NVPTX does not support Julia's address spaces

[juliagpu]

NVPTX does not support Julia's address spaces

[maleadt]

Yes, I know, CUDAnative is currently broken on master after JuliaLang/julia#21888. I haven't had the time to fix, please use release-0.6 or v0.6.0 for the time being.

[colbec]

Now that the final binary has been released for Julia 0.6 I have switched to this for my main work. Since I did not compile from source, is there a way to match up CUDA tools with the binary version of 0.6? Currently the installation calls for the custom compile approach.

[maleadt]

The source build is required for LLVM.jl, which builds a custom libLLVM_extra.so. The problem with binary dists is that 1) it doesn't ship the necessary build artifacts (llvm-config, LLVM headers, etc), and 2) there's going to be a compiler mismatch (the one used to build julia's binary LLVM, and the one you're using to build libLLVM_extra). The former you could "fix" by building libLLVM_extra once and caching that .so, the second issue you cannot fix. It might work, but there might be incompatibilities.

[maleadt]

Triaged: after JuliaLang/julia#21888, Julia emits address spaces (on function parameters, loads, addrspacecasts and in the future possibly alloca's) on points in non-generic address spaces. NVPTX's ISEL doesn't support these, and we need to strip them from the IR. LLVM.jl isn't powerful enough for this, as the C API doesn't export any of Utils/CloneFunction, so we'd need to do this as part of base (ref JuliaLang/julia#22414).