perf: optimize the label representation

[yannham]

This slightly optimizes the label representation, by using (smaller) PosIdx instead of TermPos, and a combination of Vec, SmallVec, and Vector, that was found by experimentation, trying to optimize first for runtime but also for memory consumption. On the OPL benchmark the change seems to have a very small time improvement, but does at least win a few noticeable % of leaked memory.

It was tempting to use only copy-on-write structures, such as a thin Rc<[T]> or Vector everywhere (but the latter is bigger). It turns that after measuring many label operations do succeed in taking their content mutably with content_make_mut (the operators pushing an element to the path, or a diagnostic for example), around a surprising 50%. On the other hand, some other mutable label operations do not currently try to take a mutable reference to the label, typically path operations. There, trying to take the label mutably shows that 0% of them succeeds... So for some reason, although I believe it's by accident, the current choices of when should we try to take unique access to a label in eval::operation sounds rather adapted. The data structure were chosen in consequence.

[github-actions]

Bencher Report

Branch	perf/label-repr
Testbed	ubuntu-latest

Click to view all benchmark results

Benchmark	Latency	microseconds (µs)
diagnostics-benches/inputs/goto-perf.ncl	📈 view plot 🚷 view threshold	10,091.00 µs
diagnostics-benches/inputs/large-record-tree.ncl	📈 view plot 🚷 view threshold	187,100.00 µs
diagnostics-benches/inputs/redis-replication-controller.ncl	📈 view plot 🚷 view threshold	302.24 µs
diagnostics-benches/inputs/small-record-tree.ncl	📈 view plot 🚷 view threshold	430.79 µs
fibonacci 10	📈 view plot 🚷 view threshold	219.88 µs
foldl arrays 50	📈 view plot 🚷 view threshold	457.01 µs
foldl arrays 500	📈 view plot 🚷 view threshold	2,749.80 µs
foldr strings 50	📈 view plot 🚷 view threshold	3,164.60 µs
foldr strings 500	📈 view plot 🚷 view threshold	28,938.00 µs
generate normal 250	📈 view plot 🚷 view threshold	15,600.00 µs
generate normal 50	📈 view plot 🚷 view threshold	1,058.90 µs
generate normal unchecked 1000	📈 view plot 🚷 view threshold	2,007.50 µs
generate normal unchecked 200	📈 view plot 🚷 view threshold	387.37 µs
init-diagnostics-benches/inputs/goto-perf.ncl	📈 view plot 🚷 view threshold	52,036.00 µs
init-diagnostics-benches/inputs/large-record-tree.ncl	📈 view plot 🚷 view threshold	206,600.00 µs
init-diagnostics-benches/inputs/redis-replication-controller.ncl	📈 view plot 🚷 view threshold	46,988.00 µs
init-diagnostics-benches/inputs/small-record-tree.ncl	📈 view plot 🚷 view threshold	47,016.00 µs
pidigits 100	📈 view plot 🚷 view threshold	1,897.00 µs
pipe normal 20	📈 view plot 🚷 view threshold	674.85 µs
pipe normal 200	📈 view plot 🚷 view threshold	4,531.70 µs
product 30	📈 view plot 🚷 view threshold	341.06 µs
requests-benches/inputs/goto-perf.ncl-000	📈 view plot 🚷 view threshold	3,002.60 µs
requests-benches/inputs/large-record-tree.ncl-000	📈 view plot 🚷 view threshold	581,170.00 µs
requests-benches/inputs/large-record-tree.ncl-001	📈 view plot 🚷 view threshold	88.00 µs
scalar 10	📈 view plot 🚷 view threshold	597.40 µs
sum 30	📈 view plot 🚷 view threshold	343.22 µs

🐰 View full continuous benchmarking report in Bencher