Eliminate the use of force_eval! in ceil, floor, and trunc

This commit is contained in:
Trevor Gross 2025-02-10 12:01:16 +00:00
parent 105cd79578
commit e94e987399
3 changed files with 220 additions and 37 deletions

View file

@ -7,9 +7,14 @@
//! performance seems to be better (based on icount) and it does not seem to experience rounding
//! errors on i386.
use super::super::support::{FpResult, Status};
use super::super::{Float, Int, IntTy, MinInt};
pub fn ceil<F: Float>(x: F) -> F {
ceil_status(x).val
}
pub fn ceil_status<F: Float>(x: F) -> FpResult<F> {
let zero = IntTy::<F>::ZERO;
let mut ix = x.to_bits();
@ -17,20 +22,20 @@ pub fn ceil<F: Float>(x: F) -> F {
// If the represented value has no fractional part, no truncation is needed.
if e >= F::SIG_BITS as i32 {
return x;
return FpResult::ok(x);
}
if e >= 0 {
let status;
let res = if e >= 0 {
// |x| >= 1.0
let m = F::SIG_MASK >> e.unsigned();
if (ix & m) == zero {
// Portion to be masked is already zero; no adjustment needed.
return x;
return FpResult::ok(x);
}
// Otherwise, raise an inexact exception.
force_eval!(x + F::MAX);
status = Status::INEXACT;
if x.is_sign_positive() {
ix += m;
@ -40,7 +45,11 @@ pub fn ceil<F: Float>(x: F) -> F {
F::from_bits(ix)
} else {
// |x| < 1.0, raise an inexact exception since truncation will happen (unless x == 0).
force_eval!(x + F::MAX);
if ix & F::SIG_MASK == F::Int::ZERO {
status = Status::OK;
} else {
status = Status::INEXACT;
}
if x.is_sign_negative() {
// -1.0 < x <= -0.0; rounding up goes toward -0.0.
@ -52,18 +61,30 @@ pub fn ceil<F: Float>(x: F) -> F {
// +0.0 remains unchanged
x
}
}
};
FpResult::new(res, status)
}
#[cfg(test)]
mod tests {
use super::*;
use crate::support::Hexf;
/// Test against https://en.cppreference.com/w/cpp/numeric/math/ceil
fn spec_test<F: Float>() {
// Not Asserted: that the current rounding mode has no effect.
for f in [F::ZERO, F::NEG_ZERO, F::INFINITY, F::NEG_INFINITY].iter().copied() {
assert_biteq!(ceil(f), f);
fn spec_test<F: Float>(cases: &[(F, F, Status)]) {
let roundtrip = [F::ZERO, F::ONE, F::NEG_ONE, F::NEG_ZERO, F::INFINITY, F::NEG_INFINITY];
for x in roundtrip {
let FpResult { val, status } = ceil_status(x);
assert_biteq!(val, x, "{}", Hexf(x));
assert_eq!(status, Status::OK, "{}", Hexf(x));
}
for &(x, res, res_stat) in cases {
let FpResult { val, status } = ceil_status(x);
assert_biteq!(val, res, "{}", Hexf(x));
assert_eq!(status, res_stat, "{}", Hexf(x));
}
}
@ -72,7 +93,17 @@ mod tests {
#[test]
#[cfg(f16_enabled)]
fn spec_tests_f16() {
spec_test::<f16>();
let cases = [
(0.1, 1.0, Status::INEXACT),
(-0.1, -0.0, Status::INEXACT),
(0.9, 1.0, Status::INEXACT),
(-0.9, -0.0, Status::INEXACT),
(1.1, 2.0, Status::INEXACT),
(-1.1, -1.0, Status::INEXACT),
(1.9, 2.0, Status::INEXACT),
(-1.9, -1.0, Status::INEXACT),
];
spec_test::<f16>(&cases);
}
#[test]
@ -83,7 +114,17 @@ mod tests {
#[test]
fn spec_tests_f32() {
spec_test::<f32>();
let cases = [
(0.1, 1.0, Status::INEXACT),
(-0.1, -0.0, Status::INEXACT),
(0.9, 1.0, Status::INEXACT),
(-0.9, -0.0, Status::INEXACT),
(1.1, 2.0, Status::INEXACT),
(-1.1, -1.0, Status::INEXACT),
(1.9, 2.0, Status::INEXACT),
(-1.9, -1.0, Status::INEXACT),
];
spec_test::<f32>(&cases);
}
#[test]
@ -94,12 +135,32 @@ mod tests {
#[test]
fn spec_tests_f64() {
spec_test::<f64>();
let cases = [
(0.1, 1.0, Status::INEXACT),
(-0.1, -0.0, Status::INEXACT),
(0.9, 1.0, Status::INEXACT),
(-0.9, -0.0, Status::INEXACT),
(1.1, 2.0, Status::INEXACT),
(-1.1, -1.0, Status::INEXACT),
(1.9, 2.0, Status::INEXACT),
(-1.9, -1.0, Status::INEXACT),
];
spec_test::<f64>(&cases);
}
#[test]
#[cfg(f128_enabled)]
fn spec_tests_f128() {
spec_test::<f128>();
let cases = [
(0.1, 1.0, Status::INEXACT),
(-0.1, -0.0, Status::INEXACT),
(0.9, 1.0, Status::INEXACT),
(-0.9, -0.0, Status::INEXACT),
(1.1, 2.0, Status::INEXACT),
(-1.1, -1.0, Status::INEXACT),
(1.9, 2.0, Status::INEXACT),
(-1.9, -1.0, Status::INEXACT),
];
spec_test::<f128>(&cases);
}
}

View file

@ -7,9 +7,14 @@
//! performance seems to be better (based on icount) and it does not seem to experience rounding
//! errors on i386.
use super::super::support::{FpResult, Status};
use super::super::{Float, Int, IntTy, MinInt};
pub fn floor<F: Float>(x: F) -> F {
floor_status(x).val
}
pub fn floor_status<F: Float>(x: F) -> FpResult<F> {
let zero = IntTy::<F>::ZERO;
let mut ix = x.to_bits();
@ -17,20 +22,20 @@ pub fn floor<F: Float>(x: F) -> F {
// If the represented value has no fractional part, no truncation is needed.
if e >= F::SIG_BITS as i32 {
return x;
return FpResult::ok(x);
}
if e >= 0 {
let status;
let res = if e >= 0 {
// |x| >= 1.0
let m = F::SIG_MASK >> e.unsigned();
if ix & m == zero {
// Portion to be masked is already zero; no adjustment needed.
return x;
return FpResult::ok(x);
}
// Otherwise, raise an inexact exception.
force_eval!(x + F::MAX);
status = Status::INEXACT;
if x.is_sign_negative() {
ix += m;
@ -39,8 +44,12 @@ pub fn floor<F: Float>(x: F) -> F {
ix &= !m;
F::from_bits(ix)
} else {
// |x| < 1.0, raise an inexact exception since truncation will happen (unless x == 0).
force_eval!(x + F::MAX);
// |x| < 1.0, raise an inexact exception since truncation will happen.
if ix & F::SIG_MASK == F::Int::ZERO {
status = Status::OK;
} else {
status = Status::INEXACT;
}
if x.is_sign_positive() {
// 0.0 <= x < 1.0; rounding down goes toward +0.0.
@ -52,27 +61,40 @@ pub fn floor<F: Float>(x: F) -> F {
// -0.0 remains unchanged
x
}
}
};
FpResult::new(res, status)
}
#[cfg(test)]
mod tests {
use super::*;
use crate::support::Hexf;
/// Test against https://en.cppreference.com/w/cpp/numeric/math/floor
fn spec_test<F: Float>() {
// Not Asserted: that the current rounding mode has no effect.
for f in [F::ZERO, F::NEG_ZERO, F::INFINITY, F::NEG_INFINITY].iter().copied() {
assert_biteq!(floor(f), f);
fn spec_test<F: Float>(cases: &[(F, F, Status)]) {
let roundtrip = [F::ZERO, F::ONE, F::NEG_ONE, F::NEG_ZERO, F::INFINITY, F::NEG_INFINITY];
for x in roundtrip {
let FpResult { val, status } = floor_status(x);
assert_biteq!(val, x, "{}", Hexf(x));
assert_eq!(status, Status::OK, "{}", Hexf(x));
}
for &(x, res, res_stat) in cases {
let FpResult { val, status } = floor_status(x);
assert_biteq!(val, res, "{}", Hexf(x));
assert_eq!(status, res_stat, "{}", Hexf(x));
}
}
/* Skipping f16 / f128 "sanity_check"s due to rejected literal lexing at MSRV */
/* Skipping f16 / f128 "sanity_check"s and spec cases due to rejected literal lexing at MSRV */
#[test]
#[cfg(f16_enabled)]
fn spec_tests_f16() {
spec_test::<f16>();
let cases = [];
spec_test::<f16>(&cases);
}
#[test]
@ -84,7 +106,17 @@ mod tests {
#[test]
fn spec_tests_f32() {
spec_test::<f32>();
let cases = [
(0.1, 0.0, Status::INEXACT),
(-0.1, -1.0, Status::INEXACT),
(0.9, 0.0, Status::INEXACT),
(-0.9, -1.0, Status::INEXACT),
(1.1, 1.0, Status::INEXACT),
(-1.1, -2.0, Status::INEXACT),
(1.9, 1.0, Status::INEXACT),
(-1.9, -2.0, Status::INEXACT),
];
spec_test::<f32>(&cases);
}
#[test]
@ -95,12 +127,23 @@ mod tests {
#[test]
fn spec_tests_f64() {
spec_test::<f64>();
let cases = [
(0.1, 0.0, Status::INEXACT),
(-0.1, -1.0, Status::INEXACT),
(0.9, 0.0, Status::INEXACT),
(-0.9, -1.0, Status::INEXACT),
(1.1, 1.0, Status::INEXACT),
(-1.1, -2.0, Status::INEXACT),
(1.9, 1.0, Status::INEXACT),
(-1.9, -2.0, Status::INEXACT),
];
spec_test::<f64>(&cases);
}
#[test]
#[cfg(f128_enabled)]
fn spec_tests_f128() {
spec_test::<f128>();
let cases = [];
spec_test::<f128>(&cases);
}
}

View file

@ -1,15 +1,20 @@
/* SPDX-License-Identifier: MIT
* origin: musl src/math/trunc.c */
use super::super::support::{FpResult, Status};
use super::super::{Float, Int, IntTy, MinInt};
pub fn trunc<F: Float>(x: F) -> F {
trunc_status(x).val
}
pub fn trunc_status<F: Float>(x: F) -> FpResult<F> {
let mut xi: F::Int = x.to_bits();
let e: i32 = x.exp_unbiased();
// C1: The represented value has no fractional part, so no truncation is needed
if e >= F::SIG_BITS as i32 {
return x;
return FpResult::ok(x);
}
let mask = if e < 0 {
@ -23,22 +28,68 @@ pub fn trunc<F: Float>(x: F) -> F {
// C4: If the to-be-masked-out portion is already zero, we have an exact result
if (xi & !mask) == IntTy::<F>::ZERO {
return x;
return FpResult::ok(x);
}
// C5: Otherwise the result is inexact and we will truncate. Raise `FE_INEXACT`, mask the
// result, and return.
force_eval!(x + F::MAX);
let status = if xi & F::SIG_MASK == F::Int::ZERO { Status::OK } else { Status::INEXACT };
xi &= mask;
F::from_bits(xi)
FpResult::new(F::from_bits(xi), status)
}
#[cfg(test)]
mod tests {
use super::*;
use crate::support::Hexf;
fn spec_test<F: Float>(cases: &[(F, F, Status)]) {
let roundtrip = [F::ZERO, F::ONE, F::NEG_ONE, F::NEG_ZERO, F::INFINITY, F::NEG_INFINITY];
for x in roundtrip {
let FpResult { val, status } = trunc_status(x);
assert_biteq!(val, x, "{}", Hexf(x));
assert_eq!(status, Status::OK, "{}", Hexf(x));
}
for &(x, res, res_stat) in cases {
let FpResult { val, status } = trunc_status(x);
assert_biteq!(val, res, "{}", Hexf(x));
assert_eq!(status, res_stat, "{}", Hexf(x));
}
}
/* Skipping f16 / f128 "sanity_check"s and spec cases due to rejected literal lexing at MSRV */
#[test]
fn sanity_check() {
#[cfg(f16_enabled)]
fn spec_tests_f16() {
let cases = [];
spec_test::<f16>(&cases);
}
#[test]
fn sanity_check_f32() {
assert_eq!(trunc(0.5f32), 0.0);
assert_eq!(trunc(1.1f32), 1.0);
assert_eq!(trunc(2.9f32), 2.0);
}
#[test]
fn spec_tests_f32() {
let cases = [
(0.1, 0.0, Status::INEXACT),
(-0.1, -0.0, Status::INEXACT),
(0.9, 0.0, Status::INEXACT),
(-0.9, -0.0, Status::INEXACT),
(1.1, 1.0, Status::INEXACT),
(-1.1, -1.0, Status::INEXACT),
(1.9, 1.0, Status::INEXACT),
(-1.9, -1.0, Status::INEXACT),
];
spec_test::<f32>(&cases);
assert_biteq!(trunc(1.1f32), 1.0);
assert_biteq!(trunc(1.1f64), 1.0);
@ -54,4 +105,32 @@ mod tests {
assert_biteq!(trunc(hf32!("-0x1p-1")), -0.0);
assert_biteq!(trunc(hf64!("-0x1p-1")), -0.0);
}
#[test]
fn sanity_check_f64() {
assert_eq!(trunc(1.1f64), 1.0);
assert_eq!(trunc(2.9f64), 2.0);
}
#[test]
fn spec_tests_f64() {
let cases = [
(0.1, 0.0, Status::INEXACT),
(-0.1, -0.0, Status::INEXACT),
(0.9, 0.0, Status::INEXACT),
(-0.9, -0.0, Status::INEXACT),
(1.1, 1.0, Status::INEXACT),
(-1.1, -1.0, Status::INEXACT),
(1.9, 1.0, Status::INEXACT),
(-1.9, -1.0, Status::INEXACT),
];
spec_test::<f64>(&cases);
}
#[test]
#[cfg(f128_enabled)]
fn spec_tests_f128() {
let cases = [];
spec_test::<f128>(&cases);
}
}