[PATCH] Fix canonical types w/ transparent unions (4.3 regression, PR c++/35315)

[Doug Gregor]

Here's a slightly improved patch, which moves the canonical-types fix
into build_duplicate_type. I'll commit to mainline tomorrow if there
are no objections. However, I'm pinging Mark and Jakub to ask whether
this is patch should also go into 4.3. I believe the answer is "no",
but here's some reasoning.

Any failure in the canonical types system is bad; it changes the
notion of type identity in C++, which will break programs. Thankfully,
we have a fallback with the verify-canonical-types parameter. My
initial panic over this issue was due to the use of transparent_union
in a glibc header (stdlib.h, nonetheless):

#  if !defined __GNUC__ || __GNUC__ < 2 || defined __cplusplus
#   define __WAIT_STATUS        void *
#   define __WAIT_STATUS_DEFN   void *
#  else
/* This works in GCC 2.6.1 and later.  */
typedef union
  {
    union wait *__uptr;
    int *__iptr;
  } __WAIT_STATUS __attribute__ ((__transparent_union__));
#   define __WAIT_STATUS_DEFN   int *
#  endif

But now that I've noticed the "|| defined __cplusplus", I'm not quite
as worried, because C doesn't use the canonical types system. So,
transparent_unions in C++ won't work entirely correctly in 4.3.0, but
I imagine it's a little-used feature in the C++ world. If I'm wrong,
we should accept this little patch.

  - Doug

2008-02-26  Douglas Gregor  <doug.gregor@gmail.com>

	PR c++/35315
	* tree-inline.c (build_duplicate_type): When we make a
	duplicate type, make it unique in the canonical types system.

2008-02-26  Douglas Gregor  <doug.gregor@gmail.com>

	PR c++/35315
	* g++.dg/ext/attrib32.C: Add another test case.



On Mon, Feb 25, 2008 at 11:15 PM, Doug Gregor <doug.gregor@gmail.com> wrote:
> This is a simple canonical-types fix for dealing with transparent
>  union attributes properly in the canonical-types system. We just need
>  to set the TYPE_CANONICAL for the transparent_union variant of a type
>  appropriately. This fixes PR c++/35315, a canonical-types regression.
>
>  Okay for mainline? Should it also go into the 4.3 branch?
>
>   - Doug
>
>  2008-02-25  Douglas Gregor  <doug.gregor@gmail.com>
>
>         PR c++/35315
>         * c-common.c (handle_transparent_union_attribute): When we make a
>         duplicate type, make it unique in the canonical types system.
>
>  2008-02-25  Douglas Gregor  <doug.gregor@gmail.com>
>
>         PR c++/35315
>         * g++.dg/ext/transparent_union.C: New.
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: transparent-union-canon.patch
Type: application/octet-stream
Size: 747 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20080228/80bbdfc8/attachment.obj>