diff --git a/compiler/rustc_trait_selection/src/traits/dyn_compatibility.rs b/compiler/rustc_trait_selection/src/traits/dyn_compatibility.rs
index 1d57fd9e4b285..0cc886a2c1909 100644
--- a/compiler/rustc_trait_selection/src/traits/dyn_compatibility.rs
+++ b/compiler/rustc_trait_selection/src/traits/dyn_compatibility.rs
@@ -749,14 +749,19 @@ fn receiver_is_dispatchable<'tcx>(
         let trait_predicate = ty::TraitRef::new_from_args(tcx, trait_def_id, args);
         clauses.push(trait_predicate.upcast(tcx));
 
-        // U satisfies `Trait`'s where-bounds.
+        // U satisfies Trait's trait bounds. When normalizing `U: Trait<Arg1, ..., ArgN>`, aliases
+        // in the arguments of supertrait bounds can generate trait obligations which may only be
+        // be provable using where-clauses from Trait (#161621).
         clauses.extend(
             tcx.clauses_of(trait_def_id)
                 .instantiate(tcx, args)
                 .clauses
                 .into_iter()
-                .map(Unnormalized::skip_norm_wip),
+                .map(Unnormalized::skip_norm_wip)
+                .filter(|clause| matches!(clause.kind().skip_binder(), ty::ClauseKind::Trait(_))),
         );
+        // FIXME(@dianne): we need projection clauses too in some form, but adding them all
+        // unmodified can introduce ambiguity (#163550)
 
         let meta_sized_predicate = {
             let meta_sized_did = tcx.require_lang_item(LangItem::MetaSized, DUMMY_SP);
diff --git a/tests/ui/dyn-compatibility/dispatchability-placeholder-does-not-satisfy-bounds.rs b/tests/ui/dyn-compatibility/dispatchability-placeholder-does-not-satisfy-bounds.rs
new file mode 100644
index 0000000000000..c899c66056839
--- /dev/null
+++ b/tests/ui/dyn-compatibility/dispatchability-placeholder-does-not-satisfy-bounds.rs
@@ -0,0 +1,24 @@
+//! Tests for specific instances of <https://github.com/rust-lang/rust/issues/161621> that don't
+//! pass yet. See `dispatchability-placeholder-satisfies-bounds.rs` for context.
+// FIXME(@dianne): this should be gone soon
+//@ known-bug: unknown
+
+trait HasAssoc {
+    type Assoc;
+}
+
+trait Parent<T> {}
+
+trait NotDynCompatible
+where
+    // We need to be able to normalize this for RustaceansAreAwesome in order to prove
+    // `<(RustaceansAreAwesome,) as HasAssoc>::Assoc: Owned` for the projection in the argument to
+    // the `Self: Parent<...>` supertrait bound. Currently a projection clause for it is missing
+    // from the `ParamEnv`.
+    (Self,): HasAssoc<Assoc = str>,
+    Self: Parent<<<(Self,) as HasAssoc>::Assoc as ToOwned>::Owned>,
+{
+    fn f(&self);
+}
+
+fn main() {}
diff --git a/tests/ui/dyn-compatibility/dispatchability-placeholder-does-not-satisfy-bounds.stderr b/tests/ui/dyn-compatibility/dispatchability-placeholder-does-not-satisfy-bounds.stderr
new file mode 100644
index 0000000000000..3b2b8c62e8994
--- /dev/null
+++ b/tests/ui/dyn-compatibility/dispatchability-placeholder-does-not-satisfy-bounds.stderr
@@ -0,0 +1,11 @@
+error[E0277]: the trait bound `<(RustaceansAreAwesome,) as HasAssoc>::Assoc: Clone` is not satisfied
+  --> $DIR/dispatchability-placeholder-does-not-satisfy-bounds.rs:21:5
+   |
+LL |     fn f(&self);
+   |     ^^^^^^^^^^^^ the trait `Clone` is not implemented for `<(RustaceansAreAwesome,) as HasAssoc>::Assoc`
+   |
+   = note: required for `<(RustaceansAreAwesome,) as HasAssoc>::Assoc` to implement `ToOwned`
+
+error: aborting due to 1 previous error
+
+For more information about this error, try `rustc --explain E0277`.
diff --git a/tests/ui/dyn-compatibility/dispatchability-placeholder-satisfies-bounds.rs b/tests/ui/dyn-compatibility/dispatchability-placeholder-satisfies-bounds.rs
index 421b9a512c1e4..611e1cb59d4ab 100644
--- a/tests/ui/dyn-compatibility/dispatchability-placeholder-satisfies-bounds.rs
+++ b/tests/ui/dyn-compatibility/dispatchability-placeholder-satisfies-bounds.rs
@@ -1,11 +1,15 @@
-//! Regression tests for <https://github.com/rust-lang/rust/issues/161621>. When checking whether
-//! method receivers for a trait `Trait` are dynamically dispatchable, we use a placeholder type in
-//! place of `dyn Trait` to avoid a cycle (as that would require determining whether `Trait` is dyn-
-//! compatible). This placeholder must satisfy `Trait`'s where-bounds to avoid errors in param-
-//! environment normalization.
+//! Regression tests for <https://github.com/rust-lang/rust/issues/161621> and similar. When
+//! checking whether method receivers for a trait `Trait` are dynamically dispatchable, we (at the
+//! time of writing) use a placeholder type called `RustaceansAreAwesome` in place of `dyn Trait` to
+//! avoid a cycle (as using `dyn Trait` requires determining whether `Trait` is dyn-compatible). We
+//! assume that `RustaceansAreAwesome` implements `Trait`, so when normalizing the param-environment
+//! we construct, we may have to prove obligations for aliases encountered in arguments to super-
+//! trait bounds. It's possible for these to only be provable using clauses from `Trait` or its
+//! supertraits that would be necessary for `RustaceansAreAwesome` to implement `Trait`.
 //@ check-pass
 
-// Test 1
+// This works for `Self: Super<...Ty::<RustaceansAreAwesome>::Assoc...>` after the `where`.
+// We need to assume `(Args, RustaceansAreAwesome::Output): SignatureToFnPtr`.
 
 use std::ops::Deref;
 
@@ -22,7 +26,8 @@ where
     fn method(&self);
 }
 
-// Test 2
+// This works for `Trait: Super<...Ty<RustaceansAreAwesome>::Assoc...>` before the `where`.
+// We need to assume `A: Associator<RustaceansAreAwesome>`.
 
 trait Associator<S: ?Sized> {
     type Point;
@@ -34,4 +39,60 @@ trait Trait<A: Associator<Self> + ?Sized>: Marker<A::Point> {
     fn method(&self);
 }
 
+// Regression test for <https://github.com/rust-lang/rust/issues/161621>: we need to be careful with
+// projection clauses involving `RustaceansAreAwesome`. Here, if we simply assumed that
+// `<T as Matrix2D>::RowIndex == <RustaceansAreAwesome as Matrix2D>::RowIndex`, we'd end up with
+// ambiguity, as we also assume `<T as Matrix2D>::RowIndex == <Self as Matrix2D>::RowIndex`.`Self`
+// stands in for the `Self` type of the impl we're imagining dispatching to (pretending
+// `RustaceansAreAwesome` is a trait object).
+
+pub trait Matrix {
+    type Coordinates;
+}
+
+pub trait Matrix2D: Matrix<Coordinates = <Self as Matrix2D>::RowIndex> {
+    type RowIndex;
+}
+
+pub trait Transposable<T: Matrix2D<RowIndex = Self::RowIndex>>: Matrix2D {
+    fn transpose(&self) -> T;
+}
+
+// An example where we can't add any additional projections without introducing ambiguity: we don't
+// want to assume `<() as HasAssoc>::Assoc` is both `Self` and `RustaceansAreAwesome`. At the time
+// of writing, we let `<() as HasAssoc>::Assoc` normalize to `Self` during receiver dispatchability
+// checking. This should be fine, since to call methods on a `dyn TechnicallyDynCompatible`, we'd
+// need `<() as HasAssoc>::Assoc` to be `dyn TechnicallyDynCompatible`. This is the same reasoning
+// that lets us assume `RustaceansAreAwesome` implements its trait at all: to use the trait object,
+// it must implement the trait; otherwise, receiver dispatchability doesn't mean much.
+
+trait HasAssoc {
+    type Assoc;
+}
+
+impl HasAssoc for () {
+    type Assoc = ();
+}
+
+trait Parent<T: ?Sized> {}
+
+trait TechnicallyDynCompatible
+where
+    (): HasAssoc<Assoc = Self>,
+    Self: Parent<<() as HasAssoc>::Assoc>,
+{
+    fn f(&self) {}
+}
+
+// Since we have `TechnincallyDynCompatible` defined, let's keep its dyn-compatibility from breaking
+// accidentally (and also sanity-check the above comment).
+
+impl Parent<()> for () {}
+impl TechnicallyDynCompatible for () {}
+
+fn technically_dyn_compatible_is_dyn_compatible() {
+    let x: &dyn TechnicallyDynCompatible = &();
+    // We can't call `x.f()` since `<() as HasAssoc>::Assoc` isn't `dyn TechnicallyDynCompatible`.
+}
+
 fn main() {}