diff --git a/collects/tests/typed-scheme/optimizer/tests/float-real.rkt b/collects/tests/typed-scheme/optimizer/tests/float-real.rkt
new file mode 100644
index 0000000000..7f1aadd855
--- /dev/null
+++ b/collects/tests/typed-scheme/optimizer/tests/float-real.rkt
@@ -0,0 +1,15 @@
+#;
+(
+float-real.rkt 13:1 + -- binary float
+float-real.rkt 14:1 + -- binary float
+5.3
+8.7
+14.26
+ )
+
+#lang typed/racket
+
+;; reals within float expressions should be coerced when it's safe to do so
+(+ 2.3 (ann 3 Real)) ; safe
+(+ 2.3 (* (ann 2 Integer) 3.2)) ; inner = unsafe, outer = safe
+(* 2.3 (* (ann 2 Integer) 3.1)) ; all unsafe
diff --git a/collects/typed-scheme/optimizer/float-complex.rkt b/collects/typed-scheme/optimizer/float-complex.rkt
index 534933a4aa..60f0ee0846 100644
--- a/collects/typed-scheme/optimizer/float-complex.rkt
+++ b/collects/typed-scheme/optimizer/float-complex.rkt
@@ -241,7 +241,7 @@
   
   ;; special handling of reals inside complex operations
   ;; must be after any cases that we are supposed to handle
-  (pattern e:float-coerce-expr
+  (pattern e:float-arg-expr
            #:with real-binding (unboxed-gensym 'unboxed-float-)
            #:with imag-binding #f
            #:when (log-optimization "float-coerce-expr in complex ops" #'e)
@@ -252,7 +252,7 @@
   ;; we can eliminate boxing that was introduced by the user
   (pattern (#%plain-app (~and op (~or (~literal make-rectangular)
                                       (~literal unsafe-make-flrectangular)))
-                        real:float-coerce-expr imag:float-coerce-expr)
+                        real:float-arg-expr imag:float-arg-expr)
            #:with real-binding (unboxed-gensym "unboxed-real-")
            #:with imag-binding (unboxed-gensym "unboxed-imag-")
            #:with (bindings ...)
@@ -260,7 +260,7 @@
                   #'(((real-binding) real.opt)
                      ((imag-binding) imag.opt))))
   (pattern (#%plain-app (~and op (~literal make-polar))
-                        r:float-coerce-expr theta:float-coerce-expr)
+                        r:float-arg-expr theta:float-arg-expr)
            #:with magnitude    (unboxed-gensym)
            #:with angle        (unboxed-gensym)
            #:with real-binding (unboxed-gensym "unboxed-real-")
diff --git a/collects/typed-scheme/optimizer/float.rkt b/collects/typed-scheme/optimizer/float.rkt
index 13b51f7e4f..3b3d98e52b 100644
--- a/collects/typed-scheme/optimizer/float.rkt
+++ b/collects/typed-scheme/optimizer/float.rkt
@@ -7,7 +7,7 @@
          (types numeric-tower)
          (optimizer utils fixnum))
 
-(provide float-opt-expr float-coerce-expr)
+(provide float-opt-expr float-arg-expr)
 
 
 (define (mk-float-tbl generic)
@@ -50,15 +50,6 @@
            #:with opt ((optimize) #'e)))
 
 
-;; generates coercions to floats
-(define-syntax-class float-coerce-expr
-  #:commit
-  (pattern e:float-arg-expr
-           #:with opt #'e.opt)
-  (pattern e:real-expr
-           #:with opt #'(exact->inexact e.opt)))
-
-
 ;; if the result of an operation is of type float, its non float arguments
 ;; can be promoted, and we can use unsafe float operations
 ;; note: none of the unary operations have types where non-float arguments
@@ -76,7 +67,12 @@
   (pattern e:int-expr
            #:with opt #'(->fl e.opt))
   (pattern e:float-expr
-           #:with opt #'e.opt))
+           #:with opt #'e.opt)
+  ;; reals within float expressions are not always valid to optimize because
+  ;; of the exact 0 problem, but since float-opt-expr checks whether the
+  ;; surrounding expressing is of type Float and not just Real, this is safe
+  (pattern e:real-expr
+           #:with opt #'(exact->inexact e)))
 
 (define-syntax-class float-opt-expr
   #:commit