Re: [debug-early] C++: emit early debug info for all globals, not just statics

[Aldy Hernandez <aldyh at redhat dot com>]

On 10/02/14 08:46, Jason Merrill wrote:
> On 10/01/2014 05:31 PM, Aldy Hernandez wrote:
> > +  for (tree t = level->names; t; t = TREE_CHAIN(t))
> > +    if (TREE_CODE (t) != TYPE_DECL
> > +    && TREE_CODE (t) != PARM_DECL
> > +    && !DECL_IS_BUILTIN (t))
> > +      debug_hooks->early_global_decl (t);
>
> What does this do for templates?  I think we don't want to send a
> template off to early_global_decl, but rather walk through its
> specializations and emit them.

Hmm, I'll look into this.

> Why do you need to check for PARM_DECL?  There shouldn't be any
> PARM_DECL at namespace scope.
>
> Why do you want to skip TYPE_DECL?  I think we should leave the decision
> about whether to emit a typedef to the debugging back end.

Actually, I think we/I need to rethink this whole globals thing. 
Currently we're early dumping global *_DECLs, and hoping dwarf2out 
recursion will also pick up types and any derivatives from the *_DECLs, 
but taking a closer look I've noticed a lot of things are not being 
dumped early.

For instance:


	foo()
	{
	  typedef int ITYPE;
	  ITYPE var = 5;
	}

For the above code, var's DIE gets outputted _after_ the compilation 
proper has been run here:

  if (! declaration && outer_scope && TREE_CODE (outer_scope) != ERROR_MARK
      && (!DECL_STRUCT_FUNCTION (decl)
	  || DECL_STRUCT_FUNCTION (decl)->gimple_df))
	...
	...
	decls_for_scope (outer_scope, subr_die, 0);

I think we should be outputting DIEs for locals at the end of parsing, 
so my patch going through level->names IMO is wrong.  We should be 
dumping all *_DECLs created by the FE, not just globally scoped ones.

This means that we'll need to clean up unreachable/unused DIEs after the 
flow graph has been built.

Another example currently not being dumped early is...

	function()
	{
	  class Local {
	  public:
	    void loc_foo (void) { }
	  };
	
	  Local l;
	  l.loc_foo ();
	}

...since loc_foo() is not in level->names.  Again, this seems like an 
argument for early dumping all *_DECLs directly from 
rest_of_decl_compilation() (as you've hinted) and then cleaning up 
unused DIEs after we have flow information.

Does this seem reasonable?

Aldy