This standardizes how max and min subnormals are generated. Since the
new method doesn't use powf, it also enables some of the tests for f128
that were previously disabled due to issues with powf (although it looks
like those issues were already fixed anyway). f16 signalling nan tests
previously disabled are not re-enabled, since the underlying LLVM issue
has not been closed.
Note that the behaviour of the f128 test is slightly changed to use the
same nan mask as is used in test_float_bits_conv, which is the behaviour
used by f16,f32,and f64.
float tests: deduplicate min, max, and rounding tests
Part of https://github.com/rust-lang/rust/issues/141726
Best reviewed commit-by-commit.
- Use `assert_biteq!` in the `mod.rs` tests. This requires some trickery to make shadowing macros with imports work.
- The min, max, minimum, maximum tests in `tests/floats/f*.rs` are entirely subsumed by what we already have in `tests/float/mod.rs`, so I just removed them.
- The rounding tests (floor etc) in `f*.rs` had more test points, so I copied them over. They didn't have `0.5` and `-0.5` though which seem like interesting points in particular regarding the sign of the resulting zero if that's what it sounds to, and they didn't max min/max/inf/nan tests, so this was really a merger of both tests.
r? ``@tgross35``
`assert_eq!` ignores the sign of zero, but for any tests involving zeros
we do care about this sign. Replace `assert_eq!` with `assert_biteq!`
everywhere possible for float tests to ensure we don't miss this.
`assert_biteq!` is also updated to check equality on non-NaNs, to catch
the unlikely case that bitwise equality works but our `==`
implementation is broken.
There is one notable output change: we were asserting that
`(-0.0).fract()` and `(-1.0).fract()` both return -0.0, but both
actually return +0.0.
As was mentioned at [1], we currently use `assert_approx_eq` for testing
some math functions that guarantee exact results. Replace approximate
assertions with exact ones for the following:
* `ceil`
* `floor`
* `fract`
* `from_bits`
* `mul_add`
* `round_ties_even`
* `round`
* `trunc`
This likely wasn't done in the past to avoid writing out exact decimals
that don't match the intuitive answer (e.g. 1.3 - 1.0 = 0.300...004),
but ensuring our results are accurate seems more important here.
[1]: https://github.com/rust-lang/rust/pull/138087#issuecomment-2842069281
When these functions were added in
https://github.com/rust-lang/rust/pull/138087
It made a relatively common pattern for emulating
these functions using an extension trait (which
internally uses `libm`) much more fragile.
If `core::f32` happened to be imported by the user
(to access a constant, say), then that import in
the module namespace would take precedence over
`f32` in the type namespace for resolving these
functions, running headfirst into the stability
attribute.
We ran into this in Color -
https://github.com/linebender/color - and chose to
release the remedial 0.3.1 and 0.2.4, to allow
downstream crates to build on `docs.rs`.
As these methods are perma-unstable, moving them
into a new module should not have any long-term
concerns, and ensures that this breakage doesn't
adversely impact anyone else.
The previous commit moved all test files from `std` to `core` so git
understands the move. Not all functionality is actually testable in
`core`, however, so perform move the relevant portions back. Changes
from inherent to module methods is also done since this is the form of
math operations available in `core` (as `core_float_math`).
Many float-related tests in `std` only depend on `core`, so move the
tests there. This also allows us to verify functions from
`core_float_math`.
Since the majority of test files need to be moved to `coretests`, move
the files here without any cleanup; this is done in a followup commit.
This makes git history slightly cleaner, but coretests will not build
immediately after this commit.
