The productivity of embedded software development is limited by the high fragmentation of hardware platforms. To alleviate this problem, virtualization has become an important tool in computer science; and virtual machines are used in a number of subdisciplines ranging from operating systems to processor architecture. The processor virtualization can be used to address the portability problem. While the traditional compilation flow consists of compiling program source code into binary objects that can natively executed on a given processor, processor virtualization splits that flow in two parts: the first part consists of compiling the program source code into processor-independent bytecode representation; the second part provides an execution platform that can run this bytecode in a given processor. The second part is done by a virtual machine interpreting the bytecode or by just-in-time (JIT) compiling the bytecodes of a method at run-time in order to improve the execution performance. Many applications feature real-time system requirements. The success of real-time systems relies upon their capability of producing functionally correct results within dened timing constraints. To validate these constraints, most scheduling algorithms assume that the worstcase execution time (WCET) estimation of each task is already known. The WCET of a task is the longest time it takes when it is considered in isolation. Sophisticated techniques are used in static WCET estimation (e.g. to model caches) to achieve both safe and tight estimation. Our work aims at recombining the two domains, i.e. using the JIT compilation in realtime systems. This is an ambitious goal which requires introducing the deterministic in many non-deterministic features, e.g. bound the compilation time and the overhead caused by the dynamic management of the compiled code cache, etc. Due to the limited time of the internship, this report represents a rst attempt to such combination. To obtain the WCET of a program, we have to add the compilation time to the execution time because the two phases are now mixed. Therefore, one needs to know statically how many times in the worst case a function will be compiled. It may be seemed a simple job, but if we consider a resource constraint as the limited memory size and the advanced techniques used in JIT compilation, things will be nasty. We suppose that a function is compiled at the rst time it is used, and its compiled code is cached in limited size software cache. Our objective is to find an appropriate structure cache and replacement policy which reduce the overhead of compilation in the worst case.