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/MachineFunctionPass.h" 23 #include "llvm/CodeGen/MachineInstrBuilder.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/CodeGen/RegisterScavenging.h" 26 #include "llvm/MC/MCInstBuilder.h" 27 #include "llvm/Support/ErrorHandling.h" 28 #include "llvm/Support/TargetRegistry.h" 29 30 using namespace llvm; 31 32 #define GEN_CHECK_COMPRESS_INSTR 33 #include "RISCVGenCompressInstEmitter.inc" 34 35 #define GET_INSTRINFO_CTOR_DTOR 36 #include "RISCVGenInstrInfo.inc" 37 38 namespace llvm { 39 namespace RISCVVPseudosTable { 40 41 using namespace RISCV; 42 43 #define GET_RISCVVPseudosTable_IMPL 44 #include "RISCVGenSearchableTables.inc" 45 46 } // namespace RISCVVPseudosTable 47 } // namespace llvm 48 49 RISCVInstrInfo::RISCVInstrInfo(RISCVSubtarget &STI) 50 : RISCVGenInstrInfo(RISCV::ADJCALLSTACKDOWN, RISCV::ADJCALLSTACKUP), 51 STI(STI) {} 52 53 MCInst RISCVInstrInfo::getNop() const { 54 if (STI.getFeatureBits()[RISCV::FeatureStdExtC]) 55 return MCInstBuilder(RISCV::C_NOP); 56 return MCInstBuilder(RISCV::ADDI) 57 .addReg(RISCV::X0) 58 .addReg(RISCV::X0) 59 .addImm(0); 60 } 61 62 unsigned RISCVInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 63 int &FrameIndex) const { 64 switch (MI.getOpcode()) { 65 default: 66 return 0; 67 case RISCV::LB: 68 case RISCV::LBU: 69 case RISCV::LH: 70 case RISCV::LHU: 71 case RISCV::FLH: 72 case RISCV::LW: 73 case RISCV::FLW: 74 case RISCV::LWU: 75 case RISCV::LD: 76 case RISCV::FLD: 77 break; 78 } 79 80 if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() && 81 MI.getOperand(2).getImm() == 0) { 82 FrameIndex = MI.getOperand(1).getIndex(); 83 return MI.getOperand(0).getReg(); 84 } 85 86 return 0; 87 } 88 89 unsigned RISCVInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 90 int &FrameIndex) const { 91 switch (MI.getOpcode()) { 92 default: 93 return 0; 94 case RISCV::SB: 95 case RISCV::SH: 96 case RISCV::SW: 97 case RISCV::FSH: 98 case RISCV::FSW: 99 case RISCV::SD: 100 case RISCV::FSD: 101 break; 102 } 103 104 if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() && 105 MI.getOperand(2).getImm() == 0) { 106 FrameIndex = MI.getOperand(1).getIndex(); 107 return MI.getOperand(0).getReg(); 108 } 109 110 return 0; 111 } 112 113 void RISCVInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 114 MachineBasicBlock::iterator MBBI, 115 const DebugLoc &DL, MCRegister DstReg, 116 MCRegister SrcReg, bool KillSrc) const { 117 if (RISCV::GPRRegClass.contains(DstReg, SrcReg)) { 118 BuildMI(MBB, MBBI, DL, get(RISCV::ADDI), DstReg) 119 .addReg(SrcReg, getKillRegState(KillSrc)) 120 .addImm(0); 121 return; 122 } 123 124 // FPR->FPR copies and VR->VR copies. 125 unsigned Opc; 126 bool IsScalableVector = false; 127 if (RISCV::FPR16RegClass.contains(DstReg, SrcReg)) 128 Opc = RISCV::FSGNJ_H; 129 else if (RISCV::FPR32RegClass.contains(DstReg, SrcReg)) 130 Opc = RISCV::FSGNJ_S; 131 else if (RISCV::FPR64RegClass.contains(DstReg, SrcReg)) 132 Opc = RISCV::FSGNJ_D; 133 else if (RISCV::VRRegClass.contains(DstReg, SrcReg)) { 134 Opc = RISCV::PseudoVMV1R_V; 135 IsScalableVector = true; 136 } else if (RISCV::VRM2RegClass.contains(DstReg, SrcReg)) { 137 Opc = RISCV::PseudoVMV2R_V; 138 IsScalableVector = true; 139 } else if (RISCV::VRM4RegClass.contains(DstReg, SrcReg)) { 140 Opc = RISCV::PseudoVMV4R_V; 141 IsScalableVector = true; 142 } else if (RISCV::VRM8RegClass.contains(DstReg, SrcReg)) { 143 Opc = RISCV::PseudoVMV8R_V; 144 IsScalableVector = true; 145 } else 146 llvm_unreachable("Impossible reg-to-reg copy"); 147 148 if (IsScalableVector) 149 BuildMI(MBB, MBBI, DL, get(Opc), DstReg) 150 .addReg(SrcReg, getKillRegState(KillSrc)); 151 else 152 BuildMI(MBB, MBBI, DL, get(Opc), DstReg) 153 .addReg(SrcReg, getKillRegState(KillSrc)) 154 .addReg(SrcReg, getKillRegState(KillSrc)); 155 } 156 157 void RISCVInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 158 MachineBasicBlock::iterator I, 159 Register SrcReg, bool IsKill, int FI, 160 const TargetRegisterClass *RC, 161 const TargetRegisterInfo *TRI) const { 162 DebugLoc DL; 163 if (I != MBB.end()) 164 DL = I->getDebugLoc(); 165 166 MachineFunction *MF = MBB.getParent(); 167 MachineFrameInfo &MFI = MF->getFrameInfo(); 168 169 unsigned Opcode; 170 bool IsScalableVector = true; 171 bool IsZvlsseg = true; 172 if (RISCV::GPRRegClass.hasSubClassEq(RC)) { 173 Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ? 174 RISCV::SW : RISCV::SD; 175 IsScalableVector = false; 176 } else if (RISCV::FPR16RegClass.hasSubClassEq(RC)) { 177 Opcode = RISCV::FSH; 178 IsScalableVector = false; 179 } else if (RISCV::FPR32RegClass.hasSubClassEq(RC)) { 180 Opcode = RISCV::FSW; 181 IsScalableVector = false; 182 } else if (RISCV::FPR64RegClass.hasSubClassEq(RC)) { 183 Opcode = RISCV::FSD; 184 IsScalableVector = false; 185 } else if (RISCV::VRRegClass.hasSubClassEq(RC)) { 186 Opcode = RISCV::PseudoVSPILL_M1; 187 IsZvlsseg = false; 188 } else if (RISCV::VRM2RegClass.hasSubClassEq(RC)) { 189 Opcode = RISCV::PseudoVSPILL_M2; 190 IsZvlsseg = false; 191 } else if (RISCV::VRM4RegClass.hasSubClassEq(RC)) { 192 Opcode = RISCV::PseudoVSPILL_M4; 193 IsZvlsseg = false; 194 } else if (RISCV::VRM8RegClass.hasSubClassEq(RC)) { 195 Opcode = RISCV::PseudoVSPILL_M8; 196 IsZvlsseg = false; 197 } else if (RISCV::VRN2M1RegClass.hasSubClassEq(RC)) 198 Opcode = RISCV::PseudoVSPILL2_M1; 199 else if (RISCV::VRN2M2RegClass.hasSubClassEq(RC)) 200 Opcode = RISCV::PseudoVSPILL2_M2; 201 else if (RISCV::VRN2M4RegClass.hasSubClassEq(RC)) 202 Opcode = RISCV::PseudoVSPILL2_M4; 203 else if (RISCV::VRN3M1RegClass.hasSubClassEq(RC)) 204 Opcode = RISCV::PseudoVSPILL3_M1; 205 else if (RISCV::VRN3M2RegClass.hasSubClassEq(RC)) 206 Opcode = RISCV::PseudoVSPILL3_M2; 207 else if (RISCV::VRN4M1RegClass.hasSubClassEq(RC)) 208 Opcode = RISCV::PseudoVSPILL4_M1; 209 else if (RISCV::VRN4M2RegClass.hasSubClassEq(RC)) 210 Opcode = RISCV::PseudoVSPILL4_M2; 211 else if (RISCV::VRN5M1RegClass.hasSubClassEq(RC)) 212 Opcode = RISCV::PseudoVSPILL5_M1; 213 else if (RISCV::VRN6M1RegClass.hasSubClassEq(RC)) 214 Opcode = RISCV::PseudoVSPILL6_M1; 215 else if (RISCV::VRN7M1RegClass.hasSubClassEq(RC)) 216 Opcode = RISCV::PseudoVSPILL7_M1; 217 else if (RISCV::VRN8M1RegClass.hasSubClassEq(RC)) 218 Opcode = RISCV::PseudoVSPILL8_M1; 219 else 220 llvm_unreachable("Can't store this register to stack slot"); 221 222 if (IsScalableVector) { 223 MachineMemOperand *MMO = MF->getMachineMemOperand( 224 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore, 225 MemoryLocation::UnknownSize, MFI.getObjectAlign(FI)); 226 227 MFI.setStackID(FI, TargetStackID::ScalableVector); 228 auto MIB = BuildMI(MBB, I, DL, get(Opcode)) 229 .addReg(SrcReg, getKillRegState(IsKill)) 230 .addFrameIndex(FI) 231 .addMemOperand(MMO); 232 if (IsZvlsseg) { 233 // For spilling/reloading Zvlsseg registers, append the dummy field for 234 // the scaled vector length. The argument will be used when expanding 235 // these pseudo instructions. 236 MIB.addReg(RISCV::X0); 237 } 238 } else { 239 MachineMemOperand *MMO = MF->getMachineMemOperand( 240 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore, 241 MFI.getObjectSize(FI), MFI.getObjectAlign(FI)); 242 243 BuildMI(MBB, I, DL, get(Opcode)) 244 .addReg(SrcReg, getKillRegState(IsKill)) 245 .addFrameIndex(FI) 246 .addImm(0) 247 .addMemOperand(MMO); 248 } 249 } 250 251 void RISCVInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 252 MachineBasicBlock::iterator I, 253 Register DstReg, int FI, 254 const TargetRegisterClass *RC, 255 const TargetRegisterInfo *TRI) const { 256 DebugLoc DL; 257 if (I != MBB.end()) 258 DL = I->getDebugLoc(); 259 260 MachineFunction *MF = MBB.getParent(); 261 MachineFrameInfo &MFI = MF->getFrameInfo(); 262 263 unsigned Opcode; 264 bool IsScalableVector = true; 265 bool IsZvlsseg = true; 266 if (RISCV::GPRRegClass.hasSubClassEq(RC)) { 267 Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ? 268 RISCV::LW : RISCV::LD; 269 IsScalableVector = false; 270 } else if (RISCV::FPR16RegClass.hasSubClassEq(RC)) { 271 Opcode = RISCV::FLH; 272 IsScalableVector = false; 273 } else if (RISCV::FPR32RegClass.hasSubClassEq(RC)) { 274 Opcode = RISCV::FLW; 275 IsScalableVector = false; 276 } else if (RISCV::FPR64RegClass.hasSubClassEq(RC)) { 277 Opcode = RISCV::FLD; 278 IsScalableVector = false; 279 } else if (RISCV::VRRegClass.hasSubClassEq(RC)) { 280 Opcode = RISCV::PseudoVRELOAD_M1; 281 IsZvlsseg = false; 282 } else if (RISCV::VRM2RegClass.hasSubClassEq(RC)) { 283 Opcode = RISCV::PseudoVRELOAD_M2; 284 IsZvlsseg = false; 285 } else if (RISCV::VRM4RegClass.hasSubClassEq(RC)) { 286 Opcode = RISCV::PseudoVRELOAD_M4; 287 IsZvlsseg = false; 288 } else if (RISCV::VRM8RegClass.hasSubClassEq(RC)) { 289 Opcode = RISCV::PseudoVRELOAD_M8; 290 IsZvlsseg = false; 291 } else if (RISCV::VRN2M1RegClass.hasSubClassEq(RC)) 292 Opcode = RISCV::PseudoVRELOAD2_M1; 293 else if (RISCV::VRN2M2RegClass.hasSubClassEq(RC)) 294 Opcode = RISCV::PseudoVRELOAD2_M2; 295 else if (RISCV::VRN2M4RegClass.hasSubClassEq(RC)) 296 Opcode = RISCV::PseudoVRELOAD2_M4; 297 else if (RISCV::VRN3M1RegClass.hasSubClassEq(RC)) 298 Opcode = RISCV::PseudoVRELOAD3_M1; 299 else if (RISCV::VRN3M2RegClass.hasSubClassEq(RC)) 300 Opcode = RISCV::PseudoVRELOAD3_M2; 301 else if (RISCV::VRN4M1RegClass.hasSubClassEq(RC)) 302 Opcode = RISCV::PseudoVRELOAD4_M1; 303 else if (RISCV::VRN4M2RegClass.hasSubClassEq(RC)) 304 Opcode = RISCV::PseudoVRELOAD4_M2; 305 else if (RISCV::VRN5M1RegClass.hasSubClassEq(RC)) 306 Opcode = RISCV::PseudoVRELOAD5_M1; 307 else if (RISCV::VRN6M1RegClass.hasSubClassEq(RC)) 308 Opcode = RISCV::PseudoVRELOAD6_M1; 309 else if (RISCV::VRN7M1RegClass.hasSubClassEq(RC)) 310 Opcode = RISCV::PseudoVRELOAD7_M1; 311 else if (RISCV::VRN8M1RegClass.hasSubClassEq(RC)) 312 Opcode = RISCV::PseudoVRELOAD8_M1; 313 else 314 llvm_unreachable("Can't load this register from stack slot"); 315 316 if (IsScalableVector) { 317 MachineMemOperand *MMO = MF->getMachineMemOperand( 318 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad, 319 MemoryLocation::UnknownSize, MFI.getObjectAlign(FI)); 320 321 MFI.setStackID(FI, TargetStackID::ScalableVector); 322 auto MIB = BuildMI(MBB, I, DL, get(Opcode), DstReg) 323 .addFrameIndex(FI) 324 .addMemOperand(MMO); 325 if (IsZvlsseg) { 326 // For spilling/reloading Zvlsseg registers, append the dummy field for 327 // the scaled vector length. The argument will be used when expanding 328 // these pseudo instructions. 329 MIB.addReg(RISCV::X0); 330 } 331 } else { 332 MachineMemOperand *MMO = MF->getMachineMemOperand( 333 MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad, 334 MFI.getObjectSize(FI), MFI.getObjectAlign(FI)); 335 336 BuildMI(MBB, I, DL, get(Opcode), DstReg) 337 .addFrameIndex(FI) 338 .addImm(0) 339 .addMemOperand(MMO); 340 } 341 } 342 343 void RISCVInstrInfo::movImm(MachineBasicBlock &MBB, 344 MachineBasicBlock::iterator MBBI, 345 const DebugLoc &DL, Register DstReg, uint64_t Val, 346 MachineInstr::MIFlag Flag) const { 347 MachineFunction *MF = MBB.getParent(); 348 MachineRegisterInfo &MRI = MF->getRegInfo(); 349 bool IsRV64 = MF->getSubtarget<RISCVSubtarget>().is64Bit(); 350 Register SrcReg = RISCV::X0; 351 Register Result = MRI.createVirtualRegister(&RISCV::GPRRegClass); 352 unsigned Num = 0; 353 354 if (!IsRV64 && !isInt<32>(Val)) 355 report_fatal_error("Should only materialize 32-bit constants for RV32"); 356 357 RISCVMatInt::InstSeq Seq = RISCVMatInt::generateInstSeq(Val, IsRV64); 358 assert(Seq.size() > 0); 359 360 for (RISCVMatInt::Inst &Inst : Seq) { 361 // Write the final result to DstReg if it's the last instruction in the Seq. 362 // Otherwise, write the result to the temp register. 363 if (++Num == Seq.size()) 364 Result = DstReg; 365 366 if (Inst.Opc == RISCV::LUI) { 367 BuildMI(MBB, MBBI, DL, get(RISCV::LUI), Result) 368 .addImm(Inst.Imm) 369 .setMIFlag(Flag); 370 } else { 371 BuildMI(MBB, MBBI, DL, get(Inst.Opc), Result) 372 .addReg(SrcReg, RegState::Kill) 373 .addImm(Inst.Imm) 374 .setMIFlag(Flag); 375 } 376 // Only the first instruction has X0 as its source. 377 SrcReg = Result; 378 } 379 } 380 381 // The contents of values added to Cond are not examined outside of 382 // RISCVInstrInfo, giving us flexibility in what to push to it. For RISCV, we 383 // push BranchOpcode, Reg1, Reg2. 384 static void parseCondBranch(MachineInstr &LastInst, MachineBasicBlock *&Target, 385 SmallVectorImpl<MachineOperand> &Cond) { 386 // Block ends with fall-through condbranch. 387 assert(LastInst.getDesc().isConditionalBranch() && 388 "Unknown conditional branch"); 389 Target = LastInst.getOperand(2).getMBB(); 390 Cond.push_back(MachineOperand::CreateImm(LastInst.getOpcode())); 391 Cond.push_back(LastInst.getOperand(0)); 392 Cond.push_back(LastInst.getOperand(1)); 393 } 394 395 static unsigned getOppositeBranchOpcode(int Opc) { 396 switch (Opc) { 397 default: 398 llvm_unreachable("Unrecognized conditional branch"); 399 case RISCV::BEQ: 400 return RISCV::BNE; 401 case RISCV::BNE: 402 return RISCV::BEQ; 403 case RISCV::BLT: 404 return RISCV::BGE; 405 case RISCV::BGE: 406 return RISCV::BLT; 407 case RISCV::BLTU: 408 return RISCV::BGEU; 409 case RISCV::BGEU: 410 return RISCV::BLTU; 411 } 412 } 413 414 bool RISCVInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 415 MachineBasicBlock *&TBB, 416 MachineBasicBlock *&FBB, 417 SmallVectorImpl<MachineOperand> &Cond, 418 bool AllowModify) const { 419 TBB = FBB = nullptr; 420 Cond.clear(); 421 422 // If the block has no terminators, it just falls into the block after it. 423 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 424 if (I == MBB.end() || !isUnpredicatedTerminator(*I)) 425 return false; 426 427 // Count the number of terminators and find the first unconditional or 428 // indirect branch. 429 MachineBasicBlock::iterator FirstUncondOrIndirectBr = MBB.end(); 430 int NumTerminators = 0; 431 for (auto J = I.getReverse(); J != MBB.rend() && isUnpredicatedTerminator(*J); 432 J++) { 433 NumTerminators++; 434 if (J->getDesc().isUnconditionalBranch() || 435 J->getDesc().isIndirectBranch()) { 436 FirstUncondOrIndirectBr = J.getReverse(); 437 } 438 } 439 440 // If AllowModify is true, we can erase any terminators after 441 // FirstUncondOrIndirectBR. 442 if (AllowModify && FirstUncondOrIndirectBr != MBB.end()) { 443 while (std::next(FirstUncondOrIndirectBr) != MBB.end()) { 444 std::next(FirstUncondOrIndirectBr)->eraseFromParent(); 445 NumTerminators--; 446 } 447 I = FirstUncondOrIndirectBr; 448 } 449 450 // We can't handle blocks that end in an indirect branch. 451 if (I->getDesc().isIndirectBranch()) 452 return true; 453 454 // We can't handle blocks with more than 2 terminators. 455 if (NumTerminators > 2) 456 return true; 457 458 // Handle a single unconditional branch. 459 if (NumTerminators == 1 && I->getDesc().isUnconditionalBranch()) { 460 TBB = getBranchDestBlock(*I); 461 return false; 462 } 463 464 // Handle a single conditional branch. 465 if (NumTerminators == 1 && I->getDesc().isConditionalBranch()) { 466 parseCondBranch(*I, TBB, Cond); 467 return false; 468 } 469 470 // Handle a conditional branch followed by an unconditional branch. 471 if (NumTerminators == 2 && std::prev(I)->getDesc().isConditionalBranch() && 472 I->getDesc().isUnconditionalBranch()) { 473 parseCondBranch(*std::prev(I), TBB, Cond); 474 FBB = getBranchDestBlock(*I); 475 return false; 476 } 477 478 // Otherwise, we can't handle this. 479 return true; 480 } 481 482 unsigned RISCVInstrInfo::removeBranch(MachineBasicBlock &MBB, 483 int *BytesRemoved) const { 484 if (BytesRemoved) 485 *BytesRemoved = 0; 486 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 487 if (I == MBB.end()) 488 return 0; 489 490 if (!I->getDesc().isUnconditionalBranch() && 491 !I->getDesc().isConditionalBranch()) 492 return 0; 493 494 // Remove the branch. 495 if (BytesRemoved) 496 *BytesRemoved += getInstSizeInBytes(*I); 497 I->eraseFromParent(); 498 499 I = MBB.end(); 500 501 if (I == MBB.begin()) 502 return 1; 503 --I; 504 if (!I->getDesc().isConditionalBranch()) 505 return 1; 506 507 // Remove the branch. 508 if (BytesRemoved) 509 *BytesRemoved += getInstSizeInBytes(*I); 510 I->eraseFromParent(); 511 return 2; 512 } 513 514 // Inserts a branch into the end of the specific MachineBasicBlock, returning 515 // the number of instructions inserted. 516 unsigned RISCVInstrInfo::insertBranch( 517 MachineBasicBlock &MBB, MachineBasicBlock *TBB, MachineBasicBlock *FBB, 518 ArrayRef<MachineOperand> Cond, const DebugLoc &DL, int *BytesAdded) const { 519 if (BytesAdded) 520 *BytesAdded = 0; 521 522 // Shouldn't be a fall through. 523 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 524 assert((Cond.size() == 3 || Cond.size() == 0) && 525 "RISCV branch conditions have two components!"); 526 527 // Unconditional branch. 528 if (Cond.empty()) { 529 MachineInstr &MI = *BuildMI(&MBB, DL, get(RISCV::PseudoBR)).addMBB(TBB); 530 if (BytesAdded) 531 *BytesAdded += getInstSizeInBytes(MI); 532 return 1; 533 } 534 535 // Either a one or two-way conditional branch. 536 unsigned Opc = Cond[0].getImm(); 537 MachineInstr &CondMI = 538 *BuildMI(&MBB, DL, get(Opc)).add(Cond[1]).add(Cond[2]).addMBB(TBB); 539 if (BytesAdded) 540 *BytesAdded += getInstSizeInBytes(CondMI); 541 542 // One-way conditional branch. 543 if (!FBB) 544 return 1; 545 546 // Two-way conditional branch. 547 MachineInstr &MI = *BuildMI(&MBB, DL, get(RISCV::PseudoBR)).addMBB(FBB); 548 if (BytesAdded) 549 *BytesAdded += getInstSizeInBytes(MI); 550 return 2; 551 } 552 553 unsigned RISCVInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB, 554 MachineBasicBlock &DestBB, 555 const DebugLoc &DL, 556 int64_t BrOffset, 557 RegScavenger *RS) const { 558 assert(RS && "RegScavenger required for long branching"); 559 assert(MBB.empty() && 560 "new block should be inserted for expanding unconditional branch"); 561 assert(MBB.pred_size() == 1); 562 563 MachineFunction *MF = MBB.getParent(); 564 MachineRegisterInfo &MRI = MF->getRegInfo(); 565 566 if (!isInt<32>(BrOffset)) 567 report_fatal_error( 568 "Branch offsets outside of the signed 32-bit range not supported"); 569 570 // FIXME: A virtual register must be used initially, as the register 571 // scavenger won't work with empty blocks (SIInstrInfo::insertIndirectBranch 572 // uses the same workaround). 573 Register ScratchReg = MRI.createVirtualRegister(&RISCV::GPRRegClass); 574 auto II = MBB.end(); 575 576 MachineInstr &MI = *BuildMI(MBB, II, DL, get(RISCV::PseudoJump)) 577 .addReg(ScratchReg, RegState::Define | RegState::Dead) 578 .addMBB(&DestBB, RISCVII::MO_CALL); 579 580 RS->enterBasicBlockEnd(MBB); 581 unsigned Scav = RS->scavengeRegisterBackwards(RISCV::GPRRegClass, 582 MI.getIterator(), false, 0); 583 MRI.replaceRegWith(ScratchReg, Scav); 584 MRI.clearVirtRegs(); 585 RS->setRegUsed(Scav); 586 return 8; 587 } 588 589 bool RISCVInstrInfo::reverseBranchCondition( 590 SmallVectorImpl<MachineOperand> &Cond) const { 591 assert((Cond.size() == 3) && "Invalid branch condition!"); 592 Cond[0].setImm(getOppositeBranchOpcode(Cond[0].getImm())); 593 return false; 594 } 595 596 MachineBasicBlock * 597 RISCVInstrInfo::getBranchDestBlock(const MachineInstr &MI) const { 598 assert(MI.getDesc().isBranch() && "Unexpected opcode!"); 599 // The branch target is always the last operand. 600 int NumOp = MI.getNumExplicitOperands(); 601 return MI.getOperand(NumOp - 1).getMBB(); 602 } 603 604 bool RISCVInstrInfo::isBranchOffsetInRange(unsigned BranchOp, 605 int64_t BrOffset) const { 606 unsigned XLen = STI.getXLen(); 607 // Ideally we could determine the supported branch offset from the 608 // RISCVII::FormMask, but this can't be used for Pseudo instructions like 609 // PseudoBR. 610 switch (BranchOp) { 611 default: 612 llvm_unreachable("Unexpected opcode!"); 613 case RISCV::BEQ: 614 case RISCV::BNE: 615 case RISCV::BLT: 616 case RISCV::BGE: 617 case RISCV::BLTU: 618 case RISCV::BGEU: 619 return isIntN(13, BrOffset); 620 case RISCV::JAL: 621 case RISCV::PseudoBR: 622 return isIntN(21, BrOffset); 623 case RISCV::PseudoJump: 624 return isIntN(32, SignExtend64(BrOffset + 0x800, XLen)); 625 } 626 } 627 628 unsigned RISCVInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 629 unsigned Opcode = MI.getOpcode(); 630 631 switch (Opcode) { 632 default: { 633 if (MI.getParent() && MI.getParent()->getParent()) { 634 const auto MF = MI.getMF(); 635 const auto &TM = static_cast<const RISCVTargetMachine &>(MF->getTarget()); 636 const MCRegisterInfo &MRI = *TM.getMCRegisterInfo(); 637 const MCSubtargetInfo &STI = *TM.getMCSubtargetInfo(); 638 const RISCVSubtarget &ST = MF->getSubtarget<RISCVSubtarget>(); 639 if (isCompressibleInst(MI, &ST, MRI, STI)) 640 return 2; 641 } 642 return get(Opcode).getSize(); 643 } 644 case TargetOpcode::EH_LABEL: 645 case TargetOpcode::IMPLICIT_DEF: 646 case TargetOpcode::KILL: 647 case TargetOpcode::DBG_VALUE: 648 return 0; 649 // These values are determined based on RISCVExpandAtomicPseudoInsts, 650 // RISCVExpandPseudoInsts and RISCVMCCodeEmitter, depending on where the 651 // pseudos are expanded. 652 case RISCV::PseudoCALLReg: 653 case RISCV::PseudoCALL: 654 case RISCV::PseudoJump: 655 case RISCV::PseudoTAIL: 656 case RISCV::PseudoLLA: 657 case RISCV::PseudoLA: 658 case RISCV::PseudoLA_TLS_IE: 659 case RISCV::PseudoLA_TLS_GD: 660 return 8; 661 case RISCV::PseudoAtomicLoadNand32: 662 case RISCV::PseudoAtomicLoadNand64: 663 return 20; 664 case RISCV::PseudoMaskedAtomicSwap32: 665 case RISCV::PseudoMaskedAtomicLoadAdd32: 666 case RISCV::PseudoMaskedAtomicLoadSub32: 667 return 28; 668 case RISCV::PseudoMaskedAtomicLoadNand32: 669 return 32; 670 case RISCV::PseudoMaskedAtomicLoadMax32: 671 case RISCV::PseudoMaskedAtomicLoadMin32: 672 return 44; 673 case RISCV::PseudoMaskedAtomicLoadUMax32: 674 case RISCV::PseudoMaskedAtomicLoadUMin32: 675 return 36; 676 case RISCV::PseudoCmpXchg32: 677 case RISCV::PseudoCmpXchg64: 678 return 16; 679 case RISCV::PseudoMaskedCmpXchg32: 680 return 32; 681 case TargetOpcode::INLINEASM: 682 case TargetOpcode::INLINEASM_BR: { 683 const MachineFunction &MF = *MI.getParent()->getParent(); 684 const auto &TM = static_cast<const RISCVTargetMachine &>(MF.getTarget()); 685 return getInlineAsmLength(MI.getOperand(0).getSymbolName(), 686 *TM.getMCAsmInfo()); 687 } 688 } 689 } 690 691 bool RISCVInstrInfo::isAsCheapAsAMove(const MachineInstr &MI) const { 692 const unsigned Opcode = MI.getOpcode(); 693 switch (Opcode) { 694 default: 695 break; 696 case RISCV::FSGNJ_D: 697 case RISCV::FSGNJ_S: 698 // The canonical floating-point move is fsgnj rd, rs, rs. 699 return MI.getOperand(1).isReg() && MI.getOperand(2).isReg() && 700 MI.getOperand(1).getReg() == MI.getOperand(2).getReg(); 701 case RISCV::ADDI: 702 case RISCV::ORI: 703 case RISCV::XORI: 704 return (MI.getOperand(1).isReg() && 705 MI.getOperand(1).getReg() == RISCV::X0) || 706 (MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0); 707 } 708 return MI.isAsCheapAsAMove(); 709 } 710 711 Optional<DestSourcePair> 712 RISCVInstrInfo::isCopyInstrImpl(const MachineInstr &MI) const { 713 if (MI.isMoveReg()) 714 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 715 switch (MI.getOpcode()) { 716 default: 717 break; 718 case RISCV::ADDI: 719 // Operand 1 can be a frameindex but callers expect registers 720 if (MI.getOperand(1).isReg() && MI.getOperand(2).isImm() && 721 MI.getOperand(2).getImm() == 0) 722 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 723 break; 724 case RISCV::FSGNJ_D: 725 case RISCV::FSGNJ_S: 726 // The canonical floating-point move is fsgnj rd, rs, rs. 727 if (MI.getOperand(1).isReg() && MI.getOperand(2).isReg() && 728 MI.getOperand(1).getReg() == MI.getOperand(2).getReg()) 729 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 730 break; 731 } 732 return None; 733 } 734 735 bool RISCVInstrInfo::verifyInstruction(const MachineInstr &MI, 736 StringRef &ErrInfo) const { 737 const MCInstrInfo *MCII = STI.getInstrInfo(); 738 MCInstrDesc const &Desc = MCII->get(MI.getOpcode()); 739 740 for (auto &OI : enumerate(Desc.operands())) { 741 unsigned OpType = OI.value().OperandType; 742 if (OpType >= RISCVOp::OPERAND_FIRST_RISCV_IMM && 743 OpType <= RISCVOp::OPERAND_LAST_RISCV_IMM) { 744 const MachineOperand &MO = MI.getOperand(OI.index()); 745 if (MO.isImm()) { 746 int64_t Imm = MO.getImm(); 747 bool Ok; 748 switch (OpType) { 749 default: 750 llvm_unreachable("Unexpected operand type"); 751 case RISCVOp::OPERAND_UIMM4: 752 Ok = isUInt<4>(Imm); 753 break; 754 case RISCVOp::OPERAND_UIMM5: 755 Ok = isUInt<5>(Imm); 756 break; 757 case RISCVOp::OPERAND_UIMM12: 758 Ok = isUInt<12>(Imm); 759 break; 760 case RISCVOp::OPERAND_SIMM12: 761 Ok = isInt<12>(Imm); 762 break; 763 case RISCVOp::OPERAND_UIMM20: 764 Ok = isUInt<20>(Imm); 765 break; 766 case RISCVOp::OPERAND_UIMMLOG2XLEN: 767 if (STI.getTargetTriple().isArch64Bit()) 768 Ok = isUInt<6>(Imm); 769 else 770 Ok = isUInt<5>(Imm); 771 break; 772 } 773 if (!Ok) { 774 ErrInfo = "Invalid immediate"; 775 return false; 776 } 777 } 778 } 779 } 780 781 return true; 782 } 783 784 // Return true if get the base operand, byte offset of an instruction and the 785 // memory width. Width is the size of memory that is being loaded/stored. 786 bool RISCVInstrInfo::getMemOperandWithOffsetWidth( 787 const MachineInstr &LdSt, const MachineOperand *&BaseReg, int64_t &Offset, 788 unsigned &Width, const TargetRegisterInfo *TRI) const { 789 if (!LdSt.mayLoadOrStore()) 790 return false; 791 792 // Here we assume the standard RISC-V ISA, which uses a base+offset 793 // addressing mode. You'll need to relax these conditions to support custom 794 // load/stores instructions. 795 if (LdSt.getNumExplicitOperands() != 3) 796 return false; 797 if (!LdSt.getOperand(1).isReg() || !LdSt.getOperand(2).isImm()) 798 return false; 799 800 if (!LdSt.hasOneMemOperand()) 801 return false; 802 803 Width = (*LdSt.memoperands_begin())->getSize(); 804 BaseReg = &LdSt.getOperand(1); 805 Offset = LdSt.getOperand(2).getImm(); 806 return true; 807 } 808 809 bool RISCVInstrInfo::areMemAccessesTriviallyDisjoint( 810 const MachineInstr &MIa, const MachineInstr &MIb) const { 811 assert(MIa.mayLoadOrStore() && "MIa must be a load or store."); 812 assert(MIb.mayLoadOrStore() && "MIb must be a load or store."); 813 814 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() || 815 MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef()) 816 return false; 817 818 // Retrieve the base register, offset from the base register and width. Width 819 // is the size of memory that is being loaded/stored (e.g. 1, 2, 4). If 820 // base registers are identical, and the offset of a lower memory access + 821 // the width doesn't overlap the offset of a higher memory access, 822 // then the memory accesses are different. 823 const TargetRegisterInfo *TRI = STI.getRegisterInfo(); 824 const MachineOperand *BaseOpA = nullptr, *BaseOpB = nullptr; 825 int64_t OffsetA = 0, OffsetB = 0; 826 unsigned int WidthA = 0, WidthB = 0; 827 if (getMemOperandWithOffsetWidth(MIa, BaseOpA, OffsetA, WidthA, TRI) && 828 getMemOperandWithOffsetWidth(MIb, BaseOpB, OffsetB, WidthB, TRI)) { 829 if (BaseOpA->isIdenticalTo(*BaseOpB)) { 830 int LowOffset = std::min(OffsetA, OffsetB); 831 int HighOffset = std::max(OffsetA, OffsetB); 832 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 833 if (LowOffset + LowWidth <= HighOffset) 834 return true; 835 } 836 } 837 return false; 838 } 839 840 std::pair<unsigned, unsigned> 841 RISCVInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { 842 const unsigned Mask = RISCVII::MO_DIRECT_FLAG_MASK; 843 return std::make_pair(TF & Mask, TF & ~Mask); 844 } 845 846 ArrayRef<std::pair<unsigned, const char *>> 847 RISCVInstrInfo::getSerializableDirectMachineOperandTargetFlags() const { 848 using namespace RISCVII; 849 static const std::pair<unsigned, const char *> TargetFlags[] = { 850 {MO_CALL, "riscv-call"}, 851 {MO_PLT, "riscv-plt"}, 852 {MO_LO, "riscv-lo"}, 853 {MO_HI, "riscv-hi"}, 854 {MO_PCREL_LO, "riscv-pcrel-lo"}, 855 {MO_PCREL_HI, "riscv-pcrel-hi"}, 856 {MO_GOT_HI, "riscv-got-hi"}, 857 {MO_TPREL_LO, "riscv-tprel-lo"}, 858 {MO_TPREL_HI, "riscv-tprel-hi"}, 859 {MO_TPREL_ADD, "riscv-tprel-add"}, 860 {MO_TLS_GOT_HI, "riscv-tls-got-hi"}, 861 {MO_TLS_GD_HI, "riscv-tls-gd-hi"}}; 862 return makeArrayRef(TargetFlags); 863 } 864 bool RISCVInstrInfo::isFunctionSafeToOutlineFrom( 865 MachineFunction &MF, bool OutlineFromLinkOnceODRs) const { 866 const Function &F = MF.getFunction(); 867 868 // Can F be deduplicated by the linker? If it can, don't outline from it. 869 if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage()) 870 return false; 871 872 // Don't outline from functions with section markings; the program could 873 // expect that all the code is in the named section. 874 if (F.hasSection()) 875 return false; 876 877 // It's safe to outline from MF. 878 return true; 879 } 880 881 bool RISCVInstrInfo::isMBBSafeToOutlineFrom(MachineBasicBlock &MBB, 882 unsigned &Flags) const { 883 // More accurate safety checking is done in getOutliningCandidateInfo. 884 return true; 885 } 886 887 // Enum values indicating how an outlined call should be constructed. 888 enum MachineOutlinerConstructionID { 889 MachineOutlinerDefault 890 }; 891 892 outliner::OutlinedFunction RISCVInstrInfo::getOutliningCandidateInfo( 893 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const { 894 895 // First we need to filter out candidates where the X5 register (IE t0) can't 896 // be used to setup the function call. 897 auto CannotInsertCall = [](outliner::Candidate &C) { 898 const TargetRegisterInfo *TRI = C.getMF()->getSubtarget().getRegisterInfo(); 899 900 C.initLRU(*TRI); 901 LiveRegUnits LRU = C.LRU; 902 return !LRU.available(RISCV::X5); 903 }; 904 905 llvm::erase_if(RepeatedSequenceLocs, CannotInsertCall); 906 907 // If the sequence doesn't have enough candidates left, then we're done. 908 if (RepeatedSequenceLocs.size() < 2) 909 return outliner::OutlinedFunction(); 910 911 unsigned SequenceSize = 0; 912 913 auto I = RepeatedSequenceLocs[0].front(); 914 auto E = std::next(RepeatedSequenceLocs[0].back()); 915 for (; I != E; ++I) 916 SequenceSize += getInstSizeInBytes(*I); 917 918 // call t0, function = 8 bytes. 919 unsigned CallOverhead = 8; 920 for (auto &C : RepeatedSequenceLocs) 921 C.setCallInfo(MachineOutlinerDefault, CallOverhead); 922 923 // jr t0 = 4 bytes, 2 bytes if compressed instructions are enabled. 924 unsigned FrameOverhead = 4; 925 if (RepeatedSequenceLocs[0].getMF()->getSubtarget() 926 .getFeatureBits()[RISCV::FeatureStdExtC]) 927 FrameOverhead = 2; 928 929 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 930 FrameOverhead, MachineOutlinerDefault); 931 } 932 933 outliner::InstrType 934 RISCVInstrInfo::getOutliningType(MachineBasicBlock::iterator &MBBI, 935 unsigned Flags) const { 936 MachineInstr &MI = *MBBI; 937 MachineBasicBlock *MBB = MI.getParent(); 938 const TargetRegisterInfo *TRI = 939 MBB->getParent()->getSubtarget().getRegisterInfo(); 940 941 // Positions generally can't safely be outlined. 942 if (MI.isPosition()) { 943 // We can manually strip out CFI instructions later. 944 if (MI.isCFIInstruction()) 945 return outliner::InstrType::Invisible; 946 947 return outliner::InstrType::Illegal; 948 } 949 950 // Don't trust the user to write safe inline assembly. 951 if (MI.isInlineAsm()) 952 return outliner::InstrType::Illegal; 953 954 // We can't outline branches to other basic blocks. 955 if (MI.isTerminator() && !MBB->succ_empty()) 956 return outliner::InstrType::Illegal; 957 958 // We need support for tail calls to outlined functions before return 959 // statements can be allowed. 960 if (MI.isReturn()) 961 return outliner::InstrType::Illegal; 962 963 // Don't allow modifying the X5 register which we use for return addresses for 964 // these outlined functions. 965 if (MI.modifiesRegister(RISCV::X5, TRI) || 966 MI.getDesc().hasImplicitDefOfPhysReg(RISCV::X5)) 967 return outliner::InstrType::Illegal; 968 969 // Make sure the operands don't reference something unsafe. 970 for (const auto &MO : MI.operands()) 971 if (MO.isMBB() || MO.isBlockAddress() || MO.isCPI()) 972 return outliner::InstrType::Illegal; 973 974 // Don't allow instructions which won't be materialized to impact outlining 975 // analysis. 976 if (MI.isMetaInstruction()) 977 return outliner::InstrType::Invisible; 978 979 return outliner::InstrType::Legal; 980 } 981 982 void RISCVInstrInfo::buildOutlinedFrame( 983 MachineBasicBlock &MBB, MachineFunction &MF, 984 const outliner::OutlinedFunction &OF) const { 985 986 // Strip out any CFI instructions 987 bool Changed = true; 988 while (Changed) { 989 Changed = false; 990 auto I = MBB.begin(); 991 auto E = MBB.end(); 992 for (; I != E; ++I) { 993 if (I->isCFIInstruction()) { 994 I->removeFromParent(); 995 Changed = true; 996 break; 997 } 998 } 999 } 1000 1001 MBB.addLiveIn(RISCV::X5); 1002 1003 // Add in a return instruction to the end of the outlined frame. 1004 MBB.insert(MBB.end(), BuildMI(MF, DebugLoc(), get(RISCV::JALR)) 1005 .addReg(RISCV::X0, RegState::Define) 1006 .addReg(RISCV::X5) 1007 .addImm(0)); 1008 } 1009 1010 MachineBasicBlock::iterator RISCVInstrInfo::insertOutlinedCall( 1011 Module &M, MachineBasicBlock &MBB, MachineBasicBlock::iterator &It, 1012 MachineFunction &MF, const outliner::Candidate &C) const { 1013 1014 // Add in a call instruction to the outlined function at the given location. 1015 It = MBB.insert(It, 1016 BuildMI(MF, DebugLoc(), get(RISCV::PseudoCALLReg), RISCV::X5) 1017 .addGlobalAddress(M.getNamedValue(MF.getName()), 0, 1018 RISCVII::MO_CALL)); 1019 return It; 1020 } 1021 1022 // clang-format off 1023 #define CASE_VFMA_OPCODE_COMMON(OP, TYPE, LMUL) \ 1024 RISCV::PseudoV##OP##_##TYPE##_##LMUL##_COMMUTABLE 1025 1026 #define CASE_VFMA_OPCODE_LMULS(OP, TYPE) \ 1027 CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF8): \ 1028 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF4): \ 1029 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF2): \ 1030 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M1): \ 1031 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M2): \ 1032 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M4): \ 1033 case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M8) 1034 1035 #define CASE_VFMA_SPLATS(OP) \ 1036 CASE_VFMA_OPCODE_LMULS(OP, VF16): \ 1037 case CASE_VFMA_OPCODE_LMULS(OP, VF32): \ 1038 case CASE_VFMA_OPCODE_LMULS(OP, VF64) 1039 // clang-format on 1040 1041 bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI, 1042 unsigned &SrcOpIdx1, 1043 unsigned &SrcOpIdx2) const { 1044 const MCInstrDesc &Desc = MI.getDesc(); 1045 if (!Desc.isCommutable()) 1046 return false; 1047 1048 switch (MI.getOpcode()) { 1049 case CASE_VFMA_SPLATS(FMADD): 1050 case CASE_VFMA_SPLATS(FMSUB): 1051 case CASE_VFMA_SPLATS(FMACC): 1052 case CASE_VFMA_SPLATS(FMSAC): 1053 case CASE_VFMA_SPLATS(FNMADD): 1054 case CASE_VFMA_SPLATS(FNMSUB): 1055 case CASE_VFMA_SPLATS(FNMACC): 1056 case CASE_VFMA_SPLATS(FNMSAC): 1057 case CASE_VFMA_OPCODE_LMULS(FMACC, VV): 1058 case CASE_VFMA_OPCODE_LMULS(FMSAC, VV): 1059 case CASE_VFMA_OPCODE_LMULS(FNMACC, VV): 1060 case CASE_VFMA_OPCODE_LMULS(FNMSAC, VV): { 1061 // For these instructions we can only swap operand 1 and operand 3 by 1062 // changing the opcode. 1063 unsigned CommutableOpIdx1 = 1; 1064 unsigned CommutableOpIdx2 = 3; 1065 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1, 1066 CommutableOpIdx2)) 1067 return false; 1068 return true; 1069 } 1070 case CASE_VFMA_OPCODE_LMULS(FMADD, VV): 1071 case CASE_VFMA_OPCODE_LMULS(FMSUB, VV): 1072 case CASE_VFMA_OPCODE_LMULS(FNMADD, VV): 1073 case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): { 1074 // For these instructions we have more freedom. We can commute with the 1075 // other multiplicand or with the addend/subtrahend/minuend. 1076 1077 // Any fixed operand must be from source 1, 2 or 3. 1078 if (SrcOpIdx1 != CommuteAnyOperandIndex && SrcOpIdx1 > 3) 1079 return false; 1080 if (SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx2 > 3) 1081 return false; 1082 1083 // It both ops are fixed one must be the tied source. 1084 if (SrcOpIdx1 != CommuteAnyOperandIndex && 1085 SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx1 != 1 && SrcOpIdx2 != 1) 1086 return false; 1087 1088 // Look for two different register operands assumed to be commutable 1089 // regardless of the FMA opcode. The FMA opcode is adjusted later if 1090 // needed. 1091 if (SrcOpIdx1 == CommuteAnyOperandIndex || 1092 SrcOpIdx2 == CommuteAnyOperandIndex) { 1093 // At least one of operands to be commuted is not specified and 1094 // this method is free to choose appropriate commutable operands. 1095 unsigned CommutableOpIdx1 = SrcOpIdx1; 1096 if (SrcOpIdx1 == SrcOpIdx2) { 1097 // Both of operands are not fixed. Set one of commutable 1098 // operands to the tied source. 1099 CommutableOpIdx1 = 1; 1100 } else if (SrcOpIdx1 == CommutableOpIdx1) { 1101 // Only one of the operands is not fixed. 1102 CommutableOpIdx1 = SrcOpIdx2; 1103 } 1104 1105 // CommutableOpIdx1 is well defined now. Let's choose another commutable 1106 // operand and assign its index to CommutableOpIdx2. 1107 unsigned CommutableOpIdx2; 1108 if (CommutableOpIdx1 != 1) { 1109 // If we haven't already used the tied source, we must use it now. 1110 CommutableOpIdx2 = 1; 1111 } else { 1112 Register Op1Reg = MI.getOperand(CommutableOpIdx1).getReg(); 1113 1114 // The commuted operands should have different registers. 1115 // Otherwise, the commute transformation does not change anything and 1116 // is useless. We use this as a hint to make our decision. 1117 if (Op1Reg != MI.getOperand(2).getReg()) 1118 CommutableOpIdx2 = 2; 1119 else 1120 CommutableOpIdx2 = 3; 1121 } 1122 1123 // Assign the found pair of commutable indices to SrcOpIdx1 and 1124 // SrcOpIdx2 to return those values. 1125 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1, 1126 CommutableOpIdx2)) 1127 return false; 1128 } 1129 1130 return true; 1131 } 1132 } 1133 1134 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 1135 } 1136 1137 #define CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, LMUL) \ 1138 case RISCV::PseudoV##OLDOP##_##TYPE##_##LMUL##_COMMUTABLE: \ 1139 Opc = RISCV::PseudoV##NEWOP##_##TYPE##_##LMUL##_COMMUTABLE; \ 1140 break; 1141 1142 #define CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, TYPE) \ 1143 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF8) \ 1144 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF4) \ 1145 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF2) \ 1146 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M1) \ 1147 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M2) \ 1148 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M4) \ 1149 CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M8) 1150 1151 #define CASE_VFMA_CHANGE_OPCODE_SPLATS(OLDOP, NEWOP) \ 1152 CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF16) \ 1153 CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF32) \ 1154 CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF64) 1155 1156 MachineInstr *RISCVInstrInfo::commuteInstructionImpl(MachineInstr &MI, 1157 bool NewMI, 1158 unsigned OpIdx1, 1159 unsigned OpIdx2) const { 1160 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 1161 if (NewMI) 1162 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 1163 return MI; 1164 }; 1165 1166 switch (MI.getOpcode()) { 1167 case CASE_VFMA_SPLATS(FMACC): 1168 case CASE_VFMA_SPLATS(FMADD): 1169 case CASE_VFMA_SPLATS(FMSAC): 1170 case CASE_VFMA_SPLATS(FMSUB): 1171 case CASE_VFMA_SPLATS(FNMACC): 1172 case CASE_VFMA_SPLATS(FNMADD): 1173 case CASE_VFMA_SPLATS(FNMSAC): 1174 case CASE_VFMA_SPLATS(FNMSUB): 1175 case CASE_VFMA_OPCODE_LMULS(FMACC, VV): 1176 case CASE_VFMA_OPCODE_LMULS(FMSAC, VV): 1177 case CASE_VFMA_OPCODE_LMULS(FNMACC, VV): 1178 case CASE_VFMA_OPCODE_LMULS(FNMSAC, VV): { 1179 // It only make sense to toggle these between clobbering the 1180 // addend/subtrahend/minuend one of the multiplicands. 1181 assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index"); 1182 assert((OpIdx1 == 3 || OpIdx2 == 3) && "Unexpected opcode index"); 1183 unsigned Opc; 1184 switch (MI.getOpcode()) { 1185 default: 1186 llvm_unreachable("Unexpected opcode"); 1187 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMACC, FMADD) 1188 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMADD, FMACC) 1189 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSAC, FMSUB) 1190 CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSUB, FMSAC) 1191 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMACC, FNMADD) 1192 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMADD, FNMACC) 1193 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSAC, FNMSUB) 1194 CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSUB, FNMSAC) 1195 CASE_VFMA_CHANGE_OPCODE_LMULS(FMACC, FMADD, VV) 1196 CASE_VFMA_CHANGE_OPCODE_LMULS(FMSAC, FMSUB, VV) 1197 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMACC, FNMADD, VV) 1198 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMSAC, FNMSUB, VV) 1199 } 1200 1201 auto &WorkingMI = cloneIfNew(MI); 1202 WorkingMI.setDesc(get(Opc)); 1203 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1204 OpIdx1, OpIdx2); 1205 } 1206 case CASE_VFMA_OPCODE_LMULS(FMADD, VV): 1207 case CASE_VFMA_OPCODE_LMULS(FMSUB, VV): 1208 case CASE_VFMA_OPCODE_LMULS(FNMADD, VV): 1209 case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): { 1210 assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index"); 1211 // If one of the operands, is the addend we need to change opcode. 1212 // Otherwise we're just swapping 2 of the multiplicands. 1213 if (OpIdx1 == 3 || OpIdx2 == 3) { 1214 unsigned Opc; 1215 switch (MI.getOpcode()) { 1216 default: 1217 llvm_unreachable("Unexpected opcode"); 1218 CASE_VFMA_CHANGE_OPCODE_LMULS(FMADD, FMACC, VV) 1219 CASE_VFMA_CHANGE_OPCODE_LMULS(FMSUB, FMSAC, VV) 1220 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMADD, FNMACC, VV) 1221 CASE_VFMA_CHANGE_OPCODE_LMULS(FNMSUB, FNMSAC, VV) 1222 } 1223 1224 auto &WorkingMI = cloneIfNew(MI); 1225 WorkingMI.setDesc(get(Opc)); 1226 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1227 OpIdx1, OpIdx2); 1228 } 1229 // Let the default code handle it. 1230 break; 1231 } 1232 } 1233 1234 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 1235 } 1236 1237 #undef CASE_VFMA_CHANGE_OPCODE_SPLATS 1238 #undef CASE_VFMA_CHANGE_OPCODE_LMULS 1239 #undef CASE_VFMA_CHANGE_OPCODE_COMMON 1240 #undef CASE_VFMA_SPLATS 1241 #undef CASE_VFMA_OPCODE_LMULS 1242 #undef CASE_VFMA_OPCODE_COMMON 1243 1244 Register RISCVInstrInfo::getVLENFactoredAmount(MachineFunction &MF, 1245 MachineBasicBlock &MBB, 1246 MachineBasicBlock::iterator II, 1247 int64_t Amount) const { 1248 assert(Amount > 0 && "There is no need to get VLEN scaled value."); 1249 assert(Amount % 8 == 0 && 1250 "Reserve the stack by the multiple of one vector size."); 1251 1252 MachineRegisterInfo &MRI = MF.getRegInfo(); 1253 const RISCVInstrInfo *TII = MF.getSubtarget<RISCVSubtarget>().getInstrInfo(); 1254 DebugLoc DL = II->getDebugLoc(); 1255 int64_t NumOfVReg = Amount / 8; 1256 1257 Register SizeOfVector = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1258 BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), SizeOfVector); 1259 Register FactorRegister = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1260 assert(isInt<12>(NumOfVReg) && 1261 "Expect the number of vector registers within 12-bits."); 1262 if (isPowerOf2_32(NumOfVReg)) { 1263 uint32_t ShiftAmount = Log2_32(NumOfVReg); 1264 if (ShiftAmount == 0) 1265 return SizeOfVector; 1266 BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), FactorRegister) 1267 .addReg(SizeOfVector, RegState::Kill) 1268 .addImm(ShiftAmount); 1269 } else { 1270 Register VN = MRI.createVirtualRegister(&RISCV::GPRRegClass); 1271 BuildMI(MBB, II, DL, TII->get(RISCV::ADDI), VN) 1272 .addReg(RISCV::X0) 1273 .addImm(NumOfVReg); 1274 if (!MF.getSubtarget<RISCVSubtarget>().hasStdExtM()) 1275 MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{ 1276 MF.getFunction(), 1277 "M-extension must be enabled to calculate the vscaled size/offset."}); 1278 BuildMI(MBB, II, DL, TII->get(RISCV::MUL), FactorRegister) 1279 .addReg(SizeOfVector, RegState::Kill) 1280 .addReg(VN, RegState::Kill); 1281 } 1282 1283 return FactorRegister; 1284 } 1285 1286 Optional<std::pair<unsigned, unsigned>> 1287 RISCVInstrInfo::isRVVSpillForZvlsseg(unsigned Opcode) const { 1288 switch (Opcode) { 1289 default: 1290 return None; 1291 case RISCV::PseudoVSPILL2_M1: 1292 case RISCV::PseudoVRELOAD2_M1: 1293 return std::make_pair(2u, 1u); 1294 case RISCV::PseudoVSPILL2_M2: 1295 case RISCV::PseudoVRELOAD2_M2: 1296 return std::make_pair(2u, 2u); 1297 case RISCV::PseudoVSPILL2_M4: 1298 case RISCV::PseudoVRELOAD2_M4: 1299 return std::make_pair(2u, 4u); 1300 case RISCV::PseudoVSPILL3_M1: 1301 case RISCV::PseudoVRELOAD3_M1: 1302 return std::make_pair(3u, 1u); 1303 case RISCV::PseudoVSPILL3_M2: 1304 case RISCV::PseudoVRELOAD3_M2: 1305 return std::make_pair(3u, 2u); 1306 case RISCV::PseudoVSPILL4_M1: 1307 case RISCV::PseudoVRELOAD4_M1: 1308 return std::make_pair(4u, 1u); 1309 case RISCV::PseudoVSPILL4_M2: 1310 case RISCV::PseudoVRELOAD4_M2: 1311 return std::make_pair(4u, 2u); 1312 case RISCV::PseudoVSPILL5_M1: 1313 case RISCV::PseudoVRELOAD5_M1: 1314 return std::make_pair(5u, 1u); 1315 case RISCV::PseudoVSPILL6_M1: 1316 case RISCV::PseudoVRELOAD6_M1: 1317 return std::make_pair(6u, 1u); 1318 case RISCV::PseudoVSPILL7_M1: 1319 case RISCV::PseudoVRELOAD7_M1: 1320 return std::make_pair(7u, 1u); 1321 case RISCV::PseudoVSPILL8_M1: 1322 case RISCV::PseudoVRELOAD8_M1: 1323 return std::make_pair(8u, 1u); 1324 } 1325 } 1326