Abstract?? Due to the rise of Chip multiprocessors (CMP??s) the amount of parallel computing power has increased signiﬁcantly. This is in contrast with the fact that a lot of programs are sequential and cannot exploit these parallel resources, urging the need of developing new techniques to extract parallelism from sequential programs. For this we present a new proﬁle based technique. It works in a non-speculative way, based on data dependencies between functions and ﬁnds large chunks of code to parallelize. To achieve this, we introduce the so called interprocedural data ﬂow graph and the data sharing graph. To test our technique we used the bzip2 program from the SPECcpu2000 benchmark suite. Our mechanism could signiﬁcantly speed up the compression part (3.74 times), with a global speedup of 2.45 on a quad processor system.