BUILD_SET followed by COMPARE_OP (in) can be optimized if all items are consts

[alex]

BPO	6690
Nosy	@rhettinger, @pitrou, @alex, @davidmalcolm
Files	simple-optimization-of-BUILD_SET-COMPARE_OP(IN)-to-LOAD_CONST(tuple)_COMPARE_OP(IN)-py3k.patch simple-optimization-of-BUILD_SET-COMPARE_OP(IN)-to-LOAD_CONST(frozenset)_COMPARE_OP(IN)-py3k.patch simple-optimization-of-BUILD_SET-COMPARE_OP(IN)-to-LOAD_CONST(frozenset)_COMPARE_OP(IN)-with-tests-py3k.patch optimize-BUILD_SET-v4.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2010-01-16.18:38:14.837>
created_at = <Date 2009-08-12.21:32:06.485>
labels = ['interpreter-core', 'performance']
title = 'BUILD_SET followed by COMPARE_OP (in) can be optimized if all items are consts'
updated_at = <Date 2010-01-16.18:38:14.835>
user = 'https://github.com/alex'

bugs.python.org fields:

activity = <Date 2010-01-16.18:38:14.835>
actor = 'pitrou'
assignee = 'none'
closed = True
closed_date = <Date 2010-01-16.18:38:14.837>
closer = 'pitrou'
components = ['Interpreter Core']
creation = <Date 2009-08-12.21:32:06.485>
creator = 'alex'
dependencies = []
files = ['15840', '15841', '15845', '15900']
hgrepos = []
issue_num = 6690
keywords = ['patch']
message_count = 14.0
messages = ['91506', '91614', '91626', '94324', '97640', '97650', '97651', '97653', '97654', '97671', '97856', '97858', '97859', '97894']
nosy_count = 4.0
nosy_names = ['rhettinger', 'pitrou', 'alex', 'dmalcolm']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue6690'
versions = ['Python 2.7', 'Python 3.2']

[alex]

Just like we turn BUILD_LIST; COMPARE_OP (in) into a LOAD_CONST if all
the members are consts, we can do the same for BUILD_SET, instead
turning it into a LOAD_CONST of a frozenset. The following is the
bytecode that is current produced for each datastructure.

>>> dis.dis(lambda o: o in (1,2,3))
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               3 ((1, 2, 3)) 
              6 COMPARE_OP               6 (in) 
              9 RETURN_VALUE         
>>> dis.dis(lambda o: o in [1,2,3])
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               3 ((1, 2, 3)) 
              6 COMPARE_OP               6 (in) 
              9 RETURN_VALUE         
>>> dis.dis(lambda o: o in {1,2,3})
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               0 (1) 
              6 LOAD_CONST               1 (2) 
              9 LOAD_CONST               2 (3) 
             12 BUILD_SET                3 
             15 COMPARE_OP               6 (in) 
             18 RETURN_VALUE

[pitrou]

It's certainly possible to do so, do you have a patch?

[alex]

Antoine, I hope to have some time to write a patch for this in the
coming week.

[rhettinger]

+1

[alex]

Ugh, I haven't had the time to work on this, just wanted to note that this now applies to 2.7 as well since set literals were backported.

[davidmalcolm]

Attaching a probably over-simplistic attempt at this patch, against the py3k branch.

This patch attempts to extend the replacement of
LOAD_CONST, ...., LOAD_CONST, BUILD_LIST, COMPARE_OP(in)
with
LOAD_CONST(tuple), COMPAREOP(in)

so that it also replaces:
LOAD_CONST, ...., LOAD_CONST, BUILD_SET, COMPARE_OP(in)
with
LOAD_CONST(tuple), COMPAREOP(in)

i.e. using a tuple, not a frozenset (on the grounds that the "in" conditions should at least still work); caveat: I'm relatively new to the innards of this code.

With this patch:
>>> dis.dis(lambda o: o in {1,2,3})
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               3 ((1, 2, 3)) 
              6 COMPARE_OP               6 (in) 
              9 RETURN_VALUE         
but:
>>> dis.dis(lambda o: o in {1,2,3,3,2,1})
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               3 ((1, 2, 3, 3, 2, 1)) 
              6 COMPARE_OP               6 (in) 
              9 RETURN_VALUE         

Is it worth me working on a rewrite to use a frozenset.

Hope this is helpful
Dave

[alex]

David, I think it should use frozen set since doing it this way could actually increase the time the operation takes (which is certainly not our goal!). Plus marshall.c already handles frozenset, so I don't think it's that much more work.

[davidmalcolm]

Alex: good point - thanks!

Attaching updated version of the patch (again, against py3k, likewise, I'm somewhat new to this code, so I may have missed things)

With this:
>>> dis.dis(lambda o: o in {1,2,3})
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               3 (frozenset({1, 2, 3})) 
              6 COMPARE_OP               6 (in) 
              9 RETURN_VALUE         
>>> dis.dis(lambda o: o in {1,2,3,3,2,1})
  1           0 LOAD_FAST                0 (o) 
              3 LOAD_CONST               3 (frozenset({1, 2, 3})) 
              6 COMPARE_OP               6 (in) 
              9 RETURN_VALUE         
and the tuple/list cases are again as in the original comment.

I'm in the process of running the full test suite.

[alex]

The patch looks more or less right to me (but I'm far from an expert). It needs tests in the peephole tests though.

[davidmalcolm]

Thanks for looking at the patch.

Attached is an updated version (again against py3k) which adds tests to Lib/test/test_peepholer.py, for both the new folding away of BUILD_SET, and for the pre-existing folding of BUILD_LIST (which didn't seem to have tests).

Hopefully these look good. One possible worry I had with them is with the string comparison against repr(various frozensets) for the disassembly of the bytecode: the new tests thus assume that the ordering of the repr of a frozenset is constant. Is this a reasonable assumption, or should the choice of test items be changed to ones with more robust ordering in their repr() string?

[pitrou]

No, you can't rely on the repr of a frozenset with multiple items. You should find another way of testing (if you are brave you could match the "frozenset(...)" with a regex and eval() it).

Some comments on the patch:

• there's a line or two in peephole.c which seems to use spaces for indentation; please always use tabs (for this file anyway)
• instead of self.assertTrue(X in Y), you can use self.assertIn(X, Y) (and self.assertNotIn(X, Y) for the negation)

[davidmalcolm]

Thanks for the suggestions.

Attached is a revised version of the patch.

• I believe I've fixed all tab/space issues in this version of the patch, though I may have missed some (http://www.python.org/dev/tools/ doesn't recommend an automated way of checking this).
• I've rewritten the selftests as you suggested, using re and eval
• I've rewritten the new selftests to use assertIn, assertNotIn

The existing tests don't use assertIn/assertNotIn; I'm working on a patch for that, but I'll file that as a separate bug.

[rhettinger]

Nice looking patch.

[pitrou]

The patch was committed in r77543 (py3k), thank you!