bug#68938: Emacs "master". Incorrect code generated by pcase.

[Alan Mackenzie <acm@muc.de>]

Hello, Stefan.

On Mon, Feb 05, 2024 at 13:27:21 -0500, Stefan Monnier wrote:
> > In a development version of Emacs, last synched with master in December,
> > I have added the following pcase clause to macroexp--expand-all in
> > lisp/emacs-lisp/macroexp.el:

> >             (`(,(and 'defalias d
> >                      (guard (and (null defining-symbol)
> >                                  (symbol-with-pos-p d))))
> >                ',sym . ,_)
> >              ;; Here, don't change the form; just set `defining-symbol'
> >              ;; for a (defalias 'foo ...) in the source code.
> >              ;; (when (symbol-with-pos-p d)
> >              (setq defining-symbol sym)
> >              form)

> Where did do you add it?

Just before the (`(function ,(and f `(lambda . ,_))) clause, like you
did.

> > ..  pcase expands that clause to this cond clause:

> > ((eq x0 'defalias)
> >  (cond
> >   ((let* ((d x0))
> >      (and (null defining-symbol) (symbol-with-pos-p d)))
> >    (let* ((x14 (cdr-safe form)))
> >      (cond
> >       ((consp x14)
> >        (let* ((x15 (car-safe x14)))
> >          (cond
> >           ((consp x15)
> >            (let* ((x16 (car-safe x15)))
> >              (cond
> >               ((eq x16 'quote)
> >                (let* ((x17 (cdr-safe x15)))
> >                  (cond
> >                   ((consp x17)
> >                    (let* ((x18 (car-safe x17)) (x19 (cdr-safe x17)))
> >                      (cond
> >                       ((null x19)
> >                        (let ((d x0) (sym x18))
> >                          (ignore d) (setq defining-symbol sym) form))
> >                       ((consp x0)
> >                        (let* ((x21 (car-safe x0)))
> >                          (if (eq x21 'lambda) (funcall pcase-3 x0 x14)
> >                            (funcall pcase-2 x0))))
> >                       (t (funcall pcase-2 x0)))))
> >                   ((consp x0)
> >                    (let* ((x23 (car-safe x0)))
> >                      (if (eq x23 'lambda) (funcall pcase-3 x0 x14)
> >                        (funcall pcase-2 x0))))
> >                   (t (funcall pcase-2 x0)))))
> >               ((consp x0)
> >                (let* ((x25 (car-safe x0)))
> >                  (if (eq x25 'lambda) (funcall pcase-3 x0 x14)
> >                    (funcall pcase-2 x0))))
> >               (t (funcall pcase-2 x0)))))
> >           ((consp x0)
> >            (let* ((x27 (car-safe x0)))
> >              (if (eq x27 'lambda) (funcall pcase-3 x0 x14)
> >                (funcall pcase-2 x0))))
> >           (t (funcall pcase-2 x0)))))
> >       ((consp x0)
> >        (let* ((x29 (car-safe x0)))
> >          (if (eq x29 'lambda) (funcall pcase-3 x0 x14) (funcall pcase-2 x0))))
> >       (t (funcall pcase-2 x0)))))
> >   ((consp x0)
> >    (let* ((x31 (car-safe x0)))
> >      (if (eq x31 'lambda)
> >          (let* ((x33 (cdr-safe form))) (funcall pcase-3 x0 x33))
> >        (funcall pcase-2 x0))))
> >   (t (funcall pcase-2 x0))))
> >
> > ..  This contains errors:
> >
> > (i) Although it has been established that x0 is 'defalias, there are many
> >   tests (consp x0).

> Yup, clearly some missed optimization.
> I added your code just before the

>             (`(function ,(and f `(lambda . ,_)))

> branch, and I didn't see such poor code, so it seems that it depends on
> further details.

Yes.  I think my bug report was premature at best.  I was part way
through amending backquote.el to be able to add the position information
into things like the lambda in

                      (push `(,bsym (lambda ,(mapcar #'car varvals)
                                      ,@ignores ,@code))
                            defs))

in pcase--expand.  I think this bit of code deals with creating pcase-1,
etc., so it seems highly likely my tentative changes were to blame.

> On further inspection I see a similar problem in another branch,
> where it generated:

> 	 ((pcase--flip memq '(defconst defvar) x71)
> 	  (let* ((x78 (cdr-safe form)))
> 	    (cond
> 	     ((consp x78)
> 	      (let* ((x79 (car-safe x78)))
> 		(cond
> 		 ((symbolp x79)
> 		  (let ((name x79))
> 		    (push name macroexp--dynvars)
> 		    (macroexp--all-forms form 2)))
> 		 ((consp x71)
> 		  (let* ((x81 (car-safe x71)))
> 		    (if (eq x81 'lambda) (funcall pcase-2 x71 x78)
> 		      (funcall pcase-1 x71))))
> 		 (t (funcall pcase-1 x71)))))
> 	     ((consp x71)
> 	      (let* ((x83 (car-safe x71)))
> 		(if (eq x83 'lambda) (funcall pcase-2 x71 x78)
> 		  (funcall pcase-1 x71))))
> 	     (t (funcall pcase-1 x71)))))

> where we do that same useless (consp x71) test.  I think this case is
> "normal" (the branch's test is basically (memq x71 '(defconst defvar),
> i.e. more complex than (eq x71 'defalias)) and I seem to remember
> consciously punting on handling such things in
> `pcase--mutually-exclusive-p`.

> Your case doesn't sound like one I'm aware of, OTOH.

I think I should just close the bug as not a bug.

It was one of these things I couldn't understand or do anything about,
but right after reporting it, the solution became obvious, and I
couldn't reproduce the bug easily any more.  I've got my amendments to
backquote.el working, now.

> > (ii) There are calls of the form (funcall pcase-2 x0), i.e. (funcall
> >   pcase-2 'defalias).  This causes a wrong-number-of-arguments error.

> I don't see this problem here.  In my case it's

>     (funcall pcase-1 x71)

> but the number of arguments is right since pcase-1 is defined a bit
> earlier as:

>       (lambda (func)
> 	(let ((handler (function-get func 'compiler-macro)))
> 	  (if (null handler) (macroexp--all-forms form 1)
> 	    (unless (functionp handler)
> 	      (with-demoted-errors "macroexp--expand-all: %S"
> 		(autoload-do-load (indirect-function func) func)))
> 	    (let ((newform (macroexp--compiler-macro handler form)))
> 	      (if (eq form newform)
> 		  (if
> 		      (equal form
> 			     (setq newform
> 				   (macroexp--all-forms form 1)))
> 		      form
> 		    (setq form
> 			  (macroexp--compiler-macro handler newform))
> 		    (if (eq newform form) newform
> 		      (macroexp--expand-all form)))
> 		(macroexp--expand-all newform))))))

> > (iii) There is no sign of the final `form' being returned, though this
> >   may be being done elsewhere.

> I see the following in your code sample:

>                      (cond
>                       ((null x19)
>                        (let ((d x0) (sym x18))
>                          (ignore d) (setq defining-symbol sym) form))

> which seems to be correctly returning `form`.

Yes.  Sorry about this bug report, which turned out to be a time waster.

>         Stefan

-- 
Alan Mackenzie (Nuremberg, Germany).