[clang] [Clang] Support libstdc++ workarounds when using `-E` (PR #210802)

[via cfe-commits]

Sirraide wrote:

> I think there are circumstances where the preprocessor may emit tokens that, if run through the preprocessor again, will produce _different_ tokens. e.g.,

Yes, if you preprocess twice then you will expand macros twice and things that wouldn't have been recognised as directives might be; that's not great either, but that's an orthogonal problem that we'd want something like `-fpreprocessed` for.

> IMO the 'correct' way ahead is "name-value pair" type pragma. IF it errors if it is not the `__GLIBCXX__` (or w/e), and if the value is a non-integer, that is still perfectly acceptable.

I’m fine w/ that too. The original approach was that the macro would define `__GLIBCXX__` rather than set the value directly, so we could go with that instead. That said, I would suggest that this pragma _not_ care about things like `#undef`, warnings about unused macros, or `#pragma clang final`—because essentially, users should never be writing this pragma; and conversely, any compiler-generated uses  should involve only macros that are reserved by the implementation (and a user doing anything like `#undef` to those would be UB anyway). 

I also don’t see `_MSC_VER` or `__GLIBCXX__` being redefined halfway through the TU (and even  if `__GLICXX__` were to be redefined like that, our current infrastructur for the workarounds wouldn’t account for that since it just computes the value once and then caches is), so I don’t think we need to care about this case.

https://github.com/llvm/llvm-project/pull/210802