Use 8-byte step to detect ASCII sequence in 64bit Windows builds

[animalize]

BPO	38252
Nosy	@pfmoore, @tjguk, @methane, @zware, @serhiy-storchaka, @zooba, @animalize, @sir-sigurd
PRs	bpo-38252: use 8-byte step to detect ASCII sequence in 64bit Windows build #16334

^{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 2020-10-18.16:52:39.004>
created_at = <Date 2019-09-22.11:50:02.093>
labels = ['interpreter-core', '3.10', 'OS-windows', 'performance']
title = 'Use 8-byte step to detect ASCII sequence in 64bit Windows builds'
updated_at = <Date 2020-10-18.16:52:39.004>
user = 'https://github.com/animalize'

bugs.python.org fields:

activity = <Date 2020-10-18.16:52:39.004>
actor = 'serhiy.storchaka'
assignee = 'none'
closed = True
closed_date = <Date 2020-10-18.16:52:39.004>
closer = 'serhiy.storchaka'
components = ['Interpreter Core', 'Windows']
creation = <Date 2019-09-22.11:50:02.093>
creator = 'malin'
dependencies = []
files = []
hgrepos = []
issue_num = 38252
keywords = ['patch']
message_count = 6.0
messages = ['352970', '352998', '353007', '353024', '378801', '378869']
nosy_count = 8.0
nosy_names = ['paul.moore', 'tim.golden', 'methane', 'zach.ware', 'serhiy.storchaka', 'steve.dower', 'malin', 'sir-sigurd']
pr_nums = ['16334']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue38252'
versions = ['Python 3.10']

[animalize]

C type long is 4-byte integer in 64-bit Windows build. [1]

But ucs1lib_find_max_char() function [2] uses SIZEOF_LONG, so it loses a little performance in 64-bit Windows build.

Below is the benchmark of using SIZEOF_SIZE_T and this change:

-   unsigned long value = *(unsigned long *) _p;
+   sizt_t value = *(sizt_t *) _p;

D:\dev\cpython\PCbuild\amd64\python.exe -m pyperf timeit -s "b=b'a'*10_000_000; f=b.decode;" "f('latin1')"

before: 5.83 ms +- 0.05 ms
after : 5.58 ms +- 0.06 ms

[1] https://stackoverflow.com/questions/384502

[2] https://github.com/python/cpython/blob/v3.8.0b4/Objects/stringlib/find_max_char.h#L9

Maybe there can be more optimizations, so I didn't prepare a PR for this.

[serhiy-storchaka]

This looks like a good idea. Do you mind to create a PR?

[animalize]

Maybe @sir-sigurd can find more optimizations.

FYI, _Py_bytes_isascii() function [1] also has similar code.
[1] https://github.com/python/cpython/blob/v3.8.0b4/Objects/bytes_methods.c#L104

[animalize]

There are 4 functions have the similar code, see PR 16334.
Just replaced the unsigned long type with size_t type, got these benchmarks.
Can this be backported to 3.8 branch?

1. bytes.isascii()

D:\dev\cpython\PCbuild\amd64\python.exe -m pyperf timeit -s "b = b'x' * 100_000_000; f = b.isascii;" "f()"

+-----------+-----------+------------------------------+
| Benchmark | isascii_a | isascii_b |
+===========+===========+==============================+
| timeit | 11.7 ms | 7.84 ms: 1.50x faster (-33%) |
+-----------+-----------+------------------------------+

2. bytes.decode('latin1')

D:\dev\cpython\PCbuild\amd64\python.exe -m pyperf timeit -s "b = b'x' * 100_000_000; f = b.decode;" "f('latin1')"

+-----------+----------+-----------------------------+
| Benchmark | latin1_a | latin1_b |
+===========+==========+=============================+
| timeit | 60.3 ms | 57.4 ms: 1.05x faster (-5%) |
+-----------+----------+-----------------------------+

3. bytes.decode('ascii')

D:\dev\cpython\PCbuild\amd64\python.exe -m pyperf timeit -s "b = b'x' * 100_000_000; f = b.decode;" "f('ascii')"

+-----------+---------+-----------------------------+
| Benchmark | ascii_a | ascii_b |
+===========+=========+=============================+
| timeit | 48.5 ms | 47.1 ms: 1.03x faster (-3%) |
+-----------+---------+-----------------------------+

4. bytes.decode('utf8')

D:\dev\cpython\PCbuild\amd64\python.exe -m pyperf timeit -s "b = b'x' * 100_000_000; f = b.decode;" "f('utf8')"

+-----------+---------+-----------------------------+
| Benchmark | utf8_a | utf8_b |
+===========+=========+=============================+
| timeit | 48.3 ms | 47.1 ms: 1.03x faster (-3%) |
+-----------+---------+-----------------------------+

[animalize]

Although the improvement is not great, it's a very hot code path.

Could you review the PR?

[serhiy-storchaka]

New changeset a0c603c by Ma Lin in branch 'master':
bpo-38252: Use 8-byte step to detect ASCII sequence in 64bit Windows build (GH-16334)
a0c603c