Re: [PATCH] Improve contrib/diff-highlight to highlight unevenly-sized hunks

On Fri, Jun 19, 2015 at 03:34:55AM -0400, Jeff King wrote:

And here's some more bad news. If you look at the diff for this
patch itself, it's terribly unreadable (the regular diff already is
pretty bad, but the highlights make it much worse). There are big chunks
where we take away 5 or 10 lines from the old code, and replace them
with totally unrelated lines. We end up highlighting almost the entire
thing, except for spaces and punctuation.

We might be able to solve this with a percentage heuristic similar to
the one Patrick proposed. It's not really interesting to highlight
unless we're doing it on probably 20% or less of the diff (where 20% is
a number I just made up).

That turned out to be pretty easy; patch is below (on top of what I sent
earlier). I set the percentage at 50% based on eyeballing "git log -p"
in git.git, and it seems to give good results.

So I think the big remaining issue is improved tokenizing. Maybe Patrick
will want to take a stab at it.

---

diff --git a/contrib/diff-highlight/diff-highlight b/contrib/diff-highlight/diff-highlight
index 1525ccc..9454446 100755
--- a/contrib/diff-highlight/diff-highlight
+++ b/contrib/diff-highlight/diff-highlight

@@ -114,12 +114,32 @@ sub show_hunk {
 			if $bits & 2;
 	}
 
+	my $highlighted = count_highlight(@highlight_a) +
+			  count_highlight(@highlight_b);
+	my $total = length($a) + length($b);
+	my $pct = $highlighted / $total;
+
+	if ($pct > 0.5) {
+		@highlight_a = ();
+		@highlight_b = ();
+	}
+
 	# And now show the output both with the original stripped annotations,
 	# as well as our new highlights.
 	show_image($a, [merge_annotations(\@stripped_a, \@highlight_a)]);
 	show_image($b, [merge_annotations(\@stripped_b, \@highlight_b)]);
 }
 
+sub count_highlight {
+	my $total = 0;
+	while (@_) {
+		my $from = shift;
+		my $to = shift;
+		$total += $to->[0] - $from->[0];
+	}
+	return $total;
+}
+
 # Strip out any diff syntax (i.e., leading +/-), along with any ANSI color
 # codes from the pre- or post-image of a hunk. The result is a string of text
 # suitable for diffing against the other side of the hunk.