Re: Use of C99 extra long double math functions after r236148

On 28 May 2012, at 13:30, Rainer Hurling wrote:

> On 28.05.2012 10:41 (UTC+1), David Chisnall wrote:
>> On 28 May 2012, at 05:35, Rainer Hurling wrote:
> 
> Ok, that's what I had supposed. Because the main difference between r236147 and r2136148 seems to be the define of _MATH_EXTRA_H_, the rest is more a type of binning?

Yes, it's just about making libc++'s cmath header compile, nothing more.  

> Yes, I agree. These outstanding long double math functions (like log1pl) and better unicode support are really needed for some important third party projects like R or SAGA GIS.

I very much doubt that anything is using the C11 unicode stuff yet, since no compiler or libc currently supports it...

> In the past I have read several times discussions about the correctness of long double functions on FreeBSD. Some drafts have been dismissed because of there inaccuracy in special cases. Also was discussed to get missing libm routines from NetBSD [1]. It appears as if we have to wait some more time ...

I thought we did pull in some NetBSD libm stuff recently.  Not sure what the status of that is, you'd need to check with das_at_.

> Yes, I think in most cases math/R is tolerant enough of just using doubles. But in the near future they plan to implement much more of the C99 stuff and their tolerance to offer workarounds for FreeBSD shrinks from release to release [2]. So these problems will increase :-(

Reading that email, it seems that they would prefer a function that exists and returns the wrong result to one that does not exist.  If this is really the attitude of the developers of R, then I shall make absolutely certain that I avoid using R for anything where I actually care about the results, and would strongly encourage everyone else to do the same.  

In general, (sane) people use the long double versions because they need the extra precision and care about the result.  We could easily implement the long double versions now as toy versions that just called the double versions, but that would upset a lot of people who actually care about accuracy, who are the main target audience for these functions.  

David