Thread level speculation (TLS) allows compiler to create speculative
threads for a sequential program without having to prove independence
between threads.

However, the performance of TLS may suffer from frequent inter-thread
data dependences. Previous study shows that it is beneficial to schedule
instructions for scalar variable communication. We extend that
scheduling algorithm to optimize non-scalar variable communication.
Intra-thread data/control speculations with recovery support are also
used to further increase the overlap between threads.

Even though such aggressive scheduling techniques are used,
some parallelized loops may still suffer performance loss due to
their sequential nature. It is important to identify those loops
before we apply compiler optimizations. We develop a loop selection
method that can automatically select a set of loop candidates
suitable for parallelization under TLS. It includes three
key parts: context-sensitive loop profiling, TLP estimation
and optimal loop set selection.