hyperspeed polling

#define POLLS 4
#define _data_len ( 45 * ( (POLLS) / 4 ) )

typedef struct {
    u32 init;
    volatile u8* data_pointer;
    u32 crc_state;
    u32 crc_table[ 256 ];
    u32 poll_count;
    u8  temp_data[ 24 ];
    union {
        u8  bytes[ _data_len ];
        u16 _u16[ _data_len / sizeof(u16) ];
        u32 _u32[ _data_len / sizeof(u32) ];
        s16 _s16[ _data_len / sizeof(s16) ];
        int _int[ _data_len / sizeof(int) ];
    } out;
} static_globals;
extern static_globals gvars;

typedef struct {
    u8 b1, b2, x, y;
    u16 status;
} __attribute__((packed)) raw_input_t;
// 6 bytes long
// might be bad, probably not needed
// hopefully this works if you pay attention to alignment

int pad_poll( void ) {
    u32 bits;
    u8 bits_n, i, written_bytes, padding;
    bits_n = 0;
    bits = 0ULL;
    written_bytes = 0;
    for( i = 0; i < (POLLS); ++i ) {
        raw_input_t* pad_data;
        // try to preserve the original pad input
        // *never write* to padmgr

        if( !i ) { // the first time, use pad data from the original game poll
            pad_data = (raw_input_t*)(&padmgr->pads[0]);
        } else { // afterwards, poll the pads ourselves, and discard data for pad 1
            pad_data = (raw_input_t*)(gvars.temp_data);
            osContStartReadData( queue ); // black magic
            osRecvMesg( queue, NULL, OS_MESG_BLOCK );
            osContGetReadData( pad_data ); // actually get the pad data
        }
        // pad_data now has the state of the pads for this poll

        if( pad_data[1].status == 0 || pad_data[2].status == 0 || pad_data[3].status == 0 )
            return 0;

        gvars.out.bytes[ written_bytes++ ] = pad_data[1].b1;
        gvars.out.bytes[ written_bytes++ ] = pad_data[1].x;
        gvars.out.bytes[ written_bytes++ ] = pad_data[1].y;
        gvars.out.bytes[ written_bytes++ ] = pad_data[2].b1;
        gvars.out.bytes[ written_bytes++ ] = pad_data[2].x;
        gvars.out.bytes[ written_bytes++ ] = pad_data[2].y;
        gvars.out.bytes[ written_bytes++ ] = pad_data[3].b1;
        gvars.out.bytes[ written_bytes++ ] = pad_data[3].x;
        gvars.out.bytes[ written_bytes++ ] = pad_data[3].y;
        bits <<= 6;
        bits  |= pad_data[1].b2;
        bits <<= 6;
        bits  |= pad_data[2].b2;
        bits <<= 6;
        bits  |= pad_data[3].b2;
        bits_n += 18;
        gvars.out.bytes[ written_bytes++ ] = ( bits & 0xFF );
        bits >>= 8;
        gvars.out.bytes[ written_bytes++ ] = ( bits & 0xFF );
        bits >>= 8;
        bits_n -= 16;
        if( bits_n == 8 ) {
            gvars.out.bytes[ written_bytes++ ] = ( bits & 0xFF );
            bits >>= 8;
            bits_n -= 8;
        }
    }
    return 1;
}