1 //===-- RISCVInstrInfo.cpp - RISCV Instruction Information ------*- C++ -*-===// 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 // 9 // This file contains the RISCV implementation of the TargetInstrInfo class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "RISCVInstrInfo.h" 14 #include "MCTargetDesc/RISCVMatInt.h" 15 #include "RISCV.h" 16 #include "RISCVMachineFunctionInfo.h" 17 #include "RISCVSubtarget.h" 18 #include "RISCVTargetMachine.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/Analysis/MemoryLocation.h" 22 #include "llvm/CodeGen/LiveVariables.h" 23 #include "llvm/CodeGen/MachineFunctionPass.h" 24 #include "llvm/CodeGen/MachineInstrBuilder.h" 25 #include "llvm/CodeGen/MachineRegisterInfo.h" 26 #include "llvm/CodeGen/RegisterScavenging.h" 27 #include "llvm/MC/MCInstBuilder.h" 28 #include "llvm/MC/TargetRegistry.h" 29 #include "llvm/Support/ErrorHandling.h" 30 31 using namespace llvm; 32 33 #define GEN_CHECK_COMPRESS_INSTR 34 #include "RISCVGenCompressInstEmitter.inc" 35 36 #define GET_INSTRINFO_CTOR_DTOR 37 #include "RISCVGenInstrInfo.inc" 38 39 namespace llvm { 40 namespace RISCVVPseudosTable { 41 42 using namespace RISCV; 43 44 #define GET_RISCVVPseudosTable_IMPL 45 #include "RISCVGenSearchableTables.inc" 46 47 } // namespace RISCVVPseudosTable 48 } // namespace llvm 49 50 RISCVInstrInfo::RISCVInstrInfo(RISCVSubtarget &STI) 51 : RISCVGenInstrInfo(RISCV::ADJCALLSTACKDOWN, RISCV::ADJCALLSTACKUP), 52 STI(STI) {} 53 54 MCInst RISCVInstrInfo::getNop() const { 55 if (STI.getFeatureBits()[RISCV::FeatureStdExtC]) 56 return MCInstBuilder(RISCV::C_NOP); 57 return MCInstBuilder(RISCV::ADDI) 58 .addReg(RISCV::X0) 59 .addReg(RISCV::X0) 60 .addImm(0); 61 } 62 63 unsigned RISCVInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 64 int &FrameIndex) const { 65 switch (MI.getOpcode()) { 66 default: 67 return 0; 68 case RISCV::LB: 69 case RISCV::LBU: 70 case RISCV::LH: 71 case RISCV::LHU: 72 case RISCV::FLH: 73 case RISCV::LW: 74 case RISCV::FLW: 75 case RISCV::LWU: 76 case RISCV::LD: 77 case RISCV::FLD: 78 break; 79 } 80 81 if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() && 82 MI.getOperand(2).getImm() == 0) { 83 FrameIndex = MI.getOperand(1).getIndex(); 84 return MI.getOperand(0).getReg(); 85 } 86 87 return 0; 88 } 89 90 unsigned RISCVInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 91 int &FrameIndex) const { 92 switch (MI.getOpcode()) { 93 default: 94 return 0; 95 case RISCV::SB: 96 case RISCV::SH: 97 case RISCV::SW: 98 case RISCV::FSH: 99 case RISCV::FSW: 100 case RISCV::SD: 101 case RISCV::FSD: 102 break; 103 } 104 105 if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() && 106 MI.getOperand(2).getImm() == 0) { 107 FrameIndex = MI.getOperand(1).getIndex(); 108 return MI.getOperand(0).getReg(); 109 } 110 111 return 0; 112 } 113 114 static bool forwardCopyWillClobberTuple(unsigned DstReg, unsigned SrcReg, 115 unsigned NumRegs) { 116 // We really want the positive remainder mod 32 here, that happens to be 117 // easily obtainable with a mask. 118 return ((DstReg - SrcReg) & 0x1f) < NumRegs; 119 } 120 121 void RISCVInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 122 MachineBasicBlock::iterator MBBI, 123 const DebugLoc &DL, MCRegister DstReg, 124 MCRegister SrcReg, bool KillSrc) const { 125 if (RISCV::GPRRegClass.contains(DstReg, SrcReg)) { 126 BuildMI(MBB, MBBI, DL, get(RISCV::ADDI), DstReg) 127 .addReg(SrcReg, getKillRegState(KillSrc)) 128 .addImm(0); 129 return; 130 } 131 132 // FPR->FPR copies and VR->VR copies. 133 unsigned Opc; 134 bool IsScalableVector = true; 135 unsigned NF = 1; 136 unsigned LMul = 1; 137 unsigned SubRegIdx = RISCV::sub_vrm1_0; 138 if (RISCV::FPR16RegClass.contains(DstReg, SrcReg)) { 139 Opc = RISCV::FSGNJ_H; 140 IsScalableVector = false; 141 } else if (RISCV::FPR32RegClass.contains(DstReg, SrcReg)) { 142 Opc = RISCV::FSGNJ_S; 143 IsScalableVector = false; 144 } else if (RISCV::FPR64RegClass.contains(DstReg, SrcReg)) { 145 Opc = RISCV::FSGNJ_D; 146 IsScalableVector = false; 147 } else if (RISCV::VRRegClass.contains(DstReg, SrcReg)) { 148 Opc = RISCV::PseudoVMV1R_V; 149 } else if (RISCV::VRM2RegClass.contains(DstReg, SrcReg)) { 150 Opc = RISCV::PseudoVMV2R_V; 151 } else if (RISCV::VRM4RegClass.contains(DstReg, SrcReg)) { 152 Opc = RISCV::PseudoVMV4R_V; 153 } else if (RISCV::VRM8RegClass.contains(DstReg, SrcReg)) { 154 Opc = RISCV::PseudoVMV8R_V; 155 } else if (RISCV::VRN2M1RegClass.contains(DstReg, SrcReg)) { 156 Opc = RISCV::PseudoVMV1R_V; 157 SubRegIdx = RISCV::sub_vrm1_0; 158 NF = 2; 159 LMul = 1; 160 } else if (RISCV::VRN2M2RegClass.contains(DstReg, SrcReg)) { 161 Opc = RISCV::PseudoVMV2R_V; 162 SubRegIdx = RISCV::sub_vrm2_0; 163 NF = 2; 164 LMul = 2; 165 } else if (RISCV::VRN2M4RegClass.contains(DstReg, SrcReg)) { 166 Opc = RISCV::PseudoVMV4R_V; 167 SubRegIdx = RISCV::sub_vrm4_0; 168 NF = 2; 169 LMul = 4; 170 } else if (RISCV::VRN3M1RegClass.contains(DstReg, SrcReg)) { 171 Opc = RISCV::PseudoVMV1R_V; 172 SubRegIdx = RISCV::sub_vrm1_0; 173 NF = 3; 174 LMul = 1; 175 } else if (RISCV::VRN3M2RegClass.contains(DstReg, SrcReg)) { 176 Opc = RISCV::PseudoVMV2R_V; 177 SubRegIdx = RISCV::sub_vrm2_0; 178 NF = 3; 179 LMul = 2; 180 } else if (RISCV::VRN4M1RegClass.contains(DstReg, SrcReg)) { 181 Opc = RISCV::PseudoVMV1R_V; 182 SubRegIdx = RISCV::sub_vrm1_0; 183 NF = 4; 184 LMul = 1; 185 } else if (RISCV::VRN4M2RegClass.contains(DstReg, SrcReg)) { 186 Opc = RISCV::PseudoVMV2R_V; 187 SubRegIdx = RISCV::sub_vrm2_0; 188 NF = 4; 189 LMul = 2; 190 } else if (RISCV::VRN5M1RegClass.contains(DstReg, SrcReg)) { 191 Opc = RISCV::PseudoVMV1R_V; 192 SubRegIdx = RISCV::sub_vrm1_0; 193 NF = 5; 194 LMul = 1; 195 } else if (RISCV::VRN6M1RegClass.contains(DstReg, SrcReg)) { 196 Opc = RISCV::PseudoVMV1R_V; 197 SubRegIdx = RISCV::sub_vrm1_0; 198 NF = 6; 199 LMul = 1; 200 } else if (RISCV::VRN7M1RegClass.contains(DstReg, SrcReg)) { 201 Opc = RISCV::PseudoVMV1R_V; 202 SubRegIdx = RISCV::sub_vrm1_0; 203 NF = 7; 204 LMul = 1; 205 } else if (RISCV::VRN8M1RegClass.contains(DstReg, SrcReg)) { 206 Opc = RISCV::PseudoVMV1R_V; 207 SubRegIdx = RISCV::sub_vrm1_0; 208 NF = 8; 209 LMul = 1; 210 } else { 211 llvm_unreachable("Impossible reg-to-reg copy"); 212 } 213 214 if (IsScalableVector) { 215 if (NF == 1) { 216 BuildMI(MBB, MBBI, DL, get(Opc), DstReg) 217 .addReg(SrcReg, getKillRegState(KillSrc)); 218 } else { 219 const TargetRegisterInfo *TRI = STI.getRegisterInfo(); 220 221 int I = 0, End = NF, Incr = 1; 222 unsigned SrcEncoding = TRI->getEncodingValue(SrcReg); 223 unsigned DstEncoding = TRI->getEncodingValue(DstReg); 224 if (forwardCopyWillClobberTuple(DstEncoding, SrcEncoding, NF * LMul)) { 225 I = NF - 1; 226 End = -1; 227 Incr = -1; 228 } 229 230 for (; I != End; I += Incr) { 231 BuildMI(MBB, MBBI, DL, get(Opc), TRI->getSubReg(DstReg, SubRegIdx + I)) 232 .addReg(TRI->getSubReg(SrcReg, SubRegIdx + I), 233 getKillRegState(KillSrc)); 234 } 235 } 236 } else { 237 BuildMI(MBB, MBBI, DL, get(Opc), DstReg) 238 .addReg(SrcReg, getKillRegState(KillSrc)) 239 .addReg(SrcReg, getKillRegState(KillSrc)); 240 } 241 } 242 243 void RISCVInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 244 MachineBasicBlock::iterator I, 245 Register SrcReg, bool IsKill, int FI, 246 const TargetRegisterClass *RC, 247 const TargetRegisterInfo *TRI) const { 248 DebugLoc DL; 249 if (I != MBB.end()) 250 DL = I->getDebugLoc(); 251 252 MachineFunction *MF = MBB.getParent(); 253 MachineFrameInfo &MFI = MF->getFrameInfo(); 254 255 unsigned Opcode; 256 bool IsScalableVector = true; 257 bool IsZvlsseg = true; 258 if (RISCV::GPRRegClass.hasSubClassEq(RC)) { 259 Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ? 260 RISCV::SW : RISCV::SD; 261 IsScalableVector = false; 262 } else if (RISCV::FPR16RegClass.hasSubClassEq(RC)) { 263 Opcode = RISCV::FSH; 264 IsScalableVector = false; 265 } else if (RISCV::FPR32RegClass.hasSubClassEq(RC)) { 266 Opcode = RISCV::FSW; 267 IsScalableVector = false; 268 } else if (RISCV::FPR64RegClass.hasSubClassEq(RC)) { 269 Opcode = RISCV::FSD; 270 IsScalableVector = false; 271 } else if (RISCV::VRRegClass.hasSubClassEq(RC)) { 272 Opcode = RISCV::PseudoVSPILL_M1; 273 IsZvlsseg = false; 274 } else if (RISCV::VRM2RegClass.hasSubClassEq(RC)) { 275 Opcode = RISCV::PseudoVSPILL_M2; 276 IsZvlsseg = false; 277 } else if (RISCV::VRM4RegClass.hasSubClassEq(RC)) { 278 Opcode = RISCV::PseudoVSPILL_M4; 279 IsZvlsseg = false; 280 } else if (RISCV::VRM8RegClass.hasSubClassEq(RC)) { 281 Opcode = RISCV::PseudoVSPILL_M8; 282 IsZvlsseg = false; 283 } else if (RISCV::VRN2M1RegClass.hasSubClassEq(RC)) 284 Opcode = RISCV::PseudoVSPILL2_M1; 285 else if (RISCV::VRN2M2RegClass.hasSubClassEq(RC)) 286 Opcode = RISCV::PseudoVSPILL2_M2; 287 else if (RISCV::VRN2M4RegClass.hasSubClassEq(RC)) 288 Opcode = RISCV::PseudoVSPILL2_M4; 289 else if (RISCV::VRN3M1RegClass.hasSubClassEq(RC)) 290 Opcode = RISCV::PseudoVSPILL3_M1; 291 else if (RISCV::VRN3M2RegClass.hasSubClassEq(RC)) 292 Opcode = RISCV::PseudoVSPILL3_M2; 293 else if (RISCV::VRN4M1RegClass.hasSubClassEq(RC)) 294 Opcode = RISCV::PseudoVSPILL4_M1; 295 else if (RISCV::VRN4M2RegClass.hasSubClassEq(RC)) 296 Opcode = RISCV::PseudoVSPILL4_M2; 297 else if (RISCV::VRN5M1RegClass.hasSubClassEq(RC)) 298 Opcode = RISCV::PseudoVSPILL5_M1; 299 else if (RISCV::VRN6M1RegClass.hasSubClassEq(RC)) 300 Opcode = RISCV::PseudoVSPILL6_M1; 301 else if (RISCV::VRN7M1RegClass.hasSubClassEq(RC)) 302 Opcode = RISCV::PseudoVSPILL7_M1; 303 else if (RISCV::VRN8M1RegClass.hasSubClassEq(RC)) 304 Opcode = RISCV::PseudoVSPILL8_M1; 305 else 306 llvm_unreachable("Can't store this register to stack slot"); 307 308 if (IsScalableVector) { 309 MachineMemOperand *MMO = MF->getMachineMemOperand( 310 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore, 311 MemoryLocation::UnknownSize, MFI.getObjectAlign(FI)); 312 313 MFI.setStackID(FI, TargetStackID::ScalableVector); 314 auto MIB = BuildMI(MBB, I, DL, get(Opcode)) 315 .addReg(SrcReg, getKillRegState(IsKill)) 316 .addFrameIndex(FI) 317 .addMemOperand(MMO); 318 if (IsZvlsseg) { 319 // For spilling/reloading Zvlsseg registers, append the dummy field for 320 // the scaled vector length. The argument will be used when expanding 321 // these pseudo instructions. 322 MIB.addReg(RISCV::X0); 323 } 324 } else { 325 MachineMemOperand *MMO = MF->getMachineMemOperand( 326 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore, 327 MFI.getObjectSize(FI), MFI.getObjectAlign(FI)); 328 329 BuildMI(MBB, I, DL, get(Opcode)) 330 .addReg(SrcReg, getKillRegState(IsKill)) 331 .addFrameIndex(FI) 332 .addImm(0) 333 .addMemOperand(MMO); 334 } 335 } 336 337 void RISCVInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 338 MachineBasicBlock::iterator I, 339 Register DstReg, int FI, 340 const TargetRegisterClass *RC, 341 const TargetRegisterInfo *TRI) const { 342 DebugLoc DL; 343 if (I != MBB.end()) 344 DL = I->getDebugLoc(); 345 346 MachineFunction *MF = MBB.getParent(); 347 MachineFrameInfo &MFI = MF->getFrameInfo(); 348 349 unsigned Opcode; 350 bool IsScalableVector = true; 351 bool IsZvlsseg = true; 352 if (RISCV::GPRRegClass.hasSubClassEq(RC)) { 353 Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ? 354 RISCV::LW : RISCV::LD; 355 IsScalableVector = false; 356 } else if (RISCV::FPR16RegClass.hasSubClassEq(RC)) { 357 Opcode = RISCV::FLH; 358 IsScalableVector = false; 359 } else if (RISCV::FPR32RegClass.hasSubClassEq(RC)) { 360 Opcode = RISCV::FLW; 361 IsScalableVector = false; 362 } else if (RISCV::FPR64RegClass.hasSubClassEq(RC)) { 363 Opcode = RISCV::FLD; 364 IsScalableVector = false; 365 } else if (RISCV::VRRegClass.hasSubClassEq(RC)) { 366 Opcode = RISCV::PseudoVRELOAD_M1; 367 IsZvlsseg = false; 368 } else if (RISCV::VRM2RegClass.hasSubClassEq(RC)) { 369 Opcode = RISCV::PseudoVRELOAD_M2; 370 IsZvlsseg = false; 371 } else if (RISCV::VRM4RegClass.hasSubClassEq(RC)) { 372 Opcode = RISCV::PseudoVRELOAD_M4; 373 IsZvlsseg = false; 374 } else if (RISCV::VRM8RegClass.hasSubClassEq(RC)) { 375 Opcode = RISCV::PseudoVRELOAD_M8; 376 IsZvlsseg = false; 377 } else if (RISCV::VRN2M1RegClass.hasSubClassEq(RC)) 378 Opcode = RISCV::PseudoVRELOAD2_M1; 379 else if (RISCV::VRN2M2RegClass.hasSubClassEq(RC)) 380 Opcode = RISCV::PseudoVRELOAD2_M2; 381 else if (RISCV::VRN2M4RegClass.hasSubClassEq(RC)) 382 Opcode = RISCV::PseudoVRELOAD2_M4; 383 else if (RISCV::VRN3M1RegClass.hasSubClassEq(RC)) 384 Opcode = RISCV::PseudoVRELOAD3_M1; 385 else if (RISCV::VRN3M2RegClass.hasSubClassEq(RC)) 386 Opcode = RISCV::PseudoVRELOAD3_M2; 387 else if (RISCV::VRN4M1RegClass.hasSubClassEq(RC)) 388 Opcode = RISCV::PseudoVRELOAD4_M1; 389 else if (RISCV::VRN4M2RegClass.hasSubClassEq(RC)) 390 Opcode = RISCV::PseudoVRELOAD4_M2; 391 else if (RISCV::VRN5M1RegClass.hasSubClassEq(RC)) 392 Opcode = RISCV::PseudoVRELOAD5_M1; 393 else if (RISCV::VRN6M1RegClass.hasSubClassEq(RC)) 394 Opcode = RISCV::PseudoVRELOAD6_M1; 395 else if (RISCV::VRN7M1RegClass.hasSubClassEq(RC)) 396 Opcode = RISCV::PseudoVRELOAD7_M1; 397 else if (RISCV::VRN8M1RegClass.hasSubClassEq(RC)) 398 Opcode = RISCV::PseudoVRELOAD8_M1; 399 else 400 llvm_unreachable("Can't load this register from stack slot"); 401 402 if (IsScalableVector) { 403 MachineMemOperand *MMO = MF->getMachineMemOperand( 404 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad, 405 MemoryLocation::UnknownSize, MFI.getObjectAlign(FI)); 406 407 MFI.setStackID(FI, TargetStackID::ScalableVector); 408 auto MIB = BuildMI(MBB, I, DL, get(Opcode), DstReg) 409 .addFrameIndex(FI) 410 .addMemOperand(MMO); 411 if (IsZvlsseg) { 412 // For spilling/reloading Zvlsseg registers, append the dummy field for 413 // the scaled vector length. The argument will be used when expanding 414 // these pseudo instructions. 415 MIB.addReg(RISCV::X0); 416 } 417 } else { 418 MachineMemOperand *MMO = MF->getMachineMemOperand( 419 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad, 420 MFI.getObjectSize(FI), MFI.getObjectAlign(FI)); 421 422 BuildMI(MBB, I, DL, get(Opcode), DstReg) 423 .addFrameIndex(FI) 424 .addImm(0) 425 .addMemOperand(MMO); 426 } 427 } 428 429 void RISCVInstrInfo::movImm(MachineBasicBlock &MBB, 430 MachineBasicBlock::iterator MBBI, 431 const DebugLoc &DL, Register DstReg, uint64_t Val, 432 MachineInstr::MIFlag Flag) const { 433 MachineFunction *MF = MBB.getParent(); 434 MachineRegisterInfo &MRI = MF->getRegInfo(); 435 Register SrcReg = RISCV::X0; 436 Register Result = MRI.createVirtualRegister(&RISCV::GPRRegClass); 437 unsigned Num = 0; 438 439 if (!STI.is64Bit() && !isInt<32>(Val)) 440 report_fatal_error("Should only materialize 32-bit constants for RV32"); 441 442 RISCVMatInt::InstSeq Seq = 443 RISCVMatInt::generateInstSeq(Val, STI.getFeatureBits()); 444 assert(!Seq.empty()); 445 446 for (RISCVMatInt::Inst &Inst : Seq) { 447 // Write the final result to DstReg if it's the last instruction in the Seq. 448 // Otherwise, write the result to the temp register. 449 if (++Num == Seq.size()) 450 Result = DstReg; 451 452 if (Inst.Opc == RISCV::LUI) { 453 BuildMI(MBB, MBBI, DL, get(RISCV::LUI), Result) 454 .addImm(Inst.Imm) 455 .setMIFlag(Flag); 456 } else if (Inst.Opc == RISCV::ADDUW) { 457 BuildMI(MBB, MBBI, DL, get(RISCV::ADDUW), Result) 458 .addReg(SrcReg, RegState::Kill) 459 .addReg(RISCV::X0) 460 .setMIFlag(Flag); 461 } else if (Inst.Opc == RISCV::SH1ADD || Inst.Opc == RISCV::SH2ADD || 462 Inst.Opc == RISCV::SH3ADD) { 463 BuildMI(MBB, MBBI, DL, get(Inst.Opc), Result) 464 .addReg(SrcReg, RegState::Kill) 465 .addReg(SrcReg, RegState::Kill) 466 .setMIFlag(Flag); 467 } else { 468 BuildMI(MBB, MBBI, DL, get(Inst.Opc), Result) 469 .addReg(SrcReg, RegState::Kill) 470 .addImm(Inst.Imm) 471 .setMIFlag(Flag); 472 } 473 // Only the first instruction has X0 as its source. 474 SrcReg = Result; 475 } 476 } 477 478 static RISCVCC::CondCode getCondFromBranchOpc(unsigned Opc) { 479 switch (Opc) { 480 default: 481 return RISCVCC::COND_INVALID; 482 case RISCV::BEQ: 483 return RISCVCC::COND_EQ; 484 case RISCV::BNE: 485 return RISCVCC::COND_NE; 486 case RISCV::BLT: 487 return RISCVCC::COND_LT; 488 case RISCV::BGE: 489 return RISCVCC::COND_GE; 490 case RISCV::BLTU: 491 return RISCVCC::COND_LTU; 492 case RISCV::BGEU: 493 return RISCVCC::COND_GEU; 494 } 495 } 496 497 // The contents of values added to Cond are not examined outside of 498 // RISCVInstrInfo, giving us flexibility in what to push to it. For RISCV, we 499 // push BranchOpcode, Reg1, Reg2. 500 static void parseCondBranch(MachineInstr &LastInst, MachineBasicBlock *&Target, 501 SmallVectorImpl<MachineOperand> &Cond) { 502 // Block ends with fall-through condbranch. 503 assert(LastInst.getDesc().isConditionalBranch() && 504 "Unknown conditional branch"); 505 Target = LastInst.getOperand(2).getMBB(); 506 unsigned CC = getCondFromBranchOpc(LastInst.getOpcode()); 507 Cond.push_back(MachineOperand::CreateImm(CC)); 508 Cond.push_back(LastInst.getOperand(0)); 509 Cond.push_back(LastInst.getOperand(1)); 510 } 511 512 const MCInstrDesc &RISCVInstrInfo::getBrCond(RISCVCC::CondCode CC) const { 513 switch (CC) { 514 default: 515 llvm_unreachable("Unknown condition code!"); 516 case RISCVCC::COND_EQ: 517 return get(RISCV::BEQ); 518 case RISCVCC::COND_NE: 519 return get(RISCV::BNE); 520 case RISCVCC::COND_LT: 521 return get(RISCV::BLT); 522 case RISCVCC::COND_GE: 523 return get(RISCV::BGE); 524 case RISCVCC::COND_LTU: 525 return get(RISCV::BLTU); 526 case RISCVCC::COND_GEU: 527 return get(RISCV::BGEU); 528 } 529 } 530 531 RISCVCC::CondCode RISCVCC::getOppositeBranchCondition(RISCVCC::CondCode CC) { 532 switch (CC) { 533 default: 534 llvm_unreachable("Unrecognized conditional branch"); 535 case RISCVCC::COND_EQ: 536 return RISCVCC::COND_NE; 537 case RISCVCC::COND_NE: 538 return RISCVCC::COND_EQ; 539 case RISCVCC::COND_LT: 540 return RISCVCC::COND_GE; 541 case RISCVCC::COND_GE: 542 return RISCVCC::COND_LT; 543 case RISCVCC::COND_LTU: 544 return RISCVCC::COND_GEU; 545 case RISCVCC::COND_GEU: 546 return RISCVCC::COND_LTU; 547 } 548 } 549 550 bool RISCVInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 551 MachineBasicBlock *&TBB, 552 MachineBasicBlock *&FBB, 553 SmallVectorImpl<MachineOperand> &Cond, 554 bool AllowModify) const { 555 TBB = FBB = nullptr; 556 Cond.clear(); 557 558 // If the block has no terminators, it just falls into the block after it. 559 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 560 if (I == MBB.end() || !isUnpredicatedTerminator(*I)) 561 return false; 562 563 // Count the number of terminators and find the first unconditional or 564 // indirect branch. 565 MachineBasicBlock::iterator FirstUncondOrIndirectBr = MBB.end(); 566 int NumTerminators = 0; 567 for (auto J = I.getReverse(); J != MBB.rend() && isUnpredicatedTerminator(*J); 568 J++) { 569 NumTerminators++; 570 if (J->getDesc().isUnconditionalBranch() || 571 J->getDesc().isIndirectBranch()) { 572 FirstUncondOrIndirectBr = J.getReverse(); 573 } 574 } 575 576 // If AllowModify is true, we can erase any terminators after 577 // FirstUncondOrIndirectBR. 578 if (AllowModify && FirstUncondOrIndirectBr != MBB.end()) { 579 while (std::next(FirstUncondOrIndirectBr) != MBB.end()) { 580 std::next(FirstUncondOrIndirectBr)->eraseFromParent(); 581 NumTerminators--; 582 } 583 I = FirstUncondOrIndirectBr; 584 } 585 586 // We can't handle blocks that end in an indirect branch. 587 if (I->getDesc().isIndirectBranch()) 588 return true; 589 590 // We can't handle blocks with more than 2 terminators. 591 if (NumTerminators > 2) 592 return true; 593 594 // Handle a single unconditional branch. 595 if (NumTerminators == 1 && I->getDesc().isUnconditionalBranch()) { 596 TBB = getBranchDestBlock(*I); 597 return false; 598 } 599 600 // Handle a single conditional branch. 601 if (NumTerminators == 1 && I->getDesc().isConditionalBranch()) { 602 parseCondBranch(*I, TBB, Cond); 603 return false; 604 } 605 606 // Handle a conditional branch followed by an unconditional branch. 607 if (NumTerminators == 2 && std::prev(I)->getDesc().isConditionalBranch() && 608 I->getDesc().isUnconditionalBranch()) { 609 parseCondBranch(*std::prev(I), TBB, Cond); 610 FBB = getBranchDestBlock(*I); 611 return false; 612 } 613 614 // Otherwise, we can't handle this. 615 return true; 616 } 617 618 unsigned RISCVInstrInfo::removeBranch(MachineBasicBlock &MBB, 619 int *BytesRemoved) const { 620 if (BytesRemoved) 621 *BytesRemoved = 0; 622 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 623 if (I == MBB.end()) 624 return 0; 625 626 if (!I->getDesc().isUnconditionalBranch() && 627 !I->getDesc().isConditionalBranch()) 628 return 0; 629 630 // Remove the branch. 631 if (BytesRemoved) 632 *BytesRemoved += getInstSizeInBytes(*I); 633 I->eraseFromParent(); 634 635 I = MBB.end(); 636 637 if (I == MBB.begin()) 638 return 1; 639 --I; 640 if (!I->getDesc().isConditionalBranch()) 641 return 1; 642 643 // Remove the branch. 644 if (BytesRemoved) 645 *BytesRemoved += getInstSizeInBytes(*I); 646 I->eraseFromParent(); 647 return 2; 648 } 649 650 // Inserts a branch into the end of the specific MachineBasicBlock, returning 651 // the number of instructions inserted. 652 unsigned RISCVInstrInfo::insertBranch( 653 MachineBasicBlock &MBB, MachineBasicBlock *TBB, MachineBasicBlock *FBB, 654 ArrayRef<MachineOperand> Cond, const DebugLoc &DL, int *BytesAdded) const { 655 if (BytesAdded) 656 *BytesAdded = 0; 657 658 // Shouldn't be a fall through. 659 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 660 assert((Cond.size() == 3 || Cond.size() == 0) && 661 "RISCV branch conditions have two components!"); 662 663 // Unconditional branch. 664 if (Cond.empty()) { 665 MachineInstr &MI = *BuildMI(&MBB, DL, get(RISCV::PseudoBR)).addMBB(TBB); 666 if (BytesAdded) 667 *BytesAdded += getInstSizeInBytes(MI); 668 return 1; 669 } 670 671 // Either a one or two-way conditional branch. 672 auto CC = static_cast<RISCVCC::CondCode>(Cond[0].getImm()); 673 MachineInstr &CondMI = 674 *BuildMI(&MBB, DL, getBrCond(CC)).add(Cond[1]).add(Cond[2]).addMBB(TBB); 675 if (BytesAdded) 676 *BytesAdded += getInstSizeInBytes(CondMI); 677 678 // One-way conditional branch. 679 if (!FBB) 680 return 1; 681 682 // Two-way conditional branch. 683 MachineInstr &MI = *BuildMI(&MBB, DL, get(RISCV::PseudoBR)).addMBB(FBB); 684 if (BytesAdded) 685 *BytesAdded += getInstSizeInBytes(MI); 686 return 2; 687 } 688 689 unsigned RISCVInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB, 690 MachineBasicBlock &DestBB, 691 const DebugLoc &DL, 692 int64_t BrOffset, 693 RegScavenger *RS) const { 694 assert(RS && "RegScavenger required for long branching"); 695 assert(MBB.empty() && 696 "new block should be inserted for expanding unconditional branch"); 697 assert(MBB.pred_size() == 1); 698 699 MachineFunction *MF = MBB.getParent(); 700 MachineRegisterInfo &MRI = MF->getRegInfo(); 701 702 if (!isInt<32>(BrOffset)) 703 report_fatal_error( 704 "Branch offsets outside of the signed 32-bit range not supported"); 705 706 // FIXME: A virtual register must be used initially, as the register 707 // scavenger won't work with empty blocks (SIInstrInfo::insertIndirectBranch 708 // uses the same workaround). 709 Register ScratchReg = MRI.createVirtualRegister(&RISCV::GPRRegClass); 710 auto II = MBB.end(); 711 712 MachineInstr &MI = *BuildMI(MBB, II, DL, get(RISCV::PseudoJump)) 713 .addReg(ScratchReg, RegState::Define | RegState::Dead) 714 .addMBB(&DestBB, RISCVII::MO_CALL); 715 716 RS->enterBasicBlockEnd(MBB); 717 unsigned Scav = RS->scavengeRegisterBackwards(RISCV::GPRRegClass, 718 MI.getIterator(), false, 0); 719 MRI.replaceRegWith(ScratchReg, Scav); 720 MRI.clearVirtRegs(); 721 RS->setRegUsed(Scav); 722 return 8; 723 } 724 725 bool RISCVInstrInfo::reverseBranchCondition( 726 SmallVectorImpl<MachineOperand> &Cond) const { 727 assert((Cond.size() == 3) && "Invalid branch condition!"); 728 auto CC = static_cast<RISCVCC::CondCode>(Cond[0].getImm()); 729 Cond[0].setImm(getOppositeBranchCondition(CC)); 730 return false; 731 } 732 733 MachineBasicBlock * 734 RISCVInstrInfo::getBranchDestBlock(const MachineInstr &MI) const { 735 assert(MI.getDesc().isBranch() && "Unexpected opcode!"); 736 // The branch target is always the last operand. 737 int NumOp = MI.getNumExplicitOperands(); 738 return MI.getOperand(NumOp - 1).getMBB(); 739 } 740 741 bool RISCVInstrInfo::isBranchOffsetInRange(unsigned BranchOp, 742 int64_t BrOffset) const { 743 unsigned XLen = STI.getXLen(); 744 // Ideally we could determine the supported branch offset from the 745 // RISCVII::FormMask, but this can't be used for Pseudo instructions like 746 // PseudoBR. 747 switch (BranchOp) { 748 default: 749 llvm_unreachable("Unexpected opcode!"); 750 case RISCV::BEQ: 751 case RISCV::BNE: 752 case RISCV::BLT: 753 case RISCV::BGE: 754 case RISCV::BLTU: 755 case RISCV::BGEU: 756 return isIntN(13, BrOffset); 757 case RISCV::JAL: 758 case RISCV::PseudoBR: 759 return isIntN(21, BrOffset); 760 case RISCV::PseudoJump: 761 return isIntN(32, SignExtend64(BrOffset + 0x800, XLen)); 762 } 763 } 764 765 unsigned RISCVInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 766 unsigned Opcode = MI.getOpcode(); 767 768 switch (Opcode) { 769 default: { 770 if (MI.getParent() && MI.getParent()->getParent()) { 771 const auto MF = MI.getMF(); 772 const auto &TM = static_cast<const RISCVTargetMachine &>(MF->getTarget()); 773 const MCRegisterInfo &MRI = *TM.getMCRegisterInfo(); 774 const MCSubtargetInfo &STI = *TM.getMCSubtargetInfo(); 775 const RISCVSubtarget &ST = MF->getSubtarget<RISCVSubtarget>(); 776 if (isCompressibleInst(MI, &ST, MRI, STI)) 777 return 2; 778 } 779 return get(Opcode).getSize(); 780 } 781 case TargetOpcode::EH_LABEL: 782 case TargetOpcode::IMPLICIT_DEF: 783 case TargetOpcode::KILL: 784 case TargetOpcode::DBG_VALUE: 785 return 0; 786 // These values are determined based on RISCVExpandAtomicPseudoInsts, 787 // RISCVExpandPseudoInsts and RISCVMCCodeEmitter, depending on where the 788 // pseudos are expanded. 789 case RISCV::PseudoCALLReg: 790 case RISCV::PseudoCALL: 791 case RISCV::PseudoJump: 792 case RISCV::PseudoTAIL: 793 case RISCV::PseudoLLA: 794 case RISCV::PseudoLA: 795 case RISCV::PseudoLA_TLS_IE: 796 case RISCV::PseudoLA_TLS_GD: 797 return 8; 798 case RISCV::PseudoAtomicLoadNand32: 799 case RISCV::PseudoAtomicLoadNand64: 800 return 20; 801 case RISCV::PseudoMaskedAtomicSwap32: 802 case RISCV::PseudoMaskedAtomicLoadAdd32: 803 case RISCV::PseudoMaskedAtomicLoadSub32: 804 return 28; 805 case RISCV::PseudoMaskedAtomicLoadNand32: 806 return 32; 807 case RISCV::PseudoMaskedAtomicLoadMax32: 808 case RISCV::PseudoMaskedAtomicLoadMin32: 809 return 44; 810 case RISCV::PseudoMaskedAtomicLoadUMax32: 811 case RISCV::PseudoMaskedAtomicLoadUMin32: 812 return 36; 813 case RISCV::PseudoCmpXchg32: 814 case RISCV::PseudoCmpXchg64: 815 return 16; 816 case RISCV::PseudoMaskedCmpXchg32: 817 return 32; 818 case TargetOpcode::INLINEASM: 819 case TargetOpcode::INLINEASM_BR: { 820 const MachineFunction &MF = *MI.getParent()->getParent(); 821 const auto &TM = static_cast<const RISCVTargetMachine &>(MF.getTarget()); 822 return getInlineAsmLength(MI.getOperand(0).getSymbolName(), 823 *TM.getMCAsmInfo()); 824 } 825 case RISCV::PseudoVSPILL2_M1: 826 case RISCV::PseudoVSPILL2_M2: 827 case RISCV::PseudoVSPILL2_M4: 828 case RISCV::PseudoVSPILL3_M1: 829 case RISCV::PseudoVSPILL3_M2: 830 case RISCV::PseudoVSPILL4_M1: 831 case RISCV::PseudoVSPILL4_M2: 832 case RISCV::PseudoVSPILL5_M1: 833 case RISCV::PseudoVSPILL6_M1: 834 case RISCV::PseudoVSPILL7_M1: 835 case RISCV::PseudoVSPILL8_M1: 836 case RISCV::PseudoVRELOAD2_M1: 837 case RISCV::PseudoVRELOAD2_M2: 838 case RISCV::PseudoVRELOAD2_M4: 839 case RISCV::PseudoVRELOAD3_M1: 840 case RISCV::PseudoVRELOAD3_M2: 841 case RISCV::PseudoVRELOAD4_M1: 842 case RISCV::PseudoVRELOAD4_M2: 843 case RISCV::PseudoVRELOAD5_M1: 844 case RISCV::PseudoVRELOAD6_M1: 845 case RISCV::PseudoVRELOAD7_M1: 846 case RISCV::PseudoVRELOAD8_M1: { 847 // The values are determined based on expandVSPILL and expandVRELOAD that 848 // expand the pseudos depending on NF. 849 unsigned NF = isRVVSpillForZvlsseg(Opcode)->first; 850 return 4 * (2 * NF - 1); 851 } 852 } 853 } 854 855 bool RISCVInstrInfo::isAsCheapAsAMove(const MachineInstr &MI) const { 856 const unsigned Opcode = MI.getOpcode(); 857 switch (Opcode) { 858 default: 859 break; 860 case RISCV::FSGNJ_D: 861 case RISCV::FSGNJ_S: 862 // The canonical floating-point move is fsgnj rd, rs, rs. 863 return MI.getOperand(1).isReg() && MI.getOperand(2).isReg() && 864 MI.getOperand(1).getReg() == MI.getOperand(2).getReg(); 865 case RISCV::ADDI: 866 case RISCV::ORI: 867 case RISCV::XORI: 868 return (MI.getOperand(1).isReg() && 869 MI.getOperand(1).getReg() == RISCV::X0) || 870 (MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0); 871 } 872 return MI.isAsCheapAsAMove(); 873 } 874 875 Optional<DestSourcePair> 876 RISCVInstrInfo::isCopyInstrImpl(const MachineInstr &MI) const { 877 if (MI.isMoveReg()) 878 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 879 switch (MI.getOpcode()) { 880 default: 881 break; 882 case RISCV::ADDI: 883 // Operand 1 can be a frameindex but callers expect registers 884 if (MI.getOperand(1).isReg() && MI.getOperand(2).isImm() && 885 MI.getOperand(2).getImm() == 0) 886 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 887 break; 888 case RISCV::FSGNJ_D: 889 case RISCV::FSGNJ_S: 890 // The canonical floating-point move is fsgnj rd, rs, rs. 891 if (MI.getOperand(1).isReg() && MI.getOperand(2).isReg() && 892 MI.getOperand(1).getReg() == MI.getOperand(2).getReg()) 893 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 894 break; 895 } 896 return None; 897 } 898 899 bool RISCVInstrInfo::verifyInstruction(const MachineInstr &MI, 900 StringRef &ErrInfo) const { 901 const MCInstrInfo *MCII = STI.getInstrInfo(); 902 MCInstrDesc const &Desc = MCII->get(MI.getOpcode()); 903 904 for (auto &OI : enumerate(Desc.operands())) { 905 unsigned OpType = OI.value().OperandType; 906 if (OpType >= RISCVOp::OPERAND_FIRST_RISCV_IMM && 907 OpType <= RISCVOp::OPERAND_LAST_RISCV_IMM) { 908 const MachineOperand &MO = MI.getOperand(OI.index()); 909 if (MO.isImm()) { 910 int64_t Imm = MO.getImm(); 911 bool Ok; 912 switch (OpType) { 913 default: 914 llvm_unreachable("Unexpected operand type"); 915 case RISCVOp::OPERAND_UIMM2: 916 Ok = isUInt<2>(Imm); 917 break; 918 case RISCVOp::OPERAND_UIMM3: 919 Ok = isUInt<3>(Imm); 920 break; 921 case RISCVOp::OPERAND_UIMM4: 922 Ok = isUInt<4>(Imm); 923 break; 924 case RISCVOp::OPERAND_UIMM5: 925 Ok = isUInt<5>(Imm); 926 break; 927 case RISCVOp::OPERAND_UIMM7: 928 Ok = isUInt<7>(Imm); 929 break; 930 case RISCVOp::OPERAND_UIMM12: 931 Ok = isUInt<12>(Imm); 932 break; 933 case RISCVOp::OPERAND_SIMM12: 934 Ok = isInt<12>(Imm); 935 break; 936 case RISCVOp::OPERAND_UIMM20: 937 Ok = isUInt<20>(Imm); 938 break; 939 case RISCVOp::OPERAND_UIMMLOG2XLEN: 940 if (STI.getTargetTriple().isArch64Bit()) 941 Ok = isUInt<6>(Imm); 942 else 943 Ok = isUInt<5>(Imm); 944 break; 945 } 946 if (!Ok) { 947 ErrInfo = "Invalid immediate"; 948 return false; 949 } 950 } 951 } 952 } 953 954 return true; 955 } 956 957 // Return true if get the base operand, byte offset of an instruction and the 958 // memory width. Width is the size of memory that is being loaded/stored. 959 bool RISCVInstrInfo::getMemOperandWithOffsetWidth( 960 const MachineInstr &LdSt, const MachineOperand *&BaseReg, int64_t &Offset, 961 unsigned &Width, const TargetRegisterInfo *TRI) const { 962 if (!LdSt.mayLoadOrStore()) 963 return false; 964 965 // Here we assume the standard RISC-V ISA, which uses a base+offset 966 // addressing mode. You'll need to relax these conditions to support custom 967 // load/stores instructions. 968 if (LdSt.getNumExplicitOperands() != 3) 969 return false; 970 if (!LdSt.getOperand(1).isReg() || !LdSt.getOperand(2).isImm()) 971 return false; 972 973 if (!LdSt.hasOneMemOperand()) 974 return false; 975 976 Width = (*LdSt.memoperands_begin())->getSize(); 977 BaseReg = &LdSt.getOperand(1); 978 Offset = LdSt.getOperand(2).getImm(); 979 return true; 980 } 981 982 bool RISCVInstrInfo::areMemAccessesTriviallyDisjoint( 983 const MachineInstr &MIa, const MachineInstr &MIb) const { 984 assert(MIa.mayLoadOrStore() && "MIa must be a load or store."); 985 assert(MIb.mayLoadOrStore() && "MIb must be a load or store."); 986 987 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() || 988 MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef()) 989 return false; 990 991 // Retrieve the base register, offset from the base register and width. Width 992 // is the size of memory that is being loaded/stored (e.g. 1, 2, 4). If 993 // base registers are identical, and the offset of a lower memory access + 994 // the width doesn't overlap the offset of a higher memory access, 995 // then the memory accesses are different. 996 const TargetRegisterInfo *TRI = STI.getRegisterInfo(); 997 const MachineOperand *BaseOpA = nullptr, *BaseOpB = nullptr; 998 int64_t OffsetA = 0, OffsetB = 0; 999 unsigned int WidthA = 0, WidthB = 0; 1000 if (getMemOperandWithOffsetWidth(MIa, BaseOpA, OffsetA, WidthA, TRI) && 1001 getMemOperandWithOffsetWidth(MIb, BaseOpB, OffsetB, WidthB, TRI)) { 1002 if (BaseOpA->isIdenticalTo(*BaseOpB)) { 1003 int LowOffset = std::min(OffsetA, OffsetB); 1004 int HighOffset = std::max(OffsetA, OffsetB); 1005 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1006 if (LowOffset + LowWidth <= HighOffset) 1007 return true; 1008 } 1009 } 1010 return false; 1011 } 1012 1013 std::pair<unsigned, unsigned> 1014 RISCVInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { 1015 const unsigned Mask = RISCVII::MO_DIRECT_FLAG_MASK; 1016 return std::make_pair(TF & Mask, TF & ~Mask); 1017 } 1018 1019 ArrayRef<std::pair<unsigned, const char *>> 1020 RISCVInstrInfo::getSerializableDirectMachineOperandTargetFlags() const { 1021 using namespace RISCVII; 1022 static const std::pair<unsigned, const char *> TargetFlags[] = { 1023 {MO_CALL, "riscv-call"}, 1024 {MO_PLT, "riscv-plt"}, 1025 {MO_LO, "riscv-lo"}, 1026 {MO_HI, "riscv-hi"}, 1027 {MO_PCREL_LO, "riscv-pcrel-lo"}, 1028 {MO_PCREL_HI, "riscv-pcrel-hi"}, 1029 {MO_GOT_HI, "riscv-got-hi"}, 1030 {MO_TPREL_LO, "riscv-tprel-lo"}, 1031 {MO_TPREL_HI, "riscv-tprel-hi"}, 1032 {MO_TPREL_ADD, "riscv-tprel-add"}, 1033 {MO_TLS_GOT_HI, "riscv-tls-got-hi"}, 1034 {MO_TLS_GD_HI, "riscv-tls-gd-hi"}}; 1035 return makeArrayRef(TargetFlags); 1036 } 1037 bool RISCVInstrInfo::isFunctionSafeToOutlineFrom( 1038 MachineFunction &MF, bool OutlineFromLinkOnceODRs) const { 1039 const Function &F = MF.getFunction(); 1040 1041 // Can F be deduplicated by the linker? If it can, don't outline from it. 1042 if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage()) 1043 return false; 1044 1045 // Don't outline from functions with section markings; the program could 1046 // expect that all the code is in the named section. 1047 if (F.hasSection()) 1048 return false; 1049 1050 // It's safe to outline from MF. 1051 return true; 1052 } 1053 1054 bool RISCVInstrInfo::isMBBSafeToOutlineFrom(MachineBasicBlock &MBB, 1055 unsigned &Flags) const { 1056 // More accurate safety checking is done in getOutliningCandidateInfo. 1057 return true; 1058 } 1059 1060 // Enum values indicating how an outlined call should be constructed. 1061 enum MachineOutlinerConstructionID { 1062 MachineOutlinerDefault 1063 }; 1064 1065 outliner::OutlinedFunction RISCVInstrInfo::getOutliningCandidateInfo( 1066 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const { 1067 1068 // First we need to filter out candidates where the X5 register (IE t0) can't 1069 // be used to setup the function call. 1070 auto CannotInsertCall = [](outliner::Candidate &C) { 1071 const TargetRegisterInfo *TRI = C.getMF()->getSubtarget().getRegisterInfo(); 1072 1073 C.initLRU(*TRI); 1074 LiveRegUnits LRU = C.LRU; 1075 return !LRU.available(RISCV::X5); 1076 }; 1077 1078 llvm::erase_if(RepeatedSequenceLocs, CannotInsertCall); 1079 1080 // If the sequence doesn't have enough candidates left, then we're done. 1081 if (RepeatedSequenceLocs.size() < 2) 1082 return outliner::OutlinedFunction(); 1083 1084 unsigned SequenceSize = 0; 1085 1086 auto I = RepeatedSequenceLocs[0].front(); 1087 auto E = std::next(RepeatedSequenceLocs[0].back()); 1088 for (; I != E; ++I) 1089 SequenceSize += getInstSizeInBytes(*I); 1090 1091 // call t0, function = 8 bytes. 1092 unsigned CallOverhead = 8; 1093 for (auto &C : RepeatedSequenceLocs) 1094 C.setCallInfo(MachineOutlinerDefault, CallOverhead); 1095 1096 // jr t0 = 4 bytes, 2 bytes if compressed instructions are enabled. 1097 unsigned FrameOverhead = 4; 1098 if (RepeatedSequenceLocs[0].getMF()->getSubtarget() 1099 .getFeatureBits()[RISCV::FeatureStdExtC]) 1100 FrameOverhead = 2; 1101 1102 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1103 FrameOverhead, MachineOutlinerDefault); 1104 } 1105 1106 outliner::InstrType 1107 RISCVInstrInfo::getOutliningType(MachineBasicBlock::iterator &MBBI, 1108 unsigned Flags) const { 1109 MachineInstr &MI = *MBBI; 1110 MachineBasicBlock *MBB = MI.getParent(); 1111 const TargetRegisterInfo *TRI = 1112 MBB->getParent()->getSubtarget().getRegisterInfo(); 1113 1114 // Positions generally can't safely be outlined. 1115 if (MI.isPosition()) { 1116 // We can manually strip out CFI instructions later. 1117 if (MI.isCFIInstruction()) 1118 return outliner::InstrType::Invisible; 1119 1120 return outliner::InstrType::Illegal; 1121 } 1122 1123 // Don't trust the user to write safe inline assembly. 1124 if (MI.isInlineAsm()) 1125 return outliner::InstrType::Illegal; 1126 1127 // We can't outline branches to other basic blocks. 1128 if (MI.isTerminator() && !MBB->succ_empty()) 1129 return outliner::InstrType::Illegal; 1130 1131 // We need support for tail calls to outlined functions before return 1132 // statements can be allowed. 1133 if (MI.isReturn()) 1134 return outliner::InstrType::Illegal; 1135 1136 // Don't allow modifying the X5 register which we use for return addresses for 1137 // these outlined functions. 1138 if (MI.modifiesRegister(RISCV::X5, TRI) || 1139 MI.getDesc().hasImplicitDefOfPhysReg(RISCV::X5)) 1140 return outliner::InstrType::Illegal; 1141 1142 // Make sure the operands don't reference something unsafe. 1143 for (const auto &MO : MI.operands()) 1144 if (MO.isMBB() || MO.isBlockAddress() || MO.isCPI() || MO.isJTI()) 1145 return outliner::InstrType::Illegal; 1146 1147 // Don't allow instructions which won't be materialized to impact outlining 1148 // analysis. 1149 if (MI.isMetaInstruction()) 1150 return outliner::InstrType::Invisible; 1151 1152 return outliner::InstrType::Legal; 1153 } 1154 1155 void RISCVInstrInfo::buildOutlinedFrame( 1156 MachineBasicBlock &MBB, MachineFunction &MF, 1157 const outliner::OutlinedFunction &OF) const { 1158 1159 // Strip out any CFI instructions 1160 bool Changed = true; 1161 while (Changed) { 1162 Changed = false; 1163 auto I = MBB.begin(); 1164 auto E = MBB.end(); 1165 for (; I != E; ++I) { 1166 if (I->isCFIInstruction()) { 1167 I->removeFromParent(); 1168 Changed = true; 1169 break; 1170 } 1171 } 1172 } 1173 1174 MBB.addLiveIn(RISCV::X5); 1175 1176 // Add in a return instruction to the end of the outlined frame. 1177 MBB.insert(MBB.end(), BuildMI(MF, DebugLoc(), get(RISCV::JALR)) 1178 .addReg(RISCV::X0, RegState::Define) 1179 .addReg(RISCV::X5) 1180 .addImm(0)); 1181 } 1182 1183 MachineBasicBlock::iterator RISCVInstrInfo::insertOutlinedCall( 1184 Module &M, MachineBasicBlock &MBB, MachineBasicBlock::iterator &It, 1185 MachineFunction &MF, const outliner::Candidate &C) const { 1186 1187 // Add in a call instruction to the outlined function at the given location. 1188 It = MBB.insert(It, 1189 BuildMI(MF, DebugLoc(), get(RISCV::PseudoCALLReg), RISCV::X5) 1190 .addGlobalAddress(M.getNamedValue(MF.getName()), 0, 1191 RISCVII::MO_CALL)); 1192 return It; 1193 } 1194 1195 // clang-format off 1196 #define CASE_VFMA_OPCODE_COMMON(OP, TYPE, LMUL) \ 1197 RISCV::PseudoV##OP##_##TYPE##_##LMUL 1198 1199 #define CASE_VFMA_OPCODE_LMULS(OP, TYPE) \ 1200 CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF8): \ 1201 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF4): \ 1202 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF2): \ 1203 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M1): \ 1204 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M2): \ 1205 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M4): \ 1206 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M8) 1207 1208 #define CASE_VFMA_SPLATS(OP) \ 1209 CASE_VFMA_OPCODE_LMULS(OP, VF16): \ 1210 case CASE_VFMA_OPCODE_LMULS(OP, VF32): \ 1211 case CASE_VFMA_OPCODE_LMULS(OP, VF64) 1212 // clang-format on 1213 1214 bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI, 1215 unsigned &SrcOpIdx1, 1216 unsigned &SrcOpIdx2) const { 1217 const MCInstrDesc &Desc = MI.getDesc(); 1218 if (!Desc.isCommutable()) 1219 return false; 1220 1221 switch (MI.getOpcode()) { 1222 case CASE_VFMA_SPLATS(FMADD): 1223 case CASE_VFMA_SPLATS(FMSUB): 1224 case CASE_VFMA_SPLATS(FMACC): 1225 case CASE_VFMA_SPLATS(FMSAC): 1226 case CASE_VFMA_SPLATS(FNMADD): 1227 case CASE_VFMA_SPLATS(FNMSUB): 1228 case CASE_VFMA_SPLATS(FNMACC): 1229 case CASE_VFMA_SPLATS(FNMSAC): 1230 case CASE_VFMA_OPCODE_LMULS(FMACC, VV): 1231 case CASE_VFMA_OPCODE_LMULS(FMSAC, VV): 1232 case CASE_VFMA_OPCODE_LMULS(FNMACC, VV): 1233 case CASE_VFMA_OPCODE_LMULS(FNMSAC, VV): 1234 case CASE_VFMA_OPCODE_LMULS(MADD, VX): 1235 case CASE_VFMA_OPCODE_LMULS(NMSUB, VX): 1236 case CASE_VFMA_OPCODE_LMULS(MACC, VX): 1237 case CASE_VFMA_OPCODE_LMULS(NMSAC, VX): 1238 case CASE_VFMA_OPCODE_LMULS(MACC, VV): 1239 case CASE_VFMA_OPCODE_LMULS(NMSAC, VV): { 1240 // If the tail policy is undisturbed we can't commute. 1241 assert(RISCVII::hasVecPolicyOp(MI.getDesc().TSFlags)); 1242 if ((MI.getOperand(MI.getNumExplicitOperands() - 1).getImm() & 1) == 0) 1243 return false; 1244 1245 // For these instructions we can only swap operand 1 and operand 3 by 1246 // changing the opcode. 1247 unsigned CommutableOpIdx1 = 1; 1248 unsigned CommutableOpIdx2 = 3; 1249 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1, 1250 CommutableOpIdx2)) 1251 return false; 1252 return true; 1253 } 1254 case CASE_VFMA_OPCODE_LMULS(FMADD, VV): 1255 case CASE_VFMA_OPCODE_LMULS(FMSUB, VV): 1256 case CASE_VFMA_OPCODE_LMULS(FNMADD, VV): 1257 case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): 1258 case CASE_VFMA_OPCODE_LMULS(MADD, VV): 1259 case CASE_VFMA_OPCODE_LMULS(NMSUB, VV): { 1260 // If the tail policy is undisturbed we can't commute. 1261 assert(RISCVII::hasVecPolicyOp(MI.getDesc().TSFlags)); 1262 if ((MI.getOperand(MI.getNumExplicitOperands() - 1).getImm() & 1) == 0) 1263 return false; 1264 1265 // For these instructions we have more freedom. We can commute with the 1266 // other multiplicand or with the addend/subtrahend/minuend. 1267 1268 // Any fixed operand must be from source 1, 2 or 3. 1269 if (SrcOpIdx1 != CommuteAnyOperandIndex && SrcOpIdx1 > 3) 1270 return false; 1271 if (SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx2 > 3) 1272 return false; 1273 1274 // It both ops are fixed one must be the tied source. 1275 if (SrcOpIdx1 != CommuteAnyOperandIndex && 1276 SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx1 != 1 && SrcOpIdx2 != 1) 1277 return false; 1278 1279 // Look for two different register operands assumed to be commutable 1280 // regardless of the FMA opcode. The FMA opcode is adjusted later if 1281 // needed. 1282 if (SrcOpIdx1 == CommuteAnyOperandIndex || 1283 SrcOpIdx2 == CommuteAnyOperandIndex) { 1284 // At least one of operands to be commuted is not specified and 1285 // this method is free to choose appropriate commutable operands. 1286 unsigned CommutableOpIdx1 = SrcOpIdx1; 1287 if (SrcOpIdx1 == SrcOpIdx2) { 1288 // Both of operands are not fixed. Set one of commutable 1289 // operands to the tied source. 1290 CommutableOpIdx1 = 1; 1291 } else if (SrcOpIdx1 == CommuteAnyOperandIndex) { 1292 // Only one of the operands is not fixed. 1293 CommutableOpIdx1 = SrcOpIdx2; 1294 } 1295 1296 // CommutableOpIdx1 is well defined now. Let's choose another commutable 1297 // operand and assign its index to CommutableOpIdx2. 1298 unsigned CommutableOpIdx2; 1299 if (CommutableOpIdx1 != 1) { 1300 // If we haven't already used the tied source, we must use it now. 1301 CommutableOpIdx2 = 1; 1302 } else { 1303 Register Op1Reg = MI.getOperand(CommutableOpIdx1).getReg(); 1304 1305 // The commuted operands should have different registers. 1306 // Otherwise, the commute transformation does not change anything and 1307 // is useless. We use this as a hint to make our decision. 1308 if (Op1Reg != MI.getOperand(2).getReg()) 1309 CommutableOpIdx2 = 2; 1310 else 1311 CommutableOpIdx2 = 3; 1312 } 1313 1314 // Assign the found pair of commutable indices to SrcOpIdx1 and 1315 // SrcOpIdx2 to return those values. 1316 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1, 1317 CommutableOpIdx2)) 1318 return false; 1319 } 1320 1321 return true; 1322 } 1323 } 1324 1325 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 1326 } 1327 1328 #define CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, LMUL) \ 1329 case RISCV::PseudoV##OLDOP##_##TYPE##_##LMUL: \ 1330 Opc = RISCV::PseudoV##NEWOP##_##TYPE##_##LMUL; \ 1331 break; 1332 1333 #define CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, TYPE) \ 1334 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF8) \ 1335 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF4) \ 1336 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF2) \ 1337 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M1) \ 1338 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M2) \ 1339 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M4) \ 1340 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M8) 1341 1342 #define CASE_VFMA_CHANGE_OPCODE_SPLATS(OLDOP, NEWOP) \ 1343 CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF16) \ 1344 CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF32) \ 1345 CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF64) 1346 1347 MachineInstr *RISCVInstrInfo::commuteInstructionImpl(MachineInstr &MI, 1348 bool NewMI, 1349 unsigned OpIdx1, 1350 unsigned OpIdx2) const { 1351 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 1352 if (NewMI) 1353 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 1354 return MI; 1355 }; 1356 1357 switch (MI.getOpcode()) { 1358 case CASE_VFMA_SPLATS(FMACC): 1359 case CASE_VFMA_SPLATS(FMADD): 1360 case CASE_VFMA_SPLATS(FMSAC): 1361 case CASE_VFMA_SPLATS(FMSUB): 1362 case CASE_VFMA_SPLATS(FNMACC): 1363 case CASE_VFMA_SPLATS(FNMADD): 1364 case CASE_VFMA_SPLATS(FNMSAC): 1365 case CASE_VFMA_SPLATS(FNMSUB): 1366 case CASE_VFMA_OPCODE_LMULS(FMACC, VV): 1367 case CASE_VFMA_OPCODE_LMULS(FMSAC, VV): 1368 case CASE_VFMA_OPCODE_LMULS(FNMACC, VV): 1369 case CASE_VFMA_OPCODE_LMULS(FNMSAC, VV): 1370 case CASE_VFMA_OPCODE_LMULS(MADD, VX): 1371 case CASE_VFMA_OPCODE_LMULS(NMSUB, VX): 1372 case CASE_VFMA_OPCODE_LMULS(MACC, VX): 1373 case CASE_VFMA_OPCODE_LMULS(NMSAC, VX): 1374 case CASE_VFMA_OPCODE_LMULS(MACC, VV): 1375 case CASE_VFMA_OPCODE_LMULS(NMSAC, VV): { 1376 // It only make sense to toggle these between clobbering the 1377 // addend/subtrahend/minuend one of the multiplicands. 1378 assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index"); 1379 assert((OpIdx1 == 3 || OpIdx2 == 3) && "Unexpected opcode index"); 1380 unsigned Opc; 1381 switch (MI.getOpcode()) { 1382 default: 1383 llvm_unreachable("Unexpected opcode"); 1384 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMACC, FMADD) 1385 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMADD, FMACC) 1386 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSAC, FMSUB) 1387 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSUB, FMSAC) 1388 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMACC, FNMADD) 1389 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMADD, FNMACC) 1390 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSAC, FNMSUB) 1391 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSUB, FNMSAC) 1392 CASE_VFMA_CHANGE_OPCODE_LMULS(FMACC, FMADD, VV) 1393 CASE_VFMA_CHANGE_OPCODE_LMULS(FMSAC, FMSUB, VV) 1394 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMACC, FNMADD, VV) 1395 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMSAC, FNMSUB, VV) 1396 CASE_VFMA_CHANGE_OPCODE_LMULS(MACC, MADD, VX) 1397 CASE_VFMA_CHANGE_OPCODE_LMULS(MADD, MACC, VX) 1398 CASE_VFMA_CHANGE_OPCODE_LMULS(NMSAC, NMSUB, VX) 1399 CASE_VFMA_CHANGE_OPCODE_LMULS(NMSUB, NMSAC, VX) 1400 CASE_VFMA_CHANGE_OPCODE_LMULS(MACC, MADD, VV) 1401 CASE_VFMA_CHANGE_OPCODE_LMULS(NMSAC, NMSUB, VV) 1402 } 1403 1404 auto &WorkingMI = cloneIfNew(MI); 1405 WorkingMI.setDesc(get(Opc)); 1406 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1407 OpIdx1, OpIdx2); 1408 } 1409 case CASE_VFMA_OPCODE_LMULS(FMADD, VV): 1410 case CASE_VFMA_OPCODE_LMULS(FMSUB, VV): 1411 case CASE_VFMA_OPCODE_LMULS(FNMADD, VV): 1412 case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): 1413 case CASE_VFMA_OPCODE_LMULS(MADD, VV): 1414 case CASE_VFMA_OPCODE_LMULS(NMSUB, VV): { 1415 assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index"); 1416 // If one of the operands, is the addend we need to change opcode. 1417 // Otherwise we're just swapping 2 of the multiplicands. 1418 if (OpIdx1 == 3 || OpIdx2 == 3) { 1419 unsigned Opc; 1420 switch (MI.getOpcode()) { 1421 default: 1422 llvm_unreachable("Unexpected opcode"); 1423 CASE_VFMA_CHANGE_OPCODE_LMULS(FMADD, FMACC, VV) 1424 CASE_VFMA_CHANGE_OPCODE_LMULS(FMSUB, FMSAC, VV) 1425 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMADD, FNMACC, VV) 1426 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMSUB, FNMSAC, VV) 1427 CASE_VFMA_CHANGE_OPCODE_LMULS(MADD, MACC, VV) 1428 CASE_VFMA_CHANGE_OPCODE_LMULS(NMSUB, NMSAC, VV) 1429 } 1430 1431 auto &WorkingMI = cloneIfNew(MI); 1432 WorkingMI.setDesc(get(Opc)); 1433 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1434 OpIdx1, OpIdx2); 1435 } 1436 // Let the default code handle it. 1437 break; 1438 } 1439 } 1440 1441 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 1442 } 1443 1444 #undef CASE_VFMA_CHANGE_OPCODE_SPLATS 1445 #undef CASE_VFMA_CHANGE_OPCODE_LMULS 1446 #undef CASE_VFMA_CHANGE_OPCODE_COMMON 1447 #undef CASE_VFMA_SPLATS 1448 #undef CASE_VFMA_OPCODE_LMULS 1449 #undef CASE_VFMA_OPCODE_COMMON 1450 1451 // clang-format off 1452 #define CASE_WIDEOP_OPCODE_COMMON(OP, LMUL) \ 1453 RISCV::PseudoV##OP##_##LMUL##_TIED 1454 1455 #define CASE_WIDEOP_OPCODE_LMULS(OP) \ 1456 CASE_WIDEOP_OPCODE_COMMON(OP, MF8): \ 1457 case CASE_WIDEOP_OPCODE_COMMON(OP, MF4): \ 1458 case CASE_WIDEOP_OPCODE_COMMON(OP, MF2): \ 1459 case CASE_WIDEOP_OPCODE_COMMON(OP, M1): \ 1460 case CASE_WIDEOP_OPCODE_COMMON(OP, M2): \ 1461 case CASE_WIDEOP_OPCODE_COMMON(OP, M4) 1462 // clang-format on 1463 1464 #define CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, LMUL) \ 1465 case RISCV::PseudoV##OP##_##LMUL##_TIED: \ 1466 NewOpc = RISCV::PseudoV##OP##_##LMUL; \ 1467 break; 1468 1469 #define CASE_WIDEOP_CHANGE_OPCODE_LMULS(OP) \ 1470 CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF8) \ 1471 CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF4) \ 1472 CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF2) \ 1473 CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, M1) \ 1474 CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, M2) \ 1475 CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, M4) 1476 1477 MachineInstr *RISCVInstrInfo::convertToThreeAddress(MachineInstr &MI, 1478 LiveVariables *LV) const { 1479 switch (MI.getOpcode()) { 1480 default: 1481 break; 1482 case CASE_WIDEOP_OPCODE_LMULS(FWADD_WV): 1483 case CASE_WIDEOP_OPCODE_LMULS(FWSUB_WV): 1484 case CASE_WIDEOP_OPCODE_LMULS(WADD_WV): 1485 case CASE_WIDEOP_OPCODE_LMULS(WADDU_WV): 1486 case CASE_WIDEOP_OPCODE_LMULS(WSUB_WV): 1487 case CASE_WIDEOP_OPCODE_LMULS(WSUBU_WV): { 1488 // clang-format off 1489 unsigned NewOpc; 1490 switch (MI.getOpcode()) { 1491 default: 1492 llvm_unreachable("Unexpected opcode"); 1493 CASE_WIDEOP_CHANGE_OPCODE_LMULS(FWADD_WV) 1494 CASE_WIDEOP_CHANGE_OPCODE_LMULS(FWSUB_WV) 1495 CASE_WIDEOP_CHANGE_OPCODE_LMULS(WADD_WV) 1496 CASE_WIDEOP_CHANGE_OPCODE_LMULS(WADDU_WV) 1497 CASE_WIDEOP_CHANGE_OPCODE_LMULS(WSUB_WV) 1498 CASE_WIDEOP_CHANGE_OPCODE_LMULS(WSUBU_WV) 1499 } 1500 //clang-format on 1501 1502 MachineBasicBlock &MBB = *MI.getParent(); 1503 MachineInstrBuilder MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc)) 1504 .add(MI.getOperand(0)) 1505 .add(MI.getOperand(1)) 1506 .add(MI.getOperand(2)) 1507 .add(MI.getOperand(3)) 1508 .add(MI.getOperand(4)); 1509 MIB.copyImplicitOps(MI); 1510 1511 if (LV) { 1512 unsigned NumOps = MI.getNumOperands(); 1513 for (unsigned I = 1; I < NumOps; ++I) { 1514 MachineOperand &Op = MI.getOperand(I); 1515 if (Op.isReg() && Op.isKill()) 1516 LV->replaceKillInstruction(Op.getReg(), MI, *MIB); 1517 } 1518 } 1519 1520 return MIB; 1521 } 1522 } 1523 1524 return nullptr; 1525 } 1526 1527 #undef CASE_WIDEOP_CHANGE_OPCODE_LMULS 1528 #undef CASE_WIDEOP_CHANGE_OPCODE_COMMON 1529 #undef CASE_WIDEOP_OPCODE_LMULS 1530 #undef CASE_WIDEOP_OPCODE_COMMON 1531 1532 Register RISCVInstrInfo::getVLENFactoredAmount(MachineFunction &MF, 1533 MachineBasicBlock &MBB, 1534 MachineBasicBlock::iterator II, 1535 const DebugLoc &DL, 1536 int64_t Amount, 1537 MachineInstr::MIFlag Flag) const { 1538 assert(Amount > 0 && "There is no need to get VLEN scaled value."); 1539 assert(Amount % 8 == 0 && 1540 "Reserve the stack by the multiple of one vector size."); 1541 1542 MachineRegisterInfo &MRI = MF.getRegInfo(); 1543 const RISCVInstrInfo *TII = MF.getSubtarget<RISCVSubtarget>().getInstrInfo(); 1544 int64_t NumOfVReg = Amount / 8; 1545 1546 Register VL = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1547 BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), VL) 1548 .setMIFlag(Flag); 1549 assert(isInt<32>(NumOfVReg) && 1550 "Expect the number of vector registers within 32-bits."); 1551 if (isPowerOf2_32(NumOfVReg)) { 1552 uint32_t ShiftAmount = Log2_32(NumOfVReg); 1553 if (ShiftAmount == 0) 1554 return VL; 1555 BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), VL) 1556 .addReg(VL, RegState::Kill) 1557 .addImm(ShiftAmount) 1558 .setMIFlag(Flag); 1559 } else if (isPowerOf2_32(NumOfVReg - 1)) { 1560 Register ScaledRegister = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1561 uint32_t ShiftAmount = Log2_32(NumOfVReg - 1); 1562 BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), ScaledRegister) 1563 .addReg(VL) 1564 .addImm(ShiftAmount) 1565 .setMIFlag(Flag); 1566 BuildMI(MBB, II, DL, TII->get(RISCV::ADD), VL) 1567 .addReg(ScaledRegister, RegState::Kill) 1568 .addReg(VL, RegState::Kill) 1569 .setMIFlag(Flag); 1570 } else if (isPowerOf2_32(NumOfVReg + 1)) { 1571 Register ScaledRegister = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1572 uint32_t ShiftAmount = Log2_32(NumOfVReg + 1); 1573 BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), ScaledRegister) 1574 .addReg(VL) 1575 .addImm(ShiftAmount) 1576 .setMIFlag(Flag); 1577 BuildMI(MBB, II, DL, TII->get(RISCV::SUB), VL) 1578 .addReg(ScaledRegister, RegState::Kill) 1579 .addReg(VL, RegState::Kill) 1580 .setMIFlag(Flag); 1581 } else { 1582 Register N = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1583 if (!isInt<12>(NumOfVReg)) 1584 movImm(MBB, II, DL, N, NumOfVReg); 1585 else { 1586 BuildMI(MBB, II, DL, TII->get(RISCV::ADDI), N) 1587 .addReg(RISCV::X0) 1588 .addImm(NumOfVReg) 1589 .setMIFlag(Flag); 1590 } 1591 if (!MF.getSubtarget<RISCVSubtarget>().hasStdExtM()) 1592 MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{ 1593 MF.getFunction(), 1594 "M-extension must be enabled to calculate the vscaled size/offset."}); 1595 BuildMI(MBB, II, DL, TII->get(RISCV::MUL), VL) 1596 .addReg(VL, RegState::Kill) 1597 .addReg(N, RegState::Kill) 1598 .setMIFlag(Flag); 1599 } 1600 1601 return VL; 1602 } 1603 1604 static bool isRVVWholeLoadStore(unsigned Opcode) { 1605 switch (Opcode) { 1606 default: 1607 return false; 1608 case RISCV::VS1R_V: 1609 case RISCV::VS2R_V: 1610 case RISCV::VS4R_V: 1611 case RISCV::VS8R_V: 1612 case RISCV::VL1RE8_V: 1613 case RISCV::VL2RE8_V: 1614 case RISCV::VL4RE8_V: 1615 case RISCV::VL8RE8_V: 1616 case RISCV::VL1RE16_V: 1617 case RISCV::VL2RE16_V: 1618 case RISCV::VL4RE16_V: 1619 case RISCV::VL8RE16_V: 1620 case RISCV::VL1RE32_V: 1621 case RISCV::VL2RE32_V: 1622 case RISCV::VL4RE32_V: 1623 case RISCV::VL8RE32_V: 1624 case RISCV::VL1RE64_V: 1625 case RISCV::VL2RE64_V: 1626 case RISCV::VL4RE64_V: 1627 case RISCV::VL8RE64_V: 1628 return true; 1629 } 1630 } 1631 1632 bool RISCVInstrInfo::isRVVSpill(const MachineInstr &MI, bool CheckFIs) const { 1633 // RVV lacks any support for immediate addressing for stack addresses, so be 1634 // conservative. 1635 unsigned Opcode = MI.getOpcode(); 1636 if (!RISCVVPseudosTable::getPseudoInfo(Opcode) && 1637 !isRVVWholeLoadStore(Opcode) && !isRVVSpillForZvlsseg(Opcode)) 1638 return false; 1639 return !CheckFIs || any_of(MI.operands(), [](const MachineOperand &MO) { 1640 return MO.isFI(); 1641 }); 1642 } 1643 1644 Optional<std::pair<unsigned, unsigned>> 1645 RISCVInstrInfo::isRVVSpillForZvlsseg(unsigned Opcode) const { 1646 switch (Opcode) { 1647 default: 1648 return None; 1649 case RISCV::PseudoVSPILL2_M1: 1650 case RISCV::PseudoVRELOAD2_M1: 1651 return std::make_pair(2u, 1u); 1652 case RISCV::PseudoVSPILL2_M2: 1653 case RISCV::PseudoVRELOAD2_M2: 1654 return std::make_pair(2u, 2u); 1655 case RISCV::PseudoVSPILL2_M4: 1656 case RISCV::PseudoVRELOAD2_M4: 1657 return std::make_pair(2u, 4u); 1658 case RISCV::PseudoVSPILL3_M1: 1659 case RISCV::PseudoVRELOAD3_M1: 1660 return std::make_pair(3u, 1u); 1661 case RISCV::PseudoVSPILL3_M2: 1662 case RISCV::PseudoVRELOAD3_M2: 1663 return std::make_pair(3u, 2u); 1664 case RISCV::PseudoVSPILL4_M1: 1665 case RISCV::PseudoVRELOAD4_M1: 1666 return std::make_pair(4u, 1u); 1667 case RISCV::PseudoVSPILL4_M2: 1668 case RISCV::PseudoVRELOAD4_M2: 1669 return std::make_pair(4u, 2u); 1670 case RISCV::PseudoVSPILL5_M1: 1671 case RISCV::PseudoVRELOAD5_M1: 1672 return std::make_pair(5u, 1u); 1673 case RISCV::PseudoVSPILL6_M1: 1674 case RISCV::PseudoVRELOAD6_M1: 1675 return std::make_pair(6u, 1u); 1676 case RISCV::PseudoVSPILL7_M1: 1677 case RISCV::PseudoVRELOAD7_M1: 1678 return std::make_pair(7u, 1u); 1679 case RISCV::PseudoVSPILL8_M1: 1680 case RISCV::PseudoVRELOAD8_M1: 1681 return std::make_pair(8u, 1u); 1682 } 1683 } 1684