bug#68509: 30.0.50; pcase-dolist matches backquote pattern incorrectly

[Stefan Monnier]

>> Consider the following
>> (pcase-dolist (`(,(and (pred stringp) a) .
>>               ,(and (pred stringp) b))
>>             '(("TODO") ("DONE" . "a")))
>>    (warn "%S :: %S" a b))
>> Executing the above yields
>> ⛔ Warning (emacs): "TODO" :: nil
>> ⛔ Warning (emacs): "DONE" :: "a"
>> even though ("TODO") does not match the pattern.
>
> This isn't an issue with 'pcase-dolist', but rather a known/intentional
> limitation of 'pcase-let':

Indeed, I consider the above a pilot error.
`pcase-dolist` and `pcase-let` use Pcase patterns to do *destructuring*,
which is a different task than the one done by `pcase` (which decides
whether a value matches a pattern or not).

>> Each EXP should match (i.e. be of compatible structure) to its
>> respective PATTERN; a mismatch may signal an error or may go
>> undetected, binding variables to arbitrary values, such as nil.
> I do think we should fix it somehow though.  This behavior is
> extremely confusing, and as much as I'm a fan of 'pcase', I'm
> emphatically *not* a fan of how this part works.

I'm quite happy with the way it works when you use it as intended.
But I'm not really satisfied either with the way it behaves when the
coder doesn't understand its semantics, nor about the way we document
that semantics.

The difficulty in resolving this can be illustrated with the following
pattern:

    `(a . ,b)

This pattern leads to two tests: (consp VAL) and (eq 'a (car VAL)).
When destructuring, we want to throw away both tests (we want to throw
away most tests, except those needed to choose between two `or`
branches).

We could decide to emit a warning because we silently skip
the `eq` test, which would help the coders understand that the pattern
doesn't do what they think.
But emitting that same warning because we silently skip the `consp` test
would be really annoying because rewriting the pattern to avoid this
is impractical.

For a human, it's pretty easy to distinguish those two cases.  But it's
difficult to provide a precise definition that distinguishes those two cases.

We could also keep the tests and emit a warning or even an error when
they fail, but if that's the behavior you want, then you should
arguably use `pcase(-exhaustive)` instead.


        Stefan