All good except one fill pattern render test! Rendering is super-blazing-fast.

At first I thought some sliver lines I saw were earcut artifacts, but they turn out to be present on master as well:

The fill-pattern/zoomed test looks very good actually, and there is definitely something fishy about the diff algorithm. Should we add it to ignore list @jfirebaugh? Otherwise the PR looks great.

We should update the expected rendering for all of the fill tests.

@jfirebaugh @mourner Great work! Do you have an estimate on how long it will be until this can land in a release? It resolves #930, which is blocking a client of ours.

The blocker for adopting earcut is getting confidence that the tessellation is good enough. This is a function of both the quality of our vector tile generation pipeline, and the robustness of earcut itself, and myself, @mourner, @lbud, @flippmoke, and @springmeyer are all pushing on this at the moment. We should have a better estimation next week.

Rebased this on master (with #1645) and pushed to earcut-final-final.

@lucaswoj Looks like earcut-final-final is still pointing at the same commit as earcut-final. Feel free to just force push earcut-final.

Ok. Force pushed here and deleted earcut-final-final. (Just in case, previous SHA was 44ce255)

@lucaswoj let's rebase on master to make sure it's in mergeable state

Done @mourner 🚀 (previous SHA was 494bb19)

We can close #930 when this lands

Do you have a fresh estimate on when this could be merged? I'm making heavy use of fill-layers and this branch is much more performant than master.

This is ready to ship pending

• a final code review of this PR
• a code review of Fix query tests for v2 features mapbox-gl-test-suite#91

cc @mourner @jfirebaugh @jakepruitt @ansis @springmeyer

Let's eliminate convertCoords, rewriting classifyRings and the code that generates the flattened array for earcut in terms of the {x: x, y: y} point-geometry objects. This will both simplify the code and avoid GC churning a large number of 2-element arrays.

@jfirebaugh yep, done.

The test failed because of some really weird Git/NPM issue — tried rebuilding without cache and it still failed with some "git bad object" errors. Not sure whether this is something temporary or broken in one of the subdeps.

Ah, it failed because the test suite ref it pointed to no longer exists for some reason.

OK, rebased again with updated test-suite and it passes.

👍 🚢 🎉

\o/ Thank you everyone!!