The generated `result.pal` can change between platforms, it turns out
(due to "unstable sort" differences), and creates additional churn if
changing anything related to palette generation.
tl;dr this makes the test less flaky in the long run.
- FIX: `Label & const` was not actually doing the `& const` masking
(fixes#1446)
- ADD: `LOW(Label)` can be constant if `Label` is aligned to 8 or more bits
(resolves#1444)
- ADD: `!expr` can be constant 0 if `expr` has any non-zero bits
(resolves#1447)
- `LOW()` and `HIGH()` have their own RPN operator values
(resolves#1445)
The change to RPN values means that the object file version was incremented.
This also refactors unary operators and functions, combining their
evaluation similarly to binary ones.
- Fixed-point formulas are implemented using IEEE-754 floating-point
internally, which could give infinity or NaN values whose conversion
to fixed-point integer was platform-dependent.
- Formatting fixed-point $8000_0000 (INT32_MIN, -2147483648) was
not putting the negative sign in front.
The Python Pillow image library gives different output for
`Image.getcolors()` in version 9 than version 10. This causes
libbet's SGB border to be built differently, giving a different
ROM hash and technically failing the test script.
This patches libbet to only keep the known-good colors. It also
adapts the coverage script to run all the tests, including
building the third-party projects.
This way, if a child context initializes `\@`, the parent won't
reset it. And if the child context did not initialize `\@`,
then resetting it would be redundant.
This allows us to control the order in which sections are iterated,
instead of it depending on the internals of `std::map`. (This order
is arbitrary, but should be deterministic regardless.)
This still doesn't allow identical exported label constants.
That can be addressed when or if it's requested for a real use case.
Symbols only store one source filename + line number, so this
arbitrarily keeps the last read symbol as the reported one.
