1//===-- SIInstrFormats.td - SI Instruction Encodings ----------------------===// 2// 3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4// See https://llvm.org/LICENSE.txt for license information. 5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6// 7//===----------------------------------------------------------------------===// 8// 9// SI Instruction format definitions. 10// 11//===----------------------------------------------------------------------===// 12 13class InstSI <dag outs, dag ins, string asm = "", 14 list<dag> pattern = []> : 15 AMDGPUInst<outs, ins, asm, pattern>, GCNPredicateControl { 16 // Low bits - basic encoding information. 17 field bit SALU = 0; 18 field bit VALU = 0; 19 20 // SALU instruction formats. 21 field bit SOP1 = 0; 22 field bit SOP2 = 0; 23 field bit SOPC = 0; 24 field bit SOPK = 0; 25 field bit SOPP = 0; 26 27 // VALU instruction formats. 28 field bit VOP1 = 0; 29 field bit VOP2 = 0; 30 field bit VOPC = 0; 31 field bit VOP3 = 0; 32 field bit VOP3P = 0; 33 field bit VINTRP = 0; 34 field bit SDWA = 0; 35 field bit DPP = 0; 36 field bit TRANS = 0; 37 38 // Memory instruction formats. 39 field bit MUBUF = 0; 40 field bit MTBUF = 0; 41 field bit SMRD = 0; 42 field bit MIMG = 0; 43 field bit EXP = 0; 44 field bit FLAT = 0; 45 field bit DS = 0; 46 47 // Pseudo instruction formats. 48 field bit VGPRSpill = 0; 49 field bit SGPRSpill = 0; 50 51 // LDSDIR instruction format. 52 field bit LDSDIR = 0; 53 54 // VINTERP instruction format. 55 field bit VINTERP = 0; 56 57 // High bits - other information. 58 field bit VM_CNT = 0; 59 field bit EXP_CNT = 0; 60 field bit LGKM_CNT = 0; 61 62 // Whether WQM _must_ be enabled for this instruction. 63 field bit WQM = 0; 64 65 // Whether WQM _must_ be disabled for this instruction. 66 field bit DisableWQM = 0; 67 68 field bit Gather4 = 0; 69 70 // Most sopk treat the immediate as a signed 16-bit, however some 71 // use it as unsigned. 72 field bit SOPKZext = 0; 73 74 // This is an s_store_dword* instruction that requires a cache flush 75 // on wave termination. It is necessary to distinguish from mayStore 76 // SMEM instructions like the cache flush ones. 77 field bit ScalarStore = 0; 78 79 // Whether the operands can be ignored when computing the 80 // instruction size. 81 field bit FixedSize = 0; 82 83 // This bit tells the assembler to use the 32-bit encoding in case it 84 // is unable to infer the encoding from the operands. 85 field bit VOPAsmPrefer32Bit = 0; 86 87 // This bit indicates that this is a VOP3 opcode which supports op_sel 88 // modifier (gfx9 only). 89 field bit VOP3_OPSEL = 0; 90 91 // Is it possible for this instruction to be atomic? 92 field bit maybeAtomic = 0; 93 94 // This bit indicates that this is a VI instruction which is renamed 95 // in GFX9. Required for correct mapping from pseudo to MC. 96 field bit renamedInGFX9 = 0; 97 98 // This bit indicates that this has a floating point result type, so 99 // the clamp modifier has floating point semantics. 100 field bit FPClamp = 0; 101 102 // This bit indicates that instruction may support integer clamping 103 // which depends on GPU features. 104 field bit IntClamp = 0; 105 106 // This field indicates that the clamp applies to the low component 107 // of a packed output register. 108 field bit ClampLo = 0; 109 110 // This field indicates that the clamp applies to the high component 111 // of a packed output register. 112 field bit ClampHi = 0; 113 114 // This bit indicates that this is a packed VOP3P instruction 115 field bit IsPacked = 0; 116 117 // This bit indicates that this is a D16 buffer instruction. 118 field bit D16Buf = 0; 119 120 // This field indicates that FLAT instruction accesses FLAT_GLBL segment. 121 // Must be 0 for non-FLAT instructions. 122 field bit FlatGlobal = 0; 123 124 // Reads the mode register, usually for FP environment. 125 field bit ReadsModeReg = 0; 126 127 // This bit indicates that this uses the floating point double precision 128 // rounding mode flags 129 field bit FPDPRounding = 0; 130 131 // Instruction is FP atomic. 132 field bit FPAtomic = 0; 133 134 // This bit indicates that this is one of MFMA instructions. 135 field bit IsMAI = 0; 136 137 // This bit indicates that this is one of DOT instructions. 138 field bit IsDOT = 0; 139 140 // This field indicates that FLAT instruction accesses FLAT_SCRATCH segment. 141 // Must be 0 for non-FLAT instructions. 142 field bit FlatScratch = 0; 143 144 // Atomic without a return. 145 field bit IsAtomicNoRet = 0; 146 147 // Atomic with return. 148 field bit IsAtomicRet = 0; 149 150 // These need to be kept in sync with the enum in SIInstrFlags. 151 let TSFlags{0} = SALU; 152 let TSFlags{1} = VALU; 153 154 let TSFlags{2} = SOP1; 155 let TSFlags{3} = SOP2; 156 let TSFlags{4} = SOPC; 157 let TSFlags{5} = SOPK; 158 let TSFlags{6} = SOPP; 159 160 let TSFlags{7} = VOP1; 161 let TSFlags{8} = VOP2; 162 let TSFlags{9} = VOPC; 163 let TSFlags{10} = VOP3; 164 let TSFlags{12} = VOP3P; 165 166 let TSFlags{13} = VINTRP; 167 let TSFlags{14} = SDWA; 168 let TSFlags{15} = DPP; 169 let TSFlags{16} = TRANS; 170 171 let TSFlags{17} = MUBUF; 172 let TSFlags{18} = MTBUF; 173 let TSFlags{19} = SMRD; 174 let TSFlags{20} = MIMG; 175 let TSFlags{21} = EXP; 176 let TSFlags{22} = FLAT; 177 let TSFlags{23} = DS; 178 179 let TSFlags{24} = VGPRSpill; 180 let TSFlags{25} = SGPRSpill; 181 182 let TSFlags{26} = LDSDIR; 183 let TSFlags{27} = VINTERP; 184 185 let TSFlags{32} = VM_CNT; 186 let TSFlags{33} = EXP_CNT; 187 let TSFlags{34} = LGKM_CNT; 188 189 let TSFlags{35} = WQM; 190 let TSFlags{36} = DisableWQM; 191 let TSFlags{37} = Gather4; 192 193 let TSFlags{38} = SOPKZext; 194 let TSFlags{39} = ScalarStore; 195 let TSFlags{40} = FixedSize; 196 let TSFlags{41} = VOPAsmPrefer32Bit; 197 let TSFlags{42} = VOP3_OPSEL; 198 199 let TSFlags{43} = maybeAtomic; 200 let TSFlags{44} = renamedInGFX9; 201 202 let TSFlags{45} = FPClamp; 203 let TSFlags{46} = IntClamp; 204 let TSFlags{47} = ClampLo; 205 let TSFlags{48} = ClampHi; 206 207 let TSFlags{49} = IsPacked; 208 209 let TSFlags{50} = D16Buf; 210 211 let TSFlags{51} = FlatGlobal; 212 213 let TSFlags{52} = FPDPRounding; 214 215 let TSFlags{53} = FPAtomic; 216 217 let TSFlags{54} = IsMAI; 218 219 let TSFlags{55} = IsDOT; 220 221 let TSFlags{56} = FlatScratch; 222 223 let TSFlags{57} = IsAtomicNoRet; 224 225 let TSFlags{58} = IsAtomicRet; 226 227 let SchedRW = [Write32Bit]; 228 229 let AsmVariantName = AMDGPUAsmVariants.Default; 230 231 // Avoid changing source registers in a way that violates constant bus read limitations. 232 let hasExtraSrcRegAllocReq = !or(VOP1, VOP2, VOP3, VOPC, SDWA, VALU); 233} 234 235class PseudoInstSI<dag outs, dag ins, list<dag> pattern = [], string asm = ""> 236 : InstSI<outs, ins, asm, pattern> { 237 let isPseudo = 1; 238 let isCodeGenOnly = 1; 239} 240 241class SPseudoInstSI<dag outs, dag ins, list<dag> pattern = [], string asm = ""> 242 : PseudoInstSI<outs, ins, pattern, asm> { 243 let SALU = 1; 244} 245 246class VPseudoInstSI<dag outs, dag ins, list<dag> pattern = [], string asm = ""> 247 : PseudoInstSI<outs, ins, pattern, asm> { 248 let VALU = 1; 249 let Uses = [EXEC]; 250} 251 252class CFPseudoInstSI<dag outs, dag ins, list<dag> pattern = [], 253 bit UseExec = 0, bit DefExec = 0> : 254 SPseudoInstSI<outs, ins, pattern> { 255 256 let Uses = !if(UseExec, [EXEC], []); 257 let Defs = !if(DefExec, [EXEC, SCC], [SCC]); 258 let mayLoad = 0; 259 let mayStore = 0; 260 let hasSideEffects = 0; 261} 262 263class Enc32 { 264 field bits<32> Inst; 265 int Size = 4; 266} 267 268class Enc64 { 269 field bits<64> Inst; 270 int Size = 8; 271} 272 273def CPolBit { 274 int GLC = 0; 275 int SLC = 1; 276 int DLC = 2; 277 int SCC = 4; 278} 279 280class VOPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVOPDst">; 281 282class VINTRPe <bits<2> op> : Enc32 { 283 bits<8> vdst; 284 bits<8> vsrc; 285 bits<2> attrchan; 286 bits<6> attr; 287 288 let Inst{7-0} = vsrc; 289 let Inst{9-8} = attrchan; 290 let Inst{15-10} = attr; 291 let Inst{17-16} = op; 292 let Inst{25-18} = vdst; 293 let Inst{31-26} = 0x32; // encoding 294} 295 296class MIMGe_gfxpre11 : Enc64 { 297 bits<10> vdata; 298 bits<4> dmask; 299 bits<1> unorm; 300 bits<5> cpol; 301 bits<1> r128; 302 bits<1> tfe; 303 bits<1> lwe; 304 bit d16; 305 bits<7> srsrc; 306 bits<7> ssamp; 307 308 let Inst{11-8} = dmask; 309 let Inst{12} = unorm; 310 let Inst{13} = cpol{CPolBit.GLC}; 311 let Inst{15} = r128; 312 let Inst{17} = lwe; 313 let Inst{25} = cpol{CPolBit.SLC}; 314 let Inst{31-26} = 0x3c; 315 let Inst{47-40} = vdata{7-0}; 316 let Inst{52-48} = srsrc{6-2}; 317 let Inst{57-53} = ssamp{6-2}; 318 let Inst{63} = d16; 319} 320 321class MIMGe_gfx6789 <bits<8> op> : MIMGe_gfxpre11 { 322 bits<8> vaddr; 323 bits<1> da; 324 325 let Inst{0} = op{7}; 326 let Inst{7} = cpol{CPolBit.SCC}; 327 let Inst{14} = da; 328 let Inst{16} = tfe; 329 let Inst{24-18} = op{6-0}; 330 let Inst{39-32} = vaddr; 331} 332 333class MIMGe_gfx90a <bits<8> op> : MIMGe_gfxpre11 { 334 bits<8> vaddr; 335 bits<1> da; 336 337 let Inst{0} = op{7}; 338 let Inst{7} = cpol{CPolBit.SCC}; 339 let Inst{14} = da; 340 let Inst{16} = vdata{9}; // ACC bit 341 let Inst{24-18} = op{6-0}; 342 let Inst{39-32} = vaddr; 343} 344 345class MIMGe_gfx10 <bits<8> op> : MIMGe_gfxpre11 { 346 bits<8> vaddr0; 347 bits<3> dim; 348 bits<2> nsa; 349 bits<1> a16; 350 351 let Inst{0} = op{7}; 352 let Inst{2-1} = nsa; 353 let Inst{5-3} = dim; 354 let Inst{7} = cpol{CPolBit.DLC}; 355 let Inst{16} = tfe; 356 let Inst{24-18} = op{6-0}; 357 let Inst{39-32} = vaddr0; 358 let Inst{62} = a16; 359} 360 361class MIMGe_gfx11 <bits<8> op> : Enc64 { 362 bits<8> vdata; 363 bits<4> dmask; 364 bits<1> unorm; 365 bits<5> cpol; 366 bits<1> r128; 367 bits<1> tfe; 368 bits<1> lwe; 369 bits<7> srsrc; 370 bits<7> ssamp; 371 bit d16; 372 bits<1> a16; 373 bits<8> vaddr0; 374 bits<3> dim; 375 bits<1> nsa; 376 377 let Inst{0} = nsa; 378 let Inst{4-2} = dim; 379 let Inst{7} = unorm; 380 let Inst{11-8} = dmask; 381 let Inst{12} = cpol{CPolBit.SLC}; 382 let Inst{13} = cpol{CPolBit.DLC}; 383 let Inst{14} = cpol{CPolBit.GLC}; 384 let Inst{15} = r128; 385 let Inst{16} = a16; 386 let Inst{17} = d16; 387 let Inst{25-18} = op; 388 let Inst{31-26} = 0x3c; 389 let Inst{39-32} = vaddr0; 390 let Inst{47-40} = vdata; 391 let Inst{52-48} = srsrc{6-2}; 392 let Inst{53} = tfe; 393 let Inst{54} = lwe; 394 let Inst{62-58} = ssamp{6-2}; 395} 396 397class EXPe : Enc64 { 398 bits<4> en; 399 bits<6> tgt; 400 bits<1> done; 401 bits<8> src0; 402 bits<8> src1; 403 bits<8> src2; 404 bits<8> src3; 405 406 let Inst{3-0} = en; 407 let Inst{9-4} = tgt; 408 let Inst{11} = done; 409 let Inst{31-26} = 0x3e; 410 let Inst{39-32} = src0; 411 let Inst{47-40} = src1; 412 let Inst{55-48} = src2; 413 let Inst{63-56} = src3; 414} 415 416// Pre-GFX11 encoding has compr and vm bits. 417class EXPe_ComprVM : EXPe { 418 bits<1> compr; 419 bits<1> vm; 420 421 let Inst{10} = compr; 422 let Inst{12} = vm; 423} 424 425// GFX11+ encoding has row bit. 426class EXPe_Row : EXPe { 427 bits<1> row; 428 429 let Inst{13} = row; 430} 431 432let Uses = [EXEC] in { 433 434class VINTRPCommon <dag outs, dag ins, string asm, list<dag> pattern> : 435 InstSI <outs, ins, asm, pattern> { 436 let VINTRP = 1; 437 // VINTRP instructions read parameter values from LDS, but these parameter 438 // values are stored outside of the LDS memory that is allocated to the 439 // shader for general purpose use. 440 // 441 // While it may be possible for ds_read/ds_write instructions to access 442 // the parameter values in LDS, this would essentially be an out-of-bounds 443 // memory access which we consider to be undefined behavior. 444 // 445 // So even though these instructions read memory, this memory is outside the 446 // addressable memory space for the shader, and we consider these instructions 447 // to be readnone. 448 let mayLoad = 0; 449 let mayStore = 0; 450 let hasSideEffects = 0; 451 let VALU = 1; 452} 453 454} // End Uses = [EXEC] 455