[maleadt]

The runtime's box_* functions get their type tag by declaring the data symbol jl_<type>_type as an LLVM function and loading through it. These are ordinary host methods. Sysimage builds compile them for the host: Julia 1.12+ compiles every concrete method, and on 1.11, native code instances leak in. Package images do the same on 1.11. That only links when the symbol is resolved by name at load time. On Windows, lld can't link it:

lld: error: undefined symbol: jl_int8_type
>>> referenced by GPUCompiler/src/runtime.jl:204
>>>               jl_600.tmp(text#0.o):(julia_box_uint8_29012)

This is why KernelAbstractions doesn't precompile a kernel launch on Julia 1.11 on Windows.

Nothing in these functions is device-specific, so instead of hiding them from the host, this writes the tag the way Julia's jl_set_typeof does. The types in jl_small_typeof (all the integers and Bool) get their small tag, a constant for a given build of Julia. Float32 and Float64 get the address of their type object, which goes through the existing relocation of Julia values in the runtime library. This is the same header decoding as materialize_box!.

Tested with the full suite on 1.12, plus the native tests on 1.10 and 1.13. The new test checks that the host IR of every box_* is free of jl_*_type references, and that the tags match Julia's. In a Windows VM on Julia 1.11, KernelAbstractions (with JuliaGPU/KernelAbstractions.jl#930) now precompiles with its kernel-launch workload enabled, and loads and runs kernels.

[maleadt]

CI failures unrelated, fixed downstream.