In embedded hard real-time systems, tasks must complete their executions within predefined time frames. A necessary pre- condition to achieve this requirement is predictability of their temporal behaviour. Here, the main focus is on handling exceptions in such systems. When handled in a classical way, they necessarily jeopardise the ultimate requirement, temporal predictability. Hence, it is argued that exceptions must be either prevented or avoided, as far as this is possible. For the remaining non-preventable and non-avoidable catastrophic exceptions, a technique in form of syntactic means is presented allowing to handle them in a well-structured and predictable way, and as painlessly as possible. The technique is based on recovery blocks with pre- and post-conditions. Finally, a method for the estimation of the resulting temporal behaviour (worst case execution time) is described.