[Bug ipa/60911] [4.9/4.10 Regression] wrong code with -O2 -flto -fipa-pta

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60911

--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 23 Apr 2014, jamborm at gcc dot gnu.org wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60911
> 
> --- Comment #3 from Martin Jambor <jamborm at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #2)
> > Late IPA passes need all regular IPA transforms applied and clones
> > materialized.
> 
> I agree that would make sense but it is not what we do.  Just look at
> compile() in cgraphunit.c.  When late IPA passes run, clones are
> already materialized but transformation phases of regular IPA passes
> have not been run yet.  Apart from the fact that late-IPA passes do
> not see the bodies transformed and inlining performed it also means
> that call statements have not been updated to reflect call
> redirections performed at analysis stage which is the problem here and
> which is (for reasons I have never quite grasped) done as a part of
> inlining transformation phase.

Hmm, ok.  But that makes "simple" IPA passes work on a limbo state.
I think we delay applying IPA transforms as "optimization" with
respect to compile-time/memory-use as in theory we only require
a single function body in memory during LTRANS.  I wonder why
it works without LTO ... because there I _do_ see the IPA
transform applied to main ().

> OTOH, I am all for moving the call to execute_all_ipa_transforms()
> from expand_function() to compile(), at least in trunk.

We also have, in execute_one_pass,

  /* SIPLE IPA passes do not handle callgraphs with IPA transforms in it.
     Apply all trnasforms first.  */
  if (pass->type == SIMPLE_IPA_PASS)
    {
      bool applied = false;
      do_per_function (apply_ipa_transforms, (void *)&applied);
      if (applied)
        symtab_remove_unreachable_nodes (true, dump_file);
      /* Restore current_pass.  */
      current_pass = pass;

but that doesn't seem to work ... at least with LTO?  Which is
because for main()

      FOR_EACH_DEFINED_FUNCTION (node)
        if (node->analyzed && gimple_has_body_p (node->decl)
            && (!node->clone_of || node->decl != node->clone_of->decl))
          {

this doesn't trigger (for f.constprop it does).
gimple_has_body_p (node->decl) is false.

main/1 (main) @0x7ffff6daf148
  Type: function definition analyzed
  Visibility: externally_visible prevailing_def public
  References: 
  Referring: 
  Read from file: /tmp/ccED1aQr.ltrans0.o
  Availability: available
  First run: 0
  Function flags: only_called_at_startup
  Called by: 
  Calls: abort/2 (0.00 per call) f.constprop.0/3 (1.00 per call) 

it's probably not yet streamed in?