Prepared on 2026-10-03 to identify useful reference implementations and targets while distinguishing algorithmic similarity, existing C source, and verified matching. Versioned references were inspected locally with git show at the commits below. The newer 003BEB58 contribution is linked separately; its function-level comparison and integrated full build have passed. Dates are Git committer timestamps (%cI), retaining their original UTC offsets; they are not author or publication dates.

Key	Project / credit	Exact SHA	Commit timestamp
U	vetusmagnus / ratchet-uya-decomp	f0685ac7cc859d469aced63b9b892679e33bdee6	2026-10-02T20:23:34-04:00
R	llesieur99 / rac2-decomp	8c781ebf62b5ceff9615ad20373ba14d4070abc9	2026-10-03T00:48:47-04:00
L	mateuszklysz / Lombyte	f10ca43d95edef7596bbc1b745a28aa5129703fe	2026-10-03T01:44:38+02:00
P	Veradictus / rac1-decomp, originating from Kryštof “Lynder063” Malinda's work	82cc82dd4fc81725c2d0057e28beb264c84f66d2	2026-10-02T22:19:18-06:00
O	OpenRAC / rac1-decomp (the organisation repository that P's fork merges into)	661bb60c1fb2e8f4e335a577193b8ee4a165b8fc	2026-10-06T18:07:21-06:00

O rebuilds RAC1's libgcc from GCC's own sources (O libgcc README) and builds most of its Sony library code (newlib, libkernl, libcdvd, libgraph, libmpeg and more) with Sony's 2.9-ee compiler. UYA links the same kind of code: 56 boot_elf core functions are byte-identical to libgcc.a members, and about 440 more core functions carry Sony 2.9-ee or 2.96 fingerprints (see boot_elf.md). What carries over:

What O does	Where	Use for UYA
libgcc2.c, gbl-ctors.h, longlong.h from GCC trunk 31cf01446d (1999-09-09); fp-bit.c from GCC 2.95.3 with NO_DENORMALS backported and one shared __thenan_df; stand-in build headers	src/libgcc/	The same sources build UYA's libgcc modules. Our i5bootn matrix used the 2.95.2 release libgcc2.c and Sony's no-denormals change written in place; both work. GPLv2 with the libgcc exception.
Flags -O2 -G2 -S, -DIN_LIBGCC2 -DL_<module>, one object per L_ module; FP_DEFS := -DFLOAT_BIT_ORDER_MISMATCH -DNO_DENORMALS -DUS_SOFTWARE_GOFAST, fp-bit.c built twice (-DFLOAT for fp-bit.o)	Makefile.sn	One object per module reproduces the 8-byte alignment gaps between modules. Each division module has its own static __clz_tab in rodata, which needs its own rodata placement.
Compile through the 2.9-ee driver (ee-gcc.exe), never cc1 directly	O libgcc README	The driver supplies __mips__/__R5900__, which longlong.h needs; through cc1 alone __divdi3 and __muldi3 come out wrong.
Per-file compiler: a third column ee29 in config/core_text.objects selects a static pattern rule that compiles with 2.9-ee and an explicit -I.../gcc-lib/ee/2.9-ee-991111/include (the driver under Wine doesn't find its own include dir)	Makefile.sn, config/core_text.objects	The model for our planned per-file compiler pseudo-flag in text_parts.txt, next to @ps2as.
Linker dead-stripping rule: an unreferenced function lost its first floor(size/8)*8 bytes, leaving the last word plus alignment nop when its size is 4 mod 8; fill between objects is 0xCDCDCDCD	tools/strip_dead.py, O libgcc README	Explains linker remnants inside library ranges, the same rule our remnants follow.
A function that matches a member of Sony's prebuilt .a archives is 2.9-ee code; the archives also give real names	O docs	Same method as our libgcc.a comparison; their names can label UYA's library functions.

Their three open modules match with our compiler. O keeps __moddi3, __udivdi3 and __umoddi3 as stubs: with the Windows 2.9-ee-991111b/r4 cc1 they get the right instructions but smaller stack frames, and their README notes Sony's objects were built on Linux. Compiled from O's own libgcc2.c with the Linux 2.9-ee-991111-01 driver (-O2 -G0 -DIN_LIBGCC2 -DL_<module>), _udivdi3.o, _umoddi3.o, _moddi3.o and _divdi3.o are word-for-word identical to the members of Sony's libgcc.a (same frames: 0x10, 0x30, 0x40). This agrees with our i5bootn matrix (26 of 26 libgcc functions with 2.9-ee-991111-01, 23 of 26 with the Windows 2.9-991111). _eh.o needs GCC's gthr.h/gthr-single.h from the same era, which O's tree doesn't have; frame.c isn't in O either.

Other findings in O worth knowing (from its docs, not re-verified here):

• Sony's 2.96-ee-001003-1 pads div.s/sqrt.s/rsqrt.s with nop; nop by default (-mno-handle-ee-div-pipeline-bug turns it off); SN 2.95.x and 2.9-ee never do. A function with padded divides is a 2.96 candidate.
• 2.9-ee tail-calls a void function that ends in a call but never return f(...); it has -fstrict-aliasing on by default; it saves registers with sd in 16-byte-stride slots. These match the library signs our boot_elf core pass found.
• O's 989snd (989snd.c, about 42 functions in C) is most likely an earlier build of the sound library in UYA's SN-compiled range at 0x13B330 to 0x13D420 (inferred from names and layout, not compared): a starting point for those functions.
• __sclose and the other stdio internals right after sprintf match only under 2.95.3, while sprintf.c matches only under 2.9-ee: compiler boundaries can fall inside a library.
• License: MIT for project code (keep the notice); libgcc sources GPLv2+ with the libgcc exception; include/moby_pvars.h is GPL-3.0 (from Wrench).

Relationship	Evidence in pinned sources	Use and limitations
CRC: L FUN_0020acc0 ↔ U func_00399748	Same 0xEDB88320 initialization, 0x8000 test, 0x1F45 polynomial, and eight steps per byte. L CRC, U text.	Strong algorithmic relationship; no new port. L caps length at 0x1800; U calls a capacity function. Length and accumulator signedness differ.
Linear interpolation: R FUN_002AA140 ↔ U func_003BEB48	Same expression, a + (b - a) * t. R candidates, U text.	Scalar reference for a vector operation. C is present; byte equality between games has not been measured.
Cosine interpolation: P func_00214220 ↔ U func_003BE6A8	Explicit t = 0/1 cases followed by interpolation weighted by the cosine of t * pi. P mobyutil, U text.	Semantic relationship. P's FastCos name alone does not establish the identity of U's callee.

These references are not new progress gains. Floating-point evaluation order and signedness remain part of the matching problem.

U func_003BEB58: linear interpolation of four components. The pinned U text still contains its INCLUDE_ASM and the C implementation of scalar func_003BEB48; the new local implementation is not part of that snapshot. (Both links point at the single src/text.c from before the 2026-10-03 split. On current main, func_003BEB58, func_003BEB48 and func_003BE6A8 are in src/frontbin/3BDAC0.c, and func_00399748 is in src/frontbin/3958F0.c.) Local inspection of the generated target disassembly found two lq loads, two sq copies of 16-byte vectors, four calls to the scalar, and result stores at offsets 0/4/8/12. This observation comes from local generated disassembly, not a published ASM file in the linked sources. No target bytes or derived listing are reproduced here. The useful sibling is therefore U 003BEB48, with R 002AA140 as a cross-game reference. R's cubic interpolation 002A7AA8 implements a different algorithm. The new C implementation has a strict 152-byte match, with all four jal targets resolved to 003BEB48 and zero masks, requiring no flag override or added symbol aliases. Its integrated full build reports MATCH, SHA-1 3bc94ee895e4b4af9b5602a229af599c1103b542 (0x218924 bytes), and pr_check.py reports OK. This is a new C function, unlike the already-C references above. PR #5 is a contribution awaiting maintainer review, not a claim that upstream has merged it.

ABI pattern: two aligned vectors passed by value. The matching signature takes an output pointer, two four-f32 structures explicitly aligned to 16 bytes by value, and an f32 interpolation factor. Passing the inputs as pointers describes a different ABI. Their observed lq/sq copies must not be mistaken for proof of pointer parameters. Preserve value semantics, alignment, and the four scalar calls when reproducing this pattern with the target compiler. This is evidence for this function and toolchain, not a universal rule for vector arguments. Recheck argument placement and code generation for each target. The complete PR #5 also includes 0038E730 and 003AD650: 324 newly matched C bytes across three functions. Those two functions are not claimed as cross-game ports. No PS2/gameplay validation is claimed.

The cross-game interpolation references now also lead to concrete singleplayer common-level implementations. At snapshot df121cd88669b9de716e857336ad5c44d30ef438, the opt-in common C path covers 33 canonical functions / 6060 bytes. Every body passes complete raw compiler/member/common equality and independently resolved retail equality, with zero masks and no unexpected allocated payload. Official objdiff reports every catalogued function at 100%. Count each common owner once, not once for each overlay; the canonical configuration is unchanged.

• Level vector interpolation at 00444F70 transfers the verified U 003BEB58 implementation: 152 bytes, two aligned vectors passed by value, and four fully resolved scalar calls to level 00444F60.
• Level cosine interpolation at 004421E8 transfers U 003BE6A8: 148 bytes, the explicit endpoint cases, and the actual level cosine helper 0040C3C8. P remains an algorithmic reference, not proof that its FastCos implementation or structure layout is identical.
• The quaternion caller in the same catalogue uses observed void-returning pointer interfaces, rather than unprototyped integer-returning casts. Full call resolution rejects a wrong alias even when raw size and bytes match.

PR #8 is a contribution awaiting maintainer review. These are verified UYA frontend-to-level transfers informed by the cross-game references, not new RAC1/RAC2 matches, merged upstream progress, or playable-level validation.

• L EE types declares four-float vectors, 4×4 matrices, and qword/TI storage aligned to 16 bytes; U common provides scalar types.
• Check scalar, pointer, and long widths, effective alignment, array strides, and argument/return ABI with the target compiler. The host ABI is not evidence for these properties.
• A 16-byte storage size does not establish every structure's alignment or guarantee matching lq/sq accesses. Field offsets remain specific to the game and version.
• L's resource lookup uses 64 records, a 16-byte stride, a value at +4, and rejection value -3. No UYA counterpart is established; this is a search signature, not a table to transplant.
• UYA→RAC1/RAC2: U 003E1150/003E11D0 describes a 1024-entry hash table, a tombstone, and key-dependent probing. No cross-game sibling is confirmed. Check collisions, deletion, duplicates, and saturation before functional reuse.
• Game-structure headers require verified provenance and independent offset validation before reuse; none are copied or proposed for transplantation here.

1. Pin the source commit, regional version, reference binary identity, target symbol, address, and size. A PAL name or address does not identify an NTSC/UYA function.
2. Use sizes and normalized hashes for candidate selection; then confirm instructions, constants, data accesses, and call relationships. P's mapper distinguishes us_map from size-sequence pairing.
3. U's shared-overlay tool masks j/jal destinations when comparing relocated functions. This does not establish callee identity; its data deduplication is explicitly approximate.
4. Resolve each relocation against target symbols. Preserve $gp versus lui/lo accesses, declared global sizes, and function-local aliases; see U matching patterns. Do not apply a blind address delta.
5. Keep globals extern and retain the project's per-function overrides. R compiler notes documents a GNU-EE 2.9 profile, not a compiler prescription for UYA/SN.
6. Compare linked bytes at the exact address and size, using correct endianness and fresh artifacts. R's verifier checks reference identity and symbols; adapt its dependencies before porting it.
7. UYA requires pr_check.py and a full build reporting MATCH, as specified in its contribution rules. Similarity scores and algorithmic relationships do not replace that gate.

• CRC: zero length, boundary and excess lengths, and bytes with the high bit set; retain the distinct signed/unsigned behaviors.
• Interpolation: t = 0/1, extrapolation, and destination/input aliasing; preserve copies and operation order rather than introducing FMA arbitrarily.
• Triage: preserved raw annotations, COP0/COP2 fallback, scalar SPECIAL2 exclusion from the packed-MMI bucket, and invalid-data fixtures. An opcode family does not prove that a region contains executable code.
• Matching: reject displaced aliases, missing symbols, size discrepancies, wrong regional binaries, and stale build results.

L's license, P's license, and R's license are MIT at the pinned snapshots; retain the applicable notices when reusing code. No LICENSE, COPYING, or LEGAL file was found at the root of snapshot U. Contribution instructions alone do not establish a general reuse license. This note describes relationships and methods. It distributes no new shared implementation, binary, asset, SDK, or generated game artifact. Contributions remain subject to the receiving project's permissions and matching requirements.