[Bug middle-end/55142] [4.8 Regression] internal compiler error: in plus_constant, at explow.c:88

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55142

--- Comment #27 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2012-11-08 17:17:35 UTC ---
> No, this would be one giant kludge by itself. The failure just shows that the
> controversial patch [1] for PR 49721 was wrong.
> 
> Quote from [1]:
> 
> --quote--
> I am checking in this patch, which only affects x32
> and nothing else.  This one character change, from
> 
> POINTERS_EXTEND_UNSIGNED < 0
> 
> to
> 
> POINTERS_EXTEND_UNSIGNED != 0
> 
> creates a working x32 GCC. This isn't perfect. I have
> tried many different approaches without any success.
> I will revisit it if we run into any problems with x32
> applications.
> --/qoute--
> 
> So, we run into problem.

It's not totally wrong, given the context of convert_memory_address_addr_space
which is already optimistically correct only.  The problem is that the case
POINTERS_EXTEND_UNSIGNED > 0 is trickier than POINTERS_EXTEND_UNSIGNED == 0
because RTL constants are sign-extended: in the latter case, everything is
sign-extended so this is symmetric and simple; in the former case, one part is
zero-extended and the other part sign-extended and, in order to make this work
under the same hypothesis of non-overflow, one would need to know which part is
bigger.  In the case at hand, the code would be correct if the constant was
zero-extended and the register sign-extended, not the reverse as currently.