During the final stretch, after Viridian turns off the Dimensional
Stability Generator, the map goes all psychedelic and changes colors
every 40 frames. Entities change their colors too, including conveyors,
moving platforms, and disappearing platforms.
But play around with the disappearing platforms for a bit and you'll
notice they seem a bit glitchy. If you run on them at the right time,
the tile they use while disappearing seems to abruptly change whenever
the color of the room changes. If there's a color change while they're
reappearing (when you die and respawn in the same room as them), they'll
have the wrong tile and look like a conveyor. And even if you've never
interacted with them at all, dying and respawning in the same room as
them will change their tile to something wrong and also look like a
conveyor.
So, what's the problem? Well, first off, the tile of every untouched
disappearing platform changing into a conveyor after you die and respawn
in the same room is caused by a block of code in gamelogic() that gets
run on each entity whenever you die. This block of code is the exact
same block of code that gets ran on a disappearing platform if it's in
the middle of disappearing.
As a quick primer, every entity in the game has a state, which is just a
number. You can view each entity's state in
entityclass::updateentities().
State 0 of disappearing platforms is doing nothing, and they start with
an onentity of 1, which means they turn to state 1 when they get
touched. State 1 moves to state 2. State 2 does some decrementing, then
moves to state 3 and sets the onentity to 4. State 3 also does nothing.
After being touched, state 4 makes the platform reappear and move to
state 5, but state 5 does the actual reappearing; state 5 then sets the
state back to 0 and onentity back to 1.
So, back to the copy-pasted block of code. The block of code was
originally intended to fast-forward disappearing platforms if they were
in the middle of disappearing, so the player respawn code would properly
respawn the disappearing platform, instead of leaving it disappeared.
What it does is keep updating the entity, while the state of the entity
is 2, until it is no longer in state 2, then sets it to state 4.
Crucially, the original block of code only ran if the disappearing
platform was in state 2. But the other block of code, which was
copy-pasted with slight modifications, runs on ALL disappearing
platforms in final stretch, regardless of if they are in state 2 or not.
Thus, all untouched platforms will be set to state 4, and state 4 will
do the animation of the platform reappearing, which is invalid given
that the platform never disappeared in the first place. So that's why
dying and respawning in the same room as some disappearing platforms
during final stretch will change their tiles to be conveyors.
It seems to me that doing anything with death is wrong, here. The root
cause is that map.changefinalcol() "resets" the tile of every
disappearing platform, which is a function that gets called on every
color change. The color change has nothing to do with dying, so why
fiddle with the death code?
Thus, I've deleted that entire block of code.
What I've done to fix the issue is to make it so the tile of
disappearing platforms aren't manually controlled. You see, unlike other
entities in the game, the tile of disappearing platforms gets manually
modified whenever it disappears or reappears. Other entities use the
tile as a base and store their tile offset in the separate walkingframe
attribute, which will be added to the tile attribute to produce the
drawframe, which is the final thing that gets rendered - but for
disappearing platforms, their tile gets directly incremented or
decremented whenever they disappear or reappear, so when
map.changefinalcol() gets ran to update the tile of every platform and
conveyor, it basically discards the tile offset that was manually added
in.
Instead, what I've done is make it so disappearing platforms now use
walkingframe, and thus their final drawframe will be their tile plus
their walkingframe. Whenever map.changefinalcol() gets called, it is now
free to modify the tile of disappearing platforms accordingly - after
all, the tile offset is now stored in walkingframe, so no weird
glitchiness can happen there.
Ethan, you forgot this other one.
I do have to rejiggle the control flow of the function a bit, so it
doesn't leak memory upon failure. (Although the SDL message box leaks
memory anyway because of X11 so... whatever.) Also, there's a NULL check
for if SDL_GetBasePath() fails now.
According to SDL documentation[1], the returned pointer needs to be
freed. A glance at the source code confirms that the function allocates,
and also Valgrind complains about it.
Also if it couldn't allocate, the game no longer segfaults (std::strings
do not check if the pointer is non-NULL for operator+=).
[1]: https://wiki.libsdl.org/SDL_GetClipboardText
Since mainmenu is only ever used in Input.cpp, I might as well make it
clearer by moving it into a static global variable in Input.cpp. (The
same applies to fadetolab/fadetomenu, but I didn't think much about
those at the time... that'll be a refactor for later.)
While I've decoupled fademode from gamemode starting, being faded out on
the title screen results in a black screen and you being unable to make
any input. So we'll need to store the current fademode in a temporary
variable when going to in-game options, then put it back when we return
to the pause menu. Yes, you can turn on glitchrunner mode during the
in-game options, and then immediately return to the pause menu to
instantly go back to the title screen; this is intended.
Due to frame ordering, putting the fademode back needs to be deferred to
the end of the frame to prevent a 1-frame flicker.
It's actually sufficient enough to do this temporary fademode storage to
fix the whole thing, but I also decided to decouple fademode and
gamemode starting just to be sure.
Assuming glitchrunner mode is off, if you open the pause menu while
fully faded-out and then go to Graphic Options or Game Options, then the
'mode' that you selected previously will kick in again and you'll be
suddenly warped back.
So if you previously started a new game in the main game (mode 0, also
the selected mode if you do this from command-line playtesting), and
then open the pause menu and go to in-game options, then you'll suddenly
go back to starting a new game again. If you had started a custom level,
doing this will warp you back to the start of the level again.
The problem is simple - when the title screen is fully faded out, it
calls startgamemode(). So the solution is simple as well - just decouple
the fademode from calling startgamemode(), and use a different variable
to know when to actually call startgamemode().
Custom levels can have warp lines. If you have a warp line and a warping
background in the same room, the warp line takes precedence over the
warp background.
However, whenever you enter a room with a warp line and warp background,
any entities on the warping edges will be drawn with screenwrapping for
one frame, even though they never wrapped at all.
This is due to frame ordering: when the warp line gets created,
obj.customwarpmode gets set to true. Then when the screen edges and
warping logic gets ran, the very first thing that gets checked is this
exact variable, and map.warpx/map.warpy get set appropriately - so
there's no way the entity could legitimately screenwrap.
However, that happens in gamelogic(). gamelogic() is also the one
responsible for creating entities upon room load, but that happens after
the obj.customwarpmode check - so when the game gets around to rendering
in gamerender(), it sees that map.warpx or map.warpy is on, and draws
the screenwrapping, even though map.warpx/map.warpy aren't really on at
all. Only when gamelogic() is called in the frame later do map.warpx and
map.warpy finally get set to false.
To fix this, just set map.warpx and map.warpy to false when creating
warp lines.
I just spotted this one - if vy isn't bounds-checked, this causes bogus
input from the createentity() script command to commit Undefined
Behavior. Should've spotted this one when I was adding bounds checks to
the rest of createentity() earlier, but at least it's fixed now.
Mentioning 2.3 here is going to be anachronistic when 2.4 exists. And
there's not really any point in bumping this version number every time
the game's version number is bumped as well. So it's best to just remove
it.
This makes it easier to add bounds checks to all accesses of
map.explored. Also, all manually-written existing bounds checks have
been removed, because they're going to go into the new getters and
setters.
The getter is mapclass::isexplored() and the setter is
mapclass::setexplored().
It is no longer possible to cause Undefined Behavior via accessing
out-of-bounds room properties.
What happens instead is - if you attempt to fetch an out-of-bounds room
property, you get a "blank" room property that just has all of the
defaults, plus its tileset is 1 because all tilesets that are nonzero
use tiles2.png, and it closely emulates the previous behavior where it
was some bogus value but definitely not zero. Its Direct Mode is also 1,
because the tiles contained within it are just mishmashed repeats of
existing tiles on the map, and we shouldn't autotile that.
The roomname also gets cleared in case the user attempts to set the room
name of an out-of-bounds room property.
If you attempt to set the property of an out-of-bounds room property,
then nothing happens.
This replaces all raw ed.level accesses with new setter and getter
funcs, which makes it easier to add bounds checks later. And I've also
removed all the manually-written bounds checks, since they will go into
the new getter and setter.
To get the room properties of a specific room, you use
editorclass::getroomprop(), which returns a pointer to the room
properties - then you just read off of that pointer. To set a room
property, you use editorclass::setroom<PROP>(), where <PROP> is the name
of the property. These are maintained using X macros to avoid
copy-pasting. editorclass::getroompropidx() is a helper function and
shouldn't be used directly.
This removes all traces of Undefined Behavior from getting and placing
tiles.
This mimics the previous behavior (2.2 and below) as reasonably as
possible. `vmult` was previously a vector, there was a bunch of unused
space directly after the end of the usable space of the vector, which
was all filled with zeroes. The same goes for `contents`, having
previously been a vector, and so having a bunch of zeroes immediately
following the end of the in-bounds space. That's why both are 0 if you
index them out of bounds.
This makes it easier to add bounds checks to all accesses of
ed.contents.
To do this, I've added editorclass::gettile(), editorclass::settile(),
and editorclass::getabstile() (with a helper function of
editorclass::gettileidx() that really shouldn't be used directly), and
replaced all raw accesses of ed.contents with those functions
appropriately.
This also makes the code more readable, as a side effect.
The existing bounds checks were correct sometimes but other times were
not.
The bounds check for 2x2 and 2x1 sprites only covered the top-left
sprite drawn; the other sprites could still be out of bounds. But if the
top-left sprite was out of bounds, then none of the other sprites
wouldn't be drawn - although it ought to be that the other sprites still
get attempted to be drawn. So I've updated the bounds checks
accordingly, and now an out of bounds top-left sprite won't prevent the
drawing of the rest of the sprites.
Similarly, if the sprite of a Gravitron square was out of bounds, that
would prevent its indicators from being drawn. But the indicators
weren't being bounds-checked either (2.3 lets you have less than 1200
tiles in a given tilesheet). So the bounds check has been moved to only
cover the drawframe and the indicator indexes accordingly, and an out of
bounds sprite won't prevent attempting to draw the indicators.
It is possible for any of the QueryIntAttribute()s to fail, most
commonly if the attributes don't exist. If that happens, then that part
of the temporary edentity won't be initialized, and we'll end up having
a partially-uninitialized edentity - then doing much of anything with it
will result in undefined behavior.
To fix this, just initialize the temporary edentity.
If an XML tag doesn't contain anything inside, pText will be NULL. If
that happens without being checked, then NULL will be passed to
SDL_strcmp(). SDL_strcmp() will either call libc strcmp() or use its own
implementation; both implementations will still dereference the NULL
without checking it.
This is undefined behavior, so I'm fixing it. The solution is to do what
is done with all other XML parsing functions, and to make sure pText
gets set to a safe empty string (which is just a pointer to a null
terminator) if it happens to be NULL.
PR #279 added game.gametimer solely for the editor ghosts feature. It
seems that whoever originally wrote it (Leo for the now-dead VVVVVV:
Community Edition, I believe) forgot that the game already had its own
timer, that they could use.
The game timer does increment on unfocus pause (whereas this doesn't),
but that's a separate issue, and it ought to not do that.
So #434 didn't end up solving the deltaframe flashing fully, only
reduced the chances that it could happen.
I've had the Level Complete image flash a few times when the Game Saved
text box pops up. This seems to be because the Level Complete image is
based off of the text box being at y-position 12, and the Game Saved
text box is also at y-position 12. Level Complete only gets drawn if the
text box additionally has a red channel value of 165, and the Game Saved
text box has a red channel value of 174. However, there is a check that
the text box be fully opaque first before drawing special images. So
what went wrong?
Well, after thinking about it for a while, I realized that even though
there is indeed an opaqueness check, the alpha of the text box updates
BEFORE it gets drawn. And during the deltaframes immediately after it
gets updated, the text box is considered fully opaque. It's completely
possible for the linear interpolation to end up with a red channel value
of 165 during these deltaframes, while the text box is opaque as well.
As always, it helps if you have a high refresh rate, and run the game
under 40% slowdown.
Anyways, so what's the final fix for this issue? Well, use the text box
'target' RGB values instead - its tr/tg/tb attributes instead of its
r/g/b attributes. They are not subject to interpolation and so are
completely reliable. The opaqueness check should still be kept, though,
because the target values don't account for opaqueness. And this way, we
get no more deltaframe flashes during text box fades.
An even better fix would be to not use magic RGB values to draw special
images... but that'd be something to do later.
Clang warns on this. This doesn't fix anything but it does ensure that
whoever's reading it won't be focused as to whether or not omitting the
second set of braces is legal or not.
Previously, with the wrong loop order, this kludge needed to exist so
entities in finalmode didn't have wrong colors for 1 frame when entering
a room. But now the loop order has been fixed, and so this kludge is no
longer needed.
In 2.2, at render time, the game rendered screenshakes and flashes if
their timers were above 0, and then decremented them afterwards. The
game would also update the analogue filter right before rendering it,
too.
In 2.3, this was changed so the flash and screenshake timers were
unified, and also done at the end of the frame - right before rendering
happened. This resulted in 1-frame flashes and screenshakes not
rendering at all. The other changes in this patchset don't fix this
either. The analogue filter was also in the wrong order, but that is
less of an issue than flashes and screenshakes.
So, what I've done is made the flash and screenshake timers update right
before the loop switches over to rendering, and only decrements them
when we switch back to fixed functions (after rendering). The analogue
filter is also updated right before rendering as well. This restores
1-frame flashes and screenshakes, as well as restores the correct order
of analogue filter updates.
This reintroduces 2-frame edge-flipping after the 1-frame input delay
got removed. This is because along with processing input and moving
Viridian, logical onground/onroof assignments need to processed in the
same between-render sequence as well - otherwise Viridian only gets 1
frame of edge-flipping due to frame ordering.
I will need to separate these into two different variables because I
will need to move logical onground/onroof assignments to the start of
gamelogic() - if I kept them together, however, that would change the
visuals of onground/onroof, which I want to keep consistent with 2.2.
To do this, GAMEMODE input needs to be processed, and Viridian needs to
be moved, in the same sequence between render frames. So just move
gameinput to after gamerender. Yes, this is not 2.2 order, but gameinput
only handles player input and nothing else - plus a 1-frame input delay
feels really awful to play with in over-30-mode.
In order to re-remove the 1-frame input delay, we will have to poll
input right after rendering a frame - in other words, just before an
input function gets called.
To do this, I've added a new function enum type - Func_input - that is
the same as a fixed function, but before its function gets called,
key.Poll() gets called. And all input functions have been updated to use
this enum accordingly.
This once again fixes the facing directions of crewmates upon room load,
except now it covers more cases.
So, here is the saga so far:
- 2.0 (presumably) to 2.2: crewmate direction fix is special-cased at
the end of mapclass::loadlevel(). Only covers crewmates created during
the room load, does not cover crewmates created from scripts, only
covers state 18 of crewmates.
- 2.3 currently (after #220): crewmate direction fix is moved to
entityclass::createentity(), which covers every avenue of crewmate
creation (including from scripts), but still only covers state 18.
- This commit: crewmate direction fix now covers every possible state of
the crewmate, also does not copy-paste any code.
What I've done instead is to make it so createentity() will immediately
call updateentities() on the pushed-back entity. This is kludge-y, but
is completely okay to do, because unlike other entities, crewmate
entities never change their state or have any side-effects from
double-evaluation, meaning calling updateentities() on them is
idempotent and it's okay to call their updateentities() more than once.
This does have the slight danger that if the states of crewmates were to
change in the future to no longer be idempotent, this would end up
resulting in a somewhat hard-to-track-down double-evaluation bug, but
it's worth taking that risk.
This fix is not applied to entity 14 (the supercrewmate) because it is
possible that calling updateentities() on it will immediately remove the
entity, which is not idempotent (it's changing the state of something
outside the object). Supercrewmates are a bit difficult to work with
outside of the main game anyways, and if you spawn them you could
probably just use the changedir() script command to fix their direction,
so I'm not inclined to fix this for them anyway.
This copy-pasted code only existed because the previous loop order was
incorrect and rendered entities before they would get properly updated
by the fixed render function. Now, the fixed render function is
guaranteed to be called before the render function, so we can rely on
that to update the drawframe and realcol of entities instead of
duplicating the code ourselves in createentity().
The drawframe assignment is still kept to fix the case where dying while
completestop is active (i.e. during a trinket or crewmate rescue
cutscene) and respawning in a different room won't turn everything into
Viridian sprites.
The background would change for 1 frame before sending you back to the
pause menu or editor settings. The map.nexttowercolour() call needs to
be deferred until the end of the frame.
The new loop order introduces a glitch where the menu would display
whichever menu was saved to kludge_ingametemp for 1 frame right as the
user returned to the pause menu. This happened because the
game.returntomenu() happens in titleinput(), which comes before
titlerender(). To fix this, we just need to defer it to the end of the
frame.
game.shouldreturntoeditor was added to fix a frame ordering issue that
was causing a bug where if you started playtesting in a room with a
horizontal/vertical warp background, and exited playtesting in a
different room that also had a horizontal/vertical warp background and
which was different, then the background of the room you exited in would
slowly scroll offscreen, when you re-entered the editor, instead of the
background consisting entirely of the actual background of the room.
Namely, the issue was that the game would render one more frame of
GAMEMODE after graphics.backgrounddrawn got set to false, and re-set it
to true, thus negating the background redraw, so the editor background
would be incorrect.
With defer callbacks, we can now just use a couple lines of code,
instead of having to add an extra kludge variable and putting handling
for it all over the code.
