Reported internally, if a prayer resulted in 'fix all troubles' and
one of those was TROUBLE_STUCK_IN_WALL but safe_teleds() couldn't find
any place to relocate the hero to, nothing was done and STUCK_IN_WALL
would be found again as the next trouble to fix. Since safe_teleds()
eventually resorts to trying every single spot on the map, there was
no other result possible than failing to find an available spot again,
nothing would be done, and next trouble would be STUCK_IN_WALL, ad
naseum.
I started out with a fix that looked for secret corridors to expose
and doors to open, to make more space available, then try to move a
monster off the level, then try digging out rock and/or walls and
smashing boulders. None of those guarantee success and I got bogged
down by the digging case. This was going to be a last resort if all
of those still failed to make somewhere to move the hero, but for now,
at least, I'm skipping all that other stuff and going directly to the
last resort: give the hero Passes_walls ability for a short time, and
let him or her find own way out of trouble. The next trouble to fix
won't be STUCK_IN_WALL because Passes_walls makes that a non-issue.
I'm not thrilled with the new messages involved but want to get this
behind me.
Noticed when I was looking at float_up()/float_down() vs being trapped.
(I have a substantial patch for that but it involves infrastructure
changes so will have to go into master instead of the 3.6.0 branch.)
buried_ball() was using nested loops as a very convoluted way to test
if (otmp->ox >= cc->x - 2 && otmp->ox <= cc->x + 2
&& otmp->oy >= cc->y - 2 && otmp->oy <= cc->y + 2)
I think this revised version is closer to what was intended.
There are issues. buried_ball() finds the buried iron ball nearest to
the specified location--which is always the hero's current location--
rather than the one which was 'uball' before being buried. A player
can spot that since iron balls aren't necessarily identical. Also, it
searches within a radius of two steps but a tethered hero is only
allowed to move one step away from buried ball, so something is off.
It turns out that Mjollnir before the recent aklys enhancement and
aklys since then would end up both wielded and quivered if it returned
to thrower's hand while quiver was empty. Remove it from quiver before
restoring it to hand.
Limit the throwing range of an aklys if it is expected to return (ie,
thrown while wielded as primary) since it is a "thonged club" which is
supposedly attached by a cord to the thrower's wrist or hand. I have
no idea how long a real cord like that might be. I set the range limit
to BOLT_LIM/2. Shorter makes throw-and-return uninteresting for an
item that needs to be wielded to throw--so would probably be swapped
back and forth with a stronger melee weapon or longer range missile
weapon--and longer seems absurd.
Changes to be committed:
modified: doc/fixes36.1
modified: include/unixconf.h
modified: sys/share/ioctl.c
github pull request #19 made reference to resulting code behaviour
being better when windows were resized when USE_WIN_IOCTL was defined.
The logic for including the necessary enabling code in the build in
sys/share/ioctl.c was an explicit "opt-in" strategy, so anything not
deliberately and explicitly listed was not able to take advantae
of the potentially useful code. The need to add #defines to that
list would have been perpetual as new platforms came online, and
unnecessarily restrictive for everything else.
This switches the logic to include the code by default now,
and thus
unless there is an explicit "opt-out" by uncommenting
AVOID_WIN_IOCTL in include/unixconf.h
Some platforms, and we have no way of knowing which ones, may have
to ensure that AVOID_WIN_IOCTL is #define'd.
I worked on this a while back but didn't commit it because I couldn't
figure out the dead monster which wouldn't die. Pasi beat me to that.
Clean up the Cleaver code some and make the three-target swing alternate
between counter-clockwise and clockwise to simulate normal right-to-left
swing followed by left-to-right backswing. (Alternation happens on each
swing regardless of whether it is consecutive with the most recent one.
That's suboptimal but easy....) Also, stop three-target attack early if
hero is life-saved after being killed by a passive counter-attack from
the first or second target.
Prevent rogue backstab and two-hander breaking foes' weapons when using
Cleaver hand-to-hand because getting more than one of either of those
bonuses at a time would be excessive. I think allowing those for the
primary target but not for the two adjacent ones would be better, but I
just thought of that and am not going back for another round of revising.
This doesn't incorporate the two pull requests: one to avoid hitting
tame or peaceful adjacent targets unless the primary was also tame or
peaceful, the other to avoid hitting unseen adjacent targets. I'm not
sure if that includes remembered-unseen 'I' on the spot or no monster
shown at all. (There's a third one about updating the map but it isn't
needed for the existing Cleaver code either before or after this patch.)
I'm in favor of the first one and am not sure about the second. My
original concern was that someone could use Cleaver to find/hit three
unknown monsters at a time via 'F' prefix, but forcefight aimed at
thin air doesn't reach the Cleaver code so that can't happen, nor can
attacking two known close monsters at once by targetting an empty spot
between them.
Due to inverted logic, hitting a blue jelly with the Cleaver could cause
a dmonsfree warning - the jelly would die, but then come back to life
via the passive attack. This is a post-3.6.0 bug
The comment I added to data.base--to explain that despite what the
short description says, an aklys wouldn't return to sender after being
thrown--was bugging me, so I've made aklyses behave like Mjollnir.
If thrown when wielded as primary weapon, it will usually return and
usually be re-wielded. I also added a new message when either thrown
Mjollnir or thrown aklys is expected to return and fails to do so.
Most of the diff to dothrow.c is a change in indentation level. The
amount of code needed was quite small.
Autopickup for thrown Mjollnir which had failed to return was putting
it into the quiver slot if that was empty. Note quite an outright bug,
but it started wielded and can't be thrown if quivered, so exclude it
from the stuff that will auto-fill the quiver slot when added to invent
(post-3.6.0 issue).
Reported suggested that "congratulations" was too modern for use by
your god during ascension and that the longer phrases which were
shortened to yield it were too modest for the god to use. Make the
suggested change.
Report indicated that looking up "more info?" for "kittens" would show
the data.base entry for "kitten" and then when the display window was
dismissed, another "--More--" prompt for a empty second display window
would occur. Looking up "2 arrows" from a closely-seen object on the
map behaved similarly. User correctly diagnosed that the two-pass
lookup left the 'found_in_file' flag set from the first pass during
the second but that just clearing that resulted in "I have no info
about such things" on the second pass. The code is on the convoluted
side and needed an extra flag to handle 'seen on first pass' in
addition to clearing found flag after the first pass. I also added a
check to skip the second display if primary and secondary keys don't
match each other but both find the same entry. To test it, I tried
"a +0 aklys named aclys". That found the aclys entry but failed to
find "+0 aklys", so I added another change to have the key massaging
remove +/-N after removing "a", "an", or "the".
If "Want to see more info \"" + lookup string + "\"?" was too long,
the prompt buffer passed to yn() was being left uninitialized. Also,
test for too long was based on BUFSZ but yn() complains (to paniclog)
if the prompt is longer than QBUFSZ. Make checkfile() construct a
truncated prompt if the lookup string is too long.
I untangled some spaghetti by making all the 'goto's be forward. It
didn't help a lot but did simplify a few early returns by having them
jump to a common exit instead of replicating the file close.
installing the hooks first
NHgithook.pm: add some warnings if nhversioning can't open files
make sure nhversioning fails before opening gitinfo.txt if it can't get valid
data
Incorporate some git information into NetHack so that it
is potentially visible to a player. That's useful when
collecting details about the version that they are
running and, if the gitinfo is present, it can tie the
code to a specific git commit in the repository.
This modifies 'makedefs -v' to check for the presence of a data file
called dat/gitinfo.txt and if it is there, parse out its
contents, then write additional lines to include/date.h beyond
what 'makedefs -v' was previously putting in there, similar to
this sample:
#define NETHACK_GIT_SHA "0c84e564c78e2024e562d39539376ce2e21eec8e"
#define NETHACK_GIT_BRANCH "NetHack-3.6.0"
The contents of an appropriate dat/gitinfo.txt are as follows,
and trailing/leading whitespace is not significant:
githash = 0c84e564c78e2024e562d39539376ce2e21eec8e
gitbranch = NetHack-3.6.0
It also adjusts the contents of the 'v' version information to
include the additional git info when available.
Also adds some hooks DEVEL/hooksdir and a perl file to DEVEL
for simplifying and automating the deposit of dat/gitinfo.txt
so that it generally reflects the most current git commit.
DEVEL/gitinfo.pl can be used to build dat/gitinfo.txt at any
time without doing a commit, merge, or checkout.
perl DEVEL/gitinfo.pl
command line --version and -version support
To complement the extra information being provided in the
version by the 'v' command, this also adds support for the
following new command line arguments:
--version
-version Output the NetHack version string then exit.
--version:paste Output the NetHack version string and also copy it to
-version:paste the platform's paste buffer for insertion somewhere,
then exit.
If the paste variation of -version is requested on a platform that
hasn't incorporated any support for the capability, it will deliver
the version info then an error message, prior to exiting.
To support the extended -version:paste variation, a port needs to:
- provide a port-specific routine to perform
the paste buffer copy in a port code file.
- #define RUNTIME_PASTEBUF_SUPPORT in the include/portconf.h header file.
--skeleton--
void port_insert_pastebuf(buf)
char *buf;
{
/* insert code to copy the version info from buf into
platform's paste buffer in a supported way */
}
macosx and Windows have both added support for RUNTIME_PASTEBUF_SUPPORT
Simplify docall() for object types. Adds some different complexity to
a new routine so the overall simplification is rather minimal, but we
already have a routine to construct prompt buffers involving formatted
object names without allowing overflow, so use it.
tty getlin() limits the input to COLNO characters, so 80 by default.
To get potential QBUFSZ overflow, I had to increase COLNO in global.h
and rebuild from scratch. A value greater than 127 triggers a lot of
warnings. I didn't try 127. 126 gets one warning, involving use of
FARAWAY (defined as COLNO+2) in dogmove.c.
We should change things to limit object names to much less than 80,
but this doesn't attempt to implement that.
The earlier fix for hoizontal vs vertical doors would have worked for
the Cav quest (lower left in leader's chamber) where door handling
occurs before wallification, but it wasn't working for minend-3 (east
wall of entry room) which already had walls. The more recent fix
solved the second case but broke the first one. I think this actually
solves both modes of door classification. I hope....
Forwarded to the contact form from a github "issue": in some
circumtances clairvoyance lets you move the cursor around to examine
the revealed map, and when doing so starts with "for instructions
type '?'". When extended clairvoyance periodically kicks in, as
opposed to explicitly casting the spell, there wasn't sufficient
context to figure out what it was prompting for (unless you actually
answer '?' to get instructions). Depending upon the most recent
message, it could seem like quite a strange prompt. Explicitly give
a clairvoyance-specific message prior to that.
Also, the getpos() help was including suggestions for targetting
monsters that aren't appropriate when it's being used for detection.
do_name.c has had quite a bit of formatting rot slip in since 3.6.0.
This fixes up the stuff I spotted by manual inspection.
Reorder some code in pline() so that early pline messages which use
raw_print() instead of putstr(WIN_MESSAGE) are included in the DUMPLOG
message buffer. If the game ends soon enough they'll be shown in the
final log; otherwise they'll get pushed out of the buffer once enough
later messages are delivered.
If we do it the other way round, then mimics will forget what
they're mimicking without a seemimic() call, meaning that the
line-of-sight calculations can get confused if the mimic was
mimicking something opaque.
Web contact report of a github pull request. A previous fix from
same user dealt with potential crash caused by freeing mailbox data
when the mailbox came from getenv("MAIL"). getenv() doesn't return
a value obtained by malloc so freeing it was bad. The fix was to
allocate memory to hold a copy of getenv("MAIL") so that free() was
valid. Unfortunately it didn't allocate enough space to hold the
terminating '\0' so potentially corrupted malloc/free bookkeeping
data. And the alloc+copy was being performed every time the mailbox
was checked, resulting in leaked memory from the previous check (if
MAIL came from player's environment). Fortunately the recheck only
takes place after new mail is actually detected and reported to the
player so the leak was probably small for most folks.
This compiles for the set of conditionals that apply to me (after
taking out -DNOMAIL that the hints put in my Makefile) but I can't
test that it actually works since mail is never delivered to this
machine.
I can't find the original message at the moment, but one of the things
that an analyzer complained about was the *s='\0' possibly assigning
to a Null pointer. The superfluous test of 's' in the while condition
has fooled it into thinking that's possible when it's not.
if (s) {
while (s && ...) {
*s++ = ...
}
*s = '\0';
}
- fix bug in git hooks that loses file permissions when doing variable expansion
- fix execute permissions on sys/unix/hints/macosx.sh
- explicitly call out perl as required for hooks in Developer.txt
Noticed while looking into the TROUBLE_STUCK_IN_WALL prayer bug,
place_lregion() has been using the wrong row for 'low y' in its
whole-level handling, presumeably ever since it was first introduced.
3.4.3 definitely had the same bug; I didn't check any further back.
For maze levels which only consider every other row and every other
column to be viable locations this probably didn't matter. And even
non-maze levels usually don't have anything on row 0, so this fix
isn't likely to be noticeable.
First reported two years ago, then again this week by someone else
who didn't go through the web contact page (so no new #H number or
bugzilla entry). Using vmsbuild.com to build on VAX complains about
not being able to resolve a bunch of functions--it's basically
trying to build the full program using only the code supplied by
sys/vms/vmsmain.c. The original report mentioned a workaround and
was also dealing with a second issue (already fixed post-3.6.0) that
I incorrectly guessed was responsible for the linking problem. This
report had the correct linker magic to fix the linking issue.
I'm still not sure whether the order of /Library and /Include after
the name of an object library file on a LINK command line matters.
In a linker options file, which vmsbuild.com constructs and uses,
/Include needs to come first so that the contents of the library are
searched after the explicitly included object modules are processed.
Building with the Makefiles (using DEC's MMS or some versions of
freeware MMK) doesn't collect the object files into a library so was
never affected by this. And the linker options ordering issue is
apparently specific to the VAX/VMS linker; vmsbuild.com run on Alpha
and on IA64 linked 3.6.0 successfully without this fix.
