diff: apply heuristics borrowed from GNU diff for "good enough"

This change adds some checks to decide that the search for the best
place to split the diffing process has gone for too long, or long
enough while finding a good chunk of matches.

They are based on similar heuristics that GNU diff applies and will
help in cases in which files are very long and have few common
sequences.

This brings comparing some large files (~36MB) that are very different
from ~1 hour to ~8 seconds, but it will still hit some pathological
cases, such as some very large cpp files I created for some benchmarking
that still take 1 minute.

  Benchmark 1: diff test-data/huge-base test-data/huge-very-different
    Time (mean ± σ):      2.790 s ±  0.005 s    [User: 2.714 s, System: 0.063 s]
    Range (min … max):    2.781 s …  2.798 s    10 runs

    Warning: Ignoring non-zero exit code.

  Benchmark 2: ./target/release/diffutils.no-heuristics diff test-data/huge-base test-data/huge-very-different
    Time (mean ± σ):     4755.084 s ± 172.607 s    [User: 4727.169 s, System: 0.330 s]
    Range (min … max):   4607.522 s … 5121.135 s    10 runs

    Warning: Ignoring non-zero exit code.

  Benchmark 3: ./target/release/diffutils diff test-data/huge-base test-data/huge-very-different
    Time (mean ± σ):      7.197 s ±  0.099 s    [User: 7.055 s, System: 0.094 s]
    Range (min … max):    7.143 s …  7.416 s    10 runs

    Warning: Ignoring non-zero exit code.
    Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.

  Summary
    diff test-data/huge-base test-data/huge-very-different ran
      2.58 ± 0.04 times faster than ./target/release/diffutils diff test-data/huge-base test-data/huge-very-different
   1704.04 ± 61.93 times faster than ./target/release/diffutils.no-heuristics diff test-data/huge-base test-data/huge-very-different

Note that the worse that should happen by heuristics causing the search
to end early is a suboptimal diff, but the diff will still be correct
and usable with patch.

Author: kov

src/engine.rs

@@ -16,6 +16,11 @@ struct Snake {
 }
 
 impl Snake {
+    fn is_good(&self) -> bool {
+        // This magic number comes from GNU diff.
+        self.length > 20
+    }
+
     fn maybe_update(&mut self, x: isize, y: isize, length: isize) {
         let length = length as usize;
         if length > self.length {
@@ -82,11 +87,26 @@ fn find_split_point<T: Clone + Debug + PartialEq + Into<Vec<u8>>>(
         x >= offset && y >= offset && x < left_length + offset && y < right_length + offset
     };
 
+    // This constant is the value used by GNU diff; using it should give us
+    // more similar diffs.
+    const HIGH_COST: isize = 200;
+
+    // This magic number was borrowed from GNU diff - apparently this is a
+    // good number for modern CPUs.
+    let too_expensive: isize = ((max_cost as f64).sqrt() as isize).max(4096);
+    info!(too_expensive = too_expensive);
+
     let mut best_snake = Snake::default();
 
     let forward_span = tracing::span!(Level::TRACE, "forward");
     let backward_span = tracing::span!(Level::TRACE, "backward");
-    'outer: for _ in 1..max_cost {
+    'outer: for c in 1..max_cost {
+        info!(c = c, snake_length = best_snake.length);
+        // The files appear to be large and too different. Go for good enough
+        if c > too_expensive {
+            break 'outer;
+        }
+
         // Forwards search
         forward_diagonals.expand_search();
         let fwd = forward_span.enter();
@@ -192,7 +212,21 @@ fn find_split_point<T: Clone + Debug + PartialEq + Into<Vec<u8>>>(
             }
         }
         drop(bwd);
+
+        if c > HIGH_COST && best_snake.is_good() {
+            info!("met criteria for high cost with good snake heuristic");
+            break 'outer;
+        }
+    }
+
+    // If we hit this condition, the search ran too long and found 0 matches.
+    // Get the best we can do as a split point - furthest diagonal.
+    if best_snake.length == 0 {
+        let (x, y) = forward_diagonals.get_furthest_progress();
+        best_snake.x = x;
+        best_snake.y = y;
     }
+
     info!(
         x = best_snake.x,
         y = best_snake.y,
@@ -355,6 +389,21 @@ impl Diagonals {
         actual >= 0 && (actual as usize) < self.data.len()
     }
 
+    fn get_furthest_progress(&self) -> (usize, usize) {
+        let (d, x) = self
+            .data
+            .iter()
+            .enumerate()
+            .filter(|(d, &x)| x - (*d as isize) >= 0)
+            .max_by_key(|(_, &x)| x)
+            .map(|(i, x)| (i as isize, *x))
+            .unwrap_or((0isize, 0isize));
+        let y = x - d;
+        debug_assert!(x >= 0);
+        debug_assert!(y >= 0);
+        (x as usize, y as usize)
+    }
+
     fn expand_search(&mut self) {
         let upper = if *self.search_range.end() == self.max_diag {
             self.search_range.end() - 1