[cadelass]

What

The second and third batches reserved in my #152 comment: 31 functions in total.

• 16 are matched, including all 15 members of family F7.
• 2 are classified as handwritten (ASM_FUNC).
• 13 are kept as near-miss notes.

Matched (16)

Family F7: 15 level trigger updates, 1540 bytes each, the same opcodes in every level.

• func_L06_00455BC0 and its copies: func_L00_0040EAA0, func_L01_00410AD0, func_L02_00439CD8, func_L05_00414F90, func_L10_00415A50, func_L11_0046D660, func_L12_00417690, func_L13_0041CD18, func_L14_00421BC0, func_L15_0041F2F8, func_L17_00417DB0, func_L18_00436B80, func_L20_00438F98, func_L22_0040F1C0, func_L23_00415988.
• Each copy has the same C in its own unit, <file>-f7-g8.c, built with -O2 -G8 -ffunction-sections -malign-loops=2.
• This closes the near miss that docs/MATCHING-BLOCKERS.md and the old func_L06_00455BC0 note left open, "where the 4th reference for missing comes from". Three things were needed:
  • The if (missing) func_0026FEB8(...) debug print sits in a do { } while (0), the usual shape of a debug-print macro. That counts the test at loop depth, which gives missing its fourth reference. It then outranks fail and takes $s5, as in retail. The loop is not empty: it holds the real test and the call, so it is not the empty barrier fix(tools): validate EE diagnostics and reject empty source barriers #257 removed. I would still like a reviewer's view on it.
  • That loop's label is not aligned in retail, hence -malign-loops=2 for the unit. The function's own trigger loop label already falls on an 8-byte boundary, so the flag changes nothing else.
  • func_00288BE8 is declared with its real return type, int. Other callers test its result, for example L00_00370408-edges.c. This gives retail's $v1 for the count load after it.
• The old note, its nonmatching/levels/ siblings (including alternatives/func_L14_00421BC0-count.c) and the 15 port-only files in nonmatching/functional/ are removed, as in 588f8f0.

Other match: func_L04_00418BF0, the hoverbike rail trick (Barlow and Joba), in src/levels/shared/L04_00418BF0-g8.c.

• The colour and the clamp factor share the variables that retail keeps in $a1 and $f20, so the four lerp calls cross-jump into one.
• func_00273318 is declared with its int 0x22 ahead of the float 0.0f. Each register bank keeps its order; only the argument setup order changes.

Handwritten (2)

func_L04_0032F994 and func_L04_0032F62C become ASM_FUNC: they use $at as a working pointer and a trapping addi. Per AGENTS.md this needs a maintainer to agree.

Near misses kept in nonmatching/ (13)

The src/ markers stay INCLUDE_ASM. Each file's header gives the profile, what mattered and what still differs, and each function has one row in nonmatching/README.md.

Function	Result	What differs
func_L00_003ABEC0	2588/2588	callee-saved register allocation
func_L03_003E3A00	1856/1860	allocation (path record, colour, node)
func_L04_003FBEF8	1912/1920	which saved registers hold the counters, path record and floats
func_L04_00402D90	1948/1952	saved-register roles (self, rival record, fresh)
func_L04_00419918	5132/5120	scheduling and branch-likely filling
func_L09_0037FCE8	2492/2488	frame 0x180 vs 0x1F0 and spills
func_L15_004307E0	848/848	only local-alloc's $v0/$v1 in one decrement
func_L06_00443A38	936/936	tuning-store order and two FP constant registers
func_L06_0041EF08	920/916	cse carries a constant 3 into a saved $s2
func_L05_0041EF00, func_L15_0042EFE0	3420/3420	9 rows of argument setup order
func_L25_0040DE40, func_L15_00431590 (F35)	1840/1852	per-block constant reloads

The 3420/3420 result for func_L05_0041EF00 and func_L15_0042EFE0 depends on an empty do { } while (0). That is not a landing form (#257), and their notes say so. They are kept because they show where the block boundary has to be.

Released along the way (in the #152 comment)

• Seven reserved functions that feat: USA to 84.6%, USA v2.00 to 80.9%, PAL to 54.1%, the lock-on tether family and two compiler fixes #386 had already covered.
• Seven catalogue rows that are not C:
  • Five rows start with 16 bytes of the previous function's epilogue. They need an entry correction, as v2.00 already has.
  • Two are one-word remnants in level 25.
  • The analysis is in the comment. Fixing them needs the PAL and v2.00 dumps, which I don't have.

Checked

On master fe9a7138 with the whole branch applied, USA v1.01:

• python3 tools/build.py all --strict: 28 programs OK, exit 0. This was run before the last commit; that commit's five levels (00, 01, 14, 17, 20) were then rebuilt with --strict: all OK.
• python3 -m unittest discover -s tests: 339 tests OK.
• python3 tests/check_repository.py: repository checks passed.
• The compiler was rebuilt from master before measuring.

I did not run PAL or USA v2.00, because I don't have those releases locally. func_L04_00418BF0 is in the v2.00 catalogue, so its C will also link there. progress/report.json and the README table are not touched.

🤖 Generated with Claude Code

[bytematchdev]

ByteMatch progress report

SCES_516.07

Code matched	54.15% · 3.09 MB of 5.70 MB
Fuzzy match	54.15%
Functions	6,996 / 8,672
Data matched	0.00%
Units complete	958 / 1,196

No function changes.

SCUS_972.68

Code matched	84.74% · 4.68 MB of 5.53 MB
Fuzzy match	84.74%
Functions	7,760 / 8,579
Data matched	0.00%
Units complete	924 / 1,018

No function changes.

SCUS_972.68_v2

Code matched	81.01% · 4.52 MB of 5.58 MB
Fuzzy match	81.01%
Functions	7,595 / 8,634
Data matched	0.00%
Units complete	973 / 1,156

No function changes.

_{Commit 771efdd · Ratchet & Clank: Going Commando on ByteMatch}