Re: [trans-mem] inlining fixes

[Richard Henderson <rth at redhat dot com>]

On 08/06/2009 08:48 AM, Jan Hubicka wrote:
> > This fixes up some niggly points when inlining TM_ATOMIC statements,
> > both with the ERT_TRANSACTION eh region and with marking which
> > statements can and cannot throw.
> >
> > The eh-1.C testcase still does not work at -O3 because a dominator walk
> > to find the end of an atomic region does not work with the existing
> > implementation of RESX statements.
>
> What is the problem here?
> (I am still slowly working on the EH changes, so I wonder what kind of
> problems we run into)

My code is relying on a dominator walk to maintain transaction
regions, and the current situation with EH is that it gives
incorrect dominator results.  In particular:

  struct S { ~S(); }
  void bar ();

  void foo()
  {
    try {
      S s;
      bar ();
    } catch (...) {
    }
  }

The CFG that we create does not have S::~S post-dominate bar,
even though it really does.  We create an extra edge directly
from bar to the catch handler.

I'm close to having a small patch to fix this (10 regressions left),
which simply changes reachable_next_level

     case ERT_CLEANUP:
-      /* Before landing-pad generation, we model control flow
-        directly to the individual handlers.  In this way we can
-        see that catch handler types may shadow one another.  */
       add_reachable_handler (info, region, region);
-      return RNL_MAYBE_CAUGHT;
+      return RNL_CAUGHT;

and then properly creates edges for GIMPLE_RESX statements.

I've also been thinking a good deal about a grand EH reorg; I'll try
to post something for you to comment on later today.


r~