Add collapser for sample on macOS

[jasonrhansen]

This diverges a bit in behavior from stackcollapse-sample.awk from upstream. For some reason it was implemented in a way that it outputs a bunch of erroneous stacks with 0 samples. This Rust version doesn't do that. However, the resulting flamegraphs should be the same between the two since those lines with 0 samples will be skipped over when generating the flamegraph.

One unfortunate thing about sample is it has the same problem demangling Rust names as DTrace. At least with DTrace there's a way to disable demangling so we can handle it in our collapser. I didn't add demangling to this collapser because, as far as I can tell, there's no way to disable demangling for sample.

[jonhoo]

How sure are we that everyone will always want to ignore these?

[jasonrhansen]

Not sure at all. I was just following upstream. Maybe we should add an option for this?

[jonhoo]

Depending on how the demangling is by default, it could be that it makes sense for us to just do it again?

[jasonrhansen]

For rustc_demangle to properly demangle Rust names, the names need to be mangled. Taking the improperly demangled Rust names from sample and demangling again doesn't work. At least DTrace has an option to leave names mangled, but not sample.

[jasonrhansen]

I don't know why the collapse_sample_cli test times out on Travis CI for Windows. I just tried running the tests on Windows without a problem.

[jonhoo]

Seems to fail on FreeBSD too? Is it just really slow?

[jasonrhansen]

It's not slow at all when I run it on my computer. The whole test suite runs in 4 or 5 seconds.

[jonhoo]

Hmm, that's very interesting. It seems to run just fine on my box too. The test does use assert-cmd, and I came across assert-rs/assert_cmd#42, which indicates that there's deadlock potential if stdin/stdout is particularly large, which may be the case for this test?

[jasonrhansen]

The input is 125k, but the input file we use for dtrace is 1.4M and that one doesn't deadlock.

[jonhoo]

I'll try restarting it and see if that helps. If not, maybe try adding

-- --test-threads=1 --nocapture

and some basic println! statements to that test so we can at least see where the test hangs?

[jonhoo]

@jasonrhansen hmm, nope, looks like it got stuck on the exact same test again :/

[jasonrhansen]

I haven't yet figured out why the tests are timing out on CI, but I came up with a solution to the poorly demangled Rust symbols. Using rustc_demangle on these doesn't work, but I took relevant parts of it and wrote a function that fixes these "partially demangled" Rust names. This could be used in the dtrace collapser as well, which would be much nicer than having to run dtrace with -xmangled and inferno-collapse-dtrace with --demangle.

[jasonrhansen]

I couldn't figure out how to get that test to work so I just disabled it for now.

[codecov]

Codecov Report

Merging #133 into master will decrease coverage by 0.19%.
The diff coverage is 88.41%.

@@            Coverage Diff            @@
##           master     #133     +/-   ##
=========================================
- Coverage   89.72%   89.52%   -0.2%     
=========================================
  Files           8       10      +2     
  Lines         983     1146    +163     
=========================================
+ Hits          882     1026    +144     
- Misses        101      120     +19

Impacted Files	Coverage Δ
src/collapse/mod.rs	54.54% <ø> (ø)	⬆️
src/collapse/perf.rs	92.15% <100%> (+0.07%)	⬆️
src/collapse/guess.rs	71.42% <66.66%> (-1.65%)	⬇️
src/collapse/sample.rs	87.62% <87.62%> (ø)
src/collapse/util/mod.rs	90.32% <90.32%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update eae7aa3...39712a4. Read the comment docs.

[jonhoo]

Excellent! Thank you 🎉