Repository navigation
t27o golden functions (step 4 of #9000), measured first: hand lines in converted fns of partly converting files, and which can move without changing a caller #11159
Description
Goal
Measure the per-function golden prize: the hand-written Rust lines that sit in functions t27o already converts, inside files that only partly convert, and which of those functions could be replaced by a generated function mounted into the same file without changing a caller. This is step 4 of #9000 for the files that do not convert whole.
Why
The first golden port (#11157, bootstrap/src/host/bitmask.rs, +1/-171 hand lines) was the only file in the t27 repository that converts whole; A7's analysis (#11149) found 3,006 hand lines in 80 whole files across five repositories (0.69% of the swept lines), most of them stubs or in trios crates the workspace does not build. The sweep of 2026-10-20 converts 632 of 7,038 fns in the t27 tree and 1,824 of 19,392 over all five. Those converted fns sit in files that also hold refused fns, so the file cannot be replaced, but the function can: its body moves into a spec, the file mounts the generated function with #[path], and the hand lines are deleted. Nobody has counted those lines or checked how many have callers whose types would change.
Boundary
- docs/now/2026-10-20-golden-functions.md
User Scenarios
- Given a partly converting file, when the analysis runs, then it lists each converted fn with its hand-written line span, its callers in the same file and in other files, and the type difference gen-rust would introduce.
- Given a converted fn called only by other converted fns of the same file, when it is ranked, then the row says the whole group can move together.
- Given a converted fn whose signature changes under gen-rust, when it is ranked, then it is marked with the callers it would break.
Requirements
- FR-001: The table MUST count hand-written lines the way the own-language gate does, name the sweep driver and the commits, and exclude generated, vendored and test-only paths with the reason.
- FR-002: Each ranked fn MUST be checked for valid output (
t27c test-reportof its converted spec) and for the signature gen-rust writes against the original. - FR-003: This issue is analysis only: no Rust is deleted and no spec is added before the table is on the issue.
Success Criteria
- The table is posted on this issue and held in
docs/now/2026-10-20-golden-functions.md, with at least 30 rows ranked by hand-written lines. - It names the first group of fns that deletes at least 200 lines of hand-written Rust with every caller unchanged, or states in numbers why none does.
Activity
gHashTag commentedon Oct 19, 2026
Measured, analysis only. Same content as
docs/now/2026-10-20-golden-functions.mdin draft PR #11160 (head 22a43fb), cut to the answer, the scale, the reasons, the first bundle, the top bundles and the top 120 of the 187 ranked rows; the rest, the signature appendix and the per-file list of every converted fn are in the document.valid-output: unchecked on 44 of the 187 rows (45 of the 223 specs):
t27c test-reportwas stopped when the shared Mac ran low on disk under load; those rows still carry therustccompile of their gen-rust output. The unchecked rows are 30 trios rows and 14 999-multibots-telegraf rows; every t27 row was checked (all 93 t27 specs).Answer
#[cfg(test)]; 44,864#[cfg(test)]lines sit in the same files).That is 1.51% of the 435,491 physical lines lane A4 swept (the denominator of t27o golden PRs (step 4 of #9000), measured first: the 85 Rust files that convert whole, ranked by hand lines a port would delete #11149), where the whole files of t27o golden PRs (step 4 of #9000), measured first: the 85 Rust files that convert whole, ranked by hand lines a port would delete #11149 were 0.69%.
NOT-NOW: 617 fns, 4,774 lines (73% of the hand lines in converted fns).
rings/crates 510 in 38; t27 other crates 108 in 9; trios workspace crates 421 in 49;trios crates not built 110 in 14; the three other repositories 9 in 9. Methods never move without a caller change, and a fn that names a hand-file type cannot move before the type does.
exp_f64+apply_softmaxofrings/ring-092-rust/src/lib.rs, 93 lines. No call-connected group reaches 200.rings/ring-092-rust/src/lib.rs, 201 hand lines(
abs_f649,exp_f6445,apply_softmax48,cache_kv20,compute_scores38,weighted_values30,identity_witness11), one file, one mount, five call-connected groups that do not call each other.Plainly: 201 lines deleted against a 5-line mount added, net -196; no other bundle exceeds 88 lines (ring-093 88, ffi 78, ring-099 46, cli/tri/src/fpga.rs 45). Section "First bundle" has the evidence.
cli/tri/src/fpga.rswith 45 lines; all 27 MOVABLE fns of bootstrap and cli/* hold 217 lines.t27c test-report(one spec per file, a fresh ZIG_LOCAL_CACHE_DIR deleted after each run): 100 specs hold tests (377 tests, 377 pass, 0 FAIL, 57 vacuous passes);78 specs hold no test (their file has none), so the report has nothing to run for them and those rows lean on the rustc compile of the gen-rust output. 45 specs were NOT checked
(stopped when the shared Mac ran low on disk under load): 45 files, the 44 rows marked
valid-output: uncheckedbelow (of 187 rows).Every row's gen-rust output was also compiled with
rustc --crate-type lib(stub fns removed; the ring crates also under#![no_std]).Scale (partly converting files only; fns and lines are converted non-test fns)
Why the rest cannot move now (NOT-NOW, by first reason)
Per area (fns / lines): W27: ty+ 56/367, ty- 119/836, bad 43/400, constctx 9/59, trait 11/75, dep 1/12; R27: ty+ 14/73, ty- 71/694, bad 11/92, constctx 3/12, trait 4/12; O27: ty- 10/18, attr 2/10, trait 1/3; WT: ty+ 54/416, ty- 93/632, bad 5/88, constctx 1/5, attr 16/128, trait 15/68; NT: ty+ 11/92, ty- 34/320, bad 7/128, trait 4/19; OT: ty+ 2/4, ty- 13/163, bad 2/28, attr 2/8, trait 3/12
Type::method(&self, ..)into the free fnType_method(self_, ..)and writes the type again in the generated module, soevery method and every fn that takes or returns a hand-file type needs the type moved first; types that derive
PartialEq,Eq,Hash,Defaultor serdelose those derives (gen-rust writes
Debug, Clone, Copy), which is the second row. A type-move PR is the next step for thety+fns, not a function move.&strparameters become&[u8]when the spec slices (56 free fns) or&'static str(9): the call sites need.as_bytes()or a literal, which is why those fns are CALLER-CHANGE.pub const X: Vec<u64> = [..],Vec<const u8>), const-evaluated calls (E0015),Option<T>indexing,&striteration,text indexing; these are gen-rust and converter defects, listed by rustc message in the appendix.
First bundle: rings/ring-092-rust/src/lib.rs (7 fns, 201 hand lines)
ring-092-rust(package name),rings/ring-092-rust: in the rootCargo.tomlexcludelist, built standalone by.github/workflows/rings-rust.yml(Rust 1.83,continue-on-error).The names
apply_softmax,weighted_values,compute_scoresandcache_kvoccur in no other.rsfile of the five repositories;identity_witness(17 rings),exp_f64(ring-093) andabs_f64(ring-093,compiler.rs) are defined again in independent crates that do not depend on ring-092;#![no_std],#![forbid(unsafe_code)],#![deny(missing_docs)].#[test]fns of its#[cfg(test)]module (21 of them converted into the spec as tests) and, between the fns,apply_softmax->exp_f64and
identity_witness->abs_f64. No refused fn calls any of the 7. Call-connected groups: {exp_f64, apply_softmax}, {abs_f64, identity_witness}, {cache_kv}, {compute_scores}, {weighted_values}.&[f64], which gen-rust writesimpl AsRef<[f64]>(every call form of the file compiles;&[]literals would need an annotation).The consts the fns use (
NUM_HEADS,HEAD_DIM,EMBED_DIM,CONTEXT_LEN,SACRED_SCALE) also convert and compile, but this count does not delete them from the hand file.t27c gen-rustof it (198 lines);a copy of the crate with the 7 fn spans deleted and a 5-line mount (
#[path],#[allow(missing_docs, ..)],mod golden;,pub use golden::{5 pub fns};,use golden::{abs_f64, exp_f64};).rustc --testof the original: 28 passed, 0 failed; of the mounted copy: 28 passed, 0 failed; 201 lines out, 5 in (761 -> 565 lines).abs_f64andexp_f64are used only by tests and bygenerated code in the hand file, so the
useline wants#[cfg(test)]in crates withdeny(warnings)(rings 094-099); ring-092 only warns "unused imports".exp_f64andabs_f64over4,000,022 inputs each (specials, subnormals, +-inf, NaN, the 700 cut-off, random bit patterns, uniform and wide ranges);
apply_softmax300,000 arrays andseq_lenvalues including out of range;cache_kv200,000;compute_scoresandweighted_values30,000 each, with short buffers and out-of-range positions;identity_witness. 8,560,045 comparisons, 0 disagreements, 180,769 cases whereboth panicked (out-of-contract buffers), at
-C opt-level=0(overflow checks on),-C opt-level=3 -C overflow-checks=offand-C opt-level=3 -C overflow-checks=on.python3 -I tools/dupe_scan.py --likeon the group spec: "nothing in this file is written elsewhere" (no byte-identical body). Butspecs/nn/attention.t27already declarescache_kv,compute_scores,apply_softmax,weighted_valuesandternary_matmul, with a different argument order forcompute_scoresandweighted_values; the crate says it mirrors that spec. A port must reconcile with it first (as thesix
specs/port/files do in t27o golden PRs (step 4 of #9000), measured first: the 85 Rust files that convert whole, ranked by hand lines a port would delete #11149), or it adds a second spec of the same kernel.t27c gen-rustwrites under a generated root (for examplegen/rust/rings/, mounted as#[path = "../../../gen/rust/rings/ring092_attention.rs"]), a seal and a t27b ledger row. The own-language budget sees +5 / -201 in one hand file.Bundles: the MOVABLE fns of one file, one mount per file (top 25 of 74 files)
Not all of a bundle's fns call each other: "groups" is the number of call-connected groups in it. "net" is lines minus a 5-line mount (3 if the mount attributes share one line, as in #11157).
Ranked table: every MOVABLE or MOVABLE-WITH-CALLER-CHANGE group (187 rows), ranked by hand lines
Legend. area: W27 t27 workspace crates, R27 t27 rings/, O27 other t27 crates, WT trios workspace crates, NT trios crates not built, OT trinity / trinity-fpga / 999-multibots-telegraf (repo named). A group is the
call-connected set of converted fns of one file that call only each other (a lone fn is a group of one); a group is MOVABLE only if every fn in it is. "callers in file": conv = another converted fn,
refused = a fn the converter refuses (it keeps calling the moved fn by name through a
use, so it blocks only when a signature changes), test = a#[test]or helper in the#[cfg(test)]module,other = a const/static initializer. "callers elsewhere": files outside the row that name the fn through its module path, an import or the crate name (checked by hand for every MOVABLE row that had a hit).
"gen-rust vs original": differences in the signature gen-rust writes;
const fn -> fnmatters only to a const context (checked: none for a MOVABLE row);impl AsRefwidens. "valid output":tr P/T vac Vist27c test-reportof the file's converted spec (P passes of T tests, V vacuous passes); "gen-rust compiles" isrustc --crate-type libof the generated file with the refused stubs removed.Rows 121-187, Appendix A (gen-rust signature differences) and Appendix B (every converted fn by file, with its verdict code) are in the document.
Not run
t27c test-reportof 45 specs (30 trios files, 15 files of trinity, trinity-fpga and 999-multibots-telegraf): stopped when the shared Mac ran low on disk under load. Their rows sayvalid-output: unchecked.Their gen-rust output was still compiled with
rustc.tri t27o --check(the oracle replay) on any file: not run. The only behavioural comparison is the differential run of the ring-092 bundle above; no other group was compared against its hand fns.cargo build,cargo checkorcargo testof any crate. Callers come from a token index of every.rsfile of the five repositories (comments and strings blanked), with module-path, import and crate-name matching,then
git grepby hand for every MOVABLE row that had a hit elsewhere; a caller that reaches a fn through a glob import and another name, or through a macro, would be missed. Only the ring-092 bundle was mounted andcompiled (with
rustc, the crate has no dependency).t27bledger edit, own-language gate run ordupe_scanover any spec but the ring-092 group spec; no spec was added to the repository.ty+rows are the candidates for it.docs(t27o): RFC for heap-free collections, owned strings and inlined …