diff --git a/src/concurrency_limiter.rs b/src/concurrency_limiter.rs
index 9678969134a..a73860cf18b 100644
--- a/src/concurrency_limiter.rs
+++ b/src/concurrency_limiter.rs
@@ -6,7 +6,7 @@
 // FIXME don't panic when a worker thread panics
 
 pub(super) struct ConcurrencyLimiter {
-    helper_thread: Option<HelperThread>,
+    helper_thread: Option<Mutex<HelperThread>>,
     state: Arc<Mutex<state::ConcurrencyLimiterState>>,
     available_token_condvar: Arc<Condvar>,
     finished: bool,
@@ -39,14 +39,14 @@ pub(super) fn new(sess: &Session, pending_jobs: usize) -> Self {
             })
             .unwrap();
         ConcurrencyLimiter {
-            helper_thread: Some(helper_thread),
+            helper_thread: Some(Mutex::new(helper_thread)),
             state,
             available_token_condvar,
             finished: false,
         }
     }
 
-    pub(super) fn acquire(&mut self, dcx: &rustc_errors::DiagCtxt) -> ConcurrencyLimiterToken {
+    pub(super) fn acquire(&self, dcx: &rustc_errors::DiagCtxt) -> ConcurrencyLimiterToken {
         let mut state = self.state.lock().unwrap();
         loop {
             state.assert_invariants();
@@ -73,16 +73,11 @@ pub(super) fn acquire(&mut self, dcx: &rustc_errors::DiagCtxt) -> ConcurrencyLim
                 }
             }
 
-            self.helper_thread.as_mut().unwrap().request_token();
+            self.helper_thread.as_ref().unwrap().lock().unwrap().request_token();
             state = self.available_token_condvar.wait(state).unwrap();
         }
     }
 
-    pub(super) fn job_already_done(&mut self) {
-        let mut state = self.state.lock().unwrap();
-        state.job_already_done();
-    }
-
     pub(crate) fn finished(mut self) {
         self.helper_thread.take();
 
@@ -190,14 +185,6 @@ pub(super) fn job_finished(&mut self) {
             self.assert_invariants();
         }
 
-        pub(super) fn job_already_done(&mut self) {
-            self.assert_invariants();
-            self.pending_jobs -= 1;
-            self.assert_invariants();
-            self.drop_excess_capacity();
-            self.assert_invariants();
-        }
-
         pub(super) fn poison(&mut self, error: String) {
             self.poisoned = true;
             self.stored_error = Some(error);
diff --git a/src/driver/aot.rs b/src/driver/aot.rs
index e8c96486041..2651e56cac4 100644
--- a/src/driver/aot.rs
+++ b/src/driver/aot.rs
@@ -15,6 +15,7 @@
 use rustc_codegen_ssa::{CodegenResults, CompiledModule, CrateInfo, ModuleKind};
 use rustc_data_structures::profiling::SelfProfilerRef;
 use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
+use rustc_data_structures::sync::{par_map, IntoDynSyncSend};
 use rustc_metadata::fs::copy_to_stdout;
 use rustc_metadata::EncodedMetadata;
 use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
@@ -604,39 +605,39 @@ pub(crate) fn run_aot(
 
     let global_asm_config = Arc::new(crate::global_asm::GlobalAsmConfig::new(tcx));
 
-    let mut concurrency_limiter = ConcurrencyLimiter::new(tcx.sess, cgus.len());
+    let (todo_cgus, done_cgus) =
+        cgus.into_iter().enumerate().partition::<Vec<_>, _>(|&(i, _)| match cgu_reuse[i] {
+            _ if backend_config.disable_incr_cache => true,
+            CguReuse::No => true,
+            CguReuse::PreLto | CguReuse::PostLto => false,
+        });
+
+    let concurrency_limiter = IntoDynSyncSend(ConcurrencyLimiter::new(tcx.sess, todo_cgus.len()));
 
     let modules = tcx.sess.time("codegen mono items", || {
-        cgus.iter()
-            .enumerate()
-            .map(|(i, cgu)| {
-                let cgu_reuse =
-                    if backend_config.disable_incr_cache { CguReuse::No } else { cgu_reuse[i] };
-                match cgu_reuse {
-                    CguReuse::No => {
-                        let dep_node = cgu.codegen_dep_node(tcx);
-                        tcx.dep_graph
-                            .with_task(
-                                dep_node,
-                                tcx,
-                                (
-                                    backend_config.clone(),
-                                    global_asm_config.clone(),
-                                    cgu.name(),
-                                    concurrency_limiter.acquire(tcx.dcx()),
-                                ),
-                                module_codegen,
-                                Some(rustc_middle::dep_graph::hash_result),
-                            )
-                            .0
-                    }
-                    CguReuse::PreLto | CguReuse::PostLto => {
-                        concurrency_limiter.job_already_done();
-                        OngoingModuleCodegen::Sync(reuse_workproduct_for_cgu(tcx, cgu))
-                    }
-                }
-            })
-            .collect::<Vec<_>>()
+        let mut modules: Vec<_> = par_map(todo_cgus, |(_, cgu)| {
+            let dep_node = cgu.codegen_dep_node(tcx);
+            tcx.dep_graph
+                .with_task(
+                    dep_node,
+                    tcx,
+                    (
+                        backend_config.clone(),
+                        global_asm_config.clone(),
+                        cgu.name(),
+                        concurrency_limiter.acquire(tcx.dcx()),
+                    ),
+                    module_codegen,
+                    Some(rustc_middle::dep_graph::hash_result),
+                )
+                .0
+        });
+        modules.extend(
+            done_cgus
+                .into_iter()
+                .map(|(_, cgu)| OngoingModuleCodegen::Sync(reuse_workproduct_for_cgu(tcx, cgu))),
+        );
+        modules
     });
 
     let mut allocator_module = make_module(tcx.sess, &backend_config, "allocator_shim".to_string());
@@ -705,6 +706,6 @@ pub(crate) fn run_aot(
         metadata_module,
         metadata,
         crate_info: CrateInfo::new(tcx, target_cpu),
-        concurrency_limiter,
+        concurrency_limiter: concurrency_limiter.0,
     })
 }