This fixes a bug introduced in dfb7a1b4d0 (patch-ids: stop using a
hand-rolled hashmap implementation, 2016-07-29) in which
git rev-list --cherry-pick A...B
will fail to suppress commits reachable from A even if a commit with
matching patch-id appears in B.
Around the time of that commit, the algorithm for "--cherry-pick" looked
something like this:
0. Traverse all of the commits, marking them as being on the left or
right side of the symmetric difference.
1. Iterate over the left-hand commits, inserting a patch-id struct for
each into a hashmap, and pointing commit->util to the patch-id
struct.
2. Iterate over the right-hand commits, checking which are present in
the hashmap. If so, we exclude the commit from the output _and_ we
mark the patch-id as "seen".
3. Iterate again over the left-hand commits, checking whether
commit->util->seen is set; if so, exclude them from the output.
At the end, we'll have eliminated commits from both sides that have a
matching patch-id on the other side. But there's a subtle assumption
here: for any given patch-id, we must have exactly one struct
representing it. If two commits from A both have the same patch-id and
we allow duplicates in the hashmap, then we run into a problem:
a. In step 1, we insert two patch-id structs into the hashmap.
b. In step 2, our lookups will find only one of these structs, so only
one "seen" flag is marked.
c. In step 3, one of the commits in A will have its commit->util->seen
set, but the other will not. We'll erroneously output the latter.
Prior to dfb7a1b4d0, our hashmap did not allow duplicates. Afterwards,
it used hashmap_add(), which explicitly does allow duplicates.
At that point, the solution would have been easy: when we are about to
add a duplicate, skip doing so and return the existing entry which
matches. But it gets more complicated.
In 683f17ec44 (patch-ids: replace the seen indicator with a commit
pointer, 2016-07-29), our step 3 goes away entirely. Instead, in step 2,
when the right-hand side finds a matching patch_id from the left-hand
side, we can directly mark the left-hand patch_id->commit to be omitted.
Solving that would be easy, too; there's a one-to-many relationship of
patch-ids to commits, so we just need to keep a list.
But there's more. Commit b3dfeebb92 (rebase: avoid computing unnecessary
patch IDs, 2016-07-29) built on that by lazily computing the full
patch-ids. So we don't even know when adding to the hashmap whether two
commits truly have the same id. We'd have to tentatively assign them a
list, and then possibly split them apart (possibly into N new structs)
at the moment we compute the real patch-ids. This could work, but it's
complicated and error-prone.
Instead, let's accept that we may store duplicates, and teach the lookup
side to be more clever. Rather than asking for a single matching
patch-id, it will need to iterate over all matching patch-ids. This does
mean examining every entry in a single hash bucket, but the worst-case
for a hash lookup was already doing that.
We'll keep the hashmap details out of the caller by providing a simple
iteration interface. We can retain the simple has_commit_patch_id()
interface for the other callers, but we'll simplify its return value
into an integer, rather than returning the patch_id struct. That way
they won't be tempted to look at the "commit" field of the return value
without iterating.
Reported-by: Arnaud Morin <arnaud.morin@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We weren't flushing the context each time we processed a hunk in the
patch-id generation code in diff.c, but we were doing that when we
generated "stable" patch-ids with the 'patch-id' tool. Let's port that
similar logic over from patch-id.c into diff.c so we can get the same
hash when we're generating patch-ids for 'format-patch --base=' types of
command invocations.
Cc: Xiaolong Ye <xiaolong.ye@intel.com>
Signed-off-by: Stephen Boyd <sboyd@kernel.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
I looped over the toplevel header files, creating a temporary two-line C
program for each consisting of
#include "git-compat-util.h"
#include $HEADER
This patch is the result of manually fixing errors in compiling those
tiny programs.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since we will likely be introducing a new hash function at some point,
and that hash function might be longer than 20 bytes, use the constant
GIT_MAX_RAWSZ, which is designed to be suitable for allocations, instead
of GIT_SHA1_RAWSZ. This will ease the transition down the line by
distinguishing between places where we need to allocate memory suitable
for the largest hash from those where we need to handle the current
hash.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The `rebase` family of Git commands avoid applying patches that were
already integrated upstream. They do that by using the revision walking
option that computes the patch IDs of the two sides of the rebase
(local-only patches vs upstream-only ones) and skipping those local
patches whose patch ID matches one of the upstream ones.
In many cases, this causes unnecessary churn, as already the set of
paths touched by a given commit would suffice to determine that an
upstream patch has no local equivalent.
This hurts performance in particular when there are a lot of upstream
patches, and/or large ones.
Therefore, let's introduce the concept of a "diff-header-only" patch ID,
compare those first, and only evaluate the "full" patch ID lazily.
Please note that in contrast to the "full" patch IDs, those
"diff-header-only" patch IDs are prone to collide with one another, as
adjacent commits frequently touch the very same files. Hence we now
have to be careful to allow multiple hash entries with the same hash.
We accomplish that by using the hashmap_add() function that does not even
test for hash collisions. This also allows us to evaluate the full patch ID
lazily, i.e. only when we found commits with matching diff-header-only
patch IDs.
We add a performance test that demonstrates ~1-6% improvement. In
practice this will depend on various factors such as how many upstream
changes and how big those changes are along with whether file system
caches are cold or warm. As Git's test suite has no way of catching
performance regressions, we also add a regression test that verifies
that the full patch ID computation is skipped when the diff-header-only
computation suffices.
Signed-off-by: Kevin Willford <kcwillford@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The cherry_pick_list was looping through the original side checking the
seen indicator and setting the cherry_flag on the commit. If we save
off the commit in the patch_id we can set the cherry_flag on the correct
commit when running through the other side when a patch_id match is found.
Signed-off-by: Kevin Willford <kcwillford@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This change will use the hashmap from the hashmap.h to keep track of the
patch_ids that have been encountered instead of using an internal
implementation. This simplifies the implementation of the patch ids.
Signed-off-by: Kevin Willford <kcwillford@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make commit_patch_id() available to other builtins.
Signed-off-by: Xiaolong Ye <xiaolong.ye@intel.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This implements the patch-id computation and recording library,
patch-ids.c, and rewrites the get_patch_ids() function used in
cherry and format-patch to use it, so that they do not pollute
the object namespace. Earlier code threw non-objects into the
in-core object database, and hoped for not getting bitten by
SHA-1 collisions. While it may be practically Ok, it still was
an ugly hack.
Signed-off-by: Junio C Hamano <junkio@cox.net>
