diff --git a/mats/5_3.ms b/mats/5_3.ms
index ace926700c..29f1a16305 100644
--- a/mats/5_3.ms
+++ b/mats/5_3.ms
@@ -2477,6 +2477,10 @@
     (cfl~= (expt 7.7-11.11i -2.0) (* (/ 1.0 7.7-11.11i) (/ 1.0 7.7-11.11i)))
     (~= (expt 11 1/2) (sqrt 11))
     (fl~= (expt 1.5e-20 0.5) (sqrt 1.5e-20))
+    (cfl~= -0.0+0.0i (expt -0.1 1000000000000001/3)) ; inexact approx of second arg is non-integer
+    (eqv? +inf.0+inf.0i (expt -0.1 -1000000000000001/3)) ; inexact approx of second arg is non-integer
+    (cfl~= -0.0+0.0i (expt -0.1 100000000000000000001/3)) ;  inexact approx of second arg is integer
+    (eqv? -inf.0-inf.0i (expt -0.1 -100000000000000000001/3)) ; inexact approx of second arg is integer
    ; test cp0 handling of expt
     (begin
       (define $probably-should-not-use
diff --git a/s/5_3.ss b/s/5_3.ss
index 13288beee9..ee253bb158 100644
--- a/s/5_3.ss
+++ b/s/5_3.ss
@@ -1519,7 +1519,16 @@
                   0
                   ($impoops 'expt "undefined for values ~s and ~s" x y))]
              [(eq? x 1) 1]
-             [(floatable? y) (expt x (inexact y))]
+             [(and (floatable? y)
+                   (let ([y (inexact y)])
+                     ;; Don't use this case if `(inexact y)` loses
+                     ;; precision and becomes an an integer, in which
+                     ;; case the result would be real (but should be
+                     ;; non-real complex)
+                     (and (not (and (flonum? y)
+                                    ($flinteger? y)))
+                          y)))
+              => (lambda (y) (expt x y))]
              [else (exp (* y (log x)))])]
          [else (nonnumber-error 'expt y)])))