LLVM generates max.NaN which only works on sm_80

[maleadt]

Error in testset core/device/intrinsics/math:
Error During Test at /home/tim/Julia/pkg/CUDA/test/core/device/intrinsics/math.jl:157
  Test threw exception
  Expression: isequal(maximum(cu([NaN])), NaN)
  Failed to compile PTX code (ptxas exited with code 255)
  Invocation arguments: --generate-line-info --verbose --gpu-name sm_72 --output-file /tmp/jl_Ij4FEZ6fAJ.cubin /tmp/jl_8zWhmxsv07.ptx
  ptxas /tmp/jl_8zWhmxsv07.ptx, line 274; error   : Modifier '.NaN' requires .target sm_80 or higher

This was on Tegra T194 (Xavier AGX), sm_72. However, I can't reproduce on a GTX 1080 (sm_61), so something more subtle is happening here.

[maleadt]

I guess this is llvm/llvm-project#64606.