[issue38045] enum.Flag instance creation is slow
Antony Lee
report at bugs.python.org
Fri Sep 6 14:08:17 EDT 2019
Antony Lee <anntzer.lee at gmail.com> added the comment:
Actually, microoptimizing _power_of_two is a red herring and the problem is the quadratic complexity of _decompose (_value2member_map_ becomes bigger and bigger at each iteration). Replacing the implementation of _decompose by the following fixes the performance issue (the original snippet executes in ~150ms, which may be even more optimizable but already a 30x improvement :)), and still passes test_enum.py (from cpython 3.7.4):
def _decompose(flag, value):
"""Extract all members from the value."""
# _decompose is only called if the value is not named
not_covered = value
members = []
for member in flag:
member_value = member.value
if member_value and member_value & value == member_value:
members.append(member)
not_covered &= ~member_value
if issubclass(flag, IntFlag) and value > 0:
while not_covered:
new_value = 2 ** _high_bit(not_covered)
if new_value not in flag._value2member_map_:
# construct singleton pseudo-members
pseudo_member = int.__new__(flag, value)
pseudo_member._name_ = None
pseudo_member._value_ = value
flag._value2member_map_.setdefault(value, pseudo_member)
members.append(flag(new_value))
not_covered &= ~new_value
if not members and value in flag._value2member_map_:
members.append(flag._value2member_map_[value])
members.sort(key=lambda m: m._value_, reverse=True)
if len(members) > 1 and members[0].value == value:
# we have the breakdown, don't need the value member itself
members.pop(0)
return members, not_covered
Checking for issubclass(flag, IntFlag) is a bit ugly but that's because Flag and IntFlag really need two separate implementations of _decompose -- the former wants to error out on uncovered values, the latter should create them on-the-fly.
----------
_______________________________________
Python tracker <report at bugs.python.org>
<https://bugs.python.org/issue38045>
_______________________________________
More information about the Python-bugs-list
mailing list