... out of the 64 bit result (convenienently, you can use xtea_data.l.y, for example), feed that instead of the "lrand48()" function to your program (RAND_MAX is then of course equal to UINT32_MAX = 4294967295UL).

I believe you are aware of the ineffectivity of floating point operations on a '51 (on top of the relatively high complexity of the block cipher). I really am not quite able to estimate if your application can do with fixed point operations, but you might want to do so if it turns out to be prohibitively slow.

I am certainly interested to hear about your findings.

JW