This is one way to write a compiler. One of the large tradeoffs made is extensive use of libraries on both the front and backends. It's a practical choice but a

This is one way to write a compiler. One of the large tradeoffs made is extensive use of libraries on both the front and backends. It's a practical choice but also means that the compiler itself is also likely somewhat slow. Targeting LLVM alone is a choice that will guarantee that code gen is slow (you pay for a lot of unneeded complexity in llvm with every compile). When you master the principles of parsing, it is straightforward to do in any language by hand with good performance. It is easy to write a function in any language that takes a range string, such as "_a-zA-Z" and returns a table of size 256 with 65-90, 95 and 97-122 set to 1 and the rest set to zero. You can build your parser easily on top of these tables (this is not necessarily optimal but it is more than good enough when you are starting). For the backend, you can target another language that you already know. It could be javascript, c or something else. This will be much easier than targeting llvm. Start with hello world and build up the functionality you need. In no more than 10k lines of code (and typically significantly less), you can achieve self hosting. Then you can rewrite the compiler in itself. By this point you have identified the parts of the language you like and the parts that you don't or are not carrying weight. This is the point at which maybe it makes sense to target llvm or write a custom assembly backend. The beauty of this approach is you keep the language small from the beginning without relying on a huge pile of dependencies. For a small compiler, you do not need concurrency or other fancy features because the programs it compiles are almost definitionally small. Now you have a language that is optimized for writing compilers that you understand intimately. You can use it to design a new language that has the fancy things you want like data race protection. Repeat the same process as before except this time you start with concurrent hello world and build in the data race protection from the beginning.