countme

11: Bump deps r=Veykril a=lnicola

This should let us avoid a duplicate dependency in rust-analyzer. dashmap doesn't seem to be used in the public API, so not counting it as a breaking change.

Co-authored-by: Laurențiu Nicola [email protected]

10: Optimizations r=matklad a=WaffleLapkin

Hey there! I've watched your video in the "Explaining rust-analyzer" series where you said that this crate can be optimized and decided to look into it (great series btw, thanks :>).

I've implemented some general optimizations (mostly by looking for long functions in flamegraph) as well as the first approach from #9. I've tried to make commits self-contained and write good messages for them, so reviewing commit-by-commit is probably a good idea.

For me in the end, benchmarks (see details below) show these results:

• 2.6 times faster for single-thread
• 6.8 times faster for multi-thread

That may be silly, but I've just run hyperfine on pairs of commits with a bash script like this (I'm bad at bash, I know):

#!/usr/bin/bash

if [ ! -z "$2" ]
then
  git checkout "$2" 2> /dev/null || echo "couldn't do checkout"
fi
echo "comparing:"
git log -1 --oneline

cargo br --example bench --features="enable" 2> /dev/null || echo "compilation failure"
cp ./target/release/examples/bench ./bench_current

if [ ! -z "$1" ]
then
  git checkout $1 2> /dev/null || echo "couldn't do checkout"
else
  git checkout HEAD~1 2> /dev/null || echo "couldn't do checkout"
fi
echo "to:"
git log -1 --oneline

cargo br --example bench --features="enable" 2> /dev/null || echo "compilation failure"
cp ./target/release/examples/bench ./bench_prev

git switch - 2> /dev/null || echo "couldn't switch back"

hyperfine -w 5 -m 20 -c "sleep 20" ./bench_prev ./bench_current

Results (multi-thread):

comparing:
d76b8c3 (HEAD -> optimizations) Fix docs: correctly mention when counting is and isn't enabled by default
to:
b69e96e (HEAD) Disable at runtime by default
Benchmark 1: ./bench_prev
  Time (mean ± σ):      4.102 s ±  0.214 s    [User: 30.658 s, System: 0.021 s]
  Range (min … max):    3.660 s …  4.463 s    20 runs
 
Benchmark 2: ./bench_current
  Time (mean ± σ):     604.1 ms ±  48.3 ms    [User: 4363.1 ms, System: 10.3 ms]
  Range (min … max):   541.5 ms … 731.1 ms    20 runs
 
Summary
  './bench_current' ran
    6.79 ± 0.65 times faster than './bench_prev'

Results (single-thread):

comparing:
d76b8c3 (HEAD -> optimizations) Fix docs: correctly mention when counting is and isn't enabled by default
to:
e95e4e4 (HEAD) Add single thread bench example
Benchmark 1: ./bench_prev
  Time (mean ± σ):      4.679 s ±  0.029 s    [User: 4.666 s, System: 0.007 s]
  Range (min … max):    4.626 s …  4.779 s    20 runs
 
Benchmark 2: ./bench_current
  Time (mean ± σ):      1.789 s ±  0.023 s    [User: 1.780 s, System: 0.006 s]
  Range (min … max):    1.775 s …  1.872 s    20 runs
 
  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet PC without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.
 
Summary
  './bench_current' ran
    2.62 ± 0.04 times faster than './bench_prev'

I've also tried to see if running RA_COUNT=1 cargo run -q --release -p rust-analyzer -- analysis-stats . would be faster, but it seems like noise affects performance a lot more than couting.

Co-authored-by: Maybe Waffle [email protected]

8: Disable at runtime by default r=matklad a=matklad

bors r+
🤖

Co-authored-by: Aleksey Kladov [email protected]

6: Fix formatting even more r=matklad a=lnicola

Co-authored-by: Laurențiu Nicola [email protected]

5: Fix formatting more r=matklad a=matklad

bors r+
🤖

Co-authored-by: Aleksey Kladov [email protected]

4: Nicer formatting r=matklad a=matklad

bors r+
🤖

Co-authored-by: Aleksey Kladov [email protected]

2: Overhaul API r=matklad a=matklad

bors r+
🤖

Co-authored-by: Aleksey Kladov [email protected]

This should greatly reduce the number of printouts.

The trick is stolen from dhat