[clang] [Clang] Fix parsing of reversible type traits in template arguments (PR #95969)

[Richard Smith via cfe-commits]

================
@@ -141,3 +141,15 @@ namespace r360308_regression {
     return a == b;
   }
 }
+
+namespace GH95598 {
+template<typename _Tp, bool _IsPtr = __is_pointer(_Tp)>
+struct __is_pointer {};
+// expected-warning at -1 {{keyword '__is_pointer' will be made available as an identifier for the remainder of the translation unit}}
----------------
zygoloid wrote:

I think we've incrementally ended up in a bad place. Here's my memory of what happened:

Originally, the idea was: we provide builtins such as `__is_pointer` for use by the standard library. But for a bunch of those, libstdc++ provides a type with the same name that works it out for itself. OK, we can work around that: if we see a declaration with that name, then the standard library isn't using our builtin, so we treat it as a non-keyword so that libstdc++ doesn't break.

But then it turns out that boost uses these things as keywords directly, rather than using the standard library functionality. To keep both that and libstdc++ working at the same time, we treat the keyword followed by an open paren as having the builtin meaning even when the keyword has been reverted to being a non-keyword.

And now we're noticing that we also need special behavior in other parts of the parser to treat the keyword followed by an open paren as a special case.

I don't think we should be extending this hack further. I've suggested on the libstdc++ bug report that this be fixed in libstdc++ instead, by stopping using our keywords as their type names. Maybe one day we can then remove it entirely.

https://github.com/llvm/llvm-project/pull/95969