As noted in #567, a filesystem path that contains non-ascii will break a
lot if using a clang toolchain.
Luckily, fmtlib has a lossy utf8 convert when you use it to print a path
(after including `fmt/std.h`). The vast majority of this diff is just
removing `.string()` from paths inside loggings calls.
There are some callsites I _didn't_ touch, mainly the options, because
it would be an ABI break to change those to be wide strings and I cbf
looking into settings upgrades. There are also some spots (avs mountpath
remapping, for example) where the path is guaranteed to be ascii, so I
didn't modify them.
ImGui doesn't appear to easily support wide strings (I mean, surely it
does, but I'm not gonna look too far into it) so I mostly just left
those alone too, with a few spots modified to re-use fmtlib's lossy
utf8.
Some of the changes are basically never gonna be hit IRL, like who would
put a file with a non-ascii _extension_ along with their modules? But
the diff is (I hope) pretty easy to validate as OK.
Testing has been somewhat minimal, I fired up the GCC build of spice2x
in a dodgy folder name, got mojibake (running via wine in linux so take
that as you will), ran the unmodified clang spice and crashed the same
way the reporter did. After modification, I get the exact same mojibake
so I assume if the terminal enjoys utf8 it'll display OK.
Claude (only used for review) thinks the commit is fine but is annoyed
that I use `fmt::detail` in the appdata censoring, which is part of the
private API; personally I don't care because it's pretty stable.
This was a lot more background work than expected, but that is the
nature of unnecessary hobbies I guess. Most things in here are obvious,
so here's the non-obvious thoughts and notes:
1. All the XP stuff is optional, if the toolchain isn't available it
won't even try
2. The docker build is the "known good" build and now has all the
required tools for XP and checking binaries
a. If you're OK with it, can add the XP_MUST_BUILD flag to
build_docker.sh
4. New `SPICE_XP` flag to comment out blocks
5. We could technically build using the Docker image I've made for my XP
compatible LLVM toolchain, but combining the binaries into a single zip
at the end of it is somewhat tedious, so I've chosen to grab the
binaries from my release
6. I also wrote a DLL checker to make sure the final binaries are
_actually_ XP compatible, which found some issues on x86_64 (I don't
know a single game that runs on 64 bit XP, but it wasn't too much effort
so whatever). This also won't run if missing, just to save the trouble.
7. The build script is still pretty shit, I've made some minor
improvements
a. `/usr/bin/env bash` instead of `/bin/bash` as it's more correct
b. Adding `-it` to the docker build, so Ctrl+C actually halts it
c. Main issue is that failures don't halt, because of the custom error
catching function. This has caused a lot of frustration, maybe we can
change that
## Testing
It runs on 32 and 64 bit installs of XP. My jubeat cab is on Win10 these
days and I haven't got around to testing an actual game, but I've
distributed test builds of this to people who *have* confirmed it works
as expected.
Logic was wrong; caught by compiling with a newer clang and it raising a
warning about the useless expression (since the second clause always
eval'd as true).
Changed to the intended functionality, but I have no ddr3 with p3io so I
can't test it. But it seems nobody ever really noticed this.
- Allow overriding the toolchains so I can avoid changes in my working
tree
- Squash warnings that the deps use, can't just add them as PRIVATE
compile flags because it's in the headers :(
- Squash those d3d9 warnings with a suppression instead of trying to
rewrite 400 lines of code
- Include order and forward decls were breaking button.h because a
declared but not defined struct can't be initiated (honestly this file
was cooked, crazy include order).
- `MAXINT` is GCC specific, `INT_MAX` is portable.
- InterlockedDecrement takes a LONG, not ULONG
- an imgui printf-style call using a direct string instead of %s (let's
ignore the fact that it's actually safe based on how the str is
constructed)
- missing `override` on a virtual subclass
- `CALLBACK` on a lambda threw me for a loop, but I believe moving it
after the arg list is the correct approach (see if it builds on gcc I
guess)
- `std::result_of` was removed in C++20, which the project is built with
Doing dumb things and building inside a path that contains spaces.
Quoted the relevant bits in the build script (and a couple more
shellcheck brought up that realistically won't ever be an issue, but the
squigglies annoy me.
WillandGitHub