[Bug bootstrap/63573] [5 Regression] libgo: ICE building libgo on powerpc-linux-gnu

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63573

--- Comment #16 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to Jan Hubicka from comment #15)
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63573
> > 
> > --- Comment #14 from Martin Liška <marxin at gcc dot gnu.org> ---
> > (In reply to Andreas Schwab from comment #12)
> > > This also breaks g++.dg/ipa/pr63587-2.C on powerpc -m32, but the patches in
> > > #c6 and #c8 don't fix that.
> > > 
> > > $ gcc/xg++ -Bgcc/ ../gcc/testsuite/g++.dg/ipa/pr63587-2.C -nostdinc++
> > > -Ipowerpc64-linux/32/libstdc++-v3/include/powerpc64-linux
> > > -Ipowerpc64-linux/32/libstdc++-v3/include -I../libstdc++-v3/libsupc++
> > > -I../libstdc++-v3/include/backward -I../libstdc++-v3/testsuite/util
> > > -std=gnu++11 -O2 -S -m32 -o pr63587-2.s
> > > ../gcc/testsuite/g++.dg/ipa/pr63587-2.C: In static member function ???static
> > > int
> > > boost::function_obj_invoker0<FunctionObj>::invoke(boost::function_buffer&)
> > > [with FunctionObj =
> > > boost::test_case_template_invoker<default_formatting_invoker>]???:
> > > ../gcc/testsuite/g++.dg/ipa/pr63587-2.C:21:3: internal compiler error: in
> > > expand_expr_addr_expr_1, at expr.c:7725
> > 
> > There's really problem that both suggested patches can fix just cases where
> > DECL_INCOMING_RTL is defined. In following situation:
> > 
> > static boost::log::make_output_actor<ActorT<LeftExprT>, RightT, ValueT>::type
> > boost::log::make_output_actor<ActorT<LeftExprT>, RightT,
> > ValueT>::make(ActorT<LeftExprT>, RightT&) [with ActorT = boost::actor;
> > LeftExprT =
> > boost::log::attribute_output_terminal<boost::actor<boost::log::attribute_output_terminal<boost::actor<int>,
> > boost::log::to_log_fun> >, boost::log::to_log_fun>; RightT =
> > boost::log::attribute_actor<int, boost::log::value_extractor, void,
> > boost::actor>; ValueT = int; boost::log::make_output_actor<ActorT<LeftExprT>,
> > RightT, ValueT>::type =
> > boost::actor<boost::log::attribute_output_terminal<boost::actor<boost::log::attribute_output_terminal<boost::actor<boost::log::attribute_output_terminal<boost::actor<int>,
> > boost::log::to_log_fun> >, boost::log::to_log_fun> >, boost::log::to_log_fun>
> > >] (struct actor left, struct attribute_actor & right)
> > {
> >   struct type D.4892;
> >   struct attribute_name D.4891;
> >   struct to_log_fun D.4890;
> >   struct actor left;
> > 
> >   <bb 2>:
> >   left = left;
> 
> The reason here is that thunk expansion introduce extra temporary.  
>       if (nargs)
>         for (i = 1, arg = DECL_CHAIN (a); i < nargs; i++, arg = DECL_CHAIN
> (arg))
>           {  
>             tree tmp = arg;
>             if (!is_gimple_val (arg))
>               {
>                 tmp = create_tmp_reg (TYPE_MAIN_VARIANT
>                                       (TREE_TYPE (arg)), "arg");
>                 gimple stmt = gimple_build_assign (tmp, arg);
>                 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
>               }
>             vargs.quick_push (tmp); 
>           }
> 
> the temporary obviously is not linked with the original memory expression.

Ok, that makes sense.

> 
> I supose we may want to simply drop the tailcall flag in cases copy is
> necessary, but I do not really see why a copy is desirable here: IMO whatever
> can be RHS of assignment can also be passed to function.  What makes
> left to be !is_gimple_val here?

The loops iterates from 1..(nargs-1), where 'left' is first argument and is put
to vargs here:

      if (this_adjusting)
        vargs.quick_push (thunk_adjust (&bsi, a, 1, fixed_offset,
                    virtual_offset));
      else if (nargs)
        vargs.quick_push (a);

-----------^

      if (nargs)
        for (i = 1, arg = DECL_CHAIN (a); i < nargs; i++, arg = DECL_CHAIN
(arg))

Martin

> 
> Honza