1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/EquivalenceClasses.h"
52 #include "llvm/ADT/Hashing.h"
53 #include "llvm/ADT/MapVector.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/AliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/LoopAccessAnalysis.h"
66 #include "llvm/Analysis/LoopInfo.h"
67 #include "llvm/Analysis/LoopIterator.h"
68 #include "llvm/Analysis/LoopPass.h"
69 #include "llvm/Analysis/ScalarEvolution.h"
70 #include "llvm/Analysis/ScalarEvolutionExpander.h"
71 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
72 #include "llvm/Analysis/TargetTransformInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/IR/Constants.h"
75 #include "llvm/IR/DataLayout.h"
76 #include "llvm/IR/DebugInfo.h"
77 #include "llvm/IR/DerivedTypes.h"
78 #include "llvm/IR/DiagnosticInfo.h"
79 #include "llvm/IR/Dominators.h"
80 #include "llvm/IR/Function.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/Instructions.h"
83 #include "llvm/IR/IntrinsicInst.h"
84 #include "llvm/IR/LLVMContext.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/IR/Type.h"
88 #include "llvm/IR/Value.h"
89 #include "llvm/IR/ValueHandle.h"
90 #include "llvm/IR/Verifier.h"
91 #include "llvm/Pass.h"
92 #include "llvm/Support/BranchProbability.h"
93 #include "llvm/Support/CommandLine.h"
94 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/raw_ostream.h"
96 #include "llvm/Transforms/Scalar.h"
97 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
98 #include "llvm/Transforms/Utils/Local.h"
99 #include "llvm/Analysis/VectorUtils.h"
100 #include "llvm/Transforms/Utils/LoopUtils.h"
105 using namespace llvm;
106 using namespace llvm::PatternMatch;
108 #define LV_NAME "loop-vectorize"
109 #define DEBUG_TYPE LV_NAME
111 STATISTIC(LoopsVectorized, "Number of loops vectorized");
112 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
115 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
116 cl::desc("Enable if-conversion during vectorization."));
118 /// We don't vectorize loops with a known constant trip count below this number.
119 static cl::opt<unsigned>
120 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
122 cl::desc("Don't vectorize loops with a constant "
123 "trip count that is smaller than this "
126 /// This enables versioning on the strides of symbolically striding memory
127 /// accesses in code like the following.
128 /// for (i = 0; i < N; ++i)
129 /// A[i * Stride1] += B[i * Stride2] ...
131 /// Will be roughly translated to
132 /// if (Stride1 == 1 && Stride2 == 1) {
133 /// for (i = 0; i < N; i+=4)
137 static cl::opt<bool> EnableMemAccessVersioning(
138 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
139 cl::desc("Enable symblic stride memory access versioning"));
141 static cl::opt<bool> EnableInterleavedMemAccesses(
142 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
143 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
145 /// Maximum factor for an interleaved memory access.
146 static cl::opt<unsigned> MaxInterleaveGroupFactor(
147 "max-interleave-group-factor", cl::Hidden,
148 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
151 /// We don't interleave loops with a known constant trip count below this
153 static const unsigned TinyTripCountInterleaveThreshold = 128;
155 static cl::opt<unsigned> ForceTargetNumScalarRegs(
156 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of scalar registers."));
159 static cl::opt<unsigned> ForceTargetNumVectorRegs(
160 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's number of vector registers."));
163 /// Maximum vectorization interleave count.
164 static const unsigned MaxInterleaveFactor = 16;
166 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
167 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max interleave factor for "
171 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
172 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
174 "vectorized loops."));
176 static cl::opt<unsigned> ForceTargetInstructionCost(
177 "force-target-instruction-cost", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's expected cost for "
179 "an instruction to a single constant value. Mostly "
180 "useful for getting consistent testing."));
182 static cl::opt<unsigned> SmallLoopCost(
183 "small-loop-cost", cl::init(20), cl::Hidden,
185 "The cost of a loop that is considered 'small' by the interleaver."));
187 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
188 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
189 cl::desc("Enable the use of the block frequency analysis to access PGO "
190 "heuristics minimizing code growth in cold regions and being more "
191 "aggressive in hot regions."));
193 // Runtime interleave loops for load/store throughput.
194 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
195 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
197 "Enable runtime interleaving until load/store ports are saturated"));
199 /// The number of stores in a loop that are allowed to need predication.
200 static cl::opt<unsigned> NumberOfStoresToPredicate(
201 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
202 cl::desc("Max number of stores to be predicated behind an if."));
204 static cl::opt<bool> EnableIndVarRegisterHeur(
205 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
206 cl::desc("Count the induction variable only once when interleaving"));
208 static cl::opt<bool> EnableCondStoresVectorization(
209 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
210 cl::desc("Enable if predication of stores during vectorization."));
212 static cl::opt<unsigned> MaxNestedScalarReductionIC(
213 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
214 cl::desc("The maximum interleave count to use when interleaving a scalar "
215 "reduction in a nested loop."));
219 // Forward declarations.
220 class LoopVectorizeHints;
221 class LoopVectorizationLegality;
222 class LoopVectorizationCostModel;
223 class LoopVectorizationRequirements;
225 /// \brief This modifies LoopAccessReport to initialize message with
226 /// loop-vectorizer-specific part.
227 class VectorizationReport : public LoopAccessReport {
229 VectorizationReport(Instruction *I = nullptr)
230 : LoopAccessReport("loop not vectorized: ", I) {}
232 /// \brief This allows promotion of the loop-access analysis report into the
233 /// loop-vectorizer report. It modifies the message to add the
234 /// loop-vectorizer-specific part of the message.
235 explicit VectorizationReport(const LoopAccessReport &R)
236 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
240 /// A helper function for converting Scalar types to vector types.
241 /// If the incoming type is void, we return void. If the VF is 1, we return
243 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
244 if (Scalar->isVoidTy() || VF == 1)
246 return VectorType::get(Scalar, VF);
249 /// InnerLoopVectorizer vectorizes loops which contain only one basic
250 /// block to a specified vectorization factor (VF).
251 /// This class performs the widening of scalars into vectors, or multiple
252 /// scalars. This class also implements the following features:
253 /// * It inserts an epilogue loop for handling loops that don't have iteration
254 /// counts that are known to be a multiple of the vectorization factor.
255 /// * It handles the code generation for reduction variables.
256 /// * Scalarization (implementation using scalars) of un-vectorizable
258 /// InnerLoopVectorizer does not perform any vectorization-legality
259 /// checks, and relies on the caller to check for the different legality
260 /// aspects. The InnerLoopVectorizer relies on the
261 /// LoopVectorizationLegality class to provide information about the induction
262 /// and reduction variables that were found to a given vectorization factor.
263 class InnerLoopVectorizer {
265 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
266 DominatorTree *DT, const TargetLibraryInfo *TLI,
267 const TargetTransformInfo *TTI, unsigned VecWidth,
268 unsigned UnrollFactor)
269 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
270 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
271 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
272 Legal(nullptr), AddedSafetyChecks(false) {}
274 // Perform the actual loop widening (vectorization).
275 void vectorize(LoopVectorizationLegality *L) {
277 // Create a new empty loop. Unlink the old loop and connect the new one.
279 // Widen each instruction in the old loop to a new one in the new loop.
280 // Use the Legality module to find the induction and reduction variables.
282 // Register the new loop and update the analysis passes.
286 // Return true if any runtime check is added.
287 bool IsSafetyChecksAdded() {
288 return AddedSafetyChecks;
291 virtual ~InnerLoopVectorizer() {}
294 /// A small list of PHINodes.
295 typedef SmallVector<PHINode*, 4> PhiVector;
296 /// When we unroll loops we have multiple vector values for each scalar.
297 /// This data structure holds the unrolled and vectorized values that
298 /// originated from one scalar instruction.
299 typedef SmallVector<Value*, 2> VectorParts;
301 // When we if-convert we need to create edge masks. We have to cache values
302 // so that we don't end up with exponential recursion/IR.
303 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
304 VectorParts> EdgeMaskCache;
306 /// \brief Add checks for strides that were assumed to be 1.
308 /// Returns the last check instruction and the first check instruction in the
309 /// pair as (first, last).
310 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
312 /// Create an empty loop, based on the loop ranges of the old loop.
313 void createEmptyLoop();
314 /// Copy and widen the instructions from the old loop.
315 virtual void vectorizeLoop();
317 /// \brief The Loop exit block may have single value PHI nodes where the
318 /// incoming value is 'Undef'. While vectorizing we only handled real values
319 /// that were defined inside the loop. Here we fix the 'undef case'.
323 /// A helper function that computes the predicate of the block BB, assuming
324 /// that the header block of the loop is set to True. It returns the *entry*
325 /// mask for the block BB.
326 VectorParts createBlockInMask(BasicBlock *BB);
327 /// A helper function that computes the predicate of the edge between SRC
329 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
331 /// A helper function to vectorize a single BB within the innermost loop.
332 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
334 /// Vectorize a single PHINode in a block. This method handles the induction
335 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
336 /// arbitrary length vectors.
337 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
338 unsigned UF, unsigned VF, PhiVector *PV);
340 /// Insert the new loop to the loop hierarchy and pass manager
341 /// and update the analysis passes.
342 void updateAnalysis();
344 /// This instruction is un-vectorizable. Implement it as a sequence
345 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
346 /// scalarized instruction behind an if block predicated on the control
347 /// dependence of the instruction.
348 virtual void scalarizeInstruction(Instruction *Instr,
349 bool IfPredicateStore=false);
351 /// Vectorize Load and Store instructions,
352 virtual void vectorizeMemoryInstruction(Instruction *Instr);
354 /// Create a broadcast instruction. This method generates a broadcast
355 /// instruction (shuffle) for loop invariant values and for the induction
356 /// value. If this is the induction variable then we extend it to N, N+1, ...
357 /// this is needed because each iteration in the loop corresponds to a SIMD
359 virtual Value *getBroadcastInstrs(Value *V);
361 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
362 /// to each vector element of Val. The sequence starts at StartIndex.
363 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
365 /// When we go over instructions in the basic block we rely on previous
366 /// values within the current basic block or on loop invariant values.
367 /// When we widen (vectorize) values we place them in the map. If the values
368 /// are not within the map, they have to be loop invariant, so we simply
369 /// broadcast them into a vector.
370 VectorParts &getVectorValue(Value *V);
372 /// Try to vectorize the interleaved access group that \p Instr belongs to.
373 void vectorizeInterleaveGroup(Instruction *Instr);
375 /// Generate a shuffle sequence that will reverse the vector Vec.
376 virtual Value *reverseVector(Value *Vec);
378 /// This is a helper class that holds the vectorizer state. It maps scalar
379 /// instructions to vector instructions. When the code is 'unrolled' then
380 /// then a single scalar value is mapped to multiple vector parts. The parts
381 /// are stored in the VectorPart type.
383 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
385 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
387 /// \return True if 'Key' is saved in the Value Map.
388 bool has(Value *Key) const { return MapStorage.count(Key); }
390 /// Initializes a new entry in the map. Sets all of the vector parts to the
391 /// save value in 'Val'.
392 /// \return A reference to a vector with splat values.
393 VectorParts &splat(Value *Key, Value *Val) {
394 VectorParts &Entry = MapStorage[Key];
395 Entry.assign(UF, Val);
399 ///\return A reference to the value that is stored at 'Key'.
400 VectorParts &get(Value *Key) {
401 VectorParts &Entry = MapStorage[Key];
404 assert(Entry.size() == UF);
409 /// The unroll factor. Each entry in the map stores this number of vector
413 /// Map storage. We use std::map and not DenseMap because insertions to a
414 /// dense map invalidates its iterators.
415 std::map<Value *, VectorParts> MapStorage;
418 /// The original loop.
420 /// Scev analysis to use.
428 /// Target Library Info.
429 const TargetLibraryInfo *TLI;
430 /// Target Transform Info.
431 const TargetTransformInfo *TTI;
433 /// The vectorization SIMD factor to use. Each vector will have this many
438 /// The vectorization unroll factor to use. Each scalar is vectorized to this
439 /// many different vector instructions.
442 /// The builder that we use
445 // --- Vectorization state ---
447 /// The vector-loop preheader.
448 BasicBlock *LoopVectorPreHeader;
449 /// The scalar-loop preheader.
450 BasicBlock *LoopScalarPreHeader;
451 /// Middle Block between the vector and the scalar.
452 BasicBlock *LoopMiddleBlock;
453 ///The ExitBlock of the scalar loop.
454 BasicBlock *LoopExitBlock;
455 ///The vector loop body.
456 SmallVector<BasicBlock *, 4> LoopVectorBody;
457 ///The scalar loop body.
458 BasicBlock *LoopScalarBody;
459 /// A list of all bypass blocks. The first block is the entry of the loop.
460 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
462 /// The new Induction variable which was added to the new block.
464 /// The induction variable of the old basic block.
465 PHINode *OldInduction;
466 /// Holds the extended (to the widest induction type) start index.
468 /// Maps scalars to widened vectors.
470 EdgeMaskCache MaskCache;
472 LoopVectorizationLegality *Legal;
474 // Record whether runtime check is added.
475 bool AddedSafetyChecks;
478 class InnerLoopUnroller : public InnerLoopVectorizer {
480 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
481 DominatorTree *DT, const TargetLibraryInfo *TLI,
482 const TargetTransformInfo *TTI, unsigned UnrollFactor)
483 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
486 void scalarizeInstruction(Instruction *Instr,
487 bool IfPredicateStore = false) override;
488 void vectorizeMemoryInstruction(Instruction *Instr) override;
489 Value *getBroadcastInstrs(Value *V) override;
490 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
491 Value *reverseVector(Value *Vec) override;
494 /// \brief Look for a meaningful debug location on the instruction or it's
496 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
501 if (I->getDebugLoc() != Empty)
504 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
505 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
506 if (OpInst->getDebugLoc() != Empty)
513 /// \brief Set the debug location in the builder using the debug location in the
515 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
516 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
517 B.SetCurrentDebugLocation(Inst->getDebugLoc());
519 B.SetCurrentDebugLocation(DebugLoc());
523 /// \return string containing a file name and a line # for the given loop.
524 static std::string getDebugLocString(const Loop *L) {
527 raw_string_ostream OS(Result);
528 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
529 LoopDbgLoc.print(OS);
531 // Just print the module name.
532 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
539 /// \brief Propagate known metadata from one instruction to another.
540 static void propagateMetadata(Instruction *To, const Instruction *From) {
541 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
542 From->getAllMetadataOtherThanDebugLoc(Metadata);
544 for (auto M : Metadata) {
545 unsigned Kind = M.first;
547 // These are safe to transfer (this is safe for TBAA, even when we
548 // if-convert, because should that metadata have had a control dependency
549 // on the condition, and thus actually aliased with some other
550 // non-speculated memory access when the condition was false, this would be
551 // caught by the runtime overlap checks).
552 if (Kind != LLVMContext::MD_tbaa &&
553 Kind != LLVMContext::MD_alias_scope &&
554 Kind != LLVMContext::MD_noalias &&
555 Kind != LLVMContext::MD_fpmath &&
556 Kind != LLVMContext::MD_nontemporal)
559 To->setMetadata(Kind, M.second);
563 /// \brief Propagate known metadata from one instruction to a vector of others.
564 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
566 if (Instruction *I = dyn_cast<Instruction>(V))
567 propagateMetadata(I, From);
570 /// \brief The group of interleaved loads/stores sharing the same stride and
571 /// close to each other.
573 /// Each member in this group has an index starting from 0, and the largest
574 /// index should be less than interleaved factor, which is equal to the absolute
575 /// value of the access's stride.
577 /// E.g. An interleaved load group of factor 4:
578 /// for (unsigned i = 0; i < 1024; i+=4) {
579 /// a = A[i]; // Member of index 0
580 /// b = A[i+1]; // Member of index 1
581 /// d = A[i+3]; // Member of index 3
585 /// An interleaved store group of factor 4:
586 /// for (unsigned i = 0; i < 1024; i+=4) {
588 /// A[i] = a; // Member of index 0
589 /// A[i+1] = b; // Member of index 1
590 /// A[i+2] = c; // Member of index 2
591 /// A[i+3] = d; // Member of index 3
594 /// Note: the interleaved load group could have gaps (missing members), but
595 /// the interleaved store group doesn't allow gaps.
596 class InterleaveGroup {
598 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
599 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
600 assert(Align && "The alignment should be non-zero");
602 Factor = std::abs(Stride);
603 assert(Factor > 1 && "Invalid interleave factor");
605 Reverse = Stride < 0;
609 bool isReverse() const { return Reverse; }
610 unsigned getFactor() const { return Factor; }
611 unsigned getAlignment() const { return Align; }
612 unsigned getNumMembers() const { return Members.size(); }
614 /// \brief Try to insert a new member \p Instr with index \p Index and
615 /// alignment \p NewAlign. The index is related to the leader and it could be
616 /// negative if it is the new leader.
618 /// \returns false if the instruction doesn't belong to the group.
619 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
620 assert(NewAlign && "The new member's alignment should be non-zero");
622 int Key = Index + SmallestKey;
624 // Skip if there is already a member with the same index.
625 if (Members.count(Key))
628 if (Key > LargestKey) {
629 // The largest index is always less than the interleave factor.
630 if (Index >= static_cast<int>(Factor))
634 } else if (Key < SmallestKey) {
635 // The largest index is always less than the interleave factor.
636 if (LargestKey - Key >= static_cast<int>(Factor))
642 // It's always safe to select the minimum alignment.
643 Align = std::min(Align, NewAlign);
644 Members[Key] = Instr;
648 /// \brief Get the member with the given index \p Index
650 /// \returns nullptr if contains no such member.
651 Instruction *getMember(unsigned Index) const {
652 int Key = SmallestKey + Index;
653 if (!Members.count(Key))
656 return Members.find(Key)->second;
659 /// \brief Get the index for the given member. Unlike the key in the member
660 /// map, the index starts from 0.
661 unsigned getIndex(Instruction *Instr) const {
662 for (auto I : Members)
663 if (I.second == Instr)
664 return I.first - SmallestKey;
666 llvm_unreachable("InterleaveGroup contains no such member");
669 Instruction *getInsertPos() const { return InsertPos; }
670 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
673 unsigned Factor; // Interleave Factor.
676 DenseMap<int, Instruction *> Members;
680 // To avoid breaking dependences, vectorized instructions of an interleave
681 // group should be inserted at either the first load or the last store in
684 // E.g. %even = load i32 // Insert Position
685 // %add = add i32 %even // Use of %even
689 // %odd = add i32 // Def of %odd
690 // store i32 %odd // Insert Position
691 Instruction *InsertPos;
694 /// \brief Drive the analysis of interleaved memory accesses in the loop.
696 /// Use this class to analyze interleaved accesses only when we can vectorize
697 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
698 /// on interleaved accesses is unsafe.
700 /// The analysis collects interleave groups and records the relationships
701 /// between the member and the group in a map.
702 class InterleavedAccessInfo {
704 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
705 : SE(SE), TheLoop(L), DT(DT) {}
707 ~InterleavedAccessInfo() {
708 SmallSet<InterleaveGroup *, 4> DelSet;
709 // Avoid releasing a pointer twice.
710 for (auto &I : InterleaveGroupMap)
711 DelSet.insert(I.second);
712 for (auto *Ptr : DelSet)
716 /// \brief Analyze the interleaved accesses and collect them in interleave
717 /// groups. Substitute symbolic strides using \p Strides.
718 void analyzeInterleaving(const ValueToValueMap &Strides);
720 /// \brief Check if \p Instr belongs to any interleave group.
721 bool isInterleaved(Instruction *Instr) const {
722 return InterleaveGroupMap.count(Instr);
725 /// \brief Get the interleave group that \p Instr belongs to.
727 /// \returns nullptr if doesn't have such group.
728 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
729 if (InterleaveGroupMap.count(Instr))
730 return InterleaveGroupMap.find(Instr)->second;
739 /// Holds the relationships between the members and the interleave group.
740 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
742 /// \brief The descriptor for a strided memory access.
743 struct StrideDescriptor {
744 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
746 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
748 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
750 int Stride; // The access's stride. It is negative for a reverse access.
751 const SCEV *Scev; // The scalar expression of this access
752 unsigned Size; // The size of the memory object.
753 unsigned Align; // The alignment of this access.
756 /// \brief Create a new interleave group with the given instruction \p Instr,
757 /// stride \p Stride and alignment \p Align.
759 /// \returns the newly created interleave group.
760 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
762 assert(!InterleaveGroupMap.count(Instr) &&
763 "Already in an interleaved access group");
764 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
765 return InterleaveGroupMap[Instr];
768 /// \brief Release the group and remove all the relationships.
769 void releaseGroup(InterleaveGroup *Group) {
770 for (unsigned i = 0; i < Group->getFactor(); i++)
771 if (Instruction *Member = Group->getMember(i))
772 InterleaveGroupMap.erase(Member);
777 /// \brief Collect all the accesses with a constant stride in program order.
778 void collectConstStridedAccesses(
779 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
780 const ValueToValueMap &Strides);
783 /// Utility class for getting and setting loop vectorizer hints in the form
784 /// of loop metadata.
785 /// This class keeps a number of loop annotations locally (as member variables)
786 /// and can, upon request, write them back as metadata on the loop. It will
787 /// initially scan the loop for existing metadata, and will update the local
788 /// values based on information in the loop.
789 /// We cannot write all values to metadata, as the mere presence of some info,
790 /// for example 'force', means a decision has been made. So, we need to be
791 /// careful NOT to add them if the user hasn't specifically asked so.
792 class LoopVectorizeHints {
799 /// Hint - associates name and validation with the hint value.
802 unsigned Value; // This may have to change for non-numeric values.
805 Hint(const char * Name, unsigned Value, HintKind Kind)
806 : Name(Name), Value(Value), Kind(Kind) { }
808 bool validate(unsigned Val) {
811 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
813 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
821 /// Vectorization width.
823 /// Vectorization interleave factor.
825 /// Vectorization forced
828 /// Return the loop metadata prefix.
829 static StringRef Prefix() { return "llvm.loop."; }
833 FK_Undefined = -1, ///< Not selected.
834 FK_Disabled = 0, ///< Forcing disabled.
835 FK_Enabled = 1, ///< Forcing enabled.
838 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
839 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
841 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
842 Force("vectorize.enable", FK_Undefined, HK_FORCE),
844 // Populate values with existing loop metadata.
845 getHintsFromMetadata();
847 // force-vector-interleave overrides DisableInterleaving.
848 if (VectorizerParams::isInterleaveForced())
849 Interleave.Value = VectorizerParams::VectorizationInterleave;
851 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
852 << "LV: Interleaving disabled by the pass manager\n");
855 /// Mark the loop L as already vectorized by setting the width to 1.
856 void setAlreadyVectorized() {
857 Width.Value = Interleave.Value = 1;
858 Hint Hints[] = {Width, Interleave};
859 writeHintsToMetadata(Hints);
862 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
863 if (getForce() == LoopVectorizeHints::FK_Disabled) {
864 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
865 emitOptimizationRemarkAnalysis(F->getContext(),
866 vectorizeAnalysisPassName(), *F,
867 L->getStartLoc(), emitRemark());
871 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
872 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
873 emitOptimizationRemarkAnalysis(F->getContext(),
874 vectorizeAnalysisPassName(), *F,
875 L->getStartLoc(), emitRemark());
879 if (getWidth() == 1 && getInterleave() == 1) {
880 // FIXME: Add a separate metadata to indicate when the loop has already
881 // been vectorized instead of setting width and count to 1.
882 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
883 // FIXME: Add interleave.disable metadata. This will allow
884 // vectorize.disable to be used without disabling the pass and errors
885 // to differentiate between disabled vectorization and a width of 1.
886 emitOptimizationRemarkAnalysis(
887 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
888 "loop not vectorized: vectorization and interleaving are explicitly "
889 "disabled, or vectorize width and interleave count are both set to "
897 /// Dumps all the hint information.
898 std::string emitRemark() const {
899 VectorizationReport R;
900 if (Force.Value == LoopVectorizeHints::FK_Disabled)
901 R << "vectorization is explicitly disabled";
903 R << "use -Rpass-analysis=loop-vectorize for more info";
904 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
906 if (Width.Value != 0)
907 R << ", Vector Width=" << Width.Value;
908 if (Interleave.Value != 0)
909 R << ", Interleave Count=" << Interleave.Value;
917 unsigned getWidth() const { return Width.Value; }
918 unsigned getInterleave() const { return Interleave.Value; }
919 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
920 const char *vectorizeAnalysisPassName() const {
921 // If hints are provided that don't disable vectorization use the
922 // AlwaysPrint pass name to force the frontend to print the diagnostic.
925 if (getForce() == LoopVectorizeHints::FK_Disabled)
927 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
929 return DiagnosticInfo::AlwaysPrint;
933 /// Find hints specified in the loop metadata and update local values.
934 void getHintsFromMetadata() {
935 MDNode *LoopID = TheLoop->getLoopID();
939 // First operand should refer to the loop id itself.
940 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
941 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
943 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
944 const MDString *S = nullptr;
945 SmallVector<Metadata *, 4> Args;
947 // The expected hint is either a MDString or a MDNode with the first
948 // operand a MDString.
949 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
950 if (!MD || MD->getNumOperands() == 0)
952 S = dyn_cast<MDString>(MD->getOperand(0));
953 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
954 Args.push_back(MD->getOperand(i));
956 S = dyn_cast<MDString>(LoopID->getOperand(i));
957 assert(Args.size() == 0 && "too many arguments for MDString");
963 // Check if the hint starts with the loop metadata prefix.
964 StringRef Name = S->getString();
965 if (Args.size() == 1)
966 setHint(Name, Args[0]);
970 /// Checks string hint with one operand and set value if valid.
971 void setHint(StringRef Name, Metadata *Arg) {
972 if (!Name.startswith(Prefix()))
974 Name = Name.substr(Prefix().size(), StringRef::npos);
976 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
978 unsigned Val = C->getZExtValue();
980 Hint *Hints[] = {&Width, &Interleave, &Force};
981 for (auto H : Hints) {
982 if (Name == H->Name) {
983 if (H->validate(Val))
986 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
992 /// Create a new hint from name / value pair.
993 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
994 LLVMContext &Context = TheLoop->getHeader()->getContext();
995 Metadata *MDs[] = {MDString::get(Context, Name),
996 ConstantAsMetadata::get(
997 ConstantInt::get(Type::getInt32Ty(Context), V))};
998 return MDNode::get(Context, MDs);
1001 /// Matches metadata with hint name.
1002 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1003 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1007 for (auto H : HintTypes)
1008 if (Name->getString().endswith(H.Name))
1013 /// Sets current hints into loop metadata, keeping other values intact.
1014 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1015 if (HintTypes.size() == 0)
1018 // Reserve the first element to LoopID (see below).
1019 SmallVector<Metadata *, 4> MDs(1);
1020 // If the loop already has metadata, then ignore the existing operands.
1021 MDNode *LoopID = TheLoop->getLoopID();
1023 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1024 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1025 // If node in update list, ignore old value.
1026 if (!matchesHintMetadataName(Node, HintTypes))
1027 MDs.push_back(Node);
1031 // Now, add the missing hints.
1032 for (auto H : HintTypes)
1033 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1035 // Replace current metadata node with new one.
1036 LLVMContext &Context = TheLoop->getHeader()->getContext();
1037 MDNode *NewLoopID = MDNode::get(Context, MDs);
1038 // Set operand 0 to refer to the loop id itself.
1039 NewLoopID->replaceOperandWith(0, NewLoopID);
1041 TheLoop->setLoopID(NewLoopID);
1044 /// The loop these hints belong to.
1045 const Loop *TheLoop;
1048 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1049 const LoopVectorizeHints &Hints,
1050 const LoopAccessReport &Message) {
1051 const char *Name = Hints.vectorizeAnalysisPassName();
1052 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1055 static void emitMissedWarning(Function *F, Loop *L,
1056 const LoopVectorizeHints &LH) {
1057 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1060 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1061 if (LH.getWidth() != 1)
1062 emitLoopVectorizeWarning(
1063 F->getContext(), *F, L->getStartLoc(),
1064 "failed explicitly specified loop vectorization");
1065 else if (LH.getInterleave() != 1)
1066 emitLoopInterleaveWarning(
1067 F->getContext(), *F, L->getStartLoc(),
1068 "failed explicitly specified loop interleaving");
1072 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1073 /// to what vectorization factor.
1074 /// This class does not look at the profitability of vectorization, only the
1075 /// legality. This class has two main kinds of checks:
1076 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1077 /// will change the order of memory accesses in a way that will change the
1078 /// correctness of the program.
1079 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1080 /// checks for a number of different conditions, such as the availability of a
1081 /// single induction variable, that all types are supported and vectorize-able,
1082 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1083 /// This class is also used by InnerLoopVectorizer for identifying
1084 /// induction variable and the different reduction variables.
1085 class LoopVectorizationLegality {
1087 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1088 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1089 Function *F, const TargetTransformInfo *TTI,
1090 LoopAccessAnalysis *LAA,
1091 LoopVectorizationRequirements *R,
1092 const LoopVectorizeHints *H)
1093 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1094 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
1095 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1096 Requirements(R), Hints(H) {}
1098 /// This enum represents the kinds of inductions that we support.
1099 enum InductionKind {
1100 IK_NoInduction, ///< Not an induction variable.
1101 IK_IntInduction, ///< Integer induction variable. Step = C.
1102 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
1105 /// A struct for saving information about induction variables.
1106 struct InductionInfo {
1107 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
1108 : StartValue(Start), IK(K), StepValue(Step) {
1109 assert(IK != IK_NoInduction && "Not an induction");
1110 assert(StartValue && "StartValue is null");
1111 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
1112 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
1113 "StartValue is not a pointer for pointer induction");
1114 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
1115 "StartValue is not an integer for integer induction");
1116 assert(StepValue->getType()->isIntegerTy() &&
1117 "StepValue is not an integer");
1120 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
1122 /// Get the consecutive direction. Returns:
1123 /// 0 - unknown or non-consecutive.
1124 /// 1 - consecutive and increasing.
1125 /// -1 - consecutive and decreasing.
1126 int getConsecutiveDirection() const {
1127 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
1128 return StepValue->getSExtValue();
1132 /// Compute the transformed value of Index at offset StartValue using step
1134 /// For integer induction, returns StartValue + Index * StepValue.
1135 /// For pointer induction, returns StartValue[Index * StepValue].
1136 /// FIXME: The newly created binary instructions should contain nsw/nuw
1137 /// flags, which can be found from the original scalar operations.
1138 Value *transform(IRBuilder<> &B, Value *Index) const {
1140 case IK_IntInduction:
1141 assert(Index->getType() == StartValue->getType() &&
1142 "Index type does not match StartValue type");
1143 if (StepValue->isMinusOne())
1144 return B.CreateSub(StartValue, Index);
1145 if (!StepValue->isOne())
1146 Index = B.CreateMul(Index, StepValue);
1147 return B.CreateAdd(StartValue, Index);
1149 case IK_PtrInduction:
1150 assert(Index->getType() == StepValue->getType() &&
1151 "Index type does not match StepValue type");
1152 if (StepValue->isMinusOne())
1153 Index = B.CreateNeg(Index);
1154 else if (!StepValue->isOne())
1155 Index = B.CreateMul(Index, StepValue);
1156 return B.CreateGEP(nullptr, StartValue, Index);
1158 case IK_NoInduction:
1161 llvm_unreachable("invalid enum");
1165 TrackingVH<Value> StartValue;
1169 ConstantInt *StepValue;
1172 /// ReductionList contains the reduction descriptors for all
1173 /// of the reductions that were found in the loop.
1174 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1176 /// InductionList saves induction variables and maps them to the
1177 /// induction descriptor.
1178 typedef MapVector<PHINode*, InductionInfo> InductionList;
1180 /// Returns true if it is legal to vectorize this loop.
1181 /// This does not mean that it is profitable to vectorize this
1182 /// loop, only that it is legal to do so.
1183 bool canVectorize();
1185 /// Returns the Induction variable.
1186 PHINode *getInduction() { return Induction; }
1188 /// Returns the reduction variables found in the loop.
1189 ReductionList *getReductionVars() { return &Reductions; }
1191 /// Returns the induction variables found in the loop.
1192 InductionList *getInductionVars() { return &Inductions; }
1194 /// Returns the widest induction type.
1195 Type *getWidestInductionType() { return WidestIndTy; }
1197 /// Returns True if V is an induction variable in this loop.
1198 bool isInductionVariable(const Value *V);
1200 /// Return true if the block BB needs to be predicated in order for the loop
1201 /// to be vectorized.
1202 bool blockNeedsPredication(BasicBlock *BB);
1204 /// Check if this pointer is consecutive when vectorizing. This happens
1205 /// when the last index of the GEP is the induction variable, or that the
1206 /// pointer itself is an induction variable.
1207 /// This check allows us to vectorize A[idx] into a wide load/store.
1209 /// 0 - Stride is unknown or non-consecutive.
1210 /// 1 - Address is consecutive.
1211 /// -1 - Address is consecutive, and decreasing.
1212 int isConsecutivePtr(Value *Ptr);
1214 /// Returns true if the value V is uniform within the loop.
1215 bool isUniform(Value *V);
1217 /// Returns true if this instruction will remain scalar after vectorization.
1218 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1220 /// Returns the information that we collected about runtime memory check.
1221 const RuntimePointerChecking *getRuntimePointerChecking() const {
1222 return LAI->getRuntimePointerChecking();
1225 const LoopAccessInfo *getLAI() const {
1229 /// \brief Check if \p Instr belongs to any interleaved access group.
1230 bool isAccessInterleaved(Instruction *Instr) {
1231 return InterleaveInfo.isInterleaved(Instr);
1234 /// \brief Get the interleaved access group that \p Instr belongs to.
1235 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1236 return InterleaveInfo.getInterleaveGroup(Instr);
1239 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1241 bool hasStride(Value *V) { return StrideSet.count(V); }
1242 bool mustCheckStrides() { return !StrideSet.empty(); }
1243 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1244 return StrideSet.begin();
1246 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1248 /// Returns true if the target machine supports masked store operation
1249 /// for the given \p DataType and kind of access to \p Ptr.
1250 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1251 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
1253 /// Returns true if the target machine supports masked load operation
1254 /// for the given \p DataType and kind of access to \p Ptr.
1255 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1256 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
1258 /// Returns true if vector representation of the instruction \p I
1260 bool isMaskRequired(const Instruction* I) {
1261 return (MaskedOp.count(I) != 0);
1263 unsigned getNumStores() const {
1264 return LAI->getNumStores();
1266 unsigned getNumLoads() const {
1267 return LAI->getNumLoads();
1269 unsigned getNumPredStores() const {
1270 return NumPredStores;
1273 /// Check if a single basic block loop is vectorizable.
1274 /// At this point we know that this is a loop with a constant trip count
1275 /// and we only need to check individual instructions.
1276 bool canVectorizeInstrs();
1278 /// When we vectorize loops we may change the order in which
1279 /// we read and write from memory. This method checks if it is
1280 /// legal to vectorize the code, considering only memory constrains.
1281 /// Returns true if the loop is vectorizable
1282 bool canVectorizeMemory();
1284 /// Return true if we can vectorize this loop using the IF-conversion
1286 bool canVectorizeWithIfConvert();
1288 /// Collect the variables that need to stay uniform after vectorization.
1289 void collectLoopUniforms();
1291 /// Return true if all of the instructions in the block can be speculatively
1292 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1293 /// and we know that we can read from them without segfault.
1294 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1296 /// Returns the induction kind of Phi and record the step. This function may
1297 /// return NoInduction if the PHI is not an induction variable.
1298 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
1300 /// \brief Collect memory access with loop invariant strides.
1302 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1304 void collectStridedAccess(Value *LoadOrStoreInst);
1306 /// Report an analysis message to assist the user in diagnosing loops that are
1307 /// not vectorized. These are handled as LoopAccessReport rather than
1308 /// VectorizationReport because the << operator of VectorizationReport returns
1309 /// LoopAccessReport.
1310 void emitAnalysis(const LoopAccessReport &Message) const {
1311 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1314 unsigned NumPredStores;
1316 /// The loop that we evaluate.
1319 ScalarEvolution *SE;
1320 /// Target Library Info.
1321 TargetLibraryInfo *TLI;
1323 Function *TheFunction;
1324 /// Target Transform Info
1325 const TargetTransformInfo *TTI;
1328 // LoopAccess analysis.
1329 LoopAccessAnalysis *LAA;
1330 // And the loop-accesses info corresponding to this loop. This pointer is
1331 // null until canVectorizeMemory sets it up.
1332 const LoopAccessInfo *LAI;
1334 /// The interleave access information contains groups of interleaved accesses
1335 /// with the same stride and close to each other.
1336 InterleavedAccessInfo InterleaveInfo;
1338 // --- vectorization state --- //
1340 /// Holds the integer induction variable. This is the counter of the
1343 /// Holds the reduction variables.
1344 ReductionList Reductions;
1345 /// Holds all of the induction variables that we found in the loop.
1346 /// Notice that inductions don't need to start at zero and that induction
1347 /// variables can be pointers.
1348 InductionList Inductions;
1349 /// Holds the widest induction type encountered.
1352 /// Allowed outside users. This holds the reduction
1353 /// vars which can be accessed from outside the loop.
1354 SmallPtrSet<Value*, 4> AllowedExit;
1355 /// This set holds the variables which are known to be uniform after
1357 SmallPtrSet<Instruction*, 4> Uniforms;
1359 /// Can we assume the absence of NaNs.
1360 bool HasFunNoNaNAttr;
1362 /// Vectorization requirements that will go through late-evaluation.
1363 LoopVectorizationRequirements *Requirements;
1365 /// Used to emit an analysis of any legality issues.
1366 const LoopVectorizeHints *Hints;
1368 ValueToValueMap Strides;
1369 SmallPtrSet<Value *, 8> StrideSet;
1371 /// While vectorizing these instructions we have to generate a
1372 /// call to the appropriate masked intrinsic
1373 SmallPtrSet<const Instruction*, 8> MaskedOp;
1376 /// LoopVectorizationCostModel - estimates the expected speedups due to
1378 /// In many cases vectorization is not profitable. This can happen because of
1379 /// a number of reasons. In this class we mainly attempt to predict the
1380 /// expected speedup/slowdowns due to the supported instruction set. We use the
1381 /// TargetTransformInfo to query the different backends for the cost of
1382 /// different operations.
1383 class LoopVectorizationCostModel {
1385 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1386 LoopVectorizationLegality *Legal,
1387 const TargetTransformInfo &TTI,
1388 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1389 const Function *F, const LoopVectorizeHints *Hints)
1390 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1391 TheFunction(F), Hints(Hints) {
1392 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
1395 /// Information about vectorization costs
1396 struct VectorizationFactor {
1397 unsigned Width; // Vector width with best cost
1398 unsigned Cost; // Cost of the loop with that width
1400 /// \return The most profitable vectorization factor and the cost of that VF.
1401 /// This method checks every power of two up to VF. If UserVF is not ZERO
1402 /// then this vectorization factor will be selected if vectorization is
1404 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1406 /// \return The size (in bits) of the widest type in the code that
1407 /// needs to be vectorized. We ignore values that remain scalar such as
1408 /// 64 bit loop indices.
1409 unsigned getWidestType();
1411 /// \return The desired interleave count.
1412 /// If interleave count has been specified by metadata it will be returned.
1413 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1414 /// are the selected vectorization factor and the cost of the selected VF.
1415 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1418 /// \return The most profitable unroll factor.
1419 /// This method finds the best unroll-factor based on register pressure and
1420 /// other parameters. VF and LoopCost are the selected vectorization factor
1421 /// and the cost of the selected VF.
1422 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1425 /// \brief A struct that represents some properties of the register usage
1427 struct RegisterUsage {
1428 /// Holds the number of loop invariant values that are used in the loop.
1429 unsigned LoopInvariantRegs;
1430 /// Holds the maximum number of concurrent live intervals in the loop.
1431 unsigned MaxLocalUsers;
1432 /// Holds the number of instructions in the loop.
1433 unsigned NumInstructions;
1436 /// \return information about the register usage of the loop.
1437 RegisterUsage calculateRegisterUsage();
1440 /// Returns the expected execution cost. The unit of the cost does
1441 /// not matter because we use the 'cost' units to compare different
1442 /// vector widths. The cost that is returned is *not* normalized by
1443 /// the factor width.
1444 unsigned expectedCost(unsigned VF);
1446 /// Returns the execution time cost of an instruction for a given vector
1447 /// width. Vector width of one means scalar.
1448 unsigned getInstructionCost(Instruction *I, unsigned VF);
1450 /// Returns whether the instruction is a load or store and will be a emitted
1451 /// as a vector operation.
1452 bool isConsecutiveLoadOrStore(Instruction *I);
1454 /// Report an analysis message to assist the user in diagnosing loops that are
1455 /// not vectorized. These are handled as LoopAccessReport rather than
1456 /// VectorizationReport because the << operator of VectorizationReport returns
1457 /// LoopAccessReport.
1458 void emitAnalysis(const LoopAccessReport &Message) const {
1459 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1462 /// Values used only by @llvm.assume calls.
1463 SmallPtrSet<const Value *, 32> EphValues;
1465 /// The loop that we evaluate.
1468 ScalarEvolution *SE;
1469 /// Loop Info analysis.
1471 /// Vectorization legality.
1472 LoopVectorizationLegality *Legal;
1473 /// Vector target information.
1474 const TargetTransformInfo &TTI;
1475 /// Target Library Info.
1476 const TargetLibraryInfo *TLI;
1477 const Function *TheFunction;
1478 // Loop Vectorize Hint.
1479 const LoopVectorizeHints *Hints;
1482 /// \brief This holds vectorization requirements that must be verified late in
1483 /// the process. The requirements are set by legalize and costmodel. Once
1484 /// vectorization has been determined to be possible and profitable the
1485 /// requirements can be verified by looking for metadata or compiler options.
1486 /// For example, some loops require FP commutativity which is only allowed if
1487 /// vectorization is explicitly specified or if the fast-math compiler option
1488 /// has been provided.
1489 /// Late evaluation of these requirements allows helpful diagnostics to be
1490 /// composed that tells the user what need to be done to vectorize the loop. For
1491 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1492 /// evaluation should be used only when diagnostics can generated that can be
1493 /// followed by a non-expert user.
1494 class LoopVectorizationRequirements {
1496 LoopVectorizationRequirements()
1497 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1499 void addUnsafeAlgebraInst(Instruction *I) {
1500 // First unsafe algebra instruction.
1501 if (!UnsafeAlgebraInst)
1502 UnsafeAlgebraInst = I;
1505 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1507 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1508 const char *Name = Hints.vectorizeAnalysisPassName();
1509 bool Failed = false;
1510 if (UnsafeAlgebraInst &&
1511 Hints.getForce() == LoopVectorizeHints::FK_Undefined &&
1512 Hints.getWidth() == 0) {
1513 emitOptimizationRemarkAnalysisFPCommute(
1514 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1515 VectorizationReport() << "vectorization requires changes in the "
1516 "order of operations, however IEEE 754 "
1517 "floating-point operations are not "
1522 if (NumRuntimePointerChecks >
1523 VectorizerParams::RuntimeMemoryCheckThreshold) {
1524 emitOptimizationRemarkAnalysisAliasing(
1525 F->getContext(), Name, *F, L->getStartLoc(),
1526 VectorizationReport()
1527 << "cannot prove pointers refer to independent arrays in memory. "
1528 "The loop requires "
1529 << NumRuntimePointerChecks
1530 << " runtime independence checks to vectorize the loop, but that "
1531 "would exceed the limit of "
1532 << VectorizerParams::RuntimeMemoryCheckThreshold << " checks");
1533 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1541 unsigned NumRuntimePointerChecks;
1542 Instruction *UnsafeAlgebraInst;
1545 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1547 return V.push_back(&L);
1549 for (Loop *InnerL : L)
1550 addInnerLoop(*InnerL, V);
1553 /// The LoopVectorize Pass.
1554 struct LoopVectorize : public FunctionPass {
1555 /// Pass identification, replacement for typeid
1558 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1560 DisableUnrolling(NoUnrolling),
1561 AlwaysVectorize(AlwaysVectorize) {
1562 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1565 ScalarEvolution *SE;
1567 TargetTransformInfo *TTI;
1569 BlockFrequencyInfo *BFI;
1570 TargetLibraryInfo *TLI;
1572 AssumptionCache *AC;
1573 LoopAccessAnalysis *LAA;
1574 bool DisableUnrolling;
1575 bool AlwaysVectorize;
1577 BlockFrequency ColdEntryFreq;
1579 bool runOnFunction(Function &F) override {
1580 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1581 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1582 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1583 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1584 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1585 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1586 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1587 AA = &getAnalysis<AliasAnalysis>();
1588 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1589 LAA = &getAnalysis<LoopAccessAnalysis>();
1591 // Compute some weights outside of the loop over the loops. Compute this
1592 // using a BranchProbability to re-use its scaling math.
1593 const BranchProbability ColdProb(1, 5); // 20%
1594 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1597 // 1. the target claims to have no vector registers, and
1598 // 2. interleaving won't help ILP.
1600 // The second condition is necessary because, even if the target has no
1601 // vector registers, loop vectorization may still enable scalar
1603 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1606 // Build up a worklist of inner-loops to vectorize. This is necessary as
1607 // the act of vectorizing or partially unrolling a loop creates new loops
1608 // and can invalidate iterators across the loops.
1609 SmallVector<Loop *, 8> Worklist;
1612 addInnerLoop(*L, Worklist);
1614 LoopsAnalyzed += Worklist.size();
1616 // Now walk the identified inner loops.
1617 bool Changed = false;
1618 while (!Worklist.empty())
1619 Changed |= processLoop(Worklist.pop_back_val());
1621 // Process each loop nest in the function.
1625 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1626 SmallVector<Metadata *, 4> MDs;
1627 // Reserve first location for self reference to the LoopID metadata node.
1628 MDs.push_back(nullptr);
1629 bool IsUnrollMetadata = false;
1630 MDNode *LoopID = L->getLoopID();
1632 // First find existing loop unrolling disable metadata.
1633 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1634 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1636 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1638 S && S->getString().startswith("llvm.loop.unroll.disable");
1640 MDs.push_back(LoopID->getOperand(i));
1644 if (!IsUnrollMetadata) {
1645 // Add runtime unroll disable metadata.
1646 LLVMContext &Context = L->getHeader()->getContext();
1647 SmallVector<Metadata *, 1> DisableOperands;
1648 DisableOperands.push_back(
1649 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1650 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1651 MDs.push_back(DisableNode);
1652 MDNode *NewLoopID = MDNode::get(Context, MDs);
1653 // Set operand 0 to refer to the loop id itself.
1654 NewLoopID->replaceOperandWith(0, NewLoopID);
1655 L->setLoopID(NewLoopID);
1659 bool processLoop(Loop *L) {
1660 assert(L->empty() && "Only process inner loops.");
1663 const std::string DebugLocStr = getDebugLocString(L);
1666 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1667 << L->getHeader()->getParent()->getName() << "\" from "
1668 << DebugLocStr << "\n");
1670 LoopVectorizeHints Hints(L, DisableUnrolling);
1672 DEBUG(dbgs() << "LV: Loop hints:"
1674 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1676 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1678 : "?")) << " width=" << Hints.getWidth()
1679 << " unroll=" << Hints.getInterleave() << "\n");
1681 // Function containing loop
1682 Function *F = L->getHeader()->getParent();
1684 // Looking at the diagnostic output is the only way to determine if a loop
1685 // was vectorized (other than looking at the IR or machine code), so it
1686 // is important to generate an optimization remark for each loop. Most of
1687 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1688 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1689 // less verbose reporting vectorized loops and unvectorized loops that may
1690 // benefit from vectorization, respectively.
1692 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1693 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1697 // Check the loop for a trip count threshold:
1698 // do not vectorize loops with a tiny trip count.
1699 const unsigned TC = SE->getSmallConstantTripCount(L);
1700 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1701 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1702 << "This loop is not worth vectorizing.");
1703 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1704 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1706 DEBUG(dbgs() << "\n");
1707 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1708 << "vectorization is not beneficial "
1709 "and is not explicitly forced");
1714 // Check if it is legal to vectorize the loop.
1715 LoopVectorizationRequirements Requirements;
1716 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1717 &Requirements, &Hints);
1718 if (!LVL.canVectorize()) {
1719 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1720 emitMissedWarning(F, L, Hints);
1724 // Use the cost model.
1725 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1727 // Check the function attributes to find out if this function should be
1728 // optimized for size.
1729 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1732 // Compute the weighted frequency of this loop being executed and see if it
1733 // is less than 20% of the function entry baseline frequency. Note that we
1734 // always have a canonical loop here because we think we *can* vectorize.
1735 // FIXME: This is hidden behind a flag due to pervasive problems with
1736 // exactly what block frequency models.
1737 if (LoopVectorizeWithBlockFrequency) {
1738 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1739 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1740 LoopEntryFreq < ColdEntryFreq)
1744 // Check the function attributes to see if implicit floats are allowed.
1745 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1746 // an integer loop and the vector instructions selected are purely integer
1747 // vector instructions?
1748 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1749 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1750 "attribute is used.\n");
1753 VectorizationReport()
1754 << "loop not vectorized due to NoImplicitFloat attribute");
1755 emitMissedWarning(F, L, Hints);
1759 // Select the optimal vectorization factor.
1760 const LoopVectorizationCostModel::VectorizationFactor VF =
1761 CM.selectVectorizationFactor(OptForSize);
1763 // Select the interleave count.
1764 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1766 // Get user interleave count.
1767 unsigned UserIC = Hints.getInterleave();
1769 // Identify the diagnostic messages that should be produced.
1770 std::string VecDiagMsg, IntDiagMsg;
1771 bool VectorizeLoop = true, InterleaveLoop = true;
1773 if (Requirements.doesNotMeet(F, L, Hints)) {
1774 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1776 emitMissedWarning(F, L, Hints);
1780 if (VF.Width == 1) {
1781 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1783 "the cost-model indicates that vectorization is not beneficial";
1784 VectorizeLoop = false;
1787 if (IC == 1 && UserIC <= 1) {
1788 // Tell the user interleaving is not beneficial.
1789 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1791 "the cost-model indicates that interleaving is not beneficial";
1792 InterleaveLoop = false;
1795 " and is explicitly disabled or interleave count is set to 1";
1796 } else if (IC > 1 && UserIC == 1) {
1797 // Tell the user interleaving is beneficial, but it explicitly disabled.
1799 << "LV: Interleaving is beneficial but is explicitly disabled.");
1800 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1801 "but is explicitly disabled or interleave count is set to 1";
1802 InterleaveLoop = false;
1805 // Override IC if user provided an interleave count.
1806 IC = UserIC > 0 ? UserIC : IC;
1808 // Emit diagnostic messages, if any.
1809 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1810 if (!VectorizeLoop && !InterleaveLoop) {
1811 // Do not vectorize or interleaving the loop.
1812 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1813 L->getStartLoc(), VecDiagMsg);
1814 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1815 L->getStartLoc(), IntDiagMsg);
1817 } else if (!VectorizeLoop && InterleaveLoop) {
1818 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1819 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1820 L->getStartLoc(), VecDiagMsg);
1821 } else if (VectorizeLoop && !InterleaveLoop) {
1822 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1823 << DebugLocStr << '\n');
1824 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1825 L->getStartLoc(), IntDiagMsg);
1826 } else if (VectorizeLoop && InterleaveLoop) {
1827 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1828 << DebugLocStr << '\n');
1829 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1832 if (!VectorizeLoop) {
1833 assert(IC > 1 && "interleave count should not be 1 or 0");
1834 // If we decided that it is not legal to vectorize the loop then
1836 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1837 Unroller.vectorize(&LVL);
1839 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1840 Twine("interleaved loop (interleaved count: ") +
1843 // If we decided that it is *legal* to vectorize the loop then do it.
1844 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1848 // Add metadata to disable runtime unrolling scalar loop when there's no
1849 // runtime check about strides and memory. Because at this situation,
1850 // scalar loop is rarely used not worthy to be unrolled.
1851 if (!LB.IsSafetyChecksAdded())
1852 AddRuntimeUnrollDisableMetaData(L);
1854 // Report the vectorization decision.
1855 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1856 Twine("vectorized loop (vectorization width: ") +
1857 Twine(VF.Width) + ", interleaved count: " +
1861 // Mark the loop as already vectorized to avoid vectorizing again.
1862 Hints.setAlreadyVectorized();
1864 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1868 void getAnalysisUsage(AnalysisUsage &AU) const override {
1869 AU.addRequired<AssumptionCacheTracker>();
1870 AU.addRequiredID(LoopSimplifyID);
1871 AU.addRequiredID(LCSSAID);
1872 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1873 AU.addRequired<DominatorTreeWrapperPass>();
1874 AU.addRequired<LoopInfoWrapperPass>();
1875 AU.addRequired<ScalarEvolutionWrapperPass>();
1876 AU.addRequired<TargetTransformInfoWrapperPass>();
1877 AU.addRequired<AliasAnalysis>();
1878 AU.addRequired<LoopAccessAnalysis>();
1879 AU.addPreserved<LoopInfoWrapperPass>();
1880 AU.addPreserved<DominatorTreeWrapperPass>();
1881 AU.addPreserved<AliasAnalysis>();
1886 } // end anonymous namespace
1888 //===----------------------------------------------------------------------===//
1889 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1890 // LoopVectorizationCostModel.
1891 //===----------------------------------------------------------------------===//
1893 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1894 // We need to place the broadcast of invariant variables outside the loop.
1895 Instruction *Instr = dyn_cast<Instruction>(V);
1897 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1898 Instr->getParent()) != LoopVectorBody.end());
1899 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1901 // Place the code for broadcasting invariant variables in the new preheader.
1902 IRBuilder<>::InsertPointGuard Guard(Builder);
1904 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1906 // Broadcast the scalar into all locations in the vector.
1907 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1912 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1914 assert(Val->getType()->isVectorTy() && "Must be a vector");
1915 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1916 "Elem must be an integer");
1917 assert(Step->getType() == Val->getType()->getScalarType() &&
1918 "Step has wrong type");
1919 // Create the types.
1920 Type *ITy = Val->getType()->getScalarType();
1921 VectorType *Ty = cast<VectorType>(Val->getType());
1922 int VLen = Ty->getNumElements();
1923 SmallVector<Constant*, 8> Indices;
1925 // Create a vector of consecutive numbers from zero to VF.
1926 for (int i = 0; i < VLen; ++i)
1927 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1929 // Add the consecutive indices to the vector value.
1930 Constant *Cv = ConstantVector::get(Indices);
1931 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1932 Step = Builder.CreateVectorSplat(VLen, Step);
1933 assert(Step->getType() == Val->getType() && "Invalid step vec");
1934 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1935 // which can be found from the original scalar operations.
1936 Step = Builder.CreateMul(Cv, Step);
1937 return Builder.CreateAdd(Val, Step, "induction");
1940 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1941 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1942 // Make sure that the pointer does not point to structs.
1943 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1946 // If this value is a pointer induction variable we know it is consecutive.
1947 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1948 if (Phi && Inductions.count(Phi)) {
1949 InductionInfo II = Inductions[Phi];
1950 return II.getConsecutiveDirection();
1953 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1957 unsigned NumOperands = Gep->getNumOperands();
1958 Value *GpPtr = Gep->getPointerOperand();
1959 // If this GEP value is a consecutive pointer induction variable and all of
1960 // the indices are constant then we know it is consecutive. We can
1961 Phi = dyn_cast<PHINode>(GpPtr);
1962 if (Phi && Inductions.count(Phi)) {
1964 // Make sure that the pointer does not point to structs.
1965 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1966 if (GepPtrType->getElementType()->isAggregateType())
1969 // Make sure that all of the index operands are loop invariant.
1970 for (unsigned i = 1; i < NumOperands; ++i)
1971 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1974 InductionInfo II = Inductions[Phi];
1975 return II.getConsecutiveDirection();
1978 unsigned InductionOperand = getGEPInductionOperand(Gep);
1980 // Check that all of the gep indices are uniform except for our induction
1982 for (unsigned i = 0; i != NumOperands; ++i)
1983 if (i != InductionOperand &&
1984 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1987 // We can emit wide load/stores only if the last non-zero index is the
1988 // induction variable.
1989 const SCEV *Last = nullptr;
1990 if (!Strides.count(Gep))
1991 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1993 // Because of the multiplication by a stride we can have a s/zext cast.
1994 // We are going to replace this stride by 1 so the cast is safe to ignore.
1996 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1997 // %0 = trunc i64 %indvars.iv to i32
1998 // %mul = mul i32 %0, %Stride1
1999 // %idxprom = zext i32 %mul to i64 << Safe cast.
2000 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
2002 Last = replaceSymbolicStrideSCEV(SE, Strides,
2003 Gep->getOperand(InductionOperand), Gep);
2004 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
2006 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2010 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2011 const SCEV *Step = AR->getStepRecurrence(*SE);
2013 // The memory is consecutive because the last index is consecutive
2014 // and all other indices are loop invariant.
2017 if (Step->isAllOnesValue())
2024 bool LoopVectorizationLegality::isUniform(Value *V) {
2025 return LAI->isUniform(V);
2028 InnerLoopVectorizer::VectorParts&
2029 InnerLoopVectorizer::getVectorValue(Value *V) {
2030 assert(V != Induction && "The new induction variable should not be used.");
2031 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2033 // If we have a stride that is replaced by one, do it here.
2034 if (Legal->hasStride(V))
2035 V = ConstantInt::get(V->getType(), 1);
2037 // If we have this scalar in the map, return it.
2038 if (WidenMap.has(V))
2039 return WidenMap.get(V);
2041 // If this scalar is unknown, assume that it is a constant or that it is
2042 // loop invariant. Broadcast V and save the value for future uses.
2043 Value *B = getBroadcastInstrs(V);
2044 return WidenMap.splat(V, B);
2047 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2048 assert(Vec->getType()->isVectorTy() && "Invalid type");
2049 SmallVector<Constant*, 8> ShuffleMask;
2050 for (unsigned i = 0; i < VF; ++i)
2051 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2053 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2054 ConstantVector::get(ShuffleMask),
2058 // Get a mask to interleave \p NumVec vectors into a wide vector.
2059 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2060 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2061 // <0, 4, 1, 5, 2, 6, 3, 7>
2062 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2064 SmallVector<Constant *, 16> Mask;
2065 for (unsigned i = 0; i < VF; i++)
2066 for (unsigned j = 0; j < NumVec; j++)
2067 Mask.push_back(Builder.getInt32(j * VF + i));
2069 return ConstantVector::get(Mask);
2072 // Get the strided mask starting from index \p Start.
2073 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2074 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2075 unsigned Stride, unsigned VF) {
2076 SmallVector<Constant *, 16> Mask;
2077 for (unsigned i = 0; i < VF; i++)
2078 Mask.push_back(Builder.getInt32(Start + i * Stride));
2080 return ConstantVector::get(Mask);
2083 // Get a mask of two parts: The first part consists of sequential integers
2084 // starting from 0, The second part consists of UNDEFs.
2085 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2086 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2087 unsigned NumUndef) {
2088 SmallVector<Constant *, 16> Mask;
2089 for (unsigned i = 0; i < NumInt; i++)
2090 Mask.push_back(Builder.getInt32(i));
2092 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2093 for (unsigned i = 0; i < NumUndef; i++)
2094 Mask.push_back(Undef);
2096 return ConstantVector::get(Mask);
2099 // Concatenate two vectors with the same element type. The 2nd vector should
2100 // not have more elements than the 1st vector. If the 2nd vector has less
2101 // elements, extend it with UNDEFs.
2102 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2104 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2105 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2106 assert(VecTy1 && VecTy2 &&
2107 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2108 "Expect two vectors with the same element type");
2110 unsigned NumElts1 = VecTy1->getNumElements();
2111 unsigned NumElts2 = VecTy2->getNumElements();
2112 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2114 if (NumElts1 > NumElts2) {
2115 // Extend with UNDEFs.
2117 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2118 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2121 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2122 return Builder.CreateShuffleVector(V1, V2, Mask);
2125 // Concatenate vectors in the given list. All vectors have the same type.
2126 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2127 ArrayRef<Value *> InputList) {
2128 unsigned NumVec = InputList.size();
2129 assert(NumVec > 1 && "Should be at least two vectors");
2131 SmallVector<Value *, 8> ResList;
2132 ResList.append(InputList.begin(), InputList.end());
2134 SmallVector<Value *, 8> TmpList;
2135 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2136 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2137 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2138 "Only the last vector may have a different type");
2140 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2143 // Push the last vector if the total number of vectors is odd.
2144 if (NumVec % 2 != 0)
2145 TmpList.push_back(ResList[NumVec - 1]);
2148 NumVec = ResList.size();
2149 } while (NumVec > 1);
2154 // Try to vectorize the interleave group that \p Instr belongs to.
2156 // E.g. Translate following interleaved load group (factor = 3):
2157 // for (i = 0; i < N; i+=3) {
2158 // R = Pic[i]; // Member of index 0
2159 // G = Pic[i+1]; // Member of index 1
2160 // B = Pic[i+2]; // Member of index 2
2161 // ... // do something to R, G, B
2164 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2165 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2166 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2167 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2169 // Or translate following interleaved store group (factor = 3):
2170 // for (i = 0; i < N; i+=3) {
2171 // ... do something to R, G, B
2172 // Pic[i] = R; // Member of index 0
2173 // Pic[i+1] = G; // Member of index 1
2174 // Pic[i+2] = B; // Member of index 2
2177 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2178 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2179 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2180 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2181 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2182 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2183 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2184 assert(Group && "Fail to get an interleaved access group.");
2186 // Skip if current instruction is not the insert position.
2187 if (Instr != Group->getInsertPos())
2190 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2191 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2192 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2194 // Prepare for the vector type of the interleaved load/store.
2195 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2196 unsigned InterleaveFactor = Group->getFactor();
2197 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2198 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2200 // Prepare for the new pointers.
2201 setDebugLocFromInst(Builder, Ptr);
2202 VectorParts &PtrParts = getVectorValue(Ptr);
2203 SmallVector<Value *, 2> NewPtrs;
2204 unsigned Index = Group->getIndex(Instr);
2205 for (unsigned Part = 0; Part < UF; Part++) {
2206 // Extract the pointer for current instruction from the pointer vector. A
2207 // reverse access uses the pointer in the last lane.
2208 Value *NewPtr = Builder.CreateExtractElement(
2210 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2212 // Notice current instruction could be any index. Need to adjust the address
2213 // to the member of index 0.
2215 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2216 // b = A[i]; // Member of index 0
2217 // Current pointer is pointed to A[i+1], adjust it to A[i].
2219 // E.g. A[i+1] = a; // Member of index 1
2220 // A[i] = b; // Member of index 0
2221 // A[i+2] = c; // Member of index 2 (Current instruction)
2222 // Current pointer is pointed to A[i+2], adjust it to A[i].
2223 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2225 // Cast to the vector pointer type.
2226 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2229 setDebugLocFromInst(Builder, Instr);
2230 Value *UndefVec = UndefValue::get(VecTy);
2232 // Vectorize the interleaved load group.
2234 for (unsigned Part = 0; Part < UF; Part++) {
2235 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2236 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2238 for (unsigned i = 0; i < InterleaveFactor; i++) {
2239 Instruction *Member = Group->getMember(i);
2241 // Skip the gaps in the group.
2245 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2246 Value *StridedVec = Builder.CreateShuffleVector(
2247 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2249 // If this member has different type, cast the result type.
2250 if (Member->getType() != ScalarTy) {
2251 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2252 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2255 VectorParts &Entry = WidenMap.get(Member);
2257 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2260 propagateMetadata(NewLoadInstr, Instr);
2265 // The sub vector type for current instruction.
2266 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2268 // Vectorize the interleaved store group.
2269 for (unsigned Part = 0; Part < UF; Part++) {
2270 // Collect the stored vector from each member.
2271 SmallVector<Value *, 4> StoredVecs;
2272 for (unsigned i = 0; i < InterleaveFactor; i++) {
2273 // Interleaved store group doesn't allow a gap, so each index has a member
2274 Instruction *Member = Group->getMember(i);
2275 assert(Member && "Fail to get a member from an interleaved store group");
2278 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2279 if (Group->isReverse())
2280 StoredVec = reverseVector(StoredVec);
2282 // If this member has different type, cast it to an unified type.
2283 if (StoredVec->getType() != SubVT)
2284 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2286 StoredVecs.push_back(StoredVec);
2289 // Concatenate all vectors into a wide vector.
2290 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2292 // Interleave the elements in the wide vector.
2293 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2294 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2297 Instruction *NewStoreInstr =
2298 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2299 propagateMetadata(NewStoreInstr, Instr);
2303 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2304 // Attempt to issue a wide load.
2305 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2306 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2308 assert((LI || SI) && "Invalid Load/Store instruction");
2310 // Try to vectorize the interleave group if this access is interleaved.
2311 if (Legal->isAccessInterleaved(Instr))
2312 return vectorizeInterleaveGroup(Instr);
2314 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2315 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2316 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2317 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2318 // An alignment of 0 means target abi alignment. We need to use the scalar's
2319 // target abi alignment in such a case.
2320 const DataLayout &DL = Instr->getModule()->getDataLayout();
2322 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2323 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2324 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2325 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2327 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2328 !Legal->isMaskRequired(SI))
2329 return scalarizeInstruction(Instr, true);
2331 if (ScalarAllocatedSize != VectorElementSize)
2332 return scalarizeInstruction(Instr);
2334 // If the pointer is loop invariant or if it is non-consecutive,
2335 // scalarize the load.
2336 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2337 bool Reverse = ConsecutiveStride < 0;
2338 bool UniformLoad = LI && Legal->isUniform(Ptr);
2339 if (!ConsecutiveStride || UniformLoad)
2340 return scalarizeInstruction(Instr);
2342 Constant *Zero = Builder.getInt32(0);
2343 VectorParts &Entry = WidenMap.get(Instr);
2345 // Handle consecutive loads/stores.
2346 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2347 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2348 setDebugLocFromInst(Builder, Gep);
2349 Value *PtrOperand = Gep->getPointerOperand();
2350 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2351 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2353 // Create the new GEP with the new induction variable.
2354 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2355 Gep2->setOperand(0, FirstBasePtr);
2356 Gep2->setName("gep.indvar.base");
2357 Ptr = Builder.Insert(Gep2);
2359 setDebugLocFromInst(Builder, Gep);
2360 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2361 OrigLoop) && "Base ptr must be invariant");
2363 // The last index does not have to be the induction. It can be
2364 // consecutive and be a function of the index. For example A[I+1];
2365 unsigned NumOperands = Gep->getNumOperands();
2366 unsigned InductionOperand = getGEPInductionOperand(Gep);
2367 // Create the new GEP with the new induction variable.
2368 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2370 for (unsigned i = 0; i < NumOperands; ++i) {
2371 Value *GepOperand = Gep->getOperand(i);
2372 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2374 // Update last index or loop invariant instruction anchored in loop.
2375 if (i == InductionOperand ||
2376 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2377 assert((i == InductionOperand ||
2378 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2379 "Must be last index or loop invariant");
2381 VectorParts &GEPParts = getVectorValue(GepOperand);
2382 Value *Index = GEPParts[0];
2383 Index = Builder.CreateExtractElement(Index, Zero);
2384 Gep2->setOperand(i, Index);
2385 Gep2->setName("gep.indvar.idx");
2388 Ptr = Builder.Insert(Gep2);
2390 // Use the induction element ptr.
2391 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2392 setDebugLocFromInst(Builder, Ptr);
2393 VectorParts &PtrVal = getVectorValue(Ptr);
2394 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2397 VectorParts Mask = createBlockInMask(Instr->getParent());
2400 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2401 "We do not allow storing to uniform addresses");
2402 setDebugLocFromInst(Builder, SI);
2403 // We don't want to update the value in the map as it might be used in
2404 // another expression. So don't use a reference type for "StoredVal".
2405 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2407 for (unsigned Part = 0; Part < UF; ++Part) {
2408 // Calculate the pointer for the specific unroll-part.
2410 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2413 // If we store to reverse consecutive memory locations, then we need
2414 // to reverse the order of elements in the stored value.
2415 StoredVal[Part] = reverseVector(StoredVal[Part]);
2416 // If the address is consecutive but reversed, then the
2417 // wide store needs to start at the last vector element.
2418 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2419 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2420 Mask[Part] = reverseVector(Mask[Part]);
2423 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2424 DataTy->getPointerTo(AddressSpace));
2427 if (Legal->isMaskRequired(SI))
2428 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2431 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2432 propagateMetadata(NewSI, SI);
2438 assert(LI && "Must have a load instruction");
2439 setDebugLocFromInst(Builder, LI);
2440 for (unsigned Part = 0; Part < UF; ++Part) {
2441 // Calculate the pointer for the specific unroll-part.
2443 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2446 // If the address is consecutive but reversed, then the
2447 // wide load needs to start at the last vector element.
2448 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2449 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2450 Mask[Part] = reverseVector(Mask[Part]);
2454 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2455 DataTy->getPointerTo(AddressSpace));
2456 if (Legal->isMaskRequired(LI))
2457 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2458 UndefValue::get(DataTy),
2459 "wide.masked.load");
2461 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2462 propagateMetadata(NewLI, LI);
2463 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2467 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2468 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2469 // Holds vector parameters or scalars, in case of uniform vals.
2470 SmallVector<VectorParts, 4> Params;
2472 setDebugLocFromInst(Builder, Instr);
2474 // Find all of the vectorized parameters.
2475 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2476 Value *SrcOp = Instr->getOperand(op);
2478 // If we are accessing the old induction variable, use the new one.
2479 if (SrcOp == OldInduction) {
2480 Params.push_back(getVectorValue(SrcOp));
2484 // Try using previously calculated values.
2485 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2487 // If the src is an instruction that appeared earlier in the basic block,
2488 // then it should already be vectorized.
2489 if (SrcInst && OrigLoop->contains(SrcInst)) {
2490 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2491 // The parameter is a vector value from earlier.
2492 Params.push_back(WidenMap.get(SrcInst));
2494 // The parameter is a scalar from outside the loop. Maybe even a constant.
2495 VectorParts Scalars;
2496 Scalars.append(UF, SrcOp);
2497 Params.push_back(Scalars);
2501 assert(Params.size() == Instr->getNumOperands() &&
2502 "Invalid number of operands");
2504 // Does this instruction return a value ?
2505 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2507 Value *UndefVec = IsVoidRetTy ? nullptr :
2508 UndefValue::get(VectorType::get(Instr->getType(), VF));
2509 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2510 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2512 Instruction *InsertPt = Builder.GetInsertPoint();
2513 BasicBlock *IfBlock = Builder.GetInsertBlock();
2514 BasicBlock *CondBlock = nullptr;
2517 Loop *VectorLp = nullptr;
2518 if (IfPredicateStore) {
2519 assert(Instr->getParent()->getSinglePredecessor() &&
2520 "Only support single predecessor blocks");
2521 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2522 Instr->getParent());
2523 VectorLp = LI->getLoopFor(IfBlock);
2524 assert(VectorLp && "Must have a loop for this block");
2527 // For each vector unroll 'part':
2528 for (unsigned Part = 0; Part < UF; ++Part) {
2529 // For each scalar that we create:
2530 for (unsigned Width = 0; Width < VF; ++Width) {
2533 Value *Cmp = nullptr;
2534 if (IfPredicateStore) {
2535 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2536 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2537 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2538 LoopVectorBody.push_back(CondBlock);
2539 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2540 // Update Builder with newly created basic block.
2541 Builder.SetInsertPoint(InsertPt);
2544 Instruction *Cloned = Instr->clone();
2546 Cloned->setName(Instr->getName() + ".cloned");
2547 // Replace the operands of the cloned instructions with extracted scalars.
2548 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2549 Value *Op = Params[op][Part];
2550 // Param is a vector. Need to extract the right lane.
2551 if (Op->getType()->isVectorTy())
2552 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2553 Cloned->setOperand(op, Op);
2556 // Place the cloned scalar in the new loop.
2557 Builder.Insert(Cloned);
2559 // If the original scalar returns a value we need to place it in a vector
2560 // so that future users will be able to use it.
2562 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2563 Builder.getInt32(Width));
2565 if (IfPredicateStore) {
2566 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2567 LoopVectorBody.push_back(NewIfBlock);
2568 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2569 Builder.SetInsertPoint(InsertPt);
2570 ReplaceInstWithInst(IfBlock->getTerminator(),
2571 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2572 IfBlock = NewIfBlock;
2578 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2582 if (Instruction *I = dyn_cast<Instruction>(V))
2583 return I->getParent() == Loc->getParent() ? I : nullptr;
2587 std::pair<Instruction *, Instruction *>
2588 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2589 Instruction *tnullptr = nullptr;
2590 if (!Legal->mustCheckStrides())
2591 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2593 IRBuilder<> ChkBuilder(Loc);
2596 Value *Check = nullptr;
2597 Instruction *FirstInst = nullptr;
2598 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2599 SE = Legal->strides_end();
2601 Value *Ptr = stripIntegerCast(*SI);
2602 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2604 // Store the first instruction we create.
2605 FirstInst = getFirstInst(FirstInst, C, Loc);
2607 Check = ChkBuilder.CreateOr(Check, C);
2612 // We have to do this trickery because the IRBuilder might fold the check to a
2613 // constant expression in which case there is no Instruction anchored in a
2615 LLVMContext &Ctx = Loc->getContext();
2616 Instruction *TheCheck =
2617 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2618 ChkBuilder.Insert(TheCheck, "stride.not.one");
2619 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2621 return std::make_pair(FirstInst, TheCheck);
2624 void InnerLoopVectorizer::createEmptyLoop() {
2626 In this function we generate a new loop. The new loop will contain
2627 the vectorized instructions while the old loop will continue to run the
2630 [ ] <-- loop iteration number check.
2633 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2636 || [ ] <-- vector pre header.
2640 || [ ]_| <-- vector loop.
2643 | >[ ] <--- middle-block.
2646 -|- >[ ] <--- new preheader.
2650 | [ ]_| <-- old scalar loop to handle remainder.
2653 >[ ] <-- exit block.
2657 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2658 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2659 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2660 assert(VectorPH && "Invalid loop structure");
2661 assert(ExitBlock && "Must have an exit block");
2663 // Some loops have a single integer induction variable, while other loops
2664 // don't. One example is c++ iterators that often have multiple pointer
2665 // induction variables. In the code below we also support a case where we
2666 // don't have a single induction variable.
2667 OldInduction = Legal->getInduction();
2668 Type *IdxTy = Legal->getWidestInductionType();
2670 // Find the loop boundaries.
2671 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2672 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2674 // The exit count might have the type of i64 while the phi is i32. This can
2675 // happen if we have an induction variable that is sign extended before the
2676 // compare. The only way that we get a backedge taken count is that the
2677 // induction variable was signed and as such will not overflow. In such a case
2678 // truncation is legal.
2679 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2680 IdxTy->getPrimitiveSizeInBits())
2681 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2683 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2684 // Get the total trip count from the count by adding 1.
2685 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2686 SE->getConstant(BackedgeTakeCount->getType(), 1));
2688 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2690 // Expand the trip count and place the new instructions in the preheader.
2691 // Notice that the pre-header does not change, only the loop body.
2692 SCEVExpander Exp(*SE, DL, "induction");
2694 // The loop minimum iterations check below is to ensure the loop has enough
2695 // trip count so the generated vector loop will likely be executed and the
2696 // preparation and rounding-off costs will likely be worthy.
2698 // The minimum iteration check also covers case where the backedge-taken
2699 // count is uint##_max. Adding one to it will cause overflow and an
2700 // incorrect loop trip count being generated in the vector body. In this
2701 // case we also want to directly jump to the scalar remainder loop.
2702 Value *ExitCountValue = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2703 VectorPH->getTerminator());
2704 if (ExitCountValue->getType()->isPointerTy())
2705 ExitCountValue = CastInst::CreatePointerCast(ExitCountValue, IdxTy,
2706 "exitcount.ptrcnt.to.int",
2707 VectorPH->getTerminator());
2709 Instruction *CheckMinIters =
2710 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULT, ExitCountValue,
2711 ConstantInt::get(ExitCountValue->getType(), VF * UF),
2712 "min.iters.check", VectorPH->getTerminator());
2714 // The loop index does not have to start at Zero. Find the original start
2715 // value from the induction PHI node. If we don't have an induction variable
2716 // then we know that it starts at zero.
2717 Builder.SetInsertPoint(VectorPH->getTerminator());
2718 Value *StartIdx = ExtendedIdx =
2720 ? Builder.CreateZExt(OldInduction->getIncomingValueForBlock(VectorPH),
2722 : ConstantInt::get(IdxTy, 0);
2724 // Count holds the overall loop count (N).
2725 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2726 VectorPH->getTerminator());
2728 LoopBypassBlocks.push_back(VectorPH);
2730 // Split the single block loop into the two loop structure described above.
2731 BasicBlock *VecBody =
2732 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2733 BasicBlock *MiddleBlock =
2734 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2735 BasicBlock *ScalarPH =
2736 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2738 // Create and register the new vector loop.
2739 Loop* Lp = new Loop();
2740 Loop *ParentLoop = OrigLoop->getParentLoop();
2742 // Insert the new loop into the loop nest and register the new basic blocks
2743 // before calling any utilities such as SCEV that require valid LoopInfo.
2745 ParentLoop->addChildLoop(Lp);
2746 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2747 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2749 LI->addTopLevelLoop(Lp);
2751 Lp->addBasicBlockToLoop(VecBody, *LI);
2753 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2755 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2757 // Generate the induction variable.
2758 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2759 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2760 // The loop step is equal to the vectorization factor (num of SIMD elements)
2761 // times the unroll factor (num of SIMD instructions).
2762 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2764 // Generate code to check that the loop's trip count is not less than the
2765 // minimum loop iteration number threshold.
2766 BasicBlock *NewVectorPH =
2767 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "min.iters.checked");
2769 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2770 ReplaceInstWithInst(VectorPH->getTerminator(),
2771 BranchInst::Create(ScalarPH, NewVectorPH, CheckMinIters));
2772 VectorPH = NewVectorPH;
2774 // This is the IR builder that we use to add all of the logic for bypassing
2775 // the new vector loop.
2776 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2777 setDebugLocFromInst(BypassBuilder,
2778 getDebugLocFromInstOrOperands(OldInduction));
2780 // We may need to extend the index in case there is a type mismatch.
2781 // We know that the count starts at zero and does not overflow.
2782 if (Count->getType() != IdxTy) {
2783 // The exit count can be of pointer type. Convert it to the correct
2785 if (ExitCount->getType()->isPointerTy())
2786 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2788 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2791 // Add the start index to the loop count to get the new end index.
2792 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2794 // Now we need to generate the expression for N - (N % VF), which is
2795 // the part that the vectorized body will execute.
2796 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2797 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2798 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2799 "end.idx.rnd.down");
2801 // Now, compare the new count to zero. If it is zero skip the vector loop and
2802 // jump to the scalar loop.
2804 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2806 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2808 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2809 LoopBypassBlocks.push_back(VectorPH);
2810 ReplaceInstWithInst(VectorPH->getTerminator(),
2811 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2812 VectorPH = NewVectorPH;
2814 // Generate the code to check that the strides we assumed to be one are really
2815 // one. We want the new basic block to start at the first instruction in a
2816 // sequence of instructions that form a check.
2817 Instruction *StrideCheck;
2818 Instruction *FirstCheckInst;
2819 std::tie(FirstCheckInst, StrideCheck) =
2820 addStrideCheck(VectorPH->getTerminator());
2822 AddedSafetyChecks = true;
2823 // Create a new block containing the stride check.
2824 VectorPH->setName("vector.stridecheck");
2826 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2828 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2829 LoopBypassBlocks.push_back(VectorPH);
2831 // Replace the branch into the memory check block with a conditional branch
2832 // for the "few elements case".
2833 ReplaceInstWithInst(
2834 VectorPH->getTerminator(),
2835 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2837 VectorPH = NewVectorPH;
2840 // Generate the code that checks in runtime if arrays overlap. We put the
2841 // checks into a separate block to make the more common case of few elements
2843 Instruction *MemRuntimeCheck;
2844 std::tie(FirstCheckInst, MemRuntimeCheck) =
2845 Legal->getLAI()->addRuntimeChecks(VectorPH->getTerminator());
2846 if (MemRuntimeCheck) {
2847 AddedSafetyChecks = true;
2848 // Create a new block containing the memory check.
2849 VectorPH->setName("vector.memcheck");
2851 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2853 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2854 LoopBypassBlocks.push_back(VectorPH);
2856 // Replace the branch into the memory check block with a conditional branch
2857 // for the "few elements case".
2858 ReplaceInstWithInst(
2859 VectorPH->getTerminator(),
2860 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2862 VectorPH = NewVectorPH;
2865 // We are going to resume the execution of the scalar loop.
2866 // Go over all of the induction variables that we found and fix the
2867 // PHIs that are left in the scalar version of the loop.
2868 // The starting values of PHI nodes depend on the counter of the last
2869 // iteration in the vectorized loop.
2870 // If we come from a bypass edge then we need to start from the original
2873 // This variable saves the new starting index for the scalar loop.
2874 PHINode *ResumeIndex = nullptr;
2875 LoopVectorizationLegality::InductionList::iterator I, E;
2876 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2877 // Set builder to point to last bypass block.
2878 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2879 for (I = List->begin(), E = List->end(); I != E; ++I) {
2880 PHINode *OrigPhi = I->first;
2881 LoopVectorizationLegality::InductionInfo II = I->second;
2883 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2884 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2885 MiddleBlock->getTerminator());
2886 // We might have extended the type of the induction variable but we need a
2887 // truncated version for the scalar loop.
2888 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2889 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2890 MiddleBlock->getTerminator()) : nullptr;
2892 // Create phi nodes to merge from the backedge-taken check block.
2893 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2894 ScalarPH->getTerminator());
2895 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2897 PHINode *BCTruncResumeVal = nullptr;
2898 if (OrigPhi == OldInduction) {
2900 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2901 ScalarPH->getTerminator());
2902 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2905 Value *EndValue = nullptr;
2907 case LoopVectorizationLegality::IK_NoInduction:
2908 llvm_unreachable("Unknown induction");
2909 case LoopVectorizationLegality::IK_IntInduction: {
2910 // Handle the integer induction counter.
2911 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2913 // We have the canonical induction variable.
2914 if (OrigPhi == OldInduction) {
2915 // Create a truncated version of the resume value for the scalar loop,
2916 // we might have promoted the type to a larger width.
2918 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2919 // The new PHI merges the original incoming value, in case of a bypass,
2920 // or the value at the end of the vectorized loop.
2921 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2922 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2923 TruncResumeVal->addIncoming(EndValue, VecBody);
2925 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2927 // We know what the end value is.
2928 EndValue = IdxEndRoundDown;
2929 // We also know which PHI node holds it.
2930 ResumeIndex = ResumeVal;
2934 // Not the canonical induction variable - add the vector loop count to the
2936 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2937 II.StartValue->getType(),
2939 EndValue = II.transform(BypassBuilder, CRD);
2940 EndValue->setName("ind.end");
2943 case LoopVectorizationLegality::IK_PtrInduction: {
2944 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2945 II.StepValue->getType(),
2947 EndValue = II.transform(BypassBuilder, CRD);
2948 EndValue->setName("ptr.ind.end");
2953 // The new PHI merges the original incoming value, in case of a bypass,
2954 // or the value at the end of the vectorized loop.
2955 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2956 if (OrigPhi == OldInduction)
2957 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2959 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2961 ResumeVal->addIncoming(EndValue, VecBody);
2963 // Fix the scalar body counter (PHI node).
2964 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2966 // The old induction's phi node in the scalar body needs the truncated
2968 if (OrigPhi == OldInduction) {
2969 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2970 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2972 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2973 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2977 // If we are generating a new induction variable then we also need to
2978 // generate the code that calculates the exit value. This value is not
2979 // simply the end of the counter because we may skip the vectorized body
2980 // in case of a runtime check.
2982 assert(!ResumeIndex && "Unexpected resume value found");
2983 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2984 MiddleBlock->getTerminator());
2985 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2986 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2987 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2990 // Make sure that we found the index where scalar loop needs to continue.
2991 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2992 "Invalid resume Index");
2994 // Add a check in the middle block to see if we have completed
2995 // all of the iterations in the first vector loop.
2996 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2997 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2998 ResumeIndex, "cmp.n",
2999 MiddleBlock->getTerminator());
3000 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3001 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3003 // Create i+1 and fill the PHINode.
3004 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
3005 Induction->addIncoming(StartIdx, VectorPH);
3006 Induction->addIncoming(NextIdx, VecBody);
3007 // Create the compare.
3008 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
3009 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
3011 // Now we have two terminators. Remove the old one from the block.
3012 VecBody->getTerminator()->eraseFromParent();
3014 // Get ready to start creating new instructions into the vectorized body.
3015 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
3018 LoopVectorPreHeader = VectorPH;
3019 LoopScalarPreHeader = ScalarPH;
3020 LoopMiddleBlock = MiddleBlock;
3021 LoopExitBlock = ExitBlock;
3022 LoopVectorBody.push_back(VecBody);
3023 LoopScalarBody = OldBasicBlock;
3025 LoopVectorizeHints Hints(Lp, true);
3026 Hints.setAlreadyVectorized();
3030 struct CSEDenseMapInfo {
3031 static bool canHandle(Instruction *I) {
3032 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3033 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3035 static inline Instruction *getEmptyKey() {
3036 return DenseMapInfo<Instruction *>::getEmptyKey();
3038 static inline Instruction *getTombstoneKey() {
3039 return DenseMapInfo<Instruction *>::getTombstoneKey();
3041 static unsigned getHashValue(Instruction *I) {
3042 assert(canHandle(I) && "Unknown instruction!");
3043 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3044 I->value_op_end()));
3046 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3047 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3048 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3050 return LHS->isIdenticalTo(RHS);
3055 /// \brief Check whether this block is a predicated block.
3056 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
3057 /// = ...; " blocks. We start with one vectorized basic block. For every
3058 /// conditional block we split this vectorized block. Therefore, every second
3059 /// block will be a predicated one.
3060 static bool isPredicatedBlock(unsigned BlockNum) {
3061 return BlockNum % 2;
3064 ///\brief Perform cse of induction variable instructions.
3065 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
3066 // Perform simple cse.
3067 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3068 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3069 BasicBlock *BB = BBs[i];
3070 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3071 Instruction *In = I++;
3073 if (!CSEDenseMapInfo::canHandle(In))
3076 // Check if we can replace this instruction with any of the
3077 // visited instructions.
3078 if (Instruction *V = CSEMap.lookup(In)) {
3079 In->replaceAllUsesWith(V);
3080 In->eraseFromParent();
3083 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3084 // ...;" blocks for predicated stores. Every second block is a predicated
3086 if (isPredicatedBlock(i))
3094 /// \brief Adds a 'fast' flag to floating point operations.
3095 static Value *addFastMathFlag(Value *V) {
3096 if (isa<FPMathOperator>(V)){
3097 FastMathFlags Flags;
3098 Flags.setUnsafeAlgebra();
3099 cast<Instruction>(V)->setFastMathFlags(Flags);
3104 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3105 /// the result needs to be inserted and/or extracted from vectors.
3106 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3107 const TargetTransformInfo &TTI) {
3111 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3114 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3116 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3118 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3124 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3125 // Return the cost of the instruction, including scalarization overhead if it's
3126 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3127 // i.e. either vector version isn't available, or is too expensive.
3128 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3129 const TargetTransformInfo &TTI,
3130 const TargetLibraryInfo *TLI,
3131 bool &NeedToScalarize) {
3132 Function *F = CI->getCalledFunction();
3133 StringRef FnName = CI->getCalledFunction()->getName();
3134 Type *ScalarRetTy = CI->getType();
3135 SmallVector<Type *, 4> Tys, ScalarTys;
3136 for (auto &ArgOp : CI->arg_operands())
3137 ScalarTys.push_back(ArgOp->getType());
3139 // Estimate cost of scalarized vector call. The source operands are assumed
3140 // to be vectors, so we need to extract individual elements from there,
3141 // execute VF scalar calls, and then gather the result into the vector return
3143 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3145 return ScalarCallCost;
3147 // Compute corresponding vector type for return value and arguments.
3148 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3149 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3150 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3152 // Compute costs of unpacking argument values for the scalar calls and
3153 // packing the return values to a vector.
3154 unsigned ScalarizationCost =
3155 getScalarizationOverhead(RetTy, true, false, TTI);
3156 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3157 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3159 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3161 // If we can't emit a vector call for this function, then the currently found
3162 // cost is the cost we need to return.
3163 NeedToScalarize = true;
3164 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3167 // If the corresponding vector cost is cheaper, return its cost.
3168 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3169 if (VectorCallCost < Cost) {
3170 NeedToScalarize = false;
3171 return VectorCallCost;
3176 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3177 // factor VF. Return the cost of the instruction, including scalarization
3178 // overhead if it's needed.
3179 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3180 const TargetTransformInfo &TTI,
3181 const TargetLibraryInfo *TLI) {
3182 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3183 assert(ID && "Expected intrinsic call!");
3185 Type *RetTy = ToVectorTy(CI->getType(), VF);
3186 SmallVector<Type *, 4> Tys;
3187 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3188 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3190 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3193 void InnerLoopVectorizer::vectorizeLoop() {
3194 //===------------------------------------------------===//
3196 // Notice: any optimization or new instruction that go
3197 // into the code below should be also be implemented in
3200 //===------------------------------------------------===//
3201 Constant *Zero = Builder.getInt32(0);
3203 // In order to support reduction variables we need to be able to vectorize
3204 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3205 // stages. First, we create a new vector PHI node with no incoming edges.
3206 // We use this value when we vectorize all of the instructions that use the
3207 // PHI. Next, after all of the instructions in the block are complete we
3208 // add the new incoming edges to the PHI. At this point all of the
3209 // instructions in the basic block are vectorized, so we can use them to
3210 // construct the PHI.
3211 PhiVector RdxPHIsToFix;
3213 // Scan the loop in a topological order to ensure that defs are vectorized
3215 LoopBlocksDFS DFS(OrigLoop);
3218 // Vectorize all of the blocks in the original loop.
3219 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3220 be = DFS.endRPO(); bb != be; ++bb)
3221 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3223 // At this point every instruction in the original loop is widened to
3224 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3225 // that we vectorized. The PHI nodes are currently empty because we did
3226 // not want to introduce cycles. Notice that the remaining PHI nodes
3227 // that we need to fix are reduction variables.
3229 // Create the 'reduced' values for each of the induction vars.
3230 // The reduced values are the vector values that we scalarize and combine
3231 // after the loop is finished.
3232 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3234 PHINode *RdxPhi = *it;
3235 assert(RdxPhi && "Unable to recover vectorized PHI");
3237 // Find the reduction variable descriptor.
3238 assert(Legal->getReductionVars()->count(RdxPhi) &&
3239 "Unable to find the reduction variable");
3240 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3242 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3243 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3244 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3245 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3246 RdxDesc.getMinMaxRecurrenceKind();
3247 setDebugLocFromInst(Builder, ReductionStartValue);
3249 // We need to generate a reduction vector from the incoming scalar.
3250 // To do so, we need to generate the 'identity' vector and override
3251 // one of the elements with the incoming scalar reduction. We need
3252 // to do it in the vector-loop preheader.
3253 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3255 // This is the vector-clone of the value that leaves the loop.
3256 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3257 Type *VecTy = VectorExit[0]->getType();
3259 // Find the reduction identity variable. Zero for addition, or, xor,
3260 // one for multiplication, -1 for And.
3263 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3264 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3265 // MinMax reduction have the start value as their identify.
3267 VectorStart = Identity = ReductionStartValue;
3269 VectorStart = Identity =
3270 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3273 // Handle other reduction kinds:
3274 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3275 RK, VecTy->getScalarType());
3278 // This vector is the Identity vector where the first element is the
3279 // incoming scalar reduction.
3280 VectorStart = ReductionStartValue;
3282 Identity = ConstantVector::getSplat(VF, Iden);
3284 // This vector is the Identity vector where the first element is the
3285 // incoming scalar reduction.
3287 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3291 // Fix the vector-loop phi.
3293 // Reductions do not have to start at zero. They can start with
3294 // any loop invariant values.
3295 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3296 BasicBlock *Latch = OrigLoop->getLoopLatch();
3297 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3298 VectorParts &Val = getVectorValue(LoopVal);
3299 for (unsigned part = 0; part < UF; ++part) {
3300 // Make sure to add the reduction stat value only to the
3301 // first unroll part.
3302 Value *StartVal = (part == 0) ? VectorStart : Identity;
3303 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3304 LoopVectorPreHeader);
3305 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3306 LoopVectorBody.back());
3309 // Before each round, move the insertion point right between
3310 // the PHIs and the values we are going to write.
3311 // This allows us to write both PHINodes and the extractelement
3313 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3315 VectorParts RdxParts;
3316 setDebugLocFromInst(Builder, LoopExitInst);
3317 for (unsigned part = 0; part < UF; ++part) {
3318 // This PHINode contains the vectorized reduction variable, or
3319 // the initial value vector, if we bypass the vector loop.
3320 VectorParts &RdxExitVal = getVectorValue(LoopExitInst);
3321 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3322 Value *StartVal = (part == 0) ? VectorStart : Identity;
3323 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3324 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3325 NewPhi->addIncoming(RdxExitVal[part],
3326 LoopVectorBody.back());
3327 RdxParts.push_back(NewPhi);
3330 // Reduce all of the unrolled parts into a single vector.
3331 Value *ReducedPartRdx = RdxParts[0];
3332 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3333 setDebugLocFromInst(Builder, ReducedPartRdx);
3334 for (unsigned part = 1; part < UF; ++part) {
3335 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3336 // Floating point operations had to be 'fast' to enable the reduction.
3337 ReducedPartRdx = addFastMathFlag(
3338 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3339 ReducedPartRdx, "bin.rdx"));
3341 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3342 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3346 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3347 // and vector ops, reducing the set of values being computed by half each
3349 assert(isPowerOf2_32(VF) &&
3350 "Reduction emission only supported for pow2 vectors!");
3351 Value *TmpVec = ReducedPartRdx;
3352 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3353 for (unsigned i = VF; i != 1; i >>= 1) {
3354 // Move the upper half of the vector to the lower half.
3355 for (unsigned j = 0; j != i/2; ++j)
3356 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3358 // Fill the rest of the mask with undef.
3359 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3360 UndefValue::get(Builder.getInt32Ty()));
3363 Builder.CreateShuffleVector(TmpVec,
3364 UndefValue::get(TmpVec->getType()),
3365 ConstantVector::get(ShuffleMask),
3368 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3369 // Floating point operations had to be 'fast' to enable the reduction.
3370 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3371 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3373 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3377 // The result is in the first element of the vector.
3378 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3379 Builder.getInt32(0));
3382 // Create a phi node that merges control-flow from the backedge-taken check
3383 // block and the middle block.
3384 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3385 LoopScalarPreHeader->getTerminator());
3386 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3387 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3389 // Now, we need to fix the users of the reduction variable
3390 // inside and outside of the scalar remainder loop.
3391 // We know that the loop is in LCSSA form. We need to update the
3392 // PHI nodes in the exit blocks.
3393 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3394 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3395 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3396 if (!LCSSAPhi) break;
3398 // All PHINodes need to have a single entry edge, or two if
3399 // we already fixed them.
3400 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3402 // We found our reduction value exit-PHI. Update it with the
3403 // incoming bypass edge.
3404 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3405 // Add an edge coming from the bypass.
3406 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3409 }// end of the LCSSA phi scan.
3411 // Fix the scalar loop reduction variable with the incoming reduction sum
3412 // from the vector body and from the backedge value.
3413 int IncomingEdgeBlockIdx =
3414 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3415 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3416 // Pick the other block.
3417 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3418 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3419 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3420 }// end of for each redux variable.
3424 // Remove redundant induction instructions.
3425 cse(LoopVectorBody);
3428 void InnerLoopVectorizer::fixLCSSAPHIs() {
3429 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3430 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3431 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3432 if (!LCSSAPhi) break;
3433 if (LCSSAPhi->getNumIncomingValues() == 1)
3434 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3439 InnerLoopVectorizer::VectorParts
3440 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3441 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3444 // Look for cached value.
3445 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3446 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3447 if (ECEntryIt != MaskCache.end())
3448 return ECEntryIt->second;
3450 VectorParts SrcMask = createBlockInMask(Src);
3452 // The terminator has to be a branch inst!
3453 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3454 assert(BI && "Unexpected terminator found");
3456 if (BI->isConditional()) {
3457 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3459 if (BI->getSuccessor(0) != Dst)
3460 for (unsigned part = 0; part < UF; ++part)
3461 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3463 for (unsigned part = 0; part < UF; ++part)
3464 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3466 MaskCache[Edge] = EdgeMask;
3470 MaskCache[Edge] = SrcMask;
3474 InnerLoopVectorizer::VectorParts
3475 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3476 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3478 // Loop incoming mask is all-one.
3479 if (OrigLoop->getHeader() == BB) {
3480 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3481 return getVectorValue(C);
3484 // This is the block mask. We OR all incoming edges, and with zero.
3485 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3486 VectorParts BlockMask = getVectorValue(Zero);
3489 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3490 VectorParts EM = createEdgeMask(*it, BB);
3491 for (unsigned part = 0; part < UF; ++part)
3492 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3498 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3499 InnerLoopVectorizer::VectorParts &Entry,
3500 unsigned UF, unsigned VF, PhiVector *PV) {
3501 PHINode* P = cast<PHINode>(PN);
3502 // Handle reduction variables:
3503 if (Legal->getReductionVars()->count(P)) {
3504 for (unsigned part = 0; part < UF; ++part) {
3505 // This is phase one of vectorizing PHIs.
3506 Type *VecTy = (VF == 1) ? PN->getType() :
3507 VectorType::get(PN->getType(), VF);
3508 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3509 LoopVectorBody.back()-> getFirstInsertionPt());
3515 setDebugLocFromInst(Builder, P);
3516 // Check for PHI nodes that are lowered to vector selects.
3517 if (P->getParent() != OrigLoop->getHeader()) {
3518 // We know that all PHIs in non-header blocks are converted into
3519 // selects, so we don't have to worry about the insertion order and we
3520 // can just use the builder.
3521 // At this point we generate the predication tree. There may be
3522 // duplications since this is a simple recursive scan, but future
3523 // optimizations will clean it up.
3525 unsigned NumIncoming = P->getNumIncomingValues();
3527 // Generate a sequence of selects of the form:
3528 // SELECT(Mask3, In3,
3529 // SELECT(Mask2, In2,
3531 for (unsigned In = 0; In < NumIncoming; In++) {
3532 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3534 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3536 for (unsigned part = 0; part < UF; ++part) {
3537 // We might have single edge PHIs (blocks) - use an identity
3538 // 'select' for the first PHI operand.
3540 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3543 // Select between the current value and the previous incoming edge
3544 // based on the incoming mask.
3545 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3546 Entry[part], "predphi");
3552 // This PHINode must be an induction variable.
3553 // Make sure that we know about it.
3554 assert(Legal->getInductionVars()->count(P) &&
3555 "Not an induction variable");
3557 LoopVectorizationLegality::InductionInfo II =
3558 Legal->getInductionVars()->lookup(P);
3560 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3561 // which can be found from the original scalar operations.
3563 case LoopVectorizationLegality::IK_NoInduction:
3564 llvm_unreachable("Unknown induction");
3565 case LoopVectorizationLegality::IK_IntInduction: {
3566 assert(P->getType() == II.StartValue->getType() && "Types must match");
3567 Type *PhiTy = P->getType();
3569 if (P == OldInduction) {
3570 // Handle the canonical induction variable. We might have had to
3572 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3574 // Handle other induction variables that are now based on the
3576 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3578 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3579 Broadcasted = II.transform(Builder, NormalizedIdx);
3580 Broadcasted->setName("offset.idx");
3582 Broadcasted = getBroadcastInstrs(Broadcasted);
3583 // After broadcasting the induction variable we need to make the vector
3584 // consecutive by adding 0, 1, 2, etc.
3585 for (unsigned part = 0; part < UF; ++part)
3586 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3589 case LoopVectorizationLegality::IK_PtrInduction:
3590 // Handle the pointer induction variable case.
3591 assert(P->getType()->isPointerTy() && "Unexpected type.");
3592 // This is the normalized GEP that starts counting at zero.
3593 Value *NormalizedIdx =
3594 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3596 Builder.CreateSExtOrTrunc(NormalizedIdx, II.StepValue->getType());
3597 // This is the vector of results. Notice that we don't generate
3598 // vector geps because scalar geps result in better code.
3599 for (unsigned part = 0; part < UF; ++part) {
3601 int EltIndex = part;
3602 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3603 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3604 Value *SclrGep = II.transform(Builder, GlobalIdx);
3605 SclrGep->setName("next.gep");
3606 Entry[part] = SclrGep;
3610 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3611 for (unsigned int i = 0; i < VF; ++i) {
3612 int EltIndex = i + part * VF;
3613 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3614 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3615 Value *SclrGep = II.transform(Builder, GlobalIdx);
3616 SclrGep->setName("next.gep");
3617 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3618 Builder.getInt32(i),
3621 Entry[part] = VecVal;
3627 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3628 // For each instruction in the old loop.
3629 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3630 VectorParts &Entry = WidenMap.get(it);
3631 switch (it->getOpcode()) {
3632 case Instruction::Br:
3633 // Nothing to do for PHIs and BR, since we already took care of the
3634 // loop control flow instructions.
3636 case Instruction::PHI: {
3637 // Vectorize PHINodes.
3638 widenPHIInstruction(it, Entry, UF, VF, PV);
3642 case Instruction::Add:
3643 case Instruction::FAdd:
3644 case Instruction::Sub:
3645 case Instruction::FSub:
3646 case Instruction::Mul:
3647 case Instruction::FMul:
3648 case Instruction::UDiv:
3649 case Instruction::SDiv:
3650 case Instruction::FDiv:
3651 case Instruction::URem:
3652 case Instruction::SRem:
3653 case Instruction::FRem:
3654 case Instruction::Shl:
3655 case Instruction::LShr:
3656 case Instruction::AShr:
3657 case Instruction::And:
3658 case Instruction::Or:
3659 case Instruction::Xor: {
3660 // Just widen binops.
3661 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3662 setDebugLocFromInst(Builder, BinOp);
3663 VectorParts &A = getVectorValue(it->getOperand(0));
3664 VectorParts &B = getVectorValue(it->getOperand(1));
3666 // Use this vector value for all users of the original instruction.
3667 for (unsigned Part = 0; Part < UF; ++Part) {
3668 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3670 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3671 VecOp->copyIRFlags(BinOp);
3676 propagateMetadata(Entry, it);
3679 case Instruction::Select: {
3681 // If the selector is loop invariant we can create a select
3682 // instruction with a scalar condition. Otherwise, use vector-select.
3683 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3685 setDebugLocFromInst(Builder, it);
3687 // The condition can be loop invariant but still defined inside the
3688 // loop. This means that we can't just use the original 'cond' value.
3689 // We have to take the 'vectorized' value and pick the first lane.
3690 // Instcombine will make this a no-op.
3691 VectorParts &Cond = getVectorValue(it->getOperand(0));
3692 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3693 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3695 Value *ScalarCond = (VF == 1) ? Cond[0] :
3696 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3698 for (unsigned Part = 0; Part < UF; ++Part) {
3699 Entry[Part] = Builder.CreateSelect(
3700 InvariantCond ? ScalarCond : Cond[Part],
3705 propagateMetadata(Entry, it);
3709 case Instruction::ICmp:
3710 case Instruction::FCmp: {
3711 // Widen compares. Generate vector compares.
3712 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3713 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3714 setDebugLocFromInst(Builder, it);
3715 VectorParts &A = getVectorValue(it->getOperand(0));
3716 VectorParts &B = getVectorValue(it->getOperand(1));
3717 for (unsigned Part = 0; Part < UF; ++Part) {
3720 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3722 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3726 propagateMetadata(Entry, it);
3730 case Instruction::Store:
3731 case Instruction::Load:
3732 vectorizeMemoryInstruction(it);
3734 case Instruction::ZExt:
3735 case Instruction::SExt:
3736 case Instruction::FPToUI:
3737 case Instruction::FPToSI:
3738 case Instruction::FPExt:
3739 case Instruction::PtrToInt:
3740 case Instruction::IntToPtr:
3741 case Instruction::SIToFP:
3742 case Instruction::UIToFP:
3743 case Instruction::Trunc:
3744 case Instruction::FPTrunc:
3745 case Instruction::BitCast: {
3746 CastInst *CI = dyn_cast<CastInst>(it);
3747 setDebugLocFromInst(Builder, it);
3748 /// Optimize the special case where the source is the induction
3749 /// variable. Notice that we can only optimize the 'trunc' case
3750 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3751 /// c. other casts depend on pointer size.
3752 if (CI->getOperand(0) == OldInduction &&
3753 it->getOpcode() == Instruction::Trunc) {
3754 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3756 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3757 LoopVectorizationLegality::InductionInfo II =
3758 Legal->getInductionVars()->lookup(OldInduction);
3760 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3761 for (unsigned Part = 0; Part < UF; ++Part)
3762 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3763 propagateMetadata(Entry, it);
3766 /// Vectorize casts.
3767 Type *DestTy = (VF == 1) ? CI->getType() :
3768 VectorType::get(CI->getType(), VF);
3770 VectorParts &A = getVectorValue(it->getOperand(0));
3771 for (unsigned Part = 0; Part < UF; ++Part)
3772 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3773 propagateMetadata(Entry, it);
3777 case Instruction::Call: {
3778 // Ignore dbg intrinsics.
3779 if (isa<DbgInfoIntrinsic>(it))
3781 setDebugLocFromInst(Builder, it);
3783 Module *M = BB->getParent()->getParent();
3784 CallInst *CI = cast<CallInst>(it);
3786 StringRef FnName = CI->getCalledFunction()->getName();
3787 Function *F = CI->getCalledFunction();
3788 Type *RetTy = ToVectorTy(CI->getType(), VF);
3789 SmallVector<Type *, 4> Tys;
3790 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3791 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3793 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3795 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3796 ID == Intrinsic::lifetime_start)) {
3797 scalarizeInstruction(it);
3800 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3801 // version of the instruction.
3802 // Is it beneficial to perform intrinsic call compared to lib call?
3803 bool NeedToScalarize;
3804 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3805 bool UseVectorIntrinsic =
3806 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3807 if (!UseVectorIntrinsic && NeedToScalarize) {
3808 scalarizeInstruction(it);
3812 for (unsigned Part = 0; Part < UF; ++Part) {
3813 SmallVector<Value *, 4> Args;
3814 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3815 Value *Arg = CI->getArgOperand(i);
3816 // Some intrinsics have a scalar argument - don't replace it with a
3818 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3819 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3820 Arg = VectorArg[Part];
3822 Args.push_back(Arg);
3826 if (UseVectorIntrinsic) {
3827 // Use vector version of the intrinsic.
3828 Type *TysForDecl[] = {CI->getType()};
3830 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3831 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3833 // Use vector version of the library call.
3834 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3835 assert(!VFnName.empty() && "Vector function name is empty.");
3836 VectorF = M->getFunction(VFnName);
3838 // Generate a declaration
3839 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3841 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3842 VectorF->copyAttributesFrom(F);
3845 assert(VectorF && "Can't create vector function.");
3846 Entry[Part] = Builder.CreateCall(VectorF, Args);
3849 propagateMetadata(Entry, it);
3854 // All other instructions are unsupported. Scalarize them.
3855 scalarizeInstruction(it);
3858 }// end of for_each instr.
3861 void InnerLoopVectorizer::updateAnalysis() {
3862 // Forget the original basic block.
3863 SE->forgetLoop(OrigLoop);
3865 // Update the dominator tree information.
3866 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3867 "Entry does not dominate exit.");
3869 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3870 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3871 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3873 // Due to if predication of stores we might create a sequence of "if(pred)
3874 // a[i] = ...; " blocks.
3875 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3877 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3878 else if (isPredicatedBlock(i)) {
3879 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3881 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3885 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3886 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3887 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3888 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3890 DEBUG(DT->verifyDomTree());
3893 /// \brief Check whether it is safe to if-convert this phi node.
3895 /// Phi nodes with constant expressions that can trap are not safe to if
3897 static bool canIfConvertPHINodes(BasicBlock *BB) {
3898 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3899 PHINode *Phi = dyn_cast<PHINode>(I);
3902 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3903 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3910 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3911 if (!EnableIfConversion) {
3912 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3916 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3918 // A list of pointers that we can safely read and write to.
3919 SmallPtrSet<Value *, 8> SafePointes;
3921 // Collect safe addresses.
3922 for (Loop::block_iterator BI = TheLoop->block_begin(),
3923 BE = TheLoop->block_end(); BI != BE; ++BI) {
3924 BasicBlock *BB = *BI;
3926 if (blockNeedsPredication(BB))
3929 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3930 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3931 SafePointes.insert(LI->getPointerOperand());
3932 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3933 SafePointes.insert(SI->getPointerOperand());
3937 // Collect the blocks that need predication.
3938 BasicBlock *Header = TheLoop->getHeader();
3939 for (Loop::block_iterator BI = TheLoop->block_begin(),
3940 BE = TheLoop->block_end(); BI != BE; ++BI) {
3941 BasicBlock *BB = *BI;
3943 // We don't support switch statements inside loops.
3944 if (!isa<BranchInst>(BB->getTerminator())) {
3945 emitAnalysis(VectorizationReport(BB->getTerminator())
3946 << "loop contains a switch statement");
3950 // We must be able to predicate all blocks that need to be predicated.
3951 if (blockNeedsPredication(BB)) {
3952 if (!blockCanBePredicated(BB, SafePointes)) {
3953 emitAnalysis(VectorizationReport(BB->getTerminator())
3954 << "control flow cannot be substituted for a select");
3957 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3958 emitAnalysis(VectorizationReport(BB->getTerminator())
3959 << "control flow cannot be substituted for a select");
3964 // We can if-convert this loop.
3968 bool LoopVectorizationLegality::canVectorize() {
3969 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3970 // be canonicalized.
3971 if (!TheLoop->getLoopPreheader()) {
3973 VectorizationReport() <<
3974 "loop control flow is not understood by vectorizer");
3978 // We can only vectorize innermost loops.
3979 if (!TheLoop->empty()) {
3980 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3984 // We must have a single backedge.
3985 if (TheLoop->getNumBackEdges() != 1) {
3987 VectorizationReport() <<
3988 "loop control flow is not understood by vectorizer");
3992 // We must have a single exiting block.
3993 if (!TheLoop->getExitingBlock()) {
3995 VectorizationReport() <<
3996 "loop control flow is not understood by vectorizer");
4000 // We only handle bottom-tested loops, i.e. loop in which the condition is
4001 // checked at the end of each iteration. With that we can assume that all
4002 // instructions in the loop are executed the same number of times.
4003 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
4005 VectorizationReport() <<
4006 "loop control flow is not understood by vectorizer");
4010 // We need to have a loop header.
4011 DEBUG(dbgs() << "LV: Found a loop: " <<
4012 TheLoop->getHeader()->getName() << '\n');
4014 // Check if we can if-convert non-single-bb loops.
4015 unsigned NumBlocks = TheLoop->getNumBlocks();
4016 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4017 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4021 // ScalarEvolution needs to be able to find the exit count.
4022 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
4023 if (ExitCount == SE->getCouldNotCompute()) {
4024 emitAnalysis(VectorizationReport() <<
4025 "could not determine number of loop iterations");
4026 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4030 // Check if we can vectorize the instructions and CFG in this loop.
4031 if (!canVectorizeInstrs()) {
4032 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4036 // Go over each instruction and look at memory deps.
4037 if (!canVectorizeMemory()) {
4038 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4042 // Collect all of the variables that remain uniform after vectorization.
4043 collectLoopUniforms();
4045 DEBUG(dbgs() << "LV: We can vectorize this loop"
4046 << (LAI->getRuntimePointerChecking()->Need
4047 ? " (with a runtime bound check)"
4051 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4053 // If an override option has been passed in for interleaved accesses, use it.
4054 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4055 UseInterleaved = EnableInterleavedMemAccesses;
4057 // Analyze interleaved memory accesses.
4059 InterleaveInfo.analyzeInterleaving(Strides);
4061 // Okay! We can vectorize. At this point we don't have any other mem analysis
4062 // which may limit our maximum vectorization factor, so just return true with
4067 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4068 if (Ty->isPointerTy())
4069 return DL.getIntPtrType(Ty);
4071 // It is possible that char's or short's overflow when we ask for the loop's
4072 // trip count, work around this by changing the type size.
4073 if (Ty->getScalarSizeInBits() < 32)
4074 return Type::getInt32Ty(Ty->getContext());
4079 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4080 Ty0 = convertPointerToIntegerType(DL, Ty0);
4081 Ty1 = convertPointerToIntegerType(DL, Ty1);
4082 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4087 /// \brief Check that the instruction has outside loop users and is not an
4088 /// identified reduction variable.
4089 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4090 SmallPtrSetImpl<Value *> &Reductions) {
4091 // Reduction instructions are allowed to have exit users. All other
4092 // instructions must not have external users.
4093 if (!Reductions.count(Inst))
4094 //Check that all of the users of the loop are inside the BB.
4095 for (User *U : Inst->users()) {
4096 Instruction *UI = cast<Instruction>(U);
4097 // This user may be a reduction exit value.
4098 if (!TheLoop->contains(UI)) {
4099 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4106 bool LoopVectorizationLegality::canVectorizeInstrs() {
4107 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
4108 BasicBlock *Header = TheLoop->getHeader();
4110 // Look for the attribute signaling the absence of NaNs.
4111 Function &F = *Header->getParent();
4112 const DataLayout &DL = F.getParent()->getDataLayout();
4113 if (F.hasFnAttribute("no-nans-fp-math"))
4115 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4117 // For each block in the loop.
4118 for (Loop::block_iterator bb = TheLoop->block_begin(),
4119 be = TheLoop->block_end(); bb != be; ++bb) {
4121 // Scan the instructions in the block and look for hazards.
4122 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4125 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4126 Type *PhiTy = Phi->getType();
4127 // Check that this PHI type is allowed.
4128 if (!PhiTy->isIntegerTy() &&
4129 !PhiTy->isFloatingPointTy() &&
4130 !PhiTy->isPointerTy()) {
4131 emitAnalysis(VectorizationReport(it)
4132 << "loop control flow is not understood by vectorizer");
4133 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4137 // If this PHINode is not in the header block, then we know that we
4138 // can convert it to select during if-conversion. No need to check if
4139 // the PHIs in this block are induction or reduction variables.
4140 if (*bb != Header) {
4141 // Check that this instruction has no outside users or is an
4142 // identified reduction value with an outside user.
4143 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
4145 emitAnalysis(VectorizationReport(it) <<
4146 "value could not be identified as "
4147 "an induction or reduction variable");
4151 // We only allow if-converted PHIs with exactly two incoming values.
4152 if (Phi->getNumIncomingValues() != 2) {
4153 emitAnalysis(VectorizationReport(it)
4154 << "control flow not understood by vectorizer");
4155 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4159 // This is the value coming from the preheader.
4160 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
4161 ConstantInt *StepValue = nullptr;
4162 // Check if this is an induction variable.
4163 InductionKind IK = isInductionVariable(Phi, StepValue);
4165 if (IK_NoInduction != IK) {
4166 // Get the widest type.
4168 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4170 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4172 // Int inductions are special because we only allow one IV.
4173 if (IK == IK_IntInduction && StepValue->isOne()) {
4174 // Use the phi node with the widest type as induction. Use the last
4175 // one if there are multiple (no good reason for doing this other
4176 // than it is expedient).
4177 if (!Induction || PhiTy == WidestIndTy)
4181 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4182 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
4184 // Until we explicitly handle the case of an induction variable with
4185 // an outside loop user we have to give up vectorizing this loop.
4186 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4187 emitAnalysis(VectorizationReport(it) <<
4188 "use of induction value outside of the "
4189 "loop is not handled by vectorizer");
4196 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4198 if (Reductions[Phi].hasUnsafeAlgebra())
4199 Requirements->addUnsafeAlgebraInst(
4200 Reductions[Phi].getUnsafeAlgebraInst());
4201 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4205 emitAnalysis(VectorizationReport(it) <<
4206 "value that could not be identified as "
4207 "reduction is used outside the loop");
4208 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4210 }// end of PHI handling
4212 // We handle calls that:
4213 // * Are debug info intrinsics.
4214 // * Have a mapping to an IR intrinsic.
4215 // * Have a vector version available.
4216 CallInst *CI = dyn_cast<CallInst>(it);
4217 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4218 !(CI->getCalledFunction() && TLI &&
4219 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4220 emitAnalysis(VectorizationReport(it) <<
4221 "call instruction cannot be vectorized");
4222 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4226 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4227 // second argument is the same (i.e. loop invariant)
4229 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4230 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4231 emitAnalysis(VectorizationReport(it)
4232 << "intrinsic instruction cannot be vectorized");
4233 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4238 // Check that the instruction return type is vectorizable.
4239 // Also, we can't vectorize extractelement instructions.
4240 if ((!VectorType::isValidElementType(it->getType()) &&
4241 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4242 emitAnalysis(VectorizationReport(it)
4243 << "instruction return type cannot be vectorized");
4244 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4248 // Check that the stored type is vectorizable.
4249 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4250 Type *T = ST->getValueOperand()->getType();
4251 if (!VectorType::isValidElementType(T)) {
4252 emitAnalysis(VectorizationReport(ST) <<
4253 "store instruction cannot be vectorized");
4256 if (EnableMemAccessVersioning)
4257 collectStridedAccess(ST);
4260 if (EnableMemAccessVersioning)
4261 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4262 collectStridedAccess(LI);
4264 // Reduction instructions are allowed to have exit users.
4265 // All other instructions must not have external users.
4266 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4267 emitAnalysis(VectorizationReport(it) <<
4268 "value cannot be used outside the loop");
4277 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4278 if (Inductions.empty()) {
4279 emitAnalysis(VectorizationReport()
4280 << "loop induction variable could not be identified");
4288 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4289 Value *Ptr = nullptr;
4290 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4291 Ptr = LI->getPointerOperand();
4292 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4293 Ptr = SI->getPointerOperand();
4297 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4301 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4302 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4303 Strides[Ptr] = Stride;
4304 StrideSet.insert(Stride);
4307 void LoopVectorizationLegality::collectLoopUniforms() {
4308 // We now know that the loop is vectorizable!
4309 // Collect variables that will remain uniform after vectorization.
4310 std::vector<Value*> Worklist;
4311 BasicBlock *Latch = TheLoop->getLoopLatch();
4313 // Start with the conditional branch and walk up the block.
4314 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4316 // Also add all consecutive pointer values; these values will be uniform
4317 // after vectorization (and subsequent cleanup) and, until revectorization is
4318 // supported, all dependencies must also be uniform.
4319 for (Loop::block_iterator B = TheLoop->block_begin(),
4320 BE = TheLoop->block_end(); B != BE; ++B)
4321 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4323 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4324 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4326 while (!Worklist.empty()) {
4327 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4328 Worklist.pop_back();
4330 // Look at instructions inside this loop.
4331 // Stop when reaching PHI nodes.
4332 // TODO: we need to follow values all over the loop, not only in this block.
4333 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4336 // This is a known uniform.
4339 // Insert all operands.
4340 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4344 bool LoopVectorizationLegality::canVectorizeMemory() {
4345 LAI = &LAA->getInfo(TheLoop, Strides);
4346 auto &OptionalReport = LAI->getReport();
4348 emitAnalysis(VectorizationReport(*OptionalReport));
4349 if (!LAI->canVectorizeMemory())
4352 if (LAI->hasStoreToLoopInvariantAddress()) {
4354 VectorizationReport()
4355 << "write to a loop invariant address could not be vectorized");
4356 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4360 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4365 LoopVectorizationLegality::InductionKind
4366 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4367 ConstantInt *&StepValue) {
4368 if (!isInductionPHI(Phi, SE, StepValue))
4369 return IK_NoInduction;
4371 Type *PhiTy = Phi->getType();
4372 // Found an Integer induction variable.
4373 if (PhiTy->isIntegerTy())
4374 return IK_IntInduction;
4375 // Found an Pointer induction variable.
4376 return IK_PtrInduction;
4379 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4380 Value *In0 = const_cast<Value*>(V);
4381 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4385 return Inductions.count(PN);
4388 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4389 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4392 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4393 SmallPtrSetImpl<Value *> &SafePtrs) {
4395 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4396 // Check that we don't have a constant expression that can trap as operand.
4397 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4399 if (Constant *C = dyn_cast<Constant>(*OI))
4403 // We might be able to hoist the load.
4404 if (it->mayReadFromMemory()) {
4405 LoadInst *LI = dyn_cast<LoadInst>(it);
4408 if (!SafePtrs.count(LI->getPointerOperand())) {
4409 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4410 MaskedOp.insert(LI);
4417 // We don't predicate stores at the moment.
4418 if (it->mayWriteToMemory()) {
4419 StoreInst *SI = dyn_cast<StoreInst>(it);
4420 // We only support predication of stores in basic blocks with one
4425 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4426 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4428 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4429 !isSinglePredecessor) {
4430 // Build a masked store if it is legal for the target, otherwise scalarize
4432 bool isLegalMaskedOp =
4433 isLegalMaskedStore(SI->getValueOperand()->getType(),
4434 SI->getPointerOperand());
4435 if (isLegalMaskedOp) {
4437 MaskedOp.insert(SI);
4446 // The instructions below can trap.
4447 switch (it->getOpcode()) {
4449 case Instruction::UDiv:
4450 case Instruction::SDiv:
4451 case Instruction::URem:
4452 case Instruction::SRem:
4460 void InterleavedAccessInfo::collectConstStridedAccesses(
4461 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4462 const ValueToValueMap &Strides) {
4463 // Holds load/store instructions in program order.
4464 SmallVector<Instruction *, 16> AccessList;
4466 for (auto *BB : TheLoop->getBlocks()) {
4467 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4469 for (auto &I : *BB) {
4470 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4472 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4476 AccessList.push_back(&I);
4480 if (AccessList.empty())
4483 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4484 for (auto I : AccessList) {
4485 LoadInst *LI = dyn_cast<LoadInst>(I);
4486 StoreInst *SI = dyn_cast<StoreInst>(I);
4488 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4489 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4491 // The factor of the corresponding interleave group.
4492 unsigned Factor = std::abs(Stride);
4494 // Ignore the access if the factor is too small or too large.
4495 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4498 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4499 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4500 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4502 // An alignment of 0 means target ABI alignment.
4503 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4505 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4507 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4511 // Analyze interleaved accesses and collect them into interleave groups.
4513 // Notice that the vectorization on interleaved groups will change instruction
4514 // orders and may break dependences. But the memory dependence check guarantees
4515 // that there is no overlap between two pointers of different strides, element
4516 // sizes or underlying bases.
4518 // For pointers sharing the same stride, element size and underlying base, no
4519 // need to worry about Read-After-Write dependences and Write-After-Read
4522 // E.g. The RAW dependence: A[i] = a;
4524 // This won't exist as it is a store-load forwarding conflict, which has
4525 // already been checked and forbidden in the dependence check.
4527 // E.g. The WAR dependence: a = A[i]; // (1)
4529 // The store group of (2) is always inserted at or below (2), and the load group
4530 // of (1) is always inserted at or above (1). The dependence is safe.
4531 void InterleavedAccessInfo::analyzeInterleaving(
4532 const ValueToValueMap &Strides) {
4533 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4535 // Holds all the stride accesses.
4536 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4537 collectConstStridedAccesses(StrideAccesses, Strides);
4539 if (StrideAccesses.empty())
4542 // Holds all interleaved store groups temporarily.
4543 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4545 // Search the load-load/write-write pair B-A in bottom-up order and try to
4546 // insert B into the interleave group of A according to 3 rules:
4547 // 1. A and B have the same stride.
4548 // 2. A and B have the same memory object size.
4549 // 3. B belongs to the group according to the distance.
4551 // The bottom-up order can avoid breaking the Write-After-Write dependences
4552 // between two pointers of the same base.
4553 // E.g. A[i] = a; (1)
4556 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4557 // above (1), which guarantees that (1) is always above (2).
4558 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4560 Instruction *A = I->first;
4561 StrideDescriptor DesA = I->second;
4563 InterleaveGroup *Group = getInterleaveGroup(A);
4565 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4566 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4569 if (A->mayWriteToMemory())
4570 StoreGroups.insert(Group);
4572 for (auto II = std::next(I); II != E; ++II) {
4573 Instruction *B = II->first;
4574 StrideDescriptor DesB = II->second;
4576 // Ignore if B is already in a group or B is a different memory operation.
4577 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4580 // Check the rule 1 and 2.
4581 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4584 // Calculate the distance and prepare for the rule 3.
4585 const SCEVConstant *DistToA =
4586 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4590 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4592 // Skip if the distance is not multiple of size as they are not in the
4594 if (DistanceToA % static_cast<int>(DesA.Size))
4597 // The index of B is the index of A plus the related index to A.
4599 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4601 // Try to insert B into the group.
4602 if (Group->insertMember(B, IndexB, DesB.Align)) {
4603 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4604 << " into the interleave group with" << *A << '\n');
4605 InterleaveGroupMap[B] = Group;
4607 // Set the first load in program order as the insert position.
4608 if (B->mayReadFromMemory())
4609 Group->setInsertPos(B);
4611 } // Iteration on instruction B
4612 } // Iteration on instruction A
4614 // Remove interleaved store groups with gaps.
4615 for (InterleaveGroup *Group : StoreGroups)
4616 if (Group->getNumMembers() != Group->getFactor())
4617 releaseGroup(Group);
4620 LoopVectorizationCostModel::VectorizationFactor
4621 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4622 // Width 1 means no vectorize
4623 VectorizationFactor Factor = { 1U, 0U };
4624 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4625 emitAnalysis(VectorizationReport() <<
4626 "runtime pointer checks needed. Enable vectorization of this "
4627 "loop with '#pragma clang loop vectorize(enable)' when "
4628 "compiling with -Os/-Oz");
4630 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4634 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4635 emitAnalysis(VectorizationReport() <<
4636 "store that is conditionally executed prevents vectorization");
4637 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4641 // Find the trip count.
4642 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4643 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4645 unsigned WidestType = getWidestType();
4646 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4647 unsigned MaxSafeDepDist = -1U;
4648 if (Legal->getMaxSafeDepDistBytes() != -1U)
4649 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4650 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4651 WidestRegister : MaxSafeDepDist);
4652 unsigned MaxVectorSize = WidestRegister / WidestType;
4653 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4654 DEBUG(dbgs() << "LV: The Widest register is: "
4655 << WidestRegister << " bits.\n");
4657 if (MaxVectorSize == 0) {
4658 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4662 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4663 " into one vector!");
4665 unsigned VF = MaxVectorSize;
4667 // If we optimize the program for size, avoid creating the tail loop.
4669 // If we are unable to calculate the trip count then don't try to vectorize.
4672 (VectorizationReport() <<
4673 "unable to calculate the loop count due to complex control flow");
4674 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4678 // Find the maximum SIMD width that can fit within the trip count.
4679 VF = TC % MaxVectorSize;
4684 // If the trip count that we found modulo the vectorization factor is not
4685 // zero then we require a tail.
4686 emitAnalysis(VectorizationReport() <<
4687 "cannot optimize for size and vectorize at the "
4688 "same time. Enable vectorization of this loop "
4689 "with '#pragma clang loop vectorize(enable)' "
4690 "when compiling with -Os/-Oz");
4691 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4696 int UserVF = Hints->getWidth();
4698 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4699 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4701 Factor.Width = UserVF;
4705 float Cost = expectedCost(1);
4707 const float ScalarCost = Cost;
4710 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4712 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4713 // Ignore scalar width, because the user explicitly wants vectorization.
4714 if (ForceVectorization && VF > 1) {
4716 Cost = expectedCost(Width) / (float)Width;
4719 for (unsigned i=2; i <= VF; i*=2) {
4720 // Notice that the vector loop needs to be executed less times, so
4721 // we need to divide the cost of the vector loops by the width of
4722 // the vector elements.
4723 float VectorCost = expectedCost(i) / (float)i;
4724 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4725 (int)VectorCost << ".\n");
4726 if (VectorCost < Cost) {
4732 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4733 << "LV: Vectorization seems to be not beneficial, "
4734 << "but was forced by a user.\n");
4735 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4736 Factor.Width = Width;
4737 Factor.Cost = Width * Cost;
4741 unsigned LoopVectorizationCostModel::getWidestType() {
4742 unsigned MaxWidth = 8;
4743 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4746 for (Loop::block_iterator bb = TheLoop->block_begin(),
4747 be = TheLoop->block_end(); bb != be; ++bb) {
4748 BasicBlock *BB = *bb;
4750 // For each instruction in the loop.
4751 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4752 Type *T = it->getType();
4754 // Ignore ephemeral values.
4755 if (EphValues.count(it))
4758 // Only examine Loads, Stores and PHINodes.
4759 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4762 // Examine PHI nodes that are reduction variables.
4763 if (PHINode *PN = dyn_cast<PHINode>(it))
4764 if (!Legal->getReductionVars()->count(PN))
4767 // Examine the stored values.
4768 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4769 T = ST->getValueOperand()->getType();
4771 // Ignore loaded pointer types and stored pointer types that are not
4772 // consecutive. However, we do want to take consecutive stores/loads of
4773 // pointer vectors into account.
4774 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4777 MaxWidth = std::max(MaxWidth,
4778 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4785 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4787 unsigned LoopCost) {
4789 // -- The interleave heuristics --
4790 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4791 // There are many micro-architectural considerations that we can't predict
4792 // at this level. For example, frontend pressure (on decode or fetch) due to
4793 // code size, or the number and capabilities of the execution ports.
4795 // We use the following heuristics to select the interleave count:
4796 // 1. If the code has reductions, then we interleave to break the cross
4797 // iteration dependency.
4798 // 2. If the loop is really small, then we interleave to reduce the loop
4800 // 3. We don't interleave if we think that we will spill registers to memory
4801 // due to the increased register pressure.
4803 // When we optimize for size, we don't interleave.
4807 // We used the distance for the interleave count.
4808 if (Legal->getMaxSafeDepDistBytes() != -1U)
4811 // Do not interleave loops with a relatively small trip count.
4812 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4813 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4816 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4817 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4821 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4822 TargetNumRegisters = ForceTargetNumScalarRegs;
4824 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4825 TargetNumRegisters = ForceTargetNumVectorRegs;
4828 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4829 // We divide by these constants so assume that we have at least one
4830 // instruction that uses at least one register.
4831 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4832 R.NumInstructions = std::max(R.NumInstructions, 1U);
4834 // We calculate the interleave count using the following formula.
4835 // Subtract the number of loop invariants from the number of available
4836 // registers. These registers are used by all of the interleaved instances.
4837 // Next, divide the remaining registers by the number of registers that is
4838 // required by the loop, in order to estimate how many parallel instances
4839 // fit without causing spills. All of this is rounded down if necessary to be
4840 // a power of two. We want power of two interleave count to simplify any
4841 // addressing operations or alignment considerations.
4842 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4845 // Don't count the induction variable as interleaved.
4846 if (EnableIndVarRegisterHeur)
4847 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4848 std::max(1U, (R.MaxLocalUsers - 1)));
4850 // Clamp the interleave ranges to reasonable counts.
4851 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4853 // Check if the user has overridden the max.
4855 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4856 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4858 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4859 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4862 // If we did not calculate the cost for VF (because the user selected the VF)
4863 // then we calculate the cost of VF here.
4865 LoopCost = expectedCost(VF);
4867 // Clamp the calculated IC to be between the 1 and the max interleave count
4868 // that the target allows.
4869 if (IC > MaxInterleaveCount)
4870 IC = MaxInterleaveCount;
4874 // Interleave if we vectorized this loop and there is a reduction that could
4875 // benefit from interleaving.
4876 if (VF > 1 && Legal->getReductionVars()->size()) {
4877 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4881 // Note that if we've already vectorized the loop we will have done the
4882 // runtime check and so interleaving won't require further checks.
4883 bool InterleavingRequiresRuntimePointerCheck =
4884 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4886 // We want to interleave small loops in order to reduce the loop overhead and
4887 // potentially expose ILP opportunities.
4888 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4889 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4890 // We assume that the cost overhead is 1 and we use the cost model
4891 // to estimate the cost of the loop and interleave until the cost of the
4892 // loop overhead is about 5% of the cost of the loop.
4894 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4896 // Interleave until store/load ports (estimated by max interleave count) are
4898 unsigned NumStores = Legal->getNumStores();
4899 unsigned NumLoads = Legal->getNumLoads();
4900 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4901 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4903 // If we have a scalar reduction (vector reductions are already dealt with
4904 // by this point), we can increase the critical path length if the loop
4905 // we're interleaving is inside another loop. Limit, by default to 2, so the
4906 // critical path only gets increased by one reduction operation.
4907 if (Legal->getReductionVars()->size() &&
4908 TheLoop->getLoopDepth() > 1) {
4909 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4910 SmallIC = std::min(SmallIC, F);
4911 StoresIC = std::min(StoresIC, F);
4912 LoadsIC = std::min(LoadsIC, F);
4915 if (EnableLoadStoreRuntimeInterleave &&
4916 std::max(StoresIC, LoadsIC) > SmallIC) {
4917 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4918 return std::max(StoresIC, LoadsIC);
4921 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4925 // Interleave if this is a large loop (small loops are already dealt with by
4927 // point) that could benefit from interleaving.
4928 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4929 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4930 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4934 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4938 LoopVectorizationCostModel::RegisterUsage
4939 LoopVectorizationCostModel::calculateRegisterUsage() {
4940 // This function calculates the register usage by measuring the highest number
4941 // of values that are alive at a single location. Obviously, this is a very
4942 // rough estimation. We scan the loop in a topological order in order and
4943 // assign a number to each instruction. We use RPO to ensure that defs are
4944 // met before their users. We assume that each instruction that has in-loop
4945 // users starts an interval. We record every time that an in-loop value is
4946 // used, so we have a list of the first and last occurrences of each
4947 // instruction. Next, we transpose this data structure into a multi map that
4948 // holds the list of intervals that *end* at a specific location. This multi
4949 // map allows us to perform a linear search. We scan the instructions linearly
4950 // and record each time that a new interval starts, by placing it in a set.
4951 // If we find this value in the multi-map then we remove it from the set.
4952 // The max register usage is the maximum size of the set.
4953 // We also search for instructions that are defined outside the loop, but are
4954 // used inside the loop. We need this number separately from the max-interval
4955 // usage number because when we unroll, loop-invariant values do not take
4957 LoopBlocksDFS DFS(TheLoop);
4961 R.NumInstructions = 0;
4963 // Each 'key' in the map opens a new interval. The values
4964 // of the map are the index of the 'last seen' usage of the
4965 // instruction that is the key.
4966 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4967 // Maps instruction to its index.
4968 DenseMap<unsigned, Instruction*> IdxToInstr;
4969 // Marks the end of each interval.
4970 IntervalMap EndPoint;
4971 // Saves the list of instruction indices that are used in the loop.
4972 SmallSet<Instruction*, 8> Ends;
4973 // Saves the list of values that are used in the loop but are
4974 // defined outside the loop, such as arguments and constants.
4975 SmallPtrSet<Value*, 8> LoopInvariants;
4978 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4979 be = DFS.endRPO(); bb != be; ++bb) {
4980 R.NumInstructions += (*bb)->size();
4981 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4983 Instruction *I = it;
4984 IdxToInstr[Index++] = I;
4986 // Save the end location of each USE.
4987 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4988 Value *U = I->getOperand(i);
4989 Instruction *Instr = dyn_cast<Instruction>(U);
4991 // Ignore non-instruction values such as arguments, constants, etc.
4992 if (!Instr) continue;
4994 // If this instruction is outside the loop then record it and continue.
4995 if (!TheLoop->contains(Instr)) {
4996 LoopInvariants.insert(Instr);
5000 // Overwrite previous end points.
5001 EndPoint[Instr] = Index;
5007 // Saves the list of intervals that end with the index in 'key'.
5008 typedef SmallVector<Instruction*, 2> InstrList;
5009 DenseMap<unsigned, InstrList> TransposeEnds;
5011 // Transpose the EndPoints to a list of values that end at each index.
5012 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5014 TransposeEnds[it->second].push_back(it->first);
5016 SmallSet<Instruction*, 8> OpenIntervals;
5017 unsigned MaxUsage = 0;
5020 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5021 for (unsigned int i = 0; i < Index; ++i) {
5022 Instruction *I = IdxToInstr[i];
5023 // Ignore instructions that are never used within the loop.
5024 if (!Ends.count(I)) continue;
5026 // Ignore ephemeral values.
5027 if (EphValues.count(I))
5030 // Remove all of the instructions that end at this location.
5031 InstrList &List = TransposeEnds[i];
5032 for (unsigned int j=0, e = List.size(); j < e; ++j)
5033 OpenIntervals.erase(List[j]);
5035 // Count the number of live interals.
5036 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5038 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5039 OpenIntervals.size() << '\n');
5041 // Add the current instruction to the list of open intervals.
5042 OpenIntervals.insert(I);
5045 unsigned Invariant = LoopInvariants.size();
5046 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5047 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5048 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5050 R.LoopInvariantRegs = Invariant;
5051 R.MaxLocalUsers = MaxUsage;
5055 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5059 for (Loop::block_iterator bb = TheLoop->block_begin(),
5060 be = TheLoop->block_end(); bb != be; ++bb) {
5061 unsigned BlockCost = 0;
5062 BasicBlock *BB = *bb;
5064 // For each instruction in the old loop.
5065 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5066 // Skip dbg intrinsics.
5067 if (isa<DbgInfoIntrinsic>(it))
5070 // Ignore ephemeral values.
5071 if (EphValues.count(it))
5074 unsigned C = getInstructionCost(it, VF);
5076 // Check if we should override the cost.
5077 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5078 C = ForceTargetInstructionCost;
5081 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5082 VF << " For instruction: " << *it << '\n');
5085 // We assume that if-converted blocks have a 50% chance of being executed.
5086 // When the code is scalar then some of the blocks are avoided due to CF.
5087 // When the code is vectorized we execute all code paths.
5088 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5097 /// \brief Check whether the address computation for a non-consecutive memory
5098 /// access looks like an unlikely candidate for being merged into the indexing
5101 /// We look for a GEP which has one index that is an induction variable and all
5102 /// other indices are loop invariant. If the stride of this access is also
5103 /// within a small bound we decide that this address computation can likely be
5104 /// merged into the addressing mode.
5105 /// In all other cases, we identify the address computation as complex.
5106 static bool isLikelyComplexAddressComputation(Value *Ptr,
5107 LoopVectorizationLegality *Legal,
5108 ScalarEvolution *SE,
5109 const Loop *TheLoop) {
5110 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5114 // We are looking for a gep with all loop invariant indices except for one
5115 // which should be an induction variable.
5116 unsigned NumOperands = Gep->getNumOperands();
5117 for (unsigned i = 1; i < NumOperands; ++i) {
5118 Value *Opd = Gep->getOperand(i);
5119 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5120 !Legal->isInductionVariable(Opd))
5124 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5125 // can likely be merged into the address computation.
5126 unsigned MaxMergeDistance = 64;
5128 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5132 // Check the step is constant.
5133 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5134 // Calculate the pointer stride and check if it is consecutive.
5135 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5139 const APInt &APStepVal = C->getValue()->getValue();
5141 // Huge step value - give up.
5142 if (APStepVal.getBitWidth() > 64)
5145 int64_t StepVal = APStepVal.getSExtValue();
5147 return StepVal > MaxMergeDistance;
5150 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5151 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5157 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5158 // If we know that this instruction will remain uniform, check the cost of
5159 // the scalar version.
5160 if (Legal->isUniformAfterVectorization(I))
5163 Type *RetTy = I->getType();
5164 Type *VectorTy = ToVectorTy(RetTy, VF);
5166 // TODO: We need to estimate the cost of intrinsic calls.
5167 switch (I->getOpcode()) {
5168 case Instruction::GetElementPtr:
5169 // We mark this instruction as zero-cost because the cost of GEPs in
5170 // vectorized code depends on whether the corresponding memory instruction
5171 // is scalarized or not. Therefore, we handle GEPs with the memory
5172 // instruction cost.
5174 case Instruction::Br: {
5175 return TTI.getCFInstrCost(I->getOpcode());
5177 case Instruction::PHI:
5178 //TODO: IF-converted IFs become selects.
5180 case Instruction::Add:
5181 case Instruction::FAdd:
5182 case Instruction::Sub:
5183 case Instruction::FSub:
5184 case Instruction::Mul:
5185 case Instruction::FMul:
5186 case Instruction::UDiv:
5187 case Instruction::SDiv:
5188 case Instruction::FDiv:
5189 case Instruction::URem:
5190 case Instruction::SRem:
5191 case Instruction::FRem:
5192 case Instruction::Shl:
5193 case Instruction::LShr:
5194 case Instruction::AShr:
5195 case Instruction::And:
5196 case Instruction::Or:
5197 case Instruction::Xor: {
5198 // Since we will replace the stride by 1 the multiplication should go away.
5199 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5201 // Certain instructions can be cheaper to vectorize if they have a constant
5202 // second vector operand. One example of this are shifts on x86.
5203 TargetTransformInfo::OperandValueKind Op1VK =
5204 TargetTransformInfo::OK_AnyValue;
5205 TargetTransformInfo::OperandValueKind Op2VK =
5206 TargetTransformInfo::OK_AnyValue;
5207 TargetTransformInfo::OperandValueProperties Op1VP =
5208 TargetTransformInfo::OP_None;
5209 TargetTransformInfo::OperandValueProperties Op2VP =
5210 TargetTransformInfo::OP_None;
5211 Value *Op2 = I->getOperand(1);
5213 // Check for a splat of a constant or for a non uniform vector of constants.
5214 if (isa<ConstantInt>(Op2)) {
5215 ConstantInt *CInt = cast<ConstantInt>(Op2);
5216 if (CInt && CInt->getValue().isPowerOf2())
5217 Op2VP = TargetTransformInfo::OP_PowerOf2;
5218 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5219 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5220 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5221 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5223 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5224 if (CInt && CInt->getValue().isPowerOf2())
5225 Op2VP = TargetTransformInfo::OP_PowerOf2;
5226 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5230 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5233 case Instruction::Select: {
5234 SelectInst *SI = cast<SelectInst>(I);
5235 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5236 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5237 Type *CondTy = SI->getCondition()->getType();
5239 CondTy = VectorType::get(CondTy, VF);
5241 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5243 case Instruction::ICmp:
5244 case Instruction::FCmp: {
5245 Type *ValTy = I->getOperand(0)->getType();
5246 VectorTy = ToVectorTy(ValTy, VF);
5247 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5249 case Instruction::Store:
5250 case Instruction::Load: {
5251 StoreInst *SI = dyn_cast<StoreInst>(I);
5252 LoadInst *LI = dyn_cast<LoadInst>(I);
5253 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5255 VectorTy = ToVectorTy(ValTy, VF);
5257 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5258 unsigned AS = SI ? SI->getPointerAddressSpace() :
5259 LI->getPointerAddressSpace();
5260 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5261 // We add the cost of address computation here instead of with the gep
5262 // instruction because only here we know whether the operation is
5265 return TTI.getAddressComputationCost(VectorTy) +
5266 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5268 // For an interleaved access, calculate the total cost of the whole
5269 // interleave group.
5270 if (Legal->isAccessInterleaved(I)) {
5271 auto Group = Legal->getInterleavedAccessGroup(I);
5272 assert(Group && "Fail to get an interleaved access group.");
5274 // Only calculate the cost once at the insert position.
5275 if (Group->getInsertPos() != I)
5278 unsigned InterleaveFactor = Group->getFactor();
5280 VectorType::get(VectorTy->getVectorElementType(),
5281 VectorTy->getVectorNumElements() * InterleaveFactor);
5283 // Holds the indices of existing members in an interleaved load group.
5284 // An interleaved store group doesn't need this as it dones't allow gaps.
5285 SmallVector<unsigned, 4> Indices;
5287 for (unsigned i = 0; i < InterleaveFactor; i++)
5288 if (Group->getMember(i))
5289 Indices.push_back(i);
5292 // Calculate the cost of the whole interleaved group.
5293 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5294 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5295 Group->getAlignment(), AS);
5297 if (Group->isReverse())
5299 Group->getNumMembers() *
5300 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5302 // FIXME: The interleaved load group with a huge gap could be even more
5303 // expensive than scalar operations. Then we could ignore such group and
5304 // use scalar operations instead.
5308 // Scalarized loads/stores.
5309 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5310 bool Reverse = ConsecutiveStride < 0;
5311 const DataLayout &DL = I->getModule()->getDataLayout();
5312 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5313 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5314 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5315 bool IsComplexComputation =
5316 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5318 // The cost of extracting from the value vector and pointer vector.
5319 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5320 for (unsigned i = 0; i < VF; ++i) {
5321 // The cost of extracting the pointer operand.
5322 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5323 // In case of STORE, the cost of ExtractElement from the vector.
5324 // In case of LOAD, the cost of InsertElement into the returned
5326 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5327 Instruction::InsertElement,
5331 // The cost of the scalar loads/stores.
5332 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5333 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5338 // Wide load/stores.
5339 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5340 if (Legal->isMaskRequired(I))
5341 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5344 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5347 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5351 case Instruction::ZExt:
5352 case Instruction::SExt:
5353 case Instruction::FPToUI:
5354 case Instruction::FPToSI:
5355 case Instruction::FPExt:
5356 case Instruction::PtrToInt:
5357 case Instruction::IntToPtr:
5358 case Instruction::SIToFP:
5359 case Instruction::UIToFP:
5360 case Instruction::Trunc:
5361 case Instruction::FPTrunc:
5362 case Instruction::BitCast: {
5363 // We optimize the truncation of induction variable.
5364 // The cost of these is the same as the scalar operation.
5365 if (I->getOpcode() == Instruction::Trunc &&
5366 Legal->isInductionVariable(I->getOperand(0)))
5367 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5368 I->getOperand(0)->getType());
5370 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5371 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5373 case Instruction::Call: {
5374 bool NeedToScalarize;
5375 CallInst *CI = cast<CallInst>(I);
5376 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5377 if (getIntrinsicIDForCall(CI, TLI))
5378 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5382 // We are scalarizing the instruction. Return the cost of the scalar
5383 // instruction, plus the cost of insert and extract into vector
5384 // elements, times the vector width.
5387 if (!RetTy->isVoidTy() && VF != 1) {
5388 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5390 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5393 // The cost of inserting the results plus extracting each one of the
5395 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5398 // The cost of executing VF copies of the scalar instruction. This opcode
5399 // is unknown. Assume that it is the same as 'mul'.
5400 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5406 char LoopVectorize::ID = 0;
5407 static const char lv_name[] = "Loop Vectorization";
5408 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5409 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5410 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5411 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5412 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5413 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5414 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5415 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5416 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5417 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5418 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5419 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5422 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5423 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5427 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5428 // Check for a store.
5429 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5430 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5432 // Check for a load.
5433 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5434 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5440 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5441 bool IfPredicateStore) {
5442 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5443 // Holds vector parameters or scalars, in case of uniform vals.
5444 SmallVector<VectorParts, 4> Params;
5446 setDebugLocFromInst(Builder, Instr);
5448 // Find all of the vectorized parameters.
5449 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5450 Value *SrcOp = Instr->getOperand(op);
5452 // If we are accessing the old induction variable, use the new one.
5453 if (SrcOp == OldInduction) {
5454 Params.push_back(getVectorValue(SrcOp));
5458 // Try using previously calculated values.
5459 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5461 // If the src is an instruction that appeared earlier in the basic block
5462 // then it should already be vectorized.
5463 if (SrcInst && OrigLoop->contains(SrcInst)) {
5464 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5465 // The parameter is a vector value from earlier.
5466 Params.push_back(WidenMap.get(SrcInst));
5468 // The parameter is a scalar from outside the loop. Maybe even a constant.
5469 VectorParts Scalars;
5470 Scalars.append(UF, SrcOp);
5471 Params.push_back(Scalars);
5475 assert(Params.size() == Instr->getNumOperands() &&
5476 "Invalid number of operands");
5478 // Does this instruction return a value ?
5479 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5481 Value *UndefVec = IsVoidRetTy ? nullptr :
5482 UndefValue::get(Instr->getType());
5483 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5484 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5486 Instruction *InsertPt = Builder.GetInsertPoint();
5487 BasicBlock *IfBlock = Builder.GetInsertBlock();
5488 BasicBlock *CondBlock = nullptr;
5491 Loop *VectorLp = nullptr;
5492 if (IfPredicateStore) {
5493 assert(Instr->getParent()->getSinglePredecessor() &&
5494 "Only support single predecessor blocks");
5495 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5496 Instr->getParent());
5497 VectorLp = LI->getLoopFor(IfBlock);
5498 assert(VectorLp && "Must have a loop for this block");
5501 // For each vector unroll 'part':
5502 for (unsigned Part = 0; Part < UF; ++Part) {
5503 // For each scalar that we create:
5505 // Start an "if (pred) a[i] = ..." block.
5506 Value *Cmp = nullptr;
5507 if (IfPredicateStore) {
5508 if (Cond[Part]->getType()->isVectorTy())
5510 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5511 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5512 ConstantInt::get(Cond[Part]->getType(), 1));
5513 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5514 LoopVectorBody.push_back(CondBlock);
5515 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5516 // Update Builder with newly created basic block.
5517 Builder.SetInsertPoint(InsertPt);
5520 Instruction *Cloned = Instr->clone();
5522 Cloned->setName(Instr->getName() + ".cloned");
5523 // Replace the operands of the cloned instructions with extracted scalars.
5524 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5525 Value *Op = Params[op][Part];
5526 Cloned->setOperand(op, Op);
5529 // Place the cloned scalar in the new loop.
5530 Builder.Insert(Cloned);
5532 // If the original scalar returns a value we need to place it in a vector
5533 // so that future users will be able to use it.
5535 VecResults[Part] = Cloned;
5538 if (IfPredicateStore) {
5539 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5540 LoopVectorBody.push_back(NewIfBlock);
5541 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5542 Builder.SetInsertPoint(InsertPt);
5543 ReplaceInstWithInst(IfBlock->getTerminator(),
5544 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5545 IfBlock = NewIfBlock;
5550 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5551 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5552 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5554 return scalarizeInstruction(Instr, IfPredicateStore);
5557 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5561 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5565 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5566 // When unrolling and the VF is 1, we only need to add a simple scalar.
5567 Type *ITy = Val->getType();
5568 assert(!ITy->isVectorTy() && "Val must be a scalar");
5569 Constant *C = ConstantInt::get(ITy, StartIdx);
5570 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");