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 LoopVectorizationLegality;
221 class LoopVectorizationCostModel;
222 class LoopVectorizeHints;
224 /// \brief This modifies LoopAccessReport to initialize message with
225 /// loop-vectorizer-specific part.
226 class VectorizationReport : public LoopAccessReport {
228 VectorizationReport(Instruction *I = nullptr)
229 : LoopAccessReport("loop not vectorized: ", I) {}
231 /// \brief This allows promotion of the loop-access analysis report into the
232 /// loop-vectorizer report. It modifies the message to add the
233 /// loop-vectorizer-specific part of the message.
234 explicit VectorizationReport(const LoopAccessReport &R)
235 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
239 /// A helper function for converting Scalar types to vector types.
240 /// If the incoming type is void, we return void. If the VF is 1, we return
242 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
243 if (Scalar->isVoidTy() || VF == 1)
245 return VectorType::get(Scalar, VF);
248 /// InnerLoopVectorizer vectorizes loops which contain only one basic
249 /// block to a specified vectorization factor (VF).
250 /// This class performs the widening of scalars into vectors, or multiple
251 /// scalars. This class also implements the following features:
252 /// * It inserts an epilogue loop for handling loops that don't have iteration
253 /// counts that are known to be a multiple of the vectorization factor.
254 /// * It handles the code generation for reduction variables.
255 /// * Scalarization (implementation using scalars) of un-vectorizable
257 /// InnerLoopVectorizer does not perform any vectorization-legality
258 /// checks, and relies on the caller to check for the different legality
259 /// aspects. The InnerLoopVectorizer relies on the
260 /// LoopVectorizationLegality class to provide information about the induction
261 /// and reduction variables that were found to a given vectorization factor.
262 class InnerLoopVectorizer {
264 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
265 DominatorTree *DT, const TargetLibraryInfo *TLI,
266 const TargetTransformInfo *TTI, unsigned VecWidth,
267 unsigned UnrollFactor)
268 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
269 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
270 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
271 Legal(nullptr), AddedSafetyChecks(false) {}
273 // Perform the actual loop widening (vectorization).
274 void vectorize(LoopVectorizationLegality *L) {
276 // Create a new empty loop. Unlink the old loop and connect the new one.
278 // Widen each instruction in the old loop to a new one in the new loop.
279 // Use the Legality module to find the induction and reduction variables.
281 // Register the new loop and update the analysis passes.
285 // Return true if any runtime check is added.
286 bool IsSafetyChecksAdded() {
287 return AddedSafetyChecks;
290 virtual ~InnerLoopVectorizer() {}
293 /// A small list of PHINodes.
294 typedef SmallVector<PHINode*, 4> PhiVector;
295 /// When we unroll loops we have multiple vector values for each scalar.
296 /// This data structure holds the unrolled and vectorized values that
297 /// originated from one scalar instruction.
298 typedef SmallVector<Value*, 2> VectorParts;
300 // When we if-convert we need to create edge masks. We have to cache values
301 // so that we don't end up with exponential recursion/IR.
302 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
303 VectorParts> EdgeMaskCache;
305 /// \brief Add checks for strides that were assumed to be 1.
307 /// Returns the last check instruction and the first check instruction in the
308 /// pair as (first, last).
309 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
311 /// Create an empty loop, based on the loop ranges of the old loop.
312 void createEmptyLoop();
313 /// Copy and widen the instructions from the old loop.
314 virtual void vectorizeLoop();
316 /// \brief The Loop exit block may have single value PHI nodes where the
317 /// incoming value is 'Undef'. While vectorizing we only handled real values
318 /// that were defined inside the loop. Here we fix the 'undef case'.
322 /// A helper function that computes the predicate of the block BB, assuming
323 /// that the header block of the loop is set to True. It returns the *entry*
324 /// mask for the block BB.
325 VectorParts createBlockInMask(BasicBlock *BB);
326 /// A helper function that computes the predicate of the edge between SRC
328 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
330 /// A helper function to vectorize a single BB within the innermost loop.
331 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
333 /// Vectorize a single PHINode in a block. This method handles the induction
334 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
335 /// arbitrary length vectors.
336 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
337 unsigned UF, unsigned VF, PhiVector *PV);
339 /// Insert the new loop to the loop hierarchy and pass manager
340 /// and update the analysis passes.
341 void updateAnalysis();
343 /// This instruction is un-vectorizable. Implement it as a sequence
344 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
345 /// scalarized instruction behind an if block predicated on the control
346 /// dependence of the instruction.
347 virtual void scalarizeInstruction(Instruction *Instr,
348 bool IfPredicateStore=false);
350 /// Vectorize Load and Store instructions,
351 virtual void vectorizeMemoryInstruction(Instruction *Instr);
353 /// Create a broadcast instruction. This method generates a broadcast
354 /// instruction (shuffle) for loop invariant values and for the induction
355 /// value. If this is the induction variable then we extend it to N, N+1, ...
356 /// this is needed because each iteration in the loop corresponds to a SIMD
358 virtual Value *getBroadcastInstrs(Value *V);
360 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
361 /// to each vector element of Val. The sequence starts at StartIndex.
362 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
364 /// When we go over instructions in the basic block we rely on previous
365 /// values within the current basic block or on loop invariant values.
366 /// When we widen (vectorize) values we place them in the map. If the values
367 /// are not within the map, they have to be loop invariant, so we simply
368 /// broadcast them into a vector.
369 VectorParts &getVectorValue(Value *V);
371 /// Try to vectorize the interleaved access group that \p Instr belongs to.
372 void vectorizeInterleaveGroup(Instruction *Instr);
374 /// Generate a shuffle sequence that will reverse the vector Vec.
375 virtual Value *reverseVector(Value *Vec);
377 /// This is a helper class that holds the vectorizer state. It maps scalar
378 /// instructions to vector instructions. When the code is 'unrolled' then
379 /// then a single scalar value is mapped to multiple vector parts. The parts
380 /// are stored in the VectorPart type.
382 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
384 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
386 /// \return True if 'Key' is saved in the Value Map.
387 bool has(Value *Key) const { return MapStorage.count(Key); }
389 /// Initializes a new entry in the map. Sets all of the vector parts to the
390 /// save value in 'Val'.
391 /// \return A reference to a vector with splat values.
392 VectorParts &splat(Value *Key, Value *Val) {
393 VectorParts &Entry = MapStorage[Key];
394 Entry.assign(UF, Val);
398 ///\return A reference to the value that is stored at 'Key'.
399 VectorParts &get(Value *Key) {
400 VectorParts &Entry = MapStorage[Key];
403 assert(Entry.size() == UF);
408 /// The unroll factor. Each entry in the map stores this number of vector
412 /// Map storage. We use std::map and not DenseMap because insertions to a
413 /// dense map invalidates its iterators.
414 std::map<Value *, VectorParts> MapStorage;
417 /// The original loop.
419 /// Scev analysis to use.
427 /// Target Library Info.
428 const TargetLibraryInfo *TLI;
429 /// Target Transform Info.
430 const TargetTransformInfo *TTI;
432 /// The vectorization SIMD factor to use. Each vector will have this many
437 /// The vectorization unroll factor to use. Each scalar is vectorized to this
438 /// many different vector instructions.
441 /// The builder that we use
444 // --- Vectorization state ---
446 /// The vector-loop preheader.
447 BasicBlock *LoopVectorPreHeader;
448 /// The scalar-loop preheader.
449 BasicBlock *LoopScalarPreHeader;
450 /// Middle Block between the vector and the scalar.
451 BasicBlock *LoopMiddleBlock;
452 ///The ExitBlock of the scalar loop.
453 BasicBlock *LoopExitBlock;
454 ///The vector loop body.
455 SmallVector<BasicBlock *, 4> LoopVectorBody;
456 ///The scalar loop body.
457 BasicBlock *LoopScalarBody;
458 /// A list of all bypass blocks. The first block is the entry of the loop.
459 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
461 /// The new Induction variable which was added to the new block.
463 /// The induction variable of the old basic block.
464 PHINode *OldInduction;
465 /// Holds the extended (to the widest induction type) start index.
467 /// Maps scalars to widened vectors.
469 EdgeMaskCache MaskCache;
471 LoopVectorizationLegality *Legal;
473 // Record whether runtime check is added.
474 bool AddedSafetyChecks;
477 class InnerLoopUnroller : public InnerLoopVectorizer {
479 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
480 DominatorTree *DT, const TargetLibraryInfo *TLI,
481 const TargetTransformInfo *TTI, unsigned UnrollFactor)
482 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
485 void scalarizeInstruction(Instruction *Instr,
486 bool IfPredicateStore = false) override;
487 void vectorizeMemoryInstruction(Instruction *Instr) override;
488 Value *getBroadcastInstrs(Value *V) override;
489 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
490 Value *reverseVector(Value *Vec) override;
493 /// \brief Look for a meaningful debug location on the instruction or it's
495 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
500 if (I->getDebugLoc() != Empty)
503 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
504 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
505 if (OpInst->getDebugLoc() != Empty)
512 /// \brief Set the debug location in the builder using the debug location in the
514 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
515 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
516 B.SetCurrentDebugLocation(Inst->getDebugLoc());
518 B.SetCurrentDebugLocation(DebugLoc());
522 /// \return string containing a file name and a line # for the given loop.
523 static std::string getDebugLocString(const Loop *L) {
526 raw_string_ostream OS(Result);
527 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
528 LoopDbgLoc.print(OS);
530 // Just print the module name.
531 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
538 /// \brief Propagate known metadata from one instruction to another.
539 static void propagateMetadata(Instruction *To, const Instruction *From) {
540 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
541 From->getAllMetadataOtherThanDebugLoc(Metadata);
543 for (auto M : Metadata) {
544 unsigned Kind = M.first;
546 // These are safe to transfer (this is safe for TBAA, even when we
547 // if-convert, because should that metadata have had a control dependency
548 // on the condition, and thus actually aliased with some other
549 // non-speculated memory access when the condition was false, this would be
550 // caught by the runtime overlap checks).
551 if (Kind != LLVMContext::MD_tbaa &&
552 Kind != LLVMContext::MD_alias_scope &&
553 Kind != LLVMContext::MD_noalias &&
554 Kind != LLVMContext::MD_fpmath)
557 To->setMetadata(Kind, M.second);
561 /// \brief Propagate known metadata from one instruction to a vector of others.
562 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
564 if (Instruction *I = dyn_cast<Instruction>(V))
565 propagateMetadata(I, From);
568 /// \brief The group of interleaved loads/stores sharing the same stride and
569 /// close to each other.
571 /// Each member in this group has an index starting from 0, and the largest
572 /// index should be less than interleaved factor, which is equal to the absolute
573 /// value of the access's stride.
575 /// E.g. An interleaved load group of factor 4:
576 /// for (unsigned i = 0; i < 1024; i+=4) {
577 /// a = A[i]; // Member of index 0
578 /// b = A[i+1]; // Member of index 1
579 /// d = A[i+3]; // Member of index 3
583 /// An interleaved store group of factor 4:
584 /// for (unsigned i = 0; i < 1024; i+=4) {
586 /// A[i] = a; // Member of index 0
587 /// A[i+1] = b; // Member of index 1
588 /// A[i+2] = c; // Member of index 2
589 /// A[i+3] = d; // Member of index 3
592 /// Note: the interleaved load group could have gaps (missing members), but
593 /// the interleaved store group doesn't allow gaps.
594 class InterleaveGroup {
596 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
597 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
598 assert(Align && "The alignment should be non-zero");
600 Factor = std::abs(Stride);
601 assert(Factor > 1 && "Invalid interleave factor");
603 Reverse = Stride < 0;
607 bool isReverse() const { return Reverse; }
608 unsigned getFactor() const { return Factor; }
609 unsigned getAlignment() const { return Align; }
610 unsigned getNumMembers() const { return Members.size(); }
612 /// \brief Try to insert a new member \p Instr with index \p Index and
613 /// alignment \p NewAlign. The index is related to the leader and it could be
614 /// negative if it is the new leader.
616 /// \returns false if the instruction doesn't belong to the group.
617 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
618 assert(NewAlign && "The new member's alignment should be non-zero");
620 int Key = Index + SmallestKey;
622 // Skip if there is already a member with the same index.
623 if (Members.count(Key))
626 if (Key > LargestKey) {
627 // The largest index is always less than the interleave factor.
628 if (Index >= static_cast<int>(Factor))
632 } else if (Key < SmallestKey) {
633 // The largest index is always less than the interleave factor.
634 if (LargestKey - Key >= static_cast<int>(Factor))
640 // It's always safe to select the minimum alignment.
641 Align = std::min(Align, NewAlign);
642 Members[Key] = Instr;
646 /// \brief Get the member with the given index \p Index
648 /// \returns nullptr if contains no such member.
649 Instruction *getMember(unsigned Index) const {
650 int Key = SmallestKey + Index;
651 if (!Members.count(Key))
654 return Members.find(Key)->second;
657 /// \brief Get the index for the given member. Unlike the key in the member
658 /// map, the index starts from 0.
659 unsigned getIndex(Instruction *Instr) const {
660 for (auto I : Members)
661 if (I.second == Instr)
662 return I.first - SmallestKey;
664 llvm_unreachable("InterleaveGroup contains no such member");
667 Instruction *getInsertPos() const { return InsertPos; }
668 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
671 unsigned Factor; // Interleave Factor.
674 DenseMap<int, Instruction *> Members;
678 // To avoid breaking dependences, vectorized instructions of an interleave
679 // group should be inserted at either the first load or the last store in
682 // E.g. %even = load i32 // Insert Position
683 // %add = add i32 %even // Use of %even
687 // %odd = add i32 // Def of %odd
688 // store i32 %odd // Insert Position
689 Instruction *InsertPos;
692 /// \brief Drive the analysis of interleaved memory accesses in the loop.
694 /// Use this class to analyze interleaved accesses only when we can vectorize
695 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
696 /// on interleaved accesses is unsafe.
698 /// The analysis collects interleave groups and records the relationships
699 /// between the member and the group in a map.
700 class InterleavedAccessInfo {
702 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT)
703 : SE(SE), TheLoop(L), DT(DT) {}
705 ~InterleavedAccessInfo() {
706 SmallSet<InterleaveGroup *, 4> DelSet;
707 // Avoid releasing a pointer twice.
708 for (auto &I : InterleaveGroupMap)
709 DelSet.insert(I.second);
710 for (auto *Ptr : DelSet)
714 /// \brief Analyze the interleaved accesses and collect them in interleave
715 /// groups. Substitute symbolic strides using \p Strides.
716 void analyzeInterleaving(const ValueToValueMap &Strides);
718 /// \brief Check if \p Instr belongs to any interleave group.
719 bool isInterleaved(Instruction *Instr) const {
720 return InterleaveGroupMap.count(Instr);
723 /// \brief Get the interleave group that \p Instr belongs to.
725 /// \returns nullptr if doesn't have such group.
726 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
727 if (InterleaveGroupMap.count(Instr))
728 return InterleaveGroupMap.find(Instr)->second;
737 /// Holds the relationships between the members and the interleave group.
738 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
740 /// \brief The descriptor for a strided memory access.
741 struct StrideDescriptor {
742 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
744 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
746 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
748 int Stride; // The access's stride. It is negative for a reverse access.
749 const SCEV *Scev; // The scalar expression of this access
750 unsigned Size; // The size of the memory object.
751 unsigned Align; // The alignment of this access.
754 /// \brief Create a new interleave group with the given instruction \p Instr,
755 /// stride \p Stride and alignment \p Align.
757 /// \returns the newly created interleave group.
758 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
760 assert(!InterleaveGroupMap.count(Instr) &&
761 "Already in an interleaved access group");
762 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
763 return InterleaveGroupMap[Instr];
766 /// \brief Release the group and remove all the relationships.
767 void releaseGroup(InterleaveGroup *Group) {
768 for (unsigned i = 0; i < Group->getFactor(); i++)
769 if (Instruction *Member = Group->getMember(i))
770 InterleaveGroupMap.erase(Member);
775 /// \brief Collect all the accesses with a constant stride in program order.
776 void collectConstStridedAccesses(
777 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
778 const ValueToValueMap &Strides);
781 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
782 /// to what vectorization factor.
783 /// This class does not look at the profitability of vectorization, only the
784 /// legality. This class has two main kinds of checks:
785 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
786 /// will change the order of memory accesses in a way that will change the
787 /// correctness of the program.
788 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
789 /// checks for a number of different conditions, such as the availability of a
790 /// single induction variable, that all types are supported and vectorize-able,
791 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
792 /// This class is also used by InnerLoopVectorizer for identifying
793 /// induction variable and the different reduction variables.
794 class LoopVectorizationLegality {
796 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
797 TargetLibraryInfo *TLI, AliasAnalysis *AA,
798 Function *F, const TargetTransformInfo *TTI,
799 LoopAccessAnalysis *LAA)
800 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
801 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), InterleaveInfo(SE, L, DT),
802 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
804 /// This enum represents the kinds of inductions that we support.
806 IK_NoInduction, ///< Not an induction variable.
807 IK_IntInduction, ///< Integer induction variable. Step = C.
808 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
811 /// A struct for saving information about induction variables.
812 struct InductionInfo {
813 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
814 : StartValue(Start), IK(K), StepValue(Step) {
815 assert(IK != IK_NoInduction && "Not an induction");
816 assert(StartValue && "StartValue is null");
817 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
818 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
819 "StartValue is not a pointer for pointer induction");
820 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
821 "StartValue is not an integer for integer induction");
822 assert(StepValue->getType()->isIntegerTy() &&
823 "StepValue is not an integer");
826 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
828 /// Get the consecutive direction. Returns:
829 /// 0 - unknown or non-consecutive.
830 /// 1 - consecutive and increasing.
831 /// -1 - consecutive and decreasing.
832 int getConsecutiveDirection() const {
833 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
834 return StepValue->getSExtValue();
838 /// Compute the transformed value of Index at offset StartValue using step
840 /// For integer induction, returns StartValue + Index * StepValue.
841 /// For pointer induction, returns StartValue[Index * StepValue].
842 /// FIXME: The newly created binary instructions should contain nsw/nuw
843 /// flags, which can be found from the original scalar operations.
844 Value *transform(IRBuilder<> &B, Value *Index) const {
846 case IK_IntInduction:
847 assert(Index->getType() == StartValue->getType() &&
848 "Index type does not match StartValue type");
849 if (StepValue->isMinusOne())
850 return B.CreateSub(StartValue, Index);
851 if (!StepValue->isOne())
852 Index = B.CreateMul(Index, StepValue);
853 return B.CreateAdd(StartValue, Index);
855 case IK_PtrInduction:
856 assert(Index->getType() == StepValue->getType() &&
857 "Index type does not match StepValue type");
858 if (StepValue->isMinusOne())
859 Index = B.CreateNeg(Index);
860 else if (!StepValue->isOne())
861 Index = B.CreateMul(Index, StepValue);
862 return B.CreateGEP(nullptr, StartValue, Index);
867 llvm_unreachable("invalid enum");
871 TrackingVH<Value> StartValue;
875 ConstantInt *StepValue;
878 /// ReductionList contains the reduction descriptors for all
879 /// of the reductions that were found in the loop.
880 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
882 /// InductionList saves induction variables and maps them to the
883 /// induction descriptor.
884 typedef MapVector<PHINode*, InductionInfo> InductionList;
886 /// Returns true if it is legal to vectorize this loop.
887 /// This does not mean that it is profitable to vectorize this
888 /// loop, only that it is legal to do so.
891 /// Returns the Induction variable.
892 PHINode *getInduction() { return Induction; }
894 /// Returns the reduction variables found in the loop.
895 ReductionList *getReductionVars() { return &Reductions; }
897 /// Returns the induction variables found in the loop.
898 InductionList *getInductionVars() { return &Inductions; }
900 /// Returns the widest induction type.
901 Type *getWidestInductionType() { return WidestIndTy; }
903 /// Returns True if V is an induction variable in this loop.
904 bool isInductionVariable(const Value *V);
906 /// Return true if the block BB needs to be predicated in order for the loop
907 /// to be vectorized.
908 bool blockNeedsPredication(BasicBlock *BB);
910 /// Check if this pointer is consecutive when vectorizing. This happens
911 /// when the last index of the GEP is the induction variable, or that the
912 /// pointer itself is an induction variable.
913 /// This check allows us to vectorize A[idx] into a wide load/store.
915 /// 0 - Stride is unknown or non-consecutive.
916 /// 1 - Address is consecutive.
917 /// -1 - Address is consecutive, and decreasing.
918 int isConsecutivePtr(Value *Ptr);
920 /// Returns true if the value V is uniform within the loop.
921 bool isUniform(Value *V);
923 /// Returns true if this instruction will remain scalar after vectorization.
924 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
926 /// Returns the information that we collected about runtime memory check.
927 const RuntimePointerChecking *getRuntimePointerChecking() const {
928 return LAI->getRuntimePointerChecking();
931 const LoopAccessInfo *getLAI() const {
935 /// \brief Check if \p Instr belongs to any interleaved access group.
936 bool isAccessInterleaved(Instruction *Instr) {
937 return InterleaveInfo.isInterleaved(Instr);
940 /// \brief Get the interleaved access group that \p Instr belongs to.
941 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
942 return InterleaveInfo.getInterleaveGroup(Instr);
945 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
947 bool hasStride(Value *V) { return StrideSet.count(V); }
948 bool mustCheckStrides() { return !StrideSet.empty(); }
949 SmallPtrSet<Value *, 8>::iterator strides_begin() {
950 return StrideSet.begin();
952 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
954 /// Returns true if the target machine supports masked store operation
955 /// for the given \p DataType and kind of access to \p Ptr.
956 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
957 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
959 /// Returns true if the target machine supports masked load operation
960 /// for the given \p DataType and kind of access to \p Ptr.
961 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
962 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
964 /// Returns true if vector representation of the instruction \p I
966 bool isMaskRequired(const Instruction* I) {
967 return (MaskedOp.count(I) != 0);
969 unsigned getNumStores() const {
970 return LAI->getNumStores();
972 unsigned getNumLoads() const {
973 return LAI->getNumLoads();
975 unsigned getNumPredStores() const {
976 return NumPredStores;
979 /// Check if a single basic block loop is vectorizable.
980 /// At this point we know that this is a loop with a constant trip count
981 /// and we only need to check individual instructions.
982 bool canVectorizeInstrs();
984 /// When we vectorize loops we may change the order in which
985 /// we read and write from memory. This method checks if it is
986 /// legal to vectorize the code, considering only memory constrains.
987 /// Returns true if the loop is vectorizable
988 bool canVectorizeMemory();
990 /// Return true if we can vectorize this loop using the IF-conversion
992 bool canVectorizeWithIfConvert();
994 /// Collect the variables that need to stay uniform after vectorization.
995 void collectLoopUniforms();
997 /// Return true if all of the instructions in the block can be speculatively
998 /// executed. \p SafePtrs is a list of addresses that are known to be legal
999 /// and we know that we can read from them without segfault.
1000 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1002 /// Returns the induction kind of Phi and record the step. This function may
1003 /// return NoInduction if the PHI is not an induction variable.
1004 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
1006 /// \brief Collect memory access with loop invariant strides.
1008 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1010 void collectStridedAccess(Value *LoadOrStoreInst);
1012 /// Report an analysis message to assist the user in diagnosing loops that are
1013 /// not vectorized. These are handled as LoopAccessReport rather than
1014 /// VectorizationReport because the << operator of VectorizationReport returns
1015 /// LoopAccessReport.
1016 void emitAnalysis(const LoopAccessReport &Message) {
1017 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
1020 unsigned NumPredStores;
1022 /// The loop that we evaluate.
1025 ScalarEvolution *SE;
1026 /// Target Library Info.
1027 TargetLibraryInfo *TLI;
1029 Function *TheFunction;
1030 /// Target Transform Info
1031 const TargetTransformInfo *TTI;
1034 // LoopAccess analysis.
1035 LoopAccessAnalysis *LAA;
1036 // And the loop-accesses info corresponding to this loop. This pointer is
1037 // null until canVectorizeMemory sets it up.
1038 const LoopAccessInfo *LAI;
1040 /// The interleave access information contains groups of interleaved accesses
1041 /// with the same stride and close to each other.
1042 InterleavedAccessInfo InterleaveInfo;
1044 // --- vectorization state --- //
1046 /// Holds the integer induction variable. This is the counter of the
1049 /// Holds the reduction variables.
1050 ReductionList Reductions;
1051 /// Holds all of the induction variables that we found in the loop.
1052 /// Notice that inductions don't need to start at zero and that induction
1053 /// variables can be pointers.
1054 InductionList Inductions;
1055 /// Holds the widest induction type encountered.
1058 /// Allowed outside users. This holds the reduction
1059 /// vars which can be accessed from outside the loop.
1060 SmallPtrSet<Value*, 4> AllowedExit;
1061 /// This set holds the variables which are known to be uniform after
1063 SmallPtrSet<Instruction*, 4> Uniforms;
1065 /// Can we assume the absence of NaNs.
1066 bool HasFunNoNaNAttr;
1068 ValueToValueMap Strides;
1069 SmallPtrSet<Value *, 8> StrideSet;
1071 /// While vectorizing these instructions we have to generate a
1072 /// call to the appropriate masked intrinsic
1073 SmallPtrSet<const Instruction*, 8> MaskedOp;
1076 /// LoopVectorizationCostModel - estimates the expected speedups due to
1078 /// In many cases vectorization is not profitable. This can happen because of
1079 /// a number of reasons. In this class we mainly attempt to predict the
1080 /// expected speedup/slowdowns due to the supported instruction set. We use the
1081 /// TargetTransformInfo to query the different backends for the cost of
1082 /// different operations.
1083 class LoopVectorizationCostModel {
1085 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1086 LoopVectorizationLegality *Legal,
1087 const TargetTransformInfo &TTI,
1088 const TargetLibraryInfo *TLI, AssumptionCache *AC,
1089 const Function *F, const LoopVectorizeHints *Hints)
1090 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
1091 TheFunction(F), Hints(Hints) {
1092 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
1095 /// Information about vectorization costs
1096 struct VectorizationFactor {
1097 unsigned Width; // Vector width with best cost
1098 unsigned Cost; // Cost of the loop with that width
1100 /// \return The most profitable vectorization factor and the cost of that VF.
1101 /// This method checks every power of two up to VF. If UserVF is not ZERO
1102 /// then this vectorization factor will be selected if vectorization is
1104 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1106 /// \return The size (in bits) of the widest type in the code that
1107 /// needs to be vectorized. We ignore values that remain scalar such as
1108 /// 64 bit loop indices.
1109 unsigned getWidestType();
1111 /// \return The desired interleave count.
1112 /// If interleave count has been specified by metadata it will be returned.
1113 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1114 /// are the selected vectorization factor and the cost of the selected VF.
1115 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1118 /// \return The most profitable unroll factor.
1119 /// This method finds the best unroll-factor based on register pressure and
1120 /// other parameters. VF and LoopCost are the selected vectorization factor
1121 /// and the cost of the selected VF.
1122 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1125 /// \brief A struct that represents some properties of the register usage
1127 struct RegisterUsage {
1128 /// Holds the number of loop invariant values that are used in the loop.
1129 unsigned LoopInvariantRegs;
1130 /// Holds the maximum number of concurrent live intervals in the loop.
1131 unsigned MaxLocalUsers;
1132 /// Holds the number of instructions in the loop.
1133 unsigned NumInstructions;
1136 /// \return information about the register usage of the loop.
1137 RegisterUsage calculateRegisterUsage();
1140 /// Returns the expected execution cost. The unit of the cost does
1141 /// not matter because we use the 'cost' units to compare different
1142 /// vector widths. The cost that is returned is *not* normalized by
1143 /// the factor width.
1144 unsigned expectedCost(unsigned VF);
1146 /// Returns the execution time cost of an instruction for a given vector
1147 /// width. Vector width of one means scalar.
1148 unsigned getInstructionCost(Instruction *I, unsigned VF);
1150 /// Returns whether the instruction is a load or store and will be a emitted
1151 /// as a vector operation.
1152 bool isConsecutiveLoadOrStore(Instruction *I);
1154 /// Report an analysis message to assist the user in diagnosing loops that are
1155 /// not vectorized. These are handled as LoopAccessReport rather than
1156 /// VectorizationReport because the << operator of VectorizationReport returns
1157 /// LoopAccessReport.
1158 void emitAnalysis(const LoopAccessReport &Message) {
1159 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
1162 /// Values used only by @llvm.assume calls.
1163 SmallPtrSet<const Value *, 32> EphValues;
1165 /// The loop that we evaluate.
1168 ScalarEvolution *SE;
1169 /// Loop Info analysis.
1171 /// Vectorization legality.
1172 LoopVectorizationLegality *Legal;
1173 /// Vector target information.
1174 const TargetTransformInfo &TTI;
1175 /// Target Library Info.
1176 const TargetLibraryInfo *TLI;
1177 const Function *TheFunction;
1178 // Loop Vectorize Hint.
1179 const LoopVectorizeHints *Hints;
1182 /// Utility class for getting and setting loop vectorizer hints in the form
1183 /// of loop metadata.
1184 /// This class keeps a number of loop annotations locally (as member variables)
1185 /// and can, upon request, write them back as metadata on the loop. It will
1186 /// initially scan the loop for existing metadata, and will update the local
1187 /// values based on information in the loop.
1188 /// We cannot write all values to metadata, as the mere presence of some info,
1189 /// for example 'force', means a decision has been made. So, we need to be
1190 /// careful NOT to add them if the user hasn't specifically asked so.
1191 class LoopVectorizeHints {
1198 /// Hint - associates name and validation with the hint value.
1201 unsigned Value; // This may have to change for non-numeric values.
1204 Hint(const char * Name, unsigned Value, HintKind Kind)
1205 : Name(Name), Value(Value), Kind(Kind) { }
1207 bool validate(unsigned Val) {
1210 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1212 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1220 /// Vectorization width.
1222 /// Vectorization interleave factor.
1224 /// Vectorization forced
1227 /// Return the loop metadata prefix.
1228 static StringRef Prefix() { return "llvm.loop."; }
1232 FK_Undefined = -1, ///< Not selected.
1233 FK_Disabled = 0, ///< Forcing disabled.
1234 FK_Enabled = 1, ///< Forcing enabled.
1237 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1238 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1240 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1241 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1243 // Populate values with existing loop metadata.
1244 getHintsFromMetadata();
1246 // force-vector-interleave overrides DisableInterleaving.
1247 if (VectorizerParams::isInterleaveForced())
1248 Interleave.Value = VectorizerParams::VectorizationInterleave;
1250 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1251 << "LV: Interleaving disabled by the pass manager\n");
1254 /// Mark the loop L as already vectorized by setting the width to 1.
1255 void setAlreadyVectorized() {
1256 Width.Value = Interleave.Value = 1;
1257 Hint Hints[] = {Width, Interleave};
1258 writeHintsToMetadata(Hints);
1261 /// Dumps all the hint information.
1262 std::string emitRemark() const {
1263 VectorizationReport R;
1264 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1265 R << "vectorization is explicitly disabled";
1267 R << "use -Rpass-analysis=loop-vectorize for more info";
1268 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1269 R << " (Force=true";
1270 if (Width.Value != 0)
1271 R << ", Vector Width=" << Width.Value;
1272 if (Interleave.Value != 0)
1273 R << ", Interleave Count=" << Interleave.Value;
1281 unsigned getWidth() const { return Width.Value; }
1282 unsigned getInterleave() const { return Interleave.Value; }
1283 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1286 /// Find hints specified in the loop metadata and update local values.
1287 void getHintsFromMetadata() {
1288 MDNode *LoopID = TheLoop->getLoopID();
1292 // First operand should refer to the loop id itself.
1293 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1294 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1296 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1297 const MDString *S = nullptr;
1298 SmallVector<Metadata *, 4> Args;
1300 // The expected hint is either a MDString or a MDNode with the first
1301 // operand a MDString.
1302 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1303 if (!MD || MD->getNumOperands() == 0)
1305 S = dyn_cast<MDString>(MD->getOperand(0));
1306 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1307 Args.push_back(MD->getOperand(i));
1309 S = dyn_cast<MDString>(LoopID->getOperand(i));
1310 assert(Args.size() == 0 && "too many arguments for MDString");
1316 // Check if the hint starts with the loop metadata prefix.
1317 StringRef Name = S->getString();
1318 if (Args.size() == 1)
1319 setHint(Name, Args[0]);
1323 /// Checks string hint with one operand and set value if valid.
1324 void setHint(StringRef Name, Metadata *Arg) {
1325 if (!Name.startswith(Prefix()))
1327 Name = Name.substr(Prefix().size(), StringRef::npos);
1329 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1331 unsigned Val = C->getZExtValue();
1333 Hint *Hints[] = {&Width, &Interleave, &Force};
1334 for (auto H : Hints) {
1335 if (Name == H->Name) {
1336 if (H->validate(Val))
1339 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1345 /// Create a new hint from name / value pair.
1346 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1347 LLVMContext &Context = TheLoop->getHeader()->getContext();
1348 Metadata *MDs[] = {MDString::get(Context, Name),
1349 ConstantAsMetadata::get(
1350 ConstantInt::get(Type::getInt32Ty(Context), V))};
1351 return MDNode::get(Context, MDs);
1354 /// Matches metadata with hint name.
1355 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1356 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1360 for (auto H : HintTypes)
1361 if (Name->getString().endswith(H.Name))
1366 /// Sets current hints into loop metadata, keeping other values intact.
1367 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1368 if (HintTypes.size() == 0)
1371 // Reserve the first element to LoopID (see below).
1372 SmallVector<Metadata *, 4> MDs(1);
1373 // If the loop already has metadata, then ignore the existing operands.
1374 MDNode *LoopID = TheLoop->getLoopID();
1376 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1377 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1378 // If node in update list, ignore old value.
1379 if (!matchesHintMetadataName(Node, HintTypes))
1380 MDs.push_back(Node);
1384 // Now, add the missing hints.
1385 for (auto H : HintTypes)
1386 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1388 // Replace current metadata node with new one.
1389 LLVMContext &Context = TheLoop->getHeader()->getContext();
1390 MDNode *NewLoopID = MDNode::get(Context, MDs);
1391 // Set operand 0 to refer to the loop id itself.
1392 NewLoopID->replaceOperandWith(0, NewLoopID);
1394 TheLoop->setLoopID(NewLoopID);
1397 /// The loop these hints belong to.
1398 const Loop *TheLoop;
1401 static void emitMissedWarning(Function *F, Loop *L,
1402 const LoopVectorizeHints &LH) {
1403 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1404 L->getStartLoc(), LH.emitRemark());
1406 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1407 if (LH.getWidth() != 1)
1408 emitLoopVectorizeWarning(
1409 F->getContext(), *F, L->getStartLoc(),
1410 "failed explicitly specified loop vectorization");
1411 else if (LH.getInterleave() != 1)
1412 emitLoopInterleaveWarning(
1413 F->getContext(), *F, L->getStartLoc(),
1414 "failed explicitly specified loop interleaving");
1418 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1420 return V.push_back(&L);
1422 for (Loop *InnerL : L)
1423 addInnerLoop(*InnerL, V);
1426 /// The LoopVectorize Pass.
1427 struct LoopVectorize : public FunctionPass {
1428 /// Pass identification, replacement for typeid
1431 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1433 DisableUnrolling(NoUnrolling),
1434 AlwaysVectorize(AlwaysVectorize) {
1435 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1438 ScalarEvolution *SE;
1440 TargetTransformInfo *TTI;
1442 BlockFrequencyInfo *BFI;
1443 TargetLibraryInfo *TLI;
1445 AssumptionCache *AC;
1446 LoopAccessAnalysis *LAA;
1447 bool DisableUnrolling;
1448 bool AlwaysVectorize;
1450 BlockFrequency ColdEntryFreq;
1452 bool runOnFunction(Function &F) override {
1453 SE = &getAnalysis<ScalarEvolution>();
1454 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1455 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1456 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1457 BFI = &getAnalysis<BlockFrequencyInfo>();
1458 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1459 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1460 AA = &getAnalysis<AliasAnalysis>();
1461 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1462 LAA = &getAnalysis<LoopAccessAnalysis>();
1464 // Compute some weights outside of the loop over the loops. Compute this
1465 // using a BranchProbability to re-use its scaling math.
1466 const BranchProbability ColdProb(1, 5); // 20%
1467 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1470 // 1. the target claims to have no vector registers, and
1471 // 2. interleaving won't help ILP.
1473 // The second condition is necessary because, even if the target has no
1474 // vector registers, loop vectorization may still enable scalar
1476 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1479 // Build up a worklist of inner-loops to vectorize. This is necessary as
1480 // the act of vectorizing or partially unrolling a loop creates new loops
1481 // and can invalidate iterators across the loops.
1482 SmallVector<Loop *, 8> Worklist;
1485 addInnerLoop(*L, Worklist);
1487 LoopsAnalyzed += Worklist.size();
1489 // Now walk the identified inner loops.
1490 bool Changed = false;
1491 while (!Worklist.empty())
1492 Changed |= processLoop(Worklist.pop_back_val());
1494 // Process each loop nest in the function.
1498 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1499 SmallVector<Metadata *, 4> MDs;
1500 // Reserve first location for self reference to the LoopID metadata node.
1501 MDs.push_back(nullptr);
1502 bool IsUnrollMetadata = false;
1503 MDNode *LoopID = L->getLoopID();
1505 // First find existing loop unrolling disable metadata.
1506 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1507 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1509 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1511 S && S->getString().startswith("llvm.loop.unroll.disable");
1513 MDs.push_back(LoopID->getOperand(i));
1517 if (!IsUnrollMetadata) {
1518 // Add runtime unroll disable metadata.
1519 LLVMContext &Context = L->getHeader()->getContext();
1520 SmallVector<Metadata *, 1> DisableOperands;
1521 DisableOperands.push_back(
1522 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1523 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1524 MDs.push_back(DisableNode);
1525 MDNode *NewLoopID = MDNode::get(Context, MDs);
1526 // Set operand 0 to refer to the loop id itself.
1527 NewLoopID->replaceOperandWith(0, NewLoopID);
1528 L->setLoopID(NewLoopID);
1532 bool processLoop(Loop *L) {
1533 assert(L->empty() && "Only process inner loops.");
1536 const std::string DebugLocStr = getDebugLocString(L);
1539 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1540 << L->getHeader()->getParent()->getName() << "\" from "
1541 << DebugLocStr << "\n");
1543 LoopVectorizeHints Hints(L, DisableUnrolling);
1545 DEBUG(dbgs() << "LV: Loop hints:"
1547 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1549 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1551 : "?")) << " width=" << Hints.getWidth()
1552 << " unroll=" << Hints.getInterleave() << "\n");
1554 // Function containing loop
1555 Function *F = L->getHeader()->getParent();
1557 // Looking at the diagnostic output is the only way to determine if a loop
1558 // was vectorized (other than looking at the IR or machine code), so it
1559 // is important to generate an optimization remark for each loop. Most of
1560 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1561 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1562 // less verbose reporting vectorized loops and unvectorized loops that may
1563 // benefit from vectorization, respectively.
1565 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1566 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1567 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1568 L->getStartLoc(), Hints.emitRemark());
1572 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1573 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1574 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1575 L->getStartLoc(), Hints.emitRemark());
1579 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1580 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1581 emitOptimizationRemarkAnalysis(
1582 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1583 "loop not vectorized: vector width and interleave count are "
1584 "explicitly set to 1");
1588 // Check the loop for a trip count threshold:
1589 // do not vectorize loops with a tiny trip count.
1590 const unsigned TC = SE->getSmallConstantTripCount(L);
1591 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1592 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1593 << "This loop is not worth vectorizing.");
1594 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1595 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1597 DEBUG(dbgs() << "\n");
1598 emitOptimizationRemarkAnalysis(
1599 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1600 "vectorization is not beneficial and is not explicitly forced");
1605 // Check if it is legal to vectorize the loop.
1606 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA);
1607 if (!LVL.canVectorize()) {
1608 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1609 emitMissedWarning(F, L, Hints);
1613 // Use the cost model.
1614 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1616 // Check the function attributes to find out if this function should be
1617 // optimized for size.
1618 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1619 F->hasFnAttribute(Attribute::OptimizeForSize);
1621 // Compute the weighted frequency of this loop being executed and see if it
1622 // is less than 20% of the function entry baseline frequency. Note that we
1623 // always have a canonical loop here because we think we *can* vectoriez.
1624 // FIXME: This is hidden behind a flag due to pervasive problems with
1625 // exactly what block frequency models.
1626 if (LoopVectorizeWithBlockFrequency) {
1627 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1628 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1629 LoopEntryFreq < ColdEntryFreq)
1633 // Check the function attributes to see if implicit floats are allowed.a
1634 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1635 // an integer loop and the vector instructions selected are purely integer
1636 // vector instructions?
1637 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1638 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1639 "attribute is used.\n");
1640 emitOptimizationRemarkAnalysis(
1641 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1642 "loop not vectorized due to NoImplicitFloat attribute");
1643 emitMissedWarning(F, L, Hints);
1647 // Select the optimal vectorization factor.
1648 const LoopVectorizationCostModel::VectorizationFactor VF =
1649 CM.selectVectorizationFactor(OptForSize);
1651 // Select the interleave count.
1652 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1654 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1655 << DebugLocStr << '\n');
1656 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1658 if (VF.Width == 1) {
1659 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1662 emitOptimizationRemarkAnalysis(
1663 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1664 "not beneficial to vectorize and user disabled interleaving");
1667 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1669 // Report the unrolling decision.
1670 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1671 Twine("interleaved by " + Twine(IC) +
1672 " (vectorization not beneficial)"));
1674 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC);
1675 Unroller.vectorize(&LVL);
1677 // If we decided that it is *legal* to vectorize the loop then do it.
1678 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC);
1682 // Add metadata to disable runtime unrolling scalar loop when there's no
1683 // runtime check about strides and memory. Because at this situation,
1684 // scalar loop is rarely used not worthy to be unrolled.
1685 if (!LB.IsSafetyChecksAdded())
1686 AddRuntimeUnrollDisableMetaData(L);
1688 // Report the vectorization decision.
1689 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1690 Twine("vectorized loop (vectorization width: ") +
1691 Twine(VF.Width) + ", interleaved count: " +
1695 // Mark the loop as already vectorized to avoid vectorizing again.
1696 Hints.setAlreadyVectorized();
1698 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1702 void getAnalysisUsage(AnalysisUsage &AU) const override {
1703 AU.addRequired<AssumptionCacheTracker>();
1704 AU.addRequiredID(LoopSimplifyID);
1705 AU.addRequiredID(LCSSAID);
1706 AU.addRequired<BlockFrequencyInfo>();
1707 AU.addRequired<DominatorTreeWrapperPass>();
1708 AU.addRequired<LoopInfoWrapperPass>();
1709 AU.addRequired<ScalarEvolution>();
1710 AU.addRequired<TargetTransformInfoWrapperPass>();
1711 AU.addRequired<AliasAnalysis>();
1712 AU.addRequired<LoopAccessAnalysis>();
1713 AU.addPreserved<LoopInfoWrapperPass>();
1714 AU.addPreserved<DominatorTreeWrapperPass>();
1715 AU.addPreserved<AliasAnalysis>();
1720 } // end anonymous namespace
1722 //===----------------------------------------------------------------------===//
1723 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1724 // LoopVectorizationCostModel.
1725 //===----------------------------------------------------------------------===//
1727 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1728 // We need to place the broadcast of invariant variables outside the loop.
1729 Instruction *Instr = dyn_cast<Instruction>(V);
1731 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1732 Instr->getParent()) != LoopVectorBody.end());
1733 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1735 // Place the code for broadcasting invariant variables in the new preheader.
1736 IRBuilder<>::InsertPointGuard Guard(Builder);
1738 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1740 // Broadcast the scalar into all locations in the vector.
1741 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1746 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1748 assert(Val->getType()->isVectorTy() && "Must be a vector");
1749 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1750 "Elem must be an integer");
1751 assert(Step->getType() == Val->getType()->getScalarType() &&
1752 "Step has wrong type");
1753 // Create the types.
1754 Type *ITy = Val->getType()->getScalarType();
1755 VectorType *Ty = cast<VectorType>(Val->getType());
1756 int VLen = Ty->getNumElements();
1757 SmallVector<Constant*, 8> Indices;
1759 // Create a vector of consecutive numbers from zero to VF.
1760 for (int i = 0; i < VLen; ++i)
1761 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1763 // Add the consecutive indices to the vector value.
1764 Constant *Cv = ConstantVector::get(Indices);
1765 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1766 Step = Builder.CreateVectorSplat(VLen, Step);
1767 assert(Step->getType() == Val->getType() && "Invalid step vec");
1768 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1769 // which can be found from the original scalar operations.
1770 Step = Builder.CreateMul(Cv, Step);
1771 return Builder.CreateAdd(Val, Step, "induction");
1774 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1775 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1776 // Make sure that the pointer does not point to structs.
1777 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1780 // If this value is a pointer induction variable we know it is consecutive.
1781 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1782 if (Phi && Inductions.count(Phi)) {
1783 InductionInfo II = Inductions[Phi];
1784 return II.getConsecutiveDirection();
1787 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1791 unsigned NumOperands = Gep->getNumOperands();
1792 Value *GpPtr = Gep->getPointerOperand();
1793 // If this GEP value is a consecutive pointer induction variable and all of
1794 // the indices are constant then we know it is consecutive. We can
1795 Phi = dyn_cast<PHINode>(GpPtr);
1796 if (Phi && Inductions.count(Phi)) {
1798 // Make sure that the pointer does not point to structs.
1799 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1800 if (GepPtrType->getElementType()->isAggregateType())
1803 // Make sure that all of the index operands are loop invariant.
1804 for (unsigned i = 1; i < NumOperands; ++i)
1805 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1808 InductionInfo II = Inductions[Phi];
1809 return II.getConsecutiveDirection();
1812 unsigned InductionOperand = getGEPInductionOperand(Gep);
1814 // Check that all of the gep indices are uniform except for our induction
1816 for (unsigned i = 0; i != NumOperands; ++i)
1817 if (i != InductionOperand &&
1818 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1821 // We can emit wide load/stores only if the last non-zero index is the
1822 // induction variable.
1823 const SCEV *Last = nullptr;
1824 if (!Strides.count(Gep))
1825 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1827 // Because of the multiplication by a stride we can have a s/zext cast.
1828 // We are going to replace this stride by 1 so the cast is safe to ignore.
1830 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1831 // %0 = trunc i64 %indvars.iv to i32
1832 // %mul = mul i32 %0, %Stride1
1833 // %idxprom = zext i32 %mul to i64 << Safe cast.
1834 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1836 Last = replaceSymbolicStrideSCEV(SE, Strides,
1837 Gep->getOperand(InductionOperand), Gep);
1838 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1840 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1844 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1845 const SCEV *Step = AR->getStepRecurrence(*SE);
1847 // The memory is consecutive because the last index is consecutive
1848 // and all other indices are loop invariant.
1851 if (Step->isAllOnesValue())
1858 bool LoopVectorizationLegality::isUniform(Value *V) {
1859 return LAI->isUniform(V);
1862 InnerLoopVectorizer::VectorParts&
1863 InnerLoopVectorizer::getVectorValue(Value *V) {
1864 assert(V != Induction && "The new induction variable should not be used.");
1865 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1867 // If we have a stride that is replaced by one, do it here.
1868 if (Legal->hasStride(V))
1869 V = ConstantInt::get(V->getType(), 1);
1871 // If we have this scalar in the map, return it.
1872 if (WidenMap.has(V))
1873 return WidenMap.get(V);
1875 // If this scalar is unknown, assume that it is a constant or that it is
1876 // loop invariant. Broadcast V and save the value for future uses.
1877 Value *B = getBroadcastInstrs(V);
1878 return WidenMap.splat(V, B);
1881 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1882 assert(Vec->getType()->isVectorTy() && "Invalid type");
1883 SmallVector<Constant*, 8> ShuffleMask;
1884 for (unsigned i = 0; i < VF; ++i)
1885 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1887 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1888 ConstantVector::get(ShuffleMask),
1892 // Get a mask to interleave \p NumVec vectors into a wide vector.
1893 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
1894 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
1895 // <0, 4, 1, 5, 2, 6, 3, 7>
1896 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
1898 SmallVector<Constant *, 16> Mask;
1899 for (unsigned i = 0; i < VF; i++)
1900 for (unsigned j = 0; j < NumVec; j++)
1901 Mask.push_back(Builder.getInt32(j * VF + i));
1903 return ConstantVector::get(Mask);
1906 // Get the strided mask starting from index \p Start.
1907 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
1908 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
1909 unsigned Stride, unsigned VF) {
1910 SmallVector<Constant *, 16> Mask;
1911 for (unsigned i = 0; i < VF; i++)
1912 Mask.push_back(Builder.getInt32(Start + i * Stride));
1914 return ConstantVector::get(Mask);
1917 // Get a mask of two parts: The first part consists of sequential integers
1918 // starting from 0, The second part consists of UNDEFs.
1919 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
1920 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
1921 unsigned NumUndef) {
1922 SmallVector<Constant *, 16> Mask;
1923 for (unsigned i = 0; i < NumInt; i++)
1924 Mask.push_back(Builder.getInt32(i));
1926 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
1927 for (unsigned i = 0; i < NumUndef; i++)
1928 Mask.push_back(Undef);
1930 return ConstantVector::get(Mask);
1933 // Concatenate two vectors with the same element type. The 2nd vector should
1934 // not have more elements than the 1st vector. If the 2nd vector has less
1935 // elements, extend it with UNDEFs.
1936 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
1938 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
1939 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
1940 assert(VecTy1 && VecTy2 &&
1941 VecTy1->getScalarType() == VecTy2->getScalarType() &&
1942 "Expect two vectors with the same element type");
1944 unsigned NumElts1 = VecTy1->getNumElements();
1945 unsigned NumElts2 = VecTy2->getNumElements();
1946 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
1948 if (NumElts1 > NumElts2) {
1949 // Extend with UNDEFs.
1951 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
1952 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
1955 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
1956 return Builder.CreateShuffleVector(V1, V2, Mask);
1959 // Concatenate vectors in the given list. All vectors have the same type.
1960 static Value *ConcatenateVectors(IRBuilder<> &Builder,
1961 ArrayRef<Value *> InputList) {
1962 unsigned NumVec = InputList.size();
1963 assert(NumVec > 1 && "Should be at least two vectors");
1965 SmallVector<Value *, 8> ResList;
1966 ResList.append(InputList.begin(), InputList.end());
1968 SmallVector<Value *, 8> TmpList;
1969 for (unsigned i = 0; i < NumVec - 1; i += 2) {
1970 Value *V0 = ResList[i], *V1 = ResList[i + 1];
1971 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
1972 "Only the last vector may have a different type");
1974 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
1977 // Push the last vector if the total number of vectors is odd.
1978 if (NumVec % 2 != 0)
1979 TmpList.push_back(ResList[NumVec - 1]);
1982 NumVec = ResList.size();
1983 } while (NumVec > 1);
1988 // Try to vectorize the interleave group that \p Instr belongs to.
1990 // E.g. Translate following interleaved load group (factor = 3):
1991 // for (i = 0; i < N; i+=3) {
1992 // R = Pic[i]; // Member of index 0
1993 // G = Pic[i+1]; // Member of index 1
1994 // B = Pic[i+2]; // Member of index 2
1995 // ... // do something to R, G, B
1998 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
1999 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2000 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2001 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2003 // Or translate following interleaved store group (factor = 3):
2004 // for (i = 0; i < N; i+=3) {
2005 // ... do something to R, G, B
2006 // Pic[i] = R; // Member of index 0
2007 // Pic[i+1] = G; // Member of index 1
2008 // Pic[i+2] = B; // Member of index 2
2011 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2012 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2013 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2014 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2015 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2016 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2017 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2018 assert(Group && "Fail to get an interleaved access group.");
2020 // Skip if current instruction is not the insert position.
2021 if (Instr != Group->getInsertPos())
2024 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2025 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2026 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2028 // Prepare for the vector type of the interleaved load/store.
2029 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2030 unsigned InterleaveFactor = Group->getFactor();
2031 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2032 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2034 // Prepare for the new pointers.
2035 setDebugLocFromInst(Builder, Ptr);
2036 VectorParts &PtrParts = getVectorValue(Ptr);
2037 SmallVector<Value *, 2> NewPtrs;
2038 unsigned Index = Group->getIndex(Instr);
2039 for (unsigned Part = 0; Part < UF; Part++) {
2040 // Extract the pointer for current instruction from the pointer vector. A
2041 // reverse access uses the pointer in the last lane.
2042 Value *NewPtr = Builder.CreateExtractElement(
2044 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2046 // Notice current instruction could be any index. Need to adjust the address
2047 // to the member of index 0.
2049 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2050 // b = A[i]; // Member of index 0
2051 // Current pointer is pointed to A[i+1], adjust it to A[i].
2053 // E.g. A[i+1] = a; // Member of index 1
2054 // A[i] = b; // Member of index 0
2055 // A[i+2] = c; // Member of index 2 (Current instruction)
2056 // Current pointer is pointed to A[i+2], adjust it to A[i].
2057 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2059 // Cast to the vector pointer type.
2060 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2063 setDebugLocFromInst(Builder, Instr);
2064 Value *UndefVec = UndefValue::get(VecTy);
2066 // Vectorize the interleaved load group.
2068 for (unsigned Part = 0; Part < UF; Part++) {
2069 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2070 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2072 for (unsigned i = 0; i < InterleaveFactor; i++) {
2073 Instruction *Member = Group->getMember(i);
2075 // Skip the gaps in the group.
2079 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2080 Value *StridedVec = Builder.CreateShuffleVector(
2081 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2083 // If this member has different type, cast the result type.
2084 if (Member->getType() != ScalarTy) {
2085 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2086 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2089 VectorParts &Entry = WidenMap.get(Member);
2091 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2094 propagateMetadata(NewLoadInstr, Instr);
2099 // The sub vector type for current instruction.
2100 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2102 // Vectorize the interleaved store group.
2103 for (unsigned Part = 0; Part < UF; Part++) {
2104 // Collect the stored vector from each member.
2105 SmallVector<Value *, 4> StoredVecs;
2106 for (unsigned i = 0; i < InterleaveFactor; i++) {
2107 // Interleaved store group doesn't allow a gap, so each index has a member
2108 Instruction *Member = Group->getMember(i);
2109 assert(Member && "Fail to get a member from an interleaved store group");
2112 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2113 if (Group->isReverse())
2114 StoredVec = reverseVector(StoredVec);
2116 // If this member has different type, cast it to an unified type.
2117 if (StoredVec->getType() != SubVT)
2118 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2120 StoredVecs.push_back(StoredVec);
2123 // Concatenate all vectors into a wide vector.
2124 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2126 // Interleave the elements in the wide vector.
2127 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2128 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2131 Instruction *NewStoreInstr =
2132 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2133 propagateMetadata(NewStoreInstr, Instr);
2137 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2138 // Attempt to issue a wide load.
2139 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2140 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2142 assert((LI || SI) && "Invalid Load/Store instruction");
2144 // Try to vectorize the interleave group if this access is interleaved.
2145 if (Legal->isAccessInterleaved(Instr))
2146 return vectorizeInterleaveGroup(Instr);
2148 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2149 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2150 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2151 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2152 // An alignment of 0 means target abi alignment. We need to use the scalar's
2153 // target abi alignment in such a case.
2154 const DataLayout &DL = Instr->getModule()->getDataLayout();
2156 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2157 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2158 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2159 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2161 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2162 !Legal->isMaskRequired(SI))
2163 return scalarizeInstruction(Instr, true);
2165 if (ScalarAllocatedSize != VectorElementSize)
2166 return scalarizeInstruction(Instr);
2168 // If the pointer is loop invariant or if it is non-consecutive,
2169 // scalarize the load.
2170 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2171 bool Reverse = ConsecutiveStride < 0;
2172 bool UniformLoad = LI && Legal->isUniform(Ptr);
2173 if (!ConsecutiveStride || UniformLoad)
2174 return scalarizeInstruction(Instr);
2176 Constant *Zero = Builder.getInt32(0);
2177 VectorParts &Entry = WidenMap.get(Instr);
2179 // Handle consecutive loads/stores.
2180 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
2181 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2182 setDebugLocFromInst(Builder, Gep);
2183 Value *PtrOperand = Gep->getPointerOperand();
2184 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2185 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2187 // Create the new GEP with the new induction variable.
2188 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2189 Gep2->setOperand(0, FirstBasePtr);
2190 Gep2->setName("gep.indvar.base");
2191 Ptr = Builder.Insert(Gep2);
2193 setDebugLocFromInst(Builder, Gep);
2194 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2195 OrigLoop) && "Base ptr must be invariant");
2197 // The last index does not have to be the induction. It can be
2198 // consecutive and be a function of the index. For example A[I+1];
2199 unsigned NumOperands = Gep->getNumOperands();
2200 unsigned InductionOperand = getGEPInductionOperand(Gep);
2201 // Create the new GEP with the new induction variable.
2202 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2204 for (unsigned i = 0; i < NumOperands; ++i) {
2205 Value *GepOperand = Gep->getOperand(i);
2206 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2208 // Update last index or loop invariant instruction anchored in loop.
2209 if (i == InductionOperand ||
2210 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2211 assert((i == InductionOperand ||
2212 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2213 "Must be last index or loop invariant");
2215 VectorParts &GEPParts = getVectorValue(GepOperand);
2216 Value *Index = GEPParts[0];
2217 Index = Builder.CreateExtractElement(Index, Zero);
2218 Gep2->setOperand(i, Index);
2219 Gep2->setName("gep.indvar.idx");
2222 Ptr = Builder.Insert(Gep2);
2224 // Use the induction element ptr.
2225 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2226 setDebugLocFromInst(Builder, Ptr);
2227 VectorParts &PtrVal = getVectorValue(Ptr);
2228 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2231 VectorParts Mask = createBlockInMask(Instr->getParent());
2234 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2235 "We do not allow storing to uniform addresses");
2236 setDebugLocFromInst(Builder, SI);
2237 // We don't want to update the value in the map as it might be used in
2238 // another expression. So don't use a reference type for "StoredVal".
2239 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2241 for (unsigned Part = 0; Part < UF; ++Part) {
2242 // Calculate the pointer for the specific unroll-part.
2244 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2247 // If we store to reverse consecutive memory locations then we need
2248 // to reverse the order of elements in the stored value.
2249 StoredVal[Part] = reverseVector(StoredVal[Part]);
2250 // If the address is consecutive but reversed, then the
2251 // wide store needs to start at the last vector element.
2252 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2253 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2254 Mask[Part] = reverseVector(Mask[Part]);
2257 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2258 DataTy->getPointerTo(AddressSpace));
2261 if (Legal->isMaskRequired(SI))
2262 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2265 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2266 propagateMetadata(NewSI, SI);
2272 assert(LI && "Must have a load instruction");
2273 setDebugLocFromInst(Builder, LI);
2274 for (unsigned Part = 0; Part < UF; ++Part) {
2275 // Calculate the pointer for the specific unroll-part.
2277 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2280 // If the address is consecutive but reversed, then the
2281 // wide load needs to start at the last vector element.
2282 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2283 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2284 Mask[Part] = reverseVector(Mask[Part]);
2288 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2289 DataTy->getPointerTo(AddressSpace));
2290 if (Legal->isMaskRequired(LI))
2291 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2292 UndefValue::get(DataTy),
2293 "wide.masked.load");
2295 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2296 propagateMetadata(NewLI, LI);
2297 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2301 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
2302 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2303 // Holds vector parameters or scalars, in case of uniform vals.
2304 SmallVector<VectorParts, 4> Params;
2306 setDebugLocFromInst(Builder, Instr);
2308 // Find all of the vectorized parameters.
2309 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2310 Value *SrcOp = Instr->getOperand(op);
2312 // If we are accessing the old induction variable, use the new one.
2313 if (SrcOp == OldInduction) {
2314 Params.push_back(getVectorValue(SrcOp));
2318 // Try using previously calculated values.
2319 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2321 // If the src is an instruction that appeared earlier in the basic block
2322 // then it should already be vectorized.
2323 if (SrcInst && OrigLoop->contains(SrcInst)) {
2324 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2325 // The parameter is a vector value from earlier.
2326 Params.push_back(WidenMap.get(SrcInst));
2328 // The parameter is a scalar from outside the loop. Maybe even a constant.
2329 VectorParts Scalars;
2330 Scalars.append(UF, SrcOp);
2331 Params.push_back(Scalars);
2335 assert(Params.size() == Instr->getNumOperands() &&
2336 "Invalid number of operands");
2338 // Does this instruction return a value ?
2339 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2341 Value *UndefVec = IsVoidRetTy ? nullptr :
2342 UndefValue::get(VectorType::get(Instr->getType(), VF));
2343 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2344 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2346 Instruction *InsertPt = Builder.GetInsertPoint();
2347 BasicBlock *IfBlock = Builder.GetInsertBlock();
2348 BasicBlock *CondBlock = nullptr;
2351 Loop *VectorLp = nullptr;
2352 if (IfPredicateStore) {
2353 assert(Instr->getParent()->getSinglePredecessor() &&
2354 "Only support single predecessor blocks");
2355 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2356 Instr->getParent());
2357 VectorLp = LI->getLoopFor(IfBlock);
2358 assert(VectorLp && "Must have a loop for this block");
2361 // For each vector unroll 'part':
2362 for (unsigned Part = 0; Part < UF; ++Part) {
2363 // For each scalar that we create:
2364 for (unsigned Width = 0; Width < VF; ++Width) {
2367 Value *Cmp = nullptr;
2368 if (IfPredicateStore) {
2369 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2370 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2371 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2372 LoopVectorBody.push_back(CondBlock);
2373 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
2374 // Update Builder with newly created basic block.
2375 Builder.SetInsertPoint(InsertPt);
2378 Instruction *Cloned = Instr->clone();
2380 Cloned->setName(Instr->getName() + ".cloned");
2381 // Replace the operands of the cloned instructions with extracted scalars.
2382 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2383 Value *Op = Params[op][Part];
2384 // Param is a vector. Need to extract the right lane.
2385 if (Op->getType()->isVectorTy())
2386 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2387 Cloned->setOperand(op, Op);
2390 // Place the cloned scalar in the new loop.
2391 Builder.Insert(Cloned);
2393 // If the original scalar returns a value we need to place it in a vector
2394 // so that future users will be able to use it.
2396 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2397 Builder.getInt32(Width));
2399 if (IfPredicateStore) {
2400 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2401 LoopVectorBody.push_back(NewIfBlock);
2402 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2403 Builder.SetInsertPoint(InsertPt);
2404 ReplaceInstWithInst(IfBlock->getTerminator(),
2405 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
2406 IfBlock = NewIfBlock;
2412 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2416 if (Instruction *I = dyn_cast<Instruction>(V))
2417 return I->getParent() == Loc->getParent() ? I : nullptr;
2421 std::pair<Instruction *, Instruction *>
2422 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2423 Instruction *tnullptr = nullptr;
2424 if (!Legal->mustCheckStrides())
2425 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2427 IRBuilder<> ChkBuilder(Loc);
2430 Value *Check = nullptr;
2431 Instruction *FirstInst = nullptr;
2432 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2433 SE = Legal->strides_end();
2435 Value *Ptr = stripIntegerCast(*SI);
2436 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2438 // Store the first instruction we create.
2439 FirstInst = getFirstInst(FirstInst, C, Loc);
2441 Check = ChkBuilder.CreateOr(Check, C);
2446 // We have to do this trickery because the IRBuilder might fold the check to a
2447 // constant expression in which case there is no Instruction anchored in a
2449 LLVMContext &Ctx = Loc->getContext();
2450 Instruction *TheCheck =
2451 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2452 ChkBuilder.Insert(TheCheck, "stride.not.one");
2453 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2455 return std::make_pair(FirstInst, TheCheck);
2458 void InnerLoopVectorizer::createEmptyLoop() {
2460 In this function we generate a new loop. The new loop will contain
2461 the vectorized instructions while the old loop will continue to run the
2464 [ ] <-- Back-edge taken count overflow check.
2467 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2470 || [ ] <-- vector pre header.
2474 || [ ]_| <-- vector loop.
2477 | >[ ] <--- middle-block.
2480 -|- >[ ] <--- new preheader.
2484 | [ ]_| <-- old scalar loop to handle remainder.
2487 >[ ] <-- exit block.
2491 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2492 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2493 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2494 assert(VectorPH && "Invalid loop structure");
2495 assert(ExitBlock && "Must have an exit block");
2497 // Some loops have a single integer induction variable, while other loops
2498 // don't. One example is c++ iterators that often have multiple pointer
2499 // induction variables. In the code below we also support a case where we
2500 // don't have a single induction variable.
2501 OldInduction = Legal->getInduction();
2502 Type *IdxTy = Legal->getWidestInductionType();
2504 // Find the loop boundaries.
2505 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2506 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2508 // The exit count might have the type of i64 while the phi is i32. This can
2509 // happen if we have an induction variable that is sign extended before the
2510 // compare. The only way that we get a backedge taken count is that the
2511 // induction variable was signed and as such will not overflow. In such a case
2512 // truncation is legal.
2513 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2514 IdxTy->getPrimitiveSizeInBits())
2515 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2517 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2518 // Get the total trip count from the count by adding 1.
2519 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2520 SE->getConstant(BackedgeTakeCount->getType(), 1));
2522 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2524 // Expand the trip count and place the new instructions in the preheader.
2525 // Notice that the pre-header does not change, only the loop body.
2526 SCEVExpander Exp(*SE, DL, "induction");
2528 // We need to test whether the backedge-taken count is uint##_max. Adding one
2529 // to it will cause overflow and an incorrect loop trip count in the vector
2530 // body. In case of overflow we want to directly jump to the scalar remainder
2532 Value *BackedgeCount =
2533 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2534 VectorPH->getTerminator());
2535 if (BackedgeCount->getType()->isPointerTy())
2536 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2537 "backedge.ptrcnt.to.int",
2538 VectorPH->getTerminator());
2539 Instruction *CheckBCOverflow =
2540 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2541 Constant::getAllOnesValue(BackedgeCount->getType()),
2542 "backedge.overflow", VectorPH->getTerminator());
2544 // The loop index does not have to start at Zero. Find the original start
2545 // value from the induction PHI node. If we don't have an induction variable
2546 // then we know that it starts at zero.
2547 Builder.SetInsertPoint(VectorPH->getTerminator());
2548 Value *StartIdx = ExtendedIdx =
2550 ? Builder.CreateZExt(OldInduction->getIncomingValueForBlock(VectorPH),
2552 : ConstantInt::get(IdxTy, 0);
2554 // Count holds the overall loop count (N).
2555 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2556 VectorPH->getTerminator());
2558 LoopBypassBlocks.push_back(VectorPH);
2560 // Split the single block loop into the two loop structure described above.
2561 BasicBlock *VecBody =
2562 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2563 BasicBlock *MiddleBlock =
2564 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2565 BasicBlock *ScalarPH =
2566 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2568 // Create and register the new vector loop.
2569 Loop* Lp = new Loop();
2570 Loop *ParentLoop = OrigLoop->getParentLoop();
2572 // Insert the new loop into the loop nest and register the new basic blocks
2573 // before calling any utilities such as SCEV that require valid LoopInfo.
2575 ParentLoop->addChildLoop(Lp);
2576 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2577 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2579 LI->addTopLevelLoop(Lp);
2581 Lp->addBasicBlockToLoop(VecBody, *LI);
2583 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2585 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2587 // Generate the induction variable.
2588 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2589 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2590 // The loop step is equal to the vectorization factor (num of SIMD elements)
2591 // times the unroll factor (num of SIMD instructions).
2592 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2594 // Generate code to check that the loop's trip count that we computed by
2595 // adding one to the backedge-taken count will not overflow.
2596 BasicBlock *NewVectorPH =
2597 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "overflow.checked");
2599 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2600 ReplaceInstWithInst(
2601 VectorPH->getTerminator(),
2602 BranchInst::Create(ScalarPH, NewVectorPH, CheckBCOverflow));
2603 VectorPH = NewVectorPH;
2605 // This is the IR builder that we use to add all of the logic for bypassing
2606 // the new vector loop.
2607 IRBuilder<> BypassBuilder(VectorPH->getTerminator());
2608 setDebugLocFromInst(BypassBuilder,
2609 getDebugLocFromInstOrOperands(OldInduction));
2611 // We may need to extend the index in case there is a type mismatch.
2612 // We know that the count starts at zero and does not overflow.
2613 if (Count->getType() != IdxTy) {
2614 // The exit count can be of pointer type. Convert it to the correct
2616 if (ExitCount->getType()->isPointerTy())
2617 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2619 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2622 // Add the start index to the loop count to get the new end index.
2623 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2625 // Now we need to generate the expression for N - (N % VF), which is
2626 // the part that the vectorized body will execute.
2627 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2628 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2629 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2630 "end.idx.rnd.down");
2632 // Now, compare the new count to zero. If it is zero skip the vector loop and
2633 // jump to the scalar loop.
2635 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2637 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2639 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2640 LoopBypassBlocks.push_back(VectorPH);
2641 ReplaceInstWithInst(VectorPH->getTerminator(),
2642 BranchInst::Create(MiddleBlock, NewVectorPH, Cmp));
2643 VectorPH = NewVectorPH;
2645 // Generate the code to check that the strides we assumed to be one are really
2646 // one. We want the new basic block to start at the first instruction in a
2647 // sequence of instructions that form a check.
2648 Instruction *StrideCheck;
2649 Instruction *FirstCheckInst;
2650 std::tie(FirstCheckInst, StrideCheck) =
2651 addStrideCheck(VectorPH->getTerminator());
2653 AddedSafetyChecks = true;
2654 // Create a new block containing the stride check.
2655 VectorPH->setName("vector.stridecheck");
2657 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2659 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2660 LoopBypassBlocks.push_back(VectorPH);
2662 // Replace the branch into the memory check block with a conditional branch
2663 // for the "few elements case".
2664 ReplaceInstWithInst(
2665 VectorPH->getTerminator(),
2666 BranchInst::Create(MiddleBlock, NewVectorPH, StrideCheck));
2668 VectorPH = NewVectorPH;
2671 // Generate the code that checks in runtime if arrays overlap. We put the
2672 // checks into a separate block to make the more common case of few elements
2674 Instruction *MemRuntimeCheck;
2675 std::tie(FirstCheckInst, MemRuntimeCheck) =
2676 Legal->getLAI()->addRuntimeCheck(VectorPH->getTerminator());
2677 if (MemRuntimeCheck) {
2678 AddedSafetyChecks = true;
2679 // Create a new block containing the memory check.
2680 VectorPH->setName("vector.memcheck");
2682 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.ph");
2684 ParentLoop->addBasicBlockToLoop(NewVectorPH, *LI);
2685 LoopBypassBlocks.push_back(VectorPH);
2687 // Replace the branch into the memory check block with a conditional branch
2688 // for the "few elements case".
2689 ReplaceInstWithInst(
2690 VectorPH->getTerminator(),
2691 BranchInst::Create(MiddleBlock, NewVectorPH, MemRuntimeCheck));
2693 VectorPH = NewVectorPH;
2696 // We are going to resume the execution of the scalar loop.
2697 // Go over all of the induction variables that we found and fix the
2698 // PHIs that are left in the scalar version of the loop.
2699 // The starting values of PHI nodes depend on the counter of the last
2700 // iteration in the vectorized loop.
2701 // If we come from a bypass edge then we need to start from the original
2704 // This variable saves the new starting index for the scalar loop.
2705 PHINode *ResumeIndex = nullptr;
2706 LoopVectorizationLegality::InductionList::iterator I, E;
2707 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2708 // Set builder to point to last bypass block.
2709 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2710 for (I = List->begin(), E = List->end(); I != E; ++I) {
2711 PHINode *OrigPhi = I->first;
2712 LoopVectorizationLegality::InductionInfo II = I->second;
2714 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2715 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2716 MiddleBlock->getTerminator());
2717 // We might have extended the type of the induction variable but we need a
2718 // truncated version for the scalar loop.
2719 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2720 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2721 MiddleBlock->getTerminator()) : nullptr;
2723 // Create phi nodes to merge from the backedge-taken check block.
2724 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2725 ScalarPH->getTerminator());
2726 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2728 PHINode *BCTruncResumeVal = nullptr;
2729 if (OrigPhi == OldInduction) {
2731 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2732 ScalarPH->getTerminator());
2733 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2736 Value *EndValue = nullptr;
2738 case LoopVectorizationLegality::IK_NoInduction:
2739 llvm_unreachable("Unknown induction");
2740 case LoopVectorizationLegality::IK_IntInduction: {
2741 // Handle the integer induction counter.
2742 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2744 // We have the canonical induction variable.
2745 if (OrigPhi == OldInduction) {
2746 // Create a truncated version of the resume value for the scalar loop,
2747 // we might have promoted the type to a larger width.
2749 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2750 // The new PHI merges the original incoming value, in case of a bypass,
2751 // or the value at the end of the vectorized loop.
2752 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2753 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2754 TruncResumeVal->addIncoming(EndValue, VecBody);
2756 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2758 // We know what the end value is.
2759 EndValue = IdxEndRoundDown;
2760 // We also know which PHI node holds it.
2761 ResumeIndex = ResumeVal;
2765 // Not the canonical induction variable - add the vector loop count to the
2767 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2768 II.StartValue->getType(),
2770 EndValue = II.transform(BypassBuilder, CRD);
2771 EndValue->setName("ind.end");
2774 case LoopVectorizationLegality::IK_PtrInduction: {
2775 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2776 II.StepValue->getType(),
2778 EndValue = II.transform(BypassBuilder, CRD);
2779 EndValue->setName("ptr.ind.end");
2784 // The new PHI merges the original incoming value, in case of a bypass,
2785 // or the value at the end of the vectorized loop.
2786 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2787 if (OrigPhi == OldInduction)
2788 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2790 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2792 ResumeVal->addIncoming(EndValue, VecBody);
2794 // Fix the scalar body counter (PHI node).
2795 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2797 // The old induction's phi node in the scalar body needs the truncated
2799 if (OrigPhi == OldInduction) {
2800 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2801 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2803 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2804 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2808 // If we are generating a new induction variable then we also need to
2809 // generate the code that calculates the exit value. This value is not
2810 // simply the end of the counter because we may skip the vectorized body
2811 // in case of a runtime check.
2813 assert(!ResumeIndex && "Unexpected resume value found");
2814 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2815 MiddleBlock->getTerminator());
2816 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2817 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2818 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2821 // Make sure that we found the index where scalar loop needs to continue.
2822 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2823 "Invalid resume Index");
2825 // Add a check in the middle block to see if we have completed
2826 // all of the iterations in the first vector loop.
2827 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2828 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2829 ResumeIndex, "cmp.n",
2830 MiddleBlock->getTerminator());
2831 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2832 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2834 // Create i+1 and fill the PHINode.
2835 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2836 Induction->addIncoming(StartIdx, VectorPH);
2837 Induction->addIncoming(NextIdx, VecBody);
2838 // Create the compare.
2839 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2840 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2842 // Now we have two terminators. Remove the old one from the block.
2843 VecBody->getTerminator()->eraseFromParent();
2845 // Get ready to start creating new instructions into the vectorized body.
2846 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2849 LoopVectorPreHeader = VectorPH;
2850 LoopScalarPreHeader = ScalarPH;
2851 LoopMiddleBlock = MiddleBlock;
2852 LoopExitBlock = ExitBlock;
2853 LoopVectorBody.push_back(VecBody);
2854 LoopScalarBody = OldBasicBlock;
2856 LoopVectorizeHints Hints(Lp, true);
2857 Hints.setAlreadyVectorized();
2861 struct CSEDenseMapInfo {
2862 static bool canHandle(Instruction *I) {
2863 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2864 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2866 static inline Instruction *getEmptyKey() {
2867 return DenseMapInfo<Instruction *>::getEmptyKey();
2869 static inline Instruction *getTombstoneKey() {
2870 return DenseMapInfo<Instruction *>::getTombstoneKey();
2872 static unsigned getHashValue(Instruction *I) {
2873 assert(canHandle(I) && "Unknown instruction!");
2874 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2875 I->value_op_end()));
2877 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2878 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2879 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2881 return LHS->isIdenticalTo(RHS);
2886 /// \brief Check whether this block is a predicated block.
2887 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2888 /// = ...; " blocks. We start with one vectorized basic block. For every
2889 /// conditional block we split this vectorized block. Therefore, every second
2890 /// block will be a predicated one.
2891 static bool isPredicatedBlock(unsigned BlockNum) {
2892 return BlockNum % 2;
2895 ///\brief Perform cse of induction variable instructions.
2896 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2897 // Perform simple cse.
2898 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2899 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2900 BasicBlock *BB = BBs[i];
2901 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2902 Instruction *In = I++;
2904 if (!CSEDenseMapInfo::canHandle(In))
2907 // Check if we can replace this instruction with any of the
2908 // visited instructions.
2909 if (Instruction *V = CSEMap.lookup(In)) {
2910 In->replaceAllUsesWith(V);
2911 In->eraseFromParent();
2914 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2915 // ...;" blocks for predicated stores. Every second block is a predicated
2917 if (isPredicatedBlock(i))
2925 /// \brief Adds a 'fast' flag to floating point operations.
2926 static Value *addFastMathFlag(Value *V) {
2927 if (isa<FPMathOperator>(V)){
2928 FastMathFlags Flags;
2929 Flags.setUnsafeAlgebra();
2930 cast<Instruction>(V)->setFastMathFlags(Flags);
2935 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
2936 /// the result needs to be inserted and/or extracted from vectors.
2937 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
2938 const TargetTransformInfo &TTI) {
2942 assert(Ty->isVectorTy() && "Can only scalarize vectors");
2945 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
2947 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
2949 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
2955 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
2956 // Return the cost of the instruction, including scalarization overhead if it's
2957 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
2958 // i.e. either vector version isn't available, or is too expensive.
2959 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
2960 const TargetTransformInfo &TTI,
2961 const TargetLibraryInfo *TLI,
2962 bool &NeedToScalarize) {
2963 Function *F = CI->getCalledFunction();
2964 StringRef FnName = CI->getCalledFunction()->getName();
2965 Type *ScalarRetTy = CI->getType();
2966 SmallVector<Type *, 4> Tys, ScalarTys;
2967 for (auto &ArgOp : CI->arg_operands())
2968 ScalarTys.push_back(ArgOp->getType());
2970 // Estimate cost of scalarized vector call. The source operands are assumed
2971 // to be vectors, so we need to extract individual elements from there,
2972 // execute VF scalar calls, and then gather the result into the vector return
2974 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
2976 return ScalarCallCost;
2978 // Compute corresponding vector type for return value and arguments.
2979 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
2980 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
2981 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
2983 // Compute costs of unpacking argument values for the scalar calls and
2984 // packing the return values to a vector.
2985 unsigned ScalarizationCost =
2986 getScalarizationOverhead(RetTy, true, false, TTI);
2987 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
2988 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
2990 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
2992 // If we can't emit a vector call for this function, then the currently found
2993 // cost is the cost we need to return.
2994 NeedToScalarize = true;
2995 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
2998 // If the corresponding vector cost is cheaper, return its cost.
2999 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3000 if (VectorCallCost < Cost) {
3001 NeedToScalarize = false;
3002 return VectorCallCost;
3007 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3008 // factor VF. Return the cost of the instruction, including scalarization
3009 // overhead if it's needed.
3010 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3011 const TargetTransformInfo &TTI,
3012 const TargetLibraryInfo *TLI) {
3013 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3014 assert(ID && "Expected intrinsic call!");
3016 Type *RetTy = ToVectorTy(CI->getType(), VF);
3017 SmallVector<Type *, 4> Tys;
3018 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3019 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3021 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3024 void InnerLoopVectorizer::vectorizeLoop() {
3025 //===------------------------------------------------===//
3027 // Notice: any optimization or new instruction that go
3028 // into the code below should be also be implemented in
3031 //===------------------------------------------------===//
3032 Constant *Zero = Builder.getInt32(0);
3034 // In order to support reduction variables we need to be able to vectorize
3035 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3036 // stages. First, we create a new vector PHI node with no incoming edges.
3037 // We use this value when we vectorize all of the instructions that use the
3038 // PHI. Next, after all of the instructions in the block are complete we
3039 // add the new incoming edges to the PHI. At this point all of the
3040 // instructions in the basic block are vectorized, so we can use them to
3041 // construct the PHI.
3042 PhiVector RdxPHIsToFix;
3044 // Scan the loop in a topological order to ensure that defs are vectorized
3046 LoopBlocksDFS DFS(OrigLoop);
3049 // Vectorize all of the blocks in the original loop.
3050 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3051 be = DFS.endRPO(); bb != be; ++bb)
3052 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3054 // At this point every instruction in the original loop is widened to
3055 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3056 // that we vectorized. The PHI nodes are currently empty because we did
3057 // not want to introduce cycles. Notice that the remaining PHI nodes
3058 // that we need to fix are reduction variables.
3060 // Create the 'reduced' values for each of the induction vars.
3061 // The reduced values are the vector values that we scalarize and combine
3062 // after the loop is finished.
3063 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3065 PHINode *RdxPhi = *it;
3066 assert(RdxPhi && "Unable to recover vectorized PHI");
3068 // Find the reduction variable descriptor.
3069 assert(Legal->getReductionVars()->count(RdxPhi) &&
3070 "Unable to find the reduction variable");
3071 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3073 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3074 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3075 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3076 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3077 RdxDesc.getMinMaxRecurrenceKind();
3078 setDebugLocFromInst(Builder, ReductionStartValue);
3080 // We need to generate a reduction vector from the incoming scalar.
3081 // To do so, we need to generate the 'identity' vector and override
3082 // one of the elements with the incoming scalar reduction. We need
3083 // to do it in the vector-loop preheader.
3084 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3086 // This is the vector-clone of the value that leaves the loop.
3087 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3088 Type *VecTy = VectorExit[0]->getType();
3090 // Find the reduction identity variable. Zero for addition, or, xor,
3091 // one for multiplication, -1 for And.
3094 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3095 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3096 // MinMax reduction have the start value as their identify.
3098 VectorStart = Identity = ReductionStartValue;
3100 VectorStart = Identity =
3101 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3104 // Handle other reduction kinds:
3105 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3106 RK, VecTy->getScalarType());
3109 // This vector is the Identity vector where the first element is the
3110 // incoming scalar reduction.
3111 VectorStart = ReductionStartValue;
3113 Identity = ConstantVector::getSplat(VF, Iden);
3115 // This vector is the Identity vector where the first element is the
3116 // incoming scalar reduction.
3118 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3122 // Fix the vector-loop phi.
3124 // Reductions do not have to start at zero. They can start with
3125 // any loop invariant values.
3126 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3127 BasicBlock *Latch = OrigLoop->getLoopLatch();
3128 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3129 VectorParts &Val = getVectorValue(LoopVal);
3130 for (unsigned part = 0; part < UF; ++part) {
3131 // Make sure to add the reduction stat value only to the
3132 // first unroll part.
3133 Value *StartVal = (part == 0) ? VectorStart : Identity;
3134 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3135 LoopVectorPreHeader);
3136 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3137 LoopVectorBody.back());
3140 // Before each round, move the insertion point right between
3141 // the PHIs and the values we are going to write.
3142 // This allows us to write both PHINodes and the extractelement
3144 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
3146 VectorParts RdxParts;
3147 setDebugLocFromInst(Builder, LoopExitInst);
3148 for (unsigned part = 0; part < UF; ++part) {
3149 // This PHINode contains the vectorized reduction variable, or
3150 // the initial value vector, if we bypass the vector loop.
3151 VectorParts &RdxExitVal = getVectorValue(LoopExitInst);
3152 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
3153 Value *StartVal = (part == 0) ? VectorStart : Identity;
3154 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3155 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
3156 NewPhi->addIncoming(RdxExitVal[part],
3157 LoopVectorBody.back());
3158 RdxParts.push_back(NewPhi);
3161 // Reduce all of the unrolled parts into a single vector.
3162 Value *ReducedPartRdx = RdxParts[0];
3163 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3164 setDebugLocFromInst(Builder, ReducedPartRdx);
3165 for (unsigned part = 1; part < UF; ++part) {
3166 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3167 // Floating point operations had to be 'fast' to enable the reduction.
3168 ReducedPartRdx = addFastMathFlag(
3169 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3170 ReducedPartRdx, "bin.rdx"));
3172 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3173 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3177 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3178 // and vector ops, reducing the set of values being computed by half each
3180 assert(isPowerOf2_32(VF) &&
3181 "Reduction emission only supported for pow2 vectors!");
3182 Value *TmpVec = ReducedPartRdx;
3183 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3184 for (unsigned i = VF; i != 1; i >>= 1) {
3185 // Move the upper half of the vector to the lower half.
3186 for (unsigned j = 0; j != i/2; ++j)
3187 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3189 // Fill the rest of the mask with undef.
3190 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3191 UndefValue::get(Builder.getInt32Ty()));
3194 Builder.CreateShuffleVector(TmpVec,
3195 UndefValue::get(TmpVec->getType()),
3196 ConstantVector::get(ShuffleMask),
3199 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3200 // Floating point operations had to be 'fast' to enable the reduction.
3201 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3202 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3204 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3208 // The result is in the first element of the vector.
3209 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3210 Builder.getInt32(0));
3213 // Create a phi node that merges control-flow from the backedge-taken check
3214 // block and the middle block.
3215 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3216 LoopScalarPreHeader->getTerminator());
3217 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[0]);
3218 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3220 // Now, we need to fix the users of the reduction variable
3221 // inside and outside of the scalar remainder loop.
3222 // We know that the loop is in LCSSA form. We need to update the
3223 // PHI nodes in the exit blocks.
3224 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3225 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3226 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3227 if (!LCSSAPhi) break;
3229 // All PHINodes need to have a single entry edge, or two if
3230 // we already fixed them.
3231 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3233 // We found our reduction value exit-PHI. Update it with the
3234 // incoming bypass edge.
3235 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3236 // Add an edge coming from the bypass.
3237 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3240 }// end of the LCSSA phi scan.
3242 // Fix the scalar loop reduction variable with the incoming reduction sum
3243 // from the vector body and from the backedge value.
3244 int IncomingEdgeBlockIdx =
3245 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3246 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3247 // Pick the other block.
3248 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3249 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3250 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3251 }// end of for each redux variable.
3255 // Remove redundant induction instructions.
3256 cse(LoopVectorBody);
3259 void InnerLoopVectorizer::fixLCSSAPHIs() {
3260 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3261 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3262 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3263 if (!LCSSAPhi) break;
3264 if (LCSSAPhi->getNumIncomingValues() == 1)
3265 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3270 InnerLoopVectorizer::VectorParts
3271 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3272 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3275 // Look for cached value.
3276 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3277 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3278 if (ECEntryIt != MaskCache.end())
3279 return ECEntryIt->second;
3281 VectorParts SrcMask = createBlockInMask(Src);
3283 // The terminator has to be a branch inst!
3284 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3285 assert(BI && "Unexpected terminator found");
3287 if (BI->isConditional()) {
3288 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3290 if (BI->getSuccessor(0) != Dst)
3291 for (unsigned part = 0; part < UF; ++part)
3292 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3294 for (unsigned part = 0; part < UF; ++part)
3295 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3297 MaskCache[Edge] = EdgeMask;
3301 MaskCache[Edge] = SrcMask;
3305 InnerLoopVectorizer::VectorParts
3306 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3307 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3309 // Loop incoming mask is all-one.
3310 if (OrigLoop->getHeader() == BB) {
3311 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3312 return getVectorValue(C);
3315 // This is the block mask. We OR all incoming edges, and with zero.
3316 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3317 VectorParts BlockMask = getVectorValue(Zero);
3320 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3321 VectorParts EM = createEdgeMask(*it, BB);
3322 for (unsigned part = 0; part < UF; ++part)
3323 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3329 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3330 InnerLoopVectorizer::VectorParts &Entry,
3331 unsigned UF, unsigned VF, PhiVector *PV) {
3332 PHINode* P = cast<PHINode>(PN);
3333 // Handle reduction variables:
3334 if (Legal->getReductionVars()->count(P)) {
3335 for (unsigned part = 0; part < UF; ++part) {
3336 // This is phase one of vectorizing PHIs.
3337 Type *VecTy = (VF == 1) ? PN->getType() :
3338 VectorType::get(PN->getType(), VF);
3339 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3340 LoopVectorBody.back()-> getFirstInsertionPt());
3346 setDebugLocFromInst(Builder, P);
3347 // Check for PHI nodes that are lowered to vector selects.
3348 if (P->getParent() != OrigLoop->getHeader()) {
3349 // We know that all PHIs in non-header blocks are converted into
3350 // selects, so we don't have to worry about the insertion order and we
3351 // can just use the builder.
3352 // At this point we generate the predication tree. There may be
3353 // duplications since this is a simple recursive scan, but future
3354 // optimizations will clean it up.
3356 unsigned NumIncoming = P->getNumIncomingValues();
3358 // Generate a sequence of selects of the form:
3359 // SELECT(Mask3, In3,
3360 // SELECT(Mask2, In2,
3362 for (unsigned In = 0; In < NumIncoming; In++) {
3363 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3365 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3367 for (unsigned part = 0; part < UF; ++part) {
3368 // We might have single edge PHIs (blocks) - use an identity
3369 // 'select' for the first PHI operand.
3371 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3374 // Select between the current value and the previous incoming edge
3375 // based on the incoming mask.
3376 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3377 Entry[part], "predphi");
3383 // This PHINode must be an induction variable.
3384 // Make sure that we know about it.
3385 assert(Legal->getInductionVars()->count(P) &&
3386 "Not an induction variable");
3388 LoopVectorizationLegality::InductionInfo II =
3389 Legal->getInductionVars()->lookup(P);
3391 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3392 // which can be found from the original scalar operations.
3394 case LoopVectorizationLegality::IK_NoInduction:
3395 llvm_unreachable("Unknown induction");
3396 case LoopVectorizationLegality::IK_IntInduction: {
3397 assert(P->getType() == II.StartValue->getType() && "Types must match");
3398 Type *PhiTy = P->getType();
3400 if (P == OldInduction) {
3401 // Handle the canonical induction variable. We might have had to
3403 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3405 // Handle other induction variables that are now based on the
3407 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3409 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3410 Broadcasted = II.transform(Builder, NormalizedIdx);
3411 Broadcasted->setName("offset.idx");
3413 Broadcasted = getBroadcastInstrs(Broadcasted);
3414 // After broadcasting the induction variable we need to make the vector
3415 // consecutive by adding 0, 1, 2, etc.
3416 for (unsigned part = 0; part < UF; ++part)
3417 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3420 case LoopVectorizationLegality::IK_PtrInduction:
3421 // Handle the pointer induction variable case.
3422 assert(P->getType()->isPointerTy() && "Unexpected type.");
3423 // This is the normalized GEP that starts counting at zero.
3424 Value *NormalizedIdx =
3425 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3427 Builder.CreateSExtOrTrunc(NormalizedIdx, II.StepValue->getType());
3428 // This is the vector of results. Notice that we don't generate
3429 // vector geps because scalar geps result in better code.
3430 for (unsigned part = 0; part < UF; ++part) {
3432 int EltIndex = part;
3433 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3434 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3435 Value *SclrGep = II.transform(Builder, GlobalIdx);
3436 SclrGep->setName("next.gep");
3437 Entry[part] = SclrGep;
3441 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3442 for (unsigned int i = 0; i < VF; ++i) {
3443 int EltIndex = i + part * VF;
3444 Constant *Idx = ConstantInt::get(NormalizedIdx->getType(), EltIndex);
3445 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3446 Value *SclrGep = II.transform(Builder, GlobalIdx);
3447 SclrGep->setName("next.gep");
3448 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3449 Builder.getInt32(i),
3452 Entry[part] = VecVal;
3458 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3459 // For each instruction in the old loop.
3460 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3461 VectorParts &Entry = WidenMap.get(it);
3462 switch (it->getOpcode()) {
3463 case Instruction::Br:
3464 // Nothing to do for PHIs and BR, since we already took care of the
3465 // loop control flow instructions.
3467 case Instruction::PHI: {
3468 // Vectorize PHINodes.
3469 widenPHIInstruction(it, Entry, UF, VF, PV);
3473 case Instruction::Add:
3474 case Instruction::FAdd:
3475 case Instruction::Sub:
3476 case Instruction::FSub:
3477 case Instruction::Mul:
3478 case Instruction::FMul:
3479 case Instruction::UDiv:
3480 case Instruction::SDiv:
3481 case Instruction::FDiv:
3482 case Instruction::URem:
3483 case Instruction::SRem:
3484 case Instruction::FRem:
3485 case Instruction::Shl:
3486 case Instruction::LShr:
3487 case Instruction::AShr:
3488 case Instruction::And:
3489 case Instruction::Or:
3490 case Instruction::Xor: {
3491 // Just widen binops.
3492 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3493 setDebugLocFromInst(Builder, BinOp);
3494 VectorParts &A = getVectorValue(it->getOperand(0));
3495 VectorParts &B = getVectorValue(it->getOperand(1));
3497 // Use this vector value for all users of the original instruction.
3498 for (unsigned Part = 0; Part < UF; ++Part) {
3499 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3501 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3502 VecOp->copyIRFlags(BinOp);
3507 propagateMetadata(Entry, it);
3510 case Instruction::Select: {
3512 // If the selector is loop invariant we can create a select
3513 // instruction with a scalar condition. Otherwise, use vector-select.
3514 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3516 setDebugLocFromInst(Builder, it);
3518 // The condition can be loop invariant but still defined inside the
3519 // loop. This means that we can't just use the original 'cond' value.
3520 // We have to take the 'vectorized' value and pick the first lane.
3521 // Instcombine will make this a no-op.
3522 VectorParts &Cond = getVectorValue(it->getOperand(0));
3523 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3524 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3526 Value *ScalarCond = (VF == 1) ? Cond[0] :
3527 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3529 for (unsigned Part = 0; Part < UF; ++Part) {
3530 Entry[Part] = Builder.CreateSelect(
3531 InvariantCond ? ScalarCond : Cond[Part],
3536 propagateMetadata(Entry, it);
3540 case Instruction::ICmp:
3541 case Instruction::FCmp: {
3542 // Widen compares. Generate vector compares.
3543 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3544 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3545 setDebugLocFromInst(Builder, it);
3546 VectorParts &A = getVectorValue(it->getOperand(0));
3547 VectorParts &B = getVectorValue(it->getOperand(1));
3548 for (unsigned Part = 0; Part < UF; ++Part) {
3551 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3553 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3557 propagateMetadata(Entry, it);
3561 case Instruction::Store:
3562 case Instruction::Load:
3563 vectorizeMemoryInstruction(it);
3565 case Instruction::ZExt:
3566 case Instruction::SExt:
3567 case Instruction::FPToUI:
3568 case Instruction::FPToSI:
3569 case Instruction::FPExt:
3570 case Instruction::PtrToInt:
3571 case Instruction::IntToPtr:
3572 case Instruction::SIToFP:
3573 case Instruction::UIToFP:
3574 case Instruction::Trunc:
3575 case Instruction::FPTrunc:
3576 case Instruction::BitCast: {
3577 CastInst *CI = dyn_cast<CastInst>(it);
3578 setDebugLocFromInst(Builder, it);
3579 /// Optimize the special case where the source is the induction
3580 /// variable. Notice that we can only optimize the 'trunc' case
3581 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3582 /// c. other casts depend on pointer size.
3583 if (CI->getOperand(0) == OldInduction &&
3584 it->getOpcode() == Instruction::Trunc) {
3585 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3587 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3588 LoopVectorizationLegality::InductionInfo II =
3589 Legal->getInductionVars()->lookup(OldInduction);
3591 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3592 for (unsigned Part = 0; Part < UF; ++Part)
3593 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3594 propagateMetadata(Entry, it);
3597 /// Vectorize casts.
3598 Type *DestTy = (VF == 1) ? CI->getType() :
3599 VectorType::get(CI->getType(), VF);
3601 VectorParts &A = getVectorValue(it->getOperand(0));
3602 for (unsigned Part = 0; Part < UF; ++Part)
3603 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3604 propagateMetadata(Entry, it);
3608 case Instruction::Call: {
3609 // Ignore dbg intrinsics.
3610 if (isa<DbgInfoIntrinsic>(it))
3612 setDebugLocFromInst(Builder, it);
3614 Module *M = BB->getParent()->getParent();
3615 CallInst *CI = cast<CallInst>(it);
3617 StringRef FnName = CI->getCalledFunction()->getName();
3618 Function *F = CI->getCalledFunction();
3619 Type *RetTy = ToVectorTy(CI->getType(), VF);
3620 SmallVector<Type *, 4> Tys;
3621 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3622 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3624 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3626 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3627 ID == Intrinsic::lifetime_start)) {
3628 scalarizeInstruction(it);
3631 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3632 // version of the instruction.
3633 // Is it beneficial to perform intrinsic call compared to lib call?
3634 bool NeedToScalarize;
3635 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3636 bool UseVectorIntrinsic =
3637 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3638 if (!UseVectorIntrinsic && NeedToScalarize) {
3639 scalarizeInstruction(it);
3643 for (unsigned Part = 0; Part < UF; ++Part) {
3644 SmallVector<Value *, 4> Args;
3645 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3646 Value *Arg = CI->getArgOperand(i);
3647 // Some intrinsics have a scalar argument - don't replace it with a
3649 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3650 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3651 Arg = VectorArg[Part];
3653 Args.push_back(Arg);
3657 if (UseVectorIntrinsic) {
3658 // Use vector version of the intrinsic.
3659 Type *TysForDecl[] = {CI->getType()};
3661 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3662 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3664 // Use vector version of the library call.
3665 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3666 assert(!VFnName.empty() && "Vector function name is empty.");
3667 VectorF = M->getFunction(VFnName);
3669 // Generate a declaration
3670 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3672 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3673 VectorF->copyAttributesFrom(F);
3676 assert(VectorF && "Can't create vector function.");
3677 Entry[Part] = Builder.CreateCall(VectorF, Args);
3680 propagateMetadata(Entry, it);
3685 // All other instructions are unsupported. Scalarize them.
3686 scalarizeInstruction(it);
3689 }// end of for_each instr.
3692 void InnerLoopVectorizer::updateAnalysis() {
3693 // Forget the original basic block.
3694 SE->forgetLoop(OrigLoop);
3696 // Update the dominator tree information.
3697 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3698 "Entry does not dominate exit.");
3700 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3701 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3702 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3704 // Due to if predication of stores we might create a sequence of "if(pred)
3705 // a[i] = ...; " blocks.
3706 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3708 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3709 else if (isPredicatedBlock(i)) {
3710 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3712 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3716 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3717 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3718 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3719 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3721 DEBUG(DT->verifyDomTree());
3724 /// \brief Check whether it is safe to if-convert this phi node.
3726 /// Phi nodes with constant expressions that can trap are not safe to if
3728 static bool canIfConvertPHINodes(BasicBlock *BB) {
3729 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3730 PHINode *Phi = dyn_cast<PHINode>(I);
3733 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3734 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3741 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3742 if (!EnableIfConversion) {
3743 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3747 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3749 // A list of pointers that we can safely read and write to.
3750 SmallPtrSet<Value *, 8> SafePointes;
3752 // Collect safe addresses.
3753 for (Loop::block_iterator BI = TheLoop->block_begin(),
3754 BE = TheLoop->block_end(); BI != BE; ++BI) {
3755 BasicBlock *BB = *BI;
3757 if (blockNeedsPredication(BB))
3760 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3761 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3762 SafePointes.insert(LI->getPointerOperand());
3763 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3764 SafePointes.insert(SI->getPointerOperand());
3768 // Collect the blocks that need predication.
3769 BasicBlock *Header = TheLoop->getHeader();
3770 for (Loop::block_iterator BI = TheLoop->block_begin(),
3771 BE = TheLoop->block_end(); BI != BE; ++BI) {
3772 BasicBlock *BB = *BI;
3774 // We don't support switch statements inside loops.
3775 if (!isa<BranchInst>(BB->getTerminator())) {
3776 emitAnalysis(VectorizationReport(BB->getTerminator())
3777 << "loop contains a switch statement");
3781 // We must be able to predicate all blocks that need to be predicated.
3782 if (blockNeedsPredication(BB)) {
3783 if (!blockCanBePredicated(BB, SafePointes)) {
3784 emitAnalysis(VectorizationReport(BB->getTerminator())
3785 << "control flow cannot be substituted for a select");
3788 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3789 emitAnalysis(VectorizationReport(BB->getTerminator())
3790 << "control flow cannot be substituted for a select");
3795 // We can if-convert this loop.
3799 bool LoopVectorizationLegality::canVectorize() {
3800 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3801 // be canonicalized.
3802 if (!TheLoop->getLoopPreheader()) {
3804 VectorizationReport() <<
3805 "loop control flow is not understood by vectorizer");
3809 // We can only vectorize innermost loops.
3810 if (!TheLoop->empty()) {
3811 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3815 // We must have a single backedge.
3816 if (TheLoop->getNumBackEdges() != 1) {
3818 VectorizationReport() <<
3819 "loop control flow is not understood by vectorizer");
3823 // We must have a single exiting block.
3824 if (!TheLoop->getExitingBlock()) {
3826 VectorizationReport() <<
3827 "loop control flow is not understood by vectorizer");
3831 // We only handle bottom-tested loops, i.e. loop in which the condition is
3832 // checked at the end of each iteration. With that we can assume that all
3833 // instructions in the loop are executed the same number of times.
3834 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3836 VectorizationReport() <<
3837 "loop control flow is not understood by vectorizer");
3841 // We need to have a loop header.
3842 DEBUG(dbgs() << "LV: Found a loop: " <<
3843 TheLoop->getHeader()->getName() << '\n');
3845 // Check if we can if-convert non-single-bb loops.
3846 unsigned NumBlocks = TheLoop->getNumBlocks();
3847 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3848 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3852 // ScalarEvolution needs to be able to find the exit count.
3853 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3854 if (ExitCount == SE->getCouldNotCompute()) {
3855 emitAnalysis(VectorizationReport() <<
3856 "could not determine number of loop iterations");
3857 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3861 // Check if we can vectorize the instructions and CFG in this loop.
3862 if (!canVectorizeInstrs()) {
3863 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3867 // Go over each instruction and look at memory deps.
3868 if (!canVectorizeMemory()) {
3869 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3873 // Collect all of the variables that remain uniform after vectorization.
3874 collectLoopUniforms();
3876 DEBUG(dbgs() << "LV: We can vectorize this loop"
3877 << (LAI->getRuntimePointerChecking()->Need
3878 ? " (with a runtime bound check)"
3882 // Analyze interleaved memory accesses.
3883 if (EnableInterleavedMemAccesses)
3884 InterleaveInfo.analyzeInterleaving(Strides);
3886 // Okay! We can vectorize. At this point we don't have any other mem analysis
3887 // which may limit our maximum vectorization factor, so just return true with
3892 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3893 if (Ty->isPointerTy())
3894 return DL.getIntPtrType(Ty);
3896 // It is possible that char's or short's overflow when we ask for the loop's
3897 // trip count, work around this by changing the type size.
3898 if (Ty->getScalarSizeInBits() < 32)
3899 return Type::getInt32Ty(Ty->getContext());
3904 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3905 Ty0 = convertPointerToIntegerType(DL, Ty0);
3906 Ty1 = convertPointerToIntegerType(DL, Ty1);
3907 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3912 /// \brief Check that the instruction has outside loop users and is not an
3913 /// identified reduction variable.
3914 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3915 SmallPtrSetImpl<Value *> &Reductions) {
3916 // Reduction instructions are allowed to have exit users. All other
3917 // instructions must not have external users.
3918 if (!Reductions.count(Inst))
3919 //Check that all of the users of the loop are inside the BB.
3920 for (User *U : Inst->users()) {
3921 Instruction *UI = cast<Instruction>(U);
3922 // This user may be a reduction exit value.
3923 if (!TheLoop->contains(UI)) {
3924 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3931 bool LoopVectorizationLegality::canVectorizeInstrs() {
3932 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3933 BasicBlock *Header = TheLoop->getHeader();
3935 // Look for the attribute signaling the absence of NaNs.
3936 Function &F = *Header->getParent();
3937 const DataLayout &DL = F.getParent()->getDataLayout();
3938 if (F.hasFnAttribute("no-nans-fp-math"))
3940 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3942 // For each block in the loop.
3943 for (Loop::block_iterator bb = TheLoop->block_begin(),
3944 be = TheLoop->block_end(); bb != be; ++bb) {
3946 // Scan the instructions in the block and look for hazards.
3947 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3950 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3951 Type *PhiTy = Phi->getType();
3952 // Check that this PHI type is allowed.
3953 if (!PhiTy->isIntegerTy() &&
3954 !PhiTy->isFloatingPointTy() &&
3955 !PhiTy->isPointerTy()) {
3956 emitAnalysis(VectorizationReport(it)
3957 << "loop control flow is not understood by vectorizer");
3958 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3962 // If this PHINode is not in the header block, then we know that we
3963 // can convert it to select during if-conversion. No need to check if
3964 // the PHIs in this block are induction or reduction variables.
3965 if (*bb != Header) {
3966 // Check that this instruction has no outside users or is an
3967 // identified reduction value with an outside user.
3968 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3970 emitAnalysis(VectorizationReport(it) <<
3971 "value could not be identified as "
3972 "an induction or reduction variable");
3976 // We only allow if-converted PHIs with exactly two incoming values.
3977 if (Phi->getNumIncomingValues() != 2) {
3978 emitAnalysis(VectorizationReport(it)
3979 << "control flow not understood by vectorizer");
3980 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3984 // This is the value coming from the preheader.
3985 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3986 ConstantInt *StepValue = nullptr;
3987 // Check if this is an induction variable.
3988 InductionKind IK = isInductionVariable(Phi, StepValue);
3990 if (IK_NoInduction != IK) {
3991 // Get the widest type.
3993 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
3995 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
3997 // Int inductions are special because we only allow one IV.
3998 if (IK == IK_IntInduction && StepValue->isOne()) {
3999 // Use the phi node with the widest type as induction. Use the last
4000 // one if there are multiple (no good reason for doing this other
4001 // than it is expedient).
4002 if (!Induction || PhiTy == WidestIndTy)
4006 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4007 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
4009 // Until we explicitly handle the case of an induction variable with
4010 // an outside loop user we have to give up vectorizing this loop.
4011 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4012 emitAnalysis(VectorizationReport(it) <<
4013 "use of induction value outside of the "
4014 "loop is not handled by vectorizer");
4021 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4023 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4027 emitAnalysis(VectorizationReport(it) <<
4028 "value that could not be identified as "
4029 "reduction is used outside the loop");
4030 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4032 }// end of PHI handling
4034 // We handle calls that:
4035 // * Are debug info intrinsics.
4036 // * Have a mapping to an IR intrinsic.
4037 // * Have a vector version available.
4038 CallInst *CI = dyn_cast<CallInst>(it);
4039 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4040 !(CI->getCalledFunction() && TLI &&
4041 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4042 emitAnalysis(VectorizationReport(it) <<
4043 "call instruction cannot be vectorized");
4044 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4048 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4049 // second argument is the same (i.e. loop invariant)
4051 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4052 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4053 emitAnalysis(VectorizationReport(it)
4054 << "intrinsic instruction cannot be vectorized");
4055 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4060 // Check that the instruction return type is vectorizable.
4061 // Also, we can't vectorize extractelement instructions.
4062 if ((!VectorType::isValidElementType(it->getType()) &&
4063 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4064 emitAnalysis(VectorizationReport(it)
4065 << "instruction return type cannot be vectorized");
4066 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4070 // Check that the stored type is vectorizable.
4071 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4072 Type *T = ST->getValueOperand()->getType();
4073 if (!VectorType::isValidElementType(T)) {
4074 emitAnalysis(VectorizationReport(ST) <<
4075 "store instruction cannot be vectorized");
4078 if (EnableMemAccessVersioning)
4079 collectStridedAccess(ST);
4082 if (EnableMemAccessVersioning)
4083 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4084 collectStridedAccess(LI);
4086 // Reduction instructions are allowed to have exit users.
4087 // All other instructions must not have external users.
4088 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
4089 emitAnalysis(VectorizationReport(it) <<
4090 "value cannot be used outside the loop");
4099 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4100 if (Inductions.empty()) {
4101 emitAnalysis(VectorizationReport()
4102 << "loop induction variable could not be identified");
4110 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4111 Value *Ptr = nullptr;
4112 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4113 Ptr = LI->getPointerOperand();
4114 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4115 Ptr = SI->getPointerOperand();
4119 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4123 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4124 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4125 Strides[Ptr] = Stride;
4126 StrideSet.insert(Stride);
4129 void LoopVectorizationLegality::collectLoopUniforms() {
4130 // We now know that the loop is vectorizable!
4131 // Collect variables that will remain uniform after vectorization.
4132 std::vector<Value*> Worklist;
4133 BasicBlock *Latch = TheLoop->getLoopLatch();
4135 // Start with the conditional branch and walk up the block.
4136 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4138 // Also add all consecutive pointer values; these values will be uniform
4139 // after vectorization (and subsequent cleanup) and, until revectorization is
4140 // supported, all dependencies must also be uniform.
4141 for (Loop::block_iterator B = TheLoop->block_begin(),
4142 BE = TheLoop->block_end(); B != BE; ++B)
4143 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4145 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4146 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4148 while (!Worklist.empty()) {
4149 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4150 Worklist.pop_back();
4152 // Look at instructions inside this loop.
4153 // Stop when reaching PHI nodes.
4154 // TODO: we need to follow values all over the loop, not only in this block.
4155 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4158 // This is a known uniform.
4161 // Insert all operands.
4162 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4166 bool LoopVectorizationLegality::canVectorizeMemory() {
4167 LAI = &LAA->getInfo(TheLoop, Strides);
4168 auto &OptionalReport = LAI->getReport();
4170 emitAnalysis(VectorizationReport(*OptionalReport));
4171 if (!LAI->canVectorizeMemory())
4174 if (LAI->hasStoreToLoopInvariantAddress()) {
4176 VectorizationReport()
4177 << "write to a loop invariant address could not be vectorized");
4178 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4182 if (LAI->getNumRuntimePointerChecks() >
4183 VectorizerParams::RuntimeMemoryCheckThreshold) {
4184 emitAnalysis(VectorizationReport()
4185 << LAI->getNumRuntimePointerChecks() << " exceeds limit of "
4186 << VectorizerParams::RuntimeMemoryCheckThreshold
4187 << " dependent memory operations checked at runtime");
4188 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
4194 LoopVectorizationLegality::InductionKind
4195 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4196 ConstantInt *&StepValue) {
4197 if (!isInductionPHI(Phi, SE, StepValue))
4198 return IK_NoInduction;
4200 Type *PhiTy = Phi->getType();
4201 // Found an Integer induction variable.
4202 if (PhiTy->isIntegerTy())
4203 return IK_IntInduction;
4204 // Found an Pointer induction variable.
4205 return IK_PtrInduction;
4208 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4209 Value *In0 = const_cast<Value*>(V);
4210 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4214 return Inductions.count(PN);
4217 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4218 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4221 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4222 SmallPtrSetImpl<Value *> &SafePtrs) {
4224 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4225 // Check that we don't have a constant expression that can trap as operand.
4226 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4228 if (Constant *C = dyn_cast<Constant>(*OI))
4232 // We might be able to hoist the load.
4233 if (it->mayReadFromMemory()) {
4234 LoadInst *LI = dyn_cast<LoadInst>(it);
4237 if (!SafePtrs.count(LI->getPointerOperand())) {
4238 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4239 MaskedOp.insert(LI);
4246 // We don't predicate stores at the moment.
4247 if (it->mayWriteToMemory()) {
4248 StoreInst *SI = dyn_cast<StoreInst>(it);
4249 // We only support predication of stores in basic blocks with one
4254 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4255 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4257 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4258 !isSinglePredecessor) {
4259 // Build a masked store if it is legal for the target, otherwise scalarize
4261 bool isLegalMaskedOp =
4262 isLegalMaskedStore(SI->getValueOperand()->getType(),
4263 SI->getPointerOperand());
4264 if (isLegalMaskedOp) {
4266 MaskedOp.insert(SI);
4275 // The instructions below can trap.
4276 switch (it->getOpcode()) {
4278 case Instruction::UDiv:
4279 case Instruction::SDiv:
4280 case Instruction::URem:
4281 case Instruction::SRem:
4289 void InterleavedAccessInfo::collectConstStridedAccesses(
4290 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4291 const ValueToValueMap &Strides) {
4292 // Holds load/store instructions in program order.
4293 SmallVector<Instruction *, 16> AccessList;
4295 for (auto *BB : TheLoop->getBlocks()) {
4296 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4298 for (auto &I : *BB) {
4299 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4301 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4305 AccessList.push_back(&I);
4309 if (AccessList.empty())
4312 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4313 for (auto I : AccessList) {
4314 LoadInst *LI = dyn_cast<LoadInst>(I);
4315 StoreInst *SI = dyn_cast<StoreInst>(I);
4317 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4318 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides);
4320 // The factor of the corresponding interleave group.
4321 unsigned Factor = std::abs(Stride);
4323 // Ignore the access if the factor is too small or too large.
4324 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4327 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4328 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4329 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4331 // An alignment of 0 means target ABI alignment.
4332 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4334 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4336 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4340 // Analyze interleaved accesses and collect them into interleave groups.
4342 // Notice that the vectorization on interleaved groups will change instruction
4343 // orders and may break dependences. But the memory dependence check guarantees
4344 // that there is no overlap between two pointers of different strides, element
4345 // sizes or underlying bases.
4347 // For pointers sharing the same stride, element size and underlying base, no
4348 // need to worry about Read-After-Write dependences and Write-After-Read
4351 // E.g. The RAW dependence: A[i] = a;
4353 // This won't exist as it is a store-load forwarding conflict, which has
4354 // already been checked and forbidden in the dependence check.
4356 // E.g. The WAR dependence: a = A[i]; // (1)
4358 // The store group of (2) is always inserted at or below (2), and the load group
4359 // of (1) is always inserted at or above (1). The dependence is safe.
4360 void InterleavedAccessInfo::analyzeInterleaving(
4361 const ValueToValueMap &Strides) {
4362 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4364 // Holds all the stride accesses.
4365 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4366 collectConstStridedAccesses(StrideAccesses, Strides);
4368 if (StrideAccesses.empty())
4371 // Holds all interleaved store groups temporarily.
4372 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4374 // Search the load-load/write-write pair B-A in bottom-up order and try to
4375 // insert B into the interleave group of A according to 3 rules:
4376 // 1. A and B have the same stride.
4377 // 2. A and B have the same memory object size.
4378 // 3. B belongs to the group according to the distance.
4380 // The bottom-up order can avoid breaking the Write-After-Write dependences
4381 // between two pointers of the same base.
4382 // E.g. A[i] = a; (1)
4385 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4386 // above (1), which guarantees that (1) is always above (2).
4387 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4389 Instruction *A = I->first;
4390 StrideDescriptor DesA = I->second;
4392 InterleaveGroup *Group = getInterleaveGroup(A);
4394 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4395 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4398 if (A->mayWriteToMemory())
4399 StoreGroups.insert(Group);
4401 for (auto II = std::next(I); II != E; ++II) {
4402 Instruction *B = II->first;
4403 StrideDescriptor DesB = II->second;
4405 // Ignore if B is already in a group or B is a different memory operation.
4406 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4409 // Check the rule 1 and 2.
4410 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4413 // Calculate the distance and prepare for the rule 3.
4414 const SCEVConstant *DistToA =
4415 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4419 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4421 // Skip if the distance is not multiple of size as they are not in the
4423 if (DistanceToA % static_cast<int>(DesA.Size))
4426 // The index of B is the index of A plus the related index to A.
4428 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4430 // Try to insert B into the group.
4431 if (Group->insertMember(B, IndexB, DesB.Align)) {
4432 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4433 << " into the interleave group with" << *A << '\n');
4434 InterleaveGroupMap[B] = Group;
4436 // Set the first load in program order as the insert position.
4437 if (B->mayReadFromMemory())
4438 Group->setInsertPos(B);
4440 } // Iteration on instruction B
4441 } // Iteration on instruction A
4443 // Remove interleaved store groups with gaps.
4444 for (InterleaveGroup *Group : StoreGroups)
4445 if (Group->getNumMembers() != Group->getFactor())
4446 releaseGroup(Group);
4449 LoopVectorizationCostModel::VectorizationFactor
4450 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4451 // Width 1 means no vectorize
4452 VectorizationFactor Factor = { 1U, 0U };
4453 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4454 emitAnalysis(VectorizationReport() <<
4455 "runtime pointer checks needed. Enable vectorization of this "
4456 "loop with '#pragma clang loop vectorize(enable)' when "
4457 "compiling with -Os");
4458 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4462 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4463 emitAnalysis(VectorizationReport() <<
4464 "store that is conditionally executed prevents vectorization");
4465 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4469 // Find the trip count.
4470 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4471 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4473 unsigned WidestType = getWidestType();
4474 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4475 unsigned MaxSafeDepDist = -1U;
4476 if (Legal->getMaxSafeDepDistBytes() != -1U)
4477 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4478 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4479 WidestRegister : MaxSafeDepDist);
4480 unsigned MaxVectorSize = WidestRegister / WidestType;
4481 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4482 DEBUG(dbgs() << "LV: The Widest register is: "
4483 << WidestRegister << " bits.\n");
4485 if (MaxVectorSize == 0) {
4486 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4490 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4491 " into one vector!");
4493 unsigned VF = MaxVectorSize;
4495 // If we optimize the program for size, avoid creating the tail loop.
4497 // If we are unable to calculate the trip count then don't try to vectorize.
4500 (VectorizationReport() <<
4501 "unable to calculate the loop count due to complex control flow");
4502 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4506 // Find the maximum SIMD width that can fit within the trip count.
4507 VF = TC % MaxVectorSize;
4512 // If the trip count that we found modulo the vectorization factor is not
4513 // zero then we require a tail.
4514 emitAnalysis(VectorizationReport() <<
4515 "cannot optimize for size and vectorize at the "
4516 "same time. Enable vectorization of this loop "
4517 "with '#pragma clang loop vectorize(enable)' "
4518 "when compiling with -Os");
4519 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4524 int UserVF = Hints->getWidth();
4526 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4527 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4529 Factor.Width = UserVF;
4533 float Cost = expectedCost(1);
4535 const float ScalarCost = Cost;
4538 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4540 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4541 // Ignore scalar width, because the user explicitly wants vectorization.
4542 if (ForceVectorization && VF > 1) {
4544 Cost = expectedCost(Width) / (float)Width;
4547 for (unsigned i=2; i <= VF; i*=2) {
4548 // Notice that the vector loop needs to be executed less times, so
4549 // we need to divide the cost of the vector loops by the width of
4550 // the vector elements.
4551 float VectorCost = expectedCost(i) / (float)i;
4552 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4553 (int)VectorCost << ".\n");
4554 if (VectorCost < Cost) {
4560 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4561 << "LV: Vectorization seems to be not beneficial, "
4562 << "but was forced by a user.\n");
4563 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4564 Factor.Width = Width;
4565 Factor.Cost = Width * Cost;
4569 unsigned LoopVectorizationCostModel::getWidestType() {
4570 unsigned MaxWidth = 8;
4571 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4574 for (Loop::block_iterator bb = TheLoop->block_begin(),
4575 be = TheLoop->block_end(); bb != be; ++bb) {
4576 BasicBlock *BB = *bb;
4578 // For each instruction in the loop.
4579 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4580 Type *T = it->getType();
4582 // Ignore ephemeral values.
4583 if (EphValues.count(it))
4586 // Only examine Loads, Stores and PHINodes.
4587 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4590 // Examine PHI nodes that are reduction variables.
4591 if (PHINode *PN = dyn_cast<PHINode>(it))
4592 if (!Legal->getReductionVars()->count(PN))
4595 // Examine the stored values.
4596 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4597 T = ST->getValueOperand()->getType();
4599 // Ignore loaded pointer types and stored pointer types that are not
4600 // consecutive. However, we do want to take consecutive stores/loads of
4601 // pointer vectors into account.
4602 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4605 MaxWidth = std::max(MaxWidth,
4606 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4613 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4615 unsigned LoopCost) {
4617 // -- The interleave heuristics --
4618 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4619 // There are many micro-architectural considerations that we can't predict
4620 // at this level. For example, frontend pressure (on decode or fetch) due to
4621 // code size, or the number and capabilities of the execution ports.
4623 // We use the following heuristics to select the interleave count:
4624 // 1. If the code has reductions, then we interleave to break the cross
4625 // iteration dependency.
4626 // 2. If the loop is really small, then we interleave to reduce the loop
4628 // 3. We don't interleave if we think that we will spill registers to memory
4629 // due to the increased register pressure.
4631 // Use the user preference, unless 'auto' is selected.
4632 int UserUF = Hints->getInterleave();
4636 // When we optimize for size, we don't interleave.
4640 // We used the distance for the interleave count.
4641 if (Legal->getMaxSafeDepDistBytes() != -1U)
4644 // Do not interleave loops with a relatively small trip count.
4645 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4646 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4649 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4650 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4654 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4655 TargetNumRegisters = ForceTargetNumScalarRegs;
4657 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4658 TargetNumRegisters = ForceTargetNumVectorRegs;
4661 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4662 // We divide by these constants so assume that we have at least one
4663 // instruction that uses at least one register.
4664 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4665 R.NumInstructions = std::max(R.NumInstructions, 1U);
4667 // We calculate the interleave count using the following formula.
4668 // Subtract the number of loop invariants from the number of available
4669 // registers. These registers are used by all of the interleaved instances.
4670 // Next, divide the remaining registers by the number of registers that is
4671 // required by the loop, in order to estimate how many parallel instances
4672 // fit without causing spills. All of this is rounded down if necessary to be
4673 // a power of two. We want power of two interleave count to simplify any
4674 // addressing operations or alignment considerations.
4675 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4678 // Don't count the induction variable as interleaved.
4679 if (EnableIndVarRegisterHeur)
4680 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4681 std::max(1U, (R.MaxLocalUsers - 1)));
4683 // Clamp the interleave ranges to reasonable counts.
4684 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4686 // Check if the user has overridden the max.
4688 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4689 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4691 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4692 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4695 // If we did not calculate the cost for VF (because the user selected the VF)
4696 // then we calculate the cost of VF here.
4698 LoopCost = expectedCost(VF);
4700 // Clamp the calculated IC to be between the 1 and the max interleave count
4701 // that the target allows.
4702 if (IC > MaxInterleaveCount)
4703 IC = MaxInterleaveCount;
4707 // Interleave if we vectorized this loop and there is a reduction that could
4708 // benefit from interleaving.
4709 if (VF > 1 && Legal->getReductionVars()->size()) {
4710 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
4714 // Note that if we've already vectorized the loop we will have done the
4715 // runtime check and so interleaving won't require further checks.
4716 bool InterleavingRequiresRuntimePointerCheck =
4717 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
4719 // We want to interleave small loops in order to reduce the loop overhead and
4720 // potentially expose ILP opportunities.
4721 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4722 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
4723 // We assume that the cost overhead is 1 and we use the cost model
4724 // to estimate the cost of the loop and interleave until the cost of the
4725 // loop overhead is about 5% of the cost of the loop.
4727 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4729 // Interleave until store/load ports (estimated by max interleave count) are
4731 unsigned NumStores = Legal->getNumStores();
4732 unsigned NumLoads = Legal->getNumLoads();
4733 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
4734 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
4736 // If we have a scalar reduction (vector reductions are already dealt with
4737 // by this point), we can increase the critical path length if the loop
4738 // we're interleaving is inside another loop. Limit, by default to 2, so the
4739 // critical path only gets increased by one reduction operation.
4740 if (Legal->getReductionVars()->size() &&
4741 TheLoop->getLoopDepth() > 1) {
4742 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
4743 SmallIC = std::min(SmallIC, F);
4744 StoresIC = std::min(StoresIC, F);
4745 LoadsIC = std::min(LoadsIC, F);
4748 if (EnableLoadStoreRuntimeInterleave &&
4749 std::max(StoresIC, LoadsIC) > SmallIC) {
4750 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
4751 return std::max(StoresIC, LoadsIC);
4754 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
4758 // Interleave if this is a large loop (small loops are already dealt with by
4760 // point) that could benefit from interleaving.
4761 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4762 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4763 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
4767 DEBUG(dbgs() << "LV: Not Interleaving.\n");
4771 LoopVectorizationCostModel::RegisterUsage
4772 LoopVectorizationCostModel::calculateRegisterUsage() {
4773 // This function calculates the register usage by measuring the highest number
4774 // of values that are alive at a single location. Obviously, this is a very
4775 // rough estimation. We scan the loop in a topological order in order and
4776 // assign a number to each instruction. We use RPO to ensure that defs are
4777 // met before their users. We assume that each instruction that has in-loop
4778 // users starts an interval. We record every time that an in-loop value is
4779 // used, so we have a list of the first and last occurrences of each
4780 // instruction. Next, we transpose this data structure into a multi map that
4781 // holds the list of intervals that *end* at a specific location. This multi
4782 // map allows us to perform a linear search. We scan the instructions linearly
4783 // and record each time that a new interval starts, by placing it in a set.
4784 // If we find this value in the multi-map then we remove it from the set.
4785 // The max register usage is the maximum size of the set.
4786 // We also search for instructions that are defined outside the loop, but are
4787 // used inside the loop. We need this number separately from the max-interval
4788 // usage number because when we unroll, loop-invariant values do not take
4790 LoopBlocksDFS DFS(TheLoop);
4794 R.NumInstructions = 0;
4796 // Each 'key' in the map opens a new interval. The values
4797 // of the map are the index of the 'last seen' usage of the
4798 // instruction that is the key.
4799 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4800 // Maps instruction to its index.
4801 DenseMap<unsigned, Instruction*> IdxToInstr;
4802 // Marks the end of each interval.
4803 IntervalMap EndPoint;
4804 // Saves the list of instruction indices that are used in the loop.
4805 SmallSet<Instruction*, 8> Ends;
4806 // Saves the list of values that are used in the loop but are
4807 // defined outside the loop, such as arguments and constants.
4808 SmallPtrSet<Value*, 8> LoopInvariants;
4811 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4812 be = DFS.endRPO(); bb != be; ++bb) {
4813 R.NumInstructions += (*bb)->size();
4814 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4816 Instruction *I = it;
4817 IdxToInstr[Index++] = I;
4819 // Save the end location of each USE.
4820 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4821 Value *U = I->getOperand(i);
4822 Instruction *Instr = dyn_cast<Instruction>(U);
4824 // Ignore non-instruction values such as arguments, constants, etc.
4825 if (!Instr) continue;
4827 // If this instruction is outside the loop then record it and continue.
4828 if (!TheLoop->contains(Instr)) {
4829 LoopInvariants.insert(Instr);
4833 // Overwrite previous end points.
4834 EndPoint[Instr] = Index;
4840 // Saves the list of intervals that end with the index in 'key'.
4841 typedef SmallVector<Instruction*, 2> InstrList;
4842 DenseMap<unsigned, InstrList> TransposeEnds;
4844 // Transpose the EndPoints to a list of values that end at each index.
4845 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4847 TransposeEnds[it->second].push_back(it->first);
4849 SmallSet<Instruction*, 8> OpenIntervals;
4850 unsigned MaxUsage = 0;
4853 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4854 for (unsigned int i = 0; i < Index; ++i) {
4855 Instruction *I = IdxToInstr[i];
4856 // Ignore instructions that are never used within the loop.
4857 if (!Ends.count(I)) continue;
4859 // Ignore ephemeral values.
4860 if (EphValues.count(I))
4863 // Remove all of the instructions that end at this location.
4864 InstrList &List = TransposeEnds[i];
4865 for (unsigned int j=0, e = List.size(); j < e; ++j)
4866 OpenIntervals.erase(List[j]);
4868 // Count the number of live interals.
4869 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4871 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4872 OpenIntervals.size() << '\n');
4874 // Add the current instruction to the list of open intervals.
4875 OpenIntervals.insert(I);
4878 unsigned Invariant = LoopInvariants.size();
4879 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4880 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4881 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4883 R.LoopInvariantRegs = Invariant;
4884 R.MaxLocalUsers = MaxUsage;
4888 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4892 for (Loop::block_iterator bb = TheLoop->block_begin(),
4893 be = TheLoop->block_end(); bb != be; ++bb) {
4894 unsigned BlockCost = 0;
4895 BasicBlock *BB = *bb;
4897 // For each instruction in the old loop.
4898 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4899 // Skip dbg intrinsics.
4900 if (isa<DbgInfoIntrinsic>(it))
4903 // Ignore ephemeral values.
4904 if (EphValues.count(it))
4907 unsigned C = getInstructionCost(it, VF);
4909 // Check if we should override the cost.
4910 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4911 C = ForceTargetInstructionCost;
4914 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4915 VF << " For instruction: " << *it << '\n');
4918 // We assume that if-converted blocks have a 50% chance of being executed.
4919 // When the code is scalar then some of the blocks are avoided due to CF.
4920 // When the code is vectorized we execute all code paths.
4921 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4930 /// \brief Check whether the address computation for a non-consecutive memory
4931 /// access looks like an unlikely candidate for being merged into the indexing
4934 /// We look for a GEP which has one index that is an induction variable and all
4935 /// other indices are loop invariant. If the stride of this access is also
4936 /// within a small bound we decide that this address computation can likely be
4937 /// merged into the addressing mode.
4938 /// In all other cases, we identify the address computation as complex.
4939 static bool isLikelyComplexAddressComputation(Value *Ptr,
4940 LoopVectorizationLegality *Legal,
4941 ScalarEvolution *SE,
4942 const Loop *TheLoop) {
4943 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4947 // We are looking for a gep with all loop invariant indices except for one
4948 // which should be an induction variable.
4949 unsigned NumOperands = Gep->getNumOperands();
4950 for (unsigned i = 1; i < NumOperands; ++i) {
4951 Value *Opd = Gep->getOperand(i);
4952 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4953 !Legal->isInductionVariable(Opd))
4957 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4958 // can likely be merged into the address computation.
4959 unsigned MaxMergeDistance = 64;
4961 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4965 // Check the step is constant.
4966 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4967 // Calculate the pointer stride and check if it is consecutive.
4968 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4972 const APInt &APStepVal = C->getValue()->getValue();
4974 // Huge step value - give up.
4975 if (APStepVal.getBitWidth() > 64)
4978 int64_t StepVal = APStepVal.getSExtValue();
4980 return StepVal > MaxMergeDistance;
4983 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4984 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4990 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4991 // If we know that this instruction will remain uniform, check the cost of
4992 // the scalar version.
4993 if (Legal->isUniformAfterVectorization(I))
4996 Type *RetTy = I->getType();
4997 Type *VectorTy = ToVectorTy(RetTy, VF);
4999 // TODO: We need to estimate the cost of intrinsic calls.
5000 switch (I->getOpcode()) {
5001 case Instruction::GetElementPtr:
5002 // We mark this instruction as zero-cost because the cost of GEPs in
5003 // vectorized code depends on whether the corresponding memory instruction
5004 // is scalarized or not. Therefore, we handle GEPs with the memory
5005 // instruction cost.
5007 case Instruction::Br: {
5008 return TTI.getCFInstrCost(I->getOpcode());
5010 case Instruction::PHI:
5011 //TODO: IF-converted IFs become selects.
5013 case Instruction::Add:
5014 case Instruction::FAdd:
5015 case Instruction::Sub:
5016 case Instruction::FSub:
5017 case Instruction::Mul:
5018 case Instruction::FMul:
5019 case Instruction::UDiv:
5020 case Instruction::SDiv:
5021 case Instruction::FDiv:
5022 case Instruction::URem:
5023 case Instruction::SRem:
5024 case Instruction::FRem:
5025 case Instruction::Shl:
5026 case Instruction::LShr:
5027 case Instruction::AShr:
5028 case Instruction::And:
5029 case Instruction::Or:
5030 case Instruction::Xor: {
5031 // Since we will replace the stride by 1 the multiplication should go away.
5032 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5034 // Certain instructions can be cheaper to vectorize if they have a constant
5035 // second vector operand. One example of this are shifts on x86.
5036 TargetTransformInfo::OperandValueKind Op1VK =
5037 TargetTransformInfo::OK_AnyValue;
5038 TargetTransformInfo::OperandValueKind Op2VK =
5039 TargetTransformInfo::OK_AnyValue;
5040 TargetTransformInfo::OperandValueProperties Op1VP =
5041 TargetTransformInfo::OP_None;
5042 TargetTransformInfo::OperandValueProperties Op2VP =
5043 TargetTransformInfo::OP_None;
5044 Value *Op2 = I->getOperand(1);
5046 // Check for a splat of a constant or for a non uniform vector of constants.
5047 if (isa<ConstantInt>(Op2)) {
5048 ConstantInt *CInt = cast<ConstantInt>(Op2);
5049 if (CInt && CInt->getValue().isPowerOf2())
5050 Op2VP = TargetTransformInfo::OP_PowerOf2;
5051 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5052 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5053 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5054 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5056 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5057 if (CInt && CInt->getValue().isPowerOf2())
5058 Op2VP = TargetTransformInfo::OP_PowerOf2;
5059 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5063 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5066 case Instruction::Select: {
5067 SelectInst *SI = cast<SelectInst>(I);
5068 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5069 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5070 Type *CondTy = SI->getCondition()->getType();
5072 CondTy = VectorType::get(CondTy, VF);
5074 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5076 case Instruction::ICmp:
5077 case Instruction::FCmp: {
5078 Type *ValTy = I->getOperand(0)->getType();
5079 VectorTy = ToVectorTy(ValTy, VF);
5080 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5082 case Instruction::Store:
5083 case Instruction::Load: {
5084 StoreInst *SI = dyn_cast<StoreInst>(I);
5085 LoadInst *LI = dyn_cast<LoadInst>(I);
5086 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5088 VectorTy = ToVectorTy(ValTy, VF);
5090 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5091 unsigned AS = SI ? SI->getPointerAddressSpace() :
5092 LI->getPointerAddressSpace();
5093 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5094 // We add the cost of address computation here instead of with the gep
5095 // instruction because only here we know whether the operation is
5098 return TTI.getAddressComputationCost(VectorTy) +
5099 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5101 // For an interleaved access, calculate the total cost of the whole
5102 // interleave group.
5103 if (Legal->isAccessInterleaved(I)) {
5104 auto Group = Legal->getInterleavedAccessGroup(I);
5105 assert(Group && "Fail to get an interleaved access group.");
5107 // Only calculate the cost once at the insert position.
5108 if (Group->getInsertPos() != I)
5111 unsigned InterleaveFactor = Group->getFactor();
5113 VectorType::get(VectorTy->getVectorElementType(),
5114 VectorTy->getVectorNumElements() * InterleaveFactor);
5116 // Holds the indices of existing members in an interleaved load group.
5117 // An interleaved store group doesn't need this as it dones't allow gaps.
5118 SmallVector<unsigned, 4> Indices;
5120 for (unsigned i = 0; i < InterleaveFactor; i++)
5121 if (Group->getMember(i))
5122 Indices.push_back(i);
5125 // Calculate the cost of the whole interleaved group.
5126 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5127 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5128 Group->getAlignment(), AS);
5130 if (Group->isReverse())
5132 Group->getNumMembers() *
5133 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5135 // FIXME: The interleaved load group with a huge gap could be even more
5136 // expensive than scalar operations. Then we could ignore such group and
5137 // use scalar operations instead.
5141 // Scalarized loads/stores.
5142 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5143 bool Reverse = ConsecutiveStride < 0;
5144 const DataLayout &DL = I->getModule()->getDataLayout();
5145 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5146 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5147 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5148 bool IsComplexComputation =
5149 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5151 // The cost of extracting from the value vector and pointer vector.
5152 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5153 for (unsigned i = 0; i < VF; ++i) {
5154 // The cost of extracting the pointer operand.
5155 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5156 // In case of STORE, the cost of ExtractElement from the vector.
5157 // In case of LOAD, the cost of InsertElement into the returned
5159 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5160 Instruction::InsertElement,
5164 // The cost of the scalar loads/stores.
5165 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5166 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5171 // Wide load/stores.
5172 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5173 if (Legal->isMaskRequired(I))
5174 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5177 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5180 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5184 case Instruction::ZExt:
5185 case Instruction::SExt:
5186 case Instruction::FPToUI:
5187 case Instruction::FPToSI:
5188 case Instruction::FPExt:
5189 case Instruction::PtrToInt:
5190 case Instruction::IntToPtr:
5191 case Instruction::SIToFP:
5192 case Instruction::UIToFP:
5193 case Instruction::Trunc:
5194 case Instruction::FPTrunc:
5195 case Instruction::BitCast: {
5196 // We optimize the truncation of induction variable.
5197 // The cost of these is the same as the scalar operation.
5198 if (I->getOpcode() == Instruction::Trunc &&
5199 Legal->isInductionVariable(I->getOperand(0)))
5200 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5201 I->getOperand(0)->getType());
5203 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5204 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5206 case Instruction::Call: {
5207 bool NeedToScalarize;
5208 CallInst *CI = cast<CallInst>(I);
5209 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5210 if (getIntrinsicIDForCall(CI, TLI))
5211 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5215 // We are scalarizing the instruction. Return the cost of the scalar
5216 // instruction, plus the cost of insert and extract into vector
5217 // elements, times the vector width.
5220 if (!RetTy->isVoidTy() && VF != 1) {
5221 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5223 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5226 // The cost of inserting the results plus extracting each one of the
5228 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5231 // The cost of executing VF copies of the scalar instruction. This opcode
5232 // is unknown. Assume that it is the same as 'mul'.
5233 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5239 char LoopVectorize::ID = 0;
5240 static const char lv_name[] = "Loop Vectorization";
5241 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5242 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5243 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5244 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5245 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5246 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5247 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5248 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5249 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5250 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5251 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5252 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5255 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5256 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5260 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5261 // Check for a store.
5262 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5263 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5265 // Check for a load.
5266 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5267 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5273 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5274 bool IfPredicateStore) {
5275 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5276 // Holds vector parameters or scalars, in case of uniform vals.
5277 SmallVector<VectorParts, 4> Params;
5279 setDebugLocFromInst(Builder, Instr);
5281 // Find all of the vectorized parameters.
5282 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5283 Value *SrcOp = Instr->getOperand(op);
5285 // If we are accessing the old induction variable, use the new one.
5286 if (SrcOp == OldInduction) {
5287 Params.push_back(getVectorValue(SrcOp));
5291 // Try using previously calculated values.
5292 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5294 // If the src is an instruction that appeared earlier in the basic block
5295 // then it should already be vectorized.
5296 if (SrcInst && OrigLoop->contains(SrcInst)) {
5297 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5298 // The parameter is a vector value from earlier.
5299 Params.push_back(WidenMap.get(SrcInst));
5301 // The parameter is a scalar from outside the loop. Maybe even a constant.
5302 VectorParts Scalars;
5303 Scalars.append(UF, SrcOp);
5304 Params.push_back(Scalars);
5308 assert(Params.size() == Instr->getNumOperands() &&
5309 "Invalid number of operands");
5311 // Does this instruction return a value ?
5312 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5314 Value *UndefVec = IsVoidRetTy ? nullptr :
5315 UndefValue::get(Instr->getType());
5316 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5317 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5319 Instruction *InsertPt = Builder.GetInsertPoint();
5320 BasicBlock *IfBlock = Builder.GetInsertBlock();
5321 BasicBlock *CondBlock = nullptr;
5324 Loop *VectorLp = nullptr;
5325 if (IfPredicateStore) {
5326 assert(Instr->getParent()->getSinglePredecessor() &&
5327 "Only support single predecessor blocks");
5328 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5329 Instr->getParent());
5330 VectorLp = LI->getLoopFor(IfBlock);
5331 assert(VectorLp && "Must have a loop for this block");
5334 // For each vector unroll 'part':
5335 for (unsigned Part = 0; Part < UF; ++Part) {
5336 // For each scalar that we create:
5338 // Start an "if (pred) a[i] = ..." block.
5339 Value *Cmp = nullptr;
5340 if (IfPredicateStore) {
5341 if (Cond[Part]->getType()->isVectorTy())
5343 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5344 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5345 ConstantInt::get(Cond[Part]->getType(), 1));
5346 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5347 LoopVectorBody.push_back(CondBlock);
5348 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5349 // Update Builder with newly created basic block.
5350 Builder.SetInsertPoint(InsertPt);
5353 Instruction *Cloned = Instr->clone();
5355 Cloned->setName(Instr->getName() + ".cloned");
5356 // Replace the operands of the cloned instructions with extracted scalars.
5357 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5358 Value *Op = Params[op][Part];
5359 Cloned->setOperand(op, Op);
5362 // Place the cloned scalar in the new loop.
5363 Builder.Insert(Cloned);
5365 // If the original scalar returns a value we need to place it in a vector
5366 // so that future users will be able to use it.
5368 VecResults[Part] = Cloned;
5371 if (IfPredicateStore) {
5372 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5373 LoopVectorBody.push_back(NewIfBlock);
5374 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5375 Builder.SetInsertPoint(InsertPt);
5376 ReplaceInstWithInst(IfBlock->getTerminator(),
5377 BranchInst::Create(CondBlock, NewIfBlock, Cmp));
5378 IfBlock = NewIfBlock;
5383 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5384 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5385 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5387 return scalarizeInstruction(Instr, IfPredicateStore);
5390 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5394 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5398 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5399 // When unrolling and the VF is 1, we only need to add a simple scalar.
5400 Type *ITy = Val->getType();
5401 assert(!ITy->isVectorTy() && "Val must be a scalar");
5402 Constant *C = ConstantInt::get(ITy, StartIdx);
5403 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");