1 //===- MipsFastISel.cpp - Mips FastISel implementation --------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// 10 /// \file 11 /// This file defines the MIPS-specific support for the FastISel class. 12 /// Some of the target-specific code is generated by tablegen in the file 13 /// MipsGenFastISel.inc, which is #included here. 14 /// 15 //===----------------------------------------------------------------------===// 16 17 #include "MCTargetDesc/MipsABIInfo.h" 18 #include "MCTargetDesc/MipsBaseInfo.h" 19 #include "MipsCCState.h" 20 #include "MipsISelLowering.h" 21 #include "MipsInstrInfo.h" 22 #include "MipsMachineFunction.h" 23 #include "MipsSubtarget.h" 24 #include "MipsTargetMachine.h" 25 #include "llvm/ADT/APInt.h" 26 #include "llvm/ADT/ArrayRef.h" 27 #include "llvm/ADT/DenseMap.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/Analysis/TargetLibraryInfo.h" 30 #include "llvm/CodeGen/CallingConvLower.h" 31 #include "llvm/CodeGen/FastISel.h" 32 #include "llvm/CodeGen/FunctionLoweringInfo.h" 33 #include "llvm/CodeGen/ISDOpcodes.h" 34 #include "llvm/CodeGen/MachineBasicBlock.h" 35 #include "llvm/CodeGen/MachineFrameInfo.h" 36 #include "llvm/CodeGen/MachineInstrBuilder.h" 37 #include "llvm/CodeGen/MachineMemOperand.h" 38 #include "llvm/CodeGen/MachineRegisterInfo.h" 39 #include "llvm/CodeGen/TargetInstrInfo.h" 40 #include "llvm/CodeGen/TargetLowering.h" 41 #include "llvm/CodeGen/ValueTypes.h" 42 #include "llvm/IR/Attributes.h" 43 #include "llvm/IR/CallingConv.h" 44 #include "llvm/IR/Constant.h" 45 #include "llvm/IR/Constants.h" 46 #include "llvm/IR/DataLayout.h" 47 #include "llvm/IR/Function.h" 48 #include "llvm/IR/GetElementPtrTypeIterator.h" 49 #include "llvm/IR/GlobalValue.h" 50 #include "llvm/IR/GlobalVariable.h" 51 #include "llvm/IR/InstrTypes.h" 52 #include "llvm/IR/Instruction.h" 53 #include "llvm/IR/Instructions.h" 54 #include "llvm/IR/IntrinsicInst.h" 55 #include "llvm/IR/Operator.h" 56 #include "llvm/IR/Type.h" 57 #include "llvm/IR/User.h" 58 #include "llvm/IR/Value.h" 59 #include "llvm/MC/MCInstrDesc.h" 60 #include "llvm/MC/MCRegisterInfo.h" 61 #include "llvm/MC/MCSymbol.h" 62 #include "llvm/Support/Casting.h" 63 #include "llvm/Support/Compiler.h" 64 #include "llvm/Support/Debug.h" 65 #include "llvm/Support/ErrorHandling.h" 66 #include "llvm/Support/MachineValueType.h" 67 #include "llvm/Support/MathExtras.h" 68 #include "llvm/Support/raw_ostream.h" 69 #include <algorithm> 70 #include <array> 71 #include <cassert> 72 #include <cstdint> 73 74 #define DEBUG_TYPE "mips-fastisel" 75 76 using namespace llvm; 77 78 namespace { 79 80 class MipsFastISel final : public FastISel { 81 82 // All possible address modes. 83 class Address { 84 public: 85 using BaseKind = enum { RegBase, FrameIndexBase }; 86 87 private: 88 BaseKind Kind = RegBase; 89 union { 90 unsigned Reg; 91 int FI; 92 } Base; 93 94 int64_t Offset = 0; 95 96 const GlobalValue *GV = nullptr; 97 98 public: 99 // Innocuous defaults for our address. 100 Address() { Base.Reg = 0; } 101 102 void setKind(BaseKind K) { Kind = K; } 103 BaseKind getKind() const { return Kind; } 104 bool isRegBase() const { return Kind == RegBase; } 105 bool isFIBase() const { return Kind == FrameIndexBase; } 106 107 void setReg(unsigned Reg) { 108 assert(isRegBase() && "Invalid base register access!"); 109 Base.Reg = Reg; 110 } 111 112 unsigned getReg() const { 113 assert(isRegBase() && "Invalid base register access!"); 114 return Base.Reg; 115 } 116 117 void setFI(unsigned FI) { 118 assert(isFIBase() && "Invalid base frame index access!"); 119 Base.FI = FI; 120 } 121 122 unsigned getFI() const { 123 assert(isFIBase() && "Invalid base frame index access!"); 124 return Base.FI; 125 } 126 127 void setOffset(int64_t Offset_) { Offset = Offset_; } 128 int64_t getOffset() const { return Offset; } 129 void setGlobalValue(const GlobalValue *G) { GV = G; } 130 const GlobalValue *getGlobalValue() { return GV; } 131 }; 132 133 /// Subtarget - Keep a pointer to the MipsSubtarget around so that we can 134 /// make the right decision when generating code for different targets. 135 const TargetMachine &TM; 136 const MipsSubtarget *Subtarget; 137 const TargetInstrInfo &TII; 138 const TargetLowering &TLI; 139 MipsFunctionInfo *MFI; 140 141 // Convenience variables to avoid some queries. 142 LLVMContext *Context; 143 144 bool fastLowerArguments() override; 145 bool fastLowerCall(CallLoweringInfo &CLI) override; 146 bool fastLowerIntrinsicCall(const IntrinsicInst *II) override; 147 148 bool UnsupportedFPMode; // To allow fast-isel to proceed and just not handle 149 // floating point but not reject doing fast-isel in other 150 // situations 151 152 private: 153 // Selection routines. 154 bool selectLogicalOp(const Instruction *I); 155 bool selectLoad(const Instruction *I); 156 bool selectStore(const Instruction *I); 157 bool selectBranch(const Instruction *I); 158 bool selectSelect(const Instruction *I); 159 bool selectCmp(const Instruction *I); 160 bool selectFPExt(const Instruction *I); 161 bool selectFPTrunc(const Instruction *I); 162 bool selectFPToInt(const Instruction *I, bool IsSigned); 163 bool selectRet(const Instruction *I); 164 bool selectTrunc(const Instruction *I); 165 bool selectIntExt(const Instruction *I); 166 bool selectShift(const Instruction *I); 167 bool selectDivRem(const Instruction *I, unsigned ISDOpcode); 168 169 // Utility helper routines. 170 bool isTypeLegal(Type *Ty, MVT &VT); 171 bool isTypeSupported(Type *Ty, MVT &VT); 172 bool isLoadTypeLegal(Type *Ty, MVT &VT); 173 bool computeAddress(const Value *Obj, Address &Addr); 174 bool computeCallAddress(const Value *V, Address &Addr); 175 void simplifyAddress(Address &Addr); 176 177 // Emit helper routines. 178 bool emitCmp(unsigned DestReg, const CmpInst *CI); 179 bool emitLoad(MVT VT, unsigned &ResultReg, Address &Addr, 180 unsigned Alignment = 0); 181 bool emitStore(MVT VT, unsigned SrcReg, Address Addr, 182 MachineMemOperand *MMO = nullptr); 183 bool emitStore(MVT VT, unsigned SrcReg, Address &Addr, 184 unsigned Alignment = 0); 185 unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt); 186 bool emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, unsigned DestReg, 187 188 bool IsZExt); 189 bool emitIntZExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, unsigned DestReg); 190 191 bool emitIntSExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, unsigned DestReg); 192 bool emitIntSExt32r1(MVT SrcVT, unsigned SrcReg, MVT DestVT, 193 unsigned DestReg); 194 bool emitIntSExt32r2(MVT SrcVT, unsigned SrcReg, MVT DestVT, 195 unsigned DestReg); 196 197 unsigned getRegEnsuringSimpleIntegerWidening(const Value *, bool IsUnsigned); 198 199 unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS, 200 const Value *RHS); 201 202 unsigned materializeFP(const ConstantFP *CFP, MVT VT); 203 unsigned materializeGV(const GlobalValue *GV, MVT VT); 204 unsigned materializeInt(const Constant *C, MVT VT); 205 unsigned materialize32BitInt(int64_t Imm, const TargetRegisterClass *RC); 206 unsigned materializeExternalCallSym(MCSymbol *Syn); 207 208 MachineInstrBuilder emitInst(unsigned Opc) { 209 return BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)); 210 } 211 212 MachineInstrBuilder emitInst(unsigned Opc, unsigned DstReg) { 213 return BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), 214 DstReg); 215 } 216 217 MachineInstrBuilder emitInstStore(unsigned Opc, unsigned SrcReg, 218 unsigned MemReg, int64_t MemOffset) { 219 return emitInst(Opc).addReg(SrcReg).addReg(MemReg).addImm(MemOffset); 220 } 221 222 MachineInstrBuilder emitInstLoad(unsigned Opc, unsigned DstReg, 223 unsigned MemReg, int64_t MemOffset) { 224 return emitInst(Opc, DstReg).addReg(MemReg).addImm(MemOffset); 225 } 226 227 unsigned fastEmitInst_rr(unsigned MachineInstOpcode, 228 const TargetRegisterClass *RC, 229 unsigned Op0, bool Op0IsKill, 230 unsigned Op1, bool Op1IsKill); 231 232 // for some reason, this default is not generated by tablegen 233 // so we explicitly generate it here. 234 unsigned fastEmitInst_riir(uint64_t inst, const TargetRegisterClass *RC, 235 unsigned Op0, bool Op0IsKill, uint64_t imm1, 236 uint64_t imm2, unsigned Op3, bool Op3IsKill) { 237 return 0; 238 } 239 240 // Call handling routines. 241 private: 242 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const; 243 bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs, 244 unsigned &NumBytes); 245 bool finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes); 246 247 const MipsABIInfo &getABI() const { 248 return static_cast<const MipsTargetMachine &>(TM).getABI(); 249 } 250 251 public: 252 // Backend specific FastISel code. 253 explicit MipsFastISel(FunctionLoweringInfo &funcInfo, 254 const TargetLibraryInfo *libInfo) 255 : FastISel(funcInfo, libInfo), TM(funcInfo.MF->getTarget()), 256 Subtarget(&funcInfo.MF->getSubtarget<MipsSubtarget>()), 257 TII(*Subtarget->getInstrInfo()), TLI(*Subtarget->getTargetLowering()) { 258 MFI = funcInfo.MF->getInfo<MipsFunctionInfo>(); 259 Context = &funcInfo.Fn->getContext(); 260 UnsupportedFPMode = Subtarget->isFP64bit() || Subtarget->useSoftFloat(); 261 } 262 263 unsigned fastMaterializeAlloca(const AllocaInst *AI) override; 264 unsigned fastMaterializeConstant(const Constant *C) override; 265 bool fastSelectInstruction(const Instruction *I) override; 266 267 #include "MipsGenFastISel.inc" 268 }; 269 270 } // end anonymous namespace 271 272 static bool CC_Mips(unsigned ValNo, MVT ValVT, MVT LocVT, 273 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 274 CCState &State) LLVM_ATTRIBUTE_UNUSED; 275 276 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, MVT LocVT, 277 CCValAssign::LocInfo LocInfo, 278 ISD::ArgFlagsTy ArgFlags, CCState &State) { 279 llvm_unreachable("should not be called"); 280 } 281 282 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, MVT LocVT, 283 CCValAssign::LocInfo LocInfo, 284 ISD::ArgFlagsTy ArgFlags, CCState &State) { 285 llvm_unreachable("should not be called"); 286 } 287 288 #include "MipsGenCallingConv.inc" 289 290 CCAssignFn *MipsFastISel::CCAssignFnForCall(CallingConv::ID CC) const { 291 return CC_MipsO32; 292 } 293 294 unsigned MipsFastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT, 295 const Value *LHS, const Value *RHS) { 296 // Canonicalize immediates to the RHS first. 297 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS)) 298 std::swap(LHS, RHS); 299 300 unsigned Opc; 301 switch (ISDOpc) { 302 case ISD::AND: 303 Opc = Mips::AND; 304 break; 305 case ISD::OR: 306 Opc = Mips::OR; 307 break; 308 case ISD::XOR: 309 Opc = Mips::XOR; 310 break; 311 default: 312 llvm_unreachable("unexpected opcode"); 313 } 314 315 unsigned LHSReg = getRegForValue(LHS); 316 if (!LHSReg) 317 return 0; 318 319 unsigned RHSReg; 320 if (const auto *C = dyn_cast<ConstantInt>(RHS)) 321 RHSReg = materializeInt(C, MVT::i32); 322 else 323 RHSReg = getRegForValue(RHS); 324 if (!RHSReg) 325 return 0; 326 327 unsigned ResultReg = createResultReg(&Mips::GPR32RegClass); 328 if (!ResultReg) 329 return 0; 330 331 emitInst(Opc, ResultReg).addReg(LHSReg).addReg(RHSReg); 332 return ResultReg; 333 } 334 335 unsigned MipsFastISel::fastMaterializeAlloca(const AllocaInst *AI) { 336 assert(TLI.getValueType(DL, AI->getType(), true) == MVT::i32 && 337 "Alloca should always return a pointer."); 338 339 DenseMap<const AllocaInst *, int>::iterator SI = 340 FuncInfo.StaticAllocaMap.find(AI); 341 342 if (SI != FuncInfo.StaticAllocaMap.end()) { 343 unsigned ResultReg = createResultReg(&Mips::GPR32RegClass); 344 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Mips::LEA_ADDiu), 345 ResultReg) 346 .addFrameIndex(SI->second) 347 .addImm(0); 348 return ResultReg; 349 } 350 351 return 0; 352 } 353 354 unsigned MipsFastISel::materializeInt(const Constant *C, MVT VT) { 355 if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1) 356 return 0; 357 const TargetRegisterClass *RC = &Mips::GPR32RegClass; 358 const ConstantInt *CI = cast<ConstantInt>(C); 359 return materialize32BitInt(CI->getZExtValue(), RC); 360 } 361 362 unsigned MipsFastISel::materialize32BitInt(int64_t Imm, 363 const TargetRegisterClass *RC) { 364 unsigned ResultReg = createResultReg(RC); 365 366 if (isInt<16>(Imm)) { 367 unsigned Opc = Mips::ADDiu; 368 emitInst(Opc, ResultReg).addReg(Mips::ZERO).addImm(Imm); 369 return ResultReg; 370 } else if (isUInt<16>(Imm)) { 371 emitInst(Mips::ORi, ResultReg).addReg(Mips::ZERO).addImm(Imm); 372 return ResultReg; 373 } 374 unsigned Lo = Imm & 0xFFFF; 375 unsigned Hi = (Imm >> 16) & 0xFFFF; 376 if (Lo) { 377 // Both Lo and Hi have nonzero bits. 378 unsigned TmpReg = createResultReg(RC); 379 emitInst(Mips::LUi, TmpReg).addImm(Hi); 380 emitInst(Mips::ORi, ResultReg).addReg(TmpReg).addImm(Lo); 381 } else { 382 emitInst(Mips::LUi, ResultReg).addImm(Hi); 383 } 384 return ResultReg; 385 } 386 387 unsigned MipsFastISel::materializeFP(const ConstantFP *CFP, MVT VT) { 388 if (UnsupportedFPMode) 389 return 0; 390 int64_t Imm = CFP->getValueAPF().bitcastToAPInt().getZExtValue(); 391 if (VT == MVT::f32) { 392 const TargetRegisterClass *RC = &Mips::FGR32RegClass; 393 unsigned DestReg = createResultReg(RC); 394 unsigned TempReg = materialize32BitInt(Imm, &Mips::GPR32RegClass); 395 emitInst(Mips::MTC1, DestReg).addReg(TempReg); 396 return DestReg; 397 } else if (VT == MVT::f64) { 398 const TargetRegisterClass *RC = &Mips::AFGR64RegClass; 399 unsigned DestReg = createResultReg(RC); 400 unsigned TempReg1 = materialize32BitInt(Imm >> 32, &Mips::GPR32RegClass); 401 unsigned TempReg2 = 402 materialize32BitInt(Imm & 0xFFFFFFFF, &Mips::GPR32RegClass); 403 emitInst(Mips::BuildPairF64, DestReg).addReg(TempReg2).addReg(TempReg1); 404 return DestReg; 405 } 406 return 0; 407 } 408 409 unsigned MipsFastISel::materializeGV(const GlobalValue *GV, MVT VT) { 410 // For now 32-bit only. 411 if (VT != MVT::i32) 412 return 0; 413 const TargetRegisterClass *RC = &Mips::GPR32RegClass; 414 unsigned DestReg = createResultReg(RC); 415 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV); 416 bool IsThreadLocal = GVar && GVar->isThreadLocal(); 417 // TLS not supported at this time. 418 if (IsThreadLocal) 419 return 0; 420 emitInst(Mips::LW, DestReg) 421 .addReg(MFI->getGlobalBaseReg()) 422 .addGlobalAddress(GV, 0, MipsII::MO_GOT); 423 if ((GV->hasInternalLinkage() || 424 (GV->hasLocalLinkage() && !isa<Function>(GV)))) { 425 unsigned TempReg = createResultReg(RC); 426 emitInst(Mips::ADDiu, TempReg) 427 .addReg(DestReg) 428 .addGlobalAddress(GV, 0, MipsII::MO_ABS_LO); 429 DestReg = TempReg; 430 } 431 return DestReg; 432 } 433 434 unsigned MipsFastISel::materializeExternalCallSym(MCSymbol *Sym) { 435 const TargetRegisterClass *RC = &Mips::GPR32RegClass; 436 unsigned DestReg = createResultReg(RC); 437 emitInst(Mips::LW, DestReg) 438 .addReg(MFI->getGlobalBaseReg()) 439 .addSym(Sym, MipsII::MO_GOT); 440 return DestReg; 441 } 442 443 // Materialize a constant into a register, and return the register 444 // number (or zero if we failed to handle it). 445 unsigned MipsFastISel::fastMaterializeConstant(const Constant *C) { 446 EVT CEVT = TLI.getValueType(DL, C->getType(), true); 447 448 // Only handle simple types. 449 if (!CEVT.isSimple()) 450 return 0; 451 MVT VT = CEVT.getSimpleVT(); 452 453 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) 454 return (UnsupportedFPMode) ? 0 : materializeFP(CFP, VT); 455 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) 456 return materializeGV(GV, VT); 457 else if (isa<ConstantInt>(C)) 458 return materializeInt(C, VT); 459 460 return 0; 461 } 462 463 bool MipsFastISel::computeAddress(const Value *Obj, Address &Addr) { 464 const User *U = nullptr; 465 unsigned Opcode = Instruction::UserOp1; 466 if (const Instruction *I = dyn_cast<Instruction>(Obj)) { 467 // Don't walk into other basic blocks unless the object is an alloca from 468 // another block, otherwise it may not have a virtual register assigned. 469 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) || 470 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) { 471 Opcode = I->getOpcode(); 472 U = I; 473 } 474 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) { 475 Opcode = C->getOpcode(); 476 U = C; 477 } 478 switch (Opcode) { 479 default: 480 break; 481 case Instruction::BitCast: 482 // Look through bitcasts. 483 return computeAddress(U->getOperand(0), Addr); 484 case Instruction::GetElementPtr: { 485 Address SavedAddr = Addr; 486 int64_t TmpOffset = Addr.getOffset(); 487 // Iterate through the GEP folding the constants into offsets where 488 // we can. 489 gep_type_iterator GTI = gep_type_begin(U); 490 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; 491 ++i, ++GTI) { 492 const Value *Op = *i; 493 if (StructType *STy = GTI.getStructTypeOrNull()) { 494 const StructLayout *SL = DL.getStructLayout(STy); 495 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue(); 496 TmpOffset += SL->getElementOffset(Idx); 497 } else { 498 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType()); 499 while (true) { 500 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 501 // Constant-offset addressing. 502 TmpOffset += CI->getSExtValue() * S; 503 break; 504 } 505 if (canFoldAddIntoGEP(U, Op)) { 506 // A compatible add with a constant operand. Fold the constant. 507 ConstantInt *CI = 508 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1)); 509 TmpOffset += CI->getSExtValue() * S; 510 // Iterate on the other operand. 511 Op = cast<AddOperator>(Op)->getOperand(0); 512 continue; 513 } 514 // Unsupported 515 goto unsupported_gep; 516 } 517 } 518 } 519 // Try to grab the base operand now. 520 Addr.setOffset(TmpOffset); 521 if (computeAddress(U->getOperand(0), Addr)) 522 return true; 523 // We failed, restore everything and try the other options. 524 Addr = SavedAddr; 525 unsupported_gep: 526 break; 527 } 528 case Instruction::Alloca: { 529 const AllocaInst *AI = cast<AllocaInst>(Obj); 530 DenseMap<const AllocaInst *, int>::iterator SI = 531 FuncInfo.StaticAllocaMap.find(AI); 532 if (SI != FuncInfo.StaticAllocaMap.end()) { 533 Addr.setKind(Address::FrameIndexBase); 534 Addr.setFI(SI->second); 535 return true; 536 } 537 break; 538 } 539 } 540 Addr.setReg(getRegForValue(Obj)); 541 return Addr.getReg() != 0; 542 } 543 544 bool MipsFastISel::computeCallAddress(const Value *V, Address &Addr) { 545 const User *U = nullptr; 546 unsigned Opcode = Instruction::UserOp1; 547 548 if (const auto *I = dyn_cast<Instruction>(V)) { 549 // Check if the value is defined in the same basic block. This information 550 // is crucial to know whether or not folding an operand is valid. 551 if (I->getParent() == FuncInfo.MBB->getBasicBlock()) { 552 Opcode = I->getOpcode(); 553 U = I; 554 } 555 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) { 556 Opcode = C->getOpcode(); 557 U = C; 558 } 559 560 switch (Opcode) { 561 default: 562 break; 563 case Instruction::BitCast: 564 // Look past bitcasts if its operand is in the same BB. 565 return computeCallAddress(U->getOperand(0), Addr); 566 break; 567 case Instruction::IntToPtr: 568 // Look past no-op inttoptrs if its operand is in the same BB. 569 if (TLI.getValueType(DL, U->getOperand(0)->getType()) == 570 TLI.getPointerTy(DL)) 571 return computeCallAddress(U->getOperand(0), Addr); 572 break; 573 case Instruction::PtrToInt: 574 // Look past no-op ptrtoints if its operand is in the same BB. 575 if (TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL)) 576 return computeCallAddress(U->getOperand(0), Addr); 577 break; 578 } 579 580 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 581 Addr.setGlobalValue(GV); 582 return true; 583 } 584 585 // If all else fails, try to materialize the value in a register. 586 if (!Addr.getGlobalValue()) { 587 Addr.setReg(getRegForValue(V)); 588 return Addr.getReg() != 0; 589 } 590 591 return false; 592 } 593 594 bool MipsFastISel::isTypeLegal(Type *Ty, MVT &VT) { 595 EVT evt = TLI.getValueType(DL, Ty, true); 596 // Only handle simple types. 597 if (evt == MVT::Other || !evt.isSimple()) 598 return false; 599 VT = evt.getSimpleVT(); 600 601 // Handle all legal types, i.e. a register that will directly hold this 602 // value. 603 return TLI.isTypeLegal(VT); 604 } 605 606 bool MipsFastISel::isTypeSupported(Type *Ty, MVT &VT) { 607 if (Ty->isVectorTy()) 608 return false; 609 610 if (isTypeLegal(Ty, VT)) 611 return true; 612 613 // If this is a type than can be sign or zero-extended to a basic operation 614 // go ahead and accept it now. 615 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) 616 return true; 617 618 return false; 619 } 620 621 bool MipsFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) { 622 if (isTypeLegal(Ty, VT)) 623 return true; 624 // We will extend this in a later patch: 625 // If this is a type than can be sign or zero-extended to a basic operation 626 // go ahead and accept it now. 627 if (VT == MVT::i8 || VT == MVT::i16) 628 return true; 629 return false; 630 } 631 632 // Because of how EmitCmp is called with fast-isel, you can 633 // end up with redundant "andi" instructions after the sequences emitted below. 634 // We should try and solve this issue in the future. 635 // 636 bool MipsFastISel::emitCmp(unsigned ResultReg, const CmpInst *CI) { 637 const Value *Left = CI->getOperand(0), *Right = CI->getOperand(1); 638 bool IsUnsigned = CI->isUnsigned(); 639 unsigned LeftReg = getRegEnsuringSimpleIntegerWidening(Left, IsUnsigned); 640 if (LeftReg == 0) 641 return false; 642 unsigned RightReg = getRegEnsuringSimpleIntegerWidening(Right, IsUnsigned); 643 if (RightReg == 0) 644 return false; 645 CmpInst::Predicate P = CI->getPredicate(); 646 647 switch (P) { 648 default: 649 return false; 650 case CmpInst::ICMP_EQ: { 651 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 652 emitInst(Mips::XOR, TempReg).addReg(LeftReg).addReg(RightReg); 653 emitInst(Mips::SLTiu, ResultReg).addReg(TempReg).addImm(1); 654 break; 655 } 656 case CmpInst::ICMP_NE: { 657 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 658 emitInst(Mips::XOR, TempReg).addReg(LeftReg).addReg(RightReg); 659 emitInst(Mips::SLTu, ResultReg).addReg(Mips::ZERO).addReg(TempReg); 660 break; 661 } 662 case CmpInst::ICMP_UGT: 663 emitInst(Mips::SLTu, ResultReg).addReg(RightReg).addReg(LeftReg); 664 break; 665 case CmpInst::ICMP_ULT: 666 emitInst(Mips::SLTu, ResultReg).addReg(LeftReg).addReg(RightReg); 667 break; 668 case CmpInst::ICMP_UGE: { 669 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 670 emitInst(Mips::SLTu, TempReg).addReg(LeftReg).addReg(RightReg); 671 emitInst(Mips::XORi, ResultReg).addReg(TempReg).addImm(1); 672 break; 673 } 674 case CmpInst::ICMP_ULE: { 675 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 676 emitInst(Mips::SLTu, TempReg).addReg(RightReg).addReg(LeftReg); 677 emitInst(Mips::XORi, ResultReg).addReg(TempReg).addImm(1); 678 break; 679 } 680 case CmpInst::ICMP_SGT: 681 emitInst(Mips::SLT, ResultReg).addReg(RightReg).addReg(LeftReg); 682 break; 683 case CmpInst::ICMP_SLT: 684 emitInst(Mips::SLT, ResultReg).addReg(LeftReg).addReg(RightReg); 685 break; 686 case CmpInst::ICMP_SGE: { 687 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 688 emitInst(Mips::SLT, TempReg).addReg(LeftReg).addReg(RightReg); 689 emitInst(Mips::XORi, ResultReg).addReg(TempReg).addImm(1); 690 break; 691 } 692 case CmpInst::ICMP_SLE: { 693 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 694 emitInst(Mips::SLT, TempReg).addReg(RightReg).addReg(LeftReg); 695 emitInst(Mips::XORi, ResultReg).addReg(TempReg).addImm(1); 696 break; 697 } 698 case CmpInst::FCMP_OEQ: 699 case CmpInst::FCMP_UNE: 700 case CmpInst::FCMP_OLT: 701 case CmpInst::FCMP_OLE: 702 case CmpInst::FCMP_OGT: 703 case CmpInst::FCMP_OGE: { 704 if (UnsupportedFPMode) 705 return false; 706 bool IsFloat = Left->getType()->isFloatTy(); 707 bool IsDouble = Left->getType()->isDoubleTy(); 708 if (!IsFloat && !IsDouble) 709 return false; 710 unsigned Opc, CondMovOpc; 711 switch (P) { 712 case CmpInst::FCMP_OEQ: 713 Opc = IsFloat ? Mips::C_EQ_S : Mips::C_EQ_D32; 714 CondMovOpc = Mips::MOVT_I; 715 break; 716 case CmpInst::FCMP_UNE: 717 Opc = IsFloat ? Mips::C_EQ_S : Mips::C_EQ_D32; 718 CondMovOpc = Mips::MOVF_I; 719 break; 720 case CmpInst::FCMP_OLT: 721 Opc = IsFloat ? Mips::C_OLT_S : Mips::C_OLT_D32; 722 CondMovOpc = Mips::MOVT_I; 723 break; 724 case CmpInst::FCMP_OLE: 725 Opc = IsFloat ? Mips::C_OLE_S : Mips::C_OLE_D32; 726 CondMovOpc = Mips::MOVT_I; 727 break; 728 case CmpInst::FCMP_OGT: 729 Opc = IsFloat ? Mips::C_ULE_S : Mips::C_ULE_D32; 730 CondMovOpc = Mips::MOVF_I; 731 break; 732 case CmpInst::FCMP_OGE: 733 Opc = IsFloat ? Mips::C_ULT_S : Mips::C_ULT_D32; 734 CondMovOpc = Mips::MOVF_I; 735 break; 736 default: 737 llvm_unreachable("Only switching of a subset of CCs."); 738 } 739 unsigned RegWithZero = createResultReg(&Mips::GPR32RegClass); 740 unsigned RegWithOne = createResultReg(&Mips::GPR32RegClass); 741 emitInst(Mips::ADDiu, RegWithZero).addReg(Mips::ZERO).addImm(0); 742 emitInst(Mips::ADDiu, RegWithOne).addReg(Mips::ZERO).addImm(1); 743 emitInst(Opc).addReg(Mips::FCC0, RegState::Define).addReg(LeftReg) 744 .addReg(RightReg); 745 emitInst(CondMovOpc, ResultReg) 746 .addReg(RegWithOne) 747 .addReg(Mips::FCC0) 748 .addReg(RegWithZero); 749 break; 750 } 751 } 752 return true; 753 } 754 755 bool MipsFastISel::emitLoad(MVT VT, unsigned &ResultReg, Address &Addr, 756 unsigned Alignment) { 757 // 758 // more cases will be handled here in following patches. 759 // 760 unsigned Opc; 761 switch (VT.SimpleTy) { 762 case MVT::i32: 763 ResultReg = createResultReg(&Mips::GPR32RegClass); 764 Opc = Mips::LW; 765 break; 766 case MVT::i16: 767 ResultReg = createResultReg(&Mips::GPR32RegClass); 768 Opc = Mips::LHu; 769 break; 770 case MVT::i8: 771 ResultReg = createResultReg(&Mips::GPR32RegClass); 772 Opc = Mips::LBu; 773 break; 774 case MVT::f32: 775 if (UnsupportedFPMode) 776 return false; 777 ResultReg = createResultReg(&Mips::FGR32RegClass); 778 Opc = Mips::LWC1; 779 break; 780 case MVT::f64: 781 if (UnsupportedFPMode) 782 return false; 783 ResultReg = createResultReg(&Mips::AFGR64RegClass); 784 Opc = Mips::LDC1; 785 break; 786 default: 787 return false; 788 } 789 if (Addr.isRegBase()) { 790 simplifyAddress(Addr); 791 emitInstLoad(Opc, ResultReg, Addr.getReg(), Addr.getOffset()); 792 return true; 793 } 794 if (Addr.isFIBase()) { 795 unsigned FI = Addr.getFI(); 796 unsigned Align = 4; 797 int64_t Offset = Addr.getOffset(); 798 MachineFrameInfo &MFI = MF->getFrameInfo(); 799 MachineMemOperand *MMO = MF->getMachineMemOperand( 800 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad, 801 MFI.getObjectSize(FI), Align); 802 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) 803 .addFrameIndex(FI) 804 .addImm(Offset) 805 .addMemOperand(MMO); 806 return true; 807 } 808 return false; 809 } 810 811 bool MipsFastISel::emitStore(MVT VT, unsigned SrcReg, Address &Addr, 812 unsigned Alignment) { 813 // 814 // more cases will be handled here in following patches. 815 // 816 unsigned Opc; 817 switch (VT.SimpleTy) { 818 case MVT::i8: 819 Opc = Mips::SB; 820 break; 821 case MVT::i16: 822 Opc = Mips::SH; 823 break; 824 case MVT::i32: 825 Opc = Mips::SW; 826 break; 827 case MVT::f32: 828 if (UnsupportedFPMode) 829 return false; 830 Opc = Mips::SWC1; 831 break; 832 case MVT::f64: 833 if (UnsupportedFPMode) 834 return false; 835 Opc = Mips::SDC1; 836 break; 837 default: 838 return false; 839 } 840 if (Addr.isRegBase()) { 841 simplifyAddress(Addr); 842 emitInstStore(Opc, SrcReg, Addr.getReg(), Addr.getOffset()); 843 return true; 844 } 845 if (Addr.isFIBase()) { 846 unsigned FI = Addr.getFI(); 847 unsigned Align = 4; 848 int64_t Offset = Addr.getOffset(); 849 MachineFrameInfo &MFI = MF->getFrameInfo(); 850 MachineMemOperand *MMO = MF->getMachineMemOperand( 851 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore, 852 MFI.getObjectSize(FI), Align); 853 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc)) 854 .addReg(SrcReg) 855 .addFrameIndex(FI) 856 .addImm(Offset) 857 .addMemOperand(MMO); 858 return true; 859 } 860 return false; 861 } 862 863 bool MipsFastISel::selectLogicalOp(const Instruction *I) { 864 MVT VT; 865 if (!isTypeSupported(I->getType(), VT)) 866 return false; 867 868 unsigned ResultReg; 869 switch (I->getOpcode()) { 870 default: 871 llvm_unreachable("Unexpected instruction."); 872 case Instruction::And: 873 ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1)); 874 break; 875 case Instruction::Or: 876 ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1)); 877 break; 878 case Instruction::Xor: 879 ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1)); 880 break; 881 } 882 883 if (!ResultReg) 884 return false; 885 886 updateValueMap(I, ResultReg); 887 return true; 888 } 889 890 bool MipsFastISel::selectLoad(const Instruction *I) { 891 // Atomic loads need special handling. 892 if (cast<LoadInst>(I)->isAtomic()) 893 return false; 894 895 // Verify we have a legal type before going any further. 896 MVT VT; 897 if (!isLoadTypeLegal(I->getType(), VT)) 898 return false; 899 900 // See if we can handle this address. 901 Address Addr; 902 if (!computeAddress(I->getOperand(0), Addr)) 903 return false; 904 905 unsigned ResultReg; 906 if (!emitLoad(VT, ResultReg, Addr, cast<LoadInst>(I)->getAlignment())) 907 return false; 908 updateValueMap(I, ResultReg); 909 return true; 910 } 911 912 bool MipsFastISel::selectStore(const Instruction *I) { 913 Value *Op0 = I->getOperand(0); 914 unsigned SrcReg = 0; 915 916 // Atomic stores need special handling. 917 if (cast<StoreInst>(I)->isAtomic()) 918 return false; 919 920 // Verify we have a legal type before going any further. 921 MVT VT; 922 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT)) 923 return false; 924 925 // Get the value to be stored into a register. 926 SrcReg = getRegForValue(Op0); 927 if (SrcReg == 0) 928 return false; 929 930 // See if we can handle this address. 931 Address Addr; 932 if (!computeAddress(I->getOperand(1), Addr)) 933 return false; 934 935 if (!emitStore(VT, SrcReg, Addr, cast<StoreInst>(I)->getAlignment())) 936 return false; 937 return true; 938 } 939 940 // This can cause a redundant sltiu to be generated. 941 // FIXME: try and eliminate this in a future patch. 942 bool MipsFastISel::selectBranch(const Instruction *I) { 943 const BranchInst *BI = cast<BranchInst>(I); 944 MachineBasicBlock *BrBB = FuncInfo.MBB; 945 // 946 // TBB is the basic block for the case where the comparison is true. 947 // FBB is the basic block for the case where the comparison is false. 948 // if (cond) goto TBB 949 // goto FBB 950 // TBB: 951 // 952 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)]; 953 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)]; 954 // For now, just try the simplest case where it's fed by a compare. 955 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) { 956 unsigned CondReg = getRegForValue(CI); 957 BuildMI(*BrBB, FuncInfo.InsertPt, DbgLoc, TII.get(Mips::BGTZ)) 958 .addReg(CondReg) 959 .addMBB(TBB); 960 finishCondBranch(BI->getParent(), TBB, FBB); 961 return true; 962 } 963 return false; 964 } 965 966 bool MipsFastISel::selectCmp(const Instruction *I) { 967 const CmpInst *CI = cast<CmpInst>(I); 968 unsigned ResultReg = createResultReg(&Mips::GPR32RegClass); 969 if (!emitCmp(ResultReg, CI)) 970 return false; 971 updateValueMap(I, ResultReg); 972 return true; 973 } 974 975 // Attempt to fast-select a floating-point extend instruction. 976 bool MipsFastISel::selectFPExt(const Instruction *I) { 977 if (UnsupportedFPMode) 978 return false; 979 Value *Src = I->getOperand(0); 980 EVT SrcVT = TLI.getValueType(DL, Src->getType(), true); 981 EVT DestVT = TLI.getValueType(DL, I->getType(), true); 982 983 if (SrcVT != MVT::f32 || DestVT != MVT::f64) 984 return false; 985 986 unsigned SrcReg = 987 getRegForValue(Src); // this must be a 32bit floating point register class 988 // maybe we should handle this differently 989 if (!SrcReg) 990 return false; 991 992 unsigned DestReg = createResultReg(&Mips::AFGR64RegClass); 993 emitInst(Mips::CVT_D32_S, DestReg).addReg(SrcReg); 994 updateValueMap(I, DestReg); 995 return true; 996 } 997 998 bool MipsFastISel::selectSelect(const Instruction *I) { 999 assert(isa<SelectInst>(I) && "Expected a select instruction."); 1000 1001 LLVM_DEBUG(dbgs() << "selectSelect\n"); 1002 1003 MVT VT; 1004 if (!isTypeSupported(I->getType(), VT) || UnsupportedFPMode) { 1005 LLVM_DEBUG( 1006 dbgs() << ".. .. gave up (!isTypeSupported || UnsupportedFPMode)\n"); 1007 return false; 1008 } 1009 1010 unsigned CondMovOpc; 1011 const TargetRegisterClass *RC; 1012 1013 if (VT.isInteger() && !VT.isVector() && VT.getSizeInBits() <= 32) { 1014 CondMovOpc = Mips::MOVN_I_I; 1015 RC = &Mips::GPR32RegClass; 1016 } else if (VT == MVT::f32) { 1017 CondMovOpc = Mips::MOVN_I_S; 1018 RC = &Mips::FGR32RegClass; 1019 } else if (VT == MVT::f64) { 1020 CondMovOpc = Mips::MOVN_I_D32; 1021 RC = &Mips::AFGR64RegClass; 1022 } else 1023 return false; 1024 1025 const SelectInst *SI = cast<SelectInst>(I); 1026 const Value *Cond = SI->getCondition(); 1027 unsigned Src1Reg = getRegForValue(SI->getTrueValue()); 1028 unsigned Src2Reg = getRegForValue(SI->getFalseValue()); 1029 unsigned CondReg = getRegForValue(Cond); 1030 1031 if (!Src1Reg || !Src2Reg || !CondReg) 1032 return false; 1033 1034 unsigned ZExtCondReg = createResultReg(&Mips::GPR32RegClass); 1035 if (!ZExtCondReg) 1036 return false; 1037 1038 if (!emitIntExt(MVT::i1, CondReg, MVT::i32, ZExtCondReg, true)) 1039 return false; 1040 1041 unsigned ResultReg = createResultReg(RC); 1042 unsigned TempReg = createResultReg(RC); 1043 1044 if (!ResultReg || !TempReg) 1045 return false; 1046 1047 emitInst(TargetOpcode::COPY, TempReg).addReg(Src2Reg); 1048 emitInst(CondMovOpc, ResultReg) 1049 .addReg(Src1Reg).addReg(ZExtCondReg).addReg(TempReg); 1050 updateValueMap(I, ResultReg); 1051 return true; 1052 } 1053 1054 // Attempt to fast-select a floating-point truncate instruction. 1055 bool MipsFastISel::selectFPTrunc(const Instruction *I) { 1056 if (UnsupportedFPMode) 1057 return false; 1058 Value *Src = I->getOperand(0); 1059 EVT SrcVT = TLI.getValueType(DL, Src->getType(), true); 1060 EVT DestVT = TLI.getValueType(DL, I->getType(), true); 1061 1062 if (SrcVT != MVT::f64 || DestVT != MVT::f32) 1063 return false; 1064 1065 unsigned SrcReg = getRegForValue(Src); 1066 if (!SrcReg) 1067 return false; 1068 1069 unsigned DestReg = createResultReg(&Mips::FGR32RegClass); 1070 if (!DestReg) 1071 return false; 1072 1073 emitInst(Mips::CVT_S_D32, DestReg).addReg(SrcReg); 1074 updateValueMap(I, DestReg); 1075 return true; 1076 } 1077 1078 // Attempt to fast-select a floating-point-to-integer conversion. 1079 bool MipsFastISel::selectFPToInt(const Instruction *I, bool IsSigned) { 1080 if (UnsupportedFPMode) 1081 return false; 1082 MVT DstVT, SrcVT; 1083 if (!IsSigned) 1084 return false; // We don't handle this case yet. There is no native 1085 // instruction for this but it can be synthesized. 1086 Type *DstTy = I->getType(); 1087 if (!isTypeLegal(DstTy, DstVT)) 1088 return false; 1089 1090 if (DstVT != MVT::i32) 1091 return false; 1092 1093 Value *Src = I->getOperand(0); 1094 Type *SrcTy = Src->getType(); 1095 if (!isTypeLegal(SrcTy, SrcVT)) 1096 return false; 1097 1098 if (SrcVT != MVT::f32 && SrcVT != MVT::f64) 1099 return false; 1100 1101 unsigned SrcReg = getRegForValue(Src); 1102 if (SrcReg == 0) 1103 return false; 1104 1105 // Determine the opcode for the conversion, which takes place 1106 // entirely within FPRs. 1107 unsigned DestReg = createResultReg(&Mips::GPR32RegClass); 1108 unsigned TempReg = createResultReg(&Mips::FGR32RegClass); 1109 unsigned Opc = (SrcVT == MVT::f32) ? Mips::TRUNC_W_S : Mips::TRUNC_W_D32; 1110 1111 // Generate the convert. 1112 emitInst(Opc, TempReg).addReg(SrcReg); 1113 emitInst(Mips::MFC1, DestReg).addReg(TempReg); 1114 1115 updateValueMap(I, DestReg); 1116 return true; 1117 } 1118 1119 bool MipsFastISel::processCallArgs(CallLoweringInfo &CLI, 1120 SmallVectorImpl<MVT> &OutVTs, 1121 unsigned &NumBytes) { 1122 CallingConv::ID CC = CLI.CallConv; 1123 SmallVector<CCValAssign, 16> ArgLocs; 1124 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context); 1125 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC)); 1126 // Get a count of how many bytes are to be pushed on the stack. 1127 NumBytes = CCInfo.getNextStackOffset(); 1128 // This is the minimum argument area used for A0-A3. 1129 if (NumBytes < 16) 1130 NumBytes = 16; 1131 1132 emitInst(Mips::ADJCALLSTACKDOWN).addImm(16).addImm(0); 1133 // Process the args. 1134 MVT firstMVT; 1135 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1136 CCValAssign &VA = ArgLocs[i]; 1137 const Value *ArgVal = CLI.OutVals[VA.getValNo()]; 1138 MVT ArgVT = OutVTs[VA.getValNo()]; 1139 1140 if (i == 0) { 1141 firstMVT = ArgVT; 1142 if (ArgVT == MVT::f32) { 1143 VA.convertToReg(Mips::F12); 1144 } else if (ArgVT == MVT::f64) { 1145 VA.convertToReg(Mips::D6); 1146 } 1147 } else if (i == 1) { 1148 if ((firstMVT == MVT::f32) || (firstMVT == MVT::f64)) { 1149 if (ArgVT == MVT::f32) { 1150 VA.convertToReg(Mips::F14); 1151 } else if (ArgVT == MVT::f64) { 1152 VA.convertToReg(Mips::D7); 1153 } 1154 } 1155 } 1156 if (((ArgVT == MVT::i32) || (ArgVT == MVT::f32) || (ArgVT == MVT::i16) || 1157 (ArgVT == MVT::i8)) && 1158 VA.isMemLoc()) { 1159 switch (VA.getLocMemOffset()) { 1160 case 0: 1161 VA.convertToReg(Mips::A0); 1162 break; 1163 case 4: 1164 VA.convertToReg(Mips::A1); 1165 break; 1166 case 8: 1167 VA.convertToReg(Mips::A2); 1168 break; 1169 case 12: 1170 VA.convertToReg(Mips::A3); 1171 break; 1172 default: 1173 break; 1174 } 1175 } 1176 unsigned ArgReg = getRegForValue(ArgVal); 1177 if (!ArgReg) 1178 return false; 1179 1180 // Handle arg promotion: SExt, ZExt, AExt. 1181 switch (VA.getLocInfo()) { 1182 case CCValAssign::Full: 1183 break; 1184 case CCValAssign::AExt: 1185 case CCValAssign::SExt: { 1186 MVT DestVT = VA.getLocVT(); 1187 MVT SrcVT = ArgVT; 1188 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false); 1189 if (!ArgReg) 1190 return false; 1191 break; 1192 } 1193 case CCValAssign::ZExt: { 1194 MVT DestVT = VA.getLocVT(); 1195 MVT SrcVT = ArgVT; 1196 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true); 1197 if (!ArgReg) 1198 return false; 1199 break; 1200 } 1201 default: 1202 llvm_unreachable("Unknown arg promotion!"); 1203 } 1204 1205 // Now copy/store arg to correct locations. 1206 if (VA.isRegLoc() && !VA.needsCustom()) { 1207 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1208 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg); 1209 CLI.OutRegs.push_back(VA.getLocReg()); 1210 } else if (VA.needsCustom()) { 1211 llvm_unreachable("Mips does not use custom args."); 1212 return false; 1213 } else { 1214 // 1215 // FIXME: This path will currently return false. It was copied 1216 // from the AArch64 port and should be essentially fine for Mips too. 1217 // The work to finish up this path will be done in a follow-on patch. 1218 // 1219 assert(VA.isMemLoc() && "Assuming store on stack."); 1220 // Don't emit stores for undef values. 1221 if (isa<UndefValue>(ArgVal)) 1222 continue; 1223 1224 // Need to store on the stack. 1225 // FIXME: This alignment is incorrect but this path is disabled 1226 // for now (will return false). We need to determine the right alignment 1227 // based on the normal alignment for the underlying machine type. 1228 // 1229 unsigned ArgSize = alignTo(ArgVT.getSizeInBits(), 4); 1230 1231 unsigned BEAlign = 0; 1232 if (ArgSize < 8 && !Subtarget->isLittle()) 1233 BEAlign = 8 - ArgSize; 1234 1235 Address Addr; 1236 Addr.setKind(Address::RegBase); 1237 Addr.setReg(Mips::SP); 1238 Addr.setOffset(VA.getLocMemOffset() + BEAlign); 1239 1240 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType()); 1241 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( 1242 MachinePointerInfo::getStack(*FuncInfo.MF, Addr.getOffset()), 1243 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment); 1244 (void)(MMO); 1245 // if (!emitStore(ArgVT, ArgReg, Addr, MMO)) 1246 return false; // can't store on the stack yet. 1247 } 1248 } 1249 1250 return true; 1251 } 1252 1253 bool MipsFastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT, 1254 unsigned NumBytes) { 1255 CallingConv::ID CC = CLI.CallConv; 1256 emitInst(Mips::ADJCALLSTACKUP).addImm(16).addImm(0); 1257 if (RetVT != MVT::isVoid) { 1258 SmallVector<CCValAssign, 16> RVLocs; 1259 MipsCCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context); 1260 1261 CCInfo.AnalyzeCallResult(CLI.Ins, RetCC_Mips, CLI.RetTy, 1262 CLI.Symbol ? CLI.Symbol->getName().data() 1263 : nullptr); 1264 1265 // Only handle a single return value. 1266 if (RVLocs.size() != 1) 1267 return false; 1268 // Copy all of the result registers out of their specified physreg. 1269 MVT CopyVT = RVLocs[0].getValVT(); 1270 // Special handling for extended integers. 1271 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16) 1272 CopyVT = MVT::i32; 1273 1274 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT)); 1275 if (!ResultReg) 1276 return false; 1277 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1278 TII.get(TargetOpcode::COPY), 1279 ResultReg).addReg(RVLocs[0].getLocReg()); 1280 CLI.InRegs.push_back(RVLocs[0].getLocReg()); 1281 1282 CLI.ResultReg = ResultReg; 1283 CLI.NumResultRegs = 1; 1284 } 1285 return true; 1286 } 1287 1288 bool MipsFastISel::fastLowerArguments() { 1289 LLVM_DEBUG(dbgs() << "fastLowerArguments\n"); 1290 1291 if (!FuncInfo.CanLowerReturn) { 1292 LLVM_DEBUG(dbgs() << ".. gave up (!CanLowerReturn)\n"); 1293 return false; 1294 } 1295 1296 const Function *F = FuncInfo.Fn; 1297 if (F->isVarArg()) { 1298 LLVM_DEBUG(dbgs() << ".. gave up (varargs)\n"); 1299 return false; 1300 } 1301 1302 CallingConv::ID CC = F->getCallingConv(); 1303 if (CC != CallingConv::C) { 1304 LLVM_DEBUG(dbgs() << ".. gave up (calling convention is not C)\n"); 1305 return false; 1306 } 1307 1308 std::array<MCPhysReg, 4> GPR32ArgRegs = {{Mips::A0, Mips::A1, Mips::A2, 1309 Mips::A3}}; 1310 std::array<MCPhysReg, 2> FGR32ArgRegs = {{Mips::F12, Mips::F14}}; 1311 std::array<MCPhysReg, 2> AFGR64ArgRegs = {{Mips::D6, Mips::D7}}; 1312 auto NextGPR32 = GPR32ArgRegs.begin(); 1313 auto NextFGR32 = FGR32ArgRegs.begin(); 1314 auto NextAFGR64 = AFGR64ArgRegs.begin(); 1315 1316 struct AllocatedReg { 1317 const TargetRegisterClass *RC; 1318 unsigned Reg; 1319 AllocatedReg(const TargetRegisterClass *RC, unsigned Reg) 1320 : RC(RC), Reg(Reg) {} 1321 }; 1322 1323 // Only handle simple cases. i.e. All arguments are directly mapped to 1324 // registers of the appropriate type. 1325 SmallVector<AllocatedReg, 4> Allocation; 1326 for (const auto &FormalArg : F->args()) { 1327 if (FormalArg.hasAttribute(Attribute::InReg) || 1328 FormalArg.hasAttribute(Attribute::StructRet) || 1329 FormalArg.hasAttribute(Attribute::ByVal)) { 1330 LLVM_DEBUG(dbgs() << ".. gave up (inreg, structret, byval)\n"); 1331 return false; 1332 } 1333 1334 Type *ArgTy = FormalArg.getType(); 1335 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) { 1336 LLVM_DEBUG(dbgs() << ".. gave up (struct, array, or vector)\n"); 1337 return false; 1338 } 1339 1340 EVT ArgVT = TLI.getValueType(DL, ArgTy); 1341 LLVM_DEBUG(dbgs() << ".. " << FormalArg.getArgNo() << ": " 1342 << ArgVT.getEVTString() << "\n"); 1343 if (!ArgVT.isSimple()) { 1344 LLVM_DEBUG(dbgs() << ".. .. gave up (not a simple type)\n"); 1345 return false; 1346 } 1347 1348 switch (ArgVT.getSimpleVT().SimpleTy) { 1349 case MVT::i1: 1350 case MVT::i8: 1351 case MVT::i16: 1352 if (!FormalArg.hasAttribute(Attribute::SExt) && 1353 !FormalArg.hasAttribute(Attribute::ZExt)) { 1354 // It must be any extend, this shouldn't happen for clang-generated IR 1355 // so just fall back on SelectionDAG. 1356 LLVM_DEBUG(dbgs() << ".. .. gave up (i8/i16 arg is not extended)\n"); 1357 return false; 1358 } 1359 1360 if (NextGPR32 == GPR32ArgRegs.end()) { 1361 LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of GPR32 arguments)\n"); 1362 return false; 1363 } 1364 1365 LLVM_DEBUG(dbgs() << ".. .. GPR32(" << *NextGPR32 << ")\n"); 1366 Allocation.emplace_back(&Mips::GPR32RegClass, *NextGPR32++); 1367 1368 // Allocating any GPR32 prohibits further use of floating point arguments. 1369 NextFGR32 = FGR32ArgRegs.end(); 1370 NextAFGR64 = AFGR64ArgRegs.end(); 1371 break; 1372 1373 case MVT::i32: 1374 if (FormalArg.hasAttribute(Attribute::ZExt)) { 1375 // The O32 ABI does not permit a zero-extended i32. 1376 LLVM_DEBUG(dbgs() << ".. .. gave up (i32 arg is zero extended)\n"); 1377 return false; 1378 } 1379 1380 if (NextGPR32 == GPR32ArgRegs.end()) { 1381 LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of GPR32 arguments)\n"); 1382 return false; 1383 } 1384 1385 LLVM_DEBUG(dbgs() << ".. .. GPR32(" << *NextGPR32 << ")\n"); 1386 Allocation.emplace_back(&Mips::GPR32RegClass, *NextGPR32++); 1387 1388 // Allocating any GPR32 prohibits further use of floating point arguments. 1389 NextFGR32 = FGR32ArgRegs.end(); 1390 NextAFGR64 = AFGR64ArgRegs.end(); 1391 break; 1392 1393 case MVT::f32: 1394 if (UnsupportedFPMode) { 1395 LLVM_DEBUG(dbgs() << ".. .. gave up (UnsupportedFPMode)\n"); 1396 return false; 1397 } 1398 if (NextFGR32 == FGR32ArgRegs.end()) { 1399 LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of FGR32 arguments)\n"); 1400 return false; 1401 } 1402 LLVM_DEBUG(dbgs() << ".. .. FGR32(" << *NextFGR32 << ")\n"); 1403 Allocation.emplace_back(&Mips::FGR32RegClass, *NextFGR32++); 1404 // Allocating an FGR32 also allocates the super-register AFGR64, and 1405 // ABI rules require us to skip the corresponding GPR32. 1406 if (NextGPR32 != GPR32ArgRegs.end()) 1407 NextGPR32++; 1408 if (NextAFGR64 != AFGR64ArgRegs.end()) 1409 NextAFGR64++; 1410 break; 1411 1412 case MVT::f64: 1413 if (UnsupportedFPMode) { 1414 LLVM_DEBUG(dbgs() << ".. .. gave up (UnsupportedFPMode)\n"); 1415 return false; 1416 } 1417 if (NextAFGR64 == AFGR64ArgRegs.end()) { 1418 LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of AFGR64 arguments)\n"); 1419 return false; 1420 } 1421 LLVM_DEBUG(dbgs() << ".. .. AFGR64(" << *NextAFGR64 << ")\n"); 1422 Allocation.emplace_back(&Mips::AFGR64RegClass, *NextAFGR64++); 1423 // Allocating an FGR32 also allocates the super-register AFGR64, and 1424 // ABI rules require us to skip the corresponding GPR32 pair. 1425 if (NextGPR32 != GPR32ArgRegs.end()) 1426 NextGPR32++; 1427 if (NextGPR32 != GPR32ArgRegs.end()) 1428 NextGPR32++; 1429 if (NextFGR32 != FGR32ArgRegs.end()) 1430 NextFGR32++; 1431 break; 1432 1433 default: 1434 LLVM_DEBUG(dbgs() << ".. .. gave up (unknown type)\n"); 1435 return false; 1436 } 1437 } 1438 1439 for (const auto &FormalArg : F->args()) { 1440 unsigned ArgNo = FormalArg.getArgNo(); 1441 unsigned SrcReg = Allocation[ArgNo].Reg; 1442 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, Allocation[ArgNo].RC); 1443 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. 1444 // Without this, EmitLiveInCopies may eliminate the livein if its only 1445 // use is a bitcast (which isn't turned into an instruction). 1446 unsigned ResultReg = createResultReg(Allocation[ArgNo].RC); 1447 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1448 TII.get(TargetOpcode::COPY), ResultReg) 1449 .addReg(DstReg, getKillRegState(true)); 1450 updateValueMap(&FormalArg, ResultReg); 1451 } 1452 1453 // Calculate the size of the incoming arguments area. 1454 // We currently reject all the cases where this would be non-zero. 1455 unsigned IncomingArgSizeInBytes = 0; 1456 1457 // Account for the reserved argument area on ABI's that have one (O32). 1458 // It seems strange to do this on the caller side but it's necessary in 1459 // SelectionDAG's implementation. 1460 IncomingArgSizeInBytes = std::min(getABI().GetCalleeAllocdArgSizeInBytes(CC), 1461 IncomingArgSizeInBytes); 1462 1463 MF->getInfo<MipsFunctionInfo>()->setFormalArgInfo(IncomingArgSizeInBytes, 1464 false); 1465 1466 return true; 1467 } 1468 1469 bool MipsFastISel::fastLowerCall(CallLoweringInfo &CLI) { 1470 CallingConv::ID CC = CLI.CallConv; 1471 bool IsTailCall = CLI.IsTailCall; 1472 bool IsVarArg = CLI.IsVarArg; 1473 const Value *Callee = CLI.Callee; 1474 MCSymbol *Symbol = CLI.Symbol; 1475 1476 // Do not handle FastCC. 1477 if (CC == CallingConv::Fast) 1478 return false; 1479 1480 // Allow SelectionDAG isel to handle tail calls. 1481 if (IsTailCall) 1482 return false; 1483 1484 // Let SDISel handle vararg functions. 1485 if (IsVarArg) 1486 return false; 1487 1488 // FIXME: Only handle *simple* calls for now. 1489 MVT RetVT; 1490 if (CLI.RetTy->isVoidTy()) 1491 RetVT = MVT::isVoid; 1492 else if (!isTypeSupported(CLI.RetTy, RetVT)) 1493 return false; 1494 1495 for (auto Flag : CLI.OutFlags) 1496 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal()) 1497 return false; 1498 1499 // Set up the argument vectors. 1500 SmallVector<MVT, 16> OutVTs; 1501 OutVTs.reserve(CLI.OutVals.size()); 1502 1503 for (auto *Val : CLI.OutVals) { 1504 MVT VT; 1505 if (!isTypeLegal(Val->getType(), VT) && 1506 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)) 1507 return false; 1508 1509 // We don't handle vector parameters yet. 1510 if (VT.isVector() || VT.getSizeInBits() > 64) 1511 return false; 1512 1513 OutVTs.push_back(VT); 1514 } 1515 1516 Address Addr; 1517 if (!computeCallAddress(Callee, Addr)) 1518 return false; 1519 1520 // Handle the arguments now that we've gotten them. 1521 unsigned NumBytes; 1522 if (!processCallArgs(CLI, OutVTs, NumBytes)) 1523 return false; 1524 1525 if (!Addr.getGlobalValue()) 1526 return false; 1527 1528 // Issue the call. 1529 unsigned DestAddress; 1530 if (Symbol) 1531 DestAddress = materializeExternalCallSym(Symbol); 1532 else 1533 DestAddress = materializeGV(Addr.getGlobalValue(), MVT::i32); 1534 emitInst(TargetOpcode::COPY, Mips::T9).addReg(DestAddress); 1535 MachineInstrBuilder MIB = 1536 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Mips::JALR), 1537 Mips::RA).addReg(Mips::T9); 1538 1539 // Add implicit physical register uses to the call. 1540 for (auto Reg : CLI.OutRegs) 1541 MIB.addReg(Reg, RegState::Implicit); 1542 1543 // Add a register mask with the call-preserved registers. 1544 // Proper defs for return values will be added by setPhysRegsDeadExcept(). 1545 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC)); 1546 1547 CLI.Call = MIB; 1548 1549 // Finish off the call including any return values. 1550 return finishCall(CLI, RetVT, NumBytes); 1551 } 1552 1553 bool MipsFastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) { 1554 switch (II->getIntrinsicID()) { 1555 default: 1556 return false; 1557 case Intrinsic::bswap: { 1558 Type *RetTy = II->getCalledFunction()->getReturnType(); 1559 1560 MVT VT; 1561 if (!isTypeSupported(RetTy, VT)) 1562 return false; 1563 1564 unsigned SrcReg = getRegForValue(II->getOperand(0)); 1565 if (SrcReg == 0) 1566 return false; 1567 unsigned DestReg = createResultReg(&Mips::GPR32RegClass); 1568 if (DestReg == 0) 1569 return false; 1570 if (VT == MVT::i16) { 1571 if (Subtarget->hasMips32r2()) { 1572 emitInst(Mips::WSBH, DestReg).addReg(SrcReg); 1573 updateValueMap(II, DestReg); 1574 return true; 1575 } else { 1576 unsigned TempReg[3]; 1577 for (int i = 0; i < 3; i++) { 1578 TempReg[i] = createResultReg(&Mips::GPR32RegClass); 1579 if (TempReg[i] == 0) 1580 return false; 1581 } 1582 emitInst(Mips::SLL, TempReg[0]).addReg(SrcReg).addImm(8); 1583 emitInst(Mips::SRL, TempReg[1]).addReg(SrcReg).addImm(8); 1584 emitInst(Mips::OR, TempReg[2]).addReg(TempReg[0]).addReg(TempReg[1]); 1585 emitInst(Mips::ANDi, DestReg).addReg(TempReg[2]).addImm(0xFFFF); 1586 updateValueMap(II, DestReg); 1587 return true; 1588 } 1589 } else if (VT == MVT::i32) { 1590 if (Subtarget->hasMips32r2()) { 1591 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 1592 emitInst(Mips::WSBH, TempReg).addReg(SrcReg); 1593 emitInst(Mips::ROTR, DestReg).addReg(TempReg).addImm(16); 1594 updateValueMap(II, DestReg); 1595 return true; 1596 } else { 1597 unsigned TempReg[8]; 1598 for (int i = 0; i < 8; i++) { 1599 TempReg[i] = createResultReg(&Mips::GPR32RegClass); 1600 if (TempReg[i] == 0) 1601 return false; 1602 } 1603 1604 emitInst(Mips::SRL, TempReg[0]).addReg(SrcReg).addImm(8); 1605 emitInst(Mips::SRL, TempReg[1]).addReg(SrcReg).addImm(24); 1606 emitInst(Mips::ANDi, TempReg[2]).addReg(TempReg[0]).addImm(0xFF00); 1607 emitInst(Mips::OR, TempReg[3]).addReg(TempReg[1]).addReg(TempReg[2]); 1608 1609 emitInst(Mips::ANDi, TempReg[4]).addReg(SrcReg).addImm(0xFF00); 1610 emitInst(Mips::SLL, TempReg[5]).addReg(TempReg[4]).addImm(8); 1611 1612 emitInst(Mips::SLL, TempReg[6]).addReg(SrcReg).addImm(24); 1613 emitInst(Mips::OR, TempReg[7]).addReg(TempReg[3]).addReg(TempReg[5]); 1614 emitInst(Mips::OR, DestReg).addReg(TempReg[6]).addReg(TempReg[7]); 1615 updateValueMap(II, DestReg); 1616 return true; 1617 } 1618 } 1619 return false; 1620 } 1621 case Intrinsic::memcpy: 1622 case Intrinsic::memmove: { 1623 const auto *MTI = cast<MemTransferInst>(II); 1624 // Don't handle volatile. 1625 if (MTI->isVolatile()) 1626 return false; 1627 if (!MTI->getLength()->getType()->isIntegerTy(32)) 1628 return false; 1629 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove"; 1630 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 1); 1631 } 1632 case Intrinsic::memset: { 1633 const MemSetInst *MSI = cast<MemSetInst>(II); 1634 // Don't handle volatile. 1635 if (MSI->isVolatile()) 1636 return false; 1637 if (!MSI->getLength()->getType()->isIntegerTy(32)) 1638 return false; 1639 return lowerCallTo(II, "memset", II->getNumArgOperands() - 1); 1640 } 1641 } 1642 return false; 1643 } 1644 1645 bool MipsFastISel::selectRet(const Instruction *I) { 1646 const Function &F = *I->getParent()->getParent(); 1647 const ReturnInst *Ret = cast<ReturnInst>(I); 1648 1649 LLVM_DEBUG(dbgs() << "selectRet\n"); 1650 1651 if (!FuncInfo.CanLowerReturn) 1652 return false; 1653 1654 // Build a list of return value registers. 1655 SmallVector<unsigned, 4> RetRegs; 1656 1657 if (Ret->getNumOperands() > 0) { 1658 CallingConv::ID CC = F.getCallingConv(); 1659 1660 // Do not handle FastCC. 1661 if (CC == CallingConv::Fast) 1662 return false; 1663 1664 SmallVector<ISD::OutputArg, 4> Outs; 1665 GetReturnInfo(CC, F.getReturnType(), F.getAttributes(), Outs, TLI, DL); 1666 1667 // Analyze operands of the call, assigning locations to each operand. 1668 SmallVector<CCValAssign, 16> ValLocs; 1669 MipsCCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, 1670 I->getContext()); 1671 CCAssignFn *RetCC = RetCC_Mips; 1672 CCInfo.AnalyzeReturn(Outs, RetCC); 1673 1674 // Only handle a single return value for now. 1675 if (ValLocs.size() != 1) 1676 return false; 1677 1678 CCValAssign &VA = ValLocs[0]; 1679 const Value *RV = Ret->getOperand(0); 1680 1681 // Don't bother handling odd stuff for now. 1682 if ((VA.getLocInfo() != CCValAssign::Full) && 1683 (VA.getLocInfo() != CCValAssign::BCvt)) 1684 return false; 1685 1686 // Only handle register returns for now. 1687 if (!VA.isRegLoc()) 1688 return false; 1689 1690 unsigned Reg = getRegForValue(RV); 1691 if (Reg == 0) 1692 return false; 1693 1694 unsigned SrcReg = Reg + VA.getValNo(); 1695 unsigned DestReg = VA.getLocReg(); 1696 // Avoid a cross-class copy. This is very unlikely. 1697 if (!MRI.getRegClass(SrcReg)->contains(DestReg)) 1698 return false; 1699 1700 EVT RVEVT = TLI.getValueType(DL, RV->getType()); 1701 if (!RVEVT.isSimple()) 1702 return false; 1703 1704 if (RVEVT.isVector()) 1705 return false; 1706 1707 MVT RVVT = RVEVT.getSimpleVT(); 1708 if (RVVT == MVT::f128) 1709 return false; 1710 1711 // Do not handle FGR64 returns for now. 1712 if (RVVT == MVT::f64 && UnsupportedFPMode) { 1713 LLVM_DEBUG(dbgs() << ".. .. gave up (UnsupportedFPMode\n"); 1714 return false; 1715 } 1716 1717 MVT DestVT = VA.getValVT(); 1718 // Special handling for extended integers. 1719 if (RVVT != DestVT) { 1720 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16) 1721 return false; 1722 1723 if (Outs[0].Flags.isZExt() || Outs[0].Flags.isSExt()) { 1724 bool IsZExt = Outs[0].Flags.isZExt(); 1725 SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt); 1726 if (SrcReg == 0) 1727 return false; 1728 } 1729 } 1730 1731 // Make the copy. 1732 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1733 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg); 1734 1735 // Add register to return instruction. 1736 RetRegs.push_back(VA.getLocReg()); 1737 } 1738 MachineInstrBuilder MIB = emitInst(Mips::RetRA); 1739 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i) 1740 MIB.addReg(RetRegs[i], RegState::Implicit); 1741 return true; 1742 } 1743 1744 bool MipsFastISel::selectTrunc(const Instruction *I) { 1745 // The high bits for a type smaller than the register size are assumed to be 1746 // undefined. 1747 Value *Op = I->getOperand(0); 1748 1749 EVT SrcVT, DestVT; 1750 SrcVT = TLI.getValueType(DL, Op->getType(), true); 1751 DestVT = TLI.getValueType(DL, I->getType(), true); 1752 1753 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8) 1754 return false; 1755 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1) 1756 return false; 1757 1758 unsigned SrcReg = getRegForValue(Op); 1759 if (!SrcReg) 1760 return false; 1761 1762 // Because the high bits are undefined, a truncate doesn't generate 1763 // any code. 1764 updateValueMap(I, SrcReg); 1765 return true; 1766 } 1767 1768 bool MipsFastISel::selectIntExt(const Instruction *I) { 1769 Type *DestTy = I->getType(); 1770 Value *Src = I->getOperand(0); 1771 Type *SrcTy = Src->getType(); 1772 1773 bool isZExt = isa<ZExtInst>(I); 1774 unsigned SrcReg = getRegForValue(Src); 1775 if (!SrcReg) 1776 return false; 1777 1778 EVT SrcEVT, DestEVT; 1779 SrcEVT = TLI.getValueType(DL, SrcTy, true); 1780 DestEVT = TLI.getValueType(DL, DestTy, true); 1781 if (!SrcEVT.isSimple()) 1782 return false; 1783 if (!DestEVT.isSimple()) 1784 return false; 1785 1786 MVT SrcVT = SrcEVT.getSimpleVT(); 1787 MVT DestVT = DestEVT.getSimpleVT(); 1788 unsigned ResultReg = createResultReg(&Mips::GPR32RegClass); 1789 1790 if (!emitIntExt(SrcVT, SrcReg, DestVT, ResultReg, isZExt)) 1791 return false; 1792 updateValueMap(I, ResultReg); 1793 return true; 1794 } 1795 1796 bool MipsFastISel::emitIntSExt32r1(MVT SrcVT, unsigned SrcReg, MVT DestVT, 1797 unsigned DestReg) { 1798 unsigned ShiftAmt; 1799 switch (SrcVT.SimpleTy) { 1800 default: 1801 return false; 1802 case MVT::i8: 1803 ShiftAmt = 24; 1804 break; 1805 case MVT::i16: 1806 ShiftAmt = 16; 1807 break; 1808 } 1809 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 1810 emitInst(Mips::SLL, TempReg).addReg(SrcReg).addImm(ShiftAmt); 1811 emitInst(Mips::SRA, DestReg).addReg(TempReg).addImm(ShiftAmt); 1812 return true; 1813 } 1814 1815 bool MipsFastISel::emitIntSExt32r2(MVT SrcVT, unsigned SrcReg, MVT DestVT, 1816 unsigned DestReg) { 1817 switch (SrcVT.SimpleTy) { 1818 default: 1819 return false; 1820 case MVT::i8: 1821 emitInst(Mips::SEB, DestReg).addReg(SrcReg); 1822 break; 1823 case MVT::i16: 1824 emitInst(Mips::SEH, DestReg).addReg(SrcReg); 1825 break; 1826 } 1827 return true; 1828 } 1829 1830 bool MipsFastISel::emitIntSExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, 1831 unsigned DestReg) { 1832 if ((DestVT != MVT::i32) && (DestVT != MVT::i16)) 1833 return false; 1834 if (Subtarget->hasMips32r2()) 1835 return emitIntSExt32r2(SrcVT, SrcReg, DestVT, DestReg); 1836 return emitIntSExt32r1(SrcVT, SrcReg, DestVT, DestReg); 1837 } 1838 1839 bool MipsFastISel::emitIntZExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, 1840 unsigned DestReg) { 1841 int64_t Imm; 1842 1843 switch (SrcVT.SimpleTy) { 1844 default: 1845 return false; 1846 case MVT::i1: 1847 Imm = 1; 1848 break; 1849 case MVT::i8: 1850 Imm = 0xff; 1851 break; 1852 case MVT::i16: 1853 Imm = 0xffff; 1854 break; 1855 } 1856 1857 emitInst(Mips::ANDi, DestReg).addReg(SrcReg).addImm(Imm); 1858 return true; 1859 } 1860 1861 bool MipsFastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, 1862 unsigned DestReg, bool IsZExt) { 1863 // FastISel does not have plumbing to deal with extensions where the SrcVT or 1864 // DestVT are odd things, so test to make sure that they are both types we can 1865 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise 1866 // bail out to SelectionDAG. 1867 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) && (DestVT != MVT::i32)) || 1868 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) && (SrcVT != MVT::i16))) 1869 return false; 1870 if (IsZExt) 1871 return emitIntZExt(SrcVT, SrcReg, DestVT, DestReg); 1872 return emitIntSExt(SrcVT, SrcReg, DestVT, DestReg); 1873 } 1874 1875 unsigned MipsFastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, 1876 bool isZExt) { 1877 unsigned DestReg = createResultReg(&Mips::GPR32RegClass); 1878 bool Success = emitIntExt(SrcVT, SrcReg, DestVT, DestReg, isZExt); 1879 return Success ? DestReg : 0; 1880 } 1881 1882 bool MipsFastISel::selectDivRem(const Instruction *I, unsigned ISDOpcode) { 1883 EVT DestEVT = TLI.getValueType(DL, I->getType(), true); 1884 if (!DestEVT.isSimple()) 1885 return false; 1886 1887 MVT DestVT = DestEVT.getSimpleVT(); 1888 if (DestVT != MVT::i32) 1889 return false; 1890 1891 unsigned DivOpc; 1892 switch (ISDOpcode) { 1893 default: 1894 return false; 1895 case ISD::SDIV: 1896 case ISD::SREM: 1897 DivOpc = Mips::SDIV; 1898 break; 1899 case ISD::UDIV: 1900 case ISD::UREM: 1901 DivOpc = Mips::UDIV; 1902 break; 1903 } 1904 1905 unsigned Src0Reg = getRegForValue(I->getOperand(0)); 1906 unsigned Src1Reg = getRegForValue(I->getOperand(1)); 1907 if (!Src0Reg || !Src1Reg) 1908 return false; 1909 1910 emitInst(DivOpc).addReg(Src0Reg).addReg(Src1Reg); 1911 emitInst(Mips::TEQ).addReg(Src1Reg).addReg(Mips::ZERO).addImm(7); 1912 1913 unsigned ResultReg = createResultReg(&Mips::GPR32RegClass); 1914 if (!ResultReg) 1915 return false; 1916 1917 unsigned MFOpc = (ISDOpcode == ISD::SREM || ISDOpcode == ISD::UREM) 1918 ? Mips::MFHI 1919 : Mips::MFLO; 1920 emitInst(MFOpc, ResultReg); 1921 1922 updateValueMap(I, ResultReg); 1923 return true; 1924 } 1925 1926 bool MipsFastISel::selectShift(const Instruction *I) { 1927 MVT RetVT; 1928 1929 if (!isTypeSupported(I->getType(), RetVT)) 1930 return false; 1931 1932 unsigned ResultReg = createResultReg(&Mips::GPR32RegClass); 1933 if (!ResultReg) 1934 return false; 1935 1936 unsigned Opcode = I->getOpcode(); 1937 const Value *Op0 = I->getOperand(0); 1938 unsigned Op0Reg = getRegForValue(Op0); 1939 if (!Op0Reg) 1940 return false; 1941 1942 // If AShr or LShr, then we need to make sure the operand0 is sign extended. 1943 if (Opcode == Instruction::AShr || Opcode == Instruction::LShr) { 1944 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 1945 if (!TempReg) 1946 return false; 1947 1948 MVT Op0MVT = TLI.getValueType(DL, Op0->getType(), true).getSimpleVT(); 1949 bool IsZExt = Opcode == Instruction::LShr; 1950 if (!emitIntExt(Op0MVT, Op0Reg, MVT::i32, TempReg, IsZExt)) 1951 return false; 1952 1953 Op0Reg = TempReg; 1954 } 1955 1956 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) { 1957 uint64_t ShiftVal = C->getZExtValue(); 1958 1959 switch (Opcode) { 1960 default: 1961 llvm_unreachable("Unexpected instruction."); 1962 case Instruction::Shl: 1963 Opcode = Mips::SLL; 1964 break; 1965 case Instruction::AShr: 1966 Opcode = Mips::SRA; 1967 break; 1968 case Instruction::LShr: 1969 Opcode = Mips::SRL; 1970 break; 1971 } 1972 1973 emitInst(Opcode, ResultReg).addReg(Op0Reg).addImm(ShiftVal); 1974 updateValueMap(I, ResultReg); 1975 return true; 1976 } 1977 1978 unsigned Op1Reg = getRegForValue(I->getOperand(1)); 1979 if (!Op1Reg) 1980 return false; 1981 1982 switch (Opcode) { 1983 default: 1984 llvm_unreachable("Unexpected instruction."); 1985 case Instruction::Shl: 1986 Opcode = Mips::SLLV; 1987 break; 1988 case Instruction::AShr: 1989 Opcode = Mips::SRAV; 1990 break; 1991 case Instruction::LShr: 1992 Opcode = Mips::SRLV; 1993 break; 1994 } 1995 1996 emitInst(Opcode, ResultReg).addReg(Op0Reg).addReg(Op1Reg); 1997 updateValueMap(I, ResultReg); 1998 return true; 1999 } 2000 2001 bool MipsFastISel::fastSelectInstruction(const Instruction *I) { 2002 switch (I->getOpcode()) { 2003 default: 2004 break; 2005 case Instruction::Load: 2006 return selectLoad(I); 2007 case Instruction::Store: 2008 return selectStore(I); 2009 case Instruction::SDiv: 2010 if (!selectBinaryOp(I, ISD::SDIV)) 2011 return selectDivRem(I, ISD::SDIV); 2012 return true; 2013 case Instruction::UDiv: 2014 if (!selectBinaryOp(I, ISD::UDIV)) 2015 return selectDivRem(I, ISD::UDIV); 2016 return true; 2017 case Instruction::SRem: 2018 if (!selectBinaryOp(I, ISD::SREM)) 2019 return selectDivRem(I, ISD::SREM); 2020 return true; 2021 case Instruction::URem: 2022 if (!selectBinaryOp(I, ISD::UREM)) 2023 return selectDivRem(I, ISD::UREM); 2024 return true; 2025 case Instruction::Shl: 2026 case Instruction::LShr: 2027 case Instruction::AShr: 2028 return selectShift(I); 2029 case Instruction::And: 2030 case Instruction::Or: 2031 case Instruction::Xor: 2032 return selectLogicalOp(I); 2033 case Instruction::Br: 2034 return selectBranch(I); 2035 case Instruction::Ret: 2036 return selectRet(I); 2037 case Instruction::Trunc: 2038 return selectTrunc(I); 2039 case Instruction::ZExt: 2040 case Instruction::SExt: 2041 return selectIntExt(I); 2042 case Instruction::FPTrunc: 2043 return selectFPTrunc(I); 2044 case Instruction::FPExt: 2045 return selectFPExt(I); 2046 case Instruction::FPToSI: 2047 return selectFPToInt(I, /*isSigned*/ true); 2048 case Instruction::FPToUI: 2049 return selectFPToInt(I, /*isSigned*/ false); 2050 case Instruction::ICmp: 2051 case Instruction::FCmp: 2052 return selectCmp(I); 2053 case Instruction::Select: 2054 return selectSelect(I); 2055 } 2056 return false; 2057 } 2058 2059 unsigned MipsFastISel::getRegEnsuringSimpleIntegerWidening(const Value *V, 2060 bool IsUnsigned) { 2061 unsigned VReg = getRegForValue(V); 2062 if (VReg == 0) 2063 return 0; 2064 MVT VMVT = TLI.getValueType(DL, V->getType(), true).getSimpleVT(); 2065 2066 if (VMVT == MVT::i1) 2067 return 0; 2068 2069 if ((VMVT == MVT::i8) || (VMVT == MVT::i16)) { 2070 unsigned TempReg = createResultReg(&Mips::GPR32RegClass); 2071 if (!emitIntExt(VMVT, VReg, MVT::i32, TempReg, IsUnsigned)) 2072 return 0; 2073 VReg = TempReg; 2074 } 2075 return VReg; 2076 } 2077 2078 void MipsFastISel::simplifyAddress(Address &Addr) { 2079 if (!isInt<16>(Addr.getOffset())) { 2080 unsigned TempReg = 2081 materialize32BitInt(Addr.getOffset(), &Mips::GPR32RegClass); 2082 unsigned DestReg = createResultReg(&Mips::GPR32RegClass); 2083 emitInst(Mips::ADDu, DestReg).addReg(TempReg).addReg(Addr.getReg()); 2084 Addr.setReg(DestReg); 2085 Addr.setOffset(0); 2086 } 2087 } 2088 2089 unsigned MipsFastISel::fastEmitInst_rr(unsigned MachineInstOpcode, 2090 const TargetRegisterClass *RC, 2091 unsigned Op0, bool Op0IsKill, 2092 unsigned Op1, bool Op1IsKill) { 2093 // We treat the MUL instruction in a special way because it clobbers 2094 // the HI0 & LO0 registers. The TableGen definition of this instruction can 2095 // mark these registers only as implicitly defined. As a result, the 2096 // register allocator runs out of registers when this instruction is 2097 // followed by another instruction that defines the same registers too. 2098 // We can fix this by explicitly marking those registers as dead. 2099 if (MachineInstOpcode == Mips::MUL) { 2100 unsigned ResultReg = createResultReg(RC); 2101 const MCInstrDesc &II = TII.get(MachineInstOpcode); 2102 Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs()); 2103 Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1); 2104 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg) 2105 .addReg(Op0, getKillRegState(Op0IsKill)) 2106 .addReg(Op1, getKillRegState(Op1IsKill)) 2107 .addReg(Mips::HI0, RegState::ImplicitDefine | RegState::Dead) 2108 .addReg(Mips::LO0, RegState::ImplicitDefine | RegState::Dead); 2109 return ResultReg; 2110 } 2111 2112 return FastISel::fastEmitInst_rr(MachineInstOpcode, RC, Op0, Op0IsKill, Op1, 2113 Op1IsKill); 2114 } 2115 2116 namespace llvm { 2117 2118 FastISel *Mips::createFastISel(FunctionLoweringInfo &funcInfo, 2119 const TargetLibraryInfo *libInfo) { 2120 return new MipsFastISel(funcInfo, libInfo); 2121 } 2122 2123 } // end namespace llvm 2124