union: Don't traverse sub-directories only found in one element of the union.

This significantly reduces I/O when building profiles, especially with
lots of package-specific sub-directories (such as 'share/emacs/24.3',
'texmf', etc.)

* guix/build/union.scm (union-build)[file-tree](others-have-it?): New
  procedure.  Use it in the 'enter?' parameter of 'file-system-fold';
  change 'skip' parameter accordingly.
* tests/union.scm ("union-build"): Ensure that 'include' is a symlink
  and 'bin' is a directory.

1 files changed, 11 insertions, 1 deletions

diff --git a/tests/union.scm b/tests/union.scm
index 9816882101..6287cffc38 100644
--- a/tests/union.scm
+++ b/tests/union.scm
@@ -114,7 +114,17 @@
                 (file-exists? "bin/ld")
                 (file-exists? "lib/libc.so")
                 (directory-exists? "lib/gcc")
-                (file-exists? "include/unistd.h"))))))
+                (file-exists? "include/unistd.h")
+
+                ;; The 'include' sub-directory is only found in
+                ;; glibc-bootstrap, so it should be unified in a
+                ;; straightforward way, without traversing it.
+                (eq? 'symlink (stat:type (lstat "include")))
+
+                ;; Conversely, several inputs have a 'bin' sub-directory, so
+                ;; unifying it requires traversing them all, and creating a
+                ;; new 'bin' sub-directory in the profile.
+                (eq? 'directory (stat:type (lstat "bin"))))))))
 
 (test-end)