Control: forwarded -1
https://gitlab.freedesktop.org/pulseaudio/pulseaudio/issues/619
Control: retitle -1 pulseaudio: uses -ffinite-math but might deal with
infinity
Hi,
Sorry for the delay.
On Sat, Dec 15, 2018 at 3:24 AM Michael Cree wrote:
> Source: pulseaudio
> Version: 12.2-2
> Severity: important
> Justification: fails to build from source but built in the past.
> User: debian-al...@lists.debian.org
> Usertags: alpha
> Tags: patch
>
> Pulseaudio FTBFS on alpha due to the volume-test test failing due to
> a floating-point exception which in turn is due to an infinity in
> floating-point calculations when volume-test is compiled with finite
> math options.
>
> This is bug #798248 reappearing but in a subtlely different guise.
> There the non-finite math was protected against by checking that the
> arguments are finite before performing floating point calculations,
> but it now seems that gcc takes the specification of finite math,
> being "[a]llow optimizations for floating-point arithmetic that
> assume that arguments and results are not NaNs or +-Infs" so
> pedantically true, that it is fair game to optimise away any calls
> to isfinite() because the argument must be finite: it was said so
> on the command line!
>
> Whatever, examination of the object code shows that the calls to
> isfinite() are eliminated thus the floating-point arithmetic is no
> longer protected.
>
I think if infinities might happen, then using -ffinite-math is wrong.
I have forwarded the issue upstream. I have disabled -ffast-math in the
meantime, since I think correctness is more important than speed.
>
> Fortunately we can work out whether the arguments to the offending
> arithmetic are finite by other means and I attach a patch doing
> just that. With this patch pulseaudio builds to completion on
> Alpha.
>
This is insufficient as there is a call to isnan() in the core lib.
--
Saludos,
Felipe Sateler