[dktue]

Gain: Lake Huron, the slowest relation on the planet, 103 s → 25 s (−76%); Germany wall 43.6 → 39.3 s (−10%, three rounds); planet relations phase 310 → 263 s (−15%), with the planet total unchanged within run-to-run noise. Output identical.

When a multipolygon is invalid, tilemaker repairs it with make_valid, i.e. geometry::correct() in include/geometry/correct.hpp. There, result_combine merges each corrected ring into the result by testing it with boost::geometry::intersects() against every ring added before it. That is quadratic in the number of rings, and each test walks both rings.

A few huge water and landcover relations in the planet are self-intersecting and have thousands of inner rings (islands). They dominate the relations phase, because each one runs on a single thread while the others wait. Lake Huron (r1205151, 523 outer and 14,233 inner ways) spends 88 of its 103 s in make_valid, nearly all of it in result_combine.

This PR keeps the envelope of each ring next to it and skips intersects() when the envelopes are disjoint, since rings with disjoint envelopes can't intersect. The rings are merged in the same order as before, so the output is unchanged. It also adds a unit test for correct(), which passes with and without the change.

Slowest planet relations, timed on an extract with just those relations

relation	before	after	change
r1205151 Lake Huron (14,233 inner)	103 s	25 s	−76%
r3400721 Guri Reservoir (2,574 inner)	69 s	63 s	−9%
r13985343 Lake Nasser (6,758 inner)	54 s	54 s	0%
r20516020 wood (210 inner)	44 s	44 s	0%
r13287625 grassland (1,518 inner)	40 s	30 s	−25%

Germany, 32 threads, three rounds each

	before	after	change
wall	43.45 / 43.80 / 43.60 s	39.24 / 39.04 / 39.52 s	−4.3 s (−10%)
user CPU	1039 / 1043 / 1042 s	1025 / 1023 / 1029 s	−1.5%

Most of the gain is wall time: fewer threads wait at the end of the relations phase for one large relation.

Planet

Sharded --store run, 32 threads (one run each):

	before	after	change
relations phase	310 s	263 s	−47 s (−15%)
slowest relations pass (3 of 6)	119 s	79 s	−34%
wall	2484 s	2487 s	within noise

The relations phase gains 47 s. The total doesn't show it, because the tile phase, which this PR doesn't touch, took 140 s longer in this run; identical planet runs vary by tens of seconds there.

Output

compare_area (per-tile area/length per layer and tag group) finds all layers identical, both for Germany and for the full tile output of the five relations above. On Germany only poi differs, which already differs between identical runs.

Baseline for all measurements: master + #961 + #966 + #968. It changes the same file as #974 (correct.hpp), and the two apply together cleanly.

🤖 Generated with Claude Code