Section: New Results

VMAD and LLVM

The goal is to provide a set of annotations (pragmas) that the user can insert in the source code to perform low level analyses (profiling) or optimizations (dynamic parallelization for example).

We are developing a virtual machine handling advanced dynamic analyses and transformations of programs. VMAD is organized as a sequence of basic operations, where external modules associated to specific strategies are dynamically loaded when required. The program binary files handled by VMAD are previously instrumented at compile time to include necessary data, instrumentation instructions and callbacks to the virtual machine. Dynamic information, such as memory locations of launched modules, are patched at startup in the binary file. The LLVM compiler has been extended to automatically instrument programs to meet the requirements both of VMAD and of the handled/chosen analysis and transformation strategies.

VMAD uses sampling and multi-versioning to limit the instrumentations time overhead. At runtime, targeted codes are launched by successive chunks that can be either original, instrumented or optimized/parallelized versions. After each chunk execution, decisions can be taken relatively to the current optimization strategy. At this time, VMAD is handling advanced memory access profiling through linear interpolation of the addresses, dynamic dependency analysis and version selction. The last developments are focusing on speculative polyhedral parallelization.

The profiling strategy interpolating the memory addresses accessed in a loop nest has been run on some of the SPEC2006 and Pointer Intensive benchmark suites, showing a very low time overhead, in most cases. More details are available in the research reports [46] and [47] , and publications [15] and [16] .