[perf] path::parent_path() is slow

[Tim Stack]

Describe the bug
The path::parent_path() method is unreasonably slow compared to something like POSIX's dirname(3).

To Reproduce
Using this file:

#include <libgen.h>
#include <chrono>
#include "ghc/filesystem.hpp"

int main(int argc, char *argv[])
{
  {
    std::string path_str = "/a/really/long/path/to/test/performance/of/parent_path";

    {
      ghc::filesystem::path path(path_str);
      auto start = std::chrono::system_clock::now();

      for (int lpc = 0; lpc < 1000000; lpc++) {
        auto par = path.parent_path();
      }
      auto stop = std::chrono::system_clock::now();

      printf("ghc duration: %d\n", (int32_t) (stop - start).count());
    }
    {
      auto start = std::chrono::system_clock::now();

      for (int lpc = 0; lpc < 1000000; lpc++) {
        auto par = (char *) malloc(path_str.length() + 1);

        strcpy(par, path_str.c_str());
        dirname(par);
        free(par);
      }
      auto stop = std::chrono::system_clock::now();

      printf("dirname(3) duration: %d\n", (int32_t) (stop - start).count());
    }
  }
}

Compiled with:

$ g++ --std=c++14 -O3 path_perf.cc

On my 2013 iMac with a 3.5Ghz Core i7, I get the following results:

ghc duration: 2337478
dirname(3) duration: 94162

Expected behavior
The performance of parent_path() should be a lot better.

Additional context
I'm using this library in https://github.com/tstack/lnav and parent_path() turned out to be a major source of slowness when a lot of files were open.

(This library is great, thanks for making it!)

[trapexit]

To be fair... have you done a performance comparison with some standard versions?

https://github.com/gulrak/filesystem/blob/master/include/ghc/filesystem.hpp#L2703

Might also be worth replacing the implementation with dirname and see how that compares.

[Tim Stack]

have you done a performance comparison with some standard versions?

Using the following compiler:

Configured with: --prefix=/Library/Developer/CommandLineTools/usr --with-gxx-include-dir=/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk/usr/include/c++/4.2.1
Apple clang version 11.0.3 (clang-1103.0.32.62)
Target: x86_64-apple-darwin19.6.0
Thread model: posix
InstalledDir: /Library/Developer/CommandLineTools/usr/bin

And updating the test file to use std::filesystem, the results are:

fs duration: 133115
dirname(3) duration: 97202

Which is a reasonable result.

[Steffen Schümann]

I'll look into it, thanks for reporting.

[Steffen Schümann]

Okay, I could reproduce your measurements. Thanks for delivering the nice test!

On my MacBook Pro i9, 2.4GHz I got:

ghc duration: 2261615
dirname(3) duration: 93431

I reworked the implementation and now get:

ghc duration: 177320
dirname(3) duration: 95402

It's not as fast as the libc++ std::filesystem::path::parent_path() that gets here:

fs duration: 105919
dirname(3) duration: 94327

But they use string_views internally and as this implementation needs to compile from C++11 on, and I don't want do differentiate too much of the code, I think I am quite happy with the improvement.

[Tim Stack]

That looks like a good improvement to me, thanks for taking a look.

[Steffen Schümann]

This is now part of release v1.5.0