REF: Index.__new__ (#38665)

Author: jbrockmendel

pandas/core/indexes/base.py

@@ -66,6 +66,12 @@
     validate_all_hashable,
 )
 from pandas.core.dtypes.concat import concat_compat
+from pandas.core.dtypes.dtypes import (
+    CategoricalDtype,
+    DatetimeTZDtype,
+    IntervalDtype,
+    PeriodDtype,
+)
 from pandas.core.dtypes.generic import (
     ABCDatetimeIndex,
     ABCMultiIndex,
@@ -331,12 +337,6 @@ def __new__(
 
         # index-like
         elif isinstance(data, (np.ndarray, Index, ABCSeries)):
-            # Delay import for perf. https://github.com/pandas-dev/pandas/pull/31423
-            from pandas.core.indexes.numeric import (
-                Float64Index,
-                Int64Index,
-                UInt64Index,
-            )
 
             if dtype is not None:
                 # we need to avoid having numpy coerce
@@ -347,42 +347,31 @@ def __new__(
                 data = _maybe_cast_with_dtype(data, dtype, copy)
                 dtype = data.dtype  # TODO: maybe not for object?
 
-            # maybe coerce to a sub-class
-            if is_signed_integer_dtype(data.dtype):
-                return Int64Index(data, copy=copy, dtype=dtype, name=name)
-            elif is_unsigned_integer_dtype(data.dtype):
-                return UInt64Index(data, copy=copy, dtype=dtype, name=name)
-            elif is_float_dtype(data.dtype):
-                return Float64Index(data, copy=copy, dtype=dtype, name=name)
-            elif issubclass(data.dtype.type, bool) or is_bool_dtype(data):
-                subarr = data.astype("object")
+            if data.dtype.kind in ["i", "u", "f"]:
+                # maybe coerce to a sub-class
+                arr = data
             else:
-                subarr = com.asarray_tuplesafe(data, dtype=object)
-
-            # asarray_tuplesafe does not always copy underlying data,
-            # so need to make sure that this happens
-            if copy:
-                subarr = subarr.copy()
+                arr = com.asarray_tuplesafe(data, dtype=object)
 
-            if dtype is None:
-                new_data, new_dtype = _maybe_cast_data_without_dtype(subarr)
-                if new_dtype is not None:
+                if dtype is None:
+                    new_data = _maybe_cast_data_without_dtype(arr)
+                    new_dtype = new_data.dtype
                     return cls(
-                        new_data, dtype=new_dtype, copy=False, name=name, **kwargs
+                        new_data, dtype=new_dtype, copy=copy, name=name, **kwargs
                     )
 
+            klass = cls._dtype_to_subclass(arr.dtype)
+            arr = klass._ensure_array(arr, dtype, copy)
             if kwargs:
                 raise TypeError(f"Unexpected keyword arguments {repr(set(kwargs))}")
-            if subarr.ndim > 1:
-                # GH#13601, GH#20285, GH#27125
-                raise ValueError("Index data must be 1-dimensional")
-            return cls._simple_new(subarr, name)
+            return klass._simple_new(arr, name)
 
-        elif data is None or is_scalar(data):
+        elif is_scalar(data):
             raise cls._scalar_data_error(data)
         elif hasattr(data, "__array__"):
             return Index(np.asarray(data), dtype=dtype, copy=copy, name=name, **kwargs)
         else:
+
             if tupleize_cols and is_list_like(data):
                 # GH21470: convert iterable to list before determining if empty
                 if is_iterator(data):
@@ -400,6 +389,64 @@ def __new__(
             subarr = com.asarray_tuplesafe(data, dtype=object)
             return Index(subarr, dtype=dtype, copy=copy, name=name, **kwargs)
 
+    @classmethod
+    def _ensure_array(cls, data, dtype, copy: bool):
+        """
+        Ensure we have a valid array to pass to _simple_new.
+        """
+        if data.ndim > 1:
+            # GH#13601, GH#20285, GH#27125
+            raise ValueError("Index data must be 1-dimensional")
+        if copy:
+            # asarray_tuplesafe does not always copy underlying data,
+            #  so need to make sure that this happens
+            data = data.copy()
+        return data
+
+    @classmethod
+    def _dtype_to_subclass(cls, dtype: DtypeObj):
+        # Delay import for perf. https://github.com/pandas-dev/pandas/pull/31423
+
+        if isinstance(dtype, DatetimeTZDtype) or dtype == np.dtype("M8[ns]"):
+            from pandas import DatetimeIndex
+
+            return DatetimeIndex
+        elif dtype == "m8[ns]":
+            from pandas import TimedeltaIndex
+
+            return TimedeltaIndex
+        elif isinstance(dtype, CategoricalDtype):
+            from pandas import CategoricalIndex
+
+            return CategoricalIndex
+        elif isinstance(dtype, IntervalDtype):
+            from pandas import IntervalIndex
+
+            return IntervalIndex
+        elif isinstance(dtype, PeriodDtype):
+            from pandas import PeriodIndex
+
+            return PeriodIndex
+
+        elif is_float_dtype(dtype):
+            from pandas import Float64Index
+
+            return Float64Index
+        elif is_unsigned_integer_dtype(dtype):
+            from pandas import UInt64Index
+
+            return UInt64Index
+        elif is_signed_integer_dtype(dtype):
+            from pandas import Int64Index
+
+            return Int64Index
+
+        elif dtype == object:
+            # NB: assuming away MultiIndex
+            return Index
+
+        raise NotImplementedError(dtype)
+
     """
     NOTE for new Index creation:
 
@@ -6112,25 +6159,27 @@ def _maybe_cast_data_without_dtype(subarr):
         TimedeltaArray,
     )
 
+    assert subarr.dtype == object, subarr.dtype
     inferred = lib.infer_dtype(subarr, skipna=False)
 
     if inferred == "integer":
         try:
             data = _try_convert_to_int_array(subarr, False, None)
-            return data, data.dtype
+            return data
         except ValueError:
             pass
 
-        return subarr, object
+        return subarr
 
     elif inferred in ["floating", "mixed-integer-float", "integer-na"]:
         # TODO: Returns IntegerArray for integer-na case in the future
-        return subarr, np.float64
+        data = np.asarray(subarr).astype(np.float64, copy=False)
+        return data
 
     elif inferred == "interval":
         try:
             data = IntervalArray._from_sequence(subarr, copy=False)
-            return data, data.dtype
+            return data
         except ValueError:
             # GH27172: mixed closed Intervals --> object dtype
             pass
@@ -6141,7 +6190,7 @@ def _maybe_cast_data_without_dtype(subarr):
         if inferred.startswith("datetime"):
             try:
                 data = DatetimeArray._from_sequence(subarr, copy=False)
-                return data, data.dtype
+                return data
             except (ValueError, OutOfBoundsDatetime):
                 # GH 27011
                 # If we have mixed timezones, just send it
@@ -6150,15 +6199,15 @@ def _maybe_cast_data_without_dtype(subarr):
 
         elif inferred.startswith("timedelta"):
             data = TimedeltaArray._from_sequence(subarr, copy=False)
-            return data, data.dtype
+            return data
         elif inferred == "period":
             try:
                 data = PeriodArray._from_sequence(subarr)
-                return data, data.dtype
+                return data
             except IncompatibleFrequency:
                 pass
 
-    return subarr, subarr.dtype
+    return subarr
 
 
 def _try_convert_to_int_array(

pandas/core/indexes/numeric.py

@@ -46,11 +46,20 @@ class NumericIndex(Index):
     _can_hold_strings = False
 
     def __new__(cls, data=None, dtype=None, copy=False, name=None):
-        cls._validate_dtype(dtype)
         name = maybe_extract_name(name, data, cls)
 
-        # Coerce to ndarray if not already ndarray or Index
+        subarr = cls._ensure_array(data, dtype, copy)
+        return cls._simple_new(subarr, name=name)
+
+    @classmethod
+    def _ensure_array(cls, data, dtype, copy: bool):
+        """
+        Ensure we have a valid array to pass to _simple_new.
+        """
+        cls._validate_dtype(dtype)
+
         if not isinstance(data, (np.ndarray, Index)):
+            # Coerce to ndarray if not already ndarray or Index
             if is_scalar(data):
                 raise cls._scalar_data_error(data)
 
@@ -74,7 +83,7 @@ def __new__(cls, data=None, dtype=None, copy=False, name=None):
             raise ValueError("Index data must be 1-dimensional")
 
         subarr = np.asarray(subarr)
-        return cls._simple_new(subarr, name=name)
+        return subarr
 
     @classmethod
     def _validate_dtype(cls, dtype: Dtype) -> None: