Previously diff_tree(), which is now named ll_diff_tree_sha1(), was
generating diff_filepair(s) for two trees t1 and t2, and that was
usually used for a commit as t1=HEAD~, and t2=HEAD - i.e. to see changes
a commit introduces.
In Git, however, we have fundamentally built flexibility in that a
commit can have many parents - 1 for a plain commit, 2 for a simple merge,
but also more than 2 for merging several heads at once.
For merges there is a so called combine-diff, which shows diff, a merge
introduces by itself, omitting changes done by any parent. That works
through first finding paths, that are different to all parents, and then
showing generalized diff, with separate columns for +/- for each parent.
The code lives in combine-diff.c .
There is an impedance mismatch, however, in that a commit could
generally have any number of parents, and that while diffing trees, we
divide cases for 2-tree diffs and more-than-2-tree diffs. I mean there
is no special casing for multiple parents commits in e.g.
revision-walker .
That impedance mismatch *hurts* *performance* *badly* for generating
combined diffs - in "combine-diff: optimize combine_diff_path
sets intersection" I've already removed some slowness from it, but from
the timings provided there, it could be seen, that combined diffs still
cost more than an order of magnitude more cpu time, compared to diff for
usual commits, and that would only be an optimistic estimate, if we take
into account that for e.g. linux.git there is only one merge for several
dozens of plain commits.
That slowness comes from the fact that currently, while generating
combined diff, a lot of time is spent computing diff(commit,commit^2)
just to only then intersect that huge diff to almost small set of files
from diff(commit,commit^1).
That's because at present, to compute combine-diff, for first finding
paths, that "every parent touches", we use the following combine-diff
property/definition:
D(A,P1...Pn) = D(A,P1) ^ ... ^ D(A,Pn) (w.r.t. paths)
where
D(A,P1...Pn) is combined diff between commit A, and parents Pi
and
D(A,Pi) is usual two-tree diff Pi..A
So if any of that D(A,Pi) is huge, tracting 1 n-parent combine-diff as n
1-parent diffs and intersecting results will be slow.
And usually, for linux.git and other topic-based workflows, that
D(A,P2) is huge, because, if merge-base of A and P2, is several dozens
of merges (from A, via first parent) below, that D(A,P2) will be diffing
sum of merges from several subsystems to 1 subsystem.
The solution is to avoid computing n 1-parent diffs, and to find
changed-to-all-parents paths via scanning A's and all Pi's trees
simultaneously, at each step comparing their entries, and based on that
comparison, populate paths result, and deduce we could *skip*
*recursing* into subdirectories, if at least for 1 parent, sha1 of that
dir tree is the same as in A. That would save us from doing significant
amount of needless work.
Such approach is very similar to what diff_tree() does, only there we
deal with scanning only 2 trees simultaneously, and for n+1 tree, the
logic is a bit more complex:
D(T,P1...Pn) calculation scheme
-------------------------------
D(T,P1...Pn) = D(T,P1) ^ ... ^ D(T,Pn) (regarding resulting paths set)
D(T,Pj) - diff between T..Pj
D(T,P1...Pn) - combined diff from T to parents P1,...,Pn
We start from all trees, which are sorted, and compare their entries in
lock-step:
T P1 Pn
- - -
|t| |p1| |pn|
|-| |--| ... |--| imin = argmin(p1...pn)
| | | | | |
|-| |--| |--|
|.| |. | |. |
. . .
. . .
at any time there could be 3 cases:
1) t < p[imin];
2) t > p[imin];
3) t = p[imin].
Schematic deduction of what every case means, and what to do, follows:
1) t < p[imin] -> ∀j t ∉ Pj -> "+t" ∈ D(T,Pj) -> D += "+t"; t↓
2) t > p[imin]
2.1) ∃j: pj > p[imin] -> "-p[imin]" ∉ D(T,Pj) -> D += ø; ∀ pi=p[imin] pi↓
2.2) ∀i pi = p[imin] -> pi ∉ T -> "-pi" ∈ D(T,Pi) -> D += "-p[imin]"; ∀i pi↓
3) t = p[imin]
3.1) ∃j: pj > p[imin] -> "+t" ∈ D(T,Pj) -> only pi=p[imin] remains to investigate
3.2) pi = p[imin] -> investigate δ(t,pi)
|
|
v
3.1+3.2) looking at δ(t,pi) ∀i: pi=p[imin] - if all != ø ->
⎧δ(t,pi) - if pi=p[imin]
-> D += ⎨
⎩"+t" - if pi>p[imin]
in any case t↓ ∀ pi=p[imin] pi↓
~
For comparison, here is how diff_tree() works:
D(A,B) calculation scheme
-------------------------
A B
- -
|a| |b| a < b -> a ∉ B -> D(A,B) += +a a↓
|-| |-| a > b -> b ∉ A -> D(A,B) += -b b↓
| | | | a = b -> investigate δ(a,b) a↓ b↓
|-| |-|
|.| |.|
. .
. .
~~~~~~~~
This patch generalizes diff tree-walker to work with arbitrary number of
parents as described above - i.e. now there is a resulting tree t, and
some parents trees tp[i] i=[0..nparent). The generalization builds on
the fact that usual diff
D(A,B)
is by definition the same as combined diff
D(A,[B]),
so if we could rework the code for common case and make it be not slower
for nparent=1 case, usual diff(t1,t2) generation will not be slower, and
multiparent diff tree-walker would greatly benefit generating
combine-diff.
What we do is as follows:
1) diff tree-walker ll_diff_tree_sha1() is internally reworked to be
a paths generator (new name diff_tree_paths()), with each generated path
being `struct combine_diff_path` with info for path, new sha1,mode and for
every parent which sha1,mode it was in it.
2) From that info, we can still generate usual diff queue with
struct diff_filepairs, via "exporting" generated
combine_diff_path, if we know we run for nparent=1 case.
(see emit_diff() which is now named emit_diff_first_parent_only())
3) In order for diff_can_quit_early(), which checks
DIFF_OPT_TST(opt, HAS_CHANGES))
to work, that exporting have to be happening not in bulk, but
incrementally, one diff path at a time.
For such consumers, there is a new callback in diff_options
introduced:
->pathchange(opt, struct combine_diff_path *)
which, if set to !NULL, is called for every generated path.
(see new compat ll_diff_tree_sha1() wrapper around new paths
generator for setup)
4) The paths generation itself, is reworked from previous
ll_diff_tree_sha1() code according to "D(A,P1...Pn) calculation
scheme" provided above:
On the start we allocate [nparent] arrays in place what was
earlier just for one parent tree.
then we just generalize loops, and comparison according to the
algorithm.
Some notes(*):
1) alloca(), for small arrays, is used for "runs not slower for
nparent=1 case than before" goal - if we change it to xmalloc()/free()
the timings get ~1% worse. For alloca() we use just-introduced
xalloca/xalloca_free compatibility wrappers, so it should not be a
portability problem.
2) For every parent tree, we need to keep a tag, whether entry from that
parent equals to entry from minimal parent. For performance reasons I'm
keeping that tag in entry's mode field in unused bit - see S_IFXMIN_NEQ.
Not doing so, we'd need to alloca another [nparent] array, which hurts
performance.
3) For emitted paths, memory could be reused, if we know the path was
processed via callback and will not be needed later. We use efficient
hand-made realloc-style path_appendnew(), that saves us from ~1-1.5%
of potential additional slowdown.
4) goto(s) are used in several places, as the code executes a little bit
faster with lowered register pressure.
Also
- we should now check for FIND_COPIES_HARDER not only when two entries
names are the same, and their hashes are equal, but also for a case,
when a path was removed from some of all parents having it.
The reason is, if we don't, that path won't be emitted at all (see
"a > xi" case), and we'll just skip it, and FIND_COPIES_HARDER wants
all paths - with diff or without - to be emitted, to be later analyzed
for being copies sources.
The new check is only necessary for nparent >1, as for nparent=1 case
xmin_eqtotal always =1 =nparent, and a path is always added to diff as
removal.
~~~~~~~~
Timings for
# without -c, i.e. testing only nparent=1 case
`git log --raw --no-abbrev --no-renames`
before and after the patch are as follows:
navy.git linux.git v3.10..v3.11
before 0.611s 1.889s
after 0.619s 1.907s
slowdown 1.3% 0.9%
This timings show we did no harm to usual diff(tree1,tree2) generation.
From the table we can see that we actually did ~1% slowdown, but I think
I've "earned" that 1% in the previous patch ("tree-diff: reuse base
str(buf) memory on sub-tree recursion", HEAD~~) so for nparent=1 case,
net timings stays approximately the same.
The output also stayed the same.
(*) If we revert 1)-4) to more usual techniques, for nparent=1 case,
we'll get ~2-2.5% of additional slowdown, which I've tried to avoid, as
"do no harm for nparent=1 case" rule.
For linux.git, combined diff will run an order of magnitude faster and
appropriate timings will be provided in the next commit, as we'll be
taking advantage of the new diff tree-walker for combined-diff
generation there.
P.S. and combined diff is not some exotic/for-play-only stuff - for
example for a program I write to represent Git archives as readonly
filesystem, there is initial scan with
`git log --reverse --raw --no-abbrev --no-renames -c`
to extract log of what was created/changed when, as a result building a
map
{} sha1 -> in which commit (and date) a content was added
that `-c` means also show combined diff for merges, and without them, if
a merge is non-trivial (merges changes from two parents with both having
separate changes to a file), or an evil one, the map will not be full,
i.e. some valid sha1 would be absent from it.
That case was my initial motivation for combined diffs speedup.
Signed-off-by: Kirill Smelkov <kirr@mns.spb.ru>
Signed-off-by: Junio C Hamano <gitster@pobox.com>