from_str_r moves exactly representable decimals by one ULP under TowardZero/TowardNegative (port of llvm#228928)

Start by reading the from_str_r implementation and the existing floating point conversion tests; the issue gives reproducible decimal inputs and expected bit patterns for Single and Double under directed rounding modes. Add regression cases covering the reported exact and near-boundary conversions, then verify the resulting values and status flags across the affected formats and rounding modes. The issue says the fix likely depends on an upstream LLVM change, so check that defect’s status before estimating implementation work.

This was identified by comparing floating point number conversion results across a few open-source implementations. The implementation of from_str_r incorrectly rounds some exactly representable numbers, however only in TowardZero/TowardNegative rounding modes which do not appear to be used yet. That makes this lower priority but opening for awareness - feel free to close if this is expected/intended.

Below is the full AI-generated summary:

from_str_r treats some exactly representable decimal strings as inexact. Under TowardZero and TowardNegative (and TowardPositive for negative inputs) it then returns the value one ULP below the exact one. The status is INEXACT in every mode, including NearestTiesToEven, where the value itself is right.

This is the LLVM APFloat defect reported as https://github.com/llvm/llvm-project/issues/228928, carried over by the port. rustc_apfloat 0.2.3 (+llvm-462a31f5a5ab) fails on different inputs than current LLVM does (the single-precision example in the LLVM issue converts correctly here), but the pattern is the same.

[dependencies]
rustc_apfloat = "=0.2.3"

use rustc_apfloat::ieee::{Double, Single};
use rustc_apfloat::{Float, Round};

fn check<F: Float>(label: &str, text: &str, round: Round, want: u128) -> bool {
    let r = F::from_str_r(text, round).unwrap();
    let got = r.value.to_bits();
    let ok = got == want;
    println!("{label:<24} {got:#018x}  want {want:#018x}  {:?}{}", r.status, if ok { "" } else { "  <- WRONG" });
    ok
}

fn main() {
    // (2^23 + 1) * 2^-45: exactly the binary32 value 0x34800001
    let single = "0.000000238418607523271930404007434844970703125";
    // 2^11 + 2^-41: exactly the binary64 value 0x40A0000000000001
    let double = "2048.00000000000045474735088646411895751953125";
    let mut ok = true;
    for (name, round) in [
        ("NearestTiesToEven", Round::NearestTiesToEven),
        ("TowardZero", Round::TowardZero),
        ("TowardNegative", Round::TowardNegative),
        ("TowardPositive", Round::TowardPositive),
    ] {
        ok &= check::<Single>(&format!("Single {name}"), single, round, 0x34800001);
        ok &= check::<Double>(&format!("Double {name}"), double, round, 0x40A0000000000001);
    }
    std::process::exit(if ok { 0 } else { 1 });
}

Output (rustc 1.99.0, Linux aarch64):

Single NearestTiesToEven 0x0000000034800001  want 0x0000000034800001  Status(INEXACT)
Double NearestTiesToEven 0x40a0000000000001  want 0x40a0000000000001  Status(INEXACT)
Single TowardZero        0x0000000034800000  want 0x0000000034800001  Status(INEXACT)  <- WRONG
Double TowardZero        0x40a0000000000000  want 0x40a0000000000001  Status(INEXACT)  <- WRONG
Single TowardNegative    0x0000000034800000  want 0x0000000034800001  Status(INEXACT)  <- WRONG
Double TowardNegative    0x40a0000000000000  want 0x40a0000000000001  Status(INEXACT)  <- WRONG
Single TowardPositive    0x0000000034800001  want 0x0000000034800001  Status(INEXACT)
Double TowardPositive    0x40a0000000000001  want 0x40a0000000000001  Status(INEXACT)

Scope: a sweep over decimal strings into Double, Single, Half, BFloat, Float8E5M2, Float8E4M3FN and Quad in all five modes found 1,268 exactly representable inputs moved one step in the rounding direction, in Double, Single, BFloat and Quad. It also found 84 inexact inputs within a tiny fraction of an ULP of a Quad value, rounded as if they were on the other side of it (e.g. 4.5326225836066175776471072707593525867497e-4965 under TowardZero gives 6 × 2⁻¹⁶⁴⁹⁴ instead of 7 × 2⁻¹⁶⁴⁹⁴). No other differences turned up in 151,415 conversions. Half and the FP8 types had none (LLVM 21.1.8 does fail for IEEEhalf).

Since rustc only folds literals with NearestTiesToEven, compiled code is not affected. This matters to direct users of the crate who pass a directed mode, and to anyone relying on the status flag. The fix most likely belongs in LLVM first; this issue is here so it gets ported when that happens.

Use the following links to see the expected conversion results:

• (2²³ + 1)·2⁻⁴⁵ (Double) → Single, toward zero
• 2¹¹ + 2⁻⁴¹ (binary128) → Double, toward zero