Interview
Compiler Optimization with Greta Yorsh
- Super optimization is predicted to remain infeasible for daily desktop compilation due to inherent costs, though it is expected to mature as a tool for compiler building and optimizing critical performance code, potentially enabling the integration of new architectures without rewriting intricate passes or reducing costs via partially evaluate steps.
- Feedback-directed optimization (FDO) is anticipated to stabilize performance and increase predictability for developers, even at the expense of slower execution in some cases, while highlighting the need to remove static restrictions on optimizations to allow for dynamic changes or additions.
- The server market is expected to shift away from Intel's long-term monopoly toward ARM, prompting a change in compiler optimization priorities and requiring specific improvements in the OCaml ARM backend's instruction scheduling and register allocation to match GCC performance.
- Power consumption is projected to become a growing priority for Chainstream as the CPU ecosystem evolves, alongside an aggressive push for multi-core OCaml with a multi-core garbage collector expected in an imminent first release.
- Plans include increasing backend modularity to simplify maintenance of specific components, providing infrastructure for adding custom optimization passes to handle specific code patterns, and open-sourcing performance measurement tools for high-scale users and researchers to generate community feedback.
- Initiatives to facilitate more open dialogue between core developers and users involve establishing a dedicated channel for benchmark results separate from
ocaml discussand fostering a culture where mistakes are treated as learning opportunities leading to system changes, such as adding new test suites two weeks after an error.