🚀 go-pugleaf

In article <1989Aug6.164210.11976@jarvis.csri.toronto.edu> jonah@db.toronto.edu (Jeffrey Lee) writes:

>... Having conditional delay slots for conditional branches takes
>some of the sting out of this. ...
>The delay action can be made ``conditional'' by adding silicon to
>nullify the action (e.g. prevent writeback of the results to the
>register cells or abort a load/store instruction) in the case of the
>branch not taken (or vice versa). ...
>Which choice to implement depends on the expected probability of
>branch taken versus not taken.  [Sorry, I have no data on this.] ...

The paper "Reducing the Cost of Branches" by McFarling and Hennessy (13th
Comp. Arch. Conf. pp 396-403) addresses this topic.  Data is presented that
shows the branch-taken probability to be roughly 65%.  The paper shows that
the nullify action, called "squashing", can significantly reduce the cost of
branches for a profiled program.  ("Cost" is a performance merit that
includes the performance cost of NOPs).  A following paper "Architectural
Tradeoffs in the Design of the MIPS-X" by Chow and Horowitz (14th Comp.
Arch. Conf. pp 300-308) shows that squashing melds nicely with exception
handling (i.e., the amount of added silicon is small).  They also report that
the MIPS-X, which has two branch delay slots, uses from 16-18% NOPs.

Kent Wilken <wilken@husky.ece.cmu.edu>