This is a bool, no need for != 0.

As far as I can tell, this does the same thing as before.

1. Calculate loop normals, assign to CD_NORMAL layer.
2. Use data from loop normals to split some edges. Do not modify the new loop normals
3. At the end, free the loop normals if the argument is true

I don't think code relies on the normals being calculated at first though, there are other functions to do that already.
Part of the benefit of the change is that the split_faces function doesn't also need to calculate loop normals.

Or maybe we need to add calc_normals_split() in Cycles before this is called, is that what you're referring to?

I do not understand your answer here. You describe old code behavior, not new one, afaics:

OLD CODE (in BKE_mesh_split_faces):

1. Compute split normals (aka generate CD_NORMAL temp ldata layer).
2. Split edges.
3. if free_loop_normals is true, clear temp CD_NORMAL layer from ldata.

NEW CODE:

1. If free_loop_normal is set, clear CD_NORMAL layer from ldata, else create/update it (call to BKE_mesh_calc_normals_split).
2. Split edges.

No clearing of CD_NORMAL ldata here at the end

...

The end result is the same, because BKE_mesh_split_faces didn't change the face corner normals.
That's true in the new code too. Splitting sharp edges does not change face corner normals.
Whether we do it before or after splitting edges makes no difference.

Am I really getting that wrong? It seems straightforward to me right now.