[clang] [clang][Sema] Diagnose exceptions only in non-dependent context in discarded `try/catch/throw` blocks (PR #139859)

[Erich Keane via cfe-commits]

erichkeane wrote:

> > I feel like this issue is a little too niche to be adding such complexity, maybe just covering those specific uses cases before codegen might be the way to go?
> 
> I think disabling the keywords entirely is a large ask (requires an RFC at a minimum) and so we don't need to go down that route right now.

I would be against said RFC.  If you asked me in 1996/whatever, I could be talked into it.  But we're 30+ years of this behavior.

> 
> I think `constexpr` exceptions are a red herring; if the user disabled C++ exceptions, they should be disabled at compile time same as at runtime. Some users disable exceptions because of overhead, but plenty of users also disable exceptions because of difficulties with reasoning about flow control (some style guides explicitly require exceptions to not be used, for example).

I'm coming around on this one, I think I agree.  -fno-exceptions means `no-exceptions`.  IF we get a further feature request to allow them to work at const-eval time, perhaps we can consider that, but it seems reasonable to do the disallow at the 'fully instantiated' point as I suggested above.

> 
> Since we currently validate that the exception constructs are semantically valid, I think we may as well retain them in the AST for now. It's how we currently behave: https://godbolt.org/z/xPbEPWTK8 so it would be consistent to do so in discarded statements too.



https://github.com/llvm/llvm-project/pull/139859