This is an archive of the discontinued LLVM Phabricator instance.

[obsolete][clang-tidy] Attach fixit to warning, not note, in add_new_check.py example
AbandonedPublic

Authored by PiotrZSL on Sep 2 2021, 8:31 PM.

Details

Summary

Per this commit, "Note" diagnostics should not be used to provide the default
fix behavior:
https://reviews.llvm.org/rGabbe9e227ed31e5dde9bb7567bb9f0dd047689c6

With the default clang-tidy behavior being to not "--fix-notes" in the above
commit, the add_new_check.py script would generate a check that does not pass
its own generated test (before this change)

Because the generated example clang-tidy check has a clear default FixItHint,
based on the above commit, the FixItHint would be better suited by being sent to
the warning diag(...), as opposed to keeping the separate note diag(...) and
passing "--fix-notes" in the generated test

Diff Detail

Event Timeline

mattbeardsley created this revision.Sep 2 2021, 8:31 PM
mattbeardsley requested review of this revision.Sep 2 2021, 8:31 PM
Herald added a project: Restricted Project. · View Herald TranscriptSep 2 2021, 8:31 PM
Herald added a subscriber: cfe-commits. · View Herald Transcript

Uuuh.. I get the sentiment, but this change breaks the very essence of the joke that the default generated code wants to pass on. It also does not showcase that we can emit notes, not just warnings.

How about function %0 is insufficiently awesome, mark it as 'awesome_'! in the warning text? (Note how the fix and the note's text diverged already.)

And so we should come up with a joke for the Note tag.

BTW, the test wasn't checking for the note's message at all?

Uuuh.. I get the sentiment, but this change breaks the very essence of the joke that the default generated code wants to pass on. It also does not showcase that we can emit notes, not just warnings.

How about function %0 is insufficiently awesome, mark it as 'awesome_'! in the warning text? (Note how the fix and the note's text diverged already.)

And so we should come up with a joke for the Note tag.

BTW, the test wasn't checking for the note's message at all?

I'm certainly not trying to be a buzzkill on the joke :)
Yes, I did notice that the printed diagnostic text changed from its original, but I'll add some background investigation that I should have written in the original review text, sorry about laying out quite an incomplete thought process.
At least, I'm hoping I can show there's some extent of due diligence behind-the-scenes, and that I'm not just trying to change this willy-nilly!

I read/grepped/etc through several (but certainly not all) existing checks to try to identify what the "canonical" diagnostic-and-fix pattern looks like.
I was thinking that a decent goal would be that, if someone takes the files generated by add_new_check.py as their starting point and runs with it, they'll end up with check behavior that's reasonably consistent with a "typical" existing check.

Here are my observations w.r.t "note" diagnostic usage and diagnostic message wording:

  1. < ~15% of existing checks use "note" diagnostics
$ cd clang-tools-extra/clang-tidy

# 40 checks use notes
$ find . -name '*Check.cpp' | xargs grep -nl 'DiagnosticIDs::Note' | wc -l
40

# out of 265 calling diag directly
# (there are 287 total *Check.cpp - most of the remaining 22 are android/Cloexec*.cpp which call "CloexecCheck::replaceFunc", which also doesn't use DiagnosticIDs::Note)
$ find . -name '*Check.cpp' | xargs grep -nl 'diag(' | wc -l
265
  1. Out of several messages that I skimmed through arbitrarily (subject to potentially being a misrepresentative sample, but hopefully not), the general pattern seems to be to state the problem, but not the fix in the handwritten diagnostic message text. Here are some examples from the readability module:
    • readability-redundant-string-init warns "redundant string initialization", but doesn't state "remove the initialization"
    • readability-redundant-string-cstr warns "redundant call to %0", but doesn't state "remove the call"
    • readability-named-parameter warns "all parameters should be named in a function", but doesn't state "insert the commented parameter name"
    • readability-implicit-bool-conversion warns "implicit conversion bool -> %0", but doesn't state the various insertions and removals it's going to do
    • readability-braces-around-statements warns "statement should be inside braces", but doesn't state "insert statements"

The typical pattern appeared to be to state the issue, then let the FixItHint display the suggested conversion, without writing text describing that conversion explicitly.
So I aimed to adjust the add_new_check.py script so that it would produce an example that's consistent with those observed best practices from the live code, so that a new check based on add_new_check.py would look reasonably similar to surrounding checks.
My interpretation of that was to state the issue ("insufficiently awesome"), but not describe the insertion - leaving that to the FixItHint::CreateInsertion to emit as it sees fit.

Then, since a relatively small percentage of checks use Note diagnostics, my take on it was that including the Note example would lead a newcomer to unnecessarily start adding Note diagnostics.
This is anecdotal evidence obviously, but after a couple years spent (on & off) writing about 40 clang-tidy checks in my organization, I only just learned yesterday from a git bisect that I'd been incorrectly using note diagnostics! ("why did all my tests start failing after pulling from upstream!").
And I'd been using Note diagnostics that way all along because my impression from the add_new_check.py example was that streaming the FixItHints into Note diagnostics was the right thing to do for the "standard" fix in a clang-tidy check.

Last - given relatively few (~15%) checks actually use Note diagnostics at all, it didn't seem like Note diagnostics had "earned" their place in the generated example code, and could be left as a slightly more "advanced maneuver" instead
Slightly more (42, still about 15%) Check.cpp files (using similar find/grep as above methodology) use "Lexer::getSourceText," but that's not part of the add_new_check.py example.
So I guess I'm left wondering - what drives Note diagnostics to be a crucial piece to keep as a demo in add_new_check.py, as opposed to showcasing a different maneuver? Is there something else that would be more important to demonstrate than Note diagnostics, without taking away from the helpfulness of the basicexample?

Anyway - hopefully that shows a bit more thorough of a thought process than my original brief message might have made it seem. Sorry for the mountain of text though, just wanted to explain what I was thinking.
I thought this might help overall though because then we could work on identifying which key steps/conclusions that led to the diff that there might be issues with - rather than just sending the diff and leaving you to have to wonder what in the world I'm thinking!

  • Does keeping the add_new_check.py script's generated example consistent with "typical" clang-tidy checks in the surrounding codebase seem like a suitable goal?
  • Do my observations of the diagnostic reporting patterns (namely: message wording style & note usage frequency) in the codebase seem consistent with yours?
  • Does the train of thought & conclusions seem reasonable/rational to you overall? (and of course, what differences in choices/conclusions you may have! I hear you on stating "mark it as 'awesome_', but just wanted to clarify why I'd left text like that out based on what I saw when reading through how a variety of existing checks word their messages, and get your take on it with that extra context - or if I've looked at a misrepresentative set of check implementations for their diagnostic statement patterns!)

Sorry about all the words I've just asked you read, but much appreciated if you did power through it all :). Just wanted to show some semblance of due diligence in trying to have add_new_check.py produce as useful a starter-example as possible for anyone else who's trying to get started in this wonderful codebase!

Ah, and to your direct question:

BTW, the test wasn't checking for the note's message at all?"

Right, the generated test file does not ever do // CHECK-MESSAGES: :Row:Col: note: insert 'awesome' (nor does the LIT setup here enforce that notes need to be checked - it enforces that all warning and error diagnostics must be checked, but it looks like it leaves out note diagnostics)

@mattbeardsley Thank you for the explanation. (I'd wager some of your observations could even be added to some developer docs that's versioned in the repo. 😉) I'm not opposed to the changes at all. And indeed, it looks like Notes are really an advanced feature. I'm incredibly biased because I just recently finished the development of a checker in which I extensively use notes (in some test cases, there were like 8 or 9 notes attached to a warning, albeit all without a FixIt attached).

Thanks for your reply and sorry about my very sluggish reply!
I am looking into updating the docs as you suggested, and that got me looking at this doc page. Interestingly, that doc page's version of the add_new_check.py template doesn't have the Note diagnostic. So that got me digging into the history.
It seems like 2 commits are at odds with each other about what the right thing to do is with FixItHints w.r.t. Note vs Warning diagnostics.

  1. This commit says that "fix descriptions and fixes are emitted via diagnostic::Note (rather than attaching the main warning diagnostic)." - And note how that commit changes add_new_check.py - adding the Note diagnostic!
  2. This commit seems to suggest that Note diagnostics should not be used to provide the "standard" fix (and generally enforces that by --fix-notes being off by default)

It looks like nearly all existing check implementations disagree with the former, and agree with the latter.
Those two commits appear to be at odds with each other on the semantics of Note diagnostics - any opinions on how to reconcile that apparent disagreement?

My (limited) perspective is:
Since --fix-notes is off by default, it seems like FixIts on Note diagnostics are pretty well enforced to be used for secondary/uncertain suggestions, while FixIts on Warning diagnostics are for the primary fix(es).
That seems to rule out accepting the guidance of first commit w.r.t. attaching fixes to Note diagnostics given the --fix-notes flag imposes that such note-fixes are implicitly secondary to warning-fixes.

Someone else ultimately fixed this (slightly differently) last month:
https://reviews.llvm.org/D146875

Herald added a project: Restricted Project. · View Herald TranscriptMay 1 2023, 12:05 PM
PiotrZSL retitled this revision from [clang-tidy] Attach fixit to warning, not note, in add_new_check.py example to [obsolete][clang-tidy] Attach fixit to warning, not note, in add_new_check.py example.Jun 27 2023, 11:51 PM
PiotrZSL commandeered this revision.Jul 18 2023, 12:36 AM
PiotrZSL abandoned this revision.
PiotrZSL added a reviewer: mattbeardsley.

No longer needed.