On Wed, May 11, 2022 at 9:37 AM Junio C Hamano [off-list ref] wrote:

Goss Geppert [off-list ref] writes:

quoted

diff --git a/dir.c b/dir.c
index f2b0f24210..a1886e61a3 100644
--- a/dir.c
+++ b/dir.c

@@ -1893,9 +1893,31 @@ static enum path_treatment treat_directory(struct dir_struct *dir,

[...]

quoted

+                     real_gitdir = real_pathdup(the_repository->gitdir, 0);

This function is repeatedly called during the traversal.

How expensive is it to keep calling real_pathdup() on the constant
the_repository->gitdir just in case it might be the same as our true
GIT_DIR?

I agree that treat_directory is called many times, but this
real_pathdup() call is inside the "if (nested_repo)" block, so this
new real_pathdup() invocation should occur very seldom.  Or are you
worried about cases where users have *very* large numbers of bare
repositories nested under the working directory?  Even in that case,
which seems pathological to me, I'd suspect the
is_nonbare_repository_dir() ->
read_gitfile_gently()/is_git_directory() codepath (used to determine
the value of nested_repo) would be much more expensive than this call
to real_pathdup(), so would it be worth trying to optimize this
real_pathdup() call away even in that rare case?