[PATCH] Attribute parameters and arguments

[Aaron Ballman]

On Tue, Aug 27, 2013 at 7:08 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> On Tue, Aug 27, 2013 at 3:33 PM, Aaron Ballman <aaron at aaronballman.com>
> wrote:
>>
>> On Tue, Aug 27, 2013 at 6:27 PM, Richard Smith <richard at metafoo.co.uk>
>> wrote:
>> > On Tue, Aug 27, 2013 at 3:16 PM, Aaron Ballman <aaron at aaronballman.com>
>> > wrote:
>> >>
>> >> Attributes can have a single parameter, as well as multiple (sometimes
>> >> optional) arguments.
>> >
>> >
>> > What is the distinction you're drawing between parameters and arguments
>> > here?
>>
>> Parameters are used by GNU-only attributes, arguments everywhere else.
>>  The distinction is required so that we can properly handle error
>> reporting when parameters and arguments are both required for the
>> attribute.
>>
>> > If you mean the weirdness in the GNU attribute syntax that we sometimes
>> > use
>> > different parsing rules for the first argument, I don't think modeling
>> > that
>> > as a fundamentally different kind of argument is the right way to
>> > proceed.
>> > Instead, we should look at the list of Args when parsing an attribute
>> > argument list, to determine how to parse each element.
>>
>> I do mean that weirdness, but this has very little to do with parsing.
>>  The parsing logic is identical to its previous state, except instead
>> of looking for things which don't have params and bailing out, we look
>> for the much smaller list of things which do have params and include
>> them.
>>
>> I dislike the notion of looking at the first argument in the list and
>> essentially guessing whether it's a parameter.
>
>
> It would not be a guess. We would look at the kind of the argument, and
> parse an entity of that kind.
>
>> That makes semantic
>> checking basically impossible to generalize because we don't know
>> whether we should be looking for a parameter or not.  FWIW, this
>> caused (what I think are) bugs that I rectified in the test cases once
>> the logic was made more explicit.
>
>
> The parameter/argument thing is a bogus distinction that we made up. Under
> the GNU syntax, the first argument sometimes has different parsing rules,
> that's all. The right thing to do is to change AttributeList to represent
> the "parameter" as just another argument, not to make the distinction here
> more ingrained.

Okay, after discussing it a bit on IRC, I think it's a good idea to
drop the param distinction and try to consolidate the param with the
args.  I'll do some exploration on the subject.  In the meantime, I
may also try to back out the param-specific changes from my patch and
resubmit if I can get the error reporting reasonable.

My biggest question is: are you okay with the approach I took in
AttributeList for obtaining information from the tablegen?

~Aaron