[PATCH] D29868: Recover more gracefully when __declspec is not supported as a keyword

[David Majnemer via Phabricator via cfe-commits]

majnemer added inline comments.


================
Comment at: lib/Parse/ParseDecl.cpp:2989
+
+          Diag(Loc, diag::err_ms_attributes_not_enabled);
+          continue;
----------------
aaron.ballman wrote:
> compnerd wrote:
> > aaron.ballman wrote:
> > > compnerd wrote:
> > > > I think that we want to emit the diagnostic even if there is no parenthesis as `__declspec` is a reserved identifier, and we would normally diagnose the bad `__declspec` (expected '(' after '__declspec').
> > > Yes, but it could also lead to a rejecting code that we used to accept and properly handle when __declspec is an identifier rather than a keyword. e.g.,
> > > ```
> > > struct __declspec {};
> > > 
> > > __declspec some_func(void);
> > > ```
> > > By looking for the paren, we run less risk of breaking working code, even if that code abuses the implementation namespace (after all, __declspec it not a keyword in this scenario).
> > But we would reject that code under `-fdeclspec` anyways.  I think having the code be more portable is a bit nicer.
> After discussing in IRC, I decided that I agree with @compnerd on this and have changed the patch accordingly.
What if somebody wants to use __declspec and are using the compiler in a freestanding mode? Also, we aren't the only member of the implementor's namespace.


https://reviews.llvm.org/D29868