Feel free to comment if you think that something can be done better.

See the latest update. The Python 3.4 compatibility is kinda painful but we can keep it for now.

One test that I think we need to have:

1. Compile a "deduplicable" module with compileall with the deduplication flag for all levels
2. Change the module content
3. Compile a "deduplicable" module with compileall with the deduplication flag for some levels only
4. Assert only the proper files were changed, not all

Alternate version fo the test: Replace (3) with importing the module from Python.

One test that I think we need to have:

1. Compile a "deduplicable" module with compileall with the deduplication flag for all levels

$ echo "a = 0" > module.py
$ python -m compileall2 -o 0 -o 1 -o 2 --hardlink-dupes ./
Listing './'...
Compiling './module.py'...
$ ls -i __pycache__/
8792583 module.cpython-37.opt-1.pyc  8792583 module.cpython-37.pyc
8792583 module.cpython-37.opt-2.pyc
$ md5sum __pycache__/*
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-1.pyc
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-2.pyc
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.pyc

2. Change the module content

$ echo "b = 1" > module.py

3. Compile a "deduplicable" module with compileall with the deduplication flag for some levels only

$ python -m compileall2 -o 0 -o 2 --hardlink-dupes ./
Listing './'...
Compiling './module.py'...

4. Assert only the proper files were changed, not all

$ ls -i __pycache__/
8792583 module.cpython-37.opt-1.pyc  8792590 module.cpython-37.pyc
8792590 module.cpython-37.opt-2.pyc
$ md5sum __pycache__/*
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-1.pyc
dc87c896d4f87fba2c30da386fe701f3  __pycache__/module.cpython-37.opt-2.pyc
dc87c896d4f87fba2c30da386fe701f3  __pycache__/module.cpython-37.pyc

It seems that it works. I'll create a test.

I'd check that only compiled files in step 3 have different inode and checksum and the file compiled in the step 1 has the same inode/checksum as before. WDYT?

I'd check that only compiled files in step 3 have different inode and checksum and the file compiled in the step 1 has the same inode/checksum as before. WDYT?

Sounds good.

I'd also test this:

$ echo "a = 0" > module.py
$ python -m compileall2 -o 0 -o 1 -o 2 --hardlink-dupes ./
Listing './'...
Compiling './module.py'...
$ ls -i __pycache__/
8792583 module.cpython-37.opt-1.pyc  8792583 module.cpython-37.pyc
8792583 module.cpython-37.opt-2.pyc
$ md5sum __pycache__/*
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-1.pyc
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-2.pyc
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.pyc
$ echo "b = 1" > module.py
$ python -c 'import module'
$ ls -i __pycache__/
8792583 module.cpython-37.opt-1.pyc  8792590 module.cpython-37.pyc
8792583 module.cpython-37.opt-2.pyc
$ md5sum __pycache__/*
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-1.pyc
3d62806ca2e42ea0b8e3a96dd7a1cbb8  __pycache__/module.cpython-37.opt-2.pyc
dc87c896d4f87fba2c30da386fe701f3  __pycache__/module.cpython-37.pyc

Even though I think we are testing this too much, the tests are implemented in the two latest commits.

There's no such thing as testing too much 😄

This is a response to some of the things mentioned in the discussion.

So, technically, all the tests that assert equal inode numbers will fail on Windows? This is not a big deal in Fedora, but will need some tweaks when proposing upstream. I suppose some of the following:

• disable the flag entirely on Windows, skip all the added tests on Windows
• document that the flag is a no-op on Windows, skip inode equality asserts on Windows
• upon linking the files, check inode numbers and when they differ, issue a warning and disable the flag globally, skip all the added tests on Windows, add a new one for Windows

The first option seems like the easiest one.

Anyway, for now, I think we can have this and test it in Fedora without worrying about Windows. WDYT?

Could you rebase this, so we see the tests results on Windows?

That's my plan. Rebase, clean, squash, skip tests on Windows…

The tests simply passed?

The tests simply passed?

It seems so :D I don't know how inodes work on Windows.

Let's merge and release this, so we can backport it into Fedora from a released version? Or would you prefer to test it in copr first?

It's more or less up to you. What do you prefer from minimization perspective? I'm ok with merge, release and use in Fedora directly.

Ok, let's get a merge and release please.