/* Drawn from roboticus' (1118303) & hdb's(1118256) iterative implementations (Algorithm also described by graff & bitingduck) */
void xlateBuffersI( register U64 *p, const U32 size, const U32 offset ) {
    U32 register cnt = size, dst = 0, cmp = 0;
    U64 register tmp = p[ dst ];

    assert( offset < size );
    assert( ( (U64)p & 7ull ) == 0 );

    while( cnt-- ) {
        U32 register src = ( dst + offset ) % size;
        if( src == cmp ) {
            p[ dst ] = tmp;
            dst = ++cmp;
            tmp = p[ dst ];
        }
        else {
            p[ dst ] = p[ src ];
            dst = src;
        }
    }
    return;
}

__forceinline void swapElems( register U64 *p, const U32 x, const U32 y ) {
    register U64 temp = p[ x ];
    p[ x ] = p[ y ];
    p[ y ] = temp;
}
/* Adapted from anonymonk's post 1118262 */
void xlateBuffersR( U64 *p, U32 size, U32 x, U32 y ) {
    U32 i, savedY = y;

    if( y == size ) return;

    for( i = x; i < savedY; ++i ) {
        if( y == size ) y = savedY;
        swapElems( p, x++, y++ );
    }
    xlateBuffersR( p, size, x, y );
}