OpenLibm?

OpenLibm?

[Joel Sherrill]

Hi

I was looking for a possible source of long double libm method
implementations and it looks like openlibm
(https://github.com/JuliaMath/openlibm) has a complete implementation
for multiple architectures of interest. It is based on the fdlibm
source I think most of newlib's libm came from and appears to be well
maintained.

Any thoughts on using this for at least the long double methods?

Other than inertia, is there a reason not to consider newlib just
importing this?

--joel

OpenLibm?

[Paul Zimmermann]

about Joel's mail (https://sourceware.org/pipermail/newlib/2021/018739.html).

OpenLibm long double does not implement exp10, j0, j1, y0 and y1, and has
accuracy issues with lgamma and tgamma. See Table 11 of
https://members.loria.fr/PZimmermann/papers/#accuracy.

Paul

Re: OpenLibm?

[Joel Sherrill]

On Thu, Dec 2, 2021 at 2:03 AM Paul Zimmermann <Paul.Zimmermann@inria.fr> wrote:
>
> about Joel's mail (https://sourceware.org/pipermail/newlib/2021/018739.html).
>
> OpenLibm long double does not implement exp10, j0, j1, y0 and y1, and has
> accuracy issues with lgamma and tgamma. See Table 11 of
> https://members.loria.fr/PZimmermann/papers/#accuracy.

OK both have issues. I'd like long double support in newlib's libm to
improve standards compliance. I was looking for an implementation.

Issue: newlib has no long double support when sizeof(long double) !=
sizeof(double).

Random thoughts:

+ For the missing openlibm methods, newlib's implementation of those
methods could be submitted to openlibm.

+ Does newlib have an accuracy document?

+ Is newlib's accuracy better than openlibm's?

+ Would it make sense to port long double methods from openlibm to newlib?

I'm just trying to find a technical path to get long double into newlib.
I have no vested interest in any particular path that achieves that.
If we can agree on a technical path, RTEMS could add this as a potential
Google Summer of Code 2022 project.

--joel

> Paul