On Thu, Feb 06, 2020 at 02:03:44PM -0500, Rich Felker wrote:
> On Thu, Feb 06, 2020 at 12:46:08PM -0500, Rich Felker wrote:
> > On Thu, Feb 06, 2020 at 08:15:30PM +0300, Alexander Monakov wrote:
> > > On Thu, 6 Feb 2020, Rich Felker wrote:
> > > 
> > > > I think I might like to go ahead and apply these patches now, or at
> > > > least some of them -- the ones fixing excess precision -- rather
> > > > waiting, because I got a report of a nasty bug stemming from excess
> > > > precision of the inverse trig functions:
> > > 
> > > That might be exactly the empty set of patches, as I did not yet post
> > > any for functions that might return with excess precision.
> > 
> > Indeed, I just discovered that...
> > 
> > > Be advised that I found bugs in my patches, so given that no one so far
> > > has pointed them out on the mailing list indicates that either nobody
> > > bothered to review, or people are keeping the findings to themselves.
> > 
> > I think it's just that I was planning to do further review after
> > release rather than before since I'm trying to get the release out..
> > 
> > > > If writing and testing the remaining i386 functions before release is
> > > > not practical, I wonder if just removing the asm for now, and adding
> > > > back the new code in next release cycle would be a good idea. Or I
> > > > could just leave it, but I don't like making a release with "known
> > > > bugs of consequence" like this.
> > > 
> > > I think fixing excess precision in inverse trig functions might be
> > > easier than removing the asm entirely.
> > 
> > Yes, what I'm looking at right now is fixing inverse trig and log
> > functions and removing the exp asm (since the exp logic is way too
> > messy for me to feel comfortable modifying right now) and possibly
> > re-adding it later as inline asm with the flow control in C.
> 
> FWIW nsz's new C exp seems considerably faster than the existing 386
> asm on my box (Atom S1260) (6.7s vs >8s for summing exp(x) from
> x=-2..2 stepping 0x1p-24). Test program attached in case anyone else
> wants to try it.
> 
> So I think just removing exp*.s is the right approach for now. The
> long double ones should actually be left, and that raises the issue
> that expm1l is wrongly using the exp code rather than expl code nsz
> added long ago in a8f73bb1a685dd7d67669c6f6ceb255cfa967790. I won't
> try to fix this yet but will just move the files around so we can rm
> the float/double ones and use the C for them without getting rid of
> the ld asm.

Attachment was missing.

Rich