Attached patch (issue1569291.diff) reimplements the optimization by
following Objects/stringobject.c logic as Raymond suggested.

I see two remaining issues which in my view should be addressed separately:

1. There is no check for overflow after the multiplication computing the
size of the resulting array.  Note that while newarrayobject checks that
size*itemsize does not overflow internally, there is no such check for
Py_SIZE(a) * n. A decision needs to be made whether Overflow or NoMemory
error should be raised because currently string_repeat and
newarrayobject treat overflow differently.

2. See msg64429 above.  I think both string and (patched) array codes
can be simplified.