[PATCH] Add PragmaAttr and Pragma Spelling to Tablegen

[Aaron Ballman]

On Wed, May 28, 2014 at 2:45 PM, Reid Kleckner <rnk at google.com> wrote:
> On Wed, May 28, 2014 at 6:42 AM, Aaron Ballman <aaron at aaronballman.com>
> wrote:
>>
>> On Tue, May 27, 2014 at 6:28 PM, Reid Kleckner <rnk at google.com> wrote:
>> > On Sat, May 24, 2014 at 8:28 AM, Aaron Ballman <aaron at aaronballman.com>
>> > wrote:
>> >>
>> >> > My first idea is to add a tablegen property like "HasPrettyPrinter =
>> >> > 0
>> >> > or 1"
>> >> > that supresses emission of the default printPretty method.  Then
>> >> > again,
>> >> > we
>> >> > kind of want the default version when the GNU and C++11 spellings of
>> >> > the
>> >> > attribute are used.
>> >>
>> >> That seems like the wrong way to go about it though, since this is a
>> >> new *spelling*, not a wholly new semantic kind of attribute (which I
>> >> think we really want to avoid). We could possibly attach a custom
>> >> pretty printer to the spelling itself (a la a code block), but that
>> >> seems a bit skeezy because the arguments are rather important and the
>> >> spelling shouldn't have to care about the arguments.
>> >
>> >
>> > How about this: when generating the prettyPrint implementation, if there
>> > is
>> > a pragma spelling, call MyAttr::printPrettyPragma(...).  Any attribute
>> > with
>> > a pragma spelling will have to provide this method in
>> > lib/AST/AttrImpl.cpp.
>>
>> So my ultimate goal is for pragma attributes to mirror all of the
>> other attributes, where we can infer syntax and semantics just from
>> what's in Attr.td. But it's becoming more obvious to me that the
>> amount of work required to get there is rather large (mostly due to
>> the fact that pragma argument syntax is ALL over the place).
>>
>> As a stopgap solution, it seems reasonable to me to rely on an
>> AdditionalMember being defined for pragmas (which is basically what
>> you're suggesting).
>>
>> So if an attribute has a pragma spelling, it is required to have:
>>
>> let AdditionalMembers = [{
>>   void printPrettyPragma(raw_ostream &OS, const PrintingPolicy &Policy)
>> const {
>>   }
>> }];
>
>
> Sure, but... I like writing C++ in .cpp files, not .td files.  +1 for
> defining it out of line in AttrImpl.cpp.  :)

That's an understandable desire, but a step in the wrong direction for
two reasons. 1) Anything declarative that we can put into the .td file
belongs in the .td file. There's nothing about this problem which
isn't declarative aside from the fact that we don't have the time
(yet) to make the proper declarations regarding parameter syntax. 2)
Putting it into AttrImpl.cpp decouples the logic from what gets
generated, which is far worse than writing some C++ code in a .td
file. You would have to write an out-of-line definition for
FooAttr::printPrettyPragma where the declaration gets automatically
generated for you elsewhere, which is rather ugly as sin when we
already have an alternative that suffices.

Also, this isn't a whole lot of complex C++ code either. The less of
it you write in the .td file, the more likely it is we will get the
declarative portions down without a mess of refactoring.

~Aaron