CODEWITHSUNDEEP
compiler-constructionllvmjuliacrystal-lang
Abhinav Sharma
Abhinav Sharma

Reputation: 137

How difficult would it be to add a generic code-gen backend to Julia?

How difficult would it be to add a generic code-gen backend to Julia in the vein of go and chez-schemei.e. without the LLVM.

Would LLVM be a better way to do it - given that rust and crystal are also almost entirely self hosted and they can leverage the LLVM to emit binaries, even cross compile.

For a background, I've asked a question specifically about code-gen in case of various LLVM front-ends here What are the issues faced while implementing a self-hosted language on LLVM?

Upvotes: 1

Views: 284

Answers (2)

Isaiah Norton
Isaiah Norton

Reputation: 4366

rustc links and uses LLVM, both via the LLVM C API and via an extended wrapper in C++, exposed as a C ABI. I don't know about crystal, but some languages do target textual or bytecode LLVM IR directly, which has some potential advantages, but also performance and compatibility issues.

As far as "how difficult": NaN, I guess? The most straightforward approach would be to start with Julia's type-inferred lowered form, which is already close to SSA. Someone with a good compiler background could probably write an un-optimized (spilling, no code motion, etc.), one-architecture template assembler in a short time (~months). Someone with no compiler background would probably need a fair amount of study to properly formulate the question.

Upvotes: 4

Chris Rackauckas
Chris Rackauckas

Reputation: 19162

Julia does have a form of code-gen which is its typed AST. You can see the typed AST using @code_typed in front of any function call. Theoretically you can use that to emit to any IR, and Julia chooses LLVM IR for reasons stated elsewhere. Transpiler.jl is a package which does this to output things like OpenCL code from Julia functions.

You can use the LLVM IR to emit other bytecodes. CUDANative.jl uses LLVM's .ptx backend to emit CUDA kernels directly from Julia functions. The @polly project is looking to do similar things whereby the macro would allow Julia to auto-accelerate some codes on the GPU (I admittedly have little knowledge of this, other than reading the post which suggested it and the followup. This was picked up as a GSoC project for LLVM).

Upvotes: 6

Related Questions