Merge pull request #2297 from lioncash/lps

OpcodeDispatcher: Handle VPUNPCKLBW/VPUNPCKLWD/VPUNPCKLDQ/VPUNPCKLQDQ
This commit is contained in:
Ryan Houdek authored and GitHub committed 2022-12-29 14:49:11 -08:00
commit 6bc1c3fc30
7 files changed
+168 -4

No files matched your search

@@ -5904,12 +5904,16 @@ void OpDispatchBuilder::InstallHostSpecificOpcodeHandlers() {
{OPD(1, 0b10, 0x5F), 1, &OpDispatchBuilder::AVXVectorScalarALUOp<IR::OP_VFMAX, 4>},
{OPD(1, 0b11, 0x5F), 1, &OpDispatchBuilder::AVXVectorScalarALUOp<IR::OP_VFMAX, 8>},
{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>},
{OPD(1, 0b01, 0x64), 1, &OpDispatchBuilder::AVXVectorALUOp<IR::OP_VCMPGT, 1>},
{OPD(1, 0b01, 0x65), 1, &OpDispatchBuilder::AVXVectorALUOp<IR::OP_VCMPGT, 2>},
{OPD(1, 0b01, 0x66), 1, &OpDispatchBuilder::AVXVectorALUOp<IR::OP_VCMPGT, 4>},
{OPD(1, 0b01, 0x67), 1, &OpDispatchBuilder::VPACKUSOp<2>},
{OPD(1, 0b01, 0x6B), 1, &OpDispatchBuilder::VPACKSSOp<4>},
{OPD(1, 0b01, 0x6C), 1, &OpDispatchBuilder::VPUNPCKLOp<8>},
{OPD(1, 0b01, 0x6E), 1, &OpDispatchBuilder::MOVBetweenGPR_FPR},
{OPD(1, 0b01, 0x6F), 1, &OpDispatchBuilder::VMOVAPS_VMOVAPD_Op},
@@ -905,6 +905,10 @@ void OpDispatchBuilder::VPUNPCKLOp(OpcodeArgs) {
StoreResult(FPRClass, Op, Result, -1);
}
template
void OpDispatchBuilder::VPUNPCKLOp<1>(OpcodeArgs);
template
void OpDispatchBuilder::VPUNPCKLOp<2>(OpcodeArgs);
template
void OpDispatchBuilder::VPUNPCKLOp<4>(OpcodeArgs);
template
@@ -74,9 +74,9 @@ 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, 0x61), 1, X86InstInfo{"VPUNPCKLWD", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
{OPD(1, 0b01, 0x62), 1, X86InstInfo{"VPUNPCKLDQ", 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}},
{OPD(1, 0b01, 0x64), 1, X86InstInfo{"VPCMPGTB", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}},
{OPD(1, 0b01, 0x65), 1, X86InstInfo{"VPCMPGTW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}},
@@ -175,7 +175,7 @@ void InitializeVEXTables() {
{OPD(1, 0b01, 0x69), 1, X86InstInfo{"VPUNPCKHWD", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
{OPD(1, 0b01, 0x6A), 1, X86InstInfo{"VPUNPCKHDQ", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
{OPD(1, 0b01, 0x6B), 1, X86InstInfo{"VPACKSSDW", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}},
{OPD(1, 0b01, 0x6C), 1, X86InstInfo{"VPUNPCKLQDQ", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
{OPD(1, 0b01, 0x6C), 1, X86InstInfo{"VPUNPCKLQDQ", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_VEX_1ST_SRC | FLAGS_XMM_FLAGS, 0, nullptr}},
{OPD(1, 0b01, 0x6D), 1, X86InstInfo{"VPUNPCKHQDQ", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
{OPD(1, 0b01, 0x6E), 1, X86InstInfo{"VMOV*", TYPE_INST, GenFlagsDstSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS | FLAGS_SF_SRC_GPR, 0, nullptr}},
+39
View File
@@ -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
+39
View File
@@ -0,0 +1,39 @@
%ifdef CONFIG
{
"HostFeatures": ["AVX"],
"RegData": {
"XMM2": ["0x6566676845464748", "0x6162636441424344", "0x0000000000000000", "0x0000000000000000"],
"XMM3": ["0x6566676845464748", "0x6162636441424344", "0x0000000000000000", "0x0000000000000000"],
"XMM4": ["0x6566676845464748", "0x6162636441424344", "0xAAAAAAAAEEEEEEEE", "0xBBBBBBBBFFFFFFFF"],
"XMM5": ["0x6566676845464748", "0x6162636441424344", "0xAAAAAAAAEEEEEEEE", "0xBBBBBBBBFFFFFFFF"]
},
"MemoryRegions": {
"0x100000000": "4096"
}
}
%endif
lea rdx, [rel .data]
vmovapd ymm0, [rdx]
vmovapd ymm1, [rdx + 32]
vpunpckldq xmm2, xmm0, [rdx + 32]
vpunpckldq xmm3, xmm0, xmm1
vpunpckldq ymm4, ymm0, [rdx + 32]
vpunpckldq ymm5, ymm0, ymm1
hlt
align 32
.data:
dq 0x4142434445464748
dq 0x5152535455565758
dq 0xFFFFFFFFEEEEEEEE
dq 0xDDDDDDDDCCCCCCCC
dq 0x6162636465666768
dq 0x7172737475767778
dq 0xBBBBBBBBAAAAAAAA
dq 0x9999999988888888
+39
View File
@@ -0,0 +1,39 @@
%ifdef CONFIG
{
"HostFeatures": ["AVX"],
"RegData": {
"XMM2": ["0x4142434445464748", "0x6162636465666768", "0x0000000000000000", "0x0000000000000000"],
"XMM3": ["0x4142434445464748", "0x6162636465666768", "0x0000000000000000", "0x0000000000000000"],
"XMM4": ["0x4142434445464748", "0x6162636465666768", "0xFFFFFFFFFFFFFFFF", "0xBBBBBBBBBBBBBBBB"],
"XMM5": ["0x4142434445464748", "0x6162636465666768", "0xFFFFFFFFFFFFFFFF", "0xBBBBBBBBBBBBBBBB"]
},
"MemoryRegions": {
"0x100000000": "4096"
}
}
%endif
lea rdx, [rel .data]
vmovapd ymm0, [rdx]
vmovapd ymm1, [rdx + 32]
vpunpcklqdq xmm2, xmm0, [rdx + 32]
vpunpcklqdq xmm3, xmm0, xmm1
vpunpcklqdq ymm4, ymm0, [rdx + 32]
vpunpcklqdq ymm5, ymm0, ymm1
hlt
align 32
.data:
dq 0x4142434445464748
dq 0x5152535455565758
dq 0xFFFFFFFFFFFFFFFF
dq 0xEEEEEEEEEEEEEEEE
dq 0x6162636465666768
dq 0x7172737475767778
dq 0xBBBBBBBBBBBBBBBB
dq 0xCCCCCCCCCCCCCCCC
+39
View File
@@ -0,0 +1,39 @@
%ifdef CONFIG
{
"HostFeatures": ["AVX"],
"RegData": {
"XMM2": ["0x6566454667684748", "0x6162414263644344", "0x0000000000000000", "0x0000000000000000"],
"XMM3": ["0x6566454667684748", "0x6162414263644344", "0x0000000000000000", "0x0000000000000000"],
"XMM4": ["0x6566454667684748", "0x6162414263644344", "0x5555DDDD4444CCCC", "0x7777FFFF6666EEEE"],
"XMM5": ["0x6566454667684748", "0x6162414263644344", "0x5555DDDD4444CCCC", "0x7777FFFF6666EEEE"]
},
"MemoryRegions": {
"0x100000000": "4096"
}
}
%endif
lea rdx, [rel .data]
vmovapd ymm0, [rdx]
vmovapd ymm1, [rdx + 32]
vpunpcklwd xmm2, xmm0, [rdx + 32]
vpunpcklwd xmm3, xmm0, xmm1
vpunpcklwd ymm4, ymm0, [rdx + 32]
vpunpcklwd ymm5, ymm0, ymm1
hlt
align 32
.data:
dq 0x4142434445464748
dq 0x5152535455565758
dq 0xFFFFEEEEDDDDCCCC
dq 0xBBBBAAAA99998888
dq 0x6162636465666768
dq 0x7172737475767778
dq 0x7777666655554444
dq 0x3333222211110000