[experiment] Special handling of root node in AcyclicGraph algorithm

[ppedrot]

The most expensive kind of level constraint in the current algorithm is of the form "enforce u ≤ Set". Indeed the forward pass has to crawl basically the whole graph starting from Set, as all levels are above Set. This is the worst case for the complexity of our variant of of the Tarjan algorithm, which expects the graph to be sparse. Unfortunately, in practice such constraints seem to be commonplace in universe-polymorphic code.

To work around this issue, in this PR we try to handle Set in a special way by not storing it in the graph as a normal node. Instead we assume "Set ≤ u" implicitly everywhere, and we only keep a set of constraints "Set < u". We still have to perform a kind of forward pass to merge a level with Set, but we do so hopefully efficiently by leveraging the topological order maintained by the algorithm. We add in particular a map to quickly access nodes by their topological index.

[ppedrot]

@coqbot bench

[ppedrot]

@SkySkimmer the bench died with a weird error in an unrelated script... Any idea what went wrong?

[ppedrot]

Let's retry: @coqbot bench