Static Analysis and Just-in-Time Compilation Support for Heterogeneous Multicore Architectures

Heterogeneous architectures consisting of a multicore CPU and a general-purpose graphics processing unit (GPGPU) have become the dominant computing platform for desktops, servers and handheld embedded devices. Heterogeneous multicores are notoriously difficult to program, because sequential programming languages assume a single instruction stream and a single, uniform memory image. As sequential languages provide insufficient parallel hardware abstractions, performance and portability of software on multicore architectures is greatly hampered and programmer productivity is low. The stream programming paradigm has turned out to be an effective approach for programming multicore architectures. Stream programming languages facilitate application domains characterized by regular sequences of data, such as digital signal processing, audio, video, graphics, cryptography and networking. Major obstacles stand in the way of widespread adaption of stream programming languages: (1) current special-purpose programming languages like MIT`s Streamit do not integrate with legacy code-bases in C and C++, thereby lacking acceptance by practitioners from industries. (2) compilers and run-time systems cannot adapt stream programs to varying hardware and load conditions, which results in low performance and underutilized processing capacity. To overcome these obstacles, we will conduct the following research. Novelty 1: we will research compilation support for library-level stream programming constructs with OO languages (C++ in particular). Program slicing will determine the stream-parallel portions of a program, which then become the focus of optimizations that remove bottlenecks and increase the amount of parallelism wrt. the target hardware. We extend previous work on libraries that are compiled without optimizations. Library-level stream- programming support that is tightly integrated with the host language and the compiler is essential for performance, and for the application of the stream programming paradigm with legacy code. Novelty 2: we will research static program analysis techniques for stream programming languages. Static analysis information from stream programs will enable compiler optimizations that increase the efficiency of stream programs. E.g., static analysis results on execution time and communication overhead of actors (the major building blocks of stream programs) are a pre-requisite for load-balancing stream programs across multiple, heterogeneous cores. Novelty 3: we will invest just-in-time (JIT) compilation techniques to produce native code that is tuned to the underlying architecture. Because parallel architectures differ in important ways, i.e., in the types, numbers and configurations of the available parallel execution units, a JIT compiler can perform target-specific optimizations that are not available with static compilation. We will incorporate stream program slicing, static analyses and code generation techniques as additional passes to LLVM, which is a state-of-the-art compiler infrastructure that already provides a modular and extendible frontend, an optimizing middle-end, various backends, and a JIT. The PI has previous experience with LLVM, in particular its middle and backends.

Heterogeneous architectures consisting of a multicore CPU and a general-purpose graphics processing unit (GPGPU) have become the dominant computing platform for desktops, servers and handheld embedded devices. Heterogeneous multicores are notoriously difficult to program, because sequential programming languages assume a single instruction stream and a single, uniform memory image. As sequential languages provide insufficient parallel hardware abstractions, performance and portability of software on multicore architectures is greatly hampered and programmer productivity is low.The stream programming paradigm has turned out to be an effective approach for programming multicore architectures. Stream programming languages facilitate application domains characterized by regular sequences of data, such as digital signal processing, audio, video, graphics, cryptography and networking. Major obstacles stand in the way of widespread adaption of stream programming languages: (1) current special-purpose programming languages like MITs Streamit do not integrate with legacy code-bases in C and C++, thereby lacking acceptance by practitioners from industries. (2) compilers and run-time systems cannot adapt stream programs to varying hardware and load conditions, which results in low performance and underutilized processing capacity. To overcome these obstacles, we will conduct the following research.The following methods were studied in Project SAJitcore++:1. Compilation support for library-level stream programming constructs with OO languages (C++in particular).2. Static program analysis techniques for stream programming languages. Static analysis information from stream programs will enable compiler optimizations that increase the efficiency of stream programs.3. Just-in-time (JIT) compilation techniques to produce native code that is tuned to the underlying architecture.Stream program slicing, static analyses and code generation techniques were incorporated as additional passes to LLVM.