← Back to context

Comment by mjn

13 years ago

Git's built-in diff algorithm is particularly bad for text. Since it's aimed at line-oriented code, it does line-based diffs, which is horrible for ASCII text that is reflowed, because every line in a paragraph will show up as changed for a small change.

Example: https://github.com/divegeek/uscode/commit/1fb2d83137dad1c6ca...

What's happened is that "Section 2" was moved later in the sentence, abbreviated as "Sec. 2", "of" was deleted, and "act" was capitalized:

    Section 2 of act July 30, 1947, ch. 392, 61 Stat. 674, provided...

    Act July 30, 1947, ch. 392, Sec. 2, 61 Stat. 674, provided...

The rest of the paragraph is unchanged, but git shows a 6-line diff with the entire paragraph replaced. GitHub attempts to do some word-based highlighting (see the timestamp lines), but it falls down on most of these paragraphs. Wikipedia's diffing tends to work better for this kind of thing; I'm not sure what they use. The upshot is that the number of lines changed may be a 5-10x overestimate.

7 comments

mjn

Reply
gmartres  13 years ago

> Since it's aimed at line-oriented code, it does line-based diffs

You can do word diffs with git:

    git diff --word-diff=color

  • DannyBee  13 years ago

    It's still recreating a word based diff from a line based diff.

    See diff.c line 793 for how it works.

    • azernik  13 years ago

      It may be doing that conversion, but the conversion works. For example, committing the following text (with line breaks), then joining it all into one line, shows no differences when using 'git diff --word-diff'.

        Test the first. This will check if reflowing
  text actually produces git word-diff weirdness,
  or if it's actually decent.

      The line does get reproduced on the terminal (a line diff was seen), but no text is shown in green or red to indicate an actual change.

      3 replies →

stock_toaster  13 years ago

If there are sporadic line differences, for that git diff supports different algorithms than the default. patience or histogram may work better.

as far as words in lines, you do have a point.