A Better Rust Profiler

I want a better profiler for Rust. Here’s what a rust-analyzer benchmark looks like:

#[test]
fn benchmark_syntax_highlighting_parser() {
  if skip_slow_tests() {
    return;
  }

  let fixture = bench_fixture::glorious_old_parser();
  let (analysis, file_id) = fixture::file(&fixture);

  let hash = {
    let _pt = bench("syntax highlighting parser");
    analysis
      .highlight(file_id)
      .unwrap()
      .iter()
      .filter(|it| {
        it.highlight.tag == HlTag::Symbol(SymbolKind::Function)
      })
      .count()
  };
  assert_eq!(hash, 1629);
}

Here’s how I want to profile it:

#[test]
fn benchmark_syntax_highlighting_parser() {
  if skip_slow_tests() {
    return;
  }

  let fixture = bench_fixture::glorious_old_parser();
  let (analysis, file_id) = fixture::file(&fixture);

  let hash = {
    let _b = bench("syntax highlighting parser");
    let _p = better_profiler::profile();
    analysis
      .highlight(file_id)
      .unwrap()
      .iter()
      .filter(|it| {
        it.highlight.tag == HlTag::Symbol(SymbolKind::Function)
      })
      .count()
  };
  assert_eq!(hash, 1629);
}

First, the profiler prints to stderr:

warning: run with `--release`
warning: add `debug=true` to Cargo.toml
warning: set `RUSTFLAGS="-Cforce-frame-pointers=yes"`

Otherwise, if everything is setup correctly, the output is

Output is saved to:
   ~/projects/rust-analyzer/profile-results/

The profile-results folder contains the following:

• report.txt with
  • user, cpu, sys time
  • cpu instructions
  • stats for caches & branches a-la pref-stat
  • top ten functions by cumulative time
  • top ten functions by self-time
  • top ten hot-spot
• flamegraph.svg
• data.smth, which can be fed into some existing profiler UI (kcachegrind, firefox profiler, etc).
• report.html which contains a basic interactive UI.

To tweak settings, the following API is available:

let _p = better_profiler::profile()
  .output("./other-dir/")
  .samples_per_second(999)
  .flamegraph(false);

Naturally, the following also works and produces an aggregate profile:

for _ in 0..100 {
  {
    let _p = profile();
    interesting_computation();
  }
  not_interesting_computation();
}

I don’t know how this should work. I think I would be happy with a perf-based Linux-only implementation. The perf-event crate by Jim Blandy (co-author of “Programming Rust”) is good.

Have I missed something? Does this tool already exist? Or is it impossible for some reason?

Discussion on /r/rust.