The Compilation Game: Static Analysis for AI, Code Generation for Climate Modeling, Hardware Design, and Compiler Verification

Abstract: Despite immense innovation pressure in the industry, we are held back by the slow evolution of our CPU-focused stand-alone compilation toolchains. Building a new domain-specific compiler, writing a new verification tool, optimizing an application, designing a microprocessor, or verifying some of its components takes years. While the underlying problems are inherently complex, our inability to broadly exploit synergies across seemingly disjoint communities slows us down. 

 

Deep learning, battery electric vehicles, and rocket launches have seen orders-of-magnitude improvements over the last ten years, but compiler development is still slow. We must radically change the compiler development process: break it into pieces, scale the communities involved, use verification to enable scalability, and aggressively pursue automation across the stack. Open source can serve as a platform for this change, and our research in the context of the LLVM/MLIR community takes the first steps in this direction. I show how the number of compiler abstractions exploded recently, offer insight into the new hardware design stack CIRCT, and share our most recent efforts in high-performance computing and interactive theorem proving.

 

Finally, I will make a case of how these seemingly unrelated topics seed “the Compilation Game” and potentially allow compilers to scale across communities.