The idea is to position the basic blocks of a procedure in such a way that most executions of the code will fall through branches (forward branches are typically predicted to not be taken). This minimizes the cost of branch mispredictions, and keeps commonly executed code together. The latter will have the effect of making best use of the instruction cache, and minimizing cache collisions.

Such optimal positioning of basic blocks can be very effective at speeding up programs, especially when the information about branch frequencies is obtained dynamically (i.e. as the program is running).

-- MikeVanEmmerik - 01 Dec 2001