Darwin CPU "LLVM ERROR: inconsistency in registered CommandLine options"

[joelberkeley]

When I use the CPU plugin on ARM64/AArch64 Darwin, it works with XLA HLO but when I use StableHLO it fails to even start, with error

: CommandLine Error: Option 'experimental-debuginfo-iterators' registered more than once!
LLVM ERROR: inconsistency in registered CommandLine options

which apparently happens when LLVM is linked twice. StableHLO works fully on linux. I also believe it's working on Apple Metal, though I'm seeing a different error there so I'm not completely certain.

I'm building the plugin as a shared dylib. I'm also building PJRT and the HLO builder API together in another dylib. That's two dylibs in total. The PJRT/HLO dylib has the following dependencies.

        "@llvm-project//mlir:IR",
        "@stablehlo//:register",
        "@stablehlo//:stablehlo_serialization",
        "@xla//xla/client:executable_build_options",
        "@xla//xla/hlo/builder/lib:math",
        "@xla//xla/hlo/builder/lib:matrix",
        "@xla//xla/hlo/builder/lib:prng",
        "@xla//xla/hlo/builder:xla_builder",
        "@xla//xla/hlo/ir:hlo",
        "@xla//xla/hlo/translate:stablehlo",
        "@xla//xla/mlir_hlo:hlo_dialect_registration",
        "@xla//xla/pjrt/c:pjrt_c_api_helpers",
        "@xla//xla/pjrt:pjrt_executable",
        "@xla//xla/service",
        "@xla//xla:literal",

I have also discussed this on the openxla discord with @GleasonK, where he suggested removing instances of compatible_with = get_compatible_with_portable() from BUILD files, and, separately, removing #include "xla/service/llvm_ir/llvm_util.h". I didn't get either to work, but I don't know if I did it correctly.

I forgot to mention

[joelberkeley]

Depending on how complicated this is, I might be up for fixing it.

[joelberkeley]

Note I still experience this problem when I remove my dependency

        "@llvm-project//mlir:IR",

[joelberkeley]

Is there any chance I could get an estimate of whether this will be fixed soon?