proc_macro: use crossbeam-channel for the CrossThread execution strategy

Compared to mpsc::channel, crossbeam-channel has significantly lower overhead.

Cargo.lock

@@ -3840,6 +3840,7 @@ dependencies = [
 name = "rustc_expand"
 version = "0.0.0"
 dependencies = [
+ "crossbeam-channel",
  "rustc_ast",
  "rustc_ast_passes",
  "rustc_ast_pretty",

compiler/rustc_expand/Cargo.toml

@@ -25,3 +25,4 @@ rustc_parse = { path = "../rustc_parse" }
 rustc_session = { path = "../rustc_session" }
 smallvec = { version = "1.6.1", features = ["union", "may_dangle"] }
 rustc_ast = { path = "../rustc_ast" }
+crossbeam-channel = "0.5.0"

compiler/rustc_expand/src/proc_macro.rs

@@ -12,8 +12,29 @@ use rustc_parse::parser::ForceCollect;
 use rustc_span::def_id::CrateNum;
 use rustc_span::{Span, DUMMY_SP};
 
+struct CrossbeamMessagePipe<T> {
+    tx: crossbeam_channel::Sender<T>,
+    rx: crossbeam_channel::Receiver<T>,
+}
+
+impl<T> pm::bridge::server::MessagePipe<T> for CrossbeamMessagePipe<T> {
+    fn new() -> (Self, Self) {
+        let (tx1, rx1) = crossbeam_channel::unbounded();
+        let (tx2, rx2) = crossbeam_channel::unbounded();
+        (CrossbeamMessagePipe { tx: tx1, rx: rx2 }, CrossbeamMessagePipe { tx: tx2, rx: rx1 })
+    }
+
+    fn send(&mut self, value: T) {
+        self.tx.send(value).unwrap();
+    }
+
+    fn recv(&mut self) -> Option<T> {
+        self.rx.recv().ok()
+    }
+}
+
 fn exec_strategy(ecx: &ExtCtxt<'_>) -> impl pm::bridge::server::ExecutionStrategy {
-    <pm::bridge::server::MaybeCrossThread<pm::bridge::server::StdMessagePipe<_>>>::new(
+    <pm::bridge::server::MaybeCrossThread<CrossbeamMessagePipe<_>>>::new(
         ecx.sess.opts.debugging_opts.proc_macro_cross_thread,
     )
 }

src/tools/tidy/src/deps.rs

@@ -97,6 +97,7 @@ const PERMITTED_DEPENDENCIES: &[&str] = &[
     "compiler_builtins",
     "cpuid-bool",
     "crc32fast",
+    "crossbeam-channel",
     "crossbeam-deque",
     "crossbeam-epoch",
     "crossbeam-queue",