[clang] [Clang] restrict use of attribute names reserved by the C++ standard (PR #106036)

[Aaron Ballman via cfe-commits]

================
@@ -177,6 +177,26 @@ static bool isLanguageDefinedBuiltin(const SourceManager &SourceMgr,
   return false;
 }
 
+static bool isReservedAttrName(Preprocessor &PP, IdentifierInfo *II) {
----------------
AaronBallman wrote:

These are great questions!

`__has_cpp_attribute` should return nonzero if Clang "knows" about the attribute with the current set of compilation flags. e.g., if you would not get an "unknown attribute ignored" warning for its use.

The library part of the standard reserves attribute names, including in freestanding mode, but only for purposes of use with `#define` or `#undef`. So if `__has_cpp_attribute` is zero for a standard attribute with a certain set of flags, the name is not reserved; otherwise, it is reserved and should be diagnosed if it's used as the macro name in `#define` or `#undef`.

> Should this one handle the standard versions too?

I don't think so; we expose attributes in older language modes as a conforming extension (and we fail to issue a `-pedantic` diagnostic for it, which perhaps is a mistake?). For example, `no_unique_address` is C++20, but we still expose it in C++11: https://godbolt.org/z/6TWaodjE3

So I think we ultimately want the reserved identifier warning to key off `__has_cpp_attribute` behavior; if we claim we support the attribute (even as an extension), library authors should be able to use it, and therefore we should behave as though it was reserved for purposes of `#define` and `#undef`.

CC @erichkeane for additional opinions

https://github.com/llvm/llvm-project/pull/106036