When I wrote curve25519-d... (08 Mar 2009)
When I wrote curve25519-donna I implemented many of the critical functions in x86-64 assembly. It was a lot of code, even using the C preprocessor! This got a good 20% boost in speed. This was clearly very important because it made donna-x86-64 faster than djb's version 
.
However, djb just pointed out that the 64-bit C implementation of donna was now as fast as my hand coded version. Turns out that GCC 4.3 greatly improved the quality of the code generation for this sort of code and now equals my hand crafted efforts! Well done to the GCC team because the C code is vastly smaller and easier to understand. Thus, the x86-64 of donna has been removed from the repo.