My experience of similar tests is that ifort often produces slower performance with the Fortran-77 old-style assumed-size (*) method, less so with Gfortran.
I would also like to recommend that SYSTEM_CLOCK is preferable / more accurate than CPU_TIME, especially for tests that take less than 0.015 seconds.
Based on this thread, I will now try tests of the contiguous attribute, especially to identify the array types where this may help.