[Comgr] Link with TransformUtils that defines llvm::CloneFunction

[jmmartinez]

With BUILD_SHARED_LIBS, compilation would fail with:

/usr/bin/ld:
tools/comgr/CMakeFiles/amd_comgr.dir/src/comgr-compiler.cpp.o: in function
`COMGR::AMDGPUCompiler::cloneKernelsInBitcode(COMGR::DataSet*)': /home/juamarti/llvm/_llvm/amd/comgr/src/comgr-compiler.cpp:2588: undefined reference to `llvm::CloneFunction(llvm::Function*, llvm::ValueMap<llvm::Value const*, llvm::WeakTrackingVH, llvm::ValueMapConfig<llvm::Value const*, llvm::sys::SmartMutex<false> >
>&, llvm::ClonedCodeInfo*)'

[lamb-j]

In the past I tried using CloneModule for something and it was incredibly slow. Not sure if cloneFunction has the same issues, but something to keep an eye out for

[jmmartinez]

In the past I tried using CloneModule for something and it was incredibly slow. Not sure if cloneFunction has the same issues, but something to keep an eye out for

AFAIK CloneModule is implemented through CloneFunction. But I should double check.

---

It's used when translating spirv to bitcode. When we have several block-sizes specified in the action, we have to specialize the translated kernel for every size.

I wonder if this is often used with a single block-size. In that case we could avoid the clone and reuse the kernels that already exist. @steffenlarsen do you have more context about this?