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 ds> : 56 InstSI <ds.OutOperandList, ds.InOperandList, ds.Mnemonic # ds.AsmOperands, []>, 57 Enc64 { 58 59 let isPseudo = 0; 60 let isCodeGenOnly = 0; 61 let DS = 1; 62 let UseNamedOperandTable = 1; 63 64 // copy relevant pseudo op flags 65 let SubtargetPredicate = ds.SubtargetPredicate; 66 let OtherPredicates = ds.OtherPredicates; 67 let AsmMatchConverter = ds.AsmMatchConverter; 68 let SchedRW = ds.SchedRW; 69 70 // encoding fields 71 bits<10> vdst; 72 bits<1> gds; 73 bits<8> addr; 74 bits<10> data0; 75 bits<10> data1; 76 bits<8> offset0; 77 bits<8> offset1; 78 79 bits<16> offset; 80 let offset0 = !if(ds.has_offset, offset{7-0}, ?); 81 let offset1 = !if(ds.has_offset, offset{15-8}, ?); 82 83 bits<1> acc = !if(ds.has_vdst, vdst{9}, 84 !if(!or(ds.has_data0, ds.has_gws_data0), data0{9}, 0)); 85} 86 87 88// DS Pseudo instructions 89 90class DS_0A1D_NORET<string opName, RegisterClass rc = VGPR_32> 91: DS_Pseudo<opName, 92 (outs), 93 (ins getLdStRegisterOperand<rc>.ret:$data0, offset:$offset, gds:$gds), 94 " $data0$offset$gds"> { 95 96 let has_addr = 0; 97 let has_data1 = 0; 98 let has_vdst = 0; 99} 100 101class DS_1A1D_NORET<string opName, RegisterClass rc = VGPR_32> 102: DS_Pseudo<opName, 103 (outs), 104 (ins VGPR_32:$addr, getLdStRegisterOperand<rc>.ret:$data0, offset:$offset, gds:$gds), 105 " $addr, $data0$offset$gds"> { 106 107 let has_data1 = 0; 108 let has_vdst = 0; 109 let IsAtomicNoRet = 1; 110} 111 112multiclass DS_1A1D_NORET_mc<string opName, RegisterClass rc = VGPR_32> { 113 def "" : DS_1A1D_NORET<opName, rc>, 114 AtomicNoRet<opName, 0>; 115 116 let has_m0_read = 0 in { 117 def _gfx9 : DS_1A1D_NORET<opName, rc>, 118 AtomicNoRet<opName#"_gfx9", 0>; 119 } 120} 121 122multiclass DS_1A1D_NORET_mc_gfx9<string opName, RegisterClass rc = VGPR_32> { 123 let has_m0_read = 0 in { 124 def "" : DS_1A1D_NORET<opName, rc>, 125 AtomicNoRet<opName, 0>; 126 } 127} 128 129class DS_1A2D_NORET<string opName, RegisterClass rc = VGPR_32, 130 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 131: DS_Pseudo<opName, 132 (outs), 133 (ins VGPR_32:$addr, data_op:$data0, data_op:$data1, offset:$offset, gds:$gds), 134 " $addr, $data0, $data1$offset$gds"> { 135 136 let has_vdst = 0; 137 let IsAtomicNoRet = 1; 138} 139 140multiclass DS_1A2D_NORET_mc<string opName, RegisterClass rc = VGPR_32> { 141 def "" : DS_1A2D_NORET<opName, rc>, 142 AtomicNoRet<opName, 0>; 143 144 let has_m0_read = 0 in { 145 def _gfx9 : DS_1A2D_NORET<opName, rc>, 146 AtomicNoRet<opName#"_gfx9", 0>; 147 } 148} 149 150class DS_1A2D_Off8_NORET <string opName, RegisterClass rc = VGPR_32, 151 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 152: DS_Pseudo<opName, 153 (outs), 154 (ins VGPR_32:$addr, data_op:$data0, data_op:$data1, 155 offset0:$offset0, offset1:$offset1, gds:$gds), 156 " $addr, $data0, $data1$offset0$offset1$gds"> { 157 158 let has_vdst = 0; 159 let has_offset = 0; 160 let AsmMatchConverter = "cvtDSOffset01"; 161} 162 163multiclass DS_1A2D_Off8_NORET_mc <string opName, RegisterClass rc = VGPR_32> { 164 def "" : DS_1A2D_Off8_NORET<opName, rc>; 165 166 let has_m0_read = 0 in { 167 def _gfx9 : DS_1A2D_Off8_NORET<opName, rc>; 168 } 169} 170 171class DS_1A1D_RET <string opName, RegisterClass rc = VGPR_32, 172 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 173: DS_Pseudo<opName, 174 (outs data_op:$vdst), 175 (ins VGPR_32:$addr, data_op:$data0, offset:$offset, gds:$gds), 176 " $vdst, $addr, $data0$offset$gds"> { 177 178 let hasPostISelHook = 1; 179 let has_data1 = 0; 180 let IsAtomicRet = 1; 181} 182 183multiclass DS_1A1D_RET_mc <string opName, RegisterClass rc = VGPR_32, 184 string NoRetOp = ""> { 185 def "" : DS_1A1D_RET<opName, rc>, 186 AtomicNoRet<NoRetOp, !ne(NoRetOp, "")>; 187 188 let has_m0_read = 0 in { 189 def _gfx9 : DS_1A1D_RET<opName, rc>, 190 AtomicNoRet<!if(!eq(NoRetOp, ""), "", NoRetOp#"_gfx9"), 191 !ne(NoRetOp, "")>; 192 } 193} 194 195multiclass DS_1A1D_RET_mc_gfx9 <string opName, RegisterClass rc = VGPR_32, 196 string NoRetOp = ""> { 197 let has_m0_read = 0 in { 198 def "" : DS_1A1D_RET<opName, rc>, 199 AtomicNoRet<!if(!eq(NoRetOp, ""), "", NoRetOp), 200 !if(!eq(NoRetOp, ""), 0, 1)>; 201 } 202} 203 204class DS_1A2D_RET<string opName, 205 RegisterClass rc = VGPR_32, 206 RegisterClass src = rc, 207 RegisterOperand dst_op = getLdStRegisterOperand<rc>.ret, 208 RegisterOperand src_op = getLdStRegisterOperand<src>.ret> 209: DS_Pseudo<opName, 210 (outs dst_op:$vdst), 211 (ins VGPR_32:$addr, src_op:$data0, src_op:$data1, offset:$offset, gds:$gds), 212 " $vdst, $addr, $data0, $data1$offset$gds"> { 213 214 let hasPostISelHook = 1; 215 let IsAtomicRet = 1; 216} 217 218multiclass DS_1A2D_RET_mc<string opName, 219 RegisterClass rc = VGPR_32, 220 string NoRetOp = "", 221 RegisterClass src = rc> { 222 def "" : DS_1A2D_RET<opName, rc, src>, 223 AtomicNoRet<NoRetOp, !ne(NoRetOp, "")>; 224 225 let has_m0_read = 0 in { 226 def _gfx9 : DS_1A2D_RET<opName, rc, src>, 227 AtomicNoRet<NoRetOp#"_gfx9", !ne(NoRetOp, "")>; 228 } 229} 230 231class DS_1A2D_Off8_RET<string opName, 232 RegisterClass rc = VGPR_32, 233 RegisterClass src = rc, 234 RegisterOperand dst_op = getLdStRegisterOperand<rc>.ret, 235 RegisterOperand src_op = getLdStRegisterOperand<src>.ret> 236: DS_Pseudo<opName, 237 (outs dst_op:$vdst), 238 (ins VGPR_32:$addr, src_op:$data0, src_op:$data1, offset0:$offset0, offset1:$offset1, gds:$gds), 239 " $vdst, $addr, $data0, $data1$offset0$offset1$gds"> { 240 241 let has_offset = 0; 242 let AsmMatchConverter = "cvtDSOffset01"; 243 244 let hasPostISelHook = 1; 245} 246 247multiclass DS_1A2D_Off8_RET_mc<string opName, 248 RegisterClass rc = VGPR_32, 249 RegisterClass src = rc> { 250 def "" : DS_1A2D_Off8_RET<opName, rc, src>; 251 252 let has_m0_read = 0 in { 253 def _gfx9 : DS_1A2D_Off8_RET<opName, rc, src>; 254 } 255} 256 257 258class DS_1A_RET<string opName, RegisterClass rc = VGPR_32, bit HasTiedOutput = 0, Operand ofs = offset, 259 RegisterOperand data_op = getLdStRegisterOperand<rc>.ret> 260: DS_Pseudo<opName, 261 (outs data_op:$vdst), 262 !if(HasTiedOutput, 263 (ins VGPR_32:$addr, ofs:$offset, gds:$gds, data_op:$vdst_in), 264 (ins VGPR_32:$addr, ofs:$offset, gds:$gds)), 265 " $vdst, $addr$offset$gds"> { 266 let Constraints = !if(HasTiedOutput, "$vdst = $vdst_in", ""); 267 let DisableEncoding = !if(HasTiedOutput, "$vdst_in", ""); 268 let has_data0 = 0; 269 let has_data1 = 0; 270} 271 272multiclass DS_1A_RET_mc<string opName, RegisterClass rc = VGPR_32, bit HasTiedOutput = 0, Operand ofs = offset> { 273 def "" : DS_1A_RET<opName, rc, HasTiedOutput, ofs>; 274 275 let has_m0_read = 0 in { 276 def _gfx9 : DS_1A_RET<opName, rc, HasTiedOutput, ofs>; 277 } 278} 279 280class DS_1A_RET_Tied<string opName, RegisterClass rc = VGPR_32> : 281 DS_1A_RET<opName, rc, 1>; 282 283class DS_1A_Off8_RET <string opName, RegisterClass rc = VGPR_32> 284: DS_Pseudo<opName, 285 (outs getLdStRegisterOperand<rc>.ret:$vdst), 286 (ins VGPR_32:$addr, offset0:$offset0, offset1:$offset1, gds:$gds), 287 " $vdst, $addr$offset0$offset1$gds"> { 288 289 let has_offset = 0; 290 let has_data0 = 0; 291 let has_data1 = 0; 292 let AsmMatchConverter = "cvtDSOffset01"; 293} 294 295multiclass DS_1A_Off8_RET_mc <string opName, RegisterClass rc = VGPR_32> { 296 def "" : DS_1A_Off8_RET<opName, rc>; 297 298 let has_m0_read = 0 in { 299 def _gfx9 : DS_1A_Off8_RET<opName, rc>; 300 } 301} 302 303class DS_1A_RET_GDS <string opName> : DS_Pseudo<opName, 304 (outs getLdStRegisterOperand<VGPR_32>.ret:$vdst), 305 (ins VGPR_32:$addr, offset:$offset), 306 " $vdst, $addr$offset gds"> { 307 308 let has_data0 = 0; 309 let has_data1 = 0; 310 let has_gds = 0; 311 let gdsValue = 1; 312 let AsmMatchConverter = "cvtDSGds"; 313} 314 315class DS_0A_RET <string opName> : DS_Pseudo<opName, 316 (outs getLdStRegisterOperand<VGPR_32>.ret:$vdst), 317 (ins offset:$offset, gds:$gds), 318 " $vdst$offset$gds"> { 319 320 let mayLoad = 1; 321 let mayStore = 1; 322 323 let has_addr = 0; 324 let has_data0 = 0; 325 let has_data1 = 0; 326} 327 328class DS_1A <string opName> : DS_Pseudo<opName, 329 (outs), 330 (ins VGPR_32:$addr, offset:$offset, gds:$gds), 331 " $addr$offset$gds"> { 332 333 let mayLoad = 1; 334 let mayStore = 1; 335 336 let has_vdst = 0; 337 let has_data0 = 0; 338 let has_data1 = 0; 339} 340 341multiclass DS_1A_mc <string opName> { 342 def "" : DS_1A<opName>; 343 344 let has_m0_read = 0 in { 345 def _gfx9 : DS_1A<opName>; 346 } 347} 348 349 350class DS_GWS <string opName, dag ins, string asmOps> 351: DS_Pseudo<opName, (outs), ins, asmOps> { 352 353 let has_vdst = 0; 354 let has_addr = 0; 355 let has_data0 = 0; 356 let has_data1 = 0; 357 358 let has_gds = 0; 359 let gdsValue = 1; 360 let AsmMatchConverter = "cvtDSGds"; 361} 362 363class DS_GWS_0D <string opName> 364: DS_GWS<opName, 365 (ins offset:$offset), "$offset gds"> { 366 let hasSideEffects = 1; 367} 368 369class DS_GWS_1D <string opName> 370: DS_GWS<opName, 371 (ins getLdStRegisterOperand<VGPR_32>.ret:$data0, offset:$offset), 372 " $data0$offset gds"> { 373 374 let has_gws_data0 = 1; 375 let hasSideEffects = 1; 376} 377 378class DS_VOID <string opName> : DS_Pseudo<opName, 379 (outs), (ins), ""> { 380 let mayLoad = 0; 381 let mayStore = 0; 382 let hasSideEffects = 1; 383 let UseNamedOperandTable = 0; 384 let AsmMatchConverter = ""; 385 386 let has_vdst = 0; 387 let has_addr = 0; 388 let has_data0 = 0; 389 let has_data1 = 0; 390 let has_offset = 0; 391 let has_offset0 = 0; 392 let has_offset1 = 0; 393 let has_gds = 0; 394} 395 396class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag, 397 RegisterOperand data_op = getLdStRegisterOperand<VGPR_32>.ret> 398: DS_Pseudo<opName, 399 (outs data_op:$vdst), 400 (ins VGPR_32:$addr, data_op:$data0, offset:$offset), 401 " $vdst, $addr, $data0$offset", 402 [(set i32:$vdst, 403 (node (DS1Addr1Offset i32:$addr, i16:$offset), i32:$data0))] > { 404 405 let mayLoad = 0; 406 let mayStore = 0; 407 let isConvergent = 1; 408 409 let has_data1 = 0; 410 let has_gds = 0; 411} 412 413defm DS_ADD_U32 : DS_1A1D_NORET_mc<"ds_add_u32">; 414defm DS_SUB_U32 : DS_1A1D_NORET_mc<"ds_sub_u32">; 415defm DS_RSUB_U32 : DS_1A1D_NORET_mc<"ds_rsub_u32">; 416defm DS_INC_U32 : DS_1A1D_NORET_mc<"ds_inc_u32">; 417defm DS_DEC_U32 : DS_1A1D_NORET_mc<"ds_dec_u32">; 418defm DS_MIN_I32 : DS_1A1D_NORET_mc<"ds_min_i32">; 419defm DS_MAX_I32 : DS_1A1D_NORET_mc<"ds_max_i32">; 420defm DS_MIN_U32 : DS_1A1D_NORET_mc<"ds_min_u32">; 421defm DS_MAX_U32 : DS_1A1D_NORET_mc<"ds_max_u32">; 422defm DS_AND_B32 : DS_1A1D_NORET_mc<"ds_and_b32">; 423defm DS_OR_B32 : DS_1A1D_NORET_mc<"ds_or_b32">; 424defm DS_XOR_B32 : DS_1A1D_NORET_mc<"ds_xor_b32">; 425 426let SubtargetPredicate = HasLDSFPAtomics in { 427defm DS_ADD_F32 : DS_1A1D_NORET_mc<"ds_add_f32">; 428} 429 430// FIXME: Are these really present pre-gfx8? 431defm DS_MIN_F32 : DS_1A1D_NORET_mc<"ds_min_f32">; 432defm DS_MAX_F32 : DS_1A1D_NORET_mc<"ds_max_f32">; 433 434let mayLoad = 0 in { 435defm DS_WRITE_B8 : DS_1A1D_NORET_mc<"ds_write_b8">; 436defm DS_WRITE_B16 : DS_1A1D_NORET_mc<"ds_write_b16">; 437defm DS_WRITE_B32 : DS_1A1D_NORET_mc<"ds_write_b32">; 438defm DS_WRITE2_B32 : DS_1A2D_Off8_NORET_mc<"ds_write2_b32">; 439defm DS_WRITE2ST64_B32: DS_1A2D_Off8_NORET_mc<"ds_write2st64_b32">; 440 441 442let has_m0_read = 0 in { 443 444let SubtargetPredicate = HasD16LoadStore in { 445def DS_WRITE_B8_D16_HI : DS_1A1D_NORET<"ds_write_b8_d16_hi">; 446def DS_WRITE_B16_D16_HI : DS_1A1D_NORET<"ds_write_b16_d16_hi">; 447} 448 449} // End has_m0_read = 0 450 451let SubtargetPredicate = HasDSAddTid in { 452def DS_WRITE_ADDTID_B32 : DS_0A1D_NORET<"ds_write_addtid_b32">; 453} 454 455} // End mayLoad = 0 456 457let SubtargetPredicate = isGFX90APlus in { 458 defm DS_ADD_F64 : DS_1A1D_NORET_mc_gfx9<"ds_add_f64", VReg_64>; 459 defm DS_ADD_RTN_F64 : DS_1A1D_RET_mc_gfx9<"ds_add_rtn_f64 ", VReg_64, "ds_add_f64">; 460} // End SubtargetPredicate = isGFX90APlus 461 462defm DS_MSKOR_B32 : DS_1A2D_NORET_mc<"ds_mskor_b32">; 463defm DS_CMPST_B32 : DS_1A2D_NORET_mc<"ds_cmpst_b32">; 464defm DS_CMPST_F32 : DS_1A2D_NORET_mc<"ds_cmpst_f32">; 465 466defm DS_ADD_U64 : DS_1A1D_NORET_mc<"ds_add_u64", VReg_64>; 467defm DS_SUB_U64 : DS_1A1D_NORET_mc<"ds_sub_u64", VReg_64>; 468defm DS_RSUB_U64 : DS_1A1D_NORET_mc<"ds_rsub_u64", VReg_64>; 469defm DS_INC_U64 : DS_1A1D_NORET_mc<"ds_inc_u64", VReg_64>; 470defm DS_DEC_U64 : DS_1A1D_NORET_mc<"ds_dec_u64", VReg_64>; 471defm DS_MIN_I64 : DS_1A1D_NORET_mc<"ds_min_i64", VReg_64>; 472defm DS_MAX_I64 : DS_1A1D_NORET_mc<"ds_max_i64", VReg_64>; 473defm DS_MIN_U64 : DS_1A1D_NORET_mc<"ds_min_u64", VReg_64>; 474defm DS_MAX_U64 : DS_1A1D_NORET_mc<"ds_max_u64", VReg_64>; 475defm DS_AND_B64 : DS_1A1D_NORET_mc<"ds_and_b64", VReg_64>; 476defm DS_OR_B64 : DS_1A1D_NORET_mc<"ds_or_b64", VReg_64>; 477defm DS_XOR_B64 : DS_1A1D_NORET_mc<"ds_xor_b64", VReg_64>; 478defm DS_MSKOR_B64 : DS_1A2D_NORET_mc<"ds_mskor_b64", VReg_64>; 479let mayLoad = 0 in { 480defm DS_WRITE_B64 : DS_1A1D_NORET_mc<"ds_write_b64", VReg_64>; 481defm DS_WRITE2_B64 : DS_1A2D_Off8_NORET_mc<"ds_write2_b64", VReg_64>; 482defm DS_WRITE2ST64_B64: DS_1A2D_Off8_NORET_mc<"ds_write2st64_b64", VReg_64>; 483} 484defm DS_CMPST_B64 : DS_1A2D_NORET_mc<"ds_cmpst_b64", VReg_64>; 485defm DS_CMPST_F64 : DS_1A2D_NORET_mc<"ds_cmpst_f64", VReg_64>; 486defm DS_MIN_F64 : DS_1A1D_NORET_mc<"ds_min_f64", VReg_64>; 487defm DS_MAX_F64 : DS_1A1D_NORET_mc<"ds_max_f64", VReg_64>; 488 489defm DS_ADD_RTN_U32 : DS_1A1D_RET_mc<"ds_add_rtn_u32", VGPR_32, "ds_add_u32">; 490 491let SubtargetPredicate = HasLDSFPAtomics in { 492defm DS_ADD_RTN_F32 : DS_1A1D_RET_mc<"ds_add_rtn_f32", VGPR_32, "ds_add_f32">; 493} 494defm DS_SUB_RTN_U32 : DS_1A1D_RET_mc<"ds_sub_rtn_u32", VGPR_32, "ds_sub_u32">; 495defm DS_RSUB_RTN_U32 : DS_1A1D_RET_mc<"ds_rsub_rtn_u32", VGPR_32, "ds_rsub_u32">; 496defm DS_INC_RTN_U32 : DS_1A1D_RET_mc<"ds_inc_rtn_u32", VGPR_32, "ds_inc_u32">; 497defm DS_DEC_RTN_U32 : DS_1A1D_RET_mc<"ds_dec_rtn_u32", VGPR_32, "ds_dec_u32">; 498defm DS_MIN_RTN_I32 : DS_1A1D_RET_mc<"ds_min_rtn_i32", VGPR_32, "ds_min_i32">; 499defm DS_MAX_RTN_I32 : DS_1A1D_RET_mc<"ds_max_rtn_i32", VGPR_32, "ds_max_i32">; 500defm DS_MIN_RTN_U32 : DS_1A1D_RET_mc<"ds_min_rtn_u32", VGPR_32, "ds_min_u32">; 501defm DS_MAX_RTN_U32 : DS_1A1D_RET_mc<"ds_max_rtn_u32", VGPR_32, "ds_max_u32">; 502defm DS_AND_RTN_B32 : DS_1A1D_RET_mc<"ds_and_rtn_b32", VGPR_32, "ds_and_b32">; 503defm DS_OR_RTN_B32 : DS_1A1D_RET_mc<"ds_or_rtn_b32", VGPR_32, "ds_or_b32">; 504defm DS_XOR_RTN_B32 : DS_1A1D_RET_mc<"ds_xor_rtn_b32", VGPR_32, "ds_xor_b32">; 505defm DS_MSKOR_RTN_B32 : DS_1A2D_RET_mc<"ds_mskor_rtn_b32", VGPR_32, "ds_mskor_b32">; 506defm DS_CMPST_RTN_B32 : DS_1A2D_RET_mc<"ds_cmpst_rtn_b32", VGPR_32, "ds_cmpst_b32">; 507defm DS_CMPST_RTN_F32 : DS_1A2D_RET_mc<"ds_cmpst_rtn_f32", VGPR_32, "ds_cmpst_f32">; 508defm DS_MIN_RTN_F32 : DS_1A1D_RET_mc<"ds_min_rtn_f32", VGPR_32, "ds_min_f32">; 509defm DS_MAX_RTN_F32 : DS_1A1D_RET_mc<"ds_max_rtn_f32", VGPR_32, "ds_max_f32">; 510 511defm DS_WRXCHG_RTN_B32 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b32">; 512defm DS_WRXCHG2_RTN_B32 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b32", VReg_64, VGPR_32>; 513defm DS_WRXCHG2ST64_RTN_B32 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b32", VReg_64, VGPR_32>; 514 515defm DS_ADD_RTN_U64 : DS_1A1D_RET_mc<"ds_add_rtn_u64", VReg_64, "ds_add_u64">; 516defm DS_SUB_RTN_U64 : DS_1A1D_RET_mc<"ds_sub_rtn_u64", VReg_64, "ds_sub_u64">; 517defm DS_RSUB_RTN_U64 : DS_1A1D_RET_mc<"ds_rsub_rtn_u64", VReg_64, "ds_rsub_u64">; 518defm DS_INC_RTN_U64 : DS_1A1D_RET_mc<"ds_inc_rtn_u64", VReg_64, "ds_inc_u64">; 519defm DS_DEC_RTN_U64 : DS_1A1D_RET_mc<"ds_dec_rtn_u64", VReg_64, "ds_dec_u64">; 520defm DS_MIN_RTN_I64 : DS_1A1D_RET_mc<"ds_min_rtn_i64", VReg_64, "ds_min_i64">; 521defm DS_MAX_RTN_I64 : DS_1A1D_RET_mc<"ds_max_rtn_i64", VReg_64, "ds_max_i64">; 522defm DS_MIN_RTN_U64 : DS_1A1D_RET_mc<"ds_min_rtn_u64", VReg_64, "ds_min_u64">; 523defm DS_MAX_RTN_U64 : DS_1A1D_RET_mc<"ds_max_rtn_u64", VReg_64, "ds_max_u64">; 524defm DS_AND_RTN_B64 : DS_1A1D_RET_mc<"ds_and_rtn_b64", VReg_64, "ds_and_b64">; 525defm DS_OR_RTN_B64 : DS_1A1D_RET_mc<"ds_or_rtn_b64", VReg_64, "ds_or_b64">; 526defm DS_XOR_RTN_B64 : DS_1A1D_RET_mc<"ds_xor_rtn_b64", VReg_64, "ds_xor_b64">; 527defm DS_MSKOR_RTN_B64 : DS_1A2D_RET_mc<"ds_mskor_rtn_b64", VReg_64, "ds_mskor_b64">; 528defm DS_CMPST_RTN_B64 : DS_1A2D_RET_mc<"ds_cmpst_rtn_b64", VReg_64, "ds_cmpst_b64">; 529defm DS_CMPST_RTN_F64 : DS_1A2D_RET_mc<"ds_cmpst_rtn_f64", VReg_64, "ds_cmpst_f64">; 530defm DS_MIN_RTN_F64 : DS_1A1D_RET_mc<"ds_min_rtn_f64", VReg_64, "ds_min_f64">; 531defm DS_MAX_RTN_F64 : DS_1A1D_RET_mc<"ds_max_rtn_f64", VReg_64, "ds_max_f64">; 532 533defm DS_WRXCHG_RTN_B64 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b64", VReg_64>; 534defm DS_WRXCHG2_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b64", VReg_128, VReg_64>; 535defm DS_WRXCHG2ST64_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b64", VReg_128, VReg_64>; 536 537let isConvergent = 1, usesCustomInserter = 1 in { 538def DS_GWS_INIT : DS_GWS_1D<"ds_gws_init"> { 539 let mayLoad = 0; 540} 541def DS_GWS_SEMA_V : DS_GWS_0D<"ds_gws_sema_v">; 542def DS_GWS_SEMA_BR : DS_GWS_1D<"ds_gws_sema_br">; 543def DS_GWS_SEMA_P : DS_GWS_0D<"ds_gws_sema_p">; 544def DS_GWS_BARRIER : DS_GWS_1D<"ds_gws_barrier">; 545} 546 547let SubtargetPredicate = HasDsSrc2Insts in { 548def DS_ADD_SRC2_U32 : DS_1A<"ds_add_src2_u32">; 549def DS_SUB_SRC2_U32 : DS_1A<"ds_sub_src2_u32">; 550def DS_RSUB_SRC2_U32 : DS_1A<"ds_rsub_src2_u32">; 551def DS_INC_SRC2_U32 : DS_1A<"ds_inc_src2_u32">; 552def DS_DEC_SRC2_U32 : DS_1A<"ds_dec_src2_u32">; 553def DS_MIN_SRC2_I32 : DS_1A<"ds_min_src2_i32">; 554def DS_MAX_SRC2_I32 : DS_1A<"ds_max_src2_i32">; 555def DS_MIN_SRC2_U32 : DS_1A<"ds_min_src2_u32">; 556def DS_MAX_SRC2_U32 : DS_1A<"ds_max_src2_u32">; 557def DS_AND_SRC2_B32 : DS_1A<"ds_and_src2_b32">; 558def DS_OR_SRC2_B32 : DS_1A<"ds_or_src2_b32">; 559def DS_XOR_SRC2_B32 : DS_1A<"ds_xor_src2_b32">; 560def DS_MIN_SRC2_F32 : DS_1A<"ds_min_src2_f32">; 561def DS_MAX_SRC2_F32 : DS_1A<"ds_max_src2_f32">; 562 563def DS_ADD_SRC2_U64 : DS_1A<"ds_add_src2_u64">; 564def DS_SUB_SRC2_U64 : DS_1A<"ds_sub_src2_u64">; 565def DS_RSUB_SRC2_U64 : DS_1A<"ds_rsub_src2_u64">; 566def DS_INC_SRC2_U64 : DS_1A<"ds_inc_src2_u64">; 567def DS_DEC_SRC2_U64 : DS_1A<"ds_dec_src2_u64">; 568def DS_MIN_SRC2_I64 : DS_1A<"ds_min_src2_i64">; 569def DS_MAX_SRC2_I64 : DS_1A<"ds_max_src2_i64">; 570def DS_MIN_SRC2_U64 : DS_1A<"ds_min_src2_u64">; 571def DS_MAX_SRC2_U64 : DS_1A<"ds_max_src2_u64">; 572def DS_AND_SRC2_B64 : DS_1A<"ds_and_src2_b64">; 573def DS_OR_SRC2_B64 : DS_1A<"ds_or_src2_b64">; 574def DS_XOR_SRC2_B64 : DS_1A<"ds_xor_src2_b64">; 575def DS_MIN_SRC2_F64 : DS_1A<"ds_min_src2_f64">; 576def DS_MAX_SRC2_F64 : DS_1A<"ds_max_src2_f64">; 577 578def DS_WRITE_SRC2_B32 : DS_1A<"ds_write_src2_b32">; 579def DS_WRITE_SRC2_B64 : DS_1A<"ds_write_src2_b64">; 580} // End SubtargetPredicate = HasDsSrc2Insts 581 582let Uses = [EXEC], mayLoad = 0, mayStore = 0, isConvergent = 1 in { 583def DS_SWIZZLE_B32 : DS_1A_RET <"ds_swizzle_b32", VGPR_32, 0, SwizzleImm>; 584} 585 586let mayStore = 0 in { 587defm DS_READ_I8 : DS_1A_RET_mc<"ds_read_i8">; 588defm DS_READ_U8 : DS_1A_RET_mc<"ds_read_u8">; 589defm DS_READ_I16 : DS_1A_RET_mc<"ds_read_i16">; 590defm DS_READ_U16 : DS_1A_RET_mc<"ds_read_u16">; 591defm DS_READ_B32 : DS_1A_RET_mc<"ds_read_b32">; 592defm DS_READ_B64 : DS_1A_RET_mc<"ds_read_b64", VReg_64>; 593 594defm DS_READ2_B32 : DS_1A_Off8_RET_mc<"ds_read2_b32", VReg_64>; 595defm DS_READ2ST64_B32: DS_1A_Off8_RET_mc<"ds_read2st64_b32", VReg_64>; 596 597defm DS_READ2_B64 : DS_1A_Off8_RET_mc<"ds_read2_b64", VReg_128>; 598defm DS_READ2ST64_B64: DS_1A_Off8_RET_mc<"ds_read2st64_b64", VReg_128>; 599 600let has_m0_read = 0 in { 601let SubtargetPredicate = HasD16LoadStore in { 602def DS_READ_U8_D16 : DS_1A_RET_Tied<"ds_read_u8_d16">; 603def DS_READ_U8_D16_HI : DS_1A_RET_Tied<"ds_read_u8_d16_hi">; 604def DS_READ_I8_D16 : DS_1A_RET_Tied<"ds_read_i8_d16">; 605def DS_READ_I8_D16_HI : DS_1A_RET_Tied<"ds_read_i8_d16_hi">; 606def DS_READ_U16_D16 : DS_1A_RET_Tied<"ds_read_u16_d16">; 607def DS_READ_U16_D16_HI : DS_1A_RET_Tied<"ds_read_u16_d16_hi">; 608} 609} // End has_m0_read = 0 610 611let SubtargetPredicate = HasDSAddTid in { 612def DS_READ_ADDTID_B32 : DS_0A_RET<"ds_read_addtid_b32">; 613} 614 615} // End mayStore = 0 616 617def DS_CONSUME : DS_0A_RET<"ds_consume">; 618def DS_APPEND : DS_0A_RET<"ds_append">; 619def DS_ORDERED_COUNT : DS_1A_RET_GDS<"ds_ordered_count">; 620 621//===----------------------------------------------------------------------===// 622// Instruction definitions for CI and newer. 623//===----------------------------------------------------------------------===// 624 625let SubtargetPredicate = isGFX7Plus in { 626 627defm DS_WRAP_RTN_B32 : DS_1A2D_RET_mc<"ds_wrap_rtn_b32", VGPR_32>; 628defm DS_CONDXCHG32_RTN_B64 : DS_1A1D_RET_mc<"ds_condxchg32_rtn_b64", VReg_64>; 629 630let isConvergent = 1, usesCustomInserter = 1 in { 631def DS_GWS_SEMA_RELEASE_ALL : DS_GWS_0D<"ds_gws_sema_release_all">; 632} 633 634let mayStore = 0 in { 635defm DS_READ_B96 : DS_1A_RET_mc<"ds_read_b96", VReg_96>; 636defm DS_READ_B128: DS_1A_RET_mc<"ds_read_b128", VReg_128>; 637} // End mayStore = 0 638 639let mayLoad = 0 in { 640defm DS_WRITE_B96 : DS_1A1D_NORET_mc<"ds_write_b96", VReg_96>; 641defm DS_WRITE_B128 : DS_1A1D_NORET_mc<"ds_write_b128", VReg_128>; 642} // End mayLoad = 0 643 644def DS_NOP : DS_VOID<"ds_nop">; 645 646} // let SubtargetPredicate = isGFX7Plus 647 648//===----------------------------------------------------------------------===// 649// Instruction definitions for VI and newer. 650//===----------------------------------------------------------------------===// 651 652let SubtargetPredicate = isGFX8Plus in { 653 654let Uses = [EXEC] in { 655def DS_PERMUTE_B32 : DS_1A1D_PERMUTE <"ds_permute_b32", 656 int_amdgcn_ds_permute>; 657def DS_BPERMUTE_B32 : DS_1A1D_PERMUTE <"ds_bpermute_b32", 658 int_amdgcn_ds_bpermute>; 659} 660 661} // let SubtargetPredicate = isGFX8Plus 662 663let SubtargetPredicate = HasLDSFPAtomics, OtherPredicates = [HasDsSrc2Insts] in { 664def DS_ADD_SRC2_F32 : DS_1A<"ds_add_src2_f32">; 665} 666 667//===----------------------------------------------------------------------===// 668// DS Patterns 669//===----------------------------------------------------------------------===// 670 671def : GCNPat < 672 (int_amdgcn_ds_swizzle i32:$src, timm:$offset16), 673 (DS_SWIZZLE_B32 VGPR_32:$src, (as_i16timm $offset16), (i1 0)) 674>; 675 676class DSReadPat <DS_Pseudo inst, ValueType vt, PatFrag frag, int gds=0> : GCNPat < 677 (vt (frag (DS1Addr1Offset i32:$ptr, i16:$offset))), 678 (inst $ptr, offset:$offset, (i1 gds)) 679>; 680 681multiclass DSReadPat_mc<DS_Pseudo inst, ValueType vt, string frag> { 682 683 let OtherPredicates = [LDSRequiresM0Init] in { 684 def : DSReadPat<inst, vt, !cast<PatFrag>(frag#"_m0")>; 685 } 686 687 let OtherPredicates = [NotLDSRequiresM0Init] in { 688 def : DSReadPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>; 689 } 690} 691 692class DSReadPat_D16 <DS_Pseudo inst, PatFrag frag, ValueType vt> : GCNPat < 693 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$in), 694 (inst $ptr, offset:$offset, (i1 0), $in) 695>; 696 697defm : DSReadPat_mc <DS_READ_I8, i32, "sextloadi8_local">; 698defm : DSReadPat_mc <DS_READ_I8, i16, "sextloadi8_local">; 699defm : DSReadPat_mc <DS_READ_U8, i32, "extloadi8_local">; 700defm : DSReadPat_mc <DS_READ_U8, i32, "zextloadi8_local">; 701defm : DSReadPat_mc <DS_READ_U8, i16, "extloadi8_local">; 702defm : DSReadPat_mc <DS_READ_U8, i16, "zextloadi8_local">; 703defm : DSReadPat_mc <DS_READ_I16, i32, "sextloadi16_local">; 704defm : DSReadPat_mc <DS_READ_I16, i32, "sextloadi16_local">; 705defm : DSReadPat_mc <DS_READ_U16, i32, "extloadi16_local">; 706defm : DSReadPat_mc <DS_READ_U16, i32, "zextloadi16_local">; 707defm : DSReadPat_mc <DS_READ_U16, i16, "load_local">; 708 709foreach vt = Reg32Types.types in { 710defm : DSReadPat_mc <DS_READ_B32, vt, "load_local">; 711} 712 713defm : DSReadPat_mc <DS_READ_B32, i32, "atomic_load_32_local">; 714defm : DSReadPat_mc <DS_READ_B64, i64, "atomic_load_64_local">; 715 716// Prefer ds_read over ds_read2, all other things being equal, because it has 717// a larger immediate offset range. 718let AddedComplexity = 100 in { 719 720foreach vt = VReg_64.RegTypes in { 721defm : DSReadPat_mc <DS_READ_B64, vt, "load_align8_local">; 722} 723 724let SubtargetPredicate = isGFX7Plus in { 725 726foreach vt = VReg_96.RegTypes in { 727defm : DSReadPat_mc <DS_READ_B96, vt, "load_align16_local">; 728} 729 730// For performance reasons restrict this to alignment >= 16 even with 731// unaligned-access-mode. At lower alignments ds_read2_b64 is always a better 732// choice. 733foreach vt = VReg_128.RegTypes in { 734defm : DSReadPat_mc <DS_READ_B128, vt, "load_align16_local">; 735} 736 737let SubtargetPredicate = HasUnalignedAccessMode in { 738 739foreach vt = VReg_96.RegTypes in { 740defm : DSReadPat_mc <DS_READ_B96, vt, "load_local">; 741} 742 743} // End SubtargetPredicate = HasUnalignedAccessMode 744 745} // End SubtargetPredicate = isGFX7Plus 746 747} // End AddedComplexity = 100 748 749let OtherPredicates = [D16PreservesUnusedBits] in { 750def : DSReadPat_D16<DS_READ_U16_D16_HI, load_d16_hi_local, v2i16>; 751def : DSReadPat_D16<DS_READ_U16_D16_HI, load_d16_hi_local, v2f16>; 752def : DSReadPat_D16<DS_READ_U8_D16_HI, az_extloadi8_d16_hi_local, v2i16>; 753def : DSReadPat_D16<DS_READ_U8_D16_HI, az_extloadi8_d16_hi_local, v2f16>; 754def : DSReadPat_D16<DS_READ_I8_D16_HI, sextloadi8_d16_hi_local, v2i16>; 755def : DSReadPat_D16<DS_READ_I8_D16_HI, sextloadi8_d16_hi_local, v2f16>; 756 757def : DSReadPat_D16<DS_READ_U16_D16, load_d16_lo_local, v2i16>; 758def : DSReadPat_D16<DS_READ_U16_D16, load_d16_lo_local, v2f16>; 759def : DSReadPat_D16<DS_READ_U8_D16, az_extloadi8_d16_lo_local, v2i16>; 760def : DSReadPat_D16<DS_READ_U8_D16, az_extloadi8_d16_lo_local, v2f16>; 761def : DSReadPat_D16<DS_READ_I8_D16, sextloadi8_d16_lo_local, v2i16>; 762def : DSReadPat_D16<DS_READ_I8_D16, sextloadi8_d16_lo_local, v2f16>; 763} 764 765class DSWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag, int gds=0> : GCNPat < 766 (frag vt:$value, (DS1Addr1Offset i32:$ptr, i16:$offset)), 767 (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds)) 768>; 769 770multiclass DSWritePat_mc <DS_Pseudo inst, ValueType vt, string frag> { 771 let OtherPredicates = [LDSRequiresM0Init] in { 772 def : DSWritePat<inst, vt, !cast<PatFrag>(frag#"_m0")>; 773 } 774 775 let OtherPredicates = [NotLDSRequiresM0Init] in { 776 def : DSWritePat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>; 777 } 778} 779 780// Irritatingly, atomic_store reverses the order of operands from a 781// normal store. 782class DSAtomicWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat < 783 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$value), 784 (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 0)) 785>; 786 787multiclass DSAtomicWritePat_mc <DS_Pseudo inst, ValueType vt, string frag> { 788 let OtherPredicates = [LDSRequiresM0Init] in { 789 def : DSAtomicWritePat<inst, vt, !cast<PatFrag>(frag#"_m0")>; 790 } 791 792 let OtherPredicates = [NotLDSRequiresM0Init] in { 793 def : DSAtomicWritePat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, !cast<PatFrag>(frag)>; 794 } 795} 796 797defm : DSWritePat_mc <DS_WRITE_B8, i32, "truncstorei8_local">; 798defm : DSWritePat_mc <DS_WRITE_B16, i32, "truncstorei16_local">; 799defm : DSWritePat_mc <DS_WRITE_B8, i16, "truncstorei8_local">; 800defm : DSWritePat_mc <DS_WRITE_B16, i16, "store_local">; 801 802foreach vt = Reg32Types.types in { 803defm : DSWritePat_mc <DS_WRITE_B32, vt, "store_local">; 804} 805 806defm : DSAtomicWritePat_mc <DS_WRITE_B32, i32, "atomic_store_local_32">; 807defm : DSAtomicWritePat_mc <DS_WRITE_B64, i64, "atomic_store_local_64">; 808 809let OtherPredicates = [D16PreservesUnusedBits] in { 810def : DSWritePat <DS_WRITE_B16_D16_HI, i32, store_hi16_local>; 811def : DSWritePat <DS_WRITE_B8_D16_HI, i32, truncstorei8_hi16_local>; 812} 813 814class DS64Bit4ByteAlignedReadPat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat < 815 (vt:$value (frag (DS64Bit4ByteAligned i32:$ptr, i8:$offset0, i8:$offset1))), 816 (inst $ptr, $offset0, $offset1, (i1 0)) 817>; 818 819class DS64Bit4ByteAlignedWritePat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat< 820 (frag vt:$value, (DS64Bit4ByteAligned i32:$ptr, i8:$offset0, i8:$offset1)), 821 (inst $ptr, (i32 (EXTRACT_SUBREG VReg_64:$value, sub0)), 822 (i32 (EXTRACT_SUBREG VReg_64:$value, sub1)), $offset0, $offset1, 823 (i1 0)) 824>; 825 826class DS128Bit8ByteAlignedReadPat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat < 827 (vt:$value (frag (DS128Bit8ByteAligned i32:$ptr, i8:$offset0, i8:$offset1))), 828 (inst $ptr, $offset0, $offset1, (i1 0)) 829>; 830 831class DS128Bit8ByteAlignedWritePat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat< 832 (frag vt:$value, (DS128Bit8ByteAligned i32:$ptr, i8:$offset0, i8:$offset1)), 833 (inst $ptr, (i64 (EXTRACT_SUBREG VReg_128:$value, sub0_sub1)), 834 (i64 (EXTRACT_SUBREG VReg_128:$value, sub2_sub3)), $offset0, $offset1, 835 (i1 0)) 836>; 837 838multiclass DS64Bit4ByteAlignedPat_mc<ValueType vt> { 839 let OtherPredicates = [LDSRequiresM0Init, isGFX7Plus] in { 840 def : DS64Bit4ByteAlignedReadPat<DS_READ2_B32, vt, load_local_m0>; 841 def : DS64Bit4ByteAlignedWritePat<DS_WRITE2_B32, vt, store_local_m0>; 842 } 843 844 let OtherPredicates = [NotLDSRequiresM0Init] in { 845 def : DS64Bit4ByteAlignedReadPat<DS_READ2_B32_gfx9, vt, load_local>; 846 def : DS64Bit4ByteAlignedWritePat<DS_WRITE2_B32_gfx9, vt, store_local>; 847 } 848} 849 850multiclass DS128Bit8ByteAlignedPat_mc<ValueType vt> { 851 let OtherPredicates = [LDSRequiresM0Init, isGFX7Plus] in { 852 def : DS128Bit8ByteAlignedReadPat<DS_READ2_B64, vt, load_local_m0>; 853 def : DS128Bit8ByteAlignedWritePat<DS_WRITE2_B64, vt, store_local_m0>; 854 } 855 856 let OtherPredicates = [NotLDSRequiresM0Init] in { 857 def : DS128Bit8ByteAlignedReadPat<DS_READ2_B64_gfx9, vt, load_local>; 858 def : DS128Bit8ByteAlignedWritePat<DS_WRITE2_B64_gfx9, vt, store_local>; 859 } 860} 861 862// v2i32 loads are split into i32 loads on SI during lowering, due to a bug 863// related to bounds checking. 864foreach vt = VReg_64.RegTypes in { 865defm : DS64Bit4ByteAlignedPat_mc<vt>; 866} 867 868foreach vt = VReg_128.RegTypes in { 869defm : DS128Bit8ByteAlignedPat_mc<vt>; 870} 871 872// Prefer ds_write over ds_write2, all other things being equal, because it has 873// a larger immediate offset range. 874let AddedComplexity = 100 in { 875 876foreach vt = VReg_64.RegTypes in { 877defm : DSWritePat_mc <DS_WRITE_B64, vt, "store_align8_local">; 878} 879 880let SubtargetPredicate = isGFX7Plus in { 881 882foreach vt = VReg_96.RegTypes in { 883defm : DSWritePat_mc <DS_WRITE_B96, vt, "store_align16_local">; 884} 885 886// For performance reasons restrict this to alignment >= 16 even with 887// unaligned-access-mode. At lower alignments ds_write2_b64 is always a better 888// choice. 889foreach vt = VReg_128.RegTypes in { 890defm : DSWritePat_mc <DS_WRITE_B128, vt, "store_align16_local">; 891} 892 893let SubtargetPredicate = HasUnalignedAccessMode in { 894 895foreach vt = VReg_96.RegTypes in { 896defm : DSWritePat_mc <DS_WRITE_B96, vt, "store_local">; 897} 898 899} // End SubtargetPredicate = HasUnalignedAccessMode 900 901} // End SubtargetPredicate = isGFX7Plus 902 903} // End AddedComplexity = 100 904 905class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat < 906 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$value), 907 (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds)) 908>; 909 910multiclass DSAtomicRetPat_mc<DS_Pseudo inst, ValueType vt, string frag> { 911 let OtherPredicates = [LDSRequiresM0Init] in { 912 def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_local_m0_"#vt.Size)>; 913 } 914 915 let OtherPredicates = [NotLDSRequiresM0Init] in { 916 def : DSAtomicRetPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, 917 !cast<PatFrag>(frag#"_local_"#vt.Size)>; 918 } 919 920 def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_region_m0_"#vt.Size), 1>; 921} 922 923 924 925class DSAtomicCmpXChg<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat < 926 (frag (DS1Addr1Offset i32:$ptr, i16:$offset), vt:$cmp, vt:$swap), 927 (inst $ptr, getVregSrcForVT<vt>.ret:$cmp, getVregSrcForVT<vt>.ret:$swap, offset:$offset, (i1 gds)) 928>; 929 930multiclass DSAtomicCmpXChg_mc<DS_Pseudo inst, ValueType vt, string frag> { 931 let OtherPredicates = [LDSRequiresM0Init] in { 932 def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_local_m0_"#vt.Size)>; 933 } 934 935 let OtherPredicates = [NotLDSRequiresM0Init] in { 936 def : DSAtomicCmpXChg<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt, 937 !cast<PatFrag>(frag#"_local_"#vt.Size)>; 938 } 939 940 def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_region_m0_"#vt.Size), 1>; 941} 942 943 944 945// 32-bit atomics. 946defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B32, i32, "atomic_swap">; 947defm : DSAtomicRetPat_mc<DS_ADD_RTN_U32, i32, "atomic_load_add">; 948defm : DSAtomicRetPat_mc<DS_SUB_RTN_U32, i32, "atomic_load_sub">; 949defm : DSAtomicRetPat_mc<DS_INC_RTN_U32, i32, "atomic_inc">; 950defm : DSAtomicRetPat_mc<DS_DEC_RTN_U32, i32, "atomic_dec">; 951defm : DSAtomicRetPat_mc<DS_AND_RTN_B32, i32, "atomic_load_and">; 952defm : DSAtomicRetPat_mc<DS_OR_RTN_B32, i32, "atomic_load_or">; 953defm : DSAtomicRetPat_mc<DS_XOR_RTN_B32, i32, "atomic_load_xor">; 954defm : DSAtomicRetPat_mc<DS_MIN_RTN_I32, i32, "atomic_load_min">; 955defm : DSAtomicRetPat_mc<DS_MAX_RTN_I32, i32, "atomic_load_max">; 956defm : DSAtomicRetPat_mc<DS_MIN_RTN_U32, i32, "atomic_load_umin">; 957defm : DSAtomicRetPat_mc<DS_MAX_RTN_U32, i32, "atomic_load_umax">; 958defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B32, i32, "atomic_cmp_swap">; 959 960let SubtargetPredicate = HasLDSFPAtomics in { 961defm : DSAtomicRetPat_mc<DS_MIN_RTN_F32, f32, "atomic_load_fmin">; 962defm : DSAtomicRetPat_mc<DS_MAX_RTN_F32, f32, "atomic_load_fmax">; 963defm : DSAtomicRetPat_mc<DS_ADD_RTN_F32, f32, "atomic_load_fadd">; 964} 965 966// 64-bit atomics. 967defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B64, i64, "atomic_swap">; 968defm : DSAtomicRetPat_mc<DS_ADD_RTN_U64, i64, "atomic_load_add">; 969defm : DSAtomicRetPat_mc<DS_SUB_RTN_U64, i64, "atomic_load_sub">; 970defm : DSAtomicRetPat_mc<DS_INC_RTN_U64, i64, "atomic_inc">; 971defm : DSAtomicRetPat_mc<DS_DEC_RTN_U64, i64, "atomic_dec">; 972defm : DSAtomicRetPat_mc<DS_AND_RTN_B64, i64, "atomic_load_and">; 973defm : DSAtomicRetPat_mc<DS_OR_RTN_B64, i64, "atomic_load_or">; 974defm : DSAtomicRetPat_mc<DS_XOR_RTN_B64, i64, "atomic_load_xor">; 975defm : DSAtomicRetPat_mc<DS_MIN_RTN_I64, i64, "atomic_load_min">; 976defm : DSAtomicRetPat_mc<DS_MAX_RTN_I64, i64, "atomic_load_max">; 977defm : DSAtomicRetPat_mc<DS_MIN_RTN_U64, i64, "atomic_load_umin">; 978defm : DSAtomicRetPat_mc<DS_MAX_RTN_U64, i64, "atomic_load_umax">; 979 980defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B64, i64, "atomic_cmp_swap">; 981 982let SubtargetPredicate = isGFX90APlus in { 983def : DSAtomicRetPat<DS_ADD_RTN_F64, f64, atomic_load_fadd_local_64>; 984} 985 986def : Pat < 987 (SIds_ordered_count i32:$value, i16:$offset), 988 (DS_ORDERED_COUNT $value, (as_i16imm $offset)) 989>; 990 991//===----------------------------------------------------------------------===// 992// Target-specific instruction encodings. 993//===----------------------------------------------------------------------===// 994 995//===----------------------------------------------------------------------===// 996// Base ENC_DS for GFX6, GFX7, GFX10. 997//===----------------------------------------------------------------------===// 998 999class Base_DS_Real_gfx6_gfx7_gfx10<bits<8> op, DS_Pseudo ps, int ef> : 1000 DS_Real<ps>, SIMCInstr <ps.Mnemonic, ef> { 1001 1002 let Inst{7-0} = !if(ps.has_offset0, offset0, 0); 1003 let Inst{15-8} = !if(ps.has_offset1, offset1, 0); 1004 let Inst{17} = !if(ps.has_gds, gds, ps.gdsValue); 1005 let Inst{25-18} = op; 1006 let Inst{31-26} = 0x36; 1007 let Inst{39-32} = !if(ps.has_addr, addr, !if(ps.has_gws_data0, data0{7-0}, 0)); 1008 let Inst{47-40} = !if(ps.has_data0, data0{7-0}, 0); 1009 let Inst{55-48} = !if(ps.has_data1, data1{7-0}, 0); 1010 let Inst{63-56} = !if(ps.has_vdst, vdst{7-0}, 0); 1011} 1012 1013//===----------------------------------------------------------------------===// 1014// GFX10. 1015//===----------------------------------------------------------------------===// 1016 1017let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in { 1018 multiclass DS_Real_gfx10<bits<8> op> { 1019 def _gfx10 : Base_DS_Real_gfx6_gfx7_gfx10<op, !cast<DS_Pseudo>(NAME), 1020 SIEncodingFamily.GFX10>; 1021 } 1022} // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" 1023 1024defm DS_ADD_F32 : DS_Real_gfx10<0x015>; 1025defm DS_ADD_RTN_F32 : DS_Real_gfx10<0x055>; 1026defm DS_ADD_SRC2_F32 : DS_Real_gfx10<0x095>; 1027defm DS_WRITE_B8_D16_HI : DS_Real_gfx10<0x0a0>; 1028defm DS_WRITE_B16_D16_HI : DS_Real_gfx10<0x0a1>; 1029defm DS_READ_U8_D16 : DS_Real_gfx10<0x0a2>; 1030defm DS_READ_U8_D16_HI : DS_Real_gfx10<0x0a3>; 1031defm DS_READ_I8_D16 : DS_Real_gfx10<0x0a4>; 1032defm DS_READ_I8_D16_HI : DS_Real_gfx10<0x0a5>; 1033defm DS_READ_U16_D16 : DS_Real_gfx10<0x0a6>; 1034defm DS_READ_U16_D16_HI : DS_Real_gfx10<0x0a7>; 1035defm DS_WRITE_ADDTID_B32 : DS_Real_gfx10<0x0b0>; 1036defm DS_READ_ADDTID_B32 : DS_Real_gfx10<0x0b1>; 1037defm DS_PERMUTE_B32 : DS_Real_gfx10<0x0b2>; 1038defm DS_BPERMUTE_B32 : DS_Real_gfx10<0x0b3>; 1039 1040//===----------------------------------------------------------------------===// 1041// GFX7, GFX10. 1042//===----------------------------------------------------------------------===// 1043 1044let AssemblerPredicate = isGFX7Only, DecoderNamespace = "GFX7" in { 1045 multiclass DS_Real_gfx7<bits<8> op> { 1046 def _gfx7 : Base_DS_Real_gfx6_gfx7_gfx10<op, !cast<DS_Pseudo>(NAME), 1047 SIEncodingFamily.SI>; 1048 } 1049} // End AssemblerPredicate = isGFX7Only, DecoderNamespace = "GFX7" 1050 1051multiclass DS_Real_gfx7_gfx10<bits<8> op> : 1052 DS_Real_gfx7<op>, DS_Real_gfx10<op>; 1053 1054// FIXME-GFX7: Add tests when upstreaming this part. 1055defm DS_GWS_SEMA_RELEASE_ALL : DS_Real_gfx7_gfx10<0x018>; 1056defm DS_WRAP_RTN_B32 : DS_Real_gfx7_gfx10<0x034>; 1057defm DS_CONDXCHG32_RTN_B64 : DS_Real_gfx7_gfx10<0x07e>; 1058defm DS_WRITE_B96 : DS_Real_gfx7_gfx10<0x0de>; 1059defm DS_WRITE_B128 : DS_Real_gfx7_gfx10<0x0df>; 1060defm DS_READ_B96 : DS_Real_gfx7_gfx10<0x0fe>; 1061defm DS_READ_B128 : DS_Real_gfx7_gfx10<0x0ff>; 1062 1063//===----------------------------------------------------------------------===// 1064// GFX6, GFX7, GFX10. 1065//===----------------------------------------------------------------------===// 1066 1067let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in { 1068 multiclass DS_Real_gfx6_gfx7<bits<8> op> { 1069 def _gfx6_gfx7 : Base_DS_Real_gfx6_gfx7_gfx10<op, !cast<DS_Pseudo>(NAME), 1070 SIEncodingFamily.SI>; 1071 } 1072} // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" 1073 1074multiclass DS_Real_gfx6_gfx7_gfx10<bits<8> op> : 1075 DS_Real_gfx6_gfx7<op>, DS_Real_gfx10<op>; 1076 1077defm DS_ADD_U32 : DS_Real_gfx6_gfx7_gfx10<0x000>; 1078defm DS_SUB_U32 : DS_Real_gfx6_gfx7_gfx10<0x001>; 1079defm DS_RSUB_U32 : DS_Real_gfx6_gfx7_gfx10<0x002>; 1080defm DS_INC_U32 : DS_Real_gfx6_gfx7_gfx10<0x003>; 1081defm DS_DEC_U32 : DS_Real_gfx6_gfx7_gfx10<0x004>; 1082defm DS_MIN_I32 : DS_Real_gfx6_gfx7_gfx10<0x005>; 1083defm DS_MAX_I32 : DS_Real_gfx6_gfx7_gfx10<0x006>; 1084defm DS_MIN_U32 : DS_Real_gfx6_gfx7_gfx10<0x007>; 1085defm DS_MAX_U32 : DS_Real_gfx6_gfx7_gfx10<0x008>; 1086defm DS_AND_B32 : DS_Real_gfx6_gfx7_gfx10<0x009>; 1087defm DS_OR_B32 : DS_Real_gfx6_gfx7_gfx10<0x00a>; 1088defm DS_XOR_B32 : DS_Real_gfx6_gfx7_gfx10<0x00b>; 1089defm DS_MSKOR_B32 : DS_Real_gfx6_gfx7_gfx10<0x00c>; 1090defm DS_WRITE_B32 : DS_Real_gfx6_gfx7_gfx10<0x00d>; 1091defm DS_WRITE2_B32 : DS_Real_gfx6_gfx7_gfx10<0x00e>; 1092defm DS_WRITE2ST64_B32 : DS_Real_gfx6_gfx7_gfx10<0x00f>; 1093defm DS_CMPST_B32 : DS_Real_gfx6_gfx7_gfx10<0x010>; 1094defm DS_CMPST_F32 : DS_Real_gfx6_gfx7_gfx10<0x011>; 1095defm DS_MIN_F32 : DS_Real_gfx6_gfx7_gfx10<0x012>; 1096defm DS_MAX_F32 : DS_Real_gfx6_gfx7_gfx10<0x013>; 1097defm DS_NOP : DS_Real_gfx6_gfx7_gfx10<0x014>; 1098defm DS_GWS_INIT : DS_Real_gfx6_gfx7_gfx10<0x019>; 1099defm DS_GWS_SEMA_V : DS_Real_gfx6_gfx7_gfx10<0x01a>; 1100defm DS_GWS_SEMA_BR : DS_Real_gfx6_gfx7_gfx10<0x01b>; 1101defm DS_GWS_SEMA_P : DS_Real_gfx6_gfx7_gfx10<0x01c>; 1102defm DS_GWS_BARRIER : DS_Real_gfx6_gfx7_gfx10<0x01d>; 1103defm DS_WRITE_B8 : DS_Real_gfx6_gfx7_gfx10<0x01e>; 1104defm DS_WRITE_B16 : DS_Real_gfx6_gfx7_gfx10<0x01f>; 1105defm DS_ADD_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x020>; 1106defm DS_SUB_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x021>; 1107defm DS_RSUB_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x022>; 1108defm DS_INC_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x023>; 1109defm DS_DEC_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x024>; 1110defm DS_MIN_RTN_I32 : DS_Real_gfx6_gfx7_gfx10<0x025>; 1111defm DS_MAX_RTN_I32 : DS_Real_gfx6_gfx7_gfx10<0x026>; 1112defm DS_MIN_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x027>; 1113defm DS_MAX_RTN_U32 : DS_Real_gfx6_gfx7_gfx10<0x028>; 1114defm DS_AND_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x029>; 1115defm DS_OR_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02a>; 1116defm DS_XOR_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02b>; 1117defm DS_MSKOR_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02c>; 1118defm DS_WRXCHG_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02d>; 1119defm DS_WRXCHG2_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02e>; 1120defm DS_WRXCHG2ST64_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x02f>; 1121defm DS_CMPST_RTN_B32 : DS_Real_gfx6_gfx7_gfx10<0x030>; 1122defm DS_CMPST_RTN_F32 : DS_Real_gfx6_gfx7_gfx10<0x031>; 1123defm DS_MIN_RTN_F32 : DS_Real_gfx6_gfx7_gfx10<0x032>; 1124defm DS_MAX_RTN_F32 : DS_Real_gfx6_gfx7_gfx10<0x033>; 1125defm DS_SWIZZLE_B32 : DS_Real_gfx6_gfx7_gfx10<0x035>; 1126defm DS_READ_B32 : DS_Real_gfx6_gfx7_gfx10<0x036>; 1127defm DS_READ2_B32 : DS_Real_gfx6_gfx7_gfx10<0x037>; 1128defm DS_READ2ST64_B32 : DS_Real_gfx6_gfx7_gfx10<0x038>; 1129defm DS_READ_I8 : DS_Real_gfx6_gfx7_gfx10<0x039>; 1130defm DS_READ_U8 : DS_Real_gfx6_gfx7_gfx10<0x03a>; 1131defm DS_READ_I16 : DS_Real_gfx6_gfx7_gfx10<0x03b>; 1132defm DS_READ_U16 : DS_Real_gfx6_gfx7_gfx10<0x03c>; 1133defm DS_CONSUME : DS_Real_gfx6_gfx7_gfx10<0x03d>; 1134defm DS_APPEND : DS_Real_gfx6_gfx7_gfx10<0x03e>; 1135defm DS_ORDERED_COUNT : DS_Real_gfx6_gfx7_gfx10<0x03f>; 1136defm DS_ADD_U64 : DS_Real_gfx6_gfx7_gfx10<0x040>; 1137defm DS_SUB_U64 : DS_Real_gfx6_gfx7_gfx10<0x041>; 1138defm DS_RSUB_U64 : DS_Real_gfx6_gfx7_gfx10<0x042>; 1139defm DS_INC_U64 : DS_Real_gfx6_gfx7_gfx10<0x043>; 1140defm DS_DEC_U64 : DS_Real_gfx6_gfx7_gfx10<0x044>; 1141defm DS_MIN_I64 : DS_Real_gfx6_gfx7_gfx10<0x045>; 1142defm DS_MAX_I64 : DS_Real_gfx6_gfx7_gfx10<0x046>; 1143defm DS_MIN_U64 : DS_Real_gfx6_gfx7_gfx10<0x047>; 1144defm DS_MAX_U64 : DS_Real_gfx6_gfx7_gfx10<0x048>; 1145defm DS_AND_B64 : DS_Real_gfx6_gfx7_gfx10<0x049>; 1146defm DS_OR_B64 : DS_Real_gfx6_gfx7_gfx10<0x04a>; 1147defm DS_XOR_B64 : DS_Real_gfx6_gfx7_gfx10<0x04b>; 1148defm DS_MSKOR_B64 : DS_Real_gfx6_gfx7_gfx10<0x04c>; 1149defm DS_WRITE_B64 : DS_Real_gfx6_gfx7_gfx10<0x04d>; 1150defm DS_WRITE2_B64 : DS_Real_gfx6_gfx7_gfx10<0x04e>; 1151defm DS_WRITE2ST64_B64 : DS_Real_gfx6_gfx7_gfx10<0x04f>; 1152defm DS_CMPST_B64 : DS_Real_gfx6_gfx7_gfx10<0x050>; 1153defm DS_CMPST_F64 : DS_Real_gfx6_gfx7_gfx10<0x051>; 1154defm DS_MIN_F64 : DS_Real_gfx6_gfx7_gfx10<0x052>; 1155defm DS_MAX_F64 : DS_Real_gfx6_gfx7_gfx10<0x053>; 1156defm DS_ADD_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x060>; 1157defm DS_SUB_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x061>; 1158defm DS_RSUB_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x062>; 1159defm DS_INC_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x063>; 1160defm DS_DEC_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x064>; 1161defm DS_MIN_RTN_I64 : DS_Real_gfx6_gfx7_gfx10<0x065>; 1162defm DS_MAX_RTN_I64 : DS_Real_gfx6_gfx7_gfx10<0x066>; 1163defm DS_MIN_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x067>; 1164defm DS_MAX_RTN_U64 : DS_Real_gfx6_gfx7_gfx10<0x068>; 1165defm DS_AND_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x069>; 1166defm DS_OR_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06a>; 1167defm DS_XOR_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06b>; 1168defm DS_MSKOR_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06c>; 1169defm DS_WRXCHG_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06d>; 1170defm DS_WRXCHG2_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06e>; 1171defm DS_WRXCHG2ST64_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x06f>; 1172defm DS_CMPST_RTN_B64 : DS_Real_gfx6_gfx7_gfx10<0x070>; 1173defm DS_CMPST_RTN_F64 : DS_Real_gfx6_gfx7_gfx10<0x071>; 1174defm DS_MIN_RTN_F64 : DS_Real_gfx6_gfx7_gfx10<0x072>; 1175defm DS_MAX_RTN_F64 : DS_Real_gfx6_gfx7_gfx10<0x073>; 1176defm DS_READ_B64 : DS_Real_gfx6_gfx7_gfx10<0x076>; 1177defm DS_READ2_B64 : DS_Real_gfx6_gfx7_gfx10<0x077>; 1178defm DS_READ2ST64_B64 : DS_Real_gfx6_gfx7_gfx10<0x078>; 1179defm DS_ADD_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x080>; 1180defm DS_SUB_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x081>; 1181defm DS_RSUB_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x082>; 1182defm DS_INC_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x083>; 1183defm DS_DEC_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x084>; 1184defm DS_MIN_SRC2_I32 : DS_Real_gfx6_gfx7_gfx10<0x085>; 1185defm DS_MAX_SRC2_I32 : DS_Real_gfx6_gfx7_gfx10<0x086>; 1186defm DS_MIN_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x087>; 1187defm DS_MAX_SRC2_U32 : DS_Real_gfx6_gfx7_gfx10<0x088>; 1188defm DS_AND_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x089>; 1189defm DS_OR_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x08a>; 1190defm DS_XOR_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x08b>; 1191defm DS_WRITE_SRC2_B32 : DS_Real_gfx6_gfx7_gfx10<0x08d>; 1192defm DS_MIN_SRC2_F32 : DS_Real_gfx6_gfx7_gfx10<0x092>; 1193defm DS_MAX_SRC2_F32 : DS_Real_gfx6_gfx7_gfx10<0x093>; 1194defm DS_ADD_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c0>; 1195defm DS_SUB_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c1>; 1196defm DS_RSUB_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c2>; 1197defm DS_INC_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c3>; 1198defm DS_DEC_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c4>; 1199defm DS_MIN_SRC2_I64 : DS_Real_gfx6_gfx7_gfx10<0x0c5>; 1200defm DS_MAX_SRC2_I64 : DS_Real_gfx6_gfx7_gfx10<0x0c6>; 1201defm DS_MIN_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c7>; 1202defm DS_MAX_SRC2_U64 : DS_Real_gfx6_gfx7_gfx10<0x0c8>; 1203defm DS_AND_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0c9>; 1204defm DS_OR_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0ca>; 1205defm DS_XOR_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0cb>; 1206defm DS_WRITE_SRC2_B64 : DS_Real_gfx6_gfx7_gfx10<0x0cd>; 1207defm DS_MIN_SRC2_F64 : DS_Real_gfx6_gfx7_gfx10<0x0d2>; 1208defm DS_MAX_SRC2_F64 : DS_Real_gfx6_gfx7_gfx10<0x0d3>; 1209 1210//===----------------------------------------------------------------------===// 1211// GFX8, GFX9 (VI). 1212//===----------------------------------------------------------------------===// 1213 1214class DS_Real_vi <bits<8> op, DS_Pseudo ds> : 1215 DS_Real <ds>, 1216 SIMCInstr <ds.Mnemonic, SIEncodingFamily.VI> { 1217 let AssemblerPredicate = isGFX8GFX9; 1218 let DecoderNamespace = "GFX8"; 1219 1220 // encoding 1221 let Inst{7-0} = !if(ds.has_offset0, offset0, 0); 1222 let Inst{15-8} = !if(ds.has_offset1, offset1, 0); 1223 let Inst{16} = !if(ds.has_gds, gds, ds.gdsValue); 1224 let Inst{24-17} = op; 1225 let Inst{25} = acc; 1226 let Inst{31-26} = 0x36; // ds prefix 1227 let Inst{39-32} = !if(ds.has_addr, addr, !if(ds.has_gws_data0, data0{7-0}, 0)); 1228 let Inst{47-40} = !if(ds.has_data0, data0{7-0}, 0); 1229 let Inst{55-48} = !if(ds.has_data1, data1{7-0}, 0); 1230 let Inst{63-56} = !if(ds.has_vdst, vdst{7-0}, 0); 1231} 1232 1233def DS_ADD_U32_vi : DS_Real_vi<0x0, DS_ADD_U32>; 1234def DS_SUB_U32_vi : DS_Real_vi<0x1, DS_SUB_U32>; 1235def DS_RSUB_U32_vi : DS_Real_vi<0x2, DS_RSUB_U32>; 1236def DS_INC_U32_vi : DS_Real_vi<0x3, DS_INC_U32>; 1237def DS_DEC_U32_vi : DS_Real_vi<0x4, DS_DEC_U32>; 1238def DS_MIN_I32_vi : DS_Real_vi<0x5, DS_MIN_I32>; 1239def DS_MAX_I32_vi : DS_Real_vi<0x6, DS_MAX_I32>; 1240def DS_MIN_U32_vi : DS_Real_vi<0x7, DS_MIN_U32>; 1241def DS_MAX_U32_vi : DS_Real_vi<0x8, DS_MAX_U32>; 1242def DS_AND_B32_vi : DS_Real_vi<0x9, DS_AND_B32>; 1243def DS_OR_B32_vi : DS_Real_vi<0xa, DS_OR_B32>; 1244def DS_XOR_B32_vi : DS_Real_vi<0xb, DS_XOR_B32>; 1245def DS_MSKOR_B32_vi : DS_Real_vi<0xc, DS_MSKOR_B32>; 1246def DS_WRITE_B32_vi : DS_Real_vi<0xd, DS_WRITE_B32>; 1247def DS_WRITE2_B32_vi : DS_Real_vi<0xe, DS_WRITE2_B32>; 1248def DS_WRITE2ST64_B32_vi : DS_Real_vi<0xf, DS_WRITE2ST64_B32>; 1249def DS_CMPST_B32_vi : DS_Real_vi<0x10, DS_CMPST_B32>; 1250def DS_CMPST_F32_vi : DS_Real_vi<0x11, DS_CMPST_F32>; 1251def DS_MIN_F32_vi : DS_Real_vi<0x12, DS_MIN_F32>; 1252def DS_MAX_F32_vi : DS_Real_vi<0x13, DS_MAX_F32>; 1253def DS_NOP_vi : DS_Real_vi<0x14, DS_NOP>; 1254def DS_ADD_F32_vi : DS_Real_vi<0x15, DS_ADD_F32>; 1255def DS_GWS_INIT_vi : DS_Real_vi<0x99, DS_GWS_INIT>; 1256def DS_GWS_SEMA_V_vi : DS_Real_vi<0x9a, DS_GWS_SEMA_V>; 1257def DS_GWS_SEMA_BR_vi : DS_Real_vi<0x9b, DS_GWS_SEMA_BR>; 1258def DS_GWS_SEMA_P_vi : DS_Real_vi<0x9c, DS_GWS_SEMA_P>; 1259def DS_GWS_BARRIER_vi : DS_Real_vi<0x9d, DS_GWS_BARRIER>; 1260def DS_WRITE_ADDTID_B32_vi : DS_Real_vi<0x1d, DS_WRITE_ADDTID_B32>; 1261def DS_WRITE_B8_vi : DS_Real_vi<0x1e, DS_WRITE_B8>; 1262def DS_WRITE_B16_vi : DS_Real_vi<0x1f, DS_WRITE_B16>; 1263def DS_ADD_RTN_U32_vi : DS_Real_vi<0x20, DS_ADD_RTN_U32>; 1264def DS_SUB_RTN_U32_vi : DS_Real_vi<0x21, DS_SUB_RTN_U32>; 1265def DS_RSUB_RTN_U32_vi : DS_Real_vi<0x22, DS_RSUB_RTN_U32>; 1266def DS_INC_RTN_U32_vi : DS_Real_vi<0x23, DS_INC_RTN_U32>; 1267def DS_DEC_RTN_U32_vi : DS_Real_vi<0x24, DS_DEC_RTN_U32>; 1268def DS_MIN_RTN_I32_vi : DS_Real_vi<0x25, DS_MIN_RTN_I32>; 1269def DS_MAX_RTN_I32_vi : DS_Real_vi<0x26, DS_MAX_RTN_I32>; 1270def DS_MIN_RTN_U32_vi : DS_Real_vi<0x27, DS_MIN_RTN_U32>; 1271def DS_MAX_RTN_U32_vi : DS_Real_vi<0x28, DS_MAX_RTN_U32>; 1272def DS_AND_RTN_B32_vi : DS_Real_vi<0x29, DS_AND_RTN_B32>; 1273def DS_OR_RTN_B32_vi : DS_Real_vi<0x2a, DS_OR_RTN_B32>; 1274def DS_XOR_RTN_B32_vi : DS_Real_vi<0x2b, DS_XOR_RTN_B32>; 1275def DS_MSKOR_RTN_B32_vi : DS_Real_vi<0x2c, DS_MSKOR_RTN_B32>; 1276def DS_WRXCHG_RTN_B32_vi : DS_Real_vi<0x2d, DS_WRXCHG_RTN_B32>; 1277def DS_WRXCHG2_RTN_B32_vi : DS_Real_vi<0x2e, DS_WRXCHG2_RTN_B32>; 1278def DS_WRXCHG2ST64_RTN_B32_vi : DS_Real_vi<0x2f, DS_WRXCHG2ST64_RTN_B32>; 1279def DS_CMPST_RTN_B32_vi : DS_Real_vi<0x30, DS_CMPST_RTN_B32>; 1280def DS_CMPST_RTN_F32_vi : DS_Real_vi<0x31, DS_CMPST_RTN_F32>; 1281def DS_MIN_RTN_F32_vi : DS_Real_vi<0x32, DS_MIN_RTN_F32>; 1282def DS_MAX_RTN_F32_vi : DS_Real_vi<0x33, DS_MAX_RTN_F32>; 1283def DS_WRAP_RTN_B32_vi : DS_Real_vi<0x34, DS_WRAP_RTN_B32>; 1284def DS_ADD_RTN_F32_vi : DS_Real_vi<0x35, DS_ADD_RTN_F32>; 1285def DS_READ_B32_vi : DS_Real_vi<0x36, DS_READ_B32>; 1286def DS_READ2_B32_vi : DS_Real_vi<0x37, DS_READ2_B32>; 1287def DS_READ2ST64_B32_vi : DS_Real_vi<0x38, DS_READ2ST64_B32>; 1288def DS_READ_I8_vi : DS_Real_vi<0x39, DS_READ_I8>; 1289def DS_READ_U8_vi : DS_Real_vi<0x3a, DS_READ_U8>; 1290def DS_READ_I16_vi : DS_Real_vi<0x3b, DS_READ_I16>; 1291def DS_READ_U16_vi : DS_Real_vi<0x3c, DS_READ_U16>; 1292def DS_READ_ADDTID_B32_vi : DS_Real_vi<0xb6, DS_READ_ADDTID_B32>; 1293def DS_CONSUME_vi : DS_Real_vi<0xbd, DS_CONSUME>; 1294def DS_APPEND_vi : DS_Real_vi<0xbe, DS_APPEND>; 1295def DS_ORDERED_COUNT_vi : DS_Real_vi<0xbf, DS_ORDERED_COUNT>; 1296def DS_SWIZZLE_B32_vi : DS_Real_vi<0x3d, DS_SWIZZLE_B32>; 1297def DS_PERMUTE_B32_vi : DS_Real_vi<0x3e, DS_PERMUTE_B32>; 1298def DS_BPERMUTE_B32_vi : DS_Real_vi<0x3f, DS_BPERMUTE_B32>; 1299 1300def DS_ADD_U64_vi : DS_Real_vi<0x40, DS_ADD_U64>; 1301def DS_SUB_U64_vi : DS_Real_vi<0x41, DS_SUB_U64>; 1302def DS_RSUB_U64_vi : DS_Real_vi<0x42, DS_RSUB_U64>; 1303def DS_INC_U64_vi : DS_Real_vi<0x43, DS_INC_U64>; 1304def DS_DEC_U64_vi : DS_Real_vi<0x44, DS_DEC_U64>; 1305def DS_MIN_I64_vi : DS_Real_vi<0x45, DS_MIN_I64>; 1306def DS_MAX_I64_vi : DS_Real_vi<0x46, DS_MAX_I64>; 1307def DS_MIN_U64_vi : DS_Real_vi<0x47, DS_MIN_U64>; 1308def DS_MAX_U64_vi : DS_Real_vi<0x48, DS_MAX_U64>; 1309def DS_AND_B64_vi : DS_Real_vi<0x49, DS_AND_B64>; 1310def DS_OR_B64_vi : DS_Real_vi<0x4a, DS_OR_B64>; 1311def DS_XOR_B64_vi : DS_Real_vi<0x4b, DS_XOR_B64>; 1312def DS_MSKOR_B64_vi : DS_Real_vi<0x4c, DS_MSKOR_B64>; 1313def DS_WRITE_B64_vi : DS_Real_vi<0x4d, DS_WRITE_B64>; 1314def DS_WRITE2_B64_vi : DS_Real_vi<0x4E, DS_WRITE2_B64>; 1315def DS_WRITE2ST64_B64_vi : DS_Real_vi<0x4f, DS_WRITE2ST64_B64>; 1316def DS_CMPST_B64_vi : DS_Real_vi<0x50, DS_CMPST_B64>; 1317def DS_CMPST_F64_vi : DS_Real_vi<0x51, DS_CMPST_F64>; 1318def DS_MIN_F64_vi : DS_Real_vi<0x52, DS_MIN_F64>; 1319def DS_MAX_F64_vi : DS_Real_vi<0x53, DS_MAX_F64>; 1320 1321def DS_WRITE_B8_D16_HI_vi : DS_Real_vi<0x54, DS_WRITE_B8_D16_HI>; 1322def DS_WRITE_B16_D16_HI_vi : DS_Real_vi<0x55, DS_WRITE_B16_D16_HI>; 1323 1324def DS_READ_U8_D16_vi : DS_Real_vi<0x56, DS_READ_U8_D16>; 1325def DS_READ_U8_D16_HI_vi : DS_Real_vi<0x57, DS_READ_U8_D16_HI>; 1326def DS_READ_I8_D16_vi : DS_Real_vi<0x58, DS_READ_I8_D16>; 1327def DS_READ_I8_D16_HI_vi : DS_Real_vi<0x59, DS_READ_I8_D16_HI>; 1328def DS_READ_U16_D16_vi : DS_Real_vi<0x5a, DS_READ_U16_D16>; 1329def DS_READ_U16_D16_HI_vi : DS_Real_vi<0x5b, DS_READ_U16_D16_HI>; 1330 1331def DS_ADD_RTN_U64_vi : DS_Real_vi<0x60, DS_ADD_RTN_U64>; 1332def DS_SUB_RTN_U64_vi : DS_Real_vi<0x61, DS_SUB_RTN_U64>; 1333def DS_RSUB_RTN_U64_vi : DS_Real_vi<0x62, DS_RSUB_RTN_U64>; 1334def DS_INC_RTN_U64_vi : DS_Real_vi<0x63, DS_INC_RTN_U64>; 1335def DS_DEC_RTN_U64_vi : DS_Real_vi<0x64, DS_DEC_RTN_U64>; 1336def DS_MIN_RTN_I64_vi : DS_Real_vi<0x65, DS_MIN_RTN_I64>; 1337def DS_MAX_RTN_I64_vi : DS_Real_vi<0x66, DS_MAX_RTN_I64>; 1338def DS_MIN_RTN_U64_vi : DS_Real_vi<0x67, DS_MIN_RTN_U64>; 1339def DS_MAX_RTN_U64_vi : DS_Real_vi<0x68, DS_MAX_RTN_U64>; 1340def DS_AND_RTN_B64_vi : DS_Real_vi<0x69, DS_AND_RTN_B64>; 1341def DS_OR_RTN_B64_vi : DS_Real_vi<0x6a, DS_OR_RTN_B64>; 1342def DS_XOR_RTN_B64_vi : DS_Real_vi<0x6b, DS_XOR_RTN_B64>; 1343def DS_MSKOR_RTN_B64_vi : DS_Real_vi<0x6c, DS_MSKOR_RTN_B64>; 1344def DS_WRXCHG_RTN_B64_vi : DS_Real_vi<0x6d, DS_WRXCHG_RTN_B64>; 1345def DS_WRXCHG2_RTN_B64_vi : DS_Real_vi<0x6e, DS_WRXCHG2_RTN_B64>; 1346def DS_WRXCHG2ST64_RTN_B64_vi : DS_Real_vi<0x6f, DS_WRXCHG2ST64_RTN_B64>; 1347def DS_CONDXCHG32_RTN_B64_vi : DS_Real_vi<0x7e, DS_CONDXCHG32_RTN_B64>; 1348def DS_GWS_SEMA_RELEASE_ALL_vi : DS_Real_vi<0x98, DS_GWS_SEMA_RELEASE_ALL>; 1349def DS_CMPST_RTN_B64_vi : DS_Real_vi<0x70, DS_CMPST_RTN_B64>; 1350def DS_CMPST_RTN_F64_vi : DS_Real_vi<0x71, DS_CMPST_RTN_F64>; 1351def DS_MIN_RTN_F64_vi : DS_Real_vi<0x72, DS_MIN_RTN_F64>; 1352def DS_MAX_RTN_F64_vi : DS_Real_vi<0x73, DS_MAX_RTN_F64>; 1353 1354def DS_READ_B64_vi : DS_Real_vi<0x76, DS_READ_B64>; 1355def DS_READ2_B64_vi : DS_Real_vi<0x77, DS_READ2_B64>; 1356def DS_READ2ST64_B64_vi : DS_Real_vi<0x78, DS_READ2ST64_B64>; 1357 1358def DS_ADD_SRC2_U32_vi : DS_Real_vi<0x80, DS_ADD_SRC2_U32>; 1359def DS_SUB_SRC2_U32_vi : DS_Real_vi<0x81, DS_SUB_SRC2_U32>; 1360def DS_RSUB_SRC2_U32_vi : DS_Real_vi<0x82, DS_RSUB_SRC2_U32>; 1361def DS_INC_SRC2_U32_vi : DS_Real_vi<0x83, DS_INC_SRC2_U32>; 1362def DS_DEC_SRC2_U32_vi : DS_Real_vi<0x84, DS_DEC_SRC2_U32>; 1363def DS_MIN_SRC2_I32_vi : DS_Real_vi<0x85, DS_MIN_SRC2_I32>; 1364def DS_MAX_SRC2_I32_vi : DS_Real_vi<0x86, DS_MAX_SRC2_I32>; 1365def DS_MIN_SRC2_U32_vi : DS_Real_vi<0x87, DS_MIN_SRC2_U32>; 1366def DS_MAX_SRC2_U32_vi : DS_Real_vi<0x88, DS_MAX_SRC2_U32>; 1367def DS_AND_SRC2_B32_vi : DS_Real_vi<0x89, DS_AND_SRC2_B32>; 1368def DS_OR_SRC2_B32_vi : DS_Real_vi<0x8a, DS_OR_SRC2_B32>; 1369def DS_XOR_SRC2_B32_vi : DS_Real_vi<0x8b, DS_XOR_SRC2_B32>; 1370def DS_WRITE_SRC2_B32_vi : DS_Real_vi<0x8d, DS_WRITE_SRC2_B32>; 1371def DS_MIN_SRC2_F32_vi : DS_Real_vi<0x92, DS_MIN_SRC2_F32>; 1372def DS_MAX_SRC2_F32_vi : DS_Real_vi<0x93, DS_MAX_SRC2_F32>; 1373def DS_ADD_SRC2_F32_vi : DS_Real_vi<0x95, DS_ADD_SRC2_F32>; 1374def DS_ADD_SRC2_U64_vi : DS_Real_vi<0xc0, DS_ADD_SRC2_U64>; 1375def DS_SUB_SRC2_U64_vi : DS_Real_vi<0xc1, DS_SUB_SRC2_U64>; 1376def DS_RSUB_SRC2_U64_vi : DS_Real_vi<0xc2, DS_RSUB_SRC2_U64>; 1377def DS_INC_SRC2_U64_vi : DS_Real_vi<0xc3, DS_INC_SRC2_U64>; 1378def DS_DEC_SRC2_U64_vi : DS_Real_vi<0xc4, DS_DEC_SRC2_U64>; 1379def DS_MIN_SRC2_I64_vi : DS_Real_vi<0xc5, DS_MIN_SRC2_I64>; 1380def DS_MAX_SRC2_I64_vi : DS_Real_vi<0xc6, DS_MAX_SRC2_I64>; 1381def DS_MIN_SRC2_U64_vi : DS_Real_vi<0xc7, DS_MIN_SRC2_U64>; 1382def DS_MAX_SRC2_U64_vi : DS_Real_vi<0xc8, DS_MAX_SRC2_U64>; 1383def DS_AND_SRC2_B64_vi : DS_Real_vi<0xc9, DS_AND_SRC2_B64>; 1384def DS_OR_SRC2_B64_vi : DS_Real_vi<0xca, DS_OR_SRC2_B64>; 1385def DS_XOR_SRC2_B64_vi : DS_Real_vi<0xcb, DS_XOR_SRC2_B64>; 1386def DS_WRITE_SRC2_B64_vi : DS_Real_vi<0xcd, DS_WRITE_SRC2_B64>; 1387def DS_MIN_SRC2_F64_vi : DS_Real_vi<0xd2, DS_MIN_SRC2_F64>; 1388def DS_MAX_SRC2_F64_vi : DS_Real_vi<0xd3, DS_MAX_SRC2_F64>; 1389def DS_WRITE_B96_vi : DS_Real_vi<0xde, DS_WRITE_B96>; 1390def DS_WRITE_B128_vi : DS_Real_vi<0xdf, DS_WRITE_B128>; 1391def DS_READ_B96_vi : DS_Real_vi<0xfe, DS_READ_B96>; 1392def DS_READ_B128_vi : DS_Real_vi<0xff, DS_READ_B128>; 1393 1394let SubtargetPredicate = isGFX90APlus in { 1395 def DS_ADD_F64_vi : DS_Real_vi<0x5c, DS_ADD_F64>; 1396 def DS_ADD_RTN_F64_vi : DS_Real_vi<0x7c, DS_ADD_RTN_F64>; 1397} // End SubtargetPredicate = isGFX90APlus 1398