bpo-43950: handle wide unicode characters in tracebacks

[isidentical]

Do not merge yet, only for discussion.

This PR adds support for the existing traceback machinery to work with wide unicode characters when dumping to the terminal. It uses unicodedata.east_asian_width to classify individual unicode characters.

https://bugs.python.org/issue43950

[isidentical]

CC: @pablogsal@ammaraskar

[pablogsal]

Someone should try this on Windows with "Courier New" as Terry mentioned in the issue. It will be good to check with a bunch of fonts to see what we are up against.

[isidentical]

some notes

One thing I noticed while doing the migration was that, C tokenizer doesn't match with the Python one regarding how to treat unicode symbols with multiple code points;

tokenizer: python
3xREGULAR
TokenInfo(type=3 (STRING), string="'XXX'", start=(1, 0), end=(1, 5), line="'XXX'")
3xTINY H
TokenInfo(type=3 (STRING), string="'ʰʰʰ'", start=(1, 0), end=(1, 5), line="'ʰʰʰ'")
3xEAST ASIAN
TokenInfo(type=3 (STRING), string="'该该该'", start=(1, 0), end=(1, 5), line="'该该该'")
3xEMOJIs
TokenInfo(type=3 (STRING), string="'🐍🐍🐍'", start=(1, 0), end=(1, 5), line="'🐍🐍🐍'")
tokenizer: c
3xREGULAR
TokenInfo(type=3 (STRING), string="'XXX'", start=(1, 0), end=(1, 5), line="'XXX'\n")
3xTINY H
TokenInfo(type=3 (STRING), string="'ʰʰʰ'", start=(1, 0), end=(1, 8), line="'ʰʰʰ'\n")
3xEAST ASIAN
TokenInfo(type=3 (STRING), string="'该该该'", start=(1, 0), end=(1, 11), line="'该该该'\n")
3xEMOJIs
TokenInfo(type=3 (STRING), string="'🐍🐍🐍'", start=(1, 0), end=(1, 14), line="'🐍🐍🐍'\n")

Which then directly mirrored back to AST offets, which makes stuff a bit trickier at the traceback level. We could just patch this on the traceback level though I am not really sure that 'hack' is the proper way to go. We might need to end up a similiar solution to our _PyPegen_byte_offset_to_character_offset in the AST logic too, but I am not too certain about how.

Seems to work now with my last commit

I simply applied the same functionality at the end of the specialization function, and casted the newly retrieved AST offsets back to the regular ones. There are a lot of code that needs to be cleaned away in the final version, though for the prototype at least it works (for now).

[github-actions]

This PR is stale because it has been open for 30 days with no activity.

[isidentical]

Wow, this is an old PR but @pablogsal@ammaraskar are we interested in reviving it? I think I can rebase and get it ready for 3.12.

[pablogsal]

I am. I was precisely thinking on reviving it last week so you managed to read my mind! 😁

[isidentical]

Hahahaha, perfection! I'll try to get it up to date over this week, and will ping you again for the review 💯

[pablogsal]

Hahahaha, perfection

[ammaraskar]

Sounds good, happy to re-review when you update :)

Sorry for the slow follow-ups around PEP657 stuff, I've been a little busy and inactive recently :(

[isidentical]

I've pushed the initial revision where we handle everything using the unicodedata.east_asian_width (double width on W and F labels, single on rest) as per the discussion here. Before merging this we also have to decide:

• Do we want to give users a way to opt-in/opt-out from this (considering the nature, I'd say opt-out, since I don't think anybody is going to look why behaves like this and turn this feature on)?

• Whether to do some sort of terminal support detection (non-deterministic tracebacks depending on non PYTHON* env variable configuration doesn't sound really nice but no strong opinions if we want to make this opt-In by default and add some sort of scanning)

• Is this applicable when showing the traceback in web? A lot of web frameworks use traceback.* APIs to format tracebacks and then embed them inside an HTML page on error in debug mode. Should we offer a flag in format_* APIs to allow them switch without changing it in the global Python exectuon level.

• Is this a feature or a fix? I am leaning towards a feature since this essentially changes how tracebacks are printed for certain cases and people might depend on the behavior (there are a couple of codebases where tracebacks are also part of the tests themselves in a stringized form)

I guess also looking at what other compilers are doing by default might also help us gain some insight (I recall @pablogsal mentioning rustc; maybe it might worth a shot to check out what they do to decide when to show carets).

[isidentical]

CC: @cfbolz (would also love to hear your feedback on the unicode related parts)

[cfbolz]

I'll take a look at the code!

"Amusingly" the width doesn't line up in my browser's font:

(Looks fantastic in my editor and my terminal though)

Personal opinions on some of your questions:

• imo it's a bug. for people with wide characters the position info was just confusing so far, fixing that is a bug fix.
• maybe we want a way to opt out (I have no clear opinion on that) but the new behaviour should be the default

[cfbolz]

The code looks reasonable to me.

I've been thinking about it a bit more, and it would be certainly more annoying to implement, but I am wondering whether it wouldn't be an option to use unicode chars 0x3000 (IDEOGRAPHIC SPACE) and 0xFF3E (FULLWIDTH CIRCUMFLEX ACCENT) to do the spaces/underlines under wide chars. Because even if in a font the width of two ascii spaces is not the same as a fullwidth char, the font should at least be consistent with itself and have the fullwidth space be the same width. Example:

Here are the chars:

说明说明📗a a fullwidth space
＾＾＾＾＾a fullwidth circumflex
a ascii whitespace
^^^^^^^^^^a ascii circumflex

Screenshot in my Firefox:

(in my terminal all the 'a' line up, so it doesn't matter there).

[cfbolz]

(Unfortunately it already breaks down in Chrome on my laptop, where the book emoji is even wider)

[pablogsal]

@isidentical let's push this forward. Could you rebase the PR?

[cfbolz]

yay, i'm excited for this to land :-)

[bedevere-app]

GH-111345 is a backport of this pull request to the 3.11 branch.

[bedevere-app]

GH-111346 is a backport of this pull request to the 3.12 branch.

[bedevere-app]

GH-111373 is a backport of this pull request to the 3.11 branch.