Method to extract entire e-graph

[goombamaui]

Is there a method to extract the entire e-graph in python? For more complicated cost extractions (like using partial orders and trying to extract all equivalent terms whose costs are not greater than another), it might be easier to work with a python version of the e-graph.

[saulshanabrook]

There isn't currently no. There is the serialized form, but it isn't exposed currently in a type safe manner.

Is this currently a need you have? A type safe wrapper around traversing the serialized egraph could be added.

[goombamaui]

Yes a method to traverse the serialized e-graph would be very useful; there is a lot of analysis I would like to do on the egraph once it has finished saturating the rules which would require a way to traverse it in python.

[saulshanabrook]

Sounds good. I won't have time to get to this immediately. If you wanted to have a go at adding it, these are a few thoughts I had on how it could be implemented:

1. Change the Value exported in the bindings to be a u32 instead of a value type, by using egglog_numeric_id to convert to/from u32. Also update the ValueDecl in declerations.py to just reference an int now. Add a get_value(b: BaseExpr) -> int | None in deconstruct.py that takes an expression and returns the value of it, if it is a value expression.
2. Implement serialize in rust by returning a mapping of serialized form. This should return a mapping of values to a list of tuples of SerializedNode, args (list of values) Use to_node_id from Rust so that you can lookup the function name or primitive value of an op. And use class_id_to_value to lookup the value of each class ID.
3. Add a high level .serialize() method on the e-graph which uses the low level rust .serialize() and outputs a dict[int, list[tuple[BaseExpr, int]], where the BaseExpr are either all primitives or calls whose args all points to values, which can be looked up as e-classes.

This should provide a type safe way of inspecting the resulting e-egraph, by using the methods in deconstruct to parse each term.

Maybe not the most accessible first issue, but happy to support anyone who is interested in taking it on.