1 //===- AMDGPUDisassembler.cpp - Disassembler for AMDGPU ISA ---------------===// 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 //===----------------------------------------------------------------------===// 10 // 11 /// \file 12 /// 13 /// This file contains definition for AMDGPU ISA disassembler 14 // 15 //===----------------------------------------------------------------------===// 16 17 // ToDo: What to do with instruction suffixes (v_mov_b32 vs v_mov_b32_e32)? 18 19 #include "Disassembler/AMDGPUDisassembler.h" 20 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 21 #include "TargetInfo/AMDGPUTargetInfo.h" 22 #include "Utils/AMDGPUBaseInfo.h" 23 #include "llvm-c/DisassemblerTypes.h" 24 #include "llvm/MC/MCAsmInfo.h" 25 #include "llvm/MC/MCContext.h" 26 #include "llvm/MC/MCExpr.h" 27 #include "llvm/MC/MCFixedLenDisassembler.h" 28 #include "llvm/Support/AMDHSAKernelDescriptor.h" 29 #include "llvm/Support/TargetRegistry.h" 30 31 using namespace llvm; 32 33 #define DEBUG_TYPE "amdgpu-disassembler" 34 35 #define SGPR_MAX \ 36 (isGFX10Plus() ? AMDGPU::EncValues::SGPR_MAX_GFX10 \ 37 : AMDGPU::EncValues::SGPR_MAX_SI) 38 39 using DecodeStatus = llvm::MCDisassembler::DecodeStatus; 40 41 AMDGPUDisassembler::AMDGPUDisassembler(const MCSubtargetInfo &STI, 42 MCContext &Ctx, 43 MCInstrInfo const *MCII) : 44 MCDisassembler(STI, Ctx), MCII(MCII), MRI(*Ctx.getRegisterInfo()), 45 TargetMaxInstBytes(Ctx.getAsmInfo()->getMaxInstLength(&STI)) { 46 47 // ToDo: AMDGPUDisassembler supports only VI ISA. 48 if (!STI.getFeatureBits()[AMDGPU::FeatureGCN3Encoding] && !isGFX10Plus()) 49 report_fatal_error("Disassembly not yet supported for subtarget"); 50 } 51 52 inline static MCDisassembler::DecodeStatus 53 addOperand(MCInst &Inst, const MCOperand& Opnd) { 54 Inst.addOperand(Opnd); 55 return Opnd.isValid() ? 56 MCDisassembler::Success : 57 MCDisassembler::Fail; 58 } 59 60 static int insertNamedMCOperand(MCInst &MI, const MCOperand &Op, 61 uint16_t NameIdx) { 62 int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), NameIdx); 63 if (OpIdx != -1) { 64 auto I = MI.begin(); 65 std::advance(I, OpIdx); 66 MI.insert(I, Op); 67 } 68 return OpIdx; 69 } 70 71 static DecodeStatus decodeSoppBrTarget(MCInst &Inst, unsigned Imm, 72 uint64_t Addr, const void *Decoder) { 73 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 74 75 // Our branches take a simm16, but we need two extra bits to account for the 76 // factor of 4. 77 APInt SignedOffset(18, Imm * 4, true); 78 int64_t Offset = (SignedOffset.sext(64) + 4 + Addr).getSExtValue(); 79 80 if (DAsm->tryAddingSymbolicOperand(Inst, Offset, Addr, true, 2, 2)) 81 return MCDisassembler::Success; 82 return addOperand(Inst, MCOperand::createImm(Imm)); 83 } 84 85 static DecodeStatus decodeSMEMOffset(MCInst &Inst, unsigned Imm, 86 uint64_t Addr, const void *Decoder) { 87 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 88 int64_t Offset; 89 if (DAsm->isVI()) { // VI supports 20-bit unsigned offsets. 90 Offset = Imm & 0xFFFFF; 91 } else { // GFX9+ supports 21-bit signed offsets. 92 Offset = SignExtend64<21>(Imm); 93 } 94 return addOperand(Inst, MCOperand::createImm(Offset)); 95 } 96 97 static DecodeStatus decodeBoolReg(MCInst &Inst, unsigned Val, 98 uint64_t Addr, const void *Decoder) { 99 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 100 return addOperand(Inst, DAsm->decodeBoolReg(Val)); 101 } 102 103 #define DECODE_OPERAND(StaticDecoderName, DecoderName) \ 104 static DecodeStatus StaticDecoderName(MCInst &Inst, \ 105 unsigned Imm, \ 106 uint64_t /*Addr*/, \ 107 const void *Decoder) { \ 108 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); \ 109 return addOperand(Inst, DAsm->DecoderName(Imm)); \ 110 } 111 112 #define DECODE_OPERAND_REG(RegClass) \ 113 DECODE_OPERAND(Decode##RegClass##RegisterClass, decodeOperand_##RegClass) 114 115 DECODE_OPERAND_REG(VGPR_32) 116 DECODE_OPERAND_REG(VRegOrLds_32) 117 DECODE_OPERAND_REG(VS_32) 118 DECODE_OPERAND_REG(VS_64) 119 DECODE_OPERAND_REG(VS_128) 120 121 DECODE_OPERAND_REG(VReg_64) 122 DECODE_OPERAND_REG(VReg_96) 123 DECODE_OPERAND_REG(VReg_128) 124 DECODE_OPERAND_REG(VReg_256) 125 DECODE_OPERAND_REG(VReg_512) 126 DECODE_OPERAND_REG(VReg_1024) 127 128 DECODE_OPERAND_REG(SReg_32) 129 DECODE_OPERAND_REG(SReg_32_XM0_XEXEC) 130 DECODE_OPERAND_REG(SReg_32_XEXEC_HI) 131 DECODE_OPERAND_REG(SRegOrLds_32) 132 DECODE_OPERAND_REG(SReg_64) 133 DECODE_OPERAND_REG(SReg_64_XEXEC) 134 DECODE_OPERAND_REG(SReg_128) 135 DECODE_OPERAND_REG(SReg_256) 136 DECODE_OPERAND_REG(SReg_512) 137 138 DECODE_OPERAND_REG(AGPR_32) 139 DECODE_OPERAND_REG(AReg_64) 140 DECODE_OPERAND_REG(AReg_128) 141 DECODE_OPERAND_REG(AReg_256) 142 DECODE_OPERAND_REG(AReg_512) 143 DECODE_OPERAND_REG(AReg_1024) 144 DECODE_OPERAND_REG(AV_32) 145 DECODE_OPERAND_REG(AV_64) 146 147 static DecodeStatus decodeOperand_VSrc16(MCInst &Inst, 148 unsigned Imm, 149 uint64_t Addr, 150 const void *Decoder) { 151 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 152 return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm)); 153 } 154 155 static DecodeStatus decodeOperand_VSrcV216(MCInst &Inst, 156 unsigned Imm, 157 uint64_t Addr, 158 const void *Decoder) { 159 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 160 return addOperand(Inst, DAsm->decodeOperand_VSrcV216(Imm)); 161 } 162 163 static DecodeStatus decodeOperand_VSrcV232(MCInst &Inst, 164 unsigned Imm, 165 uint64_t Addr, 166 const void *Decoder) { 167 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 168 return addOperand(Inst, DAsm->decodeOperand_VSrcV232(Imm)); 169 } 170 171 static DecodeStatus decodeOperand_VS_16(MCInst &Inst, 172 unsigned Imm, 173 uint64_t Addr, 174 const void *Decoder) { 175 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 176 return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm)); 177 } 178 179 static DecodeStatus decodeOperand_VS_32(MCInst &Inst, 180 unsigned Imm, 181 uint64_t Addr, 182 const void *Decoder) { 183 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 184 return addOperand(Inst, DAsm->decodeOperand_VS_32(Imm)); 185 } 186 187 static DecodeStatus decodeOperand_AReg_64(MCInst &Inst, 188 unsigned Imm, 189 uint64_t Addr, 190 const void *Decoder) { 191 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 192 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm | 512)); 193 } 194 195 static DecodeStatus decodeOperand_AReg_128(MCInst &Inst, 196 unsigned Imm, 197 uint64_t Addr, 198 const void *Decoder) { 199 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 200 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW128, Imm | 512)); 201 } 202 203 static DecodeStatus decodeOperand_AReg_256(MCInst &Inst, 204 unsigned Imm, 205 uint64_t Addr, 206 const void *Decoder) { 207 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 208 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW256, Imm | 512)); 209 } 210 211 static DecodeStatus decodeOperand_AReg_512(MCInst &Inst, 212 unsigned Imm, 213 uint64_t Addr, 214 const void *Decoder) { 215 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 216 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW512, Imm | 512)); 217 } 218 219 static DecodeStatus decodeOperand_AReg_1024(MCInst &Inst, 220 unsigned Imm, 221 uint64_t Addr, 222 const void *Decoder) { 223 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 224 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW1024, Imm | 512)); 225 } 226 227 static DecodeStatus decodeOperand_VReg_64(MCInst &Inst, 228 unsigned Imm, 229 uint64_t Addr, 230 const void *Decoder) { 231 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 232 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm)); 233 } 234 235 static DecodeStatus decodeOperand_VReg_128(MCInst &Inst, 236 unsigned Imm, 237 uint64_t Addr, 238 const void *Decoder) { 239 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 240 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW128, Imm)); 241 } 242 243 static DecodeStatus decodeOperand_VReg_256(MCInst &Inst, 244 unsigned Imm, 245 uint64_t Addr, 246 const void *Decoder) { 247 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 248 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW256, Imm)); 249 } 250 251 static DecodeStatus decodeOperand_VReg_512(MCInst &Inst, 252 unsigned Imm, 253 uint64_t Addr, 254 const void *Decoder) { 255 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 256 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW512, Imm)); 257 } 258 259 static DecodeStatus decodeOperand_VReg_1024(MCInst &Inst, 260 unsigned Imm, 261 uint64_t Addr, 262 const void *Decoder) { 263 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 264 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW1024, Imm)); 265 } 266 267 static bool IsAGPROperand(const MCInst &Inst, int OpIdx, 268 const MCRegisterInfo *MRI) { 269 if (OpIdx < 0) 270 return false; 271 272 const MCOperand &Op = Inst.getOperand(OpIdx); 273 if (!Op.isReg()) 274 return false; 275 276 unsigned Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0); 277 auto Reg = Sub ? Sub : Op.getReg(); 278 return Reg >= AMDGPU::AGPR0 && Reg <= AMDGPU::AGPR255; 279 } 280 281 static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, 282 unsigned Imm, 283 AMDGPUDisassembler::OpWidthTy Opw, 284 const void *Decoder) { 285 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 286 if (!DAsm->isGFX90A()) { 287 Imm &= 511; 288 } else { 289 // If atomic has both vdata and vdst their register classes are tied. 290 // The bit is decoded along with the vdst, first operand. We need to 291 // change register class to AGPR if vdst was AGPR. 292 // If a DS instruction has both data0 and data1 their register classes 293 // are also tied. 294 unsigned Opc = Inst.getOpcode(); 295 uint64_t TSFlags = DAsm->getMCII()->get(Opc).TSFlags; 296 uint16_t DataNameIdx = (TSFlags & SIInstrFlags::DS) ? AMDGPU::OpName::data0 297 : AMDGPU::OpName::vdata; 298 const MCRegisterInfo *MRI = DAsm->getContext().getRegisterInfo(); 299 int DataIdx = AMDGPU::getNamedOperandIdx(Opc, DataNameIdx); 300 if ((int)Inst.getNumOperands() == DataIdx) { 301 int DstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst); 302 if (IsAGPROperand(Inst, DstIdx, MRI)) 303 Imm |= 512; 304 } 305 306 if (TSFlags & SIInstrFlags::DS) { 307 int Data2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1); 308 if ((int)Inst.getNumOperands() == Data2Idx && 309 IsAGPROperand(Inst, DataIdx, MRI)) 310 Imm |= 512; 311 } 312 } 313 return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256)); 314 } 315 316 static DecodeStatus DecodeAVLdSt_32RegisterClass(MCInst &Inst, 317 unsigned Imm, 318 uint64_t Addr, 319 const void *Decoder) { 320 return decodeOperand_AVLdSt_Any(Inst, Imm, 321 AMDGPUDisassembler::OPW32, Decoder); 322 } 323 324 static DecodeStatus DecodeAVLdSt_64RegisterClass(MCInst &Inst, 325 unsigned Imm, 326 uint64_t Addr, 327 const void *Decoder) { 328 return decodeOperand_AVLdSt_Any(Inst, Imm, 329 AMDGPUDisassembler::OPW64, Decoder); 330 } 331 332 static DecodeStatus DecodeAVLdSt_96RegisterClass(MCInst &Inst, 333 unsigned Imm, 334 uint64_t Addr, 335 const void *Decoder) { 336 return decodeOperand_AVLdSt_Any(Inst, Imm, 337 AMDGPUDisassembler::OPW96, Decoder); 338 } 339 340 static DecodeStatus DecodeAVLdSt_128RegisterClass(MCInst &Inst, 341 unsigned Imm, 342 uint64_t Addr, 343 const void *Decoder) { 344 return decodeOperand_AVLdSt_Any(Inst, Imm, 345 AMDGPUDisassembler::OPW128, Decoder); 346 } 347 348 static DecodeStatus decodeOperand_SReg_32(MCInst &Inst, 349 unsigned Imm, 350 uint64_t Addr, 351 const void *Decoder) { 352 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 353 return addOperand(Inst, DAsm->decodeOperand_SReg_32(Imm)); 354 } 355 356 static DecodeStatus decodeOperand_VGPR_32(MCInst &Inst, 357 unsigned Imm, 358 uint64_t Addr, 359 const void *Decoder) { 360 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); 361 return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW32, Imm)); 362 } 363 364 #define DECODE_SDWA(DecName) \ 365 DECODE_OPERAND(decodeSDWA##DecName, decodeSDWA##DecName) 366 367 DECODE_SDWA(Src32) 368 DECODE_SDWA(Src16) 369 DECODE_SDWA(VopcDst) 370 371 #include "AMDGPUGenDisassemblerTables.inc" 372 373 //===----------------------------------------------------------------------===// 374 // 375 //===----------------------------------------------------------------------===// 376 377 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) { 378 assert(Bytes.size() >= sizeof(T)); 379 const auto Res = support::endian::read<T, support::endianness::little>(Bytes.data()); 380 Bytes = Bytes.slice(sizeof(T)); 381 return Res; 382 } 383 384 DecodeStatus AMDGPUDisassembler::tryDecodeInst(const uint8_t* Table, 385 MCInst &MI, 386 uint64_t Inst, 387 uint64_t Address) const { 388 assert(MI.getOpcode() == 0); 389 assert(MI.getNumOperands() == 0); 390 MCInst TmpInst; 391 HasLiteral = false; 392 const auto SavedBytes = Bytes; 393 if (decodeInstruction(Table, TmpInst, Inst, Address, this, STI)) { 394 MI = TmpInst; 395 return MCDisassembler::Success; 396 } 397 Bytes = SavedBytes; 398 return MCDisassembler::Fail; 399 } 400 401 // The disassembler is greedy, so we need to check FI operand value to 402 // not parse a dpp if the correct literal is not set. For dpp16 the 403 // autogenerated decoder checks the dpp literal 404 static bool isValidDPP8(const MCInst &MI) { 405 using namespace llvm::AMDGPU::DPP; 406 int FiIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::fi); 407 assert(FiIdx != -1); 408 if ((unsigned)FiIdx >= MI.getNumOperands()) 409 return false; 410 unsigned Fi = MI.getOperand(FiIdx).getImm(); 411 return Fi == DPP8_FI_0 || Fi == DPP8_FI_1; 412 } 413 414 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size, 415 ArrayRef<uint8_t> Bytes_, 416 uint64_t Address, 417 raw_ostream &CS) const { 418 CommentStream = &CS; 419 bool IsSDWA = false; 420 421 unsigned MaxInstBytesNum = std::min((size_t)TargetMaxInstBytes, Bytes_.size()); 422 Bytes = Bytes_.slice(0, MaxInstBytesNum); 423 424 DecodeStatus Res = MCDisassembler::Fail; 425 do { 426 // ToDo: better to switch encoding length using some bit predicate 427 // but it is unknown yet, so try all we can 428 429 // Try to decode DPP and SDWA first to solve conflict with VOP1 and VOP2 430 // encodings 431 if (Bytes.size() >= 8) { 432 const uint64_t QW = eatBytes<uint64_t>(Bytes); 433 434 if (STI.getFeatureBits()[AMDGPU::FeatureGFX10_BEncoding]) { 435 Res = tryDecodeInst(DecoderTableGFX10_B64, MI, QW, Address); 436 if (Res) { 437 if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dpp8) 438 == -1) 439 break; 440 if (convertDPP8Inst(MI) == MCDisassembler::Success) 441 break; 442 MI = MCInst(); // clear 443 } 444 } 445 446 Res = tryDecodeInst(DecoderTableDPP864, MI, QW, Address); 447 if (Res && convertDPP8Inst(MI) == MCDisassembler::Success) 448 break; 449 450 MI = MCInst(); // clear 451 452 Res = tryDecodeInst(DecoderTableDPP64, MI, QW, Address); 453 if (Res) break; 454 455 Res = tryDecodeInst(DecoderTableSDWA64, MI, QW, Address); 456 if (Res) { IsSDWA = true; break; } 457 458 Res = tryDecodeInst(DecoderTableSDWA964, MI, QW, Address); 459 if (Res) { IsSDWA = true; break; } 460 461 Res = tryDecodeInst(DecoderTableSDWA1064, MI, QW, Address); 462 if (Res) { IsSDWA = true; break; } 463 464 if (STI.getFeatureBits()[AMDGPU::FeatureUnpackedD16VMem]) { 465 Res = tryDecodeInst(DecoderTableGFX80_UNPACKED64, MI, QW, Address); 466 if (Res) 467 break; 468 } 469 470 // Some GFX9 subtargets repurposed the v_mad_mix_f32, v_mad_mixlo_f16 and 471 // v_mad_mixhi_f16 for FMA variants. Try to decode using this special 472 // table first so we print the correct name. 473 if (STI.getFeatureBits()[AMDGPU::FeatureFmaMixInsts]) { 474 Res = tryDecodeInst(DecoderTableGFX9_DL64, MI, QW, Address); 475 if (Res) 476 break; 477 } 478 } 479 480 // Reinitialize Bytes as DPP64 could have eaten too much 481 Bytes = Bytes_.slice(0, MaxInstBytesNum); 482 483 // Try decode 32-bit instruction 484 if (Bytes.size() < 4) break; 485 const uint32_t DW = eatBytes<uint32_t>(Bytes); 486 Res = tryDecodeInst(DecoderTableGFX832, MI, DW, Address); 487 if (Res) break; 488 489 Res = tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address); 490 if (Res) break; 491 492 Res = tryDecodeInst(DecoderTableGFX932, MI, DW, Address); 493 if (Res) break; 494 495 if (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]) { 496 Res = tryDecodeInst(DecoderTableGFX90A32, MI, DW, Address); 497 if (Res) 498 break; 499 } 500 501 if (STI.getFeatureBits()[AMDGPU::FeatureGFX10_BEncoding]) { 502 Res = tryDecodeInst(DecoderTableGFX10_B32, MI, DW, Address); 503 if (Res) break; 504 } 505 506 Res = tryDecodeInst(DecoderTableGFX1032, MI, DW, Address); 507 if (Res) break; 508 509 if (Bytes.size() < 4) break; 510 const uint64_t QW = ((uint64_t)eatBytes<uint32_t>(Bytes) << 32) | DW; 511 512 if (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]) { 513 Res = tryDecodeInst(DecoderTableGFX90A64, MI, QW, Address); 514 if (Res) 515 break; 516 } 517 518 Res = tryDecodeInst(DecoderTableGFX864, MI, QW, Address); 519 if (Res) break; 520 521 Res = tryDecodeInst(DecoderTableAMDGPU64, MI, QW, Address); 522 if (Res) break; 523 524 Res = tryDecodeInst(DecoderTableGFX964, MI, QW, Address); 525 if (Res) break; 526 527 Res = tryDecodeInst(DecoderTableGFX1064, MI, QW, Address); 528 } while (false); 529 530 if (Res && (MI.getOpcode() == AMDGPU::V_MAC_F32_e64_vi || 531 MI.getOpcode() == AMDGPU::V_MAC_F32_e64_gfx6_gfx7 || 532 MI.getOpcode() == AMDGPU::V_MAC_F32_e64_gfx10 || 533 MI.getOpcode() == AMDGPU::V_MAC_LEGACY_F32_e64_gfx6_gfx7 || 534 MI.getOpcode() == AMDGPU::V_MAC_LEGACY_F32_e64_gfx10 || 535 MI.getOpcode() == AMDGPU::V_MAC_F16_e64_vi || 536 MI.getOpcode() == AMDGPU::V_FMAC_F64_e64_gfx90a || 537 MI.getOpcode() == AMDGPU::V_FMAC_F32_e64_vi || 538 MI.getOpcode() == AMDGPU::V_FMAC_F32_e64_gfx10 || 539 MI.getOpcode() == AMDGPU::V_FMAC_LEGACY_F32_e64_gfx10 || 540 MI.getOpcode() == AMDGPU::V_FMAC_F16_e64_gfx10)) { 541 // Insert dummy unused src2_modifiers. 542 insertNamedMCOperand(MI, MCOperand::createImm(0), 543 AMDGPU::OpName::src2_modifiers); 544 } 545 546 if (Res && (MCII->get(MI.getOpcode()).TSFlags & 547 (SIInstrFlags::MUBUF | SIInstrFlags::FLAT | SIInstrFlags::SMRD))) { 548 int CPolPos = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 549 AMDGPU::OpName::cpol); 550 if (CPolPos != -1) { 551 unsigned CPol = 552 (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::IsAtomicRet) ? 553 AMDGPU::CPol::GLC : 0; 554 if (MI.getNumOperands() <= (unsigned)CPolPos) { 555 insertNamedMCOperand(MI, MCOperand::createImm(CPol), 556 AMDGPU::OpName::cpol); 557 } else if (CPol) { 558 MI.getOperand(CPolPos).setImm(MI.getOperand(CPolPos).getImm() | CPol); 559 } 560 } 561 } 562 563 if (Res && (MCII->get(MI.getOpcode()).TSFlags & 564 (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF)) && 565 (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts])) { 566 // GFX90A lost TFE, its place is occupied by ACC. 567 int TFEOpIdx = 568 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::tfe); 569 if (TFEOpIdx != -1) { 570 auto TFEIter = MI.begin(); 571 std::advance(TFEIter, TFEOpIdx); 572 MI.insert(TFEIter, MCOperand::createImm(0)); 573 } 574 } 575 576 if (Res && (MCII->get(MI.getOpcode()).TSFlags & 577 (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF))) { 578 int SWZOpIdx = 579 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::swz); 580 if (SWZOpIdx != -1) { 581 auto SWZIter = MI.begin(); 582 std::advance(SWZIter, SWZOpIdx); 583 MI.insert(SWZIter, MCOperand::createImm(0)); 584 } 585 } 586 587 if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::MIMG)) { 588 int VAddr0Idx = 589 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0); 590 int RsrcIdx = 591 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc); 592 unsigned NSAArgs = RsrcIdx - VAddr0Idx - 1; 593 if (VAddr0Idx >= 0 && NSAArgs > 0) { 594 unsigned NSAWords = (NSAArgs + 3) / 4; 595 if (Bytes.size() < 4 * NSAWords) { 596 Res = MCDisassembler::Fail; 597 } else { 598 for (unsigned i = 0; i < NSAArgs; ++i) { 599 MI.insert(MI.begin() + VAddr0Idx + 1 + i, 600 decodeOperand_VGPR_32(Bytes[i])); 601 } 602 Bytes = Bytes.slice(4 * NSAWords); 603 } 604 } 605 606 if (Res) 607 Res = convertMIMGInst(MI); 608 } 609 610 if (Res && IsSDWA) 611 Res = convertSDWAInst(MI); 612 613 int VDstIn_Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 614 AMDGPU::OpName::vdst_in); 615 if (VDstIn_Idx != -1) { 616 int Tied = MCII->get(MI.getOpcode()).getOperandConstraint(VDstIn_Idx, 617 MCOI::OperandConstraint::TIED_TO); 618 if (Tied != -1 && (MI.getNumOperands() <= (unsigned)VDstIn_Idx || 619 !MI.getOperand(VDstIn_Idx).isReg() || 620 MI.getOperand(VDstIn_Idx).getReg() != MI.getOperand(Tied).getReg())) { 621 if (MI.getNumOperands() > (unsigned)VDstIn_Idx) 622 MI.erase(&MI.getOperand(VDstIn_Idx)); 623 insertNamedMCOperand(MI, 624 MCOperand::createReg(MI.getOperand(Tied).getReg()), 625 AMDGPU::OpName::vdst_in); 626 } 627 } 628 629 // if the opcode was not recognized we'll assume a Size of 4 bytes 630 // (unless there are fewer bytes left) 631 Size = Res ? (MaxInstBytesNum - Bytes.size()) 632 : std::min((size_t)4, Bytes_.size()); 633 return Res; 634 } 635 636 DecodeStatus AMDGPUDisassembler::convertSDWAInst(MCInst &MI) const { 637 if (STI.getFeatureBits()[AMDGPU::FeatureGFX9] || 638 STI.getFeatureBits()[AMDGPU::FeatureGFX10]) { 639 if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst) != -1) 640 // VOPC - insert clamp 641 insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::clamp); 642 } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) { 643 int SDst = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst); 644 if (SDst != -1) { 645 // VOPC - insert VCC register as sdst 646 insertNamedMCOperand(MI, createRegOperand(AMDGPU::VCC), 647 AMDGPU::OpName::sdst); 648 } else { 649 // VOP1/2 - insert omod if present in instruction 650 insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::omod); 651 } 652 } 653 return MCDisassembler::Success; 654 } 655 656 // We must check FI == literal to reject not genuine dpp8 insts, and we must 657 // first add optional MI operands to check FI 658 DecodeStatus AMDGPUDisassembler::convertDPP8Inst(MCInst &MI) const { 659 unsigned Opc = MI.getOpcode(); 660 unsigned DescNumOps = MCII->get(Opc).getNumOperands(); 661 662 // Insert dummy unused src modifiers. 663 if (MI.getNumOperands() < DescNumOps && 664 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers) != -1) 665 insertNamedMCOperand(MI, MCOperand::createImm(0), 666 AMDGPU::OpName::src0_modifiers); 667 668 if (MI.getNumOperands() < DescNumOps && 669 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers) != -1) 670 insertNamedMCOperand(MI, MCOperand::createImm(0), 671 AMDGPU::OpName::src1_modifiers); 672 673 return isValidDPP8(MI) ? MCDisassembler::Success : MCDisassembler::SoftFail; 674 } 675 676 // Note that before gfx10, the MIMG encoding provided no information about 677 // VADDR size. Consequently, decoded instructions always show address as if it 678 // has 1 dword, which could be not really so. 679 DecodeStatus AMDGPUDisassembler::convertMIMGInst(MCInst &MI) const { 680 681 int VDstIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 682 AMDGPU::OpName::vdst); 683 684 int VDataIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 685 AMDGPU::OpName::vdata); 686 int VAddr0Idx = 687 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0); 688 int DMaskIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 689 AMDGPU::OpName::dmask); 690 691 int TFEIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 692 AMDGPU::OpName::tfe); 693 int D16Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 694 AMDGPU::OpName::d16); 695 696 assert(VDataIdx != -1); 697 if (DMaskIdx == -1 || TFEIdx == -1) {// intersect_ray 698 if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::a16) > -1) { 699 assert(MI.getOpcode() == AMDGPU::IMAGE_BVH_INTERSECT_RAY_a16_sa || 700 MI.getOpcode() == AMDGPU::IMAGE_BVH_INTERSECT_RAY_a16_nsa || 701 MI.getOpcode() == AMDGPU::IMAGE_BVH64_INTERSECT_RAY_a16_sa || 702 MI.getOpcode() == AMDGPU::IMAGE_BVH64_INTERSECT_RAY_a16_nsa); 703 addOperand(MI, MCOperand::createImm(1)); 704 } 705 return MCDisassembler::Success; 706 } 707 708 const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(MI.getOpcode()); 709 bool IsAtomic = (VDstIdx != -1); 710 bool IsGather4 = MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::Gather4; 711 712 bool IsNSA = false; 713 unsigned AddrSize = Info->VAddrDwords; 714 715 if (STI.getFeatureBits()[AMDGPU::FeatureGFX10]) { 716 unsigned DimIdx = 717 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dim); 718 int A16Idx = 719 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::a16); 720 const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode = 721 AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode); 722 const AMDGPU::MIMGDimInfo *Dim = 723 AMDGPU::getMIMGDimInfoByEncoding(MI.getOperand(DimIdx).getImm()); 724 const bool IsA16 = (A16Idx != -1 && MI.getOperand(A16Idx).getImm()); 725 726 AddrSize = 727 AMDGPU::getAddrSizeMIMGOp(BaseOpcode, Dim, IsA16, AMDGPU::hasG16(STI)); 728 729 IsNSA = Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA; 730 if (!IsNSA) { 731 if (AddrSize > 8) 732 AddrSize = 16; 733 else if (AddrSize > 4) 734 AddrSize = 8; 735 } else { 736 if (AddrSize > Info->VAddrDwords) { 737 // The NSA encoding does not contain enough operands for the combination 738 // of base opcode / dimension. Should this be an error? 739 return MCDisassembler::Success; 740 } 741 } 742 } 743 744 unsigned DMask = MI.getOperand(DMaskIdx).getImm() & 0xf; 745 unsigned DstSize = IsGather4 ? 4 : std::max(countPopulation(DMask), 1u); 746 747 bool D16 = D16Idx >= 0 && MI.getOperand(D16Idx).getImm(); 748 if (D16 && AMDGPU::hasPackedD16(STI)) { 749 DstSize = (DstSize + 1) / 2; 750 } 751 752 if (TFEIdx != -1 && MI.getOperand(TFEIdx).getImm()) 753 DstSize += 1; 754 755 if (DstSize == Info->VDataDwords && AddrSize == Info->VAddrDwords) 756 return MCDisassembler::Success; 757 758 int NewOpcode = 759 AMDGPU::getMIMGOpcode(Info->BaseOpcode, Info->MIMGEncoding, DstSize, AddrSize); 760 if (NewOpcode == -1) 761 return MCDisassembler::Success; 762 763 // Widen the register to the correct number of enabled channels. 764 unsigned NewVdata = AMDGPU::NoRegister; 765 if (DstSize != Info->VDataDwords) { 766 auto DataRCID = MCII->get(NewOpcode).OpInfo[VDataIdx].RegClass; 767 768 // Get first subregister of VData 769 unsigned Vdata0 = MI.getOperand(VDataIdx).getReg(); 770 unsigned VdataSub0 = MRI.getSubReg(Vdata0, AMDGPU::sub0); 771 Vdata0 = (VdataSub0 != 0)? VdataSub0 : Vdata0; 772 773 NewVdata = MRI.getMatchingSuperReg(Vdata0, AMDGPU::sub0, 774 &MRI.getRegClass(DataRCID)); 775 if (NewVdata == AMDGPU::NoRegister) { 776 // It's possible to encode this such that the low register + enabled 777 // components exceeds the register count. 778 return MCDisassembler::Success; 779 } 780 } 781 782 unsigned NewVAddr0 = AMDGPU::NoRegister; 783 if (STI.getFeatureBits()[AMDGPU::FeatureGFX10] && !IsNSA && 784 AddrSize != Info->VAddrDwords) { 785 unsigned VAddr0 = MI.getOperand(VAddr0Idx).getReg(); 786 unsigned VAddrSub0 = MRI.getSubReg(VAddr0, AMDGPU::sub0); 787 VAddr0 = (VAddrSub0 != 0) ? VAddrSub0 : VAddr0; 788 789 auto AddrRCID = MCII->get(NewOpcode).OpInfo[VAddr0Idx].RegClass; 790 NewVAddr0 = MRI.getMatchingSuperReg(VAddr0, AMDGPU::sub0, 791 &MRI.getRegClass(AddrRCID)); 792 if (NewVAddr0 == AMDGPU::NoRegister) 793 return MCDisassembler::Success; 794 } 795 796 MI.setOpcode(NewOpcode); 797 798 if (NewVdata != AMDGPU::NoRegister) { 799 MI.getOperand(VDataIdx) = MCOperand::createReg(NewVdata); 800 801 if (IsAtomic) { 802 // Atomic operations have an additional operand (a copy of data) 803 MI.getOperand(VDstIdx) = MCOperand::createReg(NewVdata); 804 } 805 } 806 807 if (NewVAddr0 != AMDGPU::NoRegister) { 808 MI.getOperand(VAddr0Idx) = MCOperand::createReg(NewVAddr0); 809 } else if (IsNSA) { 810 assert(AddrSize <= Info->VAddrDwords); 811 MI.erase(MI.begin() + VAddr0Idx + AddrSize, 812 MI.begin() + VAddr0Idx + Info->VAddrDwords); 813 } 814 815 return MCDisassembler::Success; 816 } 817 818 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const { 819 return getContext().getRegisterInfo()-> 820 getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]); 821 } 822 823 inline 824 MCOperand AMDGPUDisassembler::errOperand(unsigned V, 825 const Twine& ErrMsg) const { 826 *CommentStream << "Error: " + ErrMsg; 827 828 // ToDo: add support for error operands to MCInst.h 829 // return MCOperand::createError(V); 830 return MCOperand(); 831 } 832 833 inline 834 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const { 835 return MCOperand::createReg(AMDGPU::getMCReg(RegId, STI)); 836 } 837 838 inline 839 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID, 840 unsigned Val) const { 841 const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID]; 842 if (Val >= RegCl.getNumRegs()) 843 return errOperand(Val, Twine(getRegClassName(RegClassID)) + 844 ": unknown register " + Twine(Val)); 845 return createRegOperand(RegCl.getRegister(Val)); 846 } 847 848 inline 849 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID, 850 unsigned Val) const { 851 // ToDo: SI/CI have 104 SGPRs, VI - 102 852 // Valery: here we accepting as much as we can, let assembler sort it out 853 int shift = 0; 854 switch (SRegClassID) { 855 case AMDGPU::SGPR_32RegClassID: 856 case AMDGPU::TTMP_32RegClassID: 857 break; 858 case AMDGPU::SGPR_64RegClassID: 859 case AMDGPU::TTMP_64RegClassID: 860 shift = 1; 861 break; 862 case AMDGPU::SGPR_128RegClassID: 863 case AMDGPU::TTMP_128RegClassID: 864 // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in 865 // this bundle? 866 case AMDGPU::SGPR_256RegClassID: 867 case AMDGPU::TTMP_256RegClassID: 868 // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in 869 // this bundle? 870 case AMDGPU::SGPR_512RegClassID: 871 case AMDGPU::TTMP_512RegClassID: 872 shift = 2; 873 break; 874 // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in 875 // this bundle? 876 default: 877 llvm_unreachable("unhandled register class"); 878 } 879 880 if (Val % (1 << shift)) { 881 *CommentStream << "Warning: " << getRegClassName(SRegClassID) 882 << ": scalar reg isn't aligned " << Val; 883 } 884 885 return createRegOperand(SRegClassID, Val >> shift); 886 } 887 888 MCOperand AMDGPUDisassembler::decodeOperand_VS_32(unsigned Val) const { 889 return decodeSrcOp(OPW32, Val); 890 } 891 892 MCOperand AMDGPUDisassembler::decodeOperand_VS_64(unsigned Val) const { 893 return decodeSrcOp(OPW64, Val); 894 } 895 896 MCOperand AMDGPUDisassembler::decodeOperand_VS_128(unsigned Val) const { 897 return decodeSrcOp(OPW128, Val); 898 } 899 900 MCOperand AMDGPUDisassembler::decodeOperand_VSrc16(unsigned Val) const { 901 return decodeSrcOp(OPW16, Val); 902 } 903 904 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV216(unsigned Val) const { 905 return decodeSrcOp(OPWV216, Val); 906 } 907 908 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV232(unsigned Val) const { 909 return decodeSrcOp(OPWV232, Val); 910 } 911 912 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32(unsigned Val) const { 913 // Some instructions have operand restrictions beyond what the encoding 914 // allows. Some ordinarily VSrc_32 operands are VGPR_32, so clear the extra 915 // high bit. 916 Val &= 255; 917 918 return createRegOperand(AMDGPU::VGPR_32RegClassID, Val); 919 } 920 921 MCOperand AMDGPUDisassembler::decodeOperand_VRegOrLds_32(unsigned Val) const { 922 return decodeSrcOp(OPW32, Val); 923 } 924 925 MCOperand AMDGPUDisassembler::decodeOperand_AGPR_32(unsigned Val) const { 926 return createRegOperand(AMDGPU::AGPR_32RegClassID, Val & 255); 927 } 928 929 MCOperand AMDGPUDisassembler::decodeOperand_AReg_64(unsigned Val) const { 930 return createRegOperand(AMDGPU::AReg_64RegClassID, Val & 255); 931 } 932 933 MCOperand AMDGPUDisassembler::decodeOperand_AReg_128(unsigned Val) const { 934 return createRegOperand(AMDGPU::AReg_128RegClassID, Val & 255); 935 } 936 937 MCOperand AMDGPUDisassembler::decodeOperand_AReg_256(unsigned Val) const { 938 return createRegOperand(AMDGPU::AReg_256RegClassID, Val & 255); 939 } 940 941 MCOperand AMDGPUDisassembler::decodeOperand_AReg_512(unsigned Val) const { 942 return createRegOperand(AMDGPU::AReg_512RegClassID, Val & 255); 943 } 944 945 MCOperand AMDGPUDisassembler::decodeOperand_AReg_1024(unsigned Val) const { 946 return createRegOperand(AMDGPU::AReg_1024RegClassID, Val & 255); 947 } 948 949 MCOperand AMDGPUDisassembler::decodeOperand_AV_32(unsigned Val) const { 950 return decodeSrcOp(OPW32, Val); 951 } 952 953 MCOperand AMDGPUDisassembler::decodeOperand_AV_64(unsigned Val) const { 954 return decodeSrcOp(OPW64, Val); 955 } 956 957 MCOperand AMDGPUDisassembler::decodeOperand_VReg_64(unsigned Val) const { 958 return createRegOperand(AMDGPU::VReg_64RegClassID, Val); 959 } 960 961 MCOperand AMDGPUDisassembler::decodeOperand_VReg_96(unsigned Val) const { 962 return createRegOperand(AMDGPU::VReg_96RegClassID, Val); 963 } 964 965 MCOperand AMDGPUDisassembler::decodeOperand_VReg_128(unsigned Val) const { 966 return createRegOperand(AMDGPU::VReg_128RegClassID, Val); 967 } 968 969 MCOperand AMDGPUDisassembler::decodeOperand_VReg_256(unsigned Val) const { 970 return createRegOperand(AMDGPU::VReg_256RegClassID, Val); 971 } 972 973 MCOperand AMDGPUDisassembler::decodeOperand_VReg_512(unsigned Val) const { 974 return createRegOperand(AMDGPU::VReg_512RegClassID, Val); 975 } 976 977 MCOperand AMDGPUDisassembler::decodeOperand_VReg_1024(unsigned Val) const { 978 return createRegOperand(AMDGPU::VReg_1024RegClassID, Val); 979 } 980 981 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32(unsigned Val) const { 982 // table-gen generated disassembler doesn't care about operand types 983 // leaving only registry class so SSrc_32 operand turns into SReg_32 984 // and therefore we accept immediates and literals here as well 985 return decodeSrcOp(OPW32, Val); 986 } 987 988 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XM0_XEXEC( 989 unsigned Val) const { 990 // SReg_32_XM0 is SReg_32 without M0 or EXEC_LO/EXEC_HI 991 return decodeOperand_SReg_32(Val); 992 } 993 994 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XEXEC_HI( 995 unsigned Val) const { 996 // SReg_32_XM0 is SReg_32 without EXEC_HI 997 return decodeOperand_SReg_32(Val); 998 } 999 1000 MCOperand AMDGPUDisassembler::decodeOperand_SRegOrLds_32(unsigned Val) const { 1001 // table-gen generated disassembler doesn't care about operand types 1002 // leaving only registry class so SSrc_32 operand turns into SReg_32 1003 // and therefore we accept immediates and literals here as well 1004 return decodeSrcOp(OPW32, Val); 1005 } 1006 1007 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64(unsigned Val) const { 1008 return decodeSrcOp(OPW64, Val); 1009 } 1010 1011 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64_XEXEC(unsigned Val) const { 1012 return decodeSrcOp(OPW64, Val); 1013 } 1014 1015 MCOperand AMDGPUDisassembler::decodeOperand_SReg_128(unsigned Val) const { 1016 return decodeSrcOp(OPW128, Val); 1017 } 1018 1019 MCOperand AMDGPUDisassembler::decodeOperand_SReg_256(unsigned Val) const { 1020 return decodeDstOp(OPW256, Val); 1021 } 1022 1023 MCOperand AMDGPUDisassembler::decodeOperand_SReg_512(unsigned Val) const { 1024 return decodeDstOp(OPW512, Val); 1025 } 1026 1027 MCOperand AMDGPUDisassembler::decodeLiteralConstant() const { 1028 // For now all literal constants are supposed to be unsigned integer 1029 // ToDo: deal with signed/unsigned 64-bit integer constants 1030 // ToDo: deal with float/double constants 1031 if (!HasLiteral) { 1032 if (Bytes.size() < 4) { 1033 return errOperand(0, "cannot read literal, inst bytes left " + 1034 Twine(Bytes.size())); 1035 } 1036 HasLiteral = true; 1037 Literal = eatBytes<uint32_t>(Bytes); 1038 } 1039 return MCOperand::createImm(Literal); 1040 } 1041 1042 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) { 1043 using namespace AMDGPU::EncValues; 1044 1045 assert(Imm >= INLINE_INTEGER_C_MIN && Imm <= INLINE_INTEGER_C_MAX); 1046 return MCOperand::createImm((Imm <= INLINE_INTEGER_C_POSITIVE_MAX) ? 1047 (static_cast<int64_t>(Imm) - INLINE_INTEGER_C_MIN) : 1048 (INLINE_INTEGER_C_POSITIVE_MAX - static_cast<int64_t>(Imm))); 1049 // Cast prevents negative overflow. 1050 } 1051 1052 static int64_t getInlineImmVal32(unsigned Imm) { 1053 switch (Imm) { 1054 case 240: 1055 return FloatToBits(0.5f); 1056 case 241: 1057 return FloatToBits(-0.5f); 1058 case 242: 1059 return FloatToBits(1.0f); 1060 case 243: 1061 return FloatToBits(-1.0f); 1062 case 244: 1063 return FloatToBits(2.0f); 1064 case 245: 1065 return FloatToBits(-2.0f); 1066 case 246: 1067 return FloatToBits(4.0f); 1068 case 247: 1069 return FloatToBits(-4.0f); 1070 case 248: // 1 / (2 * PI) 1071 return 0x3e22f983; 1072 default: 1073 llvm_unreachable("invalid fp inline imm"); 1074 } 1075 } 1076 1077 static int64_t getInlineImmVal64(unsigned Imm) { 1078 switch (Imm) { 1079 case 240: 1080 return DoubleToBits(0.5); 1081 case 241: 1082 return DoubleToBits(-0.5); 1083 case 242: 1084 return DoubleToBits(1.0); 1085 case 243: 1086 return DoubleToBits(-1.0); 1087 case 244: 1088 return DoubleToBits(2.0); 1089 case 245: 1090 return DoubleToBits(-2.0); 1091 case 246: 1092 return DoubleToBits(4.0); 1093 case 247: 1094 return DoubleToBits(-4.0); 1095 case 248: // 1 / (2 * PI) 1096 return 0x3fc45f306dc9c882; 1097 default: 1098 llvm_unreachable("invalid fp inline imm"); 1099 } 1100 } 1101 1102 static int64_t getInlineImmVal16(unsigned Imm) { 1103 switch (Imm) { 1104 case 240: 1105 return 0x3800; 1106 case 241: 1107 return 0xB800; 1108 case 242: 1109 return 0x3C00; 1110 case 243: 1111 return 0xBC00; 1112 case 244: 1113 return 0x4000; 1114 case 245: 1115 return 0xC000; 1116 case 246: 1117 return 0x4400; 1118 case 247: 1119 return 0xC400; 1120 case 248: // 1 / (2 * PI) 1121 return 0x3118; 1122 default: 1123 llvm_unreachable("invalid fp inline imm"); 1124 } 1125 } 1126 1127 MCOperand AMDGPUDisassembler::decodeFPImmed(OpWidthTy Width, unsigned Imm) { 1128 assert(Imm >= AMDGPU::EncValues::INLINE_FLOATING_C_MIN 1129 && Imm <= AMDGPU::EncValues::INLINE_FLOATING_C_MAX); 1130 1131 // ToDo: case 248: 1/(2*PI) - is allowed only on VI 1132 switch (Width) { 1133 case OPW32: 1134 case OPW128: // splat constants 1135 case OPW512: 1136 case OPW1024: 1137 case OPWV232: 1138 return MCOperand::createImm(getInlineImmVal32(Imm)); 1139 case OPW64: 1140 case OPW256: 1141 return MCOperand::createImm(getInlineImmVal64(Imm)); 1142 case OPW16: 1143 case OPWV216: 1144 return MCOperand::createImm(getInlineImmVal16(Imm)); 1145 default: 1146 llvm_unreachable("implement me"); 1147 } 1148 } 1149 1150 unsigned AMDGPUDisassembler::getVgprClassId(const OpWidthTy Width) const { 1151 using namespace AMDGPU; 1152 1153 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_); 1154 switch (Width) { 1155 default: // fall 1156 case OPW32: 1157 case OPW16: 1158 case OPWV216: 1159 return VGPR_32RegClassID; 1160 case OPW64: 1161 case OPWV232: return VReg_64RegClassID; 1162 case OPW96: return VReg_96RegClassID; 1163 case OPW128: return VReg_128RegClassID; 1164 case OPW160: return VReg_160RegClassID; 1165 case OPW256: return VReg_256RegClassID; 1166 case OPW512: return VReg_512RegClassID; 1167 case OPW1024: return VReg_1024RegClassID; 1168 } 1169 } 1170 1171 unsigned AMDGPUDisassembler::getAgprClassId(const OpWidthTy Width) const { 1172 using namespace AMDGPU; 1173 1174 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_); 1175 switch (Width) { 1176 default: // fall 1177 case OPW32: 1178 case OPW16: 1179 case OPWV216: 1180 return AGPR_32RegClassID; 1181 case OPW64: 1182 case OPWV232: return AReg_64RegClassID; 1183 case OPW96: return AReg_96RegClassID; 1184 case OPW128: return AReg_128RegClassID; 1185 case OPW160: return AReg_160RegClassID; 1186 case OPW256: return AReg_256RegClassID; 1187 case OPW512: return AReg_512RegClassID; 1188 case OPW1024: return AReg_1024RegClassID; 1189 } 1190 } 1191 1192 1193 unsigned AMDGPUDisassembler::getSgprClassId(const OpWidthTy Width) const { 1194 using namespace AMDGPU; 1195 1196 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_); 1197 switch (Width) { 1198 default: // fall 1199 case OPW32: 1200 case OPW16: 1201 case OPWV216: 1202 return SGPR_32RegClassID; 1203 case OPW64: 1204 case OPWV232: return SGPR_64RegClassID; 1205 case OPW96: return SGPR_96RegClassID; 1206 case OPW128: return SGPR_128RegClassID; 1207 case OPW160: return SGPR_160RegClassID; 1208 case OPW256: return SGPR_256RegClassID; 1209 case OPW512: return SGPR_512RegClassID; 1210 } 1211 } 1212 1213 unsigned AMDGPUDisassembler::getTtmpClassId(const OpWidthTy Width) const { 1214 using namespace AMDGPU; 1215 1216 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_); 1217 switch (Width) { 1218 default: // fall 1219 case OPW32: 1220 case OPW16: 1221 case OPWV216: 1222 return TTMP_32RegClassID; 1223 case OPW64: 1224 case OPWV232: return TTMP_64RegClassID; 1225 case OPW128: return TTMP_128RegClassID; 1226 case OPW256: return TTMP_256RegClassID; 1227 case OPW512: return TTMP_512RegClassID; 1228 } 1229 } 1230 1231 int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const { 1232 using namespace AMDGPU::EncValues; 1233 1234 unsigned TTmpMin = isGFX9Plus() ? TTMP_GFX9PLUS_MIN : TTMP_VI_MIN; 1235 unsigned TTmpMax = isGFX9Plus() ? TTMP_GFX9PLUS_MAX : TTMP_VI_MAX; 1236 1237 return (TTmpMin <= Val && Val <= TTmpMax)? Val - TTmpMin : -1; 1238 } 1239 1240 MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val) const { 1241 using namespace AMDGPU::EncValues; 1242 1243 assert(Val < 1024); // enum10 1244 1245 bool IsAGPR = Val & 512; 1246 Val &= 511; 1247 1248 if (VGPR_MIN <= Val && Val <= VGPR_MAX) { 1249 return createRegOperand(IsAGPR ? getAgprClassId(Width) 1250 : getVgprClassId(Width), Val - VGPR_MIN); 1251 } 1252 if (Val <= SGPR_MAX) { 1253 // "SGPR_MIN <= Val" is always true and causes compilation warning. 1254 static_assert(SGPR_MIN == 0, ""); 1255 return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN); 1256 } 1257 1258 int TTmpIdx = getTTmpIdx(Val); 1259 if (TTmpIdx >= 0) { 1260 return createSRegOperand(getTtmpClassId(Width), TTmpIdx); 1261 } 1262 1263 if (INLINE_INTEGER_C_MIN <= Val && Val <= INLINE_INTEGER_C_MAX) 1264 return decodeIntImmed(Val); 1265 1266 if (INLINE_FLOATING_C_MIN <= Val && Val <= INLINE_FLOATING_C_MAX) 1267 return decodeFPImmed(Width, Val); 1268 1269 if (Val == LITERAL_CONST) 1270 return decodeLiteralConstant(); 1271 1272 switch (Width) { 1273 case OPW32: 1274 case OPW16: 1275 case OPWV216: 1276 return decodeSpecialReg32(Val); 1277 case OPW64: 1278 case OPWV232: 1279 return decodeSpecialReg64(Val); 1280 default: 1281 llvm_unreachable("unexpected immediate type"); 1282 } 1283 } 1284 1285 MCOperand AMDGPUDisassembler::decodeDstOp(const OpWidthTy Width, unsigned Val) const { 1286 using namespace AMDGPU::EncValues; 1287 1288 assert(Val < 128); 1289 assert(Width == OPW256 || Width == OPW512); 1290 1291 if (Val <= SGPR_MAX) { 1292 // "SGPR_MIN <= Val" is always true and causes compilation warning. 1293 static_assert(SGPR_MIN == 0, ""); 1294 return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN); 1295 } 1296 1297 int TTmpIdx = getTTmpIdx(Val); 1298 if (TTmpIdx >= 0) { 1299 return createSRegOperand(getTtmpClassId(Width), TTmpIdx); 1300 } 1301 1302 llvm_unreachable("unknown dst register"); 1303 } 1304 1305 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const { 1306 using namespace AMDGPU; 1307 1308 switch (Val) { 1309 case 102: return createRegOperand(FLAT_SCR_LO); 1310 case 103: return createRegOperand(FLAT_SCR_HI); 1311 case 104: return createRegOperand(XNACK_MASK_LO); 1312 case 105: return createRegOperand(XNACK_MASK_HI); 1313 case 106: return createRegOperand(VCC_LO); 1314 case 107: return createRegOperand(VCC_HI); 1315 case 108: return createRegOperand(TBA_LO); 1316 case 109: return createRegOperand(TBA_HI); 1317 case 110: return createRegOperand(TMA_LO); 1318 case 111: return createRegOperand(TMA_HI); 1319 case 124: return createRegOperand(M0); 1320 case 125: return createRegOperand(SGPR_NULL); 1321 case 126: return createRegOperand(EXEC_LO); 1322 case 127: return createRegOperand(EXEC_HI); 1323 case 235: return createRegOperand(SRC_SHARED_BASE); 1324 case 236: return createRegOperand(SRC_SHARED_LIMIT); 1325 case 237: return createRegOperand(SRC_PRIVATE_BASE); 1326 case 238: return createRegOperand(SRC_PRIVATE_LIMIT); 1327 case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID); 1328 case 251: return createRegOperand(SRC_VCCZ); 1329 case 252: return createRegOperand(SRC_EXECZ); 1330 case 253: return createRegOperand(SRC_SCC); 1331 case 254: return createRegOperand(LDS_DIRECT); 1332 default: break; 1333 } 1334 return errOperand(Val, "unknown operand encoding " + Twine(Val)); 1335 } 1336 1337 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const { 1338 using namespace AMDGPU; 1339 1340 switch (Val) { 1341 case 102: return createRegOperand(FLAT_SCR); 1342 case 104: return createRegOperand(XNACK_MASK); 1343 case 106: return createRegOperand(VCC); 1344 case 108: return createRegOperand(TBA); 1345 case 110: return createRegOperand(TMA); 1346 case 125: return createRegOperand(SGPR_NULL); 1347 case 126: return createRegOperand(EXEC); 1348 case 235: return createRegOperand(SRC_SHARED_BASE); 1349 case 236: return createRegOperand(SRC_SHARED_LIMIT); 1350 case 237: return createRegOperand(SRC_PRIVATE_BASE); 1351 case 238: return createRegOperand(SRC_PRIVATE_LIMIT); 1352 case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID); 1353 case 251: return createRegOperand(SRC_VCCZ); 1354 case 252: return createRegOperand(SRC_EXECZ); 1355 case 253: return createRegOperand(SRC_SCC); 1356 default: break; 1357 } 1358 return errOperand(Val, "unknown operand encoding " + Twine(Val)); 1359 } 1360 1361 MCOperand AMDGPUDisassembler::decodeSDWASrc(const OpWidthTy Width, 1362 const unsigned Val) const { 1363 using namespace AMDGPU::SDWA; 1364 using namespace AMDGPU::EncValues; 1365 1366 if (STI.getFeatureBits()[AMDGPU::FeatureGFX9] || 1367 STI.getFeatureBits()[AMDGPU::FeatureGFX10]) { 1368 // XXX: cast to int is needed to avoid stupid warning: 1369 // compare with unsigned is always true 1370 if (int(SDWA9EncValues::SRC_VGPR_MIN) <= int(Val) && 1371 Val <= SDWA9EncValues::SRC_VGPR_MAX) { 1372 return createRegOperand(getVgprClassId(Width), 1373 Val - SDWA9EncValues::SRC_VGPR_MIN); 1374 } 1375 if (SDWA9EncValues::SRC_SGPR_MIN <= Val && 1376 Val <= (isGFX10Plus() ? SDWA9EncValues::SRC_SGPR_MAX_GFX10 1377 : SDWA9EncValues::SRC_SGPR_MAX_SI)) { 1378 return createSRegOperand(getSgprClassId(Width), 1379 Val - SDWA9EncValues::SRC_SGPR_MIN); 1380 } 1381 if (SDWA9EncValues::SRC_TTMP_MIN <= Val && 1382 Val <= SDWA9EncValues::SRC_TTMP_MAX) { 1383 return createSRegOperand(getTtmpClassId(Width), 1384 Val - SDWA9EncValues::SRC_TTMP_MIN); 1385 } 1386 1387 const unsigned SVal = Val - SDWA9EncValues::SRC_SGPR_MIN; 1388 1389 if (INLINE_INTEGER_C_MIN <= SVal && SVal <= INLINE_INTEGER_C_MAX) 1390 return decodeIntImmed(SVal); 1391 1392 if (INLINE_FLOATING_C_MIN <= SVal && SVal <= INLINE_FLOATING_C_MAX) 1393 return decodeFPImmed(Width, SVal); 1394 1395 return decodeSpecialReg32(SVal); 1396 } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) { 1397 return createRegOperand(getVgprClassId(Width), Val); 1398 } 1399 llvm_unreachable("unsupported target"); 1400 } 1401 1402 MCOperand AMDGPUDisassembler::decodeSDWASrc16(unsigned Val) const { 1403 return decodeSDWASrc(OPW16, Val); 1404 } 1405 1406 MCOperand AMDGPUDisassembler::decodeSDWASrc32(unsigned Val) const { 1407 return decodeSDWASrc(OPW32, Val); 1408 } 1409 1410 MCOperand AMDGPUDisassembler::decodeSDWAVopcDst(unsigned Val) const { 1411 using namespace AMDGPU::SDWA; 1412 1413 assert((STI.getFeatureBits()[AMDGPU::FeatureGFX9] || 1414 STI.getFeatureBits()[AMDGPU::FeatureGFX10]) && 1415 "SDWAVopcDst should be present only on GFX9+"); 1416 1417 bool IsWave64 = STI.getFeatureBits()[AMDGPU::FeatureWavefrontSize64]; 1418 1419 if (Val & SDWA9EncValues::VOPC_DST_VCC_MASK) { 1420 Val &= SDWA9EncValues::VOPC_DST_SGPR_MASK; 1421 1422 int TTmpIdx = getTTmpIdx(Val); 1423 if (TTmpIdx >= 0) { 1424 auto TTmpClsId = getTtmpClassId(IsWave64 ? OPW64 : OPW32); 1425 return createSRegOperand(TTmpClsId, TTmpIdx); 1426 } else if (Val > SGPR_MAX) { 1427 return IsWave64 ? decodeSpecialReg64(Val) 1428 : decodeSpecialReg32(Val); 1429 } else { 1430 return createSRegOperand(getSgprClassId(IsWave64 ? OPW64 : OPW32), Val); 1431 } 1432 } else { 1433 return createRegOperand(IsWave64 ? AMDGPU::VCC : AMDGPU::VCC_LO); 1434 } 1435 } 1436 1437 MCOperand AMDGPUDisassembler::decodeBoolReg(unsigned Val) const { 1438 return STI.getFeatureBits()[AMDGPU::FeatureWavefrontSize64] ? 1439 decodeOperand_SReg_64(Val) : decodeOperand_SReg_32(Val); 1440 } 1441 1442 bool AMDGPUDisassembler::isVI() const { 1443 return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]; 1444 } 1445 1446 bool AMDGPUDisassembler::isGFX9() const { return AMDGPU::isGFX9(STI); } 1447 1448 bool AMDGPUDisassembler::isGFX90A() const { 1449 return STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]; 1450 } 1451 1452 bool AMDGPUDisassembler::isGFX9Plus() const { return AMDGPU::isGFX9Plus(STI); } 1453 1454 bool AMDGPUDisassembler::isGFX10() const { return AMDGPU::isGFX10(STI); } 1455 1456 bool AMDGPUDisassembler::isGFX10Plus() const { 1457 return AMDGPU::isGFX10Plus(STI); 1458 } 1459 1460 bool AMDGPUDisassembler::hasArchitectedFlatScratch() const { 1461 return STI.getFeatureBits()[AMDGPU::FeatureArchitectedFlatScratch]; 1462 } 1463 1464 //===----------------------------------------------------------------------===// 1465 // AMDGPU specific symbol handling 1466 //===----------------------------------------------------------------------===// 1467 #define PRINT_DIRECTIVE(DIRECTIVE, MASK) \ 1468 do { \ 1469 KdStream << Indent << DIRECTIVE " " \ 1470 << ((FourByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n'; \ 1471 } while (0) 1472 1473 // NOLINTNEXTLINE(readability-identifier-naming) 1474 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC1( 1475 uint32_t FourByteBuffer, raw_string_ostream &KdStream) const { 1476 using namespace amdhsa; 1477 StringRef Indent = "\t"; 1478 1479 // We cannot accurately backward compute #VGPRs used from 1480 // GRANULATED_WORKITEM_VGPR_COUNT. But we are concerned with getting the same 1481 // value of GRANULATED_WORKITEM_VGPR_COUNT in the reassembled binary. So we 1482 // simply calculate the inverse of what the assembler does. 1483 1484 uint32_t GranulatedWorkitemVGPRCount = 1485 (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT) >> 1486 COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT_SHIFT; 1487 1488 uint32_t NextFreeVGPR = (GranulatedWorkitemVGPRCount + 1) * 1489 AMDGPU::IsaInfo::getVGPREncodingGranule(&STI); 1490 1491 KdStream << Indent << ".amdhsa_next_free_vgpr " << NextFreeVGPR << '\n'; 1492 1493 // We cannot backward compute values used to calculate 1494 // GRANULATED_WAVEFRONT_SGPR_COUNT. Hence the original values for following 1495 // directives can't be computed: 1496 // .amdhsa_reserve_vcc 1497 // .amdhsa_reserve_flat_scratch 1498 // .amdhsa_reserve_xnack_mask 1499 // They take their respective default values if not specified in the assembly. 1500 // 1501 // GRANULATED_WAVEFRONT_SGPR_COUNT 1502 // = f(NEXT_FREE_SGPR + VCC + FLAT_SCRATCH + XNACK_MASK) 1503 // 1504 // We compute the inverse as though all directives apart from NEXT_FREE_SGPR 1505 // are set to 0. So while disassembling we consider that: 1506 // 1507 // GRANULATED_WAVEFRONT_SGPR_COUNT 1508 // = f(NEXT_FREE_SGPR + 0 + 0 + 0) 1509 // 1510 // The disassembler cannot recover the original values of those 3 directives. 1511 1512 uint32_t GranulatedWavefrontSGPRCount = 1513 (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT) >> 1514 COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT_SHIFT; 1515 1516 if (isGFX10Plus() && GranulatedWavefrontSGPRCount) 1517 return MCDisassembler::Fail; 1518 1519 uint32_t NextFreeSGPR = (GranulatedWavefrontSGPRCount + 1) * 1520 AMDGPU::IsaInfo::getSGPREncodingGranule(&STI); 1521 1522 KdStream << Indent << ".amdhsa_reserve_vcc " << 0 << '\n'; 1523 if (!hasArchitectedFlatScratch()) 1524 KdStream << Indent << ".amdhsa_reserve_flat_scratch " << 0 << '\n'; 1525 KdStream << Indent << ".amdhsa_reserve_xnack_mask " << 0 << '\n'; 1526 KdStream << Indent << ".amdhsa_next_free_sgpr " << NextFreeSGPR << "\n"; 1527 1528 if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIORITY) 1529 return MCDisassembler::Fail; 1530 1531 PRINT_DIRECTIVE(".amdhsa_float_round_mode_32", 1532 COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32); 1533 PRINT_DIRECTIVE(".amdhsa_float_round_mode_16_64", 1534 COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64); 1535 PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_32", 1536 COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32); 1537 PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_16_64", 1538 COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64); 1539 1540 if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIV) 1541 return MCDisassembler::Fail; 1542 1543 PRINT_DIRECTIVE(".amdhsa_dx10_clamp", COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP); 1544 1545 if (FourByteBuffer & COMPUTE_PGM_RSRC1_DEBUG_MODE) 1546 return MCDisassembler::Fail; 1547 1548 PRINT_DIRECTIVE(".amdhsa_ieee_mode", COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE); 1549 1550 if (FourByteBuffer & COMPUTE_PGM_RSRC1_BULKY) 1551 return MCDisassembler::Fail; 1552 1553 if (FourByteBuffer & COMPUTE_PGM_RSRC1_CDBG_USER) 1554 return MCDisassembler::Fail; 1555 1556 PRINT_DIRECTIVE(".amdhsa_fp16_overflow", COMPUTE_PGM_RSRC1_FP16_OVFL); 1557 1558 if (FourByteBuffer & COMPUTE_PGM_RSRC1_RESERVED0) 1559 return MCDisassembler::Fail; 1560 1561 if (isGFX10Plus()) { 1562 PRINT_DIRECTIVE(".amdhsa_workgroup_processor_mode", 1563 COMPUTE_PGM_RSRC1_WGP_MODE); 1564 PRINT_DIRECTIVE(".amdhsa_memory_ordered", COMPUTE_PGM_RSRC1_MEM_ORDERED); 1565 PRINT_DIRECTIVE(".amdhsa_forward_progress", COMPUTE_PGM_RSRC1_FWD_PROGRESS); 1566 } 1567 return MCDisassembler::Success; 1568 } 1569 1570 // NOLINTNEXTLINE(readability-identifier-naming) 1571 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC2( 1572 uint32_t FourByteBuffer, raw_string_ostream &KdStream) const { 1573 using namespace amdhsa; 1574 StringRef Indent = "\t"; 1575 if (hasArchitectedFlatScratch()) 1576 PRINT_DIRECTIVE(".amdhsa_enable_private_segment", 1577 COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT); 1578 else 1579 PRINT_DIRECTIVE(".amdhsa_system_sgpr_private_segment_wavefront_offset", 1580 COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT); 1581 PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_x", 1582 COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X); 1583 PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_y", 1584 COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y); 1585 PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_z", 1586 COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z); 1587 PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_info", 1588 COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO); 1589 PRINT_DIRECTIVE(".amdhsa_system_vgpr_workitem_id", 1590 COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID); 1591 1592 if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_ADDRESS_WATCH) 1593 return MCDisassembler::Fail; 1594 1595 if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_MEMORY) 1596 return MCDisassembler::Fail; 1597 1598 if (FourByteBuffer & COMPUTE_PGM_RSRC2_GRANULATED_LDS_SIZE) 1599 return MCDisassembler::Fail; 1600 1601 PRINT_DIRECTIVE( 1602 ".amdhsa_exception_fp_ieee_invalid_op", 1603 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION); 1604 PRINT_DIRECTIVE(".amdhsa_exception_fp_denorm_src", 1605 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE); 1606 PRINT_DIRECTIVE( 1607 ".amdhsa_exception_fp_ieee_div_zero", 1608 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO); 1609 PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_overflow", 1610 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW); 1611 PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_underflow", 1612 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW); 1613 PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_inexact", 1614 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT); 1615 PRINT_DIRECTIVE(".amdhsa_exception_int_div_zero", 1616 COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO); 1617 1618 if (FourByteBuffer & COMPUTE_PGM_RSRC2_RESERVED0) 1619 return MCDisassembler::Fail; 1620 1621 return MCDisassembler::Success; 1622 } 1623 1624 #undef PRINT_DIRECTIVE 1625 1626 MCDisassembler::DecodeStatus 1627 AMDGPUDisassembler::decodeKernelDescriptorDirective( 1628 DataExtractor::Cursor &Cursor, ArrayRef<uint8_t> Bytes, 1629 raw_string_ostream &KdStream) const { 1630 #define PRINT_DIRECTIVE(DIRECTIVE, MASK) \ 1631 do { \ 1632 KdStream << Indent << DIRECTIVE " " \ 1633 << ((TwoByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n'; \ 1634 } while (0) 1635 1636 uint16_t TwoByteBuffer = 0; 1637 uint32_t FourByteBuffer = 0; 1638 1639 StringRef ReservedBytes; 1640 StringRef Indent = "\t"; 1641 1642 assert(Bytes.size() == 64); 1643 DataExtractor DE(Bytes, /*IsLittleEndian=*/true, /*AddressSize=*/8); 1644 1645 switch (Cursor.tell()) { 1646 case amdhsa::GROUP_SEGMENT_FIXED_SIZE_OFFSET: 1647 FourByteBuffer = DE.getU32(Cursor); 1648 KdStream << Indent << ".amdhsa_group_segment_fixed_size " << FourByteBuffer 1649 << '\n'; 1650 return MCDisassembler::Success; 1651 1652 case amdhsa::PRIVATE_SEGMENT_FIXED_SIZE_OFFSET: 1653 FourByteBuffer = DE.getU32(Cursor); 1654 KdStream << Indent << ".amdhsa_private_segment_fixed_size " 1655 << FourByteBuffer << '\n'; 1656 return MCDisassembler::Success; 1657 1658 case amdhsa::KERNARG_SIZE_OFFSET: 1659 FourByteBuffer = DE.getU32(Cursor); 1660 KdStream << Indent << ".amdhsa_kernarg_size " 1661 << FourByteBuffer << '\n'; 1662 return MCDisassembler::Success; 1663 1664 case amdhsa::RESERVED0_OFFSET: 1665 // 4 reserved bytes, must be 0. 1666 ReservedBytes = DE.getBytes(Cursor, 4); 1667 for (int I = 0; I < 4; ++I) { 1668 if (ReservedBytes[I] != 0) { 1669 return MCDisassembler::Fail; 1670 } 1671 } 1672 return MCDisassembler::Success; 1673 1674 case amdhsa::KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET: 1675 // KERNEL_CODE_ENTRY_BYTE_OFFSET 1676 // So far no directive controls this for Code Object V3, so simply skip for 1677 // disassembly. 1678 DE.skip(Cursor, 8); 1679 return MCDisassembler::Success; 1680 1681 case amdhsa::RESERVED1_OFFSET: 1682 // 20 reserved bytes, must be 0. 1683 ReservedBytes = DE.getBytes(Cursor, 20); 1684 for (int I = 0; I < 20; ++I) { 1685 if (ReservedBytes[I] != 0) { 1686 return MCDisassembler::Fail; 1687 } 1688 } 1689 return MCDisassembler::Success; 1690 1691 case amdhsa::COMPUTE_PGM_RSRC3_OFFSET: 1692 // COMPUTE_PGM_RSRC3 1693 // - Only set for GFX10, GFX6-9 have this to be 0. 1694 // - Currently no directives directly control this. 1695 FourByteBuffer = DE.getU32(Cursor); 1696 if (!isGFX10Plus() && FourByteBuffer) { 1697 return MCDisassembler::Fail; 1698 } 1699 return MCDisassembler::Success; 1700 1701 case amdhsa::COMPUTE_PGM_RSRC1_OFFSET: 1702 FourByteBuffer = DE.getU32(Cursor); 1703 if (decodeCOMPUTE_PGM_RSRC1(FourByteBuffer, KdStream) == 1704 MCDisassembler::Fail) { 1705 return MCDisassembler::Fail; 1706 } 1707 return MCDisassembler::Success; 1708 1709 case amdhsa::COMPUTE_PGM_RSRC2_OFFSET: 1710 FourByteBuffer = DE.getU32(Cursor); 1711 if (decodeCOMPUTE_PGM_RSRC2(FourByteBuffer, KdStream) == 1712 MCDisassembler::Fail) { 1713 return MCDisassembler::Fail; 1714 } 1715 return MCDisassembler::Success; 1716 1717 case amdhsa::KERNEL_CODE_PROPERTIES_OFFSET: 1718 using namespace amdhsa; 1719 TwoByteBuffer = DE.getU16(Cursor); 1720 1721 if (!hasArchitectedFlatScratch()) 1722 PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_buffer", 1723 KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER); 1724 PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_ptr", 1725 KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR); 1726 PRINT_DIRECTIVE(".amdhsa_user_sgpr_queue_ptr", 1727 KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR); 1728 PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_segment_ptr", 1729 KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR); 1730 PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_id", 1731 KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID); 1732 if (!hasArchitectedFlatScratch()) 1733 PRINT_DIRECTIVE(".amdhsa_user_sgpr_flat_scratch_init", 1734 KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT); 1735 PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_size", 1736 KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE); 1737 1738 if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED0) 1739 return MCDisassembler::Fail; 1740 1741 // Reserved for GFX9 1742 if (isGFX9() && 1743 (TwoByteBuffer & KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32)) { 1744 return MCDisassembler::Fail; 1745 } else if (isGFX10Plus()) { 1746 PRINT_DIRECTIVE(".amdhsa_wavefront_size32", 1747 KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32); 1748 } 1749 1750 if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED1) 1751 return MCDisassembler::Fail; 1752 1753 return MCDisassembler::Success; 1754 1755 case amdhsa::RESERVED2_OFFSET: 1756 // 6 bytes from here are reserved, must be 0. 1757 ReservedBytes = DE.getBytes(Cursor, 6); 1758 for (int I = 0; I < 6; ++I) { 1759 if (ReservedBytes[I] != 0) 1760 return MCDisassembler::Fail; 1761 } 1762 return MCDisassembler::Success; 1763 1764 default: 1765 llvm_unreachable("Unhandled index. Case statements cover everything."); 1766 return MCDisassembler::Fail; 1767 } 1768 #undef PRINT_DIRECTIVE 1769 } 1770 1771 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeKernelDescriptor( 1772 StringRef KdName, ArrayRef<uint8_t> Bytes, uint64_t KdAddress) const { 1773 // CP microcode requires the kernel descriptor to be 64 aligned. 1774 if (Bytes.size() != 64 || KdAddress % 64 != 0) 1775 return MCDisassembler::Fail; 1776 1777 std::string Kd; 1778 raw_string_ostream KdStream(Kd); 1779 KdStream << ".amdhsa_kernel " << KdName << '\n'; 1780 1781 DataExtractor::Cursor C(0); 1782 while (C && C.tell() < Bytes.size()) { 1783 MCDisassembler::DecodeStatus Status = 1784 decodeKernelDescriptorDirective(C, Bytes, KdStream); 1785 1786 cantFail(C.takeError()); 1787 1788 if (Status == MCDisassembler::Fail) 1789 return MCDisassembler::Fail; 1790 } 1791 KdStream << ".end_amdhsa_kernel\n"; 1792 outs() << KdStream.str(); 1793 return MCDisassembler::Success; 1794 } 1795 1796 Optional<MCDisassembler::DecodeStatus> 1797 AMDGPUDisassembler::onSymbolStart(SymbolInfoTy &Symbol, uint64_t &Size, 1798 ArrayRef<uint8_t> Bytes, uint64_t Address, 1799 raw_ostream &CStream) const { 1800 // Right now only kernel descriptor needs to be handled. 1801 // We ignore all other symbols for target specific handling. 1802 // TODO: 1803 // Fix the spurious symbol issue for AMDGPU kernels. Exists for both Code 1804 // Object V2 and V3 when symbols are marked protected. 1805 1806 // amd_kernel_code_t for Code Object V2. 1807 if (Symbol.Type == ELF::STT_AMDGPU_HSA_KERNEL) { 1808 Size = 256; 1809 return MCDisassembler::Fail; 1810 } 1811 1812 // Code Object V3 kernel descriptors. 1813 StringRef Name = Symbol.Name; 1814 if (Symbol.Type == ELF::STT_OBJECT && Name.endswith(StringRef(".kd"))) { 1815 Size = 64; // Size = 64 regardless of success or failure. 1816 return decodeKernelDescriptor(Name.drop_back(3), Bytes, Address); 1817 } 1818 return None; 1819 } 1820 1821 //===----------------------------------------------------------------------===// 1822 // AMDGPUSymbolizer 1823 //===----------------------------------------------------------------------===// 1824 1825 // Try to find symbol name for specified label 1826 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(MCInst &Inst, 1827 raw_ostream &/*cStream*/, int64_t Value, 1828 uint64_t /*Address*/, bool IsBranch, 1829 uint64_t /*Offset*/, uint64_t /*InstSize*/) { 1830 1831 if (!IsBranch) { 1832 return false; 1833 } 1834 1835 auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo); 1836 if (!Symbols) 1837 return false; 1838 1839 auto Result = llvm::find_if(*Symbols, [Value](const SymbolInfoTy &Val) { 1840 return Val.Addr == static_cast<uint64_t>(Value) && 1841 Val.Type == ELF::STT_NOTYPE; 1842 }); 1843 if (Result != Symbols->end()) { 1844 auto *Sym = Ctx.getOrCreateSymbol(Result->Name); 1845 const auto *Add = MCSymbolRefExpr::create(Sym, Ctx); 1846 Inst.addOperand(MCOperand::createExpr(Add)); 1847 return true; 1848 } 1849 // Add to list of referenced addresses, so caller can synthesize a label. 1850 ReferencedAddresses.push_back(static_cast<uint64_t>(Value)); 1851 return false; 1852 } 1853 1854 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream, 1855 int64_t Value, 1856 uint64_t Address) { 1857 llvm_unreachable("unimplemented"); 1858 } 1859 1860 //===----------------------------------------------------------------------===// 1861 // Initialization 1862 //===----------------------------------------------------------------------===// 1863 1864 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/, 1865 LLVMOpInfoCallback /*GetOpInfo*/, 1866 LLVMSymbolLookupCallback /*SymbolLookUp*/, 1867 void *DisInfo, 1868 MCContext *Ctx, 1869 std::unique_ptr<MCRelocationInfo> &&RelInfo) { 1870 return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo); 1871 } 1872 1873 static MCDisassembler *createAMDGPUDisassembler(const Target &T, 1874 const MCSubtargetInfo &STI, 1875 MCContext &Ctx) { 1876 return new AMDGPUDisassembler(STI, Ctx, T.createMCInstrInfo()); 1877 } 1878 1879 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUDisassembler() { 1880 TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(), 1881 createAMDGPUDisassembler); 1882 TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(), 1883 createAMDGPUSymbolizer); 1884 } 1885