Const rounding methods in float types

RalfJung · May 23, 2025, 3:06pm

Indeed, that's what it is.

In fact they are all already implemented in Miri:

src/intrinsics/mod.rs

445ac3fbc


      
          "floorf16" | "ceilf16" | "truncf16" | "roundf16" | "round_ties_even_f16" => {
              let [f] = check_intrinsic_arg_count(args)?;
              let f = this.read_scalar(f)?.to_f16()?;
              let mode = match intrinsic_name {
                  "floorf16" => Round::TowardNegative,
                  "ceilf16" => Round::TowardPositive,
                  "truncf16" => Round::TowardZero,
                  "roundf16" => Round::NearestTiesToAway,
                  "round_ties_even_f16" => Round::NearestTiesToEven,
                  _ => bug!(),
              };
              let res = f.round_to_integral(mode).value;
              let res = this.adjust_nan(res, &[f]);
              this.write_scalar(res, dest)?;
          }
          "floorf32" | "ceilf32" | "truncf32" | "roundf32" | "round_ties_even_f32" => {
              let [f] = check_intrinsic_arg_count(args)?;
              let f = this.read_scalar(f)?.to_f32()?;
              let mode = match intrinsic_name {
                  "floorf32" => Round::TowardNegative,

This file has been truncated. show original

So that would mostly just have to be moved over to compiler/rustc_const_eval/src/interpret/intrinsics.rs in the compiler.

Definitely feel free to give it a shot.

Topic		Replies	Views
RFC: make all f32/f64 methods const libs	4	1756	September 22, 2020
Feature Request: const float functions	2	325	September 19, 2024
Why float consts are removed from `std::num::Float`? libs	2	1064	March 25, 2019
Idea: In the next edition, stop accepting `0.` as a valid float literal	32	3976	January 15, 2020
Implementing a Fast, Correct Float Parser internals	4	4509	September 28, 2021

Const rounding methods in float types

Related topics