diff --git a/src/stacked_borrows.rs b/src/stacked_borrows.rs
index 250def0c7ce6..bac35796ed13 100644
--- a/src/stacked_borrows.rs
+++ b/src/stacked_borrows.rs
@@ -625,6 +625,11 @@ trait EvalContextPrivExt<'a, 'mir, 'tcx: 'a+'mir>: crate::MiriEvalContextExt<'a,
         };
 
         // Reborrow.
+        // TODO: With `two_phase == true`, this performs a weak reborrow for a `Unique`. That
+        // can lead to some possibly surprising effects, if the parent permission is
+        // `SharedReadWrite` then we now have a `Unique` in the middle of them, which "splits"
+        // them in terms of what remains valid when the `Unique` gets used.  Is that really
+        // what we want?
         this.reborrow(place, size, kind, new_tag, /*force_weak:*/ two_phase, protect)?;
         let new_place = place.replace_tag(new_tag);
         // Handle two-phase borrows.
diff --git a/tests/compile-fail/stacked_borrows/illegal_read7.rs b/tests/compile-fail/stacked_borrows/illegal_read7.rs
new file mode 100644
index 000000000000..25d0878c0455
--- /dev/null
+++ b/tests/compile-fail/stacked_borrows/illegal_read7.rs
@@ -0,0 +1,20 @@
+// Creating a shared reference does not leak the data to raw pointers,
+// not even when interior mutability is involved.
+
+use std::cell::Cell;
+use std::ptr;
+
+fn main() { unsafe {
+    let x = &mut Cell::new(0);
+    let raw = x as *mut Cell<i32>;
+    let x = &mut *raw;
+    let _shr = &*x;
+    // The state here is interesting because the top of the stack is [Unique, SharedReadWrite],
+    // just like if we had done `x as *mut _`.
+    // If we said that reading from a lower item is fine if the top item is `SharedReadWrite`
+    // (one way to maybe preserve a stack discipline), then we could now read from `raw`
+    // without invalidating `x`.  That would be bad!  It would mean that creating `shr`
+    // leaked `x` to `raw`.
+    let _val = ptr::read(raw);
+    let _val = *x.get_mut(); //~ ERROR borrow stack
+} }