[geos-devel] CoordinateArraySequence == CoordinateSequence

[Paul Ramsey]

One of the things that tinkering with the SegmenString layer of overlay brought out to me was the extent to which we construct CoordinateSequence almost exclusively out of CoordinateArraySequence. Like, all the time. At yet, because we handle those CoordinateArraySequence at the API level almost exclusively as  CoordinateSequence we lose the ability to do some handy optimizations. 

Like, if one were going to (as one does on every single CoordinateSequence that enters the overlay code) 
(1) test if there are repeated points and 
(2a) remove any if there are
(2b) just return the untouched CoordinateSequence if there aren't
a useful pattern would be for ::hasRepeatedPoints() to return/populate a list of indexes at which repeated points appear and for ::removeRepeatedPoints() to do bulk copies of all the points in between those indexes. This is foreclosed by the CoordinateSequence API, you can play this trick nicely with a std::vector living underneath, but the API doesn't let us see that (in fact) that's what we have 99.9% of the time.

So, one obvious thing to do would be to remove the virtual methods in CoordinateSequence and pull the implementation up to that level, std::vector and all, and give up on the idea of an abstract interface that we don't actually use. For a handful of use cases, where data access cost is greater than computation cost (area, length, distance(?), some others (?)) this might be "bad" in some theoretical way, but note that currently we still don't actually have that abstract layer in place for a zero copy computation. Removing the virtual methods and inheritance from CoordinateSequence would foreclose an option that (a) we seem unlikely to ever deliver on and (b) has narrow performance benefits even if we did deliver on it.

Meanwhile, the flip case seems to likely have a *lot* of performance benefits just hanging around waiting to be harvested. Coordinate access without going through the inheritance structure; access to some bulk operations like the repeated points case.

For the "zero copy" crew, I feel like a big chunk of gains for them could be harvested by ensuring that point-based operations are available and don't require construction of a full Point() object. So things like PreparedGeometry->intersects(x, y). Sure, you still have to copy in your polygon feature and prepare it, but much of the overhead in that would still exist in a "zero copy" paradigm (all the internal index buildings). Meanwhile you'd no longer need to create a full Point() to do a point-in-poly test, and that would hopefully be a big win for most users.

Random thoughs on a sunny day,
P