I flipped through the list of supported Coder functions, and it appeared to me that except perhaps for some boundary conditions, they were routines that could be relatively readily implemented using the standard C library. As the resulting C code would not require any interpreter level, I would expect that the C code would usually be faster.
I would not expect that the C code would always be faster, however: the trade-off of using the standard C library is that one would not be calling upon highly tuned cpu-specific math libraries such as LAPACK -- routines which are usually not invoked until the problem is "sufficiently big" (due to the overhead of marshaling the data from MATLAB storage in to the form expected by the routines).
