Comment # 13 on bug 1092541 from Richard Biener

(In reply to Michal Srb from comment #12)
> (In reply to Richard Biener from comment #3)
> > The new situation is that it considers functions without
> > target attribute to be decorated with the effective target options of the
> > compilation.
> 
> This would explain the situation, but the gcc invocation does not have any
> target options. I tried to remove all unnecessary options and the error is
> still happening with gcc invocation looking like this (removed -D, -I and -W
> options for brevity):
> 
>     gcc -fno-strict-aliasing -fvisibility=hidden -O2 -g -c gen4_vertex.c
> -fPIC -o
>     .libs/gen4_vertex.o
> 
> (In reply to Richard Biener from comment #11)
> > I wonder whether the package
> > uses any of the -ffast-math options and thus if we could relax this
> > condition with -funsafe-math-optimizations or so.
> 
> As far as I can see it does not use -ffast-math (as parameter or attribute).
> I tried adding either -fexcess-precision=standard or
> -funsafe-math-optimizations, but it still reports the same error.

The issue is the fpmath=sse in the target attribute and the default of
fpmath=387 on the command-line (yes, default settings also count).  That's
a semantic difference (FP computation results differ due to precision).