secrets of GMP 4.3

[Wei Dai]

I noticed that GMP 4.3.0, released 2 days ago, contains new optimized
assembly code. In mpn/x86_64/addmul_2.asm, there is this comment:

C cycles/limb
C K8,K9: 2.375
C K10: ?
C P4: ?
C P6-15: 4.45

C This code is the result of running a code generation and optimization tool
C suite written by David Harvey and Torbjorn Granlund.

The claimed speeds are impressive, and I'm quite curious what optimization
techniques were used by the code generation and optimization tool suite.
Hopefully the information will be released in the near future. It will be
interesting to see whether those techniques can be used to automatically
optimize other crypto-related algorithms.