1//===---- SMInstructions.td - Scalar Memory 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 9def smrd_offset_8 : NamedOperandU32<"SMRDOffset8", 10 NamedMatchClass<"SMRDOffset8">> { 11 let OperandType = "OPERAND_IMMEDIATE"; 12} 13 14def smem_offset : NamedOperandU32<"SMEMOffset", 15 NamedMatchClass<"SMEMOffset">> { 16 let OperandType = "OPERAND_IMMEDIATE"; 17 let EncoderMethod = "getSMEMOffsetEncoding"; 18 let DecoderMethod = "decodeSMEMOffset"; 19} 20 21//===----------------------------------------------------------------------===// 22// Scalar Memory classes 23//===----------------------------------------------------------------------===// 24 25class SM_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]> : 26 InstSI <outs, ins, "", pattern>, 27 SIMCInstr<opName, SIEncodingFamily.NONE> { 28 let isPseudo = 1; 29 let isCodeGenOnly = 1; 30 31 let LGKM_CNT = 1; 32 let SMRD = 1; 33 let mayStore = 0; 34 let mayLoad = 1; 35 let hasSideEffects = 0; 36 let UseNamedOperandTable = 1; 37 let SchedRW = [WriteSMEM]; 38 39 string Mnemonic = opName; 40 string AsmOperands = asmOps; 41 42 bits<1> has_sbase = 1; 43 bits<1> has_sdst = 1; 44 bit has_glc = 0; 45 bit has_dlc = 0; 46 bits<1> has_offset = 1; 47 bits<1> offset_is_imm = 0; 48 bit is_buffer = 0; 49} 50 51class SM_Real <SM_Pseudo ps> 52 : InstSI<ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []> { 53 54 let isPseudo = 0; 55 let isCodeGenOnly = 0; 56 57 Instruction Opcode = !cast<Instruction>(NAME); 58 59 // copy relevant pseudo op flags 60 let SubtargetPredicate = ps.SubtargetPredicate; 61 let AsmMatchConverter = ps.AsmMatchConverter; 62 let UseNamedOperandTable = ps.UseNamedOperandTable; 63 let SMRD = ps.SMRD; 64 let SchedRW = ps.SchedRW; 65 66 let TSFlags = ps.TSFlags; 67 68 bit is_buffer = ps.is_buffer; 69 70 // encoding 71 bits<7> sbase; 72 bits<7> sdst; 73 bits<32> offset; 74 bits<1> imm = !if(ps.has_offset, ps.offset_is_imm, 0); 75 bits<5> cpol; 76} 77 78class SM_Probe_Pseudo <string opName, dag ins, bit isImm> 79 : SM_Pseudo<opName, (outs), ins, " $sdata, $sbase, $offset"> { 80 let mayLoad = 0; 81 let mayStore = 0; 82 let has_glc = 0; 83 let LGKM_CNT = 0; 84 let ScalarStore = 0; 85 let hasSideEffects = 1; 86 let offset_is_imm = isImm; 87 let PseudoInstr = opName # !if(isImm, "_IMM", "_SGPR"); 88} 89 90class SM_Load_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]> 91 : SM_Pseudo<opName, outs, ins, asmOps, pattern> { 92 RegisterClass BaseClass; 93 let mayLoad = 1; 94 let mayStore = 0; 95 let has_glc = 1; 96 let has_dlc = 1; 97} 98 99class SM_Store_Pseudo <string opName, dag ins, string asmOps, list<dag> pattern = []> 100 : SM_Pseudo<opName, (outs), ins, asmOps, pattern> { 101 RegisterClass BaseClass; 102 RegisterClass SrcClass; 103 let mayLoad = 0; 104 let mayStore = 1; 105 let has_glc = 1; 106 let has_dlc = 1; 107 let ScalarStore = 1; 108} 109 110class SM_Discard_Pseudo <string opName, dag ins, bit isImm> 111 : SM_Pseudo<opName, (outs), ins, " $sbase, $offset"> { 112 let mayLoad = 0; 113 let mayStore = 0; 114 let has_glc = 0; 115 let has_sdst = 0; 116 let ScalarStore = 0; 117 let hasSideEffects = 1; 118 let offset_is_imm = isImm; 119 let PseudoInstr = opName # !if(isImm, "_IMM", "_SGPR"); 120} 121 122multiclass SM_Pseudo_Loads<string opName, 123 RegisterClass baseClass, 124 RegisterClass dstClass> { 125 def _IMM : SM_Load_Pseudo <opName, 126 (outs dstClass:$sdst), 127 (ins baseClass:$sbase, i32imm:$offset, CPol:$cpol), 128 " $sdst, $sbase, $offset$cpol", []> { 129 let offset_is_imm = 1; 130 let BaseClass = baseClass; 131 let PseudoInstr = opName # "_IMM"; 132 let has_glc = 1; 133 let has_dlc = 1; 134 } 135 136 def _SGPR : SM_Load_Pseudo <opName, 137 (outs dstClass:$sdst), 138 (ins baseClass:$sbase, SReg_32:$soff, CPol:$cpol), 139 " $sdst, $sbase, $offset$cpol", []> { 140 let BaseClass = baseClass; 141 let PseudoInstr = opName # "_SGPR"; 142 let has_glc = 1; 143 let has_dlc = 1; 144 } 145} 146 147multiclass SM_Pseudo_Stores<string opName, 148 RegisterClass baseClass, 149 RegisterClass srcClass> { 150 def _IMM : SM_Store_Pseudo <opName, 151 (ins srcClass:$sdata, baseClass:$sbase, i32imm:$offset, CPol:$cpol), 152 " $sdata, $sbase, $offset$cpol", []> { 153 let offset_is_imm = 1; 154 let BaseClass = baseClass; 155 let SrcClass = srcClass; 156 let PseudoInstr = opName # "_IMM"; 157 } 158 159 def _SGPR : SM_Store_Pseudo <opName, 160 (ins srcClass:$sdata, baseClass:$sbase, SReg_32:$soff, CPol:$cpol), 161 " $sdata, $sbase, $offset$cpol", []> { 162 let BaseClass = baseClass; 163 let SrcClass = srcClass; 164 let PseudoInstr = opName # "_SGPR"; 165 } 166} 167 168multiclass SM_Pseudo_Discards<string opName> { 169 def _IMM : SM_Discard_Pseudo <opName, (ins SReg_64:$sbase, smem_offset:$offset), 1>; 170 def _SGPR : SM_Discard_Pseudo <opName, (ins SReg_64:$sbase, SReg_32:$offset), 0>; 171} 172 173class SM_Time_Pseudo<string opName, SDPatternOperator node = null_frag> : SM_Pseudo< 174 opName, (outs SReg_64_XEXEC:$sdst), (ins), 175 " $sdst", [(set i64:$sdst, (node))]> { 176 let hasSideEffects = 1; 177 178 // FIXME: This should be definitively mayStore = 0. TableGen 179 // brokenly tries to infer these based on the intrinsic properties 180 // corresponding to the IR attributes. The target intrinsics are 181 // considered as writing to memory for IR dependency purposes, but 182 // those can be modeled with hasSideEffects here. These also end up 183 // inferring differently for llvm.readcyclecounter and the amdgcn 184 // intrinsics. 185 let mayStore = ?; 186 let mayLoad = 1; 187 let has_sbase = 0; 188 let has_offset = 0; 189} 190 191class SM_Inval_Pseudo <string opName, SDPatternOperator node = null_frag> : SM_Pseudo< 192 opName, (outs), (ins), "", [(node)]> { 193 let hasSideEffects = 1; 194 let mayStore = 0; 195 let has_sdst = 0; 196 let has_sbase = 0; 197 let has_offset = 0; 198} 199 200multiclass SM_Pseudo_Probe<string opName, RegisterClass baseClass> { 201 def _IMM : SM_Probe_Pseudo <opName, (ins i8imm:$sdata, baseClass:$sbase, smem_offset:$offset), 1>; 202 def _SGPR : SM_Probe_Pseudo <opName, (ins i8imm:$sdata, baseClass:$sbase, SReg_32:$offset), 0>; 203} 204 205class SM_WaveId_Pseudo<string opName, SDPatternOperator node> : SM_Pseudo< 206 opName, (outs SReg_32_XM0_XEXEC:$sdst), (ins), 207 " $sdst", [(set i32:$sdst, (node))]> { 208 let hasSideEffects = 1; 209 let mayStore = 0; 210 let mayLoad = 1; 211 let has_sbase = 0; 212 let has_offset = 0; 213} 214 215//===----------------------------------------------------------------------===// 216// Scalar Atomic Memory Classes 217//===----------------------------------------------------------------------===// 218 219class SM_Atomic_Pseudo <string opName, 220 dag outs, dag ins, string asmOps, bit isRet> 221 : SM_Pseudo<opName, outs, ins, asmOps, []> { 222 223 bit glc = isRet; 224 225 let mayLoad = 1; 226 let mayStore = 1; 227 let has_glc = 1; 228 let has_dlc = 1; 229 230 // Should these be set? 231 let ScalarStore = 1; 232 let hasSideEffects = 1; 233 let maybeAtomic = 1; 234 235 let IsAtomicNoRet = !not(isRet); 236 let IsAtomicRet = isRet; 237 238 let AsmMatchConverter = "cvtSMEMAtomic"; 239} 240 241class SM_Pseudo_Atomic<string opName, 242 RegisterClass baseClass, 243 RegisterClass dataClass, 244 bit isImm, 245 bit isRet, 246 string opNameWithSuffix = opName # !if(isImm, 247 !if(isRet, "_IMM_RTN", "_IMM"), 248 !if(isRet, "_SGPR_RTN", "_SGPR")), 249 Operand CPolTy = !if(isRet, CPol_GLC1, CPol)> : 250 SM_Atomic_Pseudo<opName, 251 !if(isRet, (outs dataClass:$sdst), (outs)), 252 !if(isImm, 253 (ins dataClass:$sdata, baseClass:$sbase, smem_offset:$offset, CPolTy:$cpol), 254 (ins dataClass:$sdata, baseClass:$sbase, SReg_32:$offset, CPolTy:$cpol)), 255 !if(isRet, " $sdst", " $sdata") # ", $sbase, $offset$cpol", 256 isRet>, 257 AtomicNoRet <opNameWithSuffix, isRet> { 258 let offset_is_imm = isImm; 259 let PseudoInstr = opNameWithSuffix; 260 261 let Constraints = !if(isRet, "$sdst = $sdata", ""); 262 let DisableEncoding = !if(isRet, "$sdata", ""); 263} 264 265multiclass SM_Pseudo_Atomics<string opName, 266 RegisterClass baseClass, 267 RegisterClass dataClass> { 268 def _IMM : SM_Pseudo_Atomic <opName, baseClass, dataClass, 1, 0>; 269 def _SGPR : SM_Pseudo_Atomic <opName, baseClass, dataClass, 0, 0>; 270 def _IMM_RTN : SM_Pseudo_Atomic <opName, baseClass, dataClass, 1, 1>; 271 def _SGPR_RTN : SM_Pseudo_Atomic <opName, baseClass, dataClass, 0, 1>; 272} 273 274//===----------------------------------------------------------------------===// 275// Scalar Memory Instructions 276//===----------------------------------------------------------------------===// 277 278// We are using the SReg_32_XM0 and not the SReg_32 register class for 32-bit 279// SMRD instructions, because the SReg_32_XM0 register class does not include M0 280// and writing to M0 from an SMRD instruction will hang the GPU. 281 282// XXX - SMEM instructions do not allow exec for data operand, but 283// does sdst for SMRD on SI/CI? 284defm S_LOAD_DWORD : SM_Pseudo_Loads <"s_load_dword", SReg_64, SReg_32_XM0_XEXEC>; 285defm S_LOAD_DWORDX2 : SM_Pseudo_Loads <"s_load_dwordx2", SReg_64, SReg_64_XEXEC>; 286defm S_LOAD_DWORDX4 : SM_Pseudo_Loads <"s_load_dwordx4", SReg_64, SReg_128>; 287defm S_LOAD_DWORDX8 : SM_Pseudo_Loads <"s_load_dwordx8", SReg_64, SReg_256>; 288defm S_LOAD_DWORDX16 : SM_Pseudo_Loads <"s_load_dwordx16", SReg_64, SReg_512>; 289 290let is_buffer = 1 in { 291defm S_BUFFER_LOAD_DWORD : SM_Pseudo_Loads < 292 "s_buffer_load_dword", SReg_128, SReg_32_XM0_XEXEC 293>; 294 295// FIXME: exec_lo/exec_hi appear to be allowed for SMRD loads on 296// SI/CI, bit disallowed for SMEM on VI. 297defm S_BUFFER_LOAD_DWORDX2 : SM_Pseudo_Loads < 298 "s_buffer_load_dwordx2", SReg_128, SReg_64_XEXEC 299>; 300 301defm S_BUFFER_LOAD_DWORDX4 : SM_Pseudo_Loads < 302 "s_buffer_load_dwordx4", SReg_128, SReg_128 303>; 304 305defm S_BUFFER_LOAD_DWORDX8 : SM_Pseudo_Loads < 306 "s_buffer_load_dwordx8", SReg_128, SReg_256 307>; 308 309defm S_BUFFER_LOAD_DWORDX16 : SM_Pseudo_Loads < 310 "s_buffer_load_dwordx16", SReg_128, SReg_512 311>; 312} 313 314let SubtargetPredicate = HasScalarStores in { 315defm S_STORE_DWORD : SM_Pseudo_Stores <"s_store_dword", SReg_64, SReg_32_XM0_XEXEC>; 316defm S_STORE_DWORDX2 : SM_Pseudo_Stores <"s_store_dwordx2", SReg_64, SReg_64_XEXEC>; 317defm S_STORE_DWORDX4 : SM_Pseudo_Stores <"s_store_dwordx4", SReg_64, SReg_128>; 318 319let is_buffer = 1 in { 320defm S_BUFFER_STORE_DWORD : SM_Pseudo_Stores < 321 "s_buffer_store_dword", SReg_128, SReg_32_XM0_XEXEC 322>; 323 324defm S_BUFFER_STORE_DWORDX2 : SM_Pseudo_Stores < 325 "s_buffer_store_dwordx2", SReg_128, SReg_64_XEXEC 326>; 327 328defm S_BUFFER_STORE_DWORDX4 : SM_Pseudo_Stores < 329 "s_buffer_store_dwordx4", SReg_128, SReg_128 330>; 331} 332} // End SubtargetPredicate = HasScalarStores 333 334let SubtargetPredicate = HasSMemTimeInst in 335def S_MEMTIME : SM_Time_Pseudo <"s_memtime", int_amdgcn_s_memtime>; 336def S_DCACHE_INV : SM_Inval_Pseudo <"s_dcache_inv", int_amdgcn_s_dcache_inv>; 337 338let SubtargetPredicate = isGFX7GFX8GFX9 in { 339def S_DCACHE_INV_VOL : SM_Inval_Pseudo <"s_dcache_inv_vol", int_amdgcn_s_dcache_inv_vol>; 340} // let SubtargetPredicate = isGFX7GFX8GFX9 341 342let SubtargetPredicate = isGFX8Plus in { 343let OtherPredicates = [HasScalarStores] in { 344def S_DCACHE_WB : SM_Inval_Pseudo <"s_dcache_wb", int_amdgcn_s_dcache_wb>; 345def S_DCACHE_WB_VOL : SM_Inval_Pseudo <"s_dcache_wb_vol", int_amdgcn_s_dcache_wb_vol>; 346} // End OtherPredicates = [HasScalarStores] 347 348defm S_ATC_PROBE : SM_Pseudo_Probe <"s_atc_probe", SReg_64>; 349let is_buffer = 1 in { 350defm S_ATC_PROBE_BUFFER : SM_Pseudo_Probe <"s_atc_probe_buffer", SReg_128>; 351} 352} // SubtargetPredicate = isGFX8Plus 353 354let SubtargetPredicate = HasSMemRealTime in 355def S_MEMREALTIME : SM_Time_Pseudo <"s_memrealtime", int_amdgcn_s_memrealtime>; 356 357let SubtargetPredicate = isGFX10Plus in 358def S_GL1_INV : SM_Inval_Pseudo<"s_gl1_inv">; 359let SubtargetPredicate = HasGetWaveIdInst in 360def S_GET_WAVEID_IN_WORKGROUP : SM_WaveId_Pseudo <"s_get_waveid_in_workgroup", int_amdgcn_s_get_waveid_in_workgroup>; 361 362 363let SubtargetPredicate = HasScalarFlatScratchInsts, Uses = [FLAT_SCR] in { 364defm S_SCRATCH_LOAD_DWORD : SM_Pseudo_Loads <"s_scratch_load_dword", SReg_64, SReg_32_XM0_XEXEC>; 365defm S_SCRATCH_LOAD_DWORDX2 : SM_Pseudo_Loads <"s_scratch_load_dwordx2", SReg_64, SReg_64_XEXEC>; 366defm S_SCRATCH_LOAD_DWORDX4 : SM_Pseudo_Loads <"s_scratch_load_dwordx4", SReg_64, SReg_128>; 367 368defm S_SCRATCH_STORE_DWORD : SM_Pseudo_Stores <"s_scratch_store_dword", SReg_64, SReg_32_XM0_XEXEC>; 369defm S_SCRATCH_STORE_DWORDX2 : SM_Pseudo_Stores <"s_scratch_store_dwordx2", SReg_64, SReg_64_XEXEC>; 370defm S_SCRATCH_STORE_DWORDX4 : SM_Pseudo_Stores <"s_scratch_store_dwordx4", SReg_64, SReg_128>; 371} // SubtargetPredicate = HasScalarFlatScratchInsts 372 373let SubtargetPredicate = HasScalarAtomics in { 374 375let is_buffer = 1 in { 376defm S_BUFFER_ATOMIC_SWAP : SM_Pseudo_Atomics <"s_buffer_atomic_swap", SReg_128, SReg_32_XM0_XEXEC>; 377defm S_BUFFER_ATOMIC_CMPSWAP : SM_Pseudo_Atomics <"s_buffer_atomic_cmpswap", SReg_128, SReg_64_XEXEC>; 378defm S_BUFFER_ATOMIC_ADD : SM_Pseudo_Atomics <"s_buffer_atomic_add", SReg_128, SReg_32_XM0_XEXEC>; 379defm S_BUFFER_ATOMIC_SUB : SM_Pseudo_Atomics <"s_buffer_atomic_sub", SReg_128, SReg_32_XM0_XEXEC>; 380defm S_BUFFER_ATOMIC_SMIN : SM_Pseudo_Atomics <"s_buffer_atomic_smin", SReg_128, SReg_32_XM0_XEXEC>; 381defm S_BUFFER_ATOMIC_UMIN : SM_Pseudo_Atomics <"s_buffer_atomic_umin", SReg_128, SReg_32_XM0_XEXEC>; 382defm S_BUFFER_ATOMIC_SMAX : SM_Pseudo_Atomics <"s_buffer_atomic_smax", SReg_128, SReg_32_XM0_XEXEC>; 383defm S_BUFFER_ATOMIC_UMAX : SM_Pseudo_Atomics <"s_buffer_atomic_umax", SReg_128, SReg_32_XM0_XEXEC>; 384defm S_BUFFER_ATOMIC_AND : SM_Pseudo_Atomics <"s_buffer_atomic_and", SReg_128, SReg_32_XM0_XEXEC>; 385defm S_BUFFER_ATOMIC_OR : SM_Pseudo_Atomics <"s_buffer_atomic_or", SReg_128, SReg_32_XM0_XEXEC>; 386defm S_BUFFER_ATOMIC_XOR : SM_Pseudo_Atomics <"s_buffer_atomic_xor", SReg_128, SReg_32_XM0_XEXEC>; 387defm S_BUFFER_ATOMIC_INC : SM_Pseudo_Atomics <"s_buffer_atomic_inc", SReg_128, SReg_32_XM0_XEXEC>; 388defm S_BUFFER_ATOMIC_DEC : SM_Pseudo_Atomics <"s_buffer_atomic_dec", SReg_128, SReg_32_XM0_XEXEC>; 389 390defm S_BUFFER_ATOMIC_SWAP_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_swap_x2", SReg_128, SReg_64_XEXEC>; 391defm S_BUFFER_ATOMIC_CMPSWAP_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_cmpswap_x2", SReg_128, SReg_128>; 392defm S_BUFFER_ATOMIC_ADD_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_add_x2", SReg_128, SReg_64_XEXEC>; 393defm S_BUFFER_ATOMIC_SUB_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_sub_x2", SReg_128, SReg_64_XEXEC>; 394defm S_BUFFER_ATOMIC_SMIN_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_smin_x2", SReg_128, SReg_64_XEXEC>; 395defm S_BUFFER_ATOMIC_UMIN_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_umin_x2", SReg_128, SReg_64_XEXEC>; 396defm S_BUFFER_ATOMIC_SMAX_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_smax_x2", SReg_128, SReg_64_XEXEC>; 397defm S_BUFFER_ATOMIC_UMAX_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_umax_x2", SReg_128, SReg_64_XEXEC>; 398defm S_BUFFER_ATOMIC_AND_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_and_x2", SReg_128, SReg_64_XEXEC>; 399defm S_BUFFER_ATOMIC_OR_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_or_x2", SReg_128, SReg_64_XEXEC>; 400defm S_BUFFER_ATOMIC_XOR_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_xor_x2", SReg_128, SReg_64_XEXEC>; 401defm S_BUFFER_ATOMIC_INC_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_inc_x2", SReg_128, SReg_64_XEXEC>; 402defm S_BUFFER_ATOMIC_DEC_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_dec_x2", SReg_128, SReg_64_XEXEC>; 403} 404 405defm S_ATOMIC_SWAP : SM_Pseudo_Atomics <"s_atomic_swap", SReg_64, SReg_32_XM0_XEXEC>; 406defm S_ATOMIC_CMPSWAP : SM_Pseudo_Atomics <"s_atomic_cmpswap", SReg_64, SReg_64_XEXEC>; 407defm S_ATOMIC_ADD : SM_Pseudo_Atomics <"s_atomic_add", SReg_64, SReg_32_XM0_XEXEC>; 408defm S_ATOMIC_SUB : SM_Pseudo_Atomics <"s_atomic_sub", SReg_64, SReg_32_XM0_XEXEC>; 409defm S_ATOMIC_SMIN : SM_Pseudo_Atomics <"s_atomic_smin", SReg_64, SReg_32_XM0_XEXEC>; 410defm S_ATOMIC_UMIN : SM_Pseudo_Atomics <"s_atomic_umin", SReg_64, SReg_32_XM0_XEXEC>; 411defm S_ATOMIC_SMAX : SM_Pseudo_Atomics <"s_atomic_smax", SReg_64, SReg_32_XM0_XEXEC>; 412defm S_ATOMIC_UMAX : SM_Pseudo_Atomics <"s_atomic_umax", SReg_64, SReg_32_XM0_XEXEC>; 413defm S_ATOMIC_AND : SM_Pseudo_Atomics <"s_atomic_and", SReg_64, SReg_32_XM0_XEXEC>; 414defm S_ATOMIC_OR : SM_Pseudo_Atomics <"s_atomic_or", SReg_64, SReg_32_XM0_XEXEC>; 415defm S_ATOMIC_XOR : SM_Pseudo_Atomics <"s_atomic_xor", SReg_64, SReg_32_XM0_XEXEC>; 416defm S_ATOMIC_INC : SM_Pseudo_Atomics <"s_atomic_inc", SReg_64, SReg_32_XM0_XEXEC>; 417defm S_ATOMIC_DEC : SM_Pseudo_Atomics <"s_atomic_dec", SReg_64, SReg_32_XM0_XEXEC>; 418 419defm S_ATOMIC_SWAP_X2 : SM_Pseudo_Atomics <"s_atomic_swap_x2", SReg_64, SReg_64_XEXEC>; 420defm S_ATOMIC_CMPSWAP_X2 : SM_Pseudo_Atomics <"s_atomic_cmpswap_x2", SReg_64, SReg_128>; 421defm S_ATOMIC_ADD_X2 : SM_Pseudo_Atomics <"s_atomic_add_x2", SReg_64, SReg_64_XEXEC>; 422defm S_ATOMIC_SUB_X2 : SM_Pseudo_Atomics <"s_atomic_sub_x2", SReg_64, SReg_64_XEXEC>; 423defm S_ATOMIC_SMIN_X2 : SM_Pseudo_Atomics <"s_atomic_smin_x2", SReg_64, SReg_64_XEXEC>; 424defm S_ATOMIC_UMIN_X2 : SM_Pseudo_Atomics <"s_atomic_umin_x2", SReg_64, SReg_64_XEXEC>; 425defm S_ATOMIC_SMAX_X2 : SM_Pseudo_Atomics <"s_atomic_smax_x2", SReg_64, SReg_64_XEXEC>; 426defm S_ATOMIC_UMAX_X2 : SM_Pseudo_Atomics <"s_atomic_umax_x2", SReg_64, SReg_64_XEXEC>; 427defm S_ATOMIC_AND_X2 : SM_Pseudo_Atomics <"s_atomic_and_x2", SReg_64, SReg_64_XEXEC>; 428defm S_ATOMIC_OR_X2 : SM_Pseudo_Atomics <"s_atomic_or_x2", SReg_64, SReg_64_XEXEC>; 429defm S_ATOMIC_XOR_X2 : SM_Pseudo_Atomics <"s_atomic_xor_x2", SReg_64, SReg_64_XEXEC>; 430defm S_ATOMIC_INC_X2 : SM_Pseudo_Atomics <"s_atomic_inc_x2", SReg_64, SReg_64_XEXEC>; 431defm S_ATOMIC_DEC_X2 : SM_Pseudo_Atomics <"s_atomic_dec_x2", SReg_64, SReg_64_XEXEC>; 432 433} // let SubtargetPredicate = HasScalarAtomics 434 435let SubtargetPredicate = HasScalarAtomics in { 436defm S_DCACHE_DISCARD : SM_Pseudo_Discards <"s_dcache_discard">; 437defm S_DCACHE_DISCARD_X2 : SM_Pseudo_Discards <"s_dcache_discard_x2">; 438} 439 440//===----------------------------------------------------------------------===// 441// Targets 442//===----------------------------------------------------------------------===// 443 444//===----------------------------------------------------------------------===// 445// SI 446//===----------------------------------------------------------------------===// 447 448class SMRD_Real_si <bits<5> op, SM_Pseudo ps> 449 : SM_Real<ps> 450 , SIMCInstr<ps.PseudoInstr, SIEncodingFamily.SI> 451 , Enc32 { 452 453 let AssemblerPredicate = isGFX6GFX7; 454 let DecoderNamespace = "GFX6GFX7"; 455 456 let Inst{7-0} = !if(ps.has_offset, offset{7-0}, ?); 457 let Inst{8} = imm; 458 let Inst{14-9} = !if(ps.has_sbase, sbase{6-1}, ?); 459 let Inst{21-15} = !if(ps.has_sdst, sdst{6-0}, ?); 460 let Inst{26-22} = op; 461 let Inst{31-27} = 0x18; //encoding 462} 463 464// FIXME: Assembler should reject trying to use glc on SMRD 465// instructions on SI. 466multiclass SM_Real_Loads_si<bits<5> op, string ps, 467 SM_Load_Pseudo immPs = !cast<SM_Load_Pseudo>(ps#_IMM), 468 SM_Load_Pseudo sgprPs = !cast<SM_Load_Pseudo>(ps#_SGPR)> { 469 470 def _IMM_si : SMRD_Real_si <op, immPs> { 471 let InOperandList = (ins immPs.BaseClass:$sbase, smrd_offset_8:$offset, CPol:$cpol); 472 } 473 474 // FIXME: The operand name $offset is inconsistent with $soff used 475 // in the pseudo 476 def _SGPR_si : SMRD_Real_si <op, sgprPs> { 477 let InOperandList = (ins sgprPs.BaseClass:$sbase, SReg_32:$offset, CPol:$cpol); 478 } 479 480} 481 482defm S_LOAD_DWORD : SM_Real_Loads_si <0x00, "S_LOAD_DWORD">; 483defm S_LOAD_DWORDX2 : SM_Real_Loads_si <0x01, "S_LOAD_DWORDX2">; 484defm S_LOAD_DWORDX4 : SM_Real_Loads_si <0x02, "S_LOAD_DWORDX4">; 485defm S_LOAD_DWORDX8 : SM_Real_Loads_si <0x03, "S_LOAD_DWORDX8">; 486defm S_LOAD_DWORDX16 : SM_Real_Loads_si <0x04, "S_LOAD_DWORDX16">; 487defm S_BUFFER_LOAD_DWORD : SM_Real_Loads_si <0x08, "S_BUFFER_LOAD_DWORD">; 488defm S_BUFFER_LOAD_DWORDX2 : SM_Real_Loads_si <0x09, "S_BUFFER_LOAD_DWORDX2">; 489defm S_BUFFER_LOAD_DWORDX4 : SM_Real_Loads_si <0x0a, "S_BUFFER_LOAD_DWORDX4">; 490defm S_BUFFER_LOAD_DWORDX8 : SM_Real_Loads_si <0x0b, "S_BUFFER_LOAD_DWORDX8">; 491defm S_BUFFER_LOAD_DWORDX16 : SM_Real_Loads_si <0x0c, "S_BUFFER_LOAD_DWORDX16">; 492 493def S_MEMTIME_si : SMRD_Real_si <0x1e, S_MEMTIME>; 494def S_DCACHE_INV_si : SMRD_Real_si <0x1f, S_DCACHE_INV>; 495 496 497//===----------------------------------------------------------------------===// 498// VI 499//===----------------------------------------------------------------------===// 500 501class SMEM_Real_vi <bits<8> op, SM_Pseudo ps> 502 : SM_Real<ps> 503 , SIMCInstr<ps.PseudoInstr, SIEncodingFamily.VI> 504 , Enc64 { 505 let AssemblerPredicate = isGFX8GFX9; 506 let DecoderNamespace = "GFX8"; 507 508 let Inst{5-0} = !if(ps.has_sbase, sbase{6-1}, ?); 509 let Inst{12-6} = !if(ps.has_sdst, sdst{6-0}, ?); 510 511 let Inst{16} = !if(ps.has_glc, cpol{CPolBit.GLC}, ?); 512 let Inst{17} = imm; 513 let Inst{25-18} = op; 514 let Inst{31-26} = 0x30; //encoding 515 516 // VI supports 20-bit unsigned offsets while GFX9+ supports 21-bit signed. 517 // Offset value is corrected accordingly when offset is encoded/decoded. 518 let Inst{38-32} = !if(ps.has_offset, offset{6-0}, ?); 519 let Inst{52-39} = !if(ps.has_offset, !if(imm, offset{20-7}, ?), ?); 520} 521 522multiclass SM_Real_Loads_vi<bits<8> op, string ps, 523 SM_Load_Pseudo immPs = !cast<SM_Load_Pseudo>(ps#_IMM), 524 SM_Load_Pseudo sgprPs = !cast<SM_Load_Pseudo>(ps#_SGPR)> { 525 def _IMM_vi : SMEM_Real_vi <op, immPs> { 526 let InOperandList = (ins immPs.BaseClass:$sbase, smem_offset:$offset, CPol:$cpol); 527 } 528 def _SGPR_vi : SMEM_Real_vi <op, sgprPs> { 529 let InOperandList = (ins sgprPs.BaseClass:$sbase, SReg_32:$offset, CPol:$cpol); 530 } 531} 532 533class SMEM_Real_Store_vi <bits<8> op, SM_Pseudo ps> : SMEM_Real_vi <op, ps> { 534 // encoding 535 bits<7> sdata; 536 537 let sdst = ?; 538 let Inst{12-6} = !if(ps.has_sdst, sdata{6-0}, ?); 539} 540 541multiclass SM_Real_Stores_vi<bits<8> op, string ps, 542 SM_Store_Pseudo immPs = !cast<SM_Store_Pseudo>(ps#_IMM), 543 SM_Store_Pseudo sgprPs = !cast<SM_Store_Pseudo>(ps#_SGPR)> { 544 // FIXME: The operand name $offset is inconsistent with $soff used 545 // in the pseudo 546 def _IMM_vi : SMEM_Real_Store_vi <op, immPs> { 547 let InOperandList = (ins immPs.SrcClass:$sdata, immPs.BaseClass:$sbase, smem_offset:$offset, CPol:$cpol); 548 } 549 550 def _SGPR_vi : SMEM_Real_Store_vi <op, sgprPs> { 551 let InOperandList = (ins sgprPs.SrcClass:$sdata, sgprPs.BaseClass:$sbase, SReg_32:$offset, CPol:$cpol); 552 } 553} 554 555multiclass SM_Real_Probe_vi<bits<8> op, string ps> { 556 def _IMM_vi : SMEM_Real_Store_vi <op, !cast<SM_Probe_Pseudo>(ps#_IMM)>; 557 def _SGPR_vi : SMEM_Real_Store_vi <op, !cast<SM_Probe_Pseudo>(ps#_SGPR)>; 558} 559 560defm S_LOAD_DWORD : SM_Real_Loads_vi <0x00, "S_LOAD_DWORD">; 561defm S_LOAD_DWORDX2 : SM_Real_Loads_vi <0x01, "S_LOAD_DWORDX2">; 562defm S_LOAD_DWORDX4 : SM_Real_Loads_vi <0x02, "S_LOAD_DWORDX4">; 563defm S_LOAD_DWORDX8 : SM_Real_Loads_vi <0x03, "S_LOAD_DWORDX8">; 564defm S_LOAD_DWORDX16 : SM_Real_Loads_vi <0x04, "S_LOAD_DWORDX16">; 565defm S_BUFFER_LOAD_DWORD : SM_Real_Loads_vi <0x08, "S_BUFFER_LOAD_DWORD">; 566defm S_BUFFER_LOAD_DWORDX2 : SM_Real_Loads_vi <0x09, "S_BUFFER_LOAD_DWORDX2">; 567defm S_BUFFER_LOAD_DWORDX4 : SM_Real_Loads_vi <0x0a, "S_BUFFER_LOAD_DWORDX4">; 568defm S_BUFFER_LOAD_DWORDX8 : SM_Real_Loads_vi <0x0b, "S_BUFFER_LOAD_DWORDX8">; 569defm S_BUFFER_LOAD_DWORDX16 : SM_Real_Loads_vi <0x0c, "S_BUFFER_LOAD_DWORDX16">; 570 571defm S_STORE_DWORD : SM_Real_Stores_vi <0x10, "S_STORE_DWORD">; 572defm S_STORE_DWORDX2 : SM_Real_Stores_vi <0x11, "S_STORE_DWORDX2">; 573defm S_STORE_DWORDX4 : SM_Real_Stores_vi <0x12, "S_STORE_DWORDX4">; 574 575defm S_BUFFER_STORE_DWORD : SM_Real_Stores_vi <0x18, "S_BUFFER_STORE_DWORD">; 576defm S_BUFFER_STORE_DWORDX2 : SM_Real_Stores_vi <0x19, "S_BUFFER_STORE_DWORDX2">; 577defm S_BUFFER_STORE_DWORDX4 : SM_Real_Stores_vi <0x1a, "S_BUFFER_STORE_DWORDX4">; 578 579// These instructions use same encoding 580def S_DCACHE_INV_vi : SMEM_Real_vi <0x20, S_DCACHE_INV>; 581def S_DCACHE_WB_vi : SMEM_Real_vi <0x21, S_DCACHE_WB>; 582def S_DCACHE_INV_VOL_vi : SMEM_Real_vi <0x22, S_DCACHE_INV_VOL>; 583def S_DCACHE_WB_VOL_vi : SMEM_Real_vi <0x23, S_DCACHE_WB_VOL>; 584def S_MEMTIME_vi : SMEM_Real_vi <0x24, S_MEMTIME>; 585def S_MEMREALTIME_vi : SMEM_Real_vi <0x25, S_MEMREALTIME>; 586 587defm S_SCRATCH_LOAD_DWORD : SM_Real_Loads_vi <0x05, "S_SCRATCH_LOAD_DWORD">; 588defm S_SCRATCH_LOAD_DWORDX2 : SM_Real_Loads_vi <0x06, "S_SCRATCH_LOAD_DWORDX2">; 589defm S_SCRATCH_LOAD_DWORDX4 : SM_Real_Loads_vi <0x07, "S_SCRATCH_LOAD_DWORDX4">; 590 591defm S_SCRATCH_STORE_DWORD : SM_Real_Stores_vi <0x15, "S_SCRATCH_STORE_DWORD">; 592defm S_SCRATCH_STORE_DWORDX2 : SM_Real_Stores_vi <0x16, "S_SCRATCH_STORE_DWORDX2">; 593defm S_SCRATCH_STORE_DWORDX4 : SM_Real_Stores_vi <0x17, "S_SCRATCH_STORE_DWORDX4">; 594 595defm S_ATC_PROBE : SM_Real_Probe_vi <0x26, "S_ATC_PROBE">; 596defm S_ATC_PROBE_BUFFER : SM_Real_Probe_vi <0x27, "S_ATC_PROBE_BUFFER">; 597 598//===----------------------------------------------------------------------===// 599// GFX9 600//===----------------------------------------------------------------------===// 601 602class SMEM_Atomic_Real_vi <bits<8> op, SM_Atomic_Pseudo ps> 603 : SMEM_Real_vi <op, ps>, 604 AtomicNoRet <!subst("_RTN","",NAME), ps.glc> { 605 606 bits<7> sdata; 607 608 let Constraints = ps.Constraints; 609 let DisableEncoding = ps.DisableEncoding; 610 611 let cpol{CPolBit.GLC} = ps.glc; 612 let Inst{12-6} = !if(ps.glc, sdst{6-0}, sdata{6-0}); 613} 614 615multiclass SM_Real_Atomics_vi<bits<8> op, string ps> { 616 def _IMM_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_IMM)>; 617 def _SGPR_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR)>; 618 def _IMM_RTN_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_IMM_RTN)>; 619 def _SGPR_RTN_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR_RTN)>; 620} 621 622defm S_BUFFER_ATOMIC_SWAP : SM_Real_Atomics_vi <0x40, "S_BUFFER_ATOMIC_SWAP">; 623defm S_BUFFER_ATOMIC_CMPSWAP : SM_Real_Atomics_vi <0x41, "S_BUFFER_ATOMIC_CMPSWAP">; 624defm S_BUFFER_ATOMIC_ADD : SM_Real_Atomics_vi <0x42, "S_BUFFER_ATOMIC_ADD">; 625defm S_BUFFER_ATOMIC_SUB : SM_Real_Atomics_vi <0x43, "S_BUFFER_ATOMIC_SUB">; 626defm S_BUFFER_ATOMIC_SMIN : SM_Real_Atomics_vi <0x44, "S_BUFFER_ATOMIC_SMIN">; 627defm S_BUFFER_ATOMIC_UMIN : SM_Real_Atomics_vi <0x45, "S_BUFFER_ATOMIC_UMIN">; 628defm S_BUFFER_ATOMIC_SMAX : SM_Real_Atomics_vi <0x46, "S_BUFFER_ATOMIC_SMAX">; 629defm S_BUFFER_ATOMIC_UMAX : SM_Real_Atomics_vi <0x47, "S_BUFFER_ATOMIC_UMAX">; 630defm S_BUFFER_ATOMIC_AND : SM_Real_Atomics_vi <0x48, "S_BUFFER_ATOMIC_AND">; 631defm S_BUFFER_ATOMIC_OR : SM_Real_Atomics_vi <0x49, "S_BUFFER_ATOMIC_OR">; 632defm S_BUFFER_ATOMIC_XOR : SM_Real_Atomics_vi <0x4a, "S_BUFFER_ATOMIC_XOR">; 633defm S_BUFFER_ATOMIC_INC : SM_Real_Atomics_vi <0x4b, "S_BUFFER_ATOMIC_INC">; 634defm S_BUFFER_ATOMIC_DEC : SM_Real_Atomics_vi <0x4c, "S_BUFFER_ATOMIC_DEC">; 635 636defm S_BUFFER_ATOMIC_SWAP_X2 : SM_Real_Atomics_vi <0x60, "S_BUFFER_ATOMIC_SWAP_X2">; 637defm S_BUFFER_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_vi <0x61, "S_BUFFER_ATOMIC_CMPSWAP_X2">; 638defm S_BUFFER_ATOMIC_ADD_X2 : SM_Real_Atomics_vi <0x62, "S_BUFFER_ATOMIC_ADD_X2">; 639defm S_BUFFER_ATOMIC_SUB_X2 : SM_Real_Atomics_vi <0x63, "S_BUFFER_ATOMIC_SUB_X2">; 640defm S_BUFFER_ATOMIC_SMIN_X2 : SM_Real_Atomics_vi <0x64, "S_BUFFER_ATOMIC_SMIN_X2">; 641defm S_BUFFER_ATOMIC_UMIN_X2 : SM_Real_Atomics_vi <0x65, "S_BUFFER_ATOMIC_UMIN_X2">; 642defm S_BUFFER_ATOMIC_SMAX_X2 : SM_Real_Atomics_vi <0x66, "S_BUFFER_ATOMIC_SMAX_X2">; 643defm S_BUFFER_ATOMIC_UMAX_X2 : SM_Real_Atomics_vi <0x67, "S_BUFFER_ATOMIC_UMAX_X2">; 644defm S_BUFFER_ATOMIC_AND_X2 : SM_Real_Atomics_vi <0x68, "S_BUFFER_ATOMIC_AND_X2">; 645defm S_BUFFER_ATOMIC_OR_X2 : SM_Real_Atomics_vi <0x69, "S_BUFFER_ATOMIC_OR_X2">; 646defm S_BUFFER_ATOMIC_XOR_X2 : SM_Real_Atomics_vi <0x6a, "S_BUFFER_ATOMIC_XOR_X2">; 647defm S_BUFFER_ATOMIC_INC_X2 : SM_Real_Atomics_vi <0x6b, "S_BUFFER_ATOMIC_INC_X2">; 648defm S_BUFFER_ATOMIC_DEC_X2 : SM_Real_Atomics_vi <0x6c, "S_BUFFER_ATOMIC_DEC_X2">; 649 650defm S_ATOMIC_SWAP : SM_Real_Atomics_vi <0x80, "S_ATOMIC_SWAP">; 651defm S_ATOMIC_CMPSWAP : SM_Real_Atomics_vi <0x81, "S_ATOMIC_CMPSWAP">; 652defm S_ATOMIC_ADD : SM_Real_Atomics_vi <0x82, "S_ATOMIC_ADD">; 653defm S_ATOMIC_SUB : SM_Real_Atomics_vi <0x83, "S_ATOMIC_SUB">; 654defm S_ATOMIC_SMIN : SM_Real_Atomics_vi <0x84, "S_ATOMIC_SMIN">; 655defm S_ATOMIC_UMIN : SM_Real_Atomics_vi <0x85, "S_ATOMIC_UMIN">; 656defm S_ATOMIC_SMAX : SM_Real_Atomics_vi <0x86, "S_ATOMIC_SMAX">; 657defm S_ATOMIC_UMAX : SM_Real_Atomics_vi <0x87, "S_ATOMIC_UMAX">; 658defm S_ATOMIC_AND : SM_Real_Atomics_vi <0x88, "S_ATOMIC_AND">; 659defm S_ATOMIC_OR : SM_Real_Atomics_vi <0x89, "S_ATOMIC_OR">; 660defm S_ATOMIC_XOR : SM_Real_Atomics_vi <0x8a, "S_ATOMIC_XOR">; 661defm S_ATOMIC_INC : SM_Real_Atomics_vi <0x8b, "S_ATOMIC_INC">; 662defm S_ATOMIC_DEC : SM_Real_Atomics_vi <0x8c, "S_ATOMIC_DEC">; 663 664defm S_ATOMIC_SWAP_X2 : SM_Real_Atomics_vi <0xa0, "S_ATOMIC_SWAP_X2">; 665defm S_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_vi <0xa1, "S_ATOMIC_CMPSWAP_X2">; 666defm S_ATOMIC_ADD_X2 : SM_Real_Atomics_vi <0xa2, "S_ATOMIC_ADD_X2">; 667defm S_ATOMIC_SUB_X2 : SM_Real_Atomics_vi <0xa3, "S_ATOMIC_SUB_X2">; 668defm S_ATOMIC_SMIN_X2 : SM_Real_Atomics_vi <0xa4, "S_ATOMIC_SMIN_X2">; 669defm S_ATOMIC_UMIN_X2 : SM_Real_Atomics_vi <0xa5, "S_ATOMIC_UMIN_X2">; 670defm S_ATOMIC_SMAX_X2 : SM_Real_Atomics_vi <0xa6, "S_ATOMIC_SMAX_X2">; 671defm S_ATOMIC_UMAX_X2 : SM_Real_Atomics_vi <0xa7, "S_ATOMIC_UMAX_X2">; 672defm S_ATOMIC_AND_X2 : SM_Real_Atomics_vi <0xa8, "S_ATOMIC_AND_X2">; 673defm S_ATOMIC_OR_X2 : SM_Real_Atomics_vi <0xa9, "S_ATOMIC_OR_X2">; 674defm S_ATOMIC_XOR_X2 : SM_Real_Atomics_vi <0xaa, "S_ATOMIC_XOR_X2">; 675defm S_ATOMIC_INC_X2 : SM_Real_Atomics_vi <0xab, "S_ATOMIC_INC_X2">; 676defm S_ATOMIC_DEC_X2 : SM_Real_Atomics_vi <0xac, "S_ATOMIC_DEC_X2">; 677 678multiclass SM_Real_Discard_vi<bits<8> op, string ps> { 679 def _IMM_vi : SMEM_Real_vi <op, !cast<SM_Discard_Pseudo>(ps#_IMM)>; 680 def _SGPR_vi : SMEM_Real_vi <op, !cast<SM_Discard_Pseudo>(ps#_SGPR)>; 681} 682 683defm S_DCACHE_DISCARD : SM_Real_Discard_vi <0x28, "S_DCACHE_DISCARD">; 684defm S_DCACHE_DISCARD_X2 : SM_Real_Discard_vi <0x29, "S_DCACHE_DISCARD_X2">; 685 686//===----------------------------------------------------------------------===// 687// CI 688//===----------------------------------------------------------------------===// 689 690def smrd_literal_offset : NamedOperandU32<"SMRDLiteralOffset", 691 NamedMatchClass<"SMRDLiteralOffset">> { 692 let OperandType = "OPERAND_IMMEDIATE"; 693} 694 695class SMRD_Real_Load_IMM_ci <bits<5> op, SM_Load_Pseudo ps> : 696 SM_Real<ps>, 697 Enc64 { 698 699 let AssemblerPredicate = isGFX7Only; 700 let DecoderNamespace = "GFX7"; 701 let InOperandList = (ins ps.BaseClass:$sbase, smrd_literal_offset:$offset, CPol:$cpol); 702 703 let LGKM_CNT = ps.LGKM_CNT; 704 let mayLoad = ps.mayLoad; 705 let mayStore = ps.mayStore; 706 let hasSideEffects = ps.hasSideEffects; 707 let SchedRW = ps.SchedRW; 708 709 let Inst{7-0} = 0xff; 710 let Inst{8} = 0; 711 let Inst{14-9} = sbase{6-1}; 712 let Inst{21-15} = sdst{6-0}; 713 let Inst{26-22} = op; 714 let Inst{31-27} = 0x18; //encoding 715 let Inst{63-32} = offset{31-0}; 716} 717 718def S_LOAD_DWORD_IMM_ci : SMRD_Real_Load_IMM_ci <0x00, S_LOAD_DWORD_IMM>; 719def S_LOAD_DWORDX2_IMM_ci : SMRD_Real_Load_IMM_ci <0x01, S_LOAD_DWORDX2_IMM>; 720def S_LOAD_DWORDX4_IMM_ci : SMRD_Real_Load_IMM_ci <0x02, S_LOAD_DWORDX4_IMM>; 721def S_LOAD_DWORDX8_IMM_ci : SMRD_Real_Load_IMM_ci <0x03, S_LOAD_DWORDX8_IMM>; 722def S_LOAD_DWORDX16_IMM_ci : SMRD_Real_Load_IMM_ci <0x04, S_LOAD_DWORDX16_IMM>; 723def S_BUFFER_LOAD_DWORD_IMM_ci : SMRD_Real_Load_IMM_ci <0x08, S_BUFFER_LOAD_DWORD_IMM>; 724def S_BUFFER_LOAD_DWORDX2_IMM_ci : SMRD_Real_Load_IMM_ci <0x09, S_BUFFER_LOAD_DWORDX2_IMM>; 725def S_BUFFER_LOAD_DWORDX4_IMM_ci : SMRD_Real_Load_IMM_ci <0x0a, S_BUFFER_LOAD_DWORDX4_IMM>; 726def S_BUFFER_LOAD_DWORDX8_IMM_ci : SMRD_Real_Load_IMM_ci <0x0b, S_BUFFER_LOAD_DWORDX8_IMM>; 727def S_BUFFER_LOAD_DWORDX16_IMM_ci : SMRD_Real_Load_IMM_ci <0x0c, S_BUFFER_LOAD_DWORDX16_IMM>; 728 729class SMRD_Real_ci <bits<5> op, SM_Pseudo ps> 730 : SM_Real<ps> 731 , SIMCInstr<ps.PseudoInstr, SIEncodingFamily.SI> 732 , Enc32 { 733 734 let AssemblerPredicate = isGFX7Only; 735 let DecoderNamespace = "GFX7"; 736 737 let Inst{7-0} = !if(ps.has_offset, offset{7-0}, ?); 738 let Inst{8} = imm; 739 let Inst{14-9} = !if(ps.has_sbase, sbase{6-1}, ?); 740 let Inst{21-15} = !if(ps.has_sdst, sdst{6-0}, ?); 741 let Inst{26-22} = op; 742 let Inst{31-27} = 0x18; //encoding 743} 744 745def S_DCACHE_INV_VOL_ci : SMRD_Real_ci <0x1d, S_DCACHE_INV_VOL>; 746 747//===----------------------------------------------------------------------===// 748// Scalar Memory Patterns 749//===----------------------------------------------------------------------===// 750 751def smrd_load : PatFrag <(ops node:$ptr), (load node:$ptr), [{ return isUniformLoad(N);}]> { 752 let GISelPredicateCode = [{ 753 if (!MI.hasOneMemOperand()) 754 return false; 755 if (!isInstrUniform(MI)) 756 return false; 757 758 // FIXME: We should probably be caching this. 759 SmallVector<GEPInfo, 4> AddrInfo; 760 getAddrModeInfo(MI, MRI, AddrInfo); 761 762 if (hasVgprParts(AddrInfo)) 763 return false; 764 return true; 765 }]; 766} 767 768def SMRDImm : ComplexPattern<i64, 2, "SelectSMRDImm">; 769def SMRDImm32 : ComplexPattern<i64, 2, "SelectSMRDImm32">; 770def SMRDSgpr : ComplexPattern<i64, 2, "SelectSMRDSgpr">; 771def SMRDBufferImm : ComplexPattern<i32, 1, "SelectSMRDBufferImm">; 772def SMRDBufferImm32 : ComplexPattern<i32, 1, "SelectSMRDBufferImm32">; 773 774multiclass SMRD_Pattern <string Instr, ValueType vt> { 775 776 // 1. IMM offset 777 def : GCNPat < 778 (smrd_load (SMRDImm i64:$sbase, i32:$offset)), 779 (vt (!cast<SM_Pseudo>(Instr#"_IMM") $sbase, $offset, 0)) 780 >; 781 782 // 2. 32-bit IMM offset on CI 783 def : GCNPat < 784 (smrd_load (SMRDImm32 i64:$sbase, i32:$offset)), 785 (vt (!cast<InstSI>(Instr#"_IMM_ci") $sbase, $offset, 0))> { 786 let OtherPredicates = [isGFX7Only]; 787 } 788 789 // 3. SGPR offset 790 def : GCNPat < 791 (smrd_load (SMRDSgpr i64:$sbase, i32:$offset)), 792 (vt (!cast<SM_Pseudo>(Instr#"_SGPR") $sbase, $offset, 0)) 793 >; 794 795 // 4. No offset 796 def : GCNPat < 797 (vt (smrd_load (i64 SReg_64:$sbase))), 798 (vt (!cast<SM_Pseudo>(Instr#"_IMM") i64:$sbase, 0, 0)) 799 >; 800} 801 802multiclass SMLoad_Pattern <string Instr, ValueType vt> { 803 // 1. Offset as an immediate 804 def : GCNPat < 805 (SIsbuffer_load v4i32:$sbase, (SMRDBufferImm i32:$offset), timm:$cachepolicy), 806 (vt (!cast<SM_Pseudo>(Instr#"_IMM") SReg_128:$sbase, i32imm:$offset, (extract_cpol $cachepolicy)))> { 807 let AddedComplexity = 2; 808 } 809 810 // 2. 32-bit IMM offset on CI 811 def : GCNPat < 812 (vt (SIsbuffer_load v4i32:$sbase, (SMRDBufferImm32 i32:$offset), timm:$cachepolicy)), 813 (!cast<InstSI>(Instr#"_IMM_ci") SReg_128:$sbase, smrd_literal_offset:$offset, 814 (extract_cpol $cachepolicy))> { 815 let OtherPredicates = [isGFX7Only]; 816 let AddedComplexity = 1; 817 } 818 819 // 3. Offset loaded in an 32bit SGPR 820 def : GCNPat < 821 (SIsbuffer_load v4i32:$sbase, i32:$offset, timm:$cachepolicy), 822 (vt (!cast<SM_Pseudo>(Instr#"_SGPR") SReg_128:$sbase, SReg_32:$offset, (extract_cpol $cachepolicy))) 823 >; 824} 825 826// Global and constant loads can be selected to either MUBUF or SMRD 827// instructions, but SMRD instructions are faster so we want the instruction 828// selector to prefer those. 829let AddedComplexity = 100 in { 830 831foreach vt = Reg32Types.types in { 832defm : SMRD_Pattern <"S_LOAD_DWORD", vt>; 833} 834 835foreach vt = SReg_64.RegTypes in { 836defm : SMRD_Pattern <"S_LOAD_DWORDX2", vt>; 837} 838 839foreach vt = SReg_128.RegTypes in { 840defm : SMRD_Pattern <"S_LOAD_DWORDX4", vt>; 841} 842 843foreach vt = SReg_256.RegTypes in { 844defm : SMRD_Pattern <"S_LOAD_DWORDX8", vt>; 845} 846 847foreach vt = SReg_512.RegTypes in { 848defm : SMRD_Pattern <"S_LOAD_DWORDX16", vt>; 849} 850 851defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORD", i32>; 852defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX2", v2i32>; 853defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX4", v4i32>; 854defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX8", v8i32>; 855defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX16", v16i32>; 856 857defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORD", f32>; 858defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX2", v2f32>; 859defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX4", v4f32>; 860defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX8", v8f32>; 861defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX16", v16f32>; 862} // End let AddedComplexity = 100 863 864let OtherPredicates = [HasSMemTimeInst] in { 865def : GCNPat < 866 (i64 (readcyclecounter)), 867 (S_MEMTIME) 868>; 869} // let OtherPredicates = [HasSMemTimeInst] 870 871let OtherPredicates = [HasShaderCyclesRegister] in { 872def : GCNPat < 873 (i64 (readcyclecounter)), 874 (REG_SEQUENCE SReg_64, 875 (S_GETREG_B32 getHwRegImm<HWREG.SHADER_CYCLES, 0, -12>.ret), sub0, 876 (S_MOV_B32 (i32 0)), sub1)> { 877 // Prefer this to s_memtime because it has lower and more predictable latency. 878 let AddedComplexity = 1; 879} 880} // let OtherPredicates = [HasShaderCyclesRegister] 881 882//===----------------------------------------------------------------------===// 883// GFX10. 884//===----------------------------------------------------------------------===// 885 886class SMEM_Real_gfx10<bits<8> op, SM_Pseudo ps> : 887 SM_Real<ps>, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.GFX10>, Enc64 { 888 let AssemblerPredicate = isGFX10Plus; 889 let DecoderNamespace = "GFX10"; 890 891 let Inst{5-0} = !if(ps.has_sbase, sbase{6-1}, ?); 892 let Inst{12-6} = !if(ps.has_sdst, sdst{6-0}, ?); 893 let Inst{14} = !if(ps.has_dlc, cpol{CPolBit.DLC}, ?); 894 let Inst{16} = !if(ps.has_glc, cpol{CPolBit.GLC}, ?); 895 let Inst{25-18} = op; 896 let Inst{31-26} = 0x3d; 897 let Inst{52-32} = !if(ps.offset_is_imm, !if(ps.has_offset, offset{20-0}, ?), ?); 898 let Inst{63-57} = !if(ps.offset_is_imm, !cast<int>(SGPR_NULL.HWEncoding), 899 !if(ps.has_offset, offset{6-0}, ?)); 900} 901 902multiclass SM_Real_Loads_gfx10<bits<8> op, string ps, 903 SM_Load_Pseudo immPs = !cast<SM_Load_Pseudo>(ps#_IMM), 904 SM_Load_Pseudo sgprPs = !cast<SM_Load_Pseudo>(ps#_SGPR)> { 905 def _IMM_gfx10 : SMEM_Real_gfx10<op, immPs> { 906 let InOperandList = (ins immPs.BaseClass:$sbase, smem_offset:$offset, CPol:$cpol); 907 } 908 def _SGPR_gfx10 : SMEM_Real_gfx10<op, sgprPs> { 909 let InOperandList = (ins sgprPs.BaseClass:$sbase, SReg_32:$offset, CPol:$cpol); 910 } 911} 912 913class SMEM_Real_Store_gfx10<bits<8> op, SM_Pseudo ps> : SMEM_Real_gfx10<op, ps> { 914 bits<7> sdata; 915 916 let sdst = ?; 917 let Inst{12-6} = !if(ps.has_sdst, sdata{6-0}, ?); 918} 919 920multiclass SM_Real_Stores_gfx10<bits<8> op, string ps, 921 SM_Store_Pseudo immPs = !cast<SM_Store_Pseudo>(ps#_IMM), 922 SM_Store_Pseudo sgprPs = !cast<SM_Store_Pseudo>(ps#_SGPR)> { 923 // FIXME: The operand name $offset is inconsistent with $soff used 924 // in the pseudo 925 def _IMM_gfx10 : SMEM_Real_Store_gfx10 <op, immPs> { 926 let InOperandList = (ins immPs.SrcClass:$sdata, immPs.BaseClass:$sbase, smem_offset:$offset, CPol:$cpol); 927 } 928 929 def _SGPR_gfx10 : SMEM_Real_Store_gfx10 <op, sgprPs> { 930 let InOperandList = (ins sgprPs.SrcClass:$sdata, sgprPs.BaseClass:$sbase, SReg_32:$offset, CPol:$cpol); 931 } 932} 933 934defm S_LOAD_DWORD : SM_Real_Loads_gfx10<0x000, "S_LOAD_DWORD">; 935defm S_LOAD_DWORDX2 : SM_Real_Loads_gfx10<0x001, "S_LOAD_DWORDX2">; 936defm S_LOAD_DWORDX4 : SM_Real_Loads_gfx10<0x002, "S_LOAD_DWORDX4">; 937defm S_LOAD_DWORDX8 : SM_Real_Loads_gfx10<0x003, "S_LOAD_DWORDX8">; 938defm S_LOAD_DWORDX16 : SM_Real_Loads_gfx10<0x004, "S_LOAD_DWORDX16">; 939 940let SubtargetPredicate = HasScalarFlatScratchInsts in { 941defm S_SCRATCH_LOAD_DWORD : SM_Real_Loads_gfx10<0x005, "S_SCRATCH_LOAD_DWORD">; 942defm S_SCRATCH_LOAD_DWORDX2 : SM_Real_Loads_gfx10<0x006, "S_SCRATCH_LOAD_DWORDX2">; 943defm S_SCRATCH_LOAD_DWORDX4 : SM_Real_Loads_gfx10<0x007, "S_SCRATCH_LOAD_DWORDX4">; 944} // End SubtargetPredicate = HasScalarFlatScratchInsts 945 946defm S_BUFFER_LOAD_DWORD : SM_Real_Loads_gfx10<0x008, "S_BUFFER_LOAD_DWORD">; 947defm S_BUFFER_LOAD_DWORDX2 : SM_Real_Loads_gfx10<0x009, "S_BUFFER_LOAD_DWORDX2">; 948defm S_BUFFER_LOAD_DWORDX4 : SM_Real_Loads_gfx10<0x00a, "S_BUFFER_LOAD_DWORDX4">; 949defm S_BUFFER_LOAD_DWORDX8 : SM_Real_Loads_gfx10<0x00b, "S_BUFFER_LOAD_DWORDX8">; 950defm S_BUFFER_LOAD_DWORDX16 : SM_Real_Loads_gfx10<0x00c, "S_BUFFER_LOAD_DWORDX16">; 951 952let SubtargetPredicate = HasScalarStores in { 953defm S_STORE_DWORD : SM_Real_Stores_gfx10<0x010, "S_STORE_DWORD">; 954defm S_STORE_DWORDX2 : SM_Real_Stores_gfx10<0x011, "S_STORE_DWORDX2">; 955defm S_STORE_DWORDX4 : SM_Real_Stores_gfx10<0x012, "S_STORE_DWORDX4">; 956let OtherPredicates = [HasScalarFlatScratchInsts] in { 957defm S_SCRATCH_STORE_DWORD : SM_Real_Stores_gfx10<0x015, "S_SCRATCH_STORE_DWORD">; 958defm S_SCRATCH_STORE_DWORDX2 : SM_Real_Stores_gfx10<0x016, "S_SCRATCH_STORE_DWORDX2">; 959defm S_SCRATCH_STORE_DWORDX4 : SM_Real_Stores_gfx10<0x017, "S_SCRATCH_STORE_DWORDX4">; 960} // End OtherPredicates = [HasScalarFlatScratchInsts] 961defm S_BUFFER_STORE_DWORD : SM_Real_Stores_gfx10<0x018, "S_BUFFER_STORE_DWORD">; 962defm S_BUFFER_STORE_DWORDX2 : SM_Real_Stores_gfx10<0x019, "S_BUFFER_STORE_DWORDX2">; 963defm S_BUFFER_STORE_DWORDX4 : SM_Real_Stores_gfx10<0x01a, "S_BUFFER_STORE_DWORDX4">; 964} // End SubtargetPredicate = HasScalarStores 965 966def S_MEMREALTIME_gfx10 : SMEM_Real_gfx10<0x025, S_MEMREALTIME>; 967def S_MEMTIME_gfx10 : SMEM_Real_gfx10<0x024, S_MEMTIME>; 968def S_GL1_INV_gfx10 : SMEM_Real_gfx10<0x01f, S_GL1_INV>; 969def S_GET_WAVEID_IN_WORKGROUP_gfx10 : SMEM_Real_gfx10<0x02a, S_GET_WAVEID_IN_WORKGROUP>; 970def S_DCACHE_INV_gfx10 : SMEM_Real_gfx10<0x020, S_DCACHE_INV>; 971 972let SubtargetPredicate = HasScalarStores in { 973def S_DCACHE_WB_gfx10 : SMEM_Real_gfx10<0x021, S_DCACHE_WB>; 974} // End SubtargetPredicate = HasScalarStores 975 976multiclass SM_Real_Probe_gfx10<bits<8> op, string ps> { 977 def _IMM_gfx10 : SMEM_Real_Store_gfx10 <op, !cast<SM_Pseudo>(ps#_IMM)>; 978 def _SGPR_gfx10 : SMEM_Real_Store_gfx10 <op, !cast<SM_Pseudo>(ps#_SGPR)>; 979} 980 981defm S_ATC_PROBE : SM_Real_Probe_gfx10 <0x26, "S_ATC_PROBE">; 982defm S_ATC_PROBE_BUFFER : SM_Real_Probe_gfx10 <0x27, "S_ATC_PROBE_BUFFER">; 983 984class SMEM_Atomic_Real_gfx10 <bits<8> op, SM_Atomic_Pseudo ps> 985 : SMEM_Real_gfx10 <op, ps>, 986 AtomicNoRet <!subst("_RTN","",NAME), ps.glc> { 987 988 bits<7> sdata; 989 990 let Constraints = ps.Constraints; 991 let DisableEncoding = ps.DisableEncoding; 992 993 let cpol{CPolBit.GLC} = ps.glc; 994 995 let Inst{14} = !if(ps.has_dlc, cpol{CPolBit.DLC}, 0); 996 let Inst{12-6} = !if(ps.glc, sdst{6-0}, sdata{6-0}); 997} 998 999multiclass SM_Real_Atomics_gfx10<bits<8> op, string ps> { 1000 def _IMM_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_IMM)>; 1001 def _SGPR_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR)>; 1002 def _IMM_RTN_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_IMM_RTN)>; 1003 def _SGPR_RTN_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR_RTN)>; 1004} 1005 1006let SubtargetPredicate = HasScalarAtomics in { 1007 1008defm S_BUFFER_ATOMIC_SWAP : SM_Real_Atomics_gfx10 <0x40, "S_BUFFER_ATOMIC_SWAP">; 1009defm S_BUFFER_ATOMIC_CMPSWAP : SM_Real_Atomics_gfx10 <0x41, "S_BUFFER_ATOMIC_CMPSWAP">; 1010defm S_BUFFER_ATOMIC_ADD : SM_Real_Atomics_gfx10 <0x42, "S_BUFFER_ATOMIC_ADD">; 1011defm S_BUFFER_ATOMIC_SUB : SM_Real_Atomics_gfx10 <0x43, "S_BUFFER_ATOMIC_SUB">; 1012defm S_BUFFER_ATOMIC_SMIN : SM_Real_Atomics_gfx10 <0x44, "S_BUFFER_ATOMIC_SMIN">; 1013defm S_BUFFER_ATOMIC_UMIN : SM_Real_Atomics_gfx10 <0x45, "S_BUFFER_ATOMIC_UMIN">; 1014defm S_BUFFER_ATOMIC_SMAX : SM_Real_Atomics_gfx10 <0x46, "S_BUFFER_ATOMIC_SMAX">; 1015defm S_BUFFER_ATOMIC_UMAX : SM_Real_Atomics_gfx10 <0x47, "S_BUFFER_ATOMIC_UMAX">; 1016defm S_BUFFER_ATOMIC_AND : SM_Real_Atomics_gfx10 <0x48, "S_BUFFER_ATOMIC_AND">; 1017defm S_BUFFER_ATOMIC_OR : SM_Real_Atomics_gfx10 <0x49, "S_BUFFER_ATOMIC_OR">; 1018defm S_BUFFER_ATOMIC_XOR : SM_Real_Atomics_gfx10 <0x4a, "S_BUFFER_ATOMIC_XOR">; 1019defm S_BUFFER_ATOMIC_INC : SM_Real_Atomics_gfx10 <0x4b, "S_BUFFER_ATOMIC_INC">; 1020defm S_BUFFER_ATOMIC_DEC : SM_Real_Atomics_gfx10 <0x4c, "S_BUFFER_ATOMIC_DEC">; 1021 1022defm S_BUFFER_ATOMIC_SWAP_X2 : SM_Real_Atomics_gfx10 <0x60, "S_BUFFER_ATOMIC_SWAP_X2">; 1023defm S_BUFFER_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_gfx10 <0x61, "S_BUFFER_ATOMIC_CMPSWAP_X2">; 1024defm S_BUFFER_ATOMIC_ADD_X2 : SM_Real_Atomics_gfx10 <0x62, "S_BUFFER_ATOMIC_ADD_X2">; 1025defm S_BUFFER_ATOMIC_SUB_X2 : SM_Real_Atomics_gfx10 <0x63, "S_BUFFER_ATOMIC_SUB_X2">; 1026defm S_BUFFER_ATOMIC_SMIN_X2 : SM_Real_Atomics_gfx10 <0x64, "S_BUFFER_ATOMIC_SMIN_X2">; 1027defm S_BUFFER_ATOMIC_UMIN_X2 : SM_Real_Atomics_gfx10 <0x65, "S_BUFFER_ATOMIC_UMIN_X2">; 1028defm S_BUFFER_ATOMIC_SMAX_X2 : SM_Real_Atomics_gfx10 <0x66, "S_BUFFER_ATOMIC_SMAX_X2">; 1029defm S_BUFFER_ATOMIC_UMAX_X2 : SM_Real_Atomics_gfx10 <0x67, "S_BUFFER_ATOMIC_UMAX_X2">; 1030defm S_BUFFER_ATOMIC_AND_X2 : SM_Real_Atomics_gfx10 <0x68, "S_BUFFER_ATOMIC_AND_X2">; 1031defm S_BUFFER_ATOMIC_OR_X2 : SM_Real_Atomics_gfx10 <0x69, "S_BUFFER_ATOMIC_OR_X2">; 1032defm S_BUFFER_ATOMIC_XOR_X2 : SM_Real_Atomics_gfx10 <0x6a, "S_BUFFER_ATOMIC_XOR_X2">; 1033defm S_BUFFER_ATOMIC_INC_X2 : SM_Real_Atomics_gfx10 <0x6b, "S_BUFFER_ATOMIC_INC_X2">; 1034defm S_BUFFER_ATOMIC_DEC_X2 : SM_Real_Atomics_gfx10 <0x6c, "S_BUFFER_ATOMIC_DEC_X2">; 1035 1036defm S_ATOMIC_SWAP : SM_Real_Atomics_gfx10 <0x80, "S_ATOMIC_SWAP">; 1037defm S_ATOMIC_CMPSWAP : SM_Real_Atomics_gfx10 <0x81, "S_ATOMIC_CMPSWAP">; 1038defm S_ATOMIC_ADD : SM_Real_Atomics_gfx10 <0x82, "S_ATOMIC_ADD">; 1039defm S_ATOMIC_SUB : SM_Real_Atomics_gfx10 <0x83, "S_ATOMIC_SUB">; 1040defm S_ATOMIC_SMIN : SM_Real_Atomics_gfx10 <0x84, "S_ATOMIC_SMIN">; 1041defm S_ATOMIC_UMIN : SM_Real_Atomics_gfx10 <0x85, "S_ATOMIC_UMIN">; 1042defm S_ATOMIC_SMAX : SM_Real_Atomics_gfx10 <0x86, "S_ATOMIC_SMAX">; 1043defm S_ATOMIC_UMAX : SM_Real_Atomics_gfx10 <0x87, "S_ATOMIC_UMAX">; 1044defm S_ATOMIC_AND : SM_Real_Atomics_gfx10 <0x88, "S_ATOMIC_AND">; 1045defm S_ATOMIC_OR : SM_Real_Atomics_gfx10 <0x89, "S_ATOMIC_OR">; 1046defm S_ATOMIC_XOR : SM_Real_Atomics_gfx10 <0x8a, "S_ATOMIC_XOR">; 1047defm S_ATOMIC_INC : SM_Real_Atomics_gfx10 <0x8b, "S_ATOMIC_INC">; 1048defm S_ATOMIC_DEC : SM_Real_Atomics_gfx10 <0x8c, "S_ATOMIC_DEC">; 1049 1050defm S_ATOMIC_SWAP_X2 : SM_Real_Atomics_gfx10 <0xa0, "S_ATOMIC_SWAP_X2">; 1051defm S_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_gfx10 <0xa1, "S_ATOMIC_CMPSWAP_X2">; 1052defm S_ATOMIC_ADD_X2 : SM_Real_Atomics_gfx10 <0xa2, "S_ATOMIC_ADD_X2">; 1053defm S_ATOMIC_SUB_X2 : SM_Real_Atomics_gfx10 <0xa3, "S_ATOMIC_SUB_X2">; 1054defm S_ATOMIC_SMIN_X2 : SM_Real_Atomics_gfx10 <0xa4, "S_ATOMIC_SMIN_X2">; 1055defm S_ATOMIC_UMIN_X2 : SM_Real_Atomics_gfx10 <0xa5, "S_ATOMIC_UMIN_X2">; 1056defm S_ATOMIC_SMAX_X2 : SM_Real_Atomics_gfx10 <0xa6, "S_ATOMIC_SMAX_X2">; 1057defm S_ATOMIC_UMAX_X2 : SM_Real_Atomics_gfx10 <0xa7, "S_ATOMIC_UMAX_X2">; 1058defm S_ATOMIC_AND_X2 : SM_Real_Atomics_gfx10 <0xa8, "S_ATOMIC_AND_X2">; 1059defm S_ATOMIC_OR_X2 : SM_Real_Atomics_gfx10 <0xa9, "S_ATOMIC_OR_X2">; 1060defm S_ATOMIC_XOR_X2 : SM_Real_Atomics_gfx10 <0xaa, "S_ATOMIC_XOR_X2">; 1061defm S_ATOMIC_INC_X2 : SM_Real_Atomics_gfx10 <0xab, "S_ATOMIC_INC_X2">; 1062defm S_ATOMIC_DEC_X2 : SM_Real_Atomics_gfx10 <0xac, "S_ATOMIC_DEC_X2">; 1063 1064multiclass SM_Real_Discard_gfx10<bits<8> op, string ps> { 1065 def _IMM_gfx10 : SMEM_Real_gfx10 <op, !cast<SM_Pseudo>(ps#_IMM)>; 1066 def _SGPR_gfx10 : SMEM_Real_gfx10 <op, !cast<SM_Pseudo>(ps#_SGPR)>; 1067} 1068 1069defm S_DCACHE_DISCARD : SM_Real_Discard_gfx10 <0x28, "S_DCACHE_DISCARD">; 1070defm S_DCACHE_DISCARD_X2 : SM_Real_Discard_gfx10 <0x29, "S_DCACHE_DISCARD_X2">; 1071 1072} // End SubtargetPredicate = HasScalarAtomics 1073 1074def SMInfoTable : GenericTable { 1075 let FilterClass = "SM_Real"; 1076 let CppTypeName = "SMInfo"; 1077 let Fields = ["Opcode", "is_buffer"]; 1078 1079 let PrimaryKey = ["Opcode"]; 1080 let PrimaryKeyName = "getSMEMOpcodeHelper"; 1081} 1082