POCL backend: `KernelAbstractions.zeros` of a bits-union element type throws

[maleadt]

On the POCL backend (KernelAbstractions main), zeros and ones of a bits-union element type such as Union{Missing, Bool} fail while compiling the fill kernel. Allocating such an array works, and so do kernels that read or write it.

using KernelAbstractions
b = KernelAbstractions.POCLBackend()

KernelAbstractions.allocate(b, Union{Missing, Bool}, 3)   # works

@kernel function f!(y, x)
    i = @index(Global)
    @inbounds y[i] = ismissing(x[i])
end
x = copyto!(KernelAbstractions.allocate(b, Union{Missing, Bool}, 3), [missing, true, false])
y = KernelAbstractions.allocate(b, Bool, 3)
f!(b)(y, x; ndrange=3)                                     # works: Bool[1, 0, 0]

KernelAbstractions.zeros(b, Union{Missing, Bool}, 3)
# ERROR: ArgumentError: type does not have a definite number of fields

Stacktrace:
  [1] fieldcount(t::Any)                          @ Base runtime_internals.jl:1323
  [2] check_invocation(job::GPUCompiler.CompilerJob)  @ GPUCompiler src/validation.jl:99
  [3] compile_unhooked(...)                       @ GPUCompiler src/driver.jl:72
  ...
  [6] (::KernelAbstractions.POCL.var"#compile_to_obj##0#compile_to_obj##1")(ctx)  @ src/pocl/compiler/compilation.jl:230
  ...
 [13] clfunction(f::typeof(KernelAbstractions.POCL.POCLKernels.gpu_init_kernel), tt::...)  @ src/pocl/compiler/execution.jl:197
 ...
 [18] #zeros#2

KernelAbstractions main at d15c1df (0.10.0-dev), Julia 1.13.0, pocl_jll 7.2, x86_64 (AMD Ryzen 9 9950X). ones fails the same way, and so does Union{Missing, Int32}.

The POCL backend's zeros and ones launch init_kernel(arr, f, ::Type{T}) (src/pocl/backend.jl), which takes the element type as an argument. GPUCompiler's check_invocation calls fieldcount on that argument's type, which throws for Type{Union{Missing, Bool}}. Computing the fill value on the host and passing it (init_kernel(arr, zero(T))), or passing f alone, would avoid the type argument; check_invocation could also skip Type arguments, which are singletons.

AcceleratedKernels' test suite builds its test arrays with zeros, so on this backend it skips its bits-union tests (JuliaGPU/AcceleratedKernels.jl#133); that skip goes once this is fixed.