[C] Always use TRANSLATION_UNIT_DECL as context

On Mon, 27 Sep 2010, Jie Zhang wrote:

> On 09/27/2010 04:49 PM, Richard Guenther wrote:
> > On Sat, 25 Sep 2010, Jie Zhang wrote:
> > 
> > > On 09/17/2010 08:18 PM, Richard Guenther wrote:
> > > > 	* c-decl.c (pop_scope): Always set file-scope DECL_CONTEXT.
> > > 
> > > This change causes regressions on arm-none-eabi:
> > > 
> > > FAIL: gcc.c-torture/execute/builtins/abs-1.c compilation,  -O2 -flto
> > > UNRESOLVED: gcc.c-torture/execute/builtins/abs-1.c execution,  -O2 -flto
> > > 
> > > The compiler error messages are:
> > > 
> > > testsuite/gcc.c-torture/execute/builtins/abs-1.c:22:1: sorry,
> > > unimplemented:
> > > gimple bytecode streams do not support machine specific builtin functions
> > > on
> > > this target
> > > testsuite/gcc.c-torture/execute/builtins/abs-1-lib.c:22:1: sorry,
> > > unimplemented: gimple bytecode streams do not support machine specific
> > > builtin
> > > functions on this target
> > > testsuite/gcc.c-torture/execute/builtins/lib/main.c:13:1: sorry,
> > > unimplemented: gimple bytecode streams do not support machine specific
> > > builtin
> > > functions on this target
> > > compiler exited with status 1
> > > 
> > > I did some investigation, but don't know how to fix it.
> > > 
> > > The above change makes GCC always set file-scope DECL_CONTEXT. So when
> > > come to
> > > 
> > > static void
> > > lto_output_ts_decl_minimal_tree_pointers (struct output_block *ob, tree
> > > expr,
> > > 					  bool ref_p)
> > > {
> > >    lto_output_tree_or_ref (ob, DECL_NAME (expr), ref_p);
> > >    lto_output_tree_or_ref (ob, DECL_CONTEXT (expr), ref_p);
> > >    lto_output_location (ob, DECL_SOURCE_LOCATION (expr));
> > > }
> > > 
> > > with EXPR being a file scope function, DECL_CONEXT (expr) is the
> > > translation
> > > unit which contains the function. Then
> > > 
> > >    lto_output_tree_or_ref (ob, DECL_CONTEXT (expr), ref_p);
> > > 
> > > writes out the whole translation unit including all the builtin functions.
> > > Finally, in lto_output_builtin_tree, a "sorry" is called with the above
> > > error
> > > message since ARM target does not have targetm.builtin_decl.
> > > 
> > > I don't know how to fix this. Defining targetm.builtin_decl for ARM should
> > > fix
> > > it. But I'm not sure if we really want to output all builtins for each
> > > translation unit. This regression should exist for all targets which don't
> > > provide targetm.builtin_decl.
> > 
> > Yes, this writes all declarations in the TRANSLATION_UNIT_DECLs BLOCK
> > tree.  Does the following fix it?
> > 
> > Index: tree.c
> > ===================================================================
> > --- tree.c      (revision 164590)
> > +++ tree.c      (working copy)
> > @@ -4585,6 +4690,8 @@ free_lang_data_in_decl (tree decl)
> >           nesting beyond this point. */
> >         DECL_CONTEXT (decl) = NULL_TREE;
> >       }
> > +  else if (TREE_CODE (decl) == TRANSLATION_UNIT_DECL)
> > +    DECL_INITIAL (decl) = NULL_TREE;
> >   }
> > 
> Not yet. This change causes an ICE.

Ick ...

so I guess we should instead try to not push builtins to the global
scope.  Like with

Joseph might have an idea here (I also see external_scope being used,
but I'm not sure what functions actually end up in that).

Does the above help?

Thanks,
Richard.

[C] Always use TRANSLATION_UNIT_DECL as context

Commit Message

Comments

Patch