LLVM_RUNTIME_TARGETS option

[ye-luo]

LLVM_ prefix makes it sounds like global in the whole LLVM build.
Searching the code base
https://github.com/search?q=repo%3Allvm%2Fllvm-project%20LLVM_RUNTIME_TARGETS&type=code
indicates that is an option per runtime.

If I'm building 3 runtime -DLLVM_ENABLE_RUNTIMES="a;b;c"
however, if a only supports arch1 and arch2, b only support arch3, c supports arch1 and arc3,
I'm not sure how I should use LLVM_RUNTIME_TARGETS option
and what happens if I put -DLLVM_RUNTIME_TARGETS="arch1;arch2;arch3"

#136729 mentioned additional variables to control the resolution

    -DLLVM_RUNTIME_TARGETS='default;amdgcn-amd-amdhsa;nvptx64-nvidia-cuda'     \
    -DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES=openmp \
    -DRUNTIMES_amdgcn-amd-amdhsa_LLVM_ENABLE_RUNTIMES=openmp

I found it cumbersome.

Instead, I'd like to see

    -DLLVM_ENABLE_RUNTIMES=openmp \
    -DRUNTIME_openmp_TARGETS="nvptx64-nvidia-cuda" # to select a subset if added

In short, I prefer the runtime name as the primary key rather than its supported targets in the user facing CMake options.

[jhuber6]

This is something I'd need to discuss with @petrhosek. Right now LLVM_RUNTIME_TARGETS enables the targets, which then lets you enable them individually. default is kind of a weird special case I'm hoping we could improve later. So, basically -DLLVM_RUNTIME_TARGETS=a;b;c will enable separate runtimes builds with --target=a and so forth. You pass arguments to these invocations manually with -DRUNTIMES_a_<option> so, to set the runtimes list you do -DRUNTIMES_a_LLVM_ENABLE_RUNTIMES=openmp.

The normal -DLLVM_ENABLE_RUNTIMES is kind of a legacy thing that only applies to the default target. I think we should probably improve this, but I'd need to have a longer discussion. I have #136591 which (somewhat) tries to improve that, but it's broken currently.

[jhuber6]

There also really should be some documentation on this. I should spend some time detailing the LLVM runtime build.

[ye-luo]

Requiring -DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES=openmp to turn on openmp runtime is too heavy and unnecessary burden. LLVM_ENABLE_RUNTIMES really should stay. RUNTIMES_<target>_LLVM_ENABLE_RUNTIMES should really be internally derived based on LLVM_ENABLE_RUNTIMES and supported or requested triples if possible.

I also prefer RUNTIME_<runtime>_TARGETS to RUNTIMES_<target>_LLVM_ENABLE_RUNTIMES which can be derived based on RUNTIME_<runtime>_TARGETS.

In addition, the design you mentioned still doesn't allow specifying a compiler option only for a specific runtime on a specific target.
Need something like RUNTIME_openmp_TARGET_a_<option> expanded beyond semi-global options RUNTIME_TARGET_a_<option> for all the runtimes of a give target and RUNTIME_openmp_<option> for all the targets of a given runtime.

All what I described above are user facing option choices not actual CMake handling of their consumption.

[jhuber6]

Requiring -DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES=openmp to turn on openmp runtime is too heavy and unnecessary burden. LLVM_ENABLE_RUNTIMES really should stay. RUNTIMES_<target>_LLVM_ENABLE_RUNTIMES should really be internally derived based on LLVM_ENABLE_RUNTIMES and supported or requested triples if possible.

Yeah, it's definitely a little annoying to type out, but this is why we suggest people use CMake cache files if they want a 'default' experience.

I also prefer RUNTIME_<runtime>_TARGETS to RUNTIMES_<target>_LLVM_ENABLE_RUNTIMES which can be derived based on RUNTIME_<runtime>_TARGETS.

In addition, the design you mentioned still doesn't allow specifying a compiler option only for a specific runtime on a specific target. Need something like RUNTIME_openmp_TARGET_a_<option> expanded beyond semi-global options RUNTIME_TARGET_a_<option> for all the runtimes of a give target and RUNTIME_openmp_<option> for all the targets of a given runtime.

There's -DRUNTIMES_<triple>_CMAKE_ARGS="-DFOO;-DBAR" and -DRUNTIMES_<triple>_CMAKE_BUILD_TYPE=Debug for example.

All what I described above are user facing option choices not actual CMake handling of their consumption.

I think the biggest issue is the weirdness between -DLLVM_ENABLE_RUNTIMES and -DRUNTIMES_x86_64-unknown-linux-gnu_LLVM_ENABLE_RUNTIMES for example. They can mean the same thing, but users will find it confusing.