Bitslicing
An implementation technique that treats a CPU word as dozens of parallel one-bit processors and rewrites S-boxes as pure Boolean logic.
Eli Biham introduced it in 1997 for DES. Instead of encrypting one block at a time, you put bit i of 64 different blocks into the same 64-bit register, so every AND, XOR or OR processes all of them at once. S-boxes become gate networks rather than lookup tables, and bit permutations — expensive in ordinary software — become free, since they're just a matter of which register you read.
The bigger payoff turned out to be security rather than speed. No table lookups means no cache-dependent memory access, and therefore no cache-timing attack. In 2009 Emilia Käsper and Peter Schwabe published a bitsliced AES using SIMD instructions that was both fast and leak-free. The catch is that you need many blocks in flight at once, which suits CTR far better than CBC.