diff --git a/lang/src/son.rs b/lang/src/son.rs
index 627361b5..6697eb30 100644
--- a/lang/src/son.rs
+++ b/lang/src/son.rs
@@ -314,6 +314,7 @@ impl Nodes {
         );
 
         for &block in &*cfg_nodes {
+            debug_assert!(block != VOID);
             seen.clear(self.values.len());
             let mut outputs = mem::take(&mut self[block].outputs);
             self.reschedule_block(block, &mut outputs, buf, seen);
@@ -719,7 +720,6 @@ impl Nodes {
             self.bind(from, to);
         }
 
-        bind_buf.clear();
         self[VOID].outputs =
             self[VOID].outputs.iter().filter(|&&n| self[n].kind.is_at_start()).copied().collect();
 
@@ -736,8 +736,8 @@ impl Nodes {
             self[shed].outputs.push(n);
         }
 
+        bind_buf.clear();
         scratch.clear();
-
         visited.clear(self.values.len());
         self.schedule_inside_blocks(bind_buf, scratch, visited);
     }
diff --git a/lang/src/son/hbvm/regalloc.rs b/lang/src/son/hbvm/regalloc.rs
index 1aa395f7..2fe62f06 100644
--- a/lang/src/son/hbvm/regalloc.rs
+++ b/lang/src/son/hbvm/regalloc.rs
@@ -1,9 +1,12 @@
 use {
     crate::{
-        parser,
+        parser, quad_sort,
         son::{
             debug_assert_matches,
-            hbvm::{reg, reg::Reg, HbvmBackend, Nid, Nodes, PLoc},
+            hbvm::{
+                reg::{self, Reg},
+                HbvmBackend, Nid, Nodes, PLoc,
+            },
             Kind, ARG_START, MEM, VOID,
         },
         ty::{self, Arg, Loc},
@@ -11,7 +14,7 @@ use {
         Sig, Types,
     },
     alloc::{borrow::ToOwned, vec::Vec},
-    core::{mem, ops::Range},
+    core::{mem, ops::Range, u8, usize},
     hbbytecode::{self as instrs},
 };
 
@@ -167,32 +170,65 @@ impl HbvmBackend {
                             moves.len()
                         });
 
-                        moves.sort_unstable_by(|[aa, ab, _], [ba, bb, _]| {
-                            if aa == bb && ab == ba {
-                                core::cmp::Ordering::Equal
-                            } else if aa == bb {
-                                core::cmp::Ordering::Greater
-                            } else {
-                                core::cmp::Ordering::Less
-                            }
-                        });
+                        let mut graph = [u8::MAX; 256];
+                        for &[d, s, _] in moves.iter() {
+                            graph[d as usize] = s;
+                        }
 
-                        moves.dedup_by(|[aa, ab, _], [ba, bb, kind]| {
-                            if aa == bb && ab == ba {
-                                *kind = 1;
-                                true
-                            } else {
-                                false
-                            }
-                        });
+                        // [3, 2, 0]
 
-                        for [dst, src, kind] in moves {
-                            if kind == 0 {
-                                self.emit(instrs::cp(dst, src));
-                            } else {
-                                self.emit(instrs::swa(dst, src));
+                        // cp 0, 1 = 3
+                        // cp 1, 2 = 2
+                        // cp 2, 0 = 1
+                        //
+                        // swp a, b
+                        // swp b, c
+
+                        // cp a, b
+                        // cp b, c
+                        // cp c, d
+                        // cp d, a
+                        //
+                        // swp a, b
+                        // swp b, c
+                        // swp c, d
+
+                        'o: for &mut [d, s, ref mut depth] in moves.iter_mut() {
+                            let mut c = s;
+                            loop {
+                                if c == d {
+                                    break;
+                                }
+
+                                c = graph[c as usize];
+                                *depth += 1;
+
+                                if c == u8::MAX {
+                                    continue 'o;
+                                }
+                            }
+
+                            // cut the cycle
+                            graph[c as usize] = u8::MAX;
+                            // mark cycyle
+                            *depth = u8::MAX;
+                        }
+
+                        quad_sort(&mut moves, |a, b| a[2].cmp(&b[2]).reverse());
+
+                        for [mut d, mut s, depth] in moves {
+                            if depth == u8::MAX {
+                                while graph[s as usize] != u8::MAX {
+                                    self.emit(instrs::swa(d, s));
+                                    d = s;
+                                    mem::swap(&mut graph[s as usize], &mut s);
+                                }
+                                graph[s as usize] = s;
+                            } else if graph[s as usize] != s {
+                                self.emit(instrs::cp(d, s));
                             }
                         }
+
                         is_next_block = res.backrefs[nid as usize] as usize == i + 1;
                     }
                     Kind::Return { .. } => {
diff --git a/lang/tests/son_tests_fb_driver.txt b/lang/tests/son_tests_fb_driver.txt
index 76ff70fb..c26958b8 100644
--- a/lang/tests/son_tests_fb_driver.txt
+++ b/lang/tests/son_tests_fb_driver.txt
@@ -14,9 +14,9 @@ main:
     CP r35, r0
     LI64 r36, 30d
     LI64 r37, 100d
-    CP r34, r35
-    CP r33, r35
     CP r32, r35
+    CP r33, r35
+    CP r34, r35
  5: JLTU r34, r36, :0
     ADDI64 r32, r32, 1d
     CP r2, r35
diff --git a/lang/tests/son_tests_generic_types.txt b/lang/tests/son_tests_generic_types.txt
index 0801485f..6ff4f3b9 100644
--- a/lang/tests/son_tests_generic_types.txt
+++ b/lang/tests/son_tests_generic_types.txt
@@ -172,8 +172,8 @@ push:
     ADDI64 r33, r32, 1d
     LD r32, r32, 0a, 1h
     ST r32, r34, 0a, 1h
-    CP r34, r36
     CP r32, r33
+    CP r34, r36
     JMP :7
  0: LD r32, r38, 0a, 8h
     LD r33, r38, 8a, 8h
@@ -228,8 +228,8 @@ push:
     ADDI64 r33, r32, 8d
     LD r32, r32, 0a, 8h
     ST r32, r34, 0a, 8h
-    CP r34, r36
     CP r32, r33
+    CP r34, r36
     JMP :7
  0: LD r32, r38, 8a, 8h
     MULI64 r33, r32, 8d
diff --git a/lang/tests/son_tests_loops.txt b/lang/tests/son_tests_loops.txt
index de4541c5..eaf226dd 100644
--- a/lang/tests/son_tests_loops.txt
+++ b/lang/tests/son_tests_loops.txt
@@ -2,8 +2,8 @@ fib:
     CP r13, r2
     LI64 r17, 1d
     CP r15, r0
-    CP r16, r17
     CP r14, r15
+    CP r16, r17
  2: JNE r13, r15, :0
     CP r1, r14
     JMP :1
diff --git a/lang/tests/son_tests_struct_patterns.txt b/lang/tests/son_tests_struct_patterns.txt
index 55d641d3..6e8ec9cd 100644
--- a/lang/tests/son_tests_struct_patterns.txt
+++ b/lang/tests/son_tests_struct_patterns.txt
@@ -23,8 +23,8 @@ fib_iter:
     CP r13, r2
     LI64 r17, 1d
     CP r15, r0
-    CP r16, r17
     CP r14, r15
+    CP r16, r17
  2: JNE r13, r15, :0
     CP r1, r14
     JMP :1