Re: [PATCH v1 0/1] t6038-merge-text-auto.sh

Junio C Hamano [off-list ref] writes:

Imagine that you started from a history where somebody recorded a
text file with CRLF in the blob, unconverted.  Later the project
decided to express their text with LF to support cross-platform
development better, and sets up the Auto-CRLF.  Your user is working
near the tip of that history after the eol correction happened.  Now
she gets a pull-request of a branch that forked from an old point in
the history, before the eol correction and full of CRLF.  Yes, to
integrate the change being proposed, she needs to look at "theirs";
that's the whole point of a "merge".  Why should she revert the eol
correction her history has by getting fooled by the fact that the
update was based on a part of the history before the eol correction?

Thinking aloud along this line a bit further, if you really cared
(and I don't feel very strongly, as I think "safe crlf" is an ugly
workaround that lets users keep their misconfiguration by penalizing
their day-to-day operation), you may want to think about doing a
3-way merge of "eol preference" beween all stages.

That is, if the common ancestor in stage #1 and the current version
in stage #2 both have its text in LF, and the data being merged in
stage #3 is in CRLF, you sayr "do not convert; the change being
brought in wants to have CRLF endings, while our history did not
care".  Similarly, if the common ancestor in stage #1 and the data
being merged in stage #3 both have CRLF, and your version in stage
#2 has LF, you say "We wanted to fix eol since the side branch
forked, and the side branch predates that fix, so we keep the eol
fix we did since we diverged", i.e. "Do convert".

For doing this, you may want to refactor the codepath that decides
"Auto-CRLF usually wants to turn CRLF in text to LF, but should we
disable that logic now, because the user already has CRLF in the
current one?" into a function that takes a single parameter, `path`,
and returns "Yes, do convert CRLF to LF / No, do not convert"
boolean.  Having a low level function that says "What's the blob at
this path in the index" and have the caller run that logic feels
unwieldy if we want to go that route.