Abstract
As the development of hardware progresses, computers are expected to
solve increasingly complex problems. However, solving more complex
problems requires more complex software. To be able to develop these
software systems, new programming languages with new features and
higher abstraction levels are introduced. These features are designed to
ease development, but sometimes they also make the runtime behavior
unpredictable. Such features can not be used in real-time systems.
A feature that traditionally has been unpredictable is garbage collec-
tion. Moreover, even though a garbage collector frees unused memory,
almost all such methods require large amounts of additional memory. Gar-
bage collection relieves developers of the responsibility to reclaim memory
that is no longer used by the application. This is very tedious and error
prone if done manually. Since garbage collection increases productivity
and decreases programming errors, developers find it attractive, also in
the real-time domain.
This thesis presents a predictable garbage collection method, real-time
reference counting, that increases memory efficiency by about 50 % com-
pared to the most memory efficient previously presented predictable gar-
bage collector.
To increase performance, an optimization technique called object own-
ership that eliminates redundant reference count updates is presented. Ob-
ject ownership is designed for reference counters, but can also be used to
increase the performance of other incremental garbage collectors.
Finally, a static garbage collector is presented. The static garbage collec-
tor can allocate objects statically or on the runtime stack, and insert explicit
instructions to reclaim memory allocated on the heap. It makes it possi-
ble to eliminate the need for runtime garbage collection for a large class of
Java applications. The static garbage collection method can also be used
to remove costly synchronization instructions. Competing static garbage
collection methods with reasonable analysis time are restricted to stack al-
location, and thus handle a smaller class of applications.
