Janet built with 32-bit pointers saved about 20 percent of its memory and lost 50 percent of its speed; the x32 ABI would keep the cache win without that cost, but Arch doesn't compile it in and Debian ships it disabled.
A 1989 paper for the Self language compiled one copy of each procedure per receiver type, verified the guess at runtime, and doubled dynamic language performance.
A Lean formalization of Skia lets three researchers verify four rewrite patterns correct, and the optimizer built on them makes Skia programs run 18.7% faster than the library's most modern GPU backend.