Nguyen Thai Ngoc Duy [off-list ref] writes:
This patch makes this function only usable when startup_info pointer
is initialized. As "git" binary is the only caller, the change is ok.
If non-builtin commands want to use this function, they need to
initialize startup_info first.
Hrm, that explanation is understandable, but it strongly makes me suspect
that this change is making the code _more_ error prone in the longer term.
When somebody wants to add a new caller to a non-builtin, they need to
think about what prefix to pass, and would realize that they need to call
setup_git_directory() to get it. With the updated code, they can totally
forget and call it without any initialized startup_info.
Adding a totally new command is rare, new non-builtin is rarer, and adding
trace to it is even more so, so it may not be worth worrying about, but I
wonder if there is a cheap way to check such a programming mistake.