[clang] [WinEH] Check object unwinding in seh block only when -eha is used (PR #180108)

[via cfe-commits]

GkvJwa wrote:

> Sorry for the repeated questions, but I'd really like to understand this better :)
> 
> In #172287 @efriedma-quic said
> 
> > The basic issue here is that you can't generate unwind tables for both SEH and C++ exception handling in the same function.
> 
> That makes sense to me.
> 
> In our case, we are (as far as I'm aware) building with C++ exceptions disabled, so that issue should not apply to us?
> 
> That's why I don't understand why this PR gates the `err_seh_object_unwinding` diagnostic behind `getLangOpts().EHAsynch`. If the problem is mixing SEH and C++ exceptions, why not gate it behind `getLangOpts().CXXExceptions`?
> 
> > SEH is important for Windows, it can prevent many EXCEPTION_IN_PAGE_ERROR errors. **For LLVM, it only works when using /EHa**
> 
> My understanding was that LLVM has had SEH support for a long time, but only for synchronous SEH exceptions. I.e. the following should work:
> 
> ```
> __try {
>   foo();  // Bad memory access in here.
> } __except(EXCEPTION_IN_PAGE_ERROR) {
>   // ...
> }
> ```
> 
> even without `/EHa` (as long as foo is not inlined)?
> 
> > This issue(#62606) can probably answer most questions. The key point is that the handling of SEH in LLVM is not quite the same as in MSVC. Currently, it can be said that with LLVM, the code within the __except block is only explicitly generated when using /eha.
> 
> IIUC that issue is about async SEH. We don't build with `/EHa`, but my understanding was that synchronous SEH should still work?

I added this code(#62606) snippet to Chromium, and the crash still occurs even without `/EHa` enabled.
```
#include <stdio.h>
#include <windows.h>

int SafeFetch32(const int* adr, int errValue) {
  int v = 0;
  __try {
    v = *adr;
  } __except (EXCEPTION_EXECUTE_HANDLER) {
    printf("except\n");
    v = errValue;
  }
  return v;
}

// call this function
  SafeFetch32(0, -1);
```

There should be many similar codes in Chromium; for example, the code here also won't generate SEH information(like `_C_specific_handler`).
https://source.chromium.org/chromium/chromium/src/+/main:sandbox/win/src/process_thread_interception.cc;l=95

Of course, some code does indeed have SEH information because it meets certain conditions.

Since `/EHa` is not enabled(Can't open it either, has crash), it is impossible to determine exactly which parts are malfunctioning.

The `/EHa` in LLVM is not equivalent to C++ exceptions or MSVC's SEH, and the code still has many imperfections.

Ultimately, I prefer to provide a set of SEH parameters rather than putting SEH within `/EHa`(brought some problems)

https://github.com/llvm/llvm-project/pull/180108