Re: [PATCH] xdiff/xpatience: support anchoring line(s)

[Stefan Beller <sbeller@google.com>]

On Wed, Nov 22, 2017 at 3:41 PM, Jonathan Tan <jonathantanmy@google.com> wrote:
> Teach the patience diff to attempt preventing user-specified lines from
> appearing as a deletion or addition in the end result. The end user can
> use this by specifying "--anchor=<text>" one or more times when using
> Git commands like "diff" and "show".
>
> Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
> ---
> Actual patch instead of RFC.
>
> One thing that might help is to warn if --anchor is used without
> --patience, but I couldn't find a good place to put that warning. Let me
> know if you know of a good place.

Would it make sense to have `--anchor` imply patience?
(not necessarily in this patch, might be a "yes, let's do
it in a year when users complain")

> Replying to Stefan's and Junio's comments:
>
>> The solution you provide is a good thing to experiment with, but
>> longer term, I would want to have huge record of configs in which
>> humans selected the best diff, such that we can use that data
>> to reason about better automatic diff generation.
>> The diff heuristic was based on a lot of human generated data,
>> that was generated by Michael at the time. I wonder if we want to
>> permanently store the anchor so the data collection will happen
>> automatically over time.
>
> I think machine learning is beyond the scope of this patch :-)

agreed; I just wanted to share what I think we could do in the future
to select sane default. For that we'd want to collect some "most useful"
configurations.

When I proposed separate flags for the move detection regarding
ignoring whitespaces, the question "how is the user sanely select
from so many flags?" came up. And in that spirit I would want think
adding this rather fundamental flag, and then machine learn (e.g. the
weights in traversing the diff matrix) off of this collected data later
might be a viable approach.

>> or rather: "c is not moved, we don't care how the diff actually looks
>> like",
>> so maybe
>>       ! grep "+c" diff
>
> I think it's less error-prone to show "a" moving. With this, if the
> command somehow prints nothing, the test would still pass.

Makes sense.

> diff --git a/t/t4033-diff-patience.sh b/t/t4033-diff-patience.sh
> index 113304dc5..2d00d1056 100755
> --- a/t/t4033-diff-patience.sh
> +++ b/t/t4033-diff-patience.sh

I was waiting for

    test_expect_success 'one --anchor anchors many lines' '
        printf "a\nb\na\nc\na\n" >file && # many 'a's
        ....
        --anchor=a
        ...


Thanks for writing this patch,
I hope we can make use of this addition eventually a lot. :)

Stefan