[Bug middle-end/52355] [4.7 regression] address difference between array elements is not considered to be a compile time constant anymore

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52355

--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-24 
09:47:33 UTC ---
We fold end up trying to fold
(long int) &(*(a + 256))[0][0] - (long int) &(*a)[0][0] because a is of
type char[16][16] *.

I have a patch.