Bug#714915: openturns: FTBFS on any-i386: small difference in cppcheck_TrapezoidalFactory_std output

[D. Barbier]

On 2013/7/4 Samuel Thibault wrote:
> Package: openturns
> Version: 1.1-6
> Severity: serious
> Justification: FTBFS
>
> Hello,
>
> openturns currently FTBFS on any-i386, due to a small output difference:
>
> --- /«PKGBUILDDIR»/lib/test/t_TrapezoidalFactory_std.expout     2013-01-03 14:01:36.000000000 +0000
> +++ /«PKGBUILDDIR»/obj-i486-linux-gnu/lib/test/t_TrapezoidalFactory_std.out     2013-07-03 22:54:04.041954105 +0000
> @@ -1,8 +1,8 @@
>  Distribution          =class=Trapezoidal name=Trapezoidal dimension=1 a=1 b=2.3 c=4.5 d=5 h=0.322581
> -Estimated distribution=class=Trapezoidal name=Trapezoidal dimension=1 a=1.006 b=2.275 c=4.545 d=4.99 h=0.3198
> +Estimated distribution=class=Trapezoidal name=Trapezoidal dimension=1 a=1.005 b=2.279 c=4.545 d=4.99 h=0.32
>  Default distribution=class=Trapezoidal name=Trapezoidal dimension=1 a=-2 b=-1 c=1 d=2 h=0.333333
>  Distribution from parameters=class=Trapezoidal name=Trapezoidal dimension=1 a=1 b=2.3 c=4.5 d=5 h=0.322581
>  Trapezoidal          =class=Trapezoidal name=Trapezoidal dimension=1 a=1 b=2.3 c=4.5 d=5 h=0.322581
> -Estimated trapezoidal=class=Trapezoidal name=Trapezoidal dimension=1 a=1.006 b=2.275 c=4.545 d=4.99 h=0.3198
> +Estimated trapezoidal=class=Trapezoidal name=Trapezoidal dimension=1 a=1.005 b=2.279 c=4.545 d=4.99 h=0.32
>  Default trapezoidal=class=Trapezoidal name=Trapezoidal dimension=1 a=-2 b=-1 c=1 d=2 h=0.333333
>  Trapezoidal from parameters=class=Trapezoidal name=Trapezoidal dimension=1 a=1 b=2.3 c=4.5 d=5 h=0.322581
>
> I've also seen that in sphinxbase, where this is due to libc 2.17 which
> has small libmath fixes compared to libc 2.13.

Hello Samuel,

I disabled optimization on *-i386 arches because of #714411.  I
checked on fischer.d.o, but I ran two builds, one with -O0 and the
other one with -fno-cse-follow-jumps -fno-caller-saves
-fno-guess-branch-probability, and unfortunately it is likely that I
switched results, only the latter works.  I could adjust test output,
but I now wonder whether this is the best choice.
Maybe you have some ideas why those -f flags are needed on *-i386?
With GCC < 4.8, only -fno-cse-follow-jump was needed, but now we need
two more flags.

Denis