diff --git a/src/libstd/sync/once.rs b/src/libstd/sync/once.rs
index 51c42995d5e..443e2a6980d 100644
--- a/src/libstd/sync/once.rs
+++ b/src/libstd/sync/once.rs
@@ -220,7 +220,11 @@ impl Once {
     #[stable(feature = "rust1", since = "1.0.0")]
     pub fn call_once<F>(&self, f: F) where F: FnOnce() {
         // Fast path, just see if we've completed initialization.
-        if self.state.load(Ordering::SeqCst) == COMPLETE {
+        // An `Acquire` load is enough because that makes all the initialization
+        // operations visible to us. The cold path uses SeqCst consistently
+        // because the performance difference really does not matter there,
+        // and SeqCst minimizes the chances of something going wrong.
+        if self.state.load(Ordering::Acquire) == COMPLETE {
             return
         }
 
@@ -277,7 +281,11 @@ impl Once {
     #[unstable(feature = "once_poison", issue = "33577")]
     pub fn call_once_force<F>(&self, f: F) where F: FnOnce(&OnceState) {
         // same as above, just with a different parameter to `call_inner`.
-        if self.state.load(Ordering::SeqCst) == COMPLETE {
+        // An `Acquire` load is enough because that makes all the initialization
+        // operations visible to us. The cold path uses SeqCst consistently
+        // because the performance difference really does not matter there,
+        // and SeqCst minimizes the chances of something going wrong.
+        if self.state.load(Ordering::Acquire) == COMPLETE {
             return
         }