From a compiler perspective, Rusoto takes a long time to compile. Some of that is just the sheer volume of code, but also, it wouldn't surprise me if it's hitting bottlenecks in the compiler (e.g. a recently fixed issue in the performance of #[derive(...)] on a huge number of structs). Getting appropriate bits of Rusoto put into the compiler's performance testsuite could help, and profiling the compiler on it would likely turn up some areas for improvement.

I would also love to see support for using Rusoto on non-Tokio runtimes. Given the amount of code generation used for Rusoto, abstracting over the remainder (such as the choice of async traits) based on a feature flag would hopefully be feasible.