Avoid double rounding decimal literals when materializing f32 values

オープン
  • 何か問題がありますか？
#939 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
20/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
rust
領域
compilers

調査の方向性

Read the literal path through front/lexer/src/scan.rs:470, front/parser/src/expr/primary.rs:349, and front/hir/src/numeric_checks.rs:57 to see where decimal source information is lost before f32 materialization. Discuss the design with @LunaStev before implementation. Reproduce the given program at O0 and O2, then use exact-oracle tests for midpoints, adjacent decimals, negative values, and finite range boundaries; done means the reproduction yields 0x3f800001 while casts, spans, and diagnostics retain their stated semantics.

索引モデルが issue の本文から書いたものです。

説明

bug help wanted

Problem

Decimal floating literals are first converted to f64 in the lexer, then narrowed when Typed HIR selects f32. Near an f32 halfway point, rounding to f64 first can erase which side of the midpoint the original decimal occupied, producing the wrong nearest f32 value.

Starting points: front/lexer/src/scan.rs:470, front/parser/src/expr/primary.rs:349, and front/hir/src/numeric_checks.rs:57. This differs from #927's accepted f32 default and normal finite rounding: the example is rounded to the wrong f32 neighbor even with an explicit f32 declaration.

Reproduction

Audited canonical master d59bdab376bdc04edacdc8029b256cb8d55b5adf on Linux amd64 with a checkout-built Wave 0.0.1-alpha-dev, Rust 1.89.0 and LLVM 21.1.8.

import("std::math::float")::{float_to_bits_f32};
fun main() {
    var value: f32 = 1.0000000596046448;
    println("{}", float_to_bits_f32(value));
}

At both O0 and O2 the native program prints 1065353216 (0x3f800000, exactly 1.0).

The exact midpoint between 1.0 and its next f32 neighbor is 1 + 2^-24 = 1.000000059604644775390625. The source decimal 1.0000000596046448 is strictly greater than that midpoint, so its nearest f32 is the next neighbor: bits 1065353217 / 0x3f800001.

The midpoint comparison was checked using Python Decimal with precision 80; it does not use another f64 conversion as the oracle.

Scope

Preserve sufficient source information until the selected floating materialization width is known, and round the literal to that width correctly. Keep f32 as the untyped default, contextual f64 behavior, explicit conversion order, signed zero, underflow and overflow diagnostics. Do not reinterpret this as a request to reject all inexact decimals.

Please mention @LunaStev to discuss the design before starting implementation.

This touches the frontend literal representation shared by both backends, so it is not labeled a good first issue.

Completion criteria

  • The reproduction yields 0x3f800001 at O0/O2.
  • Test exact midpoints, decimal values immediately on either side, negative values and finite range boundaries with an exact oracle.
  • Preserve source spans and explicit f64-to-f32 cast semantics; distinguish directly materialized f32 literals from intentionally rounded cast chains.
主要言語
Rust
スター
56
フォーク
21
平均マージ
23時間 20分
マージ済み PR（30日）
45

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

wavefnd/Wave のほかの issue

wavefnd/Wave の issue をすべて見る

似ている issue

Rust の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。