OpcodeDispatcher: Handle VMOVSHDUP

This commit is contained in:
lioncash 2022-11-22 03:01:42 +00:00
parent bc20f1e684
commit 815cdc5b3c
5 changed files with 60 additions and 1 deletions

View File

@ -5701,6 +5701,7 @@ void OpDispatchBuilder::InstallHostSpecificOpcodeHandlers() {
{OPD(1, 0b00, 0x16), 1, &OpDispatchBuilder::VMOVHPOp},
{OPD(1, 0b01, 0x16), 1, &OpDispatchBuilder::VMOVHPOp},
{OPD(1, 0b10, 0x16), 1, &OpDispatchBuilder::VMOVSHDUPOp},
{OPD(1, 0b00, 0x17), 1, &OpDispatchBuilder::VMOVHPOp},
{OPD(1, 0b01, 0x17), 1, &OpDispatchBuilder::VMOVHPOp},

View File

@ -411,6 +411,8 @@ public:
void VMOVHPOp(OpcodeArgs);
void VMOVLPOp(OpcodeArgs);
void VMOVSHDUPOp(OpcodeArgs);
// X87 Ops
template<size_t width>
void FLD(OpcodeArgs);

View File

@ -146,6 +146,21 @@ void OpDispatchBuilder::MOVSHDUPOp(OpcodeArgs) {
StoreResult(FPRClass, Op, Result, -1);
}
void OpDispatchBuilder::VMOVSHDUPOp(OpcodeArgs) {
OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
const auto SrcSize = GetSrcSize(Op);
const auto Is256Bit = SrcSize == Core::CPUState::XMM_AVX_REG_SIZE;
OrderedNode *Result = _VInsElement(SrcSize, 4, 2, 3, Src, Src);
Result = _VInsElement(SrcSize, 4, 0, 1, Result, Src);
if (Is256Bit) {
Result = _VInsElement(SrcSize, 4, 4, 5, Result, Src);
Result = _VInsElement(SrcSize, 4, 6, 7, Result, Src);
}
StoreResult_WithOpSize(FPRClass, Op, Op->Dest, Result, 32, -1);
}
void OpDispatchBuilder::MOVSLDUPOp(OpcodeArgs) {
OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, 8);
OrderedNode *Result = _VInsElement(16, 4, 3, 2, Src, Src);

View File

@ -43,7 +43,7 @@ void InitializeVEXTables() {
{OPD(1, 0b00, 0x16), 1, X86InstInfo{"VMOVHPS", TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_XMM_FLAGS | FLAGS_VEX_1ST_SRC, 0, nullptr}},
{OPD(1, 0b01, 0x16), 1, X86InstInfo{"VMOVHPD", TYPE_INST, GenFlagsSizes(SIZE_128BIT, SIZE_64BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_XMM_FLAGS | FLAGS_VEX_1ST_SRC, 0, nullptr}},
{OPD(1, 0b10, 0x16), 1, X86InstInfo{"VMOVSHDUP", TYPE_UNDEC, FLAGS_NONE, 0, nullptr}},
{OPD(1, 0b10, 0x16), 1, X86InstInfo{"VMOVSHDUP", TYPE_INST, GenFlagsSameSize(SIZE_128BIT) | FLAGS_MODRM | FLAGS_XMM_FLAGS, 0, nullptr}},
{OPD(1, 0b00, 0x17), 1, X86InstInfo{"VMOVHPS", TYPE_INST, GenFlagsSizes(SIZE_64BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_SF_MOD_DST | FLAGS_XMM_FLAGS, 0, nullptr}},
{OPD(1, 0b01, 0x17), 1, X86InstInfo{"VMOVHPD", TYPE_INST, GenFlagsSizes(SIZE_64BIT, SIZE_128BIT) | FLAGS_MODRM | FLAGS_SF_MOD_MEM_ONLY | FLAGS_SF_MOD_DST | FLAGS_XMM_FLAGS, 0, nullptr}},

View File

@ -0,0 +1,41 @@
%ifdef CONFIG
{
"HostFeatures": ["AVX"],
"RegData": {
"XMM1": ["0xEEEEEEEEEEEEEEEE", "0xCCCCCCCCCCCCCCCC", "0xAAAAAAAAAAAAAAAA", "0x0808080808080808"],
"XMM2": ["0xEEEEEEEEEEEEEEEE", "0xCCCCCCCCCCCCCCCC", "0x0000000000000000", "0x0000000000000000"],
"XMM3": ["0xEEEEEEEEFFFFFFFF", "0xCCCCCCCCDDDDDDDD", "0xAAAAAAAABBBBBBBB", "0x0808080809090909"],
"XMM4": ["0xEEEEEEEEEEEEEEEE", "0xCCCCCCCCCCCCCCCC", "0xAAAAAAAAAAAAAAAA", "0x0808080808080808"],
"XMM5": ["0xEEEEEEEEEEEEEEEE", "0xCCCCCCCCCCCCCCCC", "0x0000000000000000", "0x0000000000000000"],
"XMM6": ["0xEEEEEEEEEEEEEEEE", "0xCCCCCCCCCCCCCCCC", "0xAAAAAAAAAAAAAAAA", "0x0808080808080808"],
"XMM7": ["0xEEEEEEEEEEEEEEEE", "0xCCCCCCCCCCCCCCCC", "0x0000000000000000", "0x0000000000000000"]
}
}
%endif
lea rdx, [rel .data]
;; Broadcast across self
vmovaps ymm1, [rdx]
vmovshdup ymm1, ymm1
; 128-bit version
vmovaps xmm2, [rdx]
vmovshdup xmm2, xmm2
;; Broadcast from different registers
vmovaps ymm3, [rdx]
vmovshdup ymm4, ymm3
; 128-bit version
vmovshdup xmm5, xmm3
;; Broadcast from memory
vmovshdup ymm6, [rdx]
; 128-bit version
vmovshdup xmm7, [rdx]
hlt
align 32
.data:
db 0xFF, 0xFF, 0xFF, 0xFF, 0xEE, 0xEE, 0xEE, 0xEE, 0xDD, 0xDD, 0xDD, 0xDD, 0xCC, 0xCC, 0xCC, 0xCC
db 0xBB, 0xBB, 0xBB, 0xBB, 0xAA, 0xAA, 0xAA, 0xAA, 0x09, 0x09, 0x09, 0x09, 0x08, 0x08, 0x08, 0x08