Use GCCs profile-guided optimizations. This option specifies the the
arguments with which to call pypy-c (and in general the translated
RPython program) to gather profile data. Example for pypy-c: "-c 'from
richards import main;main(); from test import pystone;
pystone.main()'"

NOTE: be aware of what this does in JIT-enabled executables.  What it
does is instrument and later optimize the C code that happens to run in
the example you specify, ignoring any execution of the JIT-generated
assembler.  That means that you have to choose the example wisely.  If
it is something that will just generate assembler and stay there, there
is little value.  If it is something that exercises heavily library
routines that are anyway written in C, then it will optimize that.  Most
interesting would be something that causes a lot of JIT-compilation,
like running a medium-sized test suite several times in a row, in order
to optimize the warm-up in general.