Compositional Deep Argument Flattening at ML 2025

My paper Compositional Deep Argument Flattening was accepted for presentation at the ML Family Workshop 2025 in Singapore on October 16.

Function calls can spend unnecessary time packing values into objects only for the receiving function to unpack them again. This work lets the compiler pass components of nested objects directly in registers and combine curried arguments into more efficient calls. The transformations can be applied repeatedly alongside other optimisations, even across separately compiled units. The paper establishes key correctness properties and reports on the implementation and performance benefits in MLKit.

Read the paper (PDF) or see the workshop programme.