Historically, fast JIT compilation was a black art requiring assembly knowledge. Production databases rely on LLVM or C/C++ code generation, both suffering from high compile times. AI now makes it easier to write JIT compilers with fast runtime performance.

When building pgrust, the author initially thought implementing a JIT would be difficult but found it much easier with AI assistance. This JIT compiler becomes a key reason pgrust is so fast, compiling code in around 5μs and enabling JIT compilation of every SQL query. To demonstrate, the article walks through building a simple regular expression engine using JIT compilation.

The example supports only literal strings and repetition (the regex `*`). The regex is represented as parsed Rust structures, allowing strings like `applesb(an)*` but not alternation or lookbehind. The engine uses three Node types: literal string, concatenation, and repetition.

An interpreter matches nodes against input bytes. For performance comparison, the JIT-compiled version is tested against handwritten code specific to the regex `b(an)*`. The handwritten implementation checks for the literal `b`, then loops through `a` characters and checks for a trailing `n`.

While the handwritten code is specialized, the JIT approach offers flexibility across diverse queries with minimal overhead, compiling in microseconds.