Sylvain, I did some more investigation. SM> Now I just try to understand how SM> unsigned char "type" SM> can work on other platforms. SM> SM> with unsigned char "type": SM> Rune->cursor_type offset 16 SM> Rune->type offset 17 SM> Rune->object offset 20 SM> sizeof(Rune) = 44 SM> SM> My guess is: SM> SM> * (int*) (&Rune->type) SM> SM> in lispdesc_indirect_count_1 SM> works on other platforms because all of these: SM> SM> - bytes 18-19 of rune are zero. Thats correct, memory is zeroed at allocation time. SM> - first byte of "object" member must also be zero (this I dont get). No it is not. And this leads to an uncorrent count value! The data looks like this: offset 16: 1 (the value as it is meant) offset 17: 0 offset 18: 0 offset 19: 10 (this is the first byte of "object") In binary this is: (gdb) x/t irdata 0x8b63885: 00001010000000000000000000000001 Altogether on a little endian machine leads to an int value (offset 16--19): * (int*) (&Rune->type) ==> 167772161 Later in lispdesc_process_xd_union this wrong count value randomly leads to the correct result: no variant of the XD_UNION is matched, so nothing is marked. In this case this is correct, since variant 1 only contains a char, so this variant has not to be processed. This should break for variant 2, which contains Lisp_Objects that have to be marked. Those wouldn't get marked then (at least not through this branch of the mark tree). Anyways, I couldn't produce this situation, maybe they get marked by another branch. So this really is a bug on _all_ systems, not only on sparc. I don't know why this causes a crash on sparc. Maybe it has problems with the not aligned integer cast. But this is just a guess. Anyways, it is good that it did! SM> - system has to be little endian? (sparc is big endian) SM> SM> PS: with unsigned int type, "type" gets aligned: SM> Rune->cursor_type offset 16 SM> Rune->type offset 20 SM> Rune->object offset 24 SM> sizeof(Rune) = 48 Well, I think this is the tradeoff we have to make here. The struct rune is going to be 4 bytes bigger than before. Yesterday I forgot to add a changelog entry to my patch, so here it is again, this time complete: src/ChangeLog addition: 2005-03-02 Marcus Crestani <crestani(a)informatik.uni-tuebingen.de&gt; * redisplay.h (struct rune): make type field an int, to fix KKCC marking of unions. xemacs-21.5 source patch: Diff command: cvs -q diff -u Files affected: src/redisplay.h Index: src/redisplay.h =================================================================== RCS file: /pack/xemacscvs/XEmacs/xemacs/src/redisplay.h,v retrieving revision 1.22 diff -u -r1.22 redisplay.h --- src/redisplay.h 26 Jan 2005 10:33:42 -0000 1.22 +++ src/redisplay.h 3 Mar 2005 00:46:19 -0000 ＠＠ -144,7 +144,7 ＠＠ unsigned char cursor_type; /* is this rune covered by the cursor? */ - unsigned char type; /* type of rune object */ + unsigned int type; /* type of rune object */ /* We used to do bitfields here, but if I (JV) count correctly that doesn't matter for the size of the structure. All the bit -- Marcus P.S.: Sylvain, please keep the discussion on the list.