Avoid calling expensive.toUri.toString and other micro-optimizations

[lihaoyi]

Fixes #1628

This PR micro-optimizes the top three hotspots i found in the JProfiler profiles compiling the Netty codebase using Javac in Zinc in Mill

sortBy

.sortBy(x => x.toUri().toString()) creates a new URI object and String for every comparison during the sort - that's O(n log n) object allocations, and it's blowing up the profile on the linked issue. .sorted instead uses Path's native Comparable<Path> implementation which compares paths directly without allocating objects.

Both produce deterministic ordering (which is what's needed for reproducible content hashes). The only semantic difference is the comparison method (lexicographic path comparison vs URI string comparison), but for the purpose of consistent hashing, both work.

isDirectory

toVirtualFile called Files.isDirectory() on every file from filtered list. Added toVirtualFileForRegularFile() that skips the syscall sys we know it is already a regular file

Caching contentHash

Changed all our def contentHashes to lazy vals. Looking at the code it seems to be safe - we re-create the file instances every time, and when file metadata changes ClasspathCache creates a new VirtualFile. Furthermore, contentHashStr is already a lazy val, so it seems the code already assumes the files are re-created when changed and so it should be safe to cache the hashes

Manually tested this via publishLocal and manually reproducing the original workload in the linked issue, the performance seems to have gone back to previous levels and the profiler no longer shows toURI taking up 34% of CPU time, and the isDirectory and contentHash hotspots each taking ~5% of CPU time dropped off the profiles as well. ./mill clean && time ./mill __.compile in the Netty codebase dropped from ~15s to ~8.5s

[lihaoyi]

CC @eed3si9n for review, looks like CI is green

[eed3si9n]

Thanks!

[lihaoyi]

Thanks for the merge @eed3si9n. Could I request a milestone release one of these days? I'd like to include this fix in Mill 1.1.0 which should go out Wed 21 Jan

[eed3si9n]

Sounds good.

[lihaoyi]

Ping on this @eed3si9n, since Scala 3.8.1 is out and a Zinc release with this PR is the only thing blocking Mill 1.1.0 final.

[eed3si9n]

Sorry about that. I'll try to get this out today.

[eed3si9n]

https://repo1.maven.org/maven2/org/scala-sbt/zinc_3/2.0.0-M13/ is out

[lihaoyi]

Thanks eugene!