It's quite rare, though possible, that during finalstretch you could see
a glitchy tileset that looked like this:
https://i.imgur.com/V7cYKDW.png
This happened because final_mapcol, the variable that controls which
color of finalstretch is rendered, could end up being 7. Normally, it's
in the range of 1..6, which perfectly correlates with the Warp Zone
tilesets in tiles2.png, and the higher the number the farther back in
the tileset it goes from the gray Warp Zone tileset. However, if it's 7,
then it'll start grabbing tiles from the Ship plus some unused blank
tiles, which does not look pretty in the slightest.
This happened because it's possible, though exceedingly unlikely, that
fRandom(), a function which returns a float between 0..1, could return
exactly 1. fRandom() calls rand(), which returns a result between 0 and
RAND_MAX, and divides it by RAND_MAX. This value is implementation
dependent, but required to be at least 32767, and on most systems is
2147483647. Even taking the value of 32767, that means there's a 0.003%
chance that you could get this glitchy tileset when the game cycled the
color in finalstretch. But of course, playing the game for long periods
of time will eventually increase this chance - cycling the color 1,000
times (around 17 minutes of playing) will result in the chance being 3%.
Then as the calculations in the finalstretch color cycling logic calls
fRandom(), then multiplies by 6 and adds 1, it was possible for
fRandom() to return exactly 1, then have
6 added to it, resulting in final_mapcol being 7.
To fix this, just decrement the multiplication by fRandom() to multiply
by 5 instead of 6. Now the only possible numbers that calculation can
produce would be 1..6.
This fixes a limitation where the level filename had to be the exact
same name as the name of the zip, because the game used the name of the
level to identify the zip of which to load assets, and this also made it
impossible to use assets for more than one level in a zip.
Instead, we just look up where the level came from, so we can always
load its assets regardless of its filename.
Additionally, the zip structure checks can go away too, simplifying the
code further.
This WOULD be a huge breaking change, if it weren't for the fact that no
one uses them. Which is why I'm removing them, to simplify the code.
I asked on the VVVVVV Discord whether anyone used them or was even aware
of them and basically the answer was no. I go on Distractionware and no
one uses them. And why would they, when they'd have to distribute the
level .vvvvvv file separately? Better to just distribute everything in
one zip. And it's quite a bit obscure that you have to suffix the file
with .data.zip anyway.
During review of #869, I looked at this part of the codebase again. I
have no idea how or why, but during the course of 2.4 this whole area
just became a mess.
The issues I fixed (in no particular order):
- Copy-pasting the code that loads from the binary blobs
- Making sure SDL_RWFromConstMem is used over SDL_RWFromMem wherever
possible
- Adding checks to make sure the index from the binary blob is valid
(it's possible it could not exist)
- Adding checks to make sure we gracefully handle
SDL_RWFromConstMem/PHYSFSRWOPS_openRead returning NULL
- Moving the pointer asterisk to the type instead of the name :)
So, it turns out we weren't quite done fighting CMake yet...
To accommodate #869 (and actually also #272), the C standard was raised
from C90 to C99. This turned out to require a bit of a fight with the
CentOS CI's CMake version to get it to set the flags we wanted (and to
not overwrite them later). Eventually the fix was to move the block
that sets the standards to later in the file, which was done in
24353a54bb.
As it apparently turns out, if your CMake is at least 3.1.3 and
`CMAKE_<LANG>_STANDARD` is used instead of the workaround, the standard
setting now has an effect on the third party libraries, but not on
VVVVVV itself. The cause is (probably) the phrase "if it is set when a
target is created" in the CMake documentation - the
`CMAKE_<LANG>_STANDARD` values have to come before the VVVVVV target is
defined. In other words, the compiler's default C/C++ standard will be
used, probably something like C17 and C++17. As I can confirm with
`__cplusplus` and `__STDC_VERSION__` with my recent-enough CMake. If I
force the pre-3.1.3 workaround to be used, everything is compiled with
C99/C++98 as expected; and the `-fno-exceptions` `-fno-rtti` flags
appear everywhere regardless of version.
So my fix is to make the CMakeLists a little less complex by
simplifying away the `CMAKE_<LANG>_STANDARD` and
`CMAKE_<LANG>_EXTENSIONS`, and always using the workaround regardless
of CMake version. There's nothing wrong with the workaround, the same
thing is also done for `-fno-exceptions` `-fno-rtti`, and it's good to
have a less complicated CMakeLists that doesn't do different and
unexpected things for different versions.
The previous commit f6d7a214f8 ended up
breaking CI because the workaround ended up breaking the PhysFS build
too, which was previously relying on extensions to compile.
Since #869 is going to require C99 anyways, I might as well just up the
standard now. That way the PR won't have to fight it too.
Previously, if the user had a CMake version below 3.1.3, we told them to
set `-std` themselves.
However, we are going to go to C99 soon (because of FAudio, see #869),
and CentOS 7's CMake is too old to set `-std=` automatically, defaulting
to C90. This is bad because it fails the CI.
To work around this, we set `-std=` ourselves, but first we have to
clear any existing `-std=` flag in C_FLAGS or CXX_FLAGS. Amusingly
enough, MSVC does not have `/std:` switches for either C90 or C++98, so
we just get to do nothing.
This isn't necessary, but it does silence these annoying logs if you
pass an invalid argument or don't have data.zip:
[ERROR] Could not get window size: Invalid renderer
[WARN] Stats not loaded! Not writing unlock.vvv.
[ERROR] Could not get window size: Invalid renderer
[WARN] Settings not loaded! Not writing settings.vvv.
To do this, I've added FILESYSTEM_isInit().
This means we are no longer copy-pasting PhysFS source files directly.
Since the source files reside in a src/ subdirectory, the paths in the
CMakeLists.txt have to be adjusted.
We are no longer copy-pasting LodePNG source files directly.
As we can't rename lodepng.cpp to lodepng.c in the submodule itself, we
need to make a wrapper file, lodepng_wrapper.c, that #includes
lodepng.cpp, but gets compiled as C.
This prevents writing to unlock.vvv or settings.vvv if the game hasn't
made an attempt to load them yet. Otherwise, if the game aborted via
VVV_exit() because of, say, failure to parse a graphics file, it would
overwrite perfectly existing valid save data since it hasn't loaded it
yet.
Fixes#870.
Another cause of #870 is d0ffafe117, as a
bisect tells me. What that commit did is remove screenbuffer as a
pointer, since it's a statically-allocated object that _should_ always
exist, and it removed the `screenbuffer == NULL` guards in savestats()
and savesettings(). Unfortunately, those guards did something very
important - namely, they prevented writing to the save files when the
filesystem wasn't initialized. But that wasn't made clear, because it
seemed like the point of those guards was to prevent dereferencing NULL.
So instead, explicitly make it clear that
FILESYSTEM_saveTiXml2Document() needs to fail if the filesystem isn't
initialized. I've done this by adding an isInit bool to
FileSystemUtils.cpp.
Issue #870 showed one of the problems that this game has, namely that it
only sometimes checks SDL return values, and did not do so in this case.
Part of the cause of #870 is that Screen::GetWindowSize does not check
the return value of SDL_GetRendererOutputSize, so when that function
fails (as in the case where m_renderer is NULL and does not exist), it
does not initialize the out values, so it ends up writing uninitialized
values to the save files.
We need to make sure every function's return value is checked, not just
SDL functions, but that will have to be done later.
While reviewing #272, I noticed that the PR was passing these two
arguments through a helper function, even though they really shouldn't
ever change. To obviate the need to pass these through, I'm making them
global variables.
pathSep is just a string literal from PhysFS, while basePath is a whole
complicated calculation from SDL and needs to be freed. It will be freed
upon filesystem deinit (as is done with PhysFS and the STDIN buffer).
Additionally the logic in FILESYSTEM_init is simplified by no longer
needing to keep a retval variable or use gotos to free basePath in
there.
This lets any script name use capitals and spaces all they want, while
still being able to jump to them via iftrinkets() or similar.
The issue is that whenever tokenize() is ran, all spaces are stripped
and every argument is lowercased before being put into `words`. So, the
solution here is to create a raw_words array that doesn't perform space
stripping or lowercasing, and to refer to that whenever there's a script
command that loads a script. We keep the lowercasing and space removal
elsewhere to be more forgiving to newcomers.
This is technically a forwards compatibility break, but it's only a
minor one, and all levels that utilize it can still be easily modified
to work on older versions anyway.
As reported by Dav999, Victoria and Vermilion's trophy colors are
swapped again in 2.4. He points to
37b7615b71, the commit where I fixed the
color masks of every single surface to always be RGB or RGBA.
It sounded plausible to me, because it did have to do with colors, after
all. However, it didn't make sense to me, because I was like, I didn't
touch the trophy colors at all after I originally fixed them.
After I ruled out the RGBf() function as a confounder, I decided to see
whether intentionally reversing the color order in RGBf() to be BGR
would do anything, and to my surprise it actually swapped the colors
back around and it didn't actually look bad.
And then I realized: Swapping the trophy colors between RGB and BGR
ordering results in similar colors that still look good, but are simply
wrong, but not so wrong that they take on a color that no crewmate uses,
so it'd appear as if the crewmates were swapped, when in reality the
only thing that was swapped was actually the color order of the colors.
Trying to fix this by swapping the colors again, I actively confused
colors 33 and 35 (Vermilion and Victoria) with colors 32 and 34
(Vitellary and Viridian), so I was confused when Vermilion and Victoria
weren't swapping. Then as a debugging step, I only changed 34 to 32
without swapping 32 as well, and then finally noticed that I was
swapping Vitellary and Viridian, because there were now two Vitellarys.
And then I was reminded that Vitellary and Viridian were also wrongly
swapped since 2.0 as well.
And so then I finally realized: The original comments accompanying the
colors were correct after all. The only problem was that they were fed
into a function, RGBf(), that read the colors backwards, because the
codebase habitually changed the color order on a whim and it was really
hard to reason out which color order should be used at a given time, so
it ended up reading RGB colors as BGR, while it looked like it was
passing them through as-is.
So what happened was that in the first place, RGBf() was swapping RGB to
BGR. Then I came and swapped Vermilion and Victoria, and Vitellary and
Viridian around. Then later I fixed all the color masks, so RGBf()
stopped swapping RGB and BGR around. But then this ended up swapping the
colors of Vermilion and Victoria, and Vitellary and Viridian once again!
Therefore, swapping Vermilion and Victoria, and Vitellary and Viridian
was incorrect. Or at least, not the fix to the root cause. The root
cause would be to swap the colors in RGBf(), but this would be sort of
confusing to reason about - at least if I didn't bother to just type the
RGB values into an image editor. But that doesn't fix the real issue,
which is that the game kept swapping RGB and BGR around in every corner
of the codebase.
I further confirmed that there was no more RGB or BGR swapping by
deleting the plus-one-divide-by-three transformation in RGBf() and
seeing if the colors looked okay. Now with the colors being brighter, I
could see that passing it straight through looked fine, but
intentionally reversing it to be BGR resulted in colors that at a
distance looked okay, but were either washed out or too bright. At least
finally I could use my 8 years of playing this game for something.
So in conclusion, actually, 37b7615b71
("Fix surface color masks") was the real fix, and
d271907f8c ("Fix Secret Lab Time Trial
trophies having wrong colors") was the real regression. It's just that
the regression came first, but it wasn't really a regression until I did
the other fix, so the fix isn't the regression, the regression is...
this is hurting my brain. Or the real regression was the friends we made
along the way, or something like that.
This is the most trivial bug ever caused by the technical debt of those
god-awful reversed color masks.
---
This reverts commit d271907f8c.
Fixes#862.
This replaces vcpkg with simply downloading the pre-compiled
dependencies from official upstream releases. The rationale is that
vcpkg is sometimes really slow to update to the latest SDL version when
it releases, and also that it requires the runner to compile SDL every
single time it's instantiated, which is slow and wasteful.
Instead, download the pre-compiled binaries of SDL from its release page
on GitHub. This way, we don't have to compile it ourselves, and we
aren't waiting on vcpkg whenever SDL releases a new version. And for
good measure, cache it so we aren't downloading it _every_ time, which
is even more efficient.
The same can't be done for SDL_mixer, though, because it doesn't have a
GitHub release page with pre-compiled binaries. Instead, we'll download
them from libsdl.org, which is an infrastructure that takes more strain
than if we used GitHub instead. But, it shouldn't matter anyways,
because we cahe this too. And we are going to ditch SDL_mixer for FAudio
in 2.4 anyways, so it's a moot point either way.
In hindsight, the FAudio pointer will likely be in SoundTrack since we will
want to keep the mastering voice closer to the sounds and their source voice
arrays, while the MusicTrack will likely just be one source voice that gets
PCM from different streams.
This looks redundant but will actually help in the transition to FAudio; we
mostly want to keep the game logic the same while reimplementing the current
mixer, weirdness and all. Once that's done and confirmed to be stable and
consistent we can start cutting out the workarounds and quirks.
This meant making the track vectors static, but that's kind of what we do with musicclass anyway?
In any case, this will make the transition to FAudio MUCH less invasive.
