Unless I'm mistaken, WHN in some of his seminal work added a redundant mechanism for detecting SETF macros that are defined after the existence of #'(SETF X) has been assumed.
Namely, the *SETF-ASSUMED-FBOUNDP* table contains an entry whenever IR2 emits a call to (SETF X). This differs in a subtle and not useful - I would go so far as to say "wrong" - way from the same information in globaldb that tells us that (SETF X) has been assumed to exist.

The extraneous mechanism relies on PONDER-FULL-CALL. Supposing IR1 makes a ref to #'(SETF FOO) but IR2 deals with it such that no full call occurs, how can it possibly be right not to warn about a subsequent defsetf for FOO?  Surely it's fishy that _any_ other thing whatsoever co-exists with a user-defined expander of the macro-like variety.

Following is the bulk of the patch to change the check to use (info :function :where-from).
The other 2 places to edit are in 'globaldb' and 'ir2tran' but those are trivial deletions.

Also I believe the FIXME (which I deleted) pertained to a misunderstanding about why both :INVERSE and :EXPANDER are examined. It has nothing to do with host/target issues.
I pasted in the patch below and confirmed it to warn in this very simple test case:
(defun tryit (x)
  (setf (meps x) 3))
(defsetf meps set-meps)

--- a/src/code/early-setf.lisp
+++ b/src/code/early-setf.lisp
@@ -342,24 +342,25 @@
     #+sb-xc-host (declare (ignore expander-lambda-list))
         (:symbol name "defining a setf-expander for ~A"))
-    (cond ((gethash name sb!c:*setf-assumed-fboundp*)
-           (warn
-            "defining setf macro for ~S when ~S was previously ~
+    (let ((setf-fn-name `(setf ,name)))
+      (cond ((eq (sb-c::info :function :where-from setf-fn-name) :assumed)
+     (warn
+      "defining setf macro for ~S when ~S was previously ~
              treated as a function"
-            name
-            `(setf ,name)))
-          ((not (fboundp `(setf ,name)))
-           ;; All is well, we don't need any warnings.
-           (values))
-          ((not (eq (symbol-package name) (symbol-package 'aref)))
-           (style-warn "defining setf macro for ~S when ~S is fbound"
-                       name `(setf ,name))))
-    (remhash name sb!c:*setf-assumed-fboundp*)
+      name setf-fn-name))
+    ((not (fboundp setf-fn-name))
+     ;; All is well, we don't need any warnings.
+     (values))
+    ;; Apparently this test is trying to say that we suppress
+    ;; warnings where host CL provides functional (SETF thing)
+    ;; during XC but we want the same as an expander.
+    ;; Is this the clearest way?
+    ((not (eq (symbol-package name) (symbol-package 'aref)))
+     (style-warn "defining setf macro for ~S when ~S is fbound"
+ name setf-fn-name))))
     (when expander
       (setf (%fun-lambda-list expander) expander-lambda-list))
-    ;; FIXME: It's probably possible to join these checks into one form which
-    ;; is appropriate both on the cross-compilation host and on the target.
     (when (or inverse (info :setf :inverse name))
       (setf (info :setf :inverse name) inverse))
     (when (or expander (info :setf :expander name))