[clang] [clang] Do not clear FP pragma stack when instantiating functions (PR #70646)

[John McCall via cfe-commits]

rjmccall wrote:

> > The upshot is that we don't (and shouldn't) ever late-parse at file context, which by design means we can't see stack-manipulating pragmas because they're all restricted to file context. In late parsing, we only ever observe and change the innermost state of the pragma stack, and so all we need to do is temporarily restore that innermost state to what it was when the tokens were originally captured, then restore it to its "live" value after the late parsing is complete. The rest of the stack is untouchable.
> 
> "shouldn't ever late parse at file context" is something I agree with when I think about most language constructs, but as a counter-example, folks have been agitating more and more for the ability for an attribute argument to reference a subsequent declaration. e.g.,
> 
> ```
> void func(int *ptr [[attr(len)]], size_t len);
> 
> struct S {
>   int *Ptr [[attr(len)]];
>   int Len;
> };
> ```
> 
> and the likes. I can imagine someone wanting to extend that to:
> 
> ```
> int *GlobalArray [[attr(GlobalLen)]];
> int GlobalLen;
> ```
> 
> because reordering the globals may change other semantic behavior (like initialization order for dynamic init). Should we perhaps add documentation to the internals manual so we capture the issues with file-scope late parsing?

Yes, I think that'd be good.  To me, the right way of thinking about this kind of code is that it's not truly at file context — it's not really standalone, we just don't really have the `Decl` that it's part of yet.  For example, there's a lot of things you really shouldn't be able to write in a statement-expression in a global initializer, the bound of an array type, and so on, but our checking doesn't always prohibit them because e.g. there isn't a meaningful `DeclContext` we can put in `CurContext` that will reflect that this isn't really a local context.  (The way we solved this in Swift's AST is that a `DeclContext` doesn't have to be a `Decl`, and we lazily create an `Initializer` DC when processing other kinds of top-level expressions.)  This is a longstanding issue.

https://github.com/llvm/llvm-project/pull/70646