From d91f2ed6b07bd319d1bb2fb9aa959eb78638dce4 Mon Sep 17 00:00:00 2001 From: lioncash Date: Thu, 29 Dec 2022 16:31:57 +0000 Subject: [PATCH] OpcodeDispatcher: Handle VPUNPCKLBW --- .../Interface/Core/OpcodeDispatcher.cpp | 1 + .../Core/OpcodeDispatcher/Vector.cpp | 2 + .../Interface/Core/X86Tables/VEXTables.cpp | 2 +- unittests/ASM/VEX/vpunpcklbw.asm | 39 +++++++++++++++++++ 4 files changed, 43 insertions(+), 1 deletion(-) create mode 100644 unittests/ASM/VEX/vpunpcklbw.asm diff --git a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp index 6986d445b..e82faeb75 100644 --- a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp +++ b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp @@ -5904,6 +5904,7 @@ void OpDispatchBuilder::InstallHostSpecificOpcodeHandlers() { {OPD(1, 0b10, 0x5F), 1, &OpDispatchBuilder::AVXVectorScalarALUOp}, {OPD(1, 0b11, 0x5F), 1, &OpDispatchBuilder::AVXVectorScalarALUOp}, + {OPD(1, 0b01, 0x60), 1, &OpDispatchBuilder::VPUNPCKLOp<1>}, {OPD(1, 0b01, 0x61), 1, &OpDispatchBuilder::VPUNPCKLOp<2>}, {OPD(1, 0b01, 0x62), 1, &OpDispatchBuilder::VPUNPCKLOp<4>}, {OPD(1, 0b01, 0x63), 1, &OpDispatchBuilder::VPACKSSOp<2>}, diff --git a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp index d13418da6..ba426bfa8 100644 --- a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp +++ b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp @@ -905,6 +905,8 @@ void OpDispatchBuilder::VPUNPCKLOp(OpcodeArgs) { StoreResult(FPRClass, Op, Result, -1); } +template +void OpDispatchBuilder::VPUNPCKLOp<1>(OpcodeArgs); template void OpDispatchBuilder::VPUNPCKLOp<2>(OpcodeArgs); template diff --git a/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp b/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp index 12d78e7c2..142cdabb8 100644 --- a/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp +++ b/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp @@ -74,7 +74,7 @@ void InitializeVEXTables() { {OPD(1, 0b00, 0x57), 1, X86InstInfo{"VXORPS", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}}, {OPD(1, 0b01, 0x57), 1, X86InstInfo{"VXORPD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}}, - {OPD(1, 0b01, 0x60), 1, X86InstInfo{"VPUNPCKLBW", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}}, + {OPD(1, 0b01, 0x60), 1, X86InstInfo{"VPUNPCKLBW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}}, {OPD(1, 0b01, 0x61), 1, X86InstInfo{"VPUNPCKLWD", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}}, {OPD(1, 0b01, 0x62), 1, X86InstInfo{"VPUNPCKLDQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}}, {OPD(1, 0b01, 0x63), 1, X86InstInfo{"VPACKSSWB", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}}, diff --git a/unittests/ASM/VEX/vpunpcklbw.asm b/unittests/ASM/VEX/vpunpcklbw.asm new file mode 100644 index 000000000..3acfab363 --- /dev/null +++ b/unittests/ASM/VEX/vpunpcklbw.asm @@ -0,0 +1,39 @@ +%ifdef CONFIG +{ + "HostFeatures": ["AVX"], + "RegData": { + "XMM2": ["0x6545664667476848", "0x6141624263436444", "0x0000000000000000", "0x0000000000000000"], + "XMM3": ["0x6545664667476848", "0x6141624263436444", "0x0000000000000000", "0x0000000000000000"], + "XMM4": ["0x6545664667476848", "0x6141624263436444", "0x55DD55DD44CC44CC", "0x77FF77FF66EE66EE"], + "XMM5": ["0x6545664667476848", "0x6141624263436444", "0x55DD55DD44CC44CC", "0x77FF77FF66EE66EE"] + }, + "MemoryRegions": { + "0x100000000": "4096" + } +} +%endif + +lea rdx, [rel .data] + +vmovapd ymm0, [rdx] +vmovapd ymm1, [rdx + 32] + +vpunpcklbw xmm2, xmm0, [rdx + 32] +vpunpcklbw xmm3, xmm0, xmm1 + +vpunpcklbw ymm4, ymm0, [rdx + 32] +vpunpcklbw ymm5, ymm0, ymm1 + +hlt + +align 32 +.data: +dq 0x4142434445464748 +dq 0x5152535455565758 +dq 0xFFFFEEEEDDDDCCCC +dq 0xBBBBAAAA99998888 + +dq 0x6162636465666768 +dq 0x7172737475767778 +dq 0x7777666655554444 +dq 0x3333222211110000