Puzzled!

[Martin v. Löwis]

Jakub Jelinek <jakub@redhat.com> writes:

> What exactly is required on gcc side? Verifying that without -ffast-math
> all float builtins on EDOM/ERANGE arguments either set errno, or bail to
> library functions?

It appears that the generated code need to change for all arithmetic
operations. E.g. the function

double x(double a,double b)
{
	return a+b;
}

may set errno, but currently, there is no mechanism in the code to
activate such a processing. It seems that such mechanism should be
combined with #pragma STDC FENV_ACCESS, so that the penalty for
setting errno is avoided.

Regards,
Martin