jl_error() and jl_errorf() unsafe to call if LLVM not fully initialized

[staticfloat]

If the runtime errors out early on in the bringup process, we can end up attempting to construct a backtrace and failing due to data structures not having been initialized yet. A historical example (which I am about to replace with jl_printf() and exit()):

julia/src/jitlayers.cpp

Line 823 in ef7498d

jl_errorf("Invalid CPU name \"%s\".", TheCPU.c_str());

Because jl_errorf unconditionally subs off to jl_throw, we end up attempting to construct a backtrace. jl_error, in contrast, contains a guarding if statement, however it too fails in this case, as we are far enough to have initialized jl_errorexception_type as a tag from the system image, but not far enough yet to have fully initialized LLVM.

We should have some way of denoting to the rest of the runtime that LLVM is initialized. Jeff recommended that we:

Maybe we should have jl_init_llvm handle filling in the libjulia-codegen function pointers, so the system just behaves as if codegen is not available before it's initialized?

[gbaraldi]

There's a couple options, the first thing is to just not call jl_error from there, but more generally the check inside jl_error or more precisely jl_vexceptionf is not safe, because that type get's filled in before we've actually finished initializing julia. Maybe instead of using sentinels like that we should use explicit variables, like a is_initialized global or something like it.