PERF: fix perf issue in tz conversions w/o affecting DST transitions

pandas/tseries/frequencies.py

@@ -599,7 +599,7 @@ def _period_group(freqstr):
 def _period_str_to_code(freqstr):
     # hack
     freqstr = _rule_aliases.get(freqstr, freqstr)
-    
+
     if freqstr not in _dont_uppercase:
         freqstr = _rule_aliases.get(freqstr.lower(), freqstr)
 
@@ -659,6 +659,25 @@ def infer_freq(index, warn=True):
 _ONE_HOUR = 60 * _ONE_MINUTE
 _ONE_DAY = 24 * _ONE_HOUR
 
+def _tz_convert_with_transitions(values, to_tz, from_tz):
+    """
+    convert i8 values from the specificed timezone to the to_tz zone, taking
+    into account DST transitions
+    """
+
+    # vectorization is slow, so tests if we can do this via the faster tz_convert
+    f = lambda x: tslib.tz_convert_single(x, to_tz, from_tz)
+
+    if len(values) > 2:
+        first_slow, last_slow = f(values[0]),f(values[-1])
+
+        first_fast, last_fast = tslib.tz_convert(np.array([values[0],values[-1]],dtype='i8'),to_tz,from_tz)
+
+        # don't cross a DST, so ok
+        if first_fast == first_slow and last_fast == last_slow:
+            return tslib.tz_convert(values,to_tz,from_tz)
+
+    return np.vectorize(f)(values)
 
 class _FrequencyInferer(object):
     """
@@ -670,10 +689,7 @@ def __init__(self, index, warn=True):
         self.values = np.asarray(index).view('i8')
 
         if index.tz is not None:
-            f = lambda x: tslib.tz_convert_single(x, 'UTC', index.tz)
-            self.values = np.vectorize(f)(self.values)
-            # This cant work, because of DST
-            # self.values = tslib.tz_convert(self.values, 'UTC', index.tz)
+            self.values = _tz_convert_with_transitions(self.values,'UTC',index.tz)
 
         self.warn = warn
 

pandas/tseries/index.py

@@ -14,7 +14,7 @@
 from pandas.compat import u
 from pandas.tseries.frequencies import (
     infer_freq, to_offset, get_period_alias,
-    Resolution, get_reso_string)
+    Resolution, get_reso_string, _tz_convert_with_transitions)
 from pandas.core.base import DatetimeIndexOpsMixin
 from pandas.tseries.offsets import DateOffset, generate_range, Tick, CDay
 from pandas.tseries.tools import parse_time_string, normalize_date
@@ -1376,7 +1376,10 @@ def __getitem__(self, key):
         else:
             if com._is_bool_indexer(key):
                 key = np.asarray(key)
-                key = lib.maybe_booleans_to_slice(key.view(np.uint8))
+                if key.all():
+                    key = slice(0,None,None)
+                else:
+                    key = lib.maybe_booleans_to_slice(key.view(np.uint8))
 
             new_offset = None
             if isinstance(key, slice):
@@ -1588,9 +1591,7 @@ def insert(self, loc, item):
             new_dates = np.concatenate((self[:loc].asi8, [item.view(np.int64)],
                                         self[loc:].asi8))
             if self.tz is not None:
-                f = lambda x: tslib.tz_convert_single(x, 'UTC', self.tz)
-                new_dates = np.vectorize(f)(new_dates)
-                # new_dates = tslib.tz_convert(new_dates, 'UTC', self.tz)
+                new_dates = _tz_convert_with_transitions(new_dates,'UTC',self.tz)
             return DatetimeIndex(new_dates, name=self.name, freq=freq, tz=self.tz)
 
         except (AttributeError, TypeError):