GnuWin32 diff3 per_merger test failures

[Martin Packman]

When using the GnuWin32 diff3 binary, but not when using /bin/diff3 from within cygwin, the following three tests fail with unexpected carriage returns at the end of each line.:

bt.per_merger.TestMergeImplementation.test_merge_move_and_change(diff3)
bt.per_merger.TestMergeImplementation.test_merge_specific_file(diff3)
bt.per_merger.TestMergeImplementation.test_modify_conflicts_with_delete(diff3)

Unlike diff, diff3 doesn't seem to have a --binary switch, so no obvious line endings fix to try there.

I expect this could cause real merge problems, but wasn't able to cook up an example.

[Robert Collins (lifeless)]

I wonder if we could set the fd's to binary using native windows apis before the child process gets going ?

[Gordon Tyler (doxxx)]

There is a way to do it using the msvcrt module, which I used when fixing a bug a few weeks back in the way selftest --parallel=subprocess was handled on win32.

if sys.platform == "win32":
    import msvcrt
    msvcrt.setmode(sys.stdout.fileno(), os.O_BINARY)

[Robert Collins (lifeless)]

Well, that sets it for *our* side, I mean in the spawned process side.
Might be a tad trickier there.

[Alexander Belchenko (bialix)]

That's why I've got those test failures on UDS!

I think we should create Feature and probe on small example that diff3 produce LF-output. And skip those tests otherwise.