mikros/rust
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Auto merge of #2469 - RalfJung:math, r=RalfJung

Browse Source 
implement some missing float functions

With this we support the entire float API surface of the standard library. :)

Also fixes https://github.com/rust-lang/miri/issues/2468 by using host floats to implement FMA.
This commit is contained in:
bors 2022-08-06 15:01:44 +00:00
commit f633537f3e
5 changed files with 133 additions and 57 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
src/shims/foreign_items.rs
View File

@ -575,35 +575,42 @@ fn emulate_foreign_item_by_name(
this.write_scalar(Scalar::from_machine_usize(u64::try_from(n).unwrap(), this), dest)?;
}
// math functions
// math functions (note that there are also intrinsics for some other functions)
#[rustfmt::skip]
| "cbrtf"
| "coshf"
| "sinhf"
| "tanf"
| "tanhf"
| "acosf"
| "asinf"
| "atanf"
| "log1pf"
| "expm1f"
=> {
let [f] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
// FIXME: Using host floats.
let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
let f = match link_name.as_str() {
let res = match link_name.as_str() {
"cbrtf" => f.cbrt(),
"coshf" => f.cosh(),
"sinhf" => f.sinh(),
"tanf" => f.tan(),
"tanhf" => f.tanh(),
"acosf" => f.acos(),
"asinf" => f.asin(),
"atanf" => f.atan(),
"log1pf" => f.ln_1p(),
"expm1f" => f.exp_m1(),
_ => bug!(),
};
this.write_scalar(Scalar::from_u32(f.to_bits()), dest)?;
this.write_scalar(Scalar::from_u32(res.to_bits()), dest)?;
}
#[rustfmt::skip]
| "_hypotf"
| "hypotf"
| "atan2f"
| "fdimf"
=> {
let [f1, f2] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
// underscore case for windows, here and below
@ -611,52 +618,63 @@ fn emulate_foreign_item_by_name(
// FIXME: Using host floats.
let f1 = f32::from_bits(this.read_scalar(f1)?.to_u32()?);
let f2 = f32::from_bits(this.read_scalar(f2)?.to_u32()?);
let n = match link_name.as_str() {
let res = match link_name.as_str() {
"_hypotf" | "hypotf" => f1.hypot(f2),
"atan2f" => f1.atan2(f2),
#[allow(deprecated)]
"fdimf" => f1.abs_sub(f2),
_ => bug!(),
};
this.write_scalar(Scalar::from_u32(n.to_bits()), dest)?;
this.write_scalar(Scalar::from_u32(res.to_bits()), dest)?;
}
#[rustfmt::skip]
| "cbrt"
| "cosh"
| "sinh"
| "tan"
| "tanh"
| "acos"
| "asin"
| "atan"
| "log1p"
| "expm1"
=> {
let [f] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
// FIXME: Using host floats.
let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
let f = match link_name.as_str() {
let res = match link_name.as_str() {
"cbrt" => f.cbrt(),
"cosh" => f.cosh(),
"sinh" => f.sinh(),
"tan" => f.tan(),
"tanh" => f.tanh(),
"acos" => f.acos(),
"asin" => f.asin(),
"atan" => f.atan(),
"log1p" => f.ln_1p(),
"expm1" => f.exp_m1(),
_ => bug!(),
};
this.write_scalar(Scalar::from_u64(f.to_bits()), dest)?;
this.write_scalar(Scalar::from_u64(res.to_bits()), dest)?;
}
#[rustfmt::skip]
| "_hypot"
| "hypot"
| "atan2"
| "fdim"
=> {
let [f1, f2] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
// FIXME: Using host floats.
let f1 = f64::from_bits(this.read_scalar(f1)?.to_u64()?);
let f2 = f64::from_bits(this.read_scalar(f2)?.to_u64()?);
let n = match link_name.as_str() {
let res = match link_name.as_str() {
"_hypot" | "hypot" => f1.hypot(f2),
"atan2" => f1.atan2(f2),
#[allow(deprecated)]
"fdim" => f1.abs_sub(f2),
_ => bug!(),
};
this.write_scalar(Scalar::from_u64(n.to_bits()), dest)?;
this.write_scalar(Scalar::from_u64(res.to_bits()), dest)?;
}
#[rustfmt::skip]
| "_ldexp"
@ -668,7 +686,7 @@ fn emulate_foreign_item_by_name(
let x = this.read_scalar(x)?.to_f64()?;
let exp = this.read_scalar(exp)?.to_i32()?;
// Saturating cast to i16. Even those are outside the valid exponent range to
// Saturating cast to i16. Even those are outside the valid exponent range so
// `scalbn` below will do its over/underflow handling.
let exp = if exp > i32::from(i16::MAX) {
i16::MAX

34
src/shims/intrinsics/mod.rs
View File

@ -285,7 +285,8 @@ fn emulate_intrinsic_by_name(
// FIXME: Using host floats.
let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
let f2 = f32::from_bits(this.read_scalar(f2)?.to_u32()?);
this.write_scalar(Scalar::from_u32(f.powf(f2).to_bits()), dest)?;
let res = f.powf(f2);
this.write_scalar(Scalar::from_u32(res.to_bits()), dest)?;
}
"powf64" => {
@ -293,25 +294,28 @@ fn emulate_intrinsic_by_name(
// FIXME: Using host floats.
let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
let f2 = f64::from_bits(this.read_scalar(f2)?.to_u64()?);
this.write_scalar(Scalar::from_u64(f.powf(f2).to_bits()), dest)?;
let res = f.powf(f2);
this.write_scalar(Scalar::from_u64(res.to_bits()), dest)?;
}
"fmaf32" => {
let [a, b, c] = check_arg_count(args)?;
let a = this.read_scalar(a)?.to_f32()?;
let b = this.read_scalar(b)?.to_f32()?;
let c = this.read_scalar(c)?.to_f32()?;
let res = a.mul_add(b, c).value;
this.write_scalar(Scalar::from_f32(res), dest)?;
// FIXME: Using host floats, to work around https://github.com/rust-lang/miri/issues/2468.
let a = f32::from_bits(this.read_scalar(a)?.to_u32()?);
let b = f32::from_bits(this.read_scalar(b)?.to_u32()?);
let c = f32::from_bits(this.read_scalar(c)?.to_u32()?);
let res = a.mul_add(b, c);
this.write_scalar(Scalar::from_u32(res.to_bits()), dest)?;
}
"fmaf64" => {
let [a, b, c] = check_arg_count(args)?;
let a = this.read_scalar(a)?.to_f64()?;
let b = this.read_scalar(b)?.to_f64()?;
let c = this.read_scalar(c)?.to_f64()?;
let res = a.mul_add(b, c).value;
this.write_scalar(Scalar::from_f64(res), dest)?;
// FIXME: Using host floats, to work around https://github.com/rust-lang/miri/issues/2468.
let a = f64::from_bits(this.read_scalar(a)?.to_u64()?);
let b = f64::from_bits(this.read_scalar(b)?.to_u64()?);
let c = f64::from_bits(this.read_scalar(c)?.to_u64()?);
let res = a.mul_add(b, c);
this.write_scalar(Scalar::from_u64(res.to_bits()), dest)?;
}
"powif32" => {
@ -319,7 +323,8 @@ fn emulate_intrinsic_by_name(
// FIXME: Using host floats.
let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
let i = this.read_scalar(i)?.to_i32()?;
this.write_scalar(Scalar::from_u32(f.powi(i).to_bits()), dest)?;
let res = f.powi(i);
this.write_scalar(Scalar::from_u32(res.to_bits()), dest)?;
}
"powif64" => {
@ -327,7 +332,8 @@ fn emulate_intrinsic_by_name(
// FIXME: Using host floats.
let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
let i = this.read_scalar(i)?.to_i32()?;
this.write_scalar(Scalar::from_u64(f.powi(i).to_bits()), dest)?;
let res = f.powi(i);
this.write_scalar(Scalar::from_u64(res.to_bits()), dest)?;
}
"float_to_int_unchecked" => {

19
src/shims/intrinsics/simd.rs
View File

@ -238,14 +238,25 @@ enum Op {
let dest = this.mplace_index(&dest, i)?;
// Works for f32 and f64.
// FIXME: using host floats to work around https://github.com/rust-lang/miri/issues/2468.
let ty::Float(float_ty) = dest.layout.ty.kind() else {
span_bug!(this.cur_span(), "{} operand is not a float", intrinsic_name)
};
let val = match float_ty {
FloatTy::F32 =>
Scalar::from_f32(a.to_f32()?.mul_add(b.to_f32()?, c.to_f32()?).value),
FloatTy::F64 =>
Scalar::from_f64(a.to_f64()?.mul_add(b.to_f64()?, c.to_f64()?).value),
FloatTy::F32 => {
let a = f32::from_bits(a.to_u32()?);
let b = f32::from_bits(b.to_u32()?);
let c = f32::from_bits(c.to_u32()?);
let res = a.mul_add(b, c);
Scalar::from_u32(res.to_bits())
}
FloatTy::F64 => {
let a = f64::from_bits(a.to_u64()?);
let b = f64::from_bits(b.to_u64()?);
let c = f64::from_bits(c.to_u64()?);
let res = a.mul_add(b, c);
Scalar::from_u64(res.to_bits())
}
};
this.write_scalar(val, &dest.into())?;
}

89
tests/pass/intrinsics-math.rs
View File

@ -32,24 +32,24 @@ pub fn main() {
assert_approx_eq!(25f32.powi(-2), 0.0016f32);
assert_approx_eq!(23.2f64.powi(2), 538.24f64);
assert_approx_eq!(0f32.sin(), 0f32);
assert_approx_eq!((f64::consts::PI / 2f64).sin(), 1f64);
assert_approx_eq!(0f32.cos(), 1f32);
assert_approx_eq!((f64::consts::PI * 2f64).cos(), 1f64);
assert_approx_eq!(25f32.powf(-2f32), 0.0016f32);
assert_approx_eq!(400f64.powf(0.5f64), 20f64);
assert_approx_eq!((1f32.exp() - f32::consts::E).abs(), 0f32);
assert_approx_eq!(1f32.exp(), f32::consts::E);
assert_approx_eq!(1f64.exp(), f64::consts::E);
assert_approx_eq!(1f32.exp_m1(), f32::consts::E - 1.0);
assert_approx_eq!(1f64.exp_m1(), f64::consts::E - 1.0);
assert_approx_eq!(10f32.exp2(), 1024f32);
assert_approx_eq!(50f64.exp2(), 1125899906842624f64);
assert_approx_eq!((f32::consts::E.ln() - 1f32).abs(), 0f32);
assert_approx_eq!(f32::consts::E.ln(), 1f32);
assert_approx_eq!(1f64.ln(), 0f64);
assert_approx_eq!(0f32.ln_1p(), 0f32);
assert_approx_eq!(0f64.ln_1p(), 0f64);
assert_approx_eq!(10f32.log10(), 1f32);
assert_approx_eq!(f64::consts::E.log10(), f64::consts::LOG10_E);
@ -60,10 +60,18 @@ pub fn main() {
assert_eq!(0.0f32.mul_add(-2.0, f32::consts::E), f32::consts::E);
assert_approx_eq!(3.0f64.mul_add(2.0, 5.0), 11.0);
assert_eq!(0.0f64.mul_add(-2.0f64, f64::consts::E), f64::consts::E);
assert_eq!((-3.2f32).mul_add(2.4, f32::NEG_INFINITY), f32::NEG_INFINITY);
assert_eq!((-3.2f64).mul_add(2.4, f64::NEG_INFINITY), f64::NEG_INFINITY);
assert_approx_eq!((-1.0f32).abs(), 1.0f32);
assert_approx_eq!(34.2f64.abs(), 34.2f64);
#[allow(deprecated)]
{
assert_approx_eq!(5.0f32.abs_sub(3.0), 2.0);
assert_approx_eq!(3.0f64.abs_sub(5.0), 0.0);
}
assert_approx_eq!(3.8f32.floor(), 3.0f32);
assert_approx_eq!((-1.1f64).floor(), -2.0f64);
@ -79,31 +87,54 @@ pub fn main() {
assert_approx_eq!(3.0f32.hypot(4.0f32), 5.0f32);
assert_approx_eq!(3.0f64.hypot(4.0f64), 5.0f64);
assert_approx_eq!(1.0f32.atan2(2.0f32), 0.46364761f32);
assert_approx_eq!(1.0f32.atan2(2.0f32), 0.46364761f32);
assert_approx_eq!(1.0f32.cosh(), 1.54308f32);
assert_approx_eq!(1.0f64.cosh(), 1.54308f64);
assert_approx_eq!(1.0f32.sinh(), 1.1752012f32);
assert_approx_eq!(1.0f64.sinh(), 1.1752012f64);
assert_approx_eq!(1.0f32.tan(), 1.557408f32);
assert_approx_eq!(1.0f64.tan(), 1.557408f64);
assert_approx_eq!(f32::consts::FRAC_PI_4.cos().acos(), f32::consts::FRAC_PI_4);
assert_approx_eq!(f64::consts::FRAC_PI_4.cos().acos(), f64::consts::FRAC_PI_4);
assert_approx_eq!(f32::consts::FRAC_PI_4.sin().asin(), f32::consts::FRAC_PI_4);
assert_approx_eq!(f64::consts::FRAC_PI_4.sin().asin(), f64::consts::FRAC_PI_4);
assert_approx_eq!(1.0_f32, 1.0_f32.tan().atan());
assert_approx_eq!(1.0_f64, 1.0_f64.tan().atan());
assert_eq!(3.3_f32.round(), 3.0);
assert_eq!(3.3_f64.round(), 3.0);
assert_eq!(ldexp(0.65f64, 3i32), 5.2f64);
assert_eq!(ldexp(1.42, 0xFFFF), f64::INFINITY);
assert_eq!(ldexp(1.42, -0xFFFF), 0f64);
// Trigonometric functions.
assert_approx_eq!(0f32.sin(), 0f32);
assert_approx_eq!((f64::consts::PI / 2f64).sin(), 1f64);
assert_approx_eq!(f32::consts::FRAC_PI_6.sin(), 0.5);
assert_approx_eq!(f64::consts::FRAC_PI_6.sin(), 0.5);
assert_approx_eq!(f32::consts::FRAC_PI_4.sin().asin(), f32::consts::FRAC_PI_4);
assert_approx_eq!(f64::consts::FRAC_PI_4.sin().asin(), f64::consts::FRAC_PI_4);
assert_approx_eq!(1.0f32.sinh(), 1.1752012f32);
assert_approx_eq!(1.0f64.sinh(), 1.1752012f64);
assert_approx_eq!(2.0f32.asinh(), 1.443635475178810342493276740273105f32);
assert_approx_eq!((-2.0f64).asinh(), -1.443635475178810342493276740273105f64);
assert_approx_eq!(0f32.cos(), 1f32);
assert_approx_eq!((f64::consts::PI * 2f64).cos(), 1f64);
assert_approx_eq!(f32::consts::FRAC_PI_3.cos(), 0.5);
assert_approx_eq!(f64::consts::FRAC_PI_3.cos(), 0.5);
assert_approx_eq!(f32::consts::FRAC_PI_4.cos().acos(), f32::consts::FRAC_PI_4);
assert_approx_eq!(f64::consts::FRAC_PI_4.cos().acos(), f64::consts::FRAC_PI_4);
assert_approx_eq!(1.0f32.cosh(), 1.54308f32);
assert_approx_eq!(1.0f64.cosh(), 1.54308f64);
assert_approx_eq!(2.0f32.acosh(), 1.31695789692481670862504634730796844f32);
assert_approx_eq!(3.0f64.acosh(), 1.76274717403908605046521864995958461f64);
assert_approx_eq!(1.0f32.tan(), 1.557408f32);
assert_approx_eq!(1.0f64.tan(), 1.557408f64);
assert_approx_eq!(1.0_f32, 1.0_f32.tan().atan());
assert_approx_eq!(1.0_f64, 1.0_f64.tan().atan());
assert_approx_eq!(1.0f32.atan2(2.0f32), 0.46364761f32);
assert_approx_eq!(1.0f32.atan2(2.0f32), 0.46364761f32);
assert_approx_eq!(
1.0f32.tanh(),
(1.0 - f32::consts::E.powi(-2)) / (1.0 + f32::consts::E.powi(-2))
);
assert_approx_eq!(
1.0f64.tanh(),
(1.0 - f64::consts::E.powi(-2)) / (1.0 + f64::consts::E.powi(-2))
);
assert_approx_eq!(0.5f32.atanh(), 0.54930614433405484569762261846126285f32);
assert_approx_eq!(0.5f64.atanh(), 0.54930614433405484569762261846126285f64);
}

10
tests/pass/portable-simd.rs
View File

@ -18,6 +18,11 @@ fn simd_ops_f32() {
assert_eq!(a.mul_add(b, a), (a * b) + a);
assert_eq!(b.mul_add(b, a), (b * b) + a);
assert_eq!(a.mul_add(b, b), (a * b) + b);
assert_eq!(
f32x4::splat(-3.2).mul_add(b, f32x4::splat(f32::NEG_INFINITY)),
f32x4::splat(f32::NEG_INFINITY)
);
assert_eq!((a * a).sqrt(), a);
assert_eq!((b * b).sqrt(), b.abs());
@ -67,6 +72,11 @@ fn simd_ops_f64() {
assert_eq!(a.mul_add(b, a), (a * b) + a);
assert_eq!(b.mul_add(b, a), (b * b) + a);
assert_eq!(a.mul_add(b, b), (a * b) + b);
assert_eq!(
f64x4::splat(-3.2).mul_add(b, f64x4::splat(f64::NEG_INFINITY)),
f64x4::splat(f64::NEG_INFINITY)
);
assert_eq!((a * a).sqrt(), a);
assert_eq!((b * b).sqrt(), b.abs());