Re: [RFC/PATCH 5/3] Alternative --dirstat implementation, based on diffstat analysis

On Tue, Apr 12, 2011 at 7:46 AM, Linus Torvalds
[off-list ref] wrote:

So I don't hate the idea, but I do hate the "use a config option"
part. Or rather, I hate the fact that it's the _only_ way to do it
(and the particular config name you chose).

Oh, and one thing strikes me: I think the fast dirstat gave reasonable
values when you had mixed text and binary (in the kernel tree, look
for the Documentation/logo.gif file, for example: it changed to the
Tasmanian devil in one release).

Have you checked what happens to that when you use the diffstat one?
Because binary files are done very differently (byte-based counts).

So check out

   git show --dirstat 3d4f16348b77efbf81b7fa186a18a0eb815b6b84

with and without your change. The old dirstat gives

  44.0% Documentation/
  55.9% drivers/video/logo/

which is at least not completely insane.

The reason I bring this up is because I think this was an issue at one
point, and one of the statistics things (--stat or --numstat or
--dirstat) gave absolutely horrid values (basically comparing "bytes
changed" for binaries with "lines changed" for text files). Resulting
in totally skewed statistics.

                       Linus