[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200206145156.GF1663@brightrain.aerifal.cx>
Date: Thu, 6 Feb 2020 09:51:56 -0500
From: Rich Felker <dalias@...c.org>
To: musl@...ts.openwall.com
Subject: Re: Q: dealing with missing removal of excess precision
On Tue, Jan 14, 2020 at 10:53:41PM +0300, Alexander Monakov wrote:
> On Tue, 14 Jan 2020, Rich Felker wrote:
>
> > > otoh it would be nice if there was a way to tell the compiler not to
> > > remove it (e.g. in case the asm already took care of it) even in c99
> > > standard mode.
> >
> > Perhaps this happens if the output constraint is tied to a float
> > rather than a long double?
>
> Precisely. That's what previously posted patches do, and they match
> existing hand-tuned assembly.
>
> I misspoke when saying that Glibc might return a value with excess precision.
> I was looking at fmod-like functions and missed a slightly subtle point that
> fprem does not introduce excess precision. So I don't actually have any
> example where Glibc might misbehave in that regard.
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:
https://github.com/OSGeo/PROJ/issues/1906
The problem is that, by returning excess precision, these functions
violate hard constraints on their range (either entire range, or range
for a given input domain) -- invariants like acos(x)<=M_PI or
implications like atan2(y,x)>=M_PI_2 => "(x,y) outside first quadrant"
fail to hold.
Arguably this is just a 1ulp error issue, but I don't think we have
any actual inaccuracies of that degree at "important" angles where the
result is not the correctly rounded one, even with the generic C
implementations. Rather the problem is stemming purely from wrongly
retaining excess precision and the fact that the LD80 approximation of
pi is greater than the double approximation of pi.
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.
Rich
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
