[souprock]

The most awesome was SAL, a vector math library by Mercury Systems.

Every function came in multiple varieties, specialized for the expected and desired cache hotness of the data. Suppose you wanted to multiply a couple matrices. There are two inputs and an output. The inputs could be cached, uncached, or one of each. The output might be needed soon, and thus should remain in cache, or it might be better to save the cache for other things. All of this could be specified.

That library could help a programmer to produce mighty fast programs.