git/Documentation/gitdiffcore.txt
Thomas Braun e0e7cb8080 log -G: ignore binary files
The -G<regex> option of log looks for the differences whose patch text
contains added/removed lines that match regex.

Currently -G looks also into patches of binary files (which
according to [1]) is binary as well.

This has a couple of issues:

- It makes the pickaxe search slow. In a proprietary repository of the
  author with only ~5500 commits and a total .git size of ~300MB
  searching takes ~13 seconds

    $time git log -Gwave > /dev/null

    real    0m13,241s
    user    0m12,596s
    sys     0m0,644s

  whereas when we ignore binary files with this patch it takes ~4s

    $time ~/devel/git/git log -Gwave > /dev/null

    real    0m3,713s
    user    0m3,608s
    sys     0m0,105s

  which is a speedup of more than fourfold.

- The internally used algorithm for generating patch text is based on
  xdiff and its states in [1]

  > The output format of the binary patch file is proprietary
  > (and binary) and it is basically a collection of copy and insert
  > commands [..]

  which means that the current format could change once the internal
  algorithm is changed as the format is not standardized. In addition
  the git binary patch format used for preparing patches for git apply
  is *different* from the xdiff format as can be seen by comparing

  git log -p -a

    commit 6e95bf4bafccf14650d02ab57f3affe669be10cf
    Author: A U Thor <author@example.com>
    Date:   Thu Apr 7 15:14:13 2005 -0700

        modify binary file

    diff --git a/data.bin b/data.bin
    index f414c84..edfeb6f 100644
    --- a/data.bin
    +++ b/data.bin
    @@ -1,2 +1,4 @@
     a
     a^@a
    +a
    +a^@a

  with git log --binary

    commit 6e95bf4bafccf14650d02ab57f3affe669be10cf
    Author: A U Thor <author@example.com>
    Date:   Thu Apr 7 15:14:13 2005 -0700

        modify binary file

    diff --git a/data.bin b/data.bin
    index f414c84bd3aa25fa07836bb1fb73db784635e24b..edfeb6f501[..]
    GIT binary patch
    literal 12
    QcmYe~N@Pgn0zx1O01)N^ZvX%Q

    literal 6
    NcmYe~N@Pgn0ssWg0XP5v

  which seems unexpected.

To resolve these issues this patch makes -G<regex> ignore binary files
by default. Textconv filters are supported and also -a/--text for
getting the old and broken behaviour back.

The -S<block of text> option of log looks for differences that changes
the number of occurrences of the specified block of text (i.e.
addition/deletion) in a file. As we want to keep the current behaviour,
add a test to ensure it stays that way.

[1]: http://www.xmailserver.org/xdiff.html

Signed-off-by: Thomas Braun <thomas.braun@virtuell-zuhause.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-12-26 14:59:37 -08:00

293 lines
12 KiB
Plaintext

gitdiffcore(7)
==============
NAME
----
gitdiffcore - Tweaking diff output
SYNOPSIS
--------
[verse]
'git diff' *
DESCRIPTION
-----------
The diff commands 'git diff-index', 'git diff-files', and 'git diff-tree'
can be told to manipulate differences they find in
unconventional ways before showing 'diff' output. The manipulation
is collectively called "diffcore transformation". This short note
describes what they are and how to use them to produce 'diff' output
that is easier to understand than the conventional kind.
The chain of operation
----------------------
The 'git diff-{asterisk}' family works by first comparing two sets of
files:
- 'git diff-index' compares contents of a "tree" object and the
working directory (when `--cached` flag is not used) or a
"tree" object and the index file (when `--cached` flag is
used);
- 'git diff-files' compares contents of the index file and the
working directory;
- 'git diff-tree' compares contents of two "tree" objects;
In all of these cases, the commands themselves first optionally limit
the two sets of files by any pathspecs given on their command-lines,
and compare corresponding paths in the two resulting sets of files.
The pathspecs are used to limit the world diff operates in. They remove
the filepairs outside the specified sets of pathnames. E.g. If the
input set of filepairs included:
------------------------------------------------
:100644 100644 bcd1234... 0123456... M junkfile
------------------------------------------------
but the command invocation was `git diff-files myfile`, then the
junkfile entry would be removed from the list because only "myfile"
is under consideration.
The result of comparison is passed from these commands to what is
internally called "diffcore", in a format similar to what is output
when the -p option is not used. E.g.
------------------------------------------------
in-place edit :100644 100644 bcd1234... 0123456... M file0
create :000000 100644 0000000... 1234567... A file4
delete :100644 000000 1234567... 0000000... D file5
unmerged :000000 000000 0000000... 0000000... U file6
------------------------------------------------
The diffcore mechanism is fed a list of such comparison results
(each of which is called "filepair", although at this point each
of them talks about a single file), and transforms such a list
into another list. There are currently 5 such transformations:
- diffcore-break
- diffcore-rename
- diffcore-merge-broken
- diffcore-pickaxe
- diffcore-order
These are applied in sequence. The set of filepairs 'git diff-{asterisk}'
commands find are used as the input to diffcore-break, and
the output from diffcore-break is used as the input to the
next transformation. The final result is then passed to the
output routine and generates either diff-raw format (see Output
format sections of the manual for 'git diff-{asterisk}' commands) or
diff-patch format.
diffcore-break: For Splitting Up Complete Rewrites
--------------------------------------------------
The second transformation in the chain is diffcore-break, and is
controlled by the -B option to the 'git diff-{asterisk}' commands. This is
used to detect a filepair that represents "complete rewrite" and
break such filepair into two filepairs that represent delete and
create. E.g. If the input contained this filepair:
------------------------------------------------
:100644 100644 bcd1234... 0123456... M file0
------------------------------------------------
and if it detects that the file "file0" is completely rewritten,
it changes it to:
------------------------------------------------
:100644 000000 bcd1234... 0000000... D file0
:000000 100644 0000000... 0123456... A file0
------------------------------------------------
For the purpose of breaking a filepair, diffcore-break examines
the extent of changes between the contents of the files before
and after modification (i.e. the contents that have "bcd1234..."
and "0123456..." as their SHA-1 content ID, in the above
example). The amount of deletion of original contents and
insertion of new material are added together, and if it exceeds
the "break score", the filepair is broken into two. The break
score defaults to 50% of the size of the smaller of the original
and the result (i.e. if the edit shrinks the file, the size of
the result is used; if the edit lengthens the file, the size of
the original is used), and can be customized by giving a number
after "-B" option (e.g. "-B75" to tell it to use 75%).
diffcore-rename: For Detecting Renames and Copies
-------------------------------------------------
This transformation is used to detect renames and copies, and is
controlled by the -M option (to detect renames) and the -C option
(to detect copies as well) to the 'git diff-{asterisk}' commands. If the
input contained these filepairs:
------------------------------------------------
:100644 000000 0123456... 0000000... D fileX
:000000 100644 0000000... 0123456... A file0
------------------------------------------------
and the contents of the deleted file fileX is similar enough to
the contents of the created file file0, then rename detection
merges these filepairs and creates:
------------------------------------------------
:100644 100644 0123456... 0123456... R100 fileX file0
------------------------------------------------
When the "-C" option is used, the original contents of modified files,
and deleted files (and also unmodified files, if the
"--find-copies-harder" option is used) are considered as candidates
of the source files in rename/copy operation. If the input were like
these filepairs, that talk about a modified file fileY and a newly
created file file0:
------------------------------------------------
:100644 100644 0123456... 1234567... M fileY
:000000 100644 0000000... bcd3456... A file0
------------------------------------------------
the original contents of fileY and the resulting contents of
file0 are compared, and if they are similar enough, they are
changed to:
------------------------------------------------
:100644 100644 0123456... 1234567... M fileY
:100644 100644 0123456... bcd3456... C100 fileY file0
------------------------------------------------
In both rename and copy detection, the same "extent of changes"
algorithm used in diffcore-break is used to determine if two
files are "similar enough", and can be customized to use
a similarity score different from the default of 50% by giving a
number after the "-M" or "-C" option (e.g. "-M8" to tell it to use
8/10 = 80%).
Note. When the "-C" option is used with `--find-copies-harder`
option, 'git diff-{asterisk}' commands feed unmodified filepairs to
diffcore mechanism as well as modified ones. This lets the copy
detector consider unmodified files as copy source candidates at
the expense of making it slower. Without `--find-copies-harder`,
'git diff-{asterisk}' commands can detect copies only if the file that was
copied happened to have been modified in the same changeset.
diffcore-merge-broken: For Putting Complete Rewrites Back Together
------------------------------------------------------------------
This transformation is used to merge filepairs broken by
diffcore-break, and not transformed into rename/copy by
diffcore-rename, back into a single modification. This always
runs when diffcore-break is used.
For the purpose of merging broken filepairs back, it uses a
different "extent of changes" computation from the ones used by
diffcore-break and diffcore-rename. It counts only the deletion
from the original, and does not count insertion. If you removed
only 10 lines from a 100-line document, even if you added 910
new lines to make a new 1000-line document, you did not do a
complete rewrite. diffcore-break breaks such a case in order to
help diffcore-rename to consider such filepairs as candidate of
rename/copy detection, but if filepairs broken that way were not
matched with other filepairs to create rename/copy, then this
transformation merges them back into the original
"modification".
The "extent of changes" parameter can be tweaked from the
default 80% (that is, unless more than 80% of the original
material is deleted, the broken pairs are merged back into a
single modification) by giving a second number to -B option,
like these:
* -B50/60 (give 50% "break score" to diffcore-break, use 60%
for diffcore-merge-broken).
* -B/60 (the same as above, since diffcore-break defaults to 50%).
Note that earlier implementation left a broken pair as a separate
creation and deletion patches. This was an unnecessary hack and
the latest implementation always merges all the broken pairs
back into modifications, but the resulting patch output is
formatted differently for easier review in case of such
a complete rewrite by showing the entire contents of old version
prefixed with '-', followed by the entire contents of new
version prefixed with '+'.
diffcore-pickaxe: For Detecting Addition/Deletion of Specified String
---------------------------------------------------------------------
This transformation limits the set of filepairs to those that change
specified strings between the preimage and the postimage in a certain
way. -S<block of text> and -G<regular expression> options are used to
specify different ways these strings are sought.
"-S<block of text>" detects filepairs whose preimage and postimage
have different number of occurrences of the specified block of text.
By definition, it will not detect in-file moves. Also, when a
changeset moves a file wholesale without affecting the interesting
string, diffcore-rename kicks in as usual, and `-S` omits the filepair
(since the number of occurrences of that string didn't change in that
rename-detected filepair). When used with `--pickaxe-regex`, treat
the <block of text> as an extended POSIX regular expression to match,
instead of a literal string.
"-G<regular expression>" (mnemonic: grep) detects filepairs whose
textual diff has an added or a deleted line that matches the given
regular expression. This means that it will detect in-file (or what
rename-detection considers the same file) moves, which is noise. The
implementation runs diff twice and greps, and this can be quite
expensive. To speed things up binary files without textconv filters
will be ignored.
When `-S` or `-G` are used without `--pickaxe-all`, only filepairs
that match their respective criterion are kept in the output. When
`--pickaxe-all` is used, if even one filepair matches their respective
criterion in a changeset, the entire changeset is kept. This behavior
is designed to make reviewing changes in the context of the whole
changeset easier.
diffcore-order: For Sorting the Output Based on Filenames
---------------------------------------------------------
This is used to reorder the filepairs according to the user's
(or project's) taste, and is controlled by the -O option to the
'git diff-{asterisk}' commands.
This takes a text file each of whose lines is a shell glob
pattern. Filepairs that match a glob pattern on an earlier line
in the file are output before ones that match a later line, and
filepairs that do not match any glob pattern are output last.
As an example, a typical orderfile for the core Git probably
would look like this:
------------------------------------------------
README
Makefile
Documentation
*.h
*.c
t
------------------------------------------------
SEE ALSO
--------
linkgit:git-diff[1],
linkgit:git-diff-files[1],
linkgit:git-diff-index[1],
linkgit:git-diff-tree[1],
linkgit:git-format-patch[1],
linkgit:git-log[1],
linkgit:gitglossary[7],
link:user-manual.html[The Git User's Manual]
GIT
---
Part of the linkgit:git[1] suite