You can easily beat a good compiler for many kinds of code. It just takes time, effort and a good deal of knowledge. If you're doing it outside of the hottest kernels of your code, you're probably wasting your time.
I emailed the code to the guy who originally wrote it and his response was brief, "The compiler wasn't that smart before". But yet the code sat there for like 10 years, untouched -- largely because people don't like to touch x86 assembly code that is working.
The moral, you may get a benefit today, but you don't get any future benefits of the investment at MS, Intel, the Gcc folks, etc..., who are improving the compilers. Sometimes it is the right call to hand optimize. But when you do, I suggest having a C/C++ version alongside of it that one can test aganist it whenever the compiler revs.
I was involved in writing (Fortran) code for Crays in the 90s. For the T3D, we speeded up the code with the usual set of tricks -- loop unrolling, array padding and so on. With the later T3E, it turned out that not only could the compiler now do all that, but that code hand-optimised in this way performed worse.
Or in general - if you are considering using a tool or rolling your own, what do you know about your data that the tool author doesn't?