bench: 8 more real functions per project — close the size/feature blind spot (192 → 240 rows)

[macabeus]

Adds 8 real-tier functions per project — 48 rows, 192 → 240 real, 723 total. Board: asmlift 350 match, m2c 342; no pre-existing row moved.

Why

PR #13 shipped an unconditional L3 pass and exactly one benchmark row changed. Instrumenting the pass over all 675 rows showed why: it fires on 1. The same instrumentation over a whole game's worth of klonoa functions showed it fires on 11, seven of which change the emitted C. The difference was size — the shape needs an if whose arms each carry several statements, and the real tier's medians were 9–15 source lines with maxima of 17–71. A capability could be right or wrong and the suite would report the same number either way.

How the rows were chosen

Every candidate was mined from the checkout, then compiled standalone and run through the raising tower before selection, so each entry's build status and outcome were known before it landed. Hit rates were low enough that this mattered — roughly 60/93 for kleod, 44/392 for sa3, 33/150 for pokeemerald, 15/160 for af, 16/70 for marioparty3, 13/70 for snowboardkids2. These counts come from a throwaway screening script that is not in the repo, so take them as reported, not verified; what is verifiable is every row's landed outcome, below.

Where a project had more compiling candidates than slots, the picks are balanced between functions asmlift emits (so gap size can move) and functions that expose distinct capability gaps. On the three MIPS projects that balance was not available: screening ~300 candidates produced zero that asmlift lifts, because essentially every real MIPS function contains a call or a branch-likely. (Same caveat — the screen is not committed. The consequence is visible in the tables: all 24 MIPS rows decline, and every decline is a jal/jalr or a branch-likely.)

Every prependC is verbatim project text — a source file's own #include block, its declaration preamble, or a struct copied from where the project declares it. No authored types.

All 24 new GBA rows were verified to reproduce their project's own build object byte-for-byte (relocations and Thumb bl operands masked, since a call to a same-TU symbol is already resolved in the project's object).

---

The 48 new rows

Outcome key: diff:N = compiles, objdiff distance N. declined = explicit incompleteness markers. noncompile = marker-free output that fails to build. failed = no usable output.

kleod — agbcc (prior rows: max 58 lines)

function	loc	asmlift	m2c	why this function
ProcessInputAndUpdateEntities	136	declined — jump table	failed	Largest kleod pick, and the largest kleod row in the manifest; switch dispatch over nested loops. Pure size coverage — 2.3× the previous kleod maximum.
UpdateCameraScroll	99	declined — pool word is a symbol+offset	noncompile	Passes a struct by value to ScrollBGLayer, a shape no row had, and it is the only row that declines on a symbol+offset pool word. (It was also picked as a /coalesce inhabitant, but it declines at lift, so it never reaches L3 — that motivation did not pan out.)
ConfigureEntityBehavior	94	declined — jump table	failed	A mov pc, r0 jump-table switch at 94 lines. kleod's one prior switch row (TransformSingleEntityToScreen) is also a jump table with the same decline, so this adds size, not a new gap.
CountCollectedGems	85	declined — sp-as-data	noncompile	2-D array indexing under nested loops. Declines on sp-as-data (spill slots — the function takes no addresses).
UpdateWorldMapNodeAnim	76	declined — jump table	failed	Switch + soft-division + shifts in one function; the densest arithmetic mix available.
EntityItemDrop	64	diff:148	noncompile	Switch over entity types reaching a genuine bitfield (u32 unk_204_2 : 2) — the only real bitfield row in the set.
SetupBG3WindowOverlay	56	noncompile	declined	Picked as a /coalesce inhabitant from the klonoa mining (candidate counts not recorded in the row), plus array writes into VRAM-shaped memory. Lands as noncompile — see the findings below.
EntityDeathAnimation	50	noncompile	declined	A real multi-statement do-while loop. kleod already had one do-while row (StrCpy), so this is depth, not a new feature — and the noncompile it produces is the useful part.

pokeemerald — agbcc (prior rows: max 25 lines)

function	loc	asmlift	m2c	why this function
TrySetCantSelectMoveBattleScript	105	diff:303	noncompile	The largest pokeemerald function that both compiles standalone and asmlift emits — 4× the previous maximum, and it keeps a gap-size signal at that size.
GetMoveTarget	84	declined — jump table	failed	Switch → jump table and a do-while in the same function.
SetMauvilleOldManLanguage	64	diff:59	noncompile	Takes a union OldMan * and switches on its tag — no prior row on any project (real or synthetic) mentions union. Two other new rows reach unions too (af:_MtxF_to_Mtx, sa3:sub_804DC38); this is the one that both compiles and scores.
Cmd_tryconversiontypechange	60	declined — unrecovered back-edge	noncompile	A do-while whose loop shape asmlift's recovery declines — a structuring gap rather than a lifting one.
UpdateShoalTideFlag	39	declined — sub-word .byte table	declined	Reads a static-local .byte table; the only row of its kind.
AcroBikeHandleInputTurning	38	diff:44	diff:2	u8-returning static callees whose return truncations are load-bearing; also the closest head-to-head on the new rows.
AnimTask_FlashHealthboxOnLevelUp_Step	36	declined — union overlap	noncompile	asmlift's recovered struct has overlapping fields (an artifact of s16 data[] in struct Task), which its union modelling declines. Nothing in the source is a bitfield.
DoForcedMovement	29	diff:31	declined	Indirect call through a function-pointer parameter (void (*moveFunc)(u8)), compiled to bl _call_via_r9.

sa3 — agbcc (prior rows: no switch at all, max 71 lines)

function	loc	asmlift	m2c	why this function
PackSaveSector	117	declined — sp-as-data	noncompile	Largest sa3 pick; three switches driving a field-by-field save-sector copy. Declines on sp-as-data — from spill slots, not an address-taken local.
sub_804DC38	92	declined — union overlap	noncompile	Switch + soft-division over a struct whose fields overlap (a union asmlift cannot model).
Sio32MultiLoadMain	77	declined — jump table	failed	Switch → jump table driving MMIO serial registers.
sub_802DFC8	65	declined — branch target is a data label	match	The only row whose decline is branch target is a data label. Also the clearest m2c-wins row on the new set. (Selected believing it had bitfields; it does not — see the correction note below.)
sub_806132C	54	noncompile	diff:171	Mid-size entity code asmlift emits. Picked so gap size could move; it ends up noncompile instead, which is the finding below.
sub_804D360	47	diff:57	diff:6	A switch asmlift does lift, so sa3 finally has a switch row with a score.
Sio32MultiLoadIntr	38	diff:76	noncompile	MMIO register work in an interrupt handler, and one of the four sa3 picks that produce a score.
sub_803213C	30	noncompile	match	Small enough to isolate one behaviour, and it produces a noncompile. (Also selected on a false bitfield reading — see below.)

af — ido7.1 (prior rows: max 31 lines)

Every af pick declines, with six distinct reasons. af's asm is FP-heavy and call-heavy, which is exactly what asmlift's MIPS frontend does not model yet.

function	loc	asmlift	m2c	why this function
Skin_Matrix_MulMatrix	122	declined — sdc1 store	noncompile	Largest af pick; 4×4 float matrix multiply, maximum FP register pressure.
_MtxF_to_Mtx	71	declined — lwc1 load	noncompile	MtxF and Mtx are both unions, indexed as 2-D arrays, with float↔fixed bit twiddling.
suMtxMakeSRT	67	declined — jal	declined	Float matrix construction that also calls out.
add_calc_a	53	declined — bc1tl	diff:98	FP compare feeding a branch-likely — a distinct decline from the store/load ones.
get_percent_forAccelBrake	51	declined — bc1fl	diff:78	Float division plus the opposite branch-likely sense.
Skin_Matrix_SetQuaternion	47	declined — sdc1	noncompile	Quaternion math with double-word FP stores.
sAdo_Get_KokeruLabel	43	declined — jal	diff:4	The only switch available in af's compiling set; af had none.
atans_table	36	declined — bc1f	diff:27	Quadrant dispatcher over float compares; the only af row whose decline is a plain bc1f rather than a branch-likely. Despite the name it reads no table — it calls U_GetAtanTable.

marioparty3 — gcc2.7.2 (prior rows: max 17 lines, six features total)

The narrowest project in the suite: its 32 rows only ever exercised struct, pointer, branch, global, arithmetic and loop.

function	loc	asmlift	m2c	why this function
ContDataUpdate	50	declined — blezl	diff:37	Largest marioparty3 pick — 2.9× the previous maximum — a controller-state loop.
func_80012B14_13714	41	declined — jal	noncompile	float 3-D projection with division; marioparty3 had no float row.
Hu3DCam3DToScreen	37	declined — jal	noncompile	The sibling projection path, 2-D array indexed.
func_800230F8_23CF8	36	declined — beql	noncompile	Struct/pointer loop hitting a different branch-likely form.
HuSprGrpCreate	35	declined — jal	declined	Allocation + link loop over a sprite group.
DecodeData	34	declined — jal	diff:13	Contains a MIPS jump table (jr v0), a shape this project never had — but asmlift declines earlier, on the jal at 0x40, so the table is not what this row currently exercises.
FileSeek	25	declined — beqzl	noncompile	Switch + ternary in one small function.
HuMemAllocTag	24	declined — jalr	noncompile	Indirect call through a function pointer — distinct from the direct-jal declines.

snowboardkids2 — gcc2.7.2kmc (prior rows: max 52 lines)

function	loc	asmlift	m2c	why this function
func_8005DF10_5EB10	125	declined — jal	noncompile	Largest; 2.4× the previous maximum.
func_8001E104_1ED04	114	declined — bnezl	noncompile	Branch-likely–heavy loop, a different decline from the call-blocked ones.
func_8005AE8C_5BA8C	109	declined — jal	declined	do-while + int64 + goto in one function.
func_800B2C18_A2AC8	98	declined — jal	declined	goto-based control flow.
func_8005AB58_5B758	82	declined — jal	noncompile	64-bit arithmetic over a memory buffer.
func_800B3F2C_A3DDC	80	declined — jal	declined	Switch dispatch.
func_800B3784_A3634	79	declined — beqzl	declined	A second switch of similar size that blocks on a branch-likely rather than a call, so the pair separates the call gap from the branch-likely gap. Neither reaches the switch.
func_80038090_38C90	79	declined — jal	declined	Struct/pointer loop from a fifth file — the eight picks span five source files.

---

Corrections after an adversarial review of this description

A hostile fact-check of the table above (48 rows, every claim checked against the manifests, the project checkouts and results.json) found six false rationales. They are fixed above, and one of them was a dataset bug, not just prose:

• Three rows were tagged bitfield and are not. The selection-time detector's regex matched ternaries (? 6 : 2;). struct Task is all scalars, and sa3's SPRITE_FLAG is ((value) << SF_SHIFT(name)) — a shift into a plain u32. Retagged in a separate commit; conversely kleod:EntityItemDrop reads a genuine u32 unk_204_2 : 2 and was not tagged, so it now is. Outcomes and scores are unchanged.
• kleod already had a do-while row (StrCpy) and its prior switch row (TransformSingleEntityToScreen) is already a mov pc, r0 jump table with the identical decline marker. Those two picks add size and depth, not new gaps.
• marioparty3:DecodeData declines on jal, not on its jump table — the table is there, but asmlift never reaches it.
• pokeemerald:DoForcedMovement dispatches through a function-pointer parameter, not a table.
• kleod:UpdateCameraScroll cannot exercise PR feat(l3): common-tail merge, scoped base clusters, and ranked live-range coalescing #13's coalescing lever — it declines at lift, so it never reaches L3.
• Two sp-as-data rows were described as having address-taken locals; neither takes an address. The marker's parenthetical is a disjunction and the real cause is spill slots.

What this immediately found

asmlift emits calls that do not type-check against the project's own declarations. Four new rows are noncompile — a first for asmlift on this board:

• sa3:sub_803213C — too many arguments to function VramMalloc
• kleod:SetupBG3WindowOverlay — too many arguments to function m4aSoundVSyncOff
• sa3:sub_806132C — assignment makes integer from pointer without a cast (VramMalloc returns void *)
• kleod:EntityDeathAnimation — surfaced only after the row was given its project's real u8-returning callee declarations

asmlift recovers call arity from register usage; where the project declares the real prototype and the recovered arity disagrees, the candidate fails to compile in the project's own context. The previous 192 rows never surfaced this.

11 pre-existing rows do not reproduce their project's own build. Auditing all 120 agbcc rows against the projects' build objects: kleod 40/40 equal, pokeemerald 32 equal / 8 differ, sa3 37 equal / 3 differ. Ten differ by -fprologue-bugfix (the canonical flag set has it; pokeemerald never enables it, and sa3 only for BUILD_NAME == sa1), and sa3:VerifyFlashSector_Core because the project builds lib/agb_flash at -O1. For those rows "match" means reproducing our canonical re-compile rather than what the project's build emits. That is the documented design, so changing the canonical flags — which would move rows board-wide — is left as a decision rather than folded in here.

On old_agbcc: m4aSoundVSyncOff is an old_agbcc function (defined in src/m4a.c, which kleod builds with old_agbcc -mthumb-interwork -O2), but its caller SetupBG3WindowOverlay is byte-identical to kleod's own build — the callee's compiler never reaches the caller, which only emits a bl. Only the TU a function is defined in matters for eligibility.

Verification

pnpm bench run + merge → 723 rows, meta.asmlift.dirty: false, stale-check green. pnpm bench fidelity: 1446 script runs — 1446 ok, 0 warn, 0 fail. Only one pre-existing row differs from the pre-#13 snapshot, and that is #13's own effect. pnpm typecheck, format:check, lint clean; 139 benchmark tests and 34 web tests (including the summary-consistency provenance gate) pass. A second full run on this branch reproduced every row measurement-identically.

Supersedes #14 (same content, clean two-commit history).

🤖 Generated with Claude Code