2026-09-02 - (/loop) Bulk port batch 5: the single-precision math family and byte order

documented   714 needed, 395 missing   ->   714 needed, 374 missing

Twenty-one entries. The simple ones are the standard's function under Rust's name for it; the six written by hand are the ones where that correspondence does not exist, and they are where the specification says something a plausible implementation would get wrong.

Math (ISO/IEC 9899 7.12), in orbistoun-libc/src/math.rs

By macro: acosf, asinf, atanf, cbrtf, exp2f, log10f, log2f, tanhf, hypotf, exp2. By hand: nearbyintf, frexp, modf, modff, ldexp, ldexpf, sincos.

Four details worth having written down:

• nearbyintf is ties-to-even, not round. round is ties-away-from-zero and the two disagree on exactly the halfway cases - the ones a casual test never uses. Tested at 0.5, 1.5, 2.5 and -0.5.
• frexp must answer zero, infinity and NaN unchanged with *exp = 0. A version computing the exponent from a logarithm answers negative infinity for zero and stores nonsense in the caller's int.
• modf gives both parts the argument's sign. modf(-3.5) is -0.5 and -3.0; flooring gives 0.5 and -4.0, which reassembles to the same number and is still wrong.
• ldexp splits its scaling. 2^n is representable only inside the exponent's range, so a single x * 2f64.powi(n) answers infinity for large n even when the product would be finite - and cannot reach a subnormal going the other way. Multiplying by an exact power of two is itself exact, so splitting keeps one rounding at the end. Tested at 2^1023, past it, and down to the smallest subnormal (which needs the split to reach at all).

ldexp, frexp, modf and sincos take integer arguments - an exponent or an out-parameter - which arrive in integer registers. Reading them from the float registers would read whatever the caller last put there.

sincos is a BSD extension rather than ISO C, cited to FreeBSD sincos(3); the target's C library is FreeBSD-derived, which is why it is imported at all.

Byte order, in orbistoun-fs/src/socket.rs

htonl, htons, ntohl, ntohs (POSIX.1-2008), beside the sockets that use them, with delegation rows in orbistoun-posix. Exact, and the two tests that matter say why: only the low half of the register is meaningful (the argument is a uint32_t in a 64-bit register whose top half is the caller's leftovers - reading it would answer a different number each call for the same input), and each conversion is its own inverse, which is why network order needs one operation under four names rather than two operations.

Written as a byte swap rather than a no-op, with a comment saying so: a no-op is what it would be on a big-endian host, and that is exactly the assumption that would be wrong there.

Housekeeping

Cleared five clippy warnings that were not from this batch - unnecessary qualification in orbistoun-fs/src/lib.rs, left by an earlier session's test-isolation work in this same uncommitted tree. They would have failed CI.

clippy --tests clean across libc/fs/posix, fmt clean, tests pass, nothing committed.

Next: the 259 genuinely-absent POSIX names, which is the largest remaining block. Also worth adding: a machine-checkable partial marker. The knowledge files carry known_by, returns, arity, cites and edge_cases but nothing that says "implemented, but not completely", so a deliberately partial function like getopt (which handles only the no-arguments case and says so in prose) is indistinguishable from a finished one in --static-gap. That is a real hole in the measurement, not just in the code.