bug#14281: 24.3; replace-match leaves point at wrong place

[Stefan Monnier <monnier@IRO.UMontreal.CA>]

> I reproduced the issue again and investigated. Since I'm unsure how to
> determine information about the Lisp system within GDB, I used debug
> statements in C and Lisp to arrive at the following description of what
> happens, additive to my previous description:

> My after-change-functions, in order, are:
>    semantic-change-function
>    c-after-change
>    jit-lock-after-change

> search_regs.start[0] first takes on the incorrect value inside
> c-after-change's call to save-match-data. Since c-after-change code seems
> correct, I determined that the match-data function begins returning the
> incorrect value during semantic-change-function. I am using CEDET r8557.

Not sure what CEDET r8557 does, but at least the CEDET code in Emacs's
trunk is pretty simple in this respect: semantic-change-function only
runs the semantic-change-functions hook, and grep seems to indicate that
this hook is never modified, so the whole thing should never get
anywhere near the match-data.

Note that there are other change functions than after-change-functions.
There's also before-change-functions and there are overlays's
modification-hooks.  You might like to check those as well.

>>> I suppose that caveat would pin the bug on one of the third party
>>> packages I use. However, why couldn't Emacs save off the match-data
>>> itself and restore it after the after-change-functions? Is there any
>>> legit situation where a change hook would want to change the
>>> match-data in effect after the change hook returns?

Stefan> There are many cases where an after-change-function won't use regular
Stefan> expressions at all.
> The answer doesn't seem to fit the question, so I'll rephrase: Why does
> Emacs allow after-change-functions to change the match-data beyond its
> scope? Or: why doesn't the signal_after_change C function do like
> set-match-data instead of leaving it to client change hooks to do so?

Because it wastes time in most cases (when after-change-functions
won't use regular expressions at all).

Maybe we could add code that saves just the (match-beginning 0) and
signals an error if it was not properly preserved.  This would still
require change-functions to save the match-data if they use it, but it
might catch the offenders earlier.


        Stefan