Very long analysis due to slow addr2line calls

[Mingun]

I try to use flamegraph to profile quick-xml benchmarks. I've run the following command in the quick-xml checkout:

cargo flamegraph --bench macrobenches -- --bench

perf collected over 8GB of raw data and analysis going more than a hour already. flamegraph runs perf script which runs addr2line many times, about 2-3 invocations per second. It seems to me that such a simple task consumes absurd time to complete. Probably it is better to find more performant solutions to process perf data.

[djc]

Maybe have a look at #74 and #199? I think there was an issue with some versions of the addr2line script. You may also want to try samply as a flamegraph alternative.

[Mingun]

Ah, ok, yes, this indeed a duplicate of those issues -- I'm on Ubuntu 22.04 with kernel 5.15 where perf performance is not fixed yet.

[RalfJung]

I am on binutils 2.43 and kernel 6.11 (current Debian testing), and I am still hitting this -- it's now already been around 10 minutes since perf finished (the actual program run took 3s and gathered 3k samples, which should be a trivial amount of data), and it's still spawning thousands and thousands of addr2line processes. Is there some new problem here?

[RalfJung]

Interesting, when I finally killed perf script I saw this error on the terminal printed a few hundred times:

addr2line /home/r/.debug/.build-id/37/8f73d84f9d89f55dfc18a9ab84bce321cd9b09/elf: could not read first record

No idea if that could have something to do with the performance problem.

[RalfJung]

Turns out using gimli's addr2line fixes this. Magic.