So what you actually need is a program that navigates through the huge image of the die and detects if the structure that is looking at is a 1 or a 0? This at the fundamental level is a cross between machine learning and image processing?

▲

electroly 2 hours ago | parent | next [-]

I helped out on this image-to-bits transcription, doing manual verification of the automated work. I did the whole thing by hand: I sliced the ROM images into strips that excluded parts of the image that don't encode bits, used my tablet and stylus to manually place a black dot on every 1 bit, then wrote a trivial program that detected the presence or absence of the black dot in each cell. From my perspective, the ROM is organized like a series of "ladders" where the 1 bits are missing legs of the ladder, and I was placing dots on the missing legs. I compared my results with the ML output and manually re-checked each bit where we disagreed.

http://brianluft.com/images/2026/05/386_microcode_bits.jpg -- my fully annotated result. I was working from a higher-quality PNG; this is highly compressed because it's a big image.

▲

bri3d 6 hours ago | parent | prev [-]

Yes, exactly. Historically you would make some simple image processing software that will align the grid and then look for properties at each specific bit position. Usually die shots are highly imperfect (the delayering usually leaves some artifacts or damage) so frequently merging multiple scans is important as well. Travis Goodspeed has a neat tool for this workflow at https://github.com/travisgoodspeed/maskromtool and the blog mentions John McMaster’s bitract: https://github.com/SiliconAnalysis/bitract although I think most people working on these projects usually just one-off it as the mentioned Discord users in the blog post eventually did.

More modern devices are of course more difficult due to layers, feature size, and less visually obvious ROM bit designs.

Anyway, the impressive part of this project was really understanding the undocumented microcode assembly language through inference and trace following; the 1s and 0s look like they were the easy part!

	▲	photochemsyn 3 hours ago \| parent [-]
		The full workflow seems to look something like this, with the added complications relative to the 8086 microcode being that the 80386 microcode acts as an orchestration layer on top of hardwired engines, programmable logic arrays, and fault/protection redirection. The 8086 microcode does all that algorithmically, reusing the same hardware instead of having dedicated transistors. 1. Extract the ROM bits. 2. Determine physical-to-logical bit ordering. 3. Identify microinstruction boundaries. 4. Infer field boundaries. 5. Associate fields with hardware destinations (check with die tracing). 6. Decode instruction-dispatch programmable logic arrays. 7. Associate x86 instructions with microcode entry points. 8. Infer repeated idioms: moves, ALU ops, termination, calls, tests, redirects. 9. Decode accelerator protocols. 10. Validate against known architectural behavior.