Description

The patches have been grouped into directories corresponding to each patch series that was sent to the mailing list.

The date and message-id for the cover letter email are provided here, for posterity's sake; it should be trivial to find the thread and/or the other emailed patches in the series on most common online mailing list archive.

Content listing

pre-v1

A preliminary patch series.

0000-cover-letter.patch

Date: Wed, 6 Jul 2011 14:15:42 +0800

Message-Id: 1309932945-5048-1-git-send-email-rctay89@gmail.com
0001-xdiff-xprepare-use-memset.patch
0002-xdiff-xpatience-factor-out-fall-back-diff-function.patch
0003-t4033-diff-patience-factor-out-tests.patch

pre-v2

v2 of the preliminary series.

0000-cover-letter.patch

Date: Thu, 7 Jul 2011 00:38:53 +0800

Message-Id: 1309970337-6016-1-git-send-email-rctay89@gmail.com
0001-xdiff-xprepare-use-memset.patch
0002-xdiff-xprepare-refactor-abort-cleanups.patch
0003-xdiff-xpatience-factor-out-fall-back-diff-function.patch
0004-t4033-diff-patience-factor-out-tests.patch

pre-v3

v3 of the preliminary series.

0000-cover-letter.patch

Date: Thu, 7 Jul 2011 12:23:54 +0800

Message-Id: 1310012638-3668-1-git-send-email-rctay89@gmail.com
0001-xdiff-xprepare-use-memset.patch
0002-xdiff-xprepare-refactor-abort-cleanups.patch
0003-xdiff-xpatience-factor-out-fall-back-diff-function.patch
0004-t4033-diff-patience-factor-out-tests.patch

v1

The actual histogram diff implementation.

0000-cover-letter.patch

Date: Tue, 12 Jul 2011 14:10:24 +0800

Message-Id: 1310451027-15148-1-git-send-email-rctay89@gmail.com
0001-teach-histogram-to-diff.patch
0002-xdiff-xprepare-skip-classification.patch
0003-xdiff-xprepare-use-a-smaller-sample-size-for-histogr.patch

v2

v2 of the implementation.

0000-cover-letter.patch

Date: Mon, 1 Aug 2011 11:16:40 +0800

Message-Id: 1312168608-10828-1-git-send-email-rctay89@gmail.com
0001-xdiff-xprepare-use-memset.patch
0002-do-away-with-xdl_mmfile_next.patch
0003-xdiff-xprepare-refactor-abort-cleanups.patch
0004-xdiff-xpatience-factor-out-fall-back-diff-function.patch
0005-t4033-diff-patience-factor-out-tests.patch
0006-teach-histogram-to-diff.patch
0007-xdiff-xprepare-skip-classification.patch
0008-xdiff-xprepare-use-a-smaller-sample-size-for-histogr.patch

v2-rebased

The 'v2' series rebased on 'next'. To integrate 'v2' would require re-rolling the series, but it already graduated to 'next', so this series was sent in to make the maintainer's life easier.

0000-cover-letter.patch

Date: Mon, 1 Aug 2011 12:20:06 +0800

Message-Id: 1312172410-4380-1-git-send-email-rctay89@gmail.com
0001-xdiff-do-away-with-xdl_mmfile_next.patch
0002-xdiff-xhistogram-rework-handling-of-recursed-results.patch
0003-xdiff-xhistogram-rely-on-xdl_trim_ends.patch
0004-xdiff-xhistogram-drop-need-for-additional-variable.patch

Un-mailed patches

raw-v*

These patches were a naive, line-to-line conversion of JGit's histogram diff implementation. You can see that information is stored as an integer bitmask, with shifts being used to read/write different fields.

ptr-v*

These patches provided an implementation of the histogram diff algorithm based on structs, instead of bitmasks. This reduced the complexity of reading/writing of different fields and made code more readable. However, there was a noticeable performance penalty, which held me back from submitting these series. After running the code through a profiler, I managed to pin the problem down to excessive malloc()s.

This implementation is the one that was submitted for inclusion.

xd-v*

These patches were an attempt at improving the C histogram diff implementation by reducing the calls to scanA() (the step that builds the occurrence table) for each recursive sub-diff. This, of course, ramped up the complexity of the code, but the performance gains were minimal, so I junked this in the end.

ptr2-v1

Like xd-v*; this was an initial try at reducing scanA() calls.

guilt/master

Behind the scenes, I used guilt guilt to manage changes to my patches. The folder contains a dump of my .git/patches/ repo (also available in this repo at the [guilt/master branch] guilt master.

guilt/trim-ends

I toyed with the idea of skipping hashing of common head and tail lines for a performance gain. However, as I worked on another diff-related issue (see 'Related'), I began to doubt the correctness of this.

This series is still a work-in-progress and has not been submitted for inclusion.

gists/gist-*

[gist: 1063164 [gsoc-diff] numbers] (https://gist.github.com/1063164)
[gist: 1063175 [gsoc-diff] more numbers] (https://gist.github.com/1063175)
[gist: 1066703 [gsoc-diff] alternate hashing ("golden ratio")] (https://gist.github.com/1066703)
[gist: 1066773 [gsoc-diff] try_lcs: skip chains] (https://gist.github.com/1066773)
[gist: 1066906 [gsoc-diff] diff --assume-text] (https://gist.github.com/1066906)
[gist: 1068912 [gsoc-diff] malloc()/free() once] (https://gist.github.com/1068912)
[gist: 1069099 [gsoc-diff] reduce number of rec accesses] (https://gist.github.com/1069099)
[gist: 1074709 [gsoc-diff] reduce number of mallocs for gd-ptr] (https://gist.github.com/1074709)
[gist: 1098863 [gsoc-diff] make trim tail support context] (https://gist.github.com/1098863)
[gist: 1099236 [gsoc-diff] skip hashing of common leading lines] (https://gist.github.com/1099236)
[gist: 1099515 [gsoc-diff] xdl_trim_head()] (https://gist.github.com/1099515)

Towards the close of the gsoc coding period, there was [a thread] 1 on the git mailing list regarding performance issues with git-diff. With the knowledge I had gained of git diff's machinery over the course of gsoc, I managed to diagnose the problem and [submitted a patch] 2.

rctay / gsoc-2011-git