diff --git a/docs/subset-preprocessing.md b/docs/subset-preprocessing.md
index 6f780dd4b..dec4f37e0 100644
--- a/docs/subset-preprocessing.md
+++ b/docs/subset-preprocessing.md
@@ -22,8 +22,10 @@ hb_face_t* subset = hb_subset_or_fail (preprocessed, subset_input);
    things up.
 
 *  The preprocessing operation may take longer then the time it takes to produce a subset from the
-   source font. Thus the main performance gains are made when a preprocessed face is reused for multiple
-   subsetting operations.
+   source font. Thus the main performance gains are made when a preprocessed face is reused for
+   multiple subsetting operations.
+
+*  Currently the largest performance gains are seen when using a preprocessed face for CFF subsetting.
 
 *  The preprocessed face may contain references to the memory backing the source face. If this memory
    is fully owned by a harfbuzz hb_blob_t* then it will automatically be kept alive for the lifetime
diff --git a/src/hb-subset-input.cc b/src/hb-subset-input.cc
index 8855d56f4..c94aaacd4 100644
--- a/src/hb-subset-input.cc
+++ b/src/hb-subset-input.cc
@@ -456,6 +456,8 @@ hb_subset_input_pin_axis_location (hb_subset_input_t  *input,
  * subsetter. Future subsetting operations can then use the precomputed data
  * to speed up the subsetting operation. Returns a new hb_face_t*.
  *
+ * See docs/subset-preprocessing.md for more information.
+ *
  * Note: the preprocessed face may contain sub-blobs that reference the memory
  * backing the source hb_face_t*. Therefore in the case that this memory is not
  * owned by the source face you will need to ensure that memory lives