Tracking Issue for `Duration::try_from_secs_{f32, f64}`

[mbartlett21]

Feature gate: #![feature(duration_checked_float)]

This is a tracking issue for checked methods to construct a duration from a floating-point value of seconds without panicking.

Public API

// core::time

impl Duration {
    pub const fn try_from_secs_f32(secs: f32) -> Result<Duration, FromFloatSecsError>;
    pub const fn try_from_secs_f64(secs: f64) -> Result<Duration, FromFloatSecsError>;
}

#[derive(Debug, Clone, PartialEq, Eq)]
pub struct FromFloatSecsError { ... }

impl core::fmt::Display for FromFloatSecsError { ... }

// std::error

impl std::error::Error for FromFloatSecsError { ... }

Steps / History

• Implementation
  • Add functions Duration::try_from_secs_{f32, f64} #82179
  • Improve Duration::try_from_secs_f32/64 accuracy by directly processing exponent and mantissa #90247
  • Use rounding in float to Duration conversion methods #96051
• Final commenting period (FCP)
  • Stabilize duration_checked_float #102271 (comment)
• Stabilization PR
  • Stabilize duration_checked_float #102271

Unresolved Questions

• What should the error type be called?
  • Originally FromSecsError.
  • Changed to FromFloatSecsError in Improve Duration::try_from_secs_f32/64 accuracy by directly processing exponent and mantissa #90247
  • Changed and finalised as TryFromFloatSecsError in Stabilize duration_checked_float #102271

[workingjubilee]

This needs more testing to vet its behavior in the face of various interesting floating point values.
In particular, subnormal values and values close to the limit of f32 and f64's upper range.

[mbartlett21]

This needs more testing to vet its behavior in the face of various interesting floating point values.
In particular, subnormal values and values close to the limit of f32 and f64's upper range.

This just currently checks what the othee panicking methods for the same thing do.

[newpavlov]

The new code in #90247 should handle all interesting values perfectly.

[nvzqz]

What is left to stabilize this? It is important to me that time conversions I'm using don't panic.

[mbartlett21]

There is still an unresolved question:

What should the error type be called?

• In the original PR, I called it FromSecsError.

• In Improve Duration::try_from_secs_f32/64 accuracy by directly processing exponent and mantissa #90247, @newpavlov updated the error type to FromFloatSecsError.

• It is still unclear whether a different name would be better.

Once that is resolved, we'll need a stabilisation PR and then, once that gets approved, it will be a month or two until it appears in stable

EDIT: Note: #96051 changes the rounding mode to round-to-nearest instead of truncate. (note on the note: That pr is also changing the stabilised methods from_secs_*, but it would be nice to have it correct when it is first stabilised)

[arniu]

Why not just impl TryFrom<f32 | f64> for Duration?

[mbartlett21]

Because it isn't clear from TryFrom that it is meant to be seconds. It could be milliseconds or hours or days or some other reasonable measurement.
This is the same reason why Duration doesn't impl From<u128> or something. It could likewise be a lot of other units.

[lopopolo]

I came across this unstable API when debugging a panic in a time crate I maintain:

• Time::checked_add_f64 and Time::checked_sub_f64 panic on too large input artichoke/artichoke#2194

I will prepare a stabilization PR shortly and hope we can resolve the open question on error type naming in the context of an FCP. I hope that's an ok way to approach this.