commit a2bb801f6a430f6049e5c9729a8f3bf9097d9b34
author SZEDER Gábor <szeder.dev@gmail.com> Wed Aug 21 13:04:24 2019 +0200
committer Junio C Hamano <gitster@pobox.com> Wed Aug 21 10:17:54 2019 -0700
tree 798626b2a9507806678078f85e94ef0416a15c99
parent eef5204190e6b99c9d3694fc416bd031cf253490

line-log: avoid unnecessary full tree diffs

With rename detection enabled the line-level log is able to trace the
evolution of line ranges across whole-file renames [1].  Alas, to
achieve that it uses the diff machinery very inefficiently, making the
operation very slow [2].  And since rename detection is enabled by
default, the line-level log is very slow by default.

When the line-level log processes a commit with rename detection
enabled, it currently does the following (see queue_diffs()):

  1. Computes a full tree diff between the commit and (one of) its
     parent(s), i.e. invokes diff_tree_oid() with an empty
     'diffopt->pathspec'.
  2. Checks whether any paths in the line ranges were modified.
  3. Checks whether any modified paths in the line ranges are missing
     in the parent commit's tree.
  4. If there is such a missing path, then calls diffcore_std() to
     figure out whether the path was indeed renamed based on the
     previously computed full tree diff.
  5. Continues doing stuff that are unrelated to the slowness.

So basically the line-level log computes a full tree diff for each
commit-parent pair in step (1) to be used for rename detection in step
(4) in the off chance that an interesting path is missing from the
parent.

Avoid these expensive and mostly unnecessary full tree diffs by
limiting the diffs to paths in the line ranges.  This is much cheaper,
and makes step (2) unnecessary.  If it turns out that an interesting
path is missing from the parent, then fall back and compute a full
tree diff, so the rename detection will still work.

Care must be taken when to update the pathspec used to limit the diff
in case of renames.  A path might be renamed on one branch and
modified on several parallel running branches, and while processing
commits on these branches the line-level log might have to alternate
between looking at a path's new and old name.  However, at any one
time there is only a single 'diffopt->pathspec'.

So add a step (0) to the above to ensure that the paths in the
pathspec match the paths in the line ranges associated with the
currently processed commit, and re-parse the pathspec from the paths
in the line ranges if they differ.

The new test cases include a specially crafted piece of history with
two merged branches and two files, where each branch modifies both
files, renames on of them, and then modifies both again.  Then two
separate 'git log -L' invocations check the line-level log of each of
those two files, which ensures that at least one of those invocations
have to do that back-and-forth between the file's old and new name (no
matter which branch is traversed first).  't/t4211-line-log.sh'
already contains two tests involving renames, they don't don't trigger
this back-and-forth.

Avoiding these unnecessary full tree diffs can have huge impact on
performance, especially in big repositories with big trees and mergy
history.  Tracing the evolution of a function through the whole
history:

  # git.git
  $ time git --no-pager log -L:read_alternate_refs:sha1-file.c v2.23.0

  Before:

    real    0m8.874s
    user    0m8.816s
    sys     0m0.057s

  After:

    real    0m2.516s
    user    0m2.456s
    sys     0m0.060s

  # linux.git
  $ time ~/src/git/git --no-pager log \
    -L:build_restore_work_registers:arch/mips/mm/tlbex.c v5.2

  Before:

    real    3m50.033s
    user    3m48.041s
    sys     0m0.300s

  After:

    real    0m2.599s
    user    0m2.466s
    sys     0m0.157s

That's just over 88x speedup.

[1] Line-level log's rename following is quite similar to 'git log
    --follow path', with the notable differences that it does handle
    multiple paths at once as well, and that it doesn't show the
    commit performing the rename if it's an exact rename.

[2] This slowness might not have been apparent initially, because back
    when the line-level log feature was introduced rename detection
    was not yet enabled by default; 12da1d1f6f (Implement line-history
    search (git log -L), 2013-03-28) and 5404c116aa (diff: activate
    diff.renames by default, 2016-02-25).

Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>