Update the glob documentation to reflect a semantic change made a very long time ago where glob(['**'], exclude_directories = 0) doesn't match the package's directory. Also add tests for this behavior.
Also update Skyframe globbing to have these semantics. Any discrepancy has always been problematic, but now that we have Skyframe-hybrid globbing it's a lot more dangerous and consequential.
Alternatives considered: do this the other way around (keep the stale documentation as-is and instead update legacy globbing). This would potentially require changing existing usages from stuff like 'data = glob(["**"], exclude_directories = 0)' to 'data = [x for x in glob(["**"], exclude_directories = 0) where x != '']'. I think this is too messy, so long as there is a valid use-case for globs matching directories in the first place.
--
MOS_MIGRATED_REVID=115511504
diff --git a/src/main/java/com/google/devtools/build/lib/skyframe/GlobFunction.java b/src/main/java/com/google/devtools/build/lib/skyframe/GlobFunction.java
index e1c877f..e233cdf 100644
--- a/src/main/java/com/google/devtools/build/lib/skyframe/GlobFunction.java
+++ b/src/main/java/com/google/devtools/build/lib/skyframe/GlobFunction.java
@@ -90,7 +90,8 @@
// "**" also matches an empty segment, so try the case where it is not present.
if ("**".equals(patternHead)) {
if (patternTail == null) {
- if (!glob.excludeDirs()) {
+ // Recursive globs aren't supposed to match the package's directory.
+ if (!glob.excludeDirs() && !globSubdir.equals(PathFragment.EMPTY_FRAGMENT)) {
matches.add(globSubdir);
}
} else {
