Yosys version has large impact on results

[matth2k]

I moved from Yosys 0.33 to 0.47, and I noticed that the results are significantly worse. The initial farea from Yosys is larger, and the optimization percentage is less. It is interesting that Yosys from 1 year ago seems better for FPGAs?

In any case, this isn't too unexpected on the e-graph side. Since the main configuration we use is "smart lutpacking, we don't really remove the structural bias of the input netlist. So a worse input means a worse output.

We'll have to look into ways to remove structural bias (decomposition) while still making EqSat and extraction feasible. Even without decompositions, an ILP run of some of our benchmarks can use 150GB of RAM for over 24hrs without a solution. I need to look into if a 1 or 2 hour timeout is reasonable?

Of course, we also need to look into an iterative approach instead of just using more RAM and time.