Re: TR1 Math tarball

[Benjamin Kosnik <bkoz at redhat dot com>]

> I'm putting together another patch in which the floating point 
> tolerances are replaced by usage of std::numeric_limits<>::stuff().
> There are still some integer loop limits but these are more sane.

Sounds good.

> This patch also has several substantive algorithmic improvements to some 
> of the

OK.

> The is the _Private:: to __detail:: change has is done (I noticed this 
> in <random>).
> I'll do the TR1_ for the header guards.

Great, thanks.

> I was asked to post what I had so we could talk about it.
> 
> I am finding time for this after a long break.
> 
> But I also agree that we are coming up on a point of diminishing returns 
> for me hacking on it alone and flinging tarballs around.  It would be 
> nice to have most of 4.3 to stomp on it (9 months at the going rate.)

My thoughts exactly. Thanks again for working on this. 

I'd like to see this work moved into the main gcc svn repository, after
you post this last patch and Paolo gets a chance to respond. This work
is already much further along than I expected..... it's silly to keep
it out of mainline.

The reasons are pretty simple: getting this into the repository will
increase your work's visibility, potentially shaking more contributors
out of the woodwork that can help you.

I have full confidence that refinements and work on Paolo's other
issues will proceed once this is checked in. Paolo, are you willing to
move to a more incremental style for this work?

In addition, the tr1-in-std work is potentially destabilizing, so I'd
like to get as much out-of-mainline work at least into SVN as soon as
possible, so that I screw the least number of people possible with
merging.

> Finally, I do have some issue with using std::tr1::mathfunction.  I put 
> guards around but where is _GLIBCXX_USE_C99_MATH_TR1 defined?  I'd like 
> to fix this.

I believe Paolo answered this in his email.

> There are two classes of testcase that I would like:
> 1) Checks that the nan's and throws are coming out at the right points.
> 2) Spot checks of selected numeric values.

Sounds good to me!

best,
benjamin