Improve speed of stdlib functions by replacing `re` uses

[donBarbos]

We can often find the module re in the standard library modules but it can be replaced (if it is possible). I don't suggest removing it everywhere, there are places where its use is appropriate, but there are also places where it is an unnecessary solution and leads to unpleasant consequences (they can be found below)

Cons of regular expressions and reasons to replace regular expressions with functions and methods:

1. We spend time to compile re pattern (one time, but anyway we spend it)
2. In most cases simple string methods are faster (according to my benchmarks about 2x)
3. We can remove import re which will affect import time
4. Additionally: I think for those who don't know regular expressions, the code is more difficult to read and therefore difficult to maintain.

Important

For those who want to work on the issue, please:

• Read https://devguide.python.org/getting-started/pull-request-lifecycle/ before anything else.
• Select one function to improve. It's easier to review and possibly backport.
• Always report benchmarks using pyperf, hyperfine, and tuna together with -X importtime to compare import times and execution time.
• Open a pull request with the following title: gh-130167: Improve speed of module.function by replacing re

Linked PRs

• gh-130167: Improve speed of difflib.IS_LINE_JUNK by replacing re #130170
• gh-130167: Improve speed of inspect.formatannotation by replacing re #130242
• gh-130167: Improve speed of ftplib.parse150 by replacing re #130243

[terryjreedy]

1. RE patterns are compiled once and then cached. So less relevant the more times a pattern is used.
2. Please list RE patterns, replacements, and RE timings that do and do not include compile time. This will indicate max and min time saved, depending on reuses. Should a table be added to Regular Expression HOWTO, Use String Methods. Are there string methods other than replace and translate to consider?
3. This is known, and there have been cases where we delay re imports until needed, if ever, to reduce startup time. But re import cannot be removed entirely unless all uses are removed.
4. Core developers should know regular expressions, though the point applies to some other contributors.

Unless you find and list examples of useful replacements in an stdlib module, this is not properly an issue. If you have or do, make a PR. Or suggest a revision of the HOWTO section.

[donBarbos]

@terryjreedy thank you for your comment and I know about the counterarguments you brought up but it doesn't change the increase in speed and the fact that we can get rid of unnecessary imports (sometimes or often, I don't know the statistics yet :)

I've only submitted one PR so far but I'm open to finding more examples since we use re in the stdlib all the time

[picnixz]

I removed the paragraph for new contributions as this issue is not necessary something we want to address immediately. Before opening a PR, benchmarks must be given and be convincing. If we lose the readability at the cost of an improved import time, I'm not sure it's worth unless we're like 5 or 6 times faster and if it's an important module that is usually imported and that re is not usually imported.

[donBarbos]

If we lose the readability at the cost of an improved import time, I'm not sure it's worth unless we're like 5 or 6 times faster and if it's an important module that is usually imported and that re is not usually imported.

Usually, i think, the readability of code without regular expressions is better. And from a performance standpoint, I think the speed of the function is the main improvement, although import time is also important.

you can find import re very often and only one use

[picnixz]

If you can remove re with simpler alternatives, then I'm fine. What I don't want to see is essentially PRs that would move lots of re imports to local ones if they hurt readability and don't improve start up time by much.

Note that it's also important to possibly plan for future extensions of the regex usage. Like getting the matched substring could be useful later. Now, I'm not against improving the various re usages whenever possible but benchmarks must always be given (with sufficient coverage).

[picnixz]

I've restored your guidelines but added the requirements of having benchmarks

[picnixz]

Considering I wrote #128983, I think we can treat it as a legit improvement. But let's only focus on changing re usages, not changing the imports or something else.