Fast sine approximation using SSE instructions.

fastsin.cc

char *ctable = "\x53\xBE\x25\xE6\xBC\xAB\xC4\x3E\x00\x00\x00\x00\x00\x00\x00"
"\x00\x7B\xFC\xAB\xE4\x24\xAA\xC4\x3E\x16\x46\x6F\x57\xCF\x00\xDF\x3E\xFB\x53"
"\x29\x1F\x5D\xA5\xC4\x3E\x47\x5F\xF3\x64\x6B\xFE\xEE\x3E\x2D\xE8\x53\x52\x66"
"\x9D\xC4\x3E\x3F\x66\xCA\xCB\xD3\x3B\xF7\x3E\x18\x49\x90\xB8\x41\x92\xC4\x3E"
"\x3B\xF9\x14\xB6\xDC\xF4\xFE\x3E\x53\xF8\xC0\x09\xF1\x83\xC4\x3E\x8B\xC8\x87"
"\xC9\x8F\x54\x03\x3F\xF2\x93\x02\x7B\x76\x72\xC4\x3E\x1B\x4C\x71\x20\xB6\x2B"
"\x07\x3F\xEA\xB1\x54\xBE\xD4\x5D\xC4\x3E\x84\x64\x87\xC4\x49\xFF\x0A\x3F\x64"
"\x79\x2F\x02\x0F\x46\xC4\x3E\xC1\x78\x97\xA7\xB3\xCE\x0E\x3F\x65\x0A\x06\xF1"
"\x28\x2B\xC4\x3E\x93\x8D\xE3\xAF\xAE\x4C\x11\x3F\x28\xC7\xB5\xB0\x26\x0D\xC4"
"\x3E\x9F\xE9\x62\x9F\x58\x2F\x13\x3F\x7B\x95\xE2\xE1\x0C\xEC\xC3\x3E\x1D\xFE"
"\x70\x34\x0D\x0F\x15\x3F\x5E\x31\x40\x9F\xE0\xC7\xC3\x3E\x75\x94\xFE\x75\x82"
"\xEB\x16\x3F\x12\xAD\xC8\x7C\xA7\xA0\xC3\x3E\x78\x41\x2E\xEB\x6E\xC4\x18\x3F"
"\xAD\x3D\xE0\x86\x67\x76\xC3\x3E\x58\xD0\xA8\xA6\x89\x99\x1A\x3F\x19\x76\x66"
"\x41\x27\x49\xC3\x3E\xA7\x29\xDC\x51\x8A\x6A\x1C\x3F\x6C\x15\xB5\xA6\xED\x18"
"\xC3\x3E\x8A\xFA\x22\x38\x29\x37\x1E\x3F\x1D\x90\x8C\x26\xC2\xE5\xC2\x3E\xA7"
"\x63\xD3\x51\x1F\xFF\x1F\x3F\xB5\x7E\xEE\xA4\xAC\xAF\xC2\x3E\xB1\x7D\x99\x27"
"\x13\xE1\x20\x3F\x29\x1F\xE6\x78\xB5\x76\xC2\x3E\x7C\x39\xA7\x51\xFC\xBF\x21"
"\x3F\xE9\x18\x3F\x6B\xE5\x3A\xC2\x3E\x3F\x99\x59\xC7\x28\x9C\x22\x3F\x76\xB6"
"\x2A\xB5\x45\xFC\xC1\x3E\x2B\xA2\x04\x95\x76\x75\x23\x3F\xE8\xC9\xD3\xFE\xDF"
"\xBA\xC1\x3E\xD1\xF6\x45\x38\xC4\x4B\x24\x3F\x96\x74\xE1\x5D\xBE\x76\xC1\x3E"
"\x2B\xAB\x2F\xA5\xF0\x1E\x25\x3F\xA0\x0C\xE9\x53\xEB\x2F\xC1\x3E\x78\xD4\x60"
"\x4B\xDB\xEE\x25\x3F\xC5\x5D\xCF\xCC\x71\xE6\xC0\x3E\xBF\x0B\x0B\x1B\x64\xBB"
"\x26\x3F\x79\x85\x19\x1D\x5D\x9A\xC0\x3E\xC1\x1C\xE4\x89\x6B\x84\x27\x3F\xB2"
"\xAC\x2D\x00\xB9\x4B\xC0\x3E\x27\x1E\x03\x98\xD2\x49\x28\x3F\xDC\xC8\x07\x2D"
"\x23\xF5\xBF\x3E\xEF\x32\xA8\xD4\x7A\x0B\x29\x3F\x9E\xD8\x8E\xC7\xE6\x4D\xBF"
"\x3E\x6D\x38\xEE\x62\x46\xC9\x29\x3F\x63\x5A\xD1\x99\xD6\xA1\xBE\x3E\x76\xA7"
"\x65\xFE\x17\x83\x2A\x3F\xB1\x67\x40\x2C\x0D\xF1\xBD\x3E\x11\xF3\x97\xFF\xD2"
"\x38\x2B\x3F\x7E\xF4\xC5\xC1\xA5\x3B\xBD\x3E\x62\xB2\x72\x60\x5B\xEA\x2B\x3F"
"\x85\xA0\x90\x53\xBC\x81\xBC\x3E\x76\xE7\x99\xC0\x95\x97\x2C\x3F\x57\x6D\xC3"
"\x8C\x6D\xC3\xBB\x3E\x4F\xB8\xA0\x69\x67\x40\x2D\x3F\x65\x03\x0A\xC6\xD6\x00"
"\xBB\x3E\x8B\xF3\x27\x53\xB6\xE4\x2D\x3F\x87\x34\x12\x01\x16\x3A\xBA\x3E\x09"
"\xBE\xE1\x26\x69\x84\x2E\x3F\xB3\x6E\xEB\xE3\x49\x6F\xB9\x3E\x19\xCB\x79\x44"
"\x67\x1F\x2F\x3F\x7B\xD5\x4C\xB4\x91\xA0\xB8\x3E\xE8\x84\x61\xC5\x98\xB5\x2F"
"\x3F\xF1\xBC\xC2\x52\x0D\xCE\xB7\x3E\xAC\xC7\x3F\x40\x73\x23\x30\x3F\x4C\x44"
"\xC4\x35\xDD\xF7\xB6\x3E\xD7\x09\xE1\x06\x9D\x69\x30\x3F\x5C\xD2\xB1\x64\x22"
"\x1E\xB6\x3E\x19\xA4\xC9\xE4\x3E\xAD\x30\x3F\x67\x39\xBD\x72\xFE\x40\xB5\x3E"
"\x9A\xD5\x17\x6C\x4E\xEE\x30\x3F\x86\x4B\xBC\x79\x93\x60\xB4\x3E\xF9\xB6\x6E"
"\x94\xC1\x2C\x31\x3F\xD8\xAB\xE6\x14\x04\x7D\xB3\x3E\x15\x48\x82\xBC\x8E\x68"
"\x31\x3F\x32\xAC\x7F\x5B\x73\x96\xB2\x3E\x34\x99\x93\xAB\xAC\xA1\x31\x3F\xED"
"\x09\x6D\xDB\x04\xAD\xB1\x3E\xD9\xD4\xDC\x92\x12\xD8\x31\x3F\x68\x5F\xBB\x93"
"\xDC\xC0\xB0\x3E\x25\xF2\xEC\x0E\xB8\x0B\x32\x3F\x80\x45\x22\xDE\x3D\xA4\xAF"
"\x3E\x2D\xDA\xF2\x28\x95\x3C\x32\x3F\x9B\x19\x22\x7C\xE1\xC1\xAD\x3E\x38\xCD"
"\xF7\x57\xA2\x6A\x32\x3F\x3D\x8C\x59\x63\xEE\xDA\xAB\x3E\x57\xD7\x08\x82\xD8"
"\x95\x32\x3F\x0B\x74\xD1\xAA\xAF\xEF\xA9\x3E\x8D\x26\x4F\xFD\x30\xBE\x32\x3F"
"\x56\xE4\x23\x13\x71\x00\xA8\x3E\x45\x17\x17\x91\xA5\xE3\x32\x3F\x48\xBE\xCD"
"\xFA\x7E\x0D\xA6\x3E\x7D\xCE\xC5\x76\x30\x06\x33\x3F\x51\xE9\x67\x52\x26\x17"
"\xA4\x3E\xD7\x3B\xBD\x5A\xCC\x25\x33\x3F\xB5\x03\xCA\x90\xB4\x1D\xA2\x3E\x57"
"\x60\x2E\x5D\x74\x42\x33\x3F\x8F\x5F\x18\xA7\x77\x21\xA0\x3E\x72\xB9\xD9\x12"
"\x24\x5C\x33\x3F\xCC\x47\x7E\xE9\x7B\x45\x9C\x3E\xAB\xB2\xBD\x85\xD7\x72\x33"
"\x3F\x85\xD8\xB8\x76\xAC\x43\x98\x3E\xF3\x02\xB3\x35\x8B\x86\x33\x3F\xC6\x87"
"\x36\x25\x1F\x3E\x94\x3E\x96\xDD\xF6\x18\x3C\x97\x33\x3F\xEA\xB7\x02\xB8\x72"
"\x35\x90\x3E\x78\xE1\xA2\x9C\xE7\xA4\x33\x3F\x3C\x79\xC5\xDA\x8C\x54\x88\x3E"
"\x1C\xB4\x12\xA5\x8B\xAF\x33\x3F\xA6\xFF\x8E\xCC\x73\x3A\x80\x3E\xC4\x38\x37"
"\x8E\x26\xB7\x33\x3F\x01\xBE\xBF\x36\xB4\x3B\x70\x3E\xE5\x56\xD7\x2B\xB7\xBB"
"\x33\x3F\x00\x00\x00\x00\x00\x00\x00\x00\xDE\x45\xBE\xC9\x3C\xBD\x33";

char *one_m230 = "\x00\x00\x80\xff\xff\xff\xef\x3f"
"\x00\x00\x00\x00\x00\x00\x00\x00";

char *pitable = "\x39\x79\x3E\x6D\x30\x5F\x44\x40\x48\x5F\x9D\xF0\xA7\x54\x70"
"\x3D\x39\x79\x3E\x6D\x30\x5F\x44\x40\x48\x5F\x9D\xF0\xA7\x54\x70\x3D\x88\x0C"
"\x3E\x72\x1B\xCC\xE7\x3F\xF5\xD1\x57\x27\xFC\x29\x15\x3D\x54\x70\x3D\xE4\xE4"
"\x36\x58\x3F\x9A\xFA\xE8\xAB\x13\xFE\xA4\x3C\x29\x15\x3D\x20\x72\x72\xEB\x3E"
"\x9C\xA6\x3E\xFA\xEA\x84\x4F\x3C\xF8\x83\x3C\xA9\x20\x72\x62\x3E\xE0\x6E\x9A"
"\xFA\xE8\xAB\xA3\x3B\x27\x1C\x3C\x4F\xA9\x20\xE2\x3D\x37\x70\x37\x4D\x7D\xF4"
"\x35\x3B\xF4\x35\x3B\x27\xFC\x29\x15\x3D\x29\xB6\x0D\xDC\x4D\x53\x5F\x3A\x46"
"\xFF\x3A\x75\xC2\x9F\xD2\x3C\xA5\xD8\x36\x70\x37\x4D\x3D\x3A\x50\x5F\x3A\x7D"
"\x75\xC2\x5F\x3C\xCD\x4A\xB1\x6D\xE0\x6E\xAA\x39\x4D\xD3\x39\xFA\xE8\xAB\xA3"
"\x3B\x79\x32\x2B\xC5\xB6\x81\x0B\x39\xDD\x94\x39\xA9\x8F\xBE\x6A\x3B\x88\x27"
"\xB3\x52\x6C\x1B\xC8\x38\xC0\x1D\x39\xA6\xA9\x8F\xEE\x3A\x72\x88\x27\xB3\x52"
"\x6C\x4B\x38\x60\x4B\x38\xDC\x4D\x53\x5F\x3A\x10\xE4\x10\x4F\x66\xA5\xB8\x37"
"\xC5\x06\x38\x0D\xDC\x4D\xD3\x39\xFE\x41\x90\x43\x3C\x99\x15";

float sin(float x) {

    float res;

__asm {
        movss xmm0, x
        push rsi
        //add rsp, 0x8
        stmxcsr [rsp]
        mov eax, [rsp]
        and eax, 0xffff9fff
        cmp eax, [rsp]
        jnz label_1804
label_15d6:
        pshuflw xmm1,xmm0,0x44
        movd ecx,xmm0
        cvtps2pd xmm1,xmm1
        mov rax,0x40445f306e000000
        movq xmm2,rax
        mov r10d,0x7fffffff
        mov r8d,0x49ffffff
        and r10d,ecx
        sub r8d,r10d
        sub r10d,0x39800000
        or r8d,r10d
        jl label_170D
        mov r11,0x4338000000000000
        movq xmm4,r11
        mulsd xmm2,xmm1
        mov rdx,0xbe5b1bbead603d8b
        movq xmm3,rdx
        movapd xmm5,xmm2
        addsd xmm2,xmm4
        movd r9d,xmm2
        subsd xmm2,xmm4
        mulsd xmm1,xmm3
label_164a:
        mov r10,[ctable]
        mov eax,0x180
        mov rdx,0x40c37423899a1558
        movq xmm4,rdx
        subsd xmm5,xmm2
        add r9d,r9d
        mov r8d,r9d
        sar r9b,0x7
        and eax,r8d
        add r8b,r9b
        xor r8b,r9b
        addsd xmm1,xmm5
        and r8d,0xfe
        movsd xmm3,[r10+r8*8]
        mov r11,0x40a9f02f6222c720
        movq xmm0,r11
        movddup xmm2,xmm1
        mulsd xmm1,xmm1
        movsd xmm5,[r10+r8*8+0x8]
        mov r9d,eax
        add r9d,0x80
        and r9d,0x100
        shl r9,0x37
        mulsd xmm3,xmm2
        movq xmm2,r9
        subsd xmm0,xmm1
        and eax,0x100
        shl rax,0x37
        subsd xmm4,xmm1
        movq xmm1,rax
        xorpd xmm4,xmm2
        mulsd xmm0,xmm5
        mulsd xmm3,xmm4
        xorpd xmm0,xmm1
        addsd xmm0,xmm3
        mov eax,[rsp]
        and eax,0xffff9fff
        cmp eax,[rsp]
        jnz label_1812
        cvtpd2ps xmm0,xmm0
        jmp label_182a
label_170D:
        and ecx,0x7fffffff
        cmp ecx,0x3e000000
        ja label_1741
        cvtss2sd xmm0,xmm0
        mulsd xmm0,[one_m230]
        mov eax,[rsp]
        and eax,0xffff9fff
        cmp eax,[rsp]
        jnz label_1812
label_17fe:
        cvtsd2ss xmm0,xmm0
        jmp label_182a
label_1741:
        movd eax,xmm0
        and eax,0x7f800000
        cmp eax,0x7f800000
        jz label_17d5
        shr r10d,0x17
        sub r10d,+0x1b
        and r10d,0xfff8
        mov rax, [pitable]
        movsd xmm3,[rax+r10*2]
        movsd xmm5,[rax+r10*2+0x8]
        mov r9,0xffffffffff000000
        movq xmm2,r9
        mov r11,0x4338000000000000
        movq xmm4,r11
        andpd xmm2,xmm3
        psllq xmm3,0x28
        mulsd xmm2,xmm1
        mulsd xmm3,xmm1
        mulsd xmm1,xmm5
        movapd xmm0,xmm2
        addsd xmm2,xmm3
        movapd xmm5,xmm2
        subsd xmm0,xmm2
        addsd xmm2,xmm4
        addsd xmm3,xmm0
        movd r9d,xmm2
        subsd xmm2,xmm4
        addsd xmm1,xmm3
        jmp label_164a
label_17d5:
        subss xmm0,xmm0
        mov eax,[rsp]
        and eax,0xffff9fff
        cmp eax,[rsp]
        jz label_17fc
        stmxcsr [rsp+0x4]
        mov eax,[rsp]
        and eax,0x6000
        or [rsp+0x4],eax
        ldmxcsr [rsp+0x4]
label_17fc:
        jmp label_182a
        cvtsd2ss xmm0,xmm0
        jmp label_182a
label_1804:
        mov [rsp+0x4],eax
        ldmxcsr [rsp+0x4]
        jmp label_15d6
label_1812:
        stmxcsr [rsp+0x4]
        mov eax,[rsp]
        and eax,0x6000
        or [rsp+0x4],eax
        ldmxcsr [rsp+0x4]
        jmp label_17fe
label_182a:
        pop rcx
        movss res, xmm0
    }

return res;

}