moonbitlang / core / 6905

89%

push

github

fix(math): evaluate log10 in hi/lo parts on non-JS targets

`log10(x)` scaled `ln(f)` by a single rounded 1/ln(10), up to 1.8 ulp off.
Port the tail of FreeBSD's e_log10.c (as in musl and Rust's `libm`)
instead: keep log(1+f) as a hi + lo pair, multiply by a two-part 1/ln10
and add k*log10(2) with the rounding error carried along.

On 78,522 positive finite inputs against an 80-digit reference: correctly
rounded 83.76% -> 98.41%, max error 1.782 -> 0.680 ulp, results off by more
than 1 ulp 1,009 -> 0. Results are bit-identical to musl's log10.c on
native, wasm and wasm-gc, and log10(10^k) = k still holds for every
k = -311..308. It is also faster (one inlined reduction instead of `frexp`
twice): 5.6x on native and 2.5x on wasm-gc for 1,024 calls.

The digit-exact tests live in a new non-JS test file: on JS `log10` is the
engine's `Math.log10`, whose last digit is the engine's business. The V8
tested here (14.1, Node 25) still uses the older, less accurate formula,
so the two differ in the last digit on many inputs (`log10(11.0)` moves
there for that reason).

Part of #4336.

Reviewed by Codex CLI (xhigh): "The port matches both local references in
constants, arithmetic order, and bit manipulation. Subnormal scaling,
reduction boundaries, `+0` at one, and non-JS test gating check out."

Signed-off-by: Codex CLI <codex@openai.com>

12 of 12 new or added lines in 1 file covered. (100.0%)

16698 of 18669 relevant lines covered (89.44%)

344983.09 hits per line