1 //===-- HexagonVectorCombine.cpp ------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 // HexagonVectorCombine is a utility class implementing a variety of functions
9 // that assist in vector-based optimizations.
10 //
11 // AlignVectors: replace unaligned vector loads and stores with aligned ones.
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/Analysis/VectorUtils.h"
26 #include "llvm/CodeGen/TargetPassConfig.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/IntrinsicsHexagon.h"
32 #include "llvm/IR/Metadata.h"
33 #include "llvm/InitializePasses.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/KnownBits.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetMachine.h"
39 
40 #include "HexagonSubtarget.h"
41 #include "HexagonTargetMachine.h"
42 
43 #include <algorithm>
44 #include <deque>
45 #include <map>
46 #include <set>
47 #include <utility>
48 #include <vector>
49 
50 #define DEBUG_TYPE "hexagon-vc"
51 
52 using namespace llvm;
53 
54 namespace {
55 class HexagonVectorCombine {
56 public:
HexagonVectorCombine(Function & F_,AliasAnalysis & AA_,AssumptionCache & AC_,DominatorTree & DT_,TargetLibraryInfo & TLI_,const TargetMachine & TM_)57   HexagonVectorCombine(Function &F_, AliasAnalysis &AA_, AssumptionCache &AC_,
58                        DominatorTree &DT_, TargetLibraryInfo &TLI_,
59                        const TargetMachine &TM_)
60       : F(F_), DL(F.getParent()->getDataLayout()), AA(AA_), AC(AC_), DT(DT_),
61         TLI(TLI_),
62         HST(static_cast<const HexagonSubtarget &>(*TM_.getSubtargetImpl(F))) {}
63 
64   bool run();
65 
66   // Common integer type.
67   IntegerType *getIntTy() const;
68   // Byte type: either scalar (when Length = 0), or vector with given
69   // element count.
70   Type *getByteTy(int ElemCount = 0) const;
71   // Boolean type: either scalar (when Length = 0), or vector with given
72   // element count.
73   Type *getBoolTy(int ElemCount = 0) const;
74   // Create a ConstantInt of type returned by getIntTy with the value Val.
75   ConstantInt *getConstInt(int Val) const;
76   // Get the integer value of V, if it exists.
77   Optional<APInt> getIntValue(const Value *Val) const;
78   // Is V a constant 0, or a vector of 0s?
79   bool isZero(const Value *Val) const;
80   // Is V an undef value?
81   bool isUndef(const Value *Val) const;
82 
83   int getSizeOf(const Value *Val) const;
84   int getSizeOf(const Type *Ty) const;
85   int getTypeAlignment(Type *Ty) const;
86 
87   VectorType *getByteVectorTy(int ScLen) const;
88   Constant *getNullValue(Type *Ty) const;
89   Constant *getFullValue(Type *Ty) const;
90 
91   Value *insertb(IRBuilder<> &Builder, Value *Dest, Value *Src, int Start,
92                  int Length, int Where) const;
93   Value *vlalignb(IRBuilder<> &Builder, Value *Lo, Value *Hi, Value *Amt) const;
94   Value *vralignb(IRBuilder<> &Builder, Value *Lo, Value *Hi, Value *Amt) const;
95   Value *concat(IRBuilder<> &Builder, ArrayRef<Value *> Vecs) const;
96   Value *vresize(IRBuilder<> &Builder, Value *Val, int NewSize,
97                  Value *Pad) const;
98   Value *rescale(IRBuilder<> &Builder, Value *Mask, Type *FromTy,
99                  Type *ToTy) const;
100   Value *vlsb(IRBuilder<> &Builder, Value *Val) const;
101   Value *vbytes(IRBuilder<> &Builder, Value *Val) const;
102 
103   Value *createHvxIntrinsic(IRBuilder<> &Builder, Intrinsic::ID IntID,
104                             Type *RetTy, ArrayRef<Value *> Args) const;
105 
106   Optional<int> calculatePointerDifference(Value *Ptr0, Value *Ptr1) const;
107 
108   template <typename T = std::vector<Instruction *>>
109   bool isSafeToMoveBeforeInBB(const Instruction &In,
110                               BasicBlock::const_iterator To,
111                               const T &Ignore = {}) const;
112 
113   Function &F;
114   const DataLayout &DL;
115   AliasAnalysis &AA;
116   AssumptionCache &AC;
117   DominatorTree &DT;
118   TargetLibraryInfo &TLI;
119   const HexagonSubtarget &HST;
120 
121 private:
122 #ifndef NDEBUG
123   // These two functions are only used for assertions at the moment.
124   bool isByteVecTy(Type *Ty) const;
125   bool isSectorTy(Type *Ty) const;
126 #endif
127   Value *getElementRange(IRBuilder<> &Builder, Value *Lo, Value *Hi, int Start,
128                          int Length) const;
129 };
130 
131 class AlignVectors {
132 public:
AlignVectors(HexagonVectorCombine & HVC_)133   AlignVectors(HexagonVectorCombine &HVC_) : HVC(HVC_) {}
134 
135   bool run();
136 
137 private:
138   using InstList = std::vector<Instruction *>;
139 
140   struct Segment {
141     void *Data;
142     int Start;
143     int Size;
144   };
145 
146   struct AddrInfo {
147     AddrInfo(const AddrInfo &) = default;
AddrInfo__anona9de621f0111::AlignVectors::AddrInfo148     AddrInfo(const HexagonVectorCombine &HVC, Instruction *I, Value *A, Type *T,
149              Align H)
150         : Inst(I), Addr(A), ValTy(T), HaveAlign(H),
151           NeedAlign(HVC.getTypeAlignment(ValTy)) {}
152 
153     // XXX: add Size member?
154     Instruction *Inst;
155     Value *Addr;
156     Type *ValTy;
157     Align HaveAlign;
158     Align NeedAlign;
159     int Offset = 0; // Offset (in bytes) from the first member of the
160                     // containing AddrList.
161   };
162   using AddrList = std::vector<AddrInfo>;
163 
164   struct InstrLess {
operator ()__anona9de621f0111::AlignVectors::InstrLess165     bool operator()(const Instruction *A, const Instruction *B) const {
166       return A->comesBefore(B);
167     }
168   };
169   using DepList = std::set<Instruction *, InstrLess>;
170 
171   struct MoveGroup {
MoveGroup__anona9de621f0111::AlignVectors::MoveGroup172     MoveGroup(const AddrInfo &AI, Instruction *B, bool Hvx, bool Load)
173         : Base(B), Main{AI.Inst}, IsHvx(Hvx), IsLoad(Load) {}
174     Instruction *Base; // Base instruction of the parent address group.
175     InstList Main;     // Main group of instructions.
176     InstList Deps;     // List of dependencies.
177     bool IsHvx;        // Is this group of HVX instructions?
178     bool IsLoad;       // Is this a load group?
179   };
180   using MoveList = std::vector<MoveGroup>;
181 
182   struct ByteSpan {
183     struct Segment {
184       // Segment of a Value: 'Len' bytes starting at byte 'Begin'.
Segment__anona9de621f0111::AlignVectors::ByteSpan::Segment185       Segment(Value *Val, int Begin, int Len)
186           : Val(Val), Start(Begin), Size(Len) {}
187       Segment(const Segment &Seg) = default;
188       Value *Val; // Value representable as a sequence of bytes.
189       int Start;  // First byte of the value that belongs to the segment.
190       int Size;   // Number of bytes in the segment.
191     };
192 
193     struct Block {
Block__anona9de621f0111::AlignVectors::ByteSpan::Block194       Block(Value *Val, int Len, int Pos) : Seg(Val, 0, Len), Pos(Pos) {}
Block__anona9de621f0111::AlignVectors::ByteSpan::Block195       Block(Value *Val, int Off, int Len, int Pos)
196           : Seg(Val, Off, Len), Pos(Pos) {}
197       Block(const Block &Blk) = default;
198       Segment Seg; // Value segment.
199       int Pos;     // Position (offset) of the segment in the Block.
200     };
201 
202     int extent() const;
203     ByteSpan section(int Start, int Length) const;
204     ByteSpan &shift(int Offset);
205     SmallVector<Value *, 8> values() const;
206 
size__anona9de621f0111::AlignVectors::ByteSpan207     int size() const { return Blocks.size(); }
operator []__anona9de621f0111::AlignVectors::ByteSpan208     Block &operator[](int i) { return Blocks[i]; }
209 
210     std::vector<Block> Blocks;
211 
212     using iterator = decltype(Blocks)::iterator;
begin__anona9de621f0111::AlignVectors::ByteSpan213     iterator begin() { return Blocks.begin(); }
end__anona9de621f0111::AlignVectors::ByteSpan214     iterator end() { return Blocks.end(); }
215     using const_iterator = decltype(Blocks)::const_iterator;
begin__anona9de621f0111::AlignVectors::ByteSpan216     const_iterator begin() const { return Blocks.begin(); }
end__anona9de621f0111::AlignVectors::ByteSpan217     const_iterator end() const { return Blocks.end(); }
218   };
219 
220   Align getAlignFromValue(const Value *V) const;
221   Optional<MemoryLocation> getLocation(const Instruction &In) const;
222   Optional<AddrInfo> getAddrInfo(Instruction &In) const;
223   bool isHvx(const AddrInfo &AI) const;
224 
225   Value *getPayload(Value *Val) const;
226   Value *getMask(Value *Val) const;
227   Value *getPassThrough(Value *Val) const;
228 
229   Value *createAdjustedPointer(IRBuilder<> &Builder, Value *Ptr, Type *ValTy,
230                                int Adjust) const;
231   Value *createAlignedPointer(IRBuilder<> &Builder, Value *Ptr, Type *ValTy,
232                               int Alignment) const;
233   Value *createAlignedLoad(IRBuilder<> &Builder, Type *ValTy, Value *Ptr,
234                            int Alignment, Value *Mask, Value *PassThru) const;
235   Value *createAlignedStore(IRBuilder<> &Builder, Value *Val, Value *Ptr,
236                             int Alignment, Value *Mask) const;
237 
238   bool createAddressGroups();
239   MoveList createLoadGroups(const AddrList &Group) const;
240   MoveList createStoreGroups(const AddrList &Group) const;
241   bool move(const MoveGroup &Move) const;
242   bool realignGroup(const MoveGroup &Move) const;
243 
244   friend raw_ostream &operator<<(raw_ostream &OS, const AddrInfo &AI);
245   friend raw_ostream &operator<<(raw_ostream &OS, const MoveGroup &MG);
246   friend raw_ostream &operator<<(raw_ostream &OS, const ByteSpan &BS);
247 
248   std::map<Instruction *, AddrList> AddrGroups;
249   HexagonVectorCombine &HVC;
250 };
251 
252 LLVM_ATTRIBUTE_UNUSED
operator <<(raw_ostream & OS,const AlignVectors::AddrInfo & AI)253 raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::AddrInfo &AI) {
254   OS << "Inst: " << AI.Inst << "  " << *AI.Inst << '\n';
255   OS << "Addr: " << *AI.Addr << '\n';
256   OS << "Type: " << *AI.ValTy << '\n';
257   OS << "HaveAlign: " << AI.HaveAlign.value() << '\n';
258   OS << "NeedAlign: " << AI.NeedAlign.value() << '\n';
259   OS << "Offset: " << AI.Offset;
260   return OS;
261 }
262 
263 LLVM_ATTRIBUTE_UNUSED
operator <<(raw_ostream & OS,const AlignVectors::MoveGroup & MG)264 raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::MoveGroup &MG) {
265   OS << "Main\n";
266   for (Instruction *I : MG.Main)
267     OS << "  " << *I << '\n';
268   OS << "Deps\n";
269   for (Instruction *I : MG.Deps)
270     OS << "  " << *I << '\n';
271   return OS;
272 }
273 
274 LLVM_ATTRIBUTE_UNUSED
operator <<(raw_ostream & OS,const AlignVectors::ByteSpan & BS)275 raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::ByteSpan &BS) {
276   OS << "ByteSpan[size=" << BS.size() << ", extent=" << BS.extent() << '\n';
277   for (const AlignVectors::ByteSpan::Block &B : BS) {
278     OS << "  @" << B.Pos << " [" << B.Seg.Start << ',' << B.Seg.Size << "] "
279        << *B.Seg.Val << '\n';
280   }
281   OS << ']';
282   return OS;
283 }
284 
285 } // namespace
286 
287 namespace {
288 
getIfUnordered(T * MaybeT)289 template <typename T> T *getIfUnordered(T *MaybeT) {
290   return MaybeT && MaybeT->isUnordered() ? MaybeT : nullptr;
291 }
isCandidate(Instruction * In)292 template <typename T> T *isCandidate(Instruction *In) {
293   return dyn_cast<T>(In);
294 }
isCandidate(Instruction * In)295 template <> LoadInst *isCandidate<LoadInst>(Instruction *In) {
296   return getIfUnordered(dyn_cast<LoadInst>(In));
297 }
isCandidate(Instruction * In)298 template <> StoreInst *isCandidate<StoreInst>(Instruction *In) {
299   return getIfUnordered(dyn_cast<StoreInst>(In));
300 }
301 
302 #if !defined(_MSC_VER) || _MSC_VER >= 1926
303 // VS2017 and some versions of VS2019 have trouble compiling this:
304 // error C2976: 'std::map': too few template arguments
305 // VS 2019 16.x is known to work, except for 16.4/16.5 (MSC_VER 1924/1925)
306 template <typename Pred, typename... Ts>
erase_if(std::map<Ts...> & map,Pred p)307 void erase_if(std::map<Ts...> &map, Pred p)
308 #else
309 template <typename Pred, typename T, typename U>
310 void erase_if(std::map<T, U> &map, Pred p)
311 #endif
312 {
313   for (auto i = map.begin(), e = map.end(); i != e;) {
314     if (p(*i))
315       i = map.erase(i);
316     else
317       i = std::next(i);
318   }
319 }
320 
321 // Forward other erase_ifs to the LLVM implementations.
erase_if(T && container,Pred p)322 template <typename Pred, typename T> void erase_if(T &&container, Pred p) {
323   llvm::erase_if(std::forward<T>(container), p);
324 }
325 
326 } // namespace
327 
328 // --- Begin AlignVectors
329 
extent() const330 auto AlignVectors::ByteSpan::extent() const -> int {
331   if (size() == 0)
332     return 0;
333   int Min = Blocks[0].Pos;
334   int Max = Blocks[0].Pos + Blocks[0].Seg.Size;
335   for (int i = 1, e = size(); i != e; ++i) {
336     Min = std::min(Min, Blocks[i].Pos);
337     Max = std::max(Max, Blocks[i].Pos + Blocks[i].Seg.Size);
338   }
339   return Max - Min;
340 }
341 
section(int Start,int Length) const342 auto AlignVectors::ByteSpan::section(int Start, int Length) const -> ByteSpan {
343   ByteSpan Section;
344   for (const ByteSpan::Block &B : Blocks) {
345     int L = std::max(B.Pos, Start);                       // Left end.
346     int R = std::min(B.Pos + B.Seg.Size, Start + Length); // Right end+1.
347     if (L < R) {
348       // How much to chop off the beginning of the segment:
349       int Off = L > B.Pos ? L - B.Pos : 0;
350       Section.Blocks.emplace_back(B.Seg.Val, B.Seg.Start + Off, R - L, L);
351     }
352   }
353   return Section;
354 }
355 
shift(int Offset)356 auto AlignVectors::ByteSpan::shift(int Offset) -> ByteSpan & {
357   for (Block &B : Blocks)
358     B.Pos += Offset;
359   return *this;
360 }
361 
values() const362 auto AlignVectors::ByteSpan::values() const -> SmallVector<Value *, 8> {
363   SmallVector<Value *, 8> Values(Blocks.size());
364   for (int i = 0, e = Blocks.size(); i != e; ++i)
365     Values[i] = Blocks[i].Seg.Val;
366   return Values;
367 }
368 
getAlignFromValue(const Value * V) const369 auto AlignVectors::getAlignFromValue(const Value *V) const -> Align {
370   const auto *C = dyn_cast<ConstantInt>(V);
371   assert(C && "Alignment must be a compile-time constant integer");
372   return C->getAlignValue();
373 }
374 
getAddrInfo(Instruction & In) const375 auto AlignVectors::getAddrInfo(Instruction &In) const -> Optional<AddrInfo> {
376   if (auto *L = isCandidate<LoadInst>(&In))
377     return AddrInfo(HVC, L, L->getPointerOperand(), L->getType(),
378                     L->getAlign());
379   if (auto *S = isCandidate<StoreInst>(&In))
380     return AddrInfo(HVC, S, S->getPointerOperand(),
381                     S->getValueOperand()->getType(), S->getAlign());
382   if (auto *II = isCandidate<IntrinsicInst>(&In)) {
383     Intrinsic::ID ID = II->getIntrinsicID();
384     switch (ID) {
385     case Intrinsic::masked_load:
386       return AddrInfo(HVC, II, II->getArgOperand(0), II->getType(),
387                       getAlignFromValue(II->getArgOperand(1)));
388     case Intrinsic::masked_store:
389       return AddrInfo(HVC, II, II->getArgOperand(1),
390                       II->getArgOperand(0)->getType(),
391                       getAlignFromValue(II->getArgOperand(2)));
392     }
393   }
394   return Optional<AddrInfo>();
395 }
396 
isHvx(const AddrInfo & AI) const397 auto AlignVectors::isHvx(const AddrInfo &AI) const -> bool {
398   return HVC.HST.isTypeForHVX(AI.ValTy);
399 }
400 
getPayload(Value * Val) const401 auto AlignVectors::getPayload(Value *Val) const -> Value * {
402   if (auto *In = dyn_cast<Instruction>(Val)) {
403     Intrinsic::ID ID = 0;
404     if (auto *II = dyn_cast<IntrinsicInst>(In))
405       ID = II->getIntrinsicID();
406     if (isa<StoreInst>(In) || ID == Intrinsic::masked_store)
407       return In->getOperand(0);
408   }
409   return Val;
410 }
411 
getMask(Value * Val) const412 auto AlignVectors::getMask(Value *Val) const -> Value * {
413   if (auto *II = dyn_cast<IntrinsicInst>(Val)) {
414     switch (II->getIntrinsicID()) {
415     case Intrinsic::masked_load:
416       return II->getArgOperand(2);
417     case Intrinsic::masked_store:
418       return II->getArgOperand(3);
419     }
420   }
421 
422   Type *ValTy = getPayload(Val)->getType();
423   if (auto *VecTy = dyn_cast<VectorType>(ValTy)) {
424     int ElemCount = VecTy->getElementCount().getFixedValue();
425     return HVC.getFullValue(HVC.getBoolTy(ElemCount));
426   }
427   return HVC.getFullValue(HVC.getBoolTy());
428 }
429 
getPassThrough(Value * Val) const430 auto AlignVectors::getPassThrough(Value *Val) const -> Value * {
431   if (auto *II = dyn_cast<IntrinsicInst>(Val)) {
432     if (II->getIntrinsicID() == Intrinsic::masked_load)
433       return II->getArgOperand(3);
434   }
435   return UndefValue::get(getPayload(Val)->getType());
436 }
437 
createAdjustedPointer(IRBuilder<> & Builder,Value * Ptr,Type * ValTy,int Adjust) const438 auto AlignVectors::createAdjustedPointer(IRBuilder<> &Builder, Value *Ptr,
439                                          Type *ValTy, int Adjust) const
440     -> Value * {
441   // The adjustment is in bytes, but if it's a multiple of the type size,
442   // we don't need to do pointer casts.
443   auto *PtrTy = cast<PointerType>(Ptr->getType());
444   if (!PtrTy->isOpaque()) {
445     Type *ElemTy = PtrTy->getElementType();
446     int ElemSize = HVC.getSizeOf(ElemTy);
447     if (Adjust % ElemSize == 0) {
448       Value *Tmp0 =
449           Builder.CreateGEP(ElemTy, Ptr, HVC.getConstInt(Adjust / ElemSize));
450       return Builder.CreatePointerCast(Tmp0, ValTy->getPointerTo());
451     }
452   }
453 
454   PointerType *CharPtrTy = Type::getInt8PtrTy(HVC.F.getContext());
455   Value *Tmp0 = Builder.CreatePointerCast(Ptr, CharPtrTy);
456   Value *Tmp1 = Builder.CreateGEP(Type::getInt8Ty(HVC.F.getContext()), Tmp0,
457                                   HVC.getConstInt(Adjust));
458   return Builder.CreatePointerCast(Tmp1, ValTy->getPointerTo());
459 }
460 
createAlignedPointer(IRBuilder<> & Builder,Value * Ptr,Type * ValTy,int Alignment) const461 auto AlignVectors::createAlignedPointer(IRBuilder<> &Builder, Value *Ptr,
462                                         Type *ValTy, int Alignment) const
463     -> Value * {
464   Value *AsInt = Builder.CreatePtrToInt(Ptr, HVC.getIntTy());
465   Value *Mask = HVC.getConstInt(-Alignment);
466   Value *And = Builder.CreateAnd(AsInt, Mask);
467   return Builder.CreateIntToPtr(And, ValTy->getPointerTo());
468 }
469 
createAlignedLoad(IRBuilder<> & Builder,Type * ValTy,Value * Ptr,int Alignment,Value * Mask,Value * PassThru) const470 auto AlignVectors::createAlignedLoad(IRBuilder<> &Builder, Type *ValTy,
471                                      Value *Ptr, int Alignment, Value *Mask,
472                                      Value *PassThru) const -> Value * {
473   assert(!HVC.isUndef(Mask)); // Should this be allowed?
474   if (HVC.isZero(Mask))
475     return PassThru;
476   if (Mask == ConstantInt::getTrue(Mask->getType()))
477     return Builder.CreateAlignedLoad(ValTy, Ptr, Align(Alignment));
478   return Builder.CreateMaskedLoad(ValTy, Ptr, Align(Alignment), Mask, PassThru);
479 }
480 
createAlignedStore(IRBuilder<> & Builder,Value * Val,Value * Ptr,int Alignment,Value * Mask) const481 auto AlignVectors::createAlignedStore(IRBuilder<> &Builder, Value *Val,
482                                       Value *Ptr, int Alignment,
483                                       Value *Mask) const -> Value * {
484   if (HVC.isZero(Mask) || HVC.isUndef(Val) || HVC.isUndef(Mask))
485     return UndefValue::get(Val->getType());
486   if (Mask == ConstantInt::getTrue(Mask->getType()))
487     return Builder.CreateAlignedStore(Val, Ptr, Align(Alignment));
488   return Builder.CreateMaskedStore(Val, Ptr, Align(Alignment), Mask);
489 }
490 
createAddressGroups()491 auto AlignVectors::createAddressGroups() -> bool {
492   // An address group created here may contain instructions spanning
493   // multiple basic blocks.
494   AddrList WorkStack;
495 
496   auto findBaseAndOffset = [&](AddrInfo &AI) -> std::pair<Instruction *, int> {
497     for (AddrInfo &W : WorkStack) {
498       if (auto D = HVC.calculatePointerDifference(AI.Addr, W.Addr))
499         return std::make_pair(W.Inst, *D);
500     }
501     return std::make_pair(nullptr, 0);
502   };
503 
504   auto traverseBlock = [&](DomTreeNode *DomN, auto Visit) -> void {
505     BasicBlock &Block = *DomN->getBlock();
506     for (Instruction &I : Block) {
507       auto AI = this->getAddrInfo(I); // Use this-> for gcc6.
508       if (!AI)
509         continue;
510       auto F = findBaseAndOffset(*AI);
511       Instruction *GroupInst;
512       if (Instruction *BI = F.first) {
513         AI->Offset = F.second;
514         GroupInst = BI;
515       } else {
516         WorkStack.push_back(*AI);
517         GroupInst = AI->Inst;
518       }
519       AddrGroups[GroupInst].push_back(*AI);
520     }
521 
522     for (DomTreeNode *C : DomN->children())
523       Visit(C, Visit);
524 
525     while (!WorkStack.empty() && WorkStack.back().Inst->getParent() == &Block)
526       WorkStack.pop_back();
527   };
528 
529   traverseBlock(HVC.DT.getRootNode(), traverseBlock);
530   assert(WorkStack.empty());
531 
532   // AddrGroups are formed.
533 
534   // Remove groups of size 1.
535   erase_if(AddrGroups, [](auto &G) { return G.second.size() == 1; });
536   // Remove groups that don't use HVX types.
537   erase_if(AddrGroups, [&](auto &G) {
538     return !llvm::any_of(
539         G.second, [&](auto &I) { return HVC.HST.isTypeForHVX(I.ValTy); });
540   });
541 
542   return !AddrGroups.empty();
543 }
544 
createLoadGroups(const AddrList & Group) const545 auto AlignVectors::createLoadGroups(const AddrList &Group) const -> MoveList {
546   // Form load groups.
547   // To avoid complications with moving code across basic blocks, only form
548   // groups that are contained within a single basic block.
549 
550   auto getUpwardDeps = [](Instruction *In, Instruction *Base) {
551     BasicBlock *Parent = Base->getParent();
552     assert(In->getParent() == Parent &&
553            "Base and In should be in the same block");
554     assert(Base->comesBefore(In) && "Base should come before In");
555 
556     DepList Deps;
557     std::deque<Instruction *> WorkQ = {In};
558     while (!WorkQ.empty()) {
559       Instruction *D = WorkQ.front();
560       WorkQ.pop_front();
561       Deps.insert(D);
562       for (Value *Op : D->operands()) {
563         if (auto *I = dyn_cast<Instruction>(Op)) {
564           if (I->getParent() == Parent && Base->comesBefore(I))
565             WorkQ.push_back(I);
566         }
567       }
568     }
569     return Deps;
570   };
571 
572   auto tryAddTo = [&](const AddrInfo &Info, MoveGroup &Move) {
573     assert(!Move.Main.empty() && "Move group should have non-empty Main");
574     // Don't mix HVX and non-HVX instructions.
575     if (Move.IsHvx != isHvx(Info))
576       return false;
577     // Leading instruction in the load group.
578     Instruction *Base = Move.Main.front();
579     if (Base->getParent() != Info.Inst->getParent())
580       return false;
581 
582     auto isSafeToMoveToBase = [&](const Instruction *I) {
583       return HVC.isSafeToMoveBeforeInBB(*I, Base->getIterator());
584     };
585     DepList Deps = getUpwardDeps(Info.Inst, Base);
586     if (!llvm::all_of(Deps, isSafeToMoveToBase))
587       return false;
588 
589     // The dependencies will be moved together with the load, so make sure
590     // that none of them could be moved independently in another group.
591     Deps.erase(Info.Inst);
592     auto inAddrMap = [&](Instruction *I) { return AddrGroups.count(I) > 0; };
593     if (llvm::any_of(Deps, inAddrMap))
594       return false;
595     Move.Main.push_back(Info.Inst);
596     llvm::append_range(Move.Deps, Deps);
597     return true;
598   };
599 
600   MoveList LoadGroups;
601 
602   for (const AddrInfo &Info : Group) {
603     if (!Info.Inst->mayReadFromMemory())
604       continue;
605     if (LoadGroups.empty() || !tryAddTo(Info, LoadGroups.back()))
606       LoadGroups.emplace_back(Info, Group.front().Inst, isHvx(Info), true);
607   }
608 
609   // Erase singleton groups.
610   erase_if(LoadGroups, [](const MoveGroup &G) { return G.Main.size() <= 1; });
611   return LoadGroups;
612 }
613 
createStoreGroups(const AddrList & Group) const614 auto AlignVectors::createStoreGroups(const AddrList &Group) const -> MoveList {
615   // Form store groups.
616   // To avoid complications with moving code across basic blocks, only form
617   // groups that are contained within a single basic block.
618 
619   auto tryAddTo = [&](const AddrInfo &Info, MoveGroup &Move) {
620     assert(!Move.Main.empty() && "Move group should have non-empty Main");
621     // For stores with return values we'd have to collect downward depenencies.
622     // There are no such stores that we handle at the moment, so omit that.
623     assert(Info.Inst->getType()->isVoidTy() &&
624            "Not handling stores with return values");
625     // Don't mix HVX and non-HVX instructions.
626     if (Move.IsHvx != isHvx(Info))
627       return false;
628     // For stores we need to be careful whether it's safe to move them.
629     // Stores that are otherwise safe to move together may not appear safe
630     // to move over one another (i.e. isSafeToMoveBefore may return false).
631     Instruction *Base = Move.Main.front();
632     if (Base->getParent() != Info.Inst->getParent())
633       return false;
634     if (!HVC.isSafeToMoveBeforeInBB(*Info.Inst, Base->getIterator(), Move.Main))
635       return false;
636     Move.Main.push_back(Info.Inst);
637     return true;
638   };
639 
640   MoveList StoreGroups;
641 
642   for (auto I = Group.rbegin(), E = Group.rend(); I != E; ++I) {
643     const AddrInfo &Info = *I;
644     if (!Info.Inst->mayWriteToMemory())
645       continue;
646     if (StoreGroups.empty() || !tryAddTo(Info, StoreGroups.back()))
647       StoreGroups.emplace_back(Info, Group.front().Inst, isHvx(Info), false);
648   }
649 
650   // Erase singleton groups.
651   erase_if(StoreGroups, [](const MoveGroup &G) { return G.Main.size() <= 1; });
652   return StoreGroups;
653 }
654 
move(const MoveGroup & Move) const655 auto AlignVectors::move(const MoveGroup &Move) const -> bool {
656   assert(!Move.Main.empty() && "Move group should have non-empty Main");
657   Instruction *Where = Move.Main.front();
658 
659   if (Move.IsLoad) {
660     // Move all deps to before Where, keeping order.
661     for (Instruction *D : Move.Deps)
662       D->moveBefore(Where);
663     // Move all main instructions to after Where, keeping order.
664     ArrayRef<Instruction *> Main(Move.Main);
665     for (Instruction *M : Main.drop_front(1)) {
666       M->moveAfter(Where);
667       Where = M;
668     }
669   } else {
670     // NOTE: Deps are empty for "store" groups. If they need to be
671     // non-empty, decide on the order.
672     assert(Move.Deps.empty());
673     // Move all main instructions to before Where, inverting order.
674     ArrayRef<Instruction *> Main(Move.Main);
675     for (Instruction *M : Main.drop_front(1)) {
676       M->moveBefore(Where);
677       Where = M;
678     }
679   }
680 
681   return Move.Main.size() + Move.Deps.size() > 1;
682 }
683 
realignGroup(const MoveGroup & Move) const684 auto AlignVectors::realignGroup(const MoveGroup &Move) const -> bool {
685   // TODO: Needs support for masked loads/stores of "scalar" vectors.
686   if (!Move.IsHvx)
687     return false;
688 
689   // Return the element with the maximum alignment from Range,
690   // where GetValue obtains the value to compare from an element.
691   auto getMaxOf = [](auto Range, auto GetValue) {
692     return *std::max_element(
693         Range.begin(), Range.end(),
694         [&GetValue](auto &A, auto &B) { return GetValue(A) < GetValue(B); });
695   };
696 
697   const AddrList &BaseInfos = AddrGroups.at(Move.Base);
698 
699   // Conceptually, there is a vector of N bytes covering the addresses
700   // starting from the minimum offset (i.e. Base.Addr+Start). This vector
701   // represents a contiguous memory region that spans all accessed memory
702   // locations.
703   // The correspondence between loaded or stored values will be expressed
704   // in terms of this vector. For example, the 0th element of the vector
705   // from the Base address info will start at byte Start from the beginning
706   // of this conceptual vector.
707   //
708   // This vector will be loaded/stored starting at the nearest down-aligned
709   // address and the amount od the down-alignment will be AlignVal:
710   //   valign(load_vector(align_down(Base+Start)), AlignVal)
711 
712   std::set<Instruction *> TestSet(Move.Main.begin(), Move.Main.end());
713   AddrList MoveInfos;
714   llvm::copy_if(
715       BaseInfos, std::back_inserter(MoveInfos),
716       [&TestSet](const AddrInfo &AI) { return TestSet.count(AI.Inst); });
717 
718   // Maximum alignment present in the whole address group.
719   const AddrInfo &WithMaxAlign =
720       getMaxOf(BaseInfos, [](const AddrInfo &AI) { return AI.HaveAlign; });
721   Align MaxGiven = WithMaxAlign.HaveAlign;
722 
723   // Minimum alignment present in the move address group.
724   const AddrInfo &WithMinOffset =
725       getMaxOf(MoveInfos, [](const AddrInfo &AI) { return -AI.Offset; });
726 
727   const AddrInfo &WithMaxNeeded =
728       getMaxOf(MoveInfos, [](const AddrInfo &AI) { return AI.NeedAlign; });
729   Align MinNeeded = WithMaxNeeded.NeedAlign;
730 
731   // Set the builder at the top instruction in the move group.
732   Instruction *TopIn = Move.IsLoad ? Move.Main.front() : Move.Main.back();
733   IRBuilder<> Builder(TopIn);
734   Value *AlignAddr = nullptr; // Actual aligned address.
735   Value *AlignVal = nullptr;  // Right-shift amount (for valign).
736 
737   if (MinNeeded <= MaxGiven) {
738     int Start = WithMinOffset.Offset;
739     int OffAtMax = WithMaxAlign.Offset;
740     // Shift the offset of the maximally aligned instruction (OffAtMax)
741     // back by just enough multiples of the required alignment to cover the
742     // distance from Start to OffAtMax.
743     // Calculate the address adjustment amount based on the address with the
744     // maximum alignment. This is to allow a simple gep instruction instead
745     // of potential bitcasts to i8*.
746     int Adjust = -alignTo(OffAtMax - Start, MinNeeded.value());
747     AlignAddr = createAdjustedPointer(Builder, WithMaxAlign.Addr,
748                                       WithMaxAlign.ValTy, Adjust);
749     int Diff = Start - (OffAtMax + Adjust);
750     AlignVal = HVC.getConstInt(Diff);
751     // Sanity.
752     assert(Diff >= 0);
753     assert(static_cast<decltype(MinNeeded.value())>(Diff) < MinNeeded.value());
754   } else {
755     // WithMinOffset is the lowest address in the group,
756     //   WithMinOffset.Addr = Base+Start.
757     // Align instructions for both HVX (V6_valign) and scalar (S2_valignrb)
758     // mask off unnecessary bits, so it's ok to just the original pointer as
759     // the alignment amount.
760     // Do an explicit down-alignment of the address to avoid creating an
761     // aligned instruction with an address that is not really aligned.
762     AlignAddr = createAlignedPointer(Builder, WithMinOffset.Addr,
763                                      WithMinOffset.ValTy, MinNeeded.value());
764     AlignVal = Builder.CreatePtrToInt(WithMinOffset.Addr, HVC.getIntTy());
765   }
766 
767   ByteSpan VSpan;
768   for (const AddrInfo &AI : MoveInfos) {
769     VSpan.Blocks.emplace_back(AI.Inst, HVC.getSizeOf(AI.ValTy),
770                               AI.Offset - WithMinOffset.Offset);
771   }
772 
773   // The aligned loads/stores will use blocks that are either scalars,
774   // or HVX vectors. Let "sector" be the unified term for such a block.
775   // blend(scalar, vector) -> sector...
776   int ScLen = Move.IsHvx ? HVC.HST.getVectorLength()
777                          : std::max<int>(MinNeeded.value(), 4);
778   assert(!Move.IsHvx || ScLen == 64 || ScLen == 128);
779   assert(Move.IsHvx || ScLen == 4 || ScLen == 8);
780 
781   Type *SecTy = HVC.getByteTy(ScLen);
782   int NumSectors = (VSpan.extent() + ScLen - 1) / ScLen;
783   bool DoAlign = !HVC.isZero(AlignVal);
784 
785   if (Move.IsLoad) {
786     ByteSpan ASpan;
787     auto *True = HVC.getFullValue(HVC.getBoolTy(ScLen));
788     auto *Undef = UndefValue::get(SecTy);
789 
790     for (int i = 0; i != NumSectors + DoAlign; ++i) {
791       Value *Ptr = createAdjustedPointer(Builder, AlignAddr, SecTy, i * ScLen);
792       // FIXME: generate a predicated load?
793       Value *Load = createAlignedLoad(Builder, SecTy, Ptr, ScLen, True, Undef);
794       // If vector shifting is potentially needed, accumulate metadata
795       // from source sections of twice the load width.
796       int Start = (i - DoAlign) * ScLen;
797       int Width = (1 + DoAlign) * ScLen;
798       propagateMetadata(cast<Instruction>(Load),
799                         VSpan.section(Start, Width).values());
800       ASpan.Blocks.emplace_back(Load, ScLen, i * ScLen);
801     }
802 
803     if (DoAlign) {
804       for (int j = 0; j != NumSectors; ++j) {
805         ASpan[j].Seg.Val = HVC.vralignb(Builder, ASpan[j].Seg.Val,
806                                         ASpan[j + 1].Seg.Val, AlignVal);
807       }
808     }
809 
810     for (ByteSpan::Block &B : VSpan) {
811       ByteSpan ASection = ASpan.section(B.Pos, B.Seg.Size).shift(-B.Pos);
812       Value *Accum = UndefValue::get(HVC.getByteTy(B.Seg.Size));
813       for (ByteSpan::Block &S : ASection) {
814         Value *Pay = HVC.vbytes(Builder, getPayload(S.Seg.Val));
815         Accum =
816             HVC.insertb(Builder, Accum, Pay, S.Seg.Start, S.Seg.Size, S.Pos);
817       }
818       // Instead of casting everything to bytes for the vselect, cast to the
819       // original value type. This will avoid complications with casting masks.
820       // For example, in cases when the original mask applied to i32, it could
821       // be converted to a mask applicable to i8 via pred_typecast intrinsic,
822       // but if the mask is not exactly of HVX length, extra handling would be
823       // needed to make it work.
824       Type *ValTy = getPayload(B.Seg.Val)->getType();
825       Value *Cast = Builder.CreateBitCast(Accum, ValTy);
826       Value *Sel = Builder.CreateSelect(getMask(B.Seg.Val), Cast,
827                                         getPassThrough(B.Seg.Val));
828       B.Seg.Val->replaceAllUsesWith(Sel);
829     }
830   } else {
831     // Stores.
832     ByteSpan ASpanV, ASpanM;
833 
834     // Return a vector value corresponding to the input value Val:
835     // either <1 x Val> for scalar Val, or Val itself for vector Val.
836     auto MakeVec = [](IRBuilder<> &Builder, Value *Val) -> Value * {
837       Type *Ty = Val->getType();
838       if (Ty->isVectorTy())
839         return Val;
840       auto *VecTy = VectorType::get(Ty, 1, /*Scalable*/ false);
841       return Builder.CreateBitCast(Val, VecTy);
842     };
843 
844     // Create an extra "undef" sector at the beginning and at the end.
845     // They will be used as the left/right filler in the vlalign step.
846     for (int i = (DoAlign ? -1 : 0); i != NumSectors + DoAlign; ++i) {
847       // For stores, the size of each section is an aligned vector length.
848       // Adjust the store offsets relative to the section start offset.
849       ByteSpan VSection = VSpan.section(i * ScLen, ScLen).shift(-i * ScLen);
850       Value *AccumV = UndefValue::get(SecTy);
851       Value *AccumM = HVC.getNullValue(SecTy);
852       for (ByteSpan::Block &S : VSection) {
853         Value *Pay = getPayload(S.Seg.Val);
854         Value *Mask = HVC.rescale(Builder, MakeVec(Builder, getMask(S.Seg.Val)),
855                                   Pay->getType(), HVC.getByteTy());
856         AccumM = HVC.insertb(Builder, AccumM, HVC.vbytes(Builder, Mask),
857                              S.Seg.Start, S.Seg.Size, S.Pos);
858         AccumV = HVC.insertb(Builder, AccumV, HVC.vbytes(Builder, Pay),
859                              S.Seg.Start, S.Seg.Size, S.Pos);
860       }
861       ASpanV.Blocks.emplace_back(AccumV, ScLen, i * ScLen);
862       ASpanM.Blocks.emplace_back(AccumM, ScLen, i * ScLen);
863     }
864 
865     // vlalign
866     if (DoAlign) {
867       for (int j = 1; j != NumSectors + 2; ++j) {
868         ASpanV[j - 1].Seg.Val = HVC.vlalignb(Builder, ASpanV[j - 1].Seg.Val,
869                                              ASpanV[j].Seg.Val, AlignVal);
870         ASpanM[j - 1].Seg.Val = HVC.vlalignb(Builder, ASpanM[j - 1].Seg.Val,
871                                              ASpanM[j].Seg.Val, AlignVal);
872       }
873     }
874 
875     for (int i = 0; i != NumSectors + DoAlign; ++i) {
876       Value *Ptr = createAdjustedPointer(Builder, AlignAddr, SecTy, i * ScLen);
877       Value *Val = ASpanV[i].Seg.Val;
878       Value *Mask = ASpanM[i].Seg.Val; // bytes
879       if (!HVC.isUndef(Val) && !HVC.isZero(Mask)) {
880         Value *Store = createAlignedStore(Builder, Val, Ptr, ScLen,
881                                           HVC.vlsb(Builder, Mask));
882         // If vector shifting is potentially needed, accumulate metadata
883         // from source sections of twice the store width.
884         int Start = (i - DoAlign) * ScLen;
885         int Width = (1 + DoAlign) * ScLen;
886         propagateMetadata(cast<Instruction>(Store),
887                           VSpan.section(Start, Width).values());
888       }
889     }
890   }
891 
892   for (auto *Inst : Move.Main)
893     Inst->eraseFromParent();
894 
895   return true;
896 }
897 
run()898 auto AlignVectors::run() -> bool {
899   if (!createAddressGroups())
900     return false;
901 
902   bool Changed = false;
903   MoveList LoadGroups, StoreGroups;
904 
905   for (auto &G : AddrGroups) {
906     llvm::append_range(LoadGroups, createLoadGroups(G.second));
907     llvm::append_range(StoreGroups, createStoreGroups(G.second));
908   }
909 
910   for (auto &M : LoadGroups)
911     Changed |= move(M);
912   for (auto &M : StoreGroups)
913     Changed |= move(M);
914 
915   for (auto &M : LoadGroups)
916     Changed |= realignGroup(M);
917   for (auto &M : StoreGroups)
918     Changed |= realignGroup(M);
919 
920   return Changed;
921 }
922 
923 // --- End AlignVectors
924 
run()925 auto HexagonVectorCombine::run() -> bool {
926   if (!HST.useHVXOps())
927     return false;
928 
929   bool Changed = AlignVectors(*this).run();
930   return Changed;
931 }
932 
getIntTy() const933 auto HexagonVectorCombine::getIntTy() const -> IntegerType * {
934   return Type::getInt32Ty(F.getContext());
935 }
936 
getByteTy(int ElemCount) const937 auto HexagonVectorCombine::getByteTy(int ElemCount) const -> Type * {
938   assert(ElemCount >= 0);
939   IntegerType *ByteTy = Type::getInt8Ty(F.getContext());
940   if (ElemCount == 0)
941     return ByteTy;
942   return VectorType::get(ByteTy, ElemCount, /*Scalable*/ false);
943 }
944 
getBoolTy(int ElemCount) const945 auto HexagonVectorCombine::getBoolTy(int ElemCount) const -> Type * {
946   assert(ElemCount >= 0);
947   IntegerType *BoolTy = Type::getInt1Ty(F.getContext());
948   if (ElemCount == 0)
949     return BoolTy;
950   return VectorType::get(BoolTy, ElemCount, /*Scalable*/ false);
951 }
952 
getConstInt(int Val) const953 auto HexagonVectorCombine::getConstInt(int Val) const -> ConstantInt * {
954   return ConstantInt::getSigned(getIntTy(), Val);
955 }
956 
isZero(const Value * Val) const957 auto HexagonVectorCombine::isZero(const Value *Val) const -> bool {
958   if (auto *C = dyn_cast<Constant>(Val))
959     return C->isZeroValue();
960   return false;
961 }
962 
getIntValue(const Value * Val) const963 auto HexagonVectorCombine::getIntValue(const Value *Val) const
964     -> Optional<APInt> {
965   if (auto *CI = dyn_cast<ConstantInt>(Val))
966     return CI->getValue();
967   return None;
968 }
969 
isUndef(const Value * Val) const970 auto HexagonVectorCombine::isUndef(const Value *Val) const -> bool {
971   return isa<UndefValue>(Val);
972 }
973 
getSizeOf(const Value * Val) const974 auto HexagonVectorCombine::getSizeOf(const Value *Val) const -> int {
975   return getSizeOf(Val->getType());
976 }
977 
getSizeOf(const Type * Ty) const978 auto HexagonVectorCombine::getSizeOf(const Type *Ty) const -> int {
979   return DL.getTypeStoreSize(const_cast<Type *>(Ty)).getFixedValue();
980 }
981 
getTypeAlignment(Type * Ty) const982 auto HexagonVectorCombine::getTypeAlignment(Type *Ty) const -> int {
983   // The actual type may be shorter than the HVX vector, so determine
984   // the alignment based on subtarget info.
985   if (HST.isTypeForHVX(Ty))
986     return HST.getVectorLength();
987   return DL.getABITypeAlign(Ty).value();
988 }
989 
getNullValue(Type * Ty) const990 auto HexagonVectorCombine::getNullValue(Type *Ty) const -> Constant * {
991   assert(Ty->isIntOrIntVectorTy());
992   auto Zero = ConstantInt::get(Ty->getScalarType(), 0);
993   if (auto *VecTy = dyn_cast<VectorType>(Ty))
994     return ConstantVector::getSplat(VecTy->getElementCount(), Zero);
995   return Zero;
996 }
997 
getFullValue(Type * Ty) const998 auto HexagonVectorCombine::getFullValue(Type *Ty) const -> Constant * {
999   assert(Ty->isIntOrIntVectorTy());
1000   auto Minus1 = ConstantInt::get(Ty->getScalarType(), -1);
1001   if (auto *VecTy = dyn_cast<VectorType>(Ty))
1002     return ConstantVector::getSplat(VecTy->getElementCount(), Minus1);
1003   return Minus1;
1004 }
1005 
1006 // Insert bytes [Start..Start+Length) of Src into Dst at byte Where.
insertb(IRBuilder<> & Builder,Value * Dst,Value * Src,int Start,int Length,int Where) const1007 auto HexagonVectorCombine::insertb(IRBuilder<> &Builder, Value *Dst, Value *Src,
1008                                    int Start, int Length, int Where) const
1009     -> Value * {
1010   assert(isByteVecTy(Dst->getType()) && isByteVecTy(Src->getType()));
1011   int SrcLen = getSizeOf(Src);
1012   int DstLen = getSizeOf(Dst);
1013   assert(0 <= Start && Start + Length <= SrcLen);
1014   assert(0 <= Where && Where + Length <= DstLen);
1015 
1016   int P2Len = PowerOf2Ceil(SrcLen | DstLen);
1017   auto *Undef = UndefValue::get(getByteTy());
1018   Value *P2Src = vresize(Builder, Src, P2Len, Undef);
1019   Value *P2Dst = vresize(Builder, Dst, P2Len, Undef);
1020 
1021   SmallVector<int, 256> SMask(P2Len);
1022   for (int i = 0; i != P2Len; ++i) {
1023     // If i is in [Where, Where+Length), pick Src[Start+(i-Where)].
1024     // Otherwise, pick Dst[i];
1025     SMask[i] =
1026         (Where <= i && i < Where + Length) ? P2Len + Start + (i - Where) : i;
1027   }
1028 
1029   Value *P2Insert = Builder.CreateShuffleVector(P2Dst, P2Src, SMask);
1030   return vresize(Builder, P2Insert, DstLen, Undef);
1031 }
1032 
vlalignb(IRBuilder<> & Builder,Value * Lo,Value * Hi,Value * Amt) const1033 auto HexagonVectorCombine::vlalignb(IRBuilder<> &Builder, Value *Lo, Value *Hi,
1034                                     Value *Amt) const -> Value * {
1035   assert(Lo->getType() == Hi->getType() && "Argument type mismatch");
1036   assert(isSectorTy(Hi->getType()));
1037   if (isZero(Amt))
1038     return Hi;
1039   int VecLen = getSizeOf(Hi);
1040   if (auto IntAmt = getIntValue(Amt))
1041     return getElementRange(Builder, Lo, Hi, VecLen - IntAmt->getSExtValue(),
1042                            VecLen);
1043 
1044   if (HST.isTypeForHVX(Hi->getType())) {
1045     int HwLen = HST.getVectorLength();
1046     assert(VecLen == HwLen && "Expecting an exact HVX type");
1047     Intrinsic::ID V6_vlalignb = HwLen == 64
1048                                     ? Intrinsic::hexagon_V6_vlalignb
1049                                     : Intrinsic::hexagon_V6_vlalignb_128B;
1050     return createHvxIntrinsic(Builder, V6_vlalignb, Hi->getType(),
1051                               {Hi, Lo, Amt});
1052   }
1053 
1054   if (VecLen == 4) {
1055     Value *Pair = concat(Builder, {Lo, Hi});
1056     Value *Shift = Builder.CreateLShr(Builder.CreateShl(Pair, Amt), 32);
1057     Value *Trunc = Builder.CreateTrunc(Shift, Type::getInt32Ty(F.getContext()));
1058     return Builder.CreateBitCast(Trunc, Hi->getType());
1059   }
1060   if (VecLen == 8) {
1061     Value *Sub = Builder.CreateSub(getConstInt(VecLen), Amt);
1062     return vralignb(Builder, Lo, Hi, Sub);
1063   }
1064   llvm_unreachable("Unexpected vector length");
1065 }
1066 
vralignb(IRBuilder<> & Builder,Value * Lo,Value * Hi,Value * Amt) const1067 auto HexagonVectorCombine::vralignb(IRBuilder<> &Builder, Value *Lo, Value *Hi,
1068                                     Value *Amt) const -> Value * {
1069   assert(Lo->getType() == Hi->getType() && "Argument type mismatch");
1070   assert(isSectorTy(Lo->getType()));
1071   if (isZero(Amt))
1072     return Lo;
1073   int VecLen = getSizeOf(Lo);
1074   if (auto IntAmt = getIntValue(Amt))
1075     return getElementRange(Builder, Lo, Hi, IntAmt->getSExtValue(), VecLen);
1076 
1077   if (HST.isTypeForHVX(Lo->getType())) {
1078     int HwLen = HST.getVectorLength();
1079     assert(VecLen == HwLen && "Expecting an exact HVX type");
1080     Intrinsic::ID V6_valignb = HwLen == 64 ? Intrinsic::hexagon_V6_valignb
1081                                            : Intrinsic::hexagon_V6_valignb_128B;
1082     return createHvxIntrinsic(Builder, V6_valignb, Lo->getType(),
1083                               {Hi, Lo, Amt});
1084   }
1085 
1086   if (VecLen == 4) {
1087     Value *Pair = concat(Builder, {Lo, Hi});
1088     Value *Shift = Builder.CreateLShr(Pair, Amt);
1089     Value *Trunc = Builder.CreateTrunc(Shift, Type::getInt32Ty(F.getContext()));
1090     return Builder.CreateBitCast(Trunc, Lo->getType());
1091   }
1092   if (VecLen == 8) {
1093     Type *Int64Ty = Type::getInt64Ty(F.getContext());
1094     Value *Lo64 = Builder.CreateBitCast(Lo, Int64Ty);
1095     Value *Hi64 = Builder.CreateBitCast(Hi, Int64Ty);
1096     Function *FI = Intrinsic::getDeclaration(F.getParent(),
1097                                              Intrinsic::hexagon_S2_valignrb);
1098     Value *Call = Builder.CreateCall(FI, {Hi64, Lo64, Amt});
1099     return Builder.CreateBitCast(Call, Lo->getType());
1100   }
1101   llvm_unreachable("Unexpected vector length");
1102 }
1103 
1104 // Concatenates a sequence of vectors of the same type.
concat(IRBuilder<> & Builder,ArrayRef<Value * > Vecs) const1105 auto HexagonVectorCombine::concat(IRBuilder<> &Builder,
1106                                   ArrayRef<Value *> Vecs) const -> Value * {
1107   assert(!Vecs.empty());
1108   SmallVector<int, 256> SMask;
1109   std::vector<Value *> Work[2];
1110   int ThisW = 0, OtherW = 1;
1111 
1112   Work[ThisW].assign(Vecs.begin(), Vecs.end());
1113   while (Work[ThisW].size() > 1) {
1114     auto *Ty = cast<VectorType>(Work[ThisW].front()->getType());
1115     int ElemCount = Ty->getElementCount().getFixedValue();
1116     SMask.resize(ElemCount * 2);
1117     std::iota(SMask.begin(), SMask.end(), 0);
1118 
1119     Work[OtherW].clear();
1120     if (Work[ThisW].size() % 2 != 0)
1121       Work[ThisW].push_back(UndefValue::get(Ty));
1122     for (int i = 0, e = Work[ThisW].size(); i < e; i += 2) {
1123       Value *Joined = Builder.CreateShuffleVector(Work[ThisW][i],
1124                                                   Work[ThisW][i + 1], SMask);
1125       Work[OtherW].push_back(Joined);
1126     }
1127     std::swap(ThisW, OtherW);
1128   }
1129 
1130   // Since there may have been some undefs appended to make shuffle operands
1131   // have the same type, perform the last shuffle to only pick the original
1132   // elements.
1133   SMask.resize(Vecs.size() * getSizeOf(Vecs.front()->getType()));
1134   std::iota(SMask.begin(), SMask.end(), 0);
1135   Value *Total = Work[OtherW].front();
1136   return Builder.CreateShuffleVector(Total, SMask);
1137 }
1138 
vresize(IRBuilder<> & Builder,Value * Val,int NewSize,Value * Pad) const1139 auto HexagonVectorCombine::vresize(IRBuilder<> &Builder, Value *Val,
1140                                    int NewSize, Value *Pad) const -> Value * {
1141   assert(isa<VectorType>(Val->getType()));
1142   auto *ValTy = cast<VectorType>(Val->getType());
1143   assert(ValTy->getElementType() == Pad->getType());
1144 
1145   int CurSize = ValTy->getElementCount().getFixedValue();
1146   if (CurSize == NewSize)
1147     return Val;
1148   // Truncate?
1149   if (CurSize > NewSize)
1150     return getElementRange(Builder, Val, /*Unused*/ Val, 0, NewSize);
1151   // Extend.
1152   SmallVector<int, 128> SMask(NewSize);
1153   std::iota(SMask.begin(), SMask.begin() + CurSize, 0);
1154   std::fill(SMask.begin() + CurSize, SMask.end(), CurSize);
1155   Value *PadVec = Builder.CreateVectorSplat(CurSize, Pad);
1156   return Builder.CreateShuffleVector(Val, PadVec, SMask);
1157 }
1158 
rescale(IRBuilder<> & Builder,Value * Mask,Type * FromTy,Type * ToTy) const1159 auto HexagonVectorCombine::rescale(IRBuilder<> &Builder, Value *Mask,
1160                                    Type *FromTy, Type *ToTy) const -> Value * {
1161   // Mask is a vector <N x i1>, where each element corresponds to an
1162   // element of FromTy. Remap it so that each element will correspond
1163   // to an element of ToTy.
1164   assert(isa<VectorType>(Mask->getType()));
1165 
1166   Type *FromSTy = FromTy->getScalarType();
1167   Type *ToSTy = ToTy->getScalarType();
1168   if (FromSTy == ToSTy)
1169     return Mask;
1170 
1171   int FromSize = getSizeOf(FromSTy);
1172   int ToSize = getSizeOf(ToSTy);
1173   assert(FromSize % ToSize == 0 || ToSize % FromSize == 0);
1174 
1175   auto *MaskTy = cast<VectorType>(Mask->getType());
1176   int FromCount = MaskTy->getElementCount().getFixedValue();
1177   int ToCount = (FromCount * FromSize) / ToSize;
1178   assert((FromCount * FromSize) % ToSize == 0);
1179 
1180   // Mask <N x i1> -> sext to <N x FromTy> -> bitcast to <M x ToTy> ->
1181   // -> trunc to <M x i1>.
1182   Value *Ext = Builder.CreateSExt(
1183       Mask, VectorType::get(FromSTy, FromCount, /*Scalable*/ false));
1184   Value *Cast = Builder.CreateBitCast(
1185       Ext, VectorType::get(ToSTy, ToCount, /*Scalable*/ false));
1186   return Builder.CreateTrunc(
1187       Cast, VectorType::get(getBoolTy(), ToCount, /*Scalable*/ false));
1188 }
1189 
1190 // Bitcast to bytes, and return least significant bits.
vlsb(IRBuilder<> & Builder,Value * Val) const1191 auto HexagonVectorCombine::vlsb(IRBuilder<> &Builder, Value *Val) const
1192     -> Value * {
1193   Type *ScalarTy = Val->getType()->getScalarType();
1194   if (ScalarTy == getBoolTy())
1195     return Val;
1196 
1197   Value *Bytes = vbytes(Builder, Val);
1198   if (auto *VecTy = dyn_cast<VectorType>(Bytes->getType()))
1199     return Builder.CreateTrunc(Bytes, getBoolTy(getSizeOf(VecTy)));
1200   // If Bytes is a scalar (i.e. Val was a scalar byte), return i1, not
1201   // <1 x i1>.
1202   return Builder.CreateTrunc(Bytes, getBoolTy());
1203 }
1204 
1205 // Bitcast to bytes for non-bool. For bool, convert i1 -> i8.
vbytes(IRBuilder<> & Builder,Value * Val) const1206 auto HexagonVectorCombine::vbytes(IRBuilder<> &Builder, Value *Val) const
1207     -> Value * {
1208   Type *ScalarTy = Val->getType()->getScalarType();
1209   if (ScalarTy == getByteTy())
1210     return Val;
1211 
1212   if (ScalarTy != getBoolTy())
1213     return Builder.CreateBitCast(Val, getByteTy(getSizeOf(Val)));
1214   // For bool, return a sext from i1 to i8.
1215   if (auto *VecTy = dyn_cast<VectorType>(Val->getType()))
1216     return Builder.CreateSExt(Val, VectorType::get(getByteTy(), VecTy));
1217   return Builder.CreateSExt(Val, getByteTy());
1218 }
1219 
createHvxIntrinsic(IRBuilder<> & Builder,Intrinsic::ID IntID,Type * RetTy,ArrayRef<Value * > Args) const1220 auto HexagonVectorCombine::createHvxIntrinsic(IRBuilder<> &Builder,
1221                                               Intrinsic::ID IntID, Type *RetTy,
1222                                               ArrayRef<Value *> Args) const
1223     -> Value * {
1224   int HwLen = HST.getVectorLength();
1225   Type *BoolTy = Type::getInt1Ty(F.getContext());
1226   Type *Int32Ty = Type::getInt32Ty(F.getContext());
1227   // HVX vector -> v16i32/v32i32
1228   // HVX vector predicate -> v512i1/v1024i1
1229   auto getTypeForIntrin = [&](Type *Ty) -> Type * {
1230     if (HST.isTypeForHVX(Ty, /*IncludeBool*/ true)) {
1231       Type *ElemTy = cast<VectorType>(Ty)->getElementType();
1232       if (ElemTy == Int32Ty)
1233         return Ty;
1234       if (ElemTy == BoolTy)
1235         return VectorType::get(BoolTy, 8 * HwLen, /*Scalable*/ false);
1236       return VectorType::get(Int32Ty, HwLen / 4, /*Scalable*/ false);
1237     }
1238     // Non-HVX type. It should be a scalar.
1239     assert(Ty == Int32Ty || Ty->isIntegerTy(64));
1240     return Ty;
1241   };
1242 
1243   auto getCast = [&](IRBuilder<> &Builder, Value *Val,
1244                      Type *DestTy) -> Value * {
1245     Type *SrcTy = Val->getType();
1246     if (SrcTy == DestTy)
1247       return Val;
1248     if (HST.isTypeForHVX(SrcTy, /*IncludeBool*/ true)) {
1249       if (cast<VectorType>(SrcTy)->getElementType() == BoolTy) {
1250         // This should take care of casts the other way too, for example
1251         // v1024i1 -> v32i1.
1252         Intrinsic::ID TC = HwLen == 64
1253                                ? Intrinsic::hexagon_V6_pred_typecast
1254                                : Intrinsic::hexagon_V6_pred_typecast_128B;
1255         Function *FI = Intrinsic::getDeclaration(F.getParent(), TC,
1256                                                  {DestTy, Val->getType()});
1257         return Builder.CreateCall(FI, {Val});
1258       }
1259       // Non-predicate HVX vector.
1260       return Builder.CreateBitCast(Val, DestTy);
1261     }
1262     // Non-HVX type. It should be a scalar, and it should already have
1263     // a valid type.
1264     llvm_unreachable("Unexpected type");
1265   };
1266 
1267   SmallVector<Value *, 4> IntOps;
1268   for (Value *A : Args)
1269     IntOps.push_back(getCast(Builder, A, getTypeForIntrin(A->getType())));
1270   Function *FI = Intrinsic::getDeclaration(F.getParent(), IntID);
1271   Value *Call = Builder.CreateCall(FI, IntOps);
1272 
1273   Type *CallTy = Call->getType();
1274   if (CallTy == RetTy)
1275     return Call;
1276   // Scalar types should have RetTy matching the call return type.
1277   assert(HST.isTypeForHVX(CallTy, /*IncludeBool*/ true));
1278   if (cast<VectorType>(CallTy)->getElementType() == BoolTy)
1279     return getCast(Builder, Call, RetTy);
1280   return Builder.CreateBitCast(Call, RetTy);
1281 }
1282 
calculatePointerDifference(Value * Ptr0,Value * Ptr1) const1283 auto HexagonVectorCombine::calculatePointerDifference(Value *Ptr0,
1284                                                       Value *Ptr1) const
1285     -> Optional<int> {
1286   struct Builder : IRBuilder<> {
1287     Builder(BasicBlock *B) : IRBuilder<>(B) {}
1288     ~Builder() {
1289       for (Instruction *I : llvm::reverse(ToErase))
1290         I->eraseFromParent();
1291     }
1292     SmallVector<Instruction *, 8> ToErase;
1293   };
1294 
1295 #define CallBuilder(B, F)                                                      \
1296   [&](auto &B_) {                                                              \
1297     Value *V = B_.F;                                                           \
1298     if (auto *I = dyn_cast<Instruction>(V))                                    \
1299       B_.ToErase.push_back(I);                                                 \
1300     return V;                                                                  \
1301   }(B)
1302 
1303   auto Simplify = [&](Value *V) {
1304     if (auto *I = dyn_cast<Instruction>(V)) {
1305       SimplifyQuery Q(DL, &TLI, &DT, &AC, I);
1306       if (Value *S = SimplifyInstruction(I, Q))
1307         return S;
1308     }
1309     return V;
1310   };
1311 
1312   auto StripBitCast = [](Value *V) {
1313     while (auto *C = dyn_cast<BitCastInst>(V))
1314       V = C->getOperand(0);
1315     return V;
1316   };
1317 
1318   Ptr0 = StripBitCast(Ptr0);
1319   Ptr1 = StripBitCast(Ptr1);
1320   if (!isa<GetElementPtrInst>(Ptr0) || !isa<GetElementPtrInst>(Ptr1))
1321     return None;
1322 
1323   auto *Gep0 = cast<GetElementPtrInst>(Ptr0);
1324   auto *Gep1 = cast<GetElementPtrInst>(Ptr1);
1325   if (Gep0->getPointerOperand() != Gep1->getPointerOperand())
1326     return None;
1327 
1328   Builder B(Gep0->getParent());
1329   int Scale = DL.getTypeStoreSize(Gep0->getSourceElementType());
1330 
1331   // FIXME: for now only check GEPs with a single index.
1332   if (Gep0->getNumOperands() != 2 || Gep1->getNumOperands() != 2)
1333     return None;
1334 
1335   Value *Idx0 = Gep0->getOperand(1);
1336   Value *Idx1 = Gep1->getOperand(1);
1337 
1338   // First, try to simplify the subtraction directly.
1339   if (auto *Diff = dyn_cast<ConstantInt>(
1340           Simplify(CallBuilder(B, CreateSub(Idx0, Idx1)))))
1341     return Diff->getSExtValue() * Scale;
1342 
1343   KnownBits Known0 = computeKnownBits(Idx0, DL, 0, &AC, Gep0, &DT);
1344   KnownBits Known1 = computeKnownBits(Idx1, DL, 0, &AC, Gep1, &DT);
1345   APInt Unknown = ~(Known0.Zero | Known0.One) | ~(Known1.Zero | Known1.One);
1346   if (Unknown.isAllOnesValue())
1347     return None;
1348 
1349   Value *MaskU = ConstantInt::get(Idx0->getType(), Unknown);
1350   Value *AndU0 = Simplify(CallBuilder(B, CreateAnd(Idx0, MaskU)));
1351   Value *AndU1 = Simplify(CallBuilder(B, CreateAnd(Idx1, MaskU)));
1352   Value *SubU = Simplify(CallBuilder(B, CreateSub(AndU0, AndU1)));
1353   int Diff0 = 0;
1354   if (auto *C = dyn_cast<ConstantInt>(SubU)) {
1355     Diff0 = C->getSExtValue();
1356   } else {
1357     return None;
1358   }
1359 
1360   Value *MaskK = ConstantInt::get(MaskU->getType(), ~Unknown);
1361   Value *AndK0 = Simplify(CallBuilder(B, CreateAnd(Idx0, MaskK)));
1362   Value *AndK1 = Simplify(CallBuilder(B, CreateAnd(Idx1, MaskK)));
1363   Value *SubK = Simplify(CallBuilder(B, CreateSub(AndK0, AndK1)));
1364   int Diff1 = 0;
1365   if (auto *C = dyn_cast<ConstantInt>(SubK)) {
1366     Diff1 = C->getSExtValue();
1367   } else {
1368     return None;
1369   }
1370 
1371   return (Diff0 + Diff1) * Scale;
1372 
1373 #undef CallBuilder
1374 }
1375 
1376 template <typename T>
isSafeToMoveBeforeInBB(const Instruction & In,BasicBlock::const_iterator To,const T & Ignore) const1377 auto HexagonVectorCombine::isSafeToMoveBeforeInBB(const Instruction &In,
1378                                                   BasicBlock::const_iterator To,
1379                                                   const T &Ignore) const
1380     -> bool {
1381   auto getLocOrNone = [this](const Instruction &I) -> Optional<MemoryLocation> {
1382     if (const auto *II = dyn_cast<IntrinsicInst>(&I)) {
1383       switch (II->getIntrinsicID()) {
1384       case Intrinsic::masked_load:
1385         return MemoryLocation::getForArgument(II, 0, TLI);
1386       case Intrinsic::masked_store:
1387         return MemoryLocation::getForArgument(II, 1, TLI);
1388       }
1389     }
1390     return MemoryLocation::getOrNone(&I);
1391   };
1392 
1393   // The source and the destination must be in the same basic block.
1394   const BasicBlock &Block = *In.getParent();
1395   assert(Block.begin() == To || Block.end() == To || To->getParent() == &Block);
1396   // No PHIs.
1397   if (isa<PHINode>(In) || (To != Block.end() && isa<PHINode>(*To)))
1398     return false;
1399 
1400   if (!mayBeMemoryDependent(In))
1401     return true;
1402   bool MayWrite = In.mayWriteToMemory();
1403   auto MaybeLoc = getLocOrNone(In);
1404 
1405   auto From = In.getIterator();
1406   if (From == To)
1407     return true;
1408   bool MoveUp = (To != Block.end() && To->comesBefore(&In));
1409   auto Range =
1410       MoveUp ? std::make_pair(To, From) : std::make_pair(std::next(From), To);
1411   for (auto It = Range.first; It != Range.second; ++It) {
1412     const Instruction &I = *It;
1413     if (llvm::is_contained(Ignore, &I))
1414       continue;
1415     // assume intrinsic can be ignored
1416     if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
1417       if (II->getIntrinsicID() == Intrinsic::assume)
1418         continue;
1419     }
1420     // Parts based on isSafeToMoveBefore from CoveMoverUtils.cpp.
1421     if (I.mayThrow())
1422       return false;
1423     if (auto *CB = dyn_cast<CallBase>(&I)) {
1424       if (!CB->hasFnAttr(Attribute::WillReturn))
1425         return false;
1426       if (!CB->hasFnAttr(Attribute::NoSync))
1427         return false;
1428     }
1429     if (I.mayReadOrWriteMemory()) {
1430       auto MaybeLocI = getLocOrNone(I);
1431       if (MayWrite || I.mayWriteToMemory()) {
1432         if (!MaybeLoc || !MaybeLocI)
1433           return false;
1434         if (!AA.isNoAlias(*MaybeLoc, *MaybeLocI))
1435           return false;
1436       }
1437     }
1438   }
1439   return true;
1440 }
1441 
1442 #ifndef NDEBUG
isByteVecTy(Type * Ty) const1443 auto HexagonVectorCombine::isByteVecTy(Type *Ty) const -> bool {
1444   if (auto *VecTy = dyn_cast<VectorType>(Ty))
1445     return VecTy->getElementType() == getByteTy();
1446   return false;
1447 }
1448 
isSectorTy(Type * Ty) const1449 auto HexagonVectorCombine::isSectorTy(Type *Ty) const -> bool {
1450   if (!isByteVecTy(Ty))
1451     return false;
1452   int Size = getSizeOf(Ty);
1453   if (HST.isTypeForHVX(Ty))
1454     return Size == static_cast<int>(HST.getVectorLength());
1455   return Size == 4 || Size == 8;
1456 }
1457 #endif
1458 
getElementRange(IRBuilder<> & Builder,Value * Lo,Value * Hi,int Start,int Length) const1459 auto HexagonVectorCombine::getElementRange(IRBuilder<> &Builder, Value *Lo,
1460                                            Value *Hi, int Start,
1461                                            int Length) const -> Value * {
1462   assert(0 <= Start && Start < Length);
1463   SmallVector<int, 128> SMask(Length);
1464   std::iota(SMask.begin(), SMask.end(), Start);
1465   return Builder.CreateShuffleVector(Lo, Hi, SMask);
1466 }
1467 
1468 // Pass management.
1469 
1470 namespace llvm {
1471 void initializeHexagonVectorCombineLegacyPass(PassRegistry &);
1472 FunctionPass *createHexagonVectorCombineLegacyPass();
1473 } // namespace llvm
1474 
1475 namespace {
1476 class HexagonVectorCombineLegacy : public FunctionPass {
1477 public:
1478   static char ID;
1479 
HexagonVectorCombineLegacy()1480   HexagonVectorCombineLegacy() : FunctionPass(ID) {}
1481 
getPassName() const1482   StringRef getPassName() const override { return "Hexagon Vector Combine"; }
1483 
getAnalysisUsage(AnalysisUsage & AU) const1484   void getAnalysisUsage(AnalysisUsage &AU) const override {
1485     AU.setPreservesCFG();
1486     AU.addRequired<AAResultsWrapperPass>();
1487     AU.addRequired<AssumptionCacheTracker>();
1488     AU.addRequired<DominatorTreeWrapperPass>();
1489     AU.addRequired<TargetLibraryInfoWrapperPass>();
1490     AU.addRequired<TargetPassConfig>();
1491     FunctionPass::getAnalysisUsage(AU);
1492   }
1493 
runOnFunction(Function & F)1494   bool runOnFunction(Function &F) override {
1495     if (skipFunction(F))
1496       return false;
1497     AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1498     AssumptionCache &AC =
1499         getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1500     DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1501     TargetLibraryInfo &TLI =
1502         getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1503     auto &TM = getAnalysis<TargetPassConfig>().getTM<HexagonTargetMachine>();
1504     HexagonVectorCombine HVC(F, AA, AC, DT, TLI, TM);
1505     return HVC.run();
1506   }
1507 };
1508 } // namespace
1509 
1510 char HexagonVectorCombineLegacy::ID = 0;
1511 
1512 INITIALIZE_PASS_BEGIN(HexagonVectorCombineLegacy, DEBUG_TYPE,
1513                       "Hexagon Vector Combine", false, false)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)1514 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1515 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1516 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1517 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1518 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
1519 INITIALIZE_PASS_END(HexagonVectorCombineLegacy, DEBUG_TYPE,
1520                     "Hexagon Vector Combine", false, false)
1521 
1522 FunctionPass *llvm::createHexagonVectorCombineLegacyPass() {
1523   return new HexagonVectorCombineLegacy();
1524 }
1525