------- Additional Comments From tobi at gcc dot gnu dot org 2005-07-10 16:58 ------- I had an IRC conversation with Paul Brook about this. Here's what we concluded:
<pbrook> There are two solutions. (a) Align all components of derived types to component size boundaries. (b) Express pointer strides in bytes, not elements. <tobi_s> yes, i wanted to evade the second option ... <pbrook> Right. The alignment of a derived type is the maximum alignment of all of its components. However the alignment of one component does not effect the alignment of other components within the same parent. <pbrook> so type t; integer(2) :: a; integer(1) :: b, c; end type; would still be 4 bytes. <tobi_s> so either we use byte-sized strides or we rule out packed structures <pbrook> Yep, pretty much. ... <pbrook> tobi_s: byte strides might not be so bad once we build multi-dimensional arrays properly. ... <pbrook> Anyhow, I think the solution is to make p => v(:)%c ICE/Sorry, and postpone it until we get round to using ARRAY_REFS properly for multi-d arrays. IOW, fixing the underlying bug is out-of-reach. The original bug can be fixed by forcing the function result via a temporary as Paul's patch does. I'll have to think about the correct check, though. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18022
