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: 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: 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; 148 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 { 165 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 { 172 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'. 185 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 { 194 Block(Value *Val, int Len, int Pos) : Seg(Val, 0, Len), Pos(Pos) {} 195 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; 206 207 int size() const { return Blocks.size(); } 208 Block &operator[](int i) { return Blocks[i]; } 209 210 std::vector<Block> Blocks; 211 212 using iterator = decltype(Blocks)::iterator; 213 iterator begin() { return Blocks.begin(); } 214 iterator end() { return Blocks.end(); } 215 using const_iterator = decltype(Blocks)::const_iterator; 216 const_iterator begin() const { return Blocks.begin(); } 217 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 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 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 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 { 288 289 template <typename T> T *getIfUnordered(T *MaybeT) { 290 return MaybeT && MaybeT->isUnordered() ? MaybeT : nullptr; 291 } 292 template <typename T> T *isCandidate(Instruction *In) { 293 return dyn_cast<T>(In); 294 } 295 template <> LoadInst *isCandidate<LoadInst>(Instruction *In) { 296 return getIfUnordered(dyn_cast<LoadInst>(In)); 297 } 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> 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. 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 329 330 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 } 341 342 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 } 355 356 auto AlignVectors::ByteSpan::shift(int Offset) -> ByteSpan & { 357 for (Block &B : Blocks) 358 B.Pos += Offset; 359 return *this; 360 } 361 362 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 } 368 369 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 } 374 375 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 } 396 397 auto AlignVectors::isHvx(const AddrInfo &AI) const -> bool { 398 return HVC.HST.isTypeForHVX(AI.ValTy); 399 } 400 401 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 } 411 412 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 } 429 430 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 } 437 438 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 } 460 461 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 } 469 470 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 } 480 481 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 } 490 491 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 } 544 545 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 } 613 614 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 } 654 655 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 } 683 684 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 } 897 898 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 924 925 auto HexagonVectorCombine::run() -> bool { 926 if (!HST.useHVXOps()) 927 return false; 928 929 bool Changed = AlignVectors(*this).run(); 930 return Changed; 931 } 932 933 auto HexagonVectorCombine::getIntTy() const -> IntegerType * { 934 return Type::getInt32Ty(F.getContext()); 935 } 936 937 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 } 944 945 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 } 952 953 auto HexagonVectorCombine::getConstInt(int Val) const -> ConstantInt * { 954 return ConstantInt::getSigned(getIntTy(), Val); 955 } 956 957 auto HexagonVectorCombine::isZero(const Value *Val) const -> bool { 958 if (auto *C = dyn_cast<Constant>(Val)) 959 return C->isZeroValue(); 960 return false; 961 } 962 963 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 } 969 970 auto HexagonVectorCombine::isUndef(const Value *Val) const -> bool { 971 return isa<UndefValue>(Val); 972 } 973 974 auto HexagonVectorCombine::getSizeOf(const Value *Val) const -> int { 975 return getSizeOf(Val->getType()); 976 } 977 978 auto HexagonVectorCombine::getSizeOf(const Type *Ty) const -> int { 979 return DL.getTypeStoreSize(const_cast<Type *>(Ty)).getFixedValue(); 980 } 981 982 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 } 989 990 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 } 997 998 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. 1007 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 1033 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 1067 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. 1105 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 1139 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 1159 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. 1191 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. 1206 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 1220 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 1283 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> 1377 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 1443 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 1449 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 1459 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 1480 HexagonVectorCombineLegacy() : FunctionPass(ID) {} 1481 1482 StringRef getPassName() const override { return "Hexagon Vector Combine"; } 1483 1484 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 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) 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