bug#47067: 28.0.50; [feature/native-comp] Crash while scrolling through dispnew.c

[Eli Zaretskii]

> From: Andrea Corallo <akrl@sdf.org>
> Cc: 47067@debbugs.gnu.org
> Date: Fri, 12 Mar 2021 20:10:27 +0000
> 
> >> > Just evaluating c-beginning-of-statement-1 doesn't help.  But if I
> >> > load cc-engine.el, then the crash goes away.
> >> 
> >> Okay, then probably is one of the other four c-* functions we see in the
> >> backtrace.
> >
> > Yes, but how to determine which one?
> 
> Given they are 4 one could go evaluating these one by one, if they were
> more bisection would have been the best strategy.  Yeah that's not the
> most fun...

I was wrong: it _is_ c-beginning-of-statement-1, after all.  It's just
that the procedure I followed to evaluate is was wrong.  AFAIU now, it
should be: first load-library cc-engine (which loads the .eln file),
then evaluate the function's definition in the .el file.

So now, given that this (humongous) function is the suspect, how do I
proceed with the next steps, which you described as follows:

  When the function is identified I typically construct a single function
  reproducer, for this I typically need the input parameters and I try to
  substitute all other values coming from the environment with something I
  can control.  This step involves understanding which part of the
  environment are captured by the function (say: point, current buffer
  content etc etc...).

  At that point I reduce the function searching for the minimal piece of
  code that behaves differently when native compiled.

  At this point will typically start the "smart" part of the
  investigation.

> >> To force the .elc to be loaded one has to bind `load-no-native' to
> >> non-nil.
> >
> > I think if load-file is invoked interactively, and the user actually
> > types "foo.elc", we need to bind load-no-native non-nil
> > automatically.  Otherwise users would be surprised, as it goes against
> > the logic of what we do when the user types "foo.el".
> 
> We certanly can do this if this is what we want.  This breaks a little
> the idea to have the system as much transparent as possible, I went this
> way cause this was my understanding of what we wanted but I've no strong
> feeling with that.

I don't think this will break the transparent operation, because
loading a package non-interactively (as in when the corresponding
feature is 'require'd by some code) will still load the .eln file.
Only the following 2 use cases will be affected:

  M-x load-file RET /path/to/FOO.elc RET
  M-x load-library RET FOO.elc RET

IOW, when the user loads the file/library interactively, and
explicitly uses the .elc extension, we load the file the user
specified, not the corresponding .eln file.