It is nice to try to use the existing branch prediction hw to
accelerate profiling. It is an edge profiling tech i think. The experiment
provided by this papers is good. But can we do sth for path profiling?
The alternation for HPL PlayDoh put restriction on the predicate register
usage. But if the RF is big enough, there is not big problems.
Compiler indexing should have less contentions than that of selective
indexing. In the experiment, except for benchmark li and eqntott, it
is this
case. The interesting thing is that the performance of comipler indexing
of li is still better than that of selective indexing. I think this
is
because performance is not decided soly by contention. Some kind of
correlation may works.
Shaojie