Re: [glob2-devel] Nicowar and desync

[Bradley Arsenault]

On 7/18/06, Cyrille Dunant <address@hidden> wrote:
> >
> > Yeah, Kai already emailed me with a grep. I, unfortently, am not at my
> > house, and I can't work on glob2 for atleast three weeks while my
> > family moves. Worse yet, me might have to get modem-internet in our
> > next house. Anyway, all of these usages are independant casts, so in
> > my case, atleast the same CPU will give the same results every time.
>
> no. Not even with the same compiler. The bits at the end of the mantissa are
> _random_ they depend on the system state, the weather, and the bugs in your
> hardware, and the manufacturer's breakfeast when they were built.

So why do you think that a compiler would intentionally leave bits
unitialized in the mantissa when it could initialize them and add
better determinism?

All of my usages of float are casts, temporaries, un-accumulated
values from integers. Theres no reason for the compiler to leave bits
unitialized when its being provided the same deterministic value as
input.

Float insafety is caused mainly by accumulated-error and having a
different exponent and mantissa representing the same value ( 50 * 0.1
or 5 * 1.0 ).

If your willing to explain yourself, then good, but I see no reason
for the compiler to leave unitialized bits inside a value thats being
initialized.



> Do _not_ _ever_ depend on doubles for bit perfect reproducibility.
>
> > I'll fix this as soon as I have time. Sorry everyone, my mistake.
>
> CU
>
>  -- CFD
>
>
> _______________________________________________
> glob2-devel mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/glob2-devel


--
Start and finish, Bradley Arsenault