Pull #579 was made specifically so rustc could eliminate the failure branch when using getrandom in external code. But I failed to realize that since calls to fill_inner are cross-crate, the compiler only sees the returned Result and can't remove the dead branch. This is easily solved by giving fill_inner in the windows backend the inline annotation, but I was unsure if this would be acceptable, since it seems like no other backed implementation is inlined, even when it's just a brief redirect to a system call.
Is it a policy of getrandom to have fill_inner as the crate boundary in final binaries? Would it be worth it to consider an exception for the windows implementation, since it's a unique case of not being able to fail?