问题
git log has some very useful commit-limiting options, such as --no-merges and --first-parent. I'd like to be able to use these options when generating a cumulative diff patch/stat/numstat for a range of commits.
With these commands:
git log --oneline --first-parent --no-merges --patch 29665b0..0b76a27
git log --oneline --first-parent --no-merges --stat 29665b0..0b76a27
git log --oneline --first-parent --no-merges --numstat 29665b0..0b76a27
the diff is not cumulative (the changes are listed individually for each commit).
With these commands:
git diff --patch 29665b0..0b76a27
git diff --stat 29665b0..0b76a27
git diff --numstat 29665b0..0b76a27
the diff is cumulative, but unfortunately git diff doesn't support the commit-limiting options.
So what I'd like is the cumulative diff functionality of git diff combined with the commit-limiting functionality of git log.
One idea I had is to use git log to generate a list of commit hashes, and then somehow pipe that list to git diff to generate a cumulative diff of the specified commits. Something like this (obviously this method of piping hashes to git diff doesn't actually work):
git log --pretty=format:%h --first-parent --no-merges 29665b0..0b76a27 | git diff
where --pretty=format:%h outputs the hashes of the matching commits.
Update
Thanks to @torek and @twalberg, I now understand git diff's operation more clearly. The range syntax 29665b0..0b76a27 is indeed misleading, and I now understand that it's not actually performing a cumulative diff over a range of commits. Looking through the docs, I found this:
"diff" is about comparing two endpoints, not ranges, and the range notations (
<commit>..<commit>and<commit>...<commit>) do not mean a range as defined in the "SPECIFYING RANGES" section in gitrevisions(7).
Taking this into account, I'll rephrase my question. With these commands:
git log --oneline --first-parent --no-merges --patch 29665b0..0b76a27
git log --oneline --first-parent --no-merges --stat 29665b0..0b76a27
git log --oneline --first-parent --no-merges --numstat 29665b0..0b76a27
the changes are listed individually for each matching commit. How can I combine those individual changes, to produce a cumulative patch/stat/numstat?
The answers to the linked possible duplicate question are helpful, suggesting a solution: create a temporary branch, cherry-pick the relevant commits, and then generate the diff.
I've just posted an answer which uses this technique, but I'm still interested to know if there's a solution which doesn't require a temporary branch?
回答1:
There is at least one basic misapprehension here. Specifically, git diff is not really cumulative at all: instead, it's simply pairwise.
Specifically, these two commands do the same thing:
git diff rev1 rev2
git diff rev1..rev2
That is, in git diff, there really is no such thing as a range in the first place.
With that out of the way, let's take a look behind the scenes at git log. What git log does with a range is really1 to hand the range to git rev-list, which produces a list of every rev in the range, applying the modifiers along the way:
git rev-list 29665b0..0b76a27
spits out every rev reachable from 0b76a27 that is not also reachable from 29665b0. Adding --first-parent, --max-parents=1 (aka --no-merges), and so on filters away some of the revs that would be listed here.
The final result is given back to git log, which then looks at each revision in the order git rev-list spits them out—this is also controllable via --date-order and --topo-order and so on; see the documentation for git rev-list—and shows you each log entry, perhaps along with a diff as produced by git diff-tree (which for single-parent commits, compares the commit to its parent).
What you can do, then, is invoke git rev-list yourself, directly, and then peel off the top and bottom revisions from its output. (In this particular case you probably want --topo-order too, to make sure that the last rev really is the earliest, graph-wise, regardless of dates.) For instance, in a script:
#! /bin/sh
tempfile=$(mktemp -t mydiff)
trap "rm -f $tempfile" 1 2 3 15
git rev-list 29665b0..0b76a27 --first-parent --no-merges --topo-order > $tempfile
# remember that the first rev listed is the last rev in the range
last=$(head -1 $tempfile)
first=$(tail -1 $tempfile)
rm -f $tempfile # done with it, don't leave it around while showing diff
git diff $first $last
You can get considerably fancier by using git rev-parse to parse options and split them into diff options vs rev-list options, but that's way beyond what you need here. The main thing to improve above is to get rid of the hard-coded revision-range.
1Some git commands really really do hand arguments off to git rev-list, as they're just shell scripts that use git rev-list and other git commands to handle this. Others are built together, so that git log and git rev-list are actually a single binary, and one part hands a job off to another part, but without invoking a new program.
In any case, note that git log master simply hands master off to git rev-list, which produces a list of all revs reachable from the branch-label master. If you add --no-walk, git rev-list produces just one rev, so that git log shows only that one revision.
回答2:
# Create a temporary branch to mark the start of the cherry-picked commits
git branch tmpstart
# Create and checkout a temporary branch for the cherry-picked commits
git checkout -b tmpend
# Use git log to filter the range of commits with the desired
# commit-limiting options, and then cherry-pick each matching commit
git log \
--first-parent \ # Commit-limiting
--no-merges \ # Commit-limiting
--reverse \ # Reverse the order (ascending chronological order)
--pretty=format:%h \ # Output the abbreviated hash of each matching commit
29665b0..0b76a27 \ # Range of commits
| xargs -n 1 git cherry-pick
# Generate the patch/stat/numstat of the cherry-picked commits
git diff --patch tmpstart tmpend
git diff --stat tmpstart tmpend
git diff --numstat tmpstart tmpend
来源：https://stackoverflow.com/questions/25403705/git-cumulative-diff-with-commit-limiting