[Bug middle-end/51895] [4.7 Regression] ICE in simplify_subreg

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51895

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-01-19
   Target Milestone|---                         |4.7.0
     Ever Confirmed|0                           |1

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-01-19 10:06:38 UTC ---
(In reply to comment #2)
> This starts with eipa_sra.  It changes a S argument (which has TImode
> TYPE_MODE) into char [9] (with BLKmode)) and then on both caller and callee
> side we have on one side a BLKmode type and on the other side a BLKmode MEM_REF
> with pointer to TImode on the second MEM_REF operand.
> I wonder why it does this, instead of just using type S, and if it really has
> to for some reason, why it can't at least make sure it has the same TYPE_MODE.
> Changing a TImode argument to a BLKmode argument doesn't look at least like a
> good optimization.
> 
> Or the bug is in the MEM_REF expansion, which expands a BLKmode MEM_REF into a
> TImode reg:
>                 bftype = TREE_TYPE (base);
>                 if (TYPE_MODE (TREE_TYPE (exp)) != BLKmode)
>                   bftype = TREE_TYPE (exp);
>                 return expand_expr (build3 (BIT_FIELD_REF, bftype,
>                                             base,
>                                             TYPE_SIZE (TREE_TYPE (exp)),
>                                             bit_offset),
>                                     target, tmode, modifier);
> base here is TImode (x PARM_DECL), but exp is BLKmode, so this returns a TImode
> pseudo.  Shouldn't it store it into a BLKmode temporary and return that MEM
> instead?

Using a BIT_FIELD_REF looked most convenient.  Using extract_bit_field
may also be an option (which I suppose is what the above ends up doing?)