diff options
Diffstat (limited to 'llvm/lib/Target/AMDGPU/VOP2Instructions.td')
| -rw-r--r-- | llvm/lib/Target/AMDGPU/VOP2Instructions.td | 21 |
1 files changed, 20 insertions, 1 deletions
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td index 030a6e1..9de7d6d 100644 --- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td +++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td @@ -925,6 +925,17 @@ let isAdd = 1 in { defm V_ADDC_U32 : VOP2bInst <"v_addc_u32", VOP2b_I32_I1_I32_I32_I1, null_frag, "v_addc_u32">; } +let isReMaterializable = 1 in { +let SubtargetPredicate = HasAddSubU64Insts, SchedRW = [Write64Bit] in { +defm V_ADD_U64 : VOP2Inst <"v_add_nc_u64", VOP_I64_I64_I64_ARITH>; +// We don't actually have something like V_SUBREV_U64 so V_SUB_U64 can't be treated as commutable. +let isCommutable = 0 in +defm V_SUB_U64 : VOP2Inst <"v_sub_nc_u64", VOP_I64_I64_I64_ARITH>; +} // End SubtargetPredicate = HasAddSubU64Insts, SchedRW = [Write64Bit] +let SubtargetPredicate = isGFX1250Plus, SchedRW = [WriteDouble] in +defm V_MUL_U64 : VOP2Inst <"v_mul_u64", VOP_I64_I64_I64, DivergentBinFrag<mul>>; +} // End isReMaterializable = 1 + } // End isCommutable = 1 // These are special and do not read the exec mask. @@ -1333,6 +1344,8 @@ def V_FMAAK_F64 : VOP2_Pseudo<"v_fmaak_f64", VOP_MADAK_F64, [], "">; } // End SubtargetPredicate = HasFmaakFmamkF64Insts, isReMaterializable = 1, FixedSize = 1, Size = 12, SchedRW = [Write64Bit] let SubtargetPredicate = HasPkFmacF16Inst in { +// FIXME: V_PK_FMAC_F16 is currently not used in instruction selection. +// If this changes, ensure the DPP variant is not used for GFX11+. defm V_PK_FMAC_F16 : VOP2Inst<"v_pk_fmac_f16", VOP_V2F16_V2F16_V2F16>; } // End SubtargetPredicate = HasPkFmacF16Inst @@ -1754,6 +1767,9 @@ multiclass VOP2_Real_FULL_with_name<GFXGen Gen, bits<6> op, string opName, VOP2_Realtriple_e64_with_name<Gen, op, opName, asmName>, VOP2_Real_NO_VOP3_with_name<Gen, op, opName, asmName>; +multiclass VOP2_Real_NO_DPP<GFXGen Gen, bits<6> op> : + VOP2_Real_e32<Gen, op>, VOP2_Real_e64<Gen, op>; + multiclass VOP2_Real_NO_DPP_with_name<GFXGen Gen, bits<6> op, string opName, string asmName> { defm NAME : VOP2_Real_e32_with_name<Gen, op, opName, asmName>, @@ -1843,6 +1859,9 @@ defm V_FMAC_F64 : VOP2_Real_FULL<GFX12Gen, 0x17>; defm V_FMAMK_F64 : VOP2Only_Real_MADK64<GFX1250Gen, 0x23>; defm V_FMAAK_F64 : VOP2Only_Real_MADK64<GFX1250Gen, 0x24>; +defm V_ADD_U64 : VOP2_Real_FULL<GFX1250Gen, 0x28>; +defm V_SUB_U64 : VOP2_Real_FULL<GFX1250Gen, 0x29>; +defm V_MUL_U64 : VOP2_Real_NO_DPP<GFX1250Gen, 0x2a>; //===----------------------------------------------------------------------===// // GFX11. @@ -1887,7 +1906,7 @@ multiclass VOP2_Real_FULL_with_name_gfx11_gfx12<bits<6> op, string opName, VOP2_Real_FULL_with_name<GFX12Gen, op, opName, asmName>; multiclass VOP2_Real_e32_gfx11_gfx12<bits<6> op> : - VOP2Only_Real<GFX11Gen, op>, VOP2Only_Real<GFX12Gen, op>; + VOP2Only_Real_e32<GFX11Gen, op>, VOP2Only_Real_e32<GFX12Gen, op>; multiclass VOP3Only_Realtriple_gfx11_gfx12<bits<10> op> : VOP3Only_Realtriple<GFX11Gen, op>, VOP3Only_Realtriple<GFX12Gen, op>; |