gh-116381: Remove bad specializations, add fail stats #116464
Conversation
Fidget-Spinner
requested review from
markshannon,
gvanrossum and
brandtbucher
as code owners
Fidget-Spinner
changed the title
Python/specialize.c
Outdated
|
|
||
| #ifdef Py_STATS |
There was a problem hiding this comment.
Can you move this into its own function, like load_attr_fail_kind, then this becomes SPECIALIZATION_FAIL(CONTAINS_OP, contains_fail_kind(value))
Python/specialize.c
Outdated
|
|
||
| #ifdef Py_STATS | ||
| if (PyUnicode_CheckExact(value)) { | ||
| SPECIALIZATION_FAIL(TO_BOOL, SPEC_FAIL_CONTAINS_OP_STR); |
There was a problem hiding this comment.
TO_BOOL -> CONTAINS_OP and a few more times below
Python/bytecodes.c
Outdated
| inst(CONTAINS_OP_STR, (unused/1, left, right -- b)) { | ||
| DEOPT_IF(!PyUnicode_CheckExact(right)); | ||
| int res = PyUnicode_Contains(right, left); | ||
| inst(CONTAINS_OP_MAPPINGPROXY, (unused/1, left, right -- b)) { |
There was a problem hiding this comment.
I'm surprised that this is worthwhile.
OOI, what are the numbers for this?
There was a problem hiding this comment.
It might not be. I might just remove it.
On python -m test_typing test_re test_dis test_zlib., I get:
frozenset:278
tuple:261
dict:256
list:225
set:150
str:143
mappingproxy:62
dict_values:9
EnumDict:5
collections.OrderedDict:5
collections.defaultdict:4
_Environ:3
There was a problem hiding this comment.
Note: I'm not specializing for tuple or list because it's dominated by searching up through the collection anyways so it's not worth it. The rest are O(1) so maybe worth to cut the constant factor.
|
When you're done making the requested changes, leave the comment: |
|
I'll leave this between the two of you. (TBH I didn't mean for you to merge gh-116385, I should have been clearer -- I was signing off but expected Mark to have something to say and was hoping for a benchmark.) |
…-116464) * Remove bad specializations, add fail stats
…-116464) * Remove bad specializations, add fail stats
Partially reverts some of the bad specializations in #116385. Stats for what are "bad" are from https://github.com/faster-cpython/benchmarking-public/tree/main/results/bm-20240307-3.13.0a4%2B-72dbea2.
CONTAINS_OP#116381