Compiler Optimizations for Limited Memory Embedded Systems

嵌入式系统被广泛应用于涉及普适计算的各种应用中。由于外形尺寸小(如手持设备)和其他应用需求，嵌入式系统的特点是计算能力和内存等资源有限。此外，应用需要满足实时性能要求和低功耗的要求。由于需要映射的现代应用领域的复杂性，C、C或嵌入式Java等高级语言正在取代汇编语言在嵌入式系统编程中的作用。然而，由于高级语言和底层嵌入式指令集体系结构之间的语义差距，这可能会导致代码大小和性能损失。这就是编译器优化作用变得至关重要的地方。这些优化是必须的(而不是可选的)，因为嵌入式系统上的可用资源非常有限，除非对应用程序进行优化，否则它们可能甚至不适合可用的资源(如内存)。该方案的重点是为嵌入式系统设计编译器优化，以最大限度地利用有限的资源，如可用的片上和片外内存。考虑到有限的片上存储器和寻址方式的限制，我们首先提出了针对嵌入式软件的数据和代码段的编译器优化的大小和速度。由于这些处理器上的本地资源非常有限，而且嵌入式设备通常以传感器、执行器等形式形成更大系统的一部分，因此它们通常是联网的。在这样的联网环境中，数据和代码被及时地从服务器下载到嵌入式设备上，然后执行。因此，一个非常重要的问题是如何实时动态优化下载到嵌入式设备上的代码。由于当地资源非常稀缺，这种动态优化被证明非常有用。这是我们在提案中讨论的第二组主题。我们预计，这些优化不仅可以提高映射到这些系统上的现有应用程序的性能，还可以通过满足它们的资源限制来在它们上映射新的应用程序域。

Embedded systems are widely used in a variety of applications which involve ubiquitous computing. Due to small form factors (such as in hand-held devices) and other application needs, embedded systems are characterized by limited resources such as the computing power and memory. In addition the application needs demand real time performance requirements and low power consumption. High-level languages such as C, C++ or Embedded Java are replacing assembly language in programming embedded systems due to the complexity of modern application domains that need to be mapped. This can however incur a code size and performance penalty due to the semantic gap between the high level languages and the underlying embedded instruction set architecture. This is where the role of compiler optimizations becomes critical.These optimizations are a must (and not optional) since available resources on embedded systems are quite limited and unless applications are optimized they may not even fit the available resources (such as the memory). The focus of this proposal is to devise compiler optimizations for embedded systems to maximally use the limited resources such as the available amount of on-chip and off-chip memory. We first propose compiler optimizations for data and code segments of embedded software for size and speed given the constraints on limited on-chip memory and addressing modes. Since local resources are very limited on these processors and since the embedded devices typically form a part of bigger systems in the form of sensors, actuators etc., they are often networked. In such a networked environment, data and code are down-loaded on an embedded device just-in-time from a server and then executed. A very important issue is then how to dynamically optimize codes that are down-loaded on to an embedded device just-in-time. Since local resources are very scarce, such dynamic optimizations prove immensely useful. This is the second set of topics we address in the proposal. We anticipate that these optimizations could not only improve performance of existing applications mapped on these systems but could also allow mapping new application domains on them by satisfying their resource constraints.