(1) Estimating speedup is very difficult because it depends on so many factors. Here's some of the documentation Coder provides about MEX acceleration:
http://www.mathworks.com/help/coder/ug/best-practices-for-using-mex-functions-to-accelerate-matlab-algorithms.html http://www.mathworks.com/help/coder/ug/optimize-generated-code.html
As the first link says, it is best to profile your code in MATLAB using the MATLAB profiler. This will point out expensive sections. You can then move those expensive parts into separate functions, generate code from those functions, and use the generated MEX files.
Also note, that if the runtime is dominated by built-in operations/functions like svd, lu, fft, then MATLAB Coder likely won't speed up the code since those operations are already compiled in MATLAB.
(2) You can hand write C, C++, and Fortran and compile those into MEX files as well:
(3) Could you elaborate on what you'd like to accomplish by interfacing compiled code from MATLAB? Do you want to use XCode to compile the MEX? Are you looking to use a shared library in MATLAB? Are you looking to just invoke an executable from MATLAB like you could do with the system command or ! operator?
