FEX/unittests/ASM/VEX/vrsqrtss.asm
2022-12-12 16:30:34 +00:00

81 lines
2.5 KiB
NASM

%ifdef CONFIG
{
"HostFeatures": ["AVX"],
"RegData": {
"XMM0": ["0x414243443F800000", "0xEEEEEEEEEEEEEEEE", "0x0000000000000000", "0x0000000000000000"],
"XMM1": ["0x414243443F000000", "0xDDDDDDDDDDDDDDDD", "0x0000000000000000", "0x0000000000000000"],
"XMM2": ["0x414243443EAAAAAB", "0xCCCCCCCCCCCCCCCC", "0x0000000000000000", "0x0000000000000000"],
"XMM3": ["0x414243443E800000", "0xBBBBBBBBBBBBBBBB", "0x0000000000000000", "0x0000000000000000"],
"XMM4": ["0x414243443F800000", "0xAAAAAAAAAAAAAAAA", "0x0000000000000000", "0x0000000000000000"],
"XMM5": ["0x414243443F000000", "0xAAAAAAAAAAAAAAAA", "0x0000000000000000", "0x0000000000000000"],
"XMM6": ["0x414243443EAAAAAB", "0xAAAAAAAAAAAAAAAA", "0x0000000000000000", "0x0000000000000000"],
"XMM7": ["0x414243443E800000", "0xAAAAAAAAAAAAAAAA", "0x0000000000000000", "0x0000000000000000"],
"XMM8": ["0x414243443F800000", "0xDDDDDDDDDDDDDDDD", "0x0000000000000000", "0x0000000000000000"],
"XMM9": ["0x414243443F000000", "0xCCCCCCCCCCCCCCCC", "0x0000000000000000", "0x0000000000000000"],
"XMM10": ["0x414243443EAAAAAB", "0xBBBBBBBBBBBBBBBB", "0x0000000000000000", "0x0000000000000000"],
"XMM11": ["0x414243443E800000", "0xAAAAAAAAAAAAAAAA", "0x0000000000000000", "0x0000000000000000"]
},
"MemoryRegions": {
"0x100000000": "4096"
}
}
%endif
lea rdx, [rel .data]
vmovapd ymm0, [rdx + 32 * 0]
vmovapd ymm1, [rdx + 32 * 1]
vmovapd ymm2, [rdx + 32 * 2]
vmovapd ymm3, [rdx + 32 * 3]
vmovapd ymm4, [rdx + 32 * 4]
vmovapd ymm5, [rdx + 32 * 4]
vmovapd ymm6, [rdx + 32 * 4]
vmovapd ymm7, [rdx + 32 * 4]
; Same register
vrsqrtss xmm0, xmm0, xmm0
vrsqrtss xmm1, xmm1, xmm1
vrsqrtss xmm2, xmm2, xmm2
vrsqrtss xmm3, xmm3, xmm3
; Memory operand
vrsqrtss xmm4, xmm4, [rdx + 32 * 0]
vrsqrtss xmm5, xmm5, [rdx + 32 * 1]
vrsqrtss xmm6, xmm6, [rdx + 32 * 2]
vrsqrtss xmm7, xmm7, [rdx + 32 * 3]
; Memory operand different source register
vrsqrtss xmm8, xmm1, [rdx + 32 * 0]
vrsqrtss xmm9, xmm2, [rdx + 32 * 1]
vrsqrtss xmm10, xmm3, [rdx + 32 * 2]
vrsqrtss xmm11, xmm4, [rdx + 32 * 3]
hlt
align 32
.data:
dq 0x414243443F800000 ; 1.0
dq 0xEEEEEEEEEEEEEEEE
dq 0x5152535455565758
dq 0x5152535455565758
dq 0x4142434440800000 ; 4.0
dq 0xDDDDDDDDDDDDDDDD
dq 0x5152535455565758
dq 0x5152535455565758
dq 0x4142434441100000 ; 9.0
dq 0xCCCCCCCCCCCCCCCC
dq 0x5152535455565758
dq 0x5152535455565758
dq 0x4142434441800000 ; 16.0
dq 0xBBBBBBBBBBBBBBBB
dq 0x5152535455565758
dq 0x5152535455565758
dq 0x4142434441C80000 ; 25.0
dq 0xAAAAAAAAAAAAAAAA
dq 0x5152535455565758
dq 0x5152535455565758