diff --git a/libbz2-rs-sys/src/blocksort.rs b/libbz2-rs-sys/src/blocksort.rs
index 450e4a01a..67a672c6c 100644
--- a/libbz2-rs-sys/src/blocksort.rs
+++ b/libbz2-rs-sys/src/blocksort.rs
@@ -390,209 +390,57 @@ fn mainGtU(
     nblock: u32,
     budget: &mut i32,
 ) -> bool {
-    let mut k: i32;
-    let mut c1: u8;
-    let mut c2: u8;
-    let mut s1: u16;
-    let mut s2: u16;
-
     debug_assert_ne!(i1, i2, "mainGtU");
 
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    c1 = block[i1 as usize];
-    c2 = block[i2 as usize];
-    if c1 != c2 {
-        return c1 > c2;
-    }
-    i1 = i1.wrapping_add(1);
-    i2 = i2.wrapping_add(1);
-    k = nblock.wrapping_add(8 as c_int as c_uint) as i32;
-    loop {
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
-        if c1 != c2 {
-            return c1 > c2;
-        }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
-        }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
-        c1 = block[i1 as usize];
-        c2 = block[i2 as usize];
+    let chunk1 = &block[i1 as usize..][..12];
+    let chunk2 = &block[i2 as usize..][..12];
+
+    for (c1, c2) in chunk1.chunks_exact(4).zip(chunk2.chunks_exact(4)) {
+        let c1 = u32::from_be_bytes(c1[..4].try_into().unwrap());
+        let c2 = u32::from_be_bytes(c2[..4].try_into().unwrap());
+
         if c1 != c2 {
             return c1 > c2;
         }
-        s1 = quadrant[i1 as usize];
-        s2 = quadrant[i2 as usize];
-        if s1 != s2 {
-            return s1 > s2;
+    }
+
+    i1 += 12;
+    i2 += 12;
+
+    for _ in 0..nblock.div_ceil(8) {
+        let b1 = &block[i1 as usize..][..8];
+        let b2 = &block[i2 as usize..][..8];
+
+        let q1 = &quadrant[i1 as usize..][..8];
+        let q2 = &quadrant[i2 as usize..][..8];
+
+        if b1 != b2 || q1 != q2 {
+            for (((c1, c2), s1), s2) in b1.iter().zip(b2).zip(q1).zip(q2) {
+                if c1 != c2 {
+                    return c1 > c2;
+                }
+                if s1 != s2 {
+                    return s1 > s2;
+                }
+            }
         }
-        i1 = i1.wrapping_add(1);
-        i2 = i2.wrapping_add(1);
+
+        i1 += 8;
+        i2 += 8;
+
         if i1 >= nblock {
             i1 = i1.wrapping_sub(nblock);
         }
         if i2 >= nblock {
             i2 = i2.wrapping_sub(nblock);
         }
-        k -= 8 as c_int;
+
         *budget -= 1;
-        if k < 0 as c_int {
-            break false;
-        }
     }
+
+    false
 }
+
 static INCS: [i32; 14] = [
     1 as c_int,
     4 as c_int,