I guess a benchmark suite, rather than just one benchmark program, so that different styles of coding are
covered such as integer, floating point, DSP, decision logic, function-call intensive, memory intensive, etc etc
One thorny issue is separating raw metal performance from compiler optimization performance - unless you
want to rewrite your benchmark code in assembler for each architecture you are seeing the hardware through
the veil of the compiler's ability to optimize well for the particular architecture.
I guess its pretty moot what is meant by "best way" anyhow - best from the perspective of the chip manufacturer?
from the perspective of a C programmer?