We propose a new approach of performing automatic run time program
parallelization: divide a program into traces and execute these traces in
parallel.  We describe a framework that can automatically parallelize
software using our approach.  Our framework must collect, select, and
schedule traces, as well as enforce dependencies.  The framework consists
of a baseline compiler, serial and parallel versions of a program, a trace
collection system, and an optimizing and parallelizing compiler.  We also
report on our work in progress: trying to determine the minimum overhead
of dependencies, and separating the traces from the rest of the program.