-ffast-math is apparently bad, what does synfig use it for?

[Paul Wise]

Hi all,

In addition to vorlon's comments on this debian bug:

http://bugs.debian.org/367048

I got these comments on IRC:

<pabs> vorlon: are there any other platforms than alpha where
-ffast-math is known to break code?
<Sesse> pabs: -ffast-math can break code easily
<ifvoid> pabs: many code is broken by -ffast-math
<Sesse> that's why it's not part of any -O option
<ifvoid> it used to break mozilla, for exmaple
<ifvoid> (not sure if it still does)
<pabs> so, to be avoided everywhere?
<ifvoid> pabs: except for very specialized apps, yes
<ifvoid> I tend to use it for my numerical progs
<ifvoid> it's really only useful if you have to do many cos() or pow()
or so
<pabs> hmm, I wonder if a complex 2D animation renderer counts
(http://www.synfig.com)
<ifvoid> and if you don't care about range checking
<pabs> I shall talk to upstream, thanks all
<ifvoid> I usually disable it in any package builds
<mrvn> doesn't -ffast-math only circumvent the extra redirection for
libm calls?
<Sesse> mrvn: no
<ifvoid> mrvn: it disables range checking 
<ifvoid> for negative input of log(), for example
<ifvoid> and disables checking for inf and nan and such things
<mrvn> on cpus that don't have that in hardware
<Sesse> ifvoid: it also replaces functions with odd macros
<Sesse> ifvoid: some of which have reduced argument range and precision
<pabs> Sesse/ifvoid/mrvn: mind if I paste this stuff to my upstream?
<ifvoid> pabs: go ahead
<ifvoid> pabs: the gcc info page is pretty informative, too
<ifvoid> sec 3.10
<Sesse> pabs: np

darco:

      * is there a reason -ffast-math is used by default in synfig?
      * if it is, any ideas about how to disable it on some platforms?
        linux-alpha being one of them.

-- 
bye,
pabs

http://wiki.debian.org/PaulWise
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : http://lists.alioth.debian.org/pipermail/pkg-synfig-devel/attachments/20060615/b057954e/attachment.pgp