How line ending conversions work with git core.autocrlf between different operating systems

Question

The best explanation of how core.autocrlf works is found on the gitattributes man page, in the text attribute section.

This is how core.autocrlf appears to work currently (or at least since v1.7.2 from what I am aware):

• core.autocrlf = true

1. Text files checked-out from the repository that have only LF characters are normalized to CRLF in your working tree; files that contain CRLF in the repository will not be touched
2. Text files that have only LF characters in the repository, are normalized from CRLF to LF when committed back to the repository. Files that contain CRLF in the repository will be committed untouched.

• core.autocrlf = input

1. Text files checked-out from the repository will keep original EOL characters in your working tree.
2. Text files in your working tree with CRLF characters are normalized to LF when committed back to the repository.

• core.autocrlf = false

1. core.eol dictates EOL characters in the text files of your working tree.
2. core.eol = native by default, which means working tree EOLs will depend on where git is running: CRLF on a Windows machine, or LF in *nix.
3. Repository gitattributes settings determines EOL character normalization for commits to the repository (default is normalization to LF characters).

I've only just recently researched this issue and I also find the situation to be very convoluted. The core.eol setting definitely helped clarify how EOL characters are handled by git.

The issue of EOLs in mixed-platform projects has been making my life miserable for a long time. The problems usually arise when there are already files with different and mixed EOLs already in the repo. This means that:

1. The repo may have different files with different EOLs
2. Some files in the repo may have mixed EOL, e.g. a combination of CRLF and LF in the same file.

How this happens is not the issue here, but it does happen.

I ran some conversion tests on Windows for the various modes and their combinations.
Here is what I got, in a slightly modified table:

                 | Resulting conversion when       | Resulting conversion when 
                 | committing files with various   | checking out FROM repo - 
                 | EOLs INTO repo and              | with mixed files in it and
                 |  core.autocrlf value:           | core.autocrlf value:           
--------------------------------------------------------------------------------
File             | true       | input      | false | true       | input | false
--------------------------------------------------------------------------------
Windows-CRLF     | CRLF -> LF | CRLF -> LF | as-is | as-is      | as-is | as-is
Unix -LF         | as-is      | as-is      | as-is | LF -> CRLF | as-is | as-is
Mac  -CR         | as-is      | as-is      | as-is | as-is      | as-is | as-is
Mixed-CRLF+LF    | as-is      | as-is      | as-is | as-is      | as-is | as-is
Mixed-CRLF+LF+CR | as-is      | as-is      | as-is | as-is      | as-is | as-is

As you can see, there are 2 cases when conversion happens on commit (3 left columns). In the rest of the cases the files are committed as-is.

Upon checkout (3 right columns), there is only 1 case where conversion happens when:

1. core.autocrlf is true and
2. the file in the repo has the LF EOL.

Most surprising for me, and I suspect, the cause of many EOL problems is that there is no configuration in which mixed EOL like CRLF+LF get normalized.

Note also that "old" Mac EOLs of CR only also never get converted.
This means that if a badly written EOL conversion script tries to convert a mixed ending file with CRLFs+LFs, by just converting LFs to CRLFs, then it will leave the file in a mixed mode with "lonely" CRs wherever a CRLF was converted to CRCRLF.
Git will then not convert anything, even in true mode, and EOL havoc continues. This actually happened to me and messed up my files really badly, since some editors and compilers (e.g. VS2010) don't like Mac EOLs.

I guess the only way to really handle these problems is to occasionally normalize the whole repo by checking out all the files in input or false mode, running a proper normalization and re-committing the changed files (if any). On Windows, presumably resume working with core.autocrlf true.

core.autocrlf value does not depend on OS type but on Windows default value is true and for Linux - input. I explored 3 possible values for commit and checkout cases and this is the resulting table:

╔═══════════════╦══════════════╦══════════════╦══════════════╗
║ core.autocrlf ║     false    ║     input    ║     true     ║
╠═══════════════╬══════════════╬══════════════╬══════════════╣
║               ║ LF   => LF   ║ LF   => LF   ║ LF   => LF   ║
║ git commit    ║ CR   => CR   ║ CR   => CR   ║ CR   => CR   ║
║               ║ CRLF => CRLF ║ CRLF => LF   ║ CRLF => LF   ║
╠═══════════════╬══════════════╬══════════════╬══════════════╣
║               ║ LF   => LF   ║ LF   => LF   ║ LF   => CRLF ║
║ git checkout  ║ CR   => CR   ║ CR   => CR   ║ CR   => CR   ║
║               ║ CRLF => CRLF ║ CRLF => CRLF ║ CRLF => CRLF ║
╚═══════════════╩══════════════╩══════════════╩══════════════╝

Things are about to change on the "eol conversion" front, with the upcoming Git 1.7.2:

A new config setting core.eol is being added/evolved:

This is a replacement for the 'Add "core.eol" config variable' commit that's currently in pu (the last one in my series).
Instead of implying that "core.autocrlf=true" is a replacement for "* text=auto", it makes explicit the fact that autocrlf is only for users who want to work with CRLFs in their working directory on a repository that doesn't have text file normalization.
When it is enabled, "core.eol" is ignored.

Introduce a new configuration variable, "core.eol", that allows the user to set which line endings to use for end-of-line-normalized files in the working directory.
It defaults to "native", which means CRLF on Windows and LF everywhere else. Note that "core.autocrlf" overrides core.eol.
This means that:

[core]
  autocrlf = true

puts CRLFs in the working directory even if core.eol is set to "lf".

core.eol:

Sets the line ending type to use in the working directory for files that have the text property set.
Alternatives are 'lf', 'crlf' and 'native', which uses the platform's native line ending.
The default value is native.

---

Other evolutions are being considered:

For 1.8, I would consider making core.autocrlf just turn on normalization and leave the working directory line ending decision to core.eol, but that will break people's setups.

---

git 2.8 (March 2016) improves the way core.autocrlf influences the eol:

See commit 817a0c7 (23 Feb 2016), commit 6e336a5, commit df747b8, commit df747b8 (10 Feb 2016), commit df747b8, commit df747b8 (10 Feb 2016), and commit 4b4024f, commit bb211b4, commit 92cce13, commit 320d39c, commit 4b4024f, commit bb211b4, commit 92cce13, commit 320d39c (05 Feb 2016) by Torsten Bögershausen (tboegi).
^{(Merged by Junio C Hamano -- gitster -- in commit c6b94eb, 26 Feb 2016)}

convert.c: refactor crlf_action

Refactor the determination and usage of crlf_action.
Today, when no "crlf" attribute are set on a file, crlf_action is set to CRLF_GUESS. Use CRLF_UNDEFINED instead, and search for "text" or "eol" as before.

Replace the old CRLF_GUESS usage:

CRLF_GUESS && core.autocrlf=true -> CRLF_AUTO_CRLF
CRLF_GUESS && core.autocrlf=false -> CRLF_BINARY
CRLF_GUESS && core.autocrlf=input -> CRLF_AUTO_INPUT

Make more clear, what is what, by defining:

- CRLF_UNDEFINED : No attributes set. Temparally used, until core.autocrlf
                   and core.eol is evaluated and one of CRLF_BINARY,
                   CRLF_AUTO_INPUT or CRLF_AUTO_CRLF is selected
- CRLF_BINARY    : No processing of line endings.
- CRLF_TEXT      : attribute "text" is set, line endings are processed.
- CRLF_TEXT_INPUT: attribute "input" or "eol=lf" is set. This implies text.
- CRLF_TEXT_CRLF : attribute "eol=crlf" is set. This implies text.
- CRLF_AUTO      : attribute "auto" is set.
- CRLF_AUTO_INPUT: core.autocrlf=input (no attributes)
- CRLF_AUTO_CRLF : core.autocrlf=true  (no attributes)

---

As torek adds in the comments:

all these translations (any EOL conversion from eol= or autocrlf settings, and "clean" filters) are run when files move from work-tree to index, i.e., during git add rather than at git commit time.
(Note that git commit -a or --only or --include do add files to the index at that time, though.)

For more on that, see "What is difference between autocrlf and eol".

Here is my understanding of it so far, in case it helps someone.

core.autocrlf=true and core.safecrlf = true

You have a repository where all the line endings are the same, but you work on different platforms. Git will make sure your lines endings are converted to the default for your platform. Why does this matter? Let's say you create a new file. The text editor on your platform will use its default line endings. When you check it in, if you don't have core.autocrlf set to true, you've introduced a line ending inconsistency for someone on a platform that defaults to a different line ending. I always set safecrlf too because I would like to know that the crlf operation is reversible. With these two settings, git is modifying your files, but it verifies that the modifications are reversible.

core.autocrlf=false

You have a repository that already has mixed line endings checked in and fixing the incorrect line endings could break other things. Its best not to tell git to convert line endings in this case, because then it will exacerbate the problem it was designed to solve - making diffs easier to read and merges less painful. With this setting, git doesn't modify your files.

core.autocrlf=input

I don't use this because the reason for this is to cover a use case where you created a file that has CRLF line endings on a platform that defaults to LF line endings. I prefer instead to make my text editor always save new files with the platform's line ending defaults.

No, the @jmlane answer is wrong.

For Checkin (git add, git commit):

1. if text property is Set, Set value to 'auto', the conversion happens enen the file has been committed with 'CRLF'
2. if text property is Unset:nothing happens, enen for Checkout
3. if text property is Unspecified, conversion depends on core.autocrlf
   1. if autocrlf = input or autocrlf = true, the conversion only happens when the file in the repository is 'LF', if it has been 'CRLF', nothing will happens.
   2. if autocrlf = false, nothing happens

For Checkout:

1. if text property is Unset: nothing happens.
2. if text property is Set, Set value to 'auto: it depends on core.autocrlf, core.eol.
   1. core.autocrlf = input : nothing happens
   2. core.autocrlf = true : the conversion only happens when the file in the repository is 'LF', 'LF' -> 'CRLF'
   3. core.autocrlf = false : the conversion only happens when the file in the repository is 'LF', 'LF' -> core.eol
3. if text property is Unspecified, it depends on core.autocrlf.
   1. the same as 2.1
   2. the same as 2.2
   3. None, nothing happens, core.eol is not effective when text property is Unspecified

Default behavior

So the Default behavior is text property is Unspecified and core.autocrlf = false:

1. for checkin, nothing happens
2. for checkout, nothing happens

Conclusions

1. if text property is set, checkin behavior is depends on itself, not on autocrlf
2. autocrlf or core.eol is for checkout behavior, and autocrlf > core.eol