feature_gate::MacroVisitor misses macro invocations nested in macro invocations

[pnkfelix]

Today, this program is correctly rejected:

fn main() {
    unsafe {
        let x = asm!("");
        println!("hi {:?}", x);
    }
}

due to a missing #![feature(asm)]

Unfortunately, this variant program is accepted by the compiler:

fn main() {
    unsafe {
        println!("hi {:?}", asm!(""));
    }
}

The reason (I think) is that the fn visit_mac in feature_gate::MacroVisitor needs to recursively invoke the visitor on the expression inputs to the macro.

[pnkfelix]

(investigating; I thought this would be a simple matter of calling visit::walk_mac, but that implementation is empty! well, because that's the only sane thing to do since our macro invocations are just token trees)

[pnkfelix]

also: Nominating, since this strikes me as a hole in our feature checking.

[pnkfelix]

cc @cmr @brson

[pnkfelix]

(I imagine we could just walk the token tree looking for occurrences of ` '!'`` and then assume that might be a macro invocation. Where would that break down, hmm...)

[pnkfelix]

Another potential approach to handling these scenarios could be to have the expansions inject a token that is effectively a no-op into the generated AST that marks it as being generated via asm! (or whatever other macro expansion we may want to gate), and have the post expansion feature-gate checking look for occurrences of that token. A "danger" there I suppose is that it is possible that another macro could cause the token to be dropped (i.e. a macro that discards some of its inputs), but I think we should be willing to live with that scenario.

[pnkfelix]

1.0 polish issue, P-high.