Validation uses wrong interpreter

[koenvos]

Describe the bug

Some math operations in CUDA/src/device/intrinsics/math.jl give compile errors.

It seems that operations that replace ones in Base and SpecialFunctions work fine.
Those that do not have a definition elsewhere give a "KernelError: kernel returns a value of type Union{}."
Those that redefine operations in FastMath give "invalid LLVM IR Reason: unsupported dynamic function invocation."

To reproduce

The Minimal Working Example (MWE) for this bug:

using CUDA
a = CUDA.ones(10)
function kernel(a)
    i = threadIdx().x
    a[i] = CUDA.rsqrt(a[i])
    return nothing
end
@cuda threads=length(a) kernel(a)
--> ERROR: LoadError: GPU compilation of kernel kernel(CuDeviceVector{Float32, 1}) failed KernelError: kernel 
returns a value of type `Union{}`

Manifest.toml

CUDA v3.0.0
GPUArrays v6.2.2
GPUCompiler v0.11.2
LLVM v3.6.0

Version info

Details on Julia:
Julia Version 1.6.0
Commit f9720dc2eb (2021-03-24 12:55 UTC)
Platform Info:
OS: Windows (x86_64-w64-mingw32)
CPU: Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-11.0.1 (ORCJIT, skylake)
Environment:
JULIA_EDITOR = "C:\Program Files\Microsoft VS Code\Code.exe"
JULIA_NUM_THREADS = 6

Details on CUDA:
CUDA toolkit 11.1.1, artifact installation
CUDA driver 11.1.0
NVIDIA driver 456.71.0

Libraries:

• CUBLAS: 11.3.0
• CURAND: 10.2.2
• CUFFT: 10.3.0
• CUSOLVER: 11.0.1
• CUSPARSE: 11.3.0
• CUPTI: 14.0.0
• NVML: 11.0.0+456.71
• CUDNN: 8.10.0 (for CUDA 11.2.0)
• CUTENSOR: 1.2.2 (for CUDA 11.1.0)

Toolchain:

• Julia: 1.6.0
• LLVM: 11.0.1
• PTX ISA support: 3.2, 4.0, 4.1, 4.2, 4.3, 5.0, 6.0, 6.1, 6.3, 6.4, 6.5, 7.0
• Device support: sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_62, sm_70, sm_72, sm_75, sm_80

1 device:
0: GeForce RTX 2070 (sm_75, 5.286 GiB / 8.000 GiB available)