Enable Non-determinism of float operations in Miri and change std tests
Links to [#4208](https://github.com/rust-lang/miri/issues/4208) and [#3555](https://github.com/rust-lang/miri/issues/3555) in Miri.
Non-determinism of floating point operations was disabled in rust-lang/rust#137594 because it breaks the tests and doc-tests in core/coretests and std. This PR enables some of them.
This pr includes the following changes:
- Enables the float non-determinism but with a lower relative error of 4ULP instead of 16ULP
- These operations now have a fixed output based on the C23 standard, except the pow operations, this is tracked in [#4286](https://github.com/rust-lang/miri/issues/4286#issue-3010677983)
- Changes tests that made incorrect assumptions about the operations, not to make that assumption anymore (from `assert_eq!` to `assert_approx_eq!`.
- Changed the doctests of the stdlib of these operations to compare against fixed constants instead of `f*::EPSILON`, which now succeed with Miri and `-Zmiri-many-seeds`
- Added a constant `APPROX_DELTA` in `std/tests/floats/f32.rs` which is used for approximation tests, but with a different value when run in Miri. This is to make these tests succeed.
- Added tests in the float tests of Miri to test the C23 behaviour.
Fixes https://github.com/rust-lang/miri/issues/4208
Disable f64 minimum/maximum tests for arm 32
This disables the f64 minimum/maximum tests for the arm-unknown-linux-gnueabihf job. The next release will be supporting cross-compiled doctests, and these tests fail on that platform.
It looks like this was just fixed via https://github.com/llvm/llvm-project/pull/142170, but I assume that will not trickle down to our copy of llvm in the next couple of weeks. Assuming that does get fixed when llvm is updated, then these can be removed.
cc https://github.com/rust-lang/rust/issues/141087
This disables the f64 minimum/maximum tests for the
arm-unknown-linux-gnueabihf job. The next release will be supporting
cross-compiled doctests, and these tests fail on that platform.
It looks like this was just fixed via
https://github.com/llvm/llvm-project/pull/142170, but I assume that will
not trickle down to our copy of llvm in the next couple of weeks.
Assuming that does get fixed when llvm is updated, then these can be
removed.
cc https://github.com/rust-lang/rust/issues/141087
Add const support for the float rounding methods floor, ceil, trunc,
fract, round and round_ties_even.
This works by moving the calculation logic from
src/tools/miri/src/intrinsics/mod.rs
into
compiler/rustc_const_eval/src/interpret/intrinsics.rs.
All relevant method definitions were adjusted to include the `const`
keyword for all supported float types: f16, f32, f64 and f128.
The constness is hidden behind the feature gate
feature(const_float_round_methods)
which is tracked in
https://github.com/rust-lang/rust/issues/141555
This commit is a squash of the following commits:
- test: add tests that we expect to pass when float rounding becomes const
- feat: make float rounding methods `const`
- fix: replace `rustc_allow_const_fn_unstable(core_intrinsics)` attribute with `#[rustc_const_unstable(feature = "f128", issue = "116909")]` in `library/core/src/num/f128.rs`
- revert: undo update to `library/stdarch`
- refactor: replace multiple `float_<mode>_intrinsic` rounding methods with a single, parametrized one
- fix: add `#[cfg(not(bootstrap))]` to new const method tests
- test: add extra sign tests to check `+0.0` and `-0.0`
- revert: undo accidental changes to `round` docs
- fix: gate `const` float round method behind `const_float_round_methods`
- fix: remove unnecessary `#![feature(const_float_methods)]`
- fix: remove unnecessary `#![feature(const_float_methods)]` [2]
- revert: undo changes to `tests/ui/consts/const-eval/float_methods.rs`
- fix: adjust after rebase
- test: fix float tests
- test: add tests for `fract`
- chore: add commented-out `const_float_round_methods` feature gates to `f16` and `f128`
- fix: adjust NaN when rounding floats
- chore: add FIXME comment for de-duplicating float tests
- test: remove unnecessary test file `tests/ui/consts/const-eval/float_methods.rs`
- test: fix tests after upstream simplification of how float tests are run
Implement ((un)checked_)exact_div methods for integers
tracking issue: #139911
I see that there might still be some bikeshedding to be done, so if people want changes to this implementation, I'm happy to make those. I did also see that there was a previous attempt at this PR (#116632), but I'm not sure why it got closed.
This patch enables the optimized implementation of `f32::midpoint` for
`loongarch64` targets that support the `d`feature. Targets with reliable
64-bit float support can safely use the faster and more accurate computation
via `f64`, avoiding the fallback branchy version.
Add some track_caller info to precondition panics
Currently, when you encounter a precondition check, you'll always get the caller location of the implementation of the precondition checks. But with this PR, you'll be told the location of the invalid call. Which is useful.
I thought of this while looking at https://github.com/rust-lang/rust/pull/129642#issuecomment-2311703898.
The changes to `tests/ui/const*` happen because the const-eval interpreter skips `#[track_caller]` frames in its backtraces.
The perf implications of this are:
* Increased debug binary sizes. The caller_location implementation requires that the additional data we want to display here be stored in const allocations, which are deduplicated but not across crates. There is no impact on optimized build sizes. The panic path and the caller location data get optimized out.
* The compile time hit to opt-incr-patched bitmaps happens because the patch changes the line number of some function calls with precondition checks, causing us to go from 0 dirty CGUs to 1 dirty CGU.
* The other compile time hits are marginal but real, and due to doing a handful of new queries. Adding more useful data isn't completely free.
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.
Use the existing Lemire (decimal -> float) and Dragon / Grisu algorithms
(float -> decimal) to add support for `f16`. This allows updating the
implementation for `Display` to the expected behavior for `Display`
(currently it prints the a hex bitwise representation), matching other
floats, and adds a `FromStr` implementation.
In order to avoid crashes when compiling with Cranelift or on targets
where f16 is not well supported, a fallback is used if
`cfg(target_has_reliable_f16)` is not true.
Initial implementation of `core_float_math`
Since [1], `compiler-builtins` makes a certain set of math symbols
weakly available on all platforms. This means we can begin exposing some
of the related functions in `core`, so begin this process here.
It is not possible to provide inherent methods in both `core` and `std`
while giving them different stability gates, so standalone functions are
added instead. This provides a way to experiment with the functionality
while unstable; once it is time to stabilize, they can be converted to
inherent.
For `f16` and `f128`, everything is unstable so we can move the inherent
methods.
The following are included to start:
* floor
* ceil
* round
* round_ties_even
* trunc
* fract
* mul_add
* div_euclid
* rem_euclid
* powi
* sqrt
* abs_sub
* cbrt
These mirror the set of functions that we have in `compiler-builtins`
since [1], with the exception of `powi` that has been there longer.
Details for each of the changes is in the commit messages.
Tracking issue: https://github.com/rust-lang/rust/issues/137578
[1]: https://github.com/rust-lang/compiler-builtins/pull/763
try-job: aarch64-gnu
tru-job: armhf-gnu
try-job: i686-msvc-1
try-job: test-various
try-job: x86_64-mingw-1
try-job: x86_64-mingw-2
Since [1], `compiler-builtins` makes a certain set of math symbols
weakly available on all platforms. This means we can begin exposing some
of the related functions in `core`, so begin this process here.
It is not possible to provide inherent methods in both `core` and `std`
while giving them different stability gates, so standalone functions are
added instead. This provides a way to experiment with the functionality
while unstable; once it is time to stabilize, they can be converted to
inherent.
For `f16` and `f128`, everything is unstable so we can move the inherent
methods.
The following are included to start:
* floor
* ceil
* round
* round_ties_even
* trunc
* fract
* mul_add
* div_euclid
* rem_euclid
* powi
* sqrt
* abs_sub
* cbrt
These mirror the set of functions that we have in `compiler-builtins`
since [1].
Tracking issue: https://github.com/rust-lang/rust/issues/137578
[1]: https://github.com/rust-lang/compiler-builtins/pull/763
Refactor `diy_float`
The refactor replaces bespoke algorithms with functions already inside the standard library, improving both codegen and readability.
Suggest {to,from}_ne_bytes for transmutations between arrays and integers, etc
implements #136067
Rust has helper methods for many kinds of safe transmutes, for example integer<->bytes. This is a lint against using transmute for these cases.
```rs
fn bytes_at_home(x: [u8; 4]) -> u32 {
transmute(x)
}
// other examples
transmute::<[u8; 2], u16>();
transmute::<[u8; 8], f64>();
transmute::<u32, [u8; 4]>();
transmute::<char, u32>();
transmute::<u32, char>();
```
It would be handy to suggest `u32::from_ne_bytes(x)`.
This is implemented for `[u8; _]` -> `{float int}`
This also implements the cases:
`fXX` <-> `uXX` = `{from_bits, to_bits}`
`uXX` -> `iXX` via `cast_unsigned` and `cast_signed`
{`char` -> `u32`, `bool` -> `n8`} via `from`
`u32` -> `char` via `from_u32_unchecked` (note: notes `from_u32().unwrap()`) (contested)
`u8` -> `bool` via `==` (debatable)
---
try-job: aarch64-gnu
try-job: test-various
f*::NAN: guarantee that this is a quiet NaN
I think we should guarantee that this is a quiet NaN. This then implies that programs not using `f*::from_bits` (or unsafe type conversions) are guaranteed to only work with quiet NaNs. It would be awkward if people start to write `0.0 / 0.0` instead of using the constant just because they want to get a guaranteed-quiet NaN.
This is a `@rust-lang/libs-api` change. The definition of this constant currently is `0.0 / 0.0`, which is already guaranteed to be a quiet NaN. So all this does is forward that guarantee to our users.
