This Ph.D. thesis describes a new approach for adaptive processors using a reconfigurable fabric (embedded FPGA) to implement application-specific accelerators. A novel modular Special Instruction composition is presented along with a run-time system that exploits the provided adaptivity. The approach was simulated and prototyped using and FPGA. Comparisons with state-of-the-art appl.-specific and reconf. processors demonstrate significant improvements according the performance and efficiency.