Today's compilers have many optimization options, and
it is difficult to understand all the details and the interac-
tions between them. Therefore, many application develop-
ers simply use the well-known compiler flags to compile all
their programs. On the other hand, some researchers focus
on customizing the optimizations for each program. The for-
mer method may hurt the performance of some programs
since non-optimal options are applied on them. The latter
approach takes care of the variations of programs but suffer
from the scalability problem because the optimization space
is often huge. This makes it less popular for the emerging
dynamic optimizing compilers.
    We attack this problem by exploiting two insights: first, it
is feasible to categorize the programs into different classes
according to their functionalities and characteristics; sec-
ond, dynamic profiling information is useful for recogniz-
ing the type of a program. With both insights, we present
a novel approach for program optimization: first, the pro-
grams are categorized into different classes and the class-
level optimization decision is made by the domain experts;
then, the class of an input program is recognized by the ma-
chine learning models and the optimal class-level optimiza-
tions are applied to obtain the best performance. We demon-
strate that this approach can be applied to a binary transla-
tor and great performance improvement can be achieved.