Memory allocated for a trap in getlev() wasn't being freed. There
is already one extra allocation which is supposed to get freed after
the loop, but the 'keepgoing' flag caused an extra trap allocation
before loop termination. So the unintentional one got freed but did
so by intercepting the free for the end-of-list one.
Fruit had similar code which applied to full game save and restore
rather than level save and restore so wasn't as noticeable.
phase_of_moon and friday_13th determined using rn2() instead of local
time if fuzzing. Don't reseed using init_random() if fuzzing. Allow
set_random to be called outside of hacklib. rn2_on_display_rng uses
rn2 if fuzzing so that we have a single source of random that we can
ensure is reproducible. Implement rul() that returns a random unsigned
long. Fix bug in fuzzer handling of ntposkey which would cause us to use
unitialized values for x and y. Added command line arguments to allow
auto starting and stopping of fuzzer. Add a logging facility for the
fuzzer to use to record activity. Added some scripts used to automate
fuzzer testing on windows.
A check for bad restoration (ball without chain or vice versa) issued
a warning but then left the problem around to trip up other code.
'Fix' the problem by clearing those owornmask slots and their pointers.
Doesn't try to recover memory if the one that's found is OBJ_FREE.
Also some formatting.
If a punished player picks up the iron ball, gets engulfed and
saves, then the saved game will have missed saving the dangling
chain since it was not on the floor or in the inventory. Upon
restoring the saved game, the game will be in a bad state since
the ball will be worn but the chain will be missing.
Make some progress on a couple of next minor release checklist
items, hopefully without introducing too many new bugs. This
is just the initial commit, and work continues.
Checklist items:
Savefiles compatible between Windows versions, whether 64-bit
or 32-bit in little-endian field format.
Selection of file formats:
historical (structlevel saves),
lendian (little-endian, fieldlevel saves),
and just for proof-of-concept, ascii fieldlevel saves
(the ascii is huge! 10x bigger than little-endian).
For the fieldlevel save, all complex data structures recursively
get broken down until until it is one of the simple types that
can't be broken down any further, and that gets when it gets
written to the output file.
New files needed for this build:
hand-coded:
include/sfprocs.h
src/sfbase.c - really a dispatcher to one of the
output/input format routines.
src/sflendian.c - little-endian output writer/reader.
src/sfascii.c - ascii text output writer/reader.
auto-coded (generated):
include/sfproto.h
src/sfdata.c
This is just one approach. I'm sure there are countless others
and they have different pros and cons.
For producing the auto-coded files a utility called
universal-ctags, that is actively maintained and evolving,
was used to do all the heavy-lifting of parsing the
NetHack C sources to tabulate the data fields, and store
them in an intermediate file called util/nethack.tags
(not required for building NetHack if you already have a
generated include/sfproto.h and src/sfdata.c)
util/readtags (also not required for building NetHack
itself) will decipher the nethack.tags file and produce
the functions that can deal with the NetHack struct data
fields.
You can obtain the source for universal-ctags by cloning it
from here:
https://github.com/universal-ctags/ctags.git
The combination universal-ctags + util/readtags has been
tried and tested under both Windows and Linux, so it is
not tied to a particular platform.
Note: util/readtags will work only with universal-ctags
output, so other ctags are unlikely to work as-is.
Universal-ctags can be build from source very easily
under Linux, or under Windows using visual studio.
Hero polymorphed into a vampire or v.lord can use #monster to switch
to vampire bat or fog cloud [or wolf for lord] but it was a one shot
polymorph. Remember when current form is a shape-shifted vampire and
allow #monster in shifted form to pick another shifted form or the
vampire form.
Genocide of the alternate shape forces back to base vampire. Genocide
of base vampire does too, then reverts to human (or dwarf, &c) as
vampires go away. Being killed while shafe-shifted reverts all the
way to human rather than to vampire. [Just realized: interaction
with Unchanging wasn't taken into consideration so hasn't been tested.]
Since 'youmonst' isn't saved and restored, I had to add a field to 'u'
to hold youmonst.cham during save/restore.
Tested with 3.6.2+ and seemed to be working (except saving while
shape-shifted restored as ordinary bat/cloud/wolf because new u.mcham
wasn't there to hold youmonst.cham yet). Builds with 3.7.0- but not
execution tested yet (I didn't want to clobber my current playground).
During shop repair, give a message about the shopkeeper using a spell
(if hero is close enough) before "Suddenly, <various repairs occur>."
And when shop repair is for a single untrap of landmine or bear trap
adjacent to shk (and the hero can see it happen), say "<Shk> untraps
<trap>" rather than just "Suddenly, a trap is removed from the floor!"
When deciding whether to discard interrupted lock/unlock context while
changing levels, maybe_reset_pick() checks whether xlock.box is being
carried. But it was doing so after the old level had been saved and
memory for non-carried container there had been freed.
That led to a couple of other issues. context.travelcc was using -1
for 'no cached value', but the fields of travelcc have type 'xchar' and
shouldn't be given negative values. 0 should be fine for 'no cache'.
Failed partial restore which occurred after old game's context had been
loaded would begin a new game with old game's stale context. Restoring
goes out of its way to avoid that for 'flags' but didn't for 'context'.
If hero was carrying Schroedinger's Box at end of game, disclosing
inventory converted it into an ordinary box. That interferred with
subsequent disclosure when writing DUMPLOG, which saw an empty box
if inventory had been shown or the special box with newly-determined
contents if not. I tried a couple of ways to fix it and decided
that redoing it was better in the long run.
Schroedinger's box is still flagged with box->spe = 1, but instead
of having that affect the box's weight, now there is always a cat
corpse in the box. When opened, that will already be in place for
a dead cat or be discarded for a live one, but the weight will be
standard for container+contents and when box->cknown is set it will
always be "containing 1 item" (which might turn out to be a monster).
Some temporary code fixes up old save/bones files to stay compatible.
TODO: food detection used to skip Schroedinger's Box; now it will
always find a corpse, so some fixup like the ridiculous probing code
is needed.
Bug report #H7156 listed three items, all relating to perm_invent:
1) it shouldn't persist across save/restore since restore might be
on a system which doesn't have enough room to display it (report
actually complained that config file setting was ignored when
restoring old games, which is an expected side-effect for options
that persist across save/restore);
2) permanent inventory wasn't updated when using scroll of charging;
3) attempts to update permanent inventory during restore could lead
to crash if it tries to access shop cost for unpaid items.
Items (2) and (3) have already been fixed. This fixes (1).
Replace 'flags.perm_invent' with a dummy flag, preserving save files
while removing it from flags. Add 'iflags.perm_invent' to hold the
value of the perm_invent option.
The win32 files that are updated here haven't been tested. Whichever
branch contains the curses interface needs to be updated; ditto for
any other pending/potential interfaces which support perm_invent.
Redo how updates of permanent inventory window are suppressed during
restore. Reverses part of e9f1e03271
which included a simpler attempt to deal with this.
It looks like we should have been getting impossible "unpaid_cost:
object wasn't on any bill" but segfault was reported; I haven't tried
to figure out why. The band in xname() ought to be redundant now but
is included for bulletproofing.
