On the CI machines, compiling a broadcast over a Union{Missing,Bool} array aborts the process with an LLVM verifier error instead of throwing a GPUCompiler.InvalidIRError:
x = ROCArray(Union{Missing,Bool}[missing, true])
Array(x .| false)GPUArrays 12's testsuite runs that as a probe (test/testsuite/findall.jl:105, in try/catch, to see whether the back-end supports isbits unions), so gpuarrays/reductions/any all predicates crashes its worker on #1145:
┌ Call parameter type does not match function signature!
│ ptr addrspace(1) @jl_libjulia_internal_handle
│ ptr %ijl_alloc_string.found = call ptr @ijl_load_and_lookup(ptr inttoptr (i64 3 to ptr), ptr addrspacecast (ptr addrspace(1) @_j_str_ijl_alloc_string_4 to ptr), ptr addrspace(1) @jl_libjulia_internal_handle), !dbg !435
│ [... the same for further jl_alloc_string, jl_string_to_genericmemory and jl_genericmemory_to_string lookups ...]
│ LLVM ERROR: Broken module found, compilation aborted!
│
│ [1935] signal 6 (-6): Aborted
...
│ runJuliaPasses at LLVM.jl/deps/LLVMExtra/lib/NewPM.cpp:867
│ run_passes! at LLVM/src/passes.jl:651
│ optimize! at GPUCompiler/src/optim.jl:52
│ emit_llvm at GPUCompiler/src/driver.jl:382
│ hipcompile at AMDGPU/src/compiler/codegen.jl:379
...
│ _copyto! at GPUArrays/src/host/broadcast.jl:79
So @jl_libjulia_internal_handle ends up in addrspace(1) while the ijl_load_and_lookup calls still pass it as a generic pointer, and the module fails verification during optimization, before GPUCompiler's IR validation can report the ccalls.
Where it happens:
Where it doesn't: gfx1036 (ROCm 7.2.4, artifacts or system ROCm), Julia 1.12.7 and 1.13.1, with the same package versions (AMDGPU at #1145, GPUArrays 12.0.0, GPUCompiler 2.13.0, LLVM.jl 10.1.0). There the snippet throws InvalidIRError (also when overriding HIP.gcn_arch to gfx90a or gfx942, and under Pkg.test(; coverage=true) with the CI's environment variables), and the full test suite passes (21,399 tests). So I couldn't find what the CI machines do differently.
On the CI machines, compiling a broadcast over a
Union{Missing,Bool}array aborts the process with an LLVM verifier error instead of throwing aGPUCompiler.InvalidIRError:GPUArrays 12's testsuite runs that as a probe (
test/testsuite/findall.jl:105, intry/catch, to see whether the back-end supports isbits unions), sogpuarrays/reductions/any all predicatescrashes its worker on #1145:So
@jl_libjulia_internal_handleends up inaddrspace(1)while theijl_load_and_lookupcalls still pass it as a generic pointer, and the module fails verification during optimization, before GPUCompiler's IR validation can report theccalls.Where it happens:
amdgpu1.luraess.com(system ROCm 7.2.3,JULIA_AMDGPU_DISABLE_ARTIFACTS=1), Julia 1.13.1: https://buildkite.com/julialang/amdgpu-dot-jl/builds/4380Where it doesn't: gfx1036 (ROCm 7.2.4, artifacts or system ROCm), Julia 1.12.7 and 1.13.1, with the same package versions (AMDGPU at #1145, GPUArrays 12.0.0, GPUCompiler 2.13.0, LLVM.jl 10.1.0). There the snippet throws
InvalidIRError(also when overridingHIP.gcn_archto gfx90a or gfx942, and underPkg.test(; coverage=true)with the CI's environment variables), and the full test suite passes (21,399 tests). So I couldn't find what the CI machines do differently.