Reported by Evan Shaw. The problem is that $(git --exec-path) includes a
'git' binary which is incompatible with the one in /usr/bin; if you run it,
it gives you an error about libiconv2.dll.
You might think we could just add $(git --exec-path) at the *end* of PATH,
but then if there are multiple versions of git installed, we could end up
with the wrong one; earlier versions used to put git-sh-setup in /usr/bin,
so we'd pick up that one before the new one.
So now we just set PATH back to its original value right after running
git-sh-setup, and we should be okay.
A folder in a repository that wasn't initially imported as a subtree could no longer be splitted into an entirely new subtree with no parent.
A fix and a new test to fix that regression is added here.
It's possible to specify the subdir of a subtree since Git 1.7.0 - adding
support for that functionality to make the merge more stable.
Also checking for git version - now only uses the new subtree subdir option
when on at least 1.7.
Now you can do:
git subtree add --prefix=whatever git://wherever branchname
to add a new branch, instead of rather weirdly having to do 'git fetch'
first. You can also split and push in one step:
git subtree push --prefix=whatever git://wherever newbranch
(Somewhat cleaned up by apenwarr.)
It seems that in older versions, --message="" was interpreted as "use the
default commit message" instead of "use an empty commit message", and
git-subtree was depending on this behaviour. Now we don't, so tests pass
again.
We should implement it as
git fetch ...
git subtree merge ...
But we were instead just calling
git pull -s subtree ...
because 'git subtree merge' used to be just an alias for 'git merge -s
subtree', but it no longer is.
We were trying to 'git checkout $prefix', which is ambiguous if $prefix
names a directory *and* a branch. Do 'git checkout -- $prefix' instead.
The main place this appeared was in 'git subtree add'.
Reported by several people.
If you (like me) are using a modified git straight out of its source
directory (ie. without installing), then --exec-path isn't actually correct.
Add it to the PATH instead, so if it is correct, it'll work, but if it's
not, we fall back to the previous behaviour.
As pointed out by documentation, the correct use of 'git-sh-setup' is
using $(git --exec-path) to avoid problems with not standard
installations.
Signed-off-by: gianluca.pacchiella <pacchiel@studenti.ph.unito.it>
Instead of merging in the history of the entire subproject, just squash it
all into one commit, but try to at least track which commits we used so that
we can do future merges correctly.
Bonus feature: we can actually switch branches of the subproject this way,
just by "squash merging" back and forth from one tag to another.
If any one of the parents is the same as the current one, then clearly the
other parent branch isn't important, so throw it away entirely.
Can't remember why I didn't do this before, but if I rediscover it, it
definitely needs a unit test.
This is a simplification of the previous logic. I don't *think* it'll break
anything.
Results in far fewer useless merge commmits when playing with gitweb in the
git project:
git subtree split --prefix=gitweb --annotate='(split) ' 0a8f4f0^^..f2e7330
--onto=1130ef3
...and it doesn't *seem* to eliminate anything important.
Now we only prune out a commit if it has exactly one remaining parent and
that parent's tree is identical to ours.
But I also changed the test to create the initial "-s ours" merge in one
step instead of two, and that merge can be eliminated since one of its
parents doesn't affect the subdir at all, and is thus deleted.
Now we cut out a commit if any of its parents had the same tree; just use
that parent in its place. This makes the history look nice, but I don't
think it's quite right...
