Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
While #177 is reported as being caused by a comment, the underlying behavior is a problem due to the newline that we generated (from a comment). The prior commit fixed that problem by preserving whitespace before the comment. That guarantees that a block will form there from the frontier before it will be expanded there via a "neighbors" method. Since empty lines are valid ruby code, it will be hidden and be safe. ## Problem setup This failure mode is not fixed by the prior commit, because the indentation is 0. To provide good results, we must make the algorithm less greedy. One heuristic/signal to follow is developer added newlines. If a developer puts a newline between code, it's more likely they're unrelated. For example: ``` port = rand(1000...9999) stub_request(:any, "localhost:#{port}") query = Cutlass::FunctionQuery.new( port: port ).call expect(WebMock).to have_requested(:post, "localhost:#{port}"). with(body: "{}") ``` This code is split into three chunks by the developer. Each are likely (but not guaranteed) to be intended to stand on their own (in terms of syntax). This behavior is good for scanning neighbors (same indent or higher) within a method, but bad for parsing neighbors across methods. ## Problem Code is expanded to capture all neighbors, and then it decreases indent level which allows it to capture surrounding scope (think moving from within the method to also capturing the `def/end` definition. Once the indentation level has been increased, we go back to scanning neighbors, but now neighbors also contain keywords. For example: ``` 1 def bark 2 3 end 4 5 def sit 6 end ``` In this case if lines 4, 5, and 6 are in a block when it tries to expand neighbors it will expand up. If it stops after line 2 or 3 it may cause problems since there's a valid kw/end pair, but the block will be checked without it. TLDR; It's good to stop scanning code after hitting a newline when you're in a method...it causes a problem scanning code between methods when everything inside of one of the methods is an empty line. In this case it grabs the end on line 3 and since the problem was an extra end, the program now compiles correctly. It incorrectly assumes that the block it captured was causing the problem. ## Extra bit of context One other technical detail is that after we've decided to stop scanning code for a new neighbor block expansion, we look around the block and grab any empty newlines. Basically adding empty newlines before of after a code block do not affect the parsing of that block. ## The fix Since we know that this problem only happens when there's a newline inside of a method and we know this particular failure mode is due to having an invalid block (capturing an extra end, but not it's keyword) we have all the metadata we need to detect this scenario and correct it. We know that the next line above our block must be code or empty (since we grabbed extra newlines). Same for code below it. We can count all the keywords and ends in the block. If they are balanced, it's likely (but not guaranteed) we formed the block correctly. If they're imbalanced, look above or below (depending on the nature of the imbalance), check to see if adding that line would balance the count. This concept of balance and "leaning" comes from work in #152 and has proven useful, but not been formally introduced into the main branch. ## Outcome Adding this extra check introduced no regressions and fixed the test case. It might be possible there's a mirror or similar problem that we're not handling. That will come out in time. It might also be possible that this causes a worse case in some code not under test. That too would come out in time. One other possible concern to adding logic in this area (which is a hot codepath), is performance. This extra count check will be performed for every block. In general the two most helpful performance strategies I've found are reducing total number of blocks (therefore reducing overall N internal iterations) and making better matches (the parser to determine if a close block is valid or not is a major bottleneck. If we can split valid code into valid blocks, then it's only evaluated by the parser once, where as invalid code must be continuously re-checked by the parser until it becomes valid, or is determined to be the cause of the core problem. This extra logic should very rarely result in a change, but when it does it should tend to produce slightly larger blocks (by one line) and more accurate blocks. Informally it seems to have no impact on performance: `` This branch: DEBUG_DISPLAY=1 bundle exec rspec spec/ --format=failures 3.01s user 1.62s system 113% cpu 4.076 total ``` ``` On main: DEBUG_DISPLAY=1 bundle exec rspec spec/ --format=failures 3.02s user 1.64s system 113% cpu 4.098 total ```
- Loading branch information
