PERF: enhance MultiIndex.remove_unused_levels when no levels are unused

doc/source/whatsnew/v0.23.0.txt

@@ -901,6 +901,7 @@ Performance Improvements
 - :func:`Series` / :func:`DataFrame` tab completion limits to 100 values, for better performance. (:issue:`18587`)
 - Improved performance of :func:`DataFrame.median` with ``axis=1`` when bottleneck is not installed (:issue:`16468`)
 - Improved performance of :func:`MultiIndex.get_loc` for large indexes, at the cost of a reduction in performance for small ones (:issue:`18519`)
+- Improved performance of :func:`MultiIndex.remove_unused_levels` when there are no unused levels, at the cost of a reduction in performance when there are (:issue:`19289`)
 - Improved performance of pairwise ``.rolling()`` and ``.expanding()`` with ``.cov()`` and ``.corr()`` operations (:issue:`17917`)
 - Improved performance of :func:`pandas.core.groupby.GroupBy.rank` (:issue:`15779`)
 - Improved performance of variable ``.rolling()`` on ``.min()`` and ``.max()`` (:issue:`19521`)

pandas/core/indexes/multi.py

@@ -1476,27 +1476,35 @@ def remove_unused_levels(self):
         changed = False
         for lev, lab in zip(self.levels, self.labels):
 
-            uniques = algos.unique(lab)
-            na_idx = np.where(uniques == -1)[0]
-
-            # nothing unused
-            if len(uniques) != len(lev) + len(na_idx):
+            # Since few levels are typically unused, bincount() is more
+            # efficient than unique() - however it only accepts positive values
+            # (and drops order):
+            uniques = np.where(np.bincount(lab + 1) > 0)[0] - 1
+            has_na = int(len(uniques) and (uniques[0] == -1))
+
+            if len(uniques) != len(lev) + has_na:
+                # We have unused levels
                 changed = True
 
-                if len(na_idx):
+                # Recalculate uniques, now preserving order.
+                # Can easily be cythonized by exploiting the already existing
+                # "uniques" and stop parsing "lab" when all items are found:
+                uniques = algos.unique(lab)
+                if has_na:
+                    na_idx = np.where(uniques == -1)[0]
                     # Just ensure that -1 is in first position:
                     uniques[[0, na_idx[0]]] = uniques[[na_idx[0], 0]]
 
                 # labels get mapped from uniques to 0:len(uniques)
                 # -1 (if present) is mapped to last position
-                label_mapping = np.zeros(len(lev) + len(na_idx))
+                label_mapping = np.zeros(len(lev) + has_na)
                 # ... and reassigned value -1:
-                label_mapping[uniques] = np.arange(len(uniques)) - len(na_idx)
+                label_mapping[uniques] = np.arange(len(uniques)) - has_na
 
                 lab = label_mapping[lab]
 
                 # new levels are simple
-                lev = lev.take(uniques[len(na_idx):])
+                lev = lev.take(uniques[has_na:])
 
             new_levels.append(lev)
             new_labels.append(lab)