1//===-- SIInstrInfo.td - SI Instruction Infos -------------*- tablegen -*--===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9def isCI : Predicate<"Subtarget->getGeneration() " 10 ">= SISubtarget::SEA_ISLANDS">; 11def isCIOnly : Predicate<"Subtarget->getGeneration() ==" 12 "SISubtarget::SEA_ISLANDS">, 13 AssemblerPredicate <"FeatureSeaIslands">; 14 15def DisableInst : Predicate <"false">, AssemblerPredicate<"FeatureDisable">; 16 17// Execpt for the NONE field, this must be kept in sync with the 18// SIEncodingFamily enum in AMDGPUInstrInfo.cpp 19def SIEncodingFamily { 20 int NONE = -1; 21 int SI = 0; 22 int VI = 1; 23} 24 25//===----------------------------------------------------------------------===// 26// SI DAG Nodes 27//===----------------------------------------------------------------------===// 28 29def SIload_constant : SDNode<"AMDGPUISD::LOAD_CONSTANT", 30 SDTypeProfile<1, 2, [SDTCisVT<0, f32>, SDTCisVT<1, v4i32>, SDTCisVT<2, i32>]>, 31 [SDNPMayLoad, SDNPMemOperand] 32>; 33 34def SIatomic_inc : SDNode<"AMDGPUISD::ATOMIC_INC", SDTAtomic2, 35 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain] 36>; 37 38def SIatomic_dec : SDNode<"AMDGPUISD::ATOMIC_DEC", SDTAtomic2, 39 [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain] 40>; 41 42def SItbuffer_store : SDNode<"AMDGPUISD::TBUFFER_STORE_FORMAT", 43 SDTypeProfile<0, 13, 44 [SDTCisVT<0, v4i32>, // rsrc(SGPR) 45 SDTCisVT<1, iAny>, // vdata(VGPR) 46 SDTCisVT<2, i32>, // num_channels(imm) 47 SDTCisVT<3, i32>, // vaddr(VGPR) 48 SDTCisVT<4, i32>, // soffset(SGPR) 49 SDTCisVT<5, i32>, // inst_offset(imm) 50 SDTCisVT<6, i32>, // dfmt(imm) 51 SDTCisVT<7, i32>, // nfmt(imm) 52 SDTCisVT<8, i32>, // offen(imm) 53 SDTCisVT<9, i32>, // idxen(imm) 54 SDTCisVT<10, i32>, // glc(imm) 55 SDTCisVT<11, i32>, // slc(imm) 56 SDTCisVT<12, i32> // tfe(imm) 57 ]>, 58 [SDNPMayStore, SDNPMemOperand, SDNPHasChain] 59>; 60 61def SIload_input : SDNode<"AMDGPUISD::LOAD_INPUT", 62 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisVT<1, v4i32>, SDTCisVT<2, i16>, 63 SDTCisVT<3, i32>]> 64>; 65 66class SDSample<string opcode> : SDNode <opcode, 67 SDTypeProfile<1, 4, [SDTCisVT<0, v4f32>, SDTCisVT<2, v8i32>, 68 SDTCisVT<3, v4i32>, SDTCisVT<4, i32>]> 69>; 70 71def SIsample : SDSample<"AMDGPUISD::SAMPLE">; 72def SIsampleb : SDSample<"AMDGPUISD::SAMPLEB">; 73def SIsampled : SDSample<"AMDGPUISD::SAMPLED">; 74def SIsamplel : SDSample<"AMDGPUISD::SAMPLEL">; 75 76def SIpc_add_rel_offset : SDNode<"AMDGPUISD::PC_ADD_REL_OFFSET", 77 SDTypeProfile<1, 2, [SDTCisVT<0, iPTR>, SDTCisSameAs<0,1>, SDTCisSameAs<0,2>]> 78>; 79 80//===----------------------------------------------------------------------===// 81// PatFrags for global memory operations 82//===----------------------------------------------------------------------===// 83 84def atomic_inc_global : global_binary_atomic_op<SIatomic_inc>; 85def atomic_dec_global : global_binary_atomic_op<SIatomic_dec>; 86 87//===----------------------------------------------------------------------===// 88// SDNodes and PatFrag for local loads and stores to enable s_mov_b32 m0, -1 89// to be glued to the memory instructions. 90//===----------------------------------------------------------------------===// 91 92def SIld_local : SDNode <"ISD::LOAD", SDTLoad, 93 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand, SDNPInGlue] 94>; 95 96def si_ld_local : PatFrag <(ops node:$ptr), (SIld_local node:$ptr), [{ 97 return cast<LoadSDNode>(N)->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS; 98}]>; 99 100def si_load_local : PatFrag <(ops node:$ptr), (si_ld_local node:$ptr), [{ 101 return cast<LoadSDNode>(N)->getAddressingMode() == ISD::UNINDEXED && 102 cast<LoadSDNode>(N)->getExtensionType() == ISD::NON_EXTLOAD; 103}]>; 104 105def si_load_local_align8 : Aligned8Bytes < 106 (ops node:$ptr), (si_load_local node:$ptr) 107>; 108 109def si_sextload_local : PatFrag <(ops node:$ptr), (si_ld_local node:$ptr), [{ 110 return cast<LoadSDNode>(N)->getExtensionType() == ISD::SEXTLOAD; 111}]>; 112def si_az_extload_local : AZExtLoadBase <si_ld_local>; 113 114multiclass SIExtLoadLocal <PatFrag ld_node> { 115 116 def _i8 : PatFrag <(ops node:$ptr), (ld_node node:$ptr), 117 [{return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i8;}] 118 >; 119 120 def _i16 : PatFrag <(ops node:$ptr), (ld_node node:$ptr), 121 [{return cast<LoadSDNode>(N)->getMemoryVT() == MVT::i16;}] 122 >; 123} 124 125defm si_sextload_local : SIExtLoadLocal <si_sextload_local>; 126defm si_az_extload_local : SIExtLoadLocal <si_az_extload_local>; 127 128def SIst_local : SDNode <"ISD::STORE", SDTStore, 129 [SDNPHasChain, SDNPMayStore, SDNPMemOperand, SDNPInGlue] 130>; 131 132def si_st_local : PatFrag < 133 (ops node:$val, node:$ptr), (SIst_local node:$val, node:$ptr), [{ 134 return cast<StoreSDNode>(N)->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS; 135}]>; 136 137def si_store_local : PatFrag < 138 (ops node:$val, node:$ptr), (si_st_local node:$val, node:$ptr), [{ 139 return cast<StoreSDNode>(N)->getAddressingMode() == ISD::UNINDEXED && 140 !cast<StoreSDNode>(N)->isTruncatingStore(); 141}]>; 142 143def si_store_local_align8 : Aligned8Bytes < 144 (ops node:$val, node:$ptr), (si_store_local node:$val, node:$ptr) 145>; 146 147def si_truncstore_local : PatFrag < 148 (ops node:$val, node:$ptr), (si_st_local node:$val, node:$ptr), [{ 149 return cast<StoreSDNode>(N)->isTruncatingStore(); 150}]>; 151 152def si_truncstore_local_i8 : PatFrag < 153 (ops node:$val, node:$ptr), (si_truncstore_local node:$val, node:$ptr), [{ 154 return cast<StoreSDNode>(N)->getMemoryVT() == MVT::i8; 155}]>; 156 157def si_truncstore_local_i16 : PatFrag < 158 (ops node:$val, node:$ptr), (si_truncstore_local node:$val, node:$ptr), [{ 159 return cast<StoreSDNode>(N)->getMemoryVT() == MVT::i16; 160}]>; 161 162def si_setcc_uniform : PatFrag < 163 (ops node:$lhs, node:$rhs, node:$cond), 164 (setcc node:$lhs, node:$rhs, node:$cond), [{ 165 for (SDNode *Use : N->uses()) { 166 if (Use->isMachineOpcode() || Use->getOpcode() != ISD::CopyToReg) 167 return false; 168 169 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg(); 170 if (Reg != AMDGPU::SCC) 171 return false; 172 } 173 return true; 174}]>; 175 176def si_uniform_br : PatFrag < 177 (ops node:$cond, node:$bb), (brcond node:$cond, node:$bb), [{ 178 return isUniformBr(N); 179}]>; 180 181def si_uniform_br_scc : PatFrag < 182 (ops node:$cond, node:$bb), (si_uniform_br node:$cond, node:$bb), [{ 183 return isCBranchSCC(N); 184}]>; 185 186multiclass SIAtomicM0Glue2 <string op_name, bit is_amdgpu = 0> { 187 188 def _glue : SDNode < 189 !if(is_amdgpu, "AMDGPUISD", "ISD")#"::ATOMIC_"#op_name, SDTAtomic2, 190 [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand, SDNPInGlue] 191 >; 192 193 def _local : local_binary_atomic_op <!cast<SDNode>(NAME#"_glue")>; 194} 195 196defm si_atomic_load_add : SIAtomicM0Glue2 <"LOAD_ADD">; 197defm si_atomic_load_sub : SIAtomicM0Glue2 <"LOAD_SUB">; 198defm si_atomic_inc : SIAtomicM0Glue2 <"INC", 1>; 199defm si_atomic_dec : SIAtomicM0Glue2 <"DEC", 1>; 200defm si_atomic_load_and : SIAtomicM0Glue2 <"LOAD_AND">; 201defm si_atomic_load_min : SIAtomicM0Glue2 <"LOAD_MIN">; 202defm si_atomic_load_max : SIAtomicM0Glue2 <"LOAD_MAX">; 203defm si_atomic_load_or : SIAtomicM0Glue2 <"LOAD_OR">; 204defm si_atomic_load_xor : SIAtomicM0Glue2 <"LOAD_XOR">; 205defm si_atomic_load_umin : SIAtomicM0Glue2 <"LOAD_UMIN">; 206defm si_atomic_load_umax : SIAtomicM0Glue2 <"LOAD_UMAX">; 207defm si_atomic_swap : SIAtomicM0Glue2 <"SWAP">; 208 209def si_atomic_cmp_swap_glue : SDNode <"ISD::ATOMIC_CMP_SWAP", SDTAtomic3, 210 [SDNPHasChain, SDNPMayStore, SDNPMayLoad, SDNPMemOperand, SDNPInGlue] 211>; 212 213defm si_atomic_cmp_swap : AtomicCmpSwapLocal <si_atomic_cmp_swap_glue>; 214 215def as_i1imm : SDNodeXForm<imm, [{ 216 return CurDAG->getTargetConstant(N->getZExtValue(), SDLoc(N), MVT::i1); 217}]>; 218 219def as_i8imm : SDNodeXForm<imm, [{ 220 return CurDAG->getTargetConstant(N->getZExtValue(), SDLoc(N), MVT::i8); 221}]>; 222 223def as_i16imm : SDNodeXForm<imm, [{ 224 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i16); 225}]>; 226 227def as_i32imm: SDNodeXForm<imm, [{ 228 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32); 229}]>; 230 231def as_i64imm: SDNodeXForm<imm, [{ 232 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i64); 233}]>; 234 235// Copied from the AArch64 backend: 236def bitcast_fpimm_to_i32 : SDNodeXForm<fpimm, [{ 237return CurDAG->getTargetConstant( 238 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i32); 239}]>; 240 241def frameindex_to_targetframeindex : SDNodeXForm<frameindex, [{ 242 auto FI = cast<FrameIndexSDNode>(N); 243 return CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32); 244}]>; 245 246// Copied from the AArch64 backend: 247def bitcast_fpimm_to_i64 : SDNodeXForm<fpimm, [{ 248return CurDAG->getTargetConstant( 249 N->getValueAPF().bitcastToAPInt().getZExtValue(), SDLoc(N), MVT::i64); 250}]>; 251 252def SIMM16bit : PatLeaf <(imm), 253 [{return isInt<16>(N->getSExtValue());}] 254>; 255 256def IMM20bit : PatLeaf <(imm), 257 [{return isUInt<20>(N->getZExtValue());}] 258>; 259 260class InlineImm <ValueType vt> : PatLeaf <(vt imm), [{ 261 return isInlineImmediate(N); 262}]>; 263 264class InlineFPImm <ValueType vt> : PatLeaf <(vt fpimm), [{ 265 return isInlineImmediate(N); 266}]>; 267 268class SGPRImm <dag frag> : PatLeaf<frag, [{ 269 if (Subtarget->getGeneration() < SISubtarget::SOUTHERN_ISLANDS) { 270 return false; 271 } 272 const SIRegisterInfo *SIRI = 273 static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo()); 274 for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end(); 275 U != E; ++U) { 276 const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo()); 277 if (RC && SIRI->isSGPRClass(RC)) 278 return true; 279 } 280 return false; 281}]>; 282 283//===----------------------------------------------------------------------===// 284// Custom Operands 285//===----------------------------------------------------------------------===// 286 287def SoppBrTarget : AsmOperandClass { 288 let Name = "SoppBrTarget"; 289 let ParserMethod = "parseSOppBrTarget"; 290} 291 292def sopp_brtarget : Operand<OtherVT> { 293 let EncoderMethod = "getSOPPBrEncoding"; 294 let DecoderMethod = "decodeSoppBrTarget"; 295 let OperandType = "OPERAND_PCREL"; 296 let ParserMatchClass = SoppBrTarget; 297} 298 299def si_ga : Operand<iPTR>; 300 301def InterpSlot : Operand<i32> { 302 let PrintMethod = "printInterpSlot"; 303} 304 305def SendMsgMatchClass : AsmOperandClass { 306 let Name = "SendMsg"; 307 let PredicateMethod = "isSendMsg"; 308 let ParserMethod = "parseSendMsgOp"; 309 let RenderMethod = "addImmOperands"; 310} 311 312def SendMsgImm : Operand<i32> { 313 let PrintMethod = "printSendMsg"; 314 let ParserMatchClass = SendMsgMatchClass; 315} 316 317def SWaitMatchClass : AsmOperandClass { 318 let Name = "SWaitCnt"; 319 let RenderMethod = "addImmOperands"; 320 let ParserMethod = "parseSWaitCntOps"; 321} 322 323def WAIT_FLAG : Operand <i32> { 324 let ParserMatchClass = SWaitMatchClass; 325 let PrintMethod = "printWaitFlag"; 326} 327 328include "SIInstrFormats.td" 329include "VIInstrFormats.td" 330 331class NamedMatchClass<string CName, bit Optional = 1> : AsmOperandClass { 332 let Name = "Imm"#CName; 333 let PredicateMethod = "is"#CName; 334 let ParserMethod = !if(Optional, "parseOptionalOperand", "parse"#CName); 335 let RenderMethod = "addImmOperands"; 336 let IsOptional = Optional; 337 let DefaultMethod = !if(Optional, "default"#CName, ?); 338} 339 340class NamedOperandBit<string Name, AsmOperandClass MatchClass> : Operand<i1> { 341 let PrintMethod = "print"#Name; 342 let ParserMatchClass = MatchClass; 343} 344 345class NamedOperandU8<string Name, AsmOperandClass MatchClass> : Operand<i8> { 346 let PrintMethod = "print"#Name; 347 let ParserMatchClass = MatchClass; 348} 349 350class NamedOperandU16<string Name, AsmOperandClass MatchClass> : Operand<i16> { 351 let PrintMethod = "print"#Name; 352 let ParserMatchClass = MatchClass; 353} 354 355class NamedOperandU32<string Name, AsmOperandClass MatchClass> : Operand<i32> { 356 let PrintMethod = "print"#Name; 357 let ParserMatchClass = MatchClass; 358} 359 360let OperandType = "OPERAND_IMMEDIATE" in { 361 362def offen : NamedOperandBit<"Offen", NamedMatchClass<"Offen">>; 363def idxen : NamedOperandBit<"Idxen", NamedMatchClass<"Idxen">>; 364def addr64 : NamedOperandBit<"Addr64", NamedMatchClass<"Addr64">>; 365 366def offset : NamedOperandU16<"Offset", NamedMatchClass<"Offset">>; 367def offset0 : NamedOperandU8<"Offset0", NamedMatchClass<"Offset0">>; 368def offset1 : NamedOperandU8<"Offset1", NamedMatchClass<"Offset1">>; 369 370def gds : NamedOperandBit<"GDS", NamedMatchClass<"GDS">>; 371 372def omod : NamedOperandU32<"OModSI", NamedMatchClass<"OModSI">>; 373def clampmod : NamedOperandBit<"ClampSI", NamedMatchClass<"ClampSI">>; 374 375def glc : NamedOperandBit<"GLC", NamedMatchClass<"GLC">>; 376def slc : NamedOperandBit<"SLC", NamedMatchClass<"SLC">>; 377def tfe : NamedOperandBit<"TFE", NamedMatchClass<"TFE">>; 378def unorm : NamedOperandBit<"UNorm", NamedMatchClass<"UNorm">>; 379def da : NamedOperandBit<"DA", NamedMatchClass<"DA">>; 380def r128 : NamedOperandBit<"R128", NamedMatchClass<"R128">>; 381def lwe : NamedOperandBit<"LWE", NamedMatchClass<"LWE">>; 382 383def dmask : NamedOperandU16<"DMask", NamedMatchClass<"DMask">>; 384 385def dpp_ctrl : NamedOperandU32<"DPPCtrl", NamedMatchClass<"DPPCtrl", 0>>; 386def row_mask : NamedOperandU32<"RowMask", NamedMatchClass<"RowMask">>; 387def bank_mask : NamedOperandU32<"BankMask", NamedMatchClass<"BankMask">>; 388def bound_ctrl : NamedOperandBit<"BoundCtrl", NamedMatchClass<"BoundCtrl">>; 389 390def dst_sel : NamedOperandU32<"SDWADstSel", NamedMatchClass<"SDWADstSel">>; 391def src0_sel : NamedOperandU32<"SDWASrc0Sel", NamedMatchClass<"SDWASrc0Sel">>; 392def src1_sel : NamedOperandU32<"SDWASrc1Sel", NamedMatchClass<"SDWASrc1Sel">>; 393def dst_unused : NamedOperandU32<"SDWADstUnused", NamedMatchClass<"SDWADstUnused">>; 394 395def hwreg : NamedOperandU16<"Hwreg", NamedMatchClass<"Hwreg", 0>>; 396 397} // End OperandType = "OPERAND_IMMEDIATE" 398 399 400// 32-bit VALU immediate operand that uses the constant bus. 401def KImmFP32MatchClass : AsmOperandClass { 402 let Name = "KImmFP32"; 403 let PredicateMethod = "isKImmFP32"; 404 let ParserMethod = "parseImm"; 405 let RenderMethod = "addKImmFP32Operands"; 406} 407 408def f32kimm : Operand<i32> { 409 let OperandNamespace = "AMDGPU"; 410 let OperandType = "OPERAND_KIMM32"; 411 let PrintMethod = "printU32ImmOperand"; 412 let ParserMatchClass = KImmFP32MatchClass; 413} 414 415def VOPDstS64 : VOPDstOperand <SReg_64>; 416 417class FPInputModsMatchClass <int opSize> : AsmOperandClass { 418 let Name = "RegOrImmWithFP"#opSize#"InputMods"; 419 let ParserMethod = "parseRegOrImmWithFPInputMods"; 420 let PredicateMethod = "isRegOrImmWithFP"#opSize#"InputMods"; 421} 422def FP32InputModsMatchClass : FPInputModsMatchClass<32>; 423def FP64InputModsMatchClass : FPInputModsMatchClass<64>; 424 425class InputMods <AsmOperandClass matchClass> : Operand <i32> { 426 let OperandNamespace = "AMDGPU"; 427 let OperandType = "OPERAND_INPUT_MODS"; 428 let ParserMatchClass = matchClass; 429} 430 431class FPInputMods <FPInputModsMatchClass matchClass> : InputMods <matchClass> { 432 let PrintMethod = "printOperandAndFPInputMods"; 433} 434def FP32InputMods : FPInputMods<FP32InputModsMatchClass>; 435def FP64InputMods : FPInputMods<FP64InputModsMatchClass>; 436 437class IntInputModsMatchClass <int opSize> : AsmOperandClass { 438 let Name = "RegOrImmWithInt"#opSize#"InputMods"; 439 let ParserMethod = "parseRegOrImmWithIntInputMods"; 440 let PredicateMethod = "isRegOrImmWithInt"#opSize#"InputMods"; 441} 442def Int32InputModsMatchClass : IntInputModsMatchClass<32>; 443def Int64InputModsMatchClass : IntInputModsMatchClass<64>; 444 445class IntInputMods <IntInputModsMatchClass matchClass> : InputMods <matchClass> { 446 let PrintMethod = "printOperandAndIntInputMods"; 447} 448def Int32InputMods : IntInputMods<Int32InputModsMatchClass>; 449def Int64InputMods : IntInputMods<Int64InputModsMatchClass>; 450 451//===----------------------------------------------------------------------===// 452// Complex patterns 453//===----------------------------------------------------------------------===// 454 455def DS1Addr1Offset : ComplexPattern<i32, 2, "SelectDS1Addr1Offset">; 456def DS64Bit4ByteAligned : ComplexPattern<i32, 3, "SelectDS64Bit4ByteAligned">; 457 458def MOVRELOffset : ComplexPattern<i32, 2, "SelectMOVRELOffset">; 459 460def VOP3Mods0 : ComplexPattern<untyped, 4, "SelectVOP3Mods0">; 461def VOP3NoMods0 : ComplexPattern<untyped, 4, "SelectVOP3NoMods0">; 462def VOP3Mods0Clamp : ComplexPattern<untyped, 3, "SelectVOP3Mods0Clamp">; 463def VOP3Mods0Clamp0OMod : ComplexPattern<untyped, 4, "SelectVOP3Mods0Clamp0OMod">; 464def VOP3Mods : ComplexPattern<untyped, 2, "SelectVOP3Mods">; 465def VOP3NoMods : ComplexPattern<untyped, 2, "SelectVOP3NoMods">; 466 467//===----------------------------------------------------------------------===// 468// SI assembler operands 469//===----------------------------------------------------------------------===// 470 471def SIOperand { 472 int ZERO = 0x80; 473 int VCC = 0x6A; 474 int FLAT_SCR = 0x68; 475} 476 477def SRCMODS { 478 int NONE = 0; 479 int NEG = 1; 480} 481 482def DSTCLAMP { 483 int NONE = 0; 484} 485 486def DSTOMOD { 487 int NONE = 0; 488} 489 490//===----------------------------------------------------------------------===// 491// 492// SI Instruction multiclass helpers. 493// 494// Instructions with _32 take 32-bit operands. 495// Instructions with _64 take 64-bit operands. 496// 497// VOP_* instructions can use either a 32-bit or 64-bit encoding. The 32-bit 498// encoding is the standard encoding, but instruction that make use of 499// any of the instruction modifiers must use the 64-bit encoding. 500// 501// Instructions with _e32 use the 32-bit encoding. 502// Instructions with _e64 use the 64-bit encoding. 503// 504//===----------------------------------------------------------------------===// 505 506class SIMCInstr <string pseudo, int subtarget> { 507 string PseudoInstr = pseudo; 508 int Subtarget = subtarget; 509} 510 511//===----------------------------------------------------------------------===// 512// EXP classes 513//===----------------------------------------------------------------------===// 514 515class EXPCommon : InstSI< 516 (outs), 517 (ins i32imm:$en, i32imm:$tgt, i32imm:$compr, i32imm:$done, i32imm:$vm, 518 VGPR_32:$src0, VGPR_32:$src1, VGPR_32:$src2, VGPR_32:$src3), 519 "exp $en, $tgt, $compr, $done, $vm, $src0, $src1, $src2, $src3", 520 [] > { 521 522 let EXP_CNT = 1; 523 let Uses = [EXEC]; 524 let SchedRW = [WriteExport]; 525} 526 527multiclass EXP_m { 528 529 let isPseudo = 1, isCodeGenOnly = 1 in { 530 def "" : EXPCommon, SIMCInstr <"exp", SIEncodingFamily.NONE> ; 531 } 532 533 def _si : EXPCommon, SIMCInstr <"exp", SIEncodingFamily.SI>, EXPe { 534 let DecoderNamespace="SICI"; 535 let DisableDecoder = DisableSIDecoder; 536 } 537 538 def _vi : EXPCommon, SIMCInstr <"exp", SIEncodingFamily.VI>, EXPe_vi { 539 let DecoderNamespace="VI"; 540 let DisableDecoder = DisableVIDecoder; 541 } 542} 543 544//===----------------------------------------------------------------------===// 545// Vector ALU classes 546//===----------------------------------------------------------------------===// 547 548class getNumSrcArgs<ValueType Src0, ValueType Src1, ValueType Src2> { 549 int ret = 550 !if (!eq(Src0.Value, untyped.Value), 0, 551 !if (!eq(Src1.Value, untyped.Value), 1, // VOP1 552 !if (!eq(Src2.Value, untyped.Value), 2, // VOP2 553 3))); // VOP3 554} 555 556// Returns the register class to use for the destination of VOP[123C] 557// instructions for the given VT. 558class getVALUDstForVT<ValueType VT> { 559 RegisterOperand ret = !if(!eq(VT.Size, 32), VOPDstOperand<VGPR_32>, 560 !if(!eq(VT.Size, 128), VOPDstOperand<VReg_128>, 561 !if(!eq(VT.Size, 64), VOPDstOperand<VReg_64>, 562 !if(!eq(VT.Size, 16), VOPDstOperand<VGPR_32>, 563 VOPDstOperand<SReg_64>)))); // else VT == i1 564} 565 566// Returns the register class to use for source 0 of VOP[12C] 567// instructions for the given VT. 568class getVOPSrc0ForVT<ValueType VT> { 569 bit isFP = !if(!eq(VT.Value, f16.Value), 1, 570 !if(!eq(VT.Value, f32.Value), 1, 571 !if(!eq(VT.Value, f64.Value), 1, 572 0))); 573 RegisterOperand ret = !if(isFP, 574 !if(!eq(VT.Size, 64), VSrc_f64, VSrc_f32), 575 !if(!eq(VT.Size, 64), VSrc_b64, VSrc_b32)); 576} 577 578// Returns the vreg register class to use for source operand given VT 579class getVregSrcForVT<ValueType VT> { 580 RegisterClass ret = !if(!eq(VT.Size, 128), VReg_128, 581 !if(!eq(VT.Size, 64), VReg_64, VGPR_32)); 582} 583 584 585// Returns the register class to use for sources of VOP3 instructions for the 586// given VT. 587class getVOP3SrcForVT<ValueType VT> { 588 bit isFP = !if(!eq(VT.Value, f16.Value), 1, 589 !if(!eq(VT.Value, f32.Value), 1, 590 !if(!eq(VT.Value, f64.Value), 1, 591 0))); 592 RegisterOperand ret = 593 !if(!eq(VT.Size, 128), 594 VSrc_128, 595 !if(!eq(VT.Size, 64), 596 !if(isFP, 597 VCSrc_f64, 598 VCSrc_b64), 599 !if(!eq(VT.Value, i1.Value), 600 SCSrc_b64, 601 !if(isFP, 602 VCSrc_f32, 603 VCSrc_b32) 604 ) 605 ) 606 ); 607} 608 609// Returns 1 if the source arguments have modifiers, 0 if they do not. 610// XXX - do f16 instructions? 611class isFloatType<ValueType SrcVT> { 612 bit ret = 613 !if(!eq(SrcVT.Value, f16.Value), 1, 614 !if(!eq(SrcVT.Value, f32.Value), 1, 615 !if(!eq(SrcVT.Value, f64.Value), 1, 616 0))); 617} 618 619class isIntType<ValueType SrcVT> { 620 bit ret = 621 !if(!eq(SrcVT.Value, i16.Value), 1, 622 !if(!eq(SrcVT.Value, i32.Value), 1, 623 !if(!eq(SrcVT.Value, i64.Value), 1, 624 0))); 625} 626 627 628// Return type of input modifiers operand for specified input operand 629class getSrcMod <ValueType VT> { 630 bit isFP = !if(!eq(VT.Value, f16.Value), 1, 631 !if(!eq(VT.Value, f32.Value), 1, 632 !if(!eq(VT.Value, f64.Value), 1, 633 0))); 634 Operand ret = !if(!eq(VT.Size, 64), 635 !if(isFP, FP64InputMods, Int64InputMods), 636 !if(isFP, FP32InputMods, Int32InputMods)); 637} 638 639// Returns the input arguments for VOP[12C] instructions for the given SrcVT. 640class getIns32 <RegisterOperand Src0RC, RegisterClass Src1RC, int NumSrcArgs> { 641 dag ret = !if(!eq(NumSrcArgs, 1), (ins Src0RC:$src0), // VOP1 642 !if(!eq(NumSrcArgs, 2), (ins Src0RC:$src0, Src1RC:$src1), // VOP2 643 (ins))); 644} 645 646// Returns the input arguments for VOP3 instructions for the given SrcVT. 647class getIns64 <RegisterOperand Src0RC, RegisterOperand Src1RC, 648 RegisterOperand Src2RC, int NumSrcArgs, 649 bit HasModifiers, Operand Src0Mod, Operand Src1Mod, 650 Operand Src2Mod> { 651 652 dag ret = 653 !if (!eq(NumSrcArgs, 0), 654 // VOP1 without input operands (V_NOP, V_CLREXCP) 655 (ins), 656 /* else */ 657 !if (!eq(NumSrcArgs, 1), 658 !if (!eq(HasModifiers, 1), 659 // VOP1 with modifiers 660 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 661 clampmod:$clamp, omod:$omod) 662 /* else */, 663 // VOP1 without modifiers 664 (ins Src0RC:$src0) 665 /* endif */ ), 666 !if (!eq(NumSrcArgs, 2), 667 !if (!eq(HasModifiers, 1), 668 // VOP 2 with modifiers 669 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 670 Src1Mod:$src1_modifiers, Src1RC:$src1, 671 clampmod:$clamp, omod:$omod) 672 /* else */, 673 // VOP2 without modifiers 674 (ins Src0RC:$src0, Src1RC:$src1) 675 /* endif */ ) 676 /* NumSrcArgs == 3 */, 677 !if (!eq(HasModifiers, 1), 678 // VOP3 with modifiers 679 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 680 Src1Mod:$src1_modifiers, Src1RC:$src1, 681 Src2Mod:$src2_modifiers, Src2RC:$src2, 682 clampmod:$clamp, omod:$omod) 683 /* else */, 684 // VOP3 without modifiers 685 (ins Src0RC:$src0, Src1RC:$src1, Src2RC:$src2) 686 /* endif */ )))); 687} 688 689class getInsDPP <RegisterClass Src0RC, RegisterClass Src1RC, int NumSrcArgs, 690 bit HasModifiers, Operand Src0Mod, Operand Src1Mod> { 691 692 dag ret = !if (!eq(NumSrcArgs, 0), 693 // VOP1 without input operands (V_NOP) 694 (ins dpp_ctrl:$dpp_ctrl, row_mask:$row_mask, 695 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl), 696 !if (!eq(NumSrcArgs, 1), 697 !if (!eq(HasModifiers, 1), 698 // VOP1_DPP with modifiers 699 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 700 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask, 701 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl) 702 /* else */, 703 // VOP1_DPP without modifiers 704 (ins Src0RC:$src0, dpp_ctrl:$dpp_ctrl, row_mask:$row_mask, 705 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl) 706 /* endif */) 707 /* NumSrcArgs == 2 */, 708 !if (!eq(HasModifiers, 1), 709 // VOP2_DPP with modifiers 710 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 711 Src1Mod:$src1_modifiers, Src1RC:$src1, 712 dpp_ctrl:$dpp_ctrl, row_mask:$row_mask, 713 bank_mask:$bank_mask, bound_ctrl:$bound_ctrl) 714 /* else */, 715 // VOP2_DPP without modifiers 716 (ins Src0RC:$src0, Src1RC:$src1, dpp_ctrl:$dpp_ctrl, 717 row_mask:$row_mask, bank_mask:$bank_mask, 718 bound_ctrl:$bound_ctrl) 719 /* endif */))); 720} 721 722class getInsSDWA <RegisterClass Src0RC, RegisterClass Src1RC, int NumSrcArgs, 723 bit HasFloatModifiers, Operand Src0Mod, Operand Src1Mod, 724 ValueType DstVT> { 725 726 dag ret = !if(!eq(NumSrcArgs, 0), 727 // VOP1 without input operands (V_NOP) 728 (ins), 729 !if(!eq(NumSrcArgs, 1), 730 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 731 clampmod:$clamp, dst_sel:$dst_sel, dst_unused:$dst_unused, 732 src0_sel:$src0_sel), 733 !if(!eq(NumSrcArgs, 2), 734 !if(!eq(DstVT.Size, 1), 735 // VOPC_SDWA with float modifiers 736 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 737 Src1Mod:$src1_modifiers, Src1RC:$src1, 738 clampmod:$clamp, src0_sel:$src0_sel, src1_sel:$src1_sel), 739 // VOP2_SDWA or VOPC_SDWA with float modifiers 740 (ins Src0Mod:$src0_modifiers, Src0RC:$src0, 741 Src1Mod:$src1_modifiers, Src1RC:$src1, 742 clampmod:$clamp, dst_sel:$dst_sel, dst_unused:$dst_unused, 743 src0_sel:$src0_sel, src1_sel:$src1_sel)), 744 (ins)/* endif */))); 745} 746 747// Outs for DPP and SDWA 748class getOutsExt <bit HasDst, ValueType DstVT, RegisterOperand DstRCDPP> { 749 dag ret = !if(HasDst, 750 !if(!eq(DstVT.Size, 1), 751 (outs), // no dst for VOPC, we use "vcc"-token as dst in SDWA VOPC instructions 752 (outs DstRCDPP:$vdst)), 753 (outs)); // V_NOP 754} 755 756// Returns the assembly string for the inputs and outputs of a VOP[12C] 757// instruction. This does not add the _e32 suffix, so it can be reused 758// by getAsm64. 759class getAsm32 <bit HasDst, int NumSrcArgs, ValueType DstVT = i32> { 760 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC 761 string src0 = ", $src0"; 762 string src1 = ", $src1"; 763 string src2 = ", $src2"; 764 string ret = !if(HasDst, dst, "") # 765 !if(!eq(NumSrcArgs, 1), src0, "") # 766 !if(!eq(NumSrcArgs, 2), src0#src1, "") # 767 !if(!eq(NumSrcArgs, 3), src0#src1#src2, ""); 768} 769 770// Returns the assembly string for the inputs and outputs of a VOP3 771// instruction. 772class getAsm64 <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> { 773 string dst = !if(!eq(DstVT.Size, 1), "$sdst", "$vdst"); // use $sdst for VOPC 774 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,"); 775 string src1 = !if(!eq(NumSrcArgs, 1), "", 776 !if(!eq(NumSrcArgs, 2), " $src1_modifiers", 777 " $src1_modifiers,")); 778 string src2 = !if(!eq(NumSrcArgs, 3), " $src2_modifiers", ""); 779 string ret = 780 !if(!eq(HasModifiers, 0), 781 getAsm32<HasDst, NumSrcArgs, DstVT>.ret, 782 dst#", "#src0#src1#src2#"$clamp"#"$omod"); 783} 784 785class getAsmDPP <bit HasDst, int NumSrcArgs, bit HasModifiers, ValueType DstVT = i32> { 786 string dst = !if(HasDst, 787 !if(!eq(DstVT.Size, 1), 788 "$sdst", 789 "$vdst"), 790 ""); // use $sdst for VOPC 791 string src0 = !if(!eq(NumSrcArgs, 1), "$src0_modifiers", "$src0_modifiers,"); 792 string src1 = !if(!eq(NumSrcArgs, 1), "", 793 !if(!eq(NumSrcArgs, 2), " $src1_modifiers", 794 " $src1_modifiers,")); 795 string args = !if(!eq(HasModifiers, 0), 796 getAsm32<0, NumSrcArgs, DstVT>.ret, 797 ", "#src0#src1); 798 string ret = dst#args#" $dpp_ctrl$row_mask$bank_mask$bound_ctrl"; 799} 800 801class getAsmSDWA <bit HasDst, int NumSrcArgs, bit HasFloatModifiers, 802 ValueType DstVT = i32> { 803 string dst = !if(HasDst, 804 !if(!eq(DstVT.Size, 1), 805 " vcc", // use vcc token as dst for VOPC instructioins 806 "$vdst"), 807 ""); 808 string src0 = "$src0_modifiers"; 809 string src1 = "$src1_modifiers"; 810 string args = !if(!eq(NumSrcArgs, 0), 811 "", 812 !if(!eq(NumSrcArgs, 1), 813 ", "#src0#"$clamp", 814 ", "#src0#", "#src1#"$clamp" 815 ) 816 ); 817 string sdwa = !if(!eq(NumSrcArgs, 0), 818 "", 819 !if(!eq(NumSrcArgs, 1), 820 " $dst_sel $dst_unused $src0_sel", 821 !if(!eq(DstVT.Size, 1), 822 " $src0_sel $src1_sel", // No dst_sel and dst_unused for VOPC 823 " $dst_sel $dst_unused $src0_sel $src1_sel" 824 ) 825 ) 826 ); 827 string ret = dst#args#sdwa; 828} 829 830// Function that checks if instruction supports DPP and SDWA 831class getHasExt <int NumSrcArgs, ValueType DstVT = i32, ValueType Src0VT = i32, 832 ValueType Src1VT = i32> { 833 bit ret = !if(!eq(NumSrcArgs, 3), 834 0, // NumSrcArgs == 3 - No DPP or SDWA for VOP3 835 !if(!eq(DstVT.Size, 64), 836 0, // 64-bit dst - No DPP or SDWA for 64-bit operands 837 !if(!eq(Src0VT.Size, 64), 838 0, // 64-bit src0 839 !if(!eq(Src0VT.Size, 64), 840 0, // 64-bit src2 841 1 842 ) 843 ) 844 ) 845 ); 846} 847 848class BitOr<bit a, bit b> { 849 bit ret = !if(a, 1, !if(b, 1, 0)); 850} 851 852class BitAnd<bit a, bit b> { 853 bit ret = !if(a, !if(b, 1, 0), 0); 854} 855 856class VOPProfile <list<ValueType> _ArgVT> { 857 858 field list<ValueType> ArgVT = _ArgVT; 859 860 field ValueType DstVT = ArgVT[0]; 861 field ValueType Src0VT = ArgVT[1]; 862 field ValueType Src1VT = ArgVT[2]; 863 field ValueType Src2VT = ArgVT[3]; 864 field RegisterOperand DstRC = getVALUDstForVT<DstVT>.ret; 865 field RegisterOperand DstRCDPP = getVALUDstForVT<DstVT>.ret; 866 field RegisterOperand DstRCSDWA = getVALUDstForVT<DstVT>.ret; 867 field RegisterOperand Src0RC32 = getVOPSrc0ForVT<Src0VT>.ret; 868 field RegisterClass Src1RC32 = getVregSrcForVT<Src1VT>.ret; 869 field RegisterOperand Src0RC64 = getVOP3SrcForVT<Src0VT>.ret; 870 field RegisterOperand Src1RC64 = getVOP3SrcForVT<Src1VT>.ret; 871 field RegisterOperand Src2RC64 = getVOP3SrcForVT<Src2VT>.ret; 872 field RegisterClass Src0DPP = getVregSrcForVT<Src0VT>.ret; 873 field RegisterClass Src1DPP = getVregSrcForVT<Src1VT>.ret; 874 field RegisterClass Src0SDWA = getVregSrcForVT<Src0VT>.ret; 875 field RegisterClass Src1SDWA = getVregSrcForVT<Src1VT>.ret; 876 field Operand Src0Mod = getSrcMod<Src0VT>.ret; 877 field Operand Src1Mod = getSrcMod<Src1VT>.ret; 878 field Operand Src2Mod = getSrcMod<Src2VT>.ret; 879 880 field bit HasDst = !if(!eq(DstVT.Value, untyped.Value), 0, 1); 881 field bit HasDst32 = HasDst; 882 field bit EmitDst = HasDst; // force dst encoding, see v_movreld_b32 special case 883 field int NumSrcArgs = getNumSrcArgs<Src0VT, Src1VT, Src2VT>.ret; 884 field bit HasSrc0 = !if(!eq(Src0VT.Value, untyped.Value), 0, 1); 885 field bit HasSrc1 = !if(!eq(Src1VT.Value, untyped.Value), 0, 1); 886 field bit HasSrc2 = !if(!eq(Src2VT.Value, untyped.Value), 0, 1); 887 888 // TODO: Modifiers logic is somewhat adhoc here, to be refined later 889 field bit HasModifiers = isFloatType<Src0VT>.ret; 890 891 field bit HasSrc0FloatMods = isFloatType<Src0VT>.ret; 892 field bit HasSrc1FloatMods = isFloatType<Src1VT>.ret; 893 field bit HasSrc2FloatMods = isFloatType<Src2VT>.ret; 894 895 field bit HasSrc0IntMods = isIntType<Src0VT>.ret; 896 field bit HasSrc1IntMods = isIntType<Src1VT>.ret; 897 field bit HasSrc2IntMods = isIntType<Src2VT>.ret; 898 899 field bit HasSrc0Mods = HasModifiers; 900 field bit HasSrc1Mods = !if(HasModifiers, BitOr<HasSrc1FloatMods, HasSrc1IntMods>.ret, 0); 901 field bit HasSrc2Mods = !if(HasModifiers, BitOr<HasSrc2FloatMods, HasSrc2IntMods>.ret, 0); 902 903 field bit HasOMod = HasModifiers; 904 field bit HasClamp = HasModifiers; 905 field bit HasSDWAClamp = HasSrc0; 906 907 field bit HasExt = getHasExt<NumSrcArgs, DstVT, Src0VT, Src1VT>.ret; 908 909 field dag Outs = !if(HasDst,(outs DstRC:$vdst),(outs)); 910 911 // VOP3b instructions are a special case with a second explicit 912 // output. This is manually overridden for them. 913 field dag Outs32 = Outs; 914 field dag Outs64 = Outs; 915 field dag OutsDPP = getOutsExt<HasDst, DstVT, DstRCDPP>.ret; 916 field dag OutsSDWA = getOutsExt<HasDst, DstVT, DstRCDPP>.ret; 917 918 field dag Ins32 = getIns32<Src0RC32, Src1RC32, NumSrcArgs>.ret; 919 field dag Ins64 = getIns64<Src0RC64, Src1RC64, Src2RC64, NumSrcArgs, 920 HasModifiers, Src0Mod, Src1Mod, Src2Mod>.ret; 921 field dag InsDPP = getInsDPP<Src0DPP, Src1DPP, NumSrcArgs, 922 HasModifiers, Src0Mod, Src1Mod>.ret; 923 field dag InsSDWA = getInsSDWA<Src0SDWA, Src1SDWA, NumSrcArgs, 924 HasModifiers, Src0Mod, Src1Mod, DstVT>.ret; 925 926 field string Asm32 = getAsm32<HasDst, NumSrcArgs, DstVT>.ret; 927 field string Asm64 = getAsm64<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret; 928 field string AsmDPP = getAsmDPP<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret; 929 field string AsmSDWA = getAsmSDWA<HasDst, NumSrcArgs, HasModifiers, DstVT>.ret; 930} 931 932class VOP_NO_EXT <VOPProfile p> : VOPProfile <p.ArgVT> { 933 let HasExt = 0; 934} 935 936// FIXME: I think these F16/I16 profiles will need to use f16/i16 types in order 937// for the instruction patterns to work. 938def VOP_F16_F16 : VOPProfile <[f16, f16, untyped, untyped]>; 939def VOP_F16_I16 : VOPProfile <[f16, i32, untyped, untyped]>; 940def VOP_I16_F16 : VOPProfile <[i32, f16, untyped, untyped]>; 941 942def VOP_F16_F16_F16 : VOPProfile <[f16, f16, f16, untyped]>; 943def VOP_F16_F16_I16 : VOPProfile <[f16, f16, i32, untyped]>; 944def VOP_I16_I16_I16 : VOPProfile <[i32, i32, i32, untyped]>; 945 946def VOP_I16_I16_I16_I16 : VOPProfile <[i32, i32, i32, i32, untyped]>; 947def VOP_F16_F16_F16_F16 : VOPProfile <[f16, f16, f16, f16, untyped]>; 948 949def VOP_NONE : VOPProfile <[untyped, untyped, untyped, untyped]>; 950 951def VOP_F32_F32 : VOPProfile <[f32, f32, untyped, untyped]>; 952def VOP_F32_F64 : VOPProfile <[f32, f64, untyped, untyped]>; 953def VOP_F32_I32 : VOPProfile <[f32, i32, untyped, untyped]>; 954def VOP_F64_F32 : VOPProfile <[f64, f32, untyped, untyped]>; 955def VOP_F64_F64 : VOPProfile <[f64, f64, untyped, untyped]>; 956def VOP_F64_I32 : VOPProfile <[f64, i32, untyped, untyped]>; 957def VOP_I32_F32 : VOPProfile <[i32, f32, untyped, untyped]>; 958def VOP_I32_F64 : VOPProfile <[i32, f64, untyped, untyped]>; 959def VOP_I32_I32 : VOPProfile <[i32, i32, untyped, untyped]>; 960 961def VOP_F32_F32_F32 : VOPProfile <[f32, f32, f32, untyped]>; 962def VOP_F32_F32_I32 : VOPProfile <[f32, f32, i32, untyped]>; 963def VOP_F64_F64_F64 : VOPProfile <[f64, f64, f64, untyped]>; 964def VOP_F64_F64_I32 : VOPProfile <[f64, f64, i32, untyped]>; 965def VOP_I32_F32_F32 : VOPProfile <[i32, f32, f32, untyped]>; 966def VOP_I32_F32_I32 : VOPProfile <[i32, f32, i32, untyped]>; 967def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>; 968 969def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>; 970def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>; 971def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>; 972 973def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>; 974def VOP_F64_F64_F64_F64 : VOPProfile <[f64, f64, f64, f64]>; 975def VOP_I32_I32_I32_I32 : VOPProfile <[i32, i32, i32, i32]>; 976def VOP_I64_I32_I32_I64 : VOPProfile <[i64, i32, i32, i64]>; 977def VOP_I32_F32_I32_I32 : VOPProfile <[i32, f32, i32, i32]>; 978def VOP_I64_I64_I32_I64 : VOPProfile <[i64, i64, i32, i64]>; 979def VOP_V4I32_I64_I32_V4I32 : VOPProfile <[v4i32, i64, i32, v4i32]>; 980 981class Commutable_REV <string revOp, bit isOrig> { 982 string RevOp = revOp; 983 bit IsOrig = isOrig; 984} 985 986class AtomicNoRet <string noRetOp, bit isRet> { 987 string NoRetOp = noRetOp; 988 bit IsRet = isRet; 989} 990 991//===----------------------------------------------------------------------===// 992// Interpolation opcodes 993//===----------------------------------------------------------------------===// 994 995class VINTRP_Pseudo <string opName, dag outs, dag ins, list<dag> pattern> : 996 VINTRPCommon <outs, ins, "", pattern>, 997 SIMCInstr<opName, SIEncodingFamily.NONE> { 998 let isPseudo = 1; 999 let isCodeGenOnly = 1; 1000} 1001 1002class VINTRP_Real_si <bits <2> op, string opName, dag outs, dag ins, 1003 string asm> : 1004 VINTRPCommon <outs, ins, asm, []>, 1005 VINTRPe <op>, 1006 SIMCInstr<opName, SIEncodingFamily.SI> { 1007 let AssemblerPredicate = SIAssemblerPredicate; 1008 let DecoderNamespace = "SICI"; 1009 let DisableDecoder = DisableSIDecoder; 1010} 1011 1012class VINTRP_Real_vi <bits <2> op, string opName, dag outs, dag ins, 1013 string asm> : 1014 VINTRPCommon <outs, ins, asm, []>, 1015 VINTRPe_vi <op>, 1016 SIMCInstr<opName, SIEncodingFamily.VI> { 1017 let AssemblerPredicate = VIAssemblerPredicate; 1018 let DecoderNamespace = "VI"; 1019 let DisableDecoder = DisableVIDecoder; 1020} 1021 1022multiclass VINTRP_m <bits <2> op, dag outs, dag ins, string asm, 1023 list<dag> pattern = []> { 1024 def "" : VINTRP_Pseudo <NAME, outs, ins, pattern>; 1025 1026 def _si : VINTRP_Real_si <op, NAME, outs, ins, asm>; 1027 1028 def _vi : VINTRP_Real_vi <op, NAME, outs, ins, asm>; 1029} 1030 1031//===----------------------------------------------------------------------===// 1032// Vector instruction mappings 1033//===----------------------------------------------------------------------===// 1034 1035// Maps an opcode in e32 form to its e64 equivalent 1036def getVOPe64 : InstrMapping { 1037 let FilterClass = "VOP"; 1038 let RowFields = ["OpName"]; 1039 let ColFields = ["Size"]; 1040 let KeyCol = ["4"]; 1041 let ValueCols = [["8"]]; 1042} 1043 1044// Maps an opcode in e64 form to its e32 equivalent 1045def getVOPe32 : InstrMapping { 1046 let FilterClass = "VOP"; 1047 let RowFields = ["OpName"]; 1048 let ColFields = ["Size"]; 1049 let KeyCol = ["8"]; 1050 let ValueCols = [["4"]]; 1051} 1052 1053def getMaskedMIMGOp : InstrMapping { 1054 let FilterClass = "MIMG_Mask"; 1055 let RowFields = ["Op"]; 1056 let ColFields = ["Channels"]; 1057 let KeyCol = ["4"]; 1058 let ValueCols = [["1"], ["2"], ["3"] ]; 1059} 1060 1061// Maps an commuted opcode to its original version 1062def getCommuteOrig : InstrMapping { 1063 let FilterClass = "Commutable_REV"; 1064 let RowFields = ["RevOp"]; 1065 let ColFields = ["IsOrig"]; 1066 let KeyCol = ["0"]; 1067 let ValueCols = [["1"]]; 1068} 1069 1070// Maps an original opcode to its commuted version 1071def getCommuteRev : InstrMapping { 1072 let FilterClass = "Commutable_REV"; 1073 let RowFields = ["RevOp"]; 1074 let ColFields = ["IsOrig"]; 1075 let KeyCol = ["1"]; 1076 let ValueCols = [["0"]]; 1077} 1078 1079def getMCOpcodeGen : InstrMapping { 1080 let FilterClass = "SIMCInstr"; 1081 let RowFields = ["PseudoInstr"]; 1082 let ColFields = ["Subtarget"]; 1083 let KeyCol = [!cast<string>(SIEncodingFamily.NONE)]; 1084 let ValueCols = [[!cast<string>(SIEncodingFamily.SI)], 1085 [!cast<string>(SIEncodingFamily.VI)]]; 1086} 1087 1088// Get equivalent SOPK instruction. 1089def getSOPKOp : InstrMapping { 1090 let FilterClass = "SOPKInstTable"; 1091 let RowFields = ["BaseCmpOp"]; 1092 let ColFields = ["IsSOPK"]; 1093 let KeyCol = ["0"]; 1094 let ValueCols = [["1"]]; 1095} 1096 1097def getAddr64Inst : InstrMapping { 1098 let FilterClass = "MUBUFAddr64Table"; 1099 let RowFields = ["OpName"]; 1100 let ColFields = ["IsAddr64"]; 1101 let KeyCol = ["0"]; 1102 let ValueCols = [["1"]]; 1103} 1104 1105// Maps an atomic opcode to its version with a return value. 1106def getAtomicRetOp : InstrMapping { 1107 let FilterClass = "AtomicNoRet"; 1108 let RowFields = ["NoRetOp"]; 1109 let ColFields = ["IsRet"]; 1110 let KeyCol = ["0"]; 1111 let ValueCols = [["1"]]; 1112} 1113 1114// Maps an atomic opcode to its returnless version. 1115def getAtomicNoRetOp : InstrMapping { 1116 let FilterClass = "AtomicNoRet"; 1117 let RowFields = ["NoRetOp"]; 1118 let ColFields = ["IsRet"]; 1119 let KeyCol = ["1"]; 1120 let ValueCols = [["0"]]; 1121} 1122 1123include "SIInstructions.td" 1124include "CIInstructions.td" 1125include "VIInstructions.td" 1126 1127include "DSInstructions.td" 1128include "MIMGInstructions.td" 1129