Generate original vtable and clone's vtable in the same CGU

1 files changed, 13 insertions, 0 deletions

diff --git a/library/alloc/src/task.rs b/library/alloc/src/task.rs
index b214f4946b1..b40768a52b6 100644
--- a/library/alloc/src/task.rs
+++ b/library/alloc/src/task.rs
@@ -136,6 +136,15 @@ impl<W: Wake + Send + Sync + 'static> From<Arc<W>> for RawWaker {
 #[inline(always)]
 fn raw_waker<W: Wake + Send + Sync + 'static>(waker: Arc<W>) -> RawWaker {
     // Increment the reference count of the arc to clone it.
+    //
+    // The #[inline(always)] is to ensure that raw_waker and clone_waker are
+    // always generated in the same code generation unit as one another, and
+    // therefore that the structurally identical const-promoted RawWakerVTable
+    // within both functions is deduplicated at LLVM IR code generation time.
+    // This allows optimizing Waker::will_wake to a single pointer comparison of
+    // the vtable pointers, rather than comparing all four function pointers
+    // within the vtables.
+    #[inline(always)]
     unsafe fn clone_waker<W: Wake + Send + Sync + 'static>(waker: *const ()) -> RawWaker {
         unsafe { Arc::increment_strong_count(waker as *const W) };
         RawWaker::new(
@@ -304,6 +313,10 @@ impl<W: LocalWake + 'static> From<Rc<W>> for RawWaker {
 #[inline(always)]
 fn local_raw_waker<W: LocalWake + 'static>(waker: Rc<W>) -> RawWaker {
     // Increment the reference count of the Rc to clone it.
+    //
+    // Refer to the comment on raw_waker's clone_waker regarding why this is
+    // always inline.
+    #[inline(always)]
     unsafe fn clone_waker<W: LocalWake + 'static>(waker: *const ()) -> RawWaker {
         unsafe { Rc::increment_strong_count(waker as *const W) };
         RawWaker::new(