Allow single edge paths in MLD alternatives, #4691

[oxidase]

Issue

The PR fixes #4691, by allowing alternative routes with packed.path.empty().
Also an additional check node != fst.GetData(node).parent was added to prevent infinite loops if has_plateaux_at_node(node, fst, snd) and node is the first (last) node of the route.

Tasklist

• add regression / cucumber cases (see docs/testing.md)
• review
• adjust for comments

[daniel-j-h]

Looks good - wondering if the check can be made earlier when we generate the candidate via nodes. Then we could eliminate the edge case in a single location, no?

[oxidase]

@daniel-j-h such alternatives are ok for CH, so it is better to keep them also in MLD. This requires edges vector emptiness checks in all filters that use packed paths.

[daniel-j-h]

Yes I was only talking about the MLD alternatives implementation.

What I meant was removing these candidates early on during stepping already so that we don't have to think about this edge case when adding new filters later on.

But now that we have a reproducible test case I'm fine with both approaches.

[oxidase]

Unfortunately CH alternatives results are not consistent wrt OS and the test fails on OSX and Windows builds

[oxidase]

@daniel-j-h i looked at the test fail on windows build. The problem in a loop shortcut that is not handled for CH alternative routes in computeWeightAndSharingOfViaPath so the alternative path weight instead of 7 at

osrm-backend/src/engine/routing_algorithms/alternative_path_ch.cpp

Line 308 in a53794f

*sharing_of_via_path += facade.GetEdgeData(edgeID).weight;

is the weight of the complete loop 6.8+7=13.8 and is filtered out. I did not find the reason why the loop shortcut is added with an msvc build. Both issues are completely out scope of the PR, so i added @mld-only tag to skip the test for CH.