Make sure to treat only param where clauses as inherent

compiler/rustc_hir_typeck/src/method/probe.rs

@@ -1945,6 +1945,29 @@ impl<'a, 'tcx> ProbeContext<'a, 'tcx> {
                     );
                     (xform_self_ty, xform_ret_ty) =
                         self.xform_self_ty(probe.item, trait_ref.self_ty(), trait_ref.args);
+
+                    if matches!(probe.kind, WhereClauseCandidate(_)) {
+                        // `WhereClauseCandidate` requires that the self type is a param,
+                        // because it has special behavior with candidate preference as an
+                        // inherent pick.
+                        match ocx.structurally_normalize_ty(
+                            cause,
+                            self.param_env,
+                            trait_ref.self_ty(),
+                        ) {
+                            Ok(ty) => {
+                                if !matches!(ty.kind(), ty::Param(_)) {
+                                    debug!("--> not a param ty: {xform_self_ty:?}");
+                                    return ProbeResult::NoMatch;
+                                }
+                            }
+                            Err(errors) => {
+                                debug!("--> cannot relate self-types {:?}", errors);
+                                return ProbeResult::NoMatch;
+                            }
+                        }
+                    }
+
                     xform_self_ty = ocx.normalize(cause, self.param_env, xform_self_ty);
                     match ocx.relate(cause, self.param_env, self.variance(), self_ty, xform_self_ty)
                     {

tests/ui/methods/rigid-alias-bound-is-not-inherent.rs

@@ -0,0 +1,42 @@
+// See the code below.
+//
+// We were using `DeepRejectCtxt` to ensure that `assemble_inherent_candidates_from_param`
+// did not rely on the param-env being eagerly normalized. Since aliases unify with all
+// types, this meant that a rigid param-env candidate like `<T as Deref>::Target: Trait1`
+// would be registered as a "WhereClauseCandidate", which is treated as inherent. Since
+// we evaluate these candidates for all self types in the deref chain, this candidate
+// would be satisfied for `<T as Deref>::Target`, meaning that it would be preferred over
+// an "extension" candidate like `<T as Deref>::Target: Trait2` even though it holds.
+// This is problematic, since it causes ambiguities to be broken somewhat arbitrarily.
+// And as a side-effect, it also caused our computation of "used" traits to be miscalculated
+// since inherent candidates don't count as an import usage.
+
+use std::ops::Deref;
+
+trait Trait1 {
+    fn method(&self) {
+        println!("1");
+    }
+}
+
+trait Trait2 {
+    fn method(&self) {
+        println!("2");
+    }
+}
+impl<T: Other + ?Sized> Trait2 for T {}
+
+trait Other {}
+
+fn foo<T>(x: T)
+where
+    T: Deref,
+    <T as Deref>::Target: Trait1 + Other,
+{
+    // Make sure that we don't prefer methods from where clauses for rigid aliases,
+    // just for params. We could revisit this behavior, but it would be a lang change.
+    x.method();
+    //~^ ERROR multiple applicable items in scope
+}
+
+fn main() {}

tests/ui/methods/rigid-alias-bound-is-not-inherent.stderr

@@ -0,0 +1,30 @@
+error[E0034]: multiple applicable items in scope
+  --> $DIR/rigid-alias-bound-is-not-inherent.rs:38:7
+   |
+LL |     x.method();
+   |       ^^^^^^ multiple `method` found
+   |
+note: candidate #1 is defined in the trait `Trait1`
+  --> $DIR/rigid-alias-bound-is-not-inherent.rs:17:5
+   |
+LL |     fn method(&self) {
+   |     ^^^^^^^^^^^^^^^^
+note: candidate #2 is defined in an impl of the trait `Trait2` for the type `T`
+  --> $DIR/rigid-alias-bound-is-not-inherent.rs:23:5
+   |
+LL |     fn method(&self) {
+   |     ^^^^^^^^^^^^^^^^
+help: disambiguate the method for candidate #1
+   |
+LL -     x.method();
+LL +     Trait1::method(&x);
+   |
+help: disambiguate the method for candidate #2
+   |
+LL -     x.method();
+LL +     Trait2::method(&x);
+   |
+
+error: aborting due to 1 previous error
+
+For more information about this error, try `rustc --explain E0034`.