[clang] [WinEH] Fix crash, object unwinding in the except block (PR #172287)

[via cfe-commits]

GkvJwa wrote:

> > I think this flew under the radar when this was all getting implemented because we didn't handle async exceptions, so designed the problem away by just turning off EH cleanups if SEH try was used in a function body. I suppose -EHa undoes that logic, bringing this problem back to life.
> > Personally, I don't like the design proposed here of adding a standalone recursive walk of the AST and attempting to pattern match things that might not compile. If the asynch EH numbering algorithm has bugs, I think we'd be better off improving the backend diagnostic to try to make it more actionable to users with LLVM IR debug info. We can leverage the internal / user fatal error distinction and some source locations to emit a more readable error diagnostic and that will help in all cases, instead of having this pattern match that creates additional maintenance and that we can't guarantee fixes all holes in the asynch EH numbering scheme.
> 
> This reads similar to what i tried to suggest here: [#172287 (comment)](https://github.com/llvm/llvm-project/pull/172287#discussion_r2634654818)
> 
> In this case we just handle the case for emitting a `seh.scope.begin`, which is exactly what we do not handle in the SEH state numbering (only in the CXX state numbering).

like this?
```
  if (CGF->getLangOpts().EHAsynch && IsEHCleanup && !IsLifetimeMarker &&
      CGF->getTarget().getCXXABI().isMicrosoft() && CGF->getInvokeDest()) {
    if (const auto *FD = dyn_cast<FunctionDecl>(CGF->CurFuncDecl)) {
      if (FD->usesSEHTry()) {
        CGF->CGM.getDiags().Report(CGF->CurFuncDecl->getLocation(),
                                   diag::err_seh_expected_handler);
      }
    }
    CGF->EmitSehCppScopeBegin();
  }
```

https://github.com/llvm/llvm-project/pull/172287