Michael O’Boyle
Rethinking how we build compilers: synthesis and neural machine translation
Moore’s Law has been the main driver behind the extraordinary success of computer systems. However, with the technology roadmap showing a decline in transistor scaling, computer systems are increasingly heterogeneous, specialised and diverse. As it stands, software will simply not fit and current compiler technology is strugging to bridge the gap. We need to fundamentally rethink the role and design of the compiler.
This talk presents two novel approaches to this problem. The first uses program synthesis to lift programs to MLIR, an emerging infrastructure for building high-level compilers, that can effectively target modern hardware. Our apporach gives a 25x speedup in Intel platforms and can achieve a 189x speedup when using Google's TPU.
The second uses neural machine translation to compiler and decompile code. It outperfroms state of the art decompilers and shows significant improvement over chatGPT with 3 orders of magnitude less weights.
This talk presents two novel approaches to this problem. The first uses program synthesis to lift programs to MLIR, an emerging infrastructure for building high-level compilers, that can effectively target modern hardware. Our apporach gives a 25x speedup in Intel platforms and can achieve a 189x speedup when using Google's TPU.
The second uses neural machine translation to compiler and decompile code. It outperfroms state of the art decompilers and shows significant improvement over chatGPT with 3 orders of magnitude less weights.
back to overview
Watch Recording
