stablerange: use first merge cache to skip over linear section If we are to skip over a large section of the range, skipping everything until the first merge seems like a good idea.

stablerange: add a new 'firstmerge' cache This cache store the first merge ancestors of a changesets. This is useful to avoid iteration over the changelog when building stablerange.

stablerange: drop unused `until` utility We no longer needs this function.

stablerange: use cached size data instead of walking the graph Less iteration is better.

stablesort: record previous segment size in the jump That was, we can skip over a full jump without walking it. We'll make use of this in the next changeset.

stablesort: move jump recording inside the exclusive function This will allow use to record to cache size of segment in a later changeset. If the exclusive set is empty, we need to record it outside that function.

stablerange: compute jump size after jump retrieval only This prepares the caching of jump sizes.

stablesort: warm jump cache more efficiently We no longer for a full depth iteration for all request: - we exit early for non-merge, - for merge, we only iterate over the exclusive area.

stablesort: use a regular dict for jumps This will help knowing what revision we have already queried or not.

stablerange: use the jump information for faster iteration Instead of doing a full iteration over the exclusive set, we compare the jump information instead. This perform the same kind of logic than before but will allow for much faster iteration once all the caches are in place.

stablesort: expose the jumps sequence to other code The stable range needs it to build exclusive subrange efficiently.

stablesort: use 'depth' in mergepoint tie breaker The parents with the most depth will is considered lower. It has a couple of advantages. 1) the more shallow parent probably have less exclusive revision, 2) it makes Oedipus merge behave like close to the linear case,

stablesort: rework jump gathering The old code was buggy in some case, the new code give the same result as the 'headstart' version.

stablesort: stop recording jump type There are no mature user of this "jump type" data and the current implementation is known to be buggy. So we drop the logic for now. This will make on disk storage simpler. The data will be reintroduced later

stablesort: pass a jump recording function instead of a list This seems cleaner to abstract the underlying storage to the lower level. This also makes use more flexible about jump detection.

stablesort: minor indent fix

stablesort: clarify subcall to the exclusive side Before doing deeper rework of this logic, we tackle the simplest parts.

docgraph: update test output with new output

stablesort: record, cache and reuse jump Iterating below a merge means two things: 1) iterate over the part exclusive to the higher parents, 2) iterate from the lower parents. While iterating on the exclusive part, there will be case were we just go the next natural parent, and case were we'll have to "jump" to another revision. If we record all point this "jump" happens and their target, we can easily reproduce the iteration in the future. With that information we can iterate over the exclusive part of the merge without having to compute it entirely. In addition we store the reason of the jump. This will help the stable range processing later.

stablesort: fix head start computation The recursion was wrong and provided wrong result. As a side effect, this seems to hint that `stablesort_mergepoint_bounded` is not giving the right result for complex set.

stablesort: avoid attempting to sort a tuple This seems silly.

stablerange: use the filterparents utility

stablesort: use the filtered parents utility

depthcache: use parents filter in depth cache

parents: add a utility to filter parents There is various strange case of merge out there that needs to be handled. We add an utility function to handle them all.

stablerange: abstract the bit able to store cache into sql The part about actually storing the range is independent of the algorithm, so we should extract this before using it for the next iteration.

stablerange: split pure algorithm part from the on disk cache The stable cache file is getting long, we split the logic about keeping data on disk into a dedicated file.

stablerange: add an assert to detect buggy range Nothing can be negative in there, we add an assert to make sure it is so.

stablerange: warn cache for all relevant ancestors range Warming cache in the right order will prevent recursion to happens and unlock usage on large repository.

stablesort: move parent range computation into its own method Now that we have running code, we'll need to avoid too deep recursion. We extract the piece dealing with the parent range before we write more code to warm it.