qbe-rs VS Som

You can try buildling you own backend like llvm. A good example or starting point is probably QBE since it is extremely small but very functional.

How big is the whole backend? I've heard that it is small but I wanted to compare it to QBE which is around 8 KLoC and it is quite interesting too.

I think QBE might be what you're looking for?

https://c9x.me/compile/

As far as I know, there is QBE, which is actually kinda underrated, and Cranelift, mainly designed for JIT compilation

Before that, I had spent a bit of time working with QBE, which is much simpler and really easy to write a frontend for. I switched to libgccjit though, because I got frustrated with a few of the things lacking from QBE (like the ability to easily keep track of where different variables live on the stack). I think for many hobby language projects, QBE would be a good option (my project was off the ground very fast using QBE, and I got pretty far before I ran into limitations I couldn't easily work around).

If one of your parameters is size/complexity of the backend and you prefer something smaller, have a look at qbe and cwerg

The alternatives are generally hidden inside of another compiler. The big exception seems to be qbe (https://c9x.me/compile/) however since the author appears to have written this code without peer review, it's not easy to read it's source code.

So my question is this: if a backend like QBE (~12k Loc) was added to Zig and Zig only had to compile Zig code (no C, etc) for that QBE backend -- about how many LoC would that Zig need to be?

GCC is written in C++ these days, so something like QBE(https://c9x.me/compile/) would be needed.

- https://github.com/rochus-keller/Smalltalk/ Parser, code model, interpreter and navigable browser for the original Xerox Smalltalk-80 v2 sources and virtual image file

- https://github.com/rochus-keller/Som/ Parser, code model, navigable browser and VM for the SOM Smalltalk dialect

- https://github.com/rochus-keller/Simula A Simula 67 parser written in C++ and Qt

> do you regret those endeavours?

No, not in any way; the projects were very entertaining and gave me interesting insights.

If your DSL is statically typed then I recommend that you have a look at the Mono CLR; it's compatible with the ECMA-335 standard and the IR (CIL) is well documented, even with secondary literature.

If your DSL is dynamically typed I recommend LuaJIT; the bytecode is lean and documented (not as good as CIL though). LuaJIT also works well with statically typed languages, but Mono is faster in the latter case. Even if it was originally built for Lua any compiler can generate LuaJIT bytecode.

Both approaches are lean (Mono about 8 MB, LuaJIT about 1 MB), general purpose, available on many platforms and work well (see e.g. https://github.com/rochus-keller/Oberon/ and https://github.com/rochus-keller/Som/).

At the latest when you run a benchmark suite like Are-we-fast-yet; here are some measurment results: http://software.rochus-keller.info/are-we-fast-yet_crystal_lua_node_som_pharo_i386_results_2020-12-29.pdf. See also https://github.com/rochus-keller/Som/ and https://github.com/rochus-keller/Smalltalk.

LuaJIT is well suited as a backend/runtime environment for custom languages; I did it several times (see e.g. https://github.com/rochus-keller/Smalltalk, https://github.com/rochus-keller/Som/, https://github.com/rochus-keller/Oberon/). I also implemented a bit of infrastructure to ease the reuse: https://github.com/rochus-keller/LjTools. LuaJIT has some limitations though; if you require closures you have to know that the corresponding LuaJIT FNEW bytecode is not yet supported by the JIT, i.e. switches to the interpreter; as a work-around I implemented my own closures; LuaJIT also doesn't support multi-threading, but co-routines; and there is no debugger, and the infrastructure to implement one has limitations (i.e. performance is low when running to breakpoints). For most of my projects this was no issue. Recently I switched to CIL/Mono for my Oberon+ implementation which was a good move. But still I consider LuaJIT a good choice if you can cope with the mentioned limitations. The major advantage of LuaJIT is the small footprint and impressive performance for dynamic languages.

One option is to leverage someone else's JIT: you could, for example, rewrite the interpreter to transpile to Lua source, which is then run in LuaJIT. There's a Smalltalk dialect which does this successfully; the Lua version runs in 1/12th the time of the C interpreted version. https://github.com/rochus-keller/Som You can use LuaJIT's FFI to call back into the Stunt server, or else just rewrite it completely in Lua --- large parts of the Stunt server will just go away in a native Lua implementation (e.g. the object database is just a table). Javascript would be another candidate for this.

This is a completely different kind of measurement; unfortunately this is not clear enough from my Readme. I wanted to find out, how well my naive Bluebook interpreter performs on LuaJIT (using my virtual meta tracing approach) compared to Cog, which is a dedicatd Smalltalk VM optimized with whatever genious approaches over two decades (or even longer considering the long experience record by Elliot). This experiment continues in https://github.com/rochus-keller/Som, because I didn't want to modify the original Smalltalk image. I found that my naive LuaJIT based approach is about factor seven behind the highly optimized Cog/Spur, and further improvements would require similar optimization tricks as in the latter.