r319297 - Toolchain: Normalize dwarf, sjlj and seh eh

[Martin Storsjö via cfe-commits]

On Wed, 29 Nov 2017, Reid Kleckner wrote:

> On Wed, Nov 29, 2017 at 12:21 PM, Martin Storsjö <martin at martin.st> wrote:
>       On Wed, 29 Nov 2017, Martell Malone via cfe-commits wrote:
>             Thanks for letting me know Reid.
>             I’ll in work and won’t be able to access the repo
>             until lunch time. (~3
>             hours)
>             Feel free to revert if it is not trivial.
>
>             The easy fix might be to change to == x86_64 from !=
>             x86 For is Windows in
>             the default toolchain. That should restore the old
>             behavior.
> 
>
>       My suggestion would be to just return None for all architectures
>       for the default windows (msvc) case. We didn't use to set any
>       defines to indicate EH mode there before anyway, so setting it
>       to None should make things behave just as before, right?
> 
> 
> I did this slightly differently in r319363, but maybe that's silly. My
> reasoning was that `clang -cc1 -fseh-exceptions -fexceptions` in the MSVC
> environment should still use __CxxFrameHandler3. -fseh-exceptions indicates
> what format of unwind information we should use, and we're still using the
> normal SEH .xdata opcodes.

No, I think this change makes sense - I was just about to write a comment 
pointing out this spot in the code but was waiting for a compile to finish 
before posting.

> Alternatively, you could view -fseh-exceptions, -fdwarf-exceptions, and
> -fsjlj-exceptions as choices of EH personality function,

No, I don't think that'd make sense

> in which case, my change is wrong, and we should do what you guys were 
> suggesting and stop passing this from the driver. Hard to say.

My point was that for the MSVC mode, this worked fine before since this 
was the default behaviour - avoiding passing any option (i.e. having that 
function return None) should be a safe way to keep things working as it 
did before. But then it would break again if -fseh-exceptions were passed, 
which I would expect to be a no-op in such a configuration.

My remaining concern is mostly about why we still need the workaround for 
x86 in the function getting the default (returning None instead of WinEH 
for that case). But as long as this works and the rest of this change can 
settle, we can look at that later if any further change is warranted at 
all.

// Martin