diff --git a/cranelift/codegen/src/isa/x64/lower.isle b/cranelift/codegen/src/isa/x64/lower.isle
index af11e659de8d..e0d17fe627a4 100644
--- a/cranelift/codegen/src/isa/x64/lower.isle
+++ b/cranelift/codegen/src/isa/x64/lower.isle
@@ -2082,8 +2082,7 @@
 ;; than one instruction for certain types (e.g., XMM-held, I128).
 
 (rule (lower (has_type ty (select (maybe_uextend (icmp cc a @ (value_type (fits_in_64 a_ty)) b)) x y)))
-      (let ((size OperandSize (raw_operand_size_of_type a_ty)))
-           (with_flags (x64_cmp size b a) (cmove_from_values ty cc x y))))
+      (lower_select_icmp ty (emit_cmp cc a b) x y))
 
 ;; Finally, we lower `select` from a condition value `c`. These rules are meant
 ;; to be the final, default lowerings if no other patterns matched above.
@@ -2099,6 +2098,10 @@
       (let ((cond_result IcmpCondResult (cmp_zero_i128 (CC.Z) c)))
         (select_icmp cond_result x y)))
 
+(decl lower_select_icmp (Type IcmpCondResult Value Value) InstOutput)
+(rule (lower_select_icmp ty (IcmpCondResult.Condition flags cc) x y)
+      (with_flags flags (cmove_from_values ty cc x y)))
+
 ;; Specializations for floating-point compares to generate a `mins*` or a
 ;; `maxs*` instruction. These are equivalent to the "pseudo-m{in,ax}"
 ;; specializations for vectors.