[WIP] [x86/Linux] Call EH Funclet with CallEHFunclet

[wateret]

Implement CallEHFunclet same as ARM/ARM64.'

This follows ARM/ARM64 way for calling non-filter funclets which is actually not the way discussed in #9484. This can handle simple try-catch statements but fails for nested exceptions. (e.g. throw in catch block)

[wateret]

cc: @seanshpark @parjong @YongseopKim

[parjong]

This line should be excluded.

[wateret]

OK thanks!

[parjong]

It would be better to use genPushCalleeSavedRegisters instead of explicitly adding push instructions.

[jkotas]

Is there a good reason for saving these registers here? The asm helper that calls the funclet will save them, so the funclet does not have to ... unless there is some other reason.

[wateret]

I'll remove them, thanks.

[parjong]

Similarly before, genPopCalleeSavedRegisters would be better.

[wateret]

About nested catch failure, for the test case like below:

try {
  try {
    throw A;
  } catch { // catches A
    throw B;
  }
} catch { // catches B
}

When excuting the funclet that catches A, the funclet holds ebp that is restored by CallEHFunclet. Maybe it makes the unwider starts stack-walking from the main function frame, not from the funclet frame. (Actually I'm not sure how exactly unwinder works.)
If this is the case, I guess there would be no problem if it was AMD64 style funclet. That is: there is no CallEHFunclet and the ebp is saved and restored by the funclet itself.

[wateret]

@jkotas @parjong
I think the problem with nested catch failure is, like I said above, wrong ebp value. However now I don't think just switching to AMD64 style funclet would resolve this issue because ebp value is still not valid for unwinder.
For arm it seems like CallEHFunclet stores frame pointer in r7 and unwinder uses it as frame pointer. How could we make x86 work? Maybe knowing how AMD64 works would also help.

[parjong]

@wateret While digging JIT.Methodical.casts.SEH._il_dbgcastclass_catch_neg test failure, I conclude that we need to revise CLR's managed unwinder to treat funclet frame as ESP frame.

I made a draft patch in https://github.com/parjong/coreclr/commit/a992b93ec85cf70c54558ff63a3b11ee0809b4db, it seems to work for JIT.Methodical.casts.SEH._il_dbgcastclass_catch_neg case. I'm currently running test to see its side-effect. I'll let you know when I have a result.

I'm not sure whether that patch could be aligned with your current patch as it does not use PSPSym at all, but it would be worthwhile to take a look.

[wateret]

Closing as #9601 includes this.