegality::IK_PtrInduction:
3582 // Handle the pointer induction variable case.
3583 assert(P->getType()->isPointerTy() && "Unexpected type.");
3584 // This is the normalized GEP that starts counting at zero.
3585 Value *NormalizedIdx =
3586 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3588 Builder.CreateSExtOrTrunc(NormalizedIdx, II.StepValue->getType());
3589 // This is the vector of results. Notice that we don't generate
3590 // vector geps because scalar geps result in better code.
3591 for (unsigned part = 0; part < UF; ++part) {
3593 int EltIndex = part;
3594 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3595 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3596 Value *SclrGep = II.transform(Builder, GlobalIdx);
3597 SclrGep->setName("next.gep");
3598 Entry[part] = SclrGep;
3602 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3603 for (unsigned int i = 0; i < VF; ++i) {
3604 int EltIndex = i + part * VF;
3605 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3606 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3607 Value *SclrGep = II.transform(Builder, GlobalIdx);
3608 SclrGep->setName("next.gep");
3609 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3610 Builder.getInt32(i),
3613 Entry[part] = VecVal;
3619 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3620 // For each instruction in the old loop.
3621 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3622 VectorParts &Entry = WidenMap.get(it);
3623 switch (it->getOpcode()) {
3624 case Instruction::Br:
3625 // Nothing to do for PHIs and BR, since we already took care of the
3626 // loop control flow instructions.
3628 case Instruction::PHI: {
3629 // Vectorize PHINodes.
3630 widenPHIInstruction(it, Entry, UF, VF, PV);
3634 case Instruction::Add:
3635 case Instruction::FAdd:
3636 case Instruction::Sub:
3637 case Instruction::FSub:
3638 case Instruction::Mul:
3639 case Instruction::FMul:
3640 case Instruction::UDiv:
3641 case Instruction::SDiv:
3642 case Instruction::FDiv:
3643 case Instruction::URem:
3644 case Instruction::SRem:
3645 case Instruction::FRem:
3646 case Instruction::Shl:
3647 case Instruction::LShr:
3648 case Instruction::AShr:
3649 case Instruction::And:
3650 case Instruction::Or:
3651 case Instruction::Xor: {
3652 // Just widen binops.
3653 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3654 setDebugLocFromInst(Builder, BinOp);
3655 VectorParts &A = getVectorValue(it->getOperand(0));
3656 VectorParts &B = getVectorValue(it->getOperand(1));
3658 // Use this vector value for all users of the original instruction.
3659 for (unsigned Part = 0; Part < UF; ++Part) {
3660 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3662 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3663 VecOp->copyIRFlags(BinOp);
3668 propagateMetadata(Entry, it);
3671 case Instruction::Select: {
3673 // If the selector is loop invariant we can create a select
3674 // instruction with a scalar condition. Otherwise, use vector-select.
3675 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3677 setDebugLocFromInst(Builder, it);
3679 // The condition can be loop invariant but still defined inside the
3680 // loop. This means that we can't just use the original 'cond' value.
3681 // We have to take the 'vectorized' value and pick the first lane.
3682 // Instcombine will make this a no-op.
3683 VectorParts &Cond = getVectorValue(it->getOperand(0));
3684 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3685 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3687 Value *ScalarCond = (VF == 1) ? Cond[0] :
3688 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3690 for (unsigned Part = 0; Part < UF; ++Part) {
3691 Entry[Part] = Builder.CreateSelect(
3692 InvariantCond ? ScalarCond : Cond[Part],
3697 propagateMetadata(Entry, it);
3701 case Instruction::ICmp:
3702 case Instruction::FCmp: {
3703 // Widen compares. Generate vector compares.
3704 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3705 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3706 setDebugLocFromInst(Builder, it);
3707 VectorParts &A = getVectorValue(it->getOperand(0));
3708 VectorParts &B = getVectorValue(it->getOperand(1));
3709 for (unsigned Part = 0; Part < UF; ++Part) {
3712 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3714 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3718 propagateMetadata(Entry, it);
3722 case Instruction::Store:
3723 case Instruction::Load:
3724 vectorizeMemoryInstruction(it);
3726 case Instruction::ZExt:
3727 case Instruction::SExt:
3728 case Instruction::FPToUI:
3729 case Instruction::FPToSI:
3730 case Instruction::FPExt:
3731 case Instruction::PtrToInt:
3732 case Instruction::IntToPtr:
3733 case Instruction::SIToFP:
3734 case Instruction::UIToFP:
3735 case Instruction::Trunc:
3736 case Instruction::FPTrunc:
3737 case Instruction::BitCast: {
3738 CastInst *CI = dyn_cast<CastInst>(it);
3739 setDebugLocFromInst(Builder, it);
3740 /// Optimize the special case where the source is the induction
3741 /// variable. Notice that we can only optimize the 'trunc' case
3742 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3743 /// c. other casts depend on pointer size.
3744 if (CI->getOperand(0) == OldInduction &&
3745 it->getOpcode() == Instruction::Trunc) {
3746 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3748 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3749 LoopVectorizationLegality::InductionInfo II =
3750 Legal->getInductionVars()->lookup(OldInduction);
3752 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3753 for (unsigned Part = 0; Part < UF; ++Part)
3754 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3755 propagateMetadata(Entry, it);
3758 /// Vectorize casts.
3759 Type *DestTy = (VF == 1) ? CI->getType() :
3760 VectorType::get(CI->getType(), VF);
3762 VectorParts &A = getVectorValue(it->getOperand(0));
3763 for (unsigned Part = 0; Part < UF; ++Part)
3764 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3765 propagateMetadata(Entry, it);
3769 case Instruction::Call: {
3770 // Ignore dbg intrinsics.
3771 if (isa<DbgInfoIntrinsic>(it))
3773 setDebugLocFromInst(Builder, it);
3775 Module *M = BB->getParent()->getParent();
3776 CallInst *CI = cast<CallInst>(it);
3778 StringRef FnName = CI->getCalledFunction()->getName();
3779 Function *F = CI->getCalledFunction();
3780 Type *RetTy = ToVectorTy(CI->getType(), VF);
3781 SmallVector<Type *, 4> Tys;
3782 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3783 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3785 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3787 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3788 ID == Intrinsic::lifetime_start)) {
3789 scalarizeInstruction(it);
3792 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3793 // version of the instruction.
3794 // Is it beneficial to perform intrinsic call compared to lib call?
3795 bool NeedToScalarize;
3796 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3797 bool UseVectorIntrinsic =
3798 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3799 if (!UseVectorIntrinsic && NeedToScalarize) {
3800 scalarizeInstruction(it);
3804 for (unsigned Part = 0; Part < UF; ++Part) {
3805 SmallVector<Value *, 4> Args;
3806 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3807 Value *Arg = CI->getArgOperand(i);
3808 // Some intrinsics have a scalar argument - don't replace it with a
3810 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3811 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3812 Arg = VectorArg[Part];
3814 Args.push_back(Arg);
3818 if (UseVectorIntrinsic) {
3819 // Use vector version of the intrinsic.
3820 Type *TysForDecl[] = {CI->getType()};
3822 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3823 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3825 // Use vector version of the library call.
3826 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3827 assert(!VFnName.empty() && "Vector function name is empty.");
3828 VectorF = M->getFunction(VFnName);
3830 // Generate a declaration
3831 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3833 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3834 VectorF->copyAttributesFrom(F);
3837 assert(VectorF && "Can't create vector function.");
3838 Entry[Part] = Builder.CreateCall(VectorF, Args);
3841 propagateMetadata(Entry, it);
3846 // All other instructions are unsupported. Scalarize them.
3847 scalarizeInstruction(it);
3850 }// end of for_each instr.
3853 void InnerLoopVectorizer::updateAnalysis() {
3854 // Forget the original basic block.
3855 SE->forgetLoop(OrigLoop);
3857 // Update the dominator tree information.
3858 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3859 "Entry does not dominate exit.");
3861 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3862 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3863 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3865 // Due to if predication of stores we might create a sequence of "if(pred)
3866 // a[i] = ...; " blocks.
3867 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3869 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3870 else if (isPredicatedBlock(i)) {
3871 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3873 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3877 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3878 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3879 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3880 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3882 DEBUG(DT->verifyDomTree());
3885 /// \brief Check whether it is safe to if-convert this phi node.
3887 /// Phi nodes with constant expressions that can trap are not safe to if
3889 static bool canIfConvertPHINodes(BasicBlock *BB) {
3890 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3891 PHINode *Phi = dyn_cast<PHINode>(I);
3894 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3895 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3902 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3903 if (!EnableIfConversion) {
3904 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3908 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3910 // A list of pointers that we can safely read and write to.
3911 SmallPtrSet<Value *, 8> SafePointes;
3913 // Collect safe addresses.
3914 for (Loop::block_iterator BI = TheLoop->block_begin(),
3915 BE = TheLoop->block_end(); BI != BE; ++BI) {
3916 BasicBlock *BB = *BI;
3918 if (blockNeedsPredication(BB))
3921 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3922 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3923 SafePointes.insert(LI->getPointerOperand());
3924 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3925 SafePointes.insert(SI->getPointerOperand());
3929 // Collect the blocks that need predication.
3930 BasicBlock *Header = TheLoop->getHeader();
3931 for (Loop::block_iterator BI = TheLoop->block_begin(),
3932 BE = TheLoop->block_end(); BI != BE; ++BI) {
3933 BasicBlock *BB = *BI;
3935 // We don't support switch statements inside loops.
3936 if (!isa<BranchInst>(BB->getTerminator())) {
3937 emitAnalysis(VectorizationReport(BB->getTerminator())
3938 << "loop contains a switch statement");
3942 // We must be able to predicate all blocks that need to be predicated.
3943 if (blockNeedsPredication(BB)) {
3944 if (!blockCanBePredicated(BB, SafePointes)) {
3945 emitAnalysis(VectorizationReport(BB->getTerminator())
3946 << "control flow cannot be substituted for a select");
3949 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3950 emitAnalysis(VectorizationReport(BB->getTerminator())
3951 << "control flow cannot be substituted for a select");
3956 // We can if-convert this loop.
3960 bool LoopVectorizationLegality::canVectorize() {
3961 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3962 // be canonicalized.
3963 if (!TheLoop->getLoopPreheader()) {
3965 VectorizationReport() <<
3966 "loop control flow is not understood by vectorizer");
3970 // We can only vectorize innermost loops.
3971 if (!TheLoop->empty()) {
3972 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3976 // We must have a single backedge.
3977 if (TheLoop->getNumBackEdges() != 1) {
3979 VectorizationReport() <<
3980 "loop control flow is not understood by vectorizer");
3984 // We must have a single exiting block.
3985 if (!TheLoop->getExitingBlock()) {
3987 VectorizationReport() <<
3988 "loop control flow is not understood by vectorizer");
3992 // We only handle bottom-tested loops, i.e. loop in which the condition is
3993 // checked at the end of each iteration. With that we can assume that all
3994 // instructions in the loop are executed the same number of times.
3995 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3997 VectorizationReport() <<
3998 "loop control flow is not understood by vectorizer");
4002 // We need to have a loop header.
4003 DEBUG(dbgs() << "LV: Found a loop: " <<
4004 TheLoop->getHeader()->getName() << '\n');
4006 // Check if we can if-convert non-single-bb loops.
4007 unsigned NumBlocks = TheLoop->getNumBlocks();
4008 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4009 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4013 // ScalarEvolution needs to be able to find the exit count.
4014 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
4015 if (ExitCount == SE->getCouldNotCompute()) {
4016 emitAnalysis(VectorizationReport() <<
4017 "could not determine number of loop iterations");
4018 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4022 // Check if we can vectorize the instructions and CFG in this loop.
4023 if (!canVectorizeInstrs()) {
4024 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4028 // Go over each instruction and look at memory deps.
4029 if (!canVectorizeMemory()) {
4030 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4034 // Collect all of the variables that remain uniform after vectorization.
4035 collectLoopUniforms();
4037 DEBUG(dbgs() << "LV: We can vectorize this loop"
4038 << (LAI->getRuntimePointerChecking()->Need
4039 ? " (with a runtime bound check)"
4043 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4045 // If an override option has been passed in for interleaved accesses, use it.
4046 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4047 UseInterleaved = EnableInterleavedMemAccesses;
4049 // Analyze interleaved memory accesses.
4051 InterleaveInfo.analyzeInterleaving(Strides);
4053 // Okay! We can vectorize. At this point we don't have any other mem analysis
4054 // which may limit our maximum vectorization factor, so just return true with
4059 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4060 if (Ty->isPointerTy())
4061 return DL.getIntPtrType(Ty);
4063 // It is possible that char's or short's overflow when we ask for the loop's
4064 // trip count, work around this by changing the type size.
4065 if (Ty->getScalarSizeInBits() < 32)
4066 return Type::getInt32Ty(Ty->getContext());
4071 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4072 Ty0 = convertPointerToIntegerType(DL, Ty0);
4073 Ty1 = convertPointerToIntegerType(DL, Ty1);
4074 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4079 /// \brief Check that the instruction has outside loop users and is not an
4080 /// identified reduction variable.
4081 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4082 SmallPtrSetImpl<Value *> &Reductions) {
4083 // Reduction instructions are allowed to have exit users. All other
4084 // instructions must not have external users.
4085 if (!Reductions.count(Inst))
4086 //Check that all of the users of the loop are inside the BB.
4087 for (User *U : Inst->users()) {
4088 Instruction *UI = cast<Instruction>(U);
4089 // This user may be a reduction exit value.
4090 if (!TheLoop->contains(UI)) {
4091 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4098 bool LoopVectorizationLegality::canVectorizeInstrs() {
4099 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
4100 BasicBlock *Header = TheLoop->getHeader();
4102 // Look for the attribute signaling the absence of NaNs.
4103 Function &F = *Header->getParent();
4104 const DataLayout &DL = F.getParent()->getDataLayout();
4105 if (F.hasFnAttribute("no-nans-fp-math"))
4107 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4109 // For each block in the loop.
4110 for (Loop::block_iterator bb = TheLoop->block_begin(),
4111 be = TheLoop->block_end(); bb != be; ++bb) {
4113 // Scan the instructions in the block and look for hazards.
4114 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4117 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4118 Type *PhiTy = Phi->getType();
4119 // Check that this PHI type is allowed.
4120 if (!PhiTy->isIntegerTy() &&
4121 !PhiTy->isFloatingPointTy() &&
4122 !PhiTy->isPointerTy()) {
4123 emitAnalysis(VectorizationReport(it)
4124 << "loop control flow is not understood by vectorizer");
4125 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4129 // If this PHINode is not in the header block, then we know that we
4130 // can convert it to select during if-conversion. No need to check if
4131 // the PHIs in this block are induction or reduction variables.
4132 if (*bb != Header) {
4133 // Check that this instruction has no outside users or is an
4134 // identified reduction value with an outside user.
4135 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4137 emitAnalysis(VectorizationReport(it) <<
4138 "value could not be identified as "
4139 "an induction or reduction variable");
4143 // We only allow if-converted PHIs with exactly two incoming values.
4144 if (Phi->getNumIncomingValues() != 2) {
4145 emitAnalysis(VectorizationReport(it)
4146 << "control flow not understood by vectorizer");
4147 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4151 // This is the value coming from the preheader.
4152 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
4153 ConstantInt *StepValue = nullptr;
4154 // Check if this is an induction variable.
4155 InductionKind IK = isInductionVariable(Phi, StepValue);
4157 if (IK_NoInduction != IK) {
4158 // Get the widest type.
4160 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4162 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4164 // Int inductions are special because we only allow one IV.
4165 if (IK == IK_IntInduction && StepValue->isOne()) {
4166 // Use the phi node with the widest type as induction. Use the last
4167 // one if there are multiple (no good reason for doing this other
4168 // than it is expedient).
4169 if (!Induction || PhiTy == WidestIndTy)
4173 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4174 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
4176 // Until we explicitly handle the case of an induction variable with
4177 // an outside loop user we have to give up vectorizing this loop.
4178 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4179 emitAnalysis(VectorizationReport(it) <<
4180 "use of induction value outside of the "
4181 "loop is not handled by vectorizer");
4188 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4190 if (Reductions[Phi].hasUnsafeAlgebra())
4191 Requirements->addUnsafeAlgebraInst(
4192 Reductions[Phi].getUnsafeAlgebraInst());
4193 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4197 emitAnalysis(VectorizationReport(it) <<
4198 "value that could not be identified as "
4199 "reduction is used outside the loop");
4200 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4202 }// end of PHI handling
4204 // We handle calls that:
4205 // * Are debug info intrinsics.
4206 // * Have a mapping to an IR intrinsic.
4207 // * Have a vector version available.
4208 CallInst *CI = dyn_cast<CallInst>(it);
4209 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4210 !(CI->getCalledFunction() && TLI &&
4211 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4212 emitAnalysis(VectorizationReport(it) <<
4213 "call instruction cannot be vectorized");
4214 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4218 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4219 // second argument is the same (i.e. loop invariant)
4221 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4222 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4223 emitAnalysis(VectorizationReport(it)
4224 << "intrinsic instruction cannot be vectorized");
4225 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4230 // Check that the instruction return type is vectorizable.
4231 // Also, we can't vectorize extractelement instructions.
4232 if ((!VectorType::isValidElementType(it->getType()) &&
4233 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4234 emitAnalysis(VectorizationReport(it)
4235 << "instruction return type cannot be vectorized");
4236 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4240 // Check that the stored type is vectorizable.
4241 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4242 Type *T = ST->getValueOperand()->getType();
4243 if (!VectorType::isValidElementType(T)) {
4244 emitAnalysis(VectorizationReport(ST) <<
4245 "store instruction cannot be vectorized");
4248 if (EnableMemAccessVersioning)
4249 collectStridedAccess(ST);
4252 if (EnableMemAccessVersioning)
4253 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4254 collectStridedAccess(LI);
4256 // Reduction instructions are allowed to have exit users.
4257 // All other instructions must not have external users.
4258 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4259 emitAnalysis(VectorizationReport(it) <<
4260 "value cannot be used outside the loop");
4269 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4270 if (Inductions.empty()) {
4271 emitAnalysis(VectorizationReport()
4272 << "loop induction variable could not be identified");
4280 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4281 Value *Ptr = nullptr;
4282 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4283 Ptr = LI->getPointerOperand();
4284 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4285 Ptr = SI->getPointerOperand();
4289 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4293 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4294 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4295 Strides[Ptr] = Stride;
4296 StrideSet.insert(Stride);
4299 void LoopVectorizationLegality::collectLoopUniforms() {
4300 // We now know that the loop is vectorizable!
4301 // Collect variables that will remain uniform after vectorization.
4302 std::vector<Value*> Worklist;
4303 BasicBlock *Latch = TheLoop->getLoopLatch();
4305 // Start with the conditional branch and walk up the block.
4306 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4308 // Also add all consecutive pointer values; these values will be uniform
4309 // after vectorization (and subsequent cleanup) and, until revectorization is
4310 // supported, all dependencies must also be uniform.
4311 for (Loop::block_iterator B = TheLoop->block_begin(),
4312 BE = TheLoop->block_end(); B != BE; ++B)
4313 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4315 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4316 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4318 while (!Worklist.empty()) {
4319 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4320 Worklist.pop_back();
4322 // Look at instructions inside this loop.
4323 // Stop when reaching PHI nodes.
4324 // TODO: we need to follow values all over the loop, not only in this block.
4325 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4328 // This is a known uniform.
4331 // Insert all operands.
4332 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4336 bool LoopVectorizationLegality::canVectorizeMemory() {
4337 LAI = &LAA->getInfo(TheLoop, Strides);
4338 auto &OptionalReport = LAI->getReport();
4340 emitAnalysis(VectorizationReport(*OptionalReport));
4341 if (!LAI->canVectorizeMemory())
4344 if (LAI->hasStoreToLoopInvariantAddress()) {
4346 VectorizationReport()
4347 << "write to a loop invariant address could not be vectorized");
4348 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4352 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4357 LoopVectorizationLegality::InductionKind
4358 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4359 ConstantInt *&StepValue) {
4360 if (!isInductionPHI(Phi, SE, StepValue))
4361 return IK_NoInduction;
4363 Type *PhiTy = Phi->getType();
4364 // Found an Integer induction variable.
4365 if (PhiTy->isIntegerTy())
4366 return IK_IntInduction;
4367 // Found an Pointer induction variable.
4368 return IK_PtrInduction;
4371 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4372 Value *In0 = const_cast<Value*>(V);
4373 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4377 return Inductions.count(PN);
4380 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4381 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4384 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4385 SmallPtrSetImpl<Value *> &SafePtrs) {
4387 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4388 // Check that we don't have a constant expression that can trap as operand.
4389 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4391 if (Constant *C = dyn_cast<Constant>(*OI))
4395 // We might be able to hoist the load.
4396 if (it->mayReadFromMemory()) {
4397 LoadInst *LI = dyn_cast<LoadInst>(it);
4400 if (!SafePtrs.count(LI->getPointerOperand())) {
4401 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4402 MaskedOp.insert(LI);
4409 // We don't predicate stores at the moment.
4410 if (it->mayWriteToMemory()) {
4411 StoreInst *SI = dyn_cast<StoreInst>(it);
4412 // We only support predication of stores in basic blocks with one
4417 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4418 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4420 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4421 !isSinglePredecessor) {
4422 // Build a masked store if it is legal for the target, otherwise scalarize
4424 bool isLegalMaskedOp =
4425 isLegalMaskedStore(SI->getValueOperand()->getType(),
4426 SI->getPointerOperand());
4427 if (isLegalMaskedOp) {
4429 MaskedOp.insert(SI);
4438 // The instructions below can trap.
4439 switch (it->getOpcode()) {
4441 case Instruction::UDiv:
4442 case Instruction::SDiv:
4443 case Instruction::URem:
4444 case Instruction::SRem:
4452 void InterleavedAccessInfo::collectConstStridedAccesses(
4453 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4454 const ValueToValueMap &Strides) {
4455 // Holds load/store instructions in program order.
4456 SmallVector<Instruction *, 16> AccessList;
4458 for (auto *BB : TheLoop->getBlocks()) {
4459 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4461 for (auto &I : *BB) {
4462 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4464 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4468 AccessList.push_back(&I);
4472 if (AccessList.empty())
4475 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4476 for (auto I : AccessList) {
4477 LoadInst *LI = dyn_cast<LoadInst>(I);
4478 StoreInst *SI = dyn_cast<StoreInst>(I);
4480 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4481 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4483 // The factor of the corresponding interleave group.
4484 unsigned Factor = std::abs(Stride);
4486 // Ignore the access if the factor is too small or too large.
4487 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4490 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4491 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4492 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4494 // An alignment of 0 means target ABI alignment.
4495 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4497 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4499 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4503 // Analyze interleaved accesses and collect them into interleave groups.
4505 // Notice that the vectorization on interleaved groups will change instruction
4506 // orders and may break dependences. But the memory dependence check guarantees
4507 // that there is no overlap between two pointers of different strides, element
4508 // sizes or underlying bases.
4510 // For pointers sharing the same stride, element size and underlying base, no
4511 // need to worry about Read-After-Write dependences and Write-After-Read
4514 // E.g. The RAW dependence: A[i] = a;
4516 // This won't exist as it is a store-load forwarding conflict, which has
4517 // already been checked and forbidden in the dependence check.
4519 // E.g. The WAR dependence: a = A[i]; // (1)
4521 // The store group of (2) is always inserted at or below (2), and the load group
4522 // of (1) is always inserted at or above (1). The dependence is safe.
4523 void InterleavedAccessInfo::analyzeInterleaving(
4524 const ValueToValueMap &Strides) {
4525 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4527 // Holds all the stride accesses.
4528 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4529 collectConstStridedAccesses(StrideAccesses, Strides);
4531 if (StrideAccesses.empty())
4534 // Holds all interleaved store groups temporarily.
4535 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4537 // Search the load-load/write-write pair B-A in bottom-up order and try to
4538 // insert B into the interleave group of A according to 3 rules:
4539 // 1. A and B have the same stride.
4540 // 2. A and B have the same memory object size.
4541 // 3. B belongs to the group according to the distance.
4543 // The bottom-up order can avoid breaking the Write-After-Write dependences
4544 // between two pointers of the same base.
4545 // E.g. A[i] = a; (1)
4548 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4549 // above (1), which guarantees that (1) is always above (2).
4550 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4552 Instruction *A = I->first;
4553 StrideDescriptor DesA = I->second;
4555 InterleaveGroup *Group = getInterleaveGroup(A);
4557 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4558 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4561 if (A->mayWriteToMemory())
4562 StoreGroups.insert(Group);
4564 for (auto II = std::next(I); II != E; ++II) {
4565 Instruction *B = II->first;
4566 StrideDescriptor DesB = II->second;
4568 // Ignore if B is already in a group or B is a different memory operation.
4569 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4572 // Check the rule 1 and 2.
4573 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4576 // Calculate the distance and prepare for the rule 3.
4577 const SCEVConstant *DistToA =
4578 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4582 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4584 // Skip if the distance is not multiple of size as they are not in the
4586 if (DistanceToA % static_cast<int>(DesA.Size))
4589 // The index of B is the index of A plus the related index to A.
4591 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4593 // Try to insert B into the group.
4594 if (Group->insertMember(B, IndexB, DesB.Align)) {
4595 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4596 << " into the interleave group with" << *A << '\n');
4597 InterleaveGroupMap[B] = Group;
4599 // Set the first load in program order as the insert position.
4600 if (B->mayReadFromMemory())
4601 Group->setInsertPos(B);
4603 } // Iteration on instruction B
4604 } // Iteration on instruction A
4606 // Remove interleaved store groups with gaps.
4607 for (InterleaveGroup *Group : StoreGroups)
4608 if (Group->getNumMembers() != Group->getFactor())
4609 releaseGroup(Group);
4612 LoopVectorizationCostModel::VectorizationFactor
4613 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4614 // Width 1 means no vectorize
4615 VectorizationFactor Factor = { 1U, 0U };
4616 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4617 emitAnalysis(VectorizationReport() <<
4618 "runtime pointer checks needed. Enable vectorization of this "
4619 "loop with '#pragma clang loop vectorize(enable)' when "
4620 "compiling with -Os/-Oz");
4622 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4626 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4627 emitAnalysis(VectorizationReport() <<
4628 "store that is conditionally executed prevents vectorization");
4629 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4633 // Find the trip count.
4634 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4635 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4637 unsigned WidestType = getWidestType();
4638 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4639 unsigned MaxSafeDepDist = -1U;
4640 if (Legal->getMaxSafeDepDistBytes() != -1U)
4641 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4642 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4643 WidestRegister : MaxSafeDepDist);
4644 unsigned MaxVectorSize = WidestRegister / WidestType;
4645 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4646 DEBUG(dbgs() << "LV: The Widest register is: "
4647 << WidestRegister << " bits.\n");
4649 if (MaxVectorSize == 0) {
4650 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4654 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4655 " into one vector!");
4657 unsigned VF = MaxVectorSize;
4659 // If we optimize the program for size, avoid creating the tail loop.
4661 // If we are unable to calculate the trip count then don't try to vectorize.
4664 (VectorizationReport() <<
4665 "unable to calculate the loop count due to complex control flow");
4666 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4670 // Find the maximum SIMD width that can fit within the trip count.
4671 VF = TC % MaxVectorSize;
4676 // If the trip count that we found modulo the vectorization factor is not
4677 // zero then we require a tail.
4678 emitAnalysis(VectorizationReport() <<
4679 "cannot optimize for size and vectorize at the "
4680 "same time. Enable vectorization of this loop "
4681 "with '#pragma clang loop vectorize(enable)' "
4682 "when compiling with -Os/-Oz");
4683 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4688 int UserVF = Hints->getWidth();
4690 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4691 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4693 Factor.Width = UserVF;
4697 float Cost = expectedCost(1);
4699 const float ScalarCost = Cost;
4702 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4704 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4705 // Ignore scalar width, because the user explicitly wants vectorization.
4706 if (ForceVectorization && VF > 1) {
4708 Cost = expectedCost(Width) / (float)Width;
4711 for (unsigned i=2; i <= VF; i*=2) {
4712 // Notice that the vector loop needs to be executed less times, so
4713 // we need to divide the cost of the vector loops by the width of
4714 // the vector elements.
4715 float VectorCost = expectedCost(i) / (float)i;
4716 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4717 (int)VectorCost << ".\n");
4718 if (VectorCost < Cost) {
4724 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4725 << "LV: Vectorization seems to be not beneficial, "
4726 << "but was forced by a user.\n");
4727 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4728 Factor.Width = Width;
4729 Factor.Cost = Width * Cost;
4733 unsigned LoopVectorizationCostModel::getWidestType() {
4734 unsigned MaxWidth = 8;
4735 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4738 for (Loop::block_iterator bb = TheLoop->block_begin(),
4739 be = TheLoop->block_end(); bb != be; ++bb) {
4740 BasicBlock *BB = *bb;
4742 // For each instruction in the loop.
4743 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4744 Type *T = it->getType();
4746 // Ignore ephemeral values.
4747 if (EphValues.count(it))
4750 // Only examine Loads, Stores and PHINodes.
4751 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4754 // Examine PHI nodes that are reduction variables.
4755 if (PHINode *PN = dyn_cast<PHINode>(it))
4756 if (!Legal->getReductionVars()->count(PN))
4759 // Examine the stored values.
4760 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4761 T = ST->getValueOperand()->getType();
4763 // Ignore loaded pointer types and stored pointer types that are not
4764 // consecutive. However, we do want to take consecutive stores/loads of
4765 // pointer vectors into account.
4766 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4769 MaxWidth = std::max(MaxWidth,
4770 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4777 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4779 unsigned LoopCost) {
4781 // -- The interleave heuristics --
4782 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4783 // There are many micro-architectural considerations that we can't predict
4784 // at this level. For example, frontend pressure (on decode or fetch) due to
4785 // code size, or the number and capabilities of the execution ports.
4787 // We use the following heuristics to select the interleave count:
4788 // 1. If the code has reductions, then we interleave to break the cross
4789 // iteration dependency.
4790 // 2. If the loop is really small, then we interleave to reduce the loop
4792 // 3. We don't interleave if we think that we will spill registers to memory
4793 // due to the increased register pressure.
4795 // When we optimize for size, we don't interleave.
4799 // We used the distance for the interleave count.
4800 if (Legal->getMaxSafeDepDistBytes() != -1U)
4803 // Do not interleave loops with a relatively small trip count.
4804 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4805 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4808 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4809 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4813 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4814 TargetNumRegisters = ForceTargetNumScalarRegs;
4816 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4817 TargetNumRegisters = ForceTargetNumVectorRegs;
4820 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4821 // We divide by these constants so assume that we have at least one
4822 // instruction that uses at least one register.
4823 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4824 R.NumInstructions = std::max(R.NumInstructions, 1U);
4826 // We calculate the interleave count using the following formula.
4827 // Subtract the number of loop invariants from the number of available
4828 // registers. These registers are used by all of the interleaved instances.
4829 // Next, divide the remaining registers by the number of registers that is
4830 // required by the loop, in order to estimate how many parallel instances
4831 // fit without causing spills. All of this is rounded down if necessary to be
4832 // a power of two. We want power of two interleave count to simplify any
4833 // addressing operations or alignment considerations.
4834 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4837 // Don't count the induction variable as interleaved.
4838 if (EnableIndVarRegisterHeur)
4839 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4840 std::max(1U, (R.MaxLocalUsers - 1)));
4842 // Clamp the interleave ranges to reasonable counts.
4843 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4845 // Check if the user has overridden the max.
4847 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4848 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4850 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4851 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4854 // If we did not calculate the cost for VF (because the user selected the VF)
4855 // then we calculate the cost of VF here.
4857 LoopCost = expectedCost(VF);
4859 // Clamp the calculated IC to be between the 1 and the max interleave count
4860 // that the target allows.
4861 if (IC > MaxInterleaveCount)
4862 IC = MaxInterleaveCount;
4866 // Interleave if we vectorized this loop and there is a reduction that could
4867 // benefit from interleaving.
4868 if (VF > 1 && Legal->getReductionVars()->size()) {
4869 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4873 // Note that if we've already vectorized the loop we will have done the
4874 // runtime check and so interleaving won't require further checks.
4875 bool InterleavingRequiresRuntimePointerCheck =
4876 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4878 // We want to interleave small loops in order to reduce the loop overhead and
4879 // potentially expose ILP opportunities.
4880 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4881 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4882 // We assume that the cost overhead is 1 and we use the cost model
4883 // to estimate the cost of the loop and interleave until the cost of the
4884 // loop overhead is about 5% of the cost of the loop.
4886 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4888 // Interleave until store/load ports (estimated by max interleave count) are
4890 unsigned NumStores = Legal->getNumStores();
4891 unsigned NumLoads = Legal->getNumLoads();
4892 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4893 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4895 // If we have a scalar reduction (vector reductions are already dealt with
4896 // by this point), we can increase the critical path length if the loop
4897 // we're interleaving is inside another loop. Limit, by default to 2, so the
4898 // critical path only gets increased by one reduction operation.
4899 if (Legal->getReductionVars()->size() &&
4900 TheLoop->getLoopDepth() > 1) {
4901 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4902 SmallIC = std::min(SmallIC, F);
4903 StoresIC = std::min(StoresIC, F);
4904 LoadsIC = std::min(LoadsIC, F);
4907 if (EnableLoadStoreRuntimeInterleave &&
4908 std::max(StoresIC, LoadsIC) > SmallIC) {
4909 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4910 return std::max(StoresIC, LoadsIC);
4913 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4917 // Interleave if this is a large loop (small loops are already dealt with by
4919 // point) that could benefit from interleaving.
4920 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4921 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4922 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4926 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4930 LoopVectorizationCostModel::RegisterUsage
4931 LoopVectorizationCostModel::calculateRegisterUsage() {
4932 // This function calculates the register usage by measuring the highest number
4933 // of values that are alive at a single location. Obviously, this is a very
4934 // rough estimation. We scan the loop in a topological order in order and
4935 // assign a number to each instruction. We use RPO to ensure that defs are
4936 // met before their users. We assume that each instruction that has in-loop
4937 // users starts an interval. We record every time that an in-loop value is
4938 // used, so we have a list of the first and last occurrences of each
4939 // instruction. Next, we transpose this data structure into a multi map that
4940 // holds the list of intervals that *end* at a specific location. This multi
4941 // map allows us to perform a linear search. We scan the instructions linearly
4942 // and record each time that a new interval starts, by placing it in a set.
4943 // If we find this value in the multi-map then we remove it from the set.
4944 // The max register usage is the maximum size of the set.
4945 // We also search for instructions that are defined outside the loop, but are
4946 // used inside the loop. We need this number separately from the max-interval
4947 // usage number because when we unroll, loop-invariant values do not take
4949 LoopBlocksDFS DFS(TheLoop);
4953 R.NumInstructions = 0;
4955 // Each 'key' in the map opens a new interval. The values
4956 // of the map are the index of the 'last seen' usage of the
4957 // instruction that is the key.
4958 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4959 // Maps instruction to its index.
4960 DenseMap<unsigned, Instruction*> IdxToInstr;
4961 // Marks the end of each interval.
4962 IntervalMap EndPoint;
4963 // Saves the list of instruction indices that are used in the loop.
4964 SmallSet<Instruction*, 8> Ends;
4965 // Saves the list of values that are used in the loop but are
4966 // defined outside the loop, such as arguments and constants.
4967 SmallPtrSet<Value*, 8> LoopInvariants;
4970 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4971 be = DFS.endRPO(); bb != be; ++bb) {
4972 R.NumInstructions += (*bb)->size();
4973 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4975 Instruction *I = it;
4976 IdxToInstr[Index++] = I;
4978 // Save the end location of each USE.
4979 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4980 Value *U = I->getOperand(i);
4981 Instruction *Instr = dyn_cast<Instruction>(U);
4983 // Ignore non-instruction values such as arguments, constants, etc.
4984 if (!Instr) continue;
4986 // If this instruction is outside the loop then record it and continue.
4987 if (!TheLoop->contains(Instr)) {
4988 LoopInvariants.insert(Instr);
4992 // Overwrite previous end points.
4993 EndPoint[Instr] = Index;
4999 // Saves the list of intervals that end with the index in 'key'.
5000 typedef SmallVector<Instruction*, 2> InstrList;
5001 DenseMap<unsigned, InstrList> TransposeEnds;
5003 // Transpose the EndPoints to a list of values that end at each index.
5004 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5006 TransposeEnds[it->second].push_back(it->first);
5008 SmallSet<Instruction*, 8> OpenIntervals;
5009 unsigned MaxUsage = 0;
5012 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5013 for (unsigned int i = 0; i < Index; ++i) {
5014 Instruction *I = IdxToInstr[i];
5015 // Ignore instructions that are never used within the loop.
5016 if (!Ends.count(I)) continue;
5018 // Ignore ephemeral values.
5019 if (EphValues.count(I))
5022 // Remove all of the instructions that end at this location.
5023 InstrList &List = TransposeEnds[i];
5024 for (unsigned int j=0, e = List.size(); j < e; ++j)
5025 OpenIntervals.erase(List[j]);
5027 // Count the number of live interals.
5028 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5030 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5031 OpenIntervals.size() << '\n');
5033 // Add the current instruction to the list of open intervals.
5034 OpenIntervals.insert(I);
5037 unsigned Invariant = LoopInvariants.size();
5038 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5039 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5040 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5042 R.LoopInvariantRegs = Invariant;
5043 R.MaxLocalUsers = MaxUsage;
5047 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5051 for (Loop::block_iterator bb = TheLoop->block_begin(),
5052 be = TheLoop->block_end(); bb != be; ++bb) {
5053 unsigned BlockCost = 0;
5054 BasicBlock *BB = *bb;
5056 // For each instruction in the old loop.
5057 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5058 // Skip dbg intrinsics.
5059 if (isa<DbgInfoIntrinsic>(it))
5062 // Ignore ephemeral values.
5063 if (EphValues.count(it))
5066 unsigned C = getInstructionCost(it, VF);
5068 // Check if we should override the cost.
5069 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5070 C = ForceTargetInstructionCost;
5073 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5074 VF << " For instruction: " << *it << '\n');
5077 // We assume that if-converted blocks have a 50% chance of being executed.
5078 // When the code is scalar then some of the blocks are avoided due to CF.
5079 // When the code is vectorized we execute all code paths.
5080 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5089 /// \brief Check whether the address computation for a non-consecutive memory
5090 /// access looks like an unlikely candidate for being merged into the indexing
5093 /// We look for a GEP which has one index that is an induction variable and all
5094 /// other indices are loop invariant. If the stride of this access is also
5095 /// within a small bound we decide that this address computation can likely be
5096 /// merged into the addressing mode.
5097 /// In all other cases, we identify the address computation as complex.
5098 static bool isLikelyComplexAddressComputation(Value *Ptr,
5099 LoopVectorizationLegality *Legal,
5100 ScalarEvolution *SE,
5101 const Loop *TheLoop) {
5102 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5106 // We are looking for a gep with all loop invariant indices except for one
5107 // which should be an induction variable.
5108 unsigned NumOperands = Gep->getNumOperands();
5109 for (unsigned i = 1; i < NumOperands; ++i) {
5110 Value *Opd = Gep->getOperand(i);
5111 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5112 !Legal->isInductionVariable(Opd))
5116 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5117 // can likely be merged into the address computation.
5118 unsigned MaxMergeDistance = 64;
5120 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5124 // Check the step is constant.
5125 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5126 // Calculate the pointer stride and check if it is consecutive.
5127 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5131 const APInt &APStepVal = C->getValue()->getValue();
5133 // Huge step value - give up.
5134 if (APStepVal.getBitWidth() > 64)
5137 int64_t StepVal = APStepVal.getSExtValue();
5139 return StepVal > MaxMergeDistance;
5142 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5143 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5149 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5150 // If we know that this instruction will remain uniform, check the cost of
5151 // the scalar version.
5152 if (Legal->isUniformAfterVectorization(I))
5155 Type *RetTy = I->getType();
5156 Type *VectorTy = ToVectorTy(RetTy, VF);
5158 // TODO: We need to estimate the cost of intrinsic calls.
5159 switch (I->getOpcode()) {
5160 case Instruction::GetElementPtr:
5161 // We mark this instruction as zero-cost because the cost of GEPs in
5162 // vectorized code depends on whether the corresponding memory instruction
5163 // is scalarized or not. Therefore, we handle GEPs with the memory
5164 // instruction cost.
5166 case Instruction::Br: {
5167 return TTI.getCFInstrCost(I->getOpcode());
5169 case Instruction::PHI:
5170 //TODO: IF-converted IFs become selects.
5172 case Instruction::Add:
5173 case Instruction::FAdd:
5174 case Instruction::Sub:
5175 case Instruction::FSub:
5176 case Instruction::Mul:
5177 case Instruction::FMul:
5178 case Instruction::UDiv:
5179 case Instruction::SDiv:
5180 case Instruction::FDiv:
5181 case Instruction::URem:
5182 case Instruction::SRem:
5183 case Instruction::FRem:
5184 case Instruction::Shl:
5185 case Instruction::LShr:
5186 case Instruction::AShr:
5187 case Instruction::And:
5188 case Instruction::Or:
5189 case Instruction::Xor: {
5190 // Since we will replace the stride by 1 the multiplication should go away.
5191 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5193 // Certain instructions can be cheaper to vectorize if they have a constant
5194 // second vector operand. One example of this are shifts on x86.
5195 TargetTransformInfo::OperandValueKind Op1VK =
5196 TargetTransformInfo::OK_AnyValue;
5197 TargetTransformInfo::OperandValueKind Op2VK =
5198 TargetTransformInfo::OK_AnyValue;
5199 TargetTransformInfo::OperandValueProperties Op1VP =
5200 TargetTransformInfo::OP_None;
5201 TargetTransformInfo::OperandValueProperties Op2VP =
5202 TargetTransformInfo::OP_None;
5203 Value *Op2 = I->getOperand(1);
5205 // Check for a splat of a constant or for a non uniform vector of constants.
5206 if (isa<ConstantInt>(Op2)) {
5207 ConstantInt *CInt = cast<ConstantInt>(Op2);
5208 if (CInt && CInt->getValue().isPowerOf2())
5209 Op2VP = TargetTransformInfo::OP_PowerOf2;
5210 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5211 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5212 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5213 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5215 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5216 if (CInt && CInt->getValue().isPowerOf2())
5217 Op2VP = TargetTransformInfo::OP_PowerOf2;
5218 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5222 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5225 case Instruction::Select: {
5226 SelectInst *SI = cast<SelectInst>(I);
5227 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5228 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5229 Type *CondTy = SI->getCondition()->getType();
5231 CondTy = VectorType::get(CondTy, VF);
5233 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5235 case Instruction::ICmp:
5236 case Instruction::FCmp: {
5237 Type *ValTy = I->getOperand(0)->getType();
5238 VectorTy = ToVectorTy(ValTy, VF);
5239 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5241 case Instruction::Store:
5242 case Instruction::Load: {
5243 StoreInst *SI = dyn_cast<StoreInst>(I);
5244 LoadInst *LI = dyn_cast<LoadInst>(I);
5245 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5247 VectorTy = ToVectorTy(ValTy, VF);
5249 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5250 unsigned AS = SI ? SI->getPointerAddressSpace() :
5251 LI->getPointerAddressSpace();
5252 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5253 // We add the cost of address computation here instead of with the gep
5254 // instruction because only here we know whether the operation is
5257 return TTI.getAddressComputationCost(VectorTy) +
5258 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5260 // For an interleaved access, calculate the total cost of the whole
5261 // interleave group.
5262 if (Legal->isAccessInterleaved(I)) {
5263 auto Group = Legal->getInterleavedAccessGroup(I);
5264 assert(Group && "Fail to get an interleaved access group.");
5266 // Only calculate the cost once at the insert position.
5267 if (Group->getInsertPos() != I)
5270 unsigned InterleaveFactor = Group->getFactor();
5272 VectorType::get(VectorTy->getVectorElementType(),
5273 VectorTy->getVectorNumElements() * InterleaveFactor);
5275 // Holds the indices of existing members in an interleaved load group.
5276 // An interleaved store group doesn't need this as it dones't allow gaps.
5277 SmallVector<unsigned, 4> Indices;
5279 for (unsigned i = 0; i < InterleaveFactor; i++)
5280 if (Group->getMember(i))
5281 Indices.push_back(i);
5284 // Calculate the cost of the whole interleaved group.
5285 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5286 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5287 Group->getAlignment(), AS);
5289 if (Group->isReverse())
5291 Group->getNumMembers() *
5292 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5294 // FIXME: The interleaved load group with a huge gap could be even more
5295 // expensive than scalar operations. Then we could ignore such group and
5296 // use scalar operations instead.
5300 // Scalarized loads/stores.
5301 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5302 bool Reverse = ConsecutiveStride < 0;
5303 const DataLayout &DL = I->getModule()->getDataLayout();
5304 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5305 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5306 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5307 bool IsComplexComputation =
5308 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5310 // The cost of extracting from the value vector and pointer vector.
5311 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5312 for (unsigned i = 0; i < VF; ++i) {
5313 // The cost of extracting the pointer operand.
5314 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5315 // In case of STORE, the cost of ExtractElement from the vector.
5316 // In case of LOAD, the cost of InsertElement into the returned
5318 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5319 Instruction::InsertElement,
5323 // The cost of the scalar loads/stores.
5324 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5325 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5330 // Wide load/stores.
5331 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5332 if (Legal->isMaskRequired(I))
5333 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5336 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5339 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5343 case Instruction::ZExt:
5344 case Instruction::SExt:
5345 case Instruction::FPToUI:
5346 case Instruction::FPToSI:
5347 case Instruction::FPExt:
5348 case Instruction::PtrToInt:
5349 case Instruction::IntToPtr:
5350 case Instruction::SIToFP:
5351 case Instruction::UIToFP:
5352 case Instruction::Trunc:
5353 case Instruction::FPTrunc:
5354 case Instruction::BitCast: {
5355 // We optimize the truncation of induction variable.
5356 // The cost of these is the same as the scalar operation.
5357 if (I->getOpcode() == Instruction::Trunc &&
5358 Legal->isInductionVariable(I->getOperand(0)))
5359 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5360 I->getOperand(0)->getType());
5362 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5363 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5365 case Instruction::Call: {
5366 bool NeedToScalarize;
5367 CallInst *CI = cast<CallInst>(I);
5368 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5369 if (getIntrinsicIDForCall(CI, TLI))
5370 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5374 // We are scalarizing the instruction. Return the cost of the scalar
5375 // instruction, plus the cost of insert and extract into vector
5376 // elements, times the vector width.
5379 if (!RetTy->isVoidTy() && VF != 1) {
5380 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5382 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5385 // The cost of inserting the results plus extracting each one of the
5387 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5390 // The cost of executing VF copies of the scalar instruction. This opcode
5391 // is unknown. Assume that it is the same as 'mul'.
5392 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5398 char LoopVectorize::ID = 0;
5399 static const char lv_name[] = "Loop Vectorization";
5400 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5401 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5402 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5403 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5404 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5405 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5406 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5407 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5408 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5409 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5410 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5411 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5414 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5415 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5419 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5420 // Check for a store.
5421 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5422 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5424 // Check for a load.
5425 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5426 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5432 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5433 bool IfPredicateStore) {
5434 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5435 // Holds vector parameters or scalars, in case of uniform vals.
5436 SmallVector<VectorParts, 4> Params;
5438 setDebugLocFromInst(Builder, Instr);
5440 // Find all of the vectorized parameters.
5441 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5442 Value *SrcOp = Instr->getOperand(op);
5444 // If we are accessing the old induction variable, use the new one.
5445 if (SrcOp == OldInduction) {
5446 Params.push_back(getVectorValue(SrcOp));
5450 // Try using previously calculated values.
5451 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5453 // If the src is an instruction that appeared earlier in the basic block
5454 // then it should already be vectorized.
5455 if (SrcInst && OrigLoop->contains(SrcInst)) {
5456 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5457 // The parameter is a vector value from earlier.
5458 Params.push_back(WidenMap.get(SrcInst));
5460 // The parameter is a scalar from outside the loop. Maybe even a constant.
5461 VectorParts Scalars;
5462 Scalars.append(UF, SrcOp);
5463 Params.push_back(Scalars);
5467 assert(Params.size() == Instr->getNumOperands() &&
5468 "Invalid number of operands");
5470 // Does this instruction return a value ?
5471 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5473 Value *UndefVec = IsVoidRetTy ? nullptr :
5474 UndefValue::get(Instr->getType());
5475 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5476 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5478 Instruction *InsertPt = Builder.GetInsertPoint();
5479 BasicBlock *IfBlock = Builder.GetInsertBlock();
5480 BasicBlock *CondBlock = nullptr;
5483 Loop *VectorLp = nullptr;
5484 if (IfPredicateStore) {
5485 assert(Instr->getParent()->getSinglePredecessor() &&
5486 "Only support single predecessor blocks");
5487 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5488 Instr->getParent());
5489 VectorLp = LI->getLoopFor(IfBlock);
5490 assert(VectorLp && "Must have a loop for this block");
5493 // For each vector unroll 'part':
5494 for (unsigned Part = 0; Part < UF; ++Part) {
5495 // For each scalar that we create:
5497 // Start an "if (pred) a[i] = ..." block.
5498 Value *Cmp = nullptr;
5499 if (IfPredicateStore) {
5500 if (Cond[Part]->getType()->isVectorTy())
5502 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5503 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5504 ConstantInt::get(Cond[Part]->getType(), 1));
5505 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5506 LoopVectorBody.push_back(CondBlock);
5507 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5508 // Update Builder with newly created basic block.
5509 Builder.SetInsertPoint(InsertPt);
5512 Instruction *Cloned = Instr->clone();
5514 Cloned->setName(Instr->getName() + ".cloned");
5515 // Replace the operands of the cloned instructions with extracted scalars.
5516 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5517 Value *Op = Params[op][Part];
5518 Cloned->setOperand(op, Op);
5521 // Place the cloned scalar in the new loop.
5522 Builder.Insert(Cloned);
5524 // If the original scalar returns a value we need to place it in a vector
5525 // so that future users will be able to use it.
5527 VecResults[Part] = Cloned;
5530 if (IfPredicateStore) {
5531 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5532 LoopVectorBody.push_back(NewIfBlock);
5533 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5534 Builder.SetInsertPoint(InsertPt);
5535 ReplaceInstWithInst(IfBlock->getTerminator(),
5536 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5537 IfBlock = NewIfBlock;
5542 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5543 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5544 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5546 return scalarizeInstruction(Instr, IfPredicateStore);
5549 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5553 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5557 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5558 // When unrolling and the VF is 1, we only need to add a simple scalar.
5559 Type *ITy = Val->getType();
5560 assert(!ITy->isVectorTy() && "Val must be a scalar");
5561 Constant *C = ConstantInt::get(ITy, StartIdx);
5562 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");