state.c had put file_name into the save file. The saved value was used
to overwrite file_name on restore. If the save file had been renamed,
unlink(file_name) would then fail, because file_name held the old name.
To avoid breaking savefile compatibility, file_name is still saved, but
it is read back into a temporary buffer and then ignored.
I thought I fixed this already.
In rogue5/state.c, rs_read_daemons() zeroes out the argument and delay
if the daemon slot is empty. Unfortunately that code ended up on the
wrong side of the brace that closes the for loop, so instead of running
after each daemon, it got run once after the loop exited, when the
index was of course out of bounds.
This tended to manifest, when compiled with -O2, by overwriting hw and
setting it to NULL. When inventory() next ran, hw would be passed to
wgetch(), which returns ERR when it gets a NULL argument. This made
md_readchar() think something was wrong and autosave the game.
Upon investigation, rogue3 was found to commit the same mistake.
rogue4 and srogue don't zero the data. arogue5 already does it
properly.
Someday I am going to run all this through Valgrind. Someday when I
am a kinder person who will not be driven to invoke hordes of trolls
and centaurs upon the original authors.
In all games, rs_write_room_reference() stored -1 for a nonexistent
room, but rs_read_room_reference() did not check for out-of-bounds
values, leading to pointers to rooms[-1], which sometimes caused
crashes. rs_read_room_reference() has now been modified to use NULL
instead.
Some of the games required further changes to replace NULL with the
pointer to the actual room. Others are capable of handling NULL for
objects not in any room.
