Significantly improve performance of `:reload` (#4457)

* bump ropey to 1.5.1-alpha

* significantly improve performance of :reload

1 files changed, 203 insertions, 45 deletions

diff --git a/helix-core/src/diff.rs b/helix-core/src/diff.rs
index 6960c679..c754da30 100644
--- a/helix-core/src/diff.rs
+++ b/helix-core/src/diff.rs
@@ -1,58 +1,195 @@
-use crate::{Rope, Transaction};
+use std::ops::Range;
+use std::time::Instant;
 
-/// Compares `old` and `new` to generate a [`Transaction`] describing
-/// the steps required to get from `old` to `new`.
-pub fn compare_ropes(old: &Rope, new: &Rope) -> Transaction {
-    // `similar` only works on contiguous data, so a `Rope` has
-    // to be temporarily converted into a `String`.
-    let old_converted = old.to_string();
-    let new_converted = new.to_string();
-
-    // A timeout is set so after 1 seconds, the algorithm will start
-    // approximating. This is especially important for big `Rope`s or
-    // `Rope`s that are extremely dissimilar to each other.
-    let mut config = similar::TextDiff::configure();
-    config.timeout(std::time::Duration::from_secs(1));
-
-    let diff = config.diff_chars(&old_converted, &new_converted);
-
-    // The current position of the change needs to be tracked to
-    // construct the `Change`s.
-    let mut pos = 0;
-    Transaction::change(
-        old,
-        diff.ops()
+use imara_diff::intern::InternedInput;
+use imara_diff::Algorithm;
+use ropey::RopeSlice;
+
+use crate::{ChangeSet, Rope, Tendril, Transaction};
+
+/// A `imara_diff::Sink` that builds a `ChangeSet` for a character diff of a hunk
+struct CharChangeSetBuilder<'a> {
+    res: &'a mut ChangeSet,
+    hunk: &'a InternedInput<char>,
+    pos: u32,
+}
+
+impl imara_diff::Sink for CharChangeSetBuilder<'_> {
+    type Out = ();
+    fn process_change(&mut self, before: Range<u32>, after: Range<u32>) {
+        self.res.retain((before.start - self.pos) as usize);
+        self.res.delete(before.len());
+        self.pos = before.end;
+
+        let res = self.hunk.after[after.start as usize..after.end as usize]
+            .iter()
+            .map(|&token| self.hunk.interner[token])
+            .collect();
+
+        self.res.insert(res);
+    }
+
+    fn finish(self) -> Self::Out {
+        self.res.retain(self.hunk.before.len() - self.pos as usize);
+    }
+}
+
+struct LineChangeSetBuilder<'a> {
+    res: ChangeSet,
+    after: RopeSlice<'a>,
+    file: &'a InternedInput<RopeSlice<'a>>,
+    current_hunk: InternedInput<char>,
+    pos: u32,
+}
+
+impl imara_diff::Sink for LineChangeSetBuilder<'_> {
+    type Out = ChangeSet;
+
+    fn process_change(&mut self, before: Range<u32>, after: Range<u32>) {
+        let len = self.file.before[self.pos as usize..before.start as usize]
             .iter()
-            .map(|op| op.as_tag_tuple())
-            .filter_map(|(tag, old_range, new_range)| {
-                // `old_pos..pos` is equivalent to `start..end` for where
-                // the change should be applied.
-                let old_pos = pos;
-                pos += old_range.end - old_range.start;
-
-                match tag {
-                    // Semantically, inserts and replacements are the same thing.
-                    similar::DiffTag::Insert | similar::DiffTag::Replace => {
-                        // This is the text from the `new` rope that should be
-                        // inserted into `old`.
-                        let text: &str = {
-                            let start = new.char_to_byte(new_range.start);
-                            let end = new.char_to_byte(new_range.end);
-                            &new_converted[start..end]
-                        };
-                        Some((old_pos, pos, Some(text.into())))
+            .map(|&it| self.file.interner[it].len_chars())
+            .sum();
+        self.res.retain(len);
+        self.pos = before.end;
+
+        // do not perform diffs on large hunks
+        let len_before = before.end - before.start;
+        let len_after = after.end - after.start;
+
+        // Pure insertions/removals do not require a character diff.
+        // Very large changes are ignored because their character diff is expensive to compute
+        // TODO adjust heuristic to detect large changes?
+        if len_before == 0
+            || len_after == 0
+            || len_after > 5 * len_before
+            || 5 * len_after < len_before && len_before > 10
+            || len_before + len_after > 200
+        {
+            let remove = self.file.before[before.start as usize..before.end as usize]
+                .iter()
+                .map(|&it| self.file.interner[it].len_chars())
+                .sum();
+            self.res.delete(remove);
+            let mut fragment = Tendril::new();
+            if len_after > 500 {
+                // copying a rope line by line is slower then copying the entire
+                // rope. Use to_string for very large changes instead..
+                if self.file.after.len() == after.end as usize {
+                    if after.start == 0 {
+                        fragment = self.after.to_string().into();
+                    } else {
+                        let start = self.after.line_to_char(after.start as usize);
+                        fragment = self.after.slice(start..).to_string().into();
                     }
-                    similar::DiffTag::Delete => Some((old_pos, pos, None)),
-                    similar::DiffTag::Equal => None,
+                } else if after.start == 0 {
+                    let end = self.after.line_to_char(after.end as usize);
+                    fragment = self.after.slice(..end).to_string().into();
+                } else {
+                    let start = self.after.line_to_char(after.start as usize);
+                    let end = self.after.line_to_char(after.end as usize);
+                    fragment = self.after.slice(start..end).to_string().into();
                 }
-            }),
-    )
+            } else {
+                for &line in &self.file.after[after.start as usize..after.end as usize] {
+                    for chunk in self.file.interner[line].chunks() {
+                        fragment.push_str(chunk)
+                    }
+                }
+            };
+            self.res.insert(fragment);
+        } else {
+            // for reasonably small hunks, generating a ChangeSet from char diff can save memory
+            // TODO use a tokenizer (word diff?) for improved performance
+            let hunk_before = self.file.before[before.start as usize..before.end as usize]
+                .iter()
+                .flat_map(|&it| self.file.interner[it].chars());
+            let hunk_after = self.file.after[after.start as usize..after.end as usize]
+                .iter()
+                .flat_map(|&it| self.file.interner[it].chars());
+            self.current_hunk.update_before(hunk_before);
+            self.current_hunk.update_after(hunk_after);
+
+            // the histogram heuristic does not work as well
+            // for characters because the same characters often reoccur
+            // use myer diff instead
+            imara_diff::diff(
+                Algorithm::Myers,
+                &self.current_hunk,
+                CharChangeSetBuilder {
+                    res: &mut self.res,
+                    hunk: &self.current_hunk,
+                    pos: 0,
+                },
+            );
+
+            self.current_hunk.clear();
+        }
+    }
+
+    fn finish(mut self) -> Self::Out {
+        let len = self.file.before[self.pos as usize..]
+            .iter()
+            .map(|&it| self.file.interner[it].len_chars())
+            .sum();
+
+        self.res.retain(len);
+        self.res
+    }
+}
+
+struct RopeLines<'a>(RopeSlice<'a>);
+
+impl<'a> imara_diff::intern::TokenSource for RopeLines<'a> {
+    type Token = RopeSlice<'a>;
+    // TODO: improve performance of lines iterator (https://github.com/cessen/ropey/issues/25)
+    type Tokenizer = ropey::iter::Lines<'a>;
+
+    fn tokenize(&self) -> Self::Tokenizer {
+        self.0.lines()
+    }
+
+    fn estimate_tokens(&self) -> u32 {
+        // we can provide a perfect estimate which is very nice for performance
+        self.0.len_lines() as u32
+    }
+}
+
+/// Compares `old` and `new` to generate a [`Transaction`] describing
+/// the steps required to get from `old` to `new`.
+pub fn compare_ropes(before: &Rope, after: &Rope) -> Transaction {
+    let start = Instant::now();
+    let res = ChangeSet::with_capacity(32);
+    let after = after.slice(..);
+    let file = InternedInput::new(RopeLines(before.slice(..)), RopeLines(after));
+    let builder = LineChangeSetBuilder {
+        res,
+        file: &file,
+        after,
+        pos: 0,
+        current_hunk: InternedInput::default(),
+    };
+
+    let res = imara_diff::diff(Algorithm::Histogram, &file, builder).into();
+
+    log::debug!(
+        "rope diff took {}s",
+        Instant::now().duration_since(start).as_secs_f64()
+    );
+    res
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
 
+    fn test_identity(a: &str, b: &str) {
+        let mut old = Rope::from(a);
+        let new = Rope::from(b);
+        compare_ropes(&old, &new).apply(&mut old);
+        assert_eq!(old, new);
+    }
+
     quickcheck::quickcheck! {
         fn test_compare_ropes(a: String, b: String) -> bool {
             let mut old = Rope::from(a);
@@ -61,4 +198,25 @@ mod tests {
             old == new
         }
     }
+
+    #[test]
+    fn equal_files() {
+        test_identity("foo", "foo");
+    }
+
+    #[test]
+    fn trailing_newline() {
+        test_identity("foo\n", "foo");
+        test_identity("foo", "foo\n");
+    }
+
+    #[test]
+    fn new_file() {
+        test_identity("", "foo");
+    }
+
+    #[test]
+    fn deleted_file() {
+        test_identity("foo", "");
+    }
 }