The bytecode for f-string formatting is inefficient.

[markshannon]

BPO	33092
Nosy	@pitrou, @ericvsmith, @markshannon, @serhiy-storchaka
PRs	gh-77273: Better bytecodes for f-strings #6132

^{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 = None
created_at = <Date 2018-03-17.12:33:10.371>
labels = ['interpreter-core', '3.8', 'performance']
title = 'The bytecode for f-string formatting is inefficient.'
updated_at = <Date 2021-08-26.14:31:11.408>
user = 'https://github.com/markshannon'

bugs.python.org fields:

activity = <Date 2021-08-26.14:31:11.408>
actor = 'Mark.Shannon'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2018-03-17.12:33:10.371>
creator = 'Mark.Shannon'
dependencies = []
files = []
hgrepos = []
issue_num = 33092
keywords = ['patch']
message_count = 7.0
messages = ['314000', '314001', '314004', '314312', '314317', '322379', '400348']
nosy_count = 4.0
nosy_names = ['pitrou', 'eric.smith', 'Mark.Shannon', 'serhiy.storchaka']
pr_nums = ['6132']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue33092'
versions = ['Python 3.8']

[markshannon]

f-string expressions can be formatted in four ways:
with or without a conversion
and
with or without a format specifier

Rather than have one bytecode that parses the opcode argument at runtime it would be more efficient and produce a cleaner interpreter for the compiler to produce one or two bytecode as required.
The bytecodes should be:
CONVERT_VALUE convert_fn
FORMAT_SIMPLE
FORMAT_WITH_SPEC

For simple format expressions with no conversion or format specifier,
which make up about 3/4 of all format expressions in the standard library, just the bytecode FORMAT_SIMPLE need be executed.

[ericvsmith]

Would this change really have a performance impact? Not saying it wouldn't, but if we're going to add bytecodes we should know the answer in advance.

[markshannon]

Even if doesn't speed things up by a significant amount, I would suggest that a simper interpreter with smaller, simpler bytecodes is a worthy goal in itself.

[pitrou]

I would suggest that a simper interpreter with smaller, simpler bytecodes is a worthy goal in itself.

+1 from me. Though I'm curious about performance changes as well :-)

[serhiy-storchaka]

I wouldn't say this more efficient. Instead one instruction you would need to execute two instructions.

If I implemented f-string formatting I would add four simple opcodes for str(), repr(), ascii() and format(). But Eric merged them all in the single opcode with complex argument. While this looks more complicated and less extensible, it is more efficient.

[taleinat]

Mark, since you have a working version of this, perhaps you can supply some performance benchmark results to help in making a decision?

[markshannon]

No significant change in performance https://gist.github.com/markshannon/34a780d65e69b5a573a83f3fdb0139aa

I think this merely indicates that there are little to no f-strings in the pyperformance benchmark suite.

[iritkatriel]

@markshannon are you still interested in this? should we repeat the measurements with benchmarks that have more f-strings?

[hauntsaninja]

Looks like this was implemented in #6132, thanks all!