[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Keyword precedence
Mark Hadfield wrote:
> > > Whew! Thanks for that explanation, JD. I think I understand it now.
>
> Well, I didn't, but I'm getting there.
>
> > ... I think this was a slippery slope, since overriding a
> > *value* is very different than overriding a *variable*.
> > The former is relatively more straightforward.
> > If RSI had stuck to a _REF_EXTRA used only for returning values
> > up inherited keywords in chains of calls, we wouldn't have
> > this ambiguity... it really doesn't make sense to override the
> > *location in memory* associated with a given
> > inherited keyword variable at runtime...
>
> You'll be relieved to hear that I've finally grasped this point. I was
> focussed entirely on passing information *inwards* through the chain of
> calls.
>
> But before we give up entirely, let's consider the following again:
>
> IDL> mgh_example_keywords, COLOR=12
> % MGH_EXAMPLE_KEYWORDS: Calling color-print routine directly
> % MGH_EXAMPLE_KEYWORDS_PRINT_COLOR: COLOR is 12
> % MGH_EXAMPLE_KEYWORDS: Calling color-print routine via value wrapper
> % MGH_EXAMPLE_KEYWORDS_VALUE_WRAPPER: Passing along EXTRA keywords by value:
> COLOR{ 12}
> % MGH_EXAMPLE_KEYWORDS_PRINT_COLOR: COLOR is 12
> % MGH_EXAMPLE_KEYWORDS: Calling color-print routine via reference wrapper
> % MGH_EXAMPLE_KEYWORDS_REFERENCE_WRAPPER: Passing along EXTRA keywords by
> reference: COLOR COLOR
> % MGH_EXAMPLE_KEYWORDS_PRINT_COLOR: COLOR is 0
>
> A key difference between the value wrapper and the reference wrapper is that
> the former passes only one keyword (which it sees as the structure {COLOR:
> 12}) down to the next level, whereas the the reference wrapper passes both,
> (which it sees as the string array ['COLOR', 'COLOR']).
>
> Now that you've explained it, I see that that the reference wrapper retains
> both because either of them might point to a variable to be modified on
> output. By the way, in IDL 5.1 and before, the value wrapper also passed
> both keywords as {COLOR: 0, COLOR: 12} and the color-print routine chose the
> first of them. But this was changed in 5.2 to the current behaviour.
>
> So far this is all OK and in line with your explanation, but we agree there
> has to be a rule to choose which of the keywords the reference wrapper will
> pass to the color-print routine. By abbreviating the keyword name in the
> call to MGH_EXAMPLE_KEYWORDS (the "user keyword") and/or the name of the
> keyword specified inside MGH_EXAMPLE_KEYWORDS (the "default keyword") we can
> establish empirically that IDL behaves according to the following rules:
>
> * Inside the reference wrapper the keywords are represented in the EXTRA
> string array in the order [<default>, <user>].
>
> * If both keywords are represented by identical strings (case-insensitively)
> then the first of them (default) is passed to the color-print routine. If
> the keyword-strings are not identical then the second of them (user) is
> passed on.
>
> The second rule is a pretty weird one, I think you'll agree. Maybe it's just
> a special case of some logical general rule--but I doubt it.
>
> Re your proposed rules:
>
> > 1) 2 or more by-value and 0 by-reference(arg_present==0): Default to the
> > standard _EXTRA rules for overriding and abbreviations (Longest match
> first).
> > 2) 1 or more by-value and 1 by-reference: always pass the variable,
> > by-reference. Useful for modifying a value *and* returning a result.
> > 3) 0 or more by-value and more than 1 by-reference: generate an error.
> Not too
> > burdening since you have to go out of your way to achieve this situation.
>
> First, as an aside to rule 1, I now think that abbreviation does *not*
> affect precedence.
>
> But to the main point, assigning precedence based on arg_present would mean
> that this...
>
> color = 0
> mgh_example_keywords_reference_wrapper, COLOR=color, _EXTRA=extra
>
> could give a different result from this...
>
> mgh_example_keywords_reference_wrapper, COLOR=0, _EXTRA=extra
>
> even though the programmer's intention in both cases might be identical.
> Scary!
I disagree. The programmers intentions are ambiguous. Does he wish a return
value out? Is he passing a value in? Both? More scary is the notion of:
mgh_example_keywords_reference_wrapper, COLOR=color, _EXTRA=extra
putting a return value somewhere other than in the variable "color"... maybe not
even on this level... maybe n levels up somewhere. This value-return paradigm
is what _REF_EXTRA was intended for, and if we have to choose between scary
by-value behaviour and scary by-reference behaviour, I choose the former. I
think you are imagining cases in which you don't have control over the
inheritance chain, and may not know you are using _REF_EXTRA. This is a danger
I suppose, but it seems to an uncommon situation. Just as with normal
positional parameters, the programmer must be sure to define in advance how each
will be used: for input values, for return values, or for both. This affects
their usage! You can't say:
mypro, retval
and expect
mypro, 1
to work the same if the first positional parameter is being used to return a
value... no matter what the user intends. Now, _REF_EXTRA is more complicated
since more than two calling level are involved, but if you simply regard it as a
tunnel through intermediate levels, the same rules apply.
> Can we think of by-reference keywords as a set of named pipes through IDL
> memory space, linking different levels in the calling stack? Inheritance
> lets a bundle of these pipes pass through a routine's scope without the
> routine having to worry about their names. I think that the behaviour I am
> expecting is that each routine takes a bundle from its caller and can add
> additional pipes on the inside of the bundle (the beginning of the EXTRA
> list) or take pipes by name off the outside of the bundle (the end of the
> EXTRA list).
This is a good analogy. Each individual pipe is accessible at each end only.
One addition: all pipes originate at a given level (think of a bunch of flowers
having their stems cut as a group), but may end any level up. This allows, for
example, a chained series of GetProperty methods traversing the (class)
inheritance tree up 10 levels to "stop-off" at each appropriate level and
collect the relevant information. New pipes cannot originate mid-level, in
contrast with _EXTRA. The power gained in this tradeoff is, of course, that
_REF_EXTRA pipes flow in both directions!
> > In any case, just beware of mixing your _EXTRA metaphors in the meantime.
>
> I don't think this has anything to do with mixing by-value and by-reference
> inheritance. (Proof: change _EXTRA to _REF_EXTRA in the procedure definition
> for MGH_EXAMPLE_KEYWORDS. By-value inheritance is not involved at all when
> the reference-wrapper is called but the behaviour is exactly the same.)
Unfortunately your test is not sufficient proof, since _REF_EXTRA has hidden
_EXTRA (read: by-value) functionality built inside of it (though in a perverse
way). The way to see this is to put a "print, arg_present(color)" in your
mgh_example_keywords_print_color procedure, and change the calling keyword of
the reference wrapper to COLOR=col, where col is a local variable defined in
mgh_example_keywords. Now you can change whether the by-reference or by-value
form of _REF_EXTRA gets invoked by calling mgh_example_keywords with either no
keyword/exactly "COLOR" or some abbreviation of "COLOR". Definitely not a good
situation. If the rules I suggested were followed, you'd always get the
reference variable here.
In any case, hopefully an RSI person or two will get the basic notion that this
needs to be straightened up. And for those of you who have resolved never to
include _REF_EXTRA in your programs, please be assured that this really affects
only a very limited subset of cases of use.
JD
--
J.D. Smith /*\ WORK: (607) 255-6263
Cornell University Dept. of Astronomy \*/ (607) 255-5842
304 Space Sciences Bldg. /*\ FAX: (607) 255-5875
Ithaca, NY 14853 \*/