Re: [Maxima-discuss] Changing ERRSET to support multiple values

[Raymond T. <toy...@gm...>]

>>>>> "Kris" == Kris Katterjohn <kat...@gm...> writes:

    Kris> Hey everyone,
    Kris> I'm working on a bug fix and ERRSET's behavior when it comes to multiple
    Kris> values has come up.  I'm proposing a possible change below.  Either I
    Kris> can make the change or I'll do something else for the bug fix.

    Kris> The ERRSET variable and macro are used in quite a few places in Maxima.
    Kris> When ERRSET is non-NIL, the ERRSET macro allows errors that occur during
    Kris> the evaluation of its first argument to propagate.  When ERRSET is NIL,
    Kris> the ERRSET macro currently behaves like this:

    Kris> (errset (+ 1 'doh))     => NIL
    Kris> (errset (+ 1 2))        => (3)
    Kris> (errset (values 1 2 3)) => (1)
    Kris> (errset (values))       => (NIL)

    Kris> That is, NIL is returned if an error occurs and a singleton list
    Kris> containing the (primary!) return value from the evaluated form is
    Kris> returned if no error occurs.

    Kris> Notice that ERRSET and multiple values don't mix well.  It's unclear
    Kris> whether or not extra return values got silently dropped, and there is an
    Kris> ambiguous case where we get a singleton list containing NIL if the form
    Kris> really does return NIL or if it actually returns no values at all.

    Kris> I'd like to extend ERRSET to handle multiple values by having it return
    Kris> a list of all return values from the evaluated form in the non-error
    Kris> case.  So then we'd have

    Kris> (errset (values 1 2 3)) => (1 2 3)
    Kris> (errset (floor 4/3))    => (1 1/3)

    Kris> and so on.  This is consistent with the current behavior when there is a
    Kris> single return value.

    Kris> This can be easily accomplished with a couple of changes like this:

    Kris> (defmacro errset (&rest l)
    Kris> -  `(handler-case (list ,(car l))
    Kris> +  `(handler-case (multiple-value-list ,(car l))
    Kris> (error (e) (when errset (error e)))))

    Kris> The only possible problem I see is that now we have a different
    Kris> ambiguous case: we can get NIL back if an error occurs or if the form
    Kris> returns no values at all:

    Kris> (errset (error "uhoh")) => NIL
    Kris> (errset (values))       => NIL

    Kris> That's certainly not optimal.

    Kris> The test suite runs fine with this ERRSET definition.

    Kris> To avoid this ambiguous case I could add a special case for it:

    Kris> (defmacro errset (&rest l)
    Kris> -  `(handler-case (list ,(car l))
    Kris> +  `(handler-case (or (multiple-value-list ,(car l)) (list nil))
    Kris> (error (e) (when errset (error e)))))

    Kris> Then

    Kris> (errset (error "uhoh")) => NIL
    Kris> (errset (values))       => (NIL)
    Kris> (errset (values 1 2 3)) => (1 2 3)

    Kris> That's kinda gross, but this way the new behavior is precisely like
    Kris> the current behavior when there are one or zero return values, it works
    Kris> in the desired way when there are two or more return values, and no case
    Kris> can be confused with the error case.  So really the only difference is
    Kris> that we get all return values instead of just the primary one when there
    Kris> are two or more values.

I think your approach is good.

An alternative approach, which is a lot more work, is to return two
values like cl::ignore-errors.  If there are no errors, return the
(multiple) values of the form.  If there is an error, return nil and a
second value that is the error condition. This means updating all the
call sites to handle this case, though, and that might not be
desirable.

--
Ray