bug#42701: 28.0.50; Duplicate Edebug instrumentation of lambda form in some cases

[Philipp Stephani]

)

Am Di., 4. Aug. 2020 um 22:12 Uhr schrieb Alan Mackenzie <acm@muc.de>:
>
> Hello, Philipp.
>
> In article <mailman.680.1596473284.2739.bug-gnu-emacs@gnu.org> you wrote:
>
> > Create a file /tmp/edebug.el:
>
> > $ cat /tmp/edebug.el
> > (defun f () (if-let (x (funcall (lambda (y) 1) 2)) 3 4))
>
> > Visit the file:
>
> > $ emacs -Q -l subr-x -l edebug /tmp/edebug.el
>
> > Now instrument the function `f' using C-u C-M-x.  The *Messages* buffer
> > will now contain
>
> > Edebug: edebug-anon0 [2 times]
> > Edebug: f
>
> > The [2 times] indicates the problem: Edebug has instrumented two
> > definitions with the same (generated) symbol.
>
> I don't think you're correct, here.  I think it's instrumented the
> lambda form twice, once for each arm of the edebug spec.  It discards
> the first attempt, then succeeds at the second.
>
> The pertinent edebug spec (from the if-let definition in subr-x.el)
> looks like:
>
>     ([&or (&rest [&or symbolp (symbolp form) (form)])
>           (symbolp form)]
>      form body)
>
> .  In the outer &or sub-spec, the (&rest ....) bit doesn't match, but
> the failure to match only happens after the "Edebug: edebug-anon0"
> message has been output the first time.  Edebug then tries the (symbolp
> form) alternative, which does match and outputs the "...-anon0" message
> a second time.

That is correct, but when the match fails, it's already too late:
Edebug (in `edebug-make-form-wrapper') has already performed some
global mutations such as modifying `edebug-form-data' of the `edebug'
property of the newly-generated symbol. These mutations aren't unwound
after the first failed match.

>
> > This is a problem when using Edebug for e.g. coverage instrumentation,
> > as the coverage information is attached to the symbol itself (as a
> > symbol property), and duplicate symbols when instrumenting code lead
> > to subtle errors such as mismatching vector lengths for the position
> > of the breakpoints and the hit counts.
>
> I don't think this is happening (see above), but I admit not having
> looked into it all that closely.  Do you have any further evidence of
> two distinct functions being mapped onto the same generated symbol?

I've tried to provide some context in
https://lists.gnu.org/archive/html/bug-gnu-emacs/2020-08/msg00574.html.
The most common symptom of such duplicate definitions (as defined by
"calling `edebug-make-form-wrapper' twice with the same
`edebug-def-name'") is
https://debbugs.gnu.org/cgi/bugreport.cgi?bug=41853.

>
> > I'd speculate that this issue is similar to Bug#41988 in that Edebug
> > defines instrumented symbols even when backtracking later.  In this
> > case, Edebug backtracks to the legacy (SYM VAL) form, but has already
> > partially matched the ((SYM VAL)) form, including instrumenting the
> > lambda form therein.  I guess Edebug should perform some kind of
> > two-phase instrumentation and instrument subforms only when a form has
> > been chosen after backtracking.  Since this is somewhat difficult to
> > implement without rewriting larger parts of Edebug, it might be more
> > feasible to regenerate anonymous symbols after a failed match.
>
> I don't know why Edebug re-uses the symbol.
.
My guess is that `edebug-old-def-name' becomes non-nil because the
form has already been instrumented
(https://git.savannah.gnu.org/cgit/emacs.git/tree/lisp/emacs-lisp/edebug.el?id=1a845a672dc73c8e98e6cb9bb734616e168e60ba#n1382).
It's then reused in
https://git.savannah.gnu.org/cgit/emacs.git/tree/lisp/emacs-lisp/edebug.el?id=1a845a672dc73c8e98e6cb9bb734616e168e60ba#n1250.

>  But beyond the misleading
> double message in *Messages*, is there any harm done?  Would it be less
> confusing if two distinct messages "Edebug: edebug-anon0" and "Edebug:
> edebug-anon1" were to be output?

The duplicate message is just a symptom of calling
`edebug-make-form-wrapper' twice with the same `edebug-def-name'
symbol, and attaching various instrumentation data to the symbol.
Depending on whether the two definitions are compatible, this can be
harmless or lead to subtle issues like
https://debbugs.gnu.org/cgi/bugreport.cgi?bug=41853. Therefore in my
coverage instrumentation driver I'm rejecting such duplicate
instrumentations
(https://github.com/phst/rules_elisp/commit/31d30e99c21027c5859782229aebb314ed3cb36c,
since it doesn't seem to be possible to predict whether the
redefinition would indeed be harmless or trigger
https://debbugs.gnu.org/cgi/bugreport.cgi?bug=41853.