A processor architecture supports an electrical interface for coupling the processor core to one or more coprocessor extension units executing computational instructions, with a split-instruction transaction employed to provide operands and instructions to an extension unit and retrieve results from the extension unit. The generic instructions for sending an operation and data to the extension unit and/or retrieving data from the extension unit allow new computational instructions to be introduced without regeneration of the processor architecture. Support for multiple extension units and/or multiple execution pipes within each extension unit, multi-cycle execution latencies and different execution latencies between or within extension units, extension unit instruction predicates, and for handling processor core stalls and result save/restore on interrupt is included.