I would expect it to be fast on real code: PyPy is very much about optimizing real-world idiomatic code, rather than microbenchmarks.
From my reading of the sources topaz gets a tracing JIT compiler "for free" via the RPython toolchain (magically injected via
the jitdriver.jit_merge_point in the bytecode dispatch loop here: https://github.com/topazproject/topaz/blob/master/topaz/i... )