[PATCH] Add PragmaAttr and Pragma Spelling to Tablegen

[Aaron Ballman]

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 {
  }
}];

Then in ClangAttrEmitter.cpp, the printPretty generator function can
simply call through to this for a pragma spelling in
writePrettyPrintFunction. If the attribute has no AdditionalMembers
field, the generator can spit out a friendly error, but if the author
simply forgets the signature for this, a compile-time error will have
to suffice.

Someday, once we solve the parameter syntax issue, we can rip out the
printPrettyPragma functions. That can be added as a FIXME to
writePrettyPrintFunction.

~Aaron