Implement queries that add/remove components

[Ralith]

Fixes #334.

TODO:

• [ ] Work out why this machinery is 10x slower in the iterate_mut_100k benchmark than query_mut in master. Nothing's jumping out at me in the disassembly; input would be welcome.
• [ ] Move side effects into a guard object exposed via a new API (query_with_effect?) to fix unsoundness
• [ ] Update macros

[adamreichold]

Work out why this machinery is 10x slower in the iterate_mut_100k benchmark than query_mut in master. Nothing's jumping out at me in the disassembly; input would be welcome.

This is pure speculation, but could this be cache effects of the increased working set due to the effect table?

[Ralith]

This is pure speculation, but could this be cache effects of the increased working set due to the effect table?

Surprisingly, that slowdown is for in-place/effect-free queries, for which effects are ZSTs and the effect table should allocate no memory.

[a1phyr]

These benchmarks do very few things within the iteration, which sometimes allow LLVM to optimize the loop far better... or not ! This seems to be very fragile. Adding a continue in the iteration loop or a black_box with the query values can make or break it.

For example, on #341, adding black_box around vel.0 in the benchmark code canceled the optimization.

[Ralith]

A 10-fold difference is a lot to hand-wave, even so. Is that not sustained with black_box involved? Maybe a more computationally intensive benchmark would be interesting?

[Ralith]

This would be useful for https://github.com/Ralith/hecs/pull/366, which presently must make two passes and perform a lot of redundant lookups to add/remove components to/from entities satisfying a query.