Trace-driven compilation is a dramatically new approach to
compilation, where the unit of compilation is formed dynamically, out
of execution traces. The past decade has seen significant advances in
trace compilation, especially in binary optimizers such as Dynamo and
zPDT, and in dynamic compilers such as TraceMonkey (Javascript) and
SPUR (Javascript,) and PyPy (Python).

One promise of trace compilation is its potential to form compilation
regions better than the traditional method-based selection + inlining
approach. The process of forming traces for compilation is called
trace selection . In this talk, we study the building blocks of trace
selection algorithms. We found that seemingly reasonable trace
termination conditions used in previous selection algorithms can
terminate traces prematurely:

   1. stop recording a trace when encountering a block that is the
      head of an existing trace;
   2. stop recording a trace when encountering a block that is already
      in the trace buffer;
   3. stop recording a trace when encountering a JNI call
  
We proposed techniques to relax these conditions and demonstrated
significant performance improvement on the DaCapo benchmark using a
Java trace JIT built on top of J9/Testarossa. Our results also showed
promising signs where a trace JIT can compete with a method JIT at
certain low optimization levels.