fix(es/ast): support hashing non-finite numbers

[magic-akari]

Description:

Canonicalize non-finite numeric values in Lit::Num while preserving the existing ECMA AST Hash API.

SWC currently represents Infinity and NaN in several forms. eval_global_vars rewrites Infinity to 1 / 0, while eval_numbers can rewrite division by zero back to an Infinity identifier. Other constant-evaluation paths produce Lit::Num(INFINITY), which codegen prints as Infinity. This churn makes optimization passes repeatedly recognize equivalent values in different AST forms.

Follow esbuild’s approach of using semantic constants during optimization. Allow Lit::Num to store valid JavaScript non-finite values as the canonical internal representation, and leave syntax selection to codegen. Codegen can emit readable Infinity and NaN when safe or shorter 1 / 0 and 0 / 0 when minifying.

BREAKING CHANGE:

Number now treats all NaN values as equal and distinguishes -0.0 from +0.0.

Related issue (if exists):

• fix(es/ast): support hashing non-finite numbers #12043 (current)
  • fix(es/codegen): emit non-finite numeric literals #12047
    • fix(es): preserve numeric property key identity #12050
      • fix(es/minifier): use numeric literals for non-finite values #12048
        • fix(es/optimization): preserve JSON numeric values #12051
      • fix(es/minifier): preserve non-canonical arguments access #12052
        • fix(es/minifier): bound arguments parameter injection #12053
      • fix(es/minifier): index strings by UTF-16 code unit #12054
      • fix(es/minifier): apply ToUint16 in fromCharCode #12055
      • fix(es/minifier): preserve invalid Array lengths #12056
      • fix(es/minifier): avoid inexact number radix folding #12057
    • fix(es/minifier): apply ToInt32 when folding bitwise NOT #12058
    • fix(es/transforms): use numeric literals for non-finite enum values #12049
      • fix(es/typescript): evaluate cooked enum templates #12059
        • refactor(es/typescript): track enum evaluation context #12060

[changeset-bot]

🦋 Changeset detected

Latest commit: d9854d2

The changes in this PR will be included in the next version bump.

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[codspeed-hq]

Merging this PR will not alter performance

✅ 189 untouched benchmarks
⏩ 61 skipped benchmarks¹

---

_{Comparing fix/ecma-number-hash (d9854d2) with main (54467fe)²}

Footnotes

1. 61 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on main (062c608) during the generation of this report, so 54467fe was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[Donny/강동윤 (kdy1)]

Why do we need this?

[magic-akari]

Why do we need this?

My motivation comes from a technique used by esbuild. In internal/config/globals.go, its define phase replaces the global undefined, NaN, and Infinity identifiers with semantic constants. Later transform and minification passes operate on these internal values instead of preserving their identifier-based AST representation.

I traced SWC’s current compressor and found that we already mix several representations. eval_global_vars converts Infinity to 1 / 0, while eval_numbers can convert another x / 0 back to an Infinity identifier and even has explicit branches saying “Prefer 1/0 over Infinity” and “Prefer 0/0 over NaN.” Other constant-evaluation paths produce Lit::Num(INFINITY), which codegen prints as Infinity. This is not an infinite loop within one compression pass, but it creates real representation churn across optimization, codegen, and reparsing. The same value must also be recognized repeatedly in different AST forms.

Using Lit::Num as the internal working representation gives these transformations a canonical endpoint. Codegen can then choose the appropriate syntax at the output boundary—for example, readable Infinity/NaN in normal output when safe, or shorter 1/0/0/0 when minifying.

This PR removes the artificial restriction that prevents Lit::Num from representing these valid JavaScript numeric values, while preserving the existing Hash API.

[Donny/강동윤 (kdy1)]

Thank you for the explanation! I'll merge this PR after adding the context to the PR body (and the commit message)