Besides the immediate performance improvement, reducing
the frequency of downloads should make spurious
failures (network, rate-limiting...) less bothersome.
`FetchContent` respects existing installs, and downloads and compiles the libs
if they aren't found.
This is admittedly a little finicky, since this ignores the usual `Find*`
modules provided by CMake, requiring a bit of glue on our side.
But, one upside is that this moves that logic from our CI into the build system,
which can thus benefit other downstream users.
This also opens the door to some improvements in upcoming commits.
Doing this with Bison turned out to be much more painful, however, due to
`FindBISON` providing the specific `bison_target` command; thus, it remains
installed externally, so that it can be picked up by `FindBISON`.
This also bumps our CMake version requirement slightly, though it's
possible that older versions keep working, or could be supported with small patches;
however, our CI doesn't provide anything below 3.31, so we can't check.
Set the flags in the build itself instead of trying to override the cache
(this was an artifact of when I didn't understand the CMake variable model),
which avoids them being duplicated on each re-configuration.
Also remove `-g`, since it's defined by the build type by default
(we leave it up to the user to choose Debug or RelWithDebInfo).
- `-fsanitize=undefined` encompasses multiple checks we were specifying
- "detect_leaks=1" for `__asan_default_options` checks for memory leaks
(except for with macOS clang++, which does not support LSan)
- `-fsanitize=float-divide-by-zero` is an extra UBSan check
(and reveals a UB bug to fix with fixed-point `DIV` and `LOG`)
Notable side effects:
* Use the standard-conformant MSVC preproc
* Add test for linker script INCLUDE
* Improve wording of placement conflict errors
* Fix errors from not newline-terminated files
* Teach checkdiff about the linker script doc
* Call linker script "commands" "directives" instead
---------
Co-authored-by: Rangi42 <remy.oukaour+rangi42@gmail.com>
Currently missing from the old version:
- `-f` ("fixing" the input image to be indexed)
- `-m` (the code for detecting mirrored tiles is missing, but all of the
"plumbing" is otherwise there)
- `-C`
- `-d`
- `-x` (though I need to check the exact functionality the old one has)
- Also the man page is still a draft and needs to be fleshed out
More planned features are not implemented yet either:
- Explicit palette spec
- Better error messages, also error "images"
- Better 8x16 support, as well as other "dedup unit" sizes
- Support for arbitrary number of palettes & colors per palette
- Other output formats (for example, a "full" palette map for "streaming"
use cases like gb-open-world)
- Quantization?
Some things may also be bugged:
- Transparency support
- Tile offsets (not exposed yet)
- Tile counts per bank (not exposed yet)
...and performance remains to be checked.
We need to set up some tests, honestly.
