[Openmp-commits] [clang] [clang-tools-extra] [flang] [lldb] [llvm] [openmp] [pstl] Finally formalise our defacto line-ending policy (PR #86318)

[Martin Storsjö via Openmp-commits]

mstorsjo wrote:

This breaks a number of tests on Windows.

Previously, to have tests working on Windows, one would do `git config --global core.autocrlf false` or similar, before checking out llvm-project - a number of test input files _need_ to be in LF form to work. This was brought up earlier already by @llvm-beanz in https://github.com/llvm/llvm-project/pull/86318#issuecomment-2093160376.

Now after this change, due to the added `.gitattributes` which overrides the `core.autocrlf` setting, these files get checked out with CRLF newlines (as the native form for the platform).

Based on the comment in the .gitattributes file, it seems like this is both known and intentional behaviour:
```
# Checkout as native, commit as LF except in specific circumstances
* text=auto
```

While it was already stated that blanket checkouts with CRLF _will_ fail.

Additionally, the old mechanism of getting working newlines in the files is suddenly broken.

---

To make things worse, you won't notice this thing if you're updating an existing workdir - files that aren't touched aren't rewritten. (To trigger re-checkout of files to get `.gitattributes` applied, one can do something like `git rm -r subdir && git reset && git checkout subdir`.) So I guess most buildbots will keep chugging along fine, until someone pushes changes that touch those files. If running with a fresh checkout of llvm-project, this has a much bigger impact.

---

For compiler-rt tests, a handful of the profile tests depend on the right line endings - I can push a local `.gitattributes` file to fix that (see https://github.com/mstorsjo/llvm-project/commit/99bec81c87dcd2b7a7970954882bc0e42239d381).

But the Clang, clang-tools-extra and LLVM testsuites have _many_ tests that are broken - from an initial (unverified) test run, there are around 80 tests failing due to this - sorting that out is a much bigger task that I'm not volunteering to take on right now.

CC @AaronBallman as the majority of those failing tests are in Clang.

Can we revert this until we figure out these bits? Do we really want checkouts to default to having the majority of files with CRLF on Windows? I would expect that most people working on Windows don't really want this (this wasn't the case so far anyway)? And other than that, we do need to tag the files that rely on being in LF form so that things work on Windows with either setting.

https://github.com/llvm/llvm-project/pull/86318