The output of this file

eltypestuple() = Tuple{}
eltypestuple(x, xs...) = Tuple{eltype(x), eltypes(xs...).types...}

null_safe_op(::Any, ::Type, ::Type...) = false
Base.@pure null_safe_eltype_op(f, xs...) =
    null_safe_op(f, eltypestuple(xs...).types...)

@inline function broadcast_c(f, a...)
    if null_safe_eltype_op(f, a...)
        0
    else
        1
    end
end

f() = broadcast_c(1:10, 1:10, 1:10)

code_llvm(STDOUT, f, Tuple{})

is on 0.5, quite long; on latest 0.6, also quite long.

But briefly, around when #16961 was merged, it was just

define i64 @julia_f_61750() #0 !dbg !5 {
top:
  ret i64 1
}

so something must have changed since. (The latest commit I know to have the short LLVM is 495114c, but I'll bisect to find a better bound.)

cc @JeffBezanson

Some notes:

• If @pure is removed then the generated LLVM is never short;
• @vtjnash pointed out that @pure is incorrect anyway, so we should probably still go with Refactor null_safe_op to workaround codegen changes #21290?