1//===-- VOP3Instructions.td - Vector 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 9//===----------------------------------------------------------------------===// 10// VOP3 Classes 11//===----------------------------------------------------------------------===// 12 13class getVOP3ModPat<VOPProfile P, SDPatternOperator node> { 14 dag src0 = !if(P.HasOMod, 15 (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omod), 16 (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp)); 17 18 list<dag> ret3 = [(set P.DstVT:$vdst, 19 (DivergentFragOrOp<node, P>.ret (P.Src0VT src0), 20 (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers)), 21 (P.Src2VT (VOP3Mods P.Src2VT:$src2, i32:$src2_modifiers))))]; 22 23 list<dag> ret2 = [(set P.DstVT:$vdst, 24 (DivergentFragOrOp<node, P>.ret (P.Src0VT src0), 25 (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers))))]; 26 27 list<dag> ret1 = [(set P.DstVT:$vdst, 28 (DivergentFragOrOp<node, P>.ret (P.Src0VT src0)))]; 29 30 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3, 31 !if(!eq(P.NumSrcArgs, 2), ret2, 32 ret1)); 33} 34 35class getVOP3PModPat<VOPProfile P, SDPatternOperator node, bit HasExplicitClamp> { 36 dag src0_dag = (P.Src0VT (VOP3PMods P.Src0VT:$src0, i32:$src0_modifiers)); 37 dag src1_dag = (P.Src1VT (VOP3PMods P.Src1VT:$src1, i32:$src1_modifiers)); 38 dag src2_dag = (P.Src2VT (VOP3PMods P.Src2VT:$src2, i32:$src2_modifiers)); 39 dag clamp_dag = (i1 timm:$clamp); 40 41 list<dag> ret3 = [(set P.DstVT:$vdst, 42 !if(HasExplicitClamp, 43 (DivergentFragOrOp<node, P>.ret src0_dag, src1_dag, src2_dag, clamp_dag), 44 (DivergentFragOrOp<node, P>.ret src0_dag, src1_dag, src2_dag)))]; 45 46 list<dag> ret2 = [(set P.DstVT:$vdst, 47 !if(HasExplicitClamp, 48 (DivergentFragOrOp<node, P>.ret src0_dag, src1_dag, clamp_dag), 49 (DivergentFragOrOp<node, P>.ret src0_dag, src1_dag)))]; 50 51 list<dag> ret1 = [(set P.DstVT:$vdst, 52 !if(HasExplicitClamp, 53 (DivergentFragOrOp<node, P>.ret src0_dag, clamp_dag), 54 (DivergentFragOrOp<node, P>.ret src0_dag)))]; 55 56 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3, 57 !if(!eq(P.NumSrcArgs, 2), ret2, 58 ret1)); 59} 60 61class getVOP3OpSelPat<VOPProfile P, SDPatternOperator node> { 62 list<dag> ret3 = [(set P.DstVT:$vdst, 63 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3OpSel P.Src0VT:$src0, i32:$src0_modifiers)), 64 (P.Src1VT (VOP3OpSel P.Src1VT:$src1, i32:$src1_modifiers)), 65 (P.Src2VT (VOP3OpSel P.Src2VT:$src2, i32:$src2_modifiers))))]; 66 67 list<dag> ret2 = [(set P.DstVT:$vdst, 68 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3OpSel P.Src0VT:$src0, i32:$src0_modifiers)), 69 (P.Src1VT (VOP3OpSel P.Src1VT:$src1, i32:$src1_modifiers))))]; 70 71 list<dag> ret1 = [(set P.DstVT:$vdst, 72 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3OpSel P.Src0VT:$src0, i32:$src0_modifiers))))]; 73 74 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3, 75 !if(!eq(P.NumSrcArgs, 2), ret2, 76 ret1)); 77} 78 79class getVOP3OpSelModPat<VOPProfile P, SDPatternOperator node> { 80 list<dag> ret3 = [(set P.DstVT:$vdst, 81 (DivergentFragOrOp<node, P>.ret (P.Src0VT !if(P.HasClamp, (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers), 82 (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers))), 83 (P.Src1VT (VOP3OpSelMods P.Src1VT:$src1, i32:$src1_modifiers)), 84 (P.Src2VT (VOP3OpSelMods P.Src2VT:$src2, i32:$src2_modifiers))))]; 85 86 list<dag> ret2 = [(set P.DstVT:$vdst, 87 (DivergentFragOrOp<node, P>.ret !if(P.HasClamp, (P.Src0VT (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers)), 88 (P.Src0VT (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers))), 89 (P.Src1VT (VOP3OpSelMods P.Src1VT:$src1, i32:$src1_modifiers))))]; 90 91 list<dag> ret1 = [(set P.DstVT:$vdst, 92 (DivergentFragOrOp<node, P>.ret (P.Src0VT (VOP3OpSelMods P.Src0VT:$src0, i32:$src0_modifiers))))]; 93 94 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3, 95 !if(!eq(P.NumSrcArgs, 2), ret2, 96 ret1)); 97} 98 99class getVOP3Pat<VOPProfile P, SDPatternOperator node> { 100 list<dag> ret3 = [(set P.DstVT:$vdst, (DivergentFragOrOp<node, P>.ret P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2))]; 101 list<dag> ret2 = [(set P.DstVT:$vdst, (DivergentFragOrOp<node, P>.ret P.Src0VT:$src0, P.Src1VT:$src1))]; 102 list<dag> ret1 = [(set P.DstVT:$vdst, (DivergentFragOrOp<node, P>.ret P.Src0VT:$src0))]; 103 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3, 104 !if(!eq(P.NumSrcArgs, 2), ret2, 105 ret1)); 106} 107 108class getVOP3ClampPat<VOPProfile P, SDPatternOperator node> { 109 list<dag> ret3 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2, i1:$clamp))]; 110 list<dag> ret2 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1, i1:$clamp))]; 111 list<dag> ret1 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, i1:$clamp))]; 112 list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3, 113 !if(!eq(P.NumSrcArgs, 2), ret2, 114 ret1)); 115} 116 117class getVOP3MAIPat<VOPProfile P, SDPatternOperator node> { 118 list<dag> ret = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2, 119 timm:$cbsz, timm:$abid, timm:$blgp))]; 120} 121 122// Consistently gives instructions a _e64 suffix. 123multiclass VOP3Inst_Pseudo_Wrapper<string opName, VOPProfile P, list<dag> pattern = []> { 124 def _e64 : VOP3_Pseudo<opName, P, pattern>; 125} 126 127class VOP3InstBase<string OpName, VOPProfile P, SDPatternOperator node = null_frag> : 128 VOP3_Pseudo<OpName, P, 129 !if(P.HasOpSel, 130 !if(P.HasModifiers, 131 getVOP3OpSelModPat<P, node>.ret, 132 getVOP3OpSelPat<P, node>.ret), 133 !if(P.HasModifiers, 134 getVOP3ModPat<P, node>.ret, 135 !if(P.HasIntClamp, 136 getVOP3ClampPat<P, node>.ret, 137 !if (P.IsMAI, 138 getVOP3MAIPat<P, node>.ret, 139 getVOP3Pat<P, node>.ret)))), 140 0, P.HasOpSel> { 141 142 let IntClamp = P.HasIntClamp; 143 let AsmMatchConverter = 144 !if(P.HasOpSel, 145 "cvtVOP3OpSel", 146 !if(!or(P.HasModifiers, P.HasOMod, P.HasIntClamp), 147 "cvtVOP3", 148 "")); 149} 150 151multiclass VOP3Inst<string OpName, VOPProfile P, SDPatternOperator node = null_frag> { 152 def _e64 : VOP3InstBase<OpName, P, node>; 153} 154 155// Special case for v_div_fmas_{f32|f64}, since it seems to be the 156// only VOP instruction that implicitly reads VCC. 157let Asm64 = " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$clamp$omod" in { 158def VOP_F32_F32_F32_F32_VCC : VOPProfile<[f32, f32, f32, f32]> { 159 let Outs64 = (outs DstRC.RegClass:$vdst); 160} 161def VOP_F64_F64_F64_F64_VCC : VOPProfile<[f64, f64, f64, f64]> { 162 let Outs64 = (outs DstRC.RegClass:$vdst); 163} 164} 165 166class VOP3Features<bit Clamp, bit OpSel, bit Packed, bit MAI> { 167 bit HasClamp = Clamp; 168 bit HasOpSel = OpSel; 169 bit IsPacked = Packed; 170 bit IsMAI = MAI; 171} 172 173def VOP3_REGULAR : VOP3Features<0, 0, 0, 0>; 174def VOP3_CLAMP : VOP3Features<1, 0, 0, 0>; 175def VOP3_OPSEL : VOP3Features<1, 1, 0, 0>; 176def VOP3_PACKED : VOP3Features<1, 1, 1, 0>; 177def VOP3_MAI : VOP3Features<0, 0, 0, 1>; 178 179class VOP3_Profile<VOPProfile P, VOP3Features Features = VOP3_REGULAR> : VOPProfile<P.ArgVT> { 180 181 let HasClamp = !if(Features.HasClamp, 1, P.HasClamp); 182 let HasOpSel = !if(Features.HasOpSel, 1, P.HasOpSel); 183 let IsMAI = !if(Features.IsMAI, 1, P.IsMAI); 184 let IsPacked = !if(Features.IsPacked, 1, P.IsPacked); 185 186 let HasModifiers = !if(Features.IsMAI, 0, !or(Features.IsPacked, P.HasModifiers)); 187 let IsSingle = 1; 188} 189 190class VOP3b_Profile<ValueType vt> : VOPProfile<[vt, vt, vt, vt]> { 191 let Outs64 = (outs DstRC:$vdst, VOPDstS64orS32:$sdst); 192 let Asm64 = "$vdst, $sdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$clamp$omod"; 193 let IsSingle = 1; 194} 195 196def VOP3b_F32_I1_F32_F32_F32 : VOP3b_Profile<f32>; 197def VOP3b_F64_I1_F64_F64_F64 : VOP3b_Profile<f64>; 198 199def VOP3b_I64_I1_I32_I32_I64 : VOPProfile<[i64, i32, i32, i64]> { 200 let HasClamp = 1; 201 let IsSingle = 1; 202 203 let Outs64 = (outs DstRC:$vdst, VOPDstS64orS32:$sdst); 204 let Asm64 = "$vdst, $sdst, $src0, $src1, $src2$clamp"; 205} 206 207//===----------------------------------------------------------------------===// 208// VOP3 INTERP 209//===----------------------------------------------------------------------===// 210 211class VOP3Interp<string OpName, VOPProfile P, list<dag> pattern = []> : 212 VOP3_Pseudo<OpName, P, pattern> { 213 let AsmMatchConverter = "cvtVOP3Interp"; 214 let mayRaiseFPException = 0; 215} 216 217def VOP3_INTERP : VOPProfile<[f32, f32, i32, untyped]> { 218 let Ins64 = (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0, 219 Attr:$attr, AttrChan:$attrchan, 220 clampmod0:$clamp, omod0:$omod); 221 222 let Asm64 = "$vdst, $src0_modifiers, $attr$attrchan$clamp$omod"; 223} 224 225def VOP3_INTERP_MOV : VOPProfile<[f32, i32, i32, untyped]> { 226 let Ins64 = (ins InterpSlot:$src0, 227 Attr:$attr, AttrChan:$attrchan, 228 clampmod0:$clamp, omod0:$omod); 229 230 let Asm64 = "$vdst, $src0, $attr$attrchan$clamp$omod"; 231 232 let HasClamp = 1; 233 let HasSrc0Mods = 0; 234} 235 236class getInterp16Asm <bit HasSrc2, bit HasOMod> { 237 string src2 = !if(HasSrc2, ", $src2_modifiers", ""); 238 string omod = !if(HasOMod, "$omod", ""); 239 string ret = 240 " $vdst, $src0_modifiers, $attr$attrchan"#src2#"$high$clamp"#omod; 241} 242 243class getInterp16Ins <bit HasSrc2, bit HasOMod, 244 Operand Src0Mod, Operand Src2Mod> { 245 dag ret = !if(HasSrc2, 246 !if(HasOMod, 247 (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0, 248 Attr:$attr, AttrChan:$attrchan, 249 Src2Mod:$src2_modifiers, VRegSrc_32:$src2, 250 highmod:$high, clampmod0:$clamp, omod0:$omod), 251 (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0, 252 Attr:$attr, AttrChan:$attrchan, 253 Src2Mod:$src2_modifiers, VRegSrc_32:$src2, 254 highmod:$high, clampmod0:$clamp) 255 ), 256 (ins Src0Mod:$src0_modifiers, VRegSrc_32:$src0, 257 Attr:$attr, AttrChan:$attrchan, 258 highmod:$high, clampmod0:$clamp, omod0:$omod) 259 ); 260} 261 262class VOP3_INTERP16 <list<ValueType> ArgVT> : VOPProfile<ArgVT> { 263 264 let HasOMod = !ne(DstVT.Value, f16.Value); 265 let HasHigh = 1; 266 267 let Outs64 = (outs DstRC.RegClass:$vdst); 268 let Ins64 = getInterp16Ins<HasSrc2, HasOMod, Src0Mod, Src2Mod>.ret; 269 let Asm64 = getInterp16Asm<HasSrc2, HasOMod>.ret; 270} 271 272//===----------------------------------------------------------------------===// 273// VOP3 Instructions 274//===----------------------------------------------------------------------===// 275 276let isCommutable = 1 in { 277 278let isReMaterializable = 1 in { 279let mayRaiseFPException = 0 in { 280let SubtargetPredicate = HasMadMacF32Insts in { 281defm V_MAD_LEGACY_F32 : VOP3Inst <"v_mad_legacy_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>; 282defm V_MAD_F32 : VOP3Inst <"v_mad_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fmad>; 283} // End SubtargetPredicate = HasMadMacInsts 284 285let SubtargetPredicate = HasFmaLegacy32 in 286defm V_FMA_LEGACY_F32 : VOP3Inst <"v_fma_legacy_f32", 287 VOP3_Profile<VOP_F32_F32_F32_F32>, 288 int_amdgcn_fma_legacy>; 289} 290 291defm V_MAD_I32_I24 : VOP3Inst <"v_mad_i32_i24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 292defm V_MAD_U32_U24 : VOP3Inst <"v_mad_u32_u24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 293defm V_FMA_F32 : VOP3Inst <"v_fma_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, any_fma>; 294defm V_LERP_U8 : VOP3Inst <"v_lerp_u8", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_lerp>; 295 296let SchedRW = [WriteDoubleAdd] in { 297let FPDPRounding = 1 in { 298defm V_FMA_F64 : VOP3Inst <"v_fma_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, any_fma>; 299defm V_ADD_F64 : VOP3Inst <"v_add_f64", VOP3_Profile<VOP_F64_F64_F64>, any_fadd>; 300defm V_MUL_F64 : VOP3Inst <"v_mul_f64", VOP3_Profile<VOP_F64_F64_F64>, fmul>; 301} // End FPDPRounding = 1 302defm V_MIN_F64 : VOP3Inst <"v_min_f64", VOP3_Profile<VOP_F64_F64_F64>, fminnum_like>; 303defm V_MAX_F64 : VOP3Inst <"v_max_f64", VOP3_Profile<VOP_F64_F64_F64>, fmaxnum_like>; 304} // End SchedRW = [WriteDoubleAdd] 305 306let SchedRW = [WriteIntMul] in { 307defm V_MUL_LO_U32 : VOP3Inst <"v_mul_lo_u32", VOP3_Profile<VOP_I32_I32_I32>, DivergentBinFrag<mul>>; 308defm V_MUL_HI_U32 : VOP3Inst <"v_mul_hi_u32", VOP3_Profile<VOP_I32_I32_I32>, mulhu>; 309defm V_MUL_LO_I32 : VOP3Inst <"v_mul_lo_i32", VOP3_Profile<VOP_I32_I32_I32>>; 310defm V_MUL_HI_I32 : VOP3Inst <"v_mul_hi_i32", VOP3_Profile<VOP_I32_I32_I32>, mulhs>; 311} // End SchedRW = [WriteIntMul] 312} // End isReMaterializable = 1 313 314let Uses = [MODE, VCC, EXEC] in { 315// v_div_fmas_f32: 316// result = src0 * src1 + src2 317// if (vcc) 318// result *= 2^32 319// 320let SchedRW = [WriteFloatFMA] in 321defm V_DIV_FMAS_F32 : VOP3Inst_Pseudo_Wrapper <"v_div_fmas_f32", VOP_F32_F32_F32_F32_VCC, []>; 322// v_div_fmas_f64: 323// result = src0 * src1 + src2 324// if (vcc) 325// result *= 2^64 326// 327let SchedRW = [WriteDouble], FPDPRounding = 1 in 328defm V_DIV_FMAS_F64 : VOP3Inst_Pseudo_Wrapper <"v_div_fmas_f64", VOP_F64_F64_F64_F64_VCC, []>; 329} // End Uses = [MODE, VCC, EXEC] 330 331} // End isCommutable = 1 332 333let isReMaterializable = 1 in { 334let mayRaiseFPException = 0 in { 335defm V_CUBEID_F32 : VOP3Inst <"v_cubeid_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubeid>; 336defm V_CUBESC_F32 : VOP3Inst <"v_cubesc_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubesc>; 337defm V_CUBETC_F32 : VOP3Inst <"v_cubetc_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubetc>; 338defm V_CUBEMA_F32 : VOP3Inst <"v_cubema_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubema>; 339} // End mayRaiseFPException 340 341defm V_BFE_U32 : VOP3Inst <"v_bfe_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfe_u32>; 342defm V_BFE_I32 : VOP3Inst <"v_bfe_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfe_i32>; 343defm V_BFI_B32 : VOP3Inst <"v_bfi_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfi>; 344defm V_ALIGNBIT_B32 : VOP3Inst <"v_alignbit_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, fshr>; 345defm V_ALIGNBYTE_B32 : VOP3Inst <"v_alignbyte_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_alignbyte>; 346 347// XXX - No FPException seems suspect but manual doesn't say it does 348let mayRaiseFPException = 0 in { 349 let isCommutable = 1 in { 350 defm V_MIN3_I32 : VOP3Inst <"v_min3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmin3>; 351 defm V_MIN3_U32 : VOP3Inst <"v_min3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumin3>; 352 defm V_MAX3_I32 : VOP3Inst <"v_max3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmax3>; 353 defm V_MAX3_U32 : VOP3Inst <"v_max3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumax3>; 354 defm V_MED3_I32 : VOP3Inst <"v_med3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmed3>; 355 defm V_MED3_U32 : VOP3Inst <"v_med3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumed3>; 356 } // End isCommutable = 1 357 defm V_MIN3_F32 : VOP3Inst <"v_min3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmin3>; 358 defm V_MAX3_F32 : VOP3Inst <"v_max3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmax3>; 359 defm V_MED3_F32 : VOP3Inst <"v_med3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmed3>; 360} // End mayRaiseFPException = 0 361 362let isCommutable = 1 in { 363 defm V_SAD_U8 : VOP3Inst <"v_sad_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 364 defm V_SAD_HI_U8 : VOP3Inst <"v_sad_hi_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 365 defm V_SAD_U16 : VOP3Inst <"v_sad_u16", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 366 defm V_SAD_U32 : VOP3Inst <"v_sad_u32", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 367} // End isCommutable = 1 368defm V_CVT_PK_U8_F32 : VOP3Inst<"v_cvt_pk_u8_f32", VOP3_Profile<VOP_I32_F32_I32_I32>, int_amdgcn_cvt_pk_u8_f32>; 369 370defm V_DIV_FIXUP_F32 : VOP3Inst <"v_div_fixup_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUdiv_fixup>; 371 372let SchedRW = [WriteDoubleAdd], FPDPRounding = 1 in { 373 defm V_DIV_FIXUP_F64 : VOP3Inst <"v_div_fixup_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, AMDGPUdiv_fixup>; 374 defm V_LDEXP_F64 : VOP3Inst <"v_ldexp_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUldexp>; 375} // End SchedRW = [WriteDoubleAdd], FPDPRounding = 1 376} // End isReMaterializable = 1 377 378 379let mayRaiseFPException = 0 in { // Seems suspicious but manual doesn't say it does. 380 let SchedRW = [WriteFloatFMA, WriteSALU] in 381 defm V_DIV_SCALE_F32 : VOP3Inst_Pseudo_Wrapper <"v_div_scale_f32", VOP3b_F32_I1_F32_F32_F32> ; 382 383 // Double precision division pre-scale. 384 let SchedRW = [WriteDouble, WriteSALU], FPDPRounding = 1 in 385 defm V_DIV_SCALE_F64 : VOP3Inst_Pseudo_Wrapper <"v_div_scale_f64", VOP3b_F64_I1_F64_F64_F64>; 386} // End mayRaiseFPException = 0 387 388let isReMaterializable = 1 in 389defm V_MSAD_U8 : VOP3Inst <"v_msad_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>; 390 391let Constraints = "@earlyclobber $vdst" in { 392defm V_MQSAD_PK_U16_U8 : VOP3Inst <"v_mqsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>; 393} // End Constraints = "@earlyclobber $vdst" 394 395 396let isReMaterializable = 1 in { 397let SchedRW = [WriteDouble] in { 398defm V_TRIG_PREOP_F64 : VOP3Inst <"v_trig_preop_f64", VOP3_Profile<VOP_F64_F64_I32>, int_amdgcn_trig_preop>; 399} // End SchedRW = [WriteDouble] 400 401let SchedRW = [Write64Bit] in { 402 let SubtargetPredicate = isGFX6GFX7 in { 403 defm V_LSHL_B64 : VOP3Inst <"v_lshl_b64", VOP3_Profile<VOP_I64_I64_I32>, cshl_64>; 404 defm V_LSHR_B64 : VOP3Inst <"v_lshr_b64", VOP3_Profile<VOP_I64_I64_I32>, csrl_64>; 405 defm V_ASHR_I64 : VOP3Inst <"v_ashr_i64", VOP3_Profile<VOP_I64_I64_I32>, csra_64>; 406 } // End SubtargetPredicate = isGFX6GFX7 407 408 let SubtargetPredicate = isGFX8Plus in { 409 defm V_LSHLREV_B64 : VOP3Inst <"v_lshlrev_b64", VOP3_Profile<VOP_I64_I32_I64>, clshl_rev_64>; 410 defm V_LSHRREV_B64 : VOP3Inst <"v_lshrrev_b64", VOP3_Profile<VOP_I64_I32_I64>, clshr_rev_64>; 411 defm V_ASHRREV_I64 : VOP3Inst <"v_ashrrev_i64", VOP3_Profile<VOP_I64_I32_I64>, cashr_rev_64>; 412 } // End SubtargetPredicate = isGFX8Plus 413} // End SchedRW = [Write64Bit] 414} // End isReMaterializable = 1 415 416def : GCNPat< 417 (i32 (getDivergentFrag<sext>.ret i16:$src)), 418 (i32 (V_BFE_I32_e64 $src, (S_MOV_B32 (i32 0)), (S_MOV_B32 (i32 0x10)))) 419>; 420 421let isReMaterializable = 1 in { 422let SubtargetPredicate = isGFX6GFX7GFX10 in { 423defm V_MULLIT_F32 : VOP3Inst <"v_mullit_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>; 424} // End SubtargetPredicate = isGFX6GFX7GFX10 425 426let SchedRW = [Write32Bit] in { 427let SubtargetPredicate = isGFX8Plus in { 428defm V_PERM_B32 : VOP3Inst <"v_perm_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUperm>; 429} // End SubtargetPredicate = isGFX8Plus 430} // End SchedRW = [Write32Bit] 431} // End isReMaterializable = 1 432 433let SubtargetPredicate = isGFX7Plus in { 434 435let Constraints = "@earlyclobber $vdst", SchedRW = [WriteQuarterRate32] in { 436defm V_QSAD_PK_U16_U8 : VOP3Inst <"v_qsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>; 437defm V_MQSAD_U32_U8 : VOP3Inst <"v_mqsad_u32_u8", VOP3_Profile<VOP_V4I32_I64_I32_V4I32, VOP3_CLAMP>>; 438} // End Constraints = "@earlyclobber $vdst", SchedRW = [WriteQuarterRate32] 439 440let isCommutable = 1 in { 441let SchedRW = [WriteIntMul, WriteSALU] in { 442defm V_MAD_U64_U32 : VOP3Inst <"v_mad_u64_u32", VOP3b_I64_I1_I32_I32_I64>; 443defm V_MAD_I64_I32 : VOP3Inst <"v_mad_i64_i32", VOP3b_I64_I1_I32_I32_I64>; 444} // End SchedRW = [WriteIntMul, WriteSALU] 445} // End isCommutable = 1 446 447} // End SubtargetPredicate = isGFX7Plus 448 449let FPDPRounding = 1 in { 450 let Predicates = [Has16BitInsts, isGFX8Only] in { 451 defm V_DIV_FIXUP_F16 : VOP3Inst <"v_div_fixup_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, AMDGPUdiv_fixup>; 452 defm V_FMA_F16 : VOP3Inst <"v_fma_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, any_fma>; 453 } // End Predicates = [Has16BitInsts, isGFX8Only] 454 455 let renamedInGFX9 = 1, Predicates = [Has16BitInsts, isGFX9Plus] in { 456 defm V_DIV_FIXUP_F16_gfx9 : VOP3Inst <"v_div_fixup_f16_gfx9", 457 VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUdiv_fixup>; 458 defm V_FMA_F16_gfx9 : VOP3Inst <"v_fma_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, any_fma>; 459 } // End renamedInGFX9 = 1, Predicates = [Has16BitInsts, isGFX9Plus] 460} // End FPDPRounding = 1 461 462let SubtargetPredicate = Has16BitInsts, isCommutable = 1 in { 463 464let renamedInGFX9 = 1 in { 465 defm V_MAD_U16 : VOP3Inst <"v_mad_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_CLAMP>>; 466 defm V_MAD_I16 : VOP3Inst <"v_mad_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_CLAMP>>; 467 let FPDPRounding = 1 in { 468 defm V_MAD_F16 : VOP3Inst <"v_mad_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, fmad>; 469 let Uses = [MODE, M0, EXEC] in { 470 let OtherPredicates = [isNotGFX90APlus] in 471 // For some reason the intrinsic operands are in a different order 472 // from the instruction operands. 473 def V_INTERP_P2_F16 : VOP3Interp <"v_interp_p2_f16", VOP3_INTERP16<[f16, f32, i32, f32]>, 474 [(set f16:$vdst, 475 (int_amdgcn_interp_p2_f16 (VOP3Mods f32:$src2, i32:$src2_modifiers), 476 (VOP3Mods f32:$src0, i32:$src0_modifiers), 477 (i32 timm:$attrchan), 478 (i32 timm:$attr), 479 (i1 timm:$high), 480 M0))]>; 481 } // End Uses = [M0, MODE, EXEC] 482 } // End FPDPRounding = 1 483} // End renamedInGFX9 = 1 484 485let SubtargetPredicate = isGFX9Only, FPDPRounding = 1 in { 486 defm V_MAD_F16_gfx9 : VOP3Inst <"v_mad_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>> ; 487} // End SubtargetPredicate = isGFX9Only, FPDPRounding = 1 488 489let SubtargetPredicate = isGFX9Plus in { 490defm V_MAD_U16_gfx9 : VOP3Inst <"v_mad_u16_gfx9", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>>; 491defm V_MAD_I16_gfx9 : VOP3Inst <"v_mad_i16_gfx9", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>>; 492let OtherPredicates = [isNotGFX90APlus] in 493def V_INTERP_P2_F16_gfx9 : VOP3Interp <"v_interp_p2_f16_gfx9", VOP3_INTERP16<[f16, f32, i32, f32]>>; 494} // End SubtargetPredicate = isGFX9Plus 495 496// This predicate should only apply to the selection pattern. The 497// instruction still exists and should decode on subtargets with 498// other bank counts. 499let OtherPredicates = [isNotGFX90APlus, has32BankLDS], Uses = [MODE, M0, EXEC], FPDPRounding = 1 in { 500def V_INTERP_P1LL_F16 : VOP3Interp <"v_interp_p1ll_f16", VOP3_INTERP16<[f32, f32, i32, untyped]>, 501 [(set f32:$vdst, (int_amdgcn_interp_p1_f16 (VOP3Mods f32:$src0, i32:$src0_modifiers), 502 (i32 timm:$attrchan), 503 (i32 timm:$attr), 504 (i1 timm:$high), M0))]>; 505} // End OtherPredicates = [isNotGFX90APlus, has32BankLDS], Uses = [MODE, M0, EXEC], FPDPRounding = 1 506 507let OtherPredicates = [isNotGFX90APlus], Uses = [MODE, M0, EXEC], FPDPRounding = 1 in { 508def V_INTERP_P1LV_F16 : VOP3Interp <"v_interp_p1lv_f16", VOP3_INTERP16<[f32, f32, i32, f16]>>; 509} // End OtherPredicates = [isNotGFX90APlus], Uses = [MODE, M0, EXEC], FPDPRounding = 1 510 511} // End SubtargetPredicate = Has16BitInsts, isCommutable = 1 512 513def : GCNPat< 514 (i64 (getDivergentFrag<sext>.ret i16:$src)), 515 (REG_SEQUENCE VReg_64, 516 (i32 (V_BFE_I32_e64 $src, (S_MOV_B32 (i32 0)), (S_MOV_B32 (i32 0x10)))), sub0, 517 (i32 (COPY_TO_REGCLASS 518 (V_ASHRREV_I32_e32 (S_MOV_B32 (i32 0x1f)), (i32 (V_BFE_I32_e64 $src, (S_MOV_B32 (i32 0)), (S_MOV_B32 (i32 0x10)))) 519 ), VGPR_32)), sub1) 520>; 521 522let SubtargetPredicate = isGFX8Plus, Uses = [MODE, M0, EXEC], OtherPredicates = [isNotGFX90APlus] in { 523def V_INTERP_P1_F32_e64 : VOP3Interp <"v_interp_p1_f32", VOP3_INTERP>; 524def V_INTERP_P2_F32_e64 : VOP3Interp <"v_interp_p2_f32", VOP3_INTERP>; 525def V_INTERP_MOV_F32_e64 : VOP3Interp <"v_interp_mov_f32", VOP3_INTERP_MOV>; 526} // End SubtargetPredicate = isGFX8Plus, Uses = [MODE, M0, EXEC], OtherPredicates = [isNotGFX90APlus] 527 528let Predicates = [Has16BitInsts, isGFX6GFX7GFX8GFX9] in { 529 530multiclass Ternary_i16_Pats <SDPatternOperator op1, SDPatternOperator op2, 531 Instruction inst> { 532def : GCNPat < 533 (op2 (op1 i16:$src0, i16:$src1), i16:$src2), 534 (inst i16:$src0, i16:$src1, i16:$src2, (i1 0)) 535>; 536 537} 538 539defm: Ternary_i16_Pats<mul, add, V_MAD_U16_e64>; 540defm: Ternary_i16_Pats<mul, add, V_MAD_I16_e64>; 541 542} // End Predicates = [Has16BitInsts, isGFX6GFX7GFX8GFX9] 543 544let Predicates = [Has16BitInsts, isGFX10Plus] in { 545 546multiclass Ternary_i16_Pats_gfx9<SDPatternOperator op1, SDPatternOperator op2, 547 Instruction inst> { 548def : GCNPat < 549 (op2 (op1 i16:$src0, i16:$src1), i16:$src2), 550 (inst SRCMODS.NONE, $src0, SRCMODS.NONE, $src1, SRCMODS.NONE, $src2, DSTCLAMP.NONE) 551>; 552 553} 554 555defm: Ternary_i16_Pats_gfx9<mul, add, V_MAD_U16_gfx9_e64>; 556defm: Ternary_i16_Pats_gfx9<mul, add, V_MAD_I16_gfx9_e64>; 557 558} // End Predicates = [Has16BitInsts, isGFX10Plus] 559 560class ThreeOpFrag<SDPatternOperator op1, SDPatternOperator op2> : PatFrag< 561 (ops node:$x, node:$y, node:$z), 562 // When the inner operation is used multiple times, selecting 3-op 563 // instructions may still be beneficial -- if the other users can be 564 // combined similarly. Let's be conservative for now. 565 (op2 (HasOneUseBinOp<op1> node:$x, node:$y), node:$z), 566 [{ 567 // Only use VALU ops when the result is divergent. 568 if (!N->isDivergent()) 569 return false; 570 571 // Check constant bus limitations. 572 // 573 // Note: Use !isDivergent as a conservative proxy for whether the value 574 // is in an SGPR (uniform values can end up in VGPRs as well). 575 unsigned ConstantBusUses = 0; 576 for (unsigned i = 0; i < 3; ++i) { 577 if (!Operands[i]->isDivergent() && 578 !isInlineImmediate(Operands[i].getNode())) { 579 ConstantBusUses++; 580 // This uses AMDGPU::V_ADD3_U32_e64, but all three operand instructions 581 // have the same constant bus limit. 582 if (ConstantBusUses > Subtarget->getConstantBusLimit(AMDGPU::V_ADD3_U32_e64)) 583 return false; 584 } 585 } 586 587 return true; 588 }]> { 589 let PredicateCodeUsesOperands = 1; 590 591 // The divergence predicate is irrelevant in GlobalISel, as we have 592 // proper register bank checks. We just need to verify the constant 593 // bus restriction when all the sources are considered. 594 // 595 // FIXME: With unlucky SGPR operands, we could penalize code by 596 // blocking folding SGPR->VGPR copies later. 597 // FIXME: There's no register bank verifier 598 let GISelPredicateCode = [{ 599 const int ConstantBusLimit = Subtarget->getConstantBusLimit(AMDGPU::V_ADD3_U32_e64); 600 int ConstantBusUses = 0; 601 for (unsigned i = 0; i < 3; ++i) { 602 const RegisterBank *RegBank = RBI.getRegBank(Operands[i]->getReg(), MRI, TRI); 603 if (RegBank->getID() == AMDGPU::SGPRRegBankID) { 604 if (++ConstantBusUses > ConstantBusLimit) 605 return false; 606 } 607 } 608 return true; 609 }]; 610} 611 612let SubtargetPredicate = isGFX9Plus in { 613let isCommutable = 1, isReMaterializable = 1 in { 614 defm V_ADD3_U32 : VOP3Inst <"v_add3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 615 defm V_AND_OR_B32 : VOP3Inst <"v_and_or_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 616 defm V_OR3_B32 : VOP3Inst <"v_or3_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 617 defm V_XAD_U32 : VOP3Inst <"v_xad_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 618 defm V_ADD_I32 : VOP3Inst <"v_add_i32", VOP3_Profile<VOP_I32_I32_I32_ARITH>>; 619 defm V_ADD_LSHL_U32 : VOP3Inst <"v_add_lshl_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 620} // End isCommutable = 1, isReMaterializable = 1 621// TODO src0 contains the opsel bit for dst, so if we commute, need to mask and swap this 622// to the new src0. 623defm V_MED3_F16 : VOP3Inst <"v_med3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmed3>; 624defm V_MED3_I16 : VOP3Inst <"v_med3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmed3>; 625defm V_MED3_U16 : VOP3Inst <"v_med3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumed3>; 626 627defm V_MIN3_F16 : VOP3Inst <"v_min3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmin3>; 628defm V_MIN3_I16 : VOP3Inst <"v_min3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmin3>; 629defm V_MIN3_U16 : VOP3Inst <"v_min3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumin3>; 630 631defm V_MAX3_F16 : VOP3Inst <"v_max3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmax3>; 632defm V_MAX3_I16 : VOP3Inst <"v_max3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmax3>; 633defm V_MAX3_U16 : VOP3Inst <"v_max3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumax3>; 634 635defm V_ADD_I16 : VOP3Inst <"v_add_i16", VOP3_Profile<VOP_I16_I16_I16, VOP3_OPSEL>>; 636defm V_SUB_I16 : VOP3Inst <"v_sub_i16", VOP3_Profile<VOP_I16_I16_I16, VOP3_OPSEL>>; 637 638defm V_MAD_U32_U16 : VOP3Inst <"v_mad_u32_u16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>; 639defm V_MAD_I32_I16 : VOP3Inst <"v_mad_i32_i16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>; 640 641defm V_CVT_PKNORM_I16_F16 : VOP3Inst <"v_cvt_pknorm_i16_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>; 642defm V_CVT_PKNORM_U16_F16 : VOP3Inst <"v_cvt_pknorm_u16_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>; 643 644defm V_PACK_B32_F16 : VOP3Inst <"v_pack_b32_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>; 645 646let isReMaterializable = 1 in { 647defm V_SUB_I32 : VOP3Inst <"v_sub_i32", VOP3_Profile<VOP_I32_I32_I32_ARITH>>; 648defm V_LSHL_ADD_U32 : VOP3Inst <"v_lshl_add_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 649defm V_LSHL_OR_B32 : VOP3Inst <"v_lshl_or_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 650} // End isReMaterializable = 1 651 652 653class ThreeOp_i32_Pats <SDPatternOperator op1, SDPatternOperator op2, Instruction inst> : GCNPat < 654 // This matches (op2 (op1 i32:$src0, i32:$src1), i32:$src2) with conditions. 655 (ThreeOpFrag<op1, op2> i32:$src0, i32:$src1, i32:$src2), 656 (inst VSrc_b32:$src0, VSrc_b32:$src1, VSrc_b32:$src2) 657>; 658 659def : ThreeOp_i32_Pats<cshl_32, add, V_LSHL_ADD_U32_e64>; 660def : ThreeOp_i32_Pats<add, cshl_32, V_ADD_LSHL_U32_e64>; 661def : ThreeOp_i32_Pats<add, add, V_ADD3_U32_e64>; 662def : ThreeOp_i32_Pats<cshl_32, or, V_LSHL_OR_B32_e64>; 663def : ThreeOp_i32_Pats<and, or, V_AND_OR_B32_e64>; 664def : ThreeOp_i32_Pats<or, or, V_OR3_B32_e64>; 665def : ThreeOp_i32_Pats<xor, add, V_XAD_U32_e64>; 666 667def : VOPBinOpClampPat<saddsat, V_ADD_I32_e64, i32>; 668def : VOPBinOpClampPat<ssubsat, V_SUB_I32_e64, i32>; 669 670def : GCNPat<(getDivergentFrag<or>.ret (or_oneuse i64:$src0, i64:$src1), i64:$src2), 671 (REG_SEQUENCE VReg_64, 672 (V_OR3_B32_e64 (i32 (EXTRACT_SUBREG $src0, sub0)), 673 (i32 (EXTRACT_SUBREG $src1, sub0)), 674 (i32 (EXTRACT_SUBREG $src2, sub0))), sub0, 675 (V_OR3_B32_e64 (i32 (EXTRACT_SUBREG $src0, sub1)), 676 (i32 (EXTRACT_SUBREG $src1, sub1)), 677 (i32 (EXTRACT_SUBREG $src2, sub1))), sub1)>; 678 679// FIXME: Probably should hardcode clamp bit in pseudo and avoid this. 680class OpSelBinOpClampPat<SDPatternOperator node, 681 Instruction inst> : GCNPat< 682 (node (i16 (VOP3OpSel i16:$src0, i32:$src0_modifiers)), 683 (i16 (VOP3OpSel i16:$src1, i32:$src1_modifiers))), 684 (inst $src0_modifiers, $src0, $src1_modifiers, $src1, DSTCLAMP.ENABLE, 0) 685>; 686 687def : OpSelBinOpClampPat<saddsat, V_ADD_I16_e64>; 688def : OpSelBinOpClampPat<ssubsat, V_SUB_I16_e64>; 689} // End SubtargetPredicate = isGFX9Plus 690 691def VOP3_PERMLANE_Profile : VOP3_Profile<VOPProfile <[i32, i32, i32, i32]>, VOP3_OPSEL> { 692 let Src0RC64 = VRegSrc_32; 693 let Src1RC64 = SCSrc_b32; 694 let Src2RC64 = SCSrc_b32; 695 let InsVOP3OpSel = (ins IntOpSelMods:$src0_modifiers, VRegSrc_32:$src0, 696 IntOpSelMods:$src1_modifiers, SCSrc_b32:$src1, 697 IntOpSelMods:$src2_modifiers, SCSrc_b32:$src2, 698 VGPR_32:$vdst_in, op_sel0:$op_sel); 699 let HasClamp = 0; 700} 701 702class PermlanePat<SDPatternOperator permlane, 703 Instruction inst> : GCNPat< 704 (permlane i32:$vdst_in, i32:$src0, i32:$src1, i32:$src2, 705 timm:$fi, timm:$bc), 706 (inst (as_i1timm $fi), VGPR_32:$src0, (as_i1timm $bc), 707 SCSrc_b32:$src1, 0, SCSrc_b32:$src2, VGPR_32:$vdst_in) 708>; 709 710// Permlane intrinsic that has either fetch invalid or bound control 711// fields enabled. 712class BoundControlOrFetchInvalidPermlane<SDPatternOperator permlane> : 713 PatFrag<(ops node:$vdst_in, node:$src0, node:$src1, node:$src2, 714 node:$fi, node:$bc), 715 (permlane node:$vdst_in, node:$src0, node: 716 $src1, node:$src2, node:$fi, node:$bc)> { 717 let PredicateCode = [{ return N->getConstantOperandVal(5) != 0 || 718 N->getConstantOperandVal(6) != 0; }]; 719 let GISelPredicateCode = [{ 720 return MI.getOperand(6).getImm() != 0 || 721 MI.getOperand(7).getImm() != 0; 722 }]; 723} 724 725// Drop the input value if it won't be read. 726class PermlaneDiscardVDstIn<SDPatternOperator permlane, 727 Instruction inst> : GCNPat< 728 (permlane srcvalue, i32:$src0, i32:$src1, i32:$src2, 729 timm:$fi, timm:$bc), 730 (inst (as_i1timm $fi), VGPR_32:$src0, (as_i1timm $bc), 731 SCSrc_b32:$src1, 0, SCSrc_b32:$src2, 732 (IMPLICIT_DEF)) 733>; 734 735 736let SubtargetPredicate = isGFX10Plus in { 737 let isCommutable = 1, isReMaterializable = 1 in { 738 defm V_XOR3_B32 : VOP3Inst <"v_xor3_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>; 739 } // End isCommutable = 1, isReMaterializable = 1 740 def : ThreeOp_i32_Pats<xor, xor, V_XOR3_B32_e64>; 741 742 let Constraints = "$vdst = $vdst_in", DisableEncoding="$vdst_in" in { 743 defm V_PERMLANE16_B32 : VOP3Inst<"v_permlane16_b32", VOP3_PERMLANE_Profile>; 744 defm V_PERMLANEX16_B32 : VOP3Inst<"v_permlanex16_b32", VOP3_PERMLANE_Profile>; 745 } // End $vdst = $vdst_in, DisableEncoding $vdst_in 746 747 def : PermlanePat<int_amdgcn_permlane16, V_PERMLANE16_B32_e64>; 748 def : PermlanePat<int_amdgcn_permlanex16, V_PERMLANEX16_B32_e64>; 749 750 def : PermlaneDiscardVDstIn< 751 BoundControlOrFetchInvalidPermlane<int_amdgcn_permlane16>, 752 V_PERMLANE16_B32_e64>; 753 def : PermlaneDiscardVDstIn< 754 BoundControlOrFetchInvalidPermlane<int_amdgcn_permlanex16>, 755 V_PERMLANEX16_B32_e64>; 756} // End SubtargetPredicate = isGFX10Plus 757 758class DivFmasPat<ValueType vt, Instruction inst, Register CondReg> : GCNPat< 759 (AMDGPUdiv_fmas (vt (VOP3Mods vt:$src0, i32:$src0_modifiers)), 760 (vt (VOP3Mods vt:$src1, i32:$src1_modifiers)), 761 (vt (VOP3Mods vt:$src2, i32:$src2_modifiers)), 762 (i1 CondReg)), 763 (inst $src0_modifiers, $src0, $src1_modifiers, $src1, $src2_modifiers, $src2) 764>; 765 766let WaveSizePredicate = isWave64 in { 767def : DivFmasPat<f32, V_DIV_FMAS_F32_e64, VCC>; 768def : DivFmasPat<f64, V_DIV_FMAS_F64_e64, VCC>; 769} 770 771let WaveSizePredicate = isWave32 in { 772def : DivFmasPat<f32, V_DIV_FMAS_F32_e64, VCC_LO>; 773def : DivFmasPat<f64, V_DIV_FMAS_F64_e64, VCC_LO>; 774} 775 776//===----------------------------------------------------------------------===// 777// Integer Clamp Patterns 778//===----------------------------------------------------------------------===// 779 780class getClampPat<VOPProfile P, SDPatternOperator node> { 781 dag ret3 = (P.DstVT (node P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2)); 782 dag ret2 = (P.DstVT (node P.Src0VT:$src0, P.Src1VT:$src1)); 783 dag ret1 = (P.DstVT (node P.Src0VT:$src0)); 784 dag ret = !if(!eq(P.NumSrcArgs, 3), ret3, 785 !if(!eq(P.NumSrcArgs, 2), ret2, 786 ret1)); 787} 788 789class getClampRes<VOPProfile P, Instruction inst> { 790 dag ret3 = (inst P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2, (i1 0)); 791 dag ret2 = (inst P.Src0VT:$src0, P.Src1VT:$src1, (i1 0)); 792 dag ret1 = (inst P.Src0VT:$src0, (i1 0)); 793 dag ret = !if(!eq(P.NumSrcArgs, 3), ret3, 794 !if(!eq(P.NumSrcArgs, 2), ret2, 795 ret1)); 796} 797 798class IntClampPat<VOP3InstBase inst, SDPatternOperator node> : GCNPat< 799 getClampPat<inst.Pfl, node>.ret, 800 getClampRes<inst.Pfl, inst>.ret 801>; 802 803def : IntClampPat<V_MAD_I32_I24_e64, AMDGPUmad_i24>; 804def : IntClampPat<V_MAD_U32_U24_e64, AMDGPUmad_u24>; 805 806def : IntClampPat<V_SAD_U8_e64, int_amdgcn_sad_u8>; 807def : IntClampPat<V_SAD_HI_U8_e64, int_amdgcn_sad_hi_u8>; 808def : IntClampPat<V_SAD_U16_e64, int_amdgcn_sad_u16>; 809 810def : IntClampPat<V_MSAD_U8_e64, int_amdgcn_msad_u8>; 811def : IntClampPat<V_MQSAD_PK_U16_U8_e64, int_amdgcn_mqsad_pk_u16_u8>; 812 813def : IntClampPat<V_QSAD_PK_U16_U8_e64, int_amdgcn_qsad_pk_u16_u8>; 814def : IntClampPat<V_MQSAD_U32_U8_e64, int_amdgcn_mqsad_u32_u8>; 815 816 817//===----------------------------------------------------------------------===// 818// Target-specific instruction encodings. 819//===----------------------------------------------------------------------===// 820 821//===----------------------------------------------------------------------===// 822// GFX10. 823//===----------------------------------------------------------------------===// 824 825let AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" in { 826 multiclass VOP3_Real_gfx10<bits<10> op> { 827 def _gfx10 : 828 VOP3_Real<!cast<VOP_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX10>, 829 VOP3e_gfx10<op, !cast<VOP_Pseudo>(NAME#"_e64").Pfl>; 830 } 831 multiclass VOP3_Real_No_Suffix_gfx10<bits<10> op> { 832 def _gfx10 : 833 VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.GFX10>, 834 VOP3e_gfx10<op, !cast<VOP_Pseudo>(NAME).Pfl>; 835 } 836 multiclass VOP3_Real_gfx10_with_name<bits<10> op, string opName, 837 string asmName> { 838 def _gfx10 : 839 VOP3_Real<!cast<VOP3_Pseudo>(opName#"_e64"), SIEncodingFamily.GFX10>, 840 VOP3e_gfx10<op, !cast<VOP3_Pseudo>(opName#"_e64").Pfl> { 841 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(opName#"_e64"); 842 let AsmString = asmName # ps.AsmOperands; 843 let IsSingle = 1; 844 } 845 } 846 multiclass VOP3be_Real_gfx10<bits<10> op> { 847 def _gfx10 : 848 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX10>, 849 VOP3be_gfx10<op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 850 } 851 multiclass VOP3Interp_Real_gfx10<bits<10> op> { 852 def _gfx10 : 853 VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.GFX10>, 854 VOP3Interp_gfx10<op, !cast<VOP3_Pseudo>(NAME).Pfl>; 855 } 856 multiclass VOP3OpSel_Real_gfx10<bits<10> op> { 857 def _gfx10 : 858 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX10>, 859 VOP3OpSel_gfx10<op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 860 } 861 multiclass VOP3OpSel_Real_gfx10_with_name<bits<10> op, string opName, 862 string asmName> { 863 def _gfx10 : 864 VOP3_Real<!cast<VOP3_Pseudo>(opName#"_e64"), SIEncodingFamily.GFX10>, 865 VOP3OpSel_gfx10<op, !cast<VOP3_Pseudo>(opName#"_e64").Pfl> { 866 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(opName#"_e64"); 867 let AsmString = asmName # ps.AsmOperands; 868 } 869 } 870} // End AssemblerPredicate = isGFX10Plus, DecoderNamespace = "GFX10" 871 872defm V_READLANE_B32 : VOP3_Real_No_Suffix_gfx10<0x360>; 873 874let InOperandList = (ins SSrcOrLds_b32:$src0, SCSrc_b32:$src1, VGPR_32:$vdst_in) in { 875 defm V_WRITELANE_B32 : VOP3_Real_No_Suffix_gfx10<0x361>; 876} // End InOperandList = (ins SSrcOrLds_b32:$src0, SCSrc_b32:$src1, VGPR_32:$vdst_in) 877 878let SubtargetPredicate = isGFX10Before1030 in { 879 defm V_MUL_LO_I32 : VOP3_Real_gfx10<0x16b>; 880} 881 882defm V_XOR3_B32 : VOP3_Real_gfx10<0x178>; 883defm V_LSHLREV_B64 : VOP3_Real_gfx10<0x2ff>; 884defm V_LSHRREV_B64 : VOP3_Real_gfx10<0x300>; 885defm V_ASHRREV_I64 : VOP3_Real_gfx10<0x301>; 886defm V_PERM_B32 : VOP3_Real_gfx10<0x344>; 887defm V_XAD_U32 : VOP3_Real_gfx10<0x345>; 888defm V_LSHL_ADD_U32 : VOP3_Real_gfx10<0x346>; 889defm V_ADD_LSHL_U32 : VOP3_Real_gfx10<0x347>; 890defm V_ADD3_U32 : VOP3_Real_gfx10<0x36d>; 891defm V_LSHL_OR_B32 : VOP3_Real_gfx10<0x36f>; 892defm V_AND_OR_B32 : VOP3_Real_gfx10<0x371>; 893defm V_OR3_B32 : VOP3_Real_gfx10<0x372>; 894 895// TODO-GFX10: add MC tests for v_add/sub_nc_i16 896defm V_ADD_NC_I16 : 897 VOP3OpSel_Real_gfx10_with_name<0x30d, "V_ADD_I16", "v_add_nc_i16">; 898defm V_SUB_NC_I16 : 899 VOP3OpSel_Real_gfx10_with_name<0x30e, "V_SUB_I16", "v_sub_nc_i16">; 900defm V_SUB_NC_I32 : 901 VOP3_Real_gfx10_with_name<0x376, "V_SUB_I32", "v_sub_nc_i32">; 902defm V_ADD_NC_I32 : 903 VOP3_Real_gfx10_with_name<0x37f, "V_ADD_I32", "v_add_nc_i32">; 904 905defm V_INTERP_P1_F32_e64 : VOP3Interp_Real_gfx10<0x200>; 906defm V_INTERP_P2_F32_e64 : VOP3Interp_Real_gfx10<0x201>; 907defm V_INTERP_MOV_F32_e64 : VOP3Interp_Real_gfx10<0x202>; 908 909defm V_INTERP_P1LL_F16 : VOP3Interp_Real_gfx10<0x342>; 910defm V_INTERP_P1LV_F16 : VOP3Interp_Real_gfx10<0x343>; 911defm V_INTERP_P2_F16 : VOP3Interp_Real_gfx10<0x35a>; 912 913defm V_PACK_B32_F16 : VOP3OpSel_Real_gfx10<0x311>; 914defm V_CVT_PKNORM_I16_F16 : VOP3OpSel_Real_gfx10<0x312>; 915defm V_CVT_PKNORM_U16_F16 : VOP3OpSel_Real_gfx10<0x313>; 916 917defm V_MIN3_F16 : VOP3OpSel_Real_gfx10<0x351>; 918defm V_MIN3_I16 : VOP3OpSel_Real_gfx10<0x352>; 919defm V_MIN3_U16 : VOP3OpSel_Real_gfx10<0x353>; 920defm V_MAX3_F16 : VOP3OpSel_Real_gfx10<0x354>; 921defm V_MAX3_I16 : VOP3OpSel_Real_gfx10<0x355>; 922defm V_MAX3_U16 : VOP3OpSel_Real_gfx10<0x356>; 923defm V_MED3_F16 : VOP3OpSel_Real_gfx10<0x357>; 924defm V_MED3_I16 : VOP3OpSel_Real_gfx10<0x358>; 925defm V_MED3_U16 : VOP3OpSel_Real_gfx10<0x359>; 926defm V_MAD_U32_U16 : VOP3OpSel_Real_gfx10<0x373>; 927defm V_MAD_I32_I16 : VOP3OpSel_Real_gfx10<0x375>; 928 929defm V_MAD_U16 : 930 VOP3OpSel_Real_gfx10_with_name<0x340, "V_MAD_U16_gfx9", "v_mad_u16">; 931defm V_FMA_F16 : 932 VOP3OpSel_Real_gfx10_with_name<0x34b, "V_FMA_F16_gfx9", "v_fma_f16">; 933defm V_MAD_I16 : 934 VOP3OpSel_Real_gfx10_with_name<0x35e, "V_MAD_I16_gfx9", "v_mad_i16">; 935defm V_DIV_FIXUP_F16 : 936 VOP3OpSel_Real_gfx10_with_name<0x35f, "V_DIV_FIXUP_F16_gfx9", "v_div_fixup_f16">; 937 938// FIXME-GFX10-OPSEL: Need to add "selective" opsel support to some of these 939// (they do not support SDWA or DPP). 940defm V_ADD_NC_U16 : VOP3_Real_gfx10_with_name<0x303, "V_ADD_U16", "v_add_nc_u16">; 941defm V_SUB_NC_U16 : VOP3_Real_gfx10_with_name<0x304, "V_SUB_U16", "v_sub_nc_u16">; 942defm V_MUL_LO_U16 : VOP3_Real_gfx10_with_name<0x305, "V_MUL_LO_U16", "v_mul_lo_u16">; 943defm V_LSHRREV_B16 : VOP3_Real_gfx10_with_name<0x307, "V_LSHRREV_B16", "v_lshrrev_b16">; 944defm V_ASHRREV_I16 : VOP3_Real_gfx10_with_name<0x308, "V_ASHRREV_I16", "v_ashrrev_i16">; 945defm V_MAX_U16 : VOP3_Real_gfx10_with_name<0x309, "V_MAX_U16", "v_max_u16">; 946defm V_MAX_I16 : VOP3_Real_gfx10_with_name<0x30a, "V_MAX_I16", "v_max_i16">; 947defm V_MIN_U16 : VOP3_Real_gfx10_with_name<0x30b, "V_MIN_U16", "v_min_u16">; 948defm V_MIN_I16 : VOP3_Real_gfx10_with_name<0x30c, "V_MIN_I16", "v_min_i16">; 949defm V_LSHLREV_B16 : VOP3_Real_gfx10_with_name<0x314, "V_LSHLREV_B16", "v_lshlrev_b16">; 950defm V_PERMLANE16_B32 : VOP3OpSel_Real_gfx10<0x377>; 951defm V_PERMLANEX16_B32 : VOP3OpSel_Real_gfx10<0x378>; 952 953//===----------------------------------------------------------------------===// 954// GFX7, GFX10. 955//===----------------------------------------------------------------------===// 956 957let AssemblerPredicate = isGFX7Only, DecoderNamespace = "GFX7" in { 958 multiclass VOP3_Real_gfx7<bits<10> op> { 959 def _gfx7 : 960 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>, 961 VOP3e_gfx6_gfx7<op{8-0}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 962 } 963 multiclass VOP3be_Real_gfx7<bits<10> op> { 964 def _gfx7 : 965 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>, 966 VOP3be_gfx6_gfx7<op{8-0}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 967 } 968} // End AssemblerPredicate = isGFX7Only, DecoderNamespace = "GFX7" 969 970multiclass VOP3_Real_gfx7_gfx10<bits<10> op> : 971 VOP3_Real_gfx7<op>, VOP3_Real_gfx10<op>; 972 973multiclass VOP3be_Real_gfx7_gfx10<bits<10> op> : 974 VOP3be_Real_gfx7<op>, VOP3be_Real_gfx10<op>; 975 976defm V_QSAD_PK_U16_U8 : VOP3_Real_gfx7_gfx10<0x172>; 977defm V_MQSAD_U32_U8 : VOP3_Real_gfx7_gfx10<0x175>; 978defm V_MAD_U64_U32 : VOP3be_Real_gfx7_gfx10<0x176>; 979defm V_MAD_I64_I32 : VOP3be_Real_gfx7_gfx10<0x177>; 980 981//===----------------------------------------------------------------------===// 982// GFX6, GFX7, GFX10. 983//===----------------------------------------------------------------------===// 984 985let AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" in { 986 multiclass VOP3_Real_gfx6_gfx7<bits<10> op> { 987 def _gfx6_gfx7 : 988 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>, 989 VOP3e_gfx6_gfx7<op{8-0}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 990 } 991 multiclass VOP3be_Real_gfx6_gfx7<bits<10> op> { 992 def _gfx6_gfx7 : 993 VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.SI>, 994 VOP3be_gfx6_gfx7<op{8-0}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 995 } 996} // End AssemblerPredicate = isGFX6GFX7, DecoderNamespace = "GFX6GFX7" 997 998multiclass VOP3_Real_gfx6_gfx7_gfx10<bits<10> op> : 999 VOP3_Real_gfx6_gfx7<op>, VOP3_Real_gfx10<op>; 1000 1001multiclass VOP3be_Real_gfx6_gfx7_gfx10<bits<10> op> : 1002 VOP3be_Real_gfx6_gfx7<op>, VOP3be_Real_gfx10<op>; 1003 1004defm V_LSHL_B64 : VOP3_Real_gfx6_gfx7<0x161>; 1005defm V_LSHR_B64 : VOP3_Real_gfx6_gfx7<0x162>; 1006defm V_ASHR_I64 : VOP3_Real_gfx6_gfx7<0x163>; 1007defm V_MUL_LO_I32 : VOP3_Real_gfx6_gfx7<0x16b>; 1008 1009defm V_MAD_LEGACY_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x140>; 1010defm V_MAD_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x141>; 1011defm V_MAD_I32_I24 : VOP3_Real_gfx6_gfx7_gfx10<0x142>; 1012defm V_MAD_U32_U24 : VOP3_Real_gfx6_gfx7_gfx10<0x143>; 1013defm V_CUBEID_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x144>; 1014defm V_CUBESC_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x145>; 1015defm V_CUBETC_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x146>; 1016defm V_CUBEMA_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x147>; 1017defm V_BFE_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x148>; 1018defm V_BFE_I32 : VOP3_Real_gfx6_gfx7_gfx10<0x149>; 1019defm V_BFI_B32 : VOP3_Real_gfx6_gfx7_gfx10<0x14a>; 1020defm V_FMA_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x14b>; 1021defm V_FMA_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x14c>; 1022defm V_LERP_U8 : VOP3_Real_gfx6_gfx7_gfx10<0x14d>; 1023defm V_ALIGNBIT_B32 : VOP3_Real_gfx6_gfx7_gfx10<0x14e>; 1024defm V_ALIGNBYTE_B32 : VOP3_Real_gfx6_gfx7_gfx10<0x14f>; 1025defm V_MULLIT_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x150>; 1026defm V_MIN3_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x151>; 1027defm V_MIN3_I32 : VOP3_Real_gfx6_gfx7_gfx10<0x152>; 1028defm V_MIN3_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x153>; 1029defm V_MAX3_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x154>; 1030defm V_MAX3_I32 : VOP3_Real_gfx6_gfx7_gfx10<0x155>; 1031defm V_MAX3_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x156>; 1032defm V_MED3_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x157>; 1033defm V_MED3_I32 : VOP3_Real_gfx6_gfx7_gfx10<0x158>; 1034defm V_MED3_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x159>; 1035defm V_SAD_U8 : VOP3_Real_gfx6_gfx7_gfx10<0x15a>; 1036defm V_SAD_HI_U8 : VOP3_Real_gfx6_gfx7_gfx10<0x15b>; 1037defm V_SAD_U16 : VOP3_Real_gfx6_gfx7_gfx10<0x15c>; 1038defm V_SAD_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x15d>; 1039defm V_CVT_PK_U8_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x15e>; 1040defm V_DIV_FIXUP_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x15f>; 1041defm V_DIV_FIXUP_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x160>; 1042defm V_ADD_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x164>; 1043defm V_MUL_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x165>; 1044defm V_MIN_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x166>; 1045defm V_MAX_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x167>; 1046defm V_LDEXP_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x168>; 1047defm V_MUL_LO_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x169>; 1048defm V_MUL_HI_U32 : VOP3_Real_gfx6_gfx7_gfx10<0x16a>; 1049defm V_MUL_HI_I32 : VOP3_Real_gfx6_gfx7_gfx10<0x16c>; 1050defm V_DIV_FMAS_F32 : VOP3_Real_gfx6_gfx7_gfx10<0x16f>; 1051defm V_DIV_FMAS_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x170>; 1052defm V_MSAD_U8 : VOP3_Real_gfx6_gfx7_gfx10<0x171>; 1053defm V_MQSAD_PK_U16_U8 : VOP3_Real_gfx6_gfx7_gfx10<0x173>; 1054defm V_TRIG_PREOP_F64 : VOP3_Real_gfx6_gfx7_gfx10<0x174>; 1055defm V_DIV_SCALE_F32 : VOP3be_Real_gfx6_gfx7_gfx10<0x16d>; 1056defm V_DIV_SCALE_F64 : VOP3be_Real_gfx6_gfx7_gfx10<0x16e>; 1057 1058// NB: Same opcode as v_mad_legacy_f32 1059let DecoderNamespace = "GFX10_B" in 1060defm V_FMA_LEGACY_F32 : VOP3_Real_gfx10<0x140>; 1061 1062//===----------------------------------------------------------------------===// 1063// GFX8, GFX9 (VI). 1064//===----------------------------------------------------------------------===// 1065 1066let AssemblerPredicate = isGFX8GFX9, DecoderNamespace = "GFX8" in { 1067 1068multiclass VOP3_Real_vi<bits<10> op> { 1069 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>, 1070 VOP3e_vi <op, !cast<VOP_Pseudo>(NAME#"_e64").Pfl>; 1071} 1072multiclass VOP3_Real_No_Suffix_vi<bits<10> op> { 1073 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.VI>, 1074 VOP3e_vi <op, !cast<VOP_Pseudo>(NAME).Pfl>; 1075} 1076 1077multiclass VOP3be_Real_vi<bits<10> op> { 1078 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>, 1079 VOP3be_vi <op, !cast<VOP_Pseudo>(NAME#"_e64").Pfl>; 1080} 1081 1082multiclass VOP3OpSel_Real_gfx9<bits<10> op> { 1083 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>, 1084 VOP3OpSel_gfx9 <op, !cast<VOP_Pseudo>(NAME#"_e64").Pfl>; 1085} 1086 1087multiclass VOP3Interp_Real_vi<bits<10> op> { 1088 def _vi : VOP3_Real<!cast<VOP_Pseudo>(NAME), SIEncodingFamily.VI>, 1089 VOP3Interp_vi <op, !cast<VOP_Pseudo>(NAME).Pfl>; 1090} 1091 1092} // End AssemblerPredicate = isGFX8GFX9, DecoderNamespace = "GFX8" 1093 1094let AssemblerPredicate = isGFX8Only, DecoderNamespace = "GFX8" in { 1095 1096multiclass VOP3_F16_Real_vi<bits<10> op> { 1097 def _vi : VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>, 1098 VOP3e_vi <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>; 1099} 1100 1101multiclass VOP3Interp_F16_Real_vi<bits<10> op> { 1102 def _vi : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>, 1103 VOP3Interp_vi <op, !cast<VOP3_Pseudo>(NAME).Pfl>; 1104} 1105 1106} // End AssemblerPredicate = isGFX8Only, DecoderNamespace = "GFX8" 1107 1108let AssemblerPredicate = isGFX9Only, DecoderNamespace = "GFX9" in { 1109 1110multiclass VOP3_F16_Real_gfx9<bits<10> op, string OpName, string AsmName> { 1111 def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(OpName#"_e64"), SIEncodingFamily.GFX9>, 1112 VOP3e_vi <op, !cast<VOP3_Pseudo>(OpName#"_e64").Pfl> { 1113 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName#"_e64"); 1114 let AsmString = AsmName # ps.AsmOperands; 1115 } 1116} 1117 1118multiclass VOP3OpSel_F16_Real_gfx9<bits<10> op, string AsmName> { 1119 def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX9>, 1120 VOP3OpSel_gfx9 <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl> { 1121 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(NAME#"_e64"); 1122 let AsmString = AsmName # ps.AsmOperands; 1123 } 1124} 1125 1126multiclass VOP3Interp_F16_Real_gfx9<bits<10> op, string OpName, string AsmName> { 1127 def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(OpName), SIEncodingFamily.GFX9>, 1128 VOP3Interp_vi <op, !cast<VOP3_Pseudo>(OpName).Pfl> { 1129 VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName); 1130 let AsmString = AsmName # ps.AsmOperands; 1131 } 1132} 1133 1134multiclass VOP3_Real_gfx9<bits<10> op, string AsmName> { 1135 def _gfx9 : VOP3_Real<!cast<VOP_Pseudo>(NAME#"_e64"), SIEncodingFamily.GFX9>, 1136 VOP3e_vi <op, !cast<VOP_Pseudo>(NAME#"_e64").Pfl> { 1137 VOP_Pseudo ps = !cast<VOP_Pseudo>(NAME#"_e64"); 1138 let AsmString = AsmName # ps.AsmOperands; 1139 } 1140} 1141 1142} // End AssemblerPredicate = isGFX9Only, DecoderNamespace = "GFX9" 1143 1144defm V_MAD_U64_U32 : VOP3be_Real_vi <0x1E8>; 1145defm V_MAD_I64_I32 : VOP3be_Real_vi <0x1E9>; 1146 1147defm V_MAD_LEGACY_F32 : VOP3_Real_vi <0x1c0>; 1148defm V_MAD_F32 : VOP3_Real_vi <0x1c1>; 1149defm V_MAD_I32_I24 : VOP3_Real_vi <0x1c2>; 1150defm V_MAD_U32_U24 : VOP3_Real_vi <0x1c3>; 1151defm V_CUBEID_F32 : VOP3_Real_vi <0x1c4>; 1152defm V_CUBESC_F32 : VOP3_Real_vi <0x1c5>; 1153defm V_CUBETC_F32 : VOP3_Real_vi <0x1c6>; 1154defm V_CUBEMA_F32 : VOP3_Real_vi <0x1c7>; 1155defm V_BFE_U32 : VOP3_Real_vi <0x1c8>; 1156defm V_BFE_I32 : VOP3_Real_vi <0x1c9>; 1157defm V_BFI_B32 : VOP3_Real_vi <0x1ca>; 1158defm V_FMA_F32 : VOP3_Real_vi <0x1cb>; 1159defm V_FMA_F64 : VOP3_Real_vi <0x1cc>; 1160defm V_LERP_U8 : VOP3_Real_vi <0x1cd>; 1161defm V_ALIGNBIT_B32 : VOP3_Real_vi <0x1ce>; 1162defm V_ALIGNBYTE_B32 : VOP3_Real_vi <0x1cf>; 1163defm V_MIN3_F32 : VOP3_Real_vi <0x1d0>; 1164defm V_MIN3_I32 : VOP3_Real_vi <0x1d1>; 1165defm V_MIN3_U32 : VOP3_Real_vi <0x1d2>; 1166defm V_MAX3_F32 : VOP3_Real_vi <0x1d3>; 1167defm V_MAX3_I32 : VOP3_Real_vi <0x1d4>; 1168defm V_MAX3_U32 : VOP3_Real_vi <0x1d5>; 1169defm V_MED3_F32 : VOP3_Real_vi <0x1d6>; 1170defm V_MED3_I32 : VOP3_Real_vi <0x1d7>; 1171defm V_MED3_U32 : VOP3_Real_vi <0x1d8>; 1172defm V_SAD_U8 : VOP3_Real_vi <0x1d9>; 1173defm V_SAD_HI_U8 : VOP3_Real_vi <0x1da>; 1174defm V_SAD_U16 : VOP3_Real_vi <0x1db>; 1175defm V_SAD_U32 : VOP3_Real_vi <0x1dc>; 1176defm V_CVT_PK_U8_F32 : VOP3_Real_vi <0x1dd>; 1177defm V_DIV_FIXUP_F32 : VOP3_Real_vi <0x1de>; 1178defm V_DIV_FIXUP_F64 : VOP3_Real_vi <0x1df>; 1179defm V_DIV_SCALE_F32 : VOP3be_Real_vi <0x1e0>; 1180defm V_DIV_SCALE_F64 : VOP3be_Real_vi <0x1e1>; 1181defm V_DIV_FMAS_F32 : VOP3_Real_vi <0x1e2>; 1182defm V_DIV_FMAS_F64 : VOP3_Real_vi <0x1e3>; 1183defm V_MSAD_U8 : VOP3_Real_vi <0x1e4>; 1184defm V_QSAD_PK_U16_U8 : VOP3_Real_vi <0x1e5>; 1185defm V_MQSAD_PK_U16_U8 : VOP3_Real_vi <0x1e6>; 1186defm V_MQSAD_U32_U8 : VOP3_Real_vi <0x1e7>; 1187 1188defm V_PERM_B32 : VOP3_Real_vi <0x1ed>; 1189 1190defm V_MAD_F16 : VOP3_F16_Real_vi <0x1ea>; 1191defm V_MAD_U16 : VOP3_F16_Real_vi <0x1eb>; 1192defm V_MAD_I16 : VOP3_F16_Real_vi <0x1ec>; 1193defm V_FMA_F16 : VOP3_F16_Real_vi <0x1ee>; 1194defm V_DIV_FIXUP_F16 : VOP3_F16_Real_vi <0x1ef>; 1195defm V_INTERP_P2_F16 : VOP3Interp_F16_Real_vi <0x276>; 1196 1197let FPDPRounding = 1 in { 1198defm V_MAD_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ea, "V_MAD_F16", "v_mad_legacy_f16">; 1199defm V_FMA_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ee, "V_FMA_F16", "v_fma_legacy_f16">; 1200defm V_DIV_FIXUP_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ef, "V_DIV_FIXUP_F16", "v_div_fixup_legacy_f16">; 1201defm V_INTERP_P2_LEGACY_F16 : VOP3Interp_F16_Real_gfx9 <0x276, "V_INTERP_P2_F16", "v_interp_p2_legacy_f16">; 1202} // End FPDPRounding = 1 1203 1204defm V_MAD_LEGACY_U16 : VOP3_F16_Real_gfx9 <0x1eb, "V_MAD_U16", "v_mad_legacy_u16">; 1205defm V_MAD_LEGACY_I16 : VOP3_F16_Real_gfx9 <0x1ec, "V_MAD_I16", "v_mad_legacy_i16">; 1206 1207defm V_MAD_F16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x203, "v_mad_f16">; 1208defm V_MAD_U16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x204, "v_mad_u16">; 1209defm V_MAD_I16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x205, "v_mad_i16">; 1210defm V_FMA_F16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x206, "v_fma_f16">; 1211defm V_DIV_FIXUP_F16_gfx9 : VOP3OpSel_F16_Real_gfx9 <0x207, "v_div_fixup_f16">; 1212defm V_INTERP_P2_F16_gfx9 : VOP3Interp_F16_Real_gfx9 <0x277, "V_INTERP_P2_F16_gfx9", "v_interp_p2_f16">; 1213 1214defm V_ADD_I32 : VOP3_Real_vi <0x29c>; 1215defm V_SUB_I32 : VOP3_Real_vi <0x29d>; 1216 1217defm V_INTERP_P1_F32_e64 : VOP3Interp_Real_vi <0x270>; 1218defm V_INTERP_P2_F32_e64 : VOP3Interp_Real_vi <0x271>; 1219defm V_INTERP_MOV_F32_e64 : VOP3Interp_Real_vi <0x272>; 1220 1221defm V_INTERP_P1LL_F16 : VOP3Interp_Real_vi <0x274>; 1222defm V_INTERP_P1LV_F16 : VOP3Interp_Real_vi <0x275>; 1223defm V_ADD_F64 : VOP3_Real_vi <0x280>; 1224defm V_MUL_F64 : VOP3_Real_vi <0x281>; 1225defm V_MIN_F64 : VOP3_Real_vi <0x282>; 1226defm V_MAX_F64 : VOP3_Real_vi <0x283>; 1227defm V_LDEXP_F64 : VOP3_Real_vi <0x284>; 1228defm V_MUL_LO_U32 : VOP3_Real_vi <0x285>; 1229 1230// removed from VI as identical to V_MUL_LO_U32 1231let isAsmParserOnly = 1 in { 1232defm V_MUL_LO_I32 : VOP3_Real_vi <0x285>; 1233} 1234 1235defm V_MUL_HI_U32 : VOP3_Real_vi <0x286>; 1236defm V_MUL_HI_I32 : VOP3_Real_vi <0x287>; 1237 1238defm V_READLANE_B32 : VOP3_Real_No_Suffix_vi <0x289>; 1239defm V_WRITELANE_B32 : VOP3_Real_No_Suffix_vi <0x28a>; 1240 1241defm V_LSHLREV_B64 : VOP3_Real_vi <0x28f>; 1242defm V_LSHRREV_B64 : VOP3_Real_vi <0x290>; 1243defm V_ASHRREV_I64 : VOP3_Real_vi <0x291>; 1244defm V_TRIG_PREOP_F64 : VOP3_Real_vi <0x292>; 1245 1246defm V_LSHL_ADD_U32 : VOP3_Real_vi <0x1fd>; 1247defm V_ADD_LSHL_U32 : VOP3_Real_vi <0x1fe>; 1248defm V_ADD3_U32 : VOP3_Real_vi <0x1ff>; 1249defm V_LSHL_OR_B32 : VOP3_Real_vi <0x200>; 1250defm V_AND_OR_B32 : VOP3_Real_vi <0x201>; 1251defm V_OR3_B32 : VOP3_Real_vi <0x202>; 1252defm V_PACK_B32_F16 : VOP3OpSel_Real_gfx9 <0x2a0>; 1253 1254defm V_XAD_U32 : VOP3_Real_vi <0x1f3>; 1255 1256defm V_MIN3_F16 : VOP3OpSel_Real_gfx9 <0x1f4>; 1257defm V_MIN3_I16 : VOP3OpSel_Real_gfx9 <0x1f5>; 1258defm V_MIN3_U16 : VOP3OpSel_Real_gfx9 <0x1f6>; 1259 1260defm V_MAX3_F16 : VOP3OpSel_Real_gfx9 <0x1f7>; 1261defm V_MAX3_I16 : VOP3OpSel_Real_gfx9 <0x1f8>; 1262defm V_MAX3_U16 : VOP3OpSel_Real_gfx9 <0x1f9>; 1263 1264defm V_MED3_F16 : VOP3OpSel_Real_gfx9 <0x1fa>; 1265defm V_MED3_I16 : VOP3OpSel_Real_gfx9 <0x1fb>; 1266defm V_MED3_U16 : VOP3OpSel_Real_gfx9 <0x1fc>; 1267 1268defm V_ADD_I16 : VOP3OpSel_Real_gfx9 <0x29e>; 1269defm V_SUB_I16 : VOP3OpSel_Real_gfx9 <0x29f>; 1270 1271defm V_MAD_U32_U16 : VOP3OpSel_Real_gfx9 <0x1f1>; 1272defm V_MAD_I32_I16 : VOP3OpSel_Real_gfx9 <0x1f2>; 1273 1274defm V_CVT_PKNORM_I16_F16 : VOP3OpSel_Real_gfx9 <0x299>; 1275defm V_CVT_PKNORM_U16_F16 : VOP3OpSel_Real_gfx9 <0x29a>; 1276