Thread-level speculation (TLS) is a promising, hardware-centric
technique for automatic parallelization. Software approaches to
TLS have been recently explored, however, bypassing the need for
specialized hardware designs, but tend to focus on source or VM-level
implementation aimed at a specific language and runtime environment.
This paper proposes a novel software TLS approach purely based on the
LLVM intermediate representation (IR), a language and architecture
independent IR supporting more than 10 source languages and target
architectures by many projects. This design allows speculation of
all LLVM programs on all architectures supported by LLVM, requiring
only minimal code modifications to language front-ends to enable
programmer-directed specification of speculation fork points. We
demonstrate our system using the llvm-gcc 2.9 C/C++ and Fortran
front-ends, and evaluate performance on a 40-core Intel Xeon machine.
Our work provides a practical system for further TLS research and
directly supports safe, user-directed parallelization of real-world
applications on multi-core processors.