[llvm-dev] How to do target-specific linking for OpenMP

[Johannes Doerfert via llvm-dev]

Hi Kai,

any reason you don't want to do the same as clang and use the "clang 
offload bundler"?
https://clang.llvm.org/docs/ClangOffloadBundler.html

I'm hoping to move this logic into a linker plugin but so far we haven't 
started.

~ Johannes



On 3/26/21 4:02 AM, Kai Plociennik via llvm-dev wrote:
> Dear all,
>
> we are developing a Clang/LLVM 9 Toolchain + Backend together with an 
> OpenMP offloading RTL for our own hardware accelerator. Since our 
> offload entries contain only simple, self-contained binary code which 
> does not have to be linked against any libraries, we were able to get 
> everything running for a single source file with offload entries by 
> simply copying the created file.o to a.out-openmp-<architecture>-elf, 
> which is then embedded in the host binary using the linker script. At 
> runtime, we can then read the offload entries from the unmodified .o 
> file effectively.
>
> However, if we want to combine several such .o files into a single 
> a.out, we have to combine the contents of the .omp_offloading related 
> sections in the .o files to a single one doing some actual "linking", 
> or better "relocations".
>
> I tried to employ the system linker ld to this end, e.g using several 
> versions of simple linker scripts, but the problem is that ld refuses 
> to do the work with the error "Relocations in generic ELF (EM: 248)". 
> As far as I understand, the reason is that it does not understand the 
> binary format.
>
> My question is: What would be the easiest way to do the necessary 
> combining of our .o files? Is it possible to do it via llvm lld in a 
> simple way, or by using a certain way of linker scripts for ld? My 
> problem is that I do not precisely understand the interplay of the way 
> we create our .o files using our derived form of 
> MCELFObjectTargetWriter and the necessary linking process.
>
> Any help would be greatly appreciated.
>
> Sincerely,
>
> Kai Plociennik
>