Why is 0.16-dev a lot slower than 0.15.2?

[miagi]

I used the same logic from this showcase, main branch which is on 0.15.2 and 0.16-mixed which uses the same recursion that calls processing functions and I get this result when benchmarking:

Only difference logical difference in the code is this:
0.15.2

    for (processed_args.items[0 .. processed_args.items.len - 1]) |arg| {
        var open = std.fs.cwd().openDir(
            arg,
            .{ .no_follow = true },
        ) catch |err| {
            std.log.err("{s}: {s}\n", .{ @errorName(err), arg });
            continue;
        };
        defer open.close();
        const src_path = open.stat() catch |err| {
            std.log.err("{s}: {s}\n", .{ @errorName(err), arg });
            continue;
        };

0.16.-mixed:

    for (processed_args.items[0 .. processed_args.items.len - 1]) |arg| {
        var src_path = std.Io.Dir.cwd().statFile(
            io,
            arg,
            .{ .follow_symlinks = false },
        ) catch |err| {
            std.log.err("{s}: {s}\n", .{ @errorName(err), arg });
            continue;
        };

I know, I should figure out how to stat files with out going through symlinks on 0.15.2 instead, but does that really impact the performance this much?

[IntegratedQuantum]

@miagi

Only difference logical difference in the code is this:

…and all the changes from the standard library.

Looking at the times, it appears that the program execution is on the order of a couple hundred microseconds. At this scale I expect that even things, like the creation of an Io instance, or the code that runs before even calling your main (since you are using juicy main now), could make a measurable difference.

So to figure out, what exactly is the issue, I’d suggest to try running this in a full profiler (though I’m not sure if I know any that support such short-running programs, you may need to add a for loop around the main or something like that to get reasonable data).

[chrboesch]

Today I changed my SchmidlCox-Algo to 0.16-dev 2135 on the raspberry pi:

old: 460800 samples
sc time: 18 ms (0.040 µs/Sample)

new: 460800 samples
sc time: 14 ms (0.031 µs/Sample) 

Big improvement for me!