Re: [PATCH]: Trying to make list nodes smaller, but the C++ FE keepsbeating me down

[Mark Mitchell <mark at codesourcery dot com>]

Daniel Berlin wrote:

> So i added TREE_LIST_WT (WT = With type) for the C++ FE, and was
> converting all the users over, but it appears there is magic in here
> that i don't understand that is checking in various places.  I don't
> understand where and when TREE_LIST is supposed to appear in the C++ FE,
> and when we expect TREE_TYPE to be valid in it. 

One place I suspect you didn't find is BV_VCALL_INDEX.  That's a great 
example of something that shouldn't be a tree.  You could find the uses 
of this macro and replace them with a little struct:

  struct binfo_virtual {
    tree fn;
    tree delta;
    tree vcall_index;
    struct binfo_virtual *next;
  };

Test that, get that checked in, and then try your patch again.  I'd also 
search for all uses of build_list/tree_cons/TREE_PURPOSE/TREE_VALUE and 
try to find places near there that set TREE_TYPE.  Aha, here's one in 
pt.c:instantiate_class_template:

	  base_list = tree_cons (access, base, base_list);
	  if (BINFO_VIRTUAL_P (pbase_binfo))
	    TREE_TYPE (base_list) = integer_type_node;

OK, that's another place you need a new data structure.  Look at 
xref_basetypes to see how its used.  Now replace this base_list thing 
with a VEC of:

  struct base_descriptor {
    tree base_class;
    tree access; /* For brownie points, use access_kind instead!  */
    bool is_virtual;
  };

And so forth.

Or, if you want to use a really big (and somewhat ugly) hammer, you 
could try just making every single use of tree_cons/build_list/etc. in 
the C++ front end use tree_cons_wt (after making sure that TREE_LIST_WT 
has a TREE_LIST as a prefix), changing all code in the compiler 
(including middle-end) that handles TREE_LIST to handle both types, and 
then trying that.  It still might not work, if the C++ front end depends 
on language-independent code giving it back a TREE_LIST with room for a 
TREE_TYPE, but at least you'd have a better shot.  Then, gradually back 
out handling of TREE_LIST_WT to the C++ front end.  Now, I'm not sure if 
such a patch would be acceptable, but it would be a working methodology.

--
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304