Also on ia64. 
 
Broken by this change:   
   
2005-09-24  Richard Henderson  <rth@redhat.com>  
  
	* c-common.c (handle_mode_attribute): When not modifying in place,  
	create subtypes for enumerations.  
	(sync_resolve_return): Use TYPE_MAIN_VARIANT.  
	* gimplify.c (create_tmp_from_val): Likewise.  
  

This looks like a latent bug in the Ada front-end as the only change to the middle-end would be the 
change to use TYPE_MAIN_VARIANT.

Indeed, just reverting just the gimplify.c change gets Ada in stage3 (still
running on my machine).

By Andreas.

This works at -O on x86.

> This looks like a latent bug in the Ada front-end as the only change to the
> middle-end would be the change to use TYPE_MAIN_VARIANT.

No, it isn't, please investigate a minimum before writing such a statement. 
It's a latent consistency bug in the integer-share-limit stuff.

In build_int_cst_wide, we're asserting that the type of shared constants
attached to a type is precisely that type:

      t = TREE_VEC_ELT (TYPE_CACHED_VALUES (type), ix);
      if (t)
	{
	  /* Make sure no one is clobbering the shared constant.  */
	  gcc_assert (TREE_TYPE (t) == type);

But in set_sizetype we're boldly copying TYPE_CACHED_VALUES between types:

  /* We do want to use bitsizetype's cache, as we will be replacing that
     type.  */
  TYPE_CACHED_VALUES (t) = TYPE_CACHED_VALUES (bitsizetype);
  TYPE_CACHED_VALUES_P (t) = TYPE_CACHED_VALUES_P (bitsizetype);


Nathan, who is right here?

Subject: Re:  [4.1 Regression] Ada bootstrap ICE in
 build_int_cst_wide, at tree.c:795

ebotcazou at gcc dot gnu dot org wrote:

> In build_int_cst_wide, we're asserting that the type of shared constants
> attached to a type is precisely that type:
> 
>       t = TREE_VEC_ELT (TYPE_CACHED_VALUES (type), ix);
>       if (t)
> 	{
> 	  /* Make sure no one is clobbering the shared constant.  */
> 	  gcc_assert (TREE_TYPE (t) == type);
> 
> But in set_sizetype we're boldly copying TYPE_CACHED_VALUES between types:
> 
>   /* We do want to use bitsizetype's cache, as we will be replacing that
>      type.  */
>   TYPE_CACHED_VALUES (t) = TYPE_CACHED_VALUES (bitsizetype);
>   TYPE_CACHED_VALUES_P (t) = TYPE_CACHED_VALUES_P (bitsizetype);
> 
> 
> Nathan, who is right here?

Both are correct.  However it looks like some integer constant created before 
set_sizetype is called has not got its type updated correctly.

The problem here is that we have to create some types before setting size_type, 
and in creating those types we need to set their size to some integral constant, 
and that integral constant has to have a type.  That uses the stub sizetype 
type, which we overwrite further down set_sizetype.

   /* Replace our original stub sizetype.  */
   memcpy (sizetype, t, tree_size (sizetype));
   TYPE_MAIN_VARIANT (sizetype) = sizetype;

My guess is that we've already created some type variants of the stub type at 
that point, which is not what I expected.  There should really be an
	gcc_assert (!TYPE_NEXT_VARIANT (sizetype) && !TYPE_NEXT_VARIANT (t));
just before the memcpy

nathan

I'm testing Kenner's suggestion from
http://gcc.gnu.org/ml/gcc/2005-09/msg00867.html ...

> The problem here is that we have to create some types before setting
> size_type, and in creating those types we need to set their size to some
> integral constant, and that integral constant has to have a type.  That uses
> the stub sizetype type, which we overwrite further down set_sizetype.
> 
>    /* Replace our original stub sizetype.  */
>    memcpy (sizetype, t, tree_size (sizetype));
>    TYPE_MAIN_VARIANT (sizetype) = sizetype;

OK, I totally missed that the types 't' are not supposed to escape from
set_sizetype, right?

> My guess is that we've already created some type variants of the stub type at 
> that point, which is not what I expected.  There should really be an
> 	gcc_assert (!TYPE_NEXT_VARIANT (sizetype) && !TYPE_NEXT_VARIANT (t));
> just before the memcpy

The bug appears to be that the second type 't' does escape from set_sizetype
because it is the main variant of bitsizetype, as Richard Kenner has found.

Thanks for your feedback.

(In reply to comment #8)
> I'm testing Kenner's suggestion from
> http://gcc.gnu.org/ml/gcc/2005-09/msg00867.html ...

Works for me on i686-linux
Bootstrap finished and no additional ACATS failures.

Subject: Bug 24053

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	ebotcazou@gcc.gnu.org	2005-09-29 22:11:14

Modified files:
	gcc            : ChangeLog stor-layout.c 

Log message:
	PR middle-end/24053
	* stor-layout.c (set_sizetype): Set TYPE_MAIN_VARIANT of bitsizetype.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.10052&r2=2.10053
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/stor-layout.c.diff?cvsroot=gcc&r1=1.242&r2=1.243

Fix installed on behalf of Richard.