[PATCH] Don't clobber dominator info in the combiner (PR target/77526)

[Jakub Jelinek]

On Fri, Sep 16, 2016 at 08:47:28AM +0200, Richard Biener wrote:
> On Fri, 16 Sep 2016, Jakub Jelinek wrote:
> 
> > Hi!
> > 
> > As mentioned in the PR, combiner sometimes calls
> > purge_all_dead_edges or purge_dead_edges that can invalidate the dominator
> > info if it is computed.  Other passes like CSE in that case free the
> > dominance info, this patch does the same in the combiner.
> > 
> > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
> 
> Ok.  (I wonder if/why cleanup_cfg (0) doesn't do this then?)

cleanup_cfg only frees it in some cases if it does some kind of changes:
  /* ???  We probably do this way too often.  */
  if (current_loops
      && (changed
          || (mode & CLEANUP_CFG_CHANGED)))
    {
...
      calculate_dominance_info (CDI_DOMINATORS);
But here the dominator info gets out of sync earlier, in particular on this
testcase in the
  new_direct_jump_p |= purge_all_dead_edges ();
call.

Perhaps ideally we should teach all these functions to update the dominator
info, but I'm afraid it is a lot of work.

	Jakub