Optional generators swallow errors

[maleadt]

With the optional generator syntax from #23168, there is currently no way of discovering errors that occur while executing the optional generator. In the case of a @generated function (ie. generated_only), the error will at least be noticed at run-time, but with optional generators we'll just ignore any error:

julia/base/compiler/utilities.jl

Lines 89 to 96 in bca94e4

	function get_staged(li::MethodInstance)
	try
	# user code might throw errors – ignore them
	return ccall(:jl_code_for_staged, Any, (Any,), li)::CodeInfo
	catch
	return nothing
	end
	end

Just to be clear:

julia> function foo()
           if @generated
               error("I am invisible")
           else
               42
           end
       end
foo (generic function with 1 method)

julia> foo()
42

whereas

julia> @generated function bar()
           error("I am not invisible")
       end
bar (generic function with 1 method)

julia> bar()
ERROR: I am not invisible
Stacktrace:
 [1] error at ./error.jl:33 [inlined]
 [2] #s2#3(::Any) at ./REPL[5]:2
 [3] (::Core.GeneratedFunctionStub)(::Any, ::Vararg{Any,N} where N) at ./boot.jl:466
 [4] top-level scope

I've long had a local change that prints the error along with a backtrace, since with CUDAnative we don't even get to see the error of a @generated function (because of allocs, calls to the runtime, etc). Now that there's a similar situation with regular code, I wonder if we need to do something more user friendly.

[Keno]

What it should probably do is turn in an Expr(:error), which gets thrown at runtime.