Repository navigation
In IDLE colorizer, replace re with tokenizer #140347
Description
Activity
Starting from the bottom.
In the IDLE Issues project, colorizer issues are in Highlights section. Does this fix any, or appear to make fixes easier?
- IDLE needs syntax highlighting for f-strings #73473 is fixed.
- IDLE: Wrong highlighting when previous line ends with backslash #140334 is also fixed.
The remainder of the issues in the project are not directly related to the actual colorizer logic, and there should be no change in how they would be implemented.
IDLE colorizes 10000 lines editor contents as well as single interactive lines. Will the PR colorizer do the same, with similar speed?
The PR should be faster than the regexes, although I have not yet done proper benchmarking.
The syntax categories in the REPL colorizer are not the same as IDLE's (listed on the Highlights settings page). What are they?
They can be modified to match, and in the PR it does.
The last one requires more research, for which I do not have time right now.
My main issue with the implementation is that the old colorizer would continue on errornous syntax, whereas tokenizing may fail. The repl colorizer has a fallback, but it does not always succeed. Trying to complicate the logic more is not worth it IMO, I have meddled with it a little and introduced a second line-by-line attempt, though that also causes problems (I will try fix these, the main issue with line-by-line is context, e.g. are we in a triple quoted string, is lost).
I think we would need to accept that on errornous syntax, there sometimes may be no highlighting at all.Mostly but it fails on cases like this:
That is the only bug left with the regex colorizer and it cannot be fixed without very complicated python logic or recursive regexes (which python's
redoesn't provide as far as I know).
The PR should be faster than the regexes
It will probably be much faster but unfortunately it won't really speed up IDLE's colorization. IDLE spends ~75% of the time adding the tags to the
tkinter.Textwidget after they have been computed.tkinteris a major bottleneck but can be sped upAnother case, #151471, would (I presume, since it works correctly in the repl) have been avoided if using a tokenizer.
I built a Stage-1 spike: swap IDLE's
ColorDelegator._add_tags_in_sectionto get spans from_pyrepl.utils.gen_colors, keeping IDLE's incremental engine (SYNC-boundedrecolorize_main). Findings:The glue is small. Only that one method changes. IDLE's
SYNCpoint (a newline outside strings) is exactlygen_colors's "valid start of a code block" requirement, so the incremental machinery stays as-is. Categories are compatible — pyrepl is a superset (it colors builtins too); onlynumber/opare extra. On valid code the spike is strictly more correct:match/case/typesoft keywords and f-strings the regex can't do.The real blocker is error recovery.
gen_colorsonly recovers from unterminated strings, so any otherTokenErrordrops everything after it:from _pyrepl.utils import gen_colors src = "ur'x'\nmatch point:\n case 1: pass\n" # ur'x' is an invalid prefix print(list(gen_colors(src))) # -> [] — the valid match/case below is lost too
IDLE amplifies this:
recolorize_mainaccumulates large sections, so one bad token can blank a dozen lines, where the old regex tolerated it.Where the fix belongs: in the shared
gen_colors— onTokenError, restart tokenizing at the next line (generalizing the string recovery) — so PyREPL and IDLE both benefit instead of IDLE reintroducing regex logic. Open UX question for IDLE: on an un-tokenizable line, prefer no highlight, a regex fallback, or skip-and-continue?Happy to turn the spike into a draft PR if that direction sounds right.
What is the difference between 'no highlight' and 'skip-and-continue'?
On entry of
ur"or'ru"into IDLE, the validr"oru"is highlighted and the invalid prefix not, making the error immediately visible. In REPL, there is no highlight and therefore no contrast, though user could note the string highlight absence. Interactively, a newline will likely follow soon, generating a SyntaxError with message. IDLE also used the colorizer in the editor, and a compile may be a long time away.Unlike REPL, IDLE Shell/Editors have a status bar, where a message like 'line has error' could be displayed.
How can tokenizing restart on the next line in the middle of a multiline string?
It seems to me that colorizing needs a more an error-tolerant tokenizing than compiling does. Perhaps it should be 2-stage. The first stage detects code, string (or perhaps 1-line and n-line strings and f-string variants) , and comment blocks. IDLE could use a simplified RE for this. The second stage only applies to code blocks, and IDLE could use gen_colors for that.
The current IDLE colorizer was designed when the python grammar was restricted to a context-free subtype. Parsing with RE's was sufficient or closely so to pick out substrings to be color tagged. The current grammar's sometime context-dependence make this more difficult or even impossible.
A few versions ago, the new C tokenizer used to compile Python code was exposed as a function in the tokenizer module. It replaced a python-coded tokenizer that did not always match the old C tokenizer and that must have been much slower. I assume that is was not seen as suitable for IDLE colorizer. However, the current exposed C tokenizer is being used to power the new PyREPL colorizer. The initial (draft) patch copies portions of that colorizer.
My main concerns are compatibility and speed. Some initial questions:
topic-repl.) The code is private and undocumented, I presume intentionally, and I do not expect that issues only relevant to IDLE would be welcome.Linked PRs