See this interesting post on MSRC's blog:

Building Faster AMD64 Memset Routines

In a nutshell the author replaced a significant number of unpredictable branches with overlapping stores, calculating their addresses in branch-less code. The results effectively take advantage of the very good store buffer implementations and efficient handling of unaligned accesses on modern x86 processors.

I found the article particularly interesting because in my testing of other highly data-dependent leaf functions - such as memcpy() and string manipulation functions - branch mispredictions always stood out as the biggest contributor in the time spent there. I definitely prefer this approach to the jump-happy versions that are more common (like in glibc for example).