1//===-- DSInstructions.td - DS Instruction Definitions --------------------===// 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 9class DS_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]> : 10 InstSI <outs, ins, "", pattern>, 11 SIMCInstr <opName, SIEncodingFamily.NONE> { 12 13 let LGKM_CNT = 1; 14 let DS = 1; 15 let Size = 8; 16 let UseNamedOperandTable = 1; 17 18 // Most instruction load and store data, so set this as the default. 19 let mayLoad = 1; 20 let mayStore = 1; 21 let maybeAtomic = 1; 22 23 let hasSideEffects = 0; 24 let SchedRW = [WriteLDS]; 25 26 let isPseudo = 1; 27 let isCodeGenOnly = 1; 28 29 let AsmMatchConverter = "cvtDS"; 30 31 string Mnemonic = opName; 32 string AsmOperands = asmOps; 33 34 // Well these bits a kind of hack because it would be more natural 35 // to test "outs" and "ins" dags for the presence of particular operands 36 bits<1> has_vdst = 1; 37 bits<1> has_addr = 1; 38 bits<1> has_data0 = 1; 39 bits<1> has_data1 = 1; 40 41 bits<1> has_gws_data0 = 0; // data0 is encoded as addr 42 43 bits<1> has_offset = 1; // has "offset" that should be split to offset0,1 44 bits<1> has_offset0 = 1; 45 bits<1> has_offset1 = 1; 46 47 bits<1> has_gds = 1; 48 bits<1> gdsValue = 0; // if has_gds == 0 set gds to this value 49 50 bits<1> has_m0_read = 1; 51 52 let Uses = !if(has_m0_read, [M0, EXEC], [EXEC]); 53} 54 55class DS_Real <DS_Pseudo ps, string opName = ps.Mnemonic> : 56 InstSI <ps.OutOperandList, ps.InOperandList, opName # ps.AsmOperands>, 57 Enc64 { 58 59 let isPseudo = 0; 60 let isCodeGenOnly = 0; 61 let LGKM_CNT = 1; 62 let DS = 1; 63 let UseNamedOperandTable = 1; 64 65 // copy relevant pseudo op flags 66 let SubtargetPredicate = ps.SubtargetPredicate; 67 let OtherPredicates = ps.OtherPredicates; 68 let AsmMatchConverter = ps.AsmMatchConverter; 69 let SchedRW = ps.SchedRW; 70 let mayLoad = ps.mayLoad; 71 let mayStore = ps.mayStore; 72 let IsAtomicRet = ps.IsAtomicRet; 73 let IsAtomicNoRet = ps.IsAtomicNoRet; 74 75 let Constraints = ps.Constraints; 76 let DisableEncoding = ps.DisableEncoding; 77 78 // encoding fields 79 bits<10> vdst; 80 bits<1> gds; 81 bits<8> addr; 82 bits<10> data0; 83 bits<10> data1; 84 bits<8> offset0; 85 bits<8> offset1; 86 87 bits<16> offset; 88 let offset0 = !if(ps.has_offset, offset{7-0}, ?); 89 let offset1 = !if(ps.has_offset, offset{15-8}, ?); 90 91 bits<1> acc = !if(ps.has_vdst, vdst{9}, 92 !if(!or(ps.has_data0, ps.has_gws_data0), data0{9}, 0)); 93} 94 95// DS Pseudo instructions 96 97class DS_0A1D_NORET<string opName, RegisterClass rc = VGPR_32> 98: DS_Pseudo<opName, 99 (outs), 100 (ins getLdStRegisterOperand<rc>.ret:$data0, offset:$offset, gds:$gds), 101 " $data0$offset$gds"> { 102 103 let has_addr = 0; 104 let has_data1 = 0; 105 let has_vdst = 0; 106} 107 108class DS_1A1D_NORET<string opName, RegisterClass rc = VGPR_32> 109: DS_Pseudo<opName, 110 (outs), 111 (ins VGPR_32:$addr, getLdStRegisterOperand<rc>.ret:$data0, offset:$offset, gds:$gds), 112 " $addr, $data0$offset$gds"> { 113 114 let has_data1 = 0; 115 let has_vdst = 0; 116 let IsAtomicNoRet = 1; 117} 118 119multiclass DS_1A1D_NORET_mc<string opName, RegisterClass rc = VGPR_32> { 120 def "" : DS_1A1D_NORET<opName, rc>, 121 AtomicNoRet<opName, 0>; 122 123 let has_m0_read = 0 in { 124 def _gfx9 : DS_1A1D_NORET<opName, rc>, 125 AtomicNoRet<opName#"_gfx9", 0>; 126 } 127} 128 129multiclass DS_1A1D_NORET_mc_gfx9<string opName, RegisterClass rc = VGPR_32> { 130 let has_m0_read = 0 in { 131 def "" : DS_1A1D_NORET<opName, rc>, 132 AtomicNoRet<opName, 0>; 133 } 134} 135 136class DS_1A2D_NORET<string opName, RegisterClass rc = VGPR_32, 137 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 138: DS_Pseudo<opName, 139 (outs), 140 (ins VGPR_32:$addr, data_op:$data0, data_op:$data1, offset:$offset, gds:$gds), 141 " $addr, $data0, $data1$offset$gds"> { 142 143 let has_vdst = 0; 144 let IsAtomicNoRet = 1; 145} 146 147multiclass DS_1A2D_NORET_mc<string opName, RegisterClass rc = VGPR_32> { 148 def "" : DS_1A2D_NORET<opName, rc>, 149 AtomicNoRet<opName, 0>; 150 151 let has_m0_read = 0 in { 152 def _gfx9 : DS_1A2D_NORET<opName, rc>, 153 AtomicNoRet<opName#"_gfx9", 0>; 154 } 155} 156 157class DS_1A2D_Off8_NORET <string opName, RegisterClass rc = VGPR_32, 158 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 159: DS_Pseudo<opName, 160 (outs), 161 (ins VGPR_32:$addr, data_op:$data0, data_op:$data1, 162 offset0:$offset0, offset1:$offset1, gds:$gds), 163 " $addr, $data0, $data1$offset0$offset1$gds"> { 164 165 let has_vdst = 0; 166 let has_offset = 0; 167 let AsmMatchConverter = "cvtDSOffset01"; 168} 169 170multiclass DS_1A2D_Off8_NORET_mc <string opName, RegisterClass rc = VGPR_32> { 171 def "" : DS_1A2D_Off8_NORET<opName, rc>; 172 173 let has_m0_read = 0 in { 174 def _gfx9 : DS_1A2D_Off8_NORET<opName, rc>; 175 } 176} 177 178class DS_0A1D_RET_GDS<string opName, RegisterClass rc = VGPR_32, RegisterClass src = rc, 179 RegisterOperand dst_op = getLdStRegisterOperand<rc>.ret, 180 RegisterOperand src_op = getLdStRegisterOperand<src>.ret> 181: DS_Pseudo<opName, 182 (outs dst_op:$vdst), 183 (ins src_op:$data0, offset:$offset), 184 " $vdst, $data0$offset gds"> { 185 186 let has_addr = 0; 187 let has_data1 = 0; 188 let has_gds = 0; 189 let gdsValue = 1; 190 let AsmMatchConverter = "cvtDSGds"; 191 let hasSideEffects = 1; 192} 193 194class DS_1A1D_RET <string opName, RegisterClass rc = VGPR_32, 195 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 196: DS_Pseudo<opName, 197 (outs data_op:$vdst), 198 (ins VGPR_32:$addr, data_op:$data0, offset:$offset, gds:$gds), 199 " $vdst, $addr, $data0$offset$gds"> { 200 201 let hasPostISelHook = 1; 202 let has_data1 = 0; 203 let IsAtomicRet = 1; 204} 205 206multiclass DS_1A1D_RET_mc <string opName, RegisterClass rc = VGPR_32, 207 string NoRetOp = ""> { 208 def "" : DS_1A1D_RET<opName, rc>, 209 AtomicNoRet<NoRetOp, !ne(NoRetOp, "")>; 210 211 let has_m0_read = 0 in { 212 def _gfx9 : DS_1A1D_RET<opName, rc>, 213 AtomicNoRet<!if(!eq(NoRetOp, ""), "", NoRetOp#"_gfx9"), 214 !ne(NoRetOp, "")>; 215 } 216} 217 218multiclass DS_1A1D_RET_mc_gfx9 <string opName, RegisterClass rc = VGPR_32, 219 string NoRetOp = ""> { 220 let has_m0_read = 0 in { 221 def "" : DS_1A1D_RET<opName, rc>, 222 AtomicNoRet<!if(!eq(NoRetOp, ""), "", NoRetOp), 223 !if(!eq(NoRetOp, ""), 0, 1)>; 224 } 225} 226 227class DS_1A2D_RET<string opName, 228 RegisterClass rc = VGPR_32, 229 RegisterClass src = rc, 230 RegisterOperand dst_op = getLdStRegisterOperand<rc>.ret, 231 RegisterOperand src_op = getLdStRegisterOperand<src>.ret> 232: DS_Pseudo<opName, 233 (outs dst_op:$vdst), 234 (ins VGPR_32:$addr, src_op:$data0, src_op:$data1, offset:$offset, gds:$gds), 235 " $vdst, $addr, $data0, $data1$offset$gds"> { 236 237 let hasPostISelHook = 1; 238 let IsAtomicRet = 1; 239} 240 241multiclass DS_1A2D_RET_mc<string opName, 242 RegisterClass rc = VGPR_32, 243 string NoRetOp = "", 244 RegisterClass src = rc> { 245 def "" : DS_1A2D_RET<opName, rc, src>, 246 AtomicNoRet<NoRetOp, !ne(NoRetOp, "")>; 247 248 let has_m0_read = 0 in { 249 def _gfx9 : DS_1A2D_RET<opName, rc, src>, 250 AtomicNoRet<NoRetOp#"_gfx9", !ne(NoRetOp, "")>; 251 } 252} 253 254class DS_1A2D_Off8_RET<string opName, 255 RegisterClass rc = VGPR_32, 256 RegisterClass src = rc, 257 RegisterOperand dst_op = getLdStRegisterOperand<rc>.ret, 258 RegisterOperand src_op = getLdStRegisterOperand<src>.ret> 259: DS_Pseudo<opName, 260 (outs dst_op:$vdst), 261 (ins VGPR_32:$addr, src_op:$data0, src_op:$data1, offset0:$offset0, offset1:$offset1, gds:$gds), 262 " $vdst, $addr, $data0, $data1$offset0$offset1$gds"> { 263 264 let has_offset = 0; 265 let AsmMatchConverter = "cvtDSOffset01"; 266 267 let hasPostISelHook = 1; 268} 269 270multiclass DS_1A2D_Off8_RET_mc<string opName, 271 RegisterClass rc = VGPR_32, 272 RegisterClass src = rc> { 273 def "" : DS_1A2D_Off8_RET<opName, rc, src>; 274 275 let has_m0_read = 0 in { 276 def _gfx9 : DS_1A2D_Off8_RET<opName, rc, src>; 277 } 278} 279 280 281class DS_1A_RET<string opName, RegisterClass rc = VGPR_32, bit HasTiedOutput = 0, Operand ofs = offset, 282 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 283: DS_Pseudo<opName, 284 (outs data_op:$vdst), 285 !if(HasTiedOutput, 286 (ins VGPR_32:$addr, ofs:$offset, gds:$gds, data_op:$vdst_in), 287 (ins VGPR_32:$addr, ofs:$offset, gds:$gds)), 288 " $vdst, $addr$offset$gds"> { 289 let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", ""); 290 let DisableEncoding = !if(HasTiedOutput, "$vdst_in", ""); 291 let has_data0 = 0; 292 let has_data1 = 0; 293} 294 295multiclass DS_1A_RET_mc<string opName, RegisterClass rc = VGPR_32, bit HasTiedOutput = 0, Operand ofs = offset> { 296 def "" : DS_1A_RET<opName, rc, HasTiedOutput, ofs>; 297 298 let has_m0_read = 0 in { 299 def _gfx9 : DS_1A_RET<opName, rc, HasTiedOutput, ofs>; 300 } 301} 302 303class DS_1A_RET_Tied<string opName, RegisterClass rc = VGPR_32> : 304 DS_1A_RET<opName, rc, 1>; 305 306class DS_1A_Off8_RET <string opName, RegisterClass rc = VGPR_32> 307: DS_Pseudo<opName, 308 (outs getLdStRegisterOperand<rc>.ret:$vdst), 309 (ins VGPR_32:$addr, offset0:$offset0, offset1:$offset1, gds:$gds), 310 " $vdst, $addr$offset0$offset1$gds"> { 311 312 let has_offset = 0; 313 let has_data0 = 0; 314 let has_data1 = 0; 315 let AsmMatchConverter = "cvtDSOffset01"; 316} 317 318multiclass DS_1A_Off8_RET_mc <string opName, RegisterClass rc = VGPR_32> { 319 def "" : DS_1A_Off8_RET<opName, rc>; 320 321 let has_m0_read = 0 in { 322 def _gfx9 : DS_1A_Off8_RET<opName, rc>; 323 } 324} 325 326class DS_1A_RET_GDS <string opName> : DS_Pseudo<opName, 327 (outs getLdStRegisterOperand<VGPR_32>.ret:$vdst), 328 (ins VGPR_32:$addr, offset:$offset), 329 " $vdst, $addr$offset gds"> { 330 331 let has_data0 = 0; 332 let has_data1 = 0; 333 let has_gds = 0; 334 let gdsValue = 1; 335 let AsmMatchConverter = "cvtDSGds"; 336} 337 338class DS_0A_RET <string opName> : DS_Pseudo<opName, 339 (outs getLdStRegisterOperand<VGPR_32>.ret:$vdst), 340 (ins offset:$offset, gds:$gds), 341 " $vdst$offset$gds"> { 342 343 let mayLoad = 1; 344 let mayStore = 1; 345 346 let has_addr = 0; 347 let has_data0 = 0; 348 let has_data1 = 0; 349} 350 351class DS_1A <string opName> : DS_Pseudo<opName, 352 (outs), 353 (ins VGPR_32:$addr, offset:$offset, gds:$gds), 354 " $addr$offset$gds"> { 355 356 let mayLoad = 1; 357 let mayStore = 1; 358 359 let has_vdst = 0; 360 let has_data0 = 0; 361 let has_data1 = 0; 362} 363 364multiclass DS_1A_mc <string opName> { 365 def "" : DS_1A<opName>; 366 367 let has_m0_read = 0 in { 368 def _gfx9 : DS_1A<opName>; 369 } 370} 371 372 373class DS_GWS <string opName, dag ins, string asmOps> 374: DS_Pseudo<opName, (outs), ins, asmOps> { 375 376 let has_vdst = 0; 377 let has_addr = 0; 378 let has_data0 = 0; 379 let has_data1 = 0; 380 381 let has_gds = 0; 382 let gdsValue = 1; 383 let AsmMatchConverter = "cvtDSGds"; 384} 385 386class DS_GWS_0D <string opName> 387: DS_GWS<opName, 388 (ins offset:$offset), "$offset gds"> { 389 let hasSideEffects = 1; 390} 391 392class DS_GWS_1D <string opName> 393: DS_GWS<opName, 394 (ins getLdStRegisterOperand<VGPR_32>.ret:$data0, offset:$offset), 395 " $data0$offset gds"> { 396 397 let has_gws_data0 = 1; 398 let hasSideEffects = 1; 399} 400 401class DS_VOID <string opName> : DS_Pseudo<opName, 402 (outs), (ins), ""> { 403 let mayLoad = 0; 404 let mayStore = 0; 405 let hasSideEffects = 1; 406 let UseNamedOperandTable = 0; 407 let AsmMatchConverter = ""; 408 409 let has_vdst = 0; 410 let has_addr = 0; 411 let has_data0 = 0; 412 let has_data1 = 0; 413 let has_offset = 0; 414 let has_offset0 = 0; 415 let has_offset1 = 0; 416 let has_gds = 0; 417} 418 419class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag, 420 RegisterOperand data_op = getLdStRegisterOperand<VGPR_32>.ret> 421: DS_Pseudo<opName, 422 (outs data_op:$vdst), 423 (ins VGPR_32:$addr, data_op:$data0, offset:$offset), 424 " $vdst, $addr, $data0$offset", 425 [(set i32:$vdst, 426 (node (DS1Addr1Offset i32:$addr, i16:$offset), i32:$data0))] > { 427 428 let mayLoad = 0; 429 let mayStore = 0; 430 let isConvergent = 1; 431 432 let has_data1 = 0; 433 let has_gds = 0; 434} 435 436defm DS_ADD_U32 : DS_1A1D_NORET_mc<"ds_add_u32">; 437defm DS_SUB_U32 : DS_1A1D_NORET_mc<"ds_sub_u32">; 438defm DS_RSUB_U32 : DS_1A1D_NORET_mc<"ds_rsub_u32">; 439defm DS_INC_U32 : DS_1A1D_NORET_mc<"ds_inc_u32">; 440defm DS_DEC_U32 : DS_1A1D_NORET_mc<"ds_dec_u32">; 441defm DS_MIN_I32 : DS_1A1D_NORET_mc<"ds_min_i32">; 442defm DS_MAX_I32 : DS_1A1D_NORET_mc<"ds_max_i32">; 443defm DS_MIN_U32 : DS_1A1D_NORET_mc<"ds_min_u32">; 444defm DS_MAX_U32 : DS_1A1D_NORET_mc<"ds_max_u32">; 445defm DS_AND_B32 : DS_1A1D_NORET_mc<"ds_and_b32">; 446defm DS_OR_B32 : DS_1A1D_NORET_mc<"ds_or_b32">; 447defm DS_XOR_B32 : DS_1A1D_NORET_mc<"ds_xor_b32">; 448 449let SubtargetPredicate = HasLDSFPAtomicAdd in { 450defm DS_ADD_F32 : DS_1A1D_NORET_mc<"ds_add_f32">; 451} 452 453defm DS_MIN_F32 : DS_1A1D_NORET_mc<"ds_min_f32">; 454defm DS_MAX_F32 : DS_1A1D_NORET_mc<"ds_max_f32">; 455 456let mayLoad = 0 in { 457defm DS_WRITE_B8 : DS_1A1D_NORET_mc<"ds_write_b8">; 458defm DS_WRITE_B16 : DS_1A1D_NORET_mc<"ds_write_b16">; 459defm DS_WRITE_B32 : DS_1A1D_NORET_mc<"ds_write_b32">; 460defm DS_WRITE2_B32 : DS_1A2D_Off8_NORET_mc<"ds_write2_b32">; 461defm DS_WRITE2ST64_B32: DS_1A2D_Off8_NORET_mc<"ds_write2st64_b32">; 462 463 464let has_m0_read = 0 in { 465 466let SubtargetPredicate = HasD16LoadStore in { 467def DS_WRITE_B8_D16_HI : DS_1A1D_NORET<"ds_write_b8_d16_hi">; 468def DS_WRITE_B16_D16_HI : DS_1A1D_NORET<"ds_write_b16_d16_hi">; 469} 470 471} // End has_m0_read = 0 472 473let SubtargetPredicate = HasDSAddTid in { 474def DS_WRITE_ADDTID_B32 : DS_0A1D_NORET<"ds_write_addtid_b32">; 475} 476 477} // End mayLoad = 0 478 479let SubtargetPredicate = isGFX90APlus in { 480 defm DS_ADD_F64 : DS_1A1D_NORET_mc_gfx9<"ds_add_f64", VReg_64>; 481 defm DS_ADD_RTN_F64 : DS_1A1D_RET_mc_gfx9<"ds_add_rtn_f64", VReg_64, "ds_add_f64">; 482} // End SubtargetPredicate = isGFX90APlus 483 484let SubtargetPredicate = isGFX940Plus in { 485 defm DS_PK_ADD_F16 : DS_1A1D_NORET_mc_gfx9<"ds_pk_add_f16">; 486 defm DS_PK_ADD_RTN_F16 : DS_1A1D_RET_mc_gfx9<"ds_pk_add_rtn_f16", VGPR_32, "ds_pk_add_f16">; 487 defm DS_PK_ADD_BF16 : DS_1A1D_NORET_mc_gfx9<"ds_pk_add_bf16">; 488 defm DS_PK_ADD_RTN_BF16 : DS_1A1D_RET_mc_gfx9<"ds_pk_add_rtn_bf16", VGPR_32, "ds_pk_add_bf16">; 489} // End SubtargetPredicate = isGFX940Plus 490 491defm DS_CMPSTORE_B32 : DS_1A2D_NORET_mc<"ds_cmpstore_b32">; 492defm DS_CMPSTORE_F32 : DS_1A2D_NORET_mc<"ds_cmpstore_f32">; 493defm DS_CMPSTORE_B64 : DS_1A2D_NORET_mc<"ds_cmpstore_b64", VReg_64>; 494defm DS_CMPSTORE_F64 : DS_1A2D_NORET_mc<"ds_cmpstore_f64", VReg_64>; 495defm DS_CMPSTORE_RTN_B32 : DS_1A2D_RET_mc<"ds_cmpstore_rtn_b32", VGPR_32, "ds_cmpstore_b32">; 496defm DS_CMPSTORE_RTN_F32 : DS_1A2D_RET_mc<"ds_cmpstore_rtn_f32", VGPR_32, "ds_cmpstore_f32">; 497defm DS_CMPSTORE_RTN_B64 : DS_1A2D_RET_mc<"ds_cmpstore_rtn_b64", VReg_64, "ds_cmpstore_b64">; 498defm DS_CMPSTORE_RTN_F64 : DS_1A2D_RET_mc<"ds_cmpstore_rtn_f64", VReg_64, "ds_cmpstore_f64">; 499 500defm DS_MSKOR_B32 : DS_1A2D_NORET_mc<"ds_mskor_b32">; 501defm DS_CMPST_B32 : DS_1A2D_NORET_mc<"ds_cmpst_b32">; 502defm DS_CMPST_F32 : DS_1A2D_NORET_mc<"ds_cmpst_f32">; 503 504defm DS_ADD_U64 : DS_1A1D_NORET_mc<"ds_add_u64", VReg_64>; 505defm DS_SUB_U64 : DS_1A1D_NORET_mc<"ds_sub_u64", VReg_64>; 506defm DS_RSUB_U64 : DS_1A1D_NORET_mc<"ds_rsub_u64", VReg_64>; 507defm DS_INC_U64 : DS_1A1D_NORET_mc<"ds_inc_u64", VReg_64>; 508defm DS_DEC_U64 : DS_1A1D_NORET_mc<"ds_dec_u64", VReg_64>; 509defm DS_MIN_I64 : DS_1A1D_NORET_mc<"ds_min_i64", VReg_64>; 510defm DS_MAX_I64 : DS_1A1D_NORET_mc<"ds_max_i64", VReg_64>; 511defm DS_MIN_U64 : DS_1A1D_NORET_mc<"ds_min_u64", VReg_64>; 512defm DS_MAX_U64 : DS_1A1D_NORET_mc<"ds_max_u64", VReg_64>; 513defm DS_AND_B64 : DS_1A1D_NORET_mc<"ds_and_b64", VReg_64>; 514defm DS_OR_B64 : DS_1A1D_NORET_mc<"ds_or_b64", VReg_64>; 515defm DS_XOR_B64 : DS_1A1D_NORET_mc<"ds_xor_b64", VReg_64>; 516defm DS_MSKOR_B64 : DS_1A2D_NORET_mc<"ds_mskor_b64", VReg_64>; 517let mayLoad = 0 in { 518defm DS_WRITE_B64 : DS_1A1D_NORET_mc<"ds_write_b64", VReg_64>; 519defm DS_WRITE2_B64 : DS_1A2D_Off8_NORET_mc<"ds_write2_b64", VReg_64>; 520defm DS_WRITE2ST64_B64: DS_1A2D_Off8_NORET_mc<"ds_write2st64_b64", VReg_64>; 521} 522defm DS_CMPST_B64 : DS_1A2D_NORET_mc<"ds_cmpst_b64", VReg_64>; 523defm DS_CMPST_F64 : DS_1A2D_NORET_mc<"ds_cmpst_f64", VReg_64>; 524defm DS_MIN_F64 : DS_1A1D_NORET_mc<"ds_min_f64", VReg_64>; 525defm DS_MAX_F64 : DS_1A1D_NORET_mc<"ds_max_f64", VReg_64>; 526 527defm DS_ADD_RTN_U32 : DS_1A1D_RET_mc<"ds_add_rtn_u32", VGPR_32, "ds_add_u32">; 528 529let SubtargetPredicate = HasLDSFPAtomicAdd in { 530defm DS_ADD_RTN_F32 : DS_1A1D_RET_mc<"ds_add_rtn_f32", VGPR_32, "ds_add_f32">; 531} 532defm DS_SUB_RTN_U32 : DS_1A1D_RET_mc<"ds_sub_rtn_u32", VGPR_32, "ds_sub_u32">; 533defm DS_RSUB_RTN_U32 : DS_1A1D_RET_mc<"ds_rsub_rtn_u32", VGPR_32, "ds_rsub_u32">; 534defm DS_INC_RTN_U32 : DS_1A1D_RET_mc<"ds_inc_rtn_u32", VGPR_32, "ds_inc_u32">; 535defm DS_DEC_RTN_U32 : DS_1A1D_RET_mc<"ds_dec_rtn_u32", VGPR_32, "ds_dec_u32">; 536defm DS_MIN_RTN_I32 : DS_1A1D_RET_mc<"ds_min_rtn_i32", VGPR_32, "ds_min_i32">; 537defm DS_MAX_RTN_I32 : DS_1A1D_RET_mc<"ds_max_rtn_i32", VGPR_32, "ds_max_i32">; 538defm DS_MIN_RTN_U32 : DS_1A1D_RET_mc<"ds_min_rtn_u32", VGPR_32, "ds_min_u32">; 539defm DS_MAX_RTN_U32 : DS_1A1D_RET_mc<"ds_max_rtn_u32", VGPR_32, "ds_max_u32">; 540defm DS_AND_RTN_B32 : DS_1A1D_RET_mc<"ds_and_rtn_b32", VGPR_32, "ds_and_b32">; 541defm DS_OR_RTN_B32 : DS_1A1D_RET_mc<"ds_or_rtn_b32", VGPR_32, "ds_or_b32">; 542defm DS_XOR_RTN_B32 : DS_1A1D_RET_mc<"ds_xor_rtn_b32", VGPR_32, "ds_xor_b32">; 543defm DS_MSKOR_RTN_B32 : DS_1A2D_RET_mc<"ds_mskor_rtn_b32", VGPR_32, "ds_mskor_b32">; 544defm DS_CMPST_RTN_B32 : DS_1A2D_RET_mc<"ds_cmpst_rtn_b32", VGPR_32, "ds_cmpst_b32">; 545defm DS_CMPST_RTN_F32 : DS_1A2D_RET_mc<"ds_cmpst_rtn_f32", VGPR_32, "ds_cmpst_f32">; 546defm DS_MIN_RTN_F32 : DS_1A1D_RET_mc<"ds_min_rtn_f32", VGPR_32, "ds_min_f32">; 547defm DS_MAX_RTN_F32 : DS_1A1D_RET_mc<"ds_max_rtn_f32", VGPR_32, "ds_max_f32">; 548 549defm DS_WRXCHG_RTN_B32 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b32">; 550defm DS_WRXCHG2_RTN_B32 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b32", VReg_64, VGPR_32>; 551defm DS_WRXCHG2ST64_RTN_B32 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b32", VReg_64, VGPR_32>; 552 553defm DS_ADD_RTN_U64 : DS_1A1D_RET_mc<"ds_add_rtn_u64", VReg_64, "ds_add_u64">; 554defm DS_SUB_RTN_U64 : DS_1A1D_RET_mc<"ds_sub_rtn_u64", VReg_64, "ds_sub_u64">; 555defm DS_RSUB_RTN_U64 : DS_1A1D_RET_mc<"ds_rsub_rtn_u64", VReg_64, "ds_rsub_u64">; 556defm DS_INC_RTN_U64 : DS_1A1D_RET_mc<"ds_inc_rtn_u64", VReg_64, "ds_inc_u64">; 557defm DS_DEC_RTN_U64 : DS_1A1D_RET_mc<"ds_dec_rtn_u64", VReg_64, "ds_dec_u64">; 558defm DS_MIN_RTN_I64 : DS_1A1D_RET_mc<"ds_min_rtn_i64", VReg_64, "ds_min_i64">; 559defm DS_MAX_RTN_I64 : DS_1A1D_RET_mc<"ds_max_rtn_i64", VReg_64, "ds_max_i64">; 560defm DS_MIN_RTN_U64 : DS_1A1D_RET_mc<"ds_min_rtn_u64", VReg_64, "ds_min_u64">; 561defm DS_MAX_RTN_U64 : DS_1A1D_RET_mc<"ds_max_rtn_u64", VReg_64, "ds_max_u64">; 562defm DS_AND_RTN_B64 : DS_1A1D_RET_mc<"ds_and_rtn_b64", VReg_64, "ds_and_b64">; 563defm DS_OR_RTN_B64 : DS_1A1D_RET_mc<"ds_or_rtn_b64", VReg_64, "ds_or_b64">; 564defm DS_XOR_RTN_B64 : DS_1A1D_RET_mc<"ds_xor_rtn_b64", VReg_64, "ds_xor_b64">; 565defm DS_MSKOR_RTN_B64 : DS_1A2D_RET_mc<"ds_mskor_rtn_b64", VReg_64, "ds_mskor_b64">; 566defm DS_CMPST_RTN_B64 : DS_1A2D_RET_mc<"ds_cmpst_rtn_b64", VReg_64, "ds_cmpst_b64">; 567defm DS_CMPST_RTN_F64 : DS_1A2D_RET_mc<"ds_cmpst_rtn_f64", VReg_64, "ds_cmpst_f64">; 568defm DS_MIN_RTN_F64 : DS_1A1D_RET_mc<"ds_min_rtn_f64", VReg_64, "ds_min_f64">; 569defm DS_MAX_RTN_F64 : DS_1A1D_RET_mc<"ds_max_rtn_f64", VReg_64, "ds_max_f64">; 570 571defm DS_WRXCHG_RTN_B64 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b64", VReg_64>; 572defm DS_WRXCHG2_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b64", VReg_128, VReg_64>; 573defm DS_WRXCHG2ST64_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b64", VReg_128, VReg_64>; 574 575let isConvergent = 1, usesCustomInserter = 1 in { 576def DS_GWS_INIT : DS_GWS_1D<"ds_gws_init"> { 577 let mayLoad = 0; 578} 579def DS_GWS_SEMA_V : DS_GWS_0D<"ds_gws_sema_v">; 580def DS_GWS_SEMA_BR : DS_GWS_1D<"ds_gws_sema_br">; 581def DS_GWS_SEMA_P : DS_GWS_0D<"ds_gws_sema_p">; 582def DS_GWS_BARRIER : DS_GWS_1D<"ds_gws_barrier">; 583} 584 585let SubtargetPredicate = HasDsSrc2Insts in { 586def DS_ADD_SRC2_U32 : DS_1A<"ds_add_src2_u32">; 587def DS_SUB_SRC2_U32 : DS_1A<"ds_sub_src2_u32">; 588def DS_RSUB_SRC2_U32 : DS_1A<"ds_rsub_src2_u32">; 589def DS_INC_SRC2_U32 : DS_1A<"ds_inc_src2_u32">; 590def DS_DEC_SRC2_U32 : DS_1A<"ds_dec_src2_u32">; 591def DS_MIN_SRC2_I32 : DS_1A<"ds_min_src2_i32">; 592def DS_MAX_SRC2_I32 : DS_1A<"ds_max_src2_i32">; 593def DS_MIN_SRC2_U32 : DS_1A<"ds_min_src2_u32">; 594def DS_MAX_SRC2_U32 : DS_1A<"ds_max_src2_u32">; 595def DS_AND_SRC2_B32 : DS_1A<"ds_and_src2_b32">; 596def DS_OR_SRC2_B32 : DS_1A<"ds_or_src2_b32">; 597def DS_XOR_SRC2_B32 : DS_1A<"ds_xor_src2_b32">; 598def DS_MIN_SRC2_F32 : DS_1A<"ds_min_src2_f32">; 599def DS_MAX_SRC2_F32 : DS_1A<"ds_max_src2_f32">; 600 601def DS_ADD_SRC2_U64 : DS_1A<"ds_add_src2_u64">; 602def DS_SUB_SRC2_U64 : DS_1A<"ds_sub_src2_u64">; 603def DS_RSUB_SRC2_U64 : DS_1A<"ds_rsub_src2_u64">; 604def DS_INC_SRC2_U64 : DS_1A<"ds_inc_src2_u64">; 605def DS_DEC_SRC2_U64 : DS_1A<"ds_dec_src2_u64">; 606def DS_MIN_SRC2_I64 : DS_1A<"ds_min_src2_i64">; 607def DS_MAX_SRC2_I64 : DS_1A<"ds_max_src2_i64">; 608def DS_MIN_SRC2_U64 : DS_1A<"ds_min_src2_u64">; 609def DS_MAX_SRC2_U64 : DS_1A<"ds_max_src2_u64">; 610def DS_AND_SRC2_B64 : DS_1A<"ds_and_src2_b64">; 611def DS_OR_SRC2_B64 : DS_1A<"ds_or_src2_b64">; 612def DS_XOR_SRC2_B64 : DS_1A<"ds_xor_src2_b64">; 613def DS_MIN_SRC2_F64 : DS_1A<"ds_min_src2_f64">; 614def DS_MAX_SRC2_F64 : DS_1A<"ds_max_src2_f64">; 615 616def DS_WRITE_SRC2_B32 : DS_1A<"ds_write_src2_b32">; 617def DS_WRITE_SRC2_B64 : DS_1A<"ds_write_src2_b64">; 618} // End SubtargetPredicate = HasDsSrc2Insts 619 620let Uses = [EXEC], mayLoad = 0, mayStore = 0, isConvergent = 1 in { 621def DS_SWIZZLE_B32 : DS_1A_RET <"ds_swizzle_b32", VGPR_32, 0, SwizzleImm>; 622} 623 624let mayStore = 0 in { 625defm DS_READ_I8 : DS_1A_RET_mc<"ds_read_i8">; 626defm DS_READ_U8 : DS_1A_RET_mc<"ds_read_u8">; 627defm DS_READ_I16 : DS_1A_RET_mc<"ds_read_i16">; 628defm DS_READ_U16 : DS_1A_RET_mc<"ds_read_u16">; 629defm DS_READ_B32 : DS_1A_RET_mc<"ds_read_b32">; 630defm DS_READ_B64 : DS_1A_RET_mc<"ds_read_b64", VReg_64>; 631 632defm DS_READ2_B32 : DS_1A_Off8_RET_mc<"ds_read2_b32", VReg_64>; 633defm DS_READ2ST64_B32: DS_1A_Off8_RET_mc<"ds_read2st64_b32", VReg_64>; 634 635defm DS_READ2_B64 : DS_1A_Off8_RET_mc<"ds_read2_b64", VReg_128>; 636defm DS_READ2ST64_B64: DS_1A_Off8_RET_mc<"ds_read2st64_b64", VReg_128>; 637 638let has_m0_read = 0 in { 639let SubtargetPredicate = HasD16LoadStore in { 640def DS_READ_U8_D16 : DS_1A_RET_Tied<"ds_read_u8_d16">; 641def DS_READ_U8_D16_HI : DS_1A_RET_Tied<"ds_read_u8_d16_hi">; 642def DS_READ_I8_D16 : DS_1A_RET_Tied<"ds_read_i8_d16">; 643def DS_READ_I8_D16_HI : DS_1A_RET_Tied<"ds_read_i8_d16_hi">; 644def DS_READ_U16_D16 : DS_1A_RET_Tied<"ds_read_u16_d16">; 645def DS_READ_U16_D16_HI : DS_1A_RET_Tied<"ds_read_u16_d16_hi">; 646} 647} // End has_m0_read = 0 648 649let SubtargetPredicate = HasDSAddTid in { 650def DS_READ_ADDTID_B32 : DS_0A_RET<"ds_read_addtid_b32">; 651} 652 653} // End mayStore = 0 654 655def DS_CONSUME : DS_0A_RET<"ds_consume">; 656def DS_APPEND : DS_0A_RET<"ds_append">; 657 658let SubtargetPredicate = isNotGFX90APlus in 659def DS_ORDERED_COUNT : DS_1A_RET_GDS<"ds_ordered_count">; 660 661//===----------------------------------------------------------------------===// 662// Instruction definitions for CI and newer. 663//===----------------------------------------------------------------------===// 664 665let SubtargetPredicate = isGFX7Plus in { 666 667defm DS_WRAP_RTN_B32 : DS_1A2D_RET_mc<"ds_wrap_rtn_b32", VGPR_32>; 668defm DS_CONDXCHG32_RTN_B64 : DS_1A1D_RET_mc<"ds_condxchg32_rtn_b64", VReg_64>; 669 670let isConvergent = 1, usesCustomInserter = 1 in { 671def DS_GWS_SEMA_RELEASE_ALL : DS_GWS_0D<"ds_gws_sema_release_all">; 672} 673 674let mayStore = 0 in { 675defm DS_READ_B96 : DS_1A_RET_mc<"ds_read_b96", VReg_96>; 676defm DS_READ_B128: DS_1A_RET_mc<"ds_read_b128", VReg_128>; 677} // End mayStore = 0 678 679let mayLoad = 0 in { 680defm DS_WRITE_B96 : DS_1A1D_NORET_mc<"ds_write_b96", VReg_96>; 681defm DS_WRITE_B128 : DS_1A1D_NORET_mc<"ds_write_b128", VReg_128>; 682} // End mayLoad = 0 683 684def DS_NOP : DS_VOID<"ds_nop">; 685 686} // let SubtargetPredicate = isGFX7Plus 687 688//===----------------------------------------------------------------------===// 689// Instruction definitions for VI and newer. 690//===----------------------------------------------------------------------===// 691 692let SubtargetPredicate = isGFX8Plus in { 693 694let Uses = [EXEC] in { 695def DS_PERMUTE_B32 : DS_1A1D_PERMUTE <"ds_permute_b32", 696 int_amdgcn_ds_permute>; 697def DS_BPERMUTE_B32 : DS_1A1D_PERMUTE <"ds_bpermute_b32", 698 int_amdgcn_ds_bpermute>; 699} 700 701} // let SubtargetPredicate = isGFX8Plus 702 703let SubtargetPredicate = HasLDSFPAtomicAdd, OtherPredicates = [HasDsSrc2Insts] in { 704def DS_ADD_SRC2_F32 : DS_1A<"ds_add_src2_f32">; 705} 706 707 708//===----------------------------------------------------------------------===// 709// Instruction definitions for GFX11 and newer. 710//===----------------------------------------------------------------------===// 711 712let SubtargetPredicate = isGFX11Plus in { 713 714def DS_ADD_GS_REG_RTN : DS_0A1D_RET_GDS<"ds_add_gs_reg_rtn", VReg_64, VGPR_32>; 715def DS_SUB_GS_REG_RTN : DS_0A1D_RET_GDS<"ds_sub_gs_reg_rtn", VReg_64, VGPR_32>; 716 717} // let SubtargetPredicate = isGFX11Plus 718 719//===----------------------------------------------------------------------===// 720// DS Patterns 721//===----------------------------------------------------------------------===// 722 723def : GCNPat < 724 (int_amdgcn_ds_swizzle i32:$src, timm:$offset16), 725 (DS_SWIZZLE_B32 VGPR_32:$src, (as_i16timm $offset16), (i1 0)) 726>; 727 728class DSReadPat <DS_Pseudo inst, ValueType vt, PatFrag frag, int gds=0> : GCNPat < 729 (vt (frag (DS1Addr1Offset i32:$ptr, i16:$offset))), 730 (inst $ptr, offset:$offset, (i1 gds)) 731>; 732 733multiclass DSReadPat_mc<DS_Pseudo inst, ValueType vt, string frag> { 734 735 let OtherPredicates = [LDSRequiresM0Init] in { 736 def : DSReadPat<inst, vt, !cast<PatFrag>(frag#"_m0")>; 737 } 738 739 let OtherPredicates = [NotLDSRequiresM0Init] in { 740 def : DSReadPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>; 741 } 742} 743 744class DSReadPat_D16 <DS_Pseudo inst, PatFrag frag, ValueType vt> : GCNPat < 745 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$in), 746 (inst $ptr, offset:$offset, (i1 0), $in) 747>; 748 749defm : DSReadPat_mc <DS_READ_I8, i32, "sextloadi8_local">; 750defm : DSReadPat_mc <DS_READ_I8, i16, "sextloadi8_local">; 751defm : DSReadPat_mc <DS_READ_U8, i32, "extloadi8_local">; 752defm : DSReadPat_mc <DS_READ_U8, i32, "zextloadi8_local">; 753defm : DSReadPat_mc <DS_READ_U8, i16, "extloadi8_local">; 754defm : DSReadPat_mc <DS_READ_U8, i16, "zextloadi8_local">; 755defm : DSReadPat_mc <DS_READ_I16, i32, "sextloadi16_local">; 756defm : DSReadPat_mc <DS_READ_I16, i32, "sextloadi16_local">; 757defm : DSReadPat_mc <DS_READ_U16, i32, "extloadi16_local">; 758defm : DSReadPat_mc <DS_READ_U16, i32, "zextloadi16_local">; 759defm : DSReadPat_mc <DS_READ_U16, i16, "load_local">; 760 761foreach vt = Reg32Types.types in { 762defm : DSReadPat_mc <DS_READ_B32, vt, "load_local">; 763} 764 765defm : DSReadPat_mc <DS_READ_U8, i16, "atomic_load_8_local">; 766defm : DSReadPat_mc <DS_READ_U8, i32, "atomic_load_8_local">; 767defm : DSReadPat_mc <DS_READ_U16, i16, "atomic_load_16_local">; 768defm : DSReadPat_mc <DS_READ_U16, i32, "atomic_load_16_local">; 769defm : DSReadPat_mc <DS_READ_B32, i32, "atomic_load_32_local">; 770defm : DSReadPat_mc <DS_READ_B64, i64, "atomic_load_64_local">; 771 772let OtherPredicates = [D16PreservesUnusedBits] in { 773def : DSReadPat_D16<DS_READ_U16_D16_HI, load_d16_hi_local, v2i16>; 774def : DSReadPat_D16<DS_READ_U16_D16_HI, load_d16_hi_local, v2f16>; 775def : DSReadPat_D16<DS_READ_U8_D16_HI, az_extloadi8_d16_hi_local, v2i16>; 776def : DSReadPat_D16<DS_READ_U8_D16_HI, az_extloadi8_d16_hi_local, v2f16>; 777def : DSReadPat_D16<DS_READ_I8_D16_HI, sextloadi8_d16_hi_local, v2i16>; 778def : DSReadPat_D16<DS_READ_I8_D16_HI, sextloadi8_d16_hi_local, v2f16>; 779 780def : DSReadPat_D16<DS_READ_U16_D16, load_d16_lo_local, v2i16>; 781def : DSReadPat_D16<DS_READ_U16_D16, load_d16_lo_local, v2f16>; 782def : DSReadPat_D16<DS_READ_U8_D16, az_extloadi8_d16_lo_local, v2i16>; 783def : DSReadPat_D16<DS_READ_U8_D16, az_extloadi8_d16_lo_local, v2f16>; 784def : DSReadPat_D16<DS_READ_I8_D16, sextloadi8_d16_lo_local, v2i16>; 785def : DSReadPat_D16<DS_READ_I8_D16, sextloadi8_d16_lo_local, v2f16>; 786} 787 788class DSWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag, int gds=0> : GCNPat < 789 (frag vt:$value, (DS1Addr1Offset i32:$ptr, i16:$offset)), 790 (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds)) 791>; 792 793multiclass DSWritePat_mc <DS_Pseudo inst, ValueType vt, string frag> { 794 let OtherPredicates = [LDSRequiresM0Init] in { 795 def : DSWritePat<inst, vt, !cast<PatFrag>(frag#"_m0")>; 796 } 797 798 let OtherPredicates = [NotLDSRequiresM0Init] in { 799 def : DSWritePat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>; 800 } 801} 802 803// Irritatingly, atomic_store reverses the order of operands from a 804// normal store. 805class DSAtomicWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat < 806 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$value), 807 (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 0)) 808>; 809 810multiclass DSAtomicWritePat_mc <DS_Pseudo inst, ValueType vt, string frag> { 811 let OtherPredicates = [LDSRequiresM0Init] in { 812 def : DSAtomicWritePat<inst, vt, !cast<PatFrag>(frag#"_m0")>; 813 } 814 815 let OtherPredicates = [NotLDSRequiresM0Init] in { 816 def : DSAtomicWritePat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>; 817 } 818} 819 820defm : DSWritePat_mc <DS_WRITE_B8, i32, "truncstorei8_local">; 821defm : DSWritePat_mc <DS_WRITE_B16, i32, "truncstorei16_local">; 822defm : DSWritePat_mc <DS_WRITE_B8, i16, "truncstorei8_local">; 823defm : DSWritePat_mc <DS_WRITE_B16, i16, "store_local">; 824 825foreach vt = Reg32Types.types in { 826defm : DSWritePat_mc <DS_WRITE_B32, vt, "store_local">; 827} 828 829defm : DSAtomicWritePat_mc <DS_WRITE_B8, i16, "atomic_store_8_local">; 830defm : DSAtomicWritePat_mc <DS_WRITE_B8, i32, "atomic_store_8_local">; 831defm : DSAtomicWritePat_mc <DS_WRITE_B16, i16, "atomic_store_16_local">; 832defm : DSAtomicWritePat_mc <DS_WRITE_B16, i32, "atomic_store_16_local">; 833defm : DSAtomicWritePat_mc <DS_WRITE_B32, i32, "atomic_store_32_local">; 834defm : DSAtomicWritePat_mc <DS_WRITE_B64, i64, "atomic_store_64_local">; 835 836let OtherPredicates = [HasD16LoadStore] in { 837def : DSWritePat <DS_WRITE_B16_D16_HI, i32, store_hi16_local>; 838def : DSWritePat <DS_WRITE_B8_D16_HI, i32, truncstorei8_hi16_local>; 839} 840 841class DS64Bit4ByteAlignedReadPat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat < 842 (vt:$value (frag (DS64Bit4ByteAligned i32:$ptr, i8:$offset0, i8:$offset1))), 843 (inst $ptr, $offset0, $offset1, (i1 0)) 844>; 845 846class DS64Bit4ByteAlignedWritePat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat< 847 (frag vt:$value, (DS64Bit4ByteAligned i32:$ptr, i8:$offset0, i8:$offset1)), 848 (inst $ptr, (i32 (EXTRACT_SUBREG VReg_64:$value, sub0)), 849 (i32 (EXTRACT_SUBREG VReg_64:$value, sub1)), $offset0, $offset1, 850 (i1 0)) 851>; 852 853class DS128Bit8ByteAlignedReadPat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat < 854 (vt:$value (frag (DS128Bit8ByteAligned i32:$ptr, i8:$offset0, i8:$offset1))), 855 (inst $ptr, $offset0, $offset1, (i1 0)) 856>; 857 858class DS128Bit8ByteAlignedWritePat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat< 859 (frag vt:$value, (DS128Bit8ByteAligned i32:$ptr, i8:$offset0, i8:$offset1)), 860 (inst $ptr, (i64 (EXTRACT_SUBREG VReg_128:$value, sub0_sub1)), 861 (i64 (EXTRACT_SUBREG VReg_128:$value, sub2_sub3)), $offset0, $offset1, 862 (i1 0)) 863>; 864 865multiclass DS64Bit4ByteAlignedPat_mc<ValueType vt> { 866 let OtherPredicates = [LDSRequiresM0Init, isGFX7Plus] in { 867 def : DS64Bit4ByteAlignedReadPat<DS_READ2_B32, vt, load_local_m0>; 868 def : DS64Bit4ByteAlignedWritePat<DS_WRITE2_B32, vt, store_local_m0>; 869 } 870 871 let OtherPredicates = [NotLDSRequiresM0Init] in { 872 def : DS64Bit4ByteAlignedReadPat<DS_READ2_B32_gfx9, vt, load_local>; 873 def : DS64Bit4ByteAlignedWritePat<DS_WRITE2_B32_gfx9, vt, store_local>; 874 } 875} 876 877multiclass DS128Bit8ByteAlignedPat_mc<ValueType vt> { 878 let OtherPredicates = [LDSRequiresM0Init, isGFX7Plus] in { 879 def : DS128Bit8ByteAlignedReadPat<DS_READ2_B64, vt, load_local_m0>; 880 def : DS128Bit8ByteAlignedWritePat<DS_WRITE2_B64, vt, store_local_m0>; 881 } 882 883 let OtherPredicates = [NotLDSRequiresM0Init] in { 884 def : DS128Bit8ByteAlignedReadPat<DS_READ2_B64_gfx9, vt, load_local>; 885 def : DS128Bit8ByteAlignedWritePat<DS_WRITE2_B64_gfx9, vt, store_local>; 886 } 887} 888 889// v2i32 loads are split into i32 loads on SI during lowering, due to a bug 890// related to bounds checking. 891foreach vt = VReg_64.RegTypes in { 892defm : DS64Bit4ByteAlignedPat_mc<vt>; 893} 894 895foreach vt = VReg_128.RegTypes in { 896defm : DS128Bit8ByteAlignedPat_mc<vt>; 897} 898 899// Prefer ds_read over ds_read2 and ds_write over ds_write2, all other things 900// being equal, because it has a larger immediate offset range. 901let AddedComplexity = 100 in { 902 903foreach vt = VReg_64.RegTypes in { 904defm : DSReadPat_mc <DS_READ_B64, vt, "load_align8_local">; 905defm : DSWritePat_mc <DS_WRITE_B64, vt, "store_align8_local">; 906} 907 908let SubtargetPredicate = isGFX7Plus in { 909 910foreach vt = VReg_96.RegTypes in { 911defm : DSReadPat_mc <DS_READ_B96, vt, "load_align16_local">; 912defm : DSWritePat_mc <DS_WRITE_B96, vt, "store_align16_local">; 913} 914 915foreach vt = VReg_128.RegTypes in { 916defm : DSReadPat_mc <DS_READ_B128, vt, "load_align16_local">; 917defm : DSWritePat_mc <DS_WRITE_B128, vt, "store_align16_local">; 918} 919 920let SubtargetPredicate = HasUnalignedAccessMode in { 921 922// Select 64 bit loads and stores aligned less than 4 as a single ds_read_b64/ 923// ds_write_b64 instruction as this is faster than ds_read2_b32/ds_write2_b32 924// which would be used otherwise. In this case a b32 access would still be 925// misaligned, but we will have 2 of them. 926foreach vt = VReg_64.RegTypes in { 927defm : DSReadPat_mc <DS_READ_B64, vt, "load_align_less_than_4_local">; 928defm : DSWritePat_mc <DS_WRITE_B64, vt, "store_align_less_than_4_local">; 929} 930 931// Selection will split most of the unaligned 3 dword accesses due to performance 932// reasons when beneficial. Keep these two patterns for the rest of the cases. 933foreach vt = VReg_96.RegTypes in { 934defm : DSReadPat_mc <DS_READ_B96, vt, "load_local">; 935defm : DSWritePat_mc <DS_WRITE_B96, vt, "store_local">; 936} 937 938// Select 128 bit loads and stores aligned less than 4 as a single ds_read_b128/ 939// ds_write_b128 instruction as this is faster than ds_read2_b64/ds_write2_b64 940// which would be used otherwise. In this case a b64 access would still be 941// misaligned, but we will have 2 of them. 942foreach vt = VReg_128.RegTypes in { 943defm : DSReadPat_mc <DS_READ_B128, vt, "load_align_less_than_4_local">; 944defm : DSWritePat_mc <DS_WRITE_B128, vt, "store_align_less_than_4_local">; 945} 946 947} // End SubtargetPredicate = HasUnalignedAccessMode 948 949} // End SubtargetPredicate = isGFX7Plus 950 951} // End AddedComplexity = 100 952 953class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat < 954 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$value), 955 (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds)) 956>; 957 958multiclass DSAtomicRetPat_mc<DS_Pseudo inst, ValueType vt, string frag> { 959 let OtherPredicates = [LDSRequiresM0Init] in { 960 def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_local_m0_"#vt.Size)>; 961 } 962 963 let OtherPredicates = [NotLDSRequiresM0Init] in { 964 def : DSAtomicRetPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, 965 !cast<PatFrag>(frag#"_local_"#vt.Size)>; 966 } 967 968 def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_region_m0_"#vt.Size), 1>; 969} 970 971multiclass DSAtomicRetNoRetPat_mc<DS_Pseudo inst, DS_Pseudo noRetInst, 972 ValueType vt, string frag> { 973 let OtherPredicates = [LDSRequiresM0Init] in { 974 def : DSAtomicRetPat<inst, vt, 975 !cast<PatFrag>(frag#"_local_m0_ret_"#vt.Size)>; 976 def : DSAtomicRetPat<noRetInst, vt, 977 !cast<PatFrag>(frag#"_local_m0_noret_"#vt.Size)>; 978 } 979 980 let OtherPredicates = [NotLDSRequiresM0Init] in { 981 def : DSAtomicRetPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, 982 !cast<PatFrag>(frag#"_local_ret_"#vt.Size)>; 983 def : DSAtomicRetPat<!cast<DS_Pseudo>(!cast<string>(noRetInst)#"_gfx9"), vt, 984 !cast<PatFrag>(frag#"_local_noret_"#vt.Size)>; 985 } 986 987 def : DSAtomicRetPat<inst, vt, 988 !cast<PatFrag>(frag#"_region_m0_ret_"#vt.Size), 1>; 989 def : DSAtomicRetPat<noRetInst, vt, 990 !cast<PatFrag>(frag#"_region_m0_noret_"#vt.Size), 1>; 991} 992 993 994 995let SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 in { 996// Caution, the order of src and cmp is the *opposite* of the BUFFER_ATOMIC_CMPSWAP opcode. 997class DSAtomicCmpXChgSwapped<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat < 998 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$cmp, vt:$swap), 999 (inst $ptr, getVregSrcForVT<vt>.ret:$cmp, getVregSrcForVT<vt>.ret:$swap, offset:$offset, (i1 gds)) 1000>; 1001 1002multiclass DSAtomicCmpXChgSwapped_mc<DS_Pseudo inst, DS_Pseudo noRetInst, ValueType vt, 1003 string frag> { 1004 let OtherPredicates = [LDSRequiresM0Init] in { 1005 def : DSAtomicCmpXChgSwapped<inst, vt, !cast<PatFrag>(frag#"_local_m0_ret_"#vt.Size)>; 1006 def : DSAtomicCmpXChgSwapped<noRetInst, vt, !cast<PatFrag>(frag#"_local_m0_noret_"#vt.Size)>; 1007 } 1008 1009 let OtherPredicates = [NotLDSRequiresM0Init] in { 1010 def : DSAtomicCmpXChgSwapped<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, 1011 !cast<PatFrag>(frag#"_local_ret_"#vt.Size)>; 1012 def : DSAtomicCmpXChgSwapped<!cast<DS_Pseudo>(!cast<string>(noRetInst)#"_gfx9"), vt, 1013 !cast<PatFrag>(frag#"_local_noret_"#vt.Size)>; 1014 } 1015 1016 def : DSAtomicCmpXChgSwapped<inst, vt, !cast<PatFrag>(frag#"_region_m0_ret_"#vt.Size), 1>; 1017 def : DSAtomicCmpXChgSwapped<noRetInst, vt, !cast<PatFrag>(frag#"_region_m0_noret_"#vt.Size), 1>; 1018} 1019} // End SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 1020 1021let SubtargetPredicate = isGFX11Plus in { 1022// The order of src and cmp agrees with the BUFFER_ATOMIC_CMPSWAP opcode. 1023class DSAtomicCmpXChg<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat < 1024 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$cmp, vt:$swap), 1025 (inst $ptr, getVregSrcForVT<vt>.ret:$swap, getVregSrcForVT<vt>.ret:$cmp, offset:$offset, (i1 gds)) 1026>; 1027 1028multiclass DSAtomicCmpXChg_mc<DS_Pseudo inst, DS_Pseudo noRetInst, ValueType vt, string frag> { 1029 1030 def : DSAtomicCmpXChg<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, 1031 !cast<PatFrag>(frag#"_local_ret_"#vt.Size)>; 1032 def : DSAtomicCmpXChg<!cast<DS_Pseudo>(!cast<string>(noRetInst)#"_gfx9"), vt, 1033 !cast<PatFrag>(frag#"_local_noret_"#vt.Size)>; 1034 1035 def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_region_m0_ret_"#vt.Size), 1>; 1036 def : DSAtomicCmpXChg<noRetInst, vt, !cast<PatFrag>(frag#"_region_m0_noret_"#vt.Size), 1>; 1037} 1038} // End SubtargetPredicate = isGFX11Plus 1039 1040// 32-bit atomics. 1041defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B32, i32, "atomic_swap">; 1042defm : DSAtomicRetNoRetPat_mc<DS_ADD_RTN_U32, DS_ADD_U32, i32, "atomic_load_add">; 1043defm : DSAtomicRetNoRetPat_mc<DS_SUB_RTN_U32, DS_SUB_U32, i32, "atomic_load_sub">; 1044defm : DSAtomicRetNoRetPat_mc<DS_INC_RTN_U32, DS_INC_U32, i32, "atomic_inc">; 1045defm : DSAtomicRetNoRetPat_mc<DS_DEC_RTN_U32, DS_DEC_U32, i32, "atomic_dec">; 1046defm : DSAtomicRetNoRetPat_mc<DS_AND_RTN_B32, DS_AND_B32, i32, "atomic_load_and">; 1047defm : DSAtomicRetNoRetPat_mc<DS_OR_RTN_B32, DS_OR_B32, i32, "atomic_load_or">; 1048defm : DSAtomicRetNoRetPat_mc<DS_XOR_RTN_B32, DS_XOR_B32, i32, "atomic_load_xor">; 1049defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_I32, DS_MIN_I32, i32, "atomic_load_min">; 1050defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_I32, DS_MAX_I32, i32, "atomic_load_max">; 1051defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_U32, DS_MIN_U32, i32, "atomic_load_umin">; 1052defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_U32, DS_MAX_U32, i32, "atomic_load_umax">; 1053defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_F32, DS_MIN_F32, f32, "atomic_load_fmin">; 1054defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_F32, DS_MAX_F32, f32, "atomic_load_fmax">; 1055 1056let SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 in { 1057defm : DSAtomicCmpXChgSwapped_mc<DS_CMPST_RTN_B32, DS_CMPST_B32, i32, "atomic_cmp_swap">; 1058} 1059 1060let SubtargetPredicate = isGFX11Plus in { 1061defm : DSAtomicCmpXChg_mc<DS_CMPSTORE_RTN_B32, DS_CMPSTORE_B32, i32, "atomic_cmp_swap">; 1062} 1063 1064let SubtargetPredicate = HasLDSFPAtomicAdd in { 1065defm : DSAtomicRetNoRetPat_mc<DS_ADD_RTN_F32, DS_ADD_F32, f32, "atomic_load_fadd">; 1066} 1067 1068// 64-bit atomics. 1069defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B64, i64, "atomic_swap">; 1070defm : DSAtomicRetNoRetPat_mc<DS_ADD_RTN_U64, DS_ADD_U64, i64, "atomic_load_add">; 1071defm : DSAtomicRetNoRetPat_mc<DS_SUB_RTN_U64, DS_SUB_U64, i64, "atomic_load_sub">; 1072defm : DSAtomicRetNoRetPat_mc<DS_INC_RTN_U64, DS_INC_U64, i64, "atomic_inc">; 1073defm : DSAtomicRetNoRetPat_mc<DS_DEC_RTN_U64, DS_DEC_U64, i64, "atomic_dec">; 1074defm : DSAtomicRetNoRetPat_mc<DS_AND_RTN_B64, DS_AND_B64, i64, "atomic_load_and">; 1075defm : DSAtomicRetNoRetPat_mc<DS_OR_RTN_B64, DS_OR_B64, i64, "atomic_load_or">; 1076defm : DSAtomicRetNoRetPat_mc<DS_XOR_RTN_B64, DS_XOR_B64, i64, "atomic_load_xor">; 1077defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_I64, DS_MIN_I64, i64, "atomic_load_min">; 1078defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_I64, DS_MAX_I64, i64, "atomic_load_max">; 1079defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_U64, DS_MIN_U64, i64, "atomic_load_umin">; 1080defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_U64, DS_MAX_U64, i64, "atomic_load_umax">; 1081defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_F64, DS_MIN_F64, f64, "atomic_load_fmin">; 1082defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_F64, DS_MAX_F64, f64, "atomic_load_fmax">; 1083 1084let SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 in { 1085defm : DSAtomicCmpXChgSwapped_mc<DS_CMPST_RTN_B64, DS_CMPST_B64, i64, "atomic_cmp_swap">; 1086} // End SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 1087 1088let SubtargetPredicate = isGFX11Plus in { 1089defm : DSAtomicCmpXChg_mc<DS_CMPSTORE_RTN_B64, DS_CMPSTORE_B64, i64, "atomic_cmp_swap">; 1090} // End SubtargetPredicate = isGFX11Plus 1091 1092let SubtargetPredicate = isGFX90APlus in { 1093def : DSAtomicRetPat<DS_ADD_RTN_F64, f64, atomic_load_fadd_local_ret_64>; 1094def : DSAtomicRetPat<DS_ADD_F64, f64, atomic_load_fadd_local_noret_64>; 1095} 1096 1097let SubtargetPredicate = isGFX940Plus in { 1098def : DSAtomicRetPat<DS_PK_ADD_RTN_F16, v2f16, atomic_load_fadd_v2f16_local_ret_32>; 1099def : DSAtomicRetPat<DS_PK_ADD_F16, v2f16, atomic_load_fadd_v2f16_local_noret_32>; 1100def : GCNPat < 1101 (v2i16 (int_amdgcn_ds_fadd_v2bf16_ret i32:$ptr, v2i16:$src)), 1102 (DS_PK_ADD_RTN_BF16 VGPR_32:$ptr, VGPR_32:$src, 0, 0) 1103>; 1104def : GCNPat < 1105 (v2i16 (int_amdgcn_ds_fadd_v2bf16_noret i32:$ptr, v2i16:$src)), 1106 (DS_PK_ADD_BF16 VGPR_32:$ptr, VGPR_32:$src, 0, 0) 1107>; 1108} 1109 1110def : Pat < 1111 (SIds_ordered_count i32:$value, i16:$offset), 1112 (DS_ORDERED_COUNT $value, (as_i16imm $offset)) 1113>; 1114 1115def : GCNPat < 1116 (i64 (int_amdgcn_ds_add_gs_reg_rtn i32:$src, timm:$offset32)), 1117 (DS_ADD_GS_REG_RTN VGPR_32:$src, (as_i32timm $offset32)) 1118>; 1119 1120def : GCNPat < 1121 (i32 (int_amdgcn_ds_add_gs_reg_rtn i32:$src, timm:$offset32)), 1122 (EXTRACT_SUBREG 1123 (i64 (COPY_TO_REGCLASS 1124 (DS_ADD_GS_REG_RTN VGPR_32:$src, (as_i32timm $offset32)), 1125 VReg_64)), 1126 sub0) 1127>; 1128 1129def : GCNPat < 1130 (i64 (int_amdgcn_ds_sub_gs_reg_rtn i32:$src, timm:$offset32)), 1131 (DS_SUB_GS_REG_RTN VGPR_32:$src, (as_i32timm $offset32)) 1132>; 1133 1134def : GCNPat < 1135 (i32 (int_amdgcn_ds_sub_gs_reg_rtn i32:$src, timm:$offset32)), 1136 (EXTRACT_SUBREG 1137 (i64 (COPY_TO_REGCLASS 1138 (DS_SUB_GS_REG_RTN VGPR_32:$src, (as_i32timm $offset32)), 1139 VReg_64)), 1140 sub0) 1141>; 1142 1143//===----------------------------------------------------------------------===// 1144// Target-specific instruction encodings. 1145//===----------------------------------------------------------------------===// 1146 1147//===----------------------------------------------------------------------===// 1148// Base ENC_DS for GFX6, GFX7, GFX10, GFX11. 1149//===----------------------------------------------------------------------===// 1150 1151class Base_DS_Real_gfx6_gfx7_gfx10_gfx11<bits<8> op, DS_Pseudo ps, int ef, string opName = ps.Mnemonic> : 1152 DS_Real<ps, opName>, SIMCInstr <ps.Mnemonic, ef> { 1153 1154 let Inst{7-0} = !if(ps.has_offset0, offset0, 0); 1155 let Inst{15-8} = !if(ps.has_offset1, offset1, 0); 1156 let Inst{17} = !if(ps.has_gds, gds, ps.gdsValue); 1157 let Inst{25-18} = op; 1158 let Inst{31-26} = 0x36; 1159 let Inst{39-32} = !if(ps.has_addr, addr, !if(ps.has_gws_data0, data0{7-0}, 0)); 1160 let Inst{47-40} = !if(ps.has_data0, data0{7-0}, 0); 1161 let Inst{55-48} = !if(ps.has_data1, data1{7-0}, 0); 1162 let Inst{63-56} = !if(ps.has_vdst, vdst{7-0}, 0); 1163} 1164 1165//===----------------------------------------------------------------------===// 1166// GFX11. 1167//===----------------------------------------------------------------------===// 1168 1169let AssemblerPredicate = isGFX11Plus, DecoderNamespace = "GFX11" in { 1170 multiclass DS_Real_gfx11<bits<8> op> { 1171 def _gfx11 : Base_DS_Real_gfx6_gfx7_gfx10_gfx11<op, !cast<DS_Pseudo>(NAME), 1172 SIEncodingFamily.GFX11>; 1173 } 1174 1175 multiclass DS_Real_Renamed_gfx11<bits<8> op, DS_Pseudo backing_pseudo, string real_name> { 1176 def _gfx11 : Base_DS_Real_gfx6_gfx7_gfx10_gfx11<op, backing_pseudo, SIEncodingFamily.GFX11, real_name>, 1177 MnemonicAlias<backing_pseudo.Mnemonic, real_name>, Requires<[isGFX11Plus]>; 1178 } 1179} // End AssemblerPredicate = isGFX11Plus, DecoderNamespace = "GFX11" 1180 1181defm DS_STORE_B32 : DS_Real_Renamed_gfx11<0x00d, DS_WRITE_B32, "ds_store_b32">; 1182defm DS_STORE_2ADDR_B32 : DS_Real_Renamed_gfx11<0x00e, DS_WRITE2_B32, "ds_store_2addr_b32">; 1183defm DS_STORE_2ADDR_STRIDE64_B32 : DS_Real_Renamed_gfx11<0x00f, DS_WRITE2ST64_B32, "ds_store_2addr_stride64_b32">; 1184defm DS_STORE_B8 : DS_Real_Renamed_gfx11<0x01e, DS_WRITE_B8, "ds_store_b8">; 1185defm DS_STORE_B16 : DS_Real_Renamed_gfx11<0x01f, DS_WRITE_B16, "ds_store_b16">; 1186defm DS_STOREXCHG_RTN_B32 : DS_Real_Renamed_gfx11<0x02d, DS_WRXCHG_RTN_B32, "ds_storexchg_rtn_b32">; 1187defm DS_STOREXCHG_2ADDR_RTN_B32 : DS_Real_Renamed_gfx11<0x02e, DS_WRXCHG2_RTN_B32, "ds_storexchg_2addr_rtn_b32">; 1188defm DS_STOREXCHG_2ADDR_STRIDE64_RTN_B32 : DS_Real_Renamed_gfx11<0x02f, DS_WRXCHG2ST64_RTN_B32, "ds_storexchg_2addr_stride64_rtn_b32">; 1189defm DS_LOAD_B32 : DS_Real_Renamed_gfx11<0x036, DS_READ_B32, "ds_load_b32">; 1190defm DS_LOAD_2ADDR_B32 : DS_Real_Renamed_gfx11<0x037, DS_READ2_B32, "ds_load_2addr_b32">; 1191defm DS_LOAD_2ADDR_STRIDE64_B32 : DS_Real_Renamed_gfx11<0x038, DS_READ2ST64_B32, "ds_load_2addr_stride64_b32">; 1192defm DS_LOAD_I8 : DS_Real_Renamed_gfx11<0x039, DS_READ_I8, "ds_load_i8">; 1193defm DS_LOAD_U8 : DS_Real_Renamed_gfx11<0x03a, DS_READ_U8, "ds_load_u8">; 1194defm DS_LOAD_I16 : DS_Real_Renamed_gfx11<0x03b, DS_READ_I16, "ds_load_i16">; 1195defm DS_LOAD_U16 : DS_Real_Renamed_gfx11<0x03c, DS_READ_U16, "ds_load_u16">; 1196defm DS_STORE_B64 : DS_Real_Renamed_gfx11<0x04d, DS_WRITE_B64, "ds_store_b64">; 1197defm DS_STORE_2ADDR_B64 : DS_Real_Renamed_gfx11<0x04e, DS_WRITE2_B64, "ds_store_2addr_b64">; 1198defm DS_STORE_2ADDR_STRIDE64_B64 : DS_Real_Renamed_gfx11<0x04f, DS_WRITE2ST64_B64, "ds_store_2addr_stride64_b64">; 1199defm DS_STOREXCHG_RTN_B64 : DS_Real_Renamed_gfx11<0x06d, DS_WRXCHG_RTN_B64, "ds_storexchg_rtn_b64">; 1200defm DS_STOREXCHG_2ADDR_RTN_B64 : DS_Real_Renamed_gfx11<0x06e, DS_WRXCHG2_RTN_B64, "ds_storexchg_2addr_rtn_b64">; 1201defm DS_STOREXCHG_2ADDR_STRIDE64_RTN_B64 : DS_Real_Renamed_gfx11<0x06f, DS_WRXCHG2ST64_RTN_B64, "ds_storexchg_2addr_stride64_rtn_b64">; 1202defm DS_LOAD_B64 : DS_Real_Renamed_gfx11<0x076, DS_READ_B64, "ds_load_b64">; 1203defm DS_LOAD_2ADDR_B64 : DS_Real_Renamed_gfx11<0x077, DS_READ2_B64, "ds_load_2addr_b64">; 1204defm DS_LOAD_2ADDR_STRIDE64_B64 : DS_Real_Renamed_gfx11<0x078, DS_READ2ST64_B64, "ds_load_2addr_stride64_b64">; 1205defm DS_STORE_B8_D16_HI : DS_Real_Renamed_gfx11<0x0a0, DS_WRITE_B8_D16_HI, "ds_store_b8_d16_hi">; 1206defm DS_STORE_B16_D16_HI : DS_Real_Renamed_gfx11<0x0a1, DS_WRITE_B16_D16_HI, "ds_store_b16_d16_hi">; 1207defm DS_LOAD_U8_D16 : DS_Real_Renamed_gfx11<0x0a2, DS_READ_U8_D16, "ds_load_u8_d16">; 1208defm DS_LOAD_U8_D16_HI : DS_Real_Renamed_gfx11<0x0a3, DS_READ_U8_D16_HI, "ds_load_u8_d16_hi">; 1209defm DS_LOAD_I8_D16 : DS_Real_Renamed_gfx11<0x0a4, DS_READ_I8_D16, "ds_load_i8_d16">; 1210defm DS_LOAD_I8_D16_HI : DS_Real_Renamed_gfx11<0x0a5, DS_READ_I8_D16_HI, "ds_load_i8_d16_hi">; 1211defm DS_LOAD_U16_D16 : DS_Real_Renamed_gfx11<0x0a6, DS_READ_U16_D16, "ds_load_u16_d16">; 1212defm DS_LOAD_U16_D16_HI : DS_Real_Renamed_gfx11<0x0a7, DS_READ_U16_D16_HI, "ds_load_u16_d16_hi">; 1213defm DS_STORE_ADDTID_B32 : DS_Real_Renamed_gfx11<0x0b0, DS_WRITE_ADDTID_B32, "ds_store_addtid_b32">; 1214defm DS_LOAD_ADDTID_B32 : DS_Real_Renamed_gfx11<0x0b1, DS_READ_ADDTID_B32, "ds_load_addtid_b32">; 1215defm DS_STORE_B96 : DS_Real_Renamed_gfx11<0x0de, DS_WRITE_B96, "ds_store_b96">; 1216defm DS_STORE_B128 : DS_Real_Renamed_gfx11<0x0df, DS_WRITE_B128, "ds_store_b128">; 1217defm DS_LOAD_B96 : DS_Real_Renamed_gfx11<0x0fe, DS_READ_B96, "ds_load_b96">; 1218defm DS_LOAD_B128 : DS_Real_Renamed_gfx11<0x0ff, DS_READ_B128, "ds_load_b128">; 1219 1220// DS_CMPST_* are renamed to DS_CMPSTORE_* in GFX11, but also the data operands (src and cmp) are swapped 1221// comparing to pre-GFX11. 1222// Note: the mnemonic alias is not generated to avoid a potential ambiguity due to the semantics change. 1223 1224defm DS_CMPSTORE_B32 : DS_Real_gfx11<0x010>; 1225defm DS_CMPSTORE_F32 : DS_Real_gfx11<0x011>; 1226defm DS_CMPSTORE_RTN_B32 : DS_Real_gfx11<0x030>; 1227defm DS_CMPSTORE_RTN_F32 : DS_Real_gfx11<0x031>; 1228defm DS_CMPSTORE_B64 : DS_Real_gfx11<0x050>; 1229defm DS_CMPSTORE_F64 : DS_Real_gfx11<0x051>; 1230defm DS_CMPSTORE_RTN_B64 : DS_Real_gfx11<0x070>; 1231defm DS_CMPSTORE_RTN_F64 : DS_Real_gfx11<0x071>; 1232 1233defm DS_ADD_RTN_F32 : DS_Real_gfx11<0x079>; 1234defm DS_ADD_GS_REG_RTN : DS_Real_gfx11<0x07a>; 1235defm DS_SUB_GS_REG_RTN : DS_Real_gfx11<0x07b>; 1236 1237//===----------------------------------------------------------------------===// 1238// GFX10. 1239//===----------------------------------------------------------------------===// 1240 1241let AssemblerPredicate = isGFX10Only, DecoderNamespace = "GFX10" in { 1242 multiclass DS_Real_gfx10<bits<8> op> { 1243 def _gfx10 : Base_DS_Real_gfx6_gfx7_gfx10_gfx11<op, !cast<DS_Pseudo>(NAME), 1244 SIEncodingFamily.GFX10>; 1245 } 1246} // End AssemblerPredicate = isGFX10Only, DecoderNamespace = "GFX10" 1247 1248defm DS_ADD_RTN_F32 : DS_Real_gfx10<0x055>; 1249defm DS_WRITE_B8_D16_HI : DS_Real_gfx10<0x0a0>; 1250defm DS_WRITE_B16_D16_HI : DS_Real_gfx10<0x0a1>; 1251defm DS_READ_U8_D16 : DS_Real_gfx10<0x0a2>; 1252defm DS_READ_U8_D16_HI : DS_Real_gfx10<0x0a3>; 1253defm DS_READ_I8_D16 : DS_Real_gfx10<0x0a4>; 1254defm DS_READ_I8_D16_HI : DS_Real_gfx10<0x0a5>; 1255defm DS_READ_U16_D16 : DS_Real_gfx10<0x0a6>; 1256defm DS_READ_U16_D16_HI : DS_Real_gfx10<0x0a7>; 1257defm DS_WRITE_ADDTID_B32 : DS_Real_gfx10<0x0b0>; 1258defm DS_READ_ADDTID_B32 : DS_Real_gfx10<0x0b1>; 1259 1260//===----------------------------------------------------------------------===// 1261// GFX10, GFX11. 1262//===----------------------------------------------------------------------===// 1263 1264multiclass DS_Real_gfx10_gfx11<bits<8> op> : 1265 DS_Real_gfx10<op>, DS_Real_gfx11<op>; 1266 1267defm DS_ADD_F32 : DS_Real_gfx10_gfx11<0x015>; 1268defm DS_ADD_SRC2_F32 : DS_Real_gfx10<0x095>; 1269defm DS_PERMUTE_B32 : DS_Real_gfx10_gfx11<0x0b2>; 1270defm DS_BPERMUTE_B32 : DS_Real_gfx10_gfx11<0x0b3>; 1271 1272//===----------------------------------------------------------------------===// 1273// GFX7, GFX10, GFX11. 1274//===----------------------------------------------------------------------===// 1275 1276let AssemblerPredicate = isGFX7Only, DecoderNamespace = "GFX7" in { 1277 multiclass DS_Real_gfx7<bits<8> op> { 1278 def _gfx7 : Base_DS_Real_gfx6_gfx7_gfx10_gfx11<op, !cast<DS_Pseudo>(NAME), 1279 SIEncodingFamily.SI>; 1280 } 1281} // End AssemblerPredicate = isGFX7Only, DecoderNamespace = "GFX7" 1282 1283multiclass DS_Real_gfx7_gfx10_gfx11<bits<8> op> : 1284 DS_Real_gfx7<op>, DS_Real_gfx10_gfx11<op>; 1285 1286multiclass DS_Real_gfx7_gfx10<bits<8> op> : 1287 DS_Real_gfx7<op>, DS_Real_gfx10<op>; 1288 1289// FIXME-GFX7: Add tests when upstreaming this part. 1290defm DS_GWS_SEMA_RELEASE_ALL : DS_Real_gfx7_gfx10_gfx11<0x018>; 1291defm DS_WRAP_RTN_B32 : DS_Real_gfx7_gfx10_gfx11<0x034>; 1292defm DS_CONDXCHG32_RTN_B64 : DS_Real_gfx7_gfx10_gfx11<0x07e>; 1293defm DS_WRITE_B96 : DS_Real_gfx7_gfx10<0x0de>; 1294defm DS_WRITE_B128 : DS_Real_gfx7_gfx10<0x0df>; 1295defm DS_READ_B96 : DS_Real_gfx7_gfx10<0x0fe>; 1296defm DS_READ_B128 : DS_Real_gfx7_gfx10<0x0ff>; 1297 1298//===----------------------------------------------------------------------===// 1299// GFX6, GFX7, GFX10, GFX11. 1300//===----------------------------------------------------------------------===// 1301 1302let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in { 1303 multiclass DS_Real_gfx6_gfx7<bits<8> op> { 1304 def _gfx6_gfx7 : Base_DS_Real_gfx6_gfx7_gfx10_gfx11<op, !cast<DS_Pseudo>(NAME), 1305 SIEncodingFamily.SI>; 1306 } 1307} // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" 1308 1309multiclass DS_Real_gfx6_gfx7_gfx10_gfx11<bits<8> op> : 1310 DS_Real_gfx6_gfx7<op>, DS_Real_gfx10_gfx11<op>; 1311 1312multiclass DS_Real_gfx6_gfx7_gfx10<bits<8> op> : 1313 DS_Real_gfx6_gfx7<op>, DS_Real_gfx10<op>; 1314 1315defm DS_ADD_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x000>; 1316defm DS_SUB_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x001>; 1317defm DS_RSUB_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x002>; 1318defm DS_INC_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x003>; 1319defm DS_DEC_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x004>; 1320defm DS_MIN_I32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x005>; 1321defm DS_MAX_I32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x006>; 1322defm DS_MIN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x007>; 1323defm DS_MAX_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x008>; 1324defm DS_AND_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x009>; 1325defm DS_OR_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x00a>; 1326defm DS_XOR_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x00b>; 1327defm DS_MSKOR_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x00c>; 1328 1329defm DS_WRITE_B32 : DS_Real_gfx6_gfx7_gfx10<0x00d>; 1330defm DS_WRITE2_B32 : DS_Real_gfx6_gfx7_gfx10<0x00e>; 1331defm DS_WRITE2ST64_B32 : DS_Real_gfx6_gfx7_gfx10<0x00f>; 1332defm DS_CMPST_B32 : DS_Real_gfx6_gfx7_gfx10<0x010>; 1333defm DS_CMPST_F32 : DS_Real_gfx6_gfx7_gfx10<0x011>; 1334 1335defm DS_MIN_F32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x012>; 1336defm DS_MAX_F32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x013>; 1337defm DS_NOP : DS_Real_gfx6_gfx7_gfx10_gfx11<0x014>; 1338defm DS_GWS_INIT : DS_Real_gfx6_gfx7_gfx10_gfx11<0x019>; 1339defm DS_GWS_SEMA_V : DS_Real_gfx6_gfx7_gfx10_gfx11<0x01a>; 1340defm DS_GWS_SEMA_BR : DS_Real_gfx6_gfx7_gfx10_gfx11<0x01b>; 1341defm DS_GWS_SEMA_P : DS_Real_gfx6_gfx7_gfx10_gfx11<0x01c>; 1342defm DS_GWS_BARRIER : DS_Real_gfx6_gfx7_gfx10_gfx11<0x01d>; 1343 1344defm DS_WRITE_B8 : DS_Real_gfx6_gfx7_gfx10<0x01e>; 1345defm DS_WRITE_B16 : DS_Real_gfx6_gfx7_gfx10<0x01f>; 1346 1347defm DS_ADD_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x020>; 1348defm DS_SUB_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x021>; 1349defm DS_RSUB_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x022>; 1350defm DS_INC_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x023>; 1351defm DS_DEC_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x024>; 1352defm DS_MIN_RTN_I32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x025>; 1353defm DS_MAX_RTN_I32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x026>; 1354defm DS_MIN_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x027>; 1355defm DS_MAX_RTN_U32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x028>; 1356defm DS_AND_RTN_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x029>; 1357defm DS_OR_RTN_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x02a>; 1358defm DS_XOR_RTN_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x02b>; 1359defm DS_MSKOR_RTN_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x02c>; 1360 1361defm DS_WRXCHG_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02d>; 1362defm DS_WRXCHG2_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02e>; 1363defm DS_WRXCHG2ST64_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02f>; 1364defm DS_CMPST_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x030>; 1365defm DS_CMPST_RTN_F32 : DS_Real_gfx6_gfx7_gfx10<0x031>; 1366 1367defm DS_MIN_RTN_F32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x032>; 1368defm DS_MAX_RTN_F32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x033>; 1369defm DS_SWIZZLE_B32 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x035>; 1370 1371defm DS_READ_B32 : DS_Real_gfx6_gfx7_gfx10<0x036>; 1372defm DS_READ2_B32 : DS_Real_gfx6_gfx7_gfx10<0x037>; 1373defm DS_READ2ST64_B32 : DS_Real_gfx6_gfx7_gfx10<0x038>; 1374defm DS_READ_I8 : DS_Real_gfx6_gfx7_gfx10<0x039>; 1375defm DS_READ_U8 : DS_Real_gfx6_gfx7_gfx10<0x03a>; 1376defm DS_READ_I16 : DS_Real_gfx6_gfx7_gfx10<0x03b>; 1377defm DS_READ_U16 : DS_Real_gfx6_gfx7_gfx10<0x03c>; 1378 1379defm DS_CONSUME : DS_Real_gfx6_gfx7_gfx10_gfx11<0x03d>; 1380defm DS_APPEND : DS_Real_gfx6_gfx7_gfx10_gfx11<0x03e>; 1381defm DS_ORDERED_COUNT : DS_Real_gfx6_gfx7_gfx10_gfx11<0x03f>; 1382defm DS_ADD_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x040>; 1383defm DS_SUB_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x041>; 1384defm DS_RSUB_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x042>; 1385defm DS_INC_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x043>; 1386defm DS_DEC_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x044>; 1387defm DS_MIN_I64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x045>; 1388defm DS_MAX_I64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x046>; 1389defm DS_MIN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x047>; 1390defm DS_MAX_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x048>; 1391defm DS_AND_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x049>; 1392defm DS_OR_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x04a>; 1393defm DS_XOR_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x04b>; 1394defm DS_MSKOR_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x04c>; 1395 1396defm DS_WRITE_B64 : DS_Real_gfx6_gfx7_gfx10<0x04d>; 1397defm DS_WRITE2_B64 : DS_Real_gfx6_gfx7_gfx10<0x04e>; 1398defm DS_WRITE2ST64_B64 : DS_Real_gfx6_gfx7_gfx10<0x04f>; 1399defm DS_CMPST_B64 : DS_Real_gfx6_gfx7_gfx10<0x050>; 1400defm DS_CMPST_F64 : DS_Real_gfx6_gfx7_gfx10<0x051>; 1401 1402defm DS_MIN_F64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x052>; 1403defm DS_MAX_F64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x053>; 1404defm DS_ADD_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x060>; 1405defm DS_SUB_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x061>; 1406defm DS_RSUB_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x062>; 1407defm DS_INC_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x063>; 1408defm DS_DEC_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x064>; 1409defm DS_MIN_RTN_I64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x065>; 1410defm DS_MAX_RTN_I64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x066>; 1411defm DS_MIN_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x067>; 1412defm DS_MAX_RTN_U64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x068>; 1413defm DS_AND_RTN_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x069>; 1414defm DS_OR_RTN_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x06a>; 1415defm DS_XOR_RTN_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x06b>; 1416defm DS_MSKOR_RTN_B64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x06c>; 1417 1418defm DS_WRXCHG_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06d>; 1419defm DS_WRXCHG2_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06e>; 1420defm DS_WRXCHG2ST64_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06f>; 1421defm DS_CMPST_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x070>; 1422defm DS_CMPST_RTN_F64 : DS_Real_gfx6_gfx7_gfx10<0x071>; 1423 1424defm DS_MIN_RTN_F64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x072>; 1425defm DS_MAX_RTN_F64 : DS_Real_gfx6_gfx7_gfx10_gfx11<0x073>; 1426 1427defm DS_READ_B64 : DS_Real_gfx6_gfx7_gfx10<0x076>; 1428defm DS_READ2_B64 : DS_Real_gfx6_gfx7_gfx10<0x077>; 1429defm DS_READ2ST64_B64 : DS_Real_gfx6_gfx7_gfx10<0x078>; 1430defm DS_ADD_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x080>; 1431defm DS_SUB_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x081>; 1432defm DS_RSUB_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x082>; 1433defm DS_INC_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x083>; 1434defm DS_DEC_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x084>; 1435defm DS_MIN_SRC2_I32 : DS_Real_gfx6_gfx7_gfx10<0x085>; 1436defm DS_MAX_SRC2_I32 : DS_Real_gfx6_gfx7_gfx10<0x086>; 1437defm DS_MIN_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x087>; 1438defm DS_MAX_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x088>; 1439defm DS_AND_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x089>; 1440defm DS_OR_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x08a>; 1441defm DS_XOR_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x08b>; 1442defm DS_WRITE_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x08d>; 1443defm DS_MIN_SRC2_F32 : DS_Real_gfx6_gfx7_gfx10<0x092>; 1444defm DS_MAX_SRC2_F32 : DS_Real_gfx6_gfx7_gfx10<0x093>; 1445defm DS_ADD_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c0>; 1446defm DS_SUB_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c1>; 1447defm DS_RSUB_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c2>; 1448defm DS_INC_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c3>; 1449defm DS_DEC_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c4>; 1450defm DS_MIN_SRC2_I64 : DS_Real_gfx6_gfx7_gfx10<0x0c5>; 1451defm DS_MAX_SRC2_I64 : DS_Real_gfx6_gfx7_gfx10<0x0c6>; 1452defm DS_MIN_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c7>; 1453defm DS_MAX_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c8>; 1454defm DS_AND_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0c9>; 1455defm DS_OR_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0ca>; 1456defm DS_XOR_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0cb>; 1457defm DS_WRITE_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0cd>; 1458defm DS_MIN_SRC2_F64 : DS_Real_gfx6_gfx7_gfx10<0x0d2>; 1459defm DS_MAX_SRC2_F64 : DS_Real_gfx6_gfx7_gfx10<0x0d3>; 1460 1461//===----------------------------------------------------------------------===// 1462// GFX8, GFX9 (VI). 1463//===----------------------------------------------------------------------===// 1464 1465class DS_Real_vi <bits<8> op, DS_Pseudo ps> : 1466 DS_Real <ps>, 1467 SIMCInstr <ps.Mnemonic, SIEncodingFamily.VI> { 1468 let AssemblerPredicate = isGFX8GFX9; 1469 let DecoderNamespace = "GFX8"; 1470 1471 // encoding 1472 let Inst{7-0} = !if(ps.has_offset0, offset0, 0); 1473 let Inst{15-8} = !if(ps.has_offset1, offset1, 0); 1474 let Inst{16} = !if(ps.has_gds, gds, ps.gdsValue); 1475 let Inst{24-17} = op; 1476 let Inst{25} = acc; 1477 let Inst{31-26} = 0x36; // ds prefix 1478 let Inst{39-32} = !if(ps.has_addr, addr, !if(ps.has_gws_data0, data0{7-0}, 0)); 1479 let Inst{47-40} = !if(ps.has_data0, data0{7-0}, 0); 1480 let Inst{55-48} = !if(ps.has_data1, data1{7-0}, 0); 1481 let Inst{63-56} = !if(ps.has_vdst, vdst{7-0}, 0); 1482} 1483 1484def DS_ADD_U32_vi : DS_Real_vi<0x0, DS_ADD_U32>; 1485def DS_SUB_U32_vi : DS_Real_vi<0x1, DS_SUB_U32>; 1486def DS_RSUB_U32_vi : DS_Real_vi<0x2, DS_RSUB_U32>; 1487def DS_INC_U32_vi : DS_Real_vi<0x3, DS_INC_U32>; 1488def DS_DEC_U32_vi : DS_Real_vi<0x4, DS_DEC_U32>; 1489def DS_MIN_I32_vi : DS_Real_vi<0x5, DS_MIN_I32>; 1490def DS_MAX_I32_vi : DS_Real_vi<0x6, DS_MAX_I32>; 1491def DS_MIN_U32_vi : DS_Real_vi<0x7, DS_MIN_U32>; 1492def DS_MAX_U32_vi : DS_Real_vi<0x8, DS_MAX_U32>; 1493def DS_AND_B32_vi : DS_Real_vi<0x9, DS_AND_B32>; 1494def DS_OR_B32_vi : DS_Real_vi<0xa, DS_OR_B32>; 1495def DS_XOR_B32_vi : DS_Real_vi<0xb, DS_XOR_B32>; 1496def DS_MSKOR_B32_vi : DS_Real_vi<0xc, DS_MSKOR_B32>; 1497def DS_WRITE_B32_vi : DS_Real_vi<0xd, DS_WRITE_B32>; 1498def DS_WRITE2_B32_vi : DS_Real_vi<0xe, DS_WRITE2_B32>; 1499def DS_WRITE2ST64_B32_vi : DS_Real_vi<0xf, DS_WRITE2ST64_B32>; 1500def DS_CMPST_B32_vi : DS_Real_vi<0x10, DS_CMPST_B32>; 1501def DS_CMPST_F32_vi : DS_Real_vi<0x11, DS_CMPST_F32>; 1502def DS_MIN_F32_vi : DS_Real_vi<0x12, DS_MIN_F32>; 1503def DS_MAX_F32_vi : DS_Real_vi<0x13, DS_MAX_F32>; 1504def DS_NOP_vi : DS_Real_vi<0x14, DS_NOP>; 1505def DS_ADD_F32_vi : DS_Real_vi<0x15, DS_ADD_F32>; 1506def DS_GWS_INIT_vi : DS_Real_vi<0x99, DS_GWS_INIT>; 1507def DS_GWS_SEMA_V_vi : DS_Real_vi<0x9a, DS_GWS_SEMA_V>; 1508def DS_GWS_SEMA_BR_vi : DS_Real_vi<0x9b, DS_GWS_SEMA_BR>; 1509def DS_GWS_SEMA_P_vi : DS_Real_vi<0x9c, DS_GWS_SEMA_P>; 1510def DS_GWS_BARRIER_vi : DS_Real_vi<0x9d, DS_GWS_BARRIER>; 1511def DS_WRITE_ADDTID_B32_vi : DS_Real_vi<0x1d, DS_WRITE_ADDTID_B32>; 1512def DS_WRITE_B8_vi : DS_Real_vi<0x1e, DS_WRITE_B8>; 1513def DS_WRITE_B16_vi : DS_Real_vi<0x1f, DS_WRITE_B16>; 1514def DS_ADD_RTN_U32_vi : DS_Real_vi<0x20, DS_ADD_RTN_U32>; 1515def DS_SUB_RTN_U32_vi : DS_Real_vi<0x21, DS_SUB_RTN_U32>; 1516def DS_RSUB_RTN_U32_vi : DS_Real_vi<0x22, DS_RSUB_RTN_U32>; 1517def DS_INC_RTN_U32_vi : DS_Real_vi<0x23, DS_INC_RTN_U32>; 1518def DS_DEC_RTN_U32_vi : DS_Real_vi<0x24, DS_DEC_RTN_U32>; 1519def DS_MIN_RTN_I32_vi : DS_Real_vi<0x25, DS_MIN_RTN_I32>; 1520def DS_MAX_RTN_I32_vi : DS_Real_vi<0x26, DS_MAX_RTN_I32>; 1521def DS_MIN_RTN_U32_vi : DS_Real_vi<0x27, DS_MIN_RTN_U32>; 1522def DS_MAX_RTN_U32_vi : DS_Real_vi<0x28, DS_MAX_RTN_U32>; 1523def DS_AND_RTN_B32_vi : DS_Real_vi<0x29, DS_AND_RTN_B32>; 1524def DS_OR_RTN_B32_vi : DS_Real_vi<0x2a, DS_OR_RTN_B32>; 1525def DS_XOR_RTN_B32_vi : DS_Real_vi<0x2b, DS_XOR_RTN_B32>; 1526def DS_MSKOR_RTN_B32_vi : DS_Real_vi<0x2c, DS_MSKOR_RTN_B32>; 1527def DS_WRXCHG_RTN_B32_vi : DS_Real_vi<0x2d, DS_WRXCHG_RTN_B32>; 1528def DS_WRXCHG2_RTN_B32_vi : DS_Real_vi<0x2e, DS_WRXCHG2_RTN_B32>; 1529def DS_WRXCHG2ST64_RTN_B32_vi : DS_Real_vi<0x2f, DS_WRXCHG2ST64_RTN_B32>; 1530def DS_CMPST_RTN_B32_vi : DS_Real_vi<0x30, DS_CMPST_RTN_B32>; 1531def DS_CMPST_RTN_F32_vi : DS_Real_vi<0x31, DS_CMPST_RTN_F32>; 1532def DS_MIN_RTN_F32_vi : DS_Real_vi<0x32, DS_MIN_RTN_F32>; 1533def DS_MAX_RTN_F32_vi : DS_Real_vi<0x33, DS_MAX_RTN_F32>; 1534def DS_WRAP_RTN_B32_vi : DS_Real_vi<0x34, DS_WRAP_RTN_B32>; 1535def DS_ADD_RTN_F32_vi : DS_Real_vi<0x35, DS_ADD_RTN_F32>; 1536def DS_READ_B32_vi : DS_Real_vi<0x36, DS_READ_B32>; 1537def DS_READ2_B32_vi : DS_Real_vi<0x37, DS_READ2_B32>; 1538def DS_READ2ST64_B32_vi : DS_Real_vi<0x38, DS_READ2ST64_B32>; 1539def DS_READ_I8_vi : DS_Real_vi<0x39, DS_READ_I8>; 1540def DS_READ_U8_vi : DS_Real_vi<0x3a, DS_READ_U8>; 1541def DS_READ_I16_vi : DS_Real_vi<0x3b, DS_READ_I16>; 1542def DS_READ_U16_vi : DS_Real_vi<0x3c, DS_READ_U16>; 1543def DS_READ_ADDTID_B32_vi : DS_Real_vi<0xb6, DS_READ_ADDTID_B32>; 1544def DS_CONSUME_vi : DS_Real_vi<0xbd, DS_CONSUME>; 1545def DS_APPEND_vi : DS_Real_vi<0xbe, DS_APPEND>; 1546def DS_ORDERED_COUNT_vi : DS_Real_vi<0xbf, DS_ORDERED_COUNT>; 1547def DS_SWIZZLE_B32_vi : DS_Real_vi<0x3d, DS_SWIZZLE_B32>; 1548def DS_PERMUTE_B32_vi : DS_Real_vi<0x3e, DS_PERMUTE_B32>; 1549def DS_BPERMUTE_B32_vi : DS_Real_vi<0x3f, DS_BPERMUTE_B32>; 1550 1551def DS_ADD_U64_vi : DS_Real_vi<0x40, DS_ADD_U64>; 1552def DS_SUB_U64_vi : DS_Real_vi<0x41, DS_SUB_U64>; 1553def DS_RSUB_U64_vi : DS_Real_vi<0x42, DS_RSUB_U64>; 1554def DS_INC_U64_vi : DS_Real_vi<0x43, DS_INC_U64>; 1555def DS_DEC_U64_vi : DS_Real_vi<0x44, DS_DEC_U64>; 1556def DS_MIN_I64_vi : DS_Real_vi<0x45, DS_MIN_I64>; 1557def DS_MAX_I64_vi : DS_Real_vi<0x46, DS_MAX_I64>; 1558def DS_MIN_U64_vi : DS_Real_vi<0x47, DS_MIN_U64>; 1559def DS_MAX_U64_vi : DS_Real_vi<0x48, DS_MAX_U64>; 1560def DS_AND_B64_vi : DS_Real_vi<0x49, DS_AND_B64>; 1561def DS_OR_B64_vi : DS_Real_vi<0x4a, DS_OR_B64>; 1562def DS_XOR_B64_vi : DS_Real_vi<0x4b, DS_XOR_B64>; 1563def DS_MSKOR_B64_vi : DS_Real_vi<0x4c, DS_MSKOR_B64>; 1564def DS_WRITE_B64_vi : DS_Real_vi<0x4d, DS_WRITE_B64>; 1565def DS_WRITE2_B64_vi : DS_Real_vi<0x4E, DS_WRITE2_B64>; 1566def DS_WRITE2ST64_B64_vi : DS_Real_vi<0x4f, DS_WRITE2ST64_B64>; 1567def DS_CMPST_B64_vi : DS_Real_vi<0x50, DS_CMPST_B64>; 1568def DS_CMPST_F64_vi : DS_Real_vi<0x51, DS_CMPST_F64>; 1569def DS_MIN_F64_vi : DS_Real_vi<0x52, DS_MIN_F64>; 1570def DS_MAX_F64_vi : DS_Real_vi<0x53, DS_MAX_F64>; 1571 1572def DS_WRITE_B8_D16_HI_vi : DS_Real_vi<0x54, DS_WRITE_B8_D16_HI>; 1573def DS_WRITE_B16_D16_HI_vi : DS_Real_vi<0x55, DS_WRITE_B16_D16_HI>; 1574 1575def DS_READ_U8_D16_vi : DS_Real_vi<0x56, DS_READ_U8_D16>; 1576def DS_READ_U8_D16_HI_vi : DS_Real_vi<0x57, DS_READ_U8_D16_HI>; 1577def DS_READ_I8_D16_vi : DS_Real_vi<0x58, DS_READ_I8_D16>; 1578def DS_READ_I8_D16_HI_vi : DS_Real_vi<0x59, DS_READ_I8_D16_HI>; 1579def DS_READ_U16_D16_vi : DS_Real_vi<0x5a, DS_READ_U16_D16>; 1580def DS_READ_U16_D16_HI_vi : DS_Real_vi<0x5b, DS_READ_U16_D16_HI>; 1581 1582def DS_ADD_RTN_U64_vi : DS_Real_vi<0x60, DS_ADD_RTN_U64>; 1583def DS_SUB_RTN_U64_vi : DS_Real_vi<0x61, DS_SUB_RTN_U64>; 1584def DS_RSUB_RTN_U64_vi : DS_Real_vi<0x62, DS_RSUB_RTN_U64>; 1585def DS_INC_RTN_U64_vi : DS_Real_vi<0x63, DS_INC_RTN_U64>; 1586def DS_DEC_RTN_U64_vi : DS_Real_vi<0x64, DS_DEC_RTN_U64>; 1587def DS_MIN_RTN_I64_vi : DS_Real_vi<0x65, DS_MIN_RTN_I64>; 1588def DS_MAX_RTN_I64_vi : DS_Real_vi<0x66, DS_MAX_RTN_I64>; 1589def DS_MIN_RTN_U64_vi : DS_Real_vi<0x67, DS_MIN_RTN_U64>; 1590def DS_MAX_RTN_U64_vi : DS_Real_vi<0x68, DS_MAX_RTN_U64>; 1591def DS_AND_RTN_B64_vi : DS_Real_vi<0x69, DS_AND_RTN_B64>; 1592def DS_OR_RTN_B64_vi : DS_Real_vi<0x6a, DS_OR_RTN_B64>; 1593def DS_XOR_RTN_B64_vi : DS_Real_vi<0x6b, DS_XOR_RTN_B64>; 1594def DS_MSKOR_RTN_B64_vi : DS_Real_vi<0x6c, DS_MSKOR_RTN_B64>; 1595def DS_WRXCHG_RTN_B64_vi : DS_Real_vi<0x6d, DS_WRXCHG_RTN_B64>; 1596def DS_WRXCHG2_RTN_B64_vi : DS_Real_vi<0x6e, DS_WRXCHG2_RTN_B64>; 1597def DS_WRXCHG2ST64_RTN_B64_vi : DS_Real_vi<0x6f, DS_WRXCHG2ST64_RTN_B64>; 1598def DS_CONDXCHG32_RTN_B64_vi : DS_Real_vi<0x7e, DS_CONDXCHG32_RTN_B64>; 1599def DS_GWS_SEMA_RELEASE_ALL_vi : DS_Real_vi<0x98, DS_GWS_SEMA_RELEASE_ALL>; 1600def DS_CMPST_RTN_B64_vi : DS_Real_vi<0x70, DS_CMPST_RTN_B64>; 1601def DS_CMPST_RTN_F64_vi : DS_Real_vi<0x71, DS_CMPST_RTN_F64>; 1602def DS_MIN_RTN_F64_vi : DS_Real_vi<0x72, DS_MIN_RTN_F64>; 1603def DS_MAX_RTN_F64_vi : DS_Real_vi<0x73, DS_MAX_RTN_F64>; 1604 1605def DS_READ_B64_vi : DS_Real_vi<0x76, DS_READ_B64>; 1606def DS_READ2_B64_vi : DS_Real_vi<0x77, DS_READ2_B64>; 1607def DS_READ2ST64_B64_vi : DS_Real_vi<0x78, DS_READ2ST64_B64>; 1608 1609def DS_ADD_SRC2_U32_vi : DS_Real_vi<0x80, DS_ADD_SRC2_U32>; 1610def DS_SUB_SRC2_U32_vi : DS_Real_vi<0x81, DS_SUB_SRC2_U32>; 1611def DS_RSUB_SRC2_U32_vi : DS_Real_vi<0x82, DS_RSUB_SRC2_U32>; 1612def DS_INC_SRC2_U32_vi : DS_Real_vi<0x83, DS_INC_SRC2_U32>; 1613def DS_DEC_SRC2_U32_vi : DS_Real_vi<0x84, DS_DEC_SRC2_U32>; 1614def DS_MIN_SRC2_I32_vi : DS_Real_vi<0x85, DS_MIN_SRC2_I32>; 1615def DS_MAX_SRC2_I32_vi : DS_Real_vi<0x86, DS_MAX_SRC2_I32>; 1616def DS_MIN_SRC2_U32_vi : DS_Real_vi<0x87, DS_MIN_SRC2_U32>; 1617def DS_MAX_SRC2_U32_vi : DS_Real_vi<0x88, DS_MAX_SRC2_U32>; 1618def DS_AND_SRC2_B32_vi : DS_Real_vi<0x89, DS_AND_SRC2_B32>; 1619def DS_OR_SRC2_B32_vi : DS_Real_vi<0x8a, DS_OR_SRC2_B32>; 1620def DS_XOR_SRC2_B32_vi : DS_Real_vi<0x8b, DS_XOR_SRC2_B32>; 1621def DS_WRITE_SRC2_B32_vi : DS_Real_vi<0x8d, DS_WRITE_SRC2_B32>; 1622def DS_MIN_SRC2_F32_vi : DS_Real_vi<0x92, DS_MIN_SRC2_F32>; 1623def DS_MAX_SRC2_F32_vi : DS_Real_vi<0x93, DS_MAX_SRC2_F32>; 1624def DS_ADD_SRC2_F32_vi : DS_Real_vi<0x95, DS_ADD_SRC2_F32>; 1625def DS_ADD_SRC2_U64_vi : DS_Real_vi<0xc0, DS_ADD_SRC2_U64>; 1626def DS_SUB_SRC2_U64_vi : DS_Real_vi<0xc1, DS_SUB_SRC2_U64>; 1627def DS_RSUB_SRC2_U64_vi : DS_Real_vi<0xc2, DS_RSUB_SRC2_U64>; 1628def DS_INC_SRC2_U64_vi : DS_Real_vi<0xc3, DS_INC_SRC2_U64>; 1629def DS_DEC_SRC2_U64_vi : DS_Real_vi<0xc4, DS_DEC_SRC2_U64>; 1630def DS_MIN_SRC2_I64_vi : DS_Real_vi<0xc5, DS_MIN_SRC2_I64>; 1631def DS_MAX_SRC2_I64_vi : DS_Real_vi<0xc6, DS_MAX_SRC2_I64>; 1632def DS_MIN_SRC2_U64_vi : DS_Real_vi<0xc7, DS_MIN_SRC2_U64>; 1633def DS_MAX_SRC2_U64_vi : DS_Real_vi<0xc8, DS_MAX_SRC2_U64>; 1634def DS_AND_SRC2_B64_vi : DS_Real_vi<0xc9, DS_AND_SRC2_B64>; 1635def DS_OR_SRC2_B64_vi : DS_Real_vi<0xca, DS_OR_SRC2_B64>; 1636def DS_XOR_SRC2_B64_vi : DS_Real_vi<0xcb, DS_XOR_SRC2_B64>; 1637def DS_WRITE_SRC2_B64_vi : DS_Real_vi<0xcd, DS_WRITE_SRC2_B64>; 1638def DS_MIN_SRC2_F64_vi : DS_Real_vi<0xd2, DS_MIN_SRC2_F64>; 1639def DS_MAX_SRC2_F64_vi : DS_Real_vi<0xd3, DS_MAX_SRC2_F64>; 1640def DS_WRITE_B96_vi : DS_Real_vi<0xde, DS_WRITE_B96>; 1641def DS_WRITE_B128_vi : DS_Real_vi<0xdf, DS_WRITE_B128>; 1642def DS_READ_B96_vi : DS_Real_vi<0xfe, DS_READ_B96>; 1643def DS_READ_B128_vi : DS_Real_vi<0xff, DS_READ_B128>; 1644 1645let SubtargetPredicate = isGFX90APlus in { 1646 def DS_ADD_F64_vi : DS_Real_vi<0x5c, DS_ADD_F64>; 1647 def DS_ADD_RTN_F64_vi : DS_Real_vi<0x7c, DS_ADD_RTN_F64>; 1648} // End SubtargetPredicate = isGFX90APlus 1649 1650let SubtargetPredicate = isGFX940Plus in { 1651 def DS_PK_ADD_F16_vi : DS_Real_vi<0x17, DS_PK_ADD_F16>; 1652 def DS_PK_ADD_RTN_F16_vi : DS_Real_vi<0xb7, DS_PK_ADD_RTN_F16>; 1653 def DS_PK_ADD_BF16_vi : DS_Real_vi<0x18, DS_PK_ADD_BF16>; 1654 def DS_PK_ADD_RTN_BF16_vi : DS_Real_vi<0xb8, DS_PK_ADD_RTN_BF16>; 1655} // End SubtargetPredicate = isGFX940Plus 1656