Archipelago » qemu

/* These need to return a pair of value, so still use T0/T1.  */

/* Transpose.  Argument order is rather strange to avoid special casing

   the tranlation code.

   On input T0 = rm, T1 = rd.  On output T0 = rd, T1 = rm  */

void HELPER(neon_trn_u8)(void)

{

    uint32_t rd;

    uint32_t rm;

    rd = ((T0 & 0x00ff00ff) << 8) | (T1 & 0x00ff00ff);

    rm = ((T1 & 0xff00ff00) >> 8) | (T0 & 0xff00ff00);

    T0 = rd;

    T1 = rm;

}

void HELPER(neon_trn_u16)(void)

{

    uint32_t rd;

    uint32_t rm;

    rd = (T0 << 16) | (T1 & 0xffff);

    rm = (T1 >> 16) | (T0 & 0xffff0000);

    T0 = rd;

    T1 = rm;

}

/* Worker routines for zip and unzip.  */

void HELPER(neon_unzip_u8)(void)

{

    uint32_t rd;

    uint32_t rm;

    rd = (T0 & 0xff) | ((T0 >> 8) & 0xff00)

         | ((T1 << 16) & 0xff0000) | ((T1 << 8) & 0xff000000);

    rm = ((T0 >> 8) & 0xff) | ((T0 >> 16) & 0xff00)

         | ((T1 << 8) & 0xff0000) | (T1 & 0xff000000);

    T0 = rd;

    T1 = rm;

}

void HELPER(neon_zip_u8)(void)

{

    uint32_t rd;

    uint32_t rm;

    rd = (T0 & 0xff) | ((T1 << 8) & 0xff00)

         | ((T0 << 16) & 0xff0000) | ((T1 << 24) & 0xff000000);

    rm = ((T0 >> 16) & 0xff) | ((T1 >> 8) & 0xff00)

         | ((T0 >> 8) & 0xff0000) | (T1 & 0xff000000);

    T0 = rd;

    T1 = rm;

}

void HELPER(neon_zip_u16)(void)

{

    uint32_t tmp;

    tmp = (T0 & 0xffff) | (T1 << 16);

    T1 = (T1 & 0xffff0000) | (T0 >> 16);

    T0 = tmp;

}