Accelerate 'string' % (value, ...) by using formatted string literals

[serhiy-storchaka]

BPO	28307
Nosy	@vstinner, @taleinat, @ericvsmith, @markshannon, @serhiy-storchaka, @ztane, @brandtbucher
PRs	python/cpython#5012 python/cpython#26160 python/cpython#26318
Dependencies	bpo-11549: Build-out an AST optimizer, moving some functionality out of the peephole optimizer

^{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: ```python assignee = 'https://github.com/serhiy-storchaka' closed_at = None created_at = labels = ['interpreter-core', '3.11', 'performance'] title = "Accelerate 'string' % (value, ...) by using formatted string literals" updated_at = user = 'https://github.com/serhiy-storchaka' ``` bugs.python.org fields: ```python activity = actor = 'vstinner' assignee = 'serhiy.storchaka' closed = False closed_date = None closer = None components = ['Interpreter Core'] creation = creator = 'serhiy.storchaka' dependencies = ['11549'] files = [] hgrepos = [] issue_num = 28307 keywords = ['patch'] message_count = 9.0 messages = ['277688', '277694', '277700', '277702', '277703', '309049', '324795', '393740', '402436'] nosy_count = 7.0 nosy_names = ['vstinner', 'taleinat', 'eric.smith', 'Mark.Shannon', 'serhiy.storchaka', 'ztane', 'brandtbucher'] pr_nums = ['5012', '26160', '26318'] priority = 'normal' resolution = None stage = 'patch review' status = 'open' superseder = None type = 'performance' url = 'https://bugs.python.org/issue28307' versions = ['Python 3.11'] ```

[serhiy-storchaka]

For now using formatted string literals (PEP-498) is the fastest way of formatting strings.

$ ./python -m perf timeit -s 'k = "foo"; v = "bar"' -- '"%s = %r" % (k, v)'
Median +- std dev: 2.27 us +- 0.20 us

$ ./python -m perf timeit -s 'k = "foo"; v = "bar"' -- 'f"{k!s} = {v!r}"'
Median +- std dev: 1.09 us +- 0.08 us

The compiler could translate C-style formatting with literal format string to the equivalent formatted string literal. The code '%s = %r' % (k, v) could be translated to

    t1 = k; t2 = v; f'{t1!r} = {t2!s}'; del t1, t2

or even simpler if k and v are initialized local variables.

$ ./python -m perf timeit -s 'k = "foo"; v = "bar"' -- 't1 = k; t2 = v; f"{t1!s} = {t2!r}"; del t1, t2'
Median +- std dev: 1.22 us +- 0.05 us

This is not easy issue and needs first implementing the AST optimizer.

[ericvsmith]

There isn't a direct mapping between %-formatting and __format__ format specifiers. Off the top of my head, I can think of at least one difference:

>>> '%i' % 3
'3'
>>> '{:i}'.format(3)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: Unknown format code 'i' for object of type 'int'

So you'll need to be careful with edge cases like this.

Also, for all usages of %s, remember to call str() (or add !s):

>>> '%s' % 1
'1'
>>> f'{1:s}'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: Unknown format code 's' for object of type 'int'
>>> f'{1!s:s}'
'1'

Although that also reminds me of this default alignment difference:
>>> x=0
>>> '%2s' % x
' 0'
>>> f'{x!s:2s}'
'0 '
>>> f'{x!s:>2s}'
' 0'

So, in general, the mapping will be difficult. On the other hand, if you can do it, and provide a function that maps between %-formatting codes and __format__ codes, then that might be a generally useful tool.

[serhiy-storchaka]

'%s' % x should be translated to f'{x!s}', not to f'{x:s}'. Only %s, %r and %a can be supported. Formatting with %i should left untranslated. Or maybe translate '%r: %i' % (a, x) to f'{a!r}: {"%i" % x}'.

It is possible also to introduce special opcodes that converts argument to exact int or float. Then '%06i' % x could be translated to f'{__exact_int__(x):06}'.

[ff59cd45-ebe3-4b3e-9696-65dc59a38b8c]

Serhiy, you actually did make a mistake above; '%s' % x cannot be rewritten as f'{x!s}', only '%s' % (x,) can be optimized...

(just try with x = 1, 2)

[serhiy-storchaka]

Thanks for the correction Antti. Yes, this is what I initially meant. This optimization is applicable only if the left argument of % is a literal string and the right argument is a tuple expression. Saying about '%s' % x I meant a component of the tuple.

[serhiy-storchaka]

PR 5012 implements transformation simple format strings containing only %s, %r and %a into f-strings.

[taleinat]

I'm +1 on this optimization.

[serhiy-storchaka]

PR 26160 adds support of %d, %i, %u, %o, %x, %X, %f, %e, %g, %F, %E, %G.

What is not supported:

• Formatting with a single value not wrapped into a 1-tuple (like in "%d bytes" % size). The behavior is non-trivial, it needs checking whether the value is a tuple or a mapping. It would require adding more opcodes and AST nodes and generating complex bytecode.
• Mapping keys (like in "%(name)s"). They are rarely used, and if used, it is not performance critical.
• Precision for integer formatting. Precision is not supported in new-style formatting of integers, and it is not trivial to reproduce this behavior.
• Variable width and precision (like in "%.\s"). It is possible to support them, but the code would be pretty complex, and the benefit is small, because this feature is rarely used and is not performance critical.
• Format code %c. It is relatively rarely used.
• Length modifiers "h", "l" and "L" (like in "%ld"). They ignored in Python and I did not see them in real code. While supporting them is easy, it would requires adding more than one line of code, it is not worth it.

[vstinner]

commit a0bd9e9c11f5f52c7ddd19144c8230da016b53c6 Author: Serhiy Storchaka \storchaka@gmail.com\ Date: Sat May 8 22:33:10 2021 +0300

bpo-28307: Convert simple C-style formatting with literal format into f-string. (GH-5012)

C-style formatting with literal format containing only format codes
%s, %r and %a (with optional width, precision and alignment)
will be converted to an equivalent f-string expression.

It can speed up formatting more than 2 times by eliminating
runtime parsing of the format string and creating temporary tuple.

commit 8b010673185d36d13e69e5bf7d902a0b3fa63051 Author: Serhiy Storchaka \storchaka@gmail.com\ Date: Sun May 23 19:06:48 2021 +0300

bpo-28307: Tests and fixes for optimization of C-style formatting (GH-26318)

Fix errors:
* "%10.s" should be equal to "%10.0s", not "%10s".
* Tuples with starred expressions caused a SyntaxError.