Provided by: git-man_2.32.0-1ubuntu1_all bug


       git-range-diff - Compare two commit ranges (e.g. two versions of a branch)


       git range-diff [--color=[<when>]] [--no-color] [<diff-options>]
               [--no-dual-color] [--creation-factor=<factor>]
               [--left-only | --right-only]
               ( <range1> <range2> | <rev1>...<rev2> | <base> <rev1> <rev2> )


       This command shows the differences between two versions of a patch series, or more
       generally, two commit ranges (ignoring merge commits).

       To that end, it first finds pairs of commits from both commit ranges that correspond with
       each other. Two commits are said to correspond when the diff between their patches (i.e.
       the author information, the commit message and the commit diff) is reasonably small
       compared to the patches' size. See ``Algorithm`` below for details.

       Finally, the list of matching commits is shown in the order of the second commit range,
       with unmatched commits being inserted just after all of their ancestors have been shown.

       There are three ways to specify the commit ranges:

       •   <range1> <range2>: Either commit range can be of the form <base>..<rev>, <rev>^!  or
           <rev>^-<n>. See SPECIFYING RANGES in gitrevisions(7) for more details.

       •   <rev1>...<rev2>. This is equivalent to <rev2>..<rev1> <rev1>..<rev2>.

       •   <base> <rev1> <rev2>: This is equivalent to <base>..<rev1> <base>..<rev2>.


           When the commit diffs differ, ‘git range-diff` recreates the original diffs’ coloring,
           and adds outer -/+ diff markers with the background being red/green to make it easier
           to see e.g. when there was a change in what exact lines were added.

           Additionally, the commit diff lines that are only present in the first commit range
           are shown "dimmed" (this can be overridden using the color.diff.<slot> config setting
           where <slot> is one of contextDimmed, oldDimmed and newDimmed), and the commit diff
           lines that are only present in the second commit range are shown in bold (which can be
           overridden using the config settings color.diff.<slot> with <slot> being one of
           contextBold, oldBold or newBold).

           This is known to range-diff as "dual coloring". Use --no-dual-color to revert to color
           all lines according to the outer diff markers (and completely ignore the inner diff
           when it comes to color).

           Set the creation/deletion cost fudge factor to <percent>. Defaults to 60. Try a larger
           value if git range-diff erroneously considers a large change a total rewrite (deletion
           of one commit and addition of another), and a smaller one in the reverse case. See the
           ``Algorithm`` section below for an explanation why this is needed.

           Suppress commits that are missing from the first specified range (or the "left range"
           when using the <rev1>...<rev2> format).

           Suppress commits that are missing from the second specified range (or the "right
           range" when using the <rev1>...<rev2> format).

           This flag is passed to the git log program (see git-log(1)) that generates the

       <range1> <range2>
           Compare the commits specified by the two ranges, where <range1> is considered an older
           version of <range2>.

           Equivalent to passing <rev2>..<rev1> and <rev1>..<rev2>.

       <base> <rev1> <rev2>
           Equivalent to passing <base>..<rev1> and <base>..<rev2>. Note that <base> does not
           need to be the exact branch point of the branches. Example: after rebasing a branch
           my-topic, git range-diff my-topic@{u} my-topic@{1} my-topic would show the differences
           introduced by the rebase.

       git range-diff also accepts the regular diff options (see git-diff(1)), most notably the
       --color=[<when>] and --no-color options. These options are used when generating the "diff
       between patches", i.e. to compare the author, commit message and diff of corresponding
       old/new commits. There is currently no means to tweak most of the diff options passed to
       git log when generating those patches.


       The output of the range-diff command is subject to change. It is intended to be
       human-readable porcelain output, not something that can be used across versions of Git to
       get a textually stable range-diff (as opposed to something like the --stable option to
       git-patch-id(1)). There’s also no equivalent of git-apply(1) for range-diff, the output is
       not intended to be machine-readable.

       This is particularly true when passing in diff options. Currently some options like --stat
       can, as an emergent effect, produce output that’s quite useless in the context of
       range-diff. Future versions of range-diff may learn to interpret such options in a manner
       specific to range-diff (e.g. for --stat producing human-readable output which summarizes
       how the diffstat changed).


       This command uses the diff.color.* and pager.range-diff settings (the latter is on by
       default). See git-config(1).


       When a rebase required merge conflicts to be resolved, compare the changes introduced by
       the rebase directly afterwards using:

           $ git range-diff @{u} @{1} @

       A typical output of git range-diff would look like this:

           -:  ------- > 1:  0ddba11 Prepare for the inevitable!
           1:  c0debee = 2:  cab005e Add a helpful message at the start
           2:  f00dbal ! 3:  decafe1 Describe a bug
               @@ -1,3 +1,3 @@
                Author: A U Thor <>

               -TODO: Describe a bug
               +Describe a bug
               @@ -324,5 +324,6
                 This is expected.

               -+What is unexpected is that it will also crash.
               ++Unexpectedly, it also crashes. This is a bug, and the jury is
               ++still out there how to fix it best. See ticket #314 for details.

           3:  bedead < -:  ------- TO-UNDO

       In this example, there are 3 old and 3 new commits, where the developer removed the 3rd,
       added a new one before the first two, and modified the commit message of the 2nd commit as
       well its diff.

       When the output goes to a terminal, it is color-coded by default, just like regular git
       diff's output. In addition, the first line (adding a commit) is green, the last line
       (deleting a commit) is red, the second line (with a perfect match) is yellow like the
       commit header of git show's output, and the third line colors the old commit red, the new
       one green and the rest like git show's commit header.

       A naive color-coded diff of diffs is actually a bit hard to read, though, as it colors the
       entire lines red or green. The line that added "What is unexpected" in the old commit, for
       example, is completely red, even if the intent of the old commit was to add something.

       To help with that, range uses the --dual-color mode by default. In this mode, the diff of
       diffs will retain the original diff colors, and prefix the lines with -/+ markers that
       have their background red or green, to make it more obvious that they describe how the
       diff itself changed.


       The general idea is this: we generate a cost matrix between the commits in both commit
       ranges, then solve the least-cost assignment.

       The cost matrix is populated thusly: for each pair of commits, both diffs are generated
       and the "diff of diffs" is generated, with 3 context lines, then the number of lines in
       that diff is used as cost.

       To avoid false positives (e.g. when a patch has been removed, and an unrelated patch has
       been added between two iterations of the same patch series), the cost matrix is extended
       to allow for that, by adding fixed-cost entries for wholesale deletes/adds.

       Example: Let commits 1--2 be the first iteration of a patch series and A--C the second
       iteration. Let’s assume that A is a cherry-pick of 2, and C is a cherry-pick of 1 but with
       a small modification (say, a fixed typo). Visualize the commits as a bipartite graph:

               1            A

               2            B


       We are looking for a "best" explanation of the new series in terms of the old one. We can
       represent an "explanation" as an edge in the graph:

               1            A
               2 --------'  B


       This explanation comes for "free" because there was no change. Similarly C could be
       explained using 1, but that comes at some cost c>0 because of the modification:

               1 ----.      A
                     |    /
               2 ----+---'  B
                     `----- C

       In mathematical terms, what we are looking for is some sort of a minimum cost bipartite
       matching; ‘1` is matched to C at some cost, etc. The underlying graph is in fact a
       complete bipartite graph; the cost we associate with every edge is the size of the diff
       between the two commits’ patches. To explain also new commits, we introduce dummy nodes on
       both sides:

               1 ----.      A
                     |    /
               2 ----+---'  B
               o     `----- C
               o            o

               o            o

       The cost of an edge o--C is the size of C's diff, modified by a fudge factor that should
       be smaller than 100%. The cost of an edge o--o is free. The fudge factor is necessary
       because even if 1 and C have nothing in common, they may still share a few empty lines and
       such, possibly making the assignment 1--C, o--o slightly cheaper than 1--o, o--C even if 1
       and C have nothing in common. With the fudge factor we require a much larger common part
       to consider patches as corresponding.

       The overall time needed to compute this algorithm is the time needed to compute n+m commit
       diffs and then n*m diffs of patches, plus the time needed to compute the least-cost
       assignment between n and m diffs. Git uses an implementation of the Jonker-Volgenant
       algorithm to solve the assignment problem, which has cubic runtime complexity. The
       matching found in this case will look like this:

               1 ----.      A
                     |    /
               2 ----+---'  B
               o -'  `----- C
               o ---------- o

               o ---------- o




       Part of the git(1) suite