Confirming the microoptimization here with a non-debug (opt) build:

~/oss/cpython$ sre-before/python -m timeit -s "import re; s = (b'~' * 1000000); f = re.compile(b'\s\w\d').search" "f(s)"
20 loops, best of 5: 11 msec per loop
~/oss/cpython$ sre-after/python -m timeit -s "import re; s = (b'~' * 1000000); f = re.compile(b'\s\w\d').search" "f(s)"
20 loops, best of 5: 10.1 msec per loop

Those results are consistent across reruns.