Osprey Kernels: The Best-Of Board

One column and one footnote joined the Osprey board since the first post — same suite, same machine, same oracle. koru-par is the [parallel]-annotated port (the annotation and its proof gate); koru-val is the owned-value port — real malloc/free per node through struct + obligation pairs — the only port whose loop body is hoist-clean. Every number below is from our own runs on this machine (Apple M2 Pro, hyperfine -N, warmup 3, min-runs 10). And this is our interpretation of the suite’s board, not its author’s: Osprey’s creator doesn’t run Koru or Bend — we ported the kernels ourselves and disclose every variant we measured.

The board

Each row takes the best variant a language brought to the board — openly a best-of composite, disclosed per row. For Koru that is the best of the serial, [parallel], owned-value, and koru-fold ports (s/p/v/f); for Bend the best of its serial and bend-par fork/join entries (s/p) — the same treatment Bend already gets by carrying two lanes. Every port compiles through the same koruc; the variants differ in how the kernel is spelled, not in the toolchain. The full per-lane ledger is the persisted board artifact in koru-benchmarks.

Mean wall-clock ms; lower is better. The geomean row aggregates each column’s own numbers — for koru and bend that is the disclosed best-of composite, so it answers “best spelling a language can bring,” not “one port.” * marks rows where the narrow winner is an omitted column — fib goes to clang -O3 (27.1), textstats likewise (2.4). Koru variant letters: s serial port, p the [parallel]-annotated port, v the owned-value port the hoist can prove (binarytrees is its only row), f the koru-fold telescoped-checksum variant (wordfreq). Bend variant letters: s serial, p bend-par.

The composite wins 14 of 22 rows — up from 8 on the serial port alone, which is what the first post’s table showed. The eight losses are the honest set: coins, coprime, nestedloop go to bend-par’s fork/join, fib and textstats to clang -O3, listops and pascal to C, josephus to rust. And the composite corrects a row the serial table miscredited: mutual belongs to koru-par’s 3.9, not to odin’s 6.4.

Where the new columns land

koru-par wins eight rows outright — collatz, digitsum, factorial, gcdsum, isqrt, powmod, primes, tak — not against serial Koru alone but against every binary in the suite, including the C and clang columns and Osprey on powmod, which it held before. Per-row speedup over the serial port on this board: gcdsum 6.0×, factorial 6.0×, mutual 4.3×, collatz 4.1×, powmod 4.0×, isqrt 3.3×, digitsum 2.9×, primes 2.8×, tak 1.0×, binarytrees 0.9×, fib 0.7×. Fold rows geomean 3.4× over the serial port (n=9); with the fork/join losses counted the 11-row geomean is 2.6×. fib pays the gate at 0.7× — fib(35) is below the crossover, and the annotation refuses to pretend otherwise. And binarytrees loses under [parallel] now — the hoist left the loop a bare fold, so the split’s overhead has nothing to amortize against; the serial owned-value port is simply the winner.

koru-val is not a column — one row carries a port. And it is the row Haskell took. The owned-value binarytrees port spells the tree as a recursive struct with mint-node/free-tree obligations — real allocation pressure, the thing the arena port discloses it does not measure. Its ! each body is make(d: 13) |> check(t) — it never reads the item, and hoistInvariantLoopPrefix can prove it: build once, check once, fold the checksum 1200 times, discharge on | done. 1.9ms against Haskell’s 16.9 — against the arena serial port’s 41.1 and koru-par’s 12.2 on the earlier arena spelling (2.2 on the owned-value port — splitting a bare fold buys nothing) — oracle-green at 19659600.

The honest label on that win: it is GHC’s move, not a faster allocator. What Koru adds is that the legality is decidable from the spelling — the effect system knows make is clean, knows the <!owned> consume of t must wait for | done, and refuses the hoist outright when the prefix touches the item or reaches an effect.

What the board found in the compiler

Same yield as the first run — the suite caught two defects while the columns went in:

• An unimported qualified call (std/io:print.ln with no import std/io) never resolved its declaration; the emitter name-mangled it anyway and the string arg arrived as a record literal — .@"{{…}}" = "{{…}}" — dying at the Zig stage with a nonsense error. Now a post-canonicalize pass refuses it as KORU040 with the import spelled out.
• The hoist itself nearly laundered an obligation violation: take { h: *H<!owned> } -> *H<owned!> returns owned!, which read as a mint — but a move of an outer-scope binding is not a mint. Hoisting it relocated the consume out of the loop and quietly legalized the refusal 330_082 pins. Now every <!owned> argument on a lifted step must be bound by an earlier lifted step; consumers of outer obligations stay put.

Measured scope

Both new columns are oracle-green and regression-pinned. This board is arm64 M2 Pro and stays a separate lane from the dedicated x86_64 numbers; nothing here is merged across machines. koru-val ports are .kz — the owned-value spelling needs host declarations for mint/dispose, which is the point: the benchmark’s nominal workload is allocation, and this port does it for real.