When building opensubdiv with more aggressive optimization flags (-march=native -02) the output meshes would differ a bit from what we expected in the current automated modifier test file.

The differences in vertex position is within the 1e-6 range, which I would call is acceptable for floats.
In addition to this, all the modifier test that tests the subdiv modifier in particular pass without any modifications.
I've updated two tests in the modifier test file and script to make it pass (listed below).

Updated blend file:

I had to modify the following test categories:

1. Decimate test.

Here there was a subdiv modifier applied before the actual decimate modifier.
Because the decimate modifier creates a queue of potential vertices it can remove, it is highly sensitive to even small changes as it drastically changes in which order the vertices are decimated in.

As this test should only be testing the decimate modifier, I pre-applied the subdiv modifier in the test file.

2. RandomCubeModifier

For these tests I removed the subdiv modifier as well. As with decimate, a small change in vertex position here can lead to quite different results.