We find rename candidates by computing a fingerprint hash of
each file, and then comparing those fingerprints. There are
inherently O(n^2) comparisons, so it pays in CPU time to
hoist the (rather expensive) computation of the fingerprint
out of that loop (or to cache it once we have computed it once).
Previously, we didn't keep the filespec information around
because then we had the potential to consume a great deal of
memory. However, instead of keeping all of the filespec
data, we can instead just keep the fingerprint.
This patch implements and uses diff_free_filespec_data_large
to accomplish that goal. We also have to change
estimate_similarity not to needlessly repopulate the
filespec data when we already have the hash.
Practical tests showed 4.5x speedup for a 10% memory usage
increase.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This adds more proper rename detection limits. Instead of just checking
the limit against the number of potential rename destinations, we verify
that the rename matrix (which is what really matters) doesn't grow
ridiculously large, and we also make sure that we don't overflow when
doing the matrix size calculation.
This also changes the default limits from unlimited, to a rename matrix
that is limited to 100 entries on a side. You can raise it with the config
entry, or by using the "-l<n>" command line flag, but at least the default
is now a sane number that avoids spending lots of time (and memory) in
situations that likely don't merit it.
The choice of default value is of course very debatable. Limiting the
rename matrix to a 100x100 size will mean that even if you have just one
obvious rename, but you also create (or delete) 10,000 files, the rename
matrix will be so big that we disable the heuristics. Sounds reasonable to
me, but let's see if people hit this (and, perhaps more importantly,
actually *care*) in real life.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* jc/diffcore:
diffcore-delta.c: Ignore CR in CRLF for text files
diffcore-delta.c: update the comment on the algorithm.
diffcore_filespec: add is_binary
diffcore_count_changes: pass diffcore_filespec
We may want to use richer information on the data we are dealing
with in this function, so instead of passing a buffer address
and length, just pass the diffcore_filespec structure. Existing
callers always call this function with parameters taken from a
filespec anyway, so there is no functionality changes.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This implements a suggestion from Johannes. It uses a separate field in
struct diff_score to keep the result of the file name comparison in the
rename detection logic. This reverts the value of the similarity index
to be a function of file contents, only, and basename comparison is only
used to decide between files with equal amounts of content changes.
Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When there are several candidates for a rename source, and one of them
has an identical basename to the rename target, take that one.
Noticed by Govind Salinas, posted by Shawn O. Pearce, partial patch
by Linus Torvalds.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We released the postimage candidate blobs after we are done to reduce
memory pressure. Do the same for preimage candidate blobs.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Some systems have sizeof(off_t) == 8 while sizeof(size_t) == 4.
This implies that we are able to access and work on files whose
maximum length is around 2^63-1 bytes, but we can only malloc or
mmap somewhat less than 2^32-1 bytes of memory.
On such a system an implicit conversion of off_t to size_t can cause
the size_t to wrap, resulting in unexpected and exciting behavior.
Right now we are working around all gcc warnings generated by the
-Wshorten-64-to-32 option by passing the off_t through xsize_t().
In the future we should make xsize_t on such problematic platforms
detect the wrapping and die if such a file is accessed.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
When comparing file contents during the second loop through a rename
detection attempt we can skip the expensive byte-by-byte comparsion
if both source and destination files have valid SHA1 values. This
improves performance by avoiding either an expensive open/mmap to
read the working tree copy, or an expensive inflate of a blob object.
Unfortunately we still have to at least initialize the sizes of the
source and destination files even if the SHA1 values don't match.
Failing to initialize the sizes causes a number of test cases to fail
and start reporting different copy/rename behavior than was expected.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The idea is that we are interested in renaming into only one path, so
we do not care about renames that happen elsewhere.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Introduces global inline:
hashcmp(const unsigned char *sha1, const unsigned char *sha2)
Uses memcmp for comparison and returns the result based on the length of
the hash name (a future runtime decision).
Acked-by: Alex Riesen <raa.lkml@gmail.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The final output from diff used to compare pathnames between
preimage and postimage to tell if the filepair is a rename/copy.
By explicitly marking the filepair created by diffcore_rename(),
the output routine, resolve_rename_copy(), does not have to do
so anymore. This helps feeding a filepair that has different
pathnames in one and two elements to the diff machinery (most
notably, comparing two blobs).
Signed-off-by: Junio C Hamano <junkio@cox.net>
When a broken pair is matched up by rename detector to be merged
back, we do not want to say it is "dissimilar" with the
similarity index. The output should just say they were changed,
taking the break score left by the earlier diffcore-break run if
any.
Signed-off-by: Junio C Hamano <junkio@cox.net>
The "score" calculation for diffcore-rename was totally broken.
It scaled "score" as
score = src_copied * MAX_SCORE / dst->size;
which means that you got a 100% similarity score even if src and dest were
different, if just every byte of dst was copied from src, even if source
was much larger than dst (eg we had copied 85% of the bytes, but _deleted_
the remaining 15%).
That's clearly bogus. We should do the score calculation relative not to
the destination size, but to the max size of the two.
This seems to fix it.
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
To reduce wasted memory, wait until the hash fills up more
densely before we rehash. This reduces the working set size a
bit further.
Signed-off-by: Junio C Hamano <junkio@cox.net>
This changes diffcore-rename to reuse statistics information
gathered during similarity estimation, and updates the hashtable
implementation used to keep track of the statistics to be
denser. This seems to give better performance.
Signed-off-by: Junio C Hamano <junkio@cox.net>
The "similarity" logic was giving added material way too much
negative weight. What we wanted to see was how similar the
post-change image was compared to the pre-change image, so the
natural definition of similarity is how much common things are
there, relative to the post-change image's size.
This simplifies things a lot.
Signed-off-by: Junio C Hamano <junkio@cox.net>
This is to rework diffcore break/rename/copy detection code
so that it does not affected when deltifier code gets improved.
Signed-off-by: Junio C Hamano <junkio@cox.net>
dietlibc versions of malloc, calloc and realloc all return NULL if
they're told to allocate 0 bytes, causes the x* wrappers to die().
There are several more places where these calls could end up asking
for 0 bytes, too...
Maybe simply not die()-ing in the x* wrappers if 0/NULL is returned
when the requested size is zero is a safer and easier way to go.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
Signed-off-by: Junio C Hamano <junkio@cox.net>
When the user is interested in pure renames, there is no point
doing the similarity scores. This changes the score argument
parsing to special case -M100 (otherwise, it is a precision
scaled value 0 <= v < 1 and would mean 0.1, not 1.0 --- if you
do mean 0.1, you can say -M1), and optimizes the diffcore_rename
transformation to only look at pure renames in that case.
Signed-off-by: Junio C Hamano <junkio@cox.net>
A while ago, a rename-detection limit logic was implemented as a
response to this thread:
http://marc.theaimsgroup.com/?l=git&m=112413080630175
where gitweb was found to be using a lot of time and memory to
detect renames on huge commits. git-diff family takes -l<num>
flag, and if the number of paths that are rename destination
candidates (i.e. new paths with -M, or modified paths with -C)
are larger than that number, skips rename/copy detection even
when -M or -C is specified on the command line.
This commit makes the rename detection limit easier to use. You
can have:
[diff]
renamelimit = 30
in your .git/config file to specify the default rename detection
limit. You can override this from the command line; giving 0
means 'unlimited':
git diff -M -l0
We might want to change the default behaviour, when you do not
have the configuration, to limit it to say 20 paths or so. This
would also help the diffstat generation after a big 'git pull'.
Signed-off-by: Junio C Hamano <junkio@cox.net>
When many paths are modified, rename detection takes a lot of time.
The new option -l<num> can be used to disable rename detection when
more than <num> paths are possibly created as renames.
Signed-off-by: Junio C Hamano <junkio@cox.net>
It is a bit embarrassing that it took this long for a fix since the
problem was first reported on Aug 13th.
Message-ID: <87y876gl1r.wl@mail2.atmark-techno.com>
From: Yasushi SHOJI <yashi@atmark-techno.com>
Newsgroups: gmane.comp.version-control.git
Subject: [patch] possible memory leak in diff.c::diff_free_filepair()
Date: Sat, 13 Aug 2005 19:58:56 +0900
This time I used valgrind to make sure that it does not overeagerly
discard memory that is still being used.
Signed-off-by: Junio C Hamano <junkio@cox.net>
When (A,B) ==> (B,C) rename-copy was detected, we incorrectly said
that C was created by copying B. This is because we only check if the
path of rename/copy source still exists in the resulting tree to see
if the file is renamed out of existence. In this case, the new B is
created by copying or renaming A, so the original B is lost and we
should say C is a rename of B not a copy of B.
Signed-off-by: Junio C Hamano <junkio@cox.net>
The diff_delta() interface was extended to reject generating too big a
delta while we were working on the packed GIT archive format.
Take advantage of that when generating delta in the similarity estimator
used in diffcore-rename.c
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Anything that generates a delta to see if two objects are close usually
isn't interested in the delta ends up being bigger than some specified
size, and this allows us to stop delta generation early when that
happens.
When rename/copy uses a file that was broken by diffcore-break
as the source, and the broken filepair gets merged back later,
the output was mislabeled as a rename. In this case, the source
file ends up staying in the output, so we should label it as a
copy instead.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This cleans up diff_scoreopt_parse() function that is used to
parse the fractional notation -B, -C and -M option takes. The
callers are modified to check for errors and complain. Earlier
they silently ignored malformed input and falled back on the
default.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Make it return copied source and insertion separately, so that
later implementation of heuristics can use them more flexibly.
This does not change the heuristics implemented in
diffcore-rename nor diffcore-break in any way.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
A new diffcore transformation, diffcore-break.c, is introduced.
When the -B flag is given, a patch that represents a complete
rewrite is broken into a deletion followed by a creation. This
makes it easier to review such a complete rewrite patch.
The -B flag takes the same syntax as the -M and -C flags to
specify the minimum amount of non-source material the resulting
file needs to have to be considered a complete rewrite, and
defaults to 99% if not specified.
As the new test t4008-diff-break-rewrite.sh demonstrates, if a
file is a complete rewrite, it is broken into a delete/create
pair, which can further be subjected to the usual rename
detection if -M or -C is used. For example, if file0 gets
completely rewritten to make it as if it were rather based on
file1 which itself disappeared, the following happens:
The original change looks like this:
file0 --> file0' (quite different from file0)
file1 --> /dev/null
After diffcore-break runs, it would become this:
file0 --> /dev/null
/dev/null --> file0'
file1 --> /dev/null
Then diffcore-rename matches them up:
file1 --> file0'
The internal score values are finer grained now. Earlier
maximum of 10000 has been raised to 60000; there is no user
visible changes but there is no reason to waste available bits.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The commit 15d061b435
[PATCH] Fix the way diffcore-rename records unremoved source.
still leaves unneeded delete records in its output stream by
mistake, which was covered up by having an extra check to turn
such a delete into a no-op downstream. Fix the check in the
diffcore-rename to simplify the output routine.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
A new macro, DIFF_PAIR_RENAME(), is introduced to distinguish a
filepair that is a rename/copy (the definition of which is src
and dst are different paths, of course). This removes the hack
used in the record_rename_pair() to always put a non-zero value
in the score field.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This attempts to optimize "diff-tree -[CM] --stdin", which
compares successible tree pairs. This optimization does not
make much sense for other commands in the diff-* brothers.
When reading from --stdin and using rename/copy detection, the
patch makes diff-tree to read the current index file first.
This is done to reuse the optimization used by diff-cache in the
non-cached case. Similarity estimator can avoid expanding a
blob if the index says what is in the work tree has an exact
copy of that blob already expanded.
Another optimization the patch makes is to check only file sizes
first to terminate similarity estimation early. In order for
this to work, it needs a way to tell the size of the blob
without expanding it. Since an obvious way of doing it, which
is to keep all the blobs previously used in the memory, is too
costly, it does so by keeping the filesize for each object it
has already seen in memory.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Earier version of diffcore-rename used to keep unmodified
filepair in its output so that the last stage of the processing
that tells renames from copies can make all of rename/copy to
copies. However this had a bad interaction with other diffcore
filters that wanted to run after diffcore-rename, in that such
unmodified filepair must be retained for proper distinction
between renames and copies to happen.
This patch fixes the problem by changing the way diffcore-rename
records the information needed to distinguish "all are copies"
case and "the last one is a rename" case.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This removes a field that is no longer used from diff_score
structure.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This introduces a new function to free a common data structure,
and plugs some leaks.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The math to reject delta that is too big was confused.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The second round similarity estimator simply used the size of
the xdelta itself to estimate the extent of damage. This patch
keeps that logic to detect big insertions to terminate the check
early, but otherwise looks at the generated delta in order to
estimate the extent of edit more accurately.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Earlier implementation had a major screw-up in the memory
management area. Rename/copy logic sometimes borrowed a pointer
to a structure without any provision for downstream to determine
which pointer is shared and which is not. This resulted in the
later clean-up code to sometimes double free such structure,
resulting in a segfault. This made -M and -C useless.
Another problem the earlier implementation had was that it
reordered the patches, and forced the logic to differentiate
renames and copies to depend on that particular order. This
problem was fixed by teaching rename/copy detection logic not to
do any reordering, and rename-copy differentiator not to depend
on the order of the patches. The diffs will leave rename/copy
detector in the same destination path order as the patch that
was fed into it. Some test vectors have been reordered to
accommodate this change.
It also adds a sanity check logic to the human-readable diff-raw
output to detect paths with embedded TAB and LF characters,
which cannot be expressed with that format. This idea came up
during a discussion with Chris Wedgwood.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
For later stages to reorder patches, pruning logic and rename detection
logic should not decide which delete to discard (because another entry
said it will take over the file as a rename) until the very end.
Also fix some tests that were assuming the earlier "last one is rename
or keep everything else is copy" semantics of diff-raw format, which no
longer is true.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The rename/copy detection logic in earlier round was only good
enough to show patch output and discussion on the mailing list
about the diff-raw format updates revealed many problems with
it. This patch fixes all the ones known to me, without making
things I want to do later impossible, mostly related to patch
reordering.
(1) Earlier rename/copy detector determined which one is rename
and which one is copy too early, which made it impossible
to later introduce diffcore transformers to reorder
patches. This patch fixes it by moving that logic to the
very end of the processing.
(2) Earlier output routine diff_flush() was pruning all the
"no-change" entries indiscriminatingly. This was done due
to my false assumption that one of the requirements in the
diff-raw output was not to show such an entry (which
resulted in my incorrect comment about "diff-helper never
being able to be equivalent to built-in diff driver"). My
special thanks go to Linus for correcting me about this.
When we produce diff-raw output, for the downstream to be
able to tell renames from copies, sometimes it _is_
necessary to output "no-change" entries, and this patch
adds diffcore_prune() function for doing it.
(3) Earlier diff_filepair structure was trying to be not too
specific about rename/copy operations, but the purpose of
the structure was to record one or two paths, which _was_
indeed about rename/copy. This patch discards xfrm_msg
field which was trying to be generic for this wrong reason,
and introduces a couple of fields (rename_score and
rename_rank) that are explicitly specific to rename/copy
logic. One thing to note is that the information in a
single diff_filepair structure _still_ does not distinguish
renames from copies, and it is deliberately so. This is to
allow patches to be reordered in later stages.
(4) This patch also adds some tests about diff-raw format
output and makes sure that necessary "no-change" entries
appear on the output.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Earlier round was not treating symbolic links carefully enough,
and would have produced diff output that renamed/copied then
edited the contents of a symbolic link, which made no practical
sense. Change it to detect only pure renames.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This moves the path selection logic from individual programs to a new
diffcore transformer (diff-tree still needs to have its own for
performance reasons). Also the header printing code in diff-tree was
tweaked not to produce anything when pickaxe is in effect and there is
nothing interesting to report. An interesting example is the following
in the GIT archive itself:
$ git-whatchanged -p -C -S'or something in a real script'
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
When the minimum score is specified as 0 (meaning "use default
value"), set it to the default as we are told.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
There was a screwy math bug in the estimator that confused what
-C1 meant and what -C9 meant, only in one of the early "cheap"
check, which resulted in quite confusing behaviour.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Update the diff-raw format as Linus and I discussed, except that
it does not use sequence of underscore '_' letters to express
nonexistence. All '0' mode is used for that purpose instead.
The new diff-raw format can express rename/copy, and the earlier
restriction that -M and -C _must_ be used with the patch format
output is no longer necessary. The patch makes -M and -C flags
independent of -p flag, so you need to say git-whatchanged -M -p
to get the diff/patch format.
Updated are both documentations and tests.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This does not actually supress the extra headers when pickaxe is
used, but prepares enough support for diff-tree to implement it.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The heuristics so far was to compare file size change and xdelta
size against the average of file size before and after the
change. This patch uses the smaller of pre- and post- change
file size instead.
It also makes a very small performance fix. I didn't measure
it; I do not expect it to make any practical difference, but
while scanning an already sorted list, breaking out in the
middle is the right thing.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This steals the "pickaxe" feature from JIT and make it available
to the bare Plumbing layer. From the command line, the user
gives a string he is intersted in.
Using the diff-core infrastructure previously introduced, it
filters the differences to limit the output only to the diffs
between <src> and <dst> where the string appears only in one but
not in the other. For example:
$ ./git-rev-list HEAD | ./git-diff-tree -Sdiff-tree-helper --stdin -M
would show the diffs that touch the string "diff-tree-helper".
In real software-archaeologist application, you would typically
look for a few to several lines of code and see where that code
came from.
The "pickaxe" module runs after "rename/copy detection" module,
so it even crosses the file rename boundary, as the above
example demonstrates.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This introduces the diff-core, the layer between the diff-tree
family and the external diff interface engine. The calls to the
interface diff-tree family uses (diff_change and diff_addremove)
have not changed and will not change. The purpose of the
diff-core layer is to provide an infrastructure to transform the
set of differences sent from the applications, before sending
them to the external diff interface.
The recently introduced rename detection code has been rewritten
to use the diff-core facility. When applications send in
separate creates and deletes, matching ones are transformed into
a single rename-and-edit diff, and sent out to the external diff
interface as such.
This patch also enhances the rename detection code further to be
able to detect copies. Currently this happens only as long as
copy sources appear as part of the modified files, but there
already is enough provision for callers to report unmodified
files to diff-core, so that they can be also used as copy source
candidates. Extending the callers this way will be done in a
separate patch.
Please see and marvel at how well this works by trying out the
newly added t/t4003-diff-rename-1.sh test script.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
