How to generate a reusable static LLVM IR module?

Specific Domains GPU
1

Hello, I am a beginner in Julia programming.

I need a way to generate a static LLVM IR module for functions written in Julia and modify the LLVM IR that was generated.

I saw on CUDAnative.jl , there are some features like CodegenHooks, CodegenParams to disable some incompatible language features and the function _dump_function . How can I use the same approach by using Julia 1.7.1 or a more recent version because CUDAnative was implemented for Julia 1.O ?

In the package GPUcompiler.jl the function Irgen can help me , but I don’t know how to use it. If someone could explain to me it would be great.

If there is another way to generate a static LLVM IR module by using Julia 1.7.1 or a more recent version. Don’t hesitate to explain to me.

Thanks in advance for your help.

2

While this isn’t super begginer stuff, you can use GPUCompiler.jl together with LLVM.jl to play around with the LLVM IR

1 Like
3

Do you have any examples that could help us understand how I can use GPUCompiler.jl with LLVM.jl to play with LLVM IR ?

4

I would look at the GPUCompiler.jl tests or the entire StaticCompiler.jl codebase since it’s pretty simple and should give you a nice headstart

1 Like
5

I looked at the functions of the StaticCompiler package and I finally found in the function generate_obj
the process that I had to use. The process works very well but unfortunately as soon as in my code I call global variables there is a segmentation fault in the function relocation_table!. And I don’t know how to fix it. Do you have any idea how I can fix it ?