Skip to content

implement floor and ceil in assembly on i586 #976

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 0 additions & 22 deletions libm-test/src/precision.rs
Original file line number Diff line number Diff line change
Expand Up @@ -271,18 +271,6 @@ impl MaybeOverride<(f32,)> for SpecialCase {

impl MaybeOverride<(f64,)> for SpecialCase {
fn check_float<F: Float>(input: (f64,), actual: F, expected: F, ctx: &CheckCtx) -> CheckAction {
if cfg!(x86_no_sse)
&& ctx.base_name == BaseName::Ceil
&& ctx.basis == CheckBasis::Musl
&& input.0 < 0.0
&& input.0 > -1.0
&& expected == F::ZERO
&& actual == F::ZERO
{
// musl returns -0.0, we return +0.0
return XFAIL("i586 ceil signed zero");
}

if cfg!(x86_no_sse)
&& (ctx.base_name == BaseName::Rint || ctx.base_name == BaseName::Roundeven)
&& (expected - actual).abs() <= F::ONE
Expand All @@ -292,16 +280,6 @@ impl MaybeOverride<(f64,)> for SpecialCase {
return XFAIL("i586 rint rounding mode");
}

if cfg!(x86_no_sse)
&& (ctx.fn_ident == Identifier::Ceil || ctx.fn_ident == Identifier::Floor)
&& expected.eq_repr(F::NEG_ZERO)
&& actual.eq_repr(F::ZERO)
{
// FIXME: the x87 implementations do not keep the distinction between -0.0 and 0.0.
// See https://github.com/rust-lang/libm/pull/404#issuecomment-2572399955
return XFAIL("i586 ceil/floor signed zero");
}

if cfg!(x86_no_sse)
&& (ctx.fn_ident == Identifier::Exp10 || ctx.fn_ident == Identifier::Exp2)
{
Expand Down
85 changes: 56 additions & 29 deletions libm/src/math/arch/i586.rs
Original file line number Diff line number Diff line change
@@ -1,37 +1,64 @@
//! Architecture-specific support for x86-32 without SSE2
//!
//! We use an alternative implementation on x86, because the
//! main implementation fails with the x87 FPU used by
//! debian i386, probably due to excess precision issues.
//!
//! See https://github.com/rust-lang/compiler-builtins/pull/976 for discussion on why these
//! functions are implemented in this way.
use super::super::fabs;
// FIXME: when the MSRV allows, use naked functions instead.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is there actually any reason to do this now? Being able to inline this assembly should be advantageous, and from a quick look at godbolt it looks like this function is only 13 instructions (compared to 14 from the NetBSD implementations).


/// Use an alternative implementation on x86, because the
/// main implementation fails with the x87 FPU used by
/// debian i386, probably due to excess precision issues.
/// Basic implementation taken from https://github.com/rust-lang/libm/issues/219.
pub fn ceil(x: f64) -> f64 {
if fabs(x).to_bits() < 4503599627370496.0_f64.to_bits() {
let truncated = x as i64 as f64;
if truncated < x {
return truncated + 1.0;
} else {
return truncated;
}
} else {
return x;
pub extern "C" fn ceil(mut x: f64) -> f64 {
unsafe {
core::arch::asm!(
"fld qword ptr [{x}]",
// Save the FPU control word, using `x` as scratch space.
"fstcw [{x}]",
// Set rounding control to 0b10 (+∞).
"mov word ptr [{x} + 2], 0x0b7f",
"fldcw [{x} + 2]",
// Round.
"frndint",
// Restore FPU control word.
"fldcw [{x}]",
// Save rounded value to memory.
"fstp qword ptr [{x}]",
x = in(reg) &mut x,
// All the x87 FPU stack is used, all registers must be clobbered
out("st(0)") _, out("st(1)") _,
out("st(2)") _, out("st(3)") _,
out("st(4)") _, out("st(5)") _,
out("st(6)") _, out("st(7)") _,
options(nostack),
);
}
x
}

/// Use an alternative implementation on x86, because the
/// main implementation fails with the x87 FPU used by
/// debian i386, probably due to excess precision issues.
/// Basic implementation taken from https://github.com/rust-lang/libm/issues/219.
pub fn floor(x: f64) -> f64 {
if fabs(x).to_bits() < 4503599627370496.0_f64.to_bits() {
let truncated = x as i64 as f64;
if truncated > x {
return truncated - 1.0;
} else {
return truncated;
}
} else {
return x;
pub extern "C" fn floor(mut x: f64) -> f64 {
unsafe {
core::arch::asm!(
"fld qword ptr [{x}]",
// Save the FPU control word, using `x` as scratch space.
"fstcw [{x}]",
// Set rounding control to 0b01 (-∞).
"mov word ptr [{x} + 2], 0x077f",
"fldcw [{x} + 2]",
// Round.
"frndint",
// Restore FPU control word.
"fldcw [{x}]",
// Save rounded value to memory.
"fstp qword ptr [{x}]",
x = in(reg) &mut x,
// All the x87 FPU stack is used, all registers must be clobbered
out("st(0)") _, out("st(1)") _,
out("st(2)") _, out("st(3)") _,
out("st(4)") _, out("st(5)") _,
out("st(6)") _, out("st(7)") _,
options(nostack),
);
}
x
}
4 changes: 2 additions & 2 deletions libm/src/math/rem_pio2_large.rs
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@
* ====================================================
*/

use super::{floor, scalbn};
use super::{scalbn, trunc};

// initial value for jk
const INIT_JK: [usize; 4] = [3, 4, 4, 6];
Expand Down Expand Up @@ -290,7 +290,7 @@ pub(crate) fn rem_pio2_large(x: &[f64], y: &mut [f64], e0: i32, prec: usize) ->

/* compute n */
z = scalbn(z, q0); /* actual value of z */
z -= 8.0 * floor(z * 0.125); /* trim off integer >= 8 */
z -= 8.0 * trunc(z * 0.125); /* trim off integer >= 8 */
n = z as i32;
z -= n as f64;
ih = 0;
Expand Down
Loading