Control-flow dependence is an intrinsic limiting factor for pro- gram acceleration. With the availability of instruction-level par- allel architectures, if-conversion optimization has, therefore, be- come pivotal for extracting parallelism from serial programs. While many if-conversion optimization heuristics have been proposed in the literature, most of them consider rigid criteria regardless of the underlying hardware and input programs. In this paper, we propose a novel if-conversion scheme that preforms an efficient if-conversion transformation using a machine learning technique (NEAT). This method enables if-conversion customization overall branches within a program unlike the literature that considered in- dividual branches. Our technique also provides flexibility required when compiling for heterogeneous systems. The efficacy of our approach is shown by experiments and reported results which il- lustrate that the programs can be accelerated on the same archi- tecture and without modifying the original code. Our technique applies for general purpose programming languages (e.g. C/C++) and is transparent for the programmer. We implemented our tech- nique in LLVM 3.6.1 compilation infrastructure and experimented on the kernels of SPEC-CPU2006 v1.1 benchmarks suite running on a multicore system of Intel(R) Xeon(R) 3.50GHz processors. Our findings show a performance gain up to 8.6% over the stan- dard optimized code (LLVM -O2 with if-conversion included), in- dicating the need for If-conversion compilation optimization that can adapt to the unique characteristics of every individual branch.