PERF: Fix performance regression in infer_dtype

asv_bench/benchmarks/dtypes.py

@@ -5,6 +5,7 @@
 from .pandas_vb_common import (
     datetime_dtypes,
     extension_dtypes,
+    lib,
     numeric_dtypes,
     string_dtypes,
 )
@@ -40,4 +41,25 @@ def time_pandas_dtype_invalid(self, dtype):
             pass
 
 
+class InferDtypes:
+    param_names = ["dtype"]
+    data_dict = {
+        "np-object": np.array([1] * 100000, dtype="O"),
+        "py-object": [1] * 100000,
+        "np-null": np.array([1] * 50000 + [np.nan] * 50000),
+        "py-null": [1] * 50000 + [None] * 50000,
+        "np-int": np.array([1] * 100000, dtype=int),
+        "np-floating": np.array([1.0] * 100000, dtype=float),
+        "empty": [],
+        "bytes": [b"a"] * 100000,
+    }
+    params = list(data_dict.keys())
+
+    def time_infer_skipna(self, dtype):
+        lib.infer_dtype(self.data_dict[dtype], skipna=True)
+
+    def time_infer(self, dtype):
+        lib.infer_dtype(self.data_dict[dtype], skipna=False)
+
+
 from .pandas_vb_common import setup  # noqa: F401 isort:skip

doc/source/whatsnew/v1.0.0.rst

@@ -659,6 +659,7 @@ Performance improvements
 - Performance improvement when checking if values in a :class:`Categorical` are equal, equal or larger or larger than a given scalar.
   The improvement is not present if checking if the :class:`Categorical` is less than or less than or equal than the scalar (:issue:`29820`)
 - Performance improvement in :meth:`Index.equals` and  :meth:`MultiIndex.equals` (:issue:`29134`)
+- Performance improvement in :func:`~pandas.api.types.infer_dtype` when ``skipna`` is ``True`` (:issue:`28814`)
 
 .. _whatsnew_1000.bug_fixes:
 

pandas/_libs/lib.pyx

@@ -1259,16 +1259,16 @@ def infer_dtype(value: object, skipna: bool = True) -> str:
     # make contiguous
     values = values.ravel()
 
-    if skipna:
-        values = values[~isnaobj(values)]
-
     val = _try_infer_map(values)
     if val is not None:
         return val
 
     if values.dtype != np.object_:
         values = values.astype('O')
 
+    if skipna:
+        values = values[~isnaobj(values)]
+
     n = len(values)
     if n == 0:
         return 'empty'