#899 exposes allocating throw paths on malloc-less backends

[michel2323]

GPUCompiler@v2.2.2 contains #899, which sets CC.OptimizationParams(; compilesig_invokes=false). With that, the @noinline throw1(x) = throw(DomainError(x, ...)) closure inside Base.Math.exponent now gets codegen'd into the kernel, and DomainError(x, msg) needs a heap box → gc_pool_alloc → gpu_malloc. gpu_malloc is a backend-provided runtime symbol (compile(:malloc, ...) in GPUCompiler's runtime.jl), and oneAPI.jl doesn't provide one. Under 2.2.1, the throw path never reached the IR.

I'm working on a malloc for oneAPI.jl and pinned the GPUCompiler.jl version for now. From what I understand, the latter is needed for Julia 1.14.

[christiangnrd]

See also #904, JuliaGPU/Metal.jl#928

@simeonschaub

[simeonschaub]

Sorry, I didn't realize that change would cause so much breakage! Nothing showed up on CI, so I assumed it would be fine. I'd be ok with reverting #899 for now until we out how to properly handle these throw paths. I do think we want to eventually set compilesig_invokes = false, it's just that it exposes some paths that previously didn't really get handled