Are there any courses on creating a new language, interpreter, and compiler from scratch? Like not using LLVM or similar to generate intermediate representation but to literally pick an ISA and, well, compile for it.
You can find a PDF and the associated course with a very minimal amount of following links.
If you are thinking of making your own language, it's also good to learn something about programming language theory, if you don't already. Many languages make mistakes that have been solved 25+ years ago. PLAI is good for that: https://www.plai.org/
I used what I learned from GCC output Then I wrote a basic amd64 expression compiler with simple live ranges and linear register allocation and ANF (a normal form) for postorder AST traversal for code generation. It can only evaluate nested add and mul expressions.
https://github.com/samsquire/compiler or on replit: https://replit.com/@Chronological/Compiler3#main.py
Finding opcodes is a struggle but there is this website: https://www.felixcloutier.com/x86/ and https://www.cs.uaf.edu/2016/fall/cs301/lecture/09_28_machine... (the table at the bottom is really useful)
To learn code generation not targeting amd64, I wrote an imaginary assembly language and then wrote a switch based virtual machine and a compiler for it with a frontend that looks similar to javascript.
https://github.com/samsquire/multiversion-concurrency-contro... in ProgramParser.java, LanguageInterpreter.java, and LanguageInterpreterRunner.java for where main() is.
Author has made complete draft PDFs for both the Racket and Python versions available here. https://wphomes.soic.indiana.edu/jsiek/
I tend to be biased towards what Jeremy Siek himself markets as "Proven in the classroom" when it comes to book authors in CS. Many book authors lack this experience and simply write for themselves, which is ok, but can result in bad didactics. Good teachers and authors from academia are invaluable.
I just wish textbooks didn't have their own bad tendencies: they have pablum as an attractor, because on average students just want to get through the class, not doing too much worse than average among the other students. Even without this problem, there's a more basic one: like with enterprise software, the decision to buy the book is not typically up to the user. "Will people actually want to read this on their own time?" is a strong driver of quality, even though it has pitfalls too.
Simple machine code generation with basic register allocation and evaluation which eagerly spills to the stack is not super hard. Producing fast code instead is a bottomless well.