With DEBUG suppressed, I started getting
16 warning: empty body in an if-statement
and 2 warning: empty body in an else-statement
from gcc.
Using braces for an empty block instead of just ';' avoids the warning:
if (foo)
debugpline("foo");
is bad,
if (bar) {
debugpline("bar");
}
is good. ;-)
The changes to lint.h are just precautionary.
modified:
include/lint.h
src/attrib.c, bones.c, dbridge.c, dig.c, eat.c,
makemon.c, mkmaze.c, mon.c, sp_lev.c
Shorten a function name in sp_lev.c that exceeded 31 characters.
That's a limit imposed by the VMS linker and the compiler complains
that it will be truncated.
Make all sp_lev.c functions which aren't listed in extern.h be static
and give all of them a declaration at the top of the file. I reordered
the ones already declared there in the same order as they occur in the
source, so the diff is quite a bit bigger than the actual changes.
(Once the one with the long name became static, the length of its name
no longer mattered, but I've shortened it anyway.)
Indent a couple of #pragma directives. Some pre-ANSI compiler didn't
like '#' in column 1 followed by something it didn't understand, even
when that occurred in a conditional block which was in the midst of
being excluded. (util/*_comp.y recently reminded me of that. files.c
should get a fix like this too.)
This could use a bit of reformatting. I stripped off trailing spaces
from a handful of lines but otherwise have kept it as is since the
formatting issue is minor.
Another issue from old beta-tester mail: #annotate and #overview were
missing from the list of extended commands. M-A and M-O were listed
but marked "(if supported)" even though they've become unconditional.
Same for M-R, although in its case #ride wasn't missing.
Some old beta-tester mail suggested mentioning the implicit_uncursed
option in the "Curses and Blessings" section; this patch does that.
It also mentions that option in the "Configuring Menu Colors" section
since anyone trying to specify a color for " uncursed " will want
objects to be explicitly described as "uncursed".
The changes to the LaTeX version haven't been tested. The generated
plain text version has a lot of spurious changes due to the padding
method it uses to right-justify short lines.
Reported by a beta tester months ago: it was possible to recognize
your god's temple on the Astral Plane by stepping into its doorway,
since #overview would show "temple of <your god>" (only if just one
temple had been entered and the altar in it was the only one you'd
seen and it was for your own god; #overview doesn't show "temple of
<other god>", just "a temple"). After this fix it will just show
"a temple" even when you can see the temple's altar, so #overview
can't be used as a shortcut to finding the right temple.
While testing the fix I discovered that amnesia wasn't handled when
forgetting the current level's map, only when other levels got
flagged as forgotten. The number of altars, fountains, and so on
are recalculated when #overview is executed, so current-level amnesia
worked for those. But data about known rooms is not recalculated,
so the number of temples and shops you'd visited on the level stayed
instead of being forgotten. The fix is a bit iffy for the case where
you only forget random spots scattered across the level's map rather
than the whole thing; this just wipes #overview memory of every room
even if parts of rooms are still remembered.
Change the 'make' rules for *_lex.c, *_yacc.c, and *_comp.h to put
the end files names into whatever #line directives that refer to
the generated names (lex.yy.c, y.tab.c, y.tab.h). This should not
produce any change in behavior during compilation except to provide
more accurate specifications of where any warnings or errors occur.
Add a 'make dist' target that copies generated scanner and parser
files to sys/share. (As mentioned in the previous patch, I haven't
committed any of those since I have different versions of flex and
also of bison.)
Bump the number of nodes (%e) specified in lev_comp.l so that the lex
on Solaris doesn't choke. flex doesn't use those paramters (%e, %p,
and a few others), so this change gets accepted but hasn't been
adequately tested to verify that it solves the problem.
Reformat some of the C code at the end of dgn_comp.l and lev_comp.l;
going through the whole thing isn't worth the effort, particularly
since flex generates such a mess in the rest of the code it produces.
I haven't updated sys/share/*_lex.c and they definitely need it now.
I've got an older version of flex than the one that was used on the
code that's there now, and the most recent version. Both of those
produce different code--the stuff that's there now has been through
more testing so whatever version was used for it ought to be used for
new editions.
How that there's nothing left to trigger it, put in some extra quest
messge debugging code. Could be useful if new roles are added or if
any deliver-by-pline messages get changed to be deliver-by-window.
The samurai ones leave something to be desired since they make no
attempt to include Japanese terminology or to make sure the tone stays
appropriate to a samurai's honor.
Another bit prompted by vibrating square testing:
|You see a strange vibration beneath the little dog's rear claws.
Fix up some body parts: dog, cat, and yeti-class (includes sasquatch,
monkey and ape, owlbear) already have "paws" instead of "fore claws".
Take away all 'Y' except owlbear from that list and add rodents to it.
Give them "rear paws" instead of "rear claws" for their feet; for legs,
use "foreleg" instead of "forelimb" and "read leg" instead of "rear limb".
For yeti/sasquatch/monkey/ape/carnivorous-ape, switch from paws to hands
since they have opposable thumbs, and switch to arm, leg, foot instead
of forelimb, rear limb, and rear claw. I've left "fore claw" for finger.
Noticed while testing the look-at vs vibrating square patch: I was
able to cover stairs with a trap via wizard mode wish. That wish can
achieve a lot of irregular things, but prevent this particular one.
When examining a trap with '/' or ';', show
|a trap (arrow trap)
instead of
|a trap or a vibrating square (arrow trap)
outside of Gehennom (unless the trap actually is a vibrating square,
which could happen via wizard mode wish). The extra verbosity is
distracting, and limiting mention of the vibrating square to the region
where it's relevant may give a hint to players getting that far for the
first time.
Preformat SYSCF entry 'WIZARDS' so that it can be displayed during panic
feedback without allocating memory for the formatted list at that time.
It also gets displayed for help's "support information" ('?k').
For panic(), push "it may be possible to rebuild" to a second line since
the formatted usernames might make the line long.
Mark panic() as never returning so that code analysis might be able
to do a smarter job. It required splitting done() into two routines
since the first part really can return (but not if PANICKED was the
reason it got called). done() is now much shorter and ends with a
call to new really_done(), and panic() skips done()'s might-return
part by calling really_done() directly.
Noticed in passing: the "report error to <list of SYSCF WIZARDS>"
code calls a routine which uses alloc(), which won't work very well
if the reason for panic was because malloc() ran out of memory.
On OSX 10.5, save file compression and uncompression fail when I run
under debugger control (for gdb, at least; I don't know why). I used
'#panic' to try something out. panic() calls exit_nhwindows(), then
after some messages calls dosave0(). Saving calls docompress_file()
and when that fails on tty it tries to call clear_nhwindow(WIN_MESSAGE)
but WIN_MESSAGE has been torn down already, leading to a nested panic().
Avoid the call to clear_nhwindow() if windows aren't up.
In the midst of composing a commit message about how I reorganized some
of genl_putmixed()'s code without finding any problem, I realized that
there was a problem. The character immediately after \G12345678 would
be copied directly to the output buffer without examination. If that
was the leading backslash for a second encoded sequence, the G and the
hex digits would follow their backslash as just ordinary chars, which
is not what's intended. Or if instead of a backslash the next character
was the input's terminating '\0', the latter would be copied into the
output and the pointer to the input string would be incremented, then
the next loop iteraction would examined whatever followed. If valgrind
is smart enough--and it seems to be--it would complain about accessing
a character that putmixed()'s caller hadn't initialized.
The only use of putmixed() I'm sure about is the what-is code showing
a screen symbol with its explanation, which doesn't exercise either
\G12345678\G12345678 or \G12345678\0. I didn't go hunting to see if
there was someplace that might have an encoded symbol at the end of the
string. what-is still works after this patch....
The only substantive change is adding ``continue'' but I haven't gone
back and undone the reorg that preceded it.
Fix some more of the complaints from clang's static analyzer. The one
in options.c (manipulating warnings symbols) appears to be an actual bug.
All the rest are either because the analysis isn't quite sophicated
enough or outright bogus.
Two of them appear to be because a static routine is attempting to guard
against callers in the same file failing to pass in required output
pointers. Stripping away the check for missing pointer should convince
the analyzer that those output parameters always receive a value. We'll
see once the analysis is eventually re-run....
Use an alternate fix for the complaint from clang's static analyzer
(about potentially derefencing a null pointer, which can't happen
here because alloc() panics and quits rather than return Null), plus
some reformatting and removal of a chunk of unused code (strncmpi).
Also a formatting bit for lev_comp.y, making sys/share/lev_yacc.c
be out of date. However, the generated code will be the same--except
for line numbers--so this shouldn't inhibit anybody's planned testing
waiting for the generated copy to be updated.
I think this should fix one of the valgrind complaints. Traps which
didn't use the trap->vl union field never initialized it, leaving a
bit of random garbage in the malloc'd trap structure. (And traps
which overwrote existing ones that did use it didn't reinitialize it
so kept stale data around.) Since those fields weren't in use by
the traps that don't care about them, this didn't provoke any actual
trouble.
Also reformatting....
This should avoid two of the three bogus clang complaints about
retaining the address of a stack variable after it has gone out of
scope.
Plus a recreation of some formatting I did a while back and then
accidentally clobbered before committing.
Fix a couple of the clang static analyzer's warnings.
muse.c has some reformatting. zap.c wasn't triggering any warning about
possible null pointer, but using MON_AT() to maybe avoid m_at() is not
a useful optimization since m_at() is a macro which starts out by using
MON_AT() itself.
Add missing 'sysconf' to sys/unix/ and sys/winnt/ sections of Files.
Update sys/unix/sysconf; I started out just to remove the duplicate
DEBUGFILES entry but ended up expanding several of the comments too.
Also, fix a typo in the vms build/install instructions.
Update the unix Makefiles and the older OSX hints files to handle the
pile marker tile overlay. I didn't touch hints/macosx10.10 and .11
since I think there's still a merge for them pending.
A couple of formatting tweaks for bemain.c are included, for no
compelling reason. What are the odds that anyone will every build
that again?
Confusing build failure, explained by a typo in sys/unix/Makefile.utl.
dgn_lex.o didn't get rebuilt after modifying unixconf.h to take out
the #define MONITOR_HEAP I had in place, resulting in link failure for
dgn_comp because the old object file was referencing 'nhalloc' rather
than 'alloc'. dgn_lex.o accidentally didn't care about modifications
to config.h and the other headers that pulls in, such as unixconf.h.
This typo was already present when the last cvs repository was
initialized nearly 14 years ago.
I tracked down the widest lines, which sometimes occur due to mis-indent
of block comments (see tradstdc.h for an example), and fixed those up.
For the files affected, I also converted tabs to spaces.
A couple of macros with comments in the midst of their expansions got
badly mangled by the automated reformat and one ended up with a line
that was over 150 characters wide.
Some reformatting (replacing tabs with spaces), but mostly updating some
of the comments, particularly about SYSCF.
Shouldn't GENERIC_USERNAMES be controllable via SYSCF?
When reading a novel, select a random passage which hasn't been shown
already. Once you've run through all the passages, it resets to get
them all again (with new random order that might happen to the be same
order if there aren't many passages). Switching to a different novel--
even another copy of the same one--will cause the previous passage
selection to be discarded and restarted from scratch if the prior book
is read again. Passage tracking for the most recently read novel is
kept across save and restore. (That means I needed to bump EDITLEVEL,
so it will need to be reset to 0 again before release.)
Keep window bookkeeping up to date when tty interface is shuting down.
The other interfaces should do something similar when they make windows
known to the core become unavailable.
exit_nhwindows() is called before terminate(), and the tty incarnation
destroys all windows--including 'pickinv_cache_win'--without setting
the various index variables used to access them to WIN_ERR, then
terminate() calls freedynamicdata() which calls free_pickinv_cache()
which tries to destroy 'pickinv_cache_win' since it isn't WIN_ERR (if
the perm_invent option has been enabled during that playing session).
Some of the other <interface>_exit_nhwindows() also tear things down
without resetting the variables used to track them, so fixing this in
exit_nhwindows() would have been pretty messy.
Call free_pickinv_cache() before exit_nhwindows() in done(). At the
moment it's only called from done(), so other exit paths won't release
the small chunk(s) of memory used for the alternate inventory window
(if it got created for perm_invent support).
Replace the code that uses strcat with two pointers into the same buffer.
Treated separately, they point at distinct strings (no overlap possible),
but the C standard does disallow that in order to enable optimizations
using block transfer or such, so the tool that complained about it isn't
wrong. The characters getting appended to the output pointer can end
up overlapping the beginning of the other input pointer, conceivably
breaking an implementation that didn't use simple left-to-right byte-at-
a-time copying.
Also, I noticed that wishing for "luck stone" gave me a random gem.
There's code to strip off " stone" and compare against gem types, but it
prevents other code that accepts "foo bar" as a match for "foobar" and
vice versa from finding a match, since "luck" doesn't match anything
once "stone" is gone. So add the four gray stones into the array of
alternate spellings.
Another bit: it now accepts " gem" in addition to " stone" as optional
gem suffix, so "ruby", "ruby stone", and "ruby gem" all yield ruby.
("luck gem" won't work; you'll end up with a random gem-class object.)
And a last other bit: wishing for "lamp (lit) named foo" would yield
an unlit, unnamed lamp because "(lit)" followed by anything didn't match
"(lit)" and threw away everything past the opening paren. Now it will
produce "lamp named foo (lit)"--a lit lamp named "foo". (Wishing for
"lamp named foo (lit)" produces an unlit lamp named "foo (lit)". That's
acceptable to me... I'm not crawling any farther down this hole. Maybe
object formatting should be changed to keep the lit attribute in front
of the name?)
The memory leak (monst->mextra->edog, monst->mextra->mname,
monst->mextra for some monster were not released) I noticed recently
was due to recording a pet's full monster attributes with its corpse.
During save and restore, obj->oextra->omonst was being treated as a
full-fledged monster so worked as intended, but when freed, omonst
was treated as a black box and its mextra details weren't handled.
Use casts to try to suppress a couple of assignments of long to short
that Michael's compiler warns about. 'cw->maxrow' might have a value
that's too big for 'short' (when dealing with really big menus), but
'cw->maxcol' never will (unless someone comes up with a terminal or
emulator that's wider that 32K-1 characters...).
dungeon.o depending on lev.h is the only change found by 'make depend'.
(I'm a bit suspicious about that.)
I haven't attempted to reconcile the vms Makefiles with the Unix ones,
just put in this one new dependency. I know vms/Makefile.src lacks
handling for sys/share/*regex.c and vms/Makefile.top and install.com
both lack handling for 'sysconf'.
> Somebody has changed versioning so that the game incorrectly states
> 3.6.1 in messages. It looks like someone updated patchlevel instead of
> editlevel?
Yes, that was me. I meant to increment EDITLEVEL and nobody noticed
the mistake until now....
This changes PATCHLEVEL back to the correct value of 0, and implicitly
resets EDITLEVEL to 0 for release (by not changing it to 1 as it was
supposed to have been for the past 3-4 weeks).
Data files from Oct. 18 through today are actually compatible but will
be rejected once anyone rebuilds with this fix, same as would happen
when EDITLEVEL changs. Data files from before Oct. 18 will be
incompatible but be accepted by nethack but not work correctly due to
a change in the 'context' structure.
Release some dynamically allocated memory prior to exit. These were
previously left alone due to assumed complexity (at least by me...),
but dealing with them turned out to be straightforward.
|#if FREE_ALL_MEMORY
free BASE_WINDOW -- tty-specific; other windows are drawn on top of it
free ttyDisplay -- tty's basic data structure
|#endif
free nh_HI, nh_HE -- termcap values handled differently from the rest
These are the last things that 'heaputil' always reported as not freed
for the basic Unix+tty configuration. (I've observed other things not
being freed; those are post-3.4.3 bugs that need to be found and fixed.)
This isn't urgent, but I figure that until the mac build stuff gets
merged in, the core is still fair game....
'O' command's autopickup_exceptions was freeing a menu pick-list even
when it hadn't been allocated (for the list case, and for the remove
case if nothing was chosen for removal). That code was evidently used
as the model for msgtype and menucolors; they had the same situation.
I think ANSI and ISO sanction free(NULL) as a no-op, but pre-ANSI free
implementations don't necessarily handle that benignly. Even if they
all do, freeing something--even if that 'something' is nothing--which
hasn't been allocated is a bug on our end.
README - add VMS back as a tested platform; thanks KevinS!
dat/history - add VMS update, remove trailing whitespace, two spaces;
instead of just one (recently added stuff) for sentence separation;
sys/vms/Install.vms - minimal update;
Files - reformat the win32 project section to fit within 80 columns.
Replace several 'foo = alloc(strlen(bar)+1), strcpy(foo,bar)' sequences
with 'foo = dupstr(bar)' calls.
Change 'free(foo)' into 'free((genericptr_t) foo)' to possibly pacify
'lint' and/or really old compilers.
Add braces around 'if something;' when 'else { otherwise; }' has braces.
Simplify option value formatting for 'sortloot'.
Tell git to ignore the presence of the heaputil program in util/.
(It lives in NHinternal/devteam/util/heaputil.c but working with it
from there is inconvenient.)
Free sysopt.shellers and sysopt.explorers when releasing the memory used
for other sysopt fields.
Also some formatting stuff since sys.c was previously untouched.
If makedefs.c is compiled with MONITOR_HEAP defined, attempted calls to
free() resulted in link failure. Since makedefs doesn't use alloc(),
call free() directly instead of redirecting to nhfree().
Also some assorted reformatting....
This one has a couple of code changes included, but they shouldn't
produce any change in game play. If anyone adds a new shirt or shield
they'll have to update the corresponding foo_on() and foo_off() routines
to avoid an 'impossible' when putting on or taking off the new item(s),
the same situation as already happens for other subclasses of armor.
The automated reformatting put a space in casts of the form
'(type)(expression)', yielding '(type) (expression)', but it didn't
do that for '(typedef)(expression)'. There are lots of instances of
'(boolean)(expression)'; (uchar) and (xchar) also occur. I haven't
noticed other types, but I haven't looked in very many files yet.
Plus a modest amount (really!) of reformatting.
Also got rid of 'copy_of()' since dupstr() does the same thing [except
for the 'treat Null as ""' part; when needed, that can be done in the
call: dupstr(!str ? "" : str)].
Explicitly combine adjacent string literals so that pre-ANSI compilers
still have a chance to compile the code. I thought these had already
been dealt with, but I kept stumbling across them while reformatting,
so am trying to get them all out of the way now.
Fixing up mis-indented block comments, but hit some files that hadn't
had the earlier mixture of tab replacement, etc, so it's bigger than I
expected. If I get to it, they'll be another round of this tomorrow.
The intended change was that you'd get an increase to max energy if
current was sufficiently close to max rather than only when it was
already at max. It wasn't intended that you'd fail to have current
boosted all the way to max in the case where it wasn't sufficiently
close. That's fixed here.
Last few && or || followed by end-of-line comments, plus tab replacement
and 'return' parentheses. Not as many of those; some of these files had
already had that done.
Also, tweaked non-cursed scroll of charging read while confused to be a
tiny bit more effective.
To do: find and fix block comments that immediately follow a line with
an end-of-line comment and got misindented to line up with that comment.
While the topic of strprepend() is current, make good use of it.
Simplify code which inserts "the Nth " in front of "<arrow,&c>".
I'm pretty sure there are one or two other places where I assumed that
the outpuf of xname() was a char array which is BUFSZ in length rather
than BUFSZ-PREFIX, and reused the buffer, but I don't know where they
occur. (BUFSZ-PREFIX is still big enough to hold most things, so it
might not lead to trouble.)
Fix the reported bug that attempting to cast an expired spell, which
causes confusion and/or stun, was replacing the duration of any existing
confusion or stun with the new amount rather than increasing it by that
amount.
Attempting to cast any spell while stunned will now fail immediately,
and casting an expired spell while confused will increase confusion
duration (and/or set stun duration) rather than override it.
Same sort of stuff as before: some continuation lines with operator
followed by end of line comment (only a few files with those still to
go...), plus tab replaced by spaces in comments, excess parenthesis
removal for return statements, and force function name to be in column
one in function definitions:
type name(args) /* comment */
argtype args;
{
to
/* comment */
type
name(args)
argtype args;
{
I've been spotting those by eye rather than rexexp, so probably missed
some.
When doing some more reformatting I came across something I've been
meaning to tweak for a long time, and since the change is only a couple
of lines I'm putting it in now instead of waiting. Make potions of
gain energy more useful for actually regaining energy so they might not
be relegated to alchemy all the time. The adjustment is probably too
low to really achieve that, but I didn't want to risk going too high.
Increase to max energy is only a little higher (average 10.5 vs 9 for
blessed, 7 vs 4 for uncursed) but to current energy is noticeably higher
(31.5 vs 9 for blessed, 21 vs 4 for uncursed; capped by max energy so
bigger increase only matters if current is below max when quaffing).
This should address the issue that the problem patch to display_pickinv()
was trying to deal with: releasing the inventory window before exiting
the program so Pasi's memory checker doesn't think it's a memory leak.
Not related, but in the same file:
The older qsort comparison routines are tagged with CFDECLSPEC to deal
with some C vs C++ interaction issue. I added that to the relatively
recently added 'sortloot' qsort compare callback.
I also changed worn_wield_only(), although it isn't actually called.
(display_minventory() has provisions to call it, but both of the latter's
callers pass in MINV_ALL so allow_all() gets used instead.)
Noticed while going through more reformatting: can_carry() was changed
to return a number rather than yes/no, but it's trying to return a long
value (obj->quan) as an int.
Gold is the only thing likely to exceed LARGEST_INT in actual play
(although rocks could manage it if somebody tried hard enough). This
makes sure that the value returned doesn't exceed LARGEST_INT, but only
tame monsters honor the resulting subset value (at least for gold) and
split the stack. The proper fix is to convert can_carry() and all its
uses to long, but I'd rather spend my time on other stuff.
One of the steps in the pre-release checklist: make sure the version
number in the X11 icons is up to date. nh32icon and nh56icon both have
the version number (major and minor values only, not patchlevel).
Update it from 3.4 to 3.6 for both. (nh72icon doesn't have any version
info in its design, so doesn't need any update.)
The X11 'bitmap' program strips out the header comments, so I modified
the image, saved it, used 'git diff --patch' to capture the changes,
'git checkout' to revert to the existing file, and edited in the diff
bands pertaining to the changed data. No doubt this could have been
done in some easier fashion, but it was victory just to find the bitmap
utility and achieve results using a one-button mouse--worse, a clumsy
touchpad substituting for one--when it's designed for a three-button one.
Reformat some trailing &&, || operators followed by end-of-line comment,
missed by the earlier continuation formating.
An
#if 0
something {
#else
something_else {
#endif
construct in rhack(cmd.c) confused the automated reformatter, resulting
in some code from inside a function ending up in column 1.
Move this small utility routine to hacklib.c where other such things
live and where it's feasible to find them if you need the functionality
elsewhere.
Author: PatR <rankin@nethack.org>
Date: Fri Oct 30 00:50:52 2015 -0700
more formatting
Fix up the files containing '[?:] */' to get trailing trinary operator
followed by end-of-line comment. Tab replacement and removal of excess
parentheses on return statements also done.
Another code change while reformatting: '#turn' by non-priest/non-knight
casts the "turn undead" spell if the hero has learned it, but it was
forcing the spell code to aim at self rather than ask for a direction.
Evidently nobody has ever used that while knowing the spell and able to
cast it....
Some reformatting mutated into a code change: OSX gets code from two
conditional blocks
'#if UNIX: eat apple or pear, give "core dump" message'
'#if MAC || OSX: eat apple, give "It's a Macintosh" message'
but the second wouldn't be chosen due to the if-then-else logic having
already handled obj->otyp==APPLE before it was reached. Reorder them
so that OSX gets the Macintosh message for apples and the core dump
message for pears.
Fix the problem reported by ais where it was possible for one monster
to knock the hero onto a level teleporter (or trapdoor or hole),
destination was selected and allowed-to-level-teleport checks were made,
then for another monster to knock or teleport the not-yet-relocated-hero
onto the Amulet and have auto-pickup move it into inventory. At the end
of that turn's monster movement, hero would level teleport successfully
despite carrying the Amulet.
This short-circuits monster movement if the hero is scheduled to be
moved to a different level. The monsters who haven't moved yet don't
lose their pending movement points; they'll catch up if/when the hero
returns to the level.
End of first pass, but '[&|?:][ \t]*$' doesn't catch trailing operater
followed by end-of-line comment so more needs to be done. As with the
past couple of batches, I've removed redundant parentheses from 'return'
statements but only for files that had continuation fix-ups.
I've also removed tabs from comments in some of the files, but didn't
start until part way through this subset of the sources.
Somewhere along the line I started removing redundant parentheses from
return statements, but only in files that needed continuation fixups
so it's not comprehensive.
Relatively small number of continuation fixes needed for this subset.
Quite a bit of mangling to engrave.c unrelated to continuation lines,
with three or four coding changes.
Mostly && and || at end of the first half of a continued line rather
than at the start of the second half. The automated reformat got
confused by comments in the midst of such lines.
foo ||
bar
was converted to
foo
|| bar
but
foo ||
/* comment */
bar
stayed as is.
Some excluded code [#if 0] was also manually reformatted, but this is
mainly stuff that can be found via regexp '[&|?:][ \t]*$' (with a lot
of false hits for labels whose colon ends their line).
Replace the code that Dean objected to with something a little bit more
robust. It doesn't rely on the two stacks being adjacent or having the
same inventory letter. It is still vulnerable to having another
splitobj() occur between the offending split and its attempted unsplit,
or to either of the two halves of a split being extracted from their
object chain. As before, failure to unsplit only results in the two
halves of the split remaining separate stacks, not anything more drastic
like the panic() that prompted all this.
Simplification of hallucinated currency names got mixed in with this
patch. I haven't bothered separating it back out.
Whoever reset PATCHLEVEL to 0 jumped the gun. This patch increments it
since change to the 'context' structure breaks save file compatibility,
so it will need to undergo another reset before release.
PatR