overflow checking benchmarks
Henry Cejtin
henry@sourcelight.com
Fri, 3 Nov 2000 13:43:37 -0600
As far as the cost for the jo instructions, could it be branch prediction? I
would think that for these benchmarks things loop sufficiently that it
doesn't matter, but perhaps for the cases not recently seen the default
assumption is wrong. I know that on the Alpha CPU, a branch backwards is
assumed to be taken if it isn't in the branch table. If your block ordering
is such that some of the jo's are to smaller addresses, then this could be
happening.
Can that explain the slow down?