Re: [Emacs-diffs] feature/byte-unwind-protect 916094a 2/2: Add new bytecodes for unwind-protect

Re: [Emacs-diffs] feature/byte-unwind-protect 916094a 2/2: Add new bytecodes for unwind-protect

[Stefan Monnier]

> +  (if (not byte-compile--use-old-handlers)
> +      (let ((except-tag (byte-compile-make-tag)))
[...]
> +    (pcase (cddr form)
> +      (`(:fun-body ,f)
> +       (byte-compile-form
> +        (if byte-compile--use-old-handlers `(list (list 'funcall ,f)) f)))
> +      (handlers
> +       (if byte-compile--use-old-handlers

Clearly the last two tests of byte-compile--use-old-handlers can be
dropped since they're inside a branch where we know
byte-compile--use-old-handlers is t.

BTW, when introducing byte-codes we've "traditionally" done it in two steps:
- in release N we add the code which handles the new byte-codes,
  and we add support for the new byte-code in the byte-compiler, but
  disabled by default.
- in release N+1 we enable the use of the new byte-codes in the
  byte-compiler by default.

This is done so as to reduce the pain when running a file compiled with
another Emacs version.

But we don't want to set byte-compile--use-old-handlers back to nil for
your new bytecodes, so better use another variable.

Also byte-compile--use-old-handlers can be retired, I think, because it
has played its role.

> +	CASE (Bpushunwindprotect): /* New in 27.1.  */
> +	  {
> +	    struct handler *c = push_handler (Qt, CATCHER_ALL);

Hmm... using this "catch+rethrow" method interferes with the
debug-on-error stacktrace:

    (defun sm-test-1 () (unwind-protect (sm-test-2) (message "hello")))
    (defun sm-test-2 () (message "test-1-in") (car 4))
    (setq debug-on-error t)

then call `sm-test-1` and you'll see that Emacs gives you a backtrace
that gives the impression the error was signaled in `sm-test-1`!

I also wonder if there can be places where we call something like
`unbind_to` which should run those new unwind-forms but won't find it
because it won't look at the handlers [ offhand, I can't think of any
place where that would happen, and I think it's probably OK, but
I can't quite convince myself that it definitely is.  ]

All in all, I'd really prefer if we kept the unwind forms in
the specpdl.  Was there a particular reason why you used the
handlers instead?


        Stefan

Re: [Emacs-diffs] feature/byte-unwind-protect 916094a 2/2: Add new bytecodes for unwind-protect

[Tom Tromey]

>>>>> "Stefan" == Stefan Monnier <monnier@IRO.UMontreal.CA> writes:

>> +	CASE (Bpushunwindprotect): /* New in 27.1.  */
>> +	  {
>> +	    struct handler *c = push_handler (Qt, CATCHER_ALL);

Stefan> Hmm... using this "catch+rethrow" method interferes with the
Stefan> debug-on-error stacktrace:

Hmm, yeah.

Any ideas for fixing this?

One idea would be to capture the stack trace when CATCHER_ALL is
targeted, then have a way to pass this to signal or throw.

But often the stack trace isn't needed.

Stefan> I also wonder if there can be places where we call something like
Stefan> `unbind_to` which should run those new unwind-forms but won't find it
Stefan> because it won't look at the handlers [ offhand, I can't think of any
Stefan> place where that would happen, and I think it's probably OK, but
Stefan> I can't quite convince myself that it definitely is.  ]

If this happens then surely it's already a problem for catch and
condition-case?

Stefan> All in all, I'd really prefer if we kept the unwind forms in
Stefan> the specpdl.  Was there a particular reason why you used the
Stefan> handlers instead?

This is how catch and condition-case are implemented.

I do agree there's no reason to have a separate handler list.  In fact
this puzzled me for a while.  However I think it's a separate cleanup.

Tom

Re: [Emacs-diffs] feature/byte-unwind-protect 916094a 2/2: Add new bytecodes for unwind-protect

[Stefan Monnier]

>>> +	CASE (Bpushunwindprotect): /* New in 27.1.  */
>>> +	  {
>>> +	    struct handler *c = push_handler (Qt, CATCHER_ALL);
> Stefan> Hmm... using this "catch+rethrow" method interferes with the
> Stefan> debug-on-error stacktrace:
> Hmm, yeah.
> Any ideas for fixing this?

Probably change the maybe_call_debugger code so it ignores
CATCHER_ALL somehow.

> One idea would be to capture the stack trace when CATCHER_ALL is
> targeted, then have a way to pass this to signal or throw.

In general, it'd be nice, yes, except that:

> But often the stack trace isn't needed.

Exactly.

> Stefan> I also wonder if there can be places where we call something like
> Stefan> `unbind_to` which should run those new unwind-forms but won't find it
> Stefan> because it won't look at the handlers [ offhand, I can't think of any
> Stefan> place where that would happen, and I think it's probably OK, but
> Stefan> I can't quite convince myself that it definitely is.  ]
> If this happens then surely it's already a problem for catch and
> condition-case?

But having looked a bit more carefully at the code, I think we're OK, indeed.

> Stefan> All in all, I'd really prefer if we kept the unwind forms in
> Stefan> the specpdl.  Was there a particular reason why you used the
> Stefan> handlers instead?
> This is how catch and condition-case are implemented.

I know.  But until now all the unwind forms are pushed to the specpdl,
whereas the catcher and condition-case handlers go to the
handlers stack.

Your patch changes that by moving the unwind forms to the handlers stack
(tho it keeps the "built-in unwinds" on the specpdl).

> I do agree there's no reason to have a separate handler list.  In fact
> this puzzled me for a while.  However I think it's a separate cleanup.

Yes, merging the two is a different issue (last I looked at that code
I was also tempted to go that way, but the two are used quite
differently, so it's not clear there'd be much benefit).


        Stefan