void Decompress(u16 *in, u16 *out, u32 size)
{
    u32 flag = 0x8000, jmp, len = 3;    // 3 assignment shouldn't be needed
    u16 *end = &in[size];

    while (in != end)
    {
        // cache next 16 flags
        if (flag & 0x8000) flag = (*in++ << 16) | 1;

        // not literal
        if (flag & 0x80000000)
        {
            jmp=*in++;

            // length encoded with 5 bits?
            if (jmp & 0xF800)
            {
                len = jmp >> 11;
                jmp &= 0x7FF;
            }
            // then it's using 16 bits
            else len = *in++;

            // premature end signal
            if (!(jmp | len)) return;

            // zero filled RLE case when there's nowhere to jump back
            if (!jmp)
                while (len > 0) *out++ = 0, len--;
            // jump back & copy from output
            else
            {
                u16 *bp = &out[-jmp];
                while (len > 0) *out++ = *bp++, len--;
            }
        }
        // literal, read and write as is
        else *out++ = *in++;

        // next flag
        flag <<= 1;
    }
}