The iterate loop

Matching a function is a tight cycle you'll run dozens of times an hour. It looks like this:

1. Pick a function. Usually one that's at 0% or a low partial match. 2. Read the target asm. Pull the retail disassembly and understand what the function does — the loads, the arithmetic, the calls, the control flow. 3. Write plausible C. Your best guess at the original source (more on the "plausible 2002 C" mindset later). 4. Rebuild one .o. Compile just your unit, not the whole game. Fast loops win. 5. Diff. Compare your object against the retail object. 6. Find the FIRST diverging instruction. Not all of them — the first one. Everything after the first divergence is often just downstream noise. 7. Hypothesize and repeat. Form one theory about why that instruction is wrong ("wrong signedness," "wrong struct field," "declaration order"), change the C, and go back to step 4.

The discipline that separates fast matchers from slow ones is step 6: fix the first divergence, then re-diff. Chasing the tenth difference when the first one is shifting everything below it is wasted effort.

The tools that drive the loop

• objdiff — the interactive diff viewer. You point it at the project; it

watches your source files and rebuilds automatically when you save, showing your asm next to the target side by side with mismatches highlighted. This is where you live while matching.

• function_objdump.py <unit> <symbol> — prints the full target asm for one

function straight from the retail object. Your starting point: "what am I trying to reproduce?"

• ndiff.py <unit> <symbol> — a normalized instruction diff between the

target object and your built object. It masks out branch-target addresses and label noise so you only see real codegen differences, and it groups them into regions. It locates divergence; it does not prescribe the fix — that's your job.

Build one unit, not the world

A typical inner loop in SFA rebuilds a single object and refreshes the report:

rm   build/GSAE01/src/main/<path>.o
ninja build/GSAE01/src/main/<path>.o
ninja build/GSAE01/report.json

Then read fuzzy_match_percent, or just watch objdiff update. Tight loop, real feedback, repeat.