For matching decompilations like this, pretty terrible. It can give a rough layout of a function with some branching, but it fails to create reasonable human like structs (which have to be inferred from the assembly), and matching what it has created can be the majority of the work. We did this without any AI assistance, but I relied on the Ghidra decompilation feature for outlining a functions layout (though even that had it's limitations).

The training data is just too minimal for this sort of thing. The decomp.me database would probably be really good to train a model on.