Treat source directories as directories

[cameron-martin]

Description

Treat source directories as directories in analysis by carrying a directory bit on SourceArtifact instead of classifying all source artifacts as files.

This fixes two related bugs:

• source directories now report File.is_directory == True
• ctx.actions.symlink(target_file=...) now accepts source directories with declare_directory() outputs

Motivation

Fixes #12954
Fixes #26705

Build API Changes

No

Checklist

• I have added tests for the new use cases (if any).
• I have updated the documentation (if applicable).

Release Notes

RELNOTES: None

[cameron-martin]

This likely isn't the best location for this test because it doesn't have anything to do with symlinks as the test below does. Are there any better places for this?

[tjgq]

I'm supportive of this PR, but as currently written, this incurs in an extra stat() for every source artifact, which I suspect is going to cause a performance regression at Google (and possibly also for other large Bazel projects). I wonder if we're already doing a stat() elsewhere and could piggyback on it. Another alternative is to delay the stat() until SourceArtifact.isDirectory() is called, but that doesn't help if it ends up getting called for most artifacts anyway; I'm also unsure about the consequences for exception handling.

Nevertheless, it would be worthwhile (for someone at Google) to run a benchmark on this PR as written, just to get a sense of the performance hit. I can't volunteer to do so at this time, though.

[cameron-martin]

@tjgq Presumably we have to do a stat when globbing, so it should be possible to feed this info through from that when the source of the input is a glob. Would just covering glob inputs be sufficient? How many source inputs come from globs vs not?

Do you know of anyone who can do this internal benchmarking?