Avoid .text.startup

[melonedo]

Another text section will trigger assertion failure, see https://github.com/JuliaLang/julia/blob/c609e9a22a17bef858559ca707cb985d770935aa/src/cgmemmgr.cpp#L861

[Gnimuc]

Could you elaborate a bit more about how .text.startup triggers the assertion? To me, it seems that somehow Julia fails to correctly call finalizeMemory. (EDIT: it turns out these methods are interfaces of LLVM's memory manager and Julia just implements a customized version to workaround some multi-threading issues related to libunwind)

[melonedo]

If you check what sections are generated by inspecting ELFWriter::writeObject, you will find that two text sections (.text and .text.startup) are emitted. The reason behind .text.startup is that constructors are grouped in a separate section to improve space locality.

Note: you will need to trace MCAsmBackend::createObjectWriter on other platforms to determine the writer.

[Gnimuc]

Thanks for the link. That explains why/how the .text.startup section is emitted. But I still can't find any clue on why Julia uses code_allocated and forces a single code section when implementing the memory manager.

[Gnimuc]

I think I found the reason. Just several lines above that line, there is a comment:

// allocating more than one code section can confuse libunwind.

[Gnimuc]

There is probably another way to fix the libunwind issue. But if we compromise on Julia's implementation, do we need to ignore other sections? like:

  for (auto &GF : jl_Module) {
     if (GF.getSection() != ".text") {
       GF.setSection("");
     }
   }

[melonedo]

That is a valid choice, I just want to be conservative here. I think we can delay this until more tests are passed.