feat(l3): common-tail merge, scoped base clusters, and ranked live-range coalescing

[macabeus]

Third round on kleod:UpdateHUDCounterDisplay:agbcc. diff:60 → diff:18 (m2c: noncompile). Board: asmlift 350 vs m2c 340, 0 lost, 0 gained, 0 flips.

The row still does not match. What ships is three general L3 capabilities plus two bug fixes; the remaining gap is described at the end.

What ships

mergeCommonTails (l3/tailmerge.ts, unconditional in structureChecked). A statement that ends EVERY arm of an if moves below the if. SSA destruction puts the same merge-variable write at the end of each arm because that is where each edge's copy belongs; the source wrote it once. Both arms execute it last on their own path, so moving it below runs it on the same paths in the same order — no liveness or dominance analysis needed, and it holds for a side-effecting statement. The merge direction is what is unconditionally sound: a common HEAD would move across the condition's own evaluation. Placement is measured, not taste — below the if scores 33, peeling the same statements above it scores 48.

hoistScopedBases / /scopebase gains a deepestCluster fallback: when a base's uses have no common scope, hoist at the deepest cluster that does, rather than declining. Differ-refereed, so it can never cost a match.

coalesceCandidates / /coalesce (l3/coalesce.ts). Which pair of locals the register allocator coalesced is not derivable from the L3 tree, and first-fit gets it wrong — on this row the two legal merges score 18 and 40, and declaration order picks the 40. So every legal single merge is emitted as its own candidate and the differ referees, the same idiom /regcopy already uses.

fix(rank): a lever that throws was swallowed and reported as a decline — indistinguishable from "this lever had nothing to offer". It now reports via onLeverError.

fix(l3): exprEquals compares a field's dot-vs-arrow spelling. A CSE or dedup that treats p->f and p.f as equal keeps one node and discards the other, silently respelling the access.

Generality: mined against a whole game, not just the benchmark

Only ONE benchmark row changed, which is a fair thing to be suspicious of for a branch that adds an unconditional pass. Two measurements answer it.

First, instrumenting the passes over all 675 benchmark cases with the harness's own options: mergeCommonTails fires on exactly 1, and the exprEquals dot fix has 0 divergences corpus-wide (it is defensive, unit-tested only).

Second, the same instrumentation over the klonoa decomp — 464 per-function .s plus the built translation units, 368 of which lift and structure:

pass	klonoa	benchmark
mergeCommonTails	11 distinct functions, 7 of which change the emitted C	1
hoistScopedBases	29 functions	wins 1 row
coalesceCandidates	19 candidates on 7 functions	2 candidates on 1 row

Firings outside the target row, verified by reading the before/after C: EntityCrushingBlock (a base assignment ending both arms), UpdateBootMinigame (two ifs, two statements merged out of each), UpdateUIElementAnimation (v7 = 0 merged out of a nested if and then its parent).

Why only one benchmark row moved: size. The klonoa functions where the shape occurs have a median of 198 asm lines; the benchmark's structured agbcc rows have a median of 33, a p90 of 65, and a maximum of 167 — exactly one row is ≥155 lines, and it is this one. A common tail needs an if whose arms each carry several statements plus a surviving merge phi, which essentially does not occur in a 33-line function.

Known weakness

constFed is doing 98% of coalesce's filtering. Relaxing that gate alone takes the klonoa corpus from 19 candidates on 7 functions to 1002 on 82 (one function yields 93). The gate is documented as a codegen heuristic that also bounds L(L-1)/2 growth, and both are true — but it means the pass enumerates ~2% of the real opportunity, and the surviving 2% is exactly the all-literal shape this row happens to have. That is the one place where "shaped by the target function" has real bite.

Also honest about the mining: 4 of the 11 tailmerge firings are output-neutral (they fire at the tree level but produce identical C after eliminateDeadStores and basecse); 3 are in m4a, which is old_agbcc territory and discounted above; and the mined population is biased toward what lifts (368 of 770 — the declines are computed jumps, sp-as-data, and overlapping struct fields).

Debt, deliberately not folded in

• basecse, argbase and scopebase hold three divergent copies of the base-hoist policy. The end state is one l3/basepolicy.ts plus one hoist pass parameterized by placement.
• l3/ has nine hand-rolled Stmt walkers; a stmtLists/mapStmt primitive in ast.ts would make a collector and a rewriter unable to disagree by construction. coalesce.ts's local rename duplicates regspell.ts's exhaustive renameInStmt, and worse.
• Candidate growth is L(L-1)/2 with no budget. Visible already: this row's pre-existing broken raw-globals candidate now fails 3× per signedness instead of once, because the coalesce axis multiplies it.
• Whether /scopebase-coalesce should compose as an axis rather than an enumerated re-spelling.

What is still missing for the match

~13 points are unaccounted. The bitfield member spelling (.dreamStones, worth ~5) is not built — it needs the layoutOf bitfield filter relaxed, bitOffset on the member interface, a map re-vendor, and an extract-shift recognizer. A /preinit lever was built and reverted: it scored worse AND a later audit found it unsound — 9 of 13 shapes silently miscompile, with a live inhabitant in the corpus (kleod:CopyBGScrollTiles, where the if sits inside a do/while, so hoisting the default makes the arm sticky). All core tests passed with it installed, which is its own finding: the repo has a byte differ but no semantic-equivalence gate for L3 rewrites.

Verification

pnpm bench run + pnpm bench regression → 0 lost, 0 missing, 0 gained, 0 other flips over 675 rows; stale-check green, meta.asmlift.dirty: false. pnpm test:offline 672 passing, apps/benchmark 139, apps/web 34. pnpm typecheck, pnpm lint, pnpm format:check clean. Each of the five code commits typechecks and passes the suite standalone.

🤖 Generated with Claude Code