assignment problem with doubles: getting -NaN unexpectedly

[Monty]

>> This problem has hit me, too, with gcc-2.95.2 and very recent libc and
>> linux kernel.  With -O3, simple double math (no division) results in
>> NaN sporadically, but inserting debugging print statements seems to
>> make this not happen.

Correct, the printfs appear to force storage back into memory.  BTW,
-ffloat-store will eliminate the bug (at a performance hit).

>See my post ("gcc 2.95.1 bug") earlier this month to gcc-bugs for a
>similar case, including a test program that reliably produces the
>mysterious NaNs (confirmed by at least one other person).

Confirmation here too.  All versions of egcs tested (1.1 through -r HEAD of CVS
last night) exhibit the bug.  Prior versions of classic GCC through 2.8.x do
not.

>I wish I knew how to better bring this to the attention of a gcc hacker;
>this is a real obstacle for me right now.

Major obstacle here as well.  Our next generation open audio project (read: mp3
killer) is hitting this like a brick wall. 

Monty