So the Alpha gcc bug is still present in gcc 2.6.3. I sent a test
program that reproduces it to the gcc development team, who said it
doesn't happen in their version of the code -- but I think this was
when 2.6.2 was current (when did 2.6.3 come out?).
> I also experimented with compiling python with the the F_ math
> functions, by defining _FASTMATH, and found that the tests fail in the
> test_b1.py file on line 208, the map(sqrt,...) test, as follows
[...]
> This could result from a number of possibilities, and since I haven't
> checked out the code I'm not sure of the exact cause. But it either
> suggests an error in the sqrt() and comparions code in python, or more
> probably an error in the F_ math routines.
Perhaps there's a precision error that doesn't show up in the rounded
version that's printed?
> The reason I tried out the _F math routines is because of the
> performance improvements; on a test on the above system I found that I
> could get 1.25e6 sin()'s a second or so with the ordinary version,
> whereas with the F_ routines I was getting 2e+6 sin()'s a second. I
> must test the accuracies further.
But who cares?
--Guido van Rossum, CWI, Amsterdam <mailto:Guido.van.Rossum@cwi.nl>
<http://www.cwi.nl/cwi/people/Guido.van.Rossum.html>