feat(functional): port C for hand-written assembly, checked against retail in an EE emulator

[Lynder063]

Summary

Hand-written assembly counts as finished for matching, but the native port still needs C for it. This adds behaviour-equivalent C for 45 hand-written functions that had none. The files are in nonmatching/functional/, each with a row in nonmatching/functional.tsv.

Checked against retail (35):

• Small helpers: bit length (func_001F9968); quadword copy and clear, and word fill (func_001F9A98, func_001F99D8, func_001F99B0); countdown timer tick (func_001F9938); set vector length in 3D and 2D (func_001F9DC0, func_001F9E10).
• Hardware and tables: DMA reset (func_0020C268); record-table fill (func_00212508); VIF1 interrupt handler (func_00235118).
• Matrix math: rotation about Z, X and Y (func_001253F8, func_001254A0, func_00125548).
• Draw and quad: func_001FA8F0, func_001F7C60, func_001F91B8, func_001F8B6C.
• Sky: triangle packet, vertex transform, tile visibility, sprites and tiles (func_0022D3F8, func_0022D2AC, func_0022D7E0, func_0022CEB8, func_0022D520).
• Tfrag and tie: texture usage, tfrag and tie lighting (func_00235EF0, func_002362B0, func_00238688).
• Moby: joint pose (func_00211808).
• Shadow volumes: func_00228D20, func_00229098, func_002291E8, func_00229838, func_002298B0.
• Level code: func_L00_001F34F0, func_L00_002963D8, func_L02_00250A58.

Written from the asm (10): these read COP0 or drive VU0 microcode, so they were not run.

• DI and EI (func_0011D960, func_0011D9A8), sharing a host-side enable flag.
• Four interrupt-guarded syscall wrappers (func_00119328, func_00119390, func_001193F8, func_00119460).
• The fatal trap (func_001F9978) and a busy-wait delay (func_001F9988).
• The moby skin/light job runner (func_00212578).

How the checks were done

Retail and the C ran side by side in a MIPS64 emulator (unicorn). A small instruction model emulates the R5900-only instructions: MMI, lq/sq, VU0 macro mode, the EE COP1 extras and ei/di. The C was compiled with the real SN compiler and loaded next to retail, with its references resolved to retail addresses; level functions run with their overlay dump loaded.

Each test case starts both runs from the same RAM. It then compares the return values, all of RAM outside the stack, the scratchpad, and the DMA/GS/VIF register pages. Each file's header gives its case count, and any float tolerance used for VU0-versus-FPU rounding (at most 1e-5). All files also build with host gcc, using the same short type names the existing port files use. The harness is not part of this PR.

Not included

• Kernel syscall stubs: the 85 stubs of the form li $3, n; syscall get no C, because the port supplies the kernel.
• Custom-ABI helpers: five helpers take inputs in non-ABI registers and are only called from other hand-written code. They belong inside their callers' C.
• Unfinished: the following were not done when work stopped: func_0020FC38, func_00228A58, func_L00_001F2BE8, func_L00_00252D20, func_L00_002539C8, and the remaining large VU0/MMI routines.

🤖 Generated with Claude Code

https://claude.ai/code/session_017Lr3jLwpQ8QBWgz9eA4WNx