Miri: non-deterministic floating point operations in foreign_items
#143906
Conversation
…ests. atan(+-INF, -INF) is not tested in this commit
rustbot
added
S-waiting-on-review
|
The Miri subtree was changed cc @rust-lang/miri |
|
There are some holes in the behaviour of some operations, because I did not know how I could efficiently handle them, I'll mark them and add some explanation. Also, |
| "atan2f" | "atan2" => val.clamp(pi.neg(), pi), | ||
|
|
||
| // FIXME: According to Wolfram Alpha, the range of ln(gamma) is [-0.121486, +INF]. What to do? | ||
| "lgammaf_r" | "lgamma_r" => val, |
There was a problem hiding this comment.
Do I make a method in IeeeExt trait that gives this exact value, or?
Something like:
fn lgamma_min_output() -> IeeeFloat<S>{ ... }| /// # Note | ||
| /// | ||
| /// For `powif*` operations of the form `(SNaN)^(±0)` we follow the same behaviour as `powf*`, see [`fixed_float_value`]. | ||
| fn fixed_float_int_value<S: Semantics>( |
There was a problem hiding this comment.
ldexp also takes an i32 as the second argument, is this change and name good?
|
|
||
| //atan2(±∞, −∞) = ±3π/4 | ||
| ("atan2f" | "atan2", [x, y]) if x.is_infinite() && y.is_neg_infinity() => | ||
| (pi_over_4 * three).value.copy_sign(*x), |
There was a problem hiding this comment.
I don't know if this will give exactly 3π/4. Should I add a function to IeeeExt for this? These will be hex bytes...
|
|
||
| // atan2(±∞, −∞) = ±3π/4 | ||
| // assert_eq!($float_type::atan2(INFINITY, NEG_INFINITY), 3.0 * FRAC_PI_4, "atan2(+∞, −∞) = 3π/4"); | ||
| // assert_eq!($float_type::atan2(NEG_INFINITY, NEG_INFINITY), -3.0 * FRAC_PI_4, "atan2(-∞, −∞) = -3π/4"); |
There was a problem hiding this comment.
This 3.0 * FRAC_PI_4 will never be exactly equal to the operation mentioned in comment above. Use explicit bytes here as well, or is there a cleaner way?
Tested this repeatedly.
| let pi = IeeeFloat::<S>::pi(); | ||
| let pi_over_2 = IeeeFloat::<S>::frac_pi_2(); | ||
|
|
||
| // Exclusive ranges are made using next_up/next_down. |
There was a problem hiding this comment.
I actually don't know if this is correct, but I assume so :).
|
|
||
| // erfc(+INF) = 0 | ||
| // REVIEW/HELP: for some reason the tests fail because erfc(+INF) = 1e-45. Which is weird because | ||
| // no error can be applied to 0. Checked and this does not happen without Miri. |
There was a problem hiding this comment.
I have no idea why this is, was stuck on this for like half an hour before I gave up.
Part of rust-lang/miri/#3555, this pr does the
foreign_itemswork.Some things have changed since #138062 and #142514. I moved the "helpers" used for creating fixed outputs and clamping operations to their defined ranges to
helpers.rs. These are now also extended to handle the floating-point operations inforeign_items. Tests inmiri/tests/float.rswere changed/added.Failing tests in
stdwere extracted, run under miri with-Zmiri-many-seeds=0..1000and changed accordingly. Double checked with-Zmiri-many-seeds.I noticed that the C standard doesn't specify the output ranges for all of its mathematical operations; it just specifies them as:
So I used Wolfram|Alpha.