It's actually a great example for how to write fast Nim code, there are a few pitfalls lurking about.
First of all you found the compiler flags, one can get very far with `-d:release` and then `-d:danger` but as you say, the algorithm is the most important. To start with, we can add an accessor template to make some of the grid math clearer, and create a `Sudoku` type (that's really just a string) You've written in a functional style, with many string copies, which is great for robustness, but I'm going for speed so the first thing I tried was replacing all the procs with `iterator():char` Here we run into a problem, how do you pass the new `square` iterator into `freeset`? We could use closure iterators but for speed we really want to use inline. the answer is to turn `freeset` into a template. with that done, the speed almost doubles. but more is possible: the recursive function `resolv` still makes copies, but with a few changes those can be done away with and we're left with everything operating on a single string. At that point, things are so fast that the speed of `find` starts to make a difference, and I have found that `strchr`, when coupled with `--passC:-march=native` will use SIMD vectorization and can be faster still. For me the difference is between about 160ms and 150ms. and here it is after all that: <https://play.nim-lang.org/#ix=3CFR> I compile with `nim c -d:danger` but you can also try adding `--passC:-march=native -d:useStrChr` for an extra 5%
