[Clang] Initial support for linking offloading code in tool

[jhuber6]

Summary

This patch adds the initial support for linking NVPTX offloading code
using the clang-linker-wrapper tool. This uses the extracted device
files and runs nvlink on them. Currently this is then passed to the
existing toolchain for creating linkable OpenMP offloading programs
using clang-offload-wrapper and compiling it manually using llc.
More work is required to support LTO, Bitcode linking, AMDGPU, and x86
offloading.

Depends on D116545

[jhuber6]

Small changes.

[jhuber6]

Need to tell llc to make position independent code so we can link it properly.

[jdoerfert]

Same as the other commit, where is this tested?

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

414

Unsure why we look at the linker bin dir first, PATH seems to be the right choice here.

428

If this is not addressed in a follow up, it seems trivial to add it.

460

keep exit success, makes it obvious this worked.

486

.str() above to get a key is somewhat ok, but then parsing the key again is weird.
Can we make it a densemap and provide a densemapinfo for struct DeviceFile. It can just inherit from the DenseMapInfo<std::string>. Might even work w/o explicitly calling the "super" functions with the .str() version if we have a std::string operator. Might then even work w/o DenseMapInfo by telling the DenseMap to use the std::string specialization of DenseMapInfo in the first place.

614

Try to replace auto where the type is short and helpful. E.g., here it would help to know what it is.

630

do we really want to return here?

[jhuber6]

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

414

Copied it from somewhere, should change it. Right now it's just a weird way to look through /bin/

428

It is addressed later.

460

Will do.

486

Probably a smarter idea, I needed a way to group all files with the same device file type. I'll look into it.

630

Might be a good idea to keep trying to remove temp files even if one fails, will address.

[jhuber6]

Maxing suggested changes.

[jdoerfert]

LG, as part of the patch set, and given the runtime test for coverage.

Some nits are left for later though.