Yes, when clipping the start and end points to the screen area, ZUN
uses an integer division to calculate the line slopes, rather than a
floating-point one. Doesn't seem like it actually causes any incorrect
lines to be drawn, though; that case is only hit in the Mima boss
fight, which draws a few lines with a bottom coordinate of 400 rather
than 399. It *might* also restore the wrong pixels at parts of the
YuugenMagan fight, causing weird flickering, but seriously, that's an
issue everywhere you look in this game.
Part of P0069, funded by [Anonymous] and Yanga.
Which store colors as GRB, as suggested by the structure's ID string.
Even master.lib's own functions add an additional XCHG AH, AL
instruction to get colors into and out of this format. MASTER.MAN
suggests that it's some sort of standard on PC-98. It does match the
order of ths hardware's palette register ports, after all.
(0AAh = green, 0ACh = red, 0AEh = blue)
Now we also know why __seg* wasn't used more commonly, as lamented in
c8e8e98. Turbo C++ simply doesn't support a lot of arithmetic on
segment pointers.
And then that undecompilable far call to a function within the same
segment, but inside a different translation unit…
Also, thanks again to Egor for the SCOPY@ hack that debuted in 0460072.
Would have probably struggled with this a lot more without that.
And *then* you realize that TH01 effectively doesn't even use the
resident palette. 😐
And yes, we're procrastinating the whole issue of potentially using
a single translation unit for all three binaries by using a common
segment name, because it *really* isn't that easy.
Completes P0066, funded by Keyblade Wiedling Neko and Splashman.
Meh, can't overload arithmetic operators that take a Subpixel without
generating a needless load and store, even with -Z. But heck, slightly
uglifying subpixel/subpixel arithmetic is exactly the right trade-off.
Completes P0061, funded by Touhou Patch Center.
Yes, decompilation, of something that was so obviously originally
written in ASM. We're still left with two un-decompilable instructions
here, but I'm amazed at how nicely I was able to abstract away all of
the gory register details, leading to pretty clear, readable, and dare
I say *portable* code?! Turbo C++ was once again pretty helpful here:
• `static_cast<char>(_BX) = _AL` actually compiles into `MOV BL, AL`,
as you would have intended,
• and no-op assignments like _DI = _DI are optimized away, allowing
us to leave them in for clarity, so that we can have all parameter
assignments for the SPRITE16 display call in a single place.
I love this compiler.
Part of P0060, funded by Touhou Patch Center.
Yeah… such fun pretending that the original code wasn't copy-pasted.
And yes, Reimu will have to wait until the next one.
Completes P0037, funded by zorg.
And if I don't manage to cover Reimu in this push, it's because ZUN
switched around the cases in half of the functions here… 😵 Here's
some macros instead, to make the code at least *look* as table-driven
and readable as it should have been in the first place.
Part of P0037, funded by zorg.
Ooh, shot position being determined by RNG at lower shot levels? That's
some RNG manipulation TAS potential right there! Maybe.
Part of P0037, funded by zorg.
And once again, the TH05 version is un-decompilable. :/ It was pretty
close this time, though, as the entire block between PUSH DI and POP DI
kind of resembles a separate inlined function, in accordance with Turbo
C++'s automatic backup of the DI register, as researched in 7f971a0.
Except that it contains a loop, and Turbo C++ refuses to inline any
function with `do`, `while`, `for`, or `goto`. If it didn't, it would
have totally worked.
Also, yes, C++ class methods are treated identically in this regard.
Oh well. Shot type control functions next, finally!
Completes P0035, funded by zorg.
So it's *_put(), inherited from master.lib, for everything just writing
to text RAM, and *_render() for everything more involved? But what
about master.lib's own graphics RAM functions like super_put()? Need to
fix that inconsistency some day.
Once again no decompilation, because…
Part of P0033, funded by zorg.
The TH02 version is a piece of cake…
… but TH04 starts turning it into this un-decompilable piece of
unnecessarily micro-optimized ZUN code. Couldn't have chosen anything
better for the first separate ASM translation unit.
Aside from now having to convert names of exported *variables* to
uppercase for visibility in ASM translation units, the most notable
lesson in this was the one about avoiding fixup overflows. From the
Borland C++ Version 4.0 User's Guide:
"In an assembly language program, a fixup overflow frequently
occurs if you have declared an external variable within a
segment definition, but this variable actually exists in a
different segment."
Can't be restated often enough.
Completes P0032, funded by zorg.
"Yeah, let's do this real quick, how can this possibly be hard, it's
just MOVs and a few function calls"…
…except that these MOVs access quite a lot of data, which we now all
have to declare in the C world, hooray.
Once it came to midbosses and bosses, I just turned them into C structs
after all. Despite what I said in 260edd8… after all, the ASM world
doesn't care about the representation in the C world, so they don't
necessarily have to be the same.
Since these structs can't contain everything related to midbosses and
bosses (really, why did all those variables have to be spread out like
this, ZUN?), it also made for a nice occasion to continue the "stuff"
naming scheme, describing "an obviously incomplete collection of
variables related to a thing", first seen in 160d4eb.
Also, PROCDESC apparently is the only syntactically correct option to
declare an extern near proc?
Also, that `boss_phase_timed_out` variable only needs to be here
already because TCC enforces word alignment for the .data segment…
yeah, it's technically not related to this commit, but why waste time
working around it if we can just include that one variable.
Completes P0030, funded by zorg.
I've had the idea to hide this implementation detail and improve code
readability for some time now, but it obviously must still all inline,
to be indistinguishable from a direct assignment of the correct value…
… which, amazingly, it does! Even the static_cast from float to int.
The latter allows us to exclusively implement this for float, since we
do have to express the occasional value smaller than 16.
Who needs macros anyway. Yay, C++ in TH04 and TH05 after all!
Part of P0030, funded by zorg.
Two DOS utilities were made for this:
- gensize: generates TASM macro definitions with filesizes.
- copycat: similar to copy/b a+b+c d, except a+b+c is specified in a
separate file to avoid command line length limitations.
th02/zun.com is bit-perfect
th03/zun.com is almost there, with insignificant differences in
zunsp.com and res_yume.com.
This, hands down, has been the single worst stretch of decompilation so far.
Three extremely difficult functions that each still required inline assembly.
And no, this didn't even work out with any of the optimization features in
Borland C++ that aren't included in Turbo C++.
Only one code segment left in both OP and FUUIN! its-happening.gif
Yeah, that commit is way larger than I'm comfortable with, but none of these
functions is particularly large or difficult to decompile (with the exception
of graph_putsa_fx(), which I actually did weeks ago), and OP and MAIN have
their own unique functions in between the shared ones, so…
Which, for some reason, is also found in the MAIN.EXE of every later game
in between completely unrelated hardware and file format functions.
Separate commit because it has its own segment in REIIDEN.EXE, and because
coming up with the nice function names took pretty long, since I haven't done
anything involving trigonometry in the past 5 years...
Yet another set of questionable C reimplementations of master.lib functions to
waste my time. And half of them, including z_text_(v)putsa, aren't even called
anywhere.
So apparently, TH01 isn't double-buffered in the usual sense, and instead uses
the second hardware framebuffer (page 1) exclusively to keep the background
image and any non-animated sprites, including the cards. Then, in order to
limit flickering when animating the bullet, character and boss sprites on top
of that (or just to the limit number of VRAM accesses, who knows), ZUN goes to
great lengths and tries to make sure to only copy back the pixels that were
modified on plane 0 in the last frame.
(Which doesn't work that well though. When you play the game, you still notice
tons of flickering whenever sprites overlap.)
And by "great lengths", I mean "having a separate counterpart function for
each shape and sprite animated which recalculates and copies back the same
pixels from plane 1 to plane 0", because that's what the new functions here
lead me to believe. Both of them are only called at one place: the wave
function on the second half of Elis' entrance animation, and the horizontal
masked line function for Reimu's X attack animations.
This function raises one of those essential questions about the eventual ports
we'd like to do. I'll explain everything more thoroughly here, since people
who might complain about the ports not being faithful enough need to
understand this.
----
The original plan was aim for "100% frame-perfect" ports and advertise them as
such. However, the PC-98 is not a console with fixed specs. As the name
implies, it's a computer architecture, and a plethora of different, more and
more powerful PC-98 models were released during its lifespan. Even if we only
consider the subset of products that fulfills the minimum requirements to run
the PC-98 Touhou games, that's still a sizable number of systems.
Therefore, the only true definition of a *frame* can be "everything that is
drawn between two Vsync wait calls". Such a *frame* may contain certain
expensive function calls, and certain systems may run these functions slower
than the developer expected, thus effectively leading to more *frames* than
the developer explicitly specified.
This is one of those functions.
Here, we have a scaling function that appears to be written deliberately to
run very slow, which ends up creating the rolling effect you see in the route
selection and the high score and continue screens of TH01. However, that
doesn't change the fact that the function is still CPU-bound, and neither
waits for Vsync nor is iteratively called by something that does. The faster
your CPU, the faster the rolling effect gets… until ultimately, it's faster
than one frame and therefore vanishes altogether. Mind you, this is true on
both emulators and real hardware. The final PC-98 model, the Ra43, had a CPU
clocked at 433 Mhz, and it may have even been instant there.
If you use more optimized algorithm, it also runs faster on the same CPU (I
tried this, and it worked beautifully)… you get the idea.
Still, it may very well be that this algorithm was not a deliberate choice and
simply resulted from a lack of experience, especially since this was ZUN's
first game.
That leaves us with two approaches to porting functions like these:
1) Look at the recommended system requirements ZUN specified, configure the
PC-98 emulator accordingly, measure how much of the work is done in each
frame, then rewrite the function to be bound to that specific frame rate…
2) …or just continue using a CPU-bound algorithm, which will pretty much
complete instantly on any modern system.
I'd argue that 2) is actually the more "faithful" approach. It will run faster
than the typical clock speeds people emulate the games at, and maybe draw a
bit of criticism because of that, but it seems a lot more rational than the
approximation provided by 1). Not to mention that it's undeniably easier to
implement, and hey, a faster game feels a lot better than a slower one, right?
… Oh well, maybe we'll still encounter some kind of CPU-bound animation that
is so essential to the experience that we do want to lock it to a certain
frame rate…
Fuck TH02 and above and their bizarre assembly code that indeed appears to be,
uh, playfully "optimized" in the most inadequate of places, far away from the
innermost loop. It's ALWAYS just these one or two instructions I just can't
fucking get out of the C compiler, which lead to the conclusion that these
functions must have either been first compiled to assembly, then "fine-tuned"
and then linked into the executable…
… or I'm really just missing some obscure compiler setting.
At least with TH01, you can tell that the source language must have undeniably
been C++, and the decompilation is a breeze.
MAIN.EXE shares most of the code in this segment, but I can't remove it from
there right now due to the weird ordering of the data segments in that
executable…
And yes, once again, those three seemingly random type casts in here are
*necessary* to build a bit-perfect binary.
Small detour into MAINE.EXE because it has all the juicy algorithms that will
explain the remaining unknown members of the highscore data structure, and
there's this one code segment here we need to get out of the way first.
The same function appears unused in TH02's MAINE.EXE. Separate commit because
this was painful enough and we can link the C version into FUUIN.EXE right
now.
Oh, OK, so this is what the PC-98 GRCG is all about. You call grcg_setcolor(),
and that puts the PC-98 hardware in some sort of "monochromatic mode". Then,
you just write your pixels into any *single* one of the 4 VRAM bitplanes. This
causes the hardware to automatically write to *all* bitplanes in such a way
that the final palette index for each of the 8, 16, or 32 pixels you just wrote
a 1 value to will actually end up to match the color you set earlier.
Don't forget to call grcg_off() at the end though, or you can't draw any
non-monochromatic graphics, heh.
Yes, all of it. Including the bouncing polygons, of course. And since it's
placed at the end of ZUN's code inside the executable, the code's already
position-independent and fully hackable.
Well, that became unbearable pretty quickly. Not sure whether I'm doing all
this Makefile business right, but this looks pretty nice.
It doesn't really help much at this point though because the 32-bit part is
still entirely separate and forces everything to rebuild all the time, but at
least it aborts on C compiler errors.
nmlgc