Alexandre Moine

Hi, Sorry for the delay of my response, please feel free to submit any PR! Otherwise, I can have a look during holidays.

Hi, > While 1.17 may just be noise, 1.66 feels like a proper regression The two are proper regression in this case :) About the `SPECIALIZE` pragma: It is here...

So I tried with `doctest`. Took some time to set up (the need to hide some of Prelude, etc), but I came to a partial implementation of tests of `Algebra.Graph`....

> but there may be others :-) Mmhhh, will need to implement more of the file to see if there is really a problem. > . I think these references...

> I have been wondering recently: do we even need graphs that can be empty? With the actual `Empty` constructor, I don't think so. But providing an encapsulated version using...

> Perhaps, it prevents some important optimisation? It was the case for `edges`. Habitually, there is a merge between a `foldr` and `map` (this is what we used to speedup...

In fact I need to do the change in `edges1` and use it from `Algebra.Graph.edges`. ```Haskell edges1 (x :| xs) = foldr (Overlay . uncurry edge) (uncurry edge x) xs...

The last benchs: ``` isEmpty: 0.63 (good) vertexList: 1.06 (OK) VertexCount: 1.07 (OK) hasVertex: 5.13 (bad) edgeCount: 1.09 (OK) edgeList: 0.76 (good) hasEdge: 0.40 (good) hasSelfLoop: 1.17 (bad) addEdge: 1.08...

Sorry, I wasn't clear. I actually test 4/5 vertices (for a graph with n vertices I test the search of vertex 0, n/3, 2*n/3, n+1). But the difference is actually...

Yeah indeed ^^. > Perhaps, into a separate Travis instance I totally agree, I will send a PR