<div dir="ltr"><div dir="ltr">On Mon, Nov 2, 2020 at 2:52 AM Haojian Wu <<a href="mailto:hokein@google.com">hokein@google.com</a>> wrote:<br></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr">Thanks for all the replies. They are all useful.<div><br></div><div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">A major goal of Clang's diagnostic experience is that if a fixit is suggested (such as a typo correction) then subsequent errors are exactly as-if the fixit had been applied. It sounds like your suggestion would go counter to that?</blockquote><div><br></div><div>Thanks. I didn't notice that this is a clang goal for recovery. At the same time, this would mean we can't tell the difference from ASTs for non-typo-correct & typo-correct code because they are exactly the same.</div></div></div></blockquote><div><br></div><div>Yep</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div>Speaking of clangd, I should have given more information on why we want to suppress secondary diagnostics:<br></div><div> - the provided case is from a bug report, and we have user complaints about it, see <a href="https://github.com/clangd/clangd/issues/547" target="_blank">https://github.com/clangd/clangd/issues/547</a></div><div> - the idea of the "first" and "secondary" diagnostics is UI-specific, most LSP clients just group all diagnostics together, which makes the "first" diagnostic less obvious to users. The tricky case here is that we have multiple diagnostics with fix-its at the *same* position -- different LSP clients have different behavior, and they are suboptimal :(</div><div> - vscode -- the quick-fix prompt-up widget just shows fix-it text, and there is no way to see which diagnostic does the fix-it fix</div><div> - neovim builtin lsp -- this is worse, the typo-correct "did you mean free" diag is emitted on the "void free()" line rathan than "if (!force)" line</div></div></blockquote><div><br></div><div>Any chance of improving the consumers here? Alternatively/also, perhaps a general workaround/solution to clangd that strips fixits if they're at the same position?</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div> - applying fix-its in interactive environments requires user confirmations, offering somewhat-speculative diagnostics nearby is a relatively worse tradeoff</div></div></blockquote><div><br></div><div>Not quite sure I follow this one - applying fix-its in a command line compiler situation requires user interaction too. You mean some editors automatically prompt for application of fixits as soon as they're generated? Rather than the user, say, right-clicking and asking to apply them?<br><br>Yeah, that would create a very different bar for fixits, I think. I'd potentially argue that that might not be a great editor default (like clang's -fixit isn't the default, but can be opted into).</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div> - recompiling code in clangd is usually fast -- we have preamble optimization, and fix-its usually don't touch preamble region</div><div><br></div><div>Agree that we should improve the typo-correction heuristics, "force/free" is a bad typo-correction. I will try to see what I can do to make it better with some reasonable amount of work.</div><div><br></div><div>Regarding confidence levels of typo-correction, how about having two thresholds T1, T2:</div><div> - 0 < match < T1: suggest and recover (the current one is EditDistance/(length + 2) <= 1/3)</div><div> - T1 < match < T2: suggest but don't recover</div><div> - match > T2: don't suggest </div><div><br></div><div>And we can pick different values for clang vs clangd, we can turn the recover off by setting T1 = T2.<br></div></div></blockquote><div><br></div><div>I'd still be pretty hesitant to encourage/support the direction of turning off typo correction recovery (though I guess you mean "by setting T1 = 0", right?) but not fundamentally problematic to demote all the typo correction to note fixits - why not all fixits, then? Is there something specifically about spelling fixits that make them different from other fixits? Perhaps today, given the not-so-great edit distance calculation, the false positive rate for typo corrections may be too high and so even at low values of T1 there are still too many suspect typo corrections compared to other more manually implemented non-typo fixits.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Fri, Oct 30, 2020 at 9:36 PM David Blaikie <<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Fri, Oct 30, 2020 at 1:31 PM Richard Smith <<a href="mailto:richard@metafoo.co.uk" target="_blank">richard@metafoo.co.uk</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr">On Fri, 30 Oct 2020 at 11:37, David Blaikie <<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>> wrote:<br></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr">On Fri, Oct 30, 2020 at 11:21 AM Richard Smith <<a href="mailto:richard@metafoo.co.uk" target="_blank">richard@metafoo.co.uk</a>> wrote:<br></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr">On Fri, 30 Oct 2020 at 10:15, David Blaikie <<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>> wrote:<br></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr">A major goal of Clang's diagnostic experience is that if a fixit is suggested (such as a typo correction) then subsequent errors are exactly as-if the fixit had been applied. It sounds like your suggestion would go counter to that?<br><br>I think there's probably a good argument to be made that not all typo corrections are high-confidence enough to merit a fixit on the error itself - if the fixit is on a note instead, the above requirement of recovery isn't applicable (so that's where we can put, say "did you mean if (a == b)" as well as "did you mean if ((a = b))" fixits on alternative notes on the general -Wparentheses warning) - so perhaps having some level of typo correction confidence would be useful to determine which kind of recovery we should do - full recovery as if the user wrote the code (with a fixit hint attached to the error itself) or "well, we're not sure but here's out best guess" where an invalid expr is created and the fixit hint is attached to a note with some wording that's a bit more vague/conveys the increased uncertainty compared to the former case.</div></div></blockquote><div><br></div><div>Right. It's an explicit goal to recover as if the typo-correction is applied, in the case where we're confident that it's right. Currently we get that confidence by checking the enclosing context in which the typo appears is valid once the correction is applied. But that's imperfect in various ways -- one of them is that the context we check is a little too narrow sometimes; another (the issue in this case) is that making the enclosing context be valid is not really sufficient to know that the typo correction actually makes sense.</div><div><br></div><div>Perhaps we could add some further heuristics to determine whether the result of typo correction seems reasonable before deciding we're confident it's correct; I could imagine, for example, annotating warnings with a "causes typo correction to be considered 'bad'" flag, in much the same way as we have a "causes SFINAE failure" flag, and using that to validate corrections -- that is, reject typo corrections not only if they would make the code invalid, but also if they would produce a warning that suggests the code is unlikely to be what the user intended. (In this case I think the warning is actually produced after we've finished correcting the typo, though that's probably not all that hard to fix.)</div></div></div></blockquote><div><br></div><div>Sounds plausible to me - what do you think about the typo correction itself being a bit more reserved about what constitutes a recoverable typo correction? If the edit distance is too far maybe we shouldn't be suggesting it, or should be suggesting it at lower priority (like force V free below - I mean, I appreciate the suggestion if that's the nearest thing, but even if it did make the code compile without any further warnings, I'm not sure it's a sufficiently good guess to recover with it?)</div></div></div></blockquote><div><br></div><div>I don't think we have the edit distance computation right yet. force -> free looks to the edit distance algorithm like two errors (each with a cost of 1): a spurious 'o' and a 'c' where an 'e' was meant (falling below the threshold of (length+2)/3 == 2), but to a human they look like completely different words, whereas getDist -> getDistance looks like 4 errors, which exceeds the threshold of (7+2)/3 == 3, but to a human would look like very-likely-correct suggestion.</div><div><br></div><div>We'd probably do a lot better if we treated a run of length N of consecutive additions / removals as having a lower cost than N independent additions / removals. Similarly, adjacent transposed characters should have a lower cost than a removal plus an addition / two replacements (which is how it's currently weighted) -- and should probably have a lower cost than a single addition or removal. And perhaps a doubled letter should have a lower cost than a general spurious letter (but maybe I only think that because the keyboard on my laptop is misbehaving!).</div><div><br></div><div>There are other obvious-to-a-human mistakes we make as part of the edit distance algorithm (for example, suggesting that `get` might be a typo for `set` or vice versa). There might be broadly-applicable rules we could use to detect those; for example, we could divide identifiers into "words" by splitting on underscores and lowercase -> uppercase transitions, and never treat a word as containing a typo if we've seen that word elsewhere in the compilation -- so we wouldn't suggest that "getFoo" is a typo for "setFoo" if we've seen the word "get" appearing in some identifier already, but we would treat "aetFoo" as a typo for "setFoo". We could in principle get really smart and notice that (for example) "get" and "set" mean different things (because we've seen "getFoo" and "setFoo" overloaded) but that "get" and "is" aren't known to mean different things (because we've never seen "getFoo" and "isFoo" overloaded), and so decide that "getFoo" is a typo for "isFoo" not "setFoo". But that's probably excessively complicated.</div><div><br></div><div>There are also heuristics we could apply based on keyboard layouts -- spurious letters are more likely if they're on keycaps that are adjacent to the next or previous letter, and replacements are similarly more likely if they're on adjacent keycaps -- but in general we don't know what keyboard layout the user is using.</div></div></div></blockquote><div><br></div><div>Ah, fair points all - yeah, maybe the higher prioritiy/bigger bang-for-buck today will be improvements to the edit distance algorithm itself, before we get to the point of benefiting from the sort of nuanced separation/classification of relative certainty.<br></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>But even with the best edit distance algorithm, I think you're right that we can't set a hard cutoff and say "anything better than this that's valid when substituted into the local context is definitely right; anything worse than this is definitely wrong", and having a middle ground of "we're not sure this correction is good enough so we're not going to use it for error recovery / fix-its" might make sense.<br></div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>- Dave</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Fri, Oct 30, 2020 at 7:34 AM Haojian Wu via cfe-dev <<a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div>Hello folks, </div><div><br></div><div>Given the following case:</div><div><br></div><div>void free();<br></div><div>void test() {</div><div> if (!force) {} // diagnostic 1: use of undeclared identifier 'force'; did you mean 'free'?</div><div> // diagnostic 2: warning: address of function 'free' will always evaluate to 'true'</div><div>}</div><div><br></div><div>The secondary diagnostic seems to be bogus, and it doesn't reflect the written source code, which can easily cause confusions. My idea is to use a dependent RecoveryExpr (which wraps the typo-correct AST node) to suppress all secondary diagnostics.</div><div><br></div><div>I have a prototype at <a href="https://reviews.llvm.org/D90459" target="_blank">https://reviews.llvm.org/D90459</a>. I see some improvements, but there are some regressions as well:</div><div><br></div><div>Improvements</div><div>- the resulting AST look better because the error is visible in the AST (with RecoveryExpr node)</div><div>- we emit more typo corrections for more cases, see <a href="https://reviews.llvm.org/differential/changeset/?ref=2240247" target="_blank">[1]</a>, <a href="https://reviews.llvm.org/differential/changeset/?ref=2240248" target="_blank">[2]</a></div><div><br></div><div>Regressions</div><div>- recursive/nested typo corrections, e.g. `TypoX.TypoY;`, we emit just 1 typo-correction while the old behavior emits two, see <a href="https://reviews.llvm.org/differential/changeset/?ref=2240254" target="_blank">[1]</a></div><div></div><div>- ambiguous typos, when there are ambiguous typo candidates (they have the same edit distance), the old one seems to perform better in some cases, see<a href="https://reviews.llvm.org/differential/changeset/?ref=2240246" target="_blank"> [1]</a></div><div>- other misc regressions, I think we could fix them</div><div><br></div><div>The secondary diagnostics are not wrong from the AST perspective, but they seem to be unnecessary. In clangd, we'd like to suppress all secondary diagnostics, but I'm not sure this is a right choice for clang.</div></div></blockquote><div><br></div><div>That would seem unfortunate to me - clang works pretty hard on diagnostic recovery so users can see/act on multiple diagnostics in one pass. Though I realize that model is a bit different if you're dealing with an editor that's recompiling after every textual change - is that always the case for clangd? I think it might still be useful to see more than one error in an IDE/editor's error list, and certainly if I were dealing with some code that's slow to compile or an editor that chooses to do less fine-grained recompiles.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div><br></div><div>What do people think?</div><div><br></div><div><br></div><div>Thanks,</div><div>Haojian</div><div><br></div></div>
_______________________________________________<br>
cfe-dev mailing list<br>
<a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a><br>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev</a><br>
</blockquote></div></div>
</blockquote></div></div>
</blockquote></div></div>
</blockquote></div></div>
</blockquote></div></div>
</blockquote></div>
</blockquote></div></div>