Re: Question on pcase

[Oleh Krehel <ohwoeowho@gmail.com>]

Hi Micheal,

Thanks for your work, it could be useful to many people.

Michael Heerdegen <michael_heerdegen@web.de> writes:

> Someone who wants to try to learn pcase (Oleh? Oleh!) can help by
> reading it and telling me if it is understandable, and send corrections
> - or write a better introduction ;-) - and format it nicely for
> inclusion into Elpa or Emacs when it turns out to help people.

The reason I dislike `pcase' is not because I don't know how to use it
(the basic rules are actually pretty simple), it's because I think it
leads to code that's hard to understand, maintain and transform. I
dislike the trivial `if-let' and `when-let' for the same reasons.

I generally dislike any custom macro that includes `if' or binds
variables. This is because I can't reason about the code that uses these
macros unless I know exactly what they do in terms of binding variables
and selecting branches. These macros don't follow the substitution model
for procedure application (SMPA) [1], which is a valuable debugging
technique for me.

However, I'm willing to implement some tooling that will allow `pcase'
to follow SMPA.

Take this code for example, with "|" being the point:

(pcase which
  (`all t)
  |(`safe (member fun completion--capf-safe-funs))
  (`optimist (not (member fun completion--capf-misbehave-funs))))

Here, the (`all t) branch is selected, and the `which' symbol
can be extracted from the context with `up-list'.  I've implemented a
function that prints "pcase: nil" if the selected branch doesn't match,
and "pcase: t" when it matches.

This way, when I'm debugging a code with `pcase' I can see which branch
is the correct one without evaluating the code inside the branch.  After
this, I can step into the correct branch and use SMPA.

Of course, this function would also need to bind the same variables that
a `pcase' branch would bind.

Taking your example:

(pcase x
  ('a       1)
  ("Hallo"  2)
  |(thing    (message "%s is neither equal to 'a nor to \"Hallo\"." thing)))

This is what I would like to have:

(equal (macroexpand '(eval-pcase-branch x
                      (thing (message "%s is neither equal to 'a nor to \"Hallo\"." thing))))
       '(progn
         (setq thing x)
         (message "pcase: t")))

After this, the inner body of the branch can be properly evaluated,
since `thing' is bound to `x' now.

So far, I've implemented some code that can check if each branch will be
followed, see
https://github.com/abo-abo/lispy/commit/d3ed4e4fee435a2a448ddc0722d07cd997ee59d3.

But the code that binds the variables, e.g. (setq thing x) will likely
be hard to implement. For instance, look at this example with macroexpand:

(setq test '(1 . 2))
(pcase test
  (`(,foo . ,baz)
    (cons baz foo)))
;; =>
;; (2 . 1)

(macroexpand '(pcase test
               (`(,foo . ,baz)
                (cons baz foo))))
;; =>
;; (if (consp test)
;;     (let* ((x (car test))
;;            (x (cdr test)))
;;       (let ((baz x)
;;             (foo x))
;;         (cons baz foo)))
;;   nil)

The macroexpanded code returns (2 . 2) when evaluated. This I don't
understand. Although, it still works fine with `eval':

(eval (macroexpand '(pcase test
                     (`(,foo . ,baz)
                      (cons baz foo)))))
;; =>
;; (2 . 1)

Maybe someone could explain the above, and also suggest the best way the
create variable bindings from a pcase branch.

thanks again,
Oleh

[1]: https://mitpress.mit.edu/sicp/full-text/sicp/book/node10.html