Re: isnan and exceptions

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 28/08/2020 12:23, Florian Weimer via Libc-help wrote:
> It seems that __builtin_isnanl raises Invalid Operation for sNaNs on
> x86-64, but glibc's isnanl implementation does not.
> 
> Is this expected?  Are both behaviors correct?
> 
> The background of my question is that if I replace isnanl with this:
> 
> int
> isnanl (long double x)
> {
>   return __builtin_isnanl (x);
> }
> 
> I get these test suite failures:
> 
> testing long double (without inline functions)
> Failure: isnan (sNaN): Exception "Invalid operation" set
> Failure: isnan (-sNaN): Exception "Invalid operation" set
> Failure: isnan_downward (sNaN): Exception "Invalid operation" set
> Failure: isnan_downward (-sNaN): Exception "Invalid operation" set
> Failure: isnan_towardzero (sNaN): Exception "Invalid operation" set
> Failure: isnan_towardzero (-sNaN): Exception "Invalid operation" set
> Failure: isnan_upward (sNaN): Exception "Invalid operation" set
> Failure: isnan_upward (-sNaN): Exception "Invalid operation" set

I think it is https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66462 

As Wilco has noted, using unordered floating point comparison in
isnan expansion is not fully correct and as patch has been submitted
but later revert on gcc.  The last iteration [1] seems handle the
raised issues, but it was never committed. 

[1] https://gcc.gnu.org/legacy-ml/gcc-patches/2016-11/msg02567.html