[PATCH v3] Cygwin: sigfe: Fix a bug that signal handler destroys fpu states

[Takashi Yano]

Hi Corinna,

On Fri, 25 Oct 2024 10:49:51 +0200
Corinna Vinschen wrote:
> On Oct 24 17:58, Takashi Yano wrote:
> > On Tue, 22 Oct 2024 17:58:02 +0200
> > Corinna Vinschen <corinna-cygwin@cygwin.com> wrote:
> > > Hi Takashi,
> > > 
> > > big change, so, honest question: Do you think this is safe for 3.5.5?
> > > 
> > > This certainly also requires an entry in the release text and there
> > > are just a few minor typos in comments, see below.
> > 
> > What about adopting my first idea to 3.5.5
> > https://cygwin.com/pipermail/cygwin/2024-October/256506.html
> > and applying this patch to 3.6.0 branch?
> 
> Admittedly, I'm also not deep in FPU stuff.
> 
> fnstenv/fldenv and dropping fninit look like a simple approach to fix
> the worst problem.  Did you just discuss this on the mailing list or did
> you check that it actually fixes the reported problem?

I tested locally using Christian's test case and confirmed that
the problem has been fixed as well as that no new problem occurs.

> But either way, I wonder if it's worth the effort to have two different
> solutions.  This isn't a regression, so we don't have to fix this ASAP
> in 3.5.5.  It could easily wait a version.
> 
> So I'm thinking we should go with your sleek new code in 3.6 and let
> this simmer for a while, so it's put to use by people running 3.6
> versions.
> 
> Does that sound right?

Hmm, this is not a regression but a long-standing bug which can
easily be fixed. Also, there is only a very small risk of regression
with the fnstenv/fldenv patch. Moreover, the bug causes critical
miscalculations to result in long double processing.

So, I think it is better to apply the v3_5-branch patch (I have just
submitted) for cygwin-3_5-branch, and to apply the v3 patch (I already
submitted) to the master branch. v3_5-branch patch is simple and small,
so it will not be so painful to maintain.

Don't you think so?

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>