Not sure if the C version of the jmcnamara idea got benchmarked.

Yes. It is laballed as _24to32_2() in my benchmark posts above. The _byteswap_ulong() is a compiler pragma that reduces to a single x64 assembler instruction that effectively does the same thing as your manipulations of the temp variable above. And the masking constants are hardcoded.

void _24to32_2( SV *packed ) {
    IS_VARS;
    char *pp = SvPVX( packed );
    int i;

    IS_RESET;
    for( i=0; i<24; i+=3 ) {
        unsigned int n = _byteswap_ulong( *(unsigned long*)&pp[ i -1 ]
+ );
        IS_PUSHUV( ( n & 0xfc0000 ) >> 18 );
        IS_PUSHUV( ( n & 0x03f000 ) >> 12 );
        IS_PUSHUV( ( n & 0x000fc0 ) >>  6 );
        IS_PUSHUV( ( n & 0x00003f )       );
    }
    IS_DONE;
    return;

}
[download]