diff --git a/src/machine.rs b/src/machine.rs
index 2f0aa915753..a7d62897b8f 100644
--- a/src/machine.rs
+++ b/src/machine.rs
@@ -27,7 +27,7 @@ use rustc_target::abi::{LayoutOf, Size};
 
 use crate::*;
 
-pub use crate::threads::{ThreadId, ThreadSet, ThreadLocalStorage};
+pub use crate::threads::{ThreadId, ThreadSet, ThreadState, ThreadLocalStorage};
 
 // Some global facts about the emulated machine.
 pub const PAGE_SIZE: u64 = 4 * 1024; // FIXME: adjust to target architecture
diff --git a/src/shims/tls.rs b/src/shims/tls.rs
index d16acb75003..ec8c31fe2c1 100644
--- a/src/shims/tls.rs
+++ b/src/shims/tls.rs
@@ -9,7 +9,7 @@ use rustc_middle::ty;
 use rustc_target::abi::{Size, HasDataLayout};
 
 use crate::{HelpersEvalContextExt, ThreadsEvalContextExt, InterpResult, MPlaceTy, Scalar, StackPopCleanup, Tag};
-use crate::machine::ThreadId;
+use crate::machine::{ThreadId, ThreadState};
 
 pub type TlsKey = u128;
 
@@ -174,7 +174,7 @@ impl<'tcx> TlsData<'tcx> {
 
 impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriEvalContext<'mir, 'tcx> {}
 pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx> {
-    /// Run TLS destructors for the currently active thread.
+    /// Run TLS destructors for all threads.
     fn run_tls_dtors(&mut self) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
         assert!(!this.machine.tls.dtors_running, "running TLS dtors twice");
@@ -223,7 +223,9 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         }
 
         // Now run the "keyed" destructors.
-        for thread_id in this.get_all_thread_ids() {
+        for (thread_id, thread_state) in this.get_all_thread_ids_with_states() {
+            assert!(thread_state == ThreadState::Terminated,
+                    "TLS destructors should be executed after all threads terminated.");
             this.set_active_thread(thread_id)?;
             let mut dtor = this.machine.tls.fetch_tls_dtor(None, thread_id);
             while let Some((instance, thread_id, ptr, key)) = dtor {
diff --git a/src/threads.rs b/src/threads.rs
index 9d982359bfb..4458f4410e4 100644
--- a/src/threads.rs
+++ b/src/threads.rs
@@ -97,10 +97,7 @@ impl<'mir, 'tcx> Default for ThreadSet<'mir, 'tcx> {
     fn default() -> Self {
         let mut threads = IndexVec::new();
         threads.push(Default::default());
-        Self {
-            active_thread: ThreadId::new(0),
-            threads: threads,
-        }
+        Self { active_thread: ThreadId::new(0), threads: threads }
     }
 }
 
@@ -156,8 +153,8 @@ impl<'mir, 'tcx: 'mir> ThreadSet<'mir, 'tcx> {
         }
     }
     /// Get ids of all threads ever allocated.
-    fn get_all_thread_ids(&mut self) -> Vec<ThreadId> {
-        (0..self.threads.len()).map(ThreadId::new).collect()
+    fn get_all_thread_ids_with_states(&self) -> Vec<(ThreadId, ThreadState)> {
+        self.threads.iter_enumerated().map(|(id, thread)| (id, thread.state)).collect()
     }
     /// Decide which thread to run next.
     ///
@@ -283,9 +280,9 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         let this = self.eval_context_mut();
         this.machine.threads.active_thread_stack_mut()
     }
-    fn get_all_thread_ids(&mut self) -> Vec<ThreadId> {
+    fn get_all_thread_ids_with_states(&mut self) -> Vec<(ThreadId, ThreadState)> {
         let this = self.eval_context_mut();
-        this.machine.threads.get_all_thread_ids()
+        this.machine.threads.get_all_thread_ids_with_states()
     }
     /// Decide which thread to run next.
     ///