Merge pull request #10381 from eric-wieser/fix-segfault

BUG/ENH: Improve output for structured non-void types

3 files changed, 48 insertions, 15 deletions

diff --git a/numpy/core/arrayprint.py b/numpy/core/arrayprint.py
index 0fdd6fc7f..84943cafc 100644
--- a/numpy/core/arrayprint.py
+++ b/numpy/core/arrayprint.py
@@ -402,9 +402,6 @@ def _get_format_function(data, **options):
     find the right formatting function for the dtype_
     """
     dtype_ = data.dtype
-    if dtype_.fields is not None:
-        return StructureFormat.from_data(data, **options)
-
     dtypeobj = dtype_.type
     formatdict = _get_formatdict(data, **options)
     if issubclass(dtypeobj, _nt.bool_):
@@ -431,7 +428,10 @@ def _get_format_function(data, **options):
     elif issubclass(dtypeobj, _nt.object_):
         return formatdict['object']()
     elif issubclass(dtypeobj, _nt.void):
-        return formatdict['void']()
+        if dtype_.names is not None:
+            return StructuredVoidFormat.from_data(data, **options)
+        else:
+            return formatdict['void']()
     else:
         return formatdict['numpystr']()
 
@@ -1233,14 +1233,21 @@ class SubArrayFormat(object):
         return "[" + ", ".join(self.__call__(a) for a in arr) + "]"
 
 
-class StructureFormat(object):
+class StructuredVoidFormat(object):
+    """
+    Formatter for structured np.void objects.
+
+    This does not work on structured alias types like np.dtype(('i4', 'i2,i2')),
+    as alias scalars lose their field information, and the implementation
+    relies upon np.void.__getitem__.
+    """
     def __init__(self, format_functions):
         self.format_functions = format_functions
 
     @classmethod
     def from_data(cls, data, **options):
         """
-        This is a second way to initialize StructureFormat, using the raw data
+        This is a second way to initialize StructuredVoidFormat, using the raw data
         as input. Added to avoid changing the signature of __init__.
         """
         format_functions = []
@@ -1261,13 +1268,24 @@ class StructureFormat(object):
         else:
             return "({})".format(", ".join(str_fields))
 
+
+# for backwards compatibility
+class StructureFormat(StructuredVoidFormat):
+    def __init__(self, *args, **kwargs):
+        # NumPy 1.14, 2018-02-14
+        warnings.warn(
+            "StructureFormat has been replaced by StructuredVoidFormat",
+            DeprecationWarning, stacklevel=2)
+        super(StructureFormat, self).__init__(*args, **kwargs)
+
+
 def _void_scalar_repr(x):
     """
     Implements the repr for structured-void scalars. It is called from the
     scalartypes.c.src code, and is placed here because it uses the elementwise
     formatters defined above.
     """
-    return StructureFormat.from_data(array(x), **_format_options)(x)
+    return StructuredVoidFormat.from_data(array(x), **_format_options)(x)
 
 
 _typelessdata = [int_, float_, complex_, bool_]
@@ -1305,6 +1323,11 @@ def dtype_is_implied(dtype):
     dtype = np.dtype(dtype)
     if _format_options['legacy'] == '1.13' and dtype.type == bool_:
         return False
+
+    # not just void types can be structured, and names are not part of the repr
+    if dtype.names is not None:
+        return False
+
     return dtype.type in _typelessdata
 
 
@@ -1317,12 +1340,12 @@ def dtype_short_repr(dtype):
     >>> from numpy import *
     >>> assert eval(dtype_short_repr(dt)) == dt
     """
-    # handle these separately so they don't give garbage like str256
-    if issubclass(dtype.type, flexible):
-        if dtype.names is not None:
-            return "%s" % str(dtype)
-        else:
-            return "'%s'" % str(dtype)
+    if dtype.names is not None:
+        # structured dtypes give a list or tuple repr
+        return str(dtype)
+    elif issubclass(dtype.type, flexible):
+        # handle these separately so they don't give garbage like str256
+        return "'%s'" % str(dtype)
 
     typename = dtype.name
     # quote typenames which can't be represented as python variable names
diff --git a/numpy/core/src/multiarray/descriptor.c b/numpy/core/src/multiarray/descriptor.c
index 80161b71c..897155238 100644
--- a/numpy/core/src/multiarray/descriptor.c
+++ b/numpy/core/src/multiarray/descriptor.c
@@ -3340,8 +3340,8 @@ arraydescr_struct_str(PyArray_Descr *dtype, int includealignflag)
         sub = arraydescr_struct_dict_str(dtype, includealignflag);
     }
 
-    /* If the data type has a non-void (subclassed) type, show it */
-    if (dtype->type_num == NPY_VOID && dtype->typeobj != &PyVoidArrType_Type) {
+    /* If the data type isn't the default, void, show it */
+    if (dtype->typeobj != &PyVoidArrType_Type) {
         /*
          * Note: We cannot get the type name from dtype->typeobj->tp_name
          * because its value depends on whether the type is dynamically or
diff --git a/numpy/core/tests/test_multiarray.py b/numpy/core/tests/test_multiarray.py
index e1fcf947e..3b92a0610 100644
--- a/numpy/core/tests/test_multiarray.py
+++ b/numpy/core/tests/test_multiarray.py
@@ -478,6 +478,16 @@ class TestDtypedescr(object):
         assert_(np.dtype('<i4') != np.dtype('>i4'))
         assert_(np.dtype([('a', '<i4')]) != np.dtype([('a', '>i4')]))
 
+    def test_structured_non_void(self):
+        fields = [('a', '<i2'), ('b', '<i2')]
+        dt_int = np.dtype(('i4', fields))
+        assert_equal(str(dt_int), "(numpy.int32, [('a', '<i2'), ('b', '<i2')])")
+
+        # gh-9821
+        arr_int = np.zeros(4, dt_int)
+        assert_equal(repr(arr_int),
+            "array([0, 0, 0, 0], dtype=(numpy.int32, [('a', '<i2'), ('b', '<i2')]))")
+
 
 class TestZeroRank(object):
     def setup(self):